Exploring distance measures for time series data: A comparative analysis

Lee, Jia Yee (2024) Exploring distance measures for time series data: A comparative analysis. Final Year Project, UTAR.

Preview

PDF Download (1792Kb) | Preview

Abstract

Time series similarity search is a method used to identify the identical pattern within two sets of time series data, finds widespread utility in clustering, anomaly detection, and forecasting. In real-world scenarios, vibration data are often vast, intricate, and noisy, with adjustments in time, amplitude, and phase shifting direct influence on search outcomes. Through a systematic evaluation, various distance measurement methods including Euclidean distance, Dynamic Time Warping, Fast Fourier Transform, Symbolic Aggregate Approximation, and Matrix Profile are performed under diverse conditions such as frequency shifting, amplitude scaling, state change, and noise. The comparative study encompasses not only quantitative assessments of accuracy but also considerations of computational efficiency and robustness. The findings reveal Matrix Profile generally outperforms classic measures like Euclidean distance, Dynamic Time Warping, and Fast Fourier Transform in accuracy, but performs poorly compared to Symbolic Aggregate Approximation. While Matrix Profile exhibits shorter computational time than Symbolic Aggregate Approximation, it slightly extends beyond other classic measures. Thus, Matrix Profile presents competitive advantages among ii distance measurement methodologies. By providing a comprehensive examination of similarity measurement techniques, this study equips the idea for the strength and weaknesses of distance measures, providing valuable insight for decision-making in time series data mining activities.

Actions (login required)