You are on page 1of 131

Closed-Loop Control of Electrically Stimulated Skeletal Muscle

Contractions

by

Cheryl L. Lynch

A thesis submitted in conformity with the requirements


for the degree of Doctor of Philosophy
Institute of Biomaterials and Biomedical Engineering
University of Toronto

Copyright
c 2011 by Cheryl L. Lynch
Abstract

Closed-Loop Control of Electrically Stimulated Skeletal Muscle Contractions

Cheryl L. Lynch

Doctor of Philosophy

Institute of Biomaterials and Biomedical Engineering

University of Toronto

2011

More than one million people are living with spinal cord injury (SCI) in North America alone.

Restoring lost motor function can alleviate SCI-related health problems, as well as markedly

increase the quality of life enjoyed by individuals with SCI. Functional electrical stimulation

(FES) can replace motor function in individuals with SCI by using short electrical pulses to gen-

erate contractions in paralyzed muscles. A wide range of FES applications have been proposed,

but few application are actually available for community use by SCI consumers. A major factor

contributing to this shortage of real-world FES applications is the lack of a feasible closed-loop

control algorithm. The purpose of this thesis is to develop a closed-loop control algorithm that

is suitable for use in practical FES applications.

This thesis consists of three separate studies. The first study examined existing closed-loop

control algorithms for FES applications, and showed that a method of testing FES control

algorithms under realistic conditions is needed to evaluate their likely real-world performance.

The second study provided such a testing method by developing a non-idealities block that can

be used to modify the nominal response of electrically stimulated muscle in simulations of FES

applications. Fatigue, muscle spasm, and tremor non-idealities are included in the block, which

allows the user to specify the severity level for each type of non-ideal behaviour. This non-

idealities block was tested in a simulation of electrically induced knee extension against gravity,

and showed that the nominal performance of the controllers was substantially better than

their performance in the realistic case that included the non-idealities model. The third study

concerned the development and testing of a novel observer-based sliding mode control (SMC)

algorithm that is suitable for use in real-world FES applications. This algorithm incorporated

ii
a fatigue minimization objective as well as co-contraction of the antagonist muscle group to

cause the joint stiffness to track a desired value. The SMC algorithm was tested in a simulation

of FES-based quiet standing, and the non-idealities block was used to determine the probable

performance of the controller in the real world. This novel controller performed very well in

simulation, and would be suitable for use in selected practical FES applications.

The work contained in this thesis can easily be extended to a wide range of FES applications.

This work represents a significant step forward in closed-loop control for FES applications, and

will facilitate the development of sophisticated new electrical stimulation systems for use by

consumers in their homes and communities.

iii
Dedication

This thesis is dedicated to my grandfather, Charlie Card, who first instilled in me a love of

science and engineering.

Acknowledgements

I would like to thank my supervisor, Dr. Milos Popovic, for his invaluable guidance and constant

support throughout the ups and downs of my degree. I would also like to thank the members of

the Rehabilitation Engineering Laboratory for their camaraderie and willingness to collaborate,

in particular Dr. Kei Masani, Dr. Dimitry Sayenko, and Zina Bezruk.

Thank you to my parents for believing that microscopes and math games were appropriate

toys for little girls, and for showing me that learning is a lifelong endeavour. Thank you for

playing such a large part in the life of my own daughters, which gave me the time I needed to

complete my degree. It has been a great comfort to know that they are in such good hands.

To my husband, thank you for your unconditional love, support, and patience, and for always

seeing the best in me even when I’m not able to myself.

Lastly, thank you to all those individuals with spinal cord injuries who participated in my

experiments. This thesis would not have been possible without your generous help. I hope that

this work will quickly become obsolete due to a cure for spinal cord injury.

iv
Contents

1 Introduction 1

1.1 Motivation for Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.3 Approach Taken to Address Objective . . . . . . . . . . . . . . . . . . . . . . . . 5

1.3.1 Analysis of Existing Closed-Loop FES Control Algorithms . . . . . . . . . 5

1.3.2 Construction of Non-Idealities Model . . . . . . . . . . . . . . . . . . . . . 5

1.3.3 Development of Observer-Based Sliding Mode Control Algorithm . . . . . 6

1.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.5 Organization of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Literature Review 9

2.1 Muscle Physiology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.2 Spinal Cord Injury . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.3 Functional Electrical Stimulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.4 Closed-Loop Control Algorithms for FES Applications . . . . . . . . . . . . . . . 19

2.5 Models of Seated Knee Extension Against Gravity . . . . . . . . . . . . . . . . . 22

3 Closed-Loop Control Algorithms Used in Thesis 25

3.1 Proportional-Integral-Derivative Control . . . . . . . . . . . . . . . . . . . . . . . 25

3.2 Gain Scheduling Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.3 Sliding Mode Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.3.1 Boundary Layer Sliding Mode Control . . . . . . . . . . . . . . . . . . . . 29

v
3.3.2 Observer-Based Sliding Mode Control . . . . . . . . . . . . . . . . . . . . 30

4 Analysis of Existing Closed-Loop FES Control Algorithms 31

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

4.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4.3.1 Implementation of Base Knee Model . . . . . . . . . . . . . . . . . . . . . 33

4.3.2 Enhanced Knee Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

4.3.3 Design and Implementation of the Control Algorithms . . . . . . . . . . . 37

PID Control Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

Gain Scheduling Control Algorithm . . . . . . . . . . . . . . . . . . . . . 37

Sliding Mode Control Algorithm . . . . . . . . . . . . . . . . . . . . . . . 39

4.3.4 Controller Testing Methodology . . . . . . . . . . . . . . . . . . . . . . . . 41

4.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

5 Construction of Non-Idealities Model 51

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

5.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

5.2.1 Construction of Non-Idealities Block . . . . . . . . . . . . . . . . . . . . . 53

Fatigue Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

Spasm and Tremor Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

5.2.2 Implementation Example for Non-Idealities Block . . . . . . . . . . . . . . 56

5.2.3 Control Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

5.2.4 Assessment of Controller Performance . . . . . . . . . . . . . . . . . . . . 60

5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

5.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

5.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

vi
6 Observer-Based Sliding Mode Control of Quiet Standing 71

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

6.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

6.2.1 Inverted Pendulum Model of Quiet Standing in SCI Individuals . . . . . . 72

6.2.2 FES Applications of Sliding Mode Control . . . . . . . . . . . . . . . . . . 74

6.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

6.3.1 Overview of Quiet Standing Simulation . . . . . . . . . . . . . . . . . . . 76

6.3.2 Implementation of Quiet Standing Simulation . . . . . . . . . . . . . . . . 76

6.3.3 Control Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

6.3.4 Assessment of Controller Performance . . . . . . . . . . . . . . . . . . . . 82

6.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

6.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

6.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

7 Discussion 94

7.1 Summary of Analysis of Existing Closed-Loop Control Algorithms . . . . . . . . 94

7.2 Summary of Development of Non-Idealities Block . . . . . . . . . . . . . . . . . . 95

7.3 Summary of Observer-Based Sliding Mode Control Algorithm . . . . . . . . . . . 96

7.4 Control Algorithms for FES Applications . . . . . . . . . . . . . . . . . . . . . . 97

7.5 Limitations of Thesis Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

7.6 Contributions Made by Thesis Work . . . . . . . . . . . . . . . . . . . . . . . . . 99

8 Conclusions 101

Bibliography 104

A Non-Ideality Classification Criteria 116

B List of Abbreviations 118

Index 120

vii
List of Tables

2.1 Closed-loop FES control algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.2 Models of knee extension against gravity . . . . . . . . . . . . . . . . . . . . . . . 23

4.1 Response metrics for subject P1 with mild non-idealities . . . . . . . . . . . . . . 42

5.1 Intra- and inter-rater reliability for rating of non-ideality waveforms . . . . . . . 60

5.2 Performance metrics for PID control of knee angle . . . . . . . . . . . . . . . . . 61

5.3 Performance metrics for sliding mode control of knee angle . . . . . . . . . . . . 62

6.1 Parameters of quiet standing system. . . . . . . . . . . . . . . . . . . . . . . . . . 77

6.2 Controller and observer parameters. . . . . . . . . . . . . . . . . . . . . . . . . . 81

6.3 PID and SMC performance metrics . . . . . . . . . . . . . . . . . . . . . . . . . . 83

6.4 SMC performance metrics with fatigue non-idealities . . . . . . . . . . . . . . . . 84

6.5 SMC performance metrics with spasm non-idealities . . . . . . . . . . . . . . . . 84

6.6 SMC performance metrics with tremor non-idealities . . . . . . . . . . . . . . . . 84

6.7 SMC performance metrics with all non-idealities . . . . . . . . . . . . . . . . . . 85

A.1 Classification criteria for non-ideality waveforms . . . . . . . . . . . . . . . . . . 117

viii
List of Figures

1.1 An FES system in use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2.1 Anatomy of skeletal muscle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.2 Production of tetanic contraction . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.3 Waveforms of equivalent stimulation patterns. . . . . . . . . . . . . . . . . . . . . 16

2.4 Typical fatigue waveforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.1 Anti-windup PID controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4.1 Diagram of knee extension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4.2 Step trajectory tracking for subject P1 . . . . . . . . . . . . . . . . . . . . . . . . 43

4.3 Walking trajectory tracking for subject P2 with mild non-idealities . . . . . . . . 44

4.4 Walking trajectory tracking for subject P2 with severe non-idealities . . . . . . . 45

4.5 Sensitivity analysis for subject P1 with ±20% parameter mismatch . . . . . . . . 46

4.6 Sensitivity analysis for subject P1 with ±50% parameter mismatch . . . . . . . . 47

5.1 Examples of non-ideality waveforms . . . . . . . . . . . . . . . . . . . . . . . . . 56

5.2 Diagram of non-idealities block implemented in FES simulation . . . . . . . . . . 58

5.3 Evolution of SMC error trajectory in R2 -space . . . . . . . . . . . . . . . . . . . 63

5.4 Step response of PID and sliding mode controllers . . . . . . . . . . . . . . . . . 64

5.5 Tracking performance of SMC for walking trajectory . . . . . . . . . . . . . . . . 65

5.6 RMS error of SMC at corners of non-ideality space . . . . . . . . . . . . . . . . . 66

5.7 Two-dimensional plot of RMS error at corners of non-ideality space . . . . . . . . 67

ix
6.1 Inverted pendulum model of quiet standing . . . . . . . . . . . . . . . . . . . . . 74

6.2 Block diagram of quiet standing simulation . . . . . . . . . . . . . . . . . . . . . 77

6.3 Evolution of error trajectory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

6.4 Step and sinusoidal response of PID and SMC for nominal model . . . . . . . . . 87

6.5 Step and sinusoidal response of SMC . . . . . . . . . . . . . . . . . . . . . . . . . 88

6.6 Sensitivity of SMC to model parameter mismatch error . . . . . . . . . . . . . . 89

x
Chapter 1

Introduction

1.1 Motivation for Thesis

More than one million people are living with spinal cord injury (SCI) in North America [1].

SCI can be caused by diseases that destroy the neurological tissues of the spinal cord or by

trauma that compresses, stretches, or severs this tissue. SCI is often irreversible, and can result

in partial or total loss of sensory or motor function, or both, to the parts of the body below the

level of the injury. For example, an injury to the spinal cord at the lower back usually affects

the legs, but not the arms.

SCI can cause secondary complications including decubitis ulcers (i.e. pressure ulcers),

muscle spasms, loss of muscle strength and volume, cardiovascular disease, osteoporosis, incon-

tinence, sexual dysfunction, pain, and psychiatric illnesses such as depression and adjustment

disorders. SCI can also result in a reduced level of independence in activities of daily living such

as bathing, dressing, and toilet use, as well as a reduced ability to participate in the community

and the work force.

The restoration of lost motor functions such as grasping, balance during sitting, standing,

and walking may help to alleviate some SCI-related health problems [2]. Restoring lost mo-

tor function can also markedly increase the quality of life enjoyed by individuals with SCI,

as well as their ability to be involved in the community and their employability. For these

reasons, improving motor function in individuals with SCI has been an active research topic in

1
Chapter 1. Introduction 2

rehabilitation engineering for more than 50 years.

The most commonly used technology for restoring or replacing motor function in individuals

with SCI is functional electrical stimulation (FES), which uses short electrical pulses to generate

contractions in paralyzed muscles. These contractions can be coordinated to move or stabilize

joints by stimulating one or more muscles that exert torques about the joint. The resulting joint

angle or joint torque can be controlled by modulating the intensity of stimulation delivered to

the flexor and extensor muscles, which actuate the joint in opposite directions.

FES can be used for a wide range of applications in individuals with SCI. FES-based SCI

rehabilitation modalities include muscle strengthening [3,4], gait training [5,6], and grasp reha-

bilitation programmes [7–9]. FES has also been used for neuroprostheses that replace lost motor

function in individuals with SCI, such as systems for grasping [10–12], elbow extension [13, 14],

standing [15–19], walking [20–23], cycling [24], rowing [25], and trunk stabilization [26, 27], and

sitting down from a standing position [28]. New neuroprostheses for improving orthostatic tol-

erance [29–32], relieving pressure on the ischial tuberosities of the hips during sitting [33–35],

and respiratory pacing in an acute health care setting [36, 37] have also been proposed. Figure

1.1 shows an FES grasping system in use.

Despite the wide range of FES applications that have been reported in the literature, few

FES systems are available for wide-spread use in the community. Commercial FES cycling

systems that are available for individuals with SCI include the RT300 (Restorative Therapies,

Inc.) and the Ergys 2 (Therapeutic Alliances, Inc.). Commercial FES grasping systems include

the NESS H200 (Bioness, Inc.). Commercial FES systems for correcting foot drop include

the ODFS Pace (Odstock Medical, Ltd.), the WalkAide foot drop stimulator (Innovative Neu-

rotronics, Inc.), and the L300 (Bioness, Inc.) Many commercial FES systems for providing

bladder control to individuals with SCI are also available.

The real-world use of FES neuroprostheses presents many implementation challenges. The

system must be able to run in real-time, which may limit the complexity of the control algorithm

that can be used. An algorithm that requires a lengthy daily tuning process may not be accepted

by potential neuroprosthesis users. The FES system must also be designed to fail safely, so that

the user is not in danger due to unexpected system behaviour.


Chapter 1. Introduction 3

Figure 1.1: An FES system in use. An individual with a quadriplegic spinal cord injury practices

grasping and releasing objects using an FES system. Without the FES system, this individual

cannot grasp or release objects. However, by pressing a button located on the wheelchair arm

rest with his left hand, he can trigger the stimulator (not shown) to electrically stimulate the

forearm muscles. The stimulation causes the muscles to contract in a grasping pattern, allowing

him to pick up the toothbrush. The neuroprosthesis is incorporated into the brace that the

individual wears on his right arm.


Chapter 1. Introduction 4

Sensors such as electrogoniometers and force transducers are used by closed-loop FES control

systems to gain feedback information on the behaviour of the controlled system. This feedback

data is used by the control system to modulate the control action, and allows the control system

to compensate for changes in the stimulated muscle response and exogenous disturbances.

However, external sensors often provide poor cosmesis, and may also be difficult for users to

don and doff without the assistance of an attendant. In the future, it may be possible to use

implanted sensors to transduce physiological feedback signals from the body, thereby avoiding

the problems associated with external sensors.

Neuroprosthesis systems that use surface electrodes also present a cosmetic problem for some

FES users, since some people prefer not to use obviously visible assistive devices. Moreover, the

repeatability of the electrode placement may be low, and the user may require the assistance

of an attendant for donning and doffing the system. Implanted neuroprosthesis systems avoid

some of these problems of cosmesis and use, but are invasive.

The shortage of practical FES applications is frustrating to SCI consumers and their advo-

cates. One of the major factors contributing to this situation is the lack of feasible closed-loop

FES control algorithms. Closed-loop control is of vital importance to the development of many

types of practical FES applications, because this technique allows the controller to adjust the

amount of stimulation delivered to the muscles in real-time to compensate for disturbances

and changes in the response dynamics of the stimulated muscles, based on feedback data from

sensors. The aim of the work described in this thesis is to address this issue by developing a

closed-loop FES control algorithm that is suitable for use in real-world FES applications by

individuals who have SCI.

1.2 Objective

The objective of the work described in this thesis is to develop a closed-loop control algorithm

that is suitable for implementation in a practical FES system, and is capable of regulating elec-

trically stimulated muscle contractions despite the complex behaviour exhibited by stimulated

muscle in individuals who have SCI.


Chapter 1. Introduction 5

1.3 Approach Taken to Address Objective

The approach that was taken to develop a suitable closed-loop control algorithm for real-world

FES applications consisted of three studies, each of which is represented by a thesis chapter.

1.3.1 Analysis of Existing Closed-Loop FES Control Algorithms

In the first study, a systematic analysis of the existing closed-loop control algorithms that have

been proposed for FES applications was conducted. The performance of several representative

control algorithms was evaluated in a simulation of FES-based knee extension against gravity.

The purpose of this analysis was to determine the technical and algorithm design factors con-

tributing to the shortage of practical FES applications that use closed-loop control. Particular

attention was paid to the ways in which the design and testing of these existing controllers

could be improved to increase their suitability for real-world FES applications.

The hypothesis for this study was that the discrepancy between the favourable performance

reported for most controllers in the literature and the paucity of practical applications of these

controllers could be traced to unrealistic testing conditions that were used to determine the

performance of the control algorithms. Moreover, it was anticipated that this study would

suggest that controlling electrically stimulated muscle contractions in the closed-loop would

require a nonlinear control algorithm due to the complex, nonlinear nature of the contractile

response of stimulated muscle.

1.3.2 Construction of Non-Idealities Model

In the second study, a new method of testing FES control algorithms in simulation was devel-

oped that allowed a realistic examination of the probable real-world performance of closed-loop

control algorithms for FES applications. A “non-idealities” block for use in FES simulations

was designed and implemented in Simulink (The Mathworks, USA). This block permitted ac-

tual experimentally observed muscle fatigue, spasm, and tremor behaviour to be incorporated

into the nominal simulated response of the electrically stimulated muscles. The experimental

data that were used to develop this block were recorded from individuals who have complete
Chapter 1. Introduction 6

SCI (AIS A) during seated electrically stimulated knee extension against gravity and isometric

quadriceps contraction experiments.

The non-idealities block implements fatigue as a modifier between 0 and 1 that scales the

nominal stimulated muscle response, where 1 corresponds to no fatigue, and 0 corresponds

to complete fatigue. The muscle spasm and tremor data is normalized with respect to the

amplitude of the knee movement from which the data was extracted, so that these non-idealities

can be generalized to movements of different amplitudes than the original data. The particular

waveform for each non-ideality type that is used in a simulation run is chosen randomly at

run-time by the non-idealities block from the pool of available non-ideality waveforms. The

non-ideality block also allows the user to specify the desired severity level for the fatigue,

muscle spasm, and tremor non-idealities.

This non-idealities block allowed the performance of closed-loop FES control systems to be

analyzed under conditions that accurately reflect the real-world non-ideal behaviour that can

be expected of electrically stimulated muscle in individuals with SCI. The block also facilitated

the assessment of control performance for the nominal case as well as for the “worst” cases that

can be expected from stimulated muscle in individuals with SCI, such as severe muscle spasms

or very rapid fatigue.

The hypothesis for this study was that including the non-idealities block in a FES simulation

would result in significantly degraded control performance as compared to the performance with

the nominal model only.

1.3.3 Development of Observer-Based Sliding Mode Control Algorithm

In the third study, a novel nonlinear closed-loop control algorithm was developed that was

suitable for use in real-world FES applications. This algorithm was an observer-based sliding

mode controller (SMC) that was designed to be robust to the disturbances and model parameter

changes that are inherent in controlling electrically stimulated muscle contractions. This design

also reflected the insights that were obtained from the analysis of existing FES control strate-

gies in study one. The observer-based SMC was tested in a simulation of FES-induced quiet

standing in an individual with complete SCI, and used the non-idealities block to determine
Chapter 1. Introduction 7

the probable real-world performance of this controller in the presence of non-ideal stimulated

muscle responses.

The hypothesis for this study was that the observer-based SMC would control the quiet

standing application more accurately than a standard FES control algorithm such as proportional-

integral-derivative (PID) control. Moreover, it was expected that this algorithm would be ca-

pable of regulating a wider range of non-ideal stimulated muscle behaviour than a standard

PID controller while still guaranteeing the stability of the controlled system.

1.4 Applications

The novel observer-based SMC algorithm presented in this thesis could be readily applied to

other FES systems. Moreover, this technology was developed using data from complete SCI

subjects, but could be applied equally well to other populations with a neurological disability

such as incomplete SCI or stroke. The availability of reliable closed-loop algorithms for FES

applications, such as the algorithm presented in this thesis, will facilitate the development of

practical FES applications such as systems for walking, dynamic grasping, and balance while

standing, sitting, or wheeling that can be used by neurologically impaired individuals in their

everyday lives.

This work represents a significant step forward in the state of the art for FES control

algorithms. In fact, the portion of this work that was presented at a prestigious international

conference was very well received by the FES research community, and won a best-paper award

for its contribution to the field [38]. The results contained in the papers that comprise this

thesis can be immediately applied by the FES community to other FES systems, potentially

improving the performance of many different types of FES applications. Moreover, this work

shows that nonlinear control methods can be feasible and simple to use for FES systems. It is

the author’s hope that this work will allow the end users of FES systems to benefit from the

sophisticated applications that can be developed using modern control technologies, instead of

being forced to settle for unsuitable and outdated control techniques that are not adequate for

regulating many real-world FES applications.


Chapter 1. Introduction 8

1.5 Organization of Thesis

This thesis is based on a series of three individual studies. Each study is presented in a separate

chapter that has its own introduction, methods, results, and discussion sections. The author’s

publications that are related to each chapter are cited after the description of the chapter below.

Chapter 2 provides background material on muscle physiology, SCI, FES and its applica-

tions, and models of electrically stimulated muscle behaviour. This chapter also reviews the

closed-loop control algorithms that have been used for FES applications. [39]

Chapter 3 presents the theory behind the closed-loop control algorithms that are used in

this thesis.

Chapter 4 contains an analysis of existing closed-loop control algorithms for FES appli-

cations, and presents the performance of several representative controllers while regulating a

simulation of knee extension against gravity. This chapter also discusses some of the reasons

behind the relative scarcity of practical controllers for FES applications for use in the commu-

nity. [40, 41]

Chapter 5 presents the design of the non-idealities block that was developed to provide

a way of testing FES controllers in the presence of real-world non-ideal behaviour, and also

demonstrates the implementation of this block in a simulation of electrically stimulated knee

extension against gravity. [38, 42–46]

Chapter 6 discusses the design and implementation of a novel observer-based SMC algorithm

for use in real-world FES applications. This chapter also compares the performance of this novel

controller and a standard PID controller for regulating electrically induced quiet standing in

the presence of non-ideal behaviour from the stimulated muscles [47, 48]

Chapter 7 discusses the results of these studies taken together, and Chapter 8 presents the

overall conclusions of this thesis.


Chapter 2

Literature Review

2.1 Muscle Physiology

The principal functions of skeletal muscle are to maintain body posture and produce movement.

Muscle cells are long, thin cells called muscle fibres. The muscle fibres are grouped to form

muscle fascicles. The fascicles are interspersed with blood vessels and nerves, and are bundled

together to form skeletal muscle. Collagen fibres hold the muscle fascicles together, and unite at

the ends of the muscle to form tendons. Tendons attach to the bones, providing a link between

the muscles and bones by which muscles can exert forces on the bones and torques about the

joints of the body. Figure 2.1 shows the anatomy of skeletal muscle.

The muscle cell membrane, or sarcolemma, both surrounds each muscle cell and extends into

the interior of the cell in the form of transverse tubules. This structure ensures that the diffusion

of products into the muscle cell can happen rapidly and nearly simultaneously throughout the

cell. When a contraction stimulus (i.e., action potential) arrives at a muscle cell, calcium ions

(Ca2+ ) are released into the muscle cell, triggering the muscle fibre to contract via a series of

chemical reactions. The cell contracts uniformly along its length because the Ca2+ ions are able

to reach all parts of the cell at nearly the same time, due to the network of transverse tubules

infiltrating the cells.

The main classes of muscle fibres are fast-twitch and slow-twitch fibres. Fast-twitch fibres

respond quickly to a contractile impulse but also fatigue quickly. Slow-twitch fibres are more

9
Chapter 2. Literature Review 10

Figure 2.1: Anatomy of skeletal muscle [49].


