Time-Frequency Dual-Stream Fusion Network for Bearing Fault Diagnosis under Variable Speed Condition

Jinrui Wang; Zhu Chen; Huadong Du; Yufeng Xu; Zongzhen Zhang; Baokun Han

doi:10.37965/jdmd.2026.1840

Authors

Jinrui Wang College of Mechanical and Electronic Engineering, Shandong University of Science and Technology, Qingdao 266590, China; https://orcid.org/0009-0009-3264-7873
Zhu Chen College of Mechanical and Electronic Engineering, Shandong University of Science and Technology, Qingdao 266590, China;
Huadong Du AUCMA Company Limited, Qingdao 266000, China;
Yufeng Xu AUCMA Company Limited, Qingdao 266000, China;
Zongzhen Zhang College of Mechanical and Electronic Engineering, Shandong University of Science and Technology, Qingdao 266590, China;
Baokun Han College of Mechanical and Electronic Engineering, Shandong University of Science and Technology, Qingdao 266590, China;

DOI:

https://doi.org/10.37965/jdmd.2026.1840

Abstract

Fault diagnosis in mechanical systems under variable speed condition has long been limited by single-modality signal analysis methods, which often result in incomplete feature characterization and fail to fully reflect the true health status of equipment. Time-domain and frequency-domain signals contain a wealth of redundant and complementary fault information. However, how to efficiently extract and fuse these two types of heterogeneous signals remains a research challenge in the field of rotating machinery fault diagnosis. To address this issue, this paper proposes a rolling bearing fault diagnosis method based on a time-frequency dual-stream fusion network, which relies on time-frequency dual-stream feature fusion technology to achieve comprehensive and robust extraction of bearing fault features. The proposed method constructs a parallel network architecture with time-domain and time-frequency-domain branches. It introduces a bidirectional cross-attention mechanism between the time-domain and time-frequency-domain to achieve bidirectional guidance and interactive enhancement of the two types of features. Furthermore, through a cross-domain attention fusion strategy, it accomplishes the deep integration and complementarity of time-frequency heterogeneous features, thereby strengthening the cross-domain representation capability of fault-sensitive features. This approach not only preserves the transient impact information of time-domain signals but also takes into account the frequency distribution patterns of time-frequency-domain signals. The network also incorporates regularization strategies such as label smoothing and cosine-annealed learning rates, effectively enhancing the model’s generalization ability and diagnostic stability. Finally, the effectiveness of the proposed method was validated through rolling bearing fault diagnosis experiments under variable speed condition.

Conflict of Interest Statement

The authors declare no conflicts of interest.