An adaptive fault detection and root-cause analysis scheme for complex industrial processes using moving window KPCA and information geometric causal inference

• Published in: Journal of intelligent manufacturing Vol. 32; no. 7; pp. 2007 - 2021
• Main Authors: Sun, Yanning, Qin, Wei, Zhuang, Zilong, Xu, Hongwei
• Format: Journal Article
• Language: English
• Published: New York Springer US 01-10-2021; Springer Nature B.V
• Subjects: Artificial intelligence; Blindness; Business and Management; Control; Data analysis; Dynamic characteristics; Evolutionary algorithms; False alarms; Fault detection; Fault diagnosis; Inference; Machines; Manufacturing; Mechatronics; Microprocessors; Model accuracy; Multiple objective analysis; Principal components analysis; Processes; Production; Robotics; Root cause analysis; Causal inference; Root-cause analysis; Moving window KPCA; Fault detection; Multiobjective evolutionary algorithm
• ISSN: 0956-5515; 1572-8145
• DOI: 10.1007/s10845-021-01752-9

In recent years, fault detection and diagnosis for industrial processes have been rapidly developed to minimize costs and maximize efficiency by taking advantages of cheap sensors and microprocessors, data analysis and artificial intelligence methods. However, due to the nonlinear and dynamic characteristics of industrial process data, the accuracy and efficiency of fault detection and diagnosis methods have always been an urgent problem in industry and academia. Therefore, this study proposes an adaptive fault detection and root-cause analysis scheme for complex industrial processes using moving window kernel principle component analysis (KPCA) and information geometric causal inference (IGCI). The proposed scheme has three main contributions. Firstly, a research scheme combining moving window KPCA with adaptive threshold is presented to handle the nonlinear and dynamic characteristics of complex industrial processes. Then, the multiobjective evolutionary algorithm is employed to select the optimal hyperparameters for fault detection, which not only avoids the blindness of hyperparameters selection, but also maximize model accuracy. Finally, the IGCI-based fault root-cause analysis method can help field operators to take corrective measures in time to resume the normal process. The proposed scheme is tested by the Tennessee Eastman platform. Its results show that this scheme has a good performance in reducing the faulty false alarms and missed detection rates and locating fault root-cause.