Research On Abnormal Object Intrusion Detection Algorithm In Traction Substation Based On Patch Clustering Learning

The traction substation,as an important part of the traction power supply system of electrified railway,is also one of the core systems of electrified railway.It not only takes on the responsibility of changing electricity and providing electric power for electric locomotive,but also plays an important role in ensuring the safe and reliable operation of trains.Therefore,it is of great significance and practical value to study abnormal object intrusion detection technology of traction substation and realize dynamic detection of outdoor area of traction substation.In view of the low accuracy of the existing abnormal object intrusion detection methods in traction substation and the dependence on abnormal samples for model training,this paper studied the abnormal object intrusion detection algorithm of traction substation based on Patch-SVDD.The method firstly extracted the video image of traction substation,to simulate the outdoor operating environment of traction substation,and then enhanced the data to simulate rain and fog weather on the basis of the collected images.Then image clipping was used to construct the intrusion detection data set required in this paper.Considering the weather interference,Dark channel prior algorithm was used to remove the interference.In order to solve the problem of small size and diversified distribution in the whole image,the original traction substation image was segmented into patches first,and then the rough clustering was performed by Mobile Net V2 network and K-means algorithm,and the semantic clustering results of image patches were obtained.Then,according to the clustering results and the improved Patch-SVDD method based on central difference convolution,abnormal object intrusion detection and localization were carried out.In order to solve the problem of insufficient abnormal object invasion data in traction substation,the traction substation anomaly object invasion detection data set in this paper was constructed by making anomaly object invasion images manually.The method overcame the abnormal data scarce and the shortcoming of unpredictability,get rid of the dependence on abnormal samples,and also to prevent the collapse pattern based on image reconstruction anomaly detection technology,implements the anomaly detection in clustering way,and through the abnormal points assigned to the corresponding pixel has realized the anomaly object positioning.In order to verify the validity and accuracy of the anomaly object intrusion detection method based on patch clustering learning in traction substation.On the constructed data set,different sizes of segmented image patches and sampling steps were selected through ablation experiments to obtain the optimal input size of image patches suitable for the network model in this paper,and the validity and accuracy of the proposed method were verified.The accuracy of abnormal object intrusion detection is 95.6%,and the localization accuracy is 97.2%.