Research On Defect Detection Of Aluminum Conductor Composite Core Wires Based On Semi-supervised Anomaly Detection

Aluminum Conductor Composite Core(ACCC)wire is a new type of Conductor,which has been widely used in China.However,there may be defects in ACCC,which will lead to wire breakage and critical line safety due to the influence of conductor characteristics,construction operation level,construction management and other factors.On the other hand,due to the long length of wire,it is not practical to rely on manual defect detection,so it is necessary to study the automatic defect detection method of ACCC.Existing automatic defect detection methods in the industry are mostly based on supervised deep learning,which rely on a large number of training data.However,in the case of defect detection of ACCC,it is difficult to obtain defect samples,so supervised algorithms are not applicable.Based on the above reason,this paper proposed an automatic defect detection scheme based on semi-supervised anomaly detection.In this detection scheme,the training of the deep model can be completed only with easy-to-obtain non-defect samples and a very small number of defect samples.The experiment shows that the accuracy rate of this scheme reaches96.86% and the recall rate reaches 90.00%.The accuracy rate reached 91.99 %.The work of this paper is carried out in the following three aspects:(1)This paper first uses X-ray to image the interior of carbon fiber wires.Due to the large size and large area of background parts in the collected image,the original image should be preprocessed.Preprocessing mainly includes three steps: Firstly,straightening the wires.Secondly,cutting background parts in the image.Finally,cutting out small patches from original images used for model training and testing images.(2)A semi-supervised defect detection algorithm based on image reconstruction is proposed and compared with the existing supervised defect detection algorithm.The proposed semi-supervised detection algorithm uses only normal samples to train an image generator,and then uses the error between the reconstructed image and the original image to judge whether there are defects in the image.Experiments show that the semi-supervised algorithm can detect defects without using defect training samples,and the detection effect of the new defects is better than that of the supervised algorithm.(3)The existing semi-supervised detection algorithm is optimized in this paper.Firstly,the image reconstruction quality is optimized by using hierarchical reconstruction and generator cascade.While ensuring that the normal part of the image is intact,the defect information in the image is removed as far as possible.Secondly,the siamese network is used instead of the original classifier.Experimental results show that the optimization measures proposed in this paper are effective.