A hip exoskeleton assist control method

Abstract

The present invention relates to a hip exoskeleton assistance control method, and the control method includes the following: training a second neural network using a dataset of kinematic parameters during normal human walking to obtain a neural network joint torque model for predicting the joint torque of the hip joint; the wearer wears the hip exoskeleton and real-time obtains the rotation angles of the hip joint, knee joint and ankle joint in the sagittal plane, coronal plane and horizontal plane, and uses the trained first neural network to determine whether the wearer is in the stance phase or the swing phase during normal human walking; if it is the stance phase, input the real-time kinematic parameters during normal human walking into the neural network joint torque model to obtain the joint torque of each joint for hip joint assistance; if it is the swing phase, the corresponding hip joint assistance torque is provided in real time according to the Newton-Euler joint torque model. This method can solve the problem that the ground reaction force is difficult to predict during daily walking.

Description

Technical Field

[0001] The present invention relates to the technical field of exoskeleton control, and particularly to a hip exoskeleton assistance control method for walking assistance. Background Art

[0002] Population aging is one of the most severe problems faced by today's society. Losing the ability to move will bring great inconvenience to the daily life of patients and also impose a huge burden on families and society. Nowadays, the population in our country is shrinking, and in the foreseeable future, problems such as labor shortage and the support and healthcare of the elderly will emerge. Against this background, the research and development of wearable lower limb assistance exoskeleton robots has emerged.

[0003] The overall control strategy of lower limb exoskeleton robots can be summarized as follows: the assistance controller of the exoskeleton generates an expected assistance curve according to the human motion state, and the actuator of the exoskeleton adjusts the exoskeleton assistance magnitude according to the human motion information received by the sensor, so that the exoskeleton provides corresponding auxiliary torques according to the different motion states of the wearer.

[0004] In the research on hip joint walking assistance, researchers can reduce the energy metabolism during human walking by arranging mechanical assistance devices at the human hip joint position and applying auxiliary torques to the human hip joint. The reduction amplitude is usually between 4% and 21%. In terms of selecting control strategies and methods, some researchers adopt relatively simple control methods, that is, the lower limb joints estimate the current gait of the wearer, determine the assistance period according to the divided gait phases, and thus calculate the output torque from the torque lookup table based on the current gait cycle. There are also researchers who use BP neural networks to divide the gait during human walking to determine the assistance moment and assistance period, and generate the assistance curve of the current gait based on the assistance period of the previous step. However, the assistance torque curves in these studies are pre-designed before the assistance action occurs. The assistance torque output by the exoskeleton and the motion information of the wearer only have a simple mapping relationship, and do not calculate the joint torque required by the exoskeleton online according to the real-time kinematic information of the wearer during a gait cycle. The exoskeleton cannot provide an auxiliary torque that conforms to the human joint dynamics for the wearer at a certain moment during a gait cycle. Summary of the Invention

[0005] According to the technical problems existing in the lower limb exoskeleton robot control method pointed out in the background art, the purpose of the present invention is to provide a hip exoskeleton assistance control method, which can directly and accurately calculate the required assistance torque for the current action of the wearer by using the motion information (angle, angular velocity, angular acceleration, etc.) of the wearer captured in real time by the wearable inertial sensor, calculate the joint torque required by the exoskeleton online, and thus provide an auxiliary torque for the wearer in real time.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A hip exoskeleton assistance control method, the control method includes the following:

[0008] Obtain the rotation angles of the hip joint, knee joint, and ankle joint in the sagittal plane, coronal plane, and horizontal plane under the human body's motion state, and label the rotation angles as the stance phase or swing phase to establish a motion state data set;

[0009] Obtain the angular velocity and angular acceleration of the hip joint, knee joint, and ankle joint, the centroid acceleration of the thigh, calf, and foot, the rotation angles of the hip joint, knee joint, and ankle joint in the sagittal plane, coronal plane, and horizontal plane, the joint coordinates of the hip joint, knee joint, and ankle joint, the centroid coordinates of the thigh, calf, and foot, and the joint torques of each joint to establish a kinematic parameter data set during normal human walking;

[0010] Use the motion state data set to train a first neural network for identifying whether the human body's motion state is the stance phase or the swing phase;

[0011] Use the kinematic parameter data set during normal human walking to train a second neural network to obtain a neural network joint torque model for predicting the joint torque of the hip joint;

[0012] The wearer wears the hip exoskeleton and real-time obtains the rotation angles of the hip joint, knee joint, and ankle joint in the sagittal plane, coronal plane, and horizontal plane. Use the trained first neural network to determine whether the wearer is in the stance phase or the swing phase during normal human walking; if it is the stance phase, input the real-time kinematic parameters during normal human walking into the neural network joint torque model to obtain the joint torques of each joint for hip joint assistance; if it is the swing phase, provide the corresponding hip joint assistance torque in real time according to the Newton-Euler joint torque model.

[0013] Further, the second neural network is a multi-layer classification-regression fusion neural network, and the multi-layer classification-regression fusion neural network includes: a classification layer, a regression layer, and a fusion layer;

[0014] The classification layer adopts the SVM neural network algorithm. The SVM neural network algorithm takes the rotation angles of the hip joint, knee joint, and ankle joint in the sagittal plane, coronal plane, and horizontal plane under the human body's motion state as the feature vector input, divides the gait phase of the human body in the stance phase into four phases: initial contact, loading response, mid-stance, and terminal stance, and records the four phases as k1, k2, k3, k4;

[0015] The regression layer adopts a parallel neural network jointly trained by NTM (Neural Turing Machine) and RNN (Recurrent Neural Network). The kinematic parameter data after normalization are combined into feature vectors and input into this parallel neural network. Each feature vector outputs the predicted value of the hip joint torque of the human body after passing through the regression layer;

[0016] After the training of the regression layer is completed, the output value of the regression layer will be input into the fusion layer. The role of the fusion layer is to calculate the weight ratio between the results of different regression layers under different divided phases. In this task, the backpropagation algorithm is used to calculate the weights;

[0017] The final hip joint torque prediction value M of the multi-layer classification-regression fusion neural network satisfies the following formula:

[0018] M = X(k)M N + Y(k)M R

[0019] where M N is the hip joint torque value predicted by NTM, M R is the hip joint torque value predicted by RNN, X(k) is the weight ratio of the NTM prediction value output by the fusion layer, Y(k) is the weight ratio of the RNN prediction value output by the fusion layer, and the weight values are all related to the gait phases divided by the SVM neural network algorithm, and k is the phase;

[0020] Use the kinematic parameter data set during normal human walking to train the multi-layer classification-regression fusion neural network to obtain a neural network joint torque model for predicting the joint torque of the hip joint.

[0021] Furthermore, the Newton-Euler joint torque model is:

[0022]

[0023] In the formula, M1 is the joint torque of the hip joint; r1, r2, r3 are the centroid coordinates of the thigh, calf, and foot respectively; c1, c2, c3 are the joint coordinates of the hip joint, knee joint, and ankle joint respectively, ω1, ω2, ω3 are the joint angular velocities of the hip joint, knee joint, and ankle joint respectively, α1, α2, α3 are the joint angular accelerations of the hip joint, knee joint, and ankle joint respectively, I1, I2, I3 are the moments of inertia of the thigh, calf, and foot about their respective centroids; f1, f2, f3 are the joint forces of the hip joint, knee joint, and ankle joint respectively, and the expressions are:

[0024]

[0025]

[0026]

[0027] In the formula, are respectively the rotation matrices from the origin of the human body base coordinate to the hip joint, knee joint, and ankle joint; f0 is the ground reaction force, which is 0 during the swing phase; a1, a2, and a3 are respectively the centroid accelerations of the thigh, calf, and foot; g is the acceleration due to gravity.

[0028] Furthermore, a human gait model is established in the VICON infrared optical motion capture system, and the kinematic parameter data (joint coordinates, joint angles, angular velocities, and angular accelerations of the hip joint, knee joint, and ankle joint, centroid coordinates, centroid velocities, and centroid accelerations of the thigh, calf, and foot, a total of 21*3 inputs) during normal human walking are collected by infrared cameras as the eigenvalue required for neural network training. At the same time, the hip joint torque of the human body during normal walking (obtained by experimental methods) collected by the force platform is combined as the target value for neural network training.

[0029] Further, the hip exoskeleton includes a force sensor arranged at the exoskeleton assistance point, a binding near the knee joint, a tension belt connecting the binding and the waist motor, a motor driver for driving the waist motor, a PID controller, and a wearable inertial sensor for capturing human motion information;

[0030] Taking the human motion information captured by the wearable inertial sensor as the input, the target assistance torques of the two motion states of stance-phase assistance and swing-phase assistance are output respectively, and thus the target human-machine interaction force F is calculated according to the following formula a ,

[0031]

[0032] where M is the target assistance torque, obtained by output from the neural network joint torque model or the Newton-Euler joint torque model, L is the distance between the hip joint and the exoskeleton assistance point; θ is the angle between the wearer's thigh and the tension belt of the hip exoskeleton;

[0033] Taking the target human-machine interaction force F designed in the two motion states of stance-phase assistance and swing-phase assistance a as the input, a force controller is designed: a force sensor is added at the exoskeleton assistance point, and the target human-machine interaction force F a is subtracted from the actual human-machine interaction force F measured by the force sensor b to obtain the force deviation value F e , and the calculation formula is F e = F a - F b ; the force deviation value F e is input into the PID controller for calculation to obtain the target current value I required by the motor driver t , and compared with the detected current feedback I sThe difference is taken to obtain the actual current value I required by the motor driver r , and the motor driver drives the motor of the hip exoskeleton to move according to the target current value I t , and applies an auxiliary force to the wearer by driving the binding, so as to achieve the target human-machine interaction force F designed in the two motion states of the support phase assistance force and the swing phase assistance force a , and provides the required target torque M for the wearer's hip joint.

[0034] Furthermore, in order to ensure that the hip joint torque prediction algorithm can adapt to the requirements of hip joint torque prediction tasks in different gait phases and improve the accuracy of hip joint torque prediction for the overall walking gait, a multi-layer classification-regression fusion neural network algorithm is proposed, including a classification layer, a regression layer and a fusion layer. The normalized kinematic parameter data is combined into a feature vector and input into the multi-layer classification regression fusion neural network. Each vector outputs the torque of the human hip joint through the regression layer. When the maximum number of iterations is reached, the training is completed, and the trained model is saved.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] The control method of the present invention can directly and accurately calculate the required assistance torque under the current action of the wearer by using the motion information (angle, angular velocity, angular acceleration, etc.) of the wearer captured in real time by the wearable inertial sensor as input, so as to provide the assistance torque required for the exoskeleton to walk at the current moment of the human body in real time. The exