WaveFA-Net: Bearing Fault Diagnosis Model Based on Multi-Domain Blind Deconvolution and Frequency Domain Attention Mechanism

Bearings are critical to mechanical systems, impacting reliability, safety, and efficiency. Traditional fault detection methods relying on vibration signal sensors face challenges with multi-scale, nonlinear, and noisy data, affecting performance and accuracy. This paper presents WaveFA-Net, a deep neural network integrating multi-resolution decomposition, frequency attention, multi-domain blind deconvolution, and attention mechanisms to improve diagnostic accuracy. The process begins with applying a discrete wavelet transform (DWT) for multi-resolution decomposition, extracting and refining key features. Next, frequency attention and multi-domain blind deconvolution are employed to isolate critical components. The multi-head and multi-scale attention mechanisms are then incorporated into the network ’ s backbone and neck modules to capture global dependencies and multi-scale information. Finally, a physically inspired loss function, data augmentation, and hyperparameter optimisation are applied to reinforce robustness and stability. Experimental results demonstrate that WaveFA-Net performs well across varying signal-to-noise ratios (SNR), proving its effectiveness in complex environments.