Research On Anomaly Detection Of Multidimensional Time Series

In many real-world applications such as industrial big data analysis,medical health big data analysis,meteorological data analysis and prediction,economic and financial data analysis,time series data are widely used as an important data object.Time series is a data object collected in the order of time development,which has the characteristics of high dimension,large amount of data,fast update speed and so on.As one of the most important tasks in time series analysis,anomaly detection is designed to detect outliers or abnormal sequences in time series.Effective processing of context information of multidimensional time series,feature extraction and detection of outliers,effective detection of multi-dimensional abnormal sequence is the key problem.The main contents of this paper include the following two parts:First,aiming at the problem of point anomaly detection in multidimensional time series,this paper presents a point anomaly detection algorithm based on bidirectional LSTM heterogeneous integration(MTSPAD).By processing the context information of bidirectional LSTM multi dimension time series,one can sequence characteristic of automatic extraction using neural network;on the other hand,the relationship between the dimensions of the sequence can be automatically processed,then the input time series in each sub model in heterogeneous learning,get the optimal model of heterogeneous integration,the anomaly detection more robust,finally using heterogeneous integrated model for anomaly detection of input data.Through experiments on six real datasets,the algorithm can detect anomalies more accurately.Compared with single learner,the accuracy and recall rate of anomaly detection is increased by 5% compared with other algorithms.The overall implementation results show that this method can detect the anomalies in the multidimensional time series accurately and effectively.Second,for another important detection task of multidimensional time series anomaly detection,sequential anomaly detection,this paper proposes a sequential outlier detection algorithm MTSSAD based on dynamic time warping(DTW).It is proposed that DDTW distance is used as the criterion of sequence similarity measure.The first to use KernelPCA dimensionality reduction algorithm in the normalized training set to learn the conversion matrix,then the object test data set into single dimension time series(UTS)data set,the predicted labels through calculation and normal sequence DDTW distance to get the target sequence,so as to judge whether the target sequence for abnormal sequence.The effectiveness of the algorithm for abnormal sequence detection is verified by experiments on remote pregnant women's fetal heart monitoring data sets.This paper aims at exploration of anomaly detection in multi dimension time series anomaly detection field point anomaly detection and sequence,study the processing method of multi dimension time series data and anomaly detection method,good experimental results show the effectiveness and accuracy of this method.The research work has some guiding significance and practical value for the problem of abnormal detection of multidimensional time series data.