The detection of defect severity in cable terminations plays a critical role in ensuring the safe and stable operation of high-speed trains (HSTs). However, the partial discharge (PD) characteristics of the same type of defect can appear similar across different severities, posing challenges for accurate insulation defect degree identification. Consequently, this article proposes an image transformation method, named the equivalent bandwidth matrix of relative locations (EBMRLs), coupled with the self-guided transformer (SG-Former) algorithm, which is more effective for fine-grained image recognition, to accurately identify different degrees of defects with similar PD characteristics. In the proposed approach, the original PD signals are first converted into images using EBMRL. This transformation embeds the characteristic and bandwidth information from the original PD data into the images, thereby reducing the similarity of information between classes in the transformed images and enhancing their distinguishability. Subsequently, the local and global features of the transformed EBMRL images are extracted to train the SG-Former model. The model is finally utilized to identify the severity of defects in cable terminations. The results demonstrate that the method proposed in this article achieves better performance compared with some of the state-of-the-art methods.