A gearbox fault diagnosis method based on Vision Transformer-BiGRU parallel network

Traditional fault diagnosis methods that use single-modal feature extraction suffer from issues such as loss of fault information, limited representation of fault states, and low fault identification rates. This paper proposes a multi-modal gearbox fault diagnosis method based on a Vision Transformer-Bidirectional Gated Recurrent Unit (BiGRU) parallel network. First, short-time Fourier transform (STFT) is used to convert the collected one-dimensional vibration signals into fault time-frequency pattern samples that contain both time-domain and frequency-domain features. Second, based on the idea of integrating Vision Transformer and BiGRU, a dual-branch parallel network is designed. Branch one uses the Vision Transformer network to extract time-frequency image features, while branch two uses a BiGRU network to extract temporal features from the vibration signals. Then, the time-frequency image features and temporal features are fused. Finally, a Softmax classifier is employed to classify different faults based on the fused features. To validate the effectiveness of the proposed method, experiments were conducted using the dataset from the gearbox power transmission comprehensive test bench, and comparisons were made with other intelligent diagnostic methods. The results show that the proposed method achieves the highest fault identification rate, with an average diagnostic accuracy of 99.69%, demonstrating its feasibility. This study has significant research implications for the intelligent diagnosis and practical application of gearboxes.