Semi-supervised training for infrastructure mediated sensing: disaggregated hot and cold water sensing with minimal calibration

Abstract

This thesis formalizes and evaluates the use a central sensor for detecting and classifying water usage in a home or apartment building. While the discussion largely explores the theory and implementation of the sensing, the main impact of this thesis is to address a single assertion: we are running out of water. With this sensing technology, my aim is to support initiatives in increasing awareness of water consumption and understanding of the diversity in ways that vast populations consume water. Through my sensing, I classify (1) what type of valve was activated in a home (i.e., faucet, dishwasher, toilet), (2) the temperature of water that the valve uses (i.e., hot, cold, or mixed), and (3) where in the home or apartment the valve resides (i.e., kitchen, master bathroom). This is typically known as disaggregated water sensing or end-use water sensing. In my approach, I use a central sensing point based on two pressure sensors. The sensors monitor pressure fluctuations in the home, providing a time series upon which to classify events. I formalize the feature extraction process using template matching, expertly chosen features, and sparse codebooks. Ultimately I show that few features can be created that generalize across residences, necessitating that each residence go through a calibration process. I formalize this calibration process using semi-supervised techniques such as expert knowledge through a rule-based classifier, co-training, and virtual evidence. I evaluate the system in several studies, including a longitudinal deployment of 15,000 water usages across a 1 month period in five homes. The results show that, with about 30-40 labeled examples, the system is 93% accurate at detecting the type of fixture used, 89% accurate at detecting which fixture in the home is activated, and 81% accurate at also detecting the temperature state of the valve. These accuracies are reported on a test set of 2,500-4,000 water usage events per residence. This thesis, then, has implications in the water sustainability community, activity detection community, and the machine learning and signal processing communities.