Professional Documents
Culture Documents
Mechatronics
journal homepage: www.elsevier.com/locate/mechatronics
a r t i c l e i n f o a b s t r a c t
Article history: This paper uses experimental data to model the human arm at the elbow joint, using both third order
Received 30 January 2013 integer dynamic models, commonly found in the literature, and fractional order models (involving deriv-
Revised 2 May 2013 atives of fractional order, between 1.2 and 1.4). The latter models prove to be superior, as they provide a
Accepted 6 May 2013
better adjustment, with less parameters, with lower parameter uncertainty, and without instability prob-
Available online 25 June 2013
lems. These novel results for modelling the human arm are in line with published literature suggesting
that muscles may be profitably modelled with fractional derivatives.
Keywords:
Ó 2013 Elsevier Ltd. All rights reserved.
Human arm
Modelling
Fractional models
Levy method
0957-4158/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.mechatronics.2013.05.003
806 I. Tejado et al. / Mechatronics 23 (2013) 805–812
reported in the literature. Finally, Section 5 draws the conclusions multiple are called commensurable. If the said least common multi-
of this paper. ple is 1, we have a usual, integer transfer function.
system running at 2 kHz frequency was used. This target computer sponses in the mentioned range. In particular, three types of inputs
had a MF624 Humusoft analogue–digital PCI board for digital-ana- were designed:
logue communications and a JR3 PCI-BUS dual receiver board.
Let us consider the motor–link–sensor rotational rigid body Type I – Oscillations in both directions around the zero-point.
dynamics as follows: Type II – Oscillations only in the positive side of the zero-point
I€h þ F ¼ s; ð6Þ (flexion of the elbow).
Type III – Oscillations only in the negative side of the zero-point
where I and F are the inertia and the friction term, respectively, and (extension of the elbow).
the applied torque s is
s ¼ K Torque IC ; Fig. 3 shows an example of signals involved in one of the exper-
iments for input force of type I. For each of the first 17 subjects, two
being K Torque and IC the motor torque constant and current input, trials were performed for input force type I and one trial for types II
respectively. In (6), it is desirable to eliminate all the motor–link– and III, resulting in four sets of data per subject (68 trials in total).
sensor dynamics in order to be able to measure only the human The two other subjects, one male and one female, performed 16
arm dynamics. To this end, the applied motor torque was additional trials, in a total of twenty trials each: six of type I, seven
s ¼ sDyn þ sExternal ; of type II, and seven of type III. The force measured by the men-
tioned sensor was practically identical to the force input. It was as-
in which sDyn refers to a dynamic compensation given by sumed that the bandwidth for the human arm is approximated to
10 Hz, fact which justifies the selection of 15 Hz bandwidth for the
sDyn ¼ bI €h þ bF ¼ bI €h þ Cb h_ þ Bb signðhÞ;
_ ð7Þ
force input.
where bI is the estimated inertia and Fb the estimated friction, com- Both the force applied on the subject, F applied ðtÞ, and the motor
prising a viscous friction term Cb h_ and a static friction term B
b signðhÞ.
_ angle, hðtÞ, were recorded with a sampling frequency of 2 kHz.
It is worth remarking that compensation (7) is as good as the esti- The desired dynamic human arm transfer function (corresponding
mation of bI and bF . For this experiment, a simple Coulomb friction to the block scheme of Fig. 1) was determined as
model was sufficient, since the motor in question is a direct drive hðsÞ
motor without the drawbacks of gear backlash. Moreover, to in- Garm ðsÞ ¼ :
F measured ðsÞ
crease the force control robustness, a PI controller was added as
shown in the block scheme in Fig. 1. To attenuate the importance
of the motor force controller dynamics on the identification loop, 3.4. Estimation of frequency responses
it was assumed that, for this control implementation, the motor
force controller has a higher bandwidth than the one presented The identification of a fractional model directly from a time re-
by the human arm. Therefore, the experimental PI tuning results sponse is reviewed in [25]; a survey of methods for both time and
into a motor controller bandwidth of 18 Hz, which is acceptable frequency responses can be found in [20]. In this paper, it was cho-
for an assumed human harm bandwidth of 10 Hz. sen to identify the dynamics of the human arm estimating a fre-
quency response from the experiments, and then applying Levy’s
3.3. Experiments method, because this was the alternative leading to the best, less
noisy results.
