A fault diagnosis method for rotating machinery in nuclear power plants based on long short-term memory and temporal convolutional networks

Vibration signals typically used for health monitoring of rotating machinery has highly integrated spatio-temporal correlations. However existing studies rarely explore the impact of spatial correlation features of rotating machinery internal components on their vibration signals. To identify the health condition of rotating machinery in NPPs in terms of spatial–temporal correlation, we propose a fault diagnosis method with combination of the long short-term memory and temporal convolutional networks. The spatial and temporal features in the vibration signals of rotating machinery are extracted using the two networks and then fused to diagnose its faults. The model was assessed against the Case Western Reserve University bearing dataset, University of Ottawa bearing dataset and Southeast University gearbox dataset. The results show that its diagnostic accuracy reaches up to 99.56 %, 100 %, and 100 % on the three datasets, respectively, and outperforms other five well-designed comparative models, demonstrating its effectiveness and superiority in rotating machinery fault diagnosis.