Diagnosing bearing faults through vibration signals is crucial for ensuring the safe operation of aircraft. Aiming at the problems of poor working environment, strong noise interference, large consumption of artificial diagnosis resources and low diagnosis accuracy, a variational mode decomposition method based on envelope cross-spectral entropy and a fault diagnosis model based on group convolution ConvNeXt were proposed. Addressing the challenge of accurately determining the number of operational modes when variational mode decomposition was used to process vibration signals, an envelope cross-spectral entropy index was proposed, which was used to measure the similarity of fault characteristics of adjacent modes. The number of modes was increased in turn. When the envelope cross-spectral entropy of adjacent modes was less than the threshold, it was considered to have an over-decomposition problem. To address the issue of low accuracy inherent in traditional models, a new approach has been proposed to use group convolution to extract features of each mode, force them to learn features of different frequency bands independently, and classify faults through ConvNeXt network. The experimental findings demonstrate that the envelope cross-spectral entropy index is highly effective in accurately determining the optimal number of modes for variational mode decomposition, and the group convolution ConvNeXt model can effectively extract fault features. It was verified on the misalignment bearing data set of China Coal Information Technology (Beijing) Company, and the fault diagnosis accuracy was up to 99.86%.