Unsupervised Anomaly Detection Using Light Switch Information in Smart Homes

• Main Authors: Chih-Wei Ho, 何致緯
• Other Authors: Chun-Ting Chou
• Format: Others
• Language: en_US
• Published: 2015
• Online Access: http://ndltd.ncl.edu.tw/handle/23742771750933813790

碩士 === 國立臺灣大學 === 電信工程學研究所 === 103 === Anomaly detection in smart homes has become one of the most important issues recently with the emergence of Internet of Things(IoT). By analyzing the enormous amount of data collected via IoT, anomaly detection techniques are able to detect anomalous behavior such as unusual activities, unconsciousness, falling down, health vitals, unauthorized intrusion, etc. However, current anomaly detection in smart homes has two issues that are not well addressed. First, most of deployed sensors lead to problems such as privacy violation, uncomfortable wear experiences, and a huge amount of battery replacement. Second, most anomaly detection algorithms adopt supervised or semi-supervised approaches that require users to label data, which is a heavy load especially for the elderly. To solve these problems, an unsupervised anomaly detection algorithm using light switches is proposed. It is an adapted model-based anomaly detection algorithm that can reduce the effect of outliers in training data. By adding constrains to the evaluated mixture model and recursively estimating the decision boundaries, the found decision boundaries are more representative of the normal regions where normal data frequently appear. The false alarm rate can be reduced in a given miss detection rate. Our goal is to find events that occur at unusual time point or last for unusual length and indicate where they occurred and why they are anomalous. To evaluate the proposed algorithm, 11 light switches are installed in an apartment with 4 permanent residents. The data collection is conducted in real life with 24 hours non-stopped and last for over 7-months. The detected anomalous events are compared with the ground truth provided by the residents, and the result shows that 80% anomalous events are detected with 25% false alarm rate. Our method also performs better than the existing unsupervised anomaly detection algorithms while analyzing the events with the requirement of miss detection rate less than 20%.