Research On Anomaly Detection Algorithm Based On Time Series Reconstruction

A time series is usually a set of univariate or multivariate data collected over equal time intervals for an indicator to be measured,reflecting the development law of an in-ternal pattern of the indicator to be measured.The task of time series anomaly detection focuses on identifying the data in the time series that does not conform to the normal trend.With the rapid development of the Internet of Things(Io T),the time series data collected by various sensors show complex trends such as dimensional explosion,heterogeneity and non-linearity.Traditional solutions are unable to effectively cope with the new chal-lenges and better capture the important information between the data from the temporal and spatial scales.In recent years,deep learning has been extensively researched and have achieved good anomaly detection results.At the same time,many excellent network ar-chitectures have been derived,among which the deep aotuencoder is considered as one of the effective representative models for time series anomaly detection tasks.The original sequence data is first encoded and decoded to obtain the reconstructed sequence,then a reasonable threshold or probability distribution scheme is formulated to analyse the error value between the two and finally abnormal information is found.This paper investigates the task of time series anomaly detection from both super-vised and unsupervised perspectives,based on deep learning solutions and using recon-struction ideas.The main work is as follows:(1)Traditional anomaly detection solutions have difficulty handling high-dimensional data and ignore sequential relationships and local information between data,which often leads to misjudgment and missed judgments in the detection process.To address this problem,this paper proposes a supervised solution for time series anomaly detection us-ing a combination of deep autoencoder(DAE)and a traditional classifier such as support vector machine(SVM).The deep autoencoder is built by a convolutional neural network(CNN)and a long short-term memory(LSTM)network.The encoding process extracts distinguishing features between normal and abnormal samples and establishes the relation between the features and the actual labels,which can be used to distinguish anomalies by traditional classifiers.This solution can not only better capture the local spatial and tem-poral features of the time series,but also overcomes the uncertainty problem faced by threshold screening anomaly detection solution.(2)Since data in real applications often lack effective anomaly labels,this paper proposes an unsupervised time series anomaly detection method using variational au-toencoder(VAE)combined with dynamic threshold.VAE is built by the bidirectional long short-term memory network(Bi LSTM),which is used to capture the forward and backward temporal dependencies of data.VAE generates reconstructed data close to the original data based on probability distributions and measures the difference between the original and reconstructed samples by the Mahalanobis distance.To avoid the difficulty of accurately capturing contextual intervals with artificially set fixed threshold,dynamic threshold without parameters can be used to flexibly discriminate anomalies.In this paper,the results of comparative experiments on several datasets show that the two time series anomaly detection solutions proposed both show good results.