Safety control is one of the most important tasks for power grid operation. However, safety control and risk warning models with high accuracy always have multitudinous parameters and need scaled resources for calculation. So as to sort out this problem, we proposed a novel and adaptive model lightweight method based on knowledge distillation. In our method, the intermediate features of the teacher network is utilized to guide the learning of the student network and a regressor is introduced behind the selected middle layer of teacher network to realize dimension match. The prediction confidence of teacher network is mapped to the feature space for the regressor initialization. Besides, the prediction confidence is used to construct the distillation temperature to improve the adaptability of information transfer. Experiments on ImageNet, CIFAR 100 and Grid-specific datasets demonstrate the lightweight model obtained by our method possesses powerful and adaptive feature extraction and representation ability.

Real-time estimation of the safety distance between power operation workers and energized equipment is essential for preventing electric shock, arc flash injury, and equipment accidents during live or near-energized operations. However, large-scale deployment remains difficult because conventional binocular ranging systems require accurate offline calibration, while monocular vision suffers from metric-scale ambiguity. This paper proposes a weakly calibrated multi-camera vision framework for safety distance estimation in power operation scenes. First, a cross-camera tracking strategy combining mask-guided ReID features, Kalman filtering, spatio-temporal constraints, and Hungarian matching is developed to maintain globally consistent worker identities. Second, a region-wise self-supervised depth estimation model is introduced: photometric consistency is imposed within worker regions, while cross-view feature reprojection is imposed within rigid equipment regions. Intrinsic parameter regularization and scene-inherent dimensional priors are then used to recover metric scale without placing additional calibration targets. Third, the YOLOv8-seg backbone is enhanced with a CNN-Transformer C2f_CT module to improve pixel-level segmentation of workers, energized equipment, and structural components under cluttered backgrounds. Experiments on three representative power operation scenarios show that the proposed method achieves 92.3% MOTA, 94.2% Rank-1 accuracy, 82.0% mAP, 8.7% relative distance error, and 22 FPS end-to-end speed. When using five-frame smoothing and uncertainty threshold, the proposed method can achieve 98.7% warning precision, 0.9% missed alarm rate and 1.3% false alarm rate. The results indicate that the proposed framework provides a practical balance between deployment cost, geometric accuracy, and real-time warning capability.