کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Deep Learning-Based Simultaneous Estimation of Elbow Joint Angle and Angular Velocity from Surface EMG Using RMS and dRMS Features

نویسندگان :

Pouria Sheikhi¹ Arshia Ghodsi² Mohammad Zareinejad³

1- دانشگاه صنعتی امیرکبیر(پلی تکنیک تهران) 2- دانشگاه صنعتی امیرکبیر(پلی تکنیک تهران) 3- دانشگاه صنعتی امیرکبیر(پلی تکنیک تهران)

کلمات کلیدی :

Surface electromyography (sEMG)،Recurrent neural networks،Joint kinematics regression،RMS and dRMS features،Rehabilitation robotics

چکیده :

Accurate estimation of joint kinematics from surface electromyography (sEMG) is crucial for achieving smooth and stable control in rehabilitation robots. Most existing approaches focus mainly on joint position estimation, while neglecting motion dynamics that are essential for physiologically consistent control. This study presents a framework for the simultaneous estimation of joint angle and angular velocity using multichannel sEMG signals. The proposed method captures both the magnitude and temporal evolution of muscle activation by combining root mean square (RMS) features with their first-order derivative (dRMS). While RMS reflects the level of motor unit recruitment, dRMS represents the rate of muscle activation and provides information related to movement velocity and electromechanical delay. These features are extracted over fixed-length temporal windows and used as inputs to sequential deep learning models. Several architectures, including LSTM, BiLSTM, CNN-LSTM, and Transformer-based regressors, are evaluated using a normalized Total MAE metric. Experimental results obtained from multi-subject elbow motion data show that the LSTM model achieves the lowest Total MAE (0.4470), outperforming the other architectures. The findings indicate that recurrent models with explicit temporal gating are particularly effective for EMG-to-kinematics regression over short temporal windows.