Study On Human Motion Intent Understanding Using Gaussian Process Based Autoregressive Learning And Active Compliance Control Of A Lower-limb Exoskeleton

Exoskeleton robots aiming at enhancing and reconstructing human motor functions have been widely applied in the fields,such as helping the elderly and the disabled,rehabilitation and military using.Through the integration of biology,computer science,control theory and mechanics together with the physical human-robot interaction,realtime sensing,dynamic control,humanmachine communication and interaction can be achieved by the corresponding human-robot system.However,the transparency between human and the robot has not been reached yet.Thus,Gaussian process based autoregressive model is developed to feature out continuous characters of human motion patterns,and evolving system is used to deal with the unspecified human motion so that accurate and fast human intent understanding could be achieved.Specifically,the computational efficiency of the energy kernel method is improved in the feature extraction of electromyogram(EMG)signals.Coupled human-exoskeleton dynamic model is developed from Hill's muscle model and semiphenomenological model.Based on this,a state space model is also developed for human motion intent learning.A Gaussian process(GP)autoregressive model is developed for multi-source information fusion.Integrated with the evolving system theory,system adaptivity is also improved to achieve accurate recognition and understanding of human motion intent.A hybrid deep learning control strategy for intelligent exoskeleton system is also developed based on the convolutional neural network(CNN).Preliminary experiments are also conducted to verify the effectiveness.The main contents and contributions of the work can be summarized as follows:First,the extracting algorithm efficiency of EMG signal feathers are improved.When the signal extraction accuracy is ensured,the variation trend of the angle between the long axis of the ellipse and the x-axis is analyzed from the signal phase portrait to improve its computational efficiency and provide convenience for its realtime application in human motion intent recognition.The area computation of the ellipse is transformed into that of its bounding rectangle.The experimental results suggest that its accuracy is similar to the original energy kernel method and the computational efficiency has been largely improved.Furthermore,the short-time Fourier transform(STFT)results of EMG are input into the CNN to extract the signal features,which makes the performance of feature extraction even better.Second,the biomechanical model based state space model for humanexoskeleton system is developed for human joint angle and angular velocity prediction.When interactive force is often ignored in the modeling of humanexoskeleton system for simplification,the coupled dynamic model is developed to interpret the generation of interactive force based on Hill's muscle model and semiphenomenological model.The forward dynamics of the joint is also integrated into the system to construct the system state space model when the interactive force and muscle activation level are included and the vector composed of joint angle and angular velocity is treated as the state vector.The state model is used to interpret the forward dynamics of the joint,and the measurement model is used to interpret the generation of the interactive force.When there are too many parameters in the biomechanical model,the acquisition of these parameters is extremely difficult and they are prone to errors,GP based nonlinear autoregressive with exogenous inputs(NARX)model and back propagation neural network(BPNN)are applied to the state model and measurement model respectively to increase the feasibility and adaptivity of the system.Unscented Kalman filter(UKF)is also used to develop a close-loop model so that the learning and prediction of human joint angle and angular velocity can be achieved.Third,evolving Gaussian process based human motion intent learning is proposed to deal with non-stationarity of EMG signal and the irregular motion pattern of human.Evolving system theory is integrated with GP-NARX model.The system evolving process is achieved through the elements upgrade in the most informative dataset.When GP can provide the uncertainty of the prediction results,the system must undergo the evolving process in two situations.One is when the absolute difference between the prediction mean value and the measured value is greater than the preset threshold.The other is when the prediction variance is larger than a preset value,as large variance means that the model is not confident in the prediction.Based on the proposed model,the joint force/torque learning and prediction model from EMG signal and the joint angle learning and prediction model form EMG signals and interactive force are developed.Fourth,hybrid deep learning based exoskeletal control strategy is developed to achieve better transparency between human and exoskeleton.The established human-exoskeleton dynamic model is further simplified to find out the interactive damping which can be used to evaluate the system compliance.Muscle activation is acquired from EMG signals through the application of CNN.It is then integrated with interactive force to achieve accurate learning and prediction of human motion intent based on evolving GP-NARX model so that the merits of both parametric model and nonparametric model can be integrated.Based on the prediction results,the hybrid deep learning based exoskeletal control strategy is developed.Using the selfdeveloped signal acquisition equipment and lower limb exoskeleton,preliminary experiments are conducted to verify the effectiveness of the proposed method.