Research On Infrared Diagnosis Technology For Power Substation Equipment Based On Deep Learning

With the development of infrared inspection of substation equipment in the direction of intelligence,the amount of infrared image data collected by inspection robots and UAVs with their own infrared thermal imagers has increased geometrically,and it is obviously unrealistic to carry out subsequent diagnosis of the pictures by manual means.In response to the problem of large workload and low efficiency of infrared diagnosis in substations,this paper takes the infrared image data of typical electrical equipment collected by infrared thermal imagers in the operation and inspection department of the State Grid over the years as a data set,four types of substation equipment,namely insulators,equipment clamps,bushings and disconnect switches,are used as research objects.Firstly,to address the problems of small number of defective infrared images and unbalanced data set,we propose to use SMOTE algorithm to equalize the data sample set,and adopt the data expansion method of randomly selected mixed enhancement means to enhance the infrared defective images,and finally carry out uniform cropping and the data is then uniformly cropped and labelled to construct the sample data set required for the experiment.In order to adapt to the characteristics of the infrared image dataset of power equipment and to solve the problems of small target size of substation equipment and the difficulty of recognition,this paper,based on the original YOLOv3 model,firstly adjusts the network structure and adds the NIN convolution layer to avoid excessive loss of spatial information caused by downsampling and to better adapt to the detection of abnormal heating regions of power equipment to be tested,and at the same time reduces the number of parameters and improves the Then,in order to ensure that the anchor frame size is close to the power equipment to be identified,it is necessary to recluster the anchor frames according to the k-means algorithm,set different numbers of anchor frames and calculate the intersection ratio between the target prediction frame and the real frame,and finally obtain the optimal number of anchor frames and the cluster size.Finally,the loss function in the original network model is improved by introducing Focal loss to solve the sample imbalance problem,and then replacing the category loss function of the original loss function with a balanced cross-entropy function to improve the feature learning ability of the model.In order to verify the effectiveness of the improved YOLOv3 model for recognition and localization,the improved YOLOv3 model,the original Yolo v3 model and the Faster rcnn model were compared on the test set by means of accuracy,recall,MAP and detection speed,and the results showed that the YOLOv3 algorithm proposed in this paper has higher accuracy and MAP than the original model and the Faster rcnn algorithm,and the improvement in accuracy,MAP and detection speed is quite significant.Finally,this paper proposes a fault diagnosis method for power equipment based on the grey-scale-temperature conversion formula and the improved YOLOv3 model.The method combines the recognition and location function of the improved YOLOv3 model with the grey-scale processing method,and firstly classifies and locates the wire clamps,insulators,bushings and disconnect switches to be identified in the input image respectively,and then intersects and merges the equipment areas and The intersection ratio is then calculated for the equipment area and the abnormal heat area in the input image,and the fault temperature criterion of the corresponding electrical equipment is obtained according to the network classification result.The model is tested on infrared samples of casing and the results demonstrate the feasibility and effectiveness of the method.