Multi-scale boundary-enhanced adversarial domain adaptation network: A unified information-aware framework for cross-domain fault diagnosis

Abstract

Effective fault diagnosis across different operating conditions is hindered by domain shift, and class imbalance. Existing adversarial domain adaptation methods often struggle with three key issues: insufficient feature learning from multi-resolution signals, bias induced by imbalanced data distributions, and unreliable predictions for samples near decision boundaries. To overcome these limitations, this paper presents a multi-scale and boundary-enhanced adversarial domain adaptation network (MSBE-ADAN). The proposed framework systematically mitigates these limitations through three synergistically integrated components: a multi-scale hierarchical feature extractor that captures robust patterns across frequency bands, a domain-aware focal loss that dynamically prioritizes challenging and transferable samples, and a boundary refinement module that enhances prediction consistency for uncertain target samples. Comprehensive evaluation on the bearing and gearbox dataset demonstrates that MSBE-ADAN outperforms all baseline methods by approximately 10% across cross-condition tasks.