Research Of Anomaly Detection Algorithm Based On Variational Autoencoder And Recurrent Reconstruction Strategy

The purpose of anomaly detection is to find few and different data.Due to the problems of unknown abnormal features and significant imbalance between normal and abnormal samples in anomaly detection,the current mainstream method is to train the model with normal data in an unsupervised manner and detect anomalies based on reconstruction error.There are two main problems in this kind of methods: firstly,most methods directly use the generative model to model the normal data,and do not consider the strong generalization ability of the generative model may affect the reconstruction ability of the model for abnormal data,and then lead to detection errors.Secondly,the training and detection strategy based on single reconstruction error cannot extract more comprehensive features of the data,and the anomaly score is not enough to correctly judge whether some data is normal or abnormal.This thesis proposes to add constraint network and constraint rules to Variational Auto Encoder structure.To limit the reconstruction ability of the model for abnormal data,a constraint matrix is added to the latent space of the Variational Auto Encoder.In the training phase,the matrix learns the features and representation of normal data.In the detection phase,a new latent vector based on the linear representation of the eigenvectors in the matrix is obtained by calculating the similarity between the vectors sampled from the latent distribution and the vectors in the matrix.The constraint network can maintain the ability of the model to reconstruct the normal data,at the same time,it can limit the ability of the model to reconstruct the abnormal data to a certain extent,and finally make the model produce more reasonable anomaly scores.This thesis proposes a training and detection strategy based on the idea of recurrent reconstruction.In the training phase,the strategy does not train the model only through the reconstruction samples generated by one forward propagation of the model,but takes the reconstruction samples generated by the previous model as the original samples for the next training,so as to generate new reconstruction samples and train the model iteratively.In the detection phase,the reconstruction error sequence of the test sample is obtained by iteratively inputting the test sample into the model for several times,and finally the classification result of the sample is obtained from the reconstructed error sequence by forward differential or otherwise.The recurrent strategy can not only augment original training data set,but also make the model extract more complex features of the data.The improved model based on Variational Auto Encoder is tested on different types of datasets.From the experimental results,compared with the baseline model and other models,the accuracy of the improved model is improved with AUC as the evaluation metric,and the ablation experiment shows that the constraint network and corresponding rules can indeed improve the accuracy of the model.For the training and detection strategy of recurrent reconstruction proposed in this thesis,the strategy is implemented on the model proposed in this thesis and other models,and the data set of the original model experiment is used.Compared with the results of the original strategy,the AUC value of the model using the recurrent reconstruction strategy has increased compared with the original model.