Abstract: This paper addresses the high proportion of small targets in distribution-line monitoring images, the sensitivity of IoU-based candidate matching to slight localization deviations, and the difficulties in adapting YOLO26s for deployment on the RV1126B platform. A scale-adaptive NWD-STAL YOLO26s detection method for edge deployment is proposed. First, a distribution-line hazard dataset containing 11,000 images is constructed, covering multiple categories of hazardous targets, including construction machinery, smoke and fire, and foreign objects on conductors. Second, in the positive-sample assignment branch during training, Complete Intersection over Union (CIoU) and normalized Wasserstein distance (NWD) are adaptively fused according to the ground-truth box area. For small-scale targets, the weight of NWD is increased to reduce the sensitivity of candidate ranking to slight localization deviations, whereas CIoU is used for matching larger targets. For edge deployment, a multi-scale raw-output graph is constructed, and INT8 quantization, DFL-free bounding-box decoding, classification-logit threshold mapping, and RGA-based image preprocessing are implemented. Experimental results show that, compared with the YOLO26s baseline, the proposed model improves recall and mAP@0.5 by 1.75% and 0.96%, respectively. The converted RKNN model has also been functionally verified on 4K images using the RV1126B platform. The results demonstrate that the proposed method achieves a balance between distribution-line hazard detection performance and edge-deployment requirements without introducing additional inference-structure overhead.