Rolling Bearing Fault Diagnosis with Noise-Resistant Multi-Dimensional Integrated Convolutional Residual Networks

A novel three-dimensional six-branch optimised convolutional neural network (MS-MD-CNN) is proposed to address the multimodal characteristics in rolling bearing input signals. The model combines a residual network (ResNet) and a self-attention mechanism, aiming to comprehensively capture different forms of input signals and thus improve the accuracy of rolling bearing fault diagnosis. In the data preprocessing stage, data enhancement is performed by introducing Gaussian noise and randomisation methods to simulate the real environment, increase the diversity of the data, and verify the model's ability to resist noise. Meanwhile, the multi-branch structure of MS-MD-CNN effectively mines the feature information of multimodal signals. To solve the problem of information transfer and gradient vanishing, the model introduces residual connectivity. Finally, the single-head self-attention mechanism is designed for adaptive learning of the relationship between features to enhance the model's feature extraction ability and suppression of noise in the classification task. The model is evaluated using the Western Reserve University dataset, and the experimental results show some practical value.