Scholarlink

<p>Accurate real-time estimation of torque and angular joint movements using surface electromyography (sEMG) signals is essential for advancements in biomechanics, rehabilitation, and human–machine interfaces. sEMG signals, which reflect neuromuscular activity, offer significant potential for intuitive control of robotic systems by correlating neuromuscular signals with motion dynamics. In this study, a deep learning model utilizing Long Short-Term Memory (LSTM) networks was developed to predict torque and joint angles derived from sEMG signals. The model was developed to operate on the NVIDIA Jetson Nano GPU, enabling low-latency edge computing. The LSTM model was constructed and optimized using NVIDIA TensorRT, which substantially reduced inference time without compromising precision. A comprehensive dataset was constructed, comprising sEMG signals from human subjects performing controlled arm movements, alongside corresponding torque and joint angle measurements. Experimental validation was conducted using both real-time sEMG data and publicly available datasets. The proposed model exhibited strong performance, attaining a root mean squared error (RMSE) of 0.87 Nm for torque estimation and 2.5° for joint angle prediction during slow flexion–extension tasks. A correlation coefficient greater than 0.95 further highlighted the model’s reliability. Inference latency was reduced to under 10&#xa0;ms, enabling seamless integration into wearable devices and assistive technologies on the Jetson Nano platform. With a compact model size of 12&#xa0;MB and power consumption of less than 1.3 W, the system is well-suited for deployment in battery-powered wearable devices. This research addresses the critical challenge of real-time torque and joint angle estimation from sEMG signals, achieving high accuracy with minimal delay. The proposed method enables efficient and precise real-time biomechanical estimations in wearable prosthetics and rehabilitation systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Real-time torque and joint angle estimation using electromyography signals and an LSTM deep learning model on edge computing platform

  • 期刊论文

  • 发表日期：

    2025年4月28日

查看全文
  • Ajmisha Maideen,
  • A. Mohanarathinam

摘要

Accurate real-time estimation of torque and angular joint movements using surface electromyography (sEMG) signals is essential for advancements in biomechanics, rehabilitation, and human–machine interfaces. sEMG signals, which reflect neuromuscular activity, offer significant potential for intuitive control of robotic systems by correlating neuromuscular signals with motion dynamics. In this study, a deep learning model utilizing Long Short-Term Memory (LSTM) networks was developed to predict torque and joint angles derived from sEMG signals. The model was developed to operate on the NVIDIA Jetson Nano GPU, enabling low-latency edge computing. The LSTM model was constructed and optimized using NVIDIA TensorRT, which substantially reduced inference time without compromising precision. A comprehensive dataset was constructed, comprising sEMG signals from human subjects performing controlled arm movements, alongside corresponding torque and joint angle measurements. Experimental validation was conducted using both real-time sEMG data and publicly available datasets. The proposed model exhibited strong performance, attaining a root mean squared error (RMSE) of 0.87 Nm for torque estimation and 2.5° for joint angle prediction during slow flexion–extension tasks. A correlation coefficient greater than 0.95 further highlighted the model’s reliability. Inference latency was reduced to under 10 ms, enabling seamless integration into wearable devices and assistive technologies on the Jetson Nano platform. With a compact model size of 12 MB and power consumption of less than 1.3 W, the system is well-suited for deployment in battery-powered wearable devices. This research addresses the critical challenge of real-time torque and joint angle estimation from sEMG signals, achieving high accuracy with minimal delay. The proposed method enables efficient and precise real-time biomechanical estimations in wearable prosthetics and rehabilitation systems.