ADAPTIVE NEURO-FUZZY CONTROL OF PH PROCESSES

Abstract

Control of pH has been identified as a challenging problem and has been often used as the benchmark for evaluating advanced control techniques. The difficulty in pH control arises from the fact that the process dynamics is highly nonlinear and time-varying. This problem is particularly magnified in the range of pH 6-8 (neutralization region). Among various control strategies available, fuzzy logic control was observed to possess the potential to produce good results for pH. In this study fuzzy logic is applied to pH control and, due to the nonlinearity and the changing dynamics of the pH process, an adaptation using artificial neural network is included. The parameters of a fuzzy logic controller include scaling factors and membership functions for error, change in error and change in manipulated variable and the rule table. All of these play an integral role in determining the prescribed control action. Apart from tuning these parameters, adaptation of any one or more of these can be incorporated to design a suitable adaptive fuzzy controller. The scaling factors for change in error and change in manipulated variable are chosen here for adaptation. FlexControl, a Computer-Aided-Engineering platform for fuzzy logic control, was modified to produce the adaptive neuro-fuzzy controller. A 7x7 rule table with bell-shaped membership functions for the error, change in error and change in manipulated variable was first designed. Reference experiments were conducted at different pH setpoints to determine the scaling factors for change in error and change in manipulated variable. The resulting values were then used for off-line training of a radial basis function neural network. The trained network was then employed for online adaptation of the above mentioned fuzzy control parameters. The new control scheme was evaluated on a real-time, laboratory scale continuous stirred tank reactor where an acid-base neutralization was carried out by mixing an acid and a base stream. The effluent pH was measured and fed back to the proposed fuzzy controller. Base flow was used as the manipulated variable. Extensive experimentation was conducted involving disturbances in flow, concentration and composition of the influent acid and in setpoint. The responses were compared with those obtained using an adaptive internal model control and ordinary (non-adaptive) fuzzy control techniques. The performance of the adaptive neuro-fuzzy controller was found to be superior to that of the other two techniques in terms of the settling time and maximum deviation from setpoint. In other words, performance of the ordinary fuzzy controller was enhanced with the introduction of adaptation and the resulting model-free adaptive neuro-fuzzy controller was also found to be significantly better than the model-based adaptive IMC structure.