通信与信息技术

为提升变电站通信机房隐患检测能力，本文提出基于YOLOv8-DUA的优化深度学习模型。针对隐患背景相似度高、目标尺度多变及样本不均衡等问题，模型从三方面进行改进：首先，引入DETR技术优化特征提取与检测头，增强小目标检测能力；其次，将主干网络C2f模块升级为C2f-UniRepLKNetBlock，提升特征提取效果与泛化性能；最后，在颈部网络创新融合ASF-YOLO的Attentional Scale Sequence Fusion与DySample，提出DyS-ASF机制以强化多尺度特征融合。经实验验证，优化后模型在机房隐患检测任务中表现显著提升，mAP@0.5达96.4%（较原模型+5.6%）,mAP@0.5～0.95提升4.6%，准确率与召回率分别提高4.1%和6.5%。同时开发了基于PyQt5的应用程序，支持图像、视频及摄像头实时检测，有效弥补人工检测疏漏，为变电站通信机房安全运行提供了高效技术保障。

To enhance the detection capability of potential hazards in substation communication equipment rooms, this paper proposes an optimized deep learning model based on YOLOv8-DUA. Addressing challenges such as high similarity in hazard backgrounds,multi-scale target variations, and sample imbalance, the model incorporates three key improvements: First, DETR technology is introduced to optimize feature extraction and detection heads, enhancing small target detection performance. Second, the C2f module in the backbone network is upgraded to C2f-UniRepLKNetBlock, improving feature extraction effectiveness and generalization capability.Third, a novel DyS-ASF mechanism is proposed by integrating ASF-YOLO's Attentional Scale Sequence Fusion with DySample in the neck network, strengthening multi-scale feature fusion. Experimental results demonstrate significant performance improvements: The optimized model achieves an mAP@0.5 of 96.4%(a 5.6 percentage-point increase over the original YOLOv8 model), with mAP@0.5-0.95 rising by 4.6 percentage points, while recognition accuracy and recall improve by 4.1% and 6.5%, respectively. Additionally, a PyQt5-based application is developed to support real-time image processing, video analysis, and camera-based detection, effectively compensating for human inspection oversights. This research provides an efficient technical solution to ensure the safe and stable operation of substation communication infrastructure through the integration of algorithmic optimization and engineering applications.