A normalisation approach improves the performance of inter-subject sEMG-based hand gesture recognition with a ConvNet

Recently, the subject-specific surface electromyography (sEMG)-based gesture classification with deep learning algorithms has been widely researched. However, it is not practical to obtain the training data by requiring a user to perform hand gestures many times in real life. This problem can be alleviated to a certain extent if sEMG from many other subjects could be used to train the classifier. In this paper, we propose a normalisation approach that allows implementing real-time subject-independent sEMG based hand gesture classification without training the deep learning algorithm subject specifically. We hypothesed that the amplitude ranges of sEMG across channels between forearm muscle contractions for a hand gesture recorded in the same condition do not vary significantly within each individual. Therefore, the min-max normalisation is applied to source domain data but the new maximum and minimum values of each channel used to restrict the amplitude range are calculated from a trial cycle of a new user (target domain) and assigned by the class label. A convolutional neural network (ConvNet) trained with the normalised data achieved an average 87.03 accuracy on our G. dataset (12 gestures) and 94.53 on M. dataset (7 gestures) by using the leave-one-subject-out cross-validation.