Extracting robust fault sensitive features of vibration signals remains a challenge for rotating machinery fault diagnosis under variable operating conditions. Most existing fault diagnosis methods based on the convolutional neural network (CNN) can only extract single-scale features, which not only loss fault sensitive information on other scales, but also suffer from the domain shift problem. In this work, a novel end-to-end deep learning network named adaptive weighted multiscale convolutional neural network (AWMSCNN) is proposed to adaptively extract robust and discriminative multiscale fusion features from raw vibration signals. The AWMSCNN consists of three main components: the denoising layer, the adaptive weighted multiscale convolutional (AWMSC) block, and the multiscale feature fusion layer. The AWMSC block can learn rich and complementary features on multiple scales in parallel. Then, an adaptive weight vector is introduced to modulate multiscale features to emphasize fault sensitive features and suppress features that are sensitive to operating conditions. The train wheelset bearing dataset and the bearing dataset provided by Case Western Reserve University (CWRU) are used to verify the superiority of the proposed model over the basic CNN and other multiscale CNN models. The experiment results show that the proposed model has strong fault discriminative ability and domain adaptive ability against variable operating conditions.
Existing deep learning-based fault diagnosis models for wind turbine planetary gearboxes cannot cope with the problem of poor domain generalization ability caused by differences in data distribution under different operating conditions. To address this limitation, a domain generalization fault diagnosis method based on multi-level feature alignment (DGFD-MLFA) for planetary gearbox is proposed. Firstly, the multi-level feature extraction (MLFE) module is designed, which includes a convolutional denoising autoencoder, a multi-scale separable convolution, and a transformer encoder. The MLFE module can extract both local and global features while denoising. Secondly, the multi-level feature alignment (MLFA) module includes the Local Maximum Mean Discrepancy (LMMD) and the Maximum Mean Discrepancy (MMD) is designed to align local and global features respectively, allowing the model to ultimately extract domain-invariant features and enhancing its fault diagnosis capability under unknown operating conditions. Experiment results on nine domain generalization fault diagnosis tasks based on the wind turbine planetary gearbox dataset, including ablation studies and comparisons with different feature alignment methods, show that the proposed model significantly improves fault diagnosis accuracy under both interpolation and extrapolation domain generalization tasks.