Bearing Frequency-Related Fault Detection in Pump Systems Using Inference LSTM Autoencoder With Differential Regularization

In the industrial sector, technicians typically detect equipment faults based on multi-sensor data (MSD), which often exhibit strong logical correlations. However, existing AI-based methods for MSD fault detection have not fully considered these correlations. To this end, we propose an inference LSTM autoencoder with differential regularization (IDR-LSTMAE) for bearing frequency-related fault detection in pump systems. Our LSTM Autoencoder learns normal MSD behavior patterns, while a novel differential regularization in the loss function ensures the model cannot be disturbed by noise and unlabeled faults, resulting in smoother reconstructions. Additionally, leveraging specialist knowledge, we design an inference function that considers logical relationships between MSD during fault detection. Such integration of the inference function can help improve accuracy and reduce false alarms. Comprehensive comparison experiments and ablation studies show that IDR-LSTMAE achieves superior fault detection performance and highlights the importance of regularization and the inference function.