Chapter 2. Literature Review 11

fatigue resistant than fast-twitch fibres, but respond more slowly to a contractile impulse than

fast-twitch fibres. The ratio between fast-twitch and slow-twitch fibres in a muscle depends on

the function of the muscle. Also, the composition of a muscle can change over time depending

on the type of contractions to which it is subjected [50]. For example, sprinters develop a higher

proportion of fast-twitch fibres in their leg muscles than distance runners. This difference in

muscle composition occurs because sprinters need muscles that respond quickly, but don’t need

the muscles to work for an extended period of time, whereas distance runners need muscles

that are fatigue resistant.

Skeletal muscle activity is controlled by the nervous system in neurologically intact individ-

uals. Voluntary skeletal muscle contractions are initiated by electrochemical signals from the

brain, which propagate through the nervous system to a motor neuron. Each motor neuron,

together with the muscle fibres innervated by the motor neuron, form a motor unit. Muscles

that are involved in gross movements, such as leg muscles, have few motor units, each of which

includes a large number of muscle fibres. Muscles that are involved in fine movements, such as

facial muscles, have many motor units, each of which includes a small number of muscle fibres.

When an action potential arrives at the junction between the motor neuron and the muscle

fibre, molecules of the neurotransmitter acetylcholine are released into the junction gap. The

acetylcholine binds to receptors on the muscle fibre, causing Ca2+ ions to be released into the

muscle cell by a series of chemical reactions. There is a delay of approximately 2 ms between

the action potential arriving at the motor junction and the onset of the muscle fibre contraction,

due to the chain reaction of chemical events that must take place [51]. The acetylcholine in

the junction gap is broken down very quickly by the enzyme acetylcholinesterase, meaning that

each nerve impulse results in a very short contraction of the innervated muscle fibres.

The energy required to contract muscles is stored in the muscle cells in the form of adenosine

triphosphate (ATP) and creatine phosphate. These products are rapidly exhausted when the

muscle fibre contracts, resulting in muscle fatigue. Eventually, the muscle cell runs out of

energy, and must rebuild its stores of ATP or creatine phosphate, or both, in order to contract

again.

Tension is produced in skeletal muscle by contracting muscle fibres. However, many fibres
Chapter 2. Literature Review 12

must work together in order to produce enough tension to do useful work. The body achieves

gradations of muscular effort by recruiting varying numbers of motor units. For example, a task

requiring maximum muscular effort uses almost all of the motor units in the muscle, resulting

in a large portion of the muscle fibres being activated.

A single impulse in a motor neuron results in a fast, transient contraction of a single motor

unit. Therefore, a motor neuron must deliver a train of impulses to its associated muscle fibres

to maintain a constant contraction in the portion of the muscle innervated by the motor unit.

Typically, voluntary muscle contractions involve sustained, constant tension. The body achieves

this constant tension, known as a tetanic contraction, by activating adjacent motor units at a

frequency of 6 to 8 Hz in a sequential manner, so that one motor unit delivers a contractile

impulse to its muscle fibres before the adjacent motor unit relaxes from the previous contractile

impulse [52], as shown in Figure 2.2. This method of sustaining a muscle contraction, known

as asynchronous recruitment, allows the various motor units to share the work of maintaining

a muscle contraction. Asynchronous recruitment ensures that the muscle fatigues slowly, since

each motor unit is active only part of the time. Moreover, each motor unit has an opportunity

to re-build its energy stores while it is resting, slowing the rate of fatigue exhibited by the

muscle as a whole.

2.2 Spinal Cord Injury

SCI is caused by diseases that destroy the neurological tissue of the spinal cord, or by trauma

that compresses, stretches, or severs this tissue. Traditionally, a large proportion of those

affected by SCI have been young males who sustained a traumatic SCI while engaged in risk-

taking behaviour. In recent years, the incidence of SCI in older adults has increased to the

point where this population forms a large proportion of those who sustain a SCI. These injuries

are frequently due to non-traumatic causes such as tumors or surgical side-effects, but can also

be due to trauma such as falls and motor vehicle accidents.

SCI is often irreversible, and usually results in partial or total loss of sensory or motor

function (i.e., paralysis), or both, to the parts of the body below the level of the injury. In
Chapter 2. Literature Review 13

Total Tension
in Muscle
Tension
MU #3
Tension Tension
MU #1 MU #2

Time (ms)

Figure 2.2: Production of tetanic contraction (adapted from [53]). The production of tension in

skeletal muscle is accomplished by sequentially stimulating adjacent motor units, abbreviated

“MU” in the figure. The stimulation is timed by the intact neurological system so that each

motor unit contracts before the previously stimulated motor unit relaxes completely. The

tension in the overall muscle is the sum of the tensions in the individual motor units.
Chapter 2. Literature Review 14

many cases, SCI also disrupts the autonomic nervous system, which regulates visceral functions

such as blood pressure, heart rate, body temperature, and digestive processes.

A SCI that results in total loss of sensory and motor function to the affected parts of the

body is referred to as a complete SCI, and is classified as an AIS A injury, where AIS is an

acronym for the American Spinal Injury Association (ASIA) Impairment Scale [54]. Other types

of SCI are referred to as incomplete SCI, and are classified as AIS B through D, depending on

the effects of the injury.

As mentioned in Section 1.1, SCI can also cause secondary complications such as decubitis

ulcers (i.e., pressure ulcers), muscle spasms, loss of muscle strength and volume, cardiovascular

disease, osteoporosis, sexual dysfunction, incontinence, musculoskeletal pain, neuropathic pain,

and psychiatric conditions such as depression and adjustment disorders. These SCI-related

health problems can sometimes be improved by restoring motor functions such as grasping,

standing, and walking. Restoring lost motor function can also markedly increase the quality of

life enjoyed by individuals with SCI [2].

SCI is also associated with several phenomena that affect control system design for FES

applications, including spasticity, spinal reflexes, muscle tremors, and disuse atrophy of the

affected muscles.

Spasticity is a common phenomenon in individuals with SCI that is characterized by varying

degrees of increased muscle tone [55]. In the absence of supra-spinal signals, muscles can develop

a tendency to maximally contract in response to a wide range of muscular or cutaneous stimuli.

A paralyzed muscle does not exhibit voluntary contractions because it receives no signals from

the brain. However, the muscle can still contract in response to other stimuli, such as electrical

stimulation or pain. Spasticity can be treated to some extent by physiotherapy or drug therapy,

but is often conservatively treated [56], and may be present in some FES users.

Spinal reflexes are feedback loops in the central nervous system (CNS) [52]. Some spinal

reflexes involve only a single synapse between two motor neurons, whereas others are complex,

multi-synaptic reflexes. An example of a simple reflex is a stretch reflex, where a rapid stretch

perturbation applied to a muscle results in a rapid and immediate contraction of the muscle.

An example of a complex reflex is the flexion withdrawal reflex, which causes a limb to rapidly
Chapter 2. Literature Review 15

recoil toward the body in response to a pain stimulus in an extremity. Complex CNS feedback

loops can involve neural circuits and oscillators formed by connections between neurons in the

spinal cord, and sometimes involve the brain stem and other phylogenetically primitive parts

of the brain.

Spinal reflexes that originate in the spinal cord below the level of injury are often present in

individuals who have SCI, although these reflexes may be significantly altered (i.e., heightened

or depressed) as a result of the injury [57]. These reflexes may be activated during FES-

induced muscle contractions, causing exogenous contractile signals to be sent to the paralyzed

muscles in parallel with the FES control signal. For example, walking is controlled in part by a

neuronal circuit located in the spinal cord called a central pattern generator (CPG) [52]. When

an individual with SCI uses FES for walking, a CPG might generate a walking pattern that

conflicts with the pattern of muscle contractions induced by the FES walking system [58].

Muscle tremors are typically due to incomplete tetanus of the stimulated muscle contrac-

tion. Tremors can occur when the stimulation intensity is slightly too low to elicit a tetanic

contraction, but also commonly occur when the stimulated muscle becomes so fatigued that it

can no longer produce the tension required to create a tetanic contraction, even at the maximum

stimulation intensity.

Muscle that does not receive regular exercise undergoes disuse atrophy and converts to a

higher proportion of fast-twitch fibres than is present in active muscle [50]. Many people with

SCI have extensive disuse atrophy in their affected muscles [55]. Consequently, the affected

muscles are weak and fatigue quickly, and have become mostly composed of fast-twitch fibres.

However, disuse atrophy is often a fully or partially reversible process since the affected muscles

can be re-trained with electrically stimulated weight-bearing exercise to increase their strength

and fatigue resistance [55]. This re-training process can also occur as a beneficial side effect of

regular FES use.


Chapter 2. Literature Review 16

2.3 Functional Electrical Stimulation

The most commonly used technology for replacing lost motor function in individuals with

SCI is functional electrical stimulation (FES), which involves artificially inducing a current in a

specific motor nerve to generate a skeletal muscle contraction. This is accomplished by applying

a series of short electrical pulses to the nerve using electrodes [59]. These electrodes can be

transcutaneous (i.e., temporarily affixed to the skin surface), epimysial (i.e., surgically affixed

to the surface of the muscle), percutaneous (i.e., surgically placed within the muscle), cuff (i.e.,

surgically wrapped around the nerve that innervates the muscle of interest), or intraneural (i.e.,

surgically inserted into the nerve that innervates the muscle of interest) [55].

The tension produced in electrically stimulated muscle is a function of the stimulation

waveform, the frequency of stimulation, and the stimulation intensity. The stimulation intensity

is a function of the total charge transferred to the muscle, which depends on the amplitude and

pulse width of stimulation. In Figure 2.3, the two waveforms shown result in the same charge

transfer to the muscle even though their amplitudes and pulse widths are different. A biphasic

stimulation waveform is typically used in FES applications to prevent tissue damage; biphasic

pulses induce charge transfer into the tissue and then immediately remove the charge from the

tissue to prevent galvanic processes that may cause tissue damage [59].

amplitude amplitude

time time

Figure 2.3: Waveforms of equivalent stimulation patterns.

FES can be used to induce joint movement by stimulating the flexor and/or extensor muscles

of the joint. The angle of the joint, or, alternatively, the torque about a joint, can be regulated

by varying the tension produced in the flexor and extensor muscles of the joint. Consequently,

the joint angle or joint torque can be controlled by modulating the stimulation pulse duration,
Chapter 2. Literature Review 17

amplitude, or frequency. Typically, either the pulse duration or amplitude is regulated. Also,

the resistance of the stimulated tissue can vary day to day. Therefore, to ensure that a pre-

dictable amount of charge is transferred to the stimulated tissue despite this variable resistance,

current regulated stimulation is typically used for FES applications instead of voltage regulated

stimulation.

As discussed in Section 2.1, muscles show increasing signs of fatigue as they progressively

expend their available energy resources. If muscles are artificially stimulated to the point of

fatigue, their response changes nonlinearly as their energy stores are depleted. Eventually, the

muscle will no longer be able to produce tension. Figure 2.4 shows four fatigue patterns that

were seen in a single SCI subject during different FES sessions.

50 50

40 40
Force (N)

Force (N)

30 30

20 20

10 10

0 0
0 5 10 15 20 25 0 5 10 15 20 25
Time (s) Time (s)

50 50

40 40
Force (N)

Force (N)

30 30

20 20

10 10

0 0
0 5 10 15 20 25 0 5 10 15 20 25
Time (s) Time (s)

Figure 2.4: Force exerted by quadriceps versus time. These plots show the force exerted by the

quadriceps muscle group when subjected to electrical stimulation. The data were collected on

different days from an individual who had sustained a SCI resulting in complete paralysis of

the lower extremities [60]. Four different force decay profiles are shown, representing the wide

range of force profiles that can be seen as the muscle fatigues in response to stimulation.

FES recruits motor units in a synchronous manner, unlike the asynchronous recruitment
Chapter 2. Literature Review 18

of motor units that occurs in the intact nervous system. Synchronous recruitment means that

FES stimulates all of the motor units at the same time, instead of rotating through the motor

units as is done by the nervous system. For this reason, achieving tetanic contractions with

FES stimulation requires a much higher stimulation frequency (i.e., frequencies greater than

16 Hz, with typical values being 20 to 40 Hz) than the frequency required to achieve tetanic

contractions with the asynchronous recruitment used by the intact nervous system (i.e., 6 to

8 Hz). This higher stimulation frequency is the main cause of the increased rate of fatigue

associated with stimulated muscle contractions as compared to contractions initiated by the

central nervous system [50].

Furthermore, FES has a tendency to recruit the fast-twitch muscle fibres before the slow-

twitch muscle fibres. This order of fibre recruitment, known as non-physiological recruitment,

is the opposite of natural muscle fibre recruitment order, in which the slow-twitch fibres are

recruited first. Non-physiological recruitment happens because the fast-twitch fibres are in-

nervated by axons with a larger diameter than the diameter of the axons that innervate the

slow-twitch fibres. Since fast-twitch fibres fatigue more quickly than slow-twitch fibres, the non-

physiological order of recruitment that occurs with FES also contributes to the increased rate of

fatigue seen with artificial stimulation as compared to natural stimulation. Currently, there is

no way to avoid the synchronous motor unit stimulation or non-physiological recruitment that

occurs with FES when transcutaneous electrodes are used. Several selective nerve-block stimu-

lation techniques that aim to provide more natural motor unit stimulation are available [61,62].

However, these techniques require nerve cuff electrodes, which must be surgically implanted in

the user.

Regardless of the type of electrode used, it is possible to increase the fatigue resistance of

electrically stimulated muscle by intensive muscle training using FES [9]. This training causes

an increase in the strength and volume of the muscles, which increase the fatigue resistance of

the muscle. Regular stimulation with FES causes the composition of muscle to change; over

time, some of the fast-twitch fibres change to slow-twitch fibres. This change in composition

increases the fatigue resistance of the muscle, and introduces another time-varying factor that

affects the response of muscle to FES.


Chapter 2. Literature Review 19

2.4 Closed-Loop Control Algorithms for FES Applications

Many real-world FES applications require the ability to modulate the pulse-to-pulse electrical

stimulation in real-time to compensate for fatigue, muscle spasms and tremors, muscle re-

training effects, and day-to-day changes in the stimulated muscle response, as well as modeling

errors and exogenous disturbances. Closed-loop control can be used to address these challenges.

The most common test beds for evaluating control of FES-based muscle contractions are

unsupported quiet standing and knee extension against gravity. Control algorithms for regula-

tion of cyclic movements such as gait patterns, and regulation of wrist and elbow movements

have also been reported. The important work in each of these areas is discussed below, and is

summarized in Table 2.1.

Riess and Abbas developed an adaptive neural network controller for regulating cyclic move-

ments of the electrically stimulated quadriceps muscle [63, 64]. The authors reported good re-

sults with this approach, but noted that it required 10 minutes of data collection and a lengthy

training process each time it was used. The extended calibration time required makes this

controller unsuitable for real-world FES applications.

Lemay and Crago used PID control to regulate wrist movements using FES in quadriplegic

individuals [65]. Micera, Sabatini, and Dario studied fuzzy logic control of electrically stimulated

co-contracting elbow flexion and extension muscles [66]. Adamczyk and Crago investigated the

use of a neural network to control co-contraction of electrically stimulated hand and wrist

muscles [67].

Davoodi and Andrews developed a fuzzy logic system for controlling a FES rowing ap-

plication [25], and reported that the fuzzy logic controller resulted in less fatigue and longer

rowing duration than PID control. However, the success of this control strategy was achieved

by manually tuning the rules of the fuzzy logic controller to obtain the desired response. This

labour-intensive procedure may not be feasible for real-world FES applications.

Abbas and Chizeck used a PID controller to regulate quiet standing in SCI subjects [68].

Jaime, Matjacic, and Hunt [69] and Matjacic, Hunt, Gollee, and Sinkjaer also implemented PID

controllers for quiet standing [70], although Matjacic et al. noted that the derivative action of
Chapter 2. Literature Review 20

Table 2.1: Closed-loop FES control algorithms. NN denotes neural network, FL denotes fuzzy

logic, ∗ denotes success in laboratory testing, ∗∗ denotes success in real-world applications, and

† denotes that PID control is used, which is unsuitable for practical FES applications.
Authors Year Application Controller Advantages Disadvantages

Riess et al. 2000 Cyclic NN Lengthy calibration

Davoodi et al. 2004 Rowing FL ∗∗ Lengthy tuning

Lemay et al. 1997 Wrist PID †

Micera et al. 1999 Elbow FL Co-contraction

Adamczyk et al. 2000 Wrist NN Co-contraction

Abbas et al. 1991 Standing PID ∗ †

Matjacic et al. 1998 Standing PID ∗ Voluntary bracing, †

Jaime et al. 2002 Standing PID ∗ †

Matjacic et al. 2003 Standing PID ∗ †

Hunt et al. 2000 Standing H∞ ∗ Voluntary bracing

Holderbaum et al. 2002 Standing H∞ ∗ Voluntary bracing

Gollee et al. 2004 Standing H∞ ∗ Voluntary bracing

Hatwell et al. 1991 Knee Adaptive Unsuitable for NLTV

Chang et al. 1997 Knee Neuro-PID ∗ Daily re-training

Ferrarin et al. 2001 Knee Adaptive ∗ Periodic only

Previdi et al. 2004 Knee GSC Good design, ∗

Jezernik et al. 2004 Knee SMC Good design, ∗

Ajoudani et al. 2009 Knee NN/SMC ∗ Complex design

Kobravi et al. 2009 Knee Adaptive SMC ∗ Stability ??


Chapter 2. Literature Review 21

PID controllers can amplify high-frequency system dynamics and lead to system instability.

Hunt, Jaime, and Gollee [71] and Holderbaum, Hunt, and Gollee [72] used robust H∞

control for regulating quiet standing in able-bodied and SCI subjects, respectively. Gollee,

Hunt, and Wood compared H∞ and linear quadratic Gaussian (LQG) control for regulating

quiet standing [16]. The results of this work showed that the subject was able to maintain

standing balance and to reject a perturbation of approximately 1◦ of sway angle in the anterior

direction using a strategy of combined FES stimulation and voluntary bracing using his arms.

Matjacic and Bajd also successfully controlled FES-based quiet standing [18]. This work

showed that it is possible for an individual with SCI to use FES to stabilize his or her legs

in an extended position and provide ankle stiffness, while using voluntary control of trunk

movements to maintain standing balance. The work done by Hunt and colleagues [16, 71, 72]

and Matjacic and Bajd [18] relied on voluntary torso movements to oppose the effects of the

plantarflexor muscles, however it should be possible to achieve the same result by co-contracting

the plantarflexor and dorsiflexor muscles.

Hatwell, Oderkerk, Sacher, and Inbar developed a model reference adaptive controller for

regulating knee movement in SCI subjects [73]. The controller tracked angles at the ends of

the range of motion quite well, but exhibited poor control of mid-range angles. The authors

also noted that the nonlinear recruitment characteristics of muscle and the disturbances arising

from spastic reflexes would cause problems with the adaptive control algorithm that they chose,

since the algorithm assumed a linear or linearized plant.

Vette, Masani, and Popovic demonstrated that the intact neurological system may employ

a PD-like control strategy to maintain quiet standing using plantarflexor contractions only [19].

However, Vette et al.’s colleague Tan noted that paralysis of the plantarflexor and dorsiflexor

muscles results in reduced ankle stiffness in individuals with SCI, leading to oscillation of the

standing angle with PD control of the plantarflexors only [74].

Chang et al. reported a combined neural network/PID control system for FES-based knee

joint control [75]. The neural network was trained to obtain the inverse dynamics of the knee

joint, and was then used for feedforward control. A fixed-parameter PID controller was used

as a feedback controller in parallel with the feedforward controller to compensate for residual
Chapter 2. Literature Review 22

tracking errors caused by disturbances and modeling errors. This system was tested on one

able-bodied subject and one paraplegic subject. The authors found that the combined neuro-

PID controller performed better than classic fixed-parameter PID control. However, the neural

network may have to be re-trained each time the system was used in order to generate a

sufficiently accurate model of the inverse dynamics of the knee joint.

Ferrarin, Palazzo, Riener, and Quintern compared an adaptive control algorithm with a

feedback, a feedforward, and a combined feedback/feedforward control algorithm [76]. The

algorithms were tested with SCI subjects, and showed that the adaptive controller provided

the best performance in the presence of muscle fatigue. However, the authors noted that

the adaptive controller was only applicable to periodic movements, due to the nature of the

algorithm.

Previdi and Carpanzano used a gain scheduling controller to regulate FES-induced knee

movement [77]. This controller interpolated between local linear quadratic regulators. The

regulators were defined using a nonlinear model of the knee dynamics that was linearized at a

series of operating points.

Jezernik, Wassink, and Keller used sliding mode control to regulate knee angle [78]. Jezernik

et al. used a continuous sliding control law to avoid the problem of chattering in the control,

and tested the controller using both able-bodied and SCI subjects. Good tracking of the desired

knee angle was achieved for up to 8 seconds.

Ajoudani and Erfanian show that a sliding mode controller, in combination with a neural

network controller, can be used to regulate FES-induced knee movements [79]. Kobravi and

Erfanian present an adaptive sliding mode controller for regulating ankle angle with FES that

makes use of co-activation to increase joint stiffness [80].

2.5 Models of Seated Knee Extension Against Gravity

Several models of electrically stimulated knee extension against gravity are available in the

literature [73, 81–87]. These models describe knee angle as a function of quadriceps muscle

stimulation in the seated position, and are summarized in Table 2.2. Hatwell et al. [73] proposed
Chapter 2. Literature Review 23

a deterministic, autoregressive moving average (DARMA) model of the leg dynamics during

knee movement. Previdi presented a nonlinear autoregressive exogenous (NARX) model of knee

movement that resulted from FES induced quadriceps muscle contractions [81]. Ferrarin and

Pedotti used a nonlinear second-order system to model the dynamics of the knee and lower leg,

and a one-pole transfer function to model the relationship between stimulation pulse width and

quadriceps torque [82]. Ferrarin and Pedotti tested their model with both able-bodied and SCI

subjects. Perumal, Wexler, and Binder-MacLeod proposed a model of knee angle in response

to various quadriceps stimulation parameters and different inertial loads at the ankle, and

tested the model with healthy subjects [83]. All of these models provide detailed, deterministic

information, but are not designed to capture the large amount of day-to-day and fatigue-related

variation in electrically stimulated knee angle that is seen in individuals with SCI.

Table 2.2: Models of seated knee extension against gravity. NL denotes nonlinear characteris-

tics, TV denotes time-varying characteristics, NLTV denotes nonlinear time-varying character-

istics.
Authors Year Advantages Disadvantages

Hatwell et al. 1991 Doesn’t capture NLTV char.

Previdi 2002 Doesn’t capture NLTV char.

Ferrarin & Pedotti 2000 Simple model Doesn’t capture NLTV char.

Perumal et al. 2006 Doesn’t capture NLTV char.

Schauer et al. 2002 Includes TV char. Doesn’t capture NL char.

Marion et al. 2009 Includes fatigue Doesn’t capture TV char.

Riener et al. 1996 Includes NLTV char. Complex, difficult to parameterize

A few models of electrically stimulated knee angle as a function of electrical stimulation do

include some sources of variation. Schauer et al. proposed a model that uses a time-varying

component to describe day-to-day variation in the stimulated knee angle [84]. Marion et al.

presented a series of models that predicted fatigue as a function of knee angle for electrically

stimulated quadriceps contractions in able-bodied individuals [85, 86]. Riener, Quintern, and

Schmidt proposed a knee model that included intra-session variation due to fatigue [87]. Reiner’s
Chapter 2. Literature Review 24

model described the dynamics of the knee in response to electrically stimulated contractions of

13 muscles and also accounted for tendon tissues. The nonlinear dynamics of the body segments,

muscle fatigue, and passive muscle velocity were included in this model. Riener’s model was

based on experimentally obtained data, and was tested with five untrained paraplegic subjects.
Chapter 3

Closed-Loop Control Algorithms

Used in Thesis

3.1 Proportional-Integral-Derivative Control

Proportional-integral-derivative (PID) control is a commonly used control algorithm that is

simple, reliable, and has well-established tuning methods [88]. A standard PID controller can

be described by
d
Z
P WQU ADS (t) = KP e(t) + KI e(t)dt + KD e(t),
dt

where KP is the proportional gain, KI is the integral gain, KD is the derivative gain, and e(t) is

the tracking error between the reference signal and the actual system behaviour. Several tuning

algorithms are available for PID controllers, such as the well-known Zeigler-Nichols method [88].

