One can visualize the response of a controller on a three-dimensional x/y/z plot. As shown in Figure 6, the x axis is the value of the measured error (e); the y axis is the change of the error (Δe) during the last sample period; and on the z axis gives the corresponding output (y), which the particular controller generates.
Artificial Neural Networks (ANN)
Neural networks, fuzzy logic and statistical process control are all “model free” or “black box” methods of control. They can be developed for processes, which are not defined theoretically and cannot be described by a mathematical model. The major difference between fuzzy logic and neural networks is that ANN is trained mostly by data and less so or not at all by reasoning. The fuzzy logic model is superior in this respect, because not only the gain (importance) of its inputs can be modified, but their functions can also be adjusted.
The main limitations of all model free expert systems is their long learning period, (which can be compared to the growing up of a child) and the fact that their knowledge is based only on past events. Consequently, they are less prepared to handle new situations, less suited for anticipation and therefore, if the process changes, they require re-training.
Figure 7 below shows a three-layer ANN model that on the basis of the detection of a number of measurable variables, predicts parameters that are not so easily measured, such as the boiling point of the distillate or the Reid vapor pressure of the bottoms product of a column. Such predictive ANN models can be valuable because they eliminate the need for analyzers (which can be unreliable or high-maintenance), and also minimize the dead time in obtaining the measurement.
||FIGURE 7: THREE-LAYER ANN MODEL|
Three-layer ANN model used to predict the composition ( quality) of overhead and bottom products in a distillation process.
The process model’s “knowledge” is stored in the ANN by the way the processing elements (nodes) are connected to each other and by the importance assigned to each node (weight). The ANN model is “trained” by example and, therefore, it contains an adaptive mechanism for learning from examples and to adjust its parameters based on the knowledge gained through adaptation. During the “training” of these networks, the connections and weights are adjusted until the output of the ANN model matches that of the real process. The hidden layers of nodes between input and output layers help the network to generalize and even to memorize past performance.
It should be noted that the experience of the process control engineers in the plant being controlled is even more important when using ANN-based control than it was when PID control is used. As shown in Figure 8, the artificial neural networks can be used to construct plant models and their inverses, within an internal model control (IMC) structure. The controller in effect is the inverse model of the plant. The ANN-based control system determines the difference (error em) between the process output (y) and the ANN model output and, after filtering that error to remove noise, it is returned as the feedback to the ANN controller. This reduces the uncertainty of the model. The difference between the outputs of the actual plant and the internal model can be caused not only by the model uncertainty, but also by process disturbances. Therefore, internal model control can also reduce the disturbance effect on the system output.
AI, ANN and FL Applications
|FIGURE 8: APPLICATION OF ANNs IN THE INTERNAL MODEL CONTROL STRUCTURE
|A variety of ANN control algorithm types (*10), are available to control nonlinear processes.
Artificial intelligence has been used in controlling such nonlinear processes as the production of pulp and paper, cement, detergent powders or the control of electric kilns. ANN applications
also include nonindustrial processes, such as oceanography, meteorology or the design of airfoil shapes. Successful industrial applications
have also been reported in controlling the hot rolling of steel in strip-steel production.
The fusion of neural networks and fuzzy logic in the form of neuro-fuzzy techniques is seen by many as the most promising way ahead for advanced process control applications. Fuzzy logic has been used in applications from adjusting air conditioners to controlling rice cookers. FL has also been applied in such consumer products as washing machines and in the measurement of the temperature of molten glass. In one self-verifying temperature detection application, an optical (INEEL) and a thermocouple type (AIC) sensor been combined into a single FL-based sensor that is more reliable than other detectors.
One advantage of FL-based sensors is that they can capture the knowledge of the operators in rule-based fuzzy logic statements. An example of such a statement might be, “I trust the newly installed sensors less than the old ones!” Soft sensors can also detect variables that are not measured, such as flow based on the rate of level change in a tank or on the basis of the opening of a control valve. Soft sensors are discussed in more detail chapter 7.18 in the first volume of my Instrument Engineer’s Handbook.