Research On Key Technology Of Human-robot Coexisting Of Assistance Exoskeleton

Exoskeleton is a complex human-robot system with wearable and close interaction with human.Exoskeleton is a combination of sensing,control and information technologies,involving key technologies such as bionic mechanical structure,drive system and control decision.In this field,although assistance exoskeleton research has made great progress in structure and low-level control algorithm,it still faces the problem of weak adaptability in key technologies of human-robot coexisting such as motion intention perception prediction and decision optimization.Human-robot coexisting assisted exoskeleton completes human-robot interaction and adaptation by processing the interaction of physiological,motor and force signals between human and exoskeleton.The main goal of the human-robot coexisting assisted exoskeleton is to maximize the reduction in human energy expenditure.Human-robot interaction force is the decisive factor of exoskeleton system performance.The interaction force reduces energy consumption by influencing the activation of relevant muscles.Appropriate human-robot coexisting strategy is the key to obtain stable and efficient interaction force curve.To meet the requirements of human-robot coexisting and optimal control of assisted exoskeleton,this paper mainly studies the prediction of motion intention perception and human-in-loop optimization in human-robot coexisting strategy.The advance characteristic of s EMG signal determines its importance in humanrobot coexisting intention prediction system.To mitigate the effect of low amplitude and high noise of s EMG signal,a convolution processing strategy adapted to s EMG signal is proposed in this study.A parallel and multi-field s EMG feature extraction algorithm was designed based on convolutional neural network.On public datasets,the proposed s EMG processing strategy achieves ideal classification accuracy in task-oriented applications.In order to further improve the characterization of s EMG feature,a reconfigurable s EMG feature extraction model based on convolutional autocoding was designed based on the fusion of muscle collaboration characteristics and multi-field convolution.With the help of the unique s EMG channel rearrangement strategy,the s EMG processing efficiency can be effectively improved to accelerate the model convergence process,and the high fidelity s EMG characteristics can be obtained.Based on the reconfigurable feature of feature extraction algorithm,the s EMG reconstruction error is obtained as the performance evaluation index of feature extraction model,which provides a guarantee for further solving the non-stationary s EMG problem.Accurate perception of human intention is fundamental to exoskeleton control.In order to obtain real-time and accurate information of lower limb joint angles and solve the geomagnetic interference of distributed inertial sensors,a joint angle vector solution method based on gravity was proposed in this study.By using the projection of gravity acceleration on adjacent sensors,the direction cosine matrix of adjacent sensors is directly estimated,which effectively alleviates the influence of geomagnetic deviation and realizes accurate calculation of joint angles.Predicting motion intention in advance and alleviating the delay of acquisition and control system are the key to ensure the accuracy of humanrobot interaction force and improve the control performance of exoskeleton.In this study,a motion intention prediction model based on recurrent neural network was proposed based on the preempt and pseudo-periodicity of s EMG signals.Experimental results show that the prediction model can accurately predict joint angle and heel striking time in advance.In order to alleviate non-stationary problems such as individual differences,muscle fatigue and muscle adaptive,and ensure the stability of prediction algorithm,an online adaptive parameter adjustment model was proposed in this study.The parameters of the prediction network model and feature extraction network model are updated in real time by using the prediction error and s EMG reconstruction error based on the network adjustment strategy,and the online adaptation of the algorithm and data is realized.By selecting the nodes of feature extraction network and prediction network reasonably,the fast and stable network self-adaptation can be realized to ensure the reliability of the motion prediction algorithm under data offset.Aiming at motion intention prediction in exoskeleton control system,this study formed a complete closed-loop process from s EMG signal feature extraction to motion prediction and adaptive adjustment.Experimental results show that the online adaptive intention prediction algorithm can accurately predict future motion intentions.At the same time,the algorithm can adjust the model parameters adaptively to ensure the stability of prediction results when s EMG and motion data are offset.The exoskeleton assist curve is the decisive factor affecting the effect of exoskeleton assist,and the optimization time and optimization effect are the key indexes of humanin-loop optimization algorithm.In order to obtain a fast and effective human-in-loop optimization algorithm,this study proposed an energy consumption evaluation index based on muscle activation from the perspective of energy consumption.The optimization goal was established to guide the optimization direction by means of the combined reward and punishment of muscle activation and mechanical constraint.In order to alleviate the problem of individual difference and complexity of motion in the optimization process,this study designed the network architecture of action value function and strategy function based on cyclic neural network.With the characteristic of high sampling rate of surface s EMG signal,the parameter optimization model of human-robot coexisting proposed in this study can realize gradual optimization.Experimental results show that the humanrobot interaction force parameter optimization model can obtain the optimal power curve in a short time for different individuals.Finally,the exoskeleton assisted experiment platform is built for the human-robot coexisting and optimal control of different individuals.The experimental platform is used to verify the proposed algorithm and the optimal interaction force curve.The final experimental results show that the human-robot coexisting strategy designed in this paper can effectively compensate the delay of acquisition system and control system,ensure the consistency of the actual human-robot interaction force curve and the optimal planning curve,and achieve high performance and stable power assist of the exoskeleton system.