In order to solve the problems that the fault signal is easily submerged by strong noise, the signal acquisition is not comprehensive and the training network is complex when the motor fault occurs, the transfer learning model based on multi-sensor signal fusion and the improved EfficientNetV2-B0 is introduced into the motor fault diagnosis. The sensor fusion method converts the 1D time series into images through the gramian angular field (GAF) to ensure the integrity of the feature information without time dependence, and uses the image fusion algorithm of Retinex enhancement and Laplacian pyramid decomposition to realize the image fusion of multi-source sensor signals. Adding depthwise separable convolution (DSConv) and efficient multi-scale attention (EMA) are proposed for the EfficientNetV2-B0 network, and combined with transfer learning (TL) technology to establish a motor fault diagnosis model. Various working conditions of the motor are classified and tested, and the results show that the method can effectively classify the equipment faults, and the average accuracy of the identification of various working conditions of the motor reaches 100%.