PID is by far the most commonly used control algorithm in industrial applications [89], and

has also been applied to several FES systems [65,68,69]. However, PID controllers are designed

to regulate linear time-invariant (LTI) systems, and commonly exhibit degraded performance

if the system model changes or there are unmodeled or external disturbances. Since FES

applications are nonlinear and time-varying (NLTV), and are subject to both variation in

the model parameters and exogenous disturbances, PID control may not be the best choice

for real-world FES applications. PID control is included in the simulations described in this

thesis despite its obvious drawbacks because of the high level of comfort most neuromodulation

25
Chapter 3. Closed-Loop Control Algorithms Used in Thesis 26

researchers have with this algorithm.

Practical PID control implementations often include an anti-windup feature to prevent

accumulation of the integral term when the control exceeds the saturation limit of the actuator.

Figure 3.1 illustrates one possible anti-windup implementation of a PID controller.

Figure 3.1: Block diagram of anti-windup PID controller. θ denotes feedback signal and u

denotes control signal.

3.2 Gain Scheduling Control

Gain scheduling control (GSC) involves linearizing a nonlinear system model at a series of

operating points that represent the operating region of the system. Consider a nonlinear system

of the form

Σ: ẋ = f (x, u)

y = g(x, u).

Let an operating point be denoted (x0 , u0 ). Therefore, the linearization of Σ about (x0 , u0 ) is

given by

d ∂f (x, u) ∂f (x, u)
Σlin : (x0 + δx) = f (x0 , u0 ) + · δx + · δu + h.o.t.
dt ∂x (x0 ,u0 ) ∂u (x0 ,u0 )

∂g(x, u) ∂g(x, u)
y0 + δy = g(x0 , u0 ) + · δx + · δu + h.o.t.,
∂x (x0 ,u0 ) ∂u (x0 ,u0 )

where h.o.t. denotes higher-order terms.


Chapter 3. Closed-Loop Control Algorithms Used in Thesis 27

A local linear controller is designed at each operating point. Next, a scheduling variable is

chosen that, ideally, captures the nonlinearity of the system and also varies slowly in comparison

to the system dynamics [90]. The scheduling variable is used to interpolate between the local

linear controllers when the system is between the operating points.

Gain scheduling control algorithms can be used very successfully when the operating points

and scheduling variable are chosen appropriately. However, the performance of the controller

may be reduced if the system goes outside the span of the operating points, or if the system

model is not sufficiently accurate. Also, gain scheduling controllers can involve a lengthy tuning

process due to the multiple local linear controllers.

It is particularly difficult to select a scheduling variable for FES applications that captures

the nonlinearity of the system in real-time. Typically, knee angle is used as the scheduling

variable for knee control implementations of gain scheduling control, since knee angle varies

relatively slowly and also captures nonlinearities related to knee angle and varying muscle

length. However, knee angle does not capture other nonlinearities including fatigue, muscle

fitness, and spasticity.

3.3 Sliding Mode Control

Sliding mode control (SMC) involves defining a control law that causes the state of the system

to converge toward a chosen sliding manifold in finite time, and then evolve along that manifold

toward a goal state in possibly infinite time [91–93]. In the context of this thesis, the manifold is

the n-dimensional analogue of a line in R2 −space. SMC provides good tracking performance, is

inherently adaptive, and requires that a relatively small number of parameters be tuned. This

controller also guarantees stability, provided that an accurate plant model is used to develop

the control algorithm.

The sliding manifold is designed to ensure that, once the system state has reached the

manifold, the state will remain on the manifold. This behaviour is referred to as the “sliding

mode” of the controller. Moreover, the manifold is also designed to guarantee that the state

will converge asymptotically toward the goal state in possibly infinite time. The SMC control
Chapter 3. Closed-Loop Control Algorithms Used in Thesis 28

law is designed force the system state to converge to the sliding manifold in finite time.

Consider a nonlinear control affine system of the form

ẋ = f (x) + B(x)u

y = h(x)

where the state vector x ∈ Rn , the input vector u ∈ Rm , the output vector y ∈ Rp , and

f (x), B(x), and h(x) are smooth vector fields on Rn . The sliding manifold is defined to be

m(x, w) = 0, where w ∈ Rn is the state reference vector. Since this thesis is concerned with

designing an output-feedback control law to drive the tracking error between w and x toward

zero, the manifold is defined on the error space e ∈ Rn , where e = w − x. Assuming that n = 2

for simplicity, the sliding manifold is defined to be

m(x, w) = m(e) = e2 + λe1

where ẋ1 = x2 , ẇ1 = w2 , e1 = w1 − x1 , e2 = ė1 = w2 − x2 , and λ > 0. Notice that this manifold

guarantees the stability of the sliding mode, since the definition of the sliding manifold m(e) = 0

implies that ṁ(e) = 0 as well. Moreover, once m(e) = 0, e2 = ė1 = −λe1 , which guarantees the

asymptotic convergence of e1 → 0. Taking the derivative of m(e) = 0,

ṁ(e) = ė2 + λė1 = 0.

This implies that ė2 = −λe2 , which guarantees the asymptotic convergence of e2 → 0, as well.

Therefore, this choice of manifold guarantees that the error state will remain on the manifold

once it has reached the manifold, and will then converge asymptotically toward the goal state

e = 0.

The sliding mode control law must guarantee the reachability of the sliding manifold by

forcing the system state to reach the manifold in finite time. For example, if the control law is

chosen to be

u = −κ sgn[m],

where sgn denotes the signum function, then the error state will be driven toward the sliding

manifold, since u > 0 when m < 0 and u < 0 when m > 0. When m = 0, u = 0, since the error

state is already on the sliding manifold.


Chapter 3. Closed-Loop Control Algorithms Used in Thesis 29

The major disadvantage of SMC is its susceptibility to “chattering” – switching rapidly back

and forth between different control actions as the system approaches the sliding manifold [93].

Chattering is due to the presence of parasitic actuator and sensor dynamics as well as delay in

the switching control action. Chattering is undesirable because it can result in unstable high-

frequency oscillation about the sliding manifold, which can lead to poor controller performance

or controller failure. Many approaches have been proposed to solve the chattering problem

in SMC. Two of these proposed solutions are used in this thesis: boundary layer SMC and

observer-based SMC.

3.3.1 Boundary Layer Sliding Mode Control

Boundary layer SMC involves defining a modified control law that replaces the switching element

of the SMC control law with a smooth approximation of the switching element within a small

boundary region that encompasses the manifold. Slotine and Li were the first to propose this

solution to the chattering problem, and suggested replacing the switching function with high-

gain proportional feedback [94]. Continuing the example in the preceding section, a boundary

layer SMC could be implemented by substituting the modified control law u = −κ m for the

SMC control law u = −κ sgn[m] within a boundary layer of the manifold of thickness ψ.

Therefore, the overall boundary layer control law is


 −κ1 sgn[m] ∀ d > ψ

u=
 −κ2 m

∀d≤ψ

where κ1 > 0, κ2 > 0, ψ > 0, and d is the distance from the error state to the manifold.

The disadvantage of boundary layer control is that, unless the parasitic dynamics of the

system are accurately known, a relatively large boundary region must be implemented to avoid

chattering. Since the finite-time convergence of SMC is not guaranteed by the proportional

feedback control law implemented inside the boundary region, the attractive features of SMC

may be compromised by such a wide boundary region.


Chapter 3. Closed-Loop Control Algorithms Used in Thesis 30

3.3.2 Observer-Based Sliding Mode Control

Observer-based SMC uses an asymptotic observer to estimate the system state and to close

a high-frequency by-pass loop around the controlled plant. Provided that the observer error

(i.e., the error between the actual system state and the estimated state) is less than the plant

error (i.e., the error between the actual plant dynamics and the nominal plant dynamics), and

that the control is discontinuous with respect to the observer variables only, then all of the

chattering will be located inside the by-pass loop. In this configuration, chattering will exist in

the observer, but will not excite any high-frequency dynamics in the plant [93, 95].

If the system can be represented in linear time-invariant form, then a Luenberger observer

[88] can be used to provide asymptotic convergence of the observer error x̃ = x − x̂ to zero. Let

the system dynamics be

ẋ = Ax + Bu

y = Cx.

Let the observer state be x̂ = [x̂1 x̂2 ]T , and let the observer state dynamics be

x̂˙ = Ax̂ + Bu + L (y − ŷ) = (A − LC) x̂ + Bu + LCxs

ŷ = C x̂,

where xs is the output feedback, and L is the observer gain. Choosing L such that (A − LC) is

Hurwitz guarantees asymptotic convergence of the observer error x̃ toward zero [88]. Note that

a matrix is Hurwitz if all of its eigenvalues have a strictly negative real part. The observer-based

SMC can then be constructed by defining the sliding manifold mOBS (x̂, w) = 0 on the observer

state space such that

mOBS = (w2 − x̂2 ) + λ (w1 − x̂1 ) ,

where λ > 0. Therefore, the observer-based SMC control law can be chosen to be

uOBS = −κ sgn [mOBS ]

where κ > 0.
Chapter 4

Analysis of Existing Closed-Loop

FES Control Algorithms

4.1 Introduction

Despite the wide range of FES applications that have been reported in the literature [3–27, 29–

37], few electrical stimulation systems that can generate meaningful functional outcomes are

currently available for use outside research laboratories. This relative scarcity of widely available

advanced FES systems that are capable of generating complex motor tasks such as walking,

standing, reaching, and grasping may be partly due to a limited ability to accurately control

FES-induced muscle contractions and compensate for disturbances that affect the controlled

neuromuscular system.

A large body of work on closed-loop control of FES is available [16, 18, 19, 25, 63–73, 75,

76, 78–80, 82], however most commercially available FES systems use open-loop control or very

simple closed-loop control strategies, such as rule-based state control, which severely limits the

flexibility of these systems. The purpose of this study was to elucidate the behaviour of several

closed-loop control systems, which regulate electrical stimulation from pulse to pulse in response

to feedback from the controlled system, and to identify some ways in which closed-loop control

systems for FES applications could be improved and applied to real-world FES systems.

In this study, the performance of several promising closed-loop FES controllers was com-

31
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 32

pared in simulation to determine how these algorithms perform under conditions that mimic

real-world applications. The simulated FES system was subjected to muscle fatigue, muscle

spasms, and the effects of muscle re-training. Standard control performance metrics were used

to compare the performance of the different controllers. These simulations were conducted to

inform the design of future control systems for FES applications by illuminating the ways in

which currently available FES control systems could be improved.

The challenges inherent in controlling electrically stimulated muscle contractions are sum-

marized in Section 4.2. In Section 4.3, the design of the three closed-loop control algorithms

that are evaluated in this study is described. The design and implementation of the FES system

that was used for controller testing, as well as the testing methodology, are also discussed. The

results are presented in Section 4.4 and discussed in Section 4.5. Section 4.6 contains the con-

clusions of this study, including an assessment of the implications of these results for designing

closed-loop controlled FES systems for real-world applications.

4.2 Background

The response of muscles to electrical stimulation is nonlinear and time-varying, and is also

subject to unpredictable reflexes that can generate strong perturbations [59]. Spasticity must

also be taken into account when designing FES control systems, since the controller should

ideally be able to reject disturbances and counteract the increased torque resulting from spastic

muscles. Muscle fatigue is a significant factor in the short-term time-varying nature of the

stimulated muscle response, and muscle re-training effects contribute to long-term changes in

the stimulated muscle response.

Real-world FES applications require the ability to modulate the pulse-to-pulse electrical

stimulation in real-time to compensate for fatigue, spasticity, and re-training effects, as well as

modeling errors and exogenous disturbances. Closed-loop control can be used to address these

challenges.

A realistic simulation of knee extension against gravity in response to electrical stimulation

of the quadriceps muscle group was implemented. This simulation was based on the knee model
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 33

developed by Ferrarin and Pedotti [82], which is a simple and physiologically accurate model

of the stimulated knee response that was characterized using data from complete SCI subjects.

The muscle recruitment dynamics are represented using a one-pole transfer function, given a

fixed stimulation pattern, frequency and amplitude.

The model in [82] was extended in this study to include fatigue, spasticity and muscle re-

training effects. This simulation was then used to test three closed-loop FES control algorithms:

a) proportional-integral-derivative (PID) control, b) gain scheduling control (GSC), and c)

boundary-layer sliding mode control (SMC). These algorithms were chosen to represent the

range of controllers that have been reported to be successful in the literature, namely a simple

linear controller (PID), a sophisticated linear controller (GSC), and a sophisticated nonlinear

controller (SMC).

4.3 Methods

4.3.1 Implementation of Base Knee Model

The FES control simulation was based on the model of stimulated knee extension that Ferrarin

and Pedotti proposed in [82]. This model assumes that the individual is seated and the lower

leg is free to swing. Knee extension is provided by electrical stimulation of the quadriceps

muscle group, and knee flexion is provided by gravity. The model is physiologically accurate,

and accounts for the gravitational and inertial characteristics of the lower leg and foot, as well

as the damping and stiffness characteristics of the knee. The input is the pulse width of the

stimulation waveform delivered to the quadriceps muscle, and the output is knee angle. The

parameters of the controller are specific to a particular stimulation waveform; in this study, a

square wave pulse train with constant frequency and amplitude was used, and the pulse width

was allowed to vary. The knee model in [82] was characterized using data collected from several

subjects with complete SCI who had not undergone any muscle strengthening regime prior to

data collection.

In this study, the dynamics of the knee are expressed as

JSHAN K φ̈(t) = τGRAV IT Y (t) + τST IF F (t) + τDAM P (t) + τQU ADS (t),
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 34

where φ(t) is the knee angle, φ̈(t) is the angular acceleration of the knee joint, and JSHAN K

is the moment of inertia of the leg below the knee. Positive torque is defined to be in the

direction of knee extension, that is, straightening the knee. φ(t) is defined to be the angle

between vertical (that is, the gravitational vector) and the longitudinal axis of the shank, as

illustrated in Figure 4.1. The gravitational torque is

τGRAV IT Y (t) = −mgℓ sin φ(t),

where m is the mass of the shank and foot, g is the gravitational constant, and ℓ is the distance

from the centre of mass of the shank and foot to the pivot point of the knee. The stiffness

torque is
π
 π 
τST IF F (t) = −λe−E (φ(t)+ 2 ) φ(t) + − ω ,
2
where λ and E are coefficients of the exponential term, and ω is the resting elastic knee angle.

The damping torque is

τDAM P = −B φ̇(t),

where B is the viscous damping coefficient of the knee and φ̇(t) is the angular velocity of the

knee. The relationship between the stimulation pulse width delivered to the quadriceps muscle

group and the resulting torque exerted about the knee is expressed in the frequency domain as
τQU ADS (s) G
= ,
P WQU ADS (s) 1 + ηs
where P WQU ADS (t) is the time domain pulse width of the stimulation train delivered to the

quadriceps, G is the transfer function gain for a particular stimulation frequency, and η is the

time constant of the transfer function pole for a particular stimulation pattern. A one-pole

transfer function was sufficient to describe the muscle recruitment dynamics for this study,

because the extremes of the range of motion of the knee were avoided by limiting the knee

angle to between 0 and 90 degrees of extension.

The values of the model coefficients were specific to each subject, and are given in [82].

4.3.2 Enhanced Knee Model

The base knee model described in [82] was extended in this study to include fatigue, spasticity

and the effects of muscle re-training. These extensions allowed the likely real-world performance
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 35

Figure 4.1: Diagram of knee extension. Approximate position of hamstrings muscle group is

indicated by heavy line, showing insertion point on shank. Angle between axis of shank and

vertical is denoted φ, and angle between axis of thigh and hamstrings is denoted ϕ.

of the FES control algorithms to be assessed.

Fatigue was implemented as a multiplier that modified the electrically stimulated quadriceps

torque. The fatigue function described by the author in [44] was used, which represents the

typical reduction in electrically stimulated quadriceps torque seen in individuals with SCI as the

muscles fatigue. The value of the fatigue function ranges between 0 and 1, with 1 corresponding

to no fatigue and 0 corresponding to no measurable response to stimulation.

Only intrinsic tonic spasticity was implemented, as defined by Adams and Hicks in [96],

which causes muscle tension to increase proportionally to the velocity of muscle stretch. In

the simulation presented in this study, the hamstrings muscle group stretches when the knee

extends. Therefore, including spasticity in this model resulted in the hamstrings exerting a

torque about the knee in opposition to the electrically stimulated quadriceps torque.

Several simplifying assumptions were made in the description of spasticity. The velocity

of the hamstrings stretch was assumed to be proportional to the angular velocity of the knee.

Moreover, it was assumed that there was no minimum stretch velocity required to elicit spastic

behaviour, and that there was no minimum threshold of muscle tension required to generate the

spastic torque. It was also assumed that the angle ϕ between the longitudinal axis of the thigh

and longitudinal axis of the hamstrings muscle group, as shown in Figure 4.1, was constant.
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 36

Under these assumptions, the spastic hamstrings torque can be expressed as

τSP ASM (t) = TSP ASM (t) ℓ̄ sin ρ(t),

where TSP ASM (t) is the hamstrings tension resulting from the spasm. The distance between

the centre of rotation of the knee and the insertion point of the hamstrings tendons on the

fibula (i.e. the long bone of the lower leg) was assumed to be ℓ̄ = 5 cm. The angle between the

longitudinal axis of the hamstrings and the axis of the shank can be expressed as
π 
ρ(t) = π − ϕ(t) − + φ(t) .
2
The hamstrings tension is given by




 0 ∀ φ̇(t) < 0

TSP ASM (t) = 0 ∀ φ̇(t) ≥ 0 and t < Td



 γ φ̇(t) otherwise,

where Td = 40 ms is the delay between the stretch and the onset of the resulting contrac-

tion [96], and γ is such that the maximum spastic torque is elicited at the maximum knee

angular velocity. The maximum spastic torque was set to 4 Nm, which is a typical maximum

electrically stimulated hamstrings torque seen in the trained SCI subjects who were participants

in the author’s experiments. 360 deg/s was used as the maximum angular velocity of the knee,

which corresponded to the angular velocity of the knee during normal walking in able-bodied

individuals [97]. Based on the author’s experimental work with SCI subjects and currently

available surface stimulation technology, it was assumed that the knee angular velocity of the

typical individual with SCI would not exceed 360 deg/s, even after re-conditioning the muscles.

The effects of muscle re-training were implemented by assuming that, in the best case, a

SCI user of FES would have a stimulated muscle response equivalent to that of an able-bodied

individual. The modified recruitment function was therefore given by


αG
τQU ADS (s) = P WQU ADS (s),
1 + βηs
where α and β modified the stimulated muscle response to yield the same behaviour as an

able-bodied individual, as defined in [82].

The enhanced knee model was implemented in Simulink (The Mathworks, Inc., Natick,

USA).
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 37

4.3.3 Design and Implementation of the Control Algorithms

PID Control Algorithm

A standard PID control algorithm was used to implement the PID controller, as discussed in

Section 4.1. The knee control system exhibited oscillation with KP = 1, KI = 0, KD = 0.

Since standard PID tuning algorithms such as the Zeigler-Nichols method are not applicable to

such systems [88], trial and error was used to tune the controller.

Gain Scheduling Control Algorithm

The gain scheduling controller (GSC) was developed using the method of Previdi and Carpan-

zano, who used GSC to regulate electrically stimulated knee movement with a different knee

model [77]. The operating points were selected to be {15◦ , 30◦ , 45◦ , 60◦ , 75◦ , 90◦ }. The be-

haviour of the enhanced knee model is not well-defined at the extremes of the knee’s range of

motion, so these regions were not included in the operating points [82]. For this reason, the

GSC was expected to suffer a decline in performance if the knee angle went outside the span

of the operating points.

The base knee model was then linearized at each of the operating points. The base model

was used for controller development instead of the enhanced knee model to assess the affect of

using a simple model on controller performance. The state variables of the model were defined

to be x(t) = [x1 (t) x2 (t)]T = [φ(t) φ̇(t)]T , the input variable to be u(t) = P WQU ADS (t), and

the output variable to be y(t) = φ(t). The state space knee model is given by

Σ: ẋ(t) = f (t, x, u)

y(t) = g(t, x, u)
1  π
 π 
∴Σ: ẋ(t) = −mgℓ sin x1 (t) − λe−E (x1 (t)+ 2 ) x1 (t) + − ω
JSHAN K 2
G −t/η
−Bx2 (t) + e ∗ u(t)µ(t)
η
y(t) = x1 (t),

where µ(t) is the Heaviside function. The operating point was denoted x0 = [x01 x02 ]T , where

x1 (0) = x01 , and x02 = 0 for all operating points since knee velocity was zero at a fixed operating
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 38

point. The input u(t) could vary at each operating point, so the linearization of Σ about x0

was

d ∂f (t, x, u)
Σlin : (x0 + δx(t)) = f (t, x0 , u) + · δx(t) + higher order terms
dt ∂x
(t,x0 ,u(t))

∂g(t, x, u)
y0 + δy(t) = g(t, x0 , u) + · δx(t) + higher order terms
∂x
(t,x0 ,u(t))

    
d
x01 x01

dt + δx1 (t)   0 1  + δx1 (t)  G
 ∼
=  +b+ e−t/η ∗ u(t)µ(t)

∴
J η

d
x02 + δx2 (t) x02 + δx2 (t) SHAN K

dt a21 a22
 
 
x 0 + δx (t)
 1 1
y0 + δy(t) ∼= ,,

1 0 
x02 + δx2 (t)

where
 0 π
  π 
a21 = −mgℓ cos x01 − λe−E (x1 + 2 ) 1 − E x01 + − ω /JSHAN K
2
a22 = −B/JSHAN K
 0 π
π  π  
−mgℓ sin x01 − x01 cos x01 − λe−E (x1 + 2 ) − ω + E x01 + − ω x01 /JSHAN K .

b =
2 2

The values of a21 , a22 , and b were calculated for each of the six operating points.

Integral control was used to ensure zero steady-state error for the knee angle tracking

task [88], such that the error dynamics were


Z Z
x01 − x01 + δx1 (t)

ξ(t) = e(t)dt = dt

¨ = − x02 + δx2 (t) .



∴ ξ(t)

The linearized knee dynamics were then augmented with the error dynamics. The augmented
˙ T , input v(t) = b+
system had state vector w(t) = [x01 +δx1 (t) x02 +δx2 (t) ξ(t) ξ(t)] G
JSHAN K η e
−t/η ∗

u(t)µ(t), and output z(t) = ξ(t), such that


   
0 1 0 0 0
   
   
 a21 a22 0 0   1 
Σaug : ẇ =   w +   v = A0 w + B0 v
   
 0 0 0 1  0 
   

   
0 −1 0 0 0
 
z = 0 0 1 0 w = C0 w.
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 39

Linear quadratic regulators (LQR) were used as the local control algorithm. The LQR

control law was given by

v(t) = −K0 w(t).

The MatLab (The Mathworks, Natick, USA) function K0 = lqr (A0 , B0 , Q0 , R0 ) was used to

find the gain matrix K0 at each operating point x0 . The local regulators were tuned by choosing

the weighting matrices Q0 and R0 of the Riccati equation to provide good unit step response

behaviour, that is, minimal overshoot and fast rise and settling times, with small steady-state

error. The reference trajectory for the tuning process was

θREF (t) = x01 (t) + 0.1µ(t − 1)

where the Heaviside function µ(t − 1) generated a small step at t = 1 s. The tuning pro-

cess yielded six gain matrices corresponding to the six operating points. A clamped spline

interpolation was used to find the appropriate controller gains at points between the operating

points.

The control signal u(t) at time ti was determined in real time by solving the approximate

deconvolution
G −t/η
v̄(t) = e ∗ u(t)µ(t) ∀ 0 ≤ t < ti

for u(t), where 
 v(t) − b ∀ 0 ≤ t < ti

v̄(t) =
 0

∀ ti ≤ t ≤ 2ti .

Under the test conditions described in Section 4.4, this approximate deconvolution resulted in

very close agreement between the actual and estimated control signals.

