Anomaly Detection Of Time Series Based On Similarity Metrics

Time series data analysis has been a hot field in recent years,but there are still few studies on anomaly detection of multi-dimensional time series data,and outliers will have a certain impact on the analysis of results.Therefore,this paper not only improves the anomaly detection algorithm of univariate time series data,but also studies the anomaly detection problem of multivariate time series data,and proposes three detection algorithms:To solve the problem that DTW distance may map multiple points in one time series to one point in another time series,this paper proposes an embedded technology that deals with similarity and clustering methods simultaneously to detect a group of outliers in an unsupervised way.The task of outlier detection is redescribed as a weighted clustering problem based on entropy and adaptive cost dynamic time warping distance.Through the entropy function and adaptive cost combined with dynamic time warping distance,and then optimize the new cost function to solve the problem of time series data anomaly detection,the improved anomaly detection algorithm can effectively solve the problems under the DTW distance increase at the same time,the improved method has higher accuracy than some existing methods.For the general shape of time series data,rather than a point to point function comparison,we define a kind of dynamic time warping distance based on adaptive cost derivative,and combined with k type median algorithm of multivariate time series data is used to detect the outliers,by comparing the correlation factor model to determine the effectiveness of the model,simulation results show that,this algorithm can recognize the time series of general shape effectively.In order to solve the problem of excessive complexity when using the dynamic time warping distance of adaptive cost derivative,based on the time series segmentation technology,the total extremum algorithm is used to identify extreme values in the time series,and the strict left extremum and right extremum points are retained,while all the flat extremum points are abandoned.The retained extremum points were taken as the change points to extract segments from the time series,and the Leader algorithm was used to cluster the subsequences during clustering,and the lower bound function was used to accelerate the calculation of the dynamic time warping distance of the derivative of adaptation cost.The simulation experiment proved that the algorithm could reduce the complexity of anomaly detection in time series.By comparing the experimental results of the three test methods,it is shown that the method proposed in this paper can detect anomalies more effectively in time series data sets.