CONTENTS

In the field of modern control engineering, much effort has been expended into the development and analysis of novel control strategies. A common assumption in these studies is that the required data is available. Unfortunately, this is often not the case in practice. However, the problem is gaining attention and a variety of solutions have been proposed to deal with various aspects of the difficulties associated with data.

5.1. Inferential Estimation

A major problem is the lack of on-line instrumentation to measure quantities that define product quality, e.g. stickiness of adhesives, smoothness of sheet material, melt flow index of polymers, flash points of fuels, etc. These are often provided by laboratory analyses resulting in infrequent feedback and substantial measurement delays, rendering automatic process control impossible. Inferential estimation is one method that has been designed to overcome this problem. The technique has also been called 'sensor-data fusion' and 'soft-sensing'.

Apart from the main quality variable, there are usually other variables such as temperatures, pressures, flows, etc., that are associated with a process. Changes in some of these variables are indicative of changes in product quality. Thus, by monitoring suitable secondary variables, it is often possible to 'infer' the state of the quality variable. Process operators and engineers do this on a daily basis in running process plants. However, the process may be complex and there could be many factors that affect product quality. As a result, the relationship between process conditions and product quality may not be straight forward, leading to inaccuracies in human judgement. Inferential Estimation alleviates this problem. The technique uses easily obtainable measurements of variables that are known to influence product quality, together with those of product quality when available, to generate estimates of product quality.

As with feedback control strategies, in applications to non-linear systems, the relationship between secondary variables and the primary output can be 'learnt' automatically. Thus, parameters that define the relationship are adjusted to match changes in process characteristics [Guilandoust et al, 1987, 1988; Lant et al, 1993; Montague et al, 1990, 1992; Tham, 1989, 1991a]. Alternatively, the inferential estimators may also be designed based upon the use of a neural network model. As shown in Fig. 7, estimates of the quality which are generated at the measurement frequency of the secondary variables may then be used for process monitoring or control purposes. [more]

Figure 7. Schematic of Feedback Control using Inferential Estimator

5.2. Data Conditioning and Validation

Even if appropriate instruments exist, the data may not be of sufficient quality for desired goals to be achieved. Signals from plant are often corrupted by noise of varying magnitudes. All control methods are data driven. If appropriate measures are not taken to condition and validate the measured signals, then even the most sophisticated scheme will fail. In other words, the adage 'rubbish in, rubbish out' applies in the field of control.

In safety critical systems, such as the control and monitoring of nuclear reactors and power generators, steps are taken to ensure that the 'correct' signals are used for decision making. In these cases, it is common for both software and hardware redundancy schemes to be implemented. Redundancy is provided for by configuring software or hardware modules in duplicate or triplicate. Voting systems are then employed to validate output signals, retaining only those that are considered to be correct [Warwick and Tham, 1988, 1990].

In less critical applications, duplex or triplex redundancy configurations are not cost effective. Therefore, unless there is absolute need, the smoothing of noisy signals is accomplished via hardware or software filtering to attenuate noise in measured signals. However, a penalty is incurred if the signal is subject to spikes. To remove these spikes or rogue points, heavy filtering has to be applied whereupon significant time-lags may be introduced into the filtered signal. Time-lags in the filtered signal may however be reduced by employing 'logic' filters which combine conventional filter algorithms with SPC concepts to validate and condition process measurements. This integrated approach has been shown to be very effective [Parr and Tham, 1992; Tham and Parr, 1994]

5.3. Data Analysis

Even if 'clean' data is available, there may be many variables associated with a particular process unit. The specification of an appropriate control strategy and controller design become complicated. Which variable should be manipulated to control another? What is the effect of this choice of manipulated-input controlled-output pairings? These are some of the important questions that have to be answered before a candidate configuration can be applied. Indeed, the results dictate the kind of models to employ for controller design and hence final controller types, and the overall control strategy that should be implemented. Inappropriate choice of input-output pairs exacerbate the problem of loop interactions. If interactions are significant, then a multivariable control design is necessary. If the input-output relations indicate nonlinear behaviour, then nonlinear controllers may have to be applied.

Many techniques can be used to tackle these issues. They range from simple graphical techniques (scatter plots, Box-plots), statistical multivariate analysis (e.g. Principal Component Analysis, Correlation Analysis, Cluster Analysis) to control theoretic, relative gain and singular value analyses. The latter are used to investigate control loop interactions and robustness of control strategies. However, there is currently no all embracing procedure for a systematic analysis of data right through to determining the suitability of the final control scheme.