Adaptive Gaussian Mixture Model-based Variational Autoencoder Network for Process Fault Isolation in Industrial Processes

Modern industrial data is commonly collected under various working conditions, which are usually nonlinear and multi-modal. This poses great challenges to the accurate isolation of fault variables in abnormal situations. Despite the wide exploration of fault detection methods, studies on fault variable isolation are limited due to the complex correlations between process variables. To cope with the nonlinear and multimodal characteristics of process data, an adaptive Gaussian Mixture Model-based variational autoencoder network is proposed for process fault isolation in industrial processes. Gaussian mixture distribution is introduced into variational autoencoder to fit the complex data distribution in multimode industrial processes. An adaptive mechanism is developed to dynamically adjust the number of Gaussian components according to the input data. A combined monitoring index is designed for fault detection via the learned features and residual space. Once the fault is detected, key variables related to the fault are identified by measuring the contribution of each variable to the reconstruction discrepancy. The benchmark Tennessee Eastman (TE) process is utilized to demonstrate the effectiveness of the proposed method.