You are on page 1of 14

Control Engineering Practice 62 (2017) 22–35

Contents lists available at ScienceDirect

Control Engineering Practice


journal homepage: www.elsevier.com/locate/conengprac

Proportional electro-hydraulic valves: An Embedded Model Control MARK


solution

Wilber Acuña-Bravoa, Enrico Canutob, , Marco Agostanic, Marco Bonadeic
a
Facultad de Ingeniería Electrónica y Biomédica, Universidad Antonio Nariño, Avenida 1ro de Mayo, Bogotá, Colombia
b
Dipartimento di Automatica e Informatica, Politecnico di Torino, Corso Duca degli Abruzzi, 24, Torino 10129, Italy
c
Atos SpA, Via alla Piana 57, 21018 Sesto Calende, Italy

A R T I C L E I N F O A BS T RAC T

Keywords: Hierarchical control architectures are a common approach when hydraulic systems are under study; provided
Disturbance rejection their multi-domain nature, the control scheme is commonly split into different hierarchical levels each one
Uncertainty associated with a particular physical domain. This paper presents the application of a model-based control
Observers structure called Embedded Model Control (EMC) when a hierarchical scheme is implemented on an electro-
Model-based control
hydraulic proportional valve. The overall control consists of two hierarchical loops: the inner loop is the solenoid
Hierarchical control
current regulator with a closed loop bandwidth close to 1 kHz. The outer loop is a position tracking control, in
Systems identification
Proportional valves charge of the accurate positioning of the spool with respect to valve openings. The paper addresses the outer
loop, i.e., the tracking of mechanical spool position by using the EMC. Analysis and synthesis are presented as
well as experimental results obtained from a test rig provided by an industrial manufacturer.

1. Introduction nonlinear back-stepping (Kaddissi et al., 2007), to adaptive sliding


mode control (Guan & Pan, 2008), to nonlinear control (Kayihan &
Proportional-directional hydraulic valves are widely used in most of Doyle, 2000), and heuristic optimization algorithms (Kim & Lee,
the pressure/flow control applications either for industrial or mobile 2006).
hydraulics. They are normally used as flow actuators for hydraulic Flow regulation is the result of the accurate positioning of a shaped
pistons in higher level control loops as well as for directly controlling cylinder, the spool, which is driven by a solenoid and a position sensor,
larger hydraulic valves that cannot be directly controlled by a solenoid. usually a Linear Variable Differential Transformer (LVDT). Solenoid
In the last years, many efforts have been oriented to the design and force being unidirectional requires a contrasting spring assembly for
implementation of digital controllers for hydraulic valves. Reasons for disposing of a differential force. Positioning accuracy is contrasted by
the migration are common to most of control applications. They solenoid plunger friction and affected by hydraulic forces. The paper
include lack of flexibility from production point of view, absence of aims to design and implement a position-tracking controller, with the
fault detection/tolerance schemes, rigidity for parameters tuning and aid of a model-based methodology such as EMC (Canuto, 2007), and to
reconfigurability, among others. The main advantage of digital con- achieve the performance required by this class of high precision control
trollers is that they can be fully developed and implemented as model- devices.
based controllers. Embedded Model Control (EMC) is an enabling The overall controller block-diagram is shown in Fig. 1. It consists
methodology, since it embeds a simple discrete time (DT) model, the of two loops arranged in a hierarchical layout.
Embedded Model (EM), inside the control unit itself. The outer loop corresponds to the position controller. The loop
Recent literature on hydraulic servo systems is mainly focused on tracks the position reference received from external sources, either
hydraulic actuators as in Marusak and Kuntanapreeda (2011), operators or higher-level controllers, within predefined requirements
Kaddissi, Kenne, and Saad (2007) and on manipulators as in such as the response delay and the tracking error bound. The inner
Mohanty and Yao (2011). Control design methodologies are various loop is a solenoid current regulator which tracks the current reference
and range from quantitative feedback theory (QFT) (Karpenko & computed by the outer loop. The design of the inner loop is not the goal
Sepehri, 2010, 2012), to adaptive control (Mohanty & Yao, 2011), to of the paper since it was already presented in Canuto, Acuña-Bravo,
model predictive control (Marusak & Kuntanapreeda, 2011), to Agostani, and Bonadei (2014a, 2014b). The current regulator, designed


Corresponding author.
E-mail addresses: wilber.a.b@uan.edu.co (W. Acuña-Bravo), enrico.canuto@polito.it (E. Canuto), magostani@atos.com (M. Agostani), mbonadei@atos.com (M. Bonadei).

http://dx.doi.org/10.1016/j.conengprac.2017.01.013
Received 15 January 2016; Received in revised form 24 January 2017; Accepted 25 January 2017
0967-0661/ © 2017 Elsevier Ltd. All rights reserved.
W. Acuña-Bravo et al. Control Engineering Practice 62 (2017) 22–35

Fig. 1. Block scheme of the two level control system.

using the EMC methodology, was employed as the inner loop of a


digital PID position controller, an alternative to the EMC position
controller of this paper. A third alternative was the analogue PID Fig. 3. Test rig used for experimental test. 1 – Proportional valve, 2 – Variable-stroke
controller employed by the manufacturer in the commercial devices. A load valves, 3 – ON–OFF valve, 4 – Pressure-relief valve, 5 – Pressure sensors, 6 –
Electric Motor, 7 – Tank.
comparison between analogue and digital PID is available in Canuto
et al. (2014b). Comparison of the three alternative controllers will be
discussed in Section 4. A further alternative – a future development – 1. The hydro-mechanical part performs flow regulation by an accurate
is to abandon the hierarchical structure and design in favour of spool positioning. The open-loop dynamics depends on mass, spring
controller based on a single EM. By the way, hierarchy is justified by compliance, and fluid properties.
the order of magnitude between the small signal bandwidth (BW) of 2. The solenoid part produces the plunger unidirectional force that
the position and current controller, from about 100 Hz to 1 kHz. contrasts the recovery spring. Balance of spring and solenoid forces
As a baseline, migration from analogue to digital controller and to generates the differential force which regulates the spool position
EMC controller was not driven by the compelling need of improved around the hydraulic zero (zero flow). Since plunger mass is about
performance, but by the aim of having a class of models available, ten times larger than the spool, the 50 Hz open loop BW of the force
which is coherent from simulation to controller and is capable of to position dynamics is fixed by plunger mass and spring compli-
describing uncertainty and dynamic limits of a class of manufactured ance.
items. A preliminary target of the class of models and of the relevant 3. The position transducer measures the plunger position, which is the
controller is the repetition of the current performance. This was the true controlled variable. Discrepancies between plunger and spool
goal of the control design addressed in this paper. For instance, the position set up only at very high frequencies, greater than the
high frequency non-minimal phase behaviour pointed out in Sections position BW of 100 Hz, since contact between plunger and spool is
2.3.2 and 4.4, though occurring at a frequency larger than the target ensured by the recovery spring. Low-frequency discrepancies (bias)
BW, is a key model finding for avoiding command saturation. With this are eliminated by making the position sensor to coincide with the
aim in mind, it was imperative a step-by-step progress and the hydraulic zero.
comparison in Section 4 between the three alternatives: analogue
controller, discretized analogue controller with digital current regula-
tor, and digital EMC controller. 1.2. Hydraulic experimental setup

The experimental setup in Fig. 3 has been used for modelling,


1.1. Plant description control synthesis, and tuning and for testing experimental perfor-

The proportional electro-hydraulic valve of this paper has four


inlet/outlet ports that can be interconnected: A, B, tank (T) and pump
(P). The valve control system regulates the flow-passing from one port
to another by micro-positioning the spool across the apertures. No
spool overlapping has been considered in control design. Fig. 2 shows a
simplified scheme of the electro-proportional valve. The valve is
subdivided into three main parts:

Fig. 2. Simplified scheme of the proportional valve. 1 – Recovery spring, 2 – Spool, 3 –


Sleeve, 4 – Solenoid, 5 – Plunger, 6 – Position transducer. Fig. 4. Hydraulic circuit of the test rig in Fig. 3.

23
W. Acuña-Bravo et al. Control Engineering Practice 62 (2017) 22–35

Fig. 5. Essential block-scheme of the EMC.

mance. Fig. 4 shows the hydraulic circuit of the test rig in Fig. 3. The The model error is the key and unique error feedback from the plant
elements of Fig. 4 are numbered in the same way: to the embedded model and as such encodes all the past and present
deviations between plant and model. Deviations are due to exogenous
1. The proportional valve is the element to be controlled. unpredictable signals (causal uncertainty), parametric uncertainty and
2. The variable-stroke load is a manually operated valve connecting neglected dynamics. The same signal is known as innovation in Kalman
port A with port B. filters but parametric errors and neglected dynamics should be
3. The ON–OFF valve is a hand-lever-directional valve distributing or negligible to avoid filter divergence. The same signal is the key feedback
not the hydraulic fluid to the circuit. of IMC (Francis & Wonham, 1976), but it is directly fed back to the
4. The relief-pressure valve can be manually adjusted to fix the limits of control law, and not to the internal model, thus hiding uncertainty
the fluid pressure in the hydraulic circuit. prediction and compelling robust control law design. The same signal is
5. Pressure sensors, ⑥ pump electric motor, ⑦ and tank. the key feedback of ADRC (Gao, 2006; Huang & Sira-Ramírez, 2015;
Yi et al., 2014; Zheng & Gao, 2014) and of Disturbance-Observer-
Adjustment of the different valves allows the proportional valve to Based Control (Li, Yang, Chen, & Chen, 2014; Wei & Guo, 2009), but
be operated under three different conditions: the uncertainty/disturbance dynamics is reduced to be first-order and
the model error is fed back to the whole state vector like in Kalman
NOP condition: The test rig is off. Neither pressure drop sets up be- filters, in which arrangements may disagree with the actual uncertainty
tween ports A and B (besides the one set up by the fluid layout. Such differences do not mean that the overall feedback transfer
inside the valve) nor flow occurs across ports A and B. NOP function, from measurements to command, as a result of EMC design,
conditions are obtained by closing the ON–OFF valve (OFF) cannot be synthesized by other methods, but basically only in the linear
in Fig. 3 or by switching off the pump electric motor. ⑥ case under simple uncertainty models. As matter of fact, Canuto (2015)
OPEN condition: The test rig is operating, and the load-valve ② con- has proved that a class of EMC state predictors including a dynamic
necting ports A and B are completely open (ON). The pres- feedback is not equivalent to Kalman filters whose feedback is only
sure drop at the output of both ports A and B is barely the static.
same The essential EMC block-diagram is in Fig. 5. The embedded model
CLOSE condition: The test rig is operating. The load-valve ② connect- which consists of M (controllable dynamics) and D (uncertainty/
ing ports A and B are closed (OFF). The pressure drop at the disturbance dynamics) separates the measurement law (noise estima-
output of both ports A and B is at the maximum value (line tor and reference generator) from the model-based control law.
pressure at the output of ④). One of the ports is connected to The EMC features the following characteristics:
the line pressure and the other to the tank.
• The ‘unknown’ disturbance dynamics (D in Fig. 5) may be any and is
driven by an unpredictable input vector w(i) to be real-time
estimated.
1.3. The Embedded Model Control • The noise input layout, to be designed, establishes which state
variables are directly affected by uncertainty.
The EMC is a model-based control methodology outlined in Canuto • The noise input is estimated by a dynamic feedback (the ‘uncer-
(2007) and recently in Canuto, Pérez-Montenegro, Colangelo, and tainty/noise estimator’) driven by the model error em(i).
Lotufo (2014d). A comparison between EMC, internal model control • A further design degree of freedom, the control ‘free dynamics’
(IMC) and active disturbance rejection control (ADRC) is in Canuto, similar to Youla parameterization (Maciejowski, 1989) is viable, but
Pérez-Montenegro, Colangelo, and Lotufo (2014c). not yet included in the application.
A real-time model of the plant, driven by the same command
dispatched to the plant (quantization error and saturation should be The disturbance dynamics may be designed to have any order, it
included), known as the embedded model (EM), parallels the plant and must be parameter-free unless monochromatic random processes have
provides the measured model error to be tracked, it is driven by a zero mean arbitrary signal w(i) (noise) to
be real-time estimated from model error. The noise layout, to be
em (i ) = y (i ) − ym (i ). (1)
designed, establishes which state variables are directly affected by

24
W. Acuña-Bravo et al. Control Engineering Practice 62 (2017) 22–35

uncertainty, and implies that the feedback from model error to noise to fluid compressibility sets up. It will be treated as a neglected
(the noise estimator) may be dynamic and not static. A static estimator dynamics and adequately filtered.
will be adopted in the paper for reasons to be clarified. State predictor
eigenvalues are designed for discriminating in the model error only the The fourth-order single body dynamics of the pair plunger and
residual uncertainty to be cancelled (uncertainty based design). Noise spool is given by:
estimator and embedded model constitute a stable state predictor
x˙s (t ) = vs (t ),
which supplies the control law with the model state, and allows control
gains to be model-based designed (a kind of separation principle). The xs (0) = xs0 mv˙s (t ) = Fs (xs , I , I˙) − Fa (xs , vs, ξ, ξ˙) − Ff (xs , vs, ΔP )
reference generator shapes a reference trajectory for each controllable 1
− FK (xs ), vs (0) = vs0 c˙ (t ) = (−c (t ) + bs (I , I˙)), c (0) = c0
state of M so as to meet requirements and to respect actuator range τc
and slew rate. Fs (t ) = F s (xs , c ),
The rest of the paper is organized as follows. Section 2 presents a
summary of the governing equations, as well as the system operating ξ˙ (t ) = g (ξ, vs ), ξ (0) = ξ0. (5)
boundaries. Section 3 presents the EMC basic ideas and the applica- The equation is fourth order but the relative degree between current
tion, requirements, design model, noise estimator and control law. and position is only three, since the variable ξ enters a velocity-force
Section 4 presents a summary of preliminary experimental results feedback as shown below. The variable c(t) accounts for the third pole
showing that performance requirements are satisfied. The nomencla- dynamics and increases the relative degree from two to three. Eqs. (5)
ture is reported in Appendix A. The main variables are listed in Table are valid within the stroke region, i.e., between the rest position Xmin
A1. of the spring assembly and the end stroke Xmax, defined by
Xmin ≤ xs ≤ Xmax , Xmin ≈ −2.5 mm, Xmax ≈ 1.25 mm, (6)
2. Dynamic model and identification
and within the frequency band dc to 300 Hz. The force components and
2.1. Current regulator state variables in (5) are the following:

