Research On Abnormal Detection Method Of Hawse Flash Butt Welding Based On Convolutional Adversarial Generation Network

With the rapid development of deep learning and generative models,more and more industrial problems will be solved by deep learning methods.With the development of the shipbuilding industry,the quality inspection of flash welding of anchor chain is the focus of attention in recent years.Since the abnormal state data is difficult to collect,most of the anchor chain samples are normal and qualified samples,so how to conduct anomaly detection through most of the normal data and a small amount of abnormal data becomes the main research problem of this thesis.The research contents of this thesis include:Firstly,aiming at the problem of class imbalance of industrial time series data,this thesis proposes a Deep Convolution generative adversarial network based on Wassterin distance,Wassterin Deep Convolution GAN(WDC GAN),which makes the model learn the correct distribution of input samples through unsupervised learning.At the same time,this thesis compares the data generation performance of Gans with different structures,and verifies that the samples generated by WDC GAN are closest to the original samples by comparing three distance measures: DTW,Fre,and EUC.In addition,compared with the traditional data enhancement algorithm,the mean value of the KNN model is the highest.Compared with the original imbalanced data set,the F1 value is increased by 4.61%,the AUC index is increased by 4%,and the G-mean index is increased by 1.17%.On the SVM model,compared with the original imbalanced dataset,the F1 value is increased by 8.26%,and the AUC index is increased by 7.73%.Secondly,due to the rarity and preciousness of industrial abnormal data.To this end,this chapter aims to establish a Least Squares Anchor Chain GAN(LSAC GAN)based on convolutional structure and distance training.From the semi-supervised perspective,LSAC GAN can learn normal and abnormal samples,so that the model can explicitly avoid generating abnormal samples.Through two public multi-dimensional time series data sets SWa T and SMAP,it is compared with other anomaly detection GAN models.For SWa T,LSAC GAN has a precision of 90.49% and a recall of 70.91%.For SMAP,LSAC GAN achieves 72.15% precision and 81.23% recall.In addition,we study the influence of the number of abnormal data on the performance of LSAC GAN.With the increase of the number of abnormal data,the performance of the model can be steadily improved.At the same time,in order to alleviate the problem of mode collapse in the model,an additional loss term based on DPP(determinant point process)is added to the loss function of the generator to make the samples generated by G have more diversity and avoid only generating samples of one contour distribution.It is proved by sample diversity experiment and sample quality experiment.The LSAC-DPP GAN integrated with GDPP is able to generate high-quality and diverse samples.Finally,for the electrode position(displacement)data and current signal data of flash welding of anchor chain,we train through LSAC-DPP GAN to generate normal displacement data and current signal data.By combining the displacement and current data into the combined data,a relatively stable recall rate,precision rate and F1 value are obtained on the test set.In addition,we combine the improved CAM visualization technology to locate and analyze the abnormal displacement and current signals of the abnormal anchor chain.