Cross Directional Control Graham C. Goodwin Day 4: Lecture 4 16th - - PowerPoint PPT Presentation

Cross Directional Control

Graham C. Goodwin Day 4: Lecture 4 16th September 2004 International Summer School Grenoble, France

Centre for Complex Dynamic Systems and Control

• 1. Introduction

In this lecture we describe a practical application of receding horizon control to a common industrial problem, namely web-forming processes. Web-forming processes represent a wide class of industrial processes with relevance in many different areas such as paper making, plastic film extrusion, steel rolling, coating and laminating.

Centre for Complex Dynamic Systems and Control

In a general set up, web processes (also known as film and sheet forming processes) are characterised by raw material entering one end of the process machine and a thin web or film being produced in (possibly) several stages at the other end of the machine. The raw material is fed to the machine in a continuous or semi-continuous fashion and its flow through the web-forming machine is generally referred to as the machine direction [MD].

Centre for Complex Dynamic Systems and Control

Sheet and film processes are effectively two-dimensional spatially distributed processes with several of the properties of the sheet of material varying in both the machine direction and in the direction across the sheet known as the cross direction [CD].

Centre for Complex Dynamic Systems and Control

cross direction actuators sensors machine direction

Figure: Generic web-forming process.

Centre for Complex Dynamic Systems and Control

The main objective of the control applied to sheet and film processes is to maintain both the MD and CD profiles of the sheet as flat as possible, in spite of disturbances such as variations in the composition of the raw material fed to the machine, uneven distribution of the material in the cross direction, and deviations in the cross-directional profile.

Centre for Complex Dynamic Systems and Control

In order to control the cross-directional profile of the web, several actuators are evenly distributed along the cross direction of the

• sheet. The number of actuators can vary from only 30 up to as

high as 300. The film properties, on the other hand, are either measured via an array of sensors placed in a downstream position

• r via a scanning sensor that moves back and forth in the cross
• direction. The number of measurements taken by a single scan of

the sensor can be up to 1000.

Centre for Complex Dynamic Systems and Control

Difficulties

the high dimensionality of the cross-directional system; the high cross-direction spatial interaction between actuators; the uncertainty in the model; the limited control authority of the actuators.

Centre for Complex Dynamic Systems and Control

• 2. Problem Formulation

It is generally the case that web-forming processes can be effectively modelled by assuming a decoupled spatial and dynamical response. This is equivalent to saying that the effect of

• ne single actuator movement is almost instantaneous in the cross

direction whilst its effect in the machine direction shows a certain dynamic behaviour.

Centre for Complex Dynamic Systems and Control

These observations allow one to consider a general model for a cross-directional system of the form yk = q−dh(q) ¯ Buk + dk, (1) where q−1 is the unitary shift operator.

Centre for Complex Dynamic Systems and Control

It is assumed that the system dynamics are the same across the machine and thus h(q) can be taken to be a scalar transfer

• function. In addition, h(q) is typically taken to be a low order, stable

and minimum-phase transfer function. A typical model is a simple first-order system with unit gain: h(q) = (1 − α) q − α . (2) A transport delay q−d accounts for the physical separation that exists between the actuators and the sensors in a typical cross-directional process application.

Centre for Complex Dynamic Systems and Control

The matrix ¯ B is the normalised steady state interaction matrix and represents the spatial influence of each actuator on the system

• utputs. In most applications it is reasonably assumed that the

steady state cross-directional profile generated by each actuator is

• identical. As a result, the interaction matrix ¯

B usually has the structure of a Toeplitz symmetric matrix.

Centre for Complex Dynamic Systems and Control

The main difficulties in dealing with cross-directional control problems are related to the spatial interaction between actuators and not so much to the complexity of dynamics, which could reasonably be regarded as benign. A key feature is that a single actuator movement not only affects a single sensor measurement in the downstream position but also influences sensors placed in nearby locations. Indeed, the interaction matrix ¯ B is typically poorly conditioned in most cases of practical importance.

