Full Papers

Soft-sensors for process estimation and

inferential control
Ming T. Tham, Gary A. Montague, A. Julian Morris, and Paul A. Lant
Department of Chemical and Process Engineering, University of Newcastle, Newcastle upon
Tyne, NE1 7RU, UK
(Received 13June 1990; revised 20 August 1990)

This paper presents two adaptive estimators (software based sensors or ‘soft-sensors’) for inferring process
outputs that are subject to large measurement delays, from other (secondary) outputs which may be
sampled more rapidly. In other words, these estimators utilize plant data sampled at different rates. The
parameters of the estimators can be continuously estimated and updated on-line, thus enabling the tracking
of slow variations in process characteristics. The estimators employ either an input-output or a state space
process description as the starting points for algorithm synthesis. In contrast to mechanistic model-based
estimators such as Kalman filters, the proposed adaptive techniques require minimal design effort. The key
contribution of the paper is thus the formulation of applications independent, adaptive multi-rate algor-
ithms, which can provide accurate estimates of infrequently measured process outputs, from other more
rapidly sampled secondary outputs. Theoretical developments are supported by results of recent appli­
cations to a variety of industrial scale processes: estimation of biomass concentration in an industrial
mycelial fermentation; top product composition of a large industrial distillation tower and melt flow index
on an industrial polymerization reactor. Measurements from established instruments such as off-gas carbon
dioxide in the fermenter; overheads temperature in the distillation column and hydrogen concentration in
the reactor were used as the secondary variables for the respective processes. The range of the applications is
an indication of the utility of the techniques. A significant improvement in overall process control perfor­
mance is also possible when estimated plant outputs, rather than the infrequently obtained measurements,
are used for feedback control. This is demonstrated by non-linear simulation studies.

(Keywords: estimation; feedback control; sensors)

In many process control situations, due to sampling limi- results were presented to demonstrate the superior per­
tations, the infrequent measurement of some process formance of their control strategy over conventional PID
outputs prevents the early detection of load disturbances. control.
This can result in large deviations from setpoints and A second approach is to use the information provided
long disturbance recovery times. Often, these adverse by other easily measurable variables to provide an esti-
effects cannot be acceptably overcome even by the use of mate of the controlled output. The estimated outputs can
existing advanced control algorithms and can lead to then be used for overall plant control. Control schemes
unsatisfactory system performance. The problem of con- based on the feedback of output estimates are often
trolling infrequently measured process outputs has long termed ‘inferential control’ schemes. An ideal situation
been studied and publications in this area date back to arises when the plant states are completely observable
the early 1970s. There are essentially two approaches to from the secondary outputs. Kalman filtering techniques
the problem. can then be employed to estimate plant states using
One is to formulate special algorithms to control the secondary output measurements, and hence estimates of
infrequently sampled outputs. For example, Soderstrom1 the controlled output computed from its relationship
formulated a number of minimum variance controllers with the states. Control of the plant is achieved by feed­
enabling the manipulated input to be changed between back of either the state estimates or the output estimates
the sampling intervals of the primary process output. to appropriate controllers. The literature on the abo ve
However, these control algorithms were only developed methods is extensive, and previous publications3-5 are
for first-order plant models and no comparative results examples of the more recent contributions describing the
were presented. Parrish and Brosilow2 also proposed a application issues of these techniques. However, the
controller design method based upon the reconstruction application of Kalman filtering techniques is limited to
of disturbance effects. The controller parameters are those processes that are completely observable from the
determined on-line by heuristic tuning rules. Simulation secondary outputs. For most plants, such a set of

J. Proc. Cont. 1991, Vo! 1, January 3

0959-1524/91/010003-12
Soft-sensors for process estimation and inferential control: M. T Tham et al.

secondary outputs can be difficult to determine and, in reactor feed rate and one or more reactants can be used
some cases, may not even exist. Brosilow et al.61 have to infer polymer properties that would otherwise only be
suggested an estimator design technique based upon an available by relatively slow on-line or off-line analysis.
input-output representation of the plant. The design is In fermentation processes, important internai vari-
approached by obtaining a least squares static estimator, ables such as biomass, substrate and secondary product
which can be used to infer the controlled output from concentrations characterize the state and progress of the
secondary measurements at the steady state. The estima­ fermentation. However, they present measurement diffi-
tor is then additionally made applicable to transient per- culties. In spite of recent encouraging developments in
iods by incorporating heuristically derived lead-lag ele- ion selective and enzyme sensors, optical and high fre-
ments into its structure. Methods for minimizing the quency based methods, most of the concentration vari-
steady state estimation error by appropriate choice of the ables in a fermentation liquid phase cannot be measured
secondary measurements were also proposed. However, accurately and reliably on-line. Therefore, laboratory
a set of secondary measurements for which this error is analyses/assays are usually required to support fermen­
satisfactorily small may not exist. In fact, the evaluation tation supervision and control12. Nevertheless a number
work carried out by Patke et al? indicated that an infer- of successful strategies have been reported13 23. It is inter-
ential control scheme based on the output estimation esting to observe that mechanistic model-based
technique proposed by Brosilow et al. can result in signi- approaches and extended Kalman filtering are mainly
ficant offsets. Moreover, to apply the technique, the employed, even though sufficiently accurate biochemical
gains and approximate time constants of the controlled models are required.
output and all secondary outputs, to all plant distur- This paper presents two algorithms that utilize plant
bances and manipulated inputs, must be known. data sampled at different rates to estimate the controlled
The control strategies reviewed above rely either on output. For example, overheads temperature is used to
infrequent measurements of the controlled output or on estimate top product composition of a distillation col-
the use of secondary measurements. It is, however, poss- umn; and biomass concentration of a fermentation is
ible to use both types of measurements. One such tech­ estimated using off-gas carbon dioxide (CO2) evolution
nique is to set up a ‘parallel cascade’ control strategy9, in rate. The first estimation algorithm is based upon a
which measurements of the controlled output are fed to a general input-output representation24, while the second,
controller whose output acts as the setpoint to a second­ summarized for comparison purposes, is derived from a
ary output controller. Although no offset problems arise state space representation of the plant25. The original
with this control configuration, transient performances ideas behind the two methods are also briefly assessed in
can at times be poor8. Another approach is to use both Guilandoust and Morris26. The extensions to Kalman
measurements of the controlled output and secondary filtering suggested by Kreisselmeier27, and the self-adap-
output within an adaptive inferential control framework. tive Kalman filtering approach suggested by Young28,
In such a scheme, the infrequent measurements of the use a similar model form to that proposed in this work.
controlled output are used only for parameter estimation This major difference, however, lies in the multi-rate
while output estimation, and hence plant control, is aspects addressed here.
achieved using a secondary output at its faster sampling
rate. To our knowledge, very little investigative work has
been carried out in this area, although a similar algor- Adaptive estimation using an input-output
ithm has been proposed previously10. The algorithm10 is, process representation
however, restricted to the assumption of a first-order
plant model, an equal number of disturbance inputs and Process description
secondary measurements, and the use of a dead beat The process dynamics are assumed to be represented by
control law. the following input-output (I/O) relationships:
There are a number of industrial situations where
infrequent sampling of the ‘controlled’ process output yo{t) = Gx(q~ >)u(t ~ mx) +>)w(r) (1)
can present potential operability problems. Common
examples are in product composition control of distilla- vo(0 = G2{q~ ')u(t - tn2) +12(^“ ’)w(Z) (2)
tion columns6"9 and chemical reactors11. In these cases,
the sampling delay is a direct result of the long cycle time X0=Jo(^-^) + vI(z) (3)
of on-line composition analysers. Due to the potential
problems associated with infrequent sampling, the v(0=vo(0 + v2(z) (4)
control of product quality of many industrial multicom-
ponent or high purity columns, for example, is com- where: t is a time index; j0(z) is the controlled output (e.g.
monly achieved by regulating an a priori chosen tray distillation column product composition, fermenter bio­
temperature at its setpoint. However, single-temperature mass concentration); and v0(r) is the secondary output
feedback control is not always effective as maintaining a (e.g. column temperature, fermenter off-gas CO2).Their
constant tray temperature does not necessarily result in corresponding observations, y(t) and v(r), are contami-
constant product composition8. A similar problem arises nated by measurement noises, V](z) and v2(í) respectively,
in polymerization reactors, in which measurements of which are assumed to be zero mean white sequences. The

