Bayesian-ANN controller for pH-control

Abstract

The major problems using artificial neural networks (ANN) for the controller in process control are overfitting and extrapolation. The former commonly occur when unmeasured disturbances affect the process and the ANN will average out the controller output, between the disturbance case and the case where there are no disturbances. Extrapolation problem stems from the ANN suggesting controller outputs in areas where the ANN has not been trained for. This paper proposes a Bayesian weighting approach to improve the generalization performance of ANN controllers in pH control where a feedforward ANN was trained to mimic the behavior of a Robust Model Predictive Controller (RMPC).  The proposed approach relies on separating the training data based on the presence and size of disturbances, as well as a Bayesian weighting scheme. The training data was generated from running multiple tests on the RMPC for different requirements and cases of pH control. The training algorithm used was the Levenberg-Marquardt algorithm. The proposed approach was applied to a MATLAB® simulated pH control system. Bayesian-ANN is a novel approach for enhancing the generalization performance of ANNs in pH control. It is straightforward to implement and can be utilized with any ANN architecture. The proposed Bayesian weighted ANN (ANN-MPC) method in controlling pH process is found to be superior to a single ANN and multiple model predictive control (MMPC). This is particularly true with regard to reducing overshoots and oscillations.