Impact of Noise on Deep Learning-Based Pseudo-Online Gesture Recognition with High-Density EMG

Deep neural network (DNN)–based approaches have demonstrated high accuracy in surface electromyography (sEMG)–driven gesture recognition under controlled conditions. However, estimation performance is known to degrade substantially when exposed to perturbations commonly encountered in the real-world. Here, we investigate the impact of typical noises on an autoencoder augmented recurrent neural network gesture estimator driven by engineered features and feature sets. High-density sEMG (HD-sEMG) signals offering rich information particularly suited to DNN-based algorithms were collected from thirteen participants performing wrist movements. Three types of synthetic disturbances were introduced: additive white Gaussian noise (WGN), channel loss, and electrode shift. Results indicate that when using amplitude-based features (specifically, the root mean square value and the mean absolute value), the estimator maintains robust performance under increasing WGN and channel loss, whereas its performance deteriorates markedly with features reflecting signal dynamics and fluctuations, like slope sign changes and zero crossings. Under electrode shift conditions, employing a combined feature set enhances the classifier’s resilience. Importantly, the degree of performance degradation depends on both the type and intensity of the noise. These findings confirm the need for noise-resilient architectures in order to achieve practical, everyday sEMG-driven human-machine interfaces.