4 J. Proc. Cont. 1991, Vol 1, January

 Soft-sensors for process estimation and inferential control: M. T. Tham et al.

output measurement delay is denoted by ‘d\ and is and

expressed as an integer multiple of the secondary output
sampling interval. Note that the discretization time step X0=i2(<rl)w(0+v2(z) (9)
is based upon the smaller sampling interval of the
secondary variable. Thus, values of XO only become As qdvx(t) and v2(z) are zero mean white sequences, it
available every d time steps whereas v(Z) is measureable follows that c(0 is also a zero mean and white sequence.
at each time step. The time delays in the responses of the Additionally, since w(z) is a row vector of stationary
controlled and secondary outputs, to changes in the signals, according to the spectral factorization theorem30,
manipulated input u(0 (e.g. column reflux flow, bioreac- Equations (8) and (9) may be expressed as:
tor feed rate) are ‘mf and cm2’, respectively. The term
w(z)eR[ is a vector of unmeasureable disturbances. X0 = H.(<r>(z) (10)
and G2(^_1) and all elements of the row vectors
and I2(<7-1) are discrete time transfer functions and
with the order of numerator less than or equal to that of
the denominator. The term q~x is the backward shift XO^GTMO (11)
operator. It is assumed that changes in any one of the
where Hfq~') and H2(q~ ’) are stable and proper polyno-
components of w(z) affect both j>o(0 and vo(z), otherwise
mial ratios and the sequence co(z) is zero mean and white.
vo(z) will not be suitable for estimating yo{t). Therefore an
It is noted that both XO and XO are generated by the
element of cannot be zero, unless the correspond-
same disturbance sequence <o(z). If this was not so, then
ing element in I2(^~') is also zero, and vice versa.
the two signals would clearly be uncorrelated. This can­
not be the case, because variations in load disturbances
Process disturbances and their representation
are assumed to affect both controlled and secondary
In deriving the estimators, it is important to consider outputs—see Equations (5)—(9). Similarly, neither
how disturbances are to be modelled. It has been Hfq~x) nor //2(^~’) can be zero without the other being
observed4 that process disturbances are characterized by zero also.
infrequently occurring random changes, lasting for per- Substitution of Equations (10) and (11) into Equations
iods of the order of at least the process response time. (5) and (6), and elimination of co(z), results in:
This behaviour has been modelled as Brownian motion
processes3 4’7’29, and was chosen because it can be pictured y(t + d) = [Gx(q~x)q-^+m2-
as a sequence of step changes with independent random H^q-•)]u(t - m2) + [H}(q~ ')/H2(q~ ')] v(0 + e(t + d) (12)
amplitudes, occurring at times described by a Poisson
distribution. Even if this disturbance model may well for mx>m2. For the case in which mx <m2:
represent the physical situation, its use poses an observa-
bility problem. The Brownian motion process descrip- y(t + d) = [GjGt ’) “ q~m2+m^G2(q-l)Hx(q'x)/
tion contains a pole on the unit circle and hence is non- Hi(q~')]u(t -mi) + [Hfq~ ')/H2(q~ >)]v(Z) + e(t + d) (13)
stationary. For the process to be observable from any set
of measurements, the number of measurements should Both Equations (12) and (13) relate measurements of the
therefore exceed the number of disturbance inputs4—a primary output, XO, to the secondary output, v(z), and
situation rarely met in process control. The random step- the values of the manipulated input, u(t). If Gx(q~x),
like behaviour of process disturbances can alternatively G2(q-0, Hfq~ ’) and H2(q~’) were all known, w(0 and v(z)
be modelled by the stationary exponentially correlated could be used in Equation (12) or (13) to compute:
noise (ECN) process, which does not introduce the
observability problem associated with the Brownian y(t + d) = y(t + d) — e(Z + d) (14)
motion process. Therefore, in the sequei, all process dis­
turbances will be assumed to be stationary random which is an estimate of the controlled output, yo(z). From
sequences so that each may be represented as the res­ Equation (7), it is observed that the deviations of y(t 4- d)
ponse of a stable filter to a white input sequence30. from the controlled output is determined essentially by
the variations of V|(z) and v2(z). Nevertheless, noise
Derivation of the IjO estimator effects can be reduced by judicious filtering of the
Equations (1) to (4) can be rewritten as: respective variables. Thus, given a sufficiently accurate
plant model, Equation (12) or (13) may be used to
y(J + d) = Gx(q~ l)u(t - mx) + XO + c(Z + d) (5) provide delay free estimates of the controlled output at
the sampling rate of v(z). In situations where there is a
v(0 = ^-1)w0-w2) + í(0 (6) large number of disturbance inputs or there are complex
or even unknown plant dynamics, the polynomial ratios
where Gx(q~'), <72(<7 ’), HiGz-1) and H2(q~') can only be
approximately obtained using off-line identification. In
e(z + zZ) = ^v1(z)-v2(z) (7) this work, however, these transfer functions are not
assumed to be known a priori. Instead, it is proposed that
X0 = Ii(<r>(0 + v2(Z) (8) their parameters be identified on-line.

