Improving Water Consumption Estimates from a Bottle-Attachable Sensor Using Heuristic Fusion

Abstract

This paper demonstrates a strategy for improving aggregate (i.e.: multiple drink) water consumption estimates obtained from a bottle-attachable IMU sensor through heuristic fusion. Aggregate consumption is estimated based upon residual container volume using a Gaussian process regression model trained on over 1,500 drinks. The model estimates the fill level of the bottle using hand-engineered features describing the estimated inclination during drinking. Fill level estimates are fused with an empirically parameterized heuristic consumption model. For initial proof-of-concept, fusion is performed using complementary and Kalman filtering. Both techniques are evaluated for 32 dedicated testing experiments containing 12 drinks each. Root mean square fill ratio estimation errors are reduced by 17.3% and 39.6% versus raw sensor estimates using the complementary and Kalman fusion frameworks, respectively.