Sliding Mode Control Algorithm

For the sliding mode controller (SMC), the position error was defined to be e1 (t) = φREF (t) −

φ(t), and the velocity error to be e2 (t) = φ̇REF (t) − φ̇(t). The goal state was defined to be

the origin of the error plane, that is, zero position error and zero velocity error. The sliding

manifold was defined to be a line in R2 -space that passed through the origin of the error plane.

The control law caused the system error to converge toward the sliding manifold, and then
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 40

travel along this line toward zero position and zero velocity errors. The sliding manifold was

denoted by

m(t) = e2 (t) + Λe1 (t), Λ > 0,

where Λ was a constant. Choosing the control law

ṁ(t) = −k1 sgn [m(t)] , k1 > 0

caused the error dynamics to approach the manifold in finite time [91]. Once the error dynamics

were on the manifold, the choice of manifold guaranteed that e1 (t) → 0 as t → ∞.

Since this control law will cause chattering as the error dynamics near the manifold, a second

control law

ṁ(t) = −k2 m(t), k2 > 0

was defined and applied within a boundary region of the sliding manifold. Therefore, the overall

sliding mode control law was



 −k1 sgn [m(t)] ∀ d > κ

ṁ(t) =
 −k2 m(t)

∀d≤κ

for some κ > 0, where d was the shortest distance between the current error (e1 (t), e2 (t)) and

the manifold. Equating the overall control law and the derivative of the sliding manifold
 
ṁ(t) = ė2 (t) + Λė1 (t) = φ̈2,REF (t) − φ̈(t) + Λ φ̇2,REF (t) − φ̇(t) ,

an expression was found for the knee angle


  
 φ̈REF (t) + Λ φ̇REF (t) − φ̇(t) + k1 sgn [m(t)] ∀ d > κ

φ̈(t) =  
 φ̈REF (t) + Λ φ̇REF (t) − φ̇(t) + k1 m(t)

∀d≤κ

As defined in Section 4.3.2, the state space model of the knee system was given by

ẋ1 (t) = x2 (t)


1   π  
ẋ2 (t) = −mgℓ sin (x1 (t)) − λe−E(x1 (t)+π/2) x1 (t) + − ω − Bx2 (t) + τquad (t) .
JSHAN K 2

Substituting the expression for φ̈(t) into the above equation, and changing variables such that

φ(t) = x1 (t), φREF (t) = x1,REF (t), φ̇(t) = x2 (t), and φ̇REF (t) = x2,REF (t), an expression for
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 41

the stimulated quadriceps torque τQU ADS (t) was given by



mgℓ sin x1 (t) + λe−E(x1 (t)+π/2) x1 (t) + π2 − ω + Bx2 (t)
 





+JSHAN K (ẋ2,REF (t) + Λ (x2,REF (t) − x2 (t)) + k1 sgn [m(t)]) ∀ d > κ


τquad (t) =
mgℓ sin x1 (t) + λe−E(x1 (t)+π/2) x1 (t) + π2 − ω + Bx2 (t)
 






+JSHAN K (ẋ2,REF (t) + Λ (x2,REF (t) − x2 (t)) + k2 m(t)) ∀ d ≤ κ.

The pulse width required to generative the desired τQU ADS (t) was found by approximately

deconvolving the exponential term in the muscle recruitment function from τQU ADS (t), using

the procedure described in Section 4.3.2, so that


 
G
P WQU ADS (t) = deconv τQU ADS (t), e −t/η
,
JSHAN K η

where P WQU ADS (t) was the time domain pulse width of the stimulation train delivered to the

quadriceps, as described in Section 4.1.

The values of parameters k1 , k2 , Λ, and κ we tuned by trial and error so that the evolution

of the errors in the error plane followed the sliding mode ideal of converging towards and then

evolving along the sliding manifold as closely as possible.

4.3.4 Controller Testing Methodology

The control algorithms were tested with four simulated subjects, designated H4, P1, P2, and

P3 in [82]. Subject H4 was an able-bodied subject, and the other three subjects were spinal

cord injured. Subject H4 was used as an example of the optimal case, that is, how well the

controllers might perform if the end user had paralyzed muscles that had undergone extensive

re-training. The controllers were simulated using a 50 Hz square wave stimulation pulse train

with an amplitude of 80 mA. The pulse width was regulated by the control algorithm between

0 µs and 250 µs. Two reference trajectories were used to evaluate the performance of the

control algorithms. The first was a step trajectory that had an initial value of 5 degrees of knee

extension (i.e., the approximate resting elastic knee angle), and a final value of 45 degrees of

knee extension. The second trajectory was based on the knee’s movements during walking, as

described by Perry in [98].


Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 42

Each of the three control algorithms was tested both with and without fatigue, spasticity,

and training effects. The tracking performance of the control algorithms was quantified using

standard control systems metrics [88]. For the step trajectory, these metrics were 10%-90% rise

time, 2% settling time, percent overshoot past the steady-state knee angle, and steady-state

error between the actual and desired knee angles. For the walking trajectory, these metrics

were the lag between the desired and actual knee angles, and the root-mean-squared (RMS)

error between the desired and actual knee angles.

In addition, the sensitivity of the control algorithms to model mismatch error was assessed

by recording the RMS error for ±20% and ±50% mismatch in parameters ℓ, m, B, λ, E, G,

and η.

4.4 Results

The unit step and walking trajectory response metrics for subject P1 for each controller are

reported in Table 4.1, which compares the ideal case to the non-ideal case that includes mild

fatigue, spasticity, and training effects. The steady-state error is not given for the non-ideal

case, because the knee angle did not reach steady-state in the presence of fatigue.

Table 4.1: Response metrics for unit step and walking trajectories for subject P1. Non-ideal

case includes mild fatigue, spasticity, and training effects. Settling time and overshoot for non-

ideal case are referred to steady-state values from ideal case, since fatigue prevents non-ideal

case from reaching steady-state.


trise tsettling % OS ess lag eRMS
Simulation Case
(s) (s) (deg) (s) (deg)

Ideal PID 2.24 11.07 17.35 0.00 0.21 16.48

Ideal GSC 1.11 1.93 1.53 0.00 0.24 13.07

Ideal SMC 0.75 3.06 6.38 -1.43 0.10 14.50

Non-ideal PID 2.20 7.69 14.13 0.17 16.31

Non-ideal GSC 1.04 1.71 1.28 0.26 13.47

Non-ideal SMC 0.84 3.39 13.57 0.11 14.73


Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 43

In Figure 4.2, the tracking performance of each controller for the step trajectory is shown,

for the ideal case with subject P1. In Figure 4.3, the tracking performance of each controller

for the walking trajectory is shown. In this instance, the results for subject P2 for the non-ideal

case that includes mild fatigue, spasticity, and training effects are shown. The final 15 seconds

of the trajectory are shown to illustrate the effects of fatigue. In Figure 4.4, the same results

are shown as in Figure 4.3 but with severe fatigue instead of mild fatigue, as defined in [44]. In

Figure 4.5 and 4.6, the sensitivity of the three controllers to model mismatch errors is reported,

for the ideal case with subject P1. Figure 4.5 contains the results for a ±20% mismatch, and

Figure 4.6 contains the results for a ±50% mismatch.

60

50

40
Knee Angle (deg)

30

20

10 Reference Trajectory
PID Control
Sliding Mode Control
Gain Scheduling Control
0
0 5 10 15
Time (s)

Figure 4.2: Step trajectory tracking for subject P1 under ideal simulation conditions.

4.5 Discussion

The PID controller generated large overshoot, slow rise and settling times, was most sensitive to

model mismatch errors and generally exhibited the poorest performance of the three controllers.

The GSC generated a slow response with very little overshoot and zero steady-state error

due to the use of integral control. This controller exhibited the largest lag, but was also least
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 44

80
Reference Trajectory
PID Control
70 Sliding Mode Control
Gain Scheduling Control

60
Knee Angle (deg)

50

40

30

20

10

0
5 10 15 20
Time (s)

Figure 4.3: Tracking of the walking trajectory for subject P2 under non-ideal simulation con-

ditions, including mild fatigue, spasticity, and training effects. Thick solid line is reference

trajectory, thin solid line is PID control, dashed line is sliding mode control, and dotted line is

gain scheduling control.


Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 45

80
Reference Trajectory
PID Control
70 Sliding Mode Control
Gain Scheduling Control

60
Knee Angle (deg)

50

40

30

20

10

0
5 10 15 20
Time (s)

Figure 4.4: Tracking of the walking trajectory for subject P2 under non-ideal simulation con-

ditions including severe fatigue. Thick solid line is reference trajectory, thin solid line is PID

control, dashed line is sliding mode control, and dotted line is gain scheduling control.
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 46

Figure 4.5: Sensitivity analysis for subject P1 under ideal simulation conditions for ±20%

parameter mismatch. Markers represent ratio of RMS error with mismatch in a single parameter

to RMS error with nominal parameter values. Heavy solid line represents unit circle. Thin solid

line represents results obtained with PID controller, dashed line corresponds to sliding mode

controller results, and dotted line corresponds to gain scheduling controller results. Y represents

λ, and Z represents η. The unlabelled chord is a place holder to provide convenient spacing of

the remaining chords.


Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 47

Figure 4.6: Sensitivity analysis for subject P1 under ideal simulation conditions for ±50%

parameter mismatch. Markers represent ratio of RMS error with mismatch in a single parameter

to RMS error with nominal parameter values. Heavy solid line represents unit circle. Thin solid

line represents results obtained with PID controller, dashed line corresponds to sliding mode

controller results, and dotted line corresponds to gain scheduling controller results. Y represents

λ, and Z represents η. The unlabelled chord is a place holder to provide convenient spacing of

the remaining chords.


Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 48

sensitive to parameter variation. Overall, this controller performed quite well. However, the

GSC required a lengthy tuning process due to the multiple local controllers, and also ran slowly

in simulation due to the spline interpolation that must be performed at each time step.

The SMC exhibited a fast response with little overshoot and a short lag. This controller

provided fast tuning, and also exhibited the best sinusoidal tracking performance, as illustrated

in Figure 4.3. However, this controller also yielded the largest steady-state error. This error

could be due to the use of a boundary layer to reduce chattering. The classic SMC behaviour is

suppressed within the boundary layer, so the error is not guaranteed to converge toward zero.

The SMC was also fairly sensitive to parameter mismatch, which is somewhat surprising given

the relatively robust nature of SMC. This discrepancy could also be due to the use of boundary

layer control, which compromises the convergence properties of the SMC. The performance

of this controller might be improved by using a more accurate model of the stimulated muscle

response that includes non-linearities, or by using a different method to overcome the chattering

problem.

The inclusion of fatigue in the simulations dampened the response of each controller, and

also increased the RMS error in each case. This effect is likely to be even more pronounced

for instances of stronger fatigue. The inclusion of spasticity also dampened the response of

each controller, but spasticity decreased the RMS error, unlike fatigue. This is an interesting

result, since intrinsic tonic spasticity effectively increases the joint stiffness for the duration of

the spasm. Including training effects in the simulation resulted in larger overshoot but also

less lag and decreased RMS error. These changes were likely due to the stronger contractions

improving the ability of the controllers to track the desired trajectories.

When fatigue, spasticity, and training effects were applied at the same time, the effects

seemed to cancel each other, resulting in performance that was comparable to the base case

with no non-linearities, as shown in Table 4.1. However, this effect is not seen with stronger

fatigue, as shown in Figure 4.4; while the GSC was the least susceptible to the effects of fatigue,

all the controllers showed a rapid drop-off in the simulated knee angle. Moreover, other types

of spasticity may have a less benign effect of the response of the controllers than intrinsic tonic

spasticity.
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 49

The fatigue model is based on empirical data collected from SCI subjects, and corresponds

to the fatigue profile most commonly seen in the author’s experimental work with trained,

complete SCI subjects. However, this fatigue model may not represent the FES user population

at large. Moreover, this model was not validated against an independent data set. The spasticity

model is based on published data and was also not validated by the author against independent

experimental data. Therefore, these models could be a source of error in the results reported

in this chapter.

It is also worthwhile to note that, although the GSC was able to provide zero steady-state

error through the use of integral control, it may be advantageous for real-world FES applications

for individuals with SCI to avoid a steady state so that the controller has both position and

velocity feedback information at all times. Moreover, this approach would result in continuous,

or at least frequent, muscle stimulation, thereby taking advantage of potentiation effects and

avoiding waiting for “cold” muscles to respond to stimulation when a control action is needed.

Most importantly, none of the controllers tracked the walking trajectory well enough for

real-world use, despite this trajectory representing a very slow walking pattern. The SMC per-

formed well at this task when no non-linearities were included in the simulation. However, the

performance of all of the controllers was wholly inadequate when the non-linearities were intro-

duced, as shown in Figure 4.4. Also, the sensitivity analysis showed that all of the controllers

were sensitive to errors in the anatomical model parameters m and ℓ as well as errors in the

muscle recruitment function, particularly the gain parameter G. These results suggest that an

accurate model is important for maximizing controller performance in FES applications.

4.6 Conclusions

Three representative closed-loop control algorithms for FES applications were tested in simu-

lation to uncover some of the reasons why these methods are not being used in real-world FES

systems. None of the three algorithms were found to be suitable for real-world use in the forms

presented in this study. All of the controllers exhibited significantly degraded performance when

real-world nonlinear effects (that is, fatigue, spasticity, and training effects) were included in
Chapter 4. Analysis of Existing Closed-Loop FES Control Algorithms 50

the simulation. Moreover, all of the controllers were sensitive to variation in the parameters of

the muscle recruitment function, which are subject to change during real-world FES use.

The results of this study suggest several ways to improve the performance of closed-loop

control algorithms for use in real-world FES systems. First, the user’s muscles should be

extensively trained with electrical stimulation to strengthen the muscles and minimize the

detrimental effects of fatigue on controller performance. Second, it is advisable to use an

accurate system model for controller development, including the nonlinearities that are present

in the electrically stimulated system. However, the resulting complex system model will likely

require a sophisticated control technique that can accommodate nonlinear time-varying systems.

Third, co-contracting opposing muscle groups to introduce stiffness to the joint behaviour

should be considered, as suggested by Kobravi and Erfanian in [80]. This strategy is supported

by the improvements in controller performance that were observed in those simulations that

included spasticity, since intrinsic tonic spasticity results in co-contraction through increased

tension in the agonist muscle group.

Closed-loop control algorithms can theoretically improve upon the performance of open-

loop control algorithms for FES applications. However, this study showed that the actual

performance of three representative closed-loop controllers was inadequate for real-world use in

FES systems. Closed-loop controllers have an important place in future FES applications, but

the performance of these algorithms must be greatly improved before they can be implemented

in real-world systems.
Chapter 5

Construction of Non-Idealities

Model

5.1 Introduction

Each new FES application must be thoroughly tested prior to wide-spread use. This testing

should be conducted with individuals who have complete SCI to ensure that all the control

effort is being generated by the FES system alone. If FES systems are tested using able-bodied

or incomplete SCI subjects, there is the possibility of unintentional voluntary contractions

occurring at the same time as the FES induced contractions, thereby clouding the results of the

testing process. Moreover, test subjects should undergo a period of electrical stimulation-based

muscle re-conditioning prior to exposing them to FES. This process is necessary because many

individuals with SCI have severely atrophied muscles, and this muscle conditioning will ensure

that they have sufficient muscle strength and endurance to take part in FES-based therapies

or interventions. In addition, muscle conditioning will help to achieve as consistent and as

repeatable muscle responses to FES as is possible.

Additionally, in individuals with SCI, the FES induced muscle contractions are subject

to undesirable effects such as day-to-day and fatigue-related variation, muscle spasms, and

tremors due to incomplete tetanus [43, 55]. Therefore, testing of FES systems should be done

over multiple sessions to account for fatigue and response variability.

51
Chapter 5. Construction of Non-Idealities Model 52

It can be time-consuming and expensive to recruit suitable subjects, re-condition the sub-

jects’ muscles using electrical stimulation, and then conduct exhaustive testing to verify the

performance of a particular FES system. For this reason, FES applications are typically re-

fined in simulation prior to testing with human subjects. However, such simulations of FES

applications are usually based on models of the typical or ideal response of the muscle to FES,

which frequently results in an overly optimistic assessment of the FES system’s performance in

real-world applications. Consequently, it is not unusual that FES applications that were tested

and validated in simulation are found to be unfeasible in actual real-world applications.

This study proposed an improved method of simulating FES applications that provides a

more realistic estimate of real-world performance by including actual non-ideal stimulated mus-

cle behaviour in the FES simulation. A “non-idealities” block was implemented in Simulink

(The Mathworks, USA) that modifies the nominal stimulated muscle response (i.e., the re-

sponse described by a model of the average or typical stimulated muscle response) to include

undesirable behaviour such as fatigue, muscle spasms and tremors that are commonly present

during FES-induced muscle contractions. This non-idealities block was based on actual data

collected from complete SCI subjects during electrically stimulated muscle contractions. These

data represent the range of responses to electrical stimulation that have been seen in the Re-

habilitation Engineering Laboratory at the Toronto Rehabilitation Institute and the University

of Toronto with SCI subjects.

This block can be incorporated into existing FES simulations to analyze the potential per-

formance of these systems under realistic conditions. The non-idealities block also allows the

user to specify the severity of the different facets of undesirable behaviour that are implemented

in the block. For example, in some muscles the spasms are minimal while fatigue is very pro-

nounced, and in other muscles tremors are prevalent while fatigue occurs slowly and is fairly

moderate.

This chapter includes an example of implementing the non-idealities block in a simula-

tion of closed-loop regulation of electrically stimulated knee extension against gravity. Both

a proportional-integral-derivative (PID) and a sliding mode controller (SMC) were applied in

this example.
Chapter 5. Construction of Non-Idealities Model 53

No model of the electrically stimulated muscle response is available that includes all sources

of day-to-day and intra-session variation. The non-idealities block proposed in this study ad-

dresses this deficiency in the literature by providing information about the worst-case scenarios

that can occur during electrically stimulated contractions. This study also illustrated how to

incorporate the non-idealities block into simulations of FES applications, and used this example

to show that the performance of the closed-loop controlled FES system degraded in the presence

of real-world non-idealities as compared to the performance for the nominal model.

The pilot work for this project was reported at the 2010 International Functional Electrical

Stimulation Society Conference [38]. The work reported in this chapter is an expanded version

of that pilot work based on a larger amount of experimental data, and includes additional

information on how to implement the non-idealities block for testing FES systems.

Section 5.2 discusses the data analysis methods used to extract the non-ideal behaviour from

the experimental data as well as the design of the non-idealities block. This section also describes

how the non-idealities block was implemented in a simulation of electrically stimulated knee

angle regulation. Section 5.3 contains the simulated performance of the knee angle regulation

example both with and without the non-idealities block. Section 5.4 discusses these results,

and Section 5.5 draws conclusions about these results.

5.2 Methods

5.2.1 Construction of Non-Idealities Block

The local research ethics board approved this study, and all experimental subjects provided

informed consent. All subjects had complete thoracic or low cervical SCI (AIS A classification)

and were neurologically stable. The non-idealities block was developed using data collected

from these subjects during electrically stimulated knee movements.

Fatigue Data

The fatigue waveforms were extracted from data collected from seven untrained subjects [99].

Each subject was seated with the knee and hip fixed in the flexed position. Adhesive reusable
Chapter 5. Construction of Non-Idealities Model 54

electrodes were affixed to the skin above the quadriceps muscle group. A 5 cm square electrode

was placed at the distal end of the quadriceps as the reference electrode, and a 5 cm by 10

cm electrode was placed at the proximal position to stimulate the motor points of the rectus

femoris and vastus lateralis heads of the quadriceps muscle group. The quadriceps muscle group

was electrically stimulated to elicit an isometric contraction.

A Compex Motion stimulator (Compex SA, Switzerland) delivered a biphasic bipolar pulse

train (pulse width 250 µs, frequency 40 Hz) to the muscles. The amplitude of stimulation was

individually set for each subject’s right and left legs to generate a sustained, maximum force

contraction. Stimulation was applied to one leg for 90 seconds, during which time the generated

isometric force was sampled at 1 kHz using a pancake load cell (Honeywell Sensotec, USA).

Following 10 minutes of rest, the other leg was stimulated in the same manner. A polynomial

fitting procedure was used to approximate each force curve. Each curve was then normalized

between 0 and 1, with 1 representing no fatigue, and 0 representing no force generation.

Spasm and Tremor Data

The muscle spasm and tremor waveforms were extracted from a different data set that was

collected from one subject. This subject underwent three one-hour electrical stimulation-based

muscle strengthening sessions per week for eight weeks prior to data collection, to increase the

fatigue resistance of the muscles. During data collection, the subject was seated with the hip

flexed and the knee free to swing. The quadriceps muscle group was electrically stimulated

in the same manner as the fatigue experiment described above, however isotonic contractions

were generated to elicit knee extension against gravity. In this experiment, the stimulation

amplitude was randomized for each trial between 0 mA and 95 mA, which elicited full knee

extension in the participant. Stimulation was applied to the quadriceps muscle group of one

leg for 5 seconds, by which time the knee angle had reached its steady state behaviour in each

trial. The stimulation was followed by 5 seconds of rest, during which time the knee returned

to its resting flexed position.

The knee angle was sampled at 100 Hz during stimulation using a goniometer (TDS130A,

Biopac Systems Inc., USA), and the data were processed using a three-point moving average
Chapter 5. Construction of Non-Idealities Model 55

filter. The data were further processed to remove spurious data points, which were defined to

be any data points for which the knee angular velocity exceeded 360 deg/s. This threshold

corresponded to the average velocity of the knee during normal walking in able-bodied indi-

viduals [97]. Since the experimental subject exhibited significant muscle atrophy even after

extensive muscle strengthening, it is unlikely that the subject’s knee velocity would exceed this

threshold. After identifying those trials that exhibited muscle spasms or tremors, the spasm or

tremor behaviour was extracted from the affected trials. The resulting spasm waveforms were

zero except where spasms were present, and the tremor waveforms were centred about zero

(i.e., the average value of the tremor signal was zero).

There is no standard metric for describing the severity of undesirable behaviour in the stim-

ulated muscle response. Therefore, each non-ideality waveform was classified as mild, moderate,

or severe. The classification was done by two independent raters based on a set of classification

criteria for each type of non-ideality. These criteria were established by an expert who had

extensive experience working with electrically induced muscle contractions in SCI subjects, and

are described in Appendix A. The raters were presented with all the fatigue waveforms, followed

by the spasm waveforms, and then the tremor waveforms. The order of the waveforms within

each type of non-linearity was initially randomized, but was then held constant for each rating

session. The rating process was repeated one week later by the first rater to assess intra-rater

reliability. The order of the waveforms was randomized again before the second set of ratings.

The classification for each non-ideality waveform was taken to be the classification assigned by

the first rater in the first rating session. Examples of fatigue, spasm, and tremor non-ideality

waveforms are given in Figure 5.1.

The non-idealities block was programmed into a Simulink S-function (The Mathworks,

USA). The equation of the block is

vQU ADS (t) = (1 + s(t) + r(t)) τQU ADS (t) f at(t), (5.1)

where vQU ADS (t) is the output of the non-idealities block. τQU ADS (t), the torque generated

by the stimulated quadriceps contractions about the knee, is the input to the non-idealities

block. s(t) represents the scaled torque resulting from a muscle spasm, r(t) represents the
Chapter 5. Construction of Non-Idealities Model 56

Normalized Fatigue Modifier

Normalized Fatigue Modifier

Normalized Fatigue Modifier


Mild Fatigue Moderate Fatigue Severe Fatigue
1 1 1

0.5 0.5 0.5

0 0 0
0 10 20 30 0 10 20 30 0 10 20 30
Time (s) Time (s) Time (s)
Mild Spasms Moderate Spasms Severe Spasms
Knee Angle (deg)

Knee Angle (deg)

Knee Angle (deg)


40 40 40
30 30 30
20 20 20
10 10 10
0 0 0
0 10 20 30 0 10 20 30 0 10 20 30
Time (s) Time (s) Time (s)
Mild Tremors Moderate Tremors Severe Tremors
10 10 10
Knee Angle (deg)

Knee Angle (deg)

Knee Angle (deg)


5 5 5

0 0 0

−5 −5 −5

−10 −10 −10


0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
Time (s) Time (s) Time (s)

Figure 5.1: Examples of fatigue, spasm, and tremor non-ideality waveforms.

scaled torque resulting from a muscle tremor, and f at(t) represents a fatigue waveform. The

user can separately specify the severity of the fatigue, spasm, and tremor non-idealities that

are included in the block. The non-idealities block algorithm randomly selects the actual

constituent non-ideality waveforms at run-time from the pool of available waveforms having

the desired severity level for each type of non-ideality. The algorithm also provides the ability

to set the randomization seed prior to running the simulation, so that the results generated

during any particular simulation session can be re-created at a later date.

