A Longitudinal Study of Pressure Sensing to Infer Real-World Water Usage Events in the Home

We present the first longitudinal study of pressure sensing to infer real-world water usage events in the home (e.g., dishwasher, upstairs bathroom sink, downstairs toilet). In order to study the pressure-based approach out in the wild, we deployed a ground truth sensor network for five weeks in three homes and two apartments that directly monitored valve-level water usage by fixtures and appliances. We use this data to, first, demonstrate the practical challenges in constructing water usage activity inference algorithms and, second, to inform the design of a new probabilistic-based classification approach. Inspired by algorithms in speech recognition, our novel Bayesian approach incorporates template matching, a language model, grammar, and prior probabilities. We show that with a single pressure sensor, our probabilistic algorithm can classify real-world water usage at the fixture level with 90% accuracy and at the fixture-category level with 96% accuracy. With two pressure sensors, these accuracies increase to 94% and 98%. Finally, we show how our new approach can be trained with fewer examples than a strict template-matching approach alone.