See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/243764002

Robust Adaptive Control for Hydraulic Servosystems

Article in Journal of Dynamic Systems Measurement and Control · June 1996

DOI: 10.1115/1.2802309

CITATIONS READS

140 298

2 authors:

Andrew R. Plummer Nicholas D. Vaughan

University of Bath Cranfield University
169 PUBLICATIONS 4,025 CITATIONS 74 PUBLICATIONS 1,283 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Andrew R. Plummer on 01 April 2019.

The user has requested enhancement of the downloaded file.

 Robust Adaptive Control
A. R. Plummer' for Hydraulic Servosystems
The application of an indirect (self-tuning) adaptive controller to an electro-hydraulic
N. D. Vaughan positioning system is described. The underlying control method is pole placement,
with the addition of a demand filter to allow noise effects to be reduced without
School of Mechanical Engineering, degrading closed-loop performance. Recursive least squares is used to estimate the
University of Bath,
plant parameters, but the data is pre-filtered to reduce bias. A novel covariance
Bath, U.K.
trace limiting algorithm provides estimator reliability despite periods of insufficient
excitation. Off-line system identification is employed to help controller design for the
electro-hydraulic servosystem. The resulting controller performs well, and adapts
rapidly to changes in load stiffness and supply pressure.

1 Introduction Direct (or model reference) adaptive control is the other

approach. In this method the controller coefficients required to
Electro-hydraulic servosystems can combine the high power give a prescribed model-following performance are estimated.
of hydraulic actuation with the versatility of electronic control. They are estimated directly from input-output data, and not
However in many circumstances simple analogue control calculated from a plant model. A number of applications of this
schemes are not sufficient to realize the full potential of the type of adaptive controller to electro-hydraulic position control
hydraulic system in terms of accuracy and dynamic perfor- systems exist, notably those of Edge and Figueredo (1987),
mance. There are several reasons for this. Hydraulic plant exhib- Hori et al. (1988), and Unbehauen et al. (1988).
its significant non-linearities, so a linear controller can only be Despite these numerous applications, further development is
optimised for one operating point. Even at one operating point required before adaptive control is widely adopted in industry.
the plant characteristics will often vary, for example due to The aims of the present study are threefold:
changes in load behaviour. These variations are common in a
wide range of hydraulic servo applications, including heavy • to achieve fast adaptation without large tuning transients,
duty robots, material testing equipment, and plastic injection • to develop an adaptive controller which performs reliably
moulding machines. and consistently under a wide range of conditions with any
As a result of these problems, a number of researchers have demand signal,
investigated the application of sophisticated digital controllers • to make adaptive controllers easier to design.
to electro-hydraulic servosystems, and in particular to position
control systems. These controllers usually assume a linear plant The indirect adaptive method was chosen for this work. This
model, and incorporate an adaptation mechanism to cope with has allowed the estimation and control parts of the algorithm
time-varying characteristics and some of the non-linearities. to be studied and developed separately with ease. Recursive
The first application of such adaptive controllers to electro- least squares is the basis of the estimator used to find the plant
hydrauhc positioning systems occurred over two decades ago model parameters. The underlying control method is pole place-
(Porter and Tatnall, 1970). However, the analogue computing ment, which has proven more flexible than classical compensa-
technology available at the time made the controller implementa- tors (such as PID), and more robust than optimal controllers.
tion very cumbersome. The rapid development of digital electron- The adaptive controller is similar to those proposed in the semi-
ics in recent years, especially the microprocessor, means that nal papers of Astrom and Wittenmark (1980) and Wellstead
adaptive controllers can now be applied to fast servosystems and Sanoff (1981).
relatively cheaply. The intervening years have also seen signifi-
cant advancements in the theoretical aspects of adaptive control.
Two distinct approaches have been used for the practical 2 Adaptive Control Sclieme
application of adaptive control to electro-hydraulic positioning 2.1 Pole Placement Controller. For the purpose of anal-
systems. Indirect (or self-tuning) adaptive control involves the
ysis, let the plant be represented by:
on-line estimation of a plant model, and uses a conventional
model-based controller which is repeatedly re-designed using A{z~')y, = Biz^')u, -h e, (1)
the latest estimated model. In its original form self-tuning con-
trol was only used for the initial tuning of controller parameters. where y, is the plant output, u, the plant input, e, is a noise
This form was implemented by Finney et al. (1985). For a full signal, and
adaptive version the estimator must be modified to forget old
data, for example by the inclusion of a fixed forgetting factor A{z^') = 1 + fliz"' + ... + a„z~"
as used by Vaughan and Whiting (1986). A similar method
was used by Daley (1987) for the speed control of a rotary B(z^') = h,z'' + b2z~^ + ... + h^z-'"
hydrauUc system.
Any dead time in the plant can be incorporated in the B(z~')
polynomial (for example dead time of one sample interval
would give Z?] = 0). If e, is coloured noise, then
* Now at Department of Mechanical Engineering, University of Leeds, Leeds,
U.K. e, = C(z ')v,
Contributed by tire Dynamic Systems and Control Division for publication in
(2)
the JOURNAL OF DYNAMIC SYSTEMS, MEASUREMENT, AND CONTROL. Manuscript
received by the DSCD March 23, 1990; revised manuscript received February where v, is a white noise sequence, and C(z ') is the noise
1994. Associate Technical Editor: A. Akers. model.