J. Proc. Cont. 1991, Vol 1, January 5

Soft-sensors for process estimation and inferential control: M. T Tham et al.

To facilitate on-line parameter estimation, and based using a suitable algorithm. The value of XO = 0T<bO “ d)
on the definitions of G2(q~x), H^q'x) and is then calculated and inserted in <t»(z). The following
x
H2(q~ ), Equations (12) or (13) can be rewritten as: relationship is then used to provide estimates of the
primary output at those time instants when its measure-
ments are not available:
■v(Z + </) = Ã(^)"(Z"ZM) + Ã(^)v(Z) + e(Z + í() (15)
Xr + <O = 0T<I>(O (20)
where w = Min(^i,w2) and, = - otcXz) - - w(i - (« - i)zO
4- Pju(/ — m— 1)4- ..4- PndM7“ m~ nd)
A(?_,)= 1 +a1#"1 + a2í7_2 + ... + aní7_n + xov(f)+.. + Tndv(t-nd)
The parameters of Equation (19) are again updated at
B(^-|) = b1í-' + b2^-2+... + b„^-n (16) time z 4- d when a new value of the primary output again
becomes available. The expression for estimating XO at
C(q~') = c0 + Ctq-1 + c2q~2 + ... + cnq~" the faster sampling rate of the secondary output is thus
given by Equation (20). A convergence proof for this
The value of the integer ‘n’ is chosen to achieve the algorithm, in the form of a multi-rate predictor for the
required accuracy of representation. It should be noted ‘slow’ process output given ‘fast’ manipulated input vari-
that Equation (15) is not of the normal ARMAX type, able data, has been published31.
although its form may, at first sight, suggest otherwise. Equations (19) has n(2</4-1)4-1 unknown parameters
This is because v(t) is not an unknown signal as in the and ‘t/’ may typically lie in the range 2 to 6. A first order
case of an ARMAX relationship. Indeed, both the prim- estima tor (n- 1) may therefore ha ve between 6 and 14
ary output, y(t)9 and secondary output, v(z), respond to unknown parameters and a second order estimator,
changes in control input and disturbances. In its present between 11 and 27. These figures may, at first sight, give
form, the parameters of Equation (15), cannot be esti- the impression that the tuning-in period for larger values
mated in a straightforward manner beause y(f) is only of ‘rf’ will be inconveniently long. This is, however, not
available at every W’ time steps. This problem can, the case. Although the number of unknown parameters
nevertheless, be resolved by making use of Equation (14), increases with ‘zf, each time the controlled output is
and rewriting Equation (15) as: sampled, the data being supplied to the estimation algor­
ithm is also increased [see Equation (19)]. Simulation
A(<r *)X* + d) = Bfo- 9w(r - m) 4- C(q~ ‘)v(Z) (17) studies, investigating the effects of analyser delays up to 6
sample intervals, have in fact confirmed that the initial
For a given value of ‘zf, it can be shown that the follow- tuning-in period increases only marginally for increasing
ing relationship: values of ‘d.'
It has also been observed that estimates of the
(1 + ad<? - d 4-... + CLnáq ' nd)X t + d) = controlled output can often be obtained with sufficient
("1 4- p2z? -2 + ... + pnd? ' nd)w(Z - m) accuracy using a first order estimator. Additionally, the
+ (to + M"1 + ... + Tnd?_nd)v(0 (18) parameter ‘a/ of the first-order estimator may be shown
to be related to ax by ad = (-zz])< As may be deduced
and Equation (17) are equivalent. However, Equation from Equation (12), |a(| is normally less than unity,
(18) has the desired features because all of its parameters except for the special case in which H2(q~') has a domi-
are associated with measured data values, and may there- nant unstable zero. For larger values of , it therefore
fore be estimated. Rearrangement of Equation (18) follows that ad< <1, especially when laj is significantly
results in: less than unity. As a result, fcad’ becomes negligible, and
Equation (20) simplifies to:
y(i) = - «^(í -</)-...- aniy(Z - nd) +
Piw(Z — m — d— 7)4-... 4- pndu(Z — m — (n 4-l)z/)4- y(t + d) = P|U(I“ m~ 1)+-. + $du(t — m — d)
t0v(t - d) 4-... 4- xnd v(z - (n 4-1 )d) + e(z) (19) + Tov(z) 4-... 4- xdv(Z - d) (21)

which can then be expressed more concisely as: Adaptive estimation based on a state space
process representation
XO = 0T<|>(/ ~ íO + XO
Instead of adopting an input-output design approach,
where: an estimator can also be derived based on the following
state space model:
[otd,..., 0tnd, Plvs Pnd» ^ovj ”^nd]

x(z 4-1) = Ax(z) 4- Bu(Z — m) 4- Lw(z) (22)

<|>(Z — = u(t-m-d- 1),..., v(t — d),...]
X0 = DxO“^) + vi(0 (23)
and XO is the equation error. When a measurement of
y(t) is available, the parameter vector 0 is estimated v ( z) = Hx(z) + v2(z) (24)

