Objective:
To address cross-user variability problem in the myoelectric pattern recognition, a novel method for domain generalization and adaptation using both mix-up and adversarial training strategies, termed MAT-DGA, is proposed in this paper.
Methods:
This method enables integration of domain generalization (DG) with unsupervised domain adaptation (UDA) into a unified framework. The DG process highlights user-generic information in the source domain for training a model expected to be suitable for a new user in a target domain, where the UDA process further improves the model performance with a few unlabeled testing data from the new user. In this framework, both mix-up and adversarial training strategies were also applied to each of both the DG and UDA processes by exploiting their complementary advantages towards enhanced integration of both processes. The performance of the proposed method was evaluated via experiments of classifying seven hand gestures using high-density myoelectric data recorded from extensor digitorum muscles of eight intact-limbed subjects.
Results:
It yielded a high accuracy of 95.71±4.17% and outperformed other UDA methods significantly (
p
<0.05) under cross-user testing scenarios. Moreover, it reduced the number of calibration samples required in the UDA process (
p
<0.05) after its initial performance had already been lifted by the DG process.
Conclusion:
The proposed method provides an effective and promising way of establishing cross-user myoelectric pattern recognition control systems.
Significance:
Our work helps to promote development of user-generic myoelectric interfaces, with wide applications in motor control and health.
