Research On Detection Method Of Substation Oil Leakage Equipment Based On Image Processing

Power capacitor equipment is widely used in substations,and its main function is to compensate the reactive power of power system and improve the voltage quality of power system.Due to the different manufacturing technology of capacitors and insulation aging,power capacitors will have oil leakage faults.When the tightness of the capacitor is destroyed,the gas,water and some residue will enter into the capacitor to reduce the performance of the capacitor,and the oil surface will also be reduced,so that the insulation of the capacitor will decrease,and even breakdown discharge will occur,which will lead to more serious power accidents.The traditional capacitor inspection is manual inspection,but most of the capacitor equipment in the substation is more than 2 meters,and the location of oil leakage is on the top of the capacitor,so the labor intensity of manual inspection is high.Nowadays,video surveillance equipment covers the entire substation,and smart grid technology is also applied.Power equipment operation status monitoring based on target detection has become a research hotspot.In this paper,taking the integrated shunt capacitors of two substations in Yunnan Province as the research object,a set of capacitor data set with 2570 images is established.On the basis of YOLOv3 target detection algorithm,a multi-scale YOLOv3 model which can effectively detect capacitor oil leakage fault under complex background and low resolution photos and a MobileNetV2-YOLOv3 capacitor oil leakage detection model which can be transplanted to embedded devices or hardware devices with low computing power are designed to meet the requirements of real-time detection speed.The specific research contents are as follows:(1)Establishing the capacitor oil leakage data set: by marking the capacitor photos obtained by the field photography and monitoring system,the initial data set with a small number of samples is established;the capacitor oil leakage data set is expanded by means of image enhancement to prevent overfitting during model training and enhance the generalization ability of the model.(2)Improve the YOLOv3 capacitor detection algorithm: choose GIOU loss as the location loss function to replace the original loss function of YOLOv3,and re-cluster the width and height of the capacitor data set using the K-means algorithm to obtain the boundary box size and quantity suitable for the capacitor data set.Simulation results verify the effectiveness of the two improvements.(3)A multi-scale YOLOv3 detection model is designed: the fourth 104x104 minimal scale prediction is added to the three-scale prediction of the original YOLOv3 algorithm,and the simulation results show that under any background and lowresolution photos,the model can effectively detect the oil leakage of power capacitors by 4.7% compared with the standard YOLOv3 algorithm.(4)A lightweight MobileNetV2-YOLOv3 detection model is designed: using the deep separable convolution design idea of Mobile Net,the Dark Net53 backbone feature extraction network of YOLOv3 is changed to MobileNetV2,.A SPP module is added to the output layer of the backbone extraction network,which combines global features with local features,which enriches the expression ability of feature matrix.The simulation results show that the lightweight YOLOv3 algorithm sacrifices part of the m AP(mean average accuracy),but the detection speed is nearly three times higher,and the number of parameters is reduced by 60%.It achieves the balance between detection efficiency and detection speed,and is more suitable for embedded devices or devices with low computing power.