Transfer learning in hand movement intention detection based on surface electromyography signals.

Over the past several years, electromyography (EMG) signals have been used as a natural interface to interact with computers and machines. Recently, deep learning algorithms such as Convolutional Neural Networks (CNNs) have gained interest for decoding the hand movement intention from EMG signals. However, deep networks require a large dataset to train appropriately. Creating such a database for a single subject could be very time-consuming. In this study, we addressed this issue from two perspectives: (i) we proposed a subject-transfer framework to use the knowledge learned from other subjects to compensate for a target subject's limited data; (ii) we proposed a task-transfer framework in which the knowledge learned from a set of basic hand movements is used to classify more complex movements, which include a combination of mentioned basic movements. We introduced two CNN-based architectures for hand movement intention detection and a subject-transfer learning approach. Classifiers are tested on the Nearlab dataset, a sEMG hand/wrist movement dataset including 8 movements and 11 subjects, along with their combination, and on open-source hand sEMG dataset "NinaPro DataBase 2 (DB2)." For the Nearlab database, the subject-transfer learning approach improved the average classification accuracy of the proposed deep classifier from 92.60 to 93.30% when classifier was utilizing 10 other subjects' data our proposed framework. For Ninapro DB2 exercise B (17 hand movement classes), this improvement was from 81.43 to 82.87%. Moreover, three stages of analysis in task-transfer approach proved that it is possible to classify combination hand movements using the knowledge learned from a set of basic hand movements with zero, few samples and few seconds of data from the target movement classes. First stage takes advantage of shared muscle synergies to classify combined movements, while second and third stages take advantage of novel algorithms using few-shot learning and fine-tuning to use samples from target domain to further train the classifier trained on the source database. The use of information learned from basic hand movements improved classification accuracy of combined hand movements by 10%.