Bearing Fault Diagnosis Based on Improved Multiscale Feature Fusion and CBMA Attention Mechanism

Rolling bearing fault diagnosis is crucial for the safe operation of mechanical equipment. However, vibration signals have the strong nonlinearity and strong noise intrference characteristics. Moreover, existing deep learning methods have shortcomings in feature representation and multi-scale fusion. To solve these problems, the paper proposes a fault diagnosis method based on multi-scale feature fusion and CBMA attention mechanism. This method extracts local features of different time spans through multi-branch convolutional modules. It realizes multi-level feature fusion from bottom to top and from top to bottom by using the feature pyramid network. The method introduces the CBMA attention mechanism to dynamically adjust feature weights in the channel and spatial dimensions, enhancing the expression of fault-sensitive features. The experiments are conducted on the bearing dataset of Case Western Reserve University. The experimental results show that the average accuracy of CBMA-MCNN is <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 9. 3 3 \%}$</tex>, significantly outperforming the CNN, MCNN, and CBMA-CNN models. This method effectively improves the extraction and classification performance of rolling bearing fault features under complex working conditions, providing a new solution for engineering applications.