EMG dataset augmentation approaches for improving the multi-DOF wrist movement regression accuracy and robustness

Benefiting from the powerful learning capacity of deep learning (DP), a general model for predicting 3-DOF wrist movements could achieve good prediction accuracy for those subjects even not involved in the training. This paper tends to verify this assumption. Since the quantity of training dataset in DP largely influences the model's performance, a limited dataset collected from a number of subjects should be extended first through some data-augmentation approaches. In this paper, we first summarized the possible mistakes happen in the EMG data-collection procedures. Then, according to these mistakes, we designed six data-augmentation approaches to expend our EMG dataset. Through experiments, we found the prediction accuracy can be improved when some approaches are introduced during training. As well, the model robustness could also be improved when the same mistakes occur in the predicting process. With regards to those approaches, placing all electrodes in opposite direction and random switching two channels have significant positive effect on both accuracy and robustness. These two data-augmentation approaches are highly advocated in the pre-processing of the training data for DP-based prediction models.