Runlin Dong ^* Xiaodong Zhang Hanzhe Li Zhufeng Lu Cunxin Li Aibin Zhu

• Xi'an Jiaotong University, Xi'an, China

Background and Objective: Exoskeleton robot control should be based on human voluntary movement intention. The ready potential (RP) component of motion-related cortical potentials (MRCP) in electroencephalogram (EEG) appears before movement and can be used for intention prediction. However, its single-trial features are weak and highly variable, and existing methods cannot maintain high accuracy when cross-temporal and cross-subject in practical online applications. Therefore, the goal of this study is to combine a deep convolutional neural network (CNN) framework with a transfer learning (TL) strategy to predict the lower limb voluntary movement intention, improving the accuracy while enhancing the model generalization capability to provide processing time for the response of the exoskeleton robotic system and to realize the robot control based on the intention of the human body.Methods: Signal characteristics of RP for lower limb movement were analyzed. Then, a parameter TL strategy based on CNN networks to predict the intention of voluntary lower limb movements was proposed. We recruited 10 subjects for offline and online experiments. The multivariate empirical mode decomposition (MEMD) method was used to remove artifacts and the onset moment of voluntary movement was labeled by the lower limb electromyography (EMG) signals in network training.Results: RP features can be observed after multiple data overlays before the onset of voluntary lower limb movements, and the feature has a long latency period. The offline experimental results showed that the average movement intention prediction accuracy was 95.23±1.25% for the right leg and 91.21±1.48% for the left leg, which showed good generalization at cross-temporal and cross-subject while the training time was greatly reduced. Online movement intention prediction can predict results 483.9±11.9 ms before movement onset with an average accuracy of 82.75%.The method proposed in this study has high prediction accuracy with less training time, good generalization performance in cross-temporal and cross-subject aspects, and is well-prioritized in terms of temporal response, which lays the foundation for further exoskeleton robot control.