Deep Learning-Based Torque and Joint Angle Estimation from Electromyography Signals: A Hybrid DNN-LSTM Approach

• Ajmisha Maideen,
• A. Mohanarathinam,

Vol. 19, No. 8, August 31, 2025
10.3837/tiis.2025.08.007, Download Paper (Free):

• Deep Neural Networks
• Deep Learning
• electromyography
• Torque Estimation
• Joint Angle Estimation
• DNN-LSTM Hybrid Model

Abstract

Estimating torque and joint angles from surface electromyography (sEMG) signals is essential in biomechanics, rehabilitation, and human-machine interface (HMI) applications. This study presents an innovative, reusable deep neural network (DNN) integrated with a long short-term memory (LSTM) network for accurately estimating torque and joint angles from sEMG signals. The hybrid DNN-LSTM framework leverages the DNN’s feature extraction capabilities and the LSTM’s temporal modeling strength to effectively capture the non-linear and dynamic relationships between muscle activation patterns and joint mechanics. The proposed model estimates stimulus-induced joint angles and torque using real EMG activity from agonist-antagonist muscle pairs. By learning complex non-linear relationships between EMG signals and biomechanical variables such as torque and angle, the architecture achieves high adaptability and reusability across datasets and motion types. Experiments were conducted using both real-time and publicly available sEMG datasets under controlled movement conditions. Tests with five subjects validated the effectiveness of the proposed model, achieving a high joint angle estimation accuracy with an R² score of 0.968 ± 0.016 and an RMSE of 6.69 ± 3.22. Torque estimation achieved an R² of 0.831 ± 0.123 and an RMSE of 8.52 ± 1.72. The results demonstrate that the proposed model is accurate, robust, and generalizable across diverse motion patterns. Compared to state-of-the-art machine learning and deep learning methods, the hybrid model significantly improves estimation accuracy and resilience to noise. The DNN learns to predict joint parameters by analyzing EMG signals, while the LSTM component captures long-term dependencies inherent in sequential muscle activity.

Statistics

Cite this article