Big data-driven fault diagnosis: A hierarchical subdomain adaptive network for cross-domain industrial fault diagnosis

Ensuring the safe and stable operation of industrial systems depends crucially on fault diagnosis models with strong generalization capabilities. However, there are significant differences in the distribution of data under different working conditions or equipment, which poses a fundamental challenge. Factors such as fluctuations in operating conditions result in models being ineffective for cross domain fault diagnosis, thereby limiting the practical value of intelligent diagnostic methods in engineering equipment. In addition, effectively integrating complementary but redundant multi-channel data to extract discriminative features remains a key technical obstacle. To solve these issues, this paper proposes a Boundary-Aware Hierarchical Subdomain Adaptation Network (BA-HSAN) for cross domain industrial fault diagnosis. Firstly, multi-channel vibration signals converted into greyscale images and use spatiotemporal attention mechanism next to deep convolutional layers to mine high-dimensional features, achieving robust multi-source data fusion. Secondly, a dual-classifier adversarial adaptation module is introduced to detect and reduce decision-boundary inconsistency between the source and target domains. Finally, LMMD-based subdomain alignment is applied to the outputs of the last two deep feature layers, enabling progressively improved class-level alignment of semantically similar samples across domains. Comprehensive experiments on multiple cross domain fault diagnosis tasks have shown that the proposed method consistently outperforms state-of-the-art methods in terms of diagnostic accuracy and robustness. This work provides a scalable framework for intelligent fault diagnosis in industrial big data environments.