Centre for Complex Dynamic Systems and Control

The poor conditioning of ¯ B can be quantified via a singular value decomposition

¯

B = USVT (3) where S, U, V ∈ Rm×m. S = diag{σ1, σ2, . . . , σm} is a diagonal matrix with positive singular values arranged in decreasing order, and U and V are orthogonal matrices such that UU = UU = Im and VV = VV = Im, where Im is the m × m identity matrix. If ¯ B is symmetric then U = V.

Centre for Complex Dynamic Systems and Control

If ¯ B is poorly conditioned then the last singular values on the diagonal of S are very small compared to the singular values at the top of the chain {σi}m

i=1. This characteristic implies that the control

directions associated with the smallest singular values are more difficult to control than those associated with the biggest singular values, in the sense that a larger control effort is required to compensate for disturbances acting in directions associated with small σi.

Centre for Complex Dynamic Systems and Control

This constitutes not only a problem in terms of the limited control authority usually available in the array of actuators, but it is also an indication of the sensitivity of the closed loop to uncertainties in the spatial components of the model.

Centre for Complex Dynamic Systems and Control

The control objective in cross-directional control systems is usually stated as the requirement to minimise the variations of the output profile subject to input constraints. This can be stated in terms of minimising the following objective function: V∞ =

∞

• k=0

yk2

2

subject to input constraints

uk∞ ≤ umax.

(4)

Centre for Complex Dynamic Systems and Control

Another type of constraint typical of CD control systems is a second-order bending constraint defined as1

∆ui+1

k

− ∆ui

k∞ ≤ bmax

for i = 1, . . . , m, (5) where ∆ui

k = ui k − ui−1 k

is the deviation between adjacent actuators in the input profile at a given time instant k.

1The superscript indicates the actuator number. Centre for Complex Dynamic Systems and Control

• 3. Example 1

To illustrate the ideas involved in cross-directional control, we consider a 21-by-21 interaction matrix ¯ B with a Toeplitz symmetric structure and exponential profile: bij = e−0.2|i−j| for i, j = 1, . . . , 21, (6) where bij are the entries of the matrix ¯ B.

Centre for Complex Dynamic Systems and Control

2 4 6 8 10 12 14 16 18 20 0.2 0.4 0.6 0.8 1

Cross-directional index

Figure: Cross-directional profile for a unit step in actuator number 11.

Centre for Complex Dynamic Systems and Control

We consider the transfer function h(q) = 1 − e−0.2 q − e−0.2 , (7) which is a discretised version of the first-order system

˙

y(t) = −y(t) + u(t) with sampling period Ts = 0.2 sec.

Centre for Complex Dynamic Systems and Control

The next figure shows the singular values of the interaction matrix

¯

• B. We observe that there exists a significant difference between

the largest singular value σ1 and the smallest singular value σ21, indicating that the matrix is poorly conditioned. Dealing with the poor conditioning of ¯ B is one of the main challenges in CD control problems as we will see later.

Centre for Complex Dynamic Systems and Control

2 4 6 8 10 12 14 16 18 20 1 2 3 4 5 6 7 8

Singular values index

Figure: Singular values of the interaction matrix ¯ B.

Centre for Complex Dynamic Systems and Control

In order to estimate the states of the system and the output disturbance dk, a Kalman filter is implemented as described for an extended system that includes the dynamics of a constant output disturbance: xk+1 = Axk + Buk, dk+1 = dk, yk = Cxk + dk,

Centre for Complex Dynamic Systems and Control

In our case A = diag{e−0.2, . . . , e−0.2}, B = (1 − e−0.2) ¯ B, C = Im.

Centre for Complex Dynamic Systems and Control

The state noise covariance Qn =

• Im

100Im

• ,

and output noise covariance Rn = Im, were considered in the design of the Kalman filter.

Centre for Complex Dynamic Systems and Control

We will consider the finite horizon quadratic objective function with both prediction and control horizons set equal to one, that is V1,1 = 1 2(y

0Qy0 + u 0Ru0 + x 1Px1).

(8) Q = Im, R = 0.1Im. (9)

Centre for Complex Dynamic Systems and Control

We assume the system is subject to physical constraints on the inputs of the form:

|ui

k| ≤ 1

for all k, i = 1, . . . , 21.

