Anomaly Detection and Feature Alignment for Time Series Data

Time series data are stemming from various applications that describe certain observations or quantities of interest over time. Their analysis typically involves the comparison (with reference data for anomaly detection) and feature alignment across different time series data sequences. General technique for anomaly detection via visualization is to compare a live signal along with reference sequences. Currently, the standard methods used in the industry are line/scatter plots. Due to limitations such as cluttering, lack of quantitative information etc., these plots are not effective. In this thesis, probabilistic envelope based technique is proposed for the visualization and anomaly detection of time series data. This technique provides quantitative information, is able to avoid the outliers in the reference data, and works well even with a large number of reference sequences. To demonstrate the practical use of the probabilistic envelope technique, it is applied in detection of over/under gauge of bore holes (wells). The implementation of gauge detection along with some results are also presented in this thesis. For feature alignment, the Dynamic Time Warping (DTW) is the standard approach to achieve an optimal alignment between two temporal signals. There are different variations of DTW proposed to address different needs of signal alignment or classifications. However, there lacks a comprehensive evaluation of their performance in these time series data processing tasks. Most DTW metrics are reported with good performance on certain types of time series data without a clear explanation of this performance. To address that, a synthesis framework is proposed to model the variation between two time series data sequences for comparison. The synthesis framework can produce realistic initial signal and deform it with controllable variation that mimics the real-world scenarios. With this synthesis framework, a large number of time series pairs with different but known variations can be produced, which are used to assess the performance of a number of well-known DTW measure in the tasks of alignment and classification. Their performance on different types of variations are reported and the proper DTW measure is suggested based on the type of variations between two time series sequences. This is the first time such a guideline for selecting proper DTW measure is presented.