Research On Electrical Equipment Identification And Thermal Fault Diagnosis Based On Lightweight Network

In recent years,with the rapid development of electric power industry of China,smart grid has become a major demand.The research on the intelligent detection system of electrical equipment based on deep learning platform is of great significance to realize the automation and intelligence of electric power inspection.The popular deep learning system can automatically learn the characteristics of target samples from a large amount of data,but its model structure is increasingly complex,and cause a huge amount of computing consumption,making it difficult to be applied in real-time inspection system.On the other hand,the precision of target detection is greatly affected by the limited computing power of mobile devices based on edge computing technology.In this research,a lightweight deep learning infrared image detection algorithm for electrical equipment based on embedded edge computing platform is proposed,which provides a feasible scheme for building an intelligent inspection system for electrical equipment with both real-time performance and accuracy as follows:(1)According to the infrared images of 194 110 KV and 111 35 KV substations in7 areas of Tianjin collected by inspection engineers with infrared thermal imager in recent five years,the Pascal VOC format infrared image data set of electrical equipment is marked and established,which provides a basis for the subsequent training and testing of deep learning algorithm.The dataset contains 10516 infrared images,and each contains 2-3 types of electrical equipment including current transformers,insulators,arresters,circuit breakers,disconnectors and drivepipes.In particular,the equipment and fault areas with abnormal temperature in 981 infrared images are marked.(2)This research proposes Lightweight ES,a lightweight convolutional neural network suitable for edge computing devices,to transform the classic SSD network.Firstly,Mobile Net V3 lightweight network is used as the backbone network of feature extraction to extract image features efficiently.Then the efficient channel attention module(ECA)is introduced to improve the detection accuracy of the network.Finally,the Soft Pool method is used to reduce the loss of the pooling information and improve the classification accuracy.The improved algorithm is applied to the infrared image data set of electrical equipment for testing,and the experimental results show that Lightweight ES algorithm can accurately identify electrical equipment and local temperature anomalies in real time.(3)The abnormal area of electrical equipment temperature is usually small in infrared images.Aiming at the problem that the target detection algorithm is not accurate enough to detect small targets in the abnormal area of equipment temperature,this thesis proposes a lightweight multi-scale feature fusion algorithm Lightweight ESFPN based on the original version.In this algorithm,the feature pyramid network structure is introduced,and the high resolution shallow feature layer and low resolution deep feature layer are fused to improve the overall detection accuracy of the model.Experimental results show that Lightweight ES-FPN algorithm can significantly improve the detection accuracy of electrical equipment temperature anomaly areas while ensuring the detection speed.(4)The Lightweight ES-FPN network is ported to NVIDIA Jetson TX2 edge computing platform.Experimental results show that this platform can meet the balance requirements of precision and real-time performance,providing a feasible solution for the real-time unmanned inspection system of power grid,as well as the practical application value.