Research On Lightweight Thermogram Recognition Model And Image Diagnosis Device Of Substation Equipment

Infrared live detection technology is widely used in various substation equipment.According to industry standards,substation equipment needs to perform infrared detection 1～3 times a year.The accurate detection process includes data collection,classification,storage,calculation,analysis and reporting.All of them rely on manual completion,and it has problems such as subjective diagnosis,inability to analyze accurately in real-time,and colossal workload.Different from general digital images,infrared temperature maps of substation equipment have many unique characteristics.Therefore,it is essential to research the real-time intelligent diagnosis technology based on the massive data.First of all,this paper collects and annotates 12,684 samples as the dataset and analyzes the statistical characteristics of infrared images of substation equipment.The design and improvement methods of the identification model of substation equipment are put forward pertinently.Secondly,this paper proposes a calculation method of adjacent sparse convolution(ASC)based on the grayscale sparsity of infrared images of substation equipment.Furthermore,a substation equipment detection model which improves efficiency and accuracy is designed.Then,an equipment defect diagnosis method is proposed based on the regional relative temperature difference and the local temperature map features,which automatically identifies typical equipment defects.Finally,this paper studies the realization method of equipment detection and intelligent diagnosis method on the embedded AI platform,develops a prototype of the handheld diagnosis device and carries out an application onsite.The research content and achievements of the paper are as follows:(1)A model improvement method for infrared image recognition of substation equipment is proposed.The statistical characteristics of the target images are analyzed from multiple aspects.The design and improvement methods of the equipment recognition model are proposed from three aspects according to the image characteristics: data formatting,image preprocessing and model lightweighting.It solves the problems of difficult identification caused by low contrast,less detailed information,and similar structures of the equipment.(2)An improved sparse convolution equipment recognition model with outlier filtering is designed.The improved sparse convolution calculation method is studied, which can reduce the submanifold expansion and be insensitive to the background.Based on this method,a lightweight target detection model suitable for infrared images is constructed,which significantly reduces the computational costs.(3)An intelligent diagnosis method for substation equipment based on the regional relative temperature difference and the temperature map features is proposed.Considering the comparison principles and empirical principles in the current diagnostic methods,the diagnostic method combining the regional relative temperature difference and the temperature map features is studied.The regional relative temperature difference realizes the improvement and automation of the existing indicators.The temperature map features realizes the mining of temperature information based on historical data.This method solves the problem of subjectivity in manual diagnosis.(4)The realization method of the intelligent diagnosis method of substation equipment on the embedded AI platform is studied,and the prototype of the real-time analysis device is developed.The parallel processing method of the improved sparse convolution method is studied,and the dynamic adjustment model of computing resources and task allocation is proposed.A handheld intelligent diagnosis device that integrates functions including infrared temperature measurement,equipment target detection,abnormal diagnosis,and analysis result storage is developed.The device was tested in the substations,and it solves the problem of the long period of an accurate analysis of equipment status.