Distributed process monitoring based on joint mutual information and projective dictionary pair learning

Abstract

In modern industrial processes, each subsystem interacts frequently and involves a large number of process variables with complex relations, which challenge process monitoring. In this paper, a distributed process monitoring method based on joint mutual information (JMI) and projective dictionary pair learning (DPL) is proposed for effective process monitoring in industrial systems with multimode, complex, and high-dimensional data. Firstly, considering the interactive information, redundancy and irrelevance among process variables, an automatic block division method based on JMI is proposed to divide process variables into several low dimensional blocks. Secondly, DPL-based monitoring model is established in each block of each mode. According to the multimode characteristic of industrial processes, a joint probability based on reconstruction error is proposed for mode recognition. Then, Bayesian inference method that fuses block statistics into global statistics is introduced for anomaly detection. The anomaly source is further determined by defining the block contribution coefficient and variable contribution coefficient. Finally, the effectiveness of the proposed method is demonstrated by a numerical simulation, Tennessee Eastman benchmark test, and experiments in an aluminum electrolysis industrial process.