During each experiment, subjects were kneeling or sitting on To estimate the frequency response, an approach for estimating
one side of the motor, grabbing the handle with an angle for the el- transfer functions of linear systems from spectral analysis will be
bow of about 90° as shown in Fig. 2. Although their elbow and used. Spectral estimation was applied to human arm modelling
wrists were not supported, they were asked not only to maintain in [5,26,9]. In particular, the measured human experiment fre-
a firm grip on the handle but also not move their trunk, shoulder quency response is computed by using Welch’s method – by means
and elbow. Likewise, during each trial the subjects were required of the MATLAB function tfestimate – with 11 Blackmanharris-win-
to keep the laser pointer at zero-point under muscle co-contraction. dowed segments (length of the window equal to 5000), 15,000
In order to avoid subject fatigue, the duration of each trial was samples and 25% overlap in order to attain smooth frequency re-
40 s. Human anticipatory reflexes make modelling the arm dynam- sponses. Previously, the output was filtered by a zero-phase for-
ics difficult, as it is hard to keep the arm passive to force distur- ward and reverse FIR equal ripple digital filter – applied with
bances. Fortunately, the more random and wide-band the force MATLAB function filtfilt – designed with pass-band and stop-band fre-
disturbance is, the less likely it will trigger the arm’s reflexes quencies of 15 Hz and 18 Hz, respectively, and a passband ripple of
[5,9]. For this reason, the current study used inputs generated by 1 dB with 80 dB of stop-band attenuation.
a sum of sinusoids with frequencies in the [0.12, 15] Hz range, Once the frequency response Gðjxp Þ; p ¼ 1; . . . ; f is found, Le-
and limited to not exceed 4 N, to render them unpredictable by vy’s method fits it a commensurable fractional model with fre-
the subjects under experiment and allow us to obtain frequency re- quency response given by
Fig. 2. Experimental set-up used for collecting experimental data of human arm dynamics: frontal view (left) and top view (right).
808 I. Tejado et al. / Mechatronics 23 (2013) 805–812
Fig. 3. Example of signals involved in one experiment for input type I: (a) reference force, (b) measured force and (c) elbow angle (the first 10 s of the experiment).
Pm ka
k¼0 bk ðjxp Þ Nðjxp Þ 4.2. Fractional models
b xp Þ ¼
Gðj Pn ¼ : ð8Þ
1 þ k¼1 ak ðjxp Þ
ka Dðjxp Þ
Let us consider a general model, corresponding to frequency re-
sponse (8); its (fractional) transfer function of commensurable or-
Notice that, without loss of generality, we have made 2 a0 ¼ 1. The der a is:
xÞ
adjustment might be done minimising GðjxÞ Nðj DðjxÞ
, but it is eas-
2
ier to minimise instead ðGðjxÞDðjxÞ NðjxÞÞ . For alternative for-
Pm
mulations and improvements of this identification method, see k¼0 bk s
ka
b
GðsÞ ¼ P :
[27]. n
1 þ k¼1 ak ska
It should be noticed that the use of weights to correct an exces-
sive influence of high-frequency
2 data [28,20], or the use of itera-
tions to minimise G ND rather than jGD N j2 [20,29], have
Each model was identified to accurately reflect the mean value of
been attempted, without improvement in results.
the measured frequency responses in the [0.12, 15] Hz range by
means of Levy’s method in Matlab (with levy function [30]) for frac-
tional and integer order models up to third order in accordance with
4. Measured-dynamic models [5].
Commensurable orders of fractional models were found sweep-
In this section, the dynamic models for the human arm resulting ing the a 20; 2½ range, where models are stable, with a 0.1 step,
of the identification are presented. and keeping the best result.
Different models were identified for:
inputs of type I;
4.1. Integer models inputs of type II;
inputs of type III;
Integer, third-order models result from considering the arm as a all inputs at once.
mechanical system consisting of five elements – one mass, two
springs and two dampers [5–9]. The rationale behind this model Since several frequency responses had to be taken into account
is shown in Fig. 4. to find each model, there were two options:
Table 1 Table 2
Arm structure parameters for the fractional model and their variability. Arm structure parameters for the fractional model of one male and their variability.