5.2.2 Implementation Example for Non-Idealities Block

An implementation example was included in this study to illustrate how the non-idealities

block can be used in simulations of FES applications. The example is seated knee extension

against gravity via electrically stimulated quadriceps contractions, and is based on the model of

stimulated knee angle proposed by Ferrarin and Pedotti in [82]. In this model, the knee angular
Chapter 5. Construction of Non-Idealities Model 57

acceleration φ̈(t) is given by

1   π  
φ̈(t) = −mgℓ sin φ(t) − λe−E(φ(t)+π/2) φ(t) + − ω − B φ̇(t) + τQU ADS (t) ,
JSHAN K 2
(5.2)

where φ(t) is the knee angle, φ̇(t) is the knee angular velocity, and τQU ADS (t), the electrically

elicited quadriceps torque about the knee, is the input to the model. JSHAN K is the moment

of inertia of the shank and foot, m is the mass of the shank and foot, ℓ is the distance from the

centre of rotation of the knee to the centre of mass of the shank and foot, λ and E are coefficients

of the exponential term, ω is the resting knee angle, and B is the damping coefficient. The

relationship between the pulse width of stimulation P WQU ADS (t) and the resulting quadriceps

torque is given in the frequency domain as

G
τQU ADS (s) = P WQU ADS (s), (5.3)
1 + ηs

where G and η are constants of the muscle activation function.

When the non-idealities block is incorporated into the model of stimulated knee movement

in Equation 5.2, the model becomes

1   π  
φ̈(t) = −mgℓ sin φ(t) − λe−E(φ(t)+π/2) φ(t) + − ω − B φ̇(t) + vSHAN K (t) ,
JSHAN K 2

where the non-idealities block modifies the nominal stimulated knee torque in Equation 5.3 to

reflect real-world undesirable behaviour including fatigue, muscle spasms, and tremors as shown

in Equation 5.1.

5.2.3 Control Design

Two different control algorithms were applied to this simulated FES system. Figure 5.2 shows

a block diagram of the non-idealities block implemented in the knee simulation.

The first controller was a PID controller with an anti-windup feature to prevent accu-

mulation of the integral term, as described in Section 3.1. PID control was included in this

implementation example despite its drawbacks because many researchers are familiar with this

method, and some in the FES community have been exploring its potential applications. The
Chapter 5. Construction of Non-Idealities Model 58

Figure 5.2: Block diagram of non-idealities block implemented in simulation of electrically

stimulated knee extension against gravity.

PID controller was tuned using trial and error to yield the most favourable unit step response

for the nominal muscle response, that is, with no non-idealities included in the simulation.

The second controller was a sliding mode controller (SMC) with a boundary layer to prevent

chattering in the control. SMC involves defining a control law that causes the state of the

system to converge toward a chosen sliding manifold in finite time, and then evolve along the

sliding manifold toward a goal state in possibly infinite time [92, 93]. This algorithm provides

good tracking performance, is inherently adaptive, requires that a relatively small number of

parameters be tuned, and guarantees stability, provided that an accurate plant model is used

to develop the control algorithm. SMC has also been used for regulating stimulation in FES

applications [78, 79, 84].

However, SMC is susceptible to “chattering”, that is, switching rapidly back and forth

between different control actions. Chattering can excite high frequency system dynamics and

result in controller failure [93]. To address this problem, a boundary layer was introduced into

the SMC. Within a small region encompassing the manifold, the sliding mode control law was

modified to a law that was not susceptible to chattering. The drawback to boundary layer

control is that the convergent behaviour of the standard SMC is no longer guaranteed.

The goal state was defined to be to be origin of the error plane, that is, zero position error

and zero velocity error. The position error is e1 (t) = φREF (t) − φ(t) and the velocity error is

e2 (t) = φ̇REF (t) − φ̇(t), where φREF (t) and φ̇REF (t) are the knee angle reference trajectory and

the knee angular velocity reference trajectory, respectively. The sliding manifold is defined to

be a line in R2 -space that passes through the origin of the error plane, denoted

m(t) = e2 (t) + Λe1 (t), Λ > 0,


Chapter 5. Construction of Non-Idealities Model 59

where Λ is a constant. Choosing the control law

ṁ(t) = −k1 sgn [m(t)] , k1 > 0

causes the error dynamics to approach the manifold in finite time [91]. Once the error dynamics

are on the manifold, the choice of manifold guarantees that e1 (t) → 0 as t → ∞.

Since this control law can cause chattering as the error dynamics near the manifold, the

control law within a boundary region of the manifold was chosen to be

ṁ(t) = −k2 m(t), k2 > 0.

Therefore, the overall sliding mode control law is



 −k1 sgn [m(t)] ∀ d > κ

ṁ(t) =
 −k2 m(t)

∀d≤κ

for some κ > 0, where d is the shortest distance between the current error (e1 (t), e2 (t)) and the

manifold. Equating the overall control law and the derivative of the sliding manifold
 
ṁ(t) = ė2 (t) + Λė1 (t) = φ̈REF (t) − φ̈(t) + Λ φ̇REF (t) − φ̇(t) ,

and re-arranging to find an expression for φ̈(t) yields


  
 φ̈REF (t) + Λ φ̇REF (t) − φ̇(t) + k1 sgn [m(t)] ∀ d > κ

φ̈(t) =  
 φ̈REF (t) + Λ φ̇REF (t) − φ̇(t) + k2 m(t)

∀ d ≤ κ.

An expression for the stimulated quadriceps torque τQU ADS (t) can be found by substituting

φ̈(t) into the state space equation of the knee model, yielding

−E(φ(t)+π/2) φ(t) + π − ω + B φ̇(t)

mgℓ sin φ(t) + λe




 2

    
+JSHAN K φ̈REF (t) + Λ φ̇REF (t) − φ̇(t) + k1 sgn [m(t)] ∀d>κ


τQU ADS (t) =
mgℓ sin φ(t) + λe−E(φ(t)+π/2) φ(t) + π2 − ω + B φ̇(t)
 





    
+JSHAN K φ̈REF (t) + Λ φ̇REF (t) − φ̇(t) + k2 m(t) ∀d≤κ

The values of parameters k1 , k2 , Λ, and κ were tuned by trial and error so that the evolution

of the errors in the error plane followed the sliding mode ideal of converging towards and then

evolving along the sliding manifold as closely as possible.


Chapter 5. Construction of Non-Idealities Model 60

5.2.4 Assessment of Controller Performance

The performance of the controllers was assessed through two tracking tasks. The first reference

trajectory was a 60 degree unit step knee extension beginning from the rest position (i.e.,

approximately 5 degrees knee extension). The second reference trajectory was a sinusoid that

mimicked the knee’s movements during normal walking in able-bodied individuals [98]. Figure

5.5 shows a sample of this trajectory.

For each simulation run, unit step response metrics were collected: 10%-90% rise time,

2% settling time, steady-state error, percent overshoot, and disturbance rejection time for a

10 degree disturbance lasting 0.3 s. RMS error and lag for the walking trajectory were also

collected. Simulations were run for the nominal response case as well as for every combination

of severity of the fatigue, muscle spasm, and tremor non-idealities. Both control algorithms

were tested for each simulation case.

5.3 Results

Table 5.1 lists the inter- and intra-rater reliability for the classification of the fatigue, muscle

spasm, and tremor non-linearity waveforms. Table 5.2 lists the performance metrics for the

knee control example for the PID controller. The nominal case with no non-idealities included

is listed first, followed by the fatigue only, spasm only, and tremor only cases. Last, selected

cases with fatigue, spasm, and tremor non-idealities are listed. Table 5.3 lists the same results

for the SMC.

Table 5.1: Intra- and inter-rater reliability for rating of non-ideality waveforms.
Fatigue Spasms Tremors

Inter-rater reliability 1.00 0.93 0.82

Intra-rater reliability for rater #1 0.90 0.93 0.72

Figure 5.3(a) shows an excerpt of the step response of the SMC for the nominal case. Figure

5.3(b) shows the evolution of the error trajectory in R2 -space toward the goal state (i.e., zero

position error and zero velocity error) for the same excerpt. Figure 5.4(a) shows the step
Chapter 5. Construction of Non-Idealities Model 61

Table 5.2: Performance metrics for PID control of knee angle. Blank spaces indicate the metric

could not be calculated for that case.


trise tsettling tdr ess lag eRMS
Fatigue Spasms Tremors % OS
(s) (s) (s) (deg) (s) (deg)

– – – 1.87 7.38 8.90 20.34 0.23 0.89 24.17

mild – – 1.98 8.76 20.93 0.85 24.11

mod. – – 2.66 2.59 0.79 23.13

severe – – -48.17 -32.14 0.86 23.50

– mild – 1.87 7.38 8.91 20.35 0.23 0.89 24.17

– mod. – 1.87 7.48 9.44 20.95 0.15 0.89 24.17

– severe – 1.87 7.38 8.90 20.34 0.23 0.90 24.17

– – mild 1.86 22.85 -3.73 0.65 24.19

– – mod. 1.86 30.91 0.87 24.18

– – severe 1.84 14.41 24.59 0.71 24.30

mild mild mild 2.01 12.71 28.87 0.86 24.11

mod. mod. mod. 2.75 2.26 0.59 22.91

severe severe severe -38.78 27.43


Chapter 5. Construction of Non-Idealities Model 62

Table 5.3: Performance metrics for sliding mode control of knee angle. Blank spaces indicate

the metric could not be calculated for that case.


trise tsettling tdr ess lag eRMS
Fatigue Spasms Tremors % OS
(s) (s) (s) (deg) (s) (deg)

– – – 2.59 -3.28 -1.95 0.19 14.50

mild – – 2.61 -3.44 -2.35 0.23 16.18

mod. – – 2.68 -3.61 -3.25 0.30 17.24

severe – – -47.71 -32.28 0.34 21.62

– mild – 2.58 -3.28 -2.10 0.18 15.34

– mod. – 2.79 -3.15 -2.07 0.19 14.78

– severe – 2.59 -3.28 -1.90 0.19 14.42

– – mild 2.61 -3.24 -2.19 0.16 15.39

– – mod. 2.66 -3.25 -2.30 0.40 13.85

– – severe 2.53 -3.07 -1.72 0.65 15.24

mild mild mild 2.57 -3.40 -2.42 0.15 17.69

mod. mod. mod. 2.82 -9.07 0.24 21.28

severe severe severe -22.59 0.54 22.77


Chapter 5. Construction of Non-Idealities Model 63

response of the PID controller for the nominal case as well as the case with mild fatigue, mild

spasm, and mild tremor non-idealities included. Figure 5.4(b) shows the same results for the

SMC. Figure 5.5 shows the performance of the SMC for the walking trajectory for the same

cases that are shown in Figure 5.4. The decrease in knee angle shown in Figure 5.5 at t = 4.5

s is due to the slight buckling of the knee that occurs when the body weight is transferred to

the leg.

(a)
60
Knee Angle (deg)

40

20

0
0 5 10 15
Time (s)
(b)
0.4
Velocity Error (deg/s)

0.2
t = 15 s t = 3.5 s t=2s
0

−0.2

−0.4
−10 0 10 20 30 40 50 60
Position Error (deg)

Figure 5.3: Evolution of error trajectory in R2 -space, for step response of SMC tested with

nominal knee model.

Figure 5.6 illustrates the relative magnitudes of the RMS error for the SMC, where the

diameter of each marker is proportional to the magnitude of the RMS error. The corners of the

solution space are shown, that is, (fatigue severity, spasm severity, tremor severity) = {(0,0,0),

(0,0,3), ..., (3,0,3), (3,3,3)} where 0 corresponds to no non-ideality included, 1 corresponds to

mild non-ideality included, 2 corresponds to moderate non-ideality included, and 3 corresponds

to severe non-ideality included. The cases for mild fatigue, moderate fatigue, severe fatigue,

mild spasms, moderate spasms, severe spasms, mild tremors, moderate tremors, and severe

tremors are also shown. Figure 5.7 shows a two-dimensional version of the data in Figure 5.6.
Chapter 5. Construction of Non-Idealities Model 64

(a)
80
Knee Angle (deg)

60

40

20

0
0 5 10 15 20 25
Time (s)
(b)
80
Knee Angle (deg)

60

40

20

0
0 5 10 15 20 25
Time (s)

Figure 5.4: Step response of (a) PID and (b) sliding mode controllers. Thin solid line represents

reference trajectory, dashed line represents response for nominal knee model, and thick solid

line represents response for knee model with mild fatigue, mild spasm, and mild tremor non-

idealities included. A disturbance is introduced at t = 15 seconds, as described in Section

5.2.4.
Chapter 5. Construction of Non-Idealities Model 65

70

60

50
Knee Angle (deg)

40

30

20

10

0
0 1 2 3 4 5 6 7 8 9 10
Time (s)

Figure 5.5: Response of SMC for walking trajectory. Thin solid line represents reference tra-

jectory, dashed line represents response for nominal knee model, and thick solid line represents

response for knee model with mild fatigue, mild spasm, and mild tremor non-idealities included.
Chapter 5. Construction of Non-Idealities Model 66

3
0.75
2.5

Normalized RMS Error


Tremor Severity

1.5
0.5
1

0.5

0 0.25
3
3
2 2.5
2
1 1.5
1
0.5 0
0 0
Spasm Severity Fatigue Severity

Figure 5.6: RMS error of SMC for knee model with various non-ideality combinations. RMS

error values at corners of non-ideality space are shown, as well as at intermediate positions along

axes of non-ideality space. Mild non-ideality is represented by 1 on the fatigue, spasm, and

tremor axes, 2 corresponds to moderate non-ideality, and 3 corresponds to severe non-ideality.

The case where no non-idealities are included is represented by 0 on each axis. Diameter of

marker is proportional to RMS error.


Chapter 5. Construction of Non-Idealities Model 67

Fatigue Severity = 0 Fatigue Severity = 1


3 3

2.5 2.5

Tremor Severity

Tremor Severity
2 2

1.5 1.5

1 1

0.5 0.5

0 0
3 2 1 0 3 2 1 0
Spasm Severity Spasm Severity

Fatigue Severity = 2 Fatigue Severity = 3


3 3

2.5 2.5
Tremor Severity

Tremor Severity
2 2

1.5 1.5

1 1

0.5 0.5

0 0
3 2 1 0 3 2 1 0
Spasm Severity Spasm Severity

Figure 5.7: Two-dimensional plot of data contained in Figure 5.6.

5.4 Discussion

The inter-rater reliability was greater than 0.8 for all types of non-ideality waveform, as shown

in Table 5.1, indicating that there was a high degree of agreement between the raters [100]. The

intra-rater reliability also showed good agreement for the fatigue and spasm non-idealities, but

not for the tremor non-ideality. This discrepancy was likely due to the nature of the tremors,

which can be difficult to differentiate from the underlying knee movement for large extension

movements. Overall, the rating method appears to have been reliable and repeatable.

The data in Table 5.2 show that the PID controller did not perform well at regulating the

electrically stimulated knee movements. The performance metrics are acceptable for the nomi-

nal non-idealities model, and those models with only muscle spasms included yield acceptable

results as well. In fact, the muscle spasms do not appear to have a large effect on the perfor-

mance of the PID controller. However, the cases that include tremor or fatigue non-idealities

in the simulation show degraded performance in Table 5.2, including many cases for which

performance metrics could not be calculated. For example, settling time cannot be calculated

if the response does not settle to within 2% of reference trajectory. The inclusion of tremors in
Chapter 5. Construction of Non-Idealities Model 68

the simulation prevented the system from settling and also greatly increased the disturbance

rejection time. Including fatigue in the simulation had the strongest negative effect on PID

controller performance, and resulted in large undershoot and an inability to settle or reject

disturbances. The last three cases in Table 5.2, which list the results for cases that include

all three types of non-ideality, show that the effects of fatigue dominated the performance, as

shown in Figure 5.4(a).

The performance of the SMC for the nominal case was fairly good, with a small overshoot,

steady-state error, lag, and RMS error, as shown in Table 5.3. Figure 5.3(b) shows how the

error trajectory converges toward the goal state (that is, zero position error and zero velocity

error). However, the error state does not actually reach the goal state, resulting in a non-zero

steady-state error, as reflected in Table 5.3. This result was not expected for a SMC, but may

have been due to the boundary layer compromising the typically convergent behaviour of SMC.

The SMC cases that include non-idealities in the knee model exhibited poorer performance

than the nominal case, including cases for which the values of the performance metrics could not

be determined. Again, muscle spasms did not have a strong effect on the SMC performance.

Tremors degraded the SMC performance slightly, and increased both the steady-state and

RMS errors. Fatigue also degraded the SMC performance, although this effect was not strong

except for the case that included severe fatigue. The responses shown in Figures 5.4(b) and 5.5

support the results shown in Table 5.3, with a reasonably good nominal response and a poorer

response for the case with mild fatigue, mild spasms, and mild tremors. It is possible that

the performance of the SMC in the non-ideal cases was compromised because the SMC was

based on the nominal non-idealities model, which was inaccurate for the cases that do include

non-linearities.

The performance of the PID and SMC for the nominal case could lead an investigator to

conclude that these FES controllers are ready to be tested with human subjects. However,

the performance of the controllers in the non-ideal cases indicates that human subject testing

would likely be unsuccessful. This difference illustrates the value of the non-idealities block for

refining control designs in simulation, thereby preventing premature testing in human subjects.

Figure 5.6 shows how the non-idealities model can be used to examine the effects of different
Chapter 5. Construction of Non-Idealities Model 69

parameters on the behaviour of the controlled system, and to elucidate potential sources of

performance problems. For example, Figure 5.6 shows graphically that RMS error increases

more quickly with increasing fatigue severity than with increasing tremor severity, in the absence

of other non-linearities. This insight suggests that further investigation into the behaviour of

the controller in the presence of fatigue may be warranted.

Interestingly, the inclusion of severe spasticity seemed to slightly improve the performance of

both controllers, in the absence of other non-linearities. This phenomenon was likely due to the

spasticity effectively increasing the stiffness of the knee joint, thereby making the knee dynamics

more similar to those of an able-bodied individual’s knee. This behaviour suggests that methods

of increasing the joint stiffness toward those values found in able-bodied individuals is a potential

strategy that may improve control of FES systems. This observation is supported by Vette et

al. [19], who stated that up to 90% of the torque needed to regulate quiet standing in able-bodied

individuals may be due to passive stiffness and damping of the ankle.

No recovery was implemented in the non-idealities block, even though stimulated muscles

do exhibit some recovery from fatigue during periods of no stimulation. This choice was made

because there is little information in the literature concerning rates of short-term recovery for

muscle contraction induced using electrical stimulation. The lack of recovery means that the

non-idealities block predicts a worst-case scenario. Additionally, this model uses fatigue data

recorded from isometric muscle contractions in untrained subjects, and applies that fatigue

to simulating isotonic muscle contractions. The non-idealities model could be improved by

substituting fatigue data collected during isotonic muscle contractions in trained subjects. The

non-idealities model could also be expanded to other joints by augmenting the model with

additional data.

5.5 Conclusions

The non-idealities block presented in this paper allows actual undesirable behaviour observed

during electrically stimulated muscle contractions in SCI individuals to be included in simula-

tions of FES applications. The resulting realistic simulations provide a more accurate assess-
Chapter 5. Construction of Non-Idealities Model 70

ment of how the controller might perform in the presence of these non-ideal behaviours than

is possible with currently used FES simulations methods. The non-idealities block provides

a convenient method of comprehensively testing FES systems in simulation prior to human

subject testing, thereby saving time and money. The implementation example shows how the

non-idealities block can be included in a simulation of regulating electrically stimulated knee

movement, and illustrates that the real-world performance of such a system may be vastly dif-

ferent from the nominal performance. The MatLab code for this project will be made freely

available for the use of others in the FES community in early 2011 at http://www.toronto-fes.ca

under the Products heading.


Chapter 6

Observer-Based Sliding Mode

Control of Quiet Standing

6.1 Introduction

Few FES systems are currently available for community-based use by individuals who have SCI.

One factor contributing to this scarcity of available systems is the general inability of currently

available closed-loop FES control systems to adequately regulate electrically stimulated muscle

contractions in the presence of real-world undesirable behaviour such as fatigue, muscle spasms,

and tremors.

In this study, a novel observer-based sliding mode control (SMC) algorithm for FES ap-

plications was proposed that addresses three challenges that complicate the use of current

closed-loop control algorithms in real-world FES applications. First, the response of muscle to

electrical stimulation is inherently nonlinear and time-varying. Despite this fact, many FES ap-

plications use linear closed-loop control strategies [19, 65, 68, 69, 101]. This approach may result

in sub-optimal control performance, even when the nonlinear system is linearized about some

operating point, because the parameters of the system are constantly changing [39,41,44]. This

study demonstrates that a nonlinear algorithm such as SMC can generate superior performance

in a simulated FES application as compared to a standard linear PID control algorithm.

Second, current FES technology results in rapid muscle fatigue due to the high stimulation

71
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 72

frequency that is necessary to induce artificial muscle contractions [55]. Despite this fact, most

control algorithms for FES applications do not explicitly address fatigue minimization. This

study included a fatigue minimization objective in the control algorithm, and showed that the

inclusion of this objective improved the performance of the SMC.

Third, individuals with SCI commonly exhibit reduced muscle bulk due to disuse atro-

phy in the paralyzed muscles, which results in reduced joint stiffness compared to able-bodied

individuals. This reduction in joint stiffness results in an under-damped system in which elec-

trically stimulated contractions can be difficult to control. The control strategy described in

this study used co-contraction of opposing muscle groups to artificially increase joint stiffness,

which resulted in an improvement in the performance of the SMC.

The test bed used to examine the behaviour of the SMC was a simulation of electrically

induced quiet standing in individuals with SCI. Quiet standing in SCI individuals is a commonly

used test bed for evaluating closed-loop control algorithms for FES applications [2, 16, 18, 68,

69, 101]. The non-idealities block [44, 46] described in Chapter 5 was included in the simulation

to evaluate the potential real-world performance of the SMC. This non-idealities block modifies

the behaviour of the simulated system to include actual muscle fatigue, spasms, and tremors

observed in experiments with individuals who have SCI.

Section 6.2 provides background information on modeling FES-induced quiet standing in

individuals with SCI. The methods used to design and implement the sliding mode and PID

control algorithms, as well as the quiet standing test bed are described in Section 6.3. This

section also describes the testing methodology that was used to assess the performance of

the controllers in simulation. The results of this study are presented in Section 6.4, and are

discussed in Section 6.5. Section 6.6 contains the conclusions of this study.

6.2 Background

6.2.1 Inverted Pendulum Model of Quiet Standing in SCI Individuals

During quiet standing in able-bodied individuals, the body is maintained at a slight anterior

lean of approximately 3 degrees [102]. This stance results in a gravitational force acting on
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 73

the body, and requires that the body exerts a torque about the ankles in opposition to gravity

to maintain stable standing. This torque is provided by the soleus and gastrocnemius muscles

of the calf, which contract to cause plantarflexion of the ankle. The resulting ground reaction

force opposes the force of gravity on the body. The tibialis anterior muscles of the calf are used

to oppose any disturbances that act in the posterior direction and return the body to a slight

anterior lean via dorsiflexion of the ankle.