Solenoid dynamics and digital current regulator are treated in 1. The force FK (xs ) = Kxs − FK 0 accounts for elastic forces applied by the
Canuto et al. (2014a). The current regulator tracks the driving current recovery spring, where K=10,000 N/m is the stiffness and −FK 0
demanded by the position control within a sufficiently wide BW. preload.
Requirements are expressed by fixing a maximum delay between true 2. xs(t) and vs(t ) are respectively the displacement and velocity of the
and reference currents I(t) and Ir(t) and a maximum tracking error as spool and plunger pair from the sensor zero-position. The offset
follows. from the hydraulic zero is calibrated and corrected. The overall mass
The solenoid current dynamics as well as the digital controller have m is about 0.1 kg, and accounts for both moving elements (plunger
been presented in Canuto et al. (2014a). As it was said before, the and spool). The mass–spring frequency fo = K / m /2π ≈ 50 Hz and
current regulator is the inner loop in the valve control architecture. The the viscous damping ζ > 1, being greater than one, define the open-
task of the inner loop is to enlarge the open-loop bandwidth. The closed loop bandwidth. The viscous damping is variable with the fluid
loop dynamics from current reference Ir(t) to solenoid current I(t) has temperature and the hydraulic flow/pressure conditions. The open-
been designed to approximate a 0.3 ms delay plus noise loop bandwidth may vary between fo for ζ = 1 and 2ζfo for ζ ≫ 1.
I (t ) = Ir (t − τI ) + ηI (t ) (2) 3. Ff (xs , vs, ΔP ) denotes the fluid viscoelastic forces, dependent on the
pressure drop ΔP = PA − PB . The force expression satisfies
ηI (t ) ≤ 1.5 mA, τI = 0.3 ms (3) Ff (0, 0, ΔP ) = 0 and Ff (x, 0, 0) = 0
4. Fa (xs , vs, ξ, ξ˙) denotes the friction forces, which are assumed to fit the
The above model only applies to a class of reference signals defined
LuGre model (Canudas de Wit, Olsson, Aström, & Lischinsky,
by the current range 0 ≤ Ir (t ) ≤ Imax and by the current slew rate
1995). The state variable ξ (t ) and the rate ξ˙ (t ) account for surface
I˙min (t ) ≤ I˙r (t ) ≤ I˙max . Slew rate is imposed by the PWM voltage range
(bristles) deformations. The force expression satisfies
|V (t )| ≤ Vmax = 24 V and by the high-frequency solenoid inductance.
Fa (xs , vs, 0, 0) = sgn(vs ) Fa0 . A lot of research has been done around
The slew rate range is asymmetric because of the current bias I0 of
the friction topic associated with hydraulic systems and valves, see
about 1.5 A fixing the hydraulic zero. The corresponding voltage offset
for example Garcia (2008) and Farenzena and Trierweiler (2012).
V0 = I0 R = 6 V renders the differential voltage range −30 V to 18 V and
Integration of ξ (t ) does not add a further relative degree between
the current derivative range asymmetric. The current derivative range
force and position. In fact, the shortest integration path that defines
was experimentally determined as
the relative degree between force and velocity and force and position
I˙min ≈ −640 A/s ≤ I˙ (t ) ≤ I˙max ≈ 360 A/s. (4) is one and two respectively.
5. Fs (xs , c ) is the electromagnetic solenoid force, which is affected by
magnetic hysteresis. At zero position, it is balanced by the preload
2.2. Plunger/spool dynamics
KX0 of the recovery spring and by the static friction Fa0 as
Even if spool and solenoid plunger are two different bodies, they Fs (0, c0 ) = KX0 − sgn(vs ) Fa0. (7)
may be assumed to be rigidly connected at least in the open-loop
Because of sgn(I˙) (due to magnetic hysteresis) and sgn(vs ) (due to
frequency band from dc to 50 Hz, which is dictated by mass and spring
friction), the bias current I0 cannot be uniquely defined, but it varies
compliance. Deviation from this assumption, as pointed out by
by fractions of amperes, about 1.5 A. A mean value of the current to
identification results in Section 2.3.2, is twofold.
force gain is ϕs (t ) = 30 N/A .
1. Above 50 Hz, a third order dynamics is identified, due to an
Fig. 6 shows the Bode diagrams of the current–force transfer
additional pole which accounts for a few millisecond pole between
function from two different tests corresponding to different plunger
solenoid current and plunger force, as explained below. The pole
positions (solid and dashed lines). The fitting diagrams (circled line)
cannot be neglected, and justifies the double derivative feedback of
show that the transfer function can be explained by a first order
the analogue PID.
dynamics plus a delay. The delay corresponds to a PWM cycle 0.1 ms
2. Above 300 Hz a complex non-minimum phase dynamics mainly due
and can be neglected, unlike the time constant τc = 1–2 ms (see

25
W. Acuña-Bravo et al. Control Engineering Practice 62 (2017) 22–35

Fig. 6(a)). Even though the magnitude is accurately explained by a first


order dynamics, the same cannot be said of the argument as it shows a
constant shift, which may be explained by a backlash phenomenon
(diamond line). Fig. 6(b) shows the phase plots, from experimental
data and fitting models. The overall dynamics is the same of c˙ (t ) in (5)
where bs (·) is the backlash dead-band, c(t) is the current dynamics
affected by backlash, and Fs (xs , c ) is the average force profile in Fig. 7.
The latter is plotted versus solenoid current and plunger position.
During the two-way test, current has been kept constant. The force
discrepancy in Fig. 7 is due to static friction.
The measurement equation at time ti = iT , with T=0.1 s, already
converted into length units, is
yx (iT ) = xs (iT ) + em (t ), (8)
where em(t) is the model error accounting for model discrepancies
from (5) and for measurement errors. In the following, ti will be
replaced by the integer i.

2.3. Embedded model from experimental data

2.3.1. Model classes


The goal is to convert (5) into a DT state equation from the solenoid
current 0 ≤ I (iT ) ≤ 3, to the measured plunger position y(iT) in (8). A
DT is requested by EMC. The model has been derived from experi-
mental data in order to identify the parameters of (5), and to assess the
relative degree and the frequency band. As a by product, information
about parametric uncertainty and model error is expected.
Due to a complex relationship between exogenous variables like
hydraulic pressure drop ΔP and fluid temperature θ, and state variables
like velocity vs and position xs, parameters are lumped in the vector
p (ΔP, θ , xs , vs ), which is decomposed into a nominal and constant value
p to be a priori identified and into an unknown bounded deviation δp to
be predicted in real-time:
p (ΔP, v, θ , t , ·) = p + δp (ΔP, v, θ , t , ·). (9)
The discrete linear time-invariant state equation can be written as

x (i + 1) = A ( p) x (i ) + B ( p) u (i ) + h (x, u; p), x (0) = x0


yx (i ) = Cx (i ) + em (i ) = ym (i ) + em (i )
p ≤ pmax , (10)
where ym(i) is the model output, yx(i) is the measured output already
defined in (8) and em(i) is the model error. h (·) is a vectorial term
Fig. 6. Current–force transfer function responses.
accounting for the effect of p on the state vector. It is assumed that δp
and h (·) are continuous in their arguments. One of the strategies of the
control law is to predict and reject h (·) through the command u(i) in
the same way as an unknown disturbance is cancelled. The model error
em (i ) = yx (i ) − ym (i ) can be rewritten as the sum of the measurement
error wy(i) (including random and systematic components) and the
effect of the neglected dynamics in (10):
em (i ) = wy (i ) + ∂P ( ym). (11)
In view of the reference generator design in Section 3.3, it should be
observed that (10) is asymptotically stable and without unitary zeros.
Therefore, given any output value ye within the operational limits (5),
|ye | ≤ Xmax , a unique disturbance-free equilibrium (xe , ue , he = 0) can be
found which satisfies
⎡ 0 ⎤ ⎡ A ( p) B ( p)⎤ ⎡ xe ⎤
⎢y ⎥ = ⎢ ⎥ ⎢ ⎥.
⎣ e⎦ ⎣ C 0 ⎦ ⎣ ue ⎦ (12)
The absence of unitary zeros guarantees that the system matrix in
(12) is invertible. The dynamic operator ∂P in (11), known as fractional
error dynamics, is driven by the model output. In the LTI case, the
transfer function can be written as ∂P = M−1P − I , where M is the
embedded model and P the true input–output dynamics. Since P has a
Fig. 7. Experimental position-force relationships. relative degree larger than M, ∂P is causal and has zero relative degree.

26
W. Acuña-Bravo et al. Control Engineering Practice 62 (2017) 22–35

Fig. 8. ΔP for three diverse hydraulic cases. Fig. 9. PSD of Sx /SI for three hydraulic conditions.

fluid visco-elastic forces, especially of the elastic term. The deviation


It is a delicate design task to separate h in (10) from ∂P in (11). In fact,
can be modelled as a significant component of h in (10) to be predicted
residual unpredicted components of h (at higher frequencies) combine
and rejected. NOP condition shows a deviation from the CLOSE
with ∂P and low-frequency components of ∂P must be converted into
conditions around 300 Hz which is likely due to a different friction
additive disturbance terms like h in order to be cancelled. Experimental
contribution. In this case, the deviation can be modelled as a significant
identification is an important step in this sense.
component of the neglected dynamics ∂P in (11), since it cannot be
predicted and cancelled by the controller command. As a baseline,
2.3.2. Embedded model identification
since it is a rather significant amplification of the transfer function, its
Fig. 8 shows the pressure difference ΔP for the three hydraulic
contribution to the real-time embedded model should be eliminated.
conditions described in Section 1.2. As mentioned before, when the
To fit the EM (10) to the experimental model (13), Eq. (10) is
load is OPEN, the flow is high, the pressure drop through the load is
rewritten as a Z-transform as follows:
small, and ΔP is low. When the load is CLOSE, i.e., no flow exists
through the ports A and B and the pressure jumps between the Bj (z )
xs (z ) = Mj (z ) u (z ) + dj (z ), Mj (z ) =
maximum Pp,max (pump pressure) and the minimum value −Pp,max Aj (z )
depending on the port which is connected to the pump (dash-dotted
yx (z ) = (1 + ∂Pj (z )) xs (z ) + wy (z ) (14)
curve in Fig. 8).
Identification has been made by spectral analysis (Ljung & Glad, where j=1,2 distinguishes between two different models to be discussed
1994). To this end, the solenoid driving voltage V (actually the PWM below. Aj(z) and Bj(z) are the following polynomials:
cycle) has been randomly varied around the offset V0 that moves the
plunger to the hydraulic zero and corresponds to the current I0. The Aj (z ) = a1j + a2j z+⋯+anj z l −1
current regulator is ineffective during the identification interval. The Bj (z ) = 1 + b1j z+⋯+blj z l . (15)
flat (unilateral) spectral density of the driving differential voltage
ΔV (i ) = V (i ) − V0 is the top curve in Fig. 9. The flat spectrum confirms and dj(z) is a disturbance signal which includes those components of
that the driving signal is a pseudo-random signal. h (·) that cannot enter a linear model.
The differential solenoid current yI (i ) = I (i ) − I0 + eI (i ) is measured A modified algorithm of Levi (1959) has been employed for fitting
together with the plunger position yx(i) in (8). The current spectral the polynomials in (15) to experimental data. The degrees m ≤ l and
density is the mid-decreasing curve in Fig. 9. The three bottom curves l ≤ 4 are a compromise between order and neglected dynamics
in Fig. 9 are the magnitude of the current–position transfer function minimization. In other terms, model (5) with a relative degree of three
which is valid up to 300 Hz has been extended beyond such a boundary
SxI (jf )
PxI (jf ) = . in order to minimize the neglected dynamics ∂Pj in (11). To better
SII (jf ) (13)
clarify model error minimization, two models denoted by j=1,2 have
They have been obtained from the complex spectral ratio between the been identified. Model 1, M1 (z ), is third order, that is l=3, has degree
position/current cross spectrum SxI(jf) and the current auto-spectrum r = l − m = 3 relative degree, and a larger neglected dynamics. The
Sx(f). The cross spectrum SxI(jf) and consequently the transfer identified parameters are very close to manufacturer data. Model 2,
function PxI(jf) have been derived under the three different hydraulic M2 (z ), is fourth order, that is l=4, and has r = l − m = 1 relative degree,
conditions (OPEN, CLOSED and NOP). The magnitude Bode plots of and neglected dynamics since the high-frequency domain becomes
PxI (jf ) in Fig. 9 are rather insensitive to hydraulic conditions except at accurately repeated.
very low frequencies and around 300 Hz which is the boundary Fig. 10(a) and (b) shows magnitude and argument of the experi-
between modelled and neglected dynamics. CLOSE condition has been mental transfer function Px I (jf ) in (13) for the worst-case CLOSE.
identified to be the worst-case for industrial applications. The low- Experimental data are accompanied by the magnitude and argument of
frequency deviation from the CLOSE condition mainly concerns the M1 (jf ) and M2 (jf ) in (14). |M1 (jf )| perfectly fits |PxI (jf )| up to 300 Hz,
OPEN condition (the continuous-line bottom plot with the smallest whereas the argument arg M1 (jf ) requires the additional phase shift of
profile for f < 10 Hz ). The deviation is due to a significant change in the the backlash as in Section 2.2. M2 (jf ) fits Px I (jf ) in magnitude and

27
W. Acuña-Bravo et al. Control Engineering Practice 62 (2017) 22–35

phase beyond 300 Hz up to about 2 kHz. Fitting is obtained by adding


of a single pole that increases l from 3 to 4, and three zeros that
increase m from 0 to 3 and decrease the relative degree r from 3 to 1.
M2 (jf ) is non-minimum-phase since the relative degree r=1 of a
minimum phase transfer function would require an asymptotic argu-
ment of −π /2 , instead of an argument smaller than −π . In fact, data
fitting forces one of the three zeros, the highest frequency zero, to lie
outside of the unit circle.
The resulting magnitude |∂Pj |, j = 1, 2, of the neglected dynamics is
shown in Fig. 11. As a result of the above fitting |∂P2 | ≤ |∂P1 | in the
whole frequency band, and |∂P2 | < 1 at higher frequencies.

3. Embedded Model Control

3.1. Requirements

Position tracking requirements are similar to those of the current


regulator. Given a reference position signal yr(t), the plunger position
xs(t) must be delayed and perturbed as follows:

xs (t ) = yr (t − τx ( yr ,max)) + ηx (t )
τx ( yr ,max) ≤ 5 ms, ηx (t ) ≤ 10 μm, (16)
where ηx is the residual tracking error defined as the difference
between the total tracking error ex (t ) = yr (t ) − xs (t ) and the nominal
tracking error e x (t ) imposed by the delayed reference signal yr (t − τx ),
as follows:
ηx (t ) = ex (t ) − e x (t ), er (t ) = y˙r (t ) − yr (t − τx ). (17)
The maximum delay in (16) refers to the maximum stroke 2Xmax
in (6), and becomes smaller for shorter strokes down to 3 ms.
Continuous-time notations are used since requirements should be
fulfilled by the true plunger position xs(t). The true plunger position
may be approached from real data by averaging the random compo-
nents of the measurement noise, in the absence of low-frequency
components (bias).

3.2. The embedded model

In order to cancel the parametric uncertainty encoded by h (·), the


EM (10) is completed with a disturbance dynamics. The disturbance
class + whose realization is denoted by the vector d (i ) is designed to be
a second order chain of integrators,
Fig. 10. Frequency responses of the identified model.
xd (i + 1) = Ad xd (i ) + wd (i )
d (i ) = Hc xd (i ) + wu (i ), (18)
with

xd (i ) = [xd1 xd 2 ]T (i )
wd (t ) = [wd1 wd 2 ]T (i )
wu (i ) = [wu1 wu2 wu3 wu 4 ]T (i )
⎡ ⎤ ⎡ 0 0 0 1 ⎤T
Ad = ⎢ 1 1⎥ , Hc = ⎢⎣ 0 0 0 0 ⎥⎦ .
⎣ 0 1⎦

The second order dynamics in (18) can be interpreted either in a


stochastic sense by assuming the driving noise vectors wd (i ) and wu (i ) to
be white noise, in which case their statistics should be added, or in a
deterministic sense, by assuming that the input vectors are bounded
arbitrary signals which at any step may change the state vector xd . In
this sense the class + is a class of piecewise polynomials and no
statistics must be made explicit. The disturbance d (i ) vector is a
combination of the state xd1 (i ) which synthesizes the piecewise poly-
nomial and of the arbitrary components of wu (i ). xd1 (i ) directly affects
only one controllable state, whereas wu affects all the state variables.
Fig. 11. Magnitude plot of Sm, of Vm, and of the neglected dynamics ∂P1 and ∂P2 . This is the result of the noise design, which is the delicate task deciding
which and how state variables are affected by uncertainty. Here the aim

28
W. Acuña-Bravo et al. Control Engineering Practice 62 (2017) 22–35

Fig. 12. Block-scheme of the embedded model.

is a simple feedback design and a simple eigenvalue tuning of the state 1. Current delay and tracking error in (2) and (3) must be designed to
predictor in Section 3.2.1. At the same time to correct the uncertainty be compatible with position requirements. In fact, their contribution
affecting each controllable state. The whole EM is the combination of is at least one order of magnitude smaller than position require-
(10) and (18) upon replacement in (10) of h (i ) with d (i ). If the fourth- ments in (16):
order model M2 in (13) is employed, the overall state vector is
partitioned into controllable state vector x (i ) = [x1 x2 x3 x4]T (i ) and |ηI |
τI = 0.3 ms ≪ 3–5 ms, ≈ 0.5 μm ≪ |ηx | ≤ 10 μm
disturbance state vector xd (i ). The model equation written in compact K
form is

2. The disturbance class + should include, though small, the current


command error.

3.2.1. The state predictor


The state predictor aims to predict the embedded model state with a
with bounded prediction error. Here it is preferred to use the word
prediction instead of estimation since the variables dispatched to the
⎡ − a1 1 0 0⎤ ⎡ b1 ⎤
⎢ ⎥ ⎢ ⎥ control law are one-step predictions, which in fact allows the next step
−a 0 1 0⎥ b
Ac = ⎢ 2 , Bc = ⎢ 2 ⎥ and Cc = [1 0 0 0]. command to be predicted without any delay. A bounded prediction
⎢ − a3 0 0 1⎥ ⎢ b3 ⎥ error requires a stable state predictor, which is ensured by closing the
⎢⎣− a4 0 0 0 ⎥⎦ ⎢⎣ b4 ⎥⎦
loop between the measured model error yx (i ) − ylm (i ) and the noise
vectors in (18), where ylm (i ) is the realization of the model output ym(i)
The controllable pair (Ac , Bc ) has the controllable form imposed by in (19). Since the noise vectors wu (i ) and wd (i ) perturb all the six state
the fourth order transfer function M2. The overall embedded model is variables (four controllable states and two disturbance states), the
not stabilizable because of the unstable disturbance dynamics (two stabilizing feedback is static and the input noise vectors have the
cascaded unit eigenvalues). The only stabilization strategy is a dis- following expressions:
turbance rejection. Since EMC methodology postulates that plant and
embedded should receive the same command, the question is which is ⎡ wu ⎤ ⎡Lw⎤
the current command u(i) in (10) and (19) to be adopted. The plant ⎢ w ⎥ (i ) = ⎢ L ⎥ ( yx (i ) − ylm (i )).
⎣ d⎦ ⎣ d⎦ (21)
clearly receives the true current I(t), which is unknown. The answer
comes from the expression (2) of the true current, in which less a short The ‘overbar’ notation is such as to distinguish estimation from one-
delay and a tracking error are equal to the reference. This implies the step prediction which is denoted by a ‘hat’. In fact, a static feedback
simple decision (noise estimator) depends on the current measurement, whereas state
variables are dependent on the past measurements. State correction by
u (i ) = Ir (i ). (20)
the current measurement is of no use since it does not allow prediction.
Identity (20) entrains two design objectives: Combination of (19) and (21) yields the equations of the state predictor

29
W. Acuña-Bravo et al. Control Engineering Practice 62 (2017) 22–35

of an output disturbance due to exogenous variables (say pressure drop


and temperature), the tracking error equation becomes
(1 − F (z )) ηx (z ) = S (z ) H (z ) M2 (z ) d y (z ) − V (z ) wy (z ) + F (z ) yr (z ), (26)
where
F (z ) = S (z ) H (z ) M2 (z ) − V (z )∂P2 (z ), (27)
S is the sensitivity of the whole control system, and V = 1 − S is the
complementary sensitivity. By assuming separation between the state
predictor and the control law tuning, state predictor spectrum can be
tuned by replacing S and V in (26) with the state predictor Sm and Vm,
where
yx (z ) − ylm (z ) ylm (z )
Sm (z ) = , Vm (z ) = .
yx (z ) yx (z ) (28)
A sufficient stability condition provided by small gain theorem is
that
max |F (jf )| ≤ ε < 1
|f |≤ fmax (29)
where 1/ ε is the gain margin. Since max|f | < fmax |H (jf ) M2 (jf )| ≤ 10 and
Fig. 12 is the detailed block diagram of (22). max|f |, fmax |∂P2 (jf )| < 0.1, stability is guaranteed by enlarging the BW of
The observer gains are given by the sensitivity Sm. The low-frequency asymptote of the sensitivity is
( f / fm )2 because of the second-order disturbance dynamics in (18): the
⎡ c5 − a1 − 4 ⎤
⎢ ⎥ two unitary poles become zeros of the sensitivity. Thus, in order to
c − 3 c − a + 6
Lw = ⎢ 4 5 2 ⎥ attain the same level, about 0.1, of the neglected dynamics, it is
⎢ c3 − 2c4 + 3c5 − a3 − 4 ⎥ sufficient that fm ≥ 100 Hz . In that case, the left-hand side of (26) can
⎢⎣ c − c + c − c − a + 1⎥⎦
2 3 4 5 4 be assumed to be equal to 1 − F (z ) ≈ 1 and the state predictor
⎡c ⎤ spectrum Λm can be tuned to meet requirements in (16) versus the
L d = ⎢ c1 ⎥ ,
⎣ 0⎦ (23) perturbations in the right-hand side of (26).
It should be observed that also the reference signal may become a
where c0, …, c5 are the coefficients of the characteristic polynomial perturbation through parametric uncertainty and neglected dynamics.
pd (γ ) = γ 6 + cm5 γ 5 + cm 4 γ 4 + cm3 γ 3 + cm2 γ 2 + cm1 γ + cm0 with λ = γ + 1 Assume, for instance, a sinusoidal reference of 1 mm amplitude and the
imposed by the desired sixth-order spectrum Λm = {λ m0, …, λ mnw}. frequency |H (jf ) M2 (jf )| which was equal to 10 Hz. In order to guarantee
Tuning of the state predictor spectrum Λm is the key step in the a tracking error below 10 mm as in (16), attenuation must be larger
EMC design (Canuto, Acuña-Bravo, Molano-Jiménez, & Pérez- than 1000 which requires fm > 300 Hz . The magnitude plots of Sm and
Montenegro, 2012, 2013). The relevant stability and performance Vm are in Fig. 11.
equation involve the position tracking error ηx (i ) defined in (16), and
a linear Z-transform relation between ηx (i ), the parametric uncertainty 3.3. The reference dynamics
h (i ) and the fractional neglected dynamics ∂P . To this end, a worst-case
linear dynamics of ∂P and h (i ) should be identified from experimental The reference generator provides the open-loop command u (i )
tests or physical considerations. In this case, the adoption of M2 has the which drives the state vector x (i ) of the controllable dynamics (10),
advantage of neatly separating the frequency domains of h, below few free of uncertainty, say with h = 0 , that is
tens of Hz and that of ∂P2 above 100 Hz, as Fig. 10 suggests. |∂P2 | is
plotted in Fig. 11 together with |∂P1 |, and whereas |∂P1 (jf )| > 1 for x (i + 1) = A ( p) x (i ) + B ( p) u (i ),
f > 400 Hz , |∂P2 (jf )| < 1 in the whole frequency domain from dc to the x (0) = x 0 y x (i ) = C x (i ). (30)
Nyquist frequency fmax =5 kHz . Fractional error magnitude and phase
may change with fluid temperature and pressure. Command synthesis aims that the output y x tracks the exogenous
A worst-case linear model of h can be approximated by comparing reference yr (i ) and that at the same time requirements in (16) are
the identified PxI (jf ) under CLOSE and OPEN conditions in Fig. 9, satisfied. The natural way would be a synthesis of u as an output-to-
where CLOSE has been selected for fitting the embedded model. input feedback driven by the output tracking error yr (i ) − yx (i ). The
Denote them with PCLOSE and POPEN respectively. It is not difficult output-to-input synthesis can be converted into a state feedback
to prove that synthesis if an intermediate state reference xr(i) and command
reference ur(i) are made available. The intermediate pair (xr , ur ) is
PCLOSE (s )
POPEN (s ) = , selected to be the disturbance-free equilibrium solution of Eq. (12)
1 + H (z ) PCLOSE (s ) (24) given the reference output yr(i).
where z has been replaced by s since the frequency domain is far from The equilibrium pair (xr , ur ) is tracked by the nonlinear state
the Nyquist frequency fmax. Since PCLOSE and POPEN are pretty feedback
equal at higher frequencies, that is, u (i ) = sat(ur (i ) + K (xr (i ) − x (i ))), (31)
lim POPEN (s ) = lim PCLOSE (s ), where sat(·) means
s →∞ s →∞ (25)
⎧u |u| < u max ,
the feedback H(s) only changes the poles, is a polynomial of degree sat(u ) = ⎨
⎩ sgn(u ) u max |u| > u max , (32)
μ ≤ 2 and, from Fig. 9, it is possible to estimate the upper bound
10K
maxf |H (jf )| < ϕ ≈ 3 A/mm , which is attained at lower frequencies, and where u max = Imax − I0 is the upper bound of the differential current
s
f < 10 Hz . The degree μ implies that H (z ) PCLOSE (s ) has a relative degree ΔI = I − I0 around the offset I0.
r ≥ 1. By replacing h with H (z )( ym (z ) + d y (z )), where dy plays the role The gain K is obtained by fixing the closed-loop spectrum Λr of

30
W. Acuña-Bravo et al. Control Engineering Practice 62 (2017) 22–35

Table 1
Embedded Model Control eigenvalues.

Control section Symbol Value

Uncertainty estimator Λm [0.43 0.7 0.8 0.86 0.91 0.94]


Control law Λc [0.43 0.95 0.95 0.95]
Reference dynamics Λrg [0.85 0.85 0.75 0.75]

first sampled at 100 kHz with a 14-bit accuracy by a NI PCI-6289


Fig. 13. Block-diagram of the reference generator. board, then averaged and down-sampled at 10 kHz. The pressures of
the ports A and B and of the tank are also measured for analysis and
identification, but do not enter the control algorithms. The solenoid
A ( p) − B ( p) K so that the delay requirement in (16) is satisfied. The voltage command is converted into the PWM duty-cycle at 10 kHz
resulting delayed and smoothed trajectories of the position yx(i) can be with a 12-bit accuracy, which corresponds to a time resolution of
appreciated in Fig. 15 (for triangular reference) and Fig. 18 (for square 25 ns.
wave reference). The solution of the steady-state condition xr (i ) is given 5. Current regulator and valve position controllers are updated at
by 10 kHz.Fig. 14 shows the block-diagram of the overall control
algorithm explained in Section 3. All the results reported in the paper
have been obtained from experiments on the test rig of Fig. 3. A
MATLAB/SimulinkTM valve simulator was developed for preliminary
tuning and test of control algorithms, but details and results are not
and is graphically depicted in Fig. 13
reported here for brevity's sake.
3.4. The control law
4.2. Performance analysis: time delay and tracking error
Since the reference state x (i ) satisfies the embedded model
Experimental results have been obtained with a pressure drop
equation, the state predictor state xl (i ) can track x (i ) with a near-zero
|ΔP| ≈ 80 bar in the hydraulic condition CLOSE. Two time responses to
error, that is the residual tracking error ηx (i ) in (16) which is due to
the triangular reference signal yr with a period Pr = 20 ms (50 Hz) are
residual parametric errors and to exogenous disturbance can be
shown in Fig. 15: the delayed reference yx which is the output of the
brought close to zero. This is the goal of the control law, which is the
reference generator in (30) and the position measurement yx. The
linear combination of the reference command u in (31), of the extended
stroke range is about ± 0.9Xmax ≈ ± 1. 12 mm . The time delay τx agrees
tracking error elx (i ) to be defined below and of the predicted disturbance
with the requirement (16), being τx = 3–5 ms.
xld in (22). The only complication comes from the definition of elx (i )
Fig. 16 shows the measured tracking errors, which differ from the
which cannot be just equal to elx = x − xl , but it should include a linear
true tracking errors defined (16) because the true plunger position xs(i)
combination of xd (i ) as follows:
(unknown) is replaced by the measurement yx(i). Since the magnitude
el (i ) = x (i ) − xl (i ) − Qxld (i ), (34) |yx (i ) − xs (i )| of the unbiased measurement error is less than 2 μm , the
where the matrix Q is univocally imposed by the embedded model. It measured error is a good indication of the true tracking error. Using
can be shown that only the extended tracking error in (34) can be made the same notations of (16), Fig. 16 shows the total tracking errors ex(i)
bounded and ideally brought to zero. The control law reads as (the largest signal) and the residual tracking errors ηx (i ) (the smallest
profile). Both errors are reported for each operating condition (NOP,
u (i ) = Ir (i ) = u (i ) + Kc l
e (i ) − Mxld (i ), (35) OPEN, and CLOSE) defined in Section 1.2. The solid line signal is the
where M is a companion matrix of Q, and both matrices are the nominal tracking error. Fig. 16 shows that the valve controller is robust
solution to the Sylvester equation against operating conditions. The residual tracking error magnitude
|ηx (i )| is larger than the requirement (16). The cause has been identified
⎡ Hc + QAd ⎤ ⎡ Ac ( p) Bc ( p)⎤ ⎡ Q ⎤
⎢ ⎥=⎢ ⎥ ⎢ ⎥. in an excessively noisy current signal as shown in Fig. 17. The issue will
⎣ 0 ⎦ ⎣ Cc ( p) dc ⎦ ⎣ M ⎦ (36) be addressed and solved in Section 4.4.
The state-predictor spectrum Λ m that fixes the gains in (23), the The time response to a square external input is shown in Fig. 18. A
reference-generator spectrum Λ rg that fixes the gain K in (31) and stroke of about 90% was used, the same as for the triangular case.
the control-law spectrum Λ c that fixes the gain K c in (35) are
reported in Table 1. 4.3. Performance analysis: harmonic response

