Key points are not available for this paper at this time.
The widespread integration of the Internet of Things (IoT) into daily operations has made optimizing energy consumption in low-power edge devices increasingly important. This is especially critical when battery-powered IoT platforms, like sensor networks, are provided to third parties to run custom sensing and perception applications. In such cases, two main requirements emerge: A) accurate measurement of task energy consumption to fairly distribute maintenance and operations costs among parties and B) reliable identification of tasks with high power consumption fluctuations to prevent interference with sensitive measurement campaigns. However, the need for compact, cost-effective, and energy-efficient power measurement solutions often compromises dynamic range, resolution, and overall measurement quality in these platforms. To address these challenges, we introduce TinyEP, a system that leverages machine learning on highly constrained microcontrollers (TinyML) to enhance the quality of power traces recorded by limited measurement hardware on highly constrained devices in real-time. Our approach incorporates online learning in the final layer of the model, combined with sparse feedback from an existing battery fuel gauge IC, to automatically adapt to new environments and compensate for concept drift. We evaluated the robustness of TinyEP under extreme conditions, demonstrating its effective operation within a temperature range of −40 circC to 80 circC. Our results show that TinyEP improves the Mean Absolute Percentage Error (MAPE) in recorded power traces and reduces the MAPE in task energy measurements from over 2 % to under 0.8 % across all tested scenarios. Additionally, it significantly decreases residual variance during task energy measurements.
Müller et al. (Mon,) studied this question.