Research On Key Technologies Of Time Series Analysis Based On Deep Learning

Time series is critical and ubiquitous data.With the rise of Internet of Things(Io T)and Artificial Intelligence for IT Operations(AIOps),more and more time series data are collected and analyzed.By applying proper technologies,meaningful and useful information can be retrieved from these data,e.g.,whether the target being monitored is anomalous,so as to make effective use of these data.Among them,there are three critical and basic technologies,which are time series generation,time series correlation detection,and time series anomaly detection,respectively.Firstly,the manual labeling of time series data is a labor-intensive task,to label the data,experts with domain knowledge are required.Time series anomaly detection algorithms require a lot of work to collect/mark time series data in order to test them in detail before real-world deployment,labeled time series data is very important and urgently needed.However,for privacy and security concerns,as well as the commercial value of time series data,there are few publicly available data for testing anomaly detection anomlay detection algorithms.Therefore,how to generate data that can satisfy the requirement of stress-test or be used for quantitative evaluation of time series-associated algorithms is a challenging but urgent task.Such tools will play a supporting role in the development of time series analysis techniques.Secondly,traditional statistical correlation analysis methods can hardly meet the new challenges brought by the development of various domains,especially for the state correlation in time series.As a kind of streaming data,time series can be used to monitor the physical or logical state of the target.Target in different states will produce time series data with different characteristics.Therefore,time series can also be divided into different states.These states can carry a lot of high-level semantic information.Studying the state correlation is helpful for understanding the relationship between these states.By far,a lot of work has focused on the analysis of traditional statistical correlation,but time series state correlation analysis has not attracted enough attention.Finally,time series anomaly detection plays a key and supporting role in various fields.In particular,large Internet and financial companies need to closely monitor the real-time performance of their systems because a short interruption of the network or degradation of quality may lead to huge business losses.To keep smooth businesses,it is very important for these companies to develop time series anomaly detection systems.Although a large number of anomaly detection algorithms have been proposed in recent years,it is believed that there is not a single algorithm that can be competent for all types of time series data.Selecting the appropriate algorithm for different types of data still consumes a lot of time.It is of great practical significance to effectively ensemble or selecting proper detectors for time series data with different characteristics in an automated way.This paper starts with the review of research status of the above three tasks,and based on the analysis of the defects and deficiencies of the existing technologies,we propose new solutions,and evaluate the effectiveness of the proposed methods through extensive experiments.In summary,the novelty and contribution of this paper are three-fold:(1)We porpose a time series generation method,TSAGen,which can can be used to evaluate time series anomaly detection algorithms.In particular,it is suitable for evaluating the invariance of algorithms and stress testing the algorithm.Experimental results show that effectiveness of the proposed method;(2)In this paper,we porpose and formulate a kind of time series semantic correlation problem – state correlation problem.For the porposed problem,we propose a state correlationn detection framework,State Corr.Under this framework,we porpose novel time series segmentation algorithm and metrics for measuring state correlation.Experimental results on real data show the effectiveness and superiority of the proposed framework;(3)We propose an ensemble framework for time series anomaly detection,TSAEns,which can automatically select and ensemble base detectors according to their performance on historical data,so as to reduce the time cost of manually selecting and improve the accuracy of detection result.Extensive experiments show the effectiveness of the proposed method.