Journal of Dynamic Systems, Measurement, and Control JUNE 1996, Vol. 1 1 8 / 2 3 7

Copyright © 1996 by ASME
Downloaded From: https://dynamicsystems.asmedigitalcollection.asme.org/ on 11/21/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use
 The pole placement controller is implemented using three
digital filters as shown in Fig. 1. The control signal is generated
thus:

u, = (Hiz-')r, - G{z"')y,)IF{z^') (3)

Htr') -i<g>
The resulting closed-loop response is:

B(z-')//(z-')
y, = F ( z - ' ) A ( z - ' ) + G ( z - ' ) S ( z - ' ) '

-P(z"')C(z"')
V, (4)
Fiz-')A{z~') + G(z-')B(z-')
Fig. 1 Pole placement controller
If A(z ') and B{z ') are known exactly, then F(z ') and
G(z''^) can be calculated to satisfy:

F(z~')A(z-') + G(z-'')B{z-') = kTiz-')H(z-') (5) appears in the transfer function from noise to output, and
it can be used to attenuate noise without affecting servo
Substituting Eq. (5) into ( 4 ) : performance. Thus 1/H(z'') can be any stable filter speci-
fied by the user. The choice and effects of the demand filter
BiLllr I f(z")C(z-') ^ are discussed in more detail in Vaughan and Plummer
(6) (1990).
For the simplest controller, the solution to (5) which gives
Thus r ( z ~ ' ) is the closed-loop characteristic polynomial polynomials F (z ') and G (z ') of minimum degree is required.
relating demand to output. It has a unity z° coefficient and This is achieved by using:
roots which are the closed-loop system poles specified by deg F ( z - ' ) 1 (i.e. m coefficients)
the user. The scalar k is used to give unity steady-state gain,
i.e., deg G(z ') = « - 1 (i-e. n coefficients) (8)
and deg { r ( z " ' ) / f ( z " ' ) } = n + OT - 1
k = B(l)/T(l) (7)
Thus equating equal powers of z ' in (5) gives n + m equations,
As can be seen from Eq. (6) the demand filter H(z~^) only and there are « -I- m unknown F(z~') and G(z~') coefficients.