Quiet standing can be induced in SCI individuals by electrically stimulating the plantarflexor

and dorsiflexor muscles to exert torques about the ankle. This strategy requires that the knees,

hips, and trunk are fixed in an extended position using voluntary effort, bracing, electrical

stimulation, or some combination of these methods. In this configuration, quiet standing can

be modeled as a single-link inverted pendulum, assuming that the two legs are lumped as one,

as shown in Figure 6.1. If the mass of the body is assumed to be concentrated in a point mass

at the end of the pendulum, the sway angle of the body as a function of the plantarflexor and

dorsiflexor torques can be expressed as

J θ̈(t) = τP LAN T AR (t) − τDORSI (t) + τST IF F (t) + τDAM P (t) − τGRAV IT Y (t),

where θ(t) is the sway angle of the body, θ̈(t) is the angular acceleration of the body about the

ankle, J is the moment of inertia of the body about the ankle joint, τP LAN T AR (t) is the total

torque exerted about the ankle by the plantarflexor muscles, and τDORSI (t) is the total torque

exerted about the ankle by the dorsiflexor muscles. τST IF F (t) and τDAM P (t) are the torques

about the ankle due to the stiffness and damping properties of the ankle, respectively, and can

be expressed as

τST IF F (t) = −k θ(t)

τDAM P (t) = −B θ̇(t)

where θ̇(t) is the angular velocity of the body about the ankle, and k and B are the coefficients

of stiffness and damping, respectively. τGRAV IT Y (t) is the gravitational torque about the ankle,

which is given by

τGRAV IT Y = mgℓ sin θ(t),


Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 74

where m is the point mass of the body, g is the gravitational constant, and ℓ is the length of

the pendulum (that is, the distance from the ground to the centre of mass of the body). If

the sway angle of the body is assumed to be between -5 and +10 degrees at all times, which

encompasses the range of angles for which stable standing can be maintained in able-bodied

individuals, then the gravitational torque can be linearly approximated as

τGRAV IT Y (t) ∼
= mgℓ θ(t).

Under these assumptions, the dynamics of quiet standing are given by

1
θ̈(t) = (τP LAN T AR (t) − τDORSI (t) − k θ(t) − mgℓ θ(t)) . (6.1)
J

Figure 6.1: Inverted pendulum model of quiet standing. m is the point mass of the body, ℓ is

the length of the pendulum, and θ is the sway angle of the body. The reference direction for

positive torque is shown as well.

6.2.2 FES Applications of Sliding Mode Control

SMC has shown promise in previous FES applications. Schauer, Holderbaum, and Hunt used

SMC to regulate electrically stimulated knee angle tracking in a seated subject [84]. Schauer

et al. tested their controller with one able-bodied subject, and used boundary layer control to
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 75

reduce chattering, replacing the switching function by a saturation function within the boundary

layer.

Jezernik, Wassink, and Keller also successfully used SMC for knee angle tracking in a

seated subject, but used a continuous control law instead of a discontinuous control law over

the entire error state space to avoid the issue of chattering [78]. Other authors have raised

concerns that this choice of control law compromises the finite time convergence to the manifold

that is guaranteed by traditional sliding mode control [79]. Mohammed et al. examined the

performance of a higher-order SMC for the same system in simulation [103], but did not propose

a solution to the chattering problem. Mohammed et al. also used static co-contraction of the

antagonist muscle group to improve control performance.

Ajoudani and Erfanian proposed a neuro-sliding mode controller for knee angle tracking,

which they tested in both able-bodied and trained, complete SCI subjects [79]. This imple-

mentation used boundary layer control that replaced the switching function with a neural

network controller inside the boundary layer. This controller was very successful, but the au-

thors acknowledged in [80] that this approach was too computationally intensive for real-world

multi-actuator or multi-joint applications. Kobravi and Erfanian reported a simpler approach

in [80], which used a SMC to regulate unloaded ankle movements in a seated subject and was

tested in able-bodied and trained, complete SCI subjects. In this case, an adaptive strategy

was used to avoid chattering. Kobravi and Erfanian also used a co-activation strategy in which

the activity of the antagonist muscle group was decreased as the activity of the agonist group

increased.

These controllers showed good performance, but most relied upon boundary layer control

to solve the chattering problem. The major drawback of boundary layer control is that the

advantageous features of SMC are lost, especially if the boundary layer is wide. Unfortunately,

the boundary region must often be relatively large when a system subject to large disturbances,

such as electrically stimulated muscle contractions, is considered. Therefore, these controllers

may suffer a degradation in performance if they are implemented in real-world FES applications.

In this study, a novel controller for FES applications is proposed that uses an observer-based

SMC to address the chattering problem without the need for a boundary layer control strategy.
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 76

6.3 Methods

6.3.1 Overview of Quiet Standing Simulation

The inverted pendulum model of quiet standing described in Section 6.2.1 was incorporated into

a simulation of electrically induced quiet standing in an individual with SCI. The plantarflexor

and dorsiflexor torques, τP LAN T AR and τDORSI , were generated by electrically stimulated mus-

cle contractions. The stimulation pattern was assumed to be fixed, and only the amplitude of

stimulation was varied. The stimulated response of the two muscle groups is normally coupled,

since the action of the plantarflexor muscles can affect the length of the dorsiflexor muscles by

changing the joint angle, and vice versa. However, the sway angle of the body was assumed to be

limited to the small range of angles for which stable standing can be maintained in able-bodied

individuals. It was also assumed that the length of the plantarflexor and dorsiflexor muscles

would not change significantly over this small sway range, which allowed the coupling between

the stimulated response of the plantarflexor and dorsiflexor muscle groups to be neglected.

Under these assumptions, the stimulated torque about the ankle is a nonlinear function of

stimulation amplitude and is also time-varying due to the effects of muscle fatigue and day-

to-day variation in the stimulated response. There is also a fixed delay between the onset of

stimulation and the onset of the resulting muscle contraction [55]. Additionally, the system is

subject to perturbations due to muscle spasms and tremors, as well as external disturbances.

Lastly, the voltage delivered to the stimulator unit by the controller must be limited to the

nominal input range of 0 to 5 V to ensure predictable behaviour from the stimulator (Compex

Motion, Compex SA, Switzerland).

6.3.2 Implementation of Quiet Standing Simulation

The quiet standing simulation was implemented in Simulink (The Mathworks, Natick, MA,

USA). The values of the system parameters are given in Table 6.1, and a block diagram of the

simulation is shown in Figure 6.2.

The input to the stimulator was limited between 0 and 5 V using a hyperbolic tangent

function instead of a pure saturation function so that the saturation function was invertible.
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 77

Table 6.1: Parameters of quiet standing system.


Aspect of System Parameter Value Source

Inverted pendulum m 82.7 kg Experimental data

model ℓ 0.94 m (see Section 6.3.2)

J 73.1 kg m2

k 235 Nm/rad

B 5.16 Nm s/rad [104]

Muscle recruitment {c1 , c2 , c3 } {317, 0.09, 80} Experimental data

functions {c4 , c5 , c6 } {74, 0.15, 36} (see Section 6.3.2)

Objective function kREF 859 Nm/rad [102]

Figure 6.2: Block diagram of simulation of electrically induced quiet standing.


Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 78

The saturation function is given by

exp[2u] − 150
σ(u) = 2.5 + 2.5,
exp[2u] + 150

where u was the input signal to be limited. A balanced biphasic square wave stimulation pat-

tern with frequency 40 Hz and pulse width 250 µs was used; the stimulation amplitude was

varied by the controller. The torques produced by the electrically stimulated plantarflexor and

dorsiflexor muscles were described by muscle recruitment functions, which were implemented

as look-up tables. The experimental data contained in these tables was collected by the author

from a trained, complete SCI subject. The subject’s plantarflexor or dorsiflexor muscles were

stimulated with different constant stimulation amplitudes, and the resulting torque about the

ankle was recorded. This individual was also used to determine the values of the anatomical

parameters for the quiet standing model. ℓ was assumed to be 0.53 of height [104]. The activa-

tion delay for each muscle group was implemented by delaying the input into the recruitment

function look-up table by TACT IV AT ION = 40 ms [105]. The inverted pendulum dynamics are

given in Equation 6.1, and the sensor dynamics were implemented as a pure delay.

The control algorithms were developed using a simplified version of the simulated quiet

standing system that neglected the muscle activation delay. Moreover, the plantarflexor and

dorsiflexor muscle recruitment functions were approximated by

c1
ĥplantarf lexors (v) =
1 + exp [−c2 (v − c3 )]
c4
ĥdorsif lexors (v) = ,
1 + exp [−c5 (v − c6 )]

respectively. The coefficients of these approximate activation functions were determined empir-

ically to provide a best-fit to the experimental data in the recruitment function look-up tables.

The saturation nonlinearity and muscle recruitment dynamics were cancelled by pre-multiplying

by the inverses of these functions, which are given by


 
1 150 + 150 ((u − 2.5) /2.5)
σ −1
(u) = ln
2 1 − ((u − 2.5) /2.5)
   
−1 c1
ĥ−1
plantarf lexors (u) = ln − c2 c3
c2 u−1
   
−1 c4
ĥ−1
dorsif lexors (u) = ln − c5 c6 .
c5 u−1
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 79

The non-idealities block was used to evaluate the performance of the controllers in the presence

of real-world non-ideal muscle behaviour. This block was implemented using a Simulink S-

function, as described in [44, 46].

6.3.3 Control Design

The control problem that was addressed in this study is output-feedback tracking of a reference

sway angle trajectory. The tracking error e = θREF − θ drives the action of the controller,

which regulates the stimulation intensity delivered to the agonist muscle group. When e ≤ 0,

the plantarflexor muscle group is the agonist group and the dorsiflexor group is the antagonist

group. When e > 0, the dorsiflexor group is the agonist group and the plantarflexor muscle

group provides the antagonist torque.

The quiet standing dynamics given in Equation 6.1 can be expressed in the state space. Let

the system state be x = [x1 x2 ]T = [θ θ̇]T , the system input be u = τAGON IST , and the system

output be y = x1 = θ, where θ is the sway angle during standing and θ̇ is the angular velocity

of sway. Therefore, the state-space dynamics are

ẋ1 = x2
1
ẋ2 = (−kx1 − mgℓx1 − Bx2 + u − d) (6.2)
J
y = x1 ,

where the system parameters k, m, ℓ, B, and J are defined in Section 6.2.1 and their values

are given in Table 6.1. d = τAN T AGON IST is the torque due to the antagonist muscle group.

Since the muscle activation delay and time-varying effects due to fatigue, muscle spasms, and

tremors were not included in the model that was used to develop the controllers, the system in

Equation 6.2 was linear time-invariant and could be expressed in block form as

ẋ = Āx + B̄u + Ēd

y = C̄x,
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 80

where
     
0 1 0 0
 
Ā =  , B̄ =  , Ē =  , C̄ = .
     
1 0
(−k − mgℓ) /J −B/J 1/J −1/J

The antagonist torque d was treated as a disturbance in this formulation of the problem.

The stimulation intensity delivered to the antagonist muscle group was determined by an

optimization algorithm that minimized a fatigue minimization objective function and a stiffness

regulation objective function, subject to the agonist torque value determined by the controller

and the constraint sgn[d] = −sgn[u], where sgn denotes the signum function. The fatigue

minimization objective function was


 T  
1 u u  1 2
u + d2f atigue ,

gf atigue (u, df atigue ) =  I =
 
2 2

df atigue df atigue

where I was the identity matrix, and the stiffness regulation objective function was
2
J ẋ2 2
  
1 τT OT AL 1
gstif f ness (x, u, dstif f ness ) = kREF − = kREF −
2 x1 2 x1
 2
1 1
= kREF − (−kx1 − mgℓx1 − Bx2 + u − dstif f ness ) ,
2 x1

where kREF is the desired ankle stiffness, as given in Table 6.1 and τT OT AL is the total torque

about the ankle.

Minimizing the fatigue objective function for any value of u will result in df atigue = 0.

However, minimizing the stiffness function for a particular value of x and u may result in

a non-zero value of dstif f ness . To resolve this problem, d was taken to be a weighted linear

combination of the antagonist torque values yielded by the two objective functions, such that

d(x, u) = α df atigue + β dstif f ness ,

where α = 0.5 and β = 0.5.

A PID controller and an observer-based SMC were implemented in this study. PID is a

commonly used control algorithm that is simple and reliable, but commonly exhibits degraded

performance in the presence of unmodeled disturbances or plant dynamics. The PID algorithm

in this quiet standing simulation was implemented using an anti-windup design to prevent
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 81

Table 6.2: Controller and observer parameters.


Controller Muscle Group No Co-contraction With Co-contraction

PID Plantarflexors KP = 350, Td = 5, Ti = 10 KP = 200, Td = 5, Ti = 10

Dorsiflexors KP = 350, Td = 0.35, Ti = 0.6 KP = 350, Td = 0.35, Ti = 0.6

SMC Either group κ = 45, λ = 3 κ = 30, λ = 2.7

Observer L1 = 100, L2 = 100

accumulation of the integral term, as described in Section 3.1, and the controller was tuned

using trial and error. The PID controller parameter values are given in Table 6.2.

The observer-based SMC algorithm used a simple asymptotic observer to provide the state

estimate, since the nominal plant model that was used for control design was linear time-

invariant. The use of the observer avoided introducing chattering into the agonist torque

control signal, as described in Section 3.3.2, which could otherwise have been a serious problem

due to the unmodeled plant dynamics.

Let the observer state be x̂ = [x̂1 x̂2 ]T , such that the state-space equation of the observer is
 
 L1 
x̂˙ = Āx + B̄u +   (x1 − x̂1 ) ,
L2

where y = xs and L1 , L2 > 0 are the observer gains. The disturbance d due to the antagonist

muscle torque was not included in the observer equation, since the observer was unaware of

unmeasured disturbances. The observer dynamics were therefore given by

x̂˙ 1 = x̂2 + L1 (x1 − x̂1 )


1
x̂˙ 2 = (−kx̂1 − mgℓx̂1 − B x̂2 + u) + L2 (x1 − x̂1 ) . (6.3)
J

Next, a SMC was designed for which the sliding manifold was defined on the observer state-

space. Let the sliding manifold m(x̂, w) = 0 be given by

m(x̂, w) = e2 + λe1 = w2 − x̂2 + λ(w1 − x̂1 ),

where w = [θREF θ̇REF ]T . The sliding mode control law was then designed to be

u = −κ sgn [m(x̂, w)] = −κ sgn [w2 − x̂2 + λ (w1 − x̂1 )] ,


Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 82

where κ > 0, λ > 0. The observer-based SMC was tuned using trial and error. The parameter

values for the observer and controller are given in Table 6.2.

6.3.4 Assessment of Controller Performance

The performance of the PID and sliding mode controllers was assessed while tracking two

reference trajectories. The first trajectory was a step from an initial sway angle of 0 degrees

(i.e., the vertical position) to a final sway angle of 2.5 degrees in the anterior direction, with

the step occurring at 0.01 s. The step trajectory also included a disturbance at the 15 second

mark that consisted of a 1 degree anterior sway with a duration of 0.1 s. The total duration of

the step trajectory was 30 s.

The second reference trajectory was a sinusoidal trajectory that represented the small oscil-

lations in sway angle that are exhibited during quiet standing in able-bodied individuals [106].

The equation of this trajectory in degrees was

θREF (t) = 0.5 sin (0.5πt) + 2.25,

and its duration was 20 s.

The controllers were tested both with and without co-contraction of the antagonist muscle

group. In each case, the performance of the controllers was evaluated for the nominal case

with no non-idealities included, as well as for various combinations of non-ideal fatigue, muscle

spasm, and muscle tremor behaviour. Standard step response metrics (i.e., 10%-90% rise time,

2% settling time, percent overshoot, steady-state error, and disturbance rejection time) and

sinusoidal response metrics (i.e., lag and root-mean-squared error) were calculated for each

simulation case [88]. In addition, the sensitivity of the SMC controller to ±20% and ±50%

error in the parameters of the quiet standing dynamics (i.e., m, k, ℓ, B, and J) was assessed.

The value of each parameter was varied separately while the other parameters were held at

their nominal values, as listed in Table 6.1.


Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 83

6.4 Results

Table 6.3 lists the performance metrics for the PID and sliding mode controllers for the nominal

quiet standing model with no non-idealities included, both with and without co-contraction of

the antagonist muscle group. Table 6.4 lists the step response metrics for the SMC for the

quiet standing model including mild, moderate, and severe fatigue non-idealities, both with

and without co-contraction of the antagonist muscle group. Table 6.5 lists the step response

metrics for the SMC with spasm non-idealities, and Table 6.6 lists the same metrics for the

SMC with tremor non-idealities. Table 6.7 lists the step response metrics for the SMC for the

quiet standing model with fatigue, spasm, and tremor non-idealities included, both with and

without co-contraction. Blank spaces in Tables 6.3 through 6.7 indicate that the metric could

not be calculated. For example, the settling time can only be calculated if the response settles

to within 2% of the desired steady-state value during the trial. If the response does not reach

±2% of the desired steady-state value during the trial, then the corresponding entry in the

table is left blank.

Table 6.3: PID and SMC performance metrics for nominal quiet standing model with no non-

idealities included, where trise is 10%-90% rise time, tsettling is 2% settling time, tdr is distur-

bance rejection time, % OS is percent overshoot, ess is steady-state error, and eRMS is RMS

error.
trise tsettling tdr % ess lag eRMS
Controller
(s) (s) (s) OS (deg) (s) (deg)

PID 2.21 4.05 0.00 0.00 0.35 0.688

PID with co-contraction 1.71 3.15 0.15 0.00 0.00 0.35 0.458

SMC 0.33 0.78 0.60 0.00 -0.11 0.344

SMC with co-contraction 0.32 0.98 0.71 1.89 -0.06 0.04 0.286

Figure 6.3(a) shows an excerpt of the step response of the PID and sliding mode controllers

tested with the nominal quiet standing model (i.e., with no non-idealities included), and with

co-contraction of the antagonist muscle group. Figure 6.3(b) compares the evolution of the
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 84

Table 6.4: SMC step response metrics for quiet standing model with fatigue non-idealities.
Without co-contraction With co-contraction

Fatigue trise tset tdr % ess eRMS trise tset tdr % ess eRMS

(s) (s) (s) OS (deg) (deg) (s) (s) (s) OS (deg) (deg)

None 0.33 0.78 0.60 0.00 -0.093 0.315 0.32 0.98 0.71 1.89 -0.034 0.275

Mild 0.77 5.40 0.395 0.90 5.50 0.367

Moderate 0.73 4.70 0.378 0.87 4.10 0.338

Severe 0.62 1.40 0.441 0.75 1.90 0.344

Table 6.5: SMC step response metrics for quiet standing model with spasm non-idealities.
Without co-contraction With co-contraction

Spasm trise tset tdr % ess eRMS trise tset tdr % ess eRMS

(s) (s) (s) OS (deg) (deg) (s) (s) (s) OS (deg) (deg)

None 0.33 0.78 0.60 0.00 -0.093 0.315 0.32 0.98 0.71 1.89 -0.034 0.275

Mild 0.32 0.90 0.60 0.00 -0.092 0.321 0.33 1.70 0.71 -0.046 0.275

Moderate 0.33 0.75 0.30 0.00 -0.092 0.315 0.32 1.10 0.71 -0.029 0.281

Severe 0.32 0.78 0.56 0.00 -0.092 0.315 0.33 1.10 0.80 -0.006 0.275

Table 6.6: SMC step response metrics for quiet standing model with tremor non-idealities.
Without co-contraction With co-contraction

Tremor trise tset tdr % ess eRMS trise tset tdr % ess eRMS

(s) (s) (s) OS (deg) (deg) (s) (s) (s) OS (deg) (deg)

None 0.33 0.78 0.60 0.00 -0.093 0.315 0.32 0.98 0.71 1.89 -0.034 0.275

Mild 0.30 0.00 -0.092 0.315 0.28 0.60 0.70 4.05 -0.080 0.281

Moderate 0.35 3.96 0.332 0.36 4.69 -0.006 0.298

Severe 1.09 5.45 0.327 0.51 7.12 -0.069 0.298


Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 85

Table 6.7: SMC step response metrics for quiet standing model with fatigue, spasm, and tremor

non-idealities.
Without co-contraction With co-contraction

Severity trise tset tdr % ess eRMS trise tset tdr % ess eRMS

(s) (s) (s) OS (deg) (deg) (s) (s) (s) OS (deg) (deg)

None 0.33 0.78 0.60 0.00 -0.093 0.315 0.32 0.98 0.71 1.89 -0.034 0.275

Mild 0.89 6.00 0.361 0.71 0.332

Moderate 0.63 0.481 0.52 0.384

Severe 0.55 0.567 0.55 0.424

error trajectories in R2 -space toward the sliding manifold for the same excerpt.

Figure 6.4 shows excerpts of the step and sinusoidal responses of the PID and sliding mode

controllers for the nominal quiet standing model, both with and without co-contraction. Figure

6.5 shows excerpts of the step and sinusoidal responses of the SMC for the nominal quiet

standing model, as well as the case with mild fatigue non-ideality included in the quiet standing

model. The results both with and without co-contraction are shown.

Figure 6.6 shows the sensitivity of the SMC to a ±20% and ±50% mismatch in the model

parameters m, ℓ, k, B, and J for the nominal quiet standing model (i.e., with no non-idealities

included in the model) and with co-contraction of the antagonist muscle group.

6.5 Discussion

Figure 6.3(b) shows the convergence of the error trajectories of the PID and sliding mode

controllers toward the sliding manifold. Clearly, the SMC reaches steady-state behaviour more

quickly than the PID controller, as shown in Figure 6.3(a). Typically, the error trajectory of

SMC is expected to converge to the sliding manifold in finite time, which is not the case in this

example. The atypical steady-state oscillation displayed by the SMC is likely due to the large

unmodeled disturbance caused by the antagonist muscle contractions.

The results shown in Table 6.3 and Figure 6.4(a) and (b) show that the SMC performance is
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 86

(a)

3
Knee Angle (deg)

2.5
2
1.5
1 Ref.
0.5 PID
SMC
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s)
(b)
0.01
Velocity Error (deg/s)

PID

0 SMC
t = 0.05 s
t = 3.5 s
−0.01

−0.02

0 0.5 1 1.5 2 2.5


Position Error (deg)

Figure 6.3: Evolution of error trajectory in R2 -space, for step response of PID and sliding mode

controllers tested with nominal quiet standing model (i.e., no non-idealities) and co-contraction

of antagonist muscle group. (a) Step response. (b) Corresponding error trajectory, showing

starting and ending points. Sliding manifold is shown by heavy vertical solid line.
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 87

(a) (b)
3.5 3.5
3 3
Sway Angle (deg)

Sway Angle (deg)


2.5 2.5
2 2
1.5 1.5
1 Ref. 1 Ref.
0.5 PID 0.5 PID
SMC SMC
0 0
0 2 4 6 8 10 0 1 2 3 4 5 6 7
Time (s) Time (s)
(c) (d)
3.5 3.5
3 3
Sway Angle (deg)

Sway Angle (deg)

2.5 2.5
2 2
1.5 1.5
1 Ref. 1 Ref.
0.5 PID 0.5 PID
SMC SMC
0 0
0 2 4 6 8 10 0 1 2 3 4 5 6 7
Time (s) Time (s)

Figure 6.4: Step and sinusoidal response of PID and sliding mode controllers for nominal quiet

standing model (i.e., no non-idealities). (a) Step response without co-contraction. (b) Sinusoidal

response without co-contraction. (c) Step response with co-contraction. (d) Sinusoidal response

with co-contraction.
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 88

(a) (b)
3.5 3.5
3 3
Sway Angle (deg)

Sway Angle (deg)


2.5 2.5
2 2
1.5 1.5
1 Ref. 1 Ref.
0.5 Nominal 0.5 Nominal
Fatigue Fatigue
0 0
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
Time (s) Time (s)
(c) (d)
3.5 3.5
3 3
Sway Angle (deg)

Sway Angle (deg)

2.5 2.5
2 2
1.5 1.5
1 Ref. 1 Ref.
0.5 Nominal 0.5 Nominal
Fatigue Fatigue
0 0
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
Time (s) Time (s)

Figure 6.5: Step and sinusoidal response of SMC. (a) Step response without co-contraction.

(b) Sinusoidal response without co-contraction. (c) Step response with co-contraction. (d)

Sinusoidal response with co-contraction.


Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 89

1.5

+B +k
1
+l
+J
0.5

+m
0
−m

−0.5
−J
−l

−1 −B
−k

−1.5
−1.5 −1 −0.5 0 0.5 1 1.5

Figure 6.6: Sensitivity of SMC with co-contraction to model parameter mismatch for nominal

quiet standing model (i.e., no non-idealities). Markers represent normalized RMS error, i.e. the

ratio of RMS error with mismatch in the parameter associated with a particular chord to the

RMS error with nominal parameter values. Heavy solid line represents unit circle. Solid line

shows normalized RMS error for ±20% parameter mismatch. Dashed line shows normalized

RMS error for ±50% parameter mismatch.


Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 90

better than the PID performance for the nominal quiet standing model with no co-contraction

included. The SMC exhibits a faster response and less lag when tracking the sinusoidal tra-

jectory than the PID controller does. Moreover, the PID controller exhibits poor sinusoidal

tracking, and its response is too slow to be suitable for real-world use, even without the inclu-

sion of real-world non-linearities.

