基于改进单阶段目标检测算法的换流站电气设备目标检测

Target Detection of Electrical Equipment in Converter Station Based on Improved Single-stage Object Detection Algorithm

针对换流站多种电气设备检测时背景复杂干扰性强而又需要快速准确检测出故障的实际情况,提出基于改进YOLOv5(You Only Look Once)的检测方法。首先,为提高算法的准确性和收敛速度,通过K-means聚类算法对YOLOv5模型中的锚框预设进行改进,在数据集预处理阶段得到更适用于换流站电气设备的锚框,使其更加契合换流站电力设备数据集;然后,为提高算法检测过程的识别速度,在特征提取网络添加注意力机制模块,筛选出重要的特征信息。将改进后的算法网络识别效果与YOLOv5中的原始算法网络检测结果进行对比分析。结果表明,检测平均识别精度均值由71.16%提高至92.51%,检测速度由21帧/s提升至31帧/s;同时与R-CNN(Regions with convolutional neural networks)等算法相比,检测精度与速度都有较大提升。添加可解释性分析,将识别结果通过热力图的形式显示,可以更好地应对算法的潜在风险。

In view of the fact that the background complex interference is strong and the fault needs to be detected quickly and accurately when detecting various electrical equipment in converter station,a detection method based on improved YOLOv5 is proposed.Firstly,in order to improve the accuracy and convergence speed of the algorithm,the anchor frame of YOLOv5 model is improved by K-means clustering algorithm,and the more suitable anchor frame for electrical equipment in converter station is obtained in the data set preprocessing stage to make it adapt to electrical equipment data set of converter stations.Then,in order to improve the recognition speed of the algorithm detection process,the attention mechanism module is added to the feature extraction network to filter out the important feature information.The improved algorithm network recognition effect is compared with the original algorithm network detection result in YOLOv5.The results show that the average detection accuracy is increased from 71.16%to 92.51%,and the detection speed is increased from 21 frames/s to 31 frames/s;at the same time,compared with regions with convolutional neural networks,the detection accuracy and speed are greatly improved.By adding interpretability analysis and displaying the identified results in the form of thermodynamic diagram,the potential risks of the algorithm can be better dealt with.