A Reconfigurable Near-Sensor Processor for Anomaly Detection in Limb Prostheses.

Area of Science:

• Biomedical Engineering
• Computer Engineering
• Artificial Intelligence

Background:

• Real-time monitoring of amputees with limb prostheses is crucial for detecting anomalous behaviors.
• Existing systems often lack the low-power, low-latency processing required for on-device prosthetic applications.

Purpose of the Study:

• To present a novel reconfigurable near-sensor anomaly detection processor for prosthetic limb applications.
• To develop an integrated, miniaturized system capable of real-time monitoring and anomaly detection.

Main Methods:

• The processor integrates a reconfigurable Variational Autoencoder (VAE), a scalable Self-Organizing Map (SOM) Array, and a window-size-adjustable Markov Chain.
• It supports up to 64 programmable sensor channels and utilizes time-series analysis for anomaly detection.
• Fabricated using UMC 40-nm LP technology, the design prioritizes low power and minimal area.

Main Results:

• Achieved real-time detection with high F1 Scores (0.933 for FSP, 0.956 for SFDLA-12) within microseconds.
• Demonstrated exceptional energy efficiency (43.84 nJ/feature for FSP, 55.17 nJ/feature for SFDLA-12).
• Outperformed ARM Cortex-M4/M33 microcontrollers in energy and area efficiency by significant margins (up to 257x).

Conclusions:

• The developed processor offers a highly efficient, reconfigurable solution for on-prosthesis anomaly detection.
• Its low-power and low-latency characteristics make it suitable for real-time monitoring of amputee limb behavior.
• This technology has the potential to enhance the safety and functionality of advanced prosthetic devices.