A wavelet packet transform-based deep feature transfer learning method for bearing fault diagnosis under different working conditions

Deep learning has achieved significant advances in the fault diagnosis of rotating machinery. However, it still suffers many challenges such as various working conditions, large environmental noise interference and insufficient effective data samples. Signal time–frequency analysis and feature transfer learning methods can help solve these problems. Combining wavelet packet transform (WPT) and multi-kernel maximum mean discrepancy (MK-MMD), this paper proposes a novel residual network (ResNet)-based deep transfer diagnosis model for bearing faults. Firstly, this paper devises a distinctive WPT time–frequency feature map (WPT-TFFM) construction method using WPT for time–frequency analysis on nonlinear and non-stationary vibration signals. Then, a modified multi-group parallel ResNet network is structured to extract the depth features of WPT-TFFM for the characteristics of small size and feature dispersion. Then, MK-MMD is further applied to evaluate the distribution difference between the depth features of the source and target domain data. Combining with the classification loss of the sample set with the source domain, the depth features extraction network is optimized to achieve better cross-domain invariance and fault state differentiation capability of the depth features. To evaluate the proposed method, this work conducts comparative experiments on two test rigs under different working loads and speeds. The results reveal that the proposed method offers excellent fault diagnosis and noise prevention capability for working condition transfer tasks.