Research On Deep Learning Fault Detection Method For Infrared Image Of Power Equipment

With the rapid development of China’s economy,the power system is facing more complex challenges.More and more new technologies have been applied to the power system for the intelligent transformation of power grid.Most of the researches on infrared images of electric power equipment are completed by extracting image features and summarizing experience to detect equipment faults.However,with the expansion of the scale of power grid system,the number of power equipment increases sharply,it is difficult to use the traditional method to efficiently complete the timely diagnosis of the fault.In recent years,the research direction of deep learning has been introduced into the field of machine learning in order to complete artificial intelligence tasks,and a large number of topics have been applied to power system.Based on the target detection algorithm based on deep learning technology,this paper studies the method of equipment fault detection using infrared image.The main research contents of this paper are as follows:(1)By analyzing and comparing the current research methods and results on infrared images of power equipment,the method of deep learning is chosen to solve the detection of equipment in the image.The target detection algorithm based on deep learning was studied in depth,and Mask R-CNN was selected to complete the target detection task according to the performance of each model and the degree of adaptation to the task in this paper.(2)After analyzing the structure of Mask R-CNN,considering the particularity of infrared image data set,a new Mask R-CNN model is proposed to improve the detection accuracy.Build a feature extraction of the network based on deformable convolution and channel attention mechanism,the deformable convolution is partly introduced to obtain more flexible receptive field and to adapt to the different shape and size of power equipment.Channel attention mechanism is introduced to make the network learn the weight coefficient of each channel,so that the network can filter noise more effectively and extract more information related to the target.On the constructed power equipment infrared image data set,the experiment proves that the improved model improves the detection accuracy,and verifies the effectiveness of the improved method.(3)The principle of infrared temperature measurement is studied,the fault types of power equipment and the judgment method of thermal fault are analyzed,and the subsequent fault detection methods are analyzed after the identification and segmentation of the equipment is completed.For electrorothermic eq uipment,the clustering method is adopted to determine the pixel set of abnormal temperature,calculate the relative temperature difference between the equipment and similar normal equipment at similar ambient temperature,and preliminarily realize the detection and severity level assessment of equipment fault based on the judgment basis in the industry.The feasibility of the method is proved by example analysis.