<p>Under the conditions of hybrid faults and continuously varying loads and speeds, the vibration signals of bearings become highly complex, posing new challenges to feature extraction and intelligent diagnostic methods. To address this issue, this paper designs a multi-condition hybrid fault experiment and proposes an end-to-end hybrid fault diagnosis method for rolling bearings based on a combination of one-dimensional convolutional neural networks (1D CNN), residual networks (ResNet), and Kolmogorov-Arnold Networks (KAN). First, 1D CNN is employed to extract rich temporal features from time-domain vibration signals. Then, a deep ResNet is introduced to further capture high-level abstract features, enhancing the model’s capability to represent complex patterns. Finally, the Kolmogorov-Arnold Network, with its powerful nonlinear mapping based on B-spline functions, enables accurate fault classification. To validate the effectiveness of the proposed approach, comparative experiments are conducted using both the widely recognized CWRU dataset and a hybrid fault dataset developed in this study. The experimental results demonstrate that the proposed method achieves nearly 100% diagnostic accuracy on both datasets. Furthermore, it maintains over 90% accuracy even under conditions of extremely limited samples and high noise levels. These results indicate that the proposed approach not only enables end-to-end fault diagnosis, but also exhibits excellent classification accuracy, generalization ability, and robustness to noise.</p>