Nomenclature

Ui = coefficient in A(z~') G{z ') = controller feedback path term, u, = plant input signal at sample instant
A{z^'^) = plant model denominator poly- of the form: t
nomial: u', = filtered plant input u,
Giz~')
V, = white noise signal
A(z"') = 1 + fliz"' + fl22"^
= 80 + giZ~^ + g2Z~^ + • . . y, = plant output signal at sample in-
+ ... + a„z"" stant t
H{z~^) — controller demand filter, of the y', = Filtered plant output y,
bi = coefficient in i?(z"') form: z~' = backward shift operator
fi(z"') = plant model numerator polyno-
a = covariance trace limit
mial: H(z~') (j)/ = regressor vector at sample instant t
B{z-') = b,z'' + bjz-^ = 1 + /ZiZ"' + /l2Z"^ + . . . tji, = regressor vector based on filtered
signals at sample instant t
+ ...+ bu~"' i = an integer X = forgetting factor
k = controller gain correction fac- 0 = vector of model parameters
C{z^^) = noise model, of the form: tor 0, = estimate of model parameter vector
Ciz-') k, = estimator gain vector at re- at recursion t
cursion ;
= 1 + ciz"' + C2Z"' + . . . m = degree of B(z~') polynomial Superscript
n = degree of A(z ') polynomial T = transpose
deg = degree of polynomial (i.e.
P, = normalized covariance matrix = estimate of
highest power)
estimate at recursion t
e, = noise signal at sample instant /
t = time as a number of sample
F(z~') = controller forward path term,
of the form: intervals
tr = Trace of matrix (i.e., sum of
F(z-') leading diagonal elements)
T{z~^) = polynomial with roots which
= / o + / i Z - ' +/2Z-" + . . . are desired closed-loop poles,
of the form:
nz-')
. = 1 + fiZ"' + t2Z~^ + ...

238 / Vol. 118, JUNE 1996 Transactions of the ASME

Downloaded From: https://dynamicsystems.asmedigitalcollection.asme.org/ on 11/21/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 so there is a unique solution. If the closed-loop characteristic change for a given error between (filtered) actual and model
polynomial is chosen to have the same degree as the open-loop outputs. In addition, it can be shown that
one, then:
k, = P,^, (14)
deg Tiz^') = n
(9) so P, directly influences the rate of change of the estimates, and
and deg H(z ') = m - 1 as it reduces, the estimates converge to steady values.
2.2 Estimation. In order to present a method for estimat- For time-varying plant, the reduction in P, means that the
ing plant model parameters, it is useful to rewrite Eq. (1) as a model will take progressively longer to adapt to changes in the
regression equation: plant. Thus traditionally a forgetting factor just below unity has
been used for this situation. This has most effect on equation
y, = <f,j0 + e, (10) (13c), where P, is scaled up on each recursion to counteract
its tendency to diminish. The forgetting factor can also be inter-
where preted as a means of weighting old data to have less importance
than new data; the significance of data from a particular sample
tf>J = [y,~u • • • ,y,-,„ "/- M,_ instant decays exponentially as time progresses with time con-
stant 1/(1 - k) sample intervals.
is a vector of input-output data (— the regressor vector), and
Problems are encountered with the use of a fixed forgetting
0 = [-«!, . . . , -a„, b,, . . . , b,„y factor less than unity if the data are not persistently exciting.
During periods of low excitation, new data does not drive P,
is a vector of model parameters. any lower, yet the matrix is still being scaled up by the forgetting
The model parameters can be estimated from sampled data factor. Thus P, can become very large, and the estimates are
using the least squares estimator. However it is well known that sensitive to small model output errors, often changing drasti-
if any regressor (element of <^,) is correlated with the noise cally in the presence of noise.
signal e,, then the least squares estimates are biased (Ljung As an alternative to a fixed forgetting factor, the constant
and Soderstrom, 1983, for example). In practice e, is usually trace algorithm has been used, for example by Hori et al. (1988)
autocorrelated (i.e., colored), and hence also correlated with and Unbehauen et al. (1988). In essence, the forgetting factor
the past values of output which appear as regressors. In the is set to unity in (I3b), but varied in (13c) to maintain P, at
presence of feedback the past values of input which appear as a constant size. The trace of P, is used as a scalar measure of
regressors will also be correlated with an autocorrelated e,. Thus the size of the matrix. However in practice this method does not
the least squares estimates will be biased. entirely solve the problem. During long periods of insufficient
Consider Eq. (10). If the input and output signals were fil- excitation, the parameter values which can fit the model to the
tered by the reciprocal of the noise model, they would be related data are not unique, and the estimates tend to drift.
by: To overcome these difficulties, a technique has been devel-
oped which uses a fixed forgetting factor, but in addition
y! = «A?'0 + V, (11) switches off the estimator if the excitation is insufficient. The
excitation is deemed to be insufficient if the trace of P, reaches
where some user specified limiting value. Thus trace limited RLS is
given by equation (13) with the following extension:
*f>'t = [ y / ' - i , • • • ^y'l-n, «/'- Ut-m\
if tr P, then
is the new regressor vector, and
P, = (a/trP,-,)P,-i (15)
1
y, = 0, = 0,-1
C(z-')
(12)
1 where a is the trace limit.
u, = The forgetting factor \ can now be reinterpreted: considering
C(z-') the algorithm in the trace limited state, the lower the value of
As the noise signal v, in the new regression equation is white, \ , the more exciting the data must be before adaptation occurs.
it is not correlated with the regressors, and the least squares Also the higher the value of a, the higher P, (and k,) can
become before adaptation is switched off, allowing the model
estimates will not be biased. However in practice the noise
to adapt more rapidly.
model is unknown, but using engineering judgment to design a
filter to attenuate noise in the signals is sufficient to improve
the estimates. 3 Computer Implementation
Using filtered input-output signals, the recursive least squares The digital computer system used for implementing the adap-
(RLS) estimate 0, of the model parameter vector 0 (based on tive controller has the following features:
data up to sample time t) can be shown (e.g., Ljung and Soders-
trom, 1983) to be: Main processor card: 20 MHz T800 Inmos Transputer (10
Mbits/s links) plus 2 Mbytes DRAM.
0, = 0,-, + k,iy; - tl,j0,-,) (a) Host: IBM AT-compatible providing user in-
where k, = P,-,il>,(\ + i//?T,-,i/r,)-' (b) (13) terface.
ADC: 12-bit, 8-channel, 5 fj,s conversion, 10
and P, = (P,-, - k^jF,.,)/k (c) /.ts communication (via transputer
link). No anti-aliasing filters.
P, is a normalized version of the covariance matrix of the esti- DAC: 12-bit, 4-channel, 8 fxs settling time, 10
mates, and X is the forgetting factor (0 < X. == 1). fis communication (via transputer
If the plant was not time-varying, the forgetting factor would link).
be set to unity. Data at each sample instant would then have Language: 3L Parallel C. 64 bit floating-point
the same importance, and the elements of the covariance matrix
arithmetic used throughout.
P, would reduce in size as more data is introduced, reflecting
greater confidence in the accuracy of the estimates. Notice from This computer system was also used for all off-line work,
(13a) that k, acts as a gain determining how much the estimates such as data acquisition and system identification.

