Research On Hot Spot Image Detection Method Of Photovoltaic Module Based On Residual Attention Mechanism

With the continuous increase of photovoltaic installed capacity,the fault problems affecting photovoltaic power generation efficiency and production safety have been paid more and more attention,and hot spot is one of the main factors affecting the abnormal operation of photovoltaic power station.Hot spots will affect the efficiency of photovoltaic power generation during the incubation period,and without intervention,hot spots will continue to develop,leading to a continuous rise in temperature in this area,which directly affects the safe operation of photovoltaic power stations.However,the efficiency of traditional detection methods is not high.Therefore,efficient and accurate thermal spot detection of photovoltaic modules is one of the key links to ensure the safe and normal operation of photovoltaic systems.This paper takes the infrared image of photovoltaic module collected by photovoltaic power station as the research object.based on the characteristic that convolution neural network can extract target features through training and high recognition accuracy,a dense depth separable multi-scale residual network AB-DDSMSRnet(Attention Based Dense Depth Separable Multi-Scale Residual Network)based on attention is proposed to realize hot spot detection of photovoltaic module.First of all,taking the collected infrared images of photovoltaic modules as the data source,using image segmentation,random rotation,random horizontal movement,random vertical movement and other image preprocessing methods,on the basis of removing unclear images,infrared image data sets of photovoltaic modules with 8categories are constructed and identified.Secondly,in order to fully extract the feature information of infrared image,a depth separable multi-scale residual module DSMSRB(Depthwise Separable Multiscale Residual Block),DSMSRB based on residual idea and depth separable convolution is designed.Multi-convolution kernel is used to extract features,and a smoother activation function Mish is selected to ensure that the information flow flows better in the network.Based on DSMSRB,a deep separable multi-scale residual network DSMSRnet(Depthwise Separable Multi-scale Residual Network)is proposed.By comparing with five different neural networks on the infrared image data set of photovoltaic modules,the DSMSRnet model achieves the highest accuracy of 90.98%.Finally,in order to ensure the speediness of task recognition,the module is optimized and improved based on the idea of dense connection,and a dense depth separable multi-scale residual module Dense-DSMSRB(Dense Depthwise Separable Multi-scale Residual Block),is designed to integrate image feature information and reduce the amount of model computation at the same time.According to the remarkable characteristics of color and texture of image,the transition layer is improved by channel attention mechanism.On this basis,a dense depth separable multi-scale residual network AB-DDSMSRnet(Attention Based Dense Depth Separable MultiScale Residual Network)based on attention is proposed.The experimental results show that the optimized AB-DDSMSRnet model shows better results in operation speed and accuracy.