Near-service condition testing of precision bearings: an optimized WDCNN-LSTM model for shallow faults and multi-speed diagnostics

In the field of precision manufacturing, fault diagnosis of bearings under near-service conditions faces critical challenges in detecting shallow defects and adapting to variable-speed operations. To address these challenges, this study constructs a novel vibration dataset and proposes an optimized WDCNN-LSTM framework to realize fault diagnosis. Specifically, a dedicated test platform is developed to collect vibration signals under 11 speed conditions with artificially induced shallow faults, simulating real-world machine tool environments. Building upon this dataset, a CNN-LSTM fault diagnosis model incorporating wide first-layer kernel and batch normalization is proposed to effectively capture low-frequency impact waveforms of shallow faults while maintaining diagnostic stability across speed variations. Ablation experiments were conducted under varying training sample sizes and operational conditions. Results demonstrate that with 60 samples, the WDCNN-LSTM model achieves fault diagnosis accuracy of 97.33%, 7.62%, 35.95%, 26.29%, and 42.69% higher than WDCNN, CNN, CNN-LSTM, and LSTM, respectively. Additionally, WDCNN-LSTM maintains high diagnostic accuracy under variable speed conditions. Comparative experiments further validated the superior performance of the proposed model. This research advances the precision bearing condition monitoring by simultaneously addressing shallow defect detection and multi-speed conditions, offering a robust solution for real-time equipment health management in smart manufacturing systems.