An On-Device Learning System for Estimating Liquid Consumption from Consumer-Grade Water Bottles and Its Evaluation

Area of Science:

• Embedded Systems
• Machine Learning
• Biomedical Engineering

Background:

• Accurate tracking of liquid consumption is crucial for health monitoring.
• Existing methods often lack real-time, on-device capabilities and are energy-intensive.

Purpose of the Study:

• To develop a lightweight, energy-aware, on-device system for estimating liquid consumption.
• To implement a self-correcting machine learning framework for robust hydration tracking.

Main Methods:

• Developed two on-device neural network models for sip detection from user gestures and post-refill sips.
• Incorporated a self-correction mechanism to minimize estimation errors after bottle refills.
• Characterized accuracy-complexity trade-offs of various machine learning models for sip detection.

Main Results:

• The system achieved low energy consumption (approx. 119 mJ) and computation time (max 300 μs).
• The self-correction mechanism enhances robustness against sip classification errors.
• The framework is suitable for integration into low-power, consumer-grade smart water bottles.

Conclusions:

• The proposed system offers an efficient and accurate solution for on-device liquid consumption estimation.
• This technology can be integrated into everyday consumer products for personalized health monitoring.
• The energy-aware machine learning approach paves the way for advanced wearable health devices.