Anomaly Detection Method And Its Application Based On Auto-Encoder Network

The goal of anomaly detection is to find objects different from the normal mode,which is a fundamental problem in many fields.The anomaly detection technology is widely used in industrial production and daily life.In the era of big data,anomaly detection mainly faces three difficulties: unbalanced samples,lack of labeling of data,and abnormal diversification.Anomaly detection based on unsupervised learning is one of the mainstream technologies most concerned by current researchers.This thesis studies the anomaly detection method based on auto-encoder network and its application in the scene identification of high-speed rail fasteners.The main work and innovations of the thesis are as follows:(1)An anomaly detection model LRGAN(Locality-preferred Recoding Generative Adversarial Network)based on locality-preferred encoding is proposed.In-depth research and analysis of the nature of the reconstructed samples based on the auto-encoder network anomaly detection method,found the locality of the latent vectors distribution in the middle layer,inspired the research ideas of the LRGAN model.The locality-preferred encoding module in LRGAN compulsorily uses the latent vectors of normal samples to represent abnormal samples,so that the samples reconstructed by the abnormal samples will be biased towards normal samples,and the difference in reconstruction will increase.Moreover,based on LRGAN,LRGAN+ is further proposed based on the two strategies of clustering and adaptive threshold distance.It uses clustering to obtain a more compact codebook and adaptive threshold distance to automatically determine whether to reencode the latent vector,and further improves the performance of the anomaly detection model on the basis of LRGAN.Based on two public datasets and two evaluation protocols,the LRGAN and LRGAN+ models proposed in this thesis are verified.The experimental results show that LRGAN can achieve the detection effect equivalent to the current best method,and the anomaly detection effect obtained by LRGAN+ is much better than the leading anomaly detection model.(2)An anomaly detection method for fasteners based on auto-encoder network is proposed.This method integrates an auto-encoder network and an adversarial generation network,and constructs a neural network model for fastener abnormality detection;in particular,this method uses a locality-preferred encoding module in the prediction process.This post-processing strategy not only reduces the complexity of model training and enhances the flexibility of the model,but also achieves detection performance comparable to LRGAN+.