An on-load tap changer (OLTC) regulates transformer voltage under load, while repeated portable imaging provides records of accessible component geometry. This study proposes TapBR-SSM, which combines high-resolution HRNet features, stage-selective Mamba-inspired directional refinement, degradation-conditioned correction, and semantic–boundary–radial prediction. Training uses five loss groups for region, boundary, shape, cross-path consistency, and shallow restoration supervision. The dataset contains 330 image–mask pairs allocated to training (210), validation (60), and testing (60). Across nine training and evaluation runs, TapBR-SSM yielded mean mIoU and mDice values of 0.9806 and 0.9902, respectively, on the 60-image test set. For longitudinal analysis, the test images were organized by date into 32 cases spanning 20 August 2024 to 17 April 2025. Reference- and prediction-derived ellipse-normalized contour ratios R_c decreased by approximately 0.247% and 0.272%, respectively, between the endpoints. Across the 32 date-based cases, the prediction-minus-reference signed bias, MAE, RMSE, and maximum absolute difference were +0.000992, 0.001129, 0.001431, and 0.003988, respectively; the Pearson correlation, Spearman correlation, and Lin’s concordance correlation coefficient (CCC) values were 0.9765, 0.9531, and 0.9555, respectively.