A novel fault diagnosis framework empowered by LSTM and attention: A case study on the Tennessee Eastman process

In the era of Industry 4.0, substantial research has been devoted to the field of fault detection and diagnosis (FDD), which plays a critical role in preventive maintenance of large chemical processes. However, the existing studies are primarily focused on few-shot samples of process data and without considering the role of activation functions in temporal diagnostic tasks. In this paper, an end-to-end chemical fault diagnosis framework that combines bidirectional long short-term memory (LSTM) with attention mechanism is proposed. In the preprocessing stage, a special sliding time window function is developed to integrate multivariate samples containing complex temporal information via operation such as subset extraction. Afterwards, the bidirectional LSTM is constructed to address dynamic and temporal relationship on longer series observation, and the attention mechanism is adopted to highlight key fault features by assigning different attention weights. A case application is performed on the enriched Tennessee Eastman process (TEP), which reduces the bias between sample statistics and larger population parameters compared to existing few-shot sample studies. The metric evaluation experiments for six activations show that the model configured with tanh function can achieve the optimal tradeoff in chemical process tasks, providing a strong benchmark for subsequent fault diagnosis research.