Barycenter-Based Optimal Transport Neighborhood Learning Network for Multisource Cross-Machine Bearing Fault Diagnosis

In recent years, leveraging data from related machines to learn fault knowledge has emerged as a promising approach to address faulty data scarcity in bearing fault diagnosis, i.e. multi-source cross-machine bearing fault diagnosis. However, variations in machine structure, operating condition, and measurement introduce huge domain discrepancies, often leading to significant performance degradation in diagnostic models. To address this challenge, this article proposes a signal-processing module that leverages prior diagnostic knowledge to mitigate the impact of these discrepancies. Furthermore, we propose the barycenter-based optimal transport neighborhood learning network (BOTNLN). Within BOTNLN, Wasserstein barycenter-based partial optimal transport (POT) is employed to integrate knowledge from multiple source domains, ensuring effective diagnostic knowledge transfer to the target domain. Additionally, self-adaptive neighborhood learning is incorporated to minimize intra-class variations, thereby improving inter-class distinguishability and clustering compactness. Extensive experiments conducted on four benchmark-bearing fault diagnosis datasets demonstrate that BOTNLN significantly enhances multi-source cross-machine fault diagnosis performance, consistently outperforming state-of-the-art (SOTA) methods.