Real-time Recognition Of Lower Limb Motion For Exoskeleton

Exoskeleton power suit is a kind of mechatronic devices that can help human complete some heavy physical labor or dangerous operations.It is composed of mechanical arms,motors,sensors,controllers and other components.It can sense human body movements through sensors and realize the expansion and enhancement of the human body.Accurate judgment of the wearer’s motion intention is the main condition for man-machine integration of power-assisted clothing.Therefore,the real-time recognition technology of lower limb movement patterns is very critical in the design of power-assisted clothing system.At present,the main exoskeleton power suit system extensively utilizes a range of sensors for obtaining motion pattern information.Among them,the bioelectrical signal on the surface of the human body has the characteristics of non-invasive acquisition and early motion generation,which can overcome the problem of signal lag of traditional physical sensors.Therefore,the research focus of this paper is to use this signal to complete the task of lower limb movement pattern recognition,and at the same time combine the movement start and end segmentation algorithm for continuous movement process to establish an online recognition system for lower limb movement;Propose reliable control strategies based on the recognition results to improve the real-time and reliability of the exoskeleton assist system.The contents of this article primarily comprise of the following:First of all,based on the mechanism and characteristics of lower limb movement,ten muscles such as gastrocnemius muscle,biceps femoris muscle and rectus femoris muscle were selected as research objects to preprocess the collected signals,including filtering,noise reduction,resampling and other steps.Drawing from the mechanism and attributes of lower limb movement as well as the characteristics of surface EMG signals,a signal feature extraction method of EMG signal feature texture map is proposed to replace the traditional time Construction of frequency domain feature vectors.Secondly,build a convolutional neural network to build an action classifier and optimize the network.The texture map is used as input,and 5 kinds of movements commonly used in human lower limbs are used as classification output.The results show that the average recognition rate of the five actions reached 95.5% by the deep learning model using the EMG image as input.At the same time,in contrast to conventional signal feature extraction methods,this approach is more efficient and straightforward as it reduces the computational load associated with processing vast amounts of data in the initial stage,which provides conditions for the real-time perceptual classification of motion behavior in the later stage.Thirdly,aiming at the dynamic activation characteristics of muscles,a single action is captured by the double-threshold algorithm,and a model for joint detection of lower limb motion cycles by multiple muscles is established.In this paper,a lower limb fatigue exercise experiment is also designed to analyze the performance of the model in the case of muscle in fatigue stage.The results demonstrate that based on the thresholds set in this paper,the double-threshold method not only achieves a high detection accuracy in the normal state(Dp=92.35%,St D=0.03 s,0.0325s),but also has an accurate detection accuracy for the actions in the fatigue state(Dp=90.5%,St D=0.095 s,0.125s).Finally,combining the above two models to construct a real-time recognition system for lower limb movement.Deployed in the exoskeleton power suit,it can complete the real-time judgment of the astronaut’s movement and send the result as a control signal,to the exoskeleton control system.According to the form of the control signal,the exoskeleton control strategy is designed to complete the mission of assisting the wearer’s lower limb movement.