Simulation for a Mems-Based CTRNN Ultra-Low Power Implementation of Human Activity Recognition.

Area of Science:

• Computer Science
• Electrical Engineering
• Biomedical Engineering

Background:

• Human Activity Recognition (HAR) typically demands significant computational power and energy consumption.
• Existing HAR systems rely on processors and memory, leading to high power usage.
• There is a need for energy-efficient solutions in wearable and embedded HAR applications.

Purpose of the Study:

• To present an energy-efficient classification framework for HAR.
• To demonstrate the feasibility of using microelectromechanical systems (MEMS) based Continuous Time Recurrent Neural Networks (CTRNN) for HAR.
• To evaluate the power consumption and classification accuracy of the proposed MEMS-CTRNN framework.

Main Methods:

• Implementation of a HAR classification framework utilizing MEMS-based CTRNN nodes.
• Real-world physical implementation to measure power consumption at the nano-watt scale.
• Evaluation of classification accuracy using the Human Activity Recognition Dataset (HAPT).
• Comparison of MEMS-CTRNN performance against traditional CTRNN and other Recurrent Neural Network (RNN) implementations.

Main Results:

• MEMS-CTRNN nodes achieved power consumption on the nano-watts scale, significantly lower than micro-watts state-of-the-art hardware.
• The HAR framework demonstrated classification accuracy comparable to traditional CTRNN and RNN methods on the HAPT dataset.
• Without pre-processing, the MEMS-CTRNN model achieved 77.94% accuracy for classifying 5 activities, versus 78.48% for traditional CTRNN.

Conclusions:

• The proposed MEMS-CTRNN framework offers a highly energy-efficient solution for HAR.
• Significant power reduction is achievable without compromising HAR classification performance.
• This technology holds promise for low-power, high-performance embedded HAR systems.