Fault diagnosis method of rolling bearing based on noise reduction enhanced multi-frequency scale network

Abstract Currently, data-driven deep learning methods have attracted much attention in the field of bearing fault diagnosis. Nonetheless, the existing rolling bearing fault methods suffer from insufficient fault feature extraction capability when dealing with variable operating conditions and strong noise environments. Therefore, this paper proposes a noise reduction enhanced multi-frequency scale network model-bidirectional long short-term memory network based on the collected bearing vibration data source. The noise embedded in the original vibration signals under different working conditions is effectively removed by designing an adaptive threshold noise reduction module. To comprehensively explore fault information within the vibration signals, a combined strategy of ordinary convolution and dilated convolution is proposed to cross-extract signal features across high, medium, and low multi-frequency scales. Simultaneously, a self-attention mechanism mode is integrated into the traditional channel attention mechanism to augment the model’s focus on multichannel and internal features, addressing the challenge of incomplete feature extraction under complex working conditions. Furthermore, the output mechanism is optimized and reacquired to grasp the intrinsic links between the combined fault characteristics. This process enhances the model’s discriminative power for early bearing faults and its generalization ability to accommodate data from diverse working conditions, thereby facilitating accurate diagnosis of bearing faults. Comparison and ablation experiments are conducted on multiple aero-engine rolling bearing datasets, validating the superior noise-resistant diagnostic performance of the method proposed in this paper under complex working conditions, which offers significant advantages compared to other methods.