Research And Implementation Of Time Series Anomaly Detection Algorithm Based On Isolation Score

Anomaly detection is an important research field in machine learning and data mining,which has a great impact on medicine,aerospace,gesture recognition and other fields.In terms of time series anomaly detection,aiming at problems such as high time complexity of the Matrix Profile and unclear definition of boundary between anomaly and normal sequence of global anomaly detection algorithm,This thesis proposes an optimized time series anomaly detection algorithm CR-iNNE based on isolation score iNNE point anomaly detection algorithm and MASS algorithm based on similarity measurement.The main work is as follows:(1)The point anomaly detection algorithm iNNE based on isolation score is studied.Considering its advantages such as providing more sensitive isolation measurement and selecting only part of samples to build models,the idea of applying the point anomaly detection algorithm in the direction of time series anomaly detection is proposed.The MASS_V2 algorithm developed based on sliding window is analyzed.Considering the advantage that only half of the convolution is used as the input matrix,the MASS algorithm based on sliding window is combined with the point anomaly detection algorithm,and it is demonstrated that the time series anomaly detection algorithm iNNE based on isolation scores can detect time series anomalies.(2)Considering that random selection in the point anomaly detection algorithm does not conform to the characteristics of time series data with time correlation,an algorithm of random selection of time sequence subsamples is proposed to solve the redundancy problem of taking the whole time series distance matrix as input.Based on the iNNE algorithm,the hypersphere model is established,and the storage of unnecessary distance matrix of time subsequence in the algorithm is improved to reduce the number of comparisons effectively.For CR-iNNE algorithm evaluation model,the problem of which hypersphere the selected query subsequence falls on is solved,so that the nearest neighbor distance is compared with the radius of the hypersphere to reduce the time complexity.(3)Gaussian white noise is added to the synthesized data to improve the generalization ability of the algorithm.The method of adding noise can make the algorithm more robust and better cope with the new data.The optimized time series anomaly detection algorithm based on isolation scores is applied to real data to demonstrate the performance of the classifier,and the operational efficiency of the algorithm before and after optimization is compared and evaluated.The experimental results show that the accuracy of the improved time series anomaly detection algorithm CR-iNNE based on isolation scores is equal to or even better than the classical Matrix Profile algorithm.When using real time series data for anomaly detection,the accuracy of iNNE algorithm is 97.38%,which is 1.54%higher than that of Matrix Profile algorithm.Compared with the original iNNE algorithm,the improved CR-iNNE algorithm has better operation efficiency,and the time is shortened by 2.7274 seconds at most.Moreover,the improved CR-iNNE algorithm can detect and classify anomalies accurately on the time series data in the fields of medicine and aerospace engineering.