Shapelet based clustering and anomaly detection for compromised, noisy and missing multivariate time series

Abstract

Multivariate time series (MTS) data are widely prevalent in various domains, including medical screening, system monitoring or astronomy, where each instance consists of multiple sequences with inherent temporal ordering. Detecting anomalies in MTS is a critical research area aimed at identifying time points or patterns that deviate from normal behaviour. The demand for not only detecting anomalies but also processing them representatively, has led to the development of shapelet-based anomaly detection methods, which offer both interpretability and accuracy for MTS analysis. Despite the increasing interest in anomaly detection methods, shapeletbased anomaly detection remains a relatively small research field. This thesis introduces a novel workflow for unsupervised shapelet-based anomaly detection and anomaly prototype identification in MTS data, combining established anomaly detection methods with a shapelet-based classification framework. To evaluate different shapelet detection techniques based on clustering and validate the effectiveness of the workflow, experiments were conducted on both synthetic and real-world telemetry datasets of an exploration greenhouse. It was found, that the proposed workflow successfully detected diverse anomaly types while demonstrating the shapelets interpretability, especially for the real-world case. Moreover, the potential to enhance anomaly detection appears to rely on the adapted weighting of cluster size and distance during shapelet selection and the issue of vanishing anomalies due to the Euclidean distance used for clustering.