A new multi-source information domain adaption network based on domain attributes and features transfer for cross-domain fault diagnosis

Compared to the single-source domain adaptation fault diagnosis methods, the multi-source domain adaptation methods can not only take advantage of the rich and diverse diagnostic information of multiple source domains but also draw on the feature alignment of single-source setting to reduce the domain discrepancy. However, forcing the alignment of feature distributions is challenging and may lead to negative transfer. Meanwhile, labeled data are often scarce and difficult to collect in actual production, which can be mitigated by multi-source information, but the diagnostic performance of the model is degraded by large domain differences. To tackle the above issues, a domain attribute and feature transfer network is proposed to model multi-source information domains in a unified deep network and achieve cross-domain fault diagnosis. In the transferable attributes learning section, we adopt an attention mechanism and a domain attribute loss function to extract transferable latent attributes from multi-source information. In the transferable features learning section, we apply the local maximum mean discrepancy metric to adjust the category distribution of single-source information and target domains. Then, intra-class compactness learning and pseudo-labeling learning strategies are utilized to further obtain richer feature representations. Finally, we propose the knowledge fusion module to fuse the results of multi-source information classifiers to yield a more reliable diagnosis result. Extensive experiments on three different multi-source information datasets show the superiority of our method compared to the state-of-the-art (SOTA) methods by comparing indicators from various aspects.