Journal of Dynamic Systems, Measurement, and Control JUNE 1996, Vol. 1 1 8 / 2 3 9

Downloaded From: https://dynamicsystems.asmedigitalcollection.asme.org/ on 11/21/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Dead Oil Volumes n n 0.012

Hydraulic Supply I |

41^-
Control Signal X
AA/yV\
Position

/
Mass

/y-T-T-y-^ Hydraulic Cylinder

Fig. 2 Electro-hydraulic positioning system

MODEL ORDER (n)

To implement the adaptive controller for the system described Fig. 4 Prediction error comparison for model structure selection
in the next section, about 3 ms computation time was required
per sample. However the software was written for the general
case (i.e. for any model order etc.), and contained many moni- The system is very oscillatory due to the interaction between
toring functions. Optimising the software for speed would prob- the large inertia and the oil compliance, with a resonance at
ably cut the computation time by about half. centre stroke of about 12 Hz. Two dead volumes of oil can be
switched into the hydraulic circuit either side of the cylinder.
4 Application to Electro - Hydraulic Positioning This changes the plant dynamics radically (reducing the reso-
nant frequency to about 7 Hz), and was a convenient way to
System test the adaptive controller.
4.1 Description of Positioning System. The adaptive In reality the plant is non-linear, and correct values for the
controller has been applied to an electro-hydraulic positioning parameters in a linear model do not exist. The non-linear behav-
system. The system consists of a servovalve-controlled cylinder iour arises from several sources, including:
driving a large inertial load, as shown in Fig. 2. It has the
following specification: • a directional nonlinearity due to the single-ended cylinder.