When co-contraction was added, the SMC controller performed even better, exhibiting fast

rise, settling, and disturbance rejection times, no overshoot or appreciable steady-state error

aside from a small steady-state ripple, and low RMS error. Figures 6.4(c) and (d) confirm

the numerical results listed in Table 6.3 for the co-contraction cases. The PID performance

improved slightly with co-contraction but is still unacceptable for use in practical FES applica-

tions. These results suggest that the SMC is a more suitable choice for real-world FES control

algorithms than PID control, despite the popularity of the latter technique among some FES

researchers, including some members of the Rehabilitation Engineering Laboratory at Toronto

Rehabilitation Institute.

Figures 6.4(b), 6.4(d), 6.5(b), and 6.5(d) show that the SMC exhibits less RMS error with

co-contraction than without. The results in Table 6.4 and Figure 6.5 show that including fatigue

non-idealities in the quiet standing model resulted in significantly slower rise and settling times

than the nominal case, even for mild fatigue. Moreover, co-contraction did not significantly

improve the SMC performance under fatigue conditions. The spasm non-linearities had minimal

effect of the SMC performance as compared to the nominal case, as shown in Table 6.5. This

result is logical, since SMC is known to be able to reject small disturbances such as those

associated with muscle spasms [31]. Table 6.6 illustrates that including severe tremor non-

idealities in the quiet standing model adversely affected SMC performance, but that mild and

moderate tremors were handled well by the SMC. Co-contraction of the antagonist muscle group

improved the rise time for the severe tremor case, but exacerbated the overshoot. This behaviour

indicates that the implementation of SMC that was used in this study is not able to reject large

steady-state model errors such as strong tremor non-idealities. Also, all combinations of non-

ideal behaviour resulted in fairly good performance both with and without co-contraction, as

shown in Table 6.7. However, it was not possible to calculate many of the metric values, most
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 91

likely due to the effect of the fatigue non-idealities.

The most beneficial effect of the co-contraction strategy was its ability to reduce RMS error.

Tables 6.4 through 6.7 show that co-contraction causes a marked improvement in RMS error

in every case.

Figure 6.6 shows that ±20% variation in the model parameters had little effect on the RMS

error of the SMC when co-contraction was included. This result confirms that the SMC is fairly

robust to parameter mismatch. This behaviour is expected, since SMC is known to be robust to

parameter mismatch [93]. The SMC was sensitive to large (i.e., −50%) mismatch in parameters

m and ℓ. This result suggested that particular attention should be paid to accurately identifying

these parameters.

This simulation work yielded exciting results, despite its limitations. Specifically, the single-

link inverted pendulum model of quiet standing is a simplification of actual standing. Quiet

standing involves two independently actuated legs, as well as contributions from head, arm,

torso, and foot movements. Moreover, the hips and knees are not locked during quiet standing,

and the torso and foot have many degrees of freedom, whereas the model assumes that they

are monolithic body parts. Also, the plant model that was used for controller development

neglected coupling between the agonist and antagonist muscle groups. Additionally, the non-

idealities block was developed using electrically stimulated seated knee extension against gravity

in SCI individuals, not quiet standing.

The work reported in this study makes three significant contributions to the state of the art

in FES control. First, it shows how including actual real-world non-ideal behaviour in the quiet

standing simulation permits a realistic analysis of the system’s likely real-world performance

in the presence of fatigue, muscle spasm, and tremor non-idealities. Second, the practicality of

observer-based SMC for FES applications is demonstrated, which is a more powerful technique

for handling SMC chattering than the standard boundary layer method. Third, the study

demonstrates the feasibility of dynamically adjusting the co-contraction torque of the antagonist

muscle group to minimize fatigue and track a desired joint stiffness value. Moreover, it is shown

that the performance of the SMC under this co-contraction algorithm is improved over the case

without co-contraction, particularly with respect to RMS error.


Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 92

This work also represents a significant step forward in the use of nonlinear control tech-

niques for real-world FES systems, since the techniques that were developed in this study for

quiet standing could easily be applied to other FES applications. In fact, the observer-based

SMC implementation with co-contraction that is reported in this study is suitable for certain

real-world FES applications, provided that the small steady-state oscillation exhibited by this

controller could be tolerated, and that severe non-ideal behaviour is not expected of the FES

system.

The applicability of this observer-based SMC to real-world FES systems could be further

improved by incorporating both the antagonist torque input d and a muscle fatigue model into

the quiet standing system model used for controller development. This change would likely

remove the steady-state oscillation that was observed in the SMC behaviour by allowing the

observer to provide a more accurate state estimate to the SMC. Also, FES systems in the

real-world are subject to disturbances due to non-idealities such as muscle spasms and tremors.

These disturbances are unknown bounded disturbances with bounded time derivatives, which

can be compensated by introducing a high-gain loop around the observer itself to reject this

disturbance [93]. This approach would likely improve the performance of the observer-based

SMC in the presence of muscle strong tremors.

6.6 Conclusions

A novel observer-based SMC algorithm for real-world FES applications was developed in this

study. This algorithm included a fatigue minimization strategy and used co-contraction of

the antagonist muscle group to artificially increase the ankle stiffness, thereby improving the

performance of the controller. This controller was evaluated in a simulation of electrically

induced quiet standing, and the performance of the SMC was tested in the presence of actual

non-ideal muscle fatigue, spasm, and tremor behaviour. The observer-based SMC performed

very well, even in the presence of non-ideal muscle behaviour, and was also fairly insensitive to

mismatch error in the quiet standing model parameters.

Future work on this project will include refining the design of the observer-based SMC to
Chapter 6. Observer-Based Sliding Mode Control of Quiet Standing 93

include the antagonist torque and a model of muscle fatigue in the quiet standing model, and

implementing a high-gain loop around the observer to reject steady-state disturbances due to

muscle tremors. Experimental testing of this control algorithm is also planned, to confirm its

simulated performance. These enhancements will further improve the utility of this controller

for real-world FES applications.


Chapter 7

Discussion

There are few commercial FES applications currently available for individuals who have a

neurological disability such as stroke or SCI. Those FES systems that are available use open-

loop or simple state-based control strategies, and cannot adjust the electrical stimulation on

a pulse-by-pulse basis to compensate for the nonlinear, time-varying nature of the response of

electrically stimulated muscle.

One of the major reasons for this dearth of sophisticated FES applications is the lack of

feasible closed-loop FES control algorithms. Closed-loop control is necessary for many innova-

tive applications of FES technology, such as walking over uneven ground, dynamic grasp, and

balance while standing, sitting, or wheeling. Closed-loop control allows the electrical stimula-

tion delivered to the muscles to be modulated in real-time to compensate for disturbances and

changes in the response of the stimulated muscles. The purpose of the work described in this

thesis was to develop a closed-loop control algorithm for practical FES applications.

7.1 Summary of Analysis of Existing Closed-Loop Control Al-

gorithms

Chapter 4 described how three existing closed-loop control algorithms for FES applications were

analyzed to examine how closed-loop FES controllers could be improved. These representative

controllers were tested in a simulation of seated knee extension against gravity, which was

94
Chapter 7. Discussion 95

based on the model of electrically stimulated knee extension proposed by Ferrarin and Pedotti.

This knee model was extended in Chapter 4 to include estimated muscle fatigue, spasm, and

re-training effects.

This simulation showed that the PID controller yielded the poorest performance of the three

controllers when nonlinear time-varying effects were included. The GSC performed relatively

well, but required extensive tuning due to its use of multiple local controllers. The SMC

exhibited acceptable performance, but suffered from non-zero steady-state error due to the

use of a boundary layer to avoid chattering in the control. All of the control algorithms were

sensitive to model mismatch error, and none of them performed well enough to be implemented

in a practical FES application. Overall, the results presented in Chapter 4 illustrated the need

to test control algorithms under realistic conditions to evaluate their potential performance in

the real world.

An interesting observation that resulted from the analysis of the existing closed-loop control

algorithms was that the inclusion of intrinsic tonic spasticity in the knee simulation improved

controller performance. Since the type of spasticity implemented in Chapter 4 caused co-

contraction of the antagonist muscle group, resulting in increased knee stiffness, this observation

suggested that deliberate co-contraction of the agonist and antagonist muscle groups could be

used to improve controller performance by increasing stiffness in the paralyzed joints.

7.2 Summary of Development of Non-Idealities Block

Chapter 5 described a new method of testing closed-loop FES control algorithms under realistic

conditions to assess their likely performance in the real world. A non-idealities block was de-

veloped that modified the nominal stimulated muscle response to include actual muscle fatigue,

spasm, and tremor behaviour observed during electrically stimulated muscle contractions in

SCI subjects.

This non-idealities block was tested in a simulation of electrically stimulated knee extension

against gravity with both a PID controller and a boundary-layer SMC. The results showed that

the PID controller and SMC yielded comparable performance for the nominal case. However,
Chapter 7. Discussion 96

when the non-idealities block was included in the simulation, the PID controller yielded terrible

performance, indicating that it would perform poorly in an actual FES application. The SMC

performed better than the PID controller when the non-idealities block was used, but not as

well as it performed in the absence of non-idealities.

The results presented in Chapter 5 showed that neither controller was appropriate for use in

real world FES applications, despite their misleadingly good performance in the nominal case.

These results illustrated the value of the non-idealities block for demonstrating the probable

performance of FES control algorithms in the presence of real-world muscle fatigue, spasms,

and tremors.

7.3 Summary of Observer-Based Sliding Mode Control Algo-

rithm

Chapter 6 discussed the development of a novel observer-based sliding mode control algorithm

that is suitable for use in real-world FES applications. This closed-loop control algorithm incor-

porated a fatigue minimization objective, as well as intentional co-contraction of the antagonist

muscle group to cause the joint stiffness to track a desired value. This controller was designed

to modulate the stimulation in real-time to compensate for changes in the stimulated mus-

cle response, to be robust to parameter variation (i.e., model mismatch error), and to reject

disturbances due to such factors as muscle spasms and tremors.

The observer-based SMC was tested in a simulation of electrically induced quiet standing

in an individual with SCI, and the non-idealities block was used to determine the probable

performance of the controller in the presence of actual non-ideal behaviour from the stimulated

muscles. The SMC performed relatively well for the non-ideal case, but showed markedly

improved performance when the fatigue minimization and stiffness tracking objectives were

employed to generate co-contraction of the antagonist muscle group.

On the basis of these result, the observer-based sliding mode controller proposed in Chapter

6 was deemed to be suitable for certain practical FES applications, provided that the small

steady-state ripple in the response could be tolerated, and that severe non-ideal behaviour was
Chapter 7. Discussion 97

not expected of the system.

7.4 Control Algorithms for FES Applications

Many research groups have pursued PID control for use in FES applications. The main reasons

for the FES community’s interest in PID control is its simplicity, ease of implementation and

tuning, and the high degree of familiarity that most researchers have with this type of con-

troller. However, the results presented in this thesis clearly demonstrate that PID control is an

unsuitable technique for practical FES applications. Until FES technology improves markedly

(i.e., new methods of muscle stimulation and muscle stimulation delivery are developed that will

allow asynchronous and single muscle motor unit control), PID control should not be considered

suitable for FES applications, because the probability of it successfully regulating electrically

stimulated muscle contractions in the real world is very low.

Modern control techniques such as GSC, SMC, neural networks, and fuzzy logic have shown

more promise for FES applications. GSC works acceptably well for FES systems, but it is

difficult to choose a suitable scheduling variable that captures the nonlinear behaviour of the

system. Moreover, GSC involves a lengthy tuning process due to its use of multiple local

controllers, and this tuning process may have to be repeated on a daily basis to accommodate

variation in the response of the stimulated muscles. Overall, GSC is not the best choice for

FES applications in the real world.

Neural networks and fuzzy logic are attractive control techniques for FES applications be-

cause they can mimic how the brain is assumed to regulate muscle contractions. However, these

techniques typically do not guarantee stability, which should be considered a mandatory control

requirement for a system that interfaces with human users, such as FES applications. Moreover,

attempting to replicate the control strategy used by the brain is a misguided approach, because

the artificial controller will not have access to the same range and fidelity of feedback data as

the brain, given current sensor technology.

SMC is a simple, reliable technique that guarantees stability under certain conditions. This

algorithm is also able to successfully regulate electrically stimulated muscle contractions in the
Chapter 7. Discussion 98

presence of non-ideal behaviour from the stimulated muscles, as demonstrated in this thesis. In

particular, the observer-based SMC implementation proposed in this thesis works well under

realistic test conditions, especially when the fatigue minimization and antagonist co-contraction

facets of the controller are included. For these reasons, SMC should be considered the first choice

of closed-loop control algorithm for use in real-world FES applications.

It should also be noted that it may be prudent to use a constant reference trajectory for

FES applications, such as a small amplitude sinusoid, instead of a static reference value. The

resulting dynamic steady-state behaviour would ensure that both position and velocity feed-

back are available to the control algorithm at all times, thereby improving control performance.

Moreover, this approach would result in continuous, or at least frequent, muscle stimulation,

which would take advantage of muscle potentiation effects and avoid waiting for “cold” mus-

cles to respond to stimulation when a control action is needed. In fact, the brain may use

this dynamic steady-state approach by maintaining a slow, small amplitude sway during quiet

standing in able-bodied individuals.

Interestingly, increasing joint stiffness may not necessarily increase the fatigue experienced

by the agonist and antagonist muscles; if increasing joint stiffness improves joint damping, then

the muscles may in fact expend less effort, since they will not have to take large corrective

actions to compensate for underdamped joint dynamics.

7.5 Limitations of Thesis Work

The work presented in this thesis has several limitations. The non-idealities block was based

on fatigue data that was collected from untrained SCI subjects during isometric contractions

(i.e., contractions in which the joint angle does not change). Applying this data to isotonic

contractions in trained SCI subjects may overestimate the fatigue that would be seen in the

real world. Moreover, the non-idealities block does not implement muscle recovery. Since some

small amount of recovery will occur in the stimulated muscles during periods of no stimulation,

the actual fatigue exhibited by muscles in the real world will be somewhat less than the fatigue

indicated by the non-idealities block. The non-idealities block could be improved by extending
Chapter 7. Discussion 99

it to include data on fatigue during isotonic muscle contractions as well as muscle recovery,

should this data become available. Moreover, data from other electrically actuated joints could

be added to make the block more generally applicable. The non-idealities block could also

be extended to take additional inputs, such as degree of muscle fatigue as reflected in EMG

behaviour, or knee angle, to further improve the realism of the model.

Also, the observer-based SMC discussed in Chapter 6 was based on a model of the stimulated

muscle response that neglected muscle activation delay and muscle fatigue. The performance

of this controller could likely be improved by removing these simplifications from the plant

model. This control algorithm could also be improved by providing the antagonist torque to

the observer, thereby generating a more accurate state estimate and reducing the steady-state

error. A high-gain loop could also be added around the observer to reject bounded unknown

disturbances such as strong muscle tremors.

7.6 Contributions Made by Thesis Work

The work contained in this thesis represents a significant step forward in closed-loop control for

practical FES applications. The important contributions of this work are as follows:

1. PID control was shown to be an unsuitable control strategy for practical FES applications.

This result can be immediately used by members of the FES community, who can start

focusing their development efforts on more promising control algorithms.

2. The non-idealities block can be used to simulate the probable performance of FES appli-

cations in the presence of actual non-ideal stimulated muscle behaviour. Using this block

will provide FES researchers with a clearer picture of the likely performance of their con-

trol algorithms in the real world, and will save time and money by avoiding human testing

of controllers which will be unsuitable for practical FES applications.

3. The success of the antagonist co-contraction strategy described in Chapter 6 should en-

courage other FES researchers to use this approach in their own FES systems. Since

muscle atrophy in individuals with SCI and other neurological disabilities leads to a re-
Chapter 7. Discussion 100

duction in joint stiffness, the resulting under-damped joints may be a factor in the poor

performance of many FES control algorithms in real-world FES applications.

4. The novel observer-based SMC approach was shown to be very successful in regulating

electrically stimulated muscle contractions, even in the presence of non-ideal behaviour

such as muscle fatigue. This result should encourage the FES community to seriously

consider nonlinear control strategies such as SMC for use in FES applications.

In addition to these specific contributions, the work in this thesis can be easily extended

to other stimulated joints and FES applications such as walking and grasping, as well as other

neurological diagnoses such as stroke and incomplete SCI. Also, the work in this thesis was de-

veloped using surface stimulation technology, but could be applied equally well to systems that

use implanted electrodes. The general nature of these results will allow many FES consumers

to benefit from these advances in FES control systems.


Chapter 8

Conclusions

The objective of this thesis was to develop a closed-loop control algorithm that would be suitable

for use in practical FES applications. This overall objective was approached using three studies,

which were described in Chapters 4, 5, and 6.

Chapter 4 described a systematic analysis of several existing closed-loop control algorithms

that were reported in the literature to be successful for FES applications. This study showed

that fatigue was a significant factor in the failure of existing closed-loop control algorithms.

Moreover, the results showed that using realistic testing methods is essential to determining

the probable performance of FES control algorithms in the real world, and that this testing

issue was neglected by much of the work on closed-loop FES control algorithms. This study

also suggested that co-contraction of the antagonist muscle group should be considered as a

means of increasing the stiffness of the electrically actuated joint, thereby improving controller

performance.

Chapter 5 described the construction of a non-idealities block that allowed actual undesirable

behaviour observed during electrically stimulated muscle contractions to be included in simu-

lations of FES applications. The non-idealities block was tested in a simulation of knee exten-

sion against gravity, and showed that including non-ideal stimulated muscle behaviour greatly

degraded controller performance. This non-idealities block provides a convenient method of

comprehensively testing FES systems in simulation prior to human testing, which will save

time and money, and speed the development of new FES systems for individuals with SCI and

101
Chapter 8. Conclusions 102

other neurological disabilities.

Chapter 6 detailed the development of a novel closed-loop observer-based SMC for use

in real-world FES applications. This algorithm was designed to be stable as well as robust to

model parameter variations, and to reject disturbances due to factors such as muscle spasms and

tremors. Moreover, this algorithm included a fatigue minimization objective and used dynamic

co-contraction of the antagonist muscle group to cause joint stiffness to track a desired value.

The observer-based SMC was tested in a simulation of FES-induced quiet standing, and the

non-idealities block was used to evaluate the probable performance of this controller in a real

FES standing scenario. The observer-based SMC performed very well, even in the presence of

non-ideal behaviour such as muscle fatigue, and is suitable for implementation in certain FES

applications, provided that a small steady-state oscillation could be tolerated and no severe

non-ideal behaviour is expected of the system.

Overall, it can be concluded that implementing closed-loop control in real-world FES ap-

plications is a very achievable goal, provided that several factors are taken into account by the

control system designer:

1. FES users must be trained to increase their muscle strength, thereby minimizing the

detrimental effects of muscle fatigue.

2. Accurate models of the stimulated muscle response must be used when developing control

algorithms, including the non-ideal behaviour that is seen from stimulated muscle in the

real world. Moreover, closed-loop control algorithms for FES systems must be robust

to model mismatch error and disturbances, and should be able to handle the nonlinear

time-varying nature of the stimulated muscle response.

3. PID control is unsuitable for FES applications, and should be abandoned by the FES

community in favour of more promising techniques, such as the observer-based sliding

mode control algorithm proposed in this thesis.

Nonlinear control can be successfully used for FES applications, and can provide excellent

performance despite the challenging nature of controlling electrically stimulated muscle con-

tractions. The FES research community should strongly consider becoming familiar with these
Chapter 8. Conclusions 103

powerful techniques so that sophisticated FES systems can be made available to FES consumers.

Following the completion of this PhD thesis, the author intends to test the observer-based

SMC in SCI subjects in collaboration with Dr. Kei Masani and Dr. Dimitry Sayenko of

Toronto Rehabilitation Institute. Initially, the controller will be tested in a quiet standing test

bed, but will then be extended to a knee extension test bed in which coupled actuators must

be considered. This control algorithm will then be ready for implementation in practical FES

systems. The author will also continue to lobby the FES community to use rigorous control

testing methods, and to encourage the adoption of modern control methods by showing how

these methods can be used in practical FES applications. It is the author’s hope that these

efforts will speed the development of sophisticated new FES applications that can be used by

individuals who have SCI and stroke in their homes and communities.
Bibliography

[1] Christopher and Dana Reeve Foundation. One degree of separation: Paralysis and spinal

cord injury in the United States, 2009.

[2] J. J. Abbas and J. C. Gillette. Using electrical stimulation to control standing posture.

IEEE Control Systems Magazine, 21(4):80–90, 2001.

[3] T. J. Demchak, J. K. Linderman, W. J. Mysiw, R. Jackson, J. Suun, and S. T. Devor.

Effects of functional electric stimulation cycle ergometry training on lower limb muscu-

lature in acute SCI individuals. Journal of Sports Science and Medicine, 4(3):263–271,

2005.

[4] N. Donaldson, T. A. Perkins, R. Fitzwater, D. E. Wood, and F. Middleton. FES cycling

may promote recovery of leg function after incomplete spinal cord injury. Spinal Cord,

38(11):680–682, 2000.

[5] A. Kralj, T. Bajd, and R. Turk. Enhancement of gait restoration in spinal injured patients

by functional electrical stimulation. Clinical Orthopaedics and Related Research, 233:34–

43, 1988.

[6] T. A. Thrasher, H. M. Flett, and M. R. Popovic. Gait training regimen for incomplete

spinal cord injury using functional electrical stimulation. Spinal Cord, 44:357–361, 2006.

[7] J. Perdan, R. Kamnik, P. Obreza, G. Kurillo, T. Bajd, and M. Munih. Design and eval-

uation of a functional electrical stimulation system for hand sensorimotor augmentation.

Neuromodulation, 11(3):208–215, 2008.

104
Bibliography 105

[8] D. B. Popovic, A. Stojanovic, A. Pjanovic, S. Radosavljevic, M. B. Popovic, S. Jovic, and

D. Vulovic. Clinical evaluation of the bionic glove. Archives of Physical Medicine and

Rehabilitation, 80(3):299–304, 1999.

[9] M. R. Popovic, T. A. Thrasher, M. E. Adams, V. Takes, V. Zivanovic, and M. I. Tonack.

Functional electrical therapy: retraining grasping in spinal cord injury. Spinal Cord,

44(3):143–151, 2006.

[10] K. L. Kilgore, H. A. Hoyen, A. M. Bryden, R. L. Hart, M. W. Keith, and P. H. Peckham.

An implanted upper-extremity neuroprosthesis using myoelectric control. Journal of Hand

Surgery - American Volume, 33A(4):539–550, 2008.

[11] S. Mangold, T. Keller, A. Curt, and V. Dietz. Transcutaneous functional electrical stim-

ulation for grasping in subjects with cervical spinal cord injury. Spinal Cord, 43(1):1–13,

2004.

[12] R. Thorsen, R. Spadone, and M. Ferrarin. A pilot study of myoelectrically controlled

FES of upper extremity. IEEE Transactions on Neural Systems and Rehabilitation Engi-

neering, 9(2):161–168, 2001.

[13] W. D. Memberg, P. E. Crago, and M. W. Keith. Restoration of elbow extension via

functional electrical stimulation in individuals with tetraplegia. Journal of Rehabilitation

Research and Development, 40(6):477–486, 2003.

[14] D. B. Popovic, M. B. Popovic, A. Stojanovic, A. Pjanovic, S. Radosavljevic, and

D. Vulovic. Clinical evaluation of the belgrade grasping system. In Proceedings of the

6th Annual Vienna International Workshop on Functional Electrical Stimulation, Vienna,

Austria, September 1998.

[15] L. E. Fisher, M. E. Miller, S. N. Bailey, J. A. Davis, J. S. Anderson, L. Rhode, D. J. Tyler,

and R. J. Triolo. Standing after spinal cord injury with four-contact nerve-cuff electrodes

for quadriceps stimulation. IEEE Transactions on Neural Systems and Rehabilitation

Engineering, 16(5):473–478, 2008.


Bibliography 106

[16] H. Gollee, K. J Hunt, and D. E Wood. New results in feedback control of unsupported

standing in paraplegia. IEEE Transactions on Neural Systems and Rehabilitation Engi-

neering, 12(1):73–80, 2004.

