MotionSync: Personal Energy Analytics through Motion Tags and Wearable Sensing

Individuals with the knowledge of electrical energy consumed by them can take steps to reduce their energy footprint, which can lead to energy conservation. Current energy apportionment schemes either require prohibitively large amounts of user-appliance training or performs poorly in detecting user-appliance interaction when there are multiple users and appliances in close proximity to each other. In this paper, we build MotionSync, a privacy-aware, scalable and robust personal energy analytics system. The system exploits the similarity between the motions of user's arm/hand (captured through wrist-worn wearable) and appliance interface (captured through a motion tag) to determine user-appliance interaction. We show that commonly used plugload devices can be classified in five categories based on their interfaces - button, door, free-floating, knob and switch. Based on this, it is possible to train a generic machine learning model for each category to detect user-appliance interaction with a significant lower training overhead. MotionSync is privacy-aware and allows users to measure their own energy consumption without sharing any private information with building infrastructure. MotionSync is also robust to crowded scenarios since it does not depend on user's location. We implement and evaluate our system on a real testbed, and find that it can determine user-appliance interaction with an average accuracy of 92.5% and has low average false positive rate of 8.6%. We also show that pre-trained models of five interface types provide very high accuracy even for new and untrained appliances and users, eliminating per-appliance and per-user training overhead.