To address insufficient semantic representation of dial damage, complex background interference caused by pseudo-textures such as dew and water vapor, and unstable localization of small-scale defects in long-distance substation instrument appearance defect detection, this paper proposes a distilled lightweight multi-scale feature-enhanced and regression-optimized YOLO model (DLMFR-YOLO) based on YOLO11n. First, a bidirectional multi-scale feature amalgamation propagation network (BiMFAPNet) is constructed to enable bidirectional interaction and detail compensation of multi-scale features in the neck, thereby enhancing the semantic and structural representation of oblique-view and small-target defects. Second, a star attention feature extraction module (SAFE) is designed to highlight key regions such as dial cracks and shell damage in both spatial and channel dimensions through star operation and contextual anchor attention, effectively suppressing interference from complex backgrounds and pseudo-textures including dew and water vapor. Third, WIoUv3 is introduced as the bounding box regression loss to improve regression stability and accuracy for small targets through a dynamic quality weighting mechanism. On this basis, a lightweight StarNet backbone and alightweight shared convolutional detection head (LSCD) are adopted to reduce parameters and computation. Finally, a logical distillation framework based on bridging cross-task protocol inconsistency knowledge distillation (BCKD) is constructed, where the models before and after lightweight redesign serve as the teacher and student networks, respectively, to compensate for the accuracy loss caused by lightweighting. Experimental results on a self-built substation instrument appearance defect dataset show that, compared with the YOLO11n, the DLMFR-YOLO improves mAP@0.5 by 3.2%, while reducing parameters by 31.4% and FLOPs by 15.9%, demonstrating clear advantages for engineering application and deployment.