کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Sparse Autoencoder-Based Extraction of Muscle Synergies for Hand Movement Analysis from Surface EMG Signals

نویسندگان :

Fariba Biyouki¹ Asma Jafari Ani²

1- موسسه آموزش عالی خراسان 2- موسسه آموزش عالی خراسان

کلمات کلیدی :

Hand movement analysis،Muscle synergies،Latent representation learning،Sparse autoencoder،Surface electromyography

چکیده :

Understanding the low-dimensional structure underlying human motor control is a central challenge in myoelectric signal analysis. The muscle synergy hypothesis suggests that complex movements are generated through the coordinated activation of a limited set of muscle groups, offering both physiological interpretability and dimensionality reduction. In this study, a sparse undercomplete autoencoder is proposed to extract shared muscle synergies from multichannel surface electromyography (sEMG) signals recorded during hand movements. Preprocessed sEMG signals from the Ninapro Database 2 were analysed for four hand movements: hand open, hand close, forearm supination, and forearm pronation. The autoencoder was trained independently for each subject using a leave-one-repetition-out cross-validation scheme, enforcing both dimensionality reduction and sparsity in the latent space. Muscle synergies were interpreted from the encoder weight matrix, while their functional role was investigated through the temporal evolution of latent activation coefficients. The extracted synergy vectors exhibited structured, non-uniform channel contributions and demonstrated moderate-to-strong stability across repetitions. Although reconstruction performance, quantified by the coefficient of determination (R²), was limited due to the imposed sparsity and undercomplete architecture, the learned representations revealed consistent synergy structures. Analysis of latent activations showed movement-dependent modulation of shared synergies, with distinct temporal activation patterns across tasks. High cosine similarity between movement-averaged activations further supported the existence of a common synergy space, with task-specific information primarily encoded in activation timing rather than magnitude. These findings indicate that sparse autoencoder-based models provide an interpretable and physiologically meaningful framework for synergy extraction, offering a robust alternative to traditional factorisation methods for sEMG analysis.