4. Experimental results Fig. 19 shows the harmonic responses of the position controller for
small reference signals (5% of the maximum stroke, about ± 70 μm )
4.1. Control hardware and for large signals (90% of the maximum stroke, about 1.2 mm).
Experimental responses are compared with the theoretical closed-loop
Control algorithms and identification tests have been implemented harmonic response (solid line) which does not account for voltage
on the following HW/SW platform and with the following signals: saturation, and with the two alternative control strategies: the analogue
PID which is commercially employed and a digital PID controller
1. The control code runs under LabVIEW Real-Time Operative System driven by a state predictor. The three alternative controllers behave in a
on a dedicated 2 GHz Intel processor. similar way, which confirms accurate and effective tuning. As a result,
2. The control code was developed under LabWindows CVI (C for their harmonic response reaches the BW limit imposed by the valve
Virtual Instrumentation). electro-mechanics and electronics. EMC offers better regularity and
3. The control time unit is T=0.1 s (10 kHz sampling). adherence to theoretical response both in amplitude and argument,
4. Valve position and solenoid current measurements yx and yI are which should be a desirable performance. EMC harmonic response
looks free of any overshoot. At higher frequencies ( > 100 Hz) the

31
W. Acuña-Bravo et al. Control Engineering Practice 62 (2017) 22–35

Fig. 14. Block diagram of the control unit algorithms.

Fig. 15. 90% stroke triangular reference signal response. Fig. 17. Control signal for 90% stroke triangle.

Fig. 16. 90% stroke triangle reference signal errors.


Fig. 18. 90% stroke square reference signal response.

32
W. Acuña-Bravo et al. Control Engineering Practice 62 (2017) 22–35

Fig. 20. Voltage command (saturated and not) for 10 Hz sine.

Fig. 21. Residual tracking error for a 10 Hz sine reference under saturated and non-
saturated command.
Fig. 19. Harmonic response (magnitude and argument) of the valve position controller.

magnitude sharply decreases whereas the argument tends to a constant


value, which is essential for high frequency noise attenuation.

4.4. Performance analysis: tracking accuracy

As pointed out in Section 4.2, the magnitude |ηx | of the residual


tracking error in response to a 50 Hz triangular reference is larger than
the target of 10 μm in (16). The cause was referred to a noisy current
command. A noisy current in turn may provoke voltage saturation
since the solenoid voltage is proportional to the current derivative.
Moreover, since the range of the differential voltage and of the current
derivative is asymmetric as pointed out in (4), saturation will be likely
to occur only at one side (for positive differential voltage, see Fig. 20)
with a subsequent differential voltage bias and differential current and
position drift, since low-frequency position is proportional to the
differential current through the spring compliance K and the cur-
rent–force gain ϕs. Fig. 22. Spectral density of the tracking errors in Fig. 21.
Such effects are shown in Figs. 21 and 22, where time profile and
spectral density of the residual tracking error are plotted in the case of

33
W. Acuña-Bravo et al. Control Engineering Practice 62 (2017) 22–35

a noisy and saturating voltage and in the case of a smooth and non- worst-case hydraulic condition (CLOSE) has been implemented as the
saturating voltage. The relevant voltage profiles are in Fig. 20. embedded model of the control unit and has been completed with a
Fig. 21 shows a large drift of the tracking residuals, incompatible disturbance second-order dynamics. The disturbance dynamics is
with requirement (16), which pretty disappears in the case of the non- driven by arbitrary signals (noise) affecting both disturbance and
saturating and smooth voltage. Fig. 22 shows that the high frequency controllable state variables. The model error between measurements
noise in excess above 300 Hz is accompanied by an increase of the and model output is the signal which is fed back to the noise vector for
spectrum level at lower frequencies because of the residual drift. stabilizing the state predictor. A proper tuning of the state predictor
Elimination of the high frequency noise by command filtering cleans eigenvalues guarantees closed-loop stability and the required perfor-
the tracking residuals as Fig. 22 shows. mance. The reference generator converts piecewise input position
It is now possible to see the whole line spectrum of the nominal reference into a delayed and smooth reference signal, which can be
tracking error (solid line in Fig. 15), whose fundamental frequency is better tracked by the embedded model state variables. Experimental
the reference signal frequency, 10 Hz in this case. The residual peak in results have shown that the required valve position accuracy can be
Fig. 22 at 10 Hz, as anticipated in Section 3.2.1, is due to the reference reached, if saturation of the solenoid voltage due to high-frequency
signal itself times low-frequency parametric errors. Smoothing of the dynamics and noise is prevented. This has been obtained by a suitable
current command and the consequent voltage desaturation has been filtering of the current command above the target position BW below
achieved by a low-pass filter with a cutoff frequency at about 300 Hz, 100 Hz and depending on the stroke range to be tracked. Comparison
which is the boundary between the low- and the high-frequency with alternative controllers, analogue and digital PID, shows that the
transfer function as pointed out in Section 2.3.2. As a smoothing digital EMC controller repeats the BW of the alternative controllers,
result, the magnitude of the residual tracking error approaches the but improves the harmonic response regularity.
target bound of 10 μm .

5. Conclusions Acknowledgements

A digital model-based controller for a proportional electro-hydrau- Part of the work has been funded by research contracts between
lic valve has been outlined. A modelling rationale and procedure have Atos SpA, Sesto Calende, Italy and Politecnico di Torino. The first
been presented, where the identification process plays a key role for author would like to thank Universidad Antonio Nariño, Colombia,
assessing the range of parametric errors and of neglected dynamics. under Grant number 2015086 for their support during the preparation
The model obtained from experimental data and associated with the of the manuscript.

Appendix A. Nomenclature

The main variables are listed in Table A1.

Table A1
List of variables.

