Research On Multi-scale Representation And Anomaly Detection Of Time Series Based On Correlation Analysis

As an important data type with time attribute,time series data exists in many fields.However,there are a lot of important information in these sequence data.Therefore,it is a hot issue in the field of data mining how to analyze the data from a scientific perspective so as to dig out more valuable information behind the data.As a dynamic new research field,anomaly detection of time series data aims to find samples that are significantly different from other data from massive,changeable and high-dimensional time series,which is of great scientific value for the analysis and processing of time series.Based on the concept of scale space,starting from the analysis of the correlation between the features representing time series,this thesis proposes a time series multi-scale representation method of key features,and builds a multi-scale anomaly detection framework based on the variable-order Markov model on the basis of the multi-scale representation.The main contents are as follows:Due to the time series Representation can only be a single Scale focused On a Feature,easily lead to leak and mistakenly identified phenomenon.At the same time,the correlation between the time series representation features will lead to the problem that the feature interpretation is not true and the accuracy of the representation model will be reduced.In this thesis,FCA-MSR(Multi-Scale Representation Based On Feature Correlation Analysis)is proposed.For this reason,a simulation experiment is designed.Experimental studies on the synthetic data and public data show that this method has higher accuracy and stronger robustness.Compared with single scale PAA(piecewise aggregation approximation),the optimal interval representation method and the CGMV-MSR(Multi-Scale Representation of key feature sets based on Conditional Generalized Minimization Variance),the method in F1 score increased by 127.57%,34.63%and 16.20%,respectively,and greatly improves the accuracy and recall of anomaly detection.In addition,due to the short memory characteristics of classical Markov model methods ignore the connections between data,the long memory characteristics of the higher order Markov model methods and makes the correlation between historical data and current data are fuzzy,in turn,led to the lower reliability model,this thesis from the perspective of memory length balance model was proposed based on variable order Markov model of time sequence of anomaly detection method,and introduce abnormal replacement strategy to avoid the data to be detected affected by abnormal data points have been detected.Experiments show that the anomaly detection method based on variable-order Markov model proposed in this thesis improves F1 score by 60.34%and 30.98%respectively compared with classic Markov model-based and high-order Markov model-based methods.Increased by 129.95%and 43.42%respectively,overcoming the limitations of the traditional Markov model-based anomaly detection method,so the anomaly detection effect was improved.