To address the challenges of poor fault diagnosis accuracy of wind turbine gearbox acoustic signals and poor anti-jamming performance of single signal source, a multimodal feature fusion fault diagnosis method with shrinkage denoising convolutional neural network and residual network with group attention mechanism was proposed. Firstly, the noise was filtered in a one-dimensional shrinkage denoising convolutional neural network to improve the signal quality and extract the features. Secondly, the acoustic signal was transformed into a recursive image in an improved residual network to label the valuable channel features and extract them. Finally, the multidimensional features were aligned and the residual structure was used to strengthen the output features to complete the fault recognition. The effectiveness and generalization of the method were verified using the gear eccentricity fault dataset from the University of Macau and the planetary gearbox fault dataset from the Beijing Institute of Technology. The experimental results show that the diagnostic accuracy of both datasets reaches more than 98% without noise, and more than 90% in multiple noise experiments. Compared with other cutting-edge data fusion diagnostic methods, it has advantages such as high diagnostic accuracy, strong anti-noise performance and good generalization.