This last option is the one from which an idea of the confidence
Applying Levy’s method to all frequency responses together. of the results can be more easily got. The mean and the standard
This means that, for each frequency, there are many different deviation should suffice to describe the distribution of the values,
gains and phases to which the model should be fit. The param- since the number of experiments involved is such that a normal dis-
eters of the resulting models are mentioned below as tribution can reasonably be assumed (as done implicitly in [31]).
‘‘nominal’’.
Applying Levy’s method separately to each of the frequency 4.3. Identification results
responses. From the parameters of the resulting models, mean
values, extreme values, and a standard deviation can be found. Identification results are summed up in Table 1, and graphically
represented in Fig. 5.
Fig. 5. Experimental and identified frequency responses for input force of (a) all types; (b) Type I; (c) Type II; and (d) Type III. Thin lines are the several frequency responses
corresponding to the measured data. Dash-dotted lines correspond to the frequency responses of the arm models.
810 I. Tejado et al. / Mechatronics 23 (2013) 805–812
Table 3 Table 5
Arm structure parameters for the fractional model of one female and their variability. Performance indices for the time responses of the fractional and integer models.
Input Zero Poles Gain a Model Input MSE ð105 Þ MAD ð103 Þ MD
Re Im Fractional All 6.21 4.64 0.138
All Nominal 480.89 98.84 ±40.22 0.16 1.3 Type I 7.15 5.25 0.138
Min 583.48 227.9 0 0.065 1.2 Type II 5.14 3.82 0.121
Max 388.31 66.17 54.98 0.374 1.3 Type III 5.11 3.84 0.136
Mean 513.78 125.01 31.97 0.2 1.33 Integer All 6.38 4.73 0.138
Std 48.23 46.58 18.74 0.08 0.06 Type I 7.22 5.31 0.138
Type I Nominal 553.45 117.60 ±36.66 0.15 1.3 Type II 5.23 4.01 0.12
Min 573.83 117.69 19.78 0.06 1.2 Type III 5.13 3.85 0.136
Max 510.84 68.69 53.68 0.17 1.3
Mean 548.42 101.26 38.19 0.14 1.28
Std 23.58 17.12 12.76 0.04 0.04
Type II Nominal 482.73 100.28 ±33.24 0.16 1.3
Models with two zeros and two poles, i.e. having the same number
Min 583.48 215.68 0 0.15 1.3
Max 464.44 94.90 41.51 0.29 1.4
of parameters as those in Table 1, are not fit for this purpose, as
Mean 510.27 134.77 24.66 0.19 1.33 their time responses do not start at zero. The best option has
Std 42.22 53.34 17.64 0.06 0.05 proved to be the structure more often found in the literature, as
Type III Nominal 525.73 164.45, 143.08 0 0.32 1.4 mentioned above in the state of the art of Section 1, with three
Min 563.46 227.9 0 0.15 1.3 poles and one zero. Parameters are given in Table 4.
Max 388.31 66.17 54.98 0.37 1.4 In order to compare and evaluate the goodness of the models in
Mean 487.61 135.61 33.95 0.26 1.36
terms of time domain responses, the following performance indi-
Std 55.73 54.56 23.65 0.09 0.05
ces were calculated for the elbow angle for the different inputs:
Table 4
Arm structure parameters for the integer model and their variability.
Re Im Re Im
All Nominal 21.28 ±111.38 128.63 15.11 ±30.79 0.1
Min 30.8 90.8 1684.2 25 21.9 0.6
Max 4.66 112.04 512.44 6.27 43.09 1.52
Mean 16.79 99.64 189.9 14.1 31.60 0.16
Std 8.54 6.36 411.81 4.99 5.11 0.39
Type I Nominal 19.1034 ±96.96 76.59 16.43 ±35.13 0.08
Min 34.87 89.37 268.32 23.33 25.24 0.04
Max 6.48 107.86 30.37 7.47 53.57 0.28
Mean 19.30 95.78 98.28 15.06 36.83 0.1
Std 8.83 5.8 76.63 4.09 6.91 0.06
Type II Nominal 17.16 ±101.91 170.71 12.6 ±27.61 0.13
Min 35.36 82.09 645.08 20.56 22.22 13.15
Max 2.5 123.7 7653.4 6 45.2 0.4
Mean 18.71 100.87 355.36 11.6 29.75 0.67
Std 11.0 12.6 1850.2 4.1 6 3.1
Type III Nominal 10.61 ±103.19 455.59 11.16 ±28.42 0.323
Min 33.81 90.6 887.12 22.77 21.27 0.477
Max 27.16 164.08 393.36 4.36 37.03 0.574
Mean 13.62 104.35 152.8 11.02 29.77 0.102
Std 14.35 16.5 247.21 4.79 4.27 0.207
I. Tejado et al. / Mechatronics 23 (2013) 805–812 811
x 10
−3
responses. Parts of the said responses are shown in Fig. 6; notice
8
that the input used is not the same for which models are devel-
6
oped: this corresponds to a harsher test to the models.