Centre for Complex Dynamic Systems and Control

Design 1

The first control strategy that we try on the problem is a linear quadratic Gaussian [LQG] controller designed with the same weighting matrices as above. This design clearly does not take into consideration the constraints imposed on the input profile. As might be expected, the application of such a blind (or serendipitous) approach to the problem would, in general, not achieve satisfactory performance.

Centre for Complex Dynamic Systems and Control

Cross-directional index

(a) Input profile.

Cross-directional index

(b) Output profile.

Figure: Input-output steady state profiles with input clipping.

Centre for Complex Dynamic Systems and Control

The above results illustrates a phenomenon that is well known in the area of cross-directional control, namely alternate inputs across the strip converge to large alternate values, that is, “input picketing” occurs. We will see below, when we test alternative design methods, that this picketing effect can be avoided by careful design leading to significantly improved disturbance compensation.

Centre for Complex Dynamic Systems and Control

Design 2

We next try RHC considering initially only input constraints. The achieved steady state input and output profiles are presented below where we have also included, for comparison, the profiles

• btained with the input clipping approach.

Centre for Complex Dynamic Systems and Control

Cross-directional index

(a) Input profile.

Cross-directional index

(b) Output profile.

Figure: Input-output steady state profiles using RHC (square-dashed line) and input clipping (circle-solid line).

Centre for Complex Dynamic Systems and Control

We observe, perhaps surprisingly, that the steady state response achieved with RHC does not seem to have improved significantly compared with the result obtained by just clipping the control in the LQG controller. In addition, the input profile obtained with RHC seems to be dominated by the same high spatial frequency modes as those that resulted from the input clipping approach.

Centre for Complex Dynamic Systems and Control

However, this is a reasonably well understood difficulty in CD control systems: The “picket fence” profile in the input arises from the controller trying to compensate for the components of the disturbance in the high spatial modes which, in turn, require bigger control effort, driving the inputs quickly into saturation.

Centre for Complex Dynamic Systems and Control

Design 3 (Using the Singular Value Structure of the Hessian

The commonly accepted solution to this inherent difficulty is to let the controller seek disturbance compensation only in the low spatial frequencies.

Centre for Complex Dynamic Systems and Control

When the prediction horizon chosen is N = 1 then in the vector formulation of the quadratic optimisation problem we can write

Γ = B

and the Hessian of the objective function is simply H = BQB + R

= ¯

B ¯ B + R. This implies that the singular values of the Hessian are simply the singular values squared of the interaction matrix ¯ B shifted by the weighting in the input R.

Centre for Complex Dynamic Systems and Control

We next limit the optimization to the first few singular vectors. This has two potential advantages: (i) It avoids chasing hard to control high spatial frequency distances. (ii) It may improve robustness since the gain associated with high spatial frequencies may be poorly defined.

Centre for Complex Dynamic Systems and Control

Cross-directional index

(a) Input profile.

Cross-directional index

(b) Output profile.

Figure: Input and output steady state profiles using the SVD–RHC strat- egy (circle-solid line) and RHC (square-dashed line).

Centre for Complex Dynamic Systems and Control

We observe that in this case the picket fence profile has disappeared from the input whilst the output profile has not changed significantly. Clearly a slight degradation of the output variance is to be expected owing to the suboptimality of the

• strategy. The steady state profiles obtained with RHC have been

repeated in the above figure for comparison purposes.

Centre for Complex Dynamic Systems and Control

• 4. Example 2

We conclude this lecture with a second example having 11 actuators and 11 sensors. Here we will run the simulations in real time so that you can see the picketing develop when the disturbance is applied.

Centre for Complex Dynamic Systems and Control

Centre for Complex Dynamic Systems and Control

Centre for Complex Dynamic Systems and Control

Centre for Complex Dynamic Systems and Control

Centre for Complex Dynamic Systems and Control

• 5. Conclusions

This lecture has presented a realistic application of constrained control to a difficult industrial control problem, namely cross directional control. This system has high complexity due to: i. the large number of actuators (several hundred), ii. large interactions, and

• iii. severe constraints on the actuator signals.