A method for fault diagnosis of bearings based on a multi-scale attention fusion network using dual-branch architecture

To solve the problems of diagnosing bearing rolling faults – including extracting features from a single scale, high computational complexity in modeling long-range temporal dependencies and insufficient fusion of multimodal features – this study proposes a fault diagnosis method based on the Multi-Scale Attention and Fusion Network (MSAFN). This approach uses a two-branch architecture for joint processing of time-frequency maps and statistical vector features. First, a dynamic multi-scale convolution module is developed that uses parallel multidimensional convolution kernels for adaptive reception field tuning, capturing fault features at different levels of granularity. Second, a residual dual attention mechanism and a Selective State Space Module (SSM) are introduced. This improves the perception of critical components in the frequency domain while achieving effective modeling of long-range temporal dependencies in vibration signals with linear computational complexity. Based on this, a hierarchical progressive fusion strategy is proposed. This strategy utilizes cross-modal control mechanisms to ensure deep interaction between heterogeneous features at different levels of abstraction. Combined with a multi-task collaborative learning framework, it jointly optimizes classification, reconstruction, and contrastive learning tasks to improve the generalized robustness of the model. Experimental results on the Case Western Reserve University (CWRU) and Paderborn University (PU) datasets demonstrate diagnostic accuracies of 99.67 and 96.0 %, respectively, outperforming baseline deep learning models and exhibiting high interpretability.

Authors: V. V. Potekhin, Ch. Guo

Direction: Informatics, Computer Technologies And Control

Keywords: rolling bearings, fault diagnosis, multi-scale attention, multimodal fusion, multitasking learning, time-frequency domain analysis

---

View full article