A Multi-Source Domain Adaptation Method for Fault Diagnosis of Train Axle Box Bearings Under Limited Samples

Transfer learning based on domain adaption demonstrates powerful capabilities in intelligent fault diagnosis. It enables diagnosis networks to utilize data from various sources, potentially improving the diagnostic performance of networks in practical engineering scenarios where measurement samples are limited. Nevertheless, current research faces several constraints: (1) They lack effective alignment methods to extract invariant representations of the source domain and the target domain. (2) They align distributions without taking into account the domain-specific decision boundaries that separate classes. To address the aforementioned limitations, a novel multi-source unsupervised domain adaptation framework is introduced for fault diagnosis of train axle box bearings. Firstly, it aligns the domain-specific distribution for each source-target domain pair. Then, utilizing the domain-specific decision boundaries, it aligns the classifiers for bearing fault states. Finally, we carried out two experiments to collect data, including train axle box bearing faults simulation and bearing faults finite element analysis. The proposed framework's effectiveness and superiority are validated through ablation studies and comparative method experiments.