Aiming at the problem of insufficient effective fault samples in shearer fault diagnosis process, a transfer learning method for shearer fault diagnosis under variable working conditions was proposed. The original vibration signals were transformed by wavelet to obtain the time-frequency graph, the time-frequency characteristics of the fault were highlighted by the principle of image enhancement, and the fault samples were formed after the image normalization. A large number of existing fault data under different working conditions were used to form source domain data, and the convolutional neural network was trained to obtain the initial fault diagnosis model. The trained model was transferred to the shearer fault diagnosis test bed, maximum mean discrepancy (MMD) was used as the optimization index, and the samples in the test bed were used to train the model and realize weight fine tuning. The results show that the vibration signals processed by wavelet transform and image enhancement can effectively highlight the fault characteristics, the model transfer of shearer fault diagnosis can be effectively realized by weight fine tuning of small samples on the test bed.