Dual-Branch Feature Learning and Hierarchical Domain Alignment for Multisource Adaptation Fault Diagnosis

Abstract

To address the significant distribution discrepancies among multiple source domains and the inefficient utilization of heterogeneous source information, a multisource adaptation fault diagnosis framework based on dual-branch feature learning and hierarchical domain alignment is proposed. A shared feature extractor and multiple domain-specific feature extractors are constructed to learn transferable common representations and complementary source-dependent discriminative information, respectively. The two branches are independently parameterized and guided by different optimization objectives, enabling the model to balance cross-domain transferability and source-domain individuality without assuming strict statistical disentanglement between the resulting feature spaces. A two-stage hierarchical domain alignment mechanism is further developed. The first stage performs intersource alignment in the shared feature space to reduce distribution discrepancies among multiple source domains, while the second stage conducts source–target alignment through domain-specific branches to enhance target-oriented feature transfer. In addition, a multiclassifier fusion strategy guided by the maximum-entropy principle is introduced to adaptively aggregate predictions from multiple sources, improving model stability under complex conditions and enhancing the recognition of uncertain samples. Experimental results demonstrate that the proposed method achieves superior diagnostic accuracy in multisource domain adaptation tasks and exhibits strong potential for practical engineering applications.