6 J. Proc. Cont. 1991, Vo! 1, January

 Soft-sensors for process estimation and inferential control: M. T Tham et al.

where t is again a time index, and XO and v(0 are more practical to use the former set of expressions. Since
sampled controlled primary and secondary outputs res- the state estimates can be computed directly, the filtering
pectively. The corresponding measurement noises, v,(0 process need not be performed.
and v2(z), are assumed to be independent random The matrices of the model can be parameterized in
sequences with zero mean and finite variances. The state various ways. Since no prior knowledge about the model
vector is x(z)eRn. It has an initial value x(0), which has structure is assumed, it is reasonable that the parameteri-
the following properties: zation be performed such that the identified parameters
would directly correspond to those of an input-output
E{x(0)} = xo (25) model33. This is best achieved by adopting an observer
canonical model structure. Having estimated the states,
and £{[x(0) - xo][x(0) - x0]T} = Po Equation (30) could then be used to obtain y(t + d) if the
row vector D was known. In principie, D could be esti­
w(z)eRn is a vector of random and unmeasurable load mated using values of y(t) and Equation (26), if the state
disturbances. The primary output measurement delay is estimates were not correlated. Unfortunately, this is not
‘d\ while W is the smallest of the time delays in the the case as may be noted from Equation (29). The
responses of the y(t) and v(t) to changes in the manipu- problem can, however, be resolved by noting that in the
lated input, w(0- Any difference between the two delays observer canonical structure, Equations (29) and (31) are
can be included in the model by extending the state of the form:
vector.
Here, the aim is to estimate the system states using -a,
observations of v(f) and thence obtain an estimate of y(f) -a2
from Equation (23). If the matrices A, B, L, H and D, as x(/+l) =
-an_
well as the noise covariances were known, then the state
-an
estimates, x(z), could be computed from v(t) by Kalman
filtering and used to provide estimates of y(t) using the
v(Z) = [l,0,...,0]x(Z) + e(z) (33)
expression:
From the definition, e(z) = v(Z)-v(Z), Equation (33)
Xí + <0 = Dx(0 (26) becomes:
The Kalman filter for Equations (22) and (24) is given32 v(z) = £,(/) (34)

x(z + 1) = Ax(0 + ®Xz - w) + K(0e(0 (27) Elimination of the state estimates Xj(z) to xn(z) from the
right-hand side of Equation (32) by successive substitu-
where tion yields:
x(0) = xo X](Z 4-1)= v(Z + 1)
c(z) = v(z)-v(z) (28) x2(Z + 1) = - a^t) -- anv(z - n + 2) + b2«(z - m)
+.. + bn«(z - m — n 4- 2) + k2e(z) +.. + kne(z - n + 2) (35)
v(0 = Hx(0
xn(Z + 1) = - anv(z) + bnu(z - w) + kne(z)
v(z) is the filtered value of v(z); K(t) is the Kalman gain
determined via solution of the Riccati equations and e(z) Equation (35) is then substituted into Equation (30) lead-
is the innovations sequence. Although this method is one ing to an expression with the following form:
of the more popular approaches adopted for output esti­
mation, it requires a rather detailed a priori knowledge y(t) — ^]U(t — m — d— 1)+.. + $n~iu(t — m — n — d+ 1)
about the process. This requirement can, however, be + tov(Z — d)+ .. + Tn_jv(z —« — d+ l) + 8]e(z — íZ— 1)
relaxed if the estimation is performed within an adaptive +..4-Ôn |€(Z“-n — <7+ 1) + V](z) (36)
framework. In this case, an innovations model is used in
place of Equations (22) to (24): This can be expressed more compactly as:

x(Z + 1) = Ax(z) + Bu(t-m) + Kc(z) (29) Xz) = 0T(|>(Z - d) + e(t) (37)

XO = Dx(Z-áO + V!(z) (30) with 0T=[pI,..,Pn_1,To,..,Tn-1,ôI,..,5n_1]

v(z) = Hx(z) + e(Z) (31) <|>(Z - d)J = [u(z - m - d- 1),.., v(Z - J),..,e(Z - d- 7),..]

where: K is the time invariant (limiting) Kalman gain, and e(z) is the equation error. Each time a primary mea­
and x is the optimal estimate of the states, x. Although surement becomes available, the parameter vector 0 can
Equations (29)—(31) and (22)—(24) are equivalent32, it is be estimated. Since Equation (37) is independent of state

J. Proc. Cont. 1991, Vol 1, January 7

Soft-sensors for process estimation and inferential control: M. T Tham et al.

estimates, this should not present any difficulties. After signal for any constant parameter or adaptive control
updating 0 at time t, it is used in: algorithm. Examination of Equations (19) and (36) indi-
cates that unless a proportional controller is used, closed
X/+*/)=0T<b(O loop identifiability problems due to correlated data
should not arise [e.g. Ref. 35]. Thus, for a time invariant
= Ç>\u(t — m — 1)4-..4-pn_|u(/ — m — n4- l)4-xov(Z) process, the estimated parameters should converge to
4-.. 4-Tn_|V(/ — n 4-1) 4- ÔiC(Z — 1)4-.. 4- ôn_]C(/ — n 4-1) (38)
final values after an acceptable number of sample inter-
vals. At this stage, parameter estimation could be halted
to provide estimates of the controlled output. The values if desired, with accurate estimates of the controlled out­
of e(Z) in Equations (36) and (38) can be computed from put still being available. In applications to a slowly time
Equations (32) and (33) once the parameters of Equation varying process, it may be preferable to continue with the
(32) have been determined34. 0 is again updated at time parameter estimation at all times. In this case, it may also
t + d when a new measurement of the controlled output be desirable to use an adaptive controller since the use of
becomes available. Equation (38) thus describes the rela- the adaptive inferential estimators does not preclude the
tionship which provides estimates of the controlled out­ application of adaptive control algorithms. Given a suit-
put at the faster secondary output sample rate. able choice of secondary variable, the inferential control
The number of parameters to be estimated in Equation strategy will automatically provide for ‘anticipatory’
(36) is 3w - 2. If secondary output measurements can be control action (in a feedforward sense). This is because
obtained with low noise leveis, and load disturbances disturbance effects will manifest first in measured
only occur infrequently, then the ‘ô’ terms tend to be very secondary variable response, and thus the strategy
small and may thus be ignored. The calculation of should yield significant improvements in disturbance
y(z4-íZ) becomes correspondingly simpler as identifica- rejection performance.
tion of a secondary output model is no longer necessary.
The result is a significant reduction in the number of
unknown parameters and Equation (38) simplifies to: Industrial evaluation

