Surface electromyogram (sEMG) has been widely used in hand gesture recognition. However, most previous studies focused on user-personalized models, which require a great amount of data from each new target user to learn the user-specific EMG patterns. In this work, we present a novel real-time gesture recognition framework based on multi-source domain adaptation, which learns extra knowledge from the data of other users, thereby reducing the data collection burdens on the target user. Additionally, compared with conventional domain adaptation methods which treat data from all users in the source domain as a whole, the proposed multi-source method treat data from different users as multiple separate source domains. Therefore, more detailed statistical information on the data distribution from each user can be learned effectively. High-density sEMG (256 channels) from 20 subjects was used to validate the proposed method. Importantly, we evaluated our method with a simulated real-time processing pipeline on continuous sEMG data stream, rather than well-segmented data. The false alarm rate during rest periods in an EMG data stream, which is typically neglected by previous studies performing offline analyses, was also considered. Our results showed that, with only 1 s sEMG data per gesture from the new user, the 10-gesture classification accuracy reached 87.66 $\%$ but the false alarm rate was reduced to 1.95 $\%$ . Our method can reduce the frustratingly heavy data collection burdens on each new user.

Human hands can execute intricate and dexterous control of diverse objects. Decoding hand motions, especially estimating the force of each individual finger via surface electromyography (sEMG), is an essential step in intuitive and dexterous control of prosthetics, exoskeletons and more various human-machine systems. Previous sEMG decoders lack explainability and show degraded performances in decoding finger forces with multiple degrees-of-freedom (DoFs). When developing a multi-DoF EMG decoder, the combinations of various forces levels exerted by different fingers are too numerous to be exhaustively enumerate. In our work, we utilized the data of 1-DoF finger activation to generate synthetic N-DoF sEMG data with a straightforward additive mixup data augmentation approach, which overlays 1-DoF sEMG signals and finger force labels. The basic assumption of our method is the additive property of sEMG associated with different DoFs. With the synthetic N-DoF sEMG data, we then developed N-DoF EMG-force models via the highly explainable deep forest built on simple and transparent decision trees. With data augmentation using only 1-DoF sEMG data, the regression error reduced by 20% of the baseline level (without data augmentation). More significantly, the explainability of the deep forest suggested that, the crucial electrodes in the decision making process of the 2-DoF deep forest are essentially a linear superposition of the counterparts in the 1-DoF deep forest.

Surface electromyography (sEMG) is commonly used as an interface in human–machine interaction systems due to their high signal-to-noise ratio and easy acquisition. It can intuitively reflect motion intentions of users, thus is widely applied in gesture recognition systems. However, wearable sEMG-based gesture recognition systems are susceptible to changes in environmental noise, electrode placement, and physiological characteristics. This could result in significant performance degradation of the model in inter-session scenarios, bringing a poor experience to users. Currently, for noise from environmental changes and electrode shifting from wearing variety, numerous studies have proposed various data-augmentation methods and highly generalized networks to improve inter-session gesture recognition accuracy. However, few studies have considered the impact of individual physiological states. In this study, we assumed that user exercise could cause changes in muscle conditions, leading to variations in sEMG features and subsequently affecting the recognition accuracy of model. To verify our hypothesis, we collected sEMG data from 12 participants performing the same gesture tasks before and after exercise, and then used Linear Discriminant Analysis (LDA) for gesture classification. For the non-exercise group, the inter-session accuracy declined only by 2.86%, whereas that of the exercise group decreased by 13.53%. This finding proves that exercise is indeed a critical factor contributing to the decline in inter-session model performance.

Myoelectric control models enabling accurate hand gesture recognition via electromyography (EMG) have attracted increasing attentions in rehabilitation robotics. Adapting pre-trained models to new users is a main challenge in real world applications due to the inter-user different EMG characteristics. Most previous transfer learning approaches employed a rigid model calibration process, usually in a supervised manner with ground truth labels, or in an unsupervised manner but still requiring users to perform pre-defined hand gestures to update model parameters. We argue that such a rigid model calibration process lacks flexibility and limit the translation of myoelectric control into real world practice. In this work, we gradually “flexibilize” the standard model calibration process toward a highly flexible version, which does not require the labels of calibration data, and can be performed on only a subset of pre-defined hand gestures or even unknown user-defined hand gestures. We identify those key components contributing to the performance difference along the way. Compared with the supervised method, the unsupervised model calibration even contributed to a 10% improvement (p< 0.05) in case where only a subset of gesture categories were available for model calibration. Moreover, the unsupervised model calibration achieved a highest recognition accuracy of 86.57% using unknown user-defined gestures, with no significant difference compared to the accuracy with pre-defined gestures (p> 0.05).

Recognizing hand gestures from surface electromyography (sEMG) signals is crucial for neural interfaces and human-machine interaction. However, developing subject-generic models remains challenging due to substantial inter-subject variability. Complicating matters further, the muscle groups driving gestures with varying degrees of freedom (DoFs) often overlap, producing highly convoluted feature distributions across subjects and DoFs. To address these challenges, we introduce a multi-branch autoencoder (AE) architecture that disentangles sEMG features into two latent subspaces: a DoF-specific (subject-invariant) space and a subject-specific (DoF-invariant) space. We systematically compare our approach against well-established feature projection methods: principal component analysis (PCA), kernel PCA (KPCA), linear discriminant analysis (LDA), kernel discriminant analysis (KDA), and a conventional AE, as well as two style-independent feature transformation methods: canonical correlation analysis (CCA) and spectral regression discriminant analysis (SRDA). Experimental results on 20 subjects across multiple days demonstrate that our multi-branch AE markedly improves DoF discrimination while maintaining subject invariance, leading to consistently higher inter-subject classification accuracy for all common classifiers. These findings underscore the potential of our approach for robust, user-independent sEMG-based gesture recognition.