Abstract: Aiming at the problem of noise interference and working condition differences leading to a significant reduction in the accuracy of rolling bearing fault diagnosis, this study proposes a method for rolling bearing fault diagnosis. By adopting the technical means of hierarchical multi-scale feature extraction and fusion, and exploring the design principles of the multi-scale channel coupling learning, the work advances the design methodology in the multi-scale channel coupling feature extraction module and feature screening separable convolutional residual module design method. This leads to the construction of an Inception attention residual module, culminating in a multi-scale attention residual network for rolling bearing fault diagnosis. Noise experiments and load experiments are carried out on the Case Western Reserve bearing dataset, demonstrating that the proposed method has better anti-noise performance and better generalization performance for different loads. The average diagnostic accuracy of gearbox bearings under different working conditions reaches 92.19%. Furthermore, it shows better generalization performance under different working conditions. The confusion matrix of gearbox bearing diagnosis indicates a a classification accuracy of 98% across five fault states, each with 200 samples, underscoring the high precision of the proposed approach.

Key words: fault diagnosis, rolling bearings, residual networks, channel coupled learning, attention mechanisms