y(t + d) = p|«(r — m — 1)4-.. 4- pn_!w(z — m — «4-1)4- Commercial scale plant validation of both the state—
Tov(/) + ..4-Tn_1v(r-«+1) (39) space and input—output based adaptive estimators (soft
sensors) was undertaken in association with our indus­
At this point, it is interesting to note the similarity trial collaborators. For reasons of commercial confiden-
between Equations (39) and (21). This agreement tiality, in all the figures to be presented, the ordinate
between the final form of two essentially different scales have been removed. In the following applications,
approaches provides additional confidence in the argu- a UD-factorized recursive least squares algorithm
ments presented. (UDRLS)33 was used to estimate the parameters of the
The above discussion has implicitly assumed that the estimators. In all the results to be presented, the estima­
process is completely observable with respect to v(/).The tors were initialized with zero parameters at the start of
question arises, however, as to what happens if this is not each application.
the case. From Equations (22)—(24), it is evident that
y(j) shares some of its poles with v(z). On the other hand, Application to a continuous fermenter
the zeros of y(t) and v(í) are, in general, different from In this study, the process is a continuous stirred tank
each other. Loosely speaking, substituting the state esti­ fermenter producing a fungai mycelium. Following
mates into Equation (23) may be regarded as assigning downstream processing, the mycelia are the final process
the poles of v(z) to y(t). The zeros assigned to y(t) are product. However, specifications placed upon its quality
then determined by the parameter estimation procedure. dictate certain fermenter operating conditions. In parti­
As may be expected, this operation places the zeros such cular, biomass concentration has to be tightly regulated.
that the effect of the difference between the poles of v(t) The current strategy is based upon 4 hourly laboratory
and y(t) is minimized. Thus, it is also possible that some analysis of biomass concentration and the dilution rate
of the poles of v(t) are effectively cancelled. Simulation adjusted upon the return of assay information. Although
studies carried out on systems with widely different prim- process analysis reveals that a sampling interval of 1 h
ary and secondary output dynamics indicate satisfactory would be more suitable for maintaining a constant bio­
estimation despite the differences in dynamics. mass concentration, it is however, not possible to sample
biomass at this frequency. In fact, present policy is based
upon increasing the sampling interval, so as to reduce the
Adaptive inferential control demand for 24 h laboratory support and operator super-
vision. A move to 8 or 12 h analysis would therefore be
With the availability of ‘fast’ primary output estimates, desirable. This has important consequences on the
(e.g. fermenter biomass concentration, distillation col- control strategy in that an altemative technique to off-
umn product composition), closed loop adaptive ‘infer­ line analysis must be found to regulate biomass effecti­
ential’ control of that variable becomes a feasible control vely.
option. The values of y(t + d) can be used as a feedback The adaptive estimators provide a means by which

8 J. Proc. Cont. 1991, Vol 1, January

 Soft-sensors for process estimation and inferentia! control: M. T. Tham et al.

Figure 3 Fermenter dry weight assayed approximately 8 hourly com-

Figure 1 Fermenter CO2 evolution rate and dilution rate measure- pared with estimated biomass concentration
ments

strates the ability of the ‘state—space based’ estimator to

provide very acceptable predictions of biomass transient
behaviour. The biomass assay results can be seen to
occur at intervals that are not completely regular. With
the state—space based estimator, this is not important
since the algorithm can accommodate such variations in
assay times (P.A. Lant, PhD Thesis in preparation,
University of Newcastle-upon-Tyne, UK). The perfor­
mance of the estimator during the plant upset is also
encouraging. It predicted the down-turn well, with only a
slight undershoot at around 280 h. This is due to conti-
nuing estimator parameter adaptation to the new operat-
ing conditions. At time 248 h, the laboratory assay indi-
cated a large increase in biomass concentration. It is
interesting to observe, however, that the estimator conti-
nued to predict a fali based upon measurements of dilu­
tion rate and CO2 evolution rate. Other possibly ‘suspect’
biomass assays can also be observed at times 58, 100 and
Figure 2 Fermenter dry weight assayed approximately 4 hourly com-
pared with estimated biomass concentration 183 h respectively. Nevertheless, the ability of the estima­
tor to predict the general trend of biomass concentration
response was not compromised. The effect of increasing
frequent estimates of biomass can be obtained using CO2
the interval of sampling for biomass concentration
evolution rate and dilution rate. Measurements of the
analysis from 4 to 8 h is shown in Figure 3. It can be
latter variables can be obtained frequently and with little
observed that there was only a marginal deterioration in
delays. Hence, an inferential estimator supplied with
the quality of biomass estimates. However, and as
hourly CO2 and dilution rate measurements, supported
expected, the initial estimator ‘tuning-in’ period is
by infrequent and irregular biomass concentration
slightly longer (see Figure 2).
assays, will enable hourly predictions of biomass con­
centration. In this application, the state—space based
estimator was implemented. Application to a high purity distillation column
Figure 1 shows the CO2 and dilution rate measure­ Distillation columns can also exhibit control problems,
ments o ver a 470 h period of continuous operation. A due to the delays introduced by on-line product ana-
major plant disturbance occurred at around time 285 h, lysers. The problem can be exacerbated in the case of
causing a rapid fali in CO2 evolution rate. Although this columns with high purity products, where composition
was unexpected, it provided the estimator with a test of dynamics can be extremely non-linear. In one such col­
its ability to track major process changes. Note that at umn, an industrial demethanizer, the aim of the control
approximately 360 h, CO2 measurements became strategy is to regulate the top product composition by
unavailable and was assumed constant for a period of varying reflux flow. Here, all process variables are mea-
about 20 h. The laboratory biomass assays (step-like sured every 5 min except for top product composition.
response) and inferred biomass concentration over the Being measured by an analyser, composition values are
fermentation period are compared in Figure 2. It demon- available only at 20 min intervals. The aim, therefore, is

J. Proc. Cont. 1991, Vol 1, January 9

Soft-sensors for process estimation and inferential control: M. T. Tham et al.

