Research On Time Series Anomaly Detection Method Based On Reconstructio

As science and technology are developed rapidly,there is an explosive growth of massive data.These data are essential for various industries.What’s more,more data exist in the form of time series.Time series have great value.It can be used to predict future tendencies.And anomalies in time series can be detected to find the problems and avoid bringing losses.Thus,time series anomaly detection attains more widespread attention.Among deep anomaly detection(DAD)methods,autoencoder and generative adversarial networks as the basis for reconstruction-based methods have been proven to be well suited for time series anomaly detection tasks.However,due to the increasing complexity of time series in terms of dimensionality,heterogeneity,and the presence of many noises in the real environment,as well as the highly complex temporal and spatial correlations.Many anomaly detection methods do not take these problems into account,resulting in low accuracy.This thesis is a study of reconstruction-based time series anomaly detection.The previous methods are optimized to address the above issues.The main innovative works are as follows:(1)In this thesis,it proposes an unsupervised time series anomaly detection method based on re-encoding.Firstly,a stacked LSTM-dropout RNN(Stacked Long Short-Term Memory Recurrent-dropout Neural Network)is designed as the basic framework of the generative adversarial network(Generative Adversarial Networks,GAN).The stacked network can convey neighbor cells’ information to capture long temporal correlation.Secondly,multiple generators with cycle-consistency loss and Wasserstein-loss are used to deal with the model collapse problem.Finally,recoding is proposed,using two encoders to obtain two potential spatial differences,because the anomalous time series will be smoothed by normal samples in the high-dimensional space,and the smoothing problem can be solved by feature extraction of the time series through dimensionality reduction,thereby amplifying the anomaly.And reencoding loss is served as a part of the anomaly score,and it further improves the performance of the model.(2)This thesis also proposes a Transformer generation adversarial anomaly detection method based on multihead dynamic graph attention.Firstly,a multihead dynamic attention(Multihead Dynamic Attention,MDA)mechanism is designed in the graph attention network to capture the dependencies between time series samples in both temporal and spatial dimensions simultaneously.And MDA also can dynamically allocate attention weights.Secondly,Transformer with contextual self-attention is proposed in the adversarial network.Punishment from the discriminator can be used as a regularization method.And the generator learns the temporal and spatial correlations of the graph attention network and the original data to obtain more comprehensive features,which avoids the overfitting problem.The contextual self-attention mechanism can match more relevant features to generate more real time series.Finally,the MLP(Multilayer Perceptron,MLP)is served as the prediction-based method to detect anomalies by sudden perturbations in time series.The final experimental results show that the proposed method is far superior to the latest baseline model.In this thesis,we use several real-world datasets to validate the accuracy of the two proposed anomaly detection methods.What’s more,ablation experiments are also implemented to verify the performance of each module for models.The final experiments show that both methods obtain high accuracy,and each of them has its advantages.