A multimodal EMG–IMU dataset and multi-dataset benchmark for deep learning-based human activity recognition
TL;DR
SDALLE, a publicly available multimodal dataset integrating surface electromyography and inertial measurement unit signals acquired using a DELSYS Trigno wireless system, and an exploratory Wasserstein Generative Adversarial Network with Gradient Penalty augmentation strategy for class balancing are presented, providing a dataset resource and an exploratory subject-dependent comparison of predictive and computational behavior.
Abstract
Human Activity Recognition (HAR) using wearable sensors is relevant to rehabilitation, assistive robotics, and mobile health applications. This study presents (i) SDALLE, a publicly available multimodal dataset integrating surface electromyography (EMG) and inertial measurement unit (IMU) signals acquired using a DELSYS Trigno wireless system; (ii) a harmonized, subject-dependent comparison with the public ENABL3S dataset; and (iii) an exploratory Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) augmentation strategy for class balancing. SDALLE contains recordings of nine healthy male subjects walking, jogging, climbing stairs and descending stairs. EMG and IMU signals were synchronized, standardized, and segmented using 75%-overlapping windows. In the implemented pipeline, feature standardization preceded segmentation, after which the windows were randomly divided into 70% training and 30% testing subsets without participant- or trial-level grouping. Consequently, overlapping windows, including windows from the same participant and trial, could occur in both subsets. Under this window-level protocol, attention-enhanced CNN–LSTM models recorded approximately 98–99% accuracy on ENABL3S and approximately 99% on SDALLE. These values are protocol-specific and may be optimistic; they do not measure generalization to unseen participants or establish a portable model ranking. GAN augmentation increased the SDALLE row count by approximately 12% and changed average testing accuracy from 96.9% before augmentation to 96.5% afterward under the implemented protocol. The study therefore provides a dataset resource and an exploratory subject-dependent comparison of predictive and computational behavior. Participant-independent evaluation with training-fold-only normalization, model selection, and GAN training is required before drawing conclusions about cross-subject robustness or deployment readiness.