TL;DR
This paper introduces AdaEM, an adaptive energy management system for self-sustainable wearable health devices. By combining machine learning for predicting energy harvesting and user activity with dynamic robust optimization, AdaEM significantly reduces the need for manual recharging while maintaining high device performance and accuracy.
What Problem Does This Paper Address?
Wearable devices hold immense promise for remote health monitoring, but their widespread adoption is hampered by their limited battery life and the frequent need for recharging. This is due to their small form factor, which restricts battery size, and the intermittent nature of energy harvesting from ambient sources like solar or motion. Existing solutions often struggle to balance energy availability, user activity, and critical health monitoring needs, leading to either device downtime or compromised functionality. The core challenge is to create a system that can intelligently manage energy from harvesting and battery reserves to ensure continuous, reliable operation without user intervention.
Methods
The paper proposes an adaptive energy management (EM) algorithm called AdaEM, built on two key principles:
- Machine Learning for Prediction: AdaEM uses machine learning (ML) models to predict future user activities and the potential energy that can be harvested from ambient sources. These predictions are informed by historical data and contextual information like user location and activity type.
- Dynamic Robust Optimization (DyRO): To account for the inherent uncertainty in ML predictions and energy harvesting, AdaEM formulates the EM problem as a dynamic robust optimization problem. This allows the system to make decisions that are resilient to worst-case scenarios.
To make this computationally intensive DyRO feasible on resource-constrained wearables, AdaEM employs a two-stage, light-weight approach:
- Energy Harvest Prediction: This stage uses supervised learning (specifically, an ensemble of regression trees) to predict future energy harvest and its uncertainty.
- Charging Optimization Algorithm: This stage, inspired by the "water-filling" algorithm, takes the predicted energy harvest and uncertainty, along with user activity and energy constraints, to determine the optimal charging schedule and energy consumption for the device. It iteratively adjusts these decisions at runtime based on actual harvested energy.
Datasets
- User Activity Data: The American Time Use Survey (ATUS) dataset was used to obtain user activity patterns. This dataset was pre-processed into five categories: {Sleep, work, exercise, leisure, and others}.
- Energy Harvesting Data: Solar irradiation data from the National Renewable Energy Laboratory (NREL) in Golden, Colorado, spanning five years (2016-2020), was used to simulate outdoor light intensity. Motion energy harvesting potential was derived from user activity intensity.
Key Insights and Ideas
- The Power of Prediction and Robustness: The paper highlights that simply harvesting energy isn't enough; intelligent management is crucial. Combining predictive ML with robust optimization allows the system to proactively adapt to changing conditions and uncertainties, a significant step beyond reactive approaches.
- Balancing Act for Wearables: AdaEM demonstrates a sophisticated way to balance competing demands: maximizing battery life, ensuring continuous operation, and maintaining high accuracy for critical health monitoring tasks, all within the strict energy and computational constraints of wearables.
- Light-weight DyRO is Key: The core innovation lies in developing a practical, light-weight DyRO solution that can be deployed on actual wearable devices, overcoming the computational barrier of traditional optimization methods.
- Context Matters for Energy Harvesting: Incorporating contextual information like user location (indoor/outdoor) and activity type significantly improves the accuracy of energy harvesting predictions.
Key Results
- Near-Optimal Performance: AdaEM achieves solutions that are within 5% of the optimal performance, even when dealing with uncertain energy harvesting predictions.
- Minimal Overhead: The algorithm has a negligible execution time and energy overhead, estimated at less than 0.005% of the battery capacity per day.
- Reduced Constraint Violations: Compared to a simple "on-demand" baseline algorithm, AdaEM significantly reduces energy constraint violations, ensuring more reliable operation.
- High Accuracy Maintained: AdaEM successfully maintains high accuracy for critical activities, with only occasional outliers when energy harvesting is severely limited.
- Significant Energy Savings: The use of energy harvesting and adaptive allocation leads to substantial energy savings compared to not using EH at all.
Limitations and Future Work
- Limitations:
- The paper acknowledges that the optimal solution is computed offline and uses actual energy harvest values, which is not feasible in real-time.
- AdaEM can have occasional accuracy outliers on days with very limited energy harvest, where energy consumption must be reduced.
- Predicting sudden, critical user activity changes (e.g., falls, cardiac arrest) remains a challenge, complicating energy management.
- Future Work:
- The authors plan to deploy the AdaEM solution in the field through user studies.
- They aim to develop more robust algorithms to handle sudden, unpredictable changes in user activity.
- The dynamic robust optimization formulation has potential applications beyond wearables, in areas like agricultural monitoring, wide-area sensing, and defense.
Summary updated