[17] W. Holderbaum, K. J. Hunt, and H. Gollee. Robust discrete-time H-infinity control

for unsupported paraplegic standing: Experiment results. European Journal of Control,

10(3):275–284, 2004.

[18] Z. Matjacic and T. Bajd. Arm-free paraplegic standing - Part I: Control model synthesis

and simulation. IEEE Transactions on Rehabilitation Engineering, 6(2):125138, 1998.

[19] A. H. Vette, K. Masani, J. Y. Kim, and M. R. Popovic. Closed-loop control of functional

electrical stimulation-assisted arm-free standing in individuals with spinal cord injury: A

feasibility study. Neuromodulation, 12(1):22–32, 2009.

[20] I. Cikajlo, Z. Matjacic, T. Bajd, R. Futami, and N. Hoshimiya. Sensory supported FES

control in gait training of incomplete SCI persons. Artificial Organs, 29(6):459–461, 2005.

[21] T. Fuhr, J. Quintern, R. Riener, and G. Schmidt. Walking with WALK! A coopera-

tive patient-driven neuroprosthetic system. IEEE Engineering in Medicine and Biology

Magazine, 27(1):38–48, 2008.

[22] D. Graupe, R. Davis, H. Kordylewski, and K. H. Kohn. Ambulation by traumatic T4-12

paraplegics using functional neuromuscular stimulation. Critical Reviews in Neurosurgery,

8(4):221–231, 1998.

[23] E. Hardin, R. Kobetic, L. Murray, M. Corado-Ahmed, G. Pinault, J. Sakai, S. N. Bai-

ley, C. Ho, and R. J. Triolo. Walking after incomplete spinal cord injury using an im-

planted FES system: A case report. Journal of Rehabilitation Research and Development,

44(3):333–346, 2007.

[24] T. A. Perkins, N. N. Donaldson, N. A. C. Hatcher, I. D. Swain, and D. E. Wood. Control of

leg-powered paraplegic cycling using stimulation of the lumbo-sacral anterior spinal nerve
Bibliography 107

roots. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 10(3):158–

164, 2002.

[25] R. Davoodi and B. J. Andrews. Fuzzy logic control of FES rowing exercise in paraplegia.

IEEE Transactions on Biomedical Engineering, 51(3):541–543, 2004.

[26] R. J. Triolo, L. Boggs, M. E. Miller, G. Nemunaitis, J. Nagy, and S. N. Bailey. Implanted

electrical stimulation of the trunk for seated postural stability and function after cervical

spinal cord injury: A single case study. Archives of Physical Medicine and Rehabilitation,

90(2):340–347, 2009.

[27] M. Vanoncini, W. Holderbaum, and B. Andrews. Development and experimental identifi-

cation of a biomechanical model of the trunk for functional electrical stimulation control

in paraplegia. Neuromodulation, 11(4):315–324, 2008.

[28] M. S. Poboroniuc, D. E. Wood, R. Riener, and N. D. Donaldson. A new controller for

fes-assisted sitting down in paraplegia. Advances in Electrical and Computer Engineering,

10(4):9–16, 2010.

[29] C. Y. Chao and G. L. Cheing. The effects of lower-extremity functional electric stimulation

on the orthostatic responses of people with tetraplegia. Archives of Physical Medicine and

Rehabilitation, 86(7):1427–1433, 2005.

[30] L. Chi, K. Masani, M. Miyatani, T. Adam Thrasher, K. Wayne Johnston, A. Mardimae,

C. Kessler, J. A. Fisher, and M. R. Popovic. Cardiovascular response to functional

electrical stimulation and dynamic tilt table therapy to improve orthostatic tolerance.

Journal of Electromyography and Kinesiology, 18(6):900–907, 2008.

[31] A. S. Elokda, D. H. Nielsen, and R. K. Shields. Effect of functional neuromuscular

stimulation on postural related orthostatic stress in individuals with acute spinal cord

injury. Journal of Rehabilitation Research and Development, 37(5):535–542, 2000.


Bibliography 108

[32] E. E. Sampson, R. S. Burnham, and B. J. Andrews. Functional electrical stimulation

effect on orthostatic hypotension after spinal cord injury. Archives of Physical Medicine

and Rehabilitation, 81(2):139–143, 2000.

[33] K. M. Bogie and R. J. Triolo. Effects of regular use of neuromuscular electrical stimulation

on tissue health. Journal of Rehabilitation Research and Development, 40(6):469–476,

2003.

[34] A. C. Ferguson, J. F. Keating, M. A. Delargy, and B. J. Andrews. Reduction of seating

pressure using FES in patients with spinal cord injury: A preliminary report. Paraplegia,

30(7):474–478, 1992.

[35] S. J. Hwang, H. Kaplan, G. E. Loeb, and Y. H. Kim. Weight shifts by the electrical

stimulation in seating for the prevention of pressure ulcer. In Proceedings of the World

Congress on Medical Physics and Biomedical Engineering, volume 14, pages 1–6, Seoul,

South Korea, 2006.

[36] J. A. Hoffer. Transvascular nerve stimulation apparatus and methods [Patent Applica-

tion], August 2008.

[37] J. A. Hoffer, B. D. Tran, J. K. Tang, J. T. W. Saunders, C. A. Francis, R. A. Sandoval,

R. Meyyappan, S. Seru, H. D. Y. Wang, M. A. Nolette, and A. C. Tanner. Diaphragm

pacing with endovascular electrodes. In Proc. 15th Ann. Conf. of the Int. Functional

Electrical Stimulation Society, pages 40–42, Vienna, Austria, 2010.

[38] C. L. Lynch, G. M. Graham, and M. R. Popovic. Including non-ideal behaviour in

simulations of functional electrical stimulation applications. In Proceedings of the 15th

Annual Conference of the International Functional Electrical Stimulation Society, pages

327–329, Vienna, Austria, September 2010.

[39] C. L. Lynch and M. R. Popovic. Functional electrical stimulation: Closed-loop control of

induced muscle contractions. IEEE Control Systems Magazine, 28(2):40–50, 2008.


Bibliography 109

[40] C. L. Lynch and M. R. Popovic. Closed-loop control of FES: Past work and future

directions. In Proc. 10th Ann. Conf. of the Int. Functional Electrical Stimulation Society,

pages 47–49, Montreal, Canada, 2005.

[41] C. L. Lynch and M. R. Popovic. A comparison of closed-loop control algorithms for

regulating electrically stimulated knee movements in individuals with spinal cord injury.

Submitted to IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2011.

[42] C. L. Lynch and M. R. Popovic. A model of the knee’s response to electrical stimulation

for use in FES applications. In Proc. of the National Spinal Cord Injury Conference,

Toronto, Canada, 2008.

[43] C. L. Lynch and M. R. Popovic. A stochastic model of knee joint response to electrical

stimulation of the quadriceps and hamstrings muscles. Submitted to Artificial Organs,

2011.

[44] C. L. Lynch, G. M. Graham, and M. R. Popovic. Including non-ideal behaviour in

simulations of functional electrical stimulation applications. Artificial Organs, 35(3):267–

269, 2011.

[45] C. L. Lynch, G. M. Graham, and M. R. Popovic. Including non-ideal behaviour in

simulations of functional electrical stimulation. In Proc. 5th Int. IEEE EMBS Conf. on

Neural Engineering, Cancun, Mexico, May 2011.

[46] C. L. Lynch, G. M. Graham, and M. R. Popovic. A generic model of real-world non-ideal

behaviour of FES-induced muscle contractions: Simulation tool. Submitted to Journal of

Neural Engineering, 2011.

[47] C. L. Lynch and M. R. Popovic. Sliding mode control of electrically induced quiet standing

in individuals with spinal cord injury. Submitted to IEEE Transactions on Neural Systems

and Rehabilitation Engineering, 2011.

[48] J. F. Tan, J. Zariffa, A. H. Vette, C. L. Lynch, K. Masani, and M. R. Popovic. Use of

an inverted pendulum apparatus for the study of closed-loop FES control of the ankle
Bibliography 110

joints during quiet standing. In Proc. 15th Ann. Conf. of the Int. Functional Electrical

Stimulation Society, pages 118–120, Vienna, Austria, 2010.

[49] W. Kapit and L. M. Elson. The Anatomy Coloring Book. Benjamin Cummings, USA,

2002.

[50] J. T. Mortimer. Motor prostheses. In Handbook of Physiology - The Nervous System II.

American Physiological Society, Bethesda, MD, USA, 1981.

[51] A. Nuruki, S. Uchida, K. Yunokuchi, and R. Nagaoka. Frequency response of evoked

potential in normal and diseased nerve muscle. IEEE Engineering in Medicine and Biology

Magazine, 18(6):27–32, 1999.

[52] A. C. Guyton and J. E. Hall. Textbook of Medical Physiology. Elsevier Saunders, Philadel-

phia, USA, 11th edition, 2006.

[53] L. L. Baker, C. L. Wederich, D. R. McNeal, C. Newsam, and R. L. Waters. Neuromuscular

Electrical Stimulation - A Practical Guide. Rancho Los Amigos National Rehabilitation

Center, Downey, CA, USA, 4th edition, 2000.

[54] American Spinal Injury Association. Standard neurological classification of spinal cord

injury, 2006.

[55] D. B. Popovic and T. Sinkjaer. Control of Movement for the Physically Disabled: Control

for Rehabilitation Technology. Springer Verlag, London, UK, 2000.

[56] Canadian Paraplegic Association. Complications of SCI, March 2000.

[57] K. M. Deutsch, T. G. Hornby, and B. D. Schmit. The intralimb coordination of the flexor

reflex response is altered in chronic human spinal cord injury. Neuroscience Letters,

380(3):305–310, 2005.

[58] N. Kawashima, D. Taguchi, K. Nakazawa, and M. Akai. Effect of lesion level on the

orthotic gait performance in individuals with spinal cord injuries. Spinal Cord, 44:487–

494, 2006.
Bibliography 111

[59] M. R. Popovic and T. A. Thrasher. Neuroprostheses. In Encyclopedia of Biomaterials and

Biomedical Engineering, volume 2, pages 1056–1065. Dekker Encyclopedias, New York,

USA, 2004. G. E. Wnek and G. L. Bowlin, editors.

[60] T. A. Thrasher, G. M. Graham, and M. R. Popovic. Attempts to reduce muscle fatigue by

randomizing FES parameters. In Proceedings of the 8th Vienna Workshop on Functional

Electrical Stimulation, pages 80–83, Vienna, Austria, 2004.

[61] H. Lanmuller, S. Sauermann, E. Unger, G. Schnetz, W. Mayr, M. Bijak, and W. Girsch.

Multi-functional implantable nerve stimulator for cardiac assistance by skeletal muscle.

Artificial Organs, 23(4):352–359, 1999.

[62] Z. Lertmanorat, K. J. Gustafson, and D. M. Durand. Electrode array for reversing

the recruitment order of peripheral nerve stimuation: Experimental studies. Annals of

Biomedical Engineering, 34(1):152–160, 2006.

[63] J. A. Riess and J. J. Abbas. Adaptive neural network control of cyclic movements us-

ing functional neuromuscular stimulation. IEEE Transactions on Neural Systems and

Rehabilitation Engineering, 8(1):42–52, 2000.

[64] J. A. Riess and J. J. Abbas. Adaptive control of cyclic movements as muscles fatigue

using functional neuromuscular stimulation. IEEE Transactions on Neural Systems and

Rehabilitation Engineering, 9(3):326–330, 2001.

[65] M. A. Lemay and P. E. Crago. Closed-loop wrist stabilization in C4 and C5 tetraplegia.

IEEE Transactions on Rehabilitation Engineering, 5(3):244–252, 1997.

[66] S. Micera, A. M. Sabatini, and P. Dario. Adaptive fuzzy control of electrically stimu-

lated muscles for arm movements. Medical and Biological Engineering and Computing,

37(1):680–685, 1999.

[67] M. M. Adamczyk and P. E. Crago. Simulated feedforward neural network coordination

of hand grasp and wrist angle in a neuroprosthesis. IEEE Transactions on Rehabilitation

Engineering, 8(3):297–304, 2000.


Bibliography 112

[68] J. J. Abbas and H. J. Chizeck. Feedback control of coronal plane hip angle in paraplegic

subjects using functional neuromuscular stimulation. IEEE Transactions on Biomedical

Engineering, 38(7):687–698, 1991.

[69] R. P. Jaime, Z. Matjacic, and K. J. Hunt. Paraplegic standing supported by FES-

controlled ankle stiffness. IEEE Transactions on Neural Systems and Rehabilitation En-

gineering, 10(4):239–248, 2002.

[70] Z. Matjacic, K. Hunt, H. Gollee, and T. Sinkjaer. Control of posture with FES systems.

Medical Engineering and Physics, 25(1):51–62, 2003.

[71] K. J. Hunt, R. P. Jaime, and H. Gollee. Robust control of electrically-stimulated muscle

using polynomial H-infinity design. Control Engineering Practice, 9(3):313–328, 2001.

[72] W. Holderbaum, K. J. Hunt, and H. Gollee. H-infinity robust control design for un-

supported paraplegic standing: Experimental evaluation. Control Engineering Practice,

10(11):1211–1222, 2002.

[73] M. S. Hatwell, B. J. Oderkerk, C. A. Sacher, and G. F. Inbar. The development of a model

reference adaptive controller to control the knee joint of paraplegics. IEEE Transactions

on Automatic Control, 36(6):683–691, 1991.

[74] J. F. Tan. Closed-Loop Control of the Ankle Joint Using Functional Electrical Stimulation.

MASc Thesis, University of Toronto, Toronto, Canada, 2008.

[75] G. C. Chang, J. J. Luh, G. D. Liao, J. S. Lai, C. K. Cheng, B. L. Kuo, and T. S. Kuo.

A neuro-control system for the knee joint position control with quadriceps stimulation.

IEEE Transactions on Rehabilitation Engineering, 5(1):2–11, 1997.

[76] M. Ferrarin, F. Palazzo, R. Riener, and J. Quintern. Model-based control of FES-induced

single joint movements. IEEE Transactions on Neural Systems and Rehabilitation Engi-

neering, 9(3):245–257, 2001.


Bibliography 113

[77] F. Previdi and E. Carpanzano. Design of a gain scheduling controller for knee-joint angle

control by using functional electrical stimulation. IEEE Transactions on Control Systems

Technology, 11(3):310–324, 2003.

[78] S. Jezernik, R. G. V. Wassink, and T. Keller. Sliding mode closed-loop control of

FES: Controlling the shank movement. IEEE Transactions on Biomedical Engineering,

51(2):263–272, February 2004.

[79] A. Ajoudani and A. Erfanian. A neuro-sliding-mode control with adaptive modeling of

uncertainty for control of movement in paralyzed limbs using functional electrical stimu-

lation. IEEE Transactions on Biomedical Engineering, 56(7):1771–1780, 2009.

[80] H. R. Kobravi and A. Erfanian. Decentralized adaptive robust control based on sliding

mode and nonlinear compensator for the control of ankle movement using functional

electrical stimulation of agonist-antagonist muscles. Journal of Neural Engineering, 6(4),

2009.

[81] F. Previdi. Identification of black-box nonlinear models for lower limb movement control

using functional electrical stimulation. Control Engineering Practice, 10(1):91–99, 2002.

[82] M. Ferrarin and A. Pedotti. The relationship between electrical stimulus and joint torque:

A dynamic model. IEEE Transactions on Rehabilitation Engineering, 8(3):342–352, 2000.

[83] R. Perumal, A. S. Wexler, and S. A. Binder-Macleod. Mathematical model that pre-

dicts lower leg motion in response to electrical stimulation. Journal of Biomechanics,

39(15):2826–2836, 2006.

[84] T. Schauer, W. Holderbaum, and K. J. Hunt. Sliding-mode control of knee-joint angle:

Experimental results. In Proc. 7th Annual Conf. Int. Functional Electrical Stimulation

Society, pages 316–318, Ljubljana, Slovenia, June 2002.

[85] M. S. Marion, A. S. Wexler, M. L. Hull, and S. A. Binder-Macleod. Predicting the effect of

muscle length on fatigue during electrical stimulation. Muscle and Nerve, 40(4):573–581,

2009.
Bibliography 114

[86] M. S. Marion, A. S. Wexler, and M. L. Hull. Predicting fatigue during electrically stim-

ulated non-isometric contractions. Muscle and Nerve, 41(6):857–867, 2010.

[87] R. Riener, J. Quintern, and G. Schmidt. Biomechanical model of the human knee evalu-

ated by neuromuscular stimulation. Journal of Biomechanics, 29(9):1157–1167, 1996.

[88] J. V. D. Vegte. Feedback Control Systems. Prentice Hall, Upper Saddle River, NJ, USA,

3rd edition, 1994.

[89] C. Knospe. PID control. IEEE Control Systems Magazine, 26(1):30–31, 2006.

[90] J. S. Shamma and M. Athans. Gain scheduling: Potential hazards and possible remedies.

IEEE Control Systems Magazine, 12(3):101–107, 2002.

[91] H. Khalil. Nonlinear Systems. Prentice Hall, Upper Saddle River, NJ, USA, 3rd edition,

2002.

[92] V. I. Utkin. Sliding Modes in Control and Optimization. Springer Verlag, Berlin, Germany,

1992.

[93] K. D. Young, V. I. Utkin, and U. Ozguner. A control engineer’s guide to sliding mode

control. IEEE Transactions on Control Systems Technology, 7(3):328–342, 1999.

[94] J. J. E. Slotine, J. K. Hedrick, and E. A. Misawa. On sliding observers for nonlinear

systems. ASME Journal of Dynamics, Systems, Measurement, and Control, 109(3):245–

252, 1987.

[95] A. G. Bondarev, S. A. Bondarev, N. E. Kostyleva, and V. I. Utkin. Sliding modes in

systems with asymptotic observers. Automation and Remote Control, pages 679–684,

1985.

[96] M. M. Adams and A. L. Hicks. Spasticity after spinal cord injury. Spinal cord, 43(10):577–

586, 2005.

[97] J. R. Brinkmann and J. Perry. Rate and range of knee motion during ambulation in

healthy and arthritic subjects. Physical Therapy, 65(7):1055, 1985.


Bibliography 115

[98] J. Perry. Gait Analysis: Normal and Pathological Function. McGraw-Hill, Inc., New

York, USA, 1992.

[99] G. M Graham, T. A Thrasher, and M. R Popovic. The effect of random modulation

of functional electrical stimulation parameters on muscle fatigue. IEEE Transactions on

Neural Systems and Rehabilitation Engineering, 14(1):38–45, 2006.

[100] K. S. Bordens and B. B. Abbott. Research Design and Methods: A Process Approach.

Mayfield Publishing Company, Mountain View, CA, USA, 3rd edition, 1996.

[101] J. Y. Kim, J. K. Mills, A. H. Vette, and M. R. Popovic. Optimal combination of minimum

degrees of freedom to be actuated in the lower limbs to facilitate arm-free paraplegic

standing. ASME Journal of Biomechanical Engineering, 129:838–847, 2007.

[102] D. A. Winter, A. E. Patla, S. Rietdyk, and M. G. Ishac. Ankle muscle stiffness in the

control of balance during quiet standing. Journal of Neurophysiology, 85(6):2630–2633,

2001.

[103] S. Mohammed, P. Fraisse, D. Guiraud, P. Poignet, and H. El Makssoud. Robust control

law strategy based on high order sliding mode: Towards a muscle control. In Proc. of the

IEEE/RSJ Int. Conf. on Intelligent Robots, pages 2882–2887, 2005.

[104] I. D. Loram and M. Lakie. Direct measurement of human ankle stiffness during quiet

standing: The intrinsic mechanical stiffness is insufficient for stability. Journal of Physi-

ology, 545:1041–1053, 2002.

[105] B. A. Lavoie, F. W. J. Cody, and C. Capaday. Cortical control of human soleus muscle

during volitional and postural activities studied using focal magnetic stimulation. Exper-

imental Brain Research, 103:97–107, 1995.

[106] T. Kiemel, K. S. Oie, and J. J. Jeka. Slow dynamics of postural sway are in the feedback

loop. Journal of Neurophysiology, 95(3):1410–1418, 2006.


Appendix A

Non-Ideality Classification Criteria

116
Appendix A. Non-Ideality Classification Criteria 117

Table A.1: Classification criteria for non-ideality waveforms. f at(t) represents normalized

fatigue as described in Section 5.2.1. θ(t) represents de-noised knee angle as described in

Section 5.2.1.
Non-ideality
Classification Criteria
Type

Fatigue mild f at(t) is not moderate or severe

moderate f at(t) ≤ 0.7 for all t > 20 s

severe f at(t) ≤ 0.7 for all t > 10 s

Spasm mild Absolute amplitude of all spikes in θ(t) is less than 10◦

moderate Absolute amplitude of all spikes in θ(t) is less than 20◦

and at least one spike in θ(t) has absolute amplitude ≥ 10◦

OR

Any section of θ(t) has ≥ 10 spikes/s

severe Absolute amplitude of any spike in θ(t) is ≥ 20◦

OR

Any section of θ(t) has ≥ 20 spikes/s

Tremor mild Peak-to-peak amplitude of sinusoidal component of

θ(t) < 7.5◦ ∀ t > 0.5 s

moderate Peak-to-peak amplitude of sinusoidal component of

θ(t) < 15◦ ∀ t > 0.5 s,

and θ(t) ≥ 7.5◦ for some t > 0.5 s

severe Peak-to-peak amplitude of sinusoidal component of

θ(t) ≥ 15◦ for some t > 0.5 s


Appendix B

List of Abbreviations

118
Appendix B. List of Abbreviations 119

AIS ASIA Impairment Scale

ASIA American Spinal Injury Association

ATP adenosine triphosphate

CNS central nervous system

CPG central pattern generator

eRMS root-mean-squared error

ess steady-state error

FES functional electrical stimulation

GSC gain scheduling control, gain scheduling controller

H∞ H-infinity

LQG linear quadratic Gaussian

LQR linear quadratic regulator

% OS percent overshoot

PD proportional-derivative

PID proportional-integral-derivative

RMS root-mean-squared

SCI spinal cord injury, spinal cord injured

SMC sliding mode control, sliding mode controller

tdr disturbance rejection time

trise 10%-90% rise time

tset 2% settling time

tsettling 2% settling time


Index

Ca2+ , 9 cuff electrode, 16

acetylcholine, 11 disuse atrophy, 15

acetylcholinesterase, 11
electrode
action potential, 11
cuff, 16
adenosine triphosphate, 11
epimysial, 16
anti-windup, 57
intraneural, 16
asynchronous recruitment, 12
percutaneous, 16
ATP, 11
transcutaneous, 16
atrophy, 15
epimysial electrode, 16

biphasic stimulation waveform, 16


fascicle, muscle, 9

fast-twitch muscle fibre, 9


calcium ion, 9
fibre, muscle, 9
central pattern generator, 15

chattering, 29 incomplete SCI, 14


complete SCI, 14 incomplete tetanus, 15
contraction integral control, 38
isometric, 69 intraneural electrode, 16
isotonic, 69 intrinsic tonic spasticity, 35
tetanic, 12 isometric muscle contraction, 69
control isotonic muscle contraction, 69
anti-windup, 57
junction gap, 11
integral, 38

CPG, 15 linear quadratic regulator, 39

creatine phosphate, 11 LQR, 39

120
INDEX 121

Luenberger observer, 30 recruitment

asynchronous, 12
manifold, sliding, 27
non-physiological, 18
motor neuron, 11
synchronous, 18
motor unit, 11
reflex, spinal, 14
muscle

atrophy, 15 sarcolemma, 9

contraction scheduling variable, 27

isometric, 69 SCI

isotonic, 69 complete, 14

fascicle, 9 incomplete, 14

fibre, 9 sliding manifold, 27

fast-twitch, 9 slow-twitch muscle fibre, 9

slow-twitch, 9 spasm, 14

re-training, 15 spasticity, 14

recruitment intrinsic tonic, 35

asynchronous, 12 spinal reflex, 14

non-physiological, 18 stimulation waveform, biphasic, 16

synchronous, 18 synchronous recruitment, 18

spasm, 14
tetanic contraction, 12
tremor, 15
tetanus
muscle re-training, 15
complete, 12
muscle tremor, 15
incomplete, 15

non-physiological recruitment, 18 transcutaneous electrode, 16

transverse tubules, 9
observer, Luenburger, 30
tubules, transverse, 9

percutaneous electrode, 16

potential, 11

re-training, 15

You might also like