Research on transfer diagnosis method of rolling bearings based on lightweight model under cross working conditions

Abstract

Rolling bearings are the main supporting components of large-scale electromechanical equipment, and their health status directly affects industrial safety and equipment maintenance efficiency. However, traditional diagnostic model methods suffer from bottlenecks such as large model parameters and high computational complexity, resulting in poor real-time performance and high power consumption when deployed at the edge of devices. This article proposes a transfer fault diagnosis method that integrates similarity coefficient matrix features and lightweight classification models to address this issue. Firstly, a one-dimensional sequence with small size and distinct features, the similarity coefficient matrix, is obtained through fast Fourier transform (FFT) and cosine similarity calculation. Subsequently, based on one-dimensional convolutional neural networks (1D-CNN), a lightweight fault diagnosis network model was designed by reducing the number of channels and network structure, significantly reducing the number of network parameters while ensuring feature extraction capability. Finally, based on the model transfer strategy and through the feature layer freezing method, a small amount of target domain data is used to fine tune the model parameters and achieve cross-condition diagnosis. The generalization ability of the proposed diagnostic model was verified through two datasets, and the diagnostic method provides an efficient solution for bearing fault diagnosis.