Data-driven Approach For MPC Performance Monitoring Under Complex Environment

The controller performance monitoring is a common means to ensure the smooth running of the controller and the whole industrial process. Due to the increasing complexity of industrial process, the traditional data-driven approaches for Model Predictive Control(MPC) performance monitoring can no longer meet the current demand. This paper focuses on the MPC performance monitoring under multi-operating conditions and multi-causes, considers the characteristics of the data, and proposes an effective and feasible data-driven method for performance monitoring.In the process of performance monitoring under multi-operating conditions, the performance benchmark of the controller varies when the operating condition changes. Therefore, it is important to determine the operating condition where the real-time data works and to determine the performance benchmark for performance assessment. This paper proposes an overall index, which is combined with the improved Principle Component Analysis(PCA) similarity factor and the Bhattacharyya distance similarity factor. This index takes into account the change of data in both PCA space and in distance, and compares the similarity between the historical data and real-time data, so that it can get the operating condition of real-time data.In the performance monitoring of multi-causes, which means there are several causes leading to MPC performance degradation, it is not sufficient to consider only one cause to solve the problem of performance monitoring. The influence of multi-causes makes the characteristics of input/output data of MPC system similar to each other in different situations, which increases the difficulty of performance monitoring. In this paper, a method based on Bhattacharyya distance similarity factor is proposed to assess the performance in the process of MPC performance monitoring under multi-causes. The method, using this similarity factor as the performance index, takes into account the variance of the change of data. By comparing the distance and the variance of real-time data and historical data, the fluctuation of I/O data is enlarged, so that the accuracy of performance assessment is improved. On the basis of performance assessment, for the problem of MPC performance diagnosis with multi-causes, a classification algorithm based on Twin Support Vector Machine(TWSVM) and Directed Acyclic Graph(DAG) is presented. The algorithm considers the impact of multi-causes, decomposes the steps of performance diagnosis, so as to achieve the purpose of diagnose all the multi-causes.At last, a software platform for the simulation of data-driven approach for MPC performance monitoring is established under MATLAB and GUI to promote the above methods. The software platform includes the following parts: data import module, operating condition judgment module, performance assessment module and performance diagnosis module. By importing the corresponding historical data and real-time data, the performance monitoring under multi-operating conditions and multi-causes can be done in this software platform.