A Custom Unsupervised Approach for Pipe-Freeze Online Anomaly Detection

• Published in: 2021 IEEE 7th World Forum on Internet of Things (WF-IoT) pp. 663 - 668
• Main Authors: Soliman, Ahmed, Rajasekaran, Sanguthevar, Toman, Patrick, Ravishanker, Nalini, Lally, Nathan, D'Addeo, Hunter
• Format: Conference Proceeding
• Language: English
• Published: IEEE 14.06.2021
• Subjects: Anomaly Detection (AD); Buildings; Companies; Forestry; Isolation Forest (IF); Machine learning (ML); Temperature distribution; Temperature measurement; Temperature sensors; Time Series (TS); Wireless sensor networks
• DOI: 10.1109/WF-IoT51360.2021.9595720

Currently, many companies are leveraging the advancement in IoT technologies by incorporating more Wireless Sensor Networks (WSNs)-based solutions in their services. WSNs are becoming cheaper and can be easily deployed in multiple locations to cover very wide geographic areas. These sensors are often an essential part of monitoring systems. Several applications exist for these systems. One application is to monitor temperatures inside buildings that are susceptible to pipe-freeze hazard. Hartford Steam Boiler (HSB) is one such company that uses temperature sensor nodes to monitor temperatures for thousands of institutional buildings across the United States to prevent pipe-freeze losses to its customers. For this to happen, it is essential to automate the prediction/detection of anomalies in the temperature sensor readings. In this paper, we introduce two generic techniques, namely, data thresholding and distance based filtering, that can be used to improve the performance of any anomaly detection algorithm. To illustrate the effectiveness of these techniques, we present a customized hybrid approach based on isolation forest (IF) and these two techniques. A set of hyperparameters has been proposed and carefully tuned to condition the data and extract only relevant information in a way that helps HSB detect anomalies, predict pipe-freeze, notify its customers, and hence prevent pipe-freeze related losses. Experiments carried out on real temperature sensors' data show the efficacy of our proposed techniques and algorithm in terms of achieving the business goals of HSB.