基于多路并行注意力融合网络的轴承故障诊断OA

To address the insufficient extraction of local features,limited multi-scale representation capability,and the tendency for key channel responses to be overwhelmed by noise in fault diagnosis of vibration signals from train axle-box bearings,a multi-branch parallel attention fusion network(MBPAFN)was proposed.In the network,raw one-dimensional(1D)vibration signals were taken as input,and three parallel branches at the raw,average-pooling-downsampled,and max-pooling-downsampled scales were constructed to capture global time-domain patterns,low-frequency trends,and local extrema features,respectively.In each branch,1D convolutional modules were employed to extract local temporal features,and an efficient channel attention mechanism was incorporated to adaptively recalibrate channel weights,while a Transformer encoder was introduced to model long-range temporal dependencies,enabling joint representation of local impact features and global evolution patterns.The three-way features were adaptively fused through a channel alignment and feature fusion module and then fed into a fully connected classifier to output fault types,forming an end-to-end diagnostic pipeline.Experimental results on the publicly available Case Western Reserve University(CWRU)bearing dataset and a self-built train axle-box bearing dataset show that,compared with several representative deep learning diagnostic models,MBPAFN achieves higher recognition accuracy and better robustness under strong noise and non-stationary operating conditions,demonstrating that the multi-branch parallel attention fusion architecture can effectively improve the fault diagnosis performance of rolling bearings.