A new algorithm has been developed for controlling the substrate concentration at a constant set point or a predetermined time course of set point in a fed-batch fermentor. It is a model-independent self-organizing fuzzy contoller with a backpropagation neural network. The neural network is used to classify process error patterns. The classified error patterns form a basis for rule formation and tuning of the fuzzy controller. Different tuning strategies are used for different patterns of the process error. The tuning operation is accomplished by varying scaling factors for the fuzzy values of the process error, its rate of change, and the controller output.
Through simulation studies with a model system of lysine fermentation, the performance of the self-organizing fuzzy controller was evaluated and compared with that of a best tuned PID controller. An approximately tuned simple fuzzy controller without neural network exihibited significant offsets or oscillations. The overall performance of the self-organizing fuzzy controller was observed to be satisfactory for regulatory and servo problems. The self-organizing fuzzy controller was started with arbitrary values of its tuning parameters. At the begining it showed a poor performance, especially for cases with measurement noises, because of its ill-tuned state. However, its control performance was improved as it became better tuned with time, and shortly became as good as that of the best tuned PID controller. This implies that the developed algorithm has a good auto-tuning capability and thus does not require a priori tuning of the controller parameters, which is a cumbersome and time-consuming task in using the conventional type of PID controller.