Valve: Dowty series 4551

Bandwidth greater than 50 Hz
Flow gain (at 70 bar) of 3.8 (L/min)/mA
Saturation at 9 mA 20
Cylinder: Stroke 610 mm
10
Piston area 2025 mm^
Annulus area 1380 mm^ 0
Load: 890 kg mass
Very low damping -ID
Position feedback: Wirewound potentiometer
Zero at centre stroke -20

Retract is positive direction

10
Supply pressure: Variable to 160 bar
5

0
—~r "^""""^asssi-^' ' ' ! I ! ! 1 ! ! ! '! !
-5

-10
0.6 M M
\ i (1 - 0.42-') ; i i ;
S -4 ^ ! i ; f \ j--i-:-j
O
0.126(1 + z-')'
%7t i i ! I'tt-
la
a
-6 i \ \ -, 1 •••••;-;•••; iX-^....j j i V\\-
0.512
(1 - 0.22-')" —f\ 'v. ^^ MM
~ ; ; \ ' : ; • • ; • • : • ;
\~\ i\^;"i \ i rri"
0.B33
(1 - 0.27Z-T

10" 101 10^ 3 4

FREQUENCY (Hz) TIME(s)

Fig. 3 Frequency response of some digital filters with breal( points at Fig. 5 Adaptive controller with steady 100 bar supply pressure (no dead
15 Hz (10 ms sample interval) volumes)

240 / Vol. 118, JUNE 1996 Transactions of the ASME

Downloaded From: https://dynamicsystems.asmedigitalcollection.asme.org/ on 11/21/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 bar supply pressure. A 10 ms sample interval was used, giving
a sample rate about 8 times the bandwidth of the plant (interpre-
ting the bandwidth as the point at which resonance occurs, in
this case 12 Hz). A least squares estimator processing filtered
data was used to estimate models of the plant off-line. The data
filter was chosen to attenuate noise and high order dynamics
above the plant bandwidth. The frequency responses of several
digital filters with 3 dB break points at 15 Hz are shown in
Fig. 3. The third order filter 0.125(1 + z " ' ) ' was selected as
it gives rapid roll-off, with a gain of zero at the Nyquist fre-
quency (50 Hz). Note that this particular choice of filter has
jAPAmyECONTROt; not been found to be critical for successful estimation.
Before the parameters in an off-line model could be esti-
mated, a model structure had to be selected. If n or m (the
estimated degrees of A(z"') and B ( z " ' ) respectively) are too
small, the model will not be able to represent the plant behaviour
adequately. If n and m are both too large, the estimates of
A{z'^) and B(z^') will tend to have common roots, in which
case the controller design Eq. (5) cannot be solved (assuming
the common roots are not also roots of r ( z ' ) ^ ( z ~ ' ) ) - How-
ever, to simplify the problem, n and m can be equated, and if
IPASMEffiRAiDAPTHittlON: this single order is equal to the largest of the true degrees n
and m, the plant behaviour can be modelled well without the
'-I
'^^- danger of pole-zero cancellation. To select an appropriate model
order, models for a range of orders were estimated from 400
data samples, and an RMS average of the prediction errors
T^ between model and actual output for each order were calculated.
These are plotted in Fig. 4. The rapid reduction in error up to
order 3, followed by virtually no further reduction, indicates
0.002
Gl that a third order model is appropriate (i.e., n = m = 3).
0.001
-^1^ if Thus the following third order model was identified:
, (1.07Z-' -h 3.78z-^ + 3.27z-')10-'
V, = ; ; r M; (lo)
TIME(s)
^ 1 - 2.37z'' + 2.28z"' - 0.909z"'

Fig. 6 Control action with increase in supply pressure from 40 bar to

160 bar after 2 s (no dead volumes)
FIXED CONTROL:

