Research On Key Technologies Of Lower-limb Exoskeleton-oriented Human Motor Intent Perception

With the development of intelligent and robotic technologies,a new generation of human-machine systems moves beyond simply informative interaction through speech and vision,and turns to focus on physical human-machine interaction through positions and force.Aiming to improve human motor performance,exoskeletons are the representative of physical human-machine interaction,which integrates sports physiology,computer science and robotics.Assistive exoskeletons assist human rather than replacing human,the assistive strategy of which is to provide efficiently personalized assistance according to users’ motor intent.As the reference of personalized assistance,acquiring human motor intent and kinestate in a timely and robust manner is the foundation of the assistive exoskeletons.Aiming to solve the issue under the scenario of human-machine force interaction,the thesis utilizes the inertial and surface myoelectric(s EMG)signals to develop kinematic and kinetic models of human according to the relationship between signals and human movements,and realizes a timely,accurate and robust method of decoding human motor intent,which guarantees the information and model for accurate and personalized assistance of exoskeletons.The main content can be summarized as follows.In order to IMU-based motion capture and joint angle estimation,we first analyze the characteristics of signals,build signal model and propose an improved complementary filter-based absolute orientation estimation and tracking method for inertial sensors.With the models for acceleration,angular rate and magnetic measurements,the estimation biasbased fusion coefficient iterating and updating method is proposed by summarizing the integration of angular rates and vector observation and improving the complementary filter algorithm.An adaptive Kalman filter is also designed to adaptively estimate covariance matrices.The experimental results suggest that the improved complementary filter-based absolute orientation estimation and tracking method,comparing with the adaptive Kalman filter-based method and other state-of-the-art methods,effectively improve the estimation accuracies under slow and rapid movements.Based on the absolute orientation estimation for inertial sensors,we further realize calibration-free and sensor-movement-robust joint angle estimation method for 3-Degreeof-Freedom(3-Do F)joints.After summarizing the characteristics of 3-Do F lower-limb joints,the mathematical formulation of the 3-Do F kinematic constraint is proposed and split into three sub-problems,in order to estimate the coordinates of joint axes.And the joint angles can be estimated by decoupling the joint movements into three sequential rotations around the joint axes.Given the kinematic constraint,how sensor movements affect measurements is formulated and used to propose metrics for detecting sensor movements.With the metrics,sensor movements can be detected and corrected online.The experimental results show that the proposed 3-Do F joint angle estimation method achieve similar performance compared with classical methods and avoid using calibration procedures.The results also demonstrate that the metrics can successfully detect the sensor movements and consequently correct the errors.Starting from the physiological prior of s EMG signals and the relationship between s EMG and human movements,the spinal module activation-based neural human-machine interface is proposed.After summarizing how the motor system transmit motor command from central nervous system to peripheral neural system,we propose to utilize the population-coded spinal neuronal module activations to decode motor intent.And we also combine the coding manners of neurons to formulate mathematically how the spinal module activation-based features(firing rates and discharging timings)can be extracted from s EMG signals.Then,the features are mapped to multiple motion-related tasks to evaluate performance and explain results.The results suggest that the spinal module activation-based features can achieve the best performance under cross-validation and one-subject-out validation and under multiple task,such as gait phase recognition,motion pattern recognition and joint moment regression.The results of the human identity identification reveals that the spinal module activation-based features contain lesser individual information thus enable a better cross-subject generalization ability.After estimating the kinematics and developing the motion-related features of s EMG signals,we develop an online personalized neuromusculoskeletal model to build the kinetic model of human.We introduce the neuromusculoskeletal model by introducing the muscle-tendon unit contractions and the geometric relationship between muscle-tendon units and their adhered bones.Then,the model is simplified according to the characteristics of muscle-tendon units and becomes differential.By combing the neural networks and the neuromusculoskeletal model,the method of online inferring the personalized physiological parameters of the neuromusculoskeletal model and accuracy compensation is proposed and compared with multiple variants.The experimental results show that the method can achieve the performance better than the neuromusculoskeletal model when the training data size is sufficient.And the method can infer the personalized physiological parameters to achieve an online personalization and a good cross-subject generalization ability.