A Feature Extraction Pipeline for Enhancing Lightweight Neural Networks in sEMG-based Joint Torque Estimation

Robot-assisted rehabilitation offers an effective approach, wherein exoskeletons adapt to users’ needs and provide personalized assistance. However, to deliver such assistance, accurate prediction of the user’s joint torques is essential. In this work, we propose a feature extraction pipeline using 8-channel surface electromyography (sEMG) signals to predict elbow and shoulder joint torques. For preliminary evaluation, this pipeline was integrated into two neural network models: the Multilayer Perceptron (MLP) and the Temporal Convolutional Network (TCN). Data were collected from a single subject performing elbow and shoulder movements under three load conditions (0 kg, 1.10 kg, and 1.85 kg) using three motion-capture cameras. Reference torques were estimated from center-of-mass kinematics under the assumption of static equilibrium. Our offline analyses showed that, with our feature extraction pipeline, MLP model achieved mean RMSE of 0.963 N m, 1.403 N m, and 1.434 N m (over five seeds) for elbow, front-shoulder, and side-shoulder joints, respectively, which were comparable to the TCN performance. These results demonstrate that the proposed feature extraction pipeline enables a simple MLP to achieve performance comparable to that of a network designed explicitly for temporal dependencies. This finding is particularly relevant for applications with limited training data, a common scenario patient care.

💡 Research Summary

The paper addresses the challenge of estimating human joint torques for robot‑assisted rehabilitation using surface electromyography (sEMG) while keeping the computational load low and the amount of training data minimal. The authors propose a comprehensive feature‑extraction pipeline that transforms raw 8‑channel sEMG signals (sampled at 500 Hz) into a compact, informative representation suitable for lightweight neural networks. Data were collected from a single healthy male subject performing two types of upper‑limb movements (a simple grasping motion and a more complex motion combining elbow flexion/extension with shoulder abduction/adduction) while holding three different loads (0 kg, 1.10 kg, 1.85 kg). Three motion‑capture cameras recorded marker trajectories, which were used to compute reference joint torques for the elbow (τₑ) and the shoulder projected onto sagittal (τ_sf) and frontal (τ_ss) planes. The torques were derived under a static‑equilibrium assumption, treating each arm segment as a rigid body with known mass distribution and center‑of‑mass locations.

The pipeline consists of several stages: (1) preprocessing with a zero‑phase fourth‑order Butterworth band‑pass filter (15–225 Hz) followed by a variance‑based smoothing filter; (2) amplitude normalization either globally across all conditions or separately per load/movement; (3) a low‑pass filter at 5 Hz for additional smoothing. To account for the electromechanical delay between neural activation and force production, the rectified EMG is passed through a second‑order difference equation p(t)=α·e(t‑d)+β₁·p(t‑1)+β₂·p(t‑2) with constraints ensuring p(t) stays within