Improved Wiener Filter Network for Interpretable Gear Fault Diagnosis

To address the issues of noise interference in complex signals and poor interpretability of recognition models in practical fault diagnosis using deep learning methods, this study proposes an interpretable gear fault diagnosis method based on an improved wiener filter network. The model consists of three layers: the mode filtering layer, the feature pooling layer, and the classifier. Firstly, the data samples along with rotational speed information are input into the model. Multiple improved wiener filter kernels (IWFKs) are constructed to adaptively extract fault modes, where the central frequency and bandwidth of each filter kernel are mapped with rotational speed using polynomial relationships. Then, the feature pooling layer calculates frequency domain feature indicators of the modes. Finally, the classifier, composed of two fully connected layers, makes the final decision based on the frequency domain feature indicators. This model integrates filtering and feature extraction from signal processing into the deep learning model, making the diagnostic process interpretable. Experimental results demonstrate that the proposed method achieves high recognition accuracy for gear fault diagnosis under different rotational speeds.