Research On Infrared Image Fault Diagnosis Of Instrument Transformers Based On Deep Convolutional Neural Network

To meet the needs of national economic development,the scale and voltage level of China’s power grid continues to improve,which leads to more severe test of the safety and reliability of the power system,and as a key device responsible for measurement and protection in the power system,the instrument transformers have become one of the essential and important equipment in the power system,however,because different types of instrument transformers have different fault characteristics and the instrument transformers are located in complex substation scenarios with different conditions,there are often situations where the thermal fault area of the instrument transformer equipment overlaps with the normal area,the traditional fault diagnosis methods alone can no longer meet the current demand for intelligent diagnosis.Therefore,it is necessary to explore new methods and techniques to improve the diagnostic accuracy and efficiency for fault diagnosis of instrument transformers.This paper used the infrared images of instrument transformers as data sources and build a semantic segmentation algorithm model of deep convolutional neural network to segment the device body and colorimetric bars for fault diagnosis of instrument transformers with different thermogenic types.The main research contents are as follows:(1)Image preprocessing,firstly,to address the problems of strong noise and blurring in infrared images,different algorithms such as mean,gaussian,median and bilateral filtering were used to denoise the infrared images of the transformer for comparison experiments,and the bilateral filter with higher PSNR value was selected as the instrument transformer denoising processing algorithm,on this basis,an improved Gauss-Laplace pyramid image enhancement algorithm was proposed to improve the image contrast,and the histogram and Contrast evaluation index control experiments were conducted in Py Charm integrated development environment to verify the effectiveness of the method.(2)To address the effectiveness of device body and colorimetric bar segmentation in complex backgrounds,the device body and colorimetric bar of the infrared image of the instrument transformer were first annotated using EISeg(Efficient Interactive Segmentation)annotation software on the basis of image pre-processing,and constructing semantic segmentation datasets of infrared images of mutual sensors;to address the problem of insufficient data samples that may lead to overfitting of the model,the dataset was expanded using a geometric approach,and finally a dataset of 6030 semantic segmentation of the infrared images of the instrument transformer was generated;secondly,an Odconv-stn-spm net(Ossnet)semantic segmentation model is designed based on the Unet semantic segmentation model by adding a full-dimensional attention dynamic convolution module to replace the original regular convolution,a streak pooling module to replace the maximum pooling of the original network,and a spatial transformation network module in the middle layer;finally,ablation experiments and model comparison experiments were conducted on 1206 data validation sets for Ossnet,and the results showed that the overall accuracy(overall accuracy,OA)of the Ossnet semantic segmentation model is 97.97%,and the mean intersection over Union(MIo U)ratio is85.78%,in which the F1＿score of the current transformer is 90.80%,and the intersection over Union ratio(intersection over union,Io U)is 82.40%;F1＿score of voltage transformer is 87.26%,Io U is 83.67%;F1＿score of colorimetric strip is 86.96%,Io U is 81.62%,compared to BANet,CRAUnet,A2 FPN,MAUet and Unet,Deeplapv3+ and other models,which can segment the device body and colorimetric bars from the samples more accurately.(3)For the fault diagnosis of the instrument transformer,firstly,the Ossnet model was designed to segment the device body and the colorimetric strip,and the temperature picture containing the temperature optimum was truncated based on the relative position of the colorimetric strip and the temperature optimum in the infrared image,and an infrared temperature recognition model was trained based on the Paddle OCR toolkit to achieve automatic temperature recognition,the test results show that the accuracy of the model reaches 99.96%;for the segmented instrument transformer devices,the temperature information of the maximum hot spot of each phase is automatically extracted by separating the three phases of the body and analyzed,then the fault area is segmented and located using Roberts’ edge detection algorithm and the minimum outer rectangle,and finally the fault diagnosis of current heating type and voltage heating type transformers was completed in Py Charm integrated development environment.