Multi-Condition Intelligent Fault Diagnosis Based on Tree-Structured Labels and Hierarchical Multi-Granularity Diagnostic Network

The aim of this study is to improve the cross-condition domain adaptability of bearing fault diagnosis models and their diagnostic performance under previously unknown conditions. Thus, this paper proposes a multi-condition adaptive bearing fault diagnosis method based on multi-granularity data annotation. A tree-structured labeling scheme is introduced to allow for multi-granularity fault annotation. A hierarchical multi-granularity diagnostic network is designed to automatically learn multi-level fault information from condition data using feature extractors of varying granularity, allowing for the extraction of shared fault information across conditions. Additionally, a multi-granularity fault loss function is developed to help the deep network learn tree-structured labels, improving intra-class compactness and reducing hierarchical similarity between classes. Two experimental cases demonstrate that the proposed method exhibits robust cross-condition domain adaptability and performs better in unseen conditions than state-of-the-art methods.