Teacher–Student Uncertainty Autoencoder for the Process-Relevant and Quality-Relevant Fault Detection in the Industrial Process

Abstract

Fault detection plays an important role in process monitoring, while current fault detection methods only concentrate on process-relevant or quality-relevant faults. Therefore, in this paper, a fault detection method based on the improved teacher-student network is proposed, in which both the process-relevant and quality-relevant faults are monitored. Concretely, the student network extracts representation features and the teacher network detects faults. As the features difference between the teacher and student networks can cause performance degradation when the features of teacher network are replaced by the one from student network, representation evaluation block (REB) is proposed to evaluate and reduce the features difference. As a concrete method of REB, uncertainty modeling, is proposed to quantify the features difference and alleviate the aleatoric uncertainty, modeling features difference as a central isotropic Gaussian distribution. Then, asynchronous iterative is designed to implement teacher network and student network joint training. Accordingly, REB based on uncertainty modeling is applied in the teacher-student network named as teacher-student uncertainty autoencoder (TSUAE). Finally, a fault detection framework based on TSUAE is proposed, the effectiveness of which is verified in a wastewater treatment process. <i>Impact Statement-</i>Fault detection is a crucial technology to ensure the normal operation of industrial processes. Both quality-relevant and process-relevant faults should be concerned. The proposed fault detection framework can monitor both the quality-relevant and process-relevant subspaces based on the improved teacher-student network. To improve the accuracy of fault detection and reduce the features difference between the teacher and student networks that can cause performance degradation of the student network, the method quantifies the features difference through uncertainty modeling, reduces the features difference with representation evaluation block and asynchronous iterative training. After that, the proposed method achieves higher detection rates than traditional methods in the wastewater treatment process. Additionally, the proposed method can more efficiently detect not only the largely quality-relevant faults but also the incipient quality-relevant fault in most situations, which is significant for actual industrial processes.