Research On Human Motion Intention Recognition Of Lower Limb Rehabilitation Exoskeleton

With the development of population aging,the issues of cardiovascular and cerebrovascular diseases caused by chronic illnesses and lower limb functional impairments resulting from traumatic injuries have gradually become a focal point of concern.Lower limb exoskeletons,as a type of rehabilitation equipment designed to assist patients with lower limb movement disorders,offer advantages in improving rehabilitation effects and reducing manpower and material resource costs.Currently,lower limb exoskeletons are emerging as commercial products,and their practicality is gradually being acknowledged.However,several challenges still exist in the practical application of lower limb rehabilitation exoskeletons,including the absence of human movement intention in the exoskeleton’s motion,the reliance on a single assistance mode,and the utilization of a fixed reference gait.This research aims to solve the current problems existing in the lower limb rehabilitation exoskeleton from the perspective of human motion intention recognition,focusing on the acquisition of active motion intention,gait phase recognition and gait pattern prediction,as well as the prediction of gait trajectory and gait phase in various motion scenarios.The main contributions and innovations of this research are outlined as follows:(1)To tackle the issue of weak muscle strength in lower limb paralysis patients and the limited distinguishable patterns of electroencephalography(EEG)signals,this research proposes a gesture-based intention recognition and exoskeleton gait control strategy.Firstly,to alleviate the preprocessing burden of surface electromyography(EMG)signals in gesture recognition,a method is introduced for segmenting activity segments based on the sample entropy of sEMG signals.Secondly,to overcome the impact of inter-subject variability in sEMG on gesture recognition,a sEMG signal gesture recognition method based on transfer learning is proposed,resulting in an improvement of 12.74%in recognition accuracy for new subjects and 6.25%in recognition accuracy for new gestures.Finally,a human-machine interaction interface is developed using upper limb sEMG,and the feasibility of controlling the exoskeleton’s gait trajectory through finite-state machines is verified.Through the use of transfer learning strategy,the application capacity of wearers’ gesture for exoskeleton gait control is validated and improved.(2)In order to address the requirements for diversified assistance modes and effective human-machine motion coupling,this research proposes gait phase recognition and gait pattern prediction methods.The gait phase recognition method utilizes a single inertial sensor placed on the foot,combined with hidden Markov model and long short-term memory network to achieve gait phase recognition in five different motion scenarios.This approach reduces the dependency on multiple sensors while ensuring robust recognition performance across different motion scenarios and for new subjects.During the unassisted exoskeleton walking,a sensitivity of 96.84%and specificity of 99.08%were achieved,while during the assisted exoskeleton walking,a sensitivity of 84.76%and specificity of 94.92%were attained.The gait pattern prediction method makes predictions based on the effect of stride time on gait patterns,providing corresponding reference gait patterns for the user of the exoskeleton,thereby improving the exoskeleton’s adaptability to the user’s gait change.(3)To meet the demand for the adaptability of exoskeletons to various motion scenarios,this research proposes a gait trajectory and event prediction algorithm based on reinforcement learning.The proposed algorithm combines the decision-making capability of reinforcement learning with the generative power of variational autoencoders,enabling continuous prediction of complete gait cycles while adapting to changes in motion scenarios.Initially,a variational autoencoder is utilized to learn the generation of diverse gait patterns that incorporate gait trajectory and event information.Subsequently,a reinforcement learning policy network is employed to adjust the generated gait patterns based on observations,incorporating stride time estimation and gait phase estimation to obtain the final prediction.Experimental results validate the adaptability of the gait trajectory and event prediction based on reinforcement learning to changes in exoskeleton users’ motion scenarios.The adaptability of various gait patterns predicted by the gait trajectory prediction enables lower limb rehabilitation exoskeletons to provide assistance on various terrains,while the gait event prediction provides valuable prediction information for the multi-scenario movement mode switching of lower limb rehabilitation exoskeletons.