Ultra-low-power Physical Activity Classifier for Wearables: From Generic MCUs to ASICs

Area of Science:

• Embedded systems engineering
• Wearable technology
• Signal processing and machine learning

Background:

• Intelligent wearable devices generate vast physiological data streams requiring onboard processing.
• Complex algorithms, especially machine learning, increase microcontroller (MCU) power consumption, impacting overall system energy budgets.
• Current embedded algorithms consume significant MCU resources, driving the need for more efficient processing solutions.

Purpose of the Study:

• To propose and evaluate the integration of machine learning algorithms into dedicated low-power circuits for wearable devices.
• To minimize the power consumption associated with data processing in embedded systems.
• To demonstrate the effectiveness of dedicated hardware for physical activity classification algorithms.

Main Methods:

• Implementation of a pre-trained physical activity classifier algorithm.
• Combination of signal processing techniques for feature extraction.
• Utilization of machine learning (decision trees) for classification tasks.
• In-silicon implementation of the algorithm on dedicated low-power circuits.

Main Results:

• Significant reduction in power consumption for the processing component, down to 0.1 µW.
• Comparison with a general-purpose ARM Cortex-M0 MCU, which consumed 73 µW.
• Demonstrated feasibility of dedicated circuits for complex algorithm execution in wearables.

Conclusions:

• Dedicated low-power circuits can drastically reduce the energy footprint of embedded algorithms in wearable Internet of Things (IoT) devices.
• This approach makes the power consumption of data processing negligible, enabling longer battery life and more sophisticated onboard analytics.
• The proposed method offers a viable solution for energy-efficient physical activity classification in smartwatches and wristbands.