4 Comparing Tables 1 and 4, it can also be seen that there are
2 unstable models among the integer order ones, something which
Angle (rad)
0 never happens for fractional models (even though the zone of the
−2 complex plane where poles are stable is more narrow [32]). This,
too, is indicative of the advantages of fractional models for this
−4
system.
−6
Table 6 shows the ratio between standard deviations and mean
−8 values for both integer and fractional models. It can be seen that
Experimental
Fractional model
−10 16 Integer model
17
with only one exception fractional models have lower ratios, an-
10 11 12 13 14 15 18 19 20
Time (s)
other point in favour of their greater appropriateness.
(a)
5. Conclusions
−3
x 10
8
Experimental
6 Integer model The dynamics of a human arm at the elbow joint can be mod-
Fractional model
elled with a commensurate fractional model, of order around 1.3
4
or 1.4, with one zero and two poles. Not only this model structure
2
outperforms integer order models with a similar or slightly supe-
Angle (rad)
−2
−4
References
−6 [1] Taïx M, Tran MT, Ares PS, Guigon E. Generating human-like reaching
−8 movements with a humanoid robot: a computational approach. J Comput Sci
2013;4(4):269–84.
−10 [2] Bar-Cohen Y. Humanlike robots: the upcoming revolution in robotics. Proc SPIE
10 11 12 13 14 15 16 17 18 19 20
2009;7401:7402.
Time (s) [3] Park S, Lim H, Kim Bs, Song Jb. Development of safe mechanism for surgical
(c) robots using equilibrium point control method. In: Larsen R, Nielsen M,
Sporring J, editors. Medical image computing and computer-assisted
intervention – MICCAI 2006. Lecture notes in computer science, vol.
Fig. 6. Part of time responses of obtained nominal models, compared with 4190. Berlin/Heidelberg: Springer; 2006. p. 570–7. ISBN 978-3-540-44707-8.
experimental data, for a type II input. Models are those obtained for: (a) all types; [4] Potkonjak V, Tzafestas S, Kostic D, Djordjevic G. Human-like behavior of robot
(b) Type I; and (c) Type III. arms: general considerations and the handwriting task – Part I: Mathematical
description of human-like motion: distributed positioning and virtual fatigue.
Robot Comput-Integr Manuf 2001;17(4):305–15.
[5] Fu MJ, Cavusoglu MC. Human-arm-and-hand-dynamic model with variability
Table 6 analyses for a stylus-based haptic interface. IEEE Trans Syst Man Cyber Part B:
Ratios of standard deviations to mean values of parameters. Cyber 2012;PP(99):1–12.
[6] Diaz I, Gil J. Influence of vibration modes and human operator on the stability
Input Fractional models Integer models of haptic rendering. IEEE Trans Robot 2010;26(1):60–165.
[7] Speich JE, Shao L, Goldfarb M. Modeling the human hand as it interacts with a
Average for all poles and Gain Average for all poles Gain (%)
telemanipulation system. Mechatronics 2005;15(9):1127–42.
zeros (%) (%) and zeros (%)
[8] de Vlugt E, Schouten AC, van der Helm FC. Adaptation of reflexive feedback
All 24 29 34 244 during arm posture to different environments. Biol Cyber 2002;87(1):10–26.
Type I 25 36 31 60 [9] van der Helm FC, Schouten AC, de Vlugt E, Brouwn GG. Identification of
Type II 39 38 25 463 intrinsic and reflexive components of human arm dynamics during postural
Type III 37 36 54 203 control. J Neurosci Methods 2002;119(1):1–14.
[10] Adewusi S, Rakheja S, Marcotte P. Biomechanical models of the human hand-
arm to simulate distributed biodynamic responses for different postures. Int J
Indust Ergonom 2012;42(2):249–60.
