Using Water Flow and Pressure Data for Activity Recognition in Smart Homes

Abstract:
In this study, we investigate the potential of a smart water metering system for human activity recognition (HAR) based on water usage patterns. The data, collected from eight households in the Czech Republic using Resideo smart water meters, includes flow, pressure, and temperature measurements. We implemented a machine learning pipeline to classify seven categories of water-related activities, using Random Forest as the primary classifier. Initial experiments achieved an overall accuracy of 93% on complete traces, with the ’tooth brushing’ category posing the greatest challenge (58% accuracy). To improve generalizability, we introduced a second stage focusing on segmented data, reducing the accuracy to 79 %. Finally, consolidating certain activity classes improved the classification accuracy to 84 %. Our findings suggest that further refinement of ’manual’ activities remains a challenge, which we discuss in the conclusion along with potential future directions.