A Ghost-Efficient Yolov7-Based lightweight model for Multi-Scale power operation risk identification

Abstract In the power operation scenarios, there are problems such as large differences in target scales, small targets susceptible to occlusion, and limited computational resources for edge devices. This leads to insufficient detection accuracy and real-time bottleneck in traditional target detection models. The study proposes ghost-efficient YOLOv7 power optimization model (GE-YPOM) for power operation risk identification system to address the above challenges. Deeply separable convolution and pruning techniques are used to reduce the computational effort. Moreover, the ghost module (Ghost) is introduced and combined with EfficientNet to enhance the feature expression capability through the channel attention mechanism. The experimental results indicated that the parameter count of GE-YPOM was only 4.2 M, which was 36.4% less than that of YOLOv7. The inference speed reached 28.5 frames per second, an improvement of 27.8%. The small target detection accuracy (mAP@0.5) was 0.923, which was 5.4% higher than that of YOLOv7 (0.876). Equipment fault detection accuracy reached 0.941, and multi-scale detection average accuracy was 0.932. On the Jetson Nano edge device, the continuous detection memory footprint was 65.3 MB, which was only 52.2% of YOLOv7. Its stability score was 94.56%. The results show that GE-YPOM effectively solves the problem of multi-scale target detection in power operation through lightweight architecture design and feature optimization. The method improves the real-time performance and energy efficiency of edge devices while ensuring the detection accuracy. It has important engineering application value for reducing operation risk and improving operation and maintenance efficiency.