Overheads temperature

Figure 6 Demethanizer column temperature measurement

Figure 7 Demethanizer column top produet composition analysis

Figure 5 Demethanizer column reflux flow rate measurement compared with estimated composition over 8250 min

to use ‘fast’ measurements of column overheads vapour degraded performances occurred only during large pro­
temperature, together with reflux flow rate, to provide cess transients. Recall that high purity columns possess
estimates of produet concentration at 5 min intervals. As strongly non-linear composition dynamics, and thus
a secondary variable, a tray liquid temperature would be represent a severe test of the resilience of the estimator.
preferable. The problems with the use of a vapour tem­ The lapse in performance quality is attributed to the
perature is that it is overly sensitive to vapour distur- estimator having to ‘tune’ to relatively frequent changes
bances; changes due to column boil-up control and in plant operating conditions. Nevertheless, when new
numerous other factors which need not necessarily affect steadier operating conditions are approached, the esti­
produet composition. However, the choice of a vapour mation error is significantly reduced. This is in marked
temperature was dictated by its availability at the time of contrast to the steady state errors reported in the evalu-
the tests. ation study of other inferential control schemes by Patke
Results are presented for a 137 h segment of a 20 day et al? Figure 8 shows the performance of the estimator
application using the input—output form of the estima- between 15000 and 16000 min, and provides a more
tor. The data for feed flow rate, reflux flow rate and detailed picture of estimator responses during a period
column temperature are shown in Figures 4—6 respecti- when process operating conditions are changing. It can
vely. A comparative plot of actual produet composition be observed that the estimates of top produet compo­
with its estimates is shown in Figure 7. In spite of the sition are very close to the composition analyser results.
‘bad’ choice of secondary variable, the estimator per- Indeed, the errors are within the tolerances demanded by
formed remarkably well. It can be observed that the plant operating personnel.

10 J. Proc. Cont. 1991, Vol 1, January

 Soft-sensors for process estimation and inferential control: M. T Tham et al.

Measurement and estimates

Figure 10 Polymerization reactor coolant flow rate measurement

Figure 8 Demethanizer column top product composition analysis

compared with estimated composition over 1000 min

Figure 11 Polymerization reactor hydrogen concentration measure­

Figure 9 Polymerization reactor fresh propylene feed rate measure­ ment
ment

Application to a polymerization reactor

As a further test of the utility of the estimator, the
input—output algorithm was applied to an industrial
polymerization process. The melt flow index (MFI) of
the product was estimated using measurements of reac­
tor feed rate, reactor coolant flow rate and hydrogen
concentration above the reaction mass (shown in Figures
9-11 respectively). The secondary variables were mea-
sured at 10 min intervals, while measures of MFI were
obtained from laboratory analyses every 2 h. Figure 12
shows the results of the very first trial application to this
process and compares the laboratory assayed melt flow
index (step-like response) and its estimates. The tuning-
in period, from zero initial parameter values took
approximately 300 h (equivalent to 150 laboratory sam-
ples). Although the algorithm was implemented with
minimal process knowledge used to ‘condition* the esti­
mator, its performance is good. In particular, the ability
to track the change in polymer grade at around 1085 h is Figure 12 Polymerization reactor melt flow index analysis compared
particularly encouraging. with estimated melt flow index over 14500 min

J. Proc. Cont. 1991, Vo! 1, January 11

Soft-sensors for process estimation and inferential control: M. T. Tham et al.

Implementation aspects
In the application of the adaptive estimators, several
important implementation issues arose. The data used
for parameter estimation have to be suitably ‘con-
ditioned’. Data scaling to ensure that input—output data
ranges were of the same order, is particularly important.
Care must also be taken to ensure that the measured
signals are filtered for noise rejection, anti-aliasing, sud-
den changes or spikes and ‘rogue’ data. Failure to
achieve satisfactory data conditioning will lead to poor
estimator performance. In the abo ve case studies, all
process measurements were filtered using low phase-shift
algorithms and transformed to represent deviations
about a mean operating levei. Data spikes were attenu-
tated via the use of logic filters designed based on on-line
calculated statistics of each data variable. Additionally,
to promote robust estimator performance, whenever a
primary output analysis (e.g. fermenter biomass, column
product composition) becomes available, it is used in the
algorithm to replace the corresponding estimated value.
It is in ensuring the integrity of the algorithms to
specific process problems that proved to be most
demanding. For instance, periodic re-calibration of the
infrared carbon dioxide analyser in the bioreactor appli­
cation (or the on-line chromatograph in the case of the
distillation column), unless accounted for, can cause sig-
nificant problems. Due to the adaptive nature of the
algorithm, howeverjt is capable of dealing with the slow
drifts in calibration usually experienced. The techniques
adopted here reílect the experiences with our work on
adaptive and self-tuning control18’19 36.

Adaptive inferential control of a distillation product

column
Figure 13 Schematic diagram of pilot plant distillation column: CR,
analyser recorder; FRC, flow recorder/controller; GC, gas chromato­
The adaptive estimators described above can also be graph; LC, levei controller
included as part of a feedback control system to provide
an adaptive inferential control strategy. The practicality
of such a scheme to effect distillation product compo­ variable and is sampled every 0.5 min. The UDRLS
sition control has been evaluated by non-linear simula- algorithm was again used to estimate the parameters of a
tion. The process under consideration is a 10 stage pilot lst order input—output estimator. Although the perfor­
plant column, installed at the University of Alberta, mances of estimators of various orders have been investi-
Canada. It separates a 5O-5Owt% methanol-water feed gated, in this application, it was generally found that the
mixture, which is introduced at a rate of 18.23gs_1 into advantages to be gained from the use of high order
the column on the 4th tray. Bottom and top product estimators were marginal.
compositions are to be maintained at 5wt% and 95wt% To provide control, fixed parameter PI controllers
methanol respectively. A schematic diagram of the pilot were used and their settings determined using Ziegler-
plant column is shown in Figure 13. The column is Nichols ultimate sensitivity tuning rules, followed by
modelied by a comprehensive set of dynamic heat and fine-tuning to obtain responses with minimum integral of
mass balance relationships37. Both the pilot plant and the absolute error. Compara ti ve performances of PI feed­
non-linear model have been widely used by many investi- back control and adaptive inferential control, under
gators to study different advanced control schemes3638. major disturbance conditions, were assessed by subject-
The control objective is to use steam flow rate to the ing the column to ± 25% feed flow changes from steady
reboiler to regulate bottom product composition, when state operating conditions. This disturbance sequence
the column is disturbed by step changes in feed flow rate. consisted of a 25% increase from steady state; a 25%
Bottom product composition is measured by a gas chro­ decrease back to steady state; a 25% decrease from
matograph, which is subject to an analyser delay of 3 steady state; and finally a 25% increase back to steady
min. The liquid temperature of tray 3, i.e. the tray imme- state.
diately below the feed tray, was chosen as the secondary The open loop responses of bottom product

