Dynamic Weight-Optimized Prototypical Contrastive Network for Cross-Domain Few-Shot Bearing Fault Diagnosis

Due to the limited sample size caused by preventive maintenance and the variable data distribution caused by environmental factors, the performance of well-trained laboratory models has decreased significantly when confronted with actual industrial bearing fault diagnosis. The existing domain adaptation methods using cross-domain labeled samples for classification make it difficult to resist the influence of outliers on the overall matching, which leads to a negative transfer problem. Given the above issues, a dynamic weight-optimized prototypical contrastive network is proposed. Primarily, domain adaptation is assisted by the discriminative information the classifiers convey during the prediction process. The model is prompted to find the source data that matches the distribution, removing the effect of distribution variability. Furthermore, the sample-aware weighting term evaluates the difficulty of the sample classification and suppresses the performance degeneration of domain adaptation. Subsequently, the in-domain prototypical contrastive learning evaluates dynamically the intensity of feature distribution around each prototype. The features from the same category are inspired to move closer to the prototype to enhance the consistency and discrimination of intra-class features. Meanwhile, cross-domain instance-prototype learning reduces the distribution inconsistency of the corresponding category data in the source and target domains to enable fine-grained inter-domain alignment and mitigate the negative transfer problem. Through numerous comparative experiments, this method shows superior effectiveness and engineering diagnostic feasibility under cross conditions with limited data resources. Note to Practitioners—This article addresses the cross-domain distribution variability and few-shot bearing limitation. Existing research on intelligent bearing fault diagnosis is based on the unattainable assumption that fault samples are sufficient and homogeneously distributed. Moreover, domain-adaptive methods using cross-domain labeled samples for classification are too coarse to resist the effect of the outlier of bearing operational data, leading to the negative transfer problem. In this article, a dynamic weight-optimized prototypical contrastive network is proposed. Primarily, a feature-level cross-domain alignment module pre-trains the feature extractor to generate unique features for disparate bearing fault signals. After pre-training, the model can find source data that matches the distribution, removing the effect of distribution variability. Among it, the sample-aware weighting term, which dynamically weights the samples according to their classification difficulty, prevents the model performance from degrading on the cross-domain task. Then, the prototypical contrastive learning module can construct prototypes, representing typical features of bearing faults. Each category of features is approached with similar prototypes, pushing away from disparate categories of prototypes. Relying on the prototypes to be aligned with disparate domain instances to weed out the effects of outliers. Not only does it achieve cross-domain alignment of underlying features, but it also aligns semantic structures in a shared cross-domain space. Through numerous experiments, it has been proved that the method can not only solve the cross-domain problem and the few-shot problem, but is also more applicable to the fault diagnosis of real industrial systems. Furthermore, its performance is much higher than the existing methods.