Comparative Analysis of Noisy Time Series Clustering

A comparative analysis of the clustering of sample time series was performed. The clustering sample contained time series of various types, among which atypical objects were present. In the numerical experiment, white noise with different variance was added to the time series. Clustering was performed by k-means and DBSCAN methods using various similarity functions of time series. The values of the quality functionals were quantitative measures of the quality of clustering. The best results were shown by the DBSCAN method using the Euclidean metric with a Complexity Invariant Distance. The method allows to separate a cluster with atypical series at different levels of additive noise. The results of the clustering of real time series confirmed the applicability of the DBSCAN method for detecting anomaly.