Performance Monitoring For Model Predictive Controller Based On Data-driven Method

The process industry is the pillar of China's economic industry.The safety of process control systems and the requirements of product standards are extremely strict.Moreover,the industrial process becomes larger and more complexity,so that the performance monitoring of the control system become more difficulty.Therefore,it is significance to find timely and accurate evaluation and diagnosis method of controller performance for ensuring the safety of industry process and enterprise efficiency.In this paper,the slow feature analysis method is introduced at first,and the comprehensive performance indexes are designed to evaluate the controller performance to solve the problem that the current MPC controller performance evaluation methods haven't monitored the steady state characteristics and the dynamic characteristics separately.In order to improve the accuracy of MPC controller performance diagnosis,a controller performance diagnosis method based on Stacked-SAE(Stacked Sparse Auto Encoder)classifier for performance diagnosis is proposed.Aiming at solving the problem that the performance diagnosis accuracy of deep learning network classifier is susceptible to unbalanced samples,a performance diagnosis method based on CSDL-KCRC is proposed for controller performance diagnosis.The main work of this paper is summarized as follows:In view of the current problem of predictive controller performance evaluation methods,the steady-state characteristics and dynamic characteristics are not effectively stripped,resulting in inaccurate evaluation results.The slow feature analysis method is applied to the performance evaluation of the MPC controller.In this paper,slow feature analysis method gives four evaluation indicators for monitoring steady-state and dynamic separately.The operators can distinguish controllable process changes(such as changes in output set-points)and truly uncontrollable controller performance degradation correctly.The comprehensive performance assessment indexes are proposed in this paper to improve the monitoring efficiency and accuracy.Simulation studies show that this method can identify the deterioration of controller performance correctly,reducing the false alarm rate and the leakage alarm rate,and it is expected to greatly reduce the non-production time caused by misjudgment.Most process industry predictive control systems are large-scale safety demanding system.The deep learning network classification model is applied to the performance diagnosis,to improve the accuracy of the controller performance diagnosis result.In the diagnosis of predictive controller performance,the existing shallow classifier and multivariate statistical diagnosis methods can not extract the deep fault feature information of sample data,so that the performance diagnosis method based on Stacked-SAE classifier is proposed in this paper.Simulation studies show that the MPC controller performance diagnosis method based on Stacked-SAE classifier can get higher accuracy.In order to solve the problem that the method based on deep learning network classifier can be affected by the unbalanced data samples,the performance diagnosis on the prediction controller based on CSDL-KCRC method is proposed in this paper.Combining the collaborative representation classifier,dictionary learning and kernel method to apply to the MPC controller performance diagnosis,to improve the accuracy of performance diagnosis.Simulation studies show that the performance diagnosis based on CSDL-KCRC method can accurately diagnose the controller performance degradation and have higher diagnostic accuracy under the condition of unbalance sample data.