Fault diagnosis and analysis of industrial bearings based on enhanced attention multilayer bidirectional LSTM

Abstract

Industrial bearing fault diagnosis is crucial for equipment safety and maintenance efficiency. Conventional methods have difficulty extracting weak fault features under strong noise or variable loads, especially for incipient faults. To solve this, we propose an enhanced attention-based multi-layer bidirectional LSTM (MLB-LSTM) model with two modules: (1) a Multi-Layer Bidirectional fusion (MLB) module that fuses time-series features across LSTM layers, reducing information redundancy or loss; (2) an Adaptive Wavelet Transform (AWT) module that optimizes wavelet parameters for multi-scale frequency-domain representation of non-stationary vibration signals, complementing temporal features for weak-fault detection. Experiments show that adding AWT raises the F1-score for 0.007- inch artificial defects to 98.5% and maintains 93.8% classification accuracy at SNR=5dB. The model has better adaptability to complex industrial conditions. This work promotes intelligent early fault diagnosis for bearings in real industrial settings and expands deep learning applications in time-series signal analysis.