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

TL;DR

This work tackles the problem of predicting upper-limb joint torques from sEMG to support robot-assisted rehabilitation. It introduces a dedicated EMG feature extraction pipeline that transforms 8-channel sEMG into rich input features, enabling a simple MLP to achieve torque prediction performance comparable to a Temporal Convolutional Network, even with limited training data. Evaluation on a single subject performing elbow and shoulder tasks under varying loads shows that condition-specific normalization and time-frequency features improve accuracy, particularly for shoulder torques, and that the pipeline helps lightweight models capture temporal dynamics without complex architectures. The approach has practical significance for online, data-efficient rehabilitation systems and motivates future work in inverse dynamics, online deployment, and user studies, with segmentation windows of $L_{window}=100$ ms and 50% overlap.

Abstract

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.

Figures (5)

• Figure 1: Experimental setup for data collection. (a) Participant prepared with EMG electrodes lifting an object; two of the three Qualisys motion-capture cameras are visible. (b) Weighted objects used in the experiment with measured masses of 1.10kg and 1.85kg. The 0kg weight is not shown.
• Figure 2: Flowchart of the experimental procedure. In each repetition, the participant performed the movements with weights chosen in ascending order. Movement 1 denotes grasping, and Movement 2 denotes complex movement.
• Figure 3: FBD of right arm. Hand is omitted for clarity.
• Figure 4: Structure of the MLP feedforward model.
• Figure 5: Band plot of predicted vs. reference joint torques for MLP and TCN models. Top row ((a)-(c)) shows MLP predictions, and bottom row ((d)-(f)) shows TCN predictions. The shaded band represents $\pm$ 1 standard deviation across five seeds around the mean torque.