Traditional fault diagnosis methods relying on single-modality feature extraction suffer from issues such as loss of fault information, limited representation of fault states and low fault recognition rates. Fault diagnosis method of engine gear transmission was hence proposed based on an optimized bidirectional gated recurrent unit (BiGRU) and Swin Transform. The Gramian Angular Difference Field (GADF) was first employed to transform the acquired one-dimensional vibration signal into a two-dimensional image with time-frequency features. Then a dual-channel feature extraction network was designed, Channel 1 employed a Swin Transformer network to extract time-frequency image features from the signal, while Channel 2 utilized an improved BiGRU-ATT network with an attention mechanism to capture deep temporal features of the signal. Subsequently, the feature fusion between time-frequency image features and temporal features was conducted. Softmax classifier was finally applied to the fused features for fault classification. To validate the proposed method, experiments were conducted using a dataset from a comprehensive powertrain fault diagnosis test bench. Comparative experiments with other intelligent diagnostic methods demonstrated that the proposed approach achieved the highest fault recognition rate with an average diagnostic accuracy of 99.68%. This confirmed its feasibility and provided valuable insights for the intelligent diagnosis and practical application of engine gear transmission mechanisms.