A Human Motor Intent Recognition Method For Exoskeleton Robot Based On EEG And EMG

With the rising demand for rehabilitation robots,the lower exoskeletons are developing rapidly.Natural interaction mode is an important subject of walking aid robots.There is a big gap between the traditional brain-computer interface experiment and the actual application situation.On the one hand,many BCI experiments are performed in an ideal environment without additional stimulation or electromagnetic interference.These experiments also require subjects to minimize physical movement,which is very different from the motion-assisted context.On the other hand,there are many brain-computer interfaces that are used in real-world situations.However,these applications tend to distinguish only a few classic action patterns,and the behavior of the subject to generate the instruction is not corresponding to the meaning of the instruction.For example,the two-handed grip represents backward movement.This paper attempts to use dynamic natural biological signals(EEG and EMG)to identify motion intentions of exoskeleton robot wearers.The main research contents are as follows.Firstly,this paper studies the methods of collecting EEG and EMG of subjects wearing exoskeleton for different actions,explores different action designs and signal acquisition schemes,and finally forms effective acquisition methods of dynamic EEG and EMG in complex environment.Seven healthy adult male subjects aged 20-27 were recruited in the data acquisition experiment and 70 groups were conducted.The movements included standing up,sitting down,walking(left foot),walking(right foot)and so on.A total of 7200 samples were generated.The whole experiment lasted more than one year,which provided a solid data basis for the study of motor intentions in the following chapters.Secondly,the feature extraction and binary classification of motion intention recognition are studied.Binary classification study can analyze the EEG and EMG patterns of subjects during different movements,and provide support for online motion intention recognition methods.This paper designs an algorithm based on IMU to complete action calibration and event segmentation.We paired 10 kinds of events and classified them.A total of 45 pairs of events are classified into two categories.We have tried a variety of feature extraction and classification methods,including PCA,ICA,CSP,LDA,SVM,perceptron,logarithmic probability regression,etc.In this paper,we focus on the comparison of CSP or ICA-and LDA or SVM methods.The model of CSP and LDA method is simple and suitable for intention recognition task.The average balance accuracy is 82.87%.Finally,the framework and method of online motion intention recognition are proposed.The effects of multi-classifier and binary classifier under this online framework are emphatically studied.Considering the characteristics of EEG and EMG in intention recognition and their temporal relationships,we first complete multi-classification of states based on EEG signals,and then complete precise two-classification of sub-actions based on EMG signals.The multi-classification of intention recognition is accomplished by neural network or two-classification ensemble,and the two-classification is accomplished by neural network or CSP linear model.After verification,this method can accomplish the task of online motion intention recognition well,which proves that the EEG-EMG patterns of wearer's natural motion can distinguish different motions,and also indicates the application prospect of EEG-EMG joint interface in the wearable equipment.