Identification of unknown faults in chemical processes using few-shot learning

Abstract

This study investigated the application of deep learning for fault diagnosis of chemical processes involving unknown classification ability. A few-shot learning-based unknown recognition and classification (FSLB-UR&C) was presented to avoid the requirement of big data in training and to deal with the problem with small sample training data using the Tennessee Eastman Process (TEP). The model uses a gated recurrent unit with a dot-product attention layer to extract features and evaluate the feature similarity score with support and query data as inputs. Semi-sequential and random sampling methods were applied to enhance the training efficiency. The results showed that the FSLB-UR&C could distinguish known and unknown labels. Furthermore, t-SNE visualization proved that the model could gather unknown faults as a clustering, whereas known faults were clustered separately from the unknown faults. Compared with the other method, FSLB-UR&C can maintain a higher accuracy of the known classification to distinguish unknown faults.