Fault diagnosis of automotive transmission bearings based on improved CNN and transformer

To address the challenges of fault feature extraction and the weak adaptability of diagnostic models for automotive gearbox bearings under complex operating conditions, this study proposes an improved intelligent diagnostic model that integrates Convolutional Neural Networks (CNN) and Transformers with a dual-stage dynamic sparse activation and three-dimensional attention mechanism. First, to overcome the limitations of traditional CNN with fixed architectures and limited perception of multi-domain fault features, a dual-stage dynamic sparse activation mechanism is designed. It enables adaptive computation path selection based on the complexity of input features. Then, to enhance the perception of multidimensional time-frequency-phase fault information, the Hilbert transform is applied to construct a three-dimensional feature tensor containing instantaneous amplitude, frequency, and phase. A 3D self-attention module is embedded to achieve multi-domain feature fusion. Finally, the proposed method is validated using experimental data collected under various gearbox bearing fault states and operating conditions. The results show that the model achieves an accuracy of 99.73 %, with precision, recall, and F1-score of 99.64 %, 99.63 %, and 99.68 %, respectively – all outperforming state-of-the-art methods such as GDS-YOLOv5s. Moreover, the model maintains stable recognition performance under noise and variable load conditions. These findings demonstrate that the proposed approach effectively captures subtle multi-domain fault features and exhibits strong adaptability and robustness, providing a reliable solution for intelligent operation and maintenance of gearbox bearings.