Toward understandable semi-supervised learning fault diagnosis of chemical processes based on long short-term memory ladder autoencoder (LSTM-LAE) and self-attention (SA)

Abstract

Fault diagnosis and localization play vital role in chemical process monitoring. Finding the root cause of fault accurately and timely is the key to avoid serious accidents and ensure process safety. Unfortunately, most deep learning-based models only involves in predicting fault state and interpretability is still not fully explored. Besides, lack of labeled samples in practical situations makes supervised learning difficult to implement. To circumvent the obstacle, semi-supervised learning based on long short-term memory ladder autoencoder is combined with self-attention mechanism, which is intended to establish interpretable model by explicitly clarifying the corresponding relationship between abnormal variables and faults. Using Tennessee Eastman process and practical high-purity carbonate production process as benchmark, fault diagnosis and identification performance of proposed LSTM-LAE-SA is validated, and interpretability analysis is performed to demonstrate its capabilities in abnormal variable locations. The developed understandable model could improve operator's trust and industrial application in fault diagnosis system.