Research On On-line Monitoring Technology Of Moa Condition For Small Sample Infrared Images

The safe and stable operation of power transmission and transformation equipment is directly related to the normal operation of power grid system.Statistical results show that the national electricity consumption is increasing year by year,so the load of power transmission and transformation equipment is bound to have an impact.How to find the abnormal state of power transmission and transformation equipment in time through effective detection and monitoring means is of great significance to ensure the safety of power grid operation.With the development of power Internet of things,the operation and maintenance of power transmission and transformation equipment is gradually developing from traditional manual inspection to intelligent detection.However,due to the complexity of the environment and the small number of samples,it is a difficult problem for automatic inspection.Therefore,this paper takes the metal oxide arrester(MOA)as the research object,and proposes a small sample oriented on-line monitoring scheme of the MOA infrared state.Firstly,the automatic recognition method of MOA at edge end is studied.The SSD in one stage algorithm is selected,and the lightweight mobile net structure,Mobile Net,is used to replace the VGG16 feature extraction network in the original network.What’ more,the anchor frame size is modified according to the shape of MOA.The final experimental results show that the improved SSD algorithm can identify and locate the MOA quickly and accurately in different background,which provides a good foundation for the subsequent cloud state identification.Secondly,single convolutional neural network is often limited by its own characteristics in different background applications,and the existing fault data is less,so it is difficult to extract deep fault features.Therefore,a MOA infrared thermal fault detection model based on multi model fusion is proposed in the cloud.The multi-level features of image are extracted by multiple pre trained deep convolution neural networks,and then the feature vectors are used to train multiple relevance vector machines(RVM).Finally,the combination strategy is used to fuse the detection results,so as to improve the accuracy and generalization ability of the detection model under the condition of small samples.The experimental results show that the ensemble classifier based on multiple features can greatly improve the accuracy of MOA thermal fault recognition.Finally,aiming at the problem that the model accuracy is not high enough due to data imbalance in the previous chapter,a data expansion model based on transfer learning and deep convolution generation adversarial network(DCGAN)is proposed.The number of infrared images of normal MOA is much more than that of fault MOA.Therefore,the idea of migration is introduced.Firstly,the model DCGAN1 of normal MOA image can be generated by weight training of classic network,and then the model DCGAN2 of fault MOA image can be generated by weight training of DCGAN1 again.Finally,using the expanded data set to train the defect detection model proposed in the previous chapter,the results show that it can effectively improve the recognition rate of MOA thermal fault detection model,which proves that the proposed data expansion method is helpful to improve the accuracy of fault recognition.