Intelligent Fault Diagnosis of Bearings in Unsupervised Dynamic Domain Adaptation Networks Under Variable Conditions

Effective fault diagnosis is crucial for ensuring the safe and reliable operation of machinery. Despite satisfactory achievements of deep learning in fault diagnosis, acquiring large labeled datasets remains a challenge in practical industrial scenarios. In addition, machinery often operates under varying load or speed conditions, which leads to diverse distributions of collected samples. Consequently, a model trained on one distribution (source domain) encounters difficulties in adapting to another distribution (target domain). To solve the problem, this paper introduces unsupervised domain adaptation method, and also proposed unsupervised dynamic domain adaptation network (UDDAN) for fault diagnosis in bearings under variable working conditions. Specifically, the multi-scale dilated convolution module is integrated into Squeeze-and-Excitation ResNeXt (SE-ResNeXt) as a feature extractor, which not only allows the model to introduce an attention mechanism, but also enhances its multi-scale information extraction capability. The domain adaptation is achieved by minimizing the difference between the two distributions through maximum mean difference (MMD). Furthermore, the innovatively designed loss function enables the model to dynamically allocate weights between the source and target domain knowledge during the domain adaptation. Extensive experiments are conducted on two datasets, and the proposed method achieves a diagnostic accuracy of up to 99.16% in typical scenarios, while also demonstrating significant robustness in noisy environments.