12 J. Proc. Cont. 1991, Vo! 1, January

 Soft-sensors for process estimation and inferential control: M. T. Tham et al.

Figure 14 Methanol distillation column bottom product composition

(PI control of analyser composition) ± 25% feed flow rate disturbances

Figure 16 Methanol distillation column bottom product composition

(Adaptive inferential control using reboiler temperature, ± 25% feed
flow rate disturbances)

back control can be regarded as implicitly providing for

dead-time compensation in the closed loop, and thus
allows the use of a controller with higher gains. This was
indeed the case, and the PI settings used in the inferential
scheme were Á^= -175 and T1 = 2.5 min. As a result,
disturbance rejection responses are superior to those
obtained using conventional feedback control. It is also
suggested that fine tuning of the PI settings in the adap­
Figure 15 Methanol distillation column bottom product composition. tive inferential control scheme could well provide for
(Adaptive inferential control using tray 3 temperature, ±25% feed further improvements.
flow rate disturbances)
It is noted that although the estimator tracks the
actual composition response closely during the first part
sition and tray 3 liquid temperature to ± 10% step of the transient, following the +25% increase in feed
changes in steam and feed fio ws (not shown), indicate rate, the actual composition response exhibits slow offset
that both bottom product composition and tray 3 tem­ removal as the estimator parameters tune in under low
perature exhibit quite different gains and time constants excitation conditions. This is particularly apparent
for both positive and negative steps in steam and feed during the two major disturbances occurring at times 20
flow rates. Indeed, the dynamics of the column become and 160 min, respectively. In practice, a continuously
severely non-linear especially when it is subject to large repeating sequence of disturbances, such as used in this
(> ±20%) changes in reflux, steam or feed flow rates. study, would not be expected to occur. Therefore, a
The linear estimator therefore has to cope with the resul- longer and richer adaptation period would be available.
tant non-linearities of all these responses. This is an The use of various tray temperatures as the secondary
important factor when assessing the adaptation proper- output ha ve also been investigated. Tray 3 temperature
ties of the estimator to changing process characteristics. was selected as the secondary variable to demonstrate the
Note that in figures to be presented, the composition performance of the estimator (soft-sensor) based inferen­
plots correspond to ‘actual’ bottoms composition, j>0(í) tial control strategy using a ‘badly’ chosen secondary
rather than its analyser delayed value, y(t). The perfor­ measurement. If the secondary measurement was made
mance of the PI controller, using the measured product closer to the sampling point of the primary output, then
composition as the feedback signal is shown in Figure 14, the performance of the estimator would be significantly
while Figure 15 shows the performance of the adaptive enhanced. This is illustrated in Figure 16, which shows
inferential control scheme. Deviations of actual compo­ the performance of the adaptive inferential control
sition from its steady state value (full line) are plotted scheme when the temperature of the liquid in the reboiler
along with the estimated composition (broken line). Par­ was used as the secondary variable. As expected, the
ticular note should be made of the difference in ordinate estimated composition values (broken line) accurately
scales in Figures 14 and 15. Clearly, an overall improved reflects the state of the actual composition (full line).
transient behaviour is obtained with the adaptive infer­ Other comparative studies have been performed using
ential control strategy. the multicomponent distillation column model described
The PI settings used in the conventional control stra­ in Patke et al* A much improved, offset free control
tegy were K= -60 and Tj = 20 min. It may be argued performance, has been reported using the estimators des­
that the proportional gain is too high, hence the results in cribed in this paper24 26. The ability to anticipate distur­
Figure 14. However, reducing the gain of the PI con­ bance effects results in a significantly improved control
troller, and increasing the integral time to account for the behaviour. This has been achieved here despite a poor
measurement delay, would result in a more sluggish res­ choice of secondary measurement location. The studies
ponse to that shown in Figure 14. Furthermore, there also hopefully indicate the potential of these estimators
was an increase in peak overshoot of about 0.6 wt% in providing for improved product quality regulation of
(response not shown). The use of the estimates for feed- multicomponent columns. It is also worth noting that in

J. Proc. Cont. 1991, Vol 1, January 13

Soft-sensors for process estimation and inferentia/ control: M. T. Tham et al.

