Hardware-Based Hopfield Neuromorphic Computing for Fall Detection.

Area of Science:

• Embedded Systems Engineering
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Smart wearable systems require efficient sensor signal processing for machine learning in embedded scenarios.
• Traditional machine learning methods for real-time data classification are computationally intensive, challenging power-efficient embedded systems.
• Accurate fall detection using Artificial Intelligence (AI) algorithms on embedded hardware is a significant challenge.

Purpose of the Study:

• To design a hardware module simulating a Hopfield Neural Network (HNN) for power-efficient, real-time fall detection.
• To leverage the associative memory feature of HNN for sensor data integration and classification.
• To implement a hardware-based solution that overcomes the computational intensity of software-based fall detection algorithms.

Main Methods:

• Designed a hardware module to simulate the HNN algorithm for fall recognition.
• Employed Hebbian learning for training the neural network and obtaining human activity feature weights.
• Implemented these weights into the hardware design by mapping them to the amplification factor setting.
• Validated the design through simulation scenarios and experiments using an analog HNN module.

Main Results:

• Achieved a classification accuracy of 88.9% for fall detection data through hardware simulation.
• The designed system successfully performs complex signal calculations using hardware feedback, replacing software methods.
• Demonstrated the feasibility of the hardware design by comparing its performance favorably with software-based machine learning algorithms.

Conclusions:

• The developed analog Hopfield Neural Network hardware module offers a feasible and power-efficient solution for fall detection in wearable systems.
• The straightforward circuit design maximizes reusability and flexibility, enabling effective weight setting from the HNN.
• This hardware-based approach provides a competitive alternative to computationally intensive software methods for embedded AI applications.