Fusion innovation: Multi-scale dilated collaborative model of ConvNeXt and MSDA for fault diagnosis

Abstract

Bearings and gears are critical components in modern industry, and cross-domain diagnosis of these elements is of great significance. However, in practical applications, challenges such as insufficient training data and variability between equipment arise. To address this issue, this study proposes an innovative neural network structure, ConvNeXt, and a Multi-Scale Dilated Attention (MSDA) mechanism to improve the accuracy problem caused by inadequate feature extraction. ConvNeXt improves upon traditional convolutional neural networks by introducing a multi-scale attention mechanism to enhance the model's performance and expressiveness. Through parallel multi-channel convolution operations, ConvNeXt can capture dependencies between different channels and reduce the number of parameters. Meanwhile, the MSDA mechanism allows signals to interact and exchange information at different scales, effectively extracting complex features in one-dimensional signals. Experimental results demonstrate a significant performance improvement in one-dimensional signal processing using ConvNeXt and MSDA, better capturing relationships between global and local features in one-dimensional signals and enhancing model accuracy. The joint application of ConvNeXt and MSDA brings new solutions to one-dimensional signal processing, offering potential opportunities for effective monitoring of critical components in rotating machinery. Experimental results show that this method achieves high diagnostic accuracy in various transfer tasks, with an average accuracy of 94.28%, providing reliable support for bearing fault diagnosis.