No. Symbol Description Equation Comments

1 {I , yI ∈  Imin ≤ I , yI ≤ Imax} True, measured solenoid current (2) –


2 Ir Solenoid current reference (2), (35), (20) Control law of outer loop
3 ηI Residual tracking error (current) (3) –
4 {V ∈  Vmin ≤ V ≤ Vmax} Voltage – Control law of inner loop
5 {xs ∈  Xmin ≤ xs ≤ Xmax} True spool position (5), (6) –
6 vs Spool velocity (5) –
7 Fa, Ff , FK Friction, fluid, elastic forces (5) –
8 ξ Friction state (5) –
9 ΔP Pressure drop PA − PB (5) –
10 bs Solenoid hysteresis (5) –
11 c Solenoid current affected by backslash (5) –
12 Fs, F s Solenoid force and average force profile (5), (7) –
13 p , p , δp parameters vector, nominal part, bounded uncertainty (10) –
14 x ∈ nc LTI controllable model dynamics (10), (19)
15 u ∈ nu Discrete position control law (10), (19) –
16 yx ∈ ny Measured position (1), (10), (14), (8) –
17 h ∈ nc Parameters uncertainty (10) –
18 ym ∈ ny Model output (1) –
19 em ∈ ny Model error (1) –
20 wy ∈ ny Measurement noise (11), (14) –
21 ∂P Fractional error dynamics (11) –
22 Sx, Sx, x; Sx, y Power spectral density: auto-function; cross-function (13) –
23 {yr ∈ ny yr ,min ≤yr ≤ yr ,max} Reference signal (position) (16), (33) –
24 τx Position reference delay (16) Depends on reference amplitude
25 ηx ∈ ny Residual tracking error (16), (17) Outer loop requirements
26 ex ∈ ny, e x ∈ ny Total and nominal tracking error (16) –
27 xd , xld ∈ xd Disturbance states, prediction (18) Disturbance dynamics
28 wu , wu ∈ wu Noise on the controllable states, estimation (18), (21) Noise estimator
(continued on next page)

34
W. Acuña-Bravo et al. Control Engineering Practice 62 (2017) 22–35

Table A1 (continued)

No. Symbol Description Equation Comments

29 wd , wd ∈ wd Noise on the disturbance states, estimation (18), (21) –


30 d ∈ nd Disturbance signals (18) –
31 Vm,Sm Predictor sensitivity, complementary sensitivity (28) –
32 xr ∈ nc, ur ∈ nu Static reference trajectories, control (31) Reference generator
33 x ∈ nc, u ∈ nu Nominal trajectories, control (31), (35) –
34 e ∈ nc Extended tracking error (34) Control law

References with nonlinear unknown parameters. Control Engineering Practice, 16(11),


1275–1284.
Huang, C., & Sira-Ramírez, H. (2015). Flatness-based active disturbance rejection
Canudas de Wit, C., Olsson, H., Aström, K., & Lischinsky, P. (1995). A new model for control for linear systems with unknown time-varying coefficients. International
control of systems with friction. IEEE Transactions on Automatic Control, 40, Journal of Control, 1–10.
419–425. Kaddissi, C., Kenne, J.-P., & Saad, M. (2007). Identification and real-time control of an
Canuto, E. (2007). Embedded model control: Outline of the theory. ISA Transactions, electrohydraulic servo system based on nonlinear backstepping. IEEE/ASME
46(3), 363–377. Transactions on Mechatronics, 12(February (1)), 12–22.
Canuto, E. (2015). On dynamic uncertainty estimators. In Proceedings of the American Karpenko, M., & Sepehri, N. (2010). On quantitative feedback design for robust position
control conference (ACC) (pp. 3968–3973). control of hydraulic actuators. Control Engineering Practice, 18(3), 289–299.
Canuto, E., Acuña-Bravo, W., Agostani, M., & Bonadei, M. (2014). Digital current Karpenko, M., & Sepehri, N. (2012). Electrohydraulic force control design of a hardware-
regulator for proportional electro-hydraulic valves with unknown disturbance in-the-loop load emulator using a nonlinear QFT technique. Control Engineering
rejection. ISA Transactions, 53(4), 909–919. Practice, 20(6), 598–609.
Canuto, E., Acuña-Bravo, W., Agostani, M., & Bonadei, M. (2014). Proportional electro- Kayihan, A., & Doyle, F. (2000). Friction compensation for a process control valve.
hydraulic valves: From analogue to digital control. International Journal of Control Engineering Practice, 8(7), 799–812.
Mechatronics and Automation, 4(2), 93–103. Kim, M. Y., & Lee, C.-O. (2006). An experimental study on the optimization of controller
Canuto, E., Acuña-Bravo, W., Molano-Jiménez, A., & Pérez-Montenegro, C. (2012). gains for an electro-hydraulic servo system using evolution strategies. Control
Embedded model control calls for disturbance modeling and rejection. ISA Engineering Practice, 14(2), 137–147.
Transactions, 51(5), 584–595. Levi, E. (1959). Complex-curve fitting. IRE Transactions on Automatic Control, AC-4,
Canuto, E., Acuña-Bravo, W., & Pérez-Montenegro, C. (2013). Robust control stability 37–44.
using the error loop. International Journal of Mechatronics and Automation, 3(2), Li, S., Yang, J., Chen, W.-H., & Chen, X. (2014). Disturbance observer-based control
94–109. Boca Raton: CRC Press.
Canuto, E., Pérez-Montenegro, C., Colangelo, L., & Lotufo, M. (2014c). Active Ljung, L., & Glad, T. (1994). Modeling of dynamic systems Englewood Cliffs: PTR
disturbance rejection control and embedded model control: A case study Prentice Hall.
comparison. In Proceedings of the 33rd Chinese control conference (pp. 3697– Maciejowski, J. (1989). Multivariable feedback design Wokingham: Addison Wesley.
3702), Nanjing, China. Marusak, P., & Kuntanapreeda, S. (2011). Constrained model predictive force control of
Canuto, E., Pérez-Montenegro, C., Colangelo, L., & Lotufo, M. (2014d). Embedded an electrohydraulic actuator. Control Engineering Practice, 19(1), 62–73.
model control: Design separation under uncertainty. In Proceedings of the 33rd Mohanty, A., & Yao, B. (2011). Integrated direct/indirect adaptive robust control of
Chinese control conference (pp. 3637–3643), Nanjing, China. hydraulic manipulators with valve deadband. IEEE/ASME Transactions on
Farenzena, M., & Trierweiler, J. (2012). Valve stiction estimation using global Mechatronics, 16(August (4)), 707–715.
optimisation. Control Engineering Practice, 20(4), Wei, X., & Guo, L. (2009). Composite disturbance-observer-based control and terminal
379–385 (Special Section: IFAC symposium on advanced control of chemical sliding mode control for non-linear systems with disturbances. International
processes – ADCHEM 2009). Journal of Control, 82(6), 1082–1098.
Francis, B. A., & Wonham, W. M. (1976). The internal model principle of control theory. Yi, H., Wenchao, X., Gao, Z., Sira-Ramirez, H., Dan, W., & Mingwei, S. (2014). Active
Automatica, 12(5), 457–465. disturbance rejection control: Methodology, practice and analysis. In Proceedings of
Gao, Z. (2006). Active disturbance rejection control: A paradigm shift in feedback control the 33rd Chinese control conference (CCC) (pp. 1–5), July.
system design. In Proceedings of American control conference (pp. 1399–2405). Zheng, Q., & Gao, Z. (2014). Predictive active disturbance rejection control for processes
Garcia, C. (2008). Comparison of friction models applied to a control valve. Control with time delay. ISA Transactions, 53(4), 873–881.
Engineering Practice, 16(10), 1231–1243.
Guan, C., & Pan, S. (2008). Adaptive sliding mode control of electro-hydraulic system

35

You might also like