Research On Lower Limb Motor Imagery Classification Algorithm For Walking Assistance Exoskeleton Brain-computer Interface

Brain-computer interface(BCI)is a direct connection pathway established between the brain and external devices,which decodes brain signals to and give commands for external devices.The motor imagery paradigm,due to its characteristics of not requiring external stimulation and promoting neural rehabilitation,can be combined with exoskeletons for rehabilitation training.The core of motor imagery brain-computer interface is the motor imagery classification algorithm.In recent years,the motor imagery classification algorithm has started to focus on the graph neural network that preserves the EEG topology,but the current algorithms in this field have the following bottlenecks:(1)they are greatly affected by the individual differences in EEG signal characteristics;(2)the existing way of constructing EEG topology tends to lead to its distortion.To address the above problems,this paper investigates the motion imagery classification algorithm based on graph neural networks.The specific work is as follows:(1)To address the problem of large individual differences in EEG signal features.A motion imagery classification algorithm based on a band-tuning graph convolution model is proposed.The model is based on the band tuning mechanism to obtain the optimal band weights of individuals from the data to reduce the influence of individual variability.According to the inter-subject results: the accuracies are improved by more than 5% on the public dataset and more than 10% on the self-built dataset compared to other models.The cross-subject results show that the method outperforms the compared models on both the public and self-built datasets.Subsequent visualization of the band-tuning optimal weights demonstrates the variability of the band weights across individuals,thus along with the classification results,demonstrate the adaptability of the proposed band tuning mechanism to inter-individual differences.(2)Existing graph neural network motion imagery classification algorithms usually build graph structures based on raw EEG data or simply on a priori knowledge,which ignores individual differences in neural connections while methods based on raw EEG data are prone to introduce spurious connections due to noise.To address this problem,this paper proposes an iterative graph update mechanism that generates connection probabilities from Euclidean distances and graph embeddings,and updates the graph structure during training.The classification results within a single subject show that the method improves the accuracy by 2.8% on the common dataset.Ablation experiments of the proposed graph update mechanism on the same back-bone network results in about1% improvement in accuracy,and visualization of the EEG topology during graph updating demonstrates the effectiveness of the graph update module and the topological feature extraction module.(3)To verify the effectiveness of the proposed motor imagery classification algorithm,the EEG data of 10 subjects performing imaginary rising,sitting and walking motor imagery tasks were collected to construct a lower limb motor imagery EEG dataset in this paper.The proposed motor imagery EEG classification algorithm was validated offline on its own dataset to demonstrate the validity of the proposed motor imagery classification algorithm,and then validated online in combination with a walking aid exoskeleton.This study proposed a frequency band tuning mechanism and an iterative update map mechanism for the problem of individual variability and map structure distortion of motor imagery EEG signals,acquire lower limb motor imagery EEG data,and systematically validate the motor imagery classification algorithm based on the two mechanisms.This study contributes to the practical application of brain-computer interface on mobility-assisted exoskeleton system to promote the development of neurological and motor rehabilitation.