Multi-source imbalanced generalization wavelet network for heterogeneous imbalanced mechanical fault diagnosis under...

In recent years, fault diagnosis of rotating machinery under unknown working conditions has been investigated. However, in real industrial scenarios, due to the complexity of equipment and the diversity of operating conditions, problems such as class imbalance and domain shift frequently occur under unknown working conditions. In addition, there are interferences from non-stationary time-varying rotational speed conditions and insufficient revelation of fault characteristics, which severely limit the cross-domain generalization performance of models. Therefore, this paper proposes a Multi-Source Imbalanced Generalization Wavelet Network (MS-IGWN). Firstly, a wavelet-enhanced residual network is constructed as the feature extractor. A db4-initialized learnable wavelet filter bank is embedded in the residual network to obtain complementary frequency-selective representations and learn multi-scale discriminative features, and the residual network is combined to learn deep discriminative features. Then, Multi-Head Self-Attention (MHSA) and Multi-order Gated Aggregation Block (MogaBlock) are introduced to strengthen the spatiotemporal correlation of features and the focusing of key information. Secondly, a Semantic-Consistent Mixup (SCMix) mechanism is introduced to enrich minority-class representations by interpolating same-label samples from different source domains. In addition, Cross-Domain Invariant Triplet Loss (CITL) promotes cross-domain intra-class compactness and inter-class separation. Performance tests of MS-IGWN are conducted on the University of Ottawa (Ottawa) bearing dataset and the Mechanical Diagnosis Gear dataset (LJS). Imbalanced domain generalization experiments under unknown time-varying rotational speed conditions show that the proposed MS-IGWN method has better clustering performance, higher classification accuracy, as well as superior generalization and robustness.