In the absence of sufficient escalator motor bearing failure data, to address the issue of unstable bearing fault characteristics during frequent load and speed variations in escalator operation, this paper proposed a transfer diagnosis method for escalator motor bearings using Stockwell (S) transformation combined with subdomain adaptation. Firstly, for the fault characteristics of escalator motor bearings, a time-frequency image of vibration signals was generated using the S transform combined with bilinear interpolation. This time-frequency image effectively reflects bearing fault features and is subsequently aligned with the requirements of the feature extraction network. Secondly, local maximum mean discrepancy (LMMD) was introduced at the output end of the feature extraction network layer based on the deep residual neural network ResNet-50. It incorporates the confidence of bearing fault sample categories as weights in the mapped maximum mean discrepancy (MMD), aligning the dis-tributions of subdomains belonging to the same category, thereby expanding the scope of transfer learning. Next, the network was constructed to minimize both LMMD and cross-entropy loss functions, and network training was performed using mini-batch gradient descent. Consequently, by refining the feature differences between different fault categories, fault subdomain self-adaptation was achieved, overcoming the problem of low transfer diagnosis accuracy. Finally, based on two publicly available bearing fault datasets and a limited amount of escalator motor bearing fault data, the S-transformed time-frequency dataset was constructed, and transfer diagnosis experiments were conducted. The results demonstrate that the proposed method achieves an average accuracy of 99.1% and 95.49% for transfer diagnosis in two different source-to-target domain scenarios of escalator bearings, outper-forming five commonly used diagnostic methods in terms of recognition accuracy and robustness.