SHAPE-BASED TIME SERIES MINING FOR PROCESS MONITORING AND ANOMALY DETECTION
Degree granted by The University of Texas at Arlington
Abstract
Due to the rapid development of computing and sensing technology, Internet of Things (IoT)-enabled monitoring plays a crucial role for people suffering from cardiac problems. It is important to detect the abnormal ECG cycles during the cardiac monitoring for the early treatment. However, most existing methods focused on the full reading of time series, for the cycle-based time series, it is wasting time to read the whole time series while we can find the characteristic patterns instead. Characteristic patterns named shapelets are time series subsequences, which are explainable and discriminative features that can best classify time series. Shapelet-based classification that uses the similarity between a shapelet and a time series has been widely used recently in many applications. In this research, we extract the statistically significant shapelets from the cycle-based ECG data, and apply the support vector data description (SVDD) algorithm to statistical process control problem for the cardiac monitoring. The experimental results on the real-world MIT-BIH dataset demonstrate the effectiveness of proposed method. Positive and unlabeled learning has attracted increasing interest in recent years. The setting of the positive and unlabeled learning is that we only access the positive and unlabeled training data sets. Many methods have been proposed for the positive and unlabeled learning, however, only a few papers integrate the shapelet features into the positive and unlabeled learning. In this paper, we proposed the positive and unlabeled shapelet learning model for the time series classification, and the experiment results from the real-world data sets demonstrate the effectiveness of our proposed method.