Abstract: In order to solve the problems that the fault signal is easy to be flooded by strong noise, the collected signal is not comprehensive and the training network is complicated, this paper introduces the residual neural network with multi-sensor signal fusion and attention mechanism into gearbox fault diagnosis. Firstly, the signals collected by multiple sensors are fused based on the variance contribution rate of the vibration signals to obtain more comprehensive fault information of a gearbox. Then, a time-frequency diagram of the signal is obtained by wavelet transform, and the two-dimensional time-frequency information of the fault signal is constructed. Finally, the residual neural network (ResNet) with local cross-channel interaction strategy (ECA module) is used to learn and classify different fault states. After the global channel-level average pooling without reducing dimension, the classification effect is obviously improved. Through the identification and analysis of the gearbox fault signals under different fault types, different signal-to-noise ratios and different working conditions, and compared with different diagnosis methods, it is proved that the proposed method is feasible and a fast recognition rate.