首页|变电站多尺度异常入侵目标轻量化检测方法

变电站多尺度异常入侵目标轻量化检测方法

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
智能电网的建设思路决定了远郊变电站异常入侵监测的无人化趋势,促进了变电站异常入侵智能检测方法的快速发展.现阶段尚未拥有该场景下异常入侵目标数据集,且现有的目标检测方法也未针对变电站边缘计算端进行轻量化优化设计,不适用于需要全天候实时监测的变电站边缘设备.针对上述问题,从实际应用需求出发,构建变电站异常入侵目标数据集(Dataset for Anomaly Invasion Targets in Substations,SAITD),基于YOLOv5s模型提出适用于变电站边缘检测设备的轻量化异常入侵目标检测网络YOLOv5-Substation.添加微小尺度目标特征提取层与上采样轻量化算子CARAFE,在扩大感受野的同时,充分保留特征图中多尺度目标的语义信息,从架构端提高原有模型的检测精度.基于知识蒸馏模型,使用网络剪枝(Network-slimming)策略对原有模型进行轻量化改进,在保证原模型检测精度的同时,加速模型推理.仿真实验表明,轻量化后的边缘端计算模型精度相较于YOLOv5s提高了 3.3％,推理速度提升了 41.9％,可为智能电网的全速运行提供强有力的数据基础、技术支撑与安全保障.
Lightweight Detection Method for Multi-scale Anomaly Invasion Targets in Substations
The construction concept of smart grids has determined the trend of unmanned monitoring of anomaly invasion targets in suburban substations,promoting the rapid development of intelligent detection methods for anomaly invasion targets in substations.However,at present,there is no dataset specifically designed for anomaly invasion targets in this scenario,and the existing target detection methods have not been optimized for lightweight design on the edge computing end of substations,which is not suitable for real-time monitoring of substation edge devices that require round-the-clock monitoring.To address these issues,starting from practical application requirements,a Dataset for Anomaly Invasion Targets in Substations(SAITD)is constructed,and a lightweight anomaly invasion target detection network,YOLOv5-Substation,which is suitable for substation edge detection devices is proposed based on the YOLOv5s model.A micro-scale target feature extraction layer and an upsampling lightweight operator CARAFE are added to expand the receptive field while fully preserving the semantic information of multi-scale targets in the feature map,improving the detection accuracy of the original model from the architecture end.Based on the knowledge distillation model,the original model is lightweighted using Network-slimming strategies to ensure the detection accuracy of the original model while accelerating model inference.Simulation experiments show that the accuracy of the lightweighted edge-end computing model is 3.3％higher than that of YOLOv5s,and the inference speed is 41.9％faster,providing a strong data foundation,technical support and security guarantee for the full-speed operation of smart grids.

anomaly invasion target detectionNetwork-slimmingknowledge distillationedge computing platformlightweight model

潘磊、赵枳晴、傅强、郑远、田俊

展开 >

中国民用航空飞行学院计算机学院,四川德阳 618307

异常入侵目标检测 网络剪枝 知识蒸馏 边缘计算平台 轻量化模型

中国民用航空飞行学院智慧民航专项民航飞行技术与飞行安全重点实验室开放基金民航飞行技术与飞行安全重点实验室自主研究项目中国民用航空飞行学院重点面上项目中国民用航空飞行学院研究生创新项目(2023)

ZHMM2022-005FZ2022KF10FZ2022ZZ06ZJ2021-11X2023-29

2024

10.3969/j.issn.1003-3106.2024.06.030

无线电工程
中国电子科技集团公司第五十四研究所

无线电工程

影响因子：0.667
ISSN：1003-3106
年,卷(期)：2024.54(6)
  • 3