Few-shot chemical process fault diagnosis based on fused self-supervised contrastive learning

Abstract

In recent years, the field of chemical process fault diagnosis based on deep learning has grown rapidly. Compared with traditional methods, deep learning models are able to learn more complex data patterns and are more suitable for modern complex industrial systems. However, deep learning in the field of chemical process fault diagnosis still faces the challenge of insufficient sample size of chemical fault data. To address the problem of insufficient fault data samples in real chemical processes, this paper proposes a Fusion Self-Supervised Contrastive Learning for Fault Diagnosis (FSSCL). Firstly, this method proposes a self-supervised model for feature recovery and a contrastive learning model for sample classification, which are pre-trained for extracting intra-sample data features and inter-sample data discrepancy features, respectively; then, the trained model is fused using feature fusion technique to stitch and merge the extracted features from the two models to deliver them to the classifier for classification. The experiments on the Coke furnace process and the Tennessee Eastman chemical process show that the FSSCL method can still achieve high fault diagnosis accuracy with a small number of samples, which effectively solves the problem that the traditional fault diagnosis model is difficult to be trained in the face of a few-shot dataset and is easy to be overfitted.