Defect Detection And Segmentation Of Substation Equipment Based On Deep Learning

The normal operation of substation equipment is the prerequisite for the stable operation of power system,so the defect detection of substation equipment is very necessary.How to find the hidden danger of substation equipment quickly and accurately has become an important problem.Traditional inspection mainly relies on manual inspection,which has many disadvantages in terms of time,efficiency and safety,and cannot meet the needs of substation safety inspection.With the development of artificial intelligence and image processing technology,deep learning algorithm provides new technical solutions for automatic inspection of substation equipment.Based on the actual project requirements,this thesis researches and improves the deep learning algorithm and finally applies it to the defect detection and classification of substation equipment.The main research work of this thesis is as follows:(1)Image preprocessing and data augmentation.Due to the small Dataset of substation defective equipment,based on the self-collection Dataset,the geometric transformation method is used for Dataset augmentation,and then the Dataset is enhanced by the Cycle GAN algorithm,which effectively solves the problem of small sample data.(2)In the defect detection stage,a defect detection algorithm based on Anchorfree bidirectional cross-scale feature fusion is proposed for substation equipment defect detection.Using Efficient Net as the feature extraction network,different models can be selected according to the application scenarios.In order to improve the detection accuracy,the improved Bi FPN is used for feature fusion to solve the problem that the size of the defective equipment is greatly different,and the small target object is difficult to locate.To overcome the problems of hyperparameter setting,Io U computing takes up a lot of resources and the difference of the number of positive and negative samples.Therefore,the Anchor-Free mechanism is used to perform object classification and position regression on the feature map.At the same time,the adaptive training collection method is used to further optimize the positive and negative sample classification problem,which effectively improves the detection accuracy and speed.(3)In the defect segmentation stage,a defect segmentation algorithm based on bottleneck attention mechanism is proposed.To solve the problem of unfriendly segmentation of small target objects in the segmentation process,the bottleneck attention mechanism is used.Keeping as many key features as possible while reducing parameters and increasing the receptive field.Therefore,four different sampling rates of void convolution are used in the improved ASPP.In order to achieve multi-scale and efficient segmentation of image targets,the encoder-decoder idea is used to combine the improved ASPP structure with Atrous Convolution with the Encoder-Decoder structure,and finally the edge of the segmented model is more accurate.(4)Based on the QT framework and combined with the above detection and segmentation algorithm models,a set of substation equipment monitoring and management system is designed.To achieve real-time video monitoring,video playback,defect detection and segmentation,upload records,log query,system settings and other functions.After the system is put into use in the project entrusted by the enterprise,it realizes the new operation mode of "less staffed,unmanned duty,regional maintenance" of the substation,which is of great significance to promote the transformation of substation equipment from traditional to monitoring.