Semi-Supervised Fault Classification Based on Pseudo-Label Learning in Chemical Processes

In recent years, deep learning methods have been widely applied to fault classification in chemical processes, but previous studies have encountered significant challenges. Firstly, actual chemical processes often have only a small number of labeled samples. Effectively utilizing these limited labeled samples along with a large amount of unlabeled data to train models, avoid loss of fault information, and improve the accuracy of fault classification is challenging. Secondly, performing fault classification on chemical process data, which is nonlinear, multivariate, and strongly time-correlated, is of great importance. To address these issues, a semi-supervised fault classification method based on pseudo-labels is proposed. Within the framework of pseudo-label learning, C-GRU is used to extract the spatiotemporal features of fault information. The model dynamically adjusts the label assignments based on the learning time of fault types, improving sample utilization, uncovering fault information, and enhancing fault diagnosis performance. The effectiveness of the proposed method is demonstrated through numerical simulation processes and the Tennessee Eastman (TE) process.