Multi-Source Domain Adaptation for Fault Diagnosis: A Unified Framework for Feature and Relation Transfer

Transfer learning methods have demonstrated significant success in cross-domain fault diagnosis due to their strong domain adaptation capabilities. However, single-source domain adaptation, as a mainstream framework is commonly bounded by limited diversity and information, making it unsuitable for complex, real-world engineering applications. Therefore, in this paper, we proposed a novel multi-source domain adaptation framework to achieve domain-invariant and discriminative feature and relation transfer, simultaneously. First, we found that diagnostic knowledge extracted from single-source domain data fails to achieve satisfactory cross-domain diagnostic results. To address this, we proposed a novel multi-source domain adaptation framework that synthesizes knowledge across source domains, which not only eliminates negative transfer but also improves model performance. Moreover, we found that the most existing methods only consider features or relations, leading to incomplete fault-based information, which may affect final identification. Therefore, we proposed the graph convolutional network and convolutional neural network to simultaneously capture features and relations, which are essential for bridging multi-source and target domains to extract sufficient fault-based representations. Finally, four types of alignment metrics are integrated into a unified deep network to achieve more effective cross-domain fault diagnosis. Experimental results on four datasets demonstrate that the proposed method not only achieves optimal diagnostic performance compared to state-of-the-art methods but also captures transferable, domain-invariant, and discriminative representations for domain adaptation.