Design And Implementation Of Anomaly Detection System From X-ray Scans For Aluminum Conductor Composite Core Wire

The Aluminum Conductor Composite Core(ACCC)is a new type of transmission lines,which has the characteristics of light weight and large carrying capacity of electric.It has been widely used.However,it is not resistant to bending,and is possible to be damaged while installing,which leaving a huge safety hazard.The damaged part often appears in the core area and is difficult to be found by manual detection.Therefore,the research on the anomaly detection system of ACCC is of great significance in industry.In the previous work,a relatively complete anomaly detection system has been developed.But the system currently has some shortcomings,such as the detection performance of algorithm.On the basis,this paper researches the defect detection algorithm,and proposes some improvement strategies.The experiments show that the detection accuracy and recall rate of this scheme have been greatly improved.At the same time,the stability and functionality of the software system are improved.The training module is developed and the defect detection algorithm is integrated as well.The work of this paper is mainly carried out from the following three aspects:1.The defect detection algorithms based on anomaly detection is studied.The image reconstruction method only uses normal samples for training,and the model can reconstruct normal samples well by learning the data distribution,while the reconstruction result of abnormal samples is bad.In the testing phase,the difference between the original image and the reconstructed image is used to determine whether the image contains defects.This paper compares the detection performance of three types of algorithms: object detection,classification network and image reconstruction through experiments,and adjusts the network data accuracy and input image size.2.Two optimization strategies of the defect detection algorithm is studied and proposed.First,the advantages and disadvantages of the Skip-connection structure in the image reconstruction task are considered,and a special analysis is made for the scene in this paper;then,the Skip-connection structure of the original network is adjusted and experiments are carried out.The experimental results show that the optimization method greatly improves the detection performanc;finally,the classification network is researched,and the task of the model is changed from learning the difference between the original image and the reconstructed image to learning whether the subtract image contains defect,and compared through experiments of Siamese network,Resnet50 and SVM.3.The stability of the software system was tested and problems were fixed,and a training module was developed in terms of function to facilitate the optimization of the model after collecting new data.At the same time,the integration method of the image reconstruction algorithm was designed.