[11] Nagarsheth H, Savsani P, Patel M. Modeling and dynamics of human arm. In:
age of all results. As can be seen, the responses of the fractional or- Proceedings of the 2008 IEEE international conference on automation science
der models are better. Fractional order models were approximated and engineering (CASE’08); 2008, p. 924–8.
[12] Mobasser F, Hashtrudi-Zaad K. A method for online estimation of human arm
by a 10th order transfer function in the [0.1, 100] rad/s frequency dynamics. In: Proceedings of th 28th annual international conference of the
range, using the Crone method [20], to obtain their time domain IEEE engineering in medicine and biology society (EMBS’06); 2006. p. 2412–6.
812 I. Tejado et al. / Mechatronics 23 (2013) 805–812
[13] Venture G, Yamane K, Nakamura Y. Identification of human musculo-tendon [23] Miller KS, Ross B. An introduction to the fractional calculus and fractional
subject specific dynamics using musculo-skeletal computations and non linear differential equations. New York: John Wiley and Sons; 1993.
least square. In: Proceedings of the 1st IEEE/RAS-EMBS international [24] Samko SG, Kilbas AA, Marichev OI. Fractional integrals and
conference on biomedical robotics and biomechatronics (BioRob’06); 2006. derivatives. Yverdon: Gordon and Breach; 1993.
p. 211–6. [25] Malti R, Victor S, Oustaloup A. Advances in system identification using
[14] Magin RL. Fractional calculus in bioengineering. Begell House; 2004. fractional models. ASME J Comput Nonlinear Dyn 2008;3:021401.
[15] Sommacal L, Melchior P, Oustaloup A, Cabelguen JM, Ijspeert AJ. Fractional [26] Palazzolo J, Ferraro M, Krebs H, Lynch D, Volpe B, et al. Stochastic estimation of
multi-models of the frog gastrocnemius muscle. J Vib Control 2008;14(9– arm mechanical impedance during robotic stroke rehabilitation. IEEE Trans
10):1415–30. Neural Syst Rehabilit Eng 2007;15(1):94–103.
[16] Sommacal L, Melchior P, Dossat A, Cabelguen J, Petit JM, Oustaloup A, et al. [27] Valério D, Ortigueira MD, Sá da Costa J. Identifying a transfer function from a
Improvement of the muscle fractional multimodel for low-rate stimulation. frequency response. ASME J Comput Nonlinear Dyn 2008;3(2):021207.
Biomed Signal Process Control 2007;2:226–33. [28] Vinagre B. Modelado y control de sistemas dinámicos caracterizados por
[17] Sommacal L, Melchior P, Cabelguen JM, Oustaloup A, Ijspeert AJ. Advances in ecuaciones íntegro-diferenciales de orden fraccional. PhD thesis, Universidad
fractional calculus: theoretical developments and applications in physics and Nacional de Educación a Distancia; Madrid; 2001.
engineering. Springer; 2007. p. 271–85 [Chap. Fractional Multimodels of the [29] Sanathanan CK, Koerner J. Transfer function synthesis as a ratio of two
Gastrocnemius Muscle for Tetanus Pattern]. complex polynomials. IEEE Trans Autom Control 1963;8:56–8.
[18] Djordjevic VD, Jaric J, Fabry B, Fredberg JJ, Stamenovic D. Fractional derivatives [30] Valério D, Sá da Costa J. Ninteger: a fractional control toolbox for Matlab. In:
embody essential features of cell rheological behavior. Ann Biomed Eng Proceedings of the 1st IFAC workshop on fractional differentiation and its
2003;31:692–9. applications, Bordeaux; 2004.
[19] Mainardi F. Fractional calculus and waves in linear viscoelasticity. an [31] Selen LPJ, Beek PJ, van Dieën JH. Impedance is modulated to meet accuracy
introduction to mathematical models. London: Imperial College Press; 2010. demands during goal-directed arm movements. Exp Brain Res
[20] Valério D, Sá da Costa J. An introduction to fractional control. IET, Stevenage; 2006;172(1):129–38.
2013. [32] Matignon D. Stability properties for generalized fractional differential systems.
[21] Valério D, Sá daCosta J. An introduction to single-input, single-output ESAIM Proc 1998;5:145–58.
fractional control. IET Control Theory Appl 2011;5(8):1033–57.
[22] Podlubny I. Fractional differential equations: an introduction to fractional
derivatives fractional differential equations to methods of their solution and
some of their applications. San Diego: Academic Press; 1999.