practice, it is unlikely that all available secondary out­ Chemicals and Polymers, the UK SERC and the Univer-
puts will exhibit highly non-linear dynamics. Neither sity of Newcastle upon Tyne.
would the disturbances affecting the plant change as
drastically as the steps used in this study. Hence the
References
ability of the adaptive algorithm to track slow changes in
plant dynamics should be adequate. As has been shown, 1 Soderstrom, T, in ‘Methods and Applications in Adaptive Control’,
the adaptive inferential control technique described (Ed. H. Unbehaven), Springer-Verlag, 1980
above appears to possess such capabilities. 2 Parrish, J. R. and Brosilow, C. B. Automático 1985, 21 (5), 527
3 Morari, M. and Stephanopoulos, G. AIChE Journol 1980, 26 (2),
247
4 Morari, M. and Stephanopoulos, G. Int. J. Control 1980,31 (2), 367
Conclusions 5 Morari, M. and Fung, K. W. Computers and Chemical Engineering
1982,6(4), 271
In this paper, two adaptive estimators for inferring infre- 6 Cwiklinski, R. R. and Brosilow, C. B. Proc. JACC 1977, 1530
7 Joseph, B., and Brosilow, C. B. AIChE Journal 1978,24 (3), 485 and
quently measured plant outputs, using more frequently 500
measured secondary variables, ha ve been derived. In one 8 Patke, N. G., Deshpande, P. B. and Chou, A. C. Ind. Eng. Chem.
approach, a general input—output representation of the Process Des. Dev., 1982, 21, 266
9 Luyden, W. L. Ind. Eng. Chem. Fundam 1973, 12 (4), 463
plant formed the starting point, while the other utilizes a 10 D’Hulster, F. M. and van Cauwenberghe, A. R., Proc. 2nd World
state space model. These algorithms ha ve been referred Congress on Chemical Engineering, Montreal, Canada, 1981, 378
to as ‘soft-sensors’ since they are software based rather 11 Wright, J. D., MacGregor, J. F., Jutan, A., Tremblay, J. P. and
Wong, A. Proc. JACC 1977, 1516
than being hardware based. Both estimators approxi- 12 Montague, G. A. Morris, A. J. and Bush, J. R. IEEE Control
mate the complex relationship between the two sets of Systems Mag. 1988, 44
output variables. The parameters of the estimators are 13 Bastin, G. and Gevers, M. R. IEEE Trans. Auto. Control 1988, 33
(7), 650
identified on-line using values of the infrequently sam- 14 Dochain, D. and Bastin, G. Automática 1984, 20, 621
pled plant output and the relatively faster sampled 15 Dochain, D. and Bastin, G. Automática 1986, 22, 705
secondary outputs and inputs. The technique enables the 16 Halme, A. and Seklainaho, A. Proc. 6th. IFAC Symp. on Identifi-
cation and System Parameter Estimation, Washington, USA, 1982
controlled output to be inferred at the same rate at which 17 Halme, A. Proc. lst. IFAC Symp. on Modelling and Control of
the secondary output is being sampled. Biotechnological Processes, Noordwijkerhout, Holland, 1985, 179
The algorithms ha ve been successfully applied to 18 Montague, G. A., Morris, A. J., Wright, A. R., Aynsley, M. and
Ward, A. C. Proc. IEE, Part D 1986,133 (5), 240
several industrial processes. Fermenter biomass concent- 19 Montague, G. A., Morris, A. J., Wright, A. R., Aynsley, M. and
ration was estimated using off-gas CO2 and fermenter Ward, A. C. Can. J. Chem. Eng. 1986, 64, 567
dilution rate measurements; top product composition of 20 Nihtila, M., Harmo, P. and Perttula, M. Proc. 9th. IFAC World
Congress, Budapest, Hungary, 1984
a demethanizer column was estimated using column 21 San, K. Y. and Stephanopouíos, G. Biotech. & Bioeng. 1984, 26,
temperature and reflux flow rate measurements; and 1189
polymer melt flow index was estimated using measure­ 22 Stephanopoulos, G. and San, K. Y. Biotechnol. & Bioeng, 1984, 26,
1176
ments of reactor feed rate, reactor coolant flow rate and 23 Swiniarski, R., Lesniewski, A., Dewski, M. A. M., Ng, M. H. and
hydrogen concentration above the reaction mass. The Leigh, J. R. Proc. lst. IFAC Workshop on Modelling and Control
results highlight the potential of the proposed adaptive of Biotechnical Processes, Espoo, Finland, 1982, 231
24 Guilandoust, M. T., Morris, A. J. and Tham, M. T. Ind. Eng. Chem.
estimation schemes for solving ‘difficult’ industrial Res. 1988, 27, 1658
control problems caused by instrumentation limitations. 25 Guilandoust, M. T., Morris, A. J. and Tham, M. T. Proc. IEE Part
With regard to closed loop ‘adaptive inferential’ D 1987, 134(3), 171
26 Guilandoust, M. T. and Morris, A. J. Proc. 35th Canadian Chemi­
control, early detection of load disturbances effects (in a cal Eng. Conf., Calgary, Canada, 1985, p. 213
feedforward sense) permits the achievement of signifi- 27 Kreisselmeier, G. IEEE Trans. Auto. Control 1977, 22, 2
cantly faster disturbance rejection. Consequently, it is 28 Young, P. C. Electronic Letters, 1979, 15 (12), 358
29 Tuffs P. S. and Clarke, D. W. Proc. IEE, Part D, 1985,128 (3), 93
expected that control schemes using the estimated output 30 Astrom, K. J. in Tntroduction to Stochastic Control Theory’, Aca-
in conjunction with fixed parameter PI or PID con- demic Press, London, UK, 1970
trollers will out-perform other conventional feedback PI/ 31 Lu, W., and Fisher, D. G. Int. J. of Control, 1988, 48 (1), 149
32 Anderson, B. D. O.and Moore, J. L. in ‘Optimal Filtering’, Prentice
PID controllers, or indeed self-tuning control schemes, Hall Publishers, Hemel Hempstead, Herts, UK, 1979
which rely solely on infrequent output measurements. 33 Ljung, L. and Soderstrom, T. ‘Theory and Practice of Recursive
This potential for adaptive inferential control was also Identification’, MIT Press, London, UK, 1983
34 Tsay, Y. T. and Shieh, L.S. Proc. IEE, Part D, 1981, 128 (3), 93
demonstrated. 35 Isermann, R. in ‘Digital Control Systems’, Springer-Verlag, FRG
36 Morris, A. J., Nazer, Y. and Wood, R. K. Optimal Control and
Applications 1982, 3, 363
Acknowledgements 37 Simonsmier, U. ‘Non-linear Binary Distillation Column Models’,
MSc Thesis, University of Alberta, Canada, 1977
38 Morris, A. J., Nazer, Y., Wood, R. K. and Lieuson, H. Proc. IFAC
The authors would like to acknowledge support from Conf. on Digital Computer App. to Process Control, (Ed. Isermann
ICI Biological Products Business, ICI Engineering, ICI and Haltenecker), Dusseldorf, FRG, 1981, p. 345

14 J. Proc. Cont. 1991, Vol 1, January