Aiming at the problems that fault diagnosis data of wind turbine gearboxes have time series and the single-channel model is difficult to effectively extract the composite fault feature information, a fault diagnosis method integrating an improved orthogonal convolutional capsule network (OCCN) and a bidirectional long short-term memory neural network (BiLSTM) is proposed. First, the original signals are preprocessed. Then, the preprocessed signals are fed into a constructed OCCN-BiLSTM dual-branch model to extract the spatial features and time domain features of composite faults, respectively. Finally, the extracted spatiotemporal features are fused through a cross-attention mechanism and input into a fully connected layer for signal classification, enabling intelligent fault diagnosis of wind turbine gearboxes. Experimental results show that the proposed diagnosis method can effectively achieve intelligent fault diagnosis for wind turbine gearboxes, with an accuracy of 99.53% on the test set.