A novel spatio‐temporal feature fusion‐based fault detection method for complex industrial processes

Abstract As modern industrial processes develop towards greater integration and intelligence, process monitoring is also becoming increasingly complex. The high‐dimensional, nonlinear time‐series data in industrial processes pose significant challenges for effective monitoring. In order to improve the performance of process monitoring, we propose a novel fault detection model based on recurrent neural network–convolutional auto encoding–Gaussian mixture model (RNN‐CAE‐GMM). Firstly, the temporal characteristics of the data are extracted using a recurrent neural network based on autoencoder while convolutional autoencoder captures its spatial characteristics. Secondly, the Gaussian mixture model (GMM) is constructed to model the distribution of the normal data by leveraging the extracted temporal and spatial features through end‐to‐end training. Then, fault detection is achieved by constructing the negative log‐likelihood function monitoring statistic based on the GMM. Finally, the effectiveness of the proposed model for fault detection is validated using datasets from both the Tennessee Eastman process and the hot strip mill process. The results demonstrate that the RNN‐CAE‐GMM model significantly enhances fault detection performance in complex industrial processes.