Research On Time Series Anomaly Detection Method Based On Auto Encoder

In recent years,time series anomaly detection has received more and more attention.Long-Short Term Memory Network(LSTM)can extract contextual information in sequence data and it is always used to process time series data.LSTM Auto Encoder(LSTM-AE)does well in detecting anomalies on single-dimensional time series.But there is a complex coupling relationship between the various dimensions of Multivariate Time Series(MTS),and it is the main reason why LSTM-AE can’t detect the anomaly samples well.In addition,most of the semi-supervised anomaly detection models based on LSTM Auto Encoders only use normal samples to train the models.And the generalization ability of the these models is poor.Focusing on the above issues,the specific work as follows:(1)An ensemble network based on LSTM Auto Encoder(LAE)is proposed.LAE integrates multiple LSTM-AEs to reconstructs subsequences of the normal MTS sample.LAE treats the reconstruction error of each LSTM-AE as the local feature of the MTS;The Fully Connected Network Auto Encoder(FCAE)is used to reconstruct the local features and performs anomaly detection based on the reconstruction error of FCAE.Experiments on several public datasets show that compared with the benchmark method,the accuracy,recall and F1 value of the LAE model are significantly improved.(2)An Iterative Self-learning Time Series Anomaly Detection Model Based on LSTM Auto Encoder(ISLAE)is proposed.ISLAE consists of an Initial Anomaly Detection Module and a Self-learning Module.The Initial Anomaly Detection Module obtains the feature representation of the training set samples by the LSTM-AE,and gets the anomaly metric value of each feature representation by OC-SVM and i Forest.According to the anomaly metric value,a sub-training set with pseudo labels is generated.The Self-learning Module is composed of the encoding network of the LSTM-AE in the Initial Anomaly Detection Module and the Fully Connected Classification Network.Self-learning Module’s training process:(1)ISLAE uses the Self-learning Module to classify the sub-training set samples,and updates the parameters of the self-learning module once according to the classification error;(2)The Self-learning Module is used to calculate the anomaly scores of all training set samples.According to the anomaly scores,a new sub-training set with pseudo-labels is generated.The process of(1)-(2)is performed iteratively until the Self-learning Module is convergent.Experiments on multiple datasets prove that compared with the benchmark method,the F1 value of ISLAE is significantly improved.