Aiming at the low precision of gearbox fault diagnosis and the high requirement of computer hardware for deep neural network model, a Shuffle-ECANet network model was proposed for gearbox fault diagnosis. The model was based on the lightweight neural network ShuffleNet V2, which optimized the network model while retaining the lightweight structure of the network. The Gelu activation function was used to enhance the nonlinear transformation capability of the model, and the efficient channel attention (ECA) module was embedded to improve network performance. Depthwise separable convolution improved the computational efficiency of the network model, and channel shuffling technology maken information more fluid and improved feature expression capabilities. The experimental results show that the network model proposed in this paper is suitable for gearbox fault diagnosis under different noise conditions while ensuring light weight. The diagnostic accuracy can reach 99.6% under the original signal, and the signal-to-noise ratio of -8 dB is added. The diagnostic accuracy can reach 92.7% under Gaussian white noise. The method proposed provides a new way for the neural network to be better applied to gearbox fault diagnosis.