A Machine Learning Framework Based on ACO-EMD-CKAN for Bearing Fault Detection

Rolling bearings are the core components of rotating machinery, so their health condition is directly related to safe production. However, there are still some problems in the traditional fault diagnosis of rolling bearings with low accuracy, which is difficult to adaptively extract typical traits. Therefore, a new method for motor rolling bearing fault diagnosis based on ACO-EMD-CKAN is proposed in this work. Firstly, the data is optimized by Ant Colony Optimization (ACO) to determine the optimal number of Intrinsic Mode Function (IMF) values. Then, denoising is performed through Empirical Mode Decomposition (EMD), and more prominent feature vectors are extracted. Next, these vectors are sent to the Convolution Kolmogorov-Arnold Network (CKAN). After passing through three layers of convolution and three layers of maximum pooling, they are finally sent to the Kolmogorov-Arnold Network (KAN) layer to perform adaptive feature extraction of the signal. Compared with the common bearing fault diagnosis methods, the proposed ACO-EMD-CKAN model has the highest accuracy, reaching $\mathbf{9 8. 8 8 \%}$. The experimental results verify the feasibility of the model for effective classification of bearing faults, and provide important application value for efficient and reliable bearing fault detection in the future.