Gait Planning For An Exoskeleton Robot Based On Multi-modal Machine Learning Method

Exoskeleton robot is a kind of new wearable device,which combines the wearer's intelligence and the exoskeleton's strength.It integrates many advanced technology,such as sensing,control,communication,medical,human-computer interaction and so on,and has been the research hotspot in the field of international robot.And gait planning is the key technology of exoskeleton robot and becomes the basis of other rich functions.However,using single source motion information can hardly take into account the accuracy,global and real-time.Therefore,this paper will collect wearers' EEG signal and physical signals,including joint position and foot pressure.The data of these two modes is fed into the machine learning model,to study the gait planning method of the exoskeleton robot.First,this paper takes the exoskeleton robot of Shenzhen Institutes of Advanced Technology as the research platform and designs the gait data collection system,which is the research basis of gait planning algorithm.The exoskeleton robot has ten degrees of freedom and the flexion and extension movement of hip and knee joints is driven by motors.Besides,the length of its legs can be adjusted to fit different wearers with various height.In order to collect gait data,angle sensors are installed on the hip,knee and ankle of the exoskeleton robot.And several pressure sensors are mounted on an insole.We invited over twenty volunteers to participate in the acquisition experiment,to ensure the diversity and reliability of gait data.Second,based on Long Short Term Memory(LSTM)architecture,a deep locomotion mode identification model is established.Different locomotion modes have disparate gait features and control characteristics,so it is a fundamental requirement for gait planning to recognize locomotion modes correctly.Five common locomotion modes,including sitting down,standing up,levelground walking,ascending stairs,and descending stairs,are taking into account in this paper.The deep locomotion mode identification machine learning model is built to dig the inherent characteristics of gait curves.In the case of using only hip,knee and ankle joint position information,our proposed model achieves accurate,robust and real-time recognition of motion modes.Besides,experimental results verify that it is far more accurate than other common machine learning models.Third,machine learning models are established from two different perspectives,to predict gait phases of the current locomotion mode and compare experimental results.Gait phase prediction is the key technology of gait planning,for its predictive results determine the output torque of the motor and the stability of gait control.According to foot pressure distribution,a complete gait is divided into four different phases in this paper,including heel-off,toe-high,heel-contact and foot-flat.Then gait phase prediction is carried out from two perspectives,including one perspective of spatial features and the other of spatio-temporal features.For the former,support vector machine(SVM)optimized by particle swarm optimization(PSO)algorithm is established,and it only focuses on joint information.For the latter,nonlinear autoregressive models with external inputs(NARX)is built,and it keeps the data of previous time in record and adds them to predict gait phases of next moment.Experimental results show that these two perspectives are capable of predicting gait phases,but in terms of accuracy,it is better to predict gait phases based on space and time dimensions simultaneously.Fourth,multi-modal data,including the EEG signal,the joint position and foot pressure,are fed into machine learning model,to improve the real-time,accuracy and security of gait planning.The brain-computer interface system based on steady state visual evoked potential is designed to collect EEG signal of subjects.After filtering by Porterworth,the Canonical Correlation Analysis method is used to recognize the motion intention of wearers.Then,the control system of the exoskeleton robot will combine the motion intention identified by EEG signal with machine state judged by physical signal to determine whether to execute relevant gait commands.Finally,the research work of gait planning for the exoskeleton robot based on multi-modal machine learning method is summarized.Besides,the future research is planned.