• the square law relationship between pressure and flow in

the valve
• different characteristics in different parts of the stroke due
to volume changes,
• saturation in the servovalve,
• backlash in the mountings of the cylinder body and rod.
The main effect of the directional non-linearity is a larger
gain when the load is driven in the extend direction as opposed ADAPTIVE CONTROIl:

to the retract direction. This can be compensated for in software

by scaling the valve control signal differently according to its
sign (Vaughan and Whiting, 1986). Note that this nonlinear
compensation is undertaken just before the control signal is sent
to the plant, and the signal passing into this compensation stage
can be considered as the plant input. Thus conceptually the
compensation is part of the plant itself.
The servovalve saturation non-linearity can be neglected if
the valve is never allowed to saturate. This can be achieved by
"IPARAMijrfeR AiDArrXtlON:
limiting the control signal in software to be just within the level
which saturates the valve. Hence the signal used for the plant
input by the estimator and for calculating the next control signal
will always be closely related to the actual opening of the valve.
4.2 Setting up the Adaptive Controller. Before the
adaptive controller can be applied to any servosystem, choices 0

have to be made regarding the structure of the plant model, the 0.002
sample rate, the forgetting factor and trace limit for the estimator h
0.001
forgetting strategy, the position of the closed-loop poles etc. To
help with these choices it is very desirable to undertake some
off-line system identification work.
TIME(s)
Input-output data were collected for the positioning system
using a pseudo-random binary sequence (PRBS) input signal, Fig. 7 Control action with decrease in supply pressure from 160 bar to
with the plant operating in open-loop near mid-stroke at 100 40 bar after 2 s (no dead volumes)

Journal of Dynamic Systems, Measurement, and Control JUNE 1996, Vol. 118 / 241

Downloaded From: https://dynamicsystems.asmedigitalcollection.asme.org/ on 11/21/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Si = —1.37
RXBD CONTROL: 4 = 0.907
S, = 1.07 X 10^'
51.3 -34.4 -5.66 X 1 0 - '
P, = -34.4 42.3 2.61 X 10^'
g -10 -5.66 X 10"' 2.61 X 10- 3.66 X 10"*

The trace of the covariance matrix above is 93.6. This was used
as a guide to choosing the trace limit ( a ) . In this case a = 200
A D A m y B CQNTROIL:
was selected. The forgetting factor ( \ ) was chosen experimen-
tally to be 0.92. This value prevented adaptation when the plant
was driven by any insufficiently exciting signal.
To simplify the choice of closed-loop pole positions, three
first order poles in the same place were used. Placing the poles
atz = 0.6 gave an acceptable speed of response. Similar criteria
apply to the choice of demand filter as to the estimation filter;
two first order roots were used, both at z = 0.3. Thus
T(z~') = (1 - 0 . 6 z - ' ) ^
"HRAMFiijR AiSAiPtffi
and
H(z~') = (1 - 0 . 3 z - ' ) '
The selection of these filters is discussed in more detail in
Vaughan and Plummer (1990). The same data filter was used
for on-line estimation as off-line estimation.
0.002 4.3 Experimental Results. Figure 5 shows the response
61
of the adaptive controller to a square wave demand input. When
the actual load position reaches the demand position in the

TIME(s)

Fig. 8 Control action with dead volumes switched in after 2s (100 bar
supply pressure) FIXED CONTROL:

Factorizing the numerator and denominator polynomials gives:

. 1.07z"'(I + l.SOz'')(l +2.04;-')10-'
^' ^ (1 - 1.00z-')(l - (0.684 + 0.664j)z-')(l - (0.684 - 0.664;)z"') "'
(17)
In order to design a controller which can adapt as rapidly as
possible to changes in the plant, the number of parameters A D A P T I V E CONTROL:
estimated on-line should be reduced to a minimum. The pole
at z = 1 in the above model represents the inherent integrating
nature of the plant, and is unlikely to change. Also the zeros
play a very minor role in modeling the dynamics of the plant,
so they can be fixed without seriously restricting the time-
variations to which the model can adapt. Thus the on-line
estimator processes (1 -I- 1.50z~')(l -I- 2.04z~')u', as the
input signal and (I - z^^)y'i as the output to estimate a
model of the form:
.^ARAMEtfiR ADAPTAtiON;
biz-
(18)
1 -I- fl,z ' + oiz^
Note that the position on the z-plane of the fastest poles in
the model (17) can be used to check the suitability of the sample
rate used. In this case the poles are at z = 0.684 ± 0.664 j , which
0
indicates a sample rate of 8.13 times the natural frequency of
the plant (giving the latter as 12.3 Hz). 0.002
Off-line estimation yielded initial values for the parameters 81
0.001
and the covariance matrix, so that the adaptive controller was : 1 ^

well behaved at start-up. To obtain the covariance matrix for

the subset of the model to be estimated on-line, the estimator
TIME (s)
was re-run for the model (18), prefiltering u', by the known
zeros and y,' by the known pole as described. The off-line Fig. 9 Control action with dead volumes switched out after 2 s (100 bar
estimates (from 400 samples of data) were: supply pressure)

242 / Vol. 118, JUNE 1996 Transactions of the ASME

Downloaded From: https://dynamicsystems.asmedigitalcollection.asme.org/ on 11/21/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 steady state, the covariance trace quickly reaches its limiting
value and adaptation is switched off. Figure 6 illustrates the
behaviour of the adaptive controller when there is an increase
in supply pressure. A fixed pole placement controller (with the
same desired closed-loop poles and demand filter) is also shown
for comparison. Both controllers are initially well tuned for the
starting pressure of 40 bar. After 2 s the pressure is suddenly
increased to 160 bar. The fixed controller becomes very oscilla-
tory, but the adaptive controller fully adapts to the new pressure
within one period of the demand cycle. Note that the control
signal saturates for part of the cycle, limiting the maximum
speed of the actuator. This maximum speed is greater with the
increased supply pressure, allowing the desired response to be
followed more closely.
Figure 7 illustrates a similar effect when the pressure is re-
duced from 160 bar to 40 bar, with both fixed and adaptive
controllers starting well tuned for the higher pressure. Note that
changing the supply pressure is just a convenient means of
changing the plant characteristics.
In Fig. 8 the supply pressure is kept constant at 100 bar.
After 2 s the dead oil volumes are switched into the circuit.
This reduces the natural frequency of the plant in much the
same way as an increase of load mass would. The adaptation
is again very rapid. Figure 9 shows the reverse situation, where
the extra oil volume is initially present, and is removed after
2 s. The fixed controller develops a high frequency limit cycle
(the control signal is oscillating between the two saturation
levels), but the adaptive controller performs well. Notice that
the estimates do show some adaptation to the difference in plant
"1
behaviour between the cylinder extending and retracting. 0.001
Figure 10 illustrates the performance of the adaptive control-
ler when the input and output signals are not filtered before
estimation. The parameter estimates exhibit more bias and •nME(s)
slower convergence, resulting in the poor response of the con-
troller. Note that increasing X to 0.99 was found to be most Fig. 11 No demand filter {H(z^) = 1): control action with increase in
supply pressure from 40 bar to 160 bar after 2 s (no dead volumes)
appropriate with no estimation filtering, and so this value was
used for the results shown.
In Fig. 11 the estimation filter is restored (and \ = 0.92) estimation still progresses satisfactorily, but the position re-
but the demand filter is omitted (i.e. //(z ') = 1 ) . Parameter sponse is slightly more sensitive to modelling errors. Also noise
has a greater effect on the control signal, and the vigorous
nature of the signal causes excessive vibration of the plant
structure.
To demonstrate the effectiveness of the trace limiting forget-
ting strategy, Fig. 12 compares different strategies when the
demand signal is stationary at zero. The value of a typical
parameter estimate (a,) is plotted every second over a 50 s
period. For the constant trace estimator, the estimate drifts until

-0.4

Constant Trace (150) ; Fixed Foigetting Factor (0.96) i ,",

Trace litiiiting

0 5 10 15 20 25 30 35 40 45 50

TIME (s)

Fig. 10 No estimation filter: control action with increase in supply pres- Fig. 12 Behavior of adaptive controllers with different forgetting strate-
sure from 40 bar to 160 bar after 2 s (no dead volumes) gies when demand is stationary at zero

Journal of Dynamic Systems, IVIeasurement, and Control JUNE 1996, Vol. 1 1 8 / 2 4 3

Downloaded From: https://dynamicsystems.asmedigitalcollection.asme.org/ on 11/21/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 after 18 s the corresponding controller becomes unstable, and Acknowledgments
the resulting load movement is sufficient to partly re-tune the
model and controller. The same drifting occurs with a fixed The authors are grateful for the cooperation and financial
forgetting factor, but the cycle of drifting and re-tuning is much support received from Dowty Rotol Ltd., and the additional
faster. The covariance trace soon reaches its maximum value financial support received from the Science and Engineering
using the trace limiting algorithm, and so adaptation is switched Research Council.
off and the estimates do not drift.
References
5. Conclusions Astrom, K. J., and Wittenmarjj, B., 1980, "Self-Tuning Controllers Based on
Pole-Zero Placement," lEE Proc. Pt D, Vol. 127, No. 3, pp. 120-130.
An indirect (self-tuning) adaptive control algorithm based Daley, S., 1987, "Application of a Fast Self-Tuning Control Algorithm to a
on pole placement control has been designed and applied to an Hydraulic Test Rig," Proc. Instn. Mech. Engrs., Vol. 201, No. C4, pp. 285-295.
Edge, K. A., and Figueredo, K. R. A., 1987, "An Adaptively Controlled Elec-
electro-hydraulic positioning system. The controller shows trohydraulic Servo-Mechanism, Part 2: Implementation," Proc. Instn. Mech.
rapid adaptation to changes in plant characteristics, even when Engrs., Vol. 201, No. B3, pp. 181-189.
the changes are very significant and virtually instantaneous. An Finney, J. M., de Pennington, A., Bloor, M. S., and Gill, G. S., 1985, "A Pole-
equivalent fixed coefficient controller breaks down in the same Assignment Controller for an Electrohydraulic Cylinder Drive," ASME JOURNAL
OF DYNAMIC SYSTEMS, MEASUREMENT, AND CONTROL, Vol. 107, pp. 145-150.
sitijation. The good performance is achieved despite the pres-
Hori, N., Ukrainetz, P. R., Nikiforuk, P. N., and Bitner, D. V., 1988, "Robust
ence of many non-linear characteristics. Discrete-Time Adaptive Control of an Electrohydraulic Servo Actuator," Proc.
Data filtering in the recursive least squares estimator, and a 8th Fluid Power Symposium, Birmingham, UK, pp. 495-514.
demand filter in the pole placement controller, are additional Ljung, L„ and Soderstrom, T., 1983, Theory and Practice of Recursive Identifi-
features compared with adaptive control algorithms previously cation, MIT Press, Cambridge, MA.
Porter, B., and Tatnall, M. L., 1970, "Performance Characteristics of an Adap-
applied to similar systems. Furthermore, a strategy which tive Hydraulic Servo-Mechanism," Int. J. Control, Vol. 11, No. 5, pp. 741-757.
switches off adaptation if the trace of the covariance matrix Unbehauen, H., Du, P., and Keuchel, U., 1988, "Application of a Digital
becomes too great is used. This overcomes the difficulties Adaptive Controller to a Hydraulic System," Int. Conf. on Control 88, Oxford,
caused by lack of excitation, and allows the system to perform UK, pp. 177-182.
Vaughan, N. D„ and Plummer, A. R., 1990, "Some Issues in Pole Placement
reliably with a wide range of demand signals. Control of Electro-HydrauUc Servos," Proc. Third Bath International Fluid
Off-line system identification is shown to be extremely useful Power Workshop, Bath, UK.
as a basis for adaptive controller design. In particular it can be Vaughan, N. D., and Whiting, I. M., 1986, "Microprocessor Control Applied
used to determine the structure of the plant model, to reduce to a Nonlinear Electrohydraulic Position Servo System," Proc. 7th Fluid Power
Sympo.iium, Bath, UK, pp. 187-198.
the number of parameters to be adapted, and to provide initial Wellstead, P. E., and Sanoff, S. P., 1981, "Extended Self-Tuning Algorithm,"
estimates and variances for controller start-up. Int. J. Control, Vol. 34, No. 3, pp. 433-455.

244 / Vol. 118, JUNE 1996 Transactions of the ASME

DownloadedViewFrom:
publicationhttps://dynamicsystems.asmedigitalcollection.asme.org/
stats on 11/21/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use