Rolling bearings are critical components in rotating machinery, and their health condition directly affects the safety and stability of equipment. To enhance the accuracy of bearing fault diagnosis, the paper proposes an innovative method called the Multi-Size Kernel-based Adaptive Convolutional Neural Network (MSKACNN). Unlike traditional Convolutional Neural Networks (CNN), MSKACNN employs a multi-size kernel structure, which allows for the automatic extraction of richer features from raw vibration signals, leading to more precise fault classification. The approach eliminates the need for manual feature ex- traction, which is often required in traditional methods. Experimental results show that MSKACNN achieves an impressive 99.83% accuracy in identifying "ball mixing" faults, outperforming methods such as Deep Neural Networks (DNN), Long Short-Term Memory networks (LSTM), and Wide First-layer Convolutional Neural Networks (WDCNN). Finally, MSKACNN has been successfully integrated into a LabView-based bearing vibration signal acquisition system, enabling fault diagnosis. The method not only improves fault recognition accuracy but also helps assess bearing quality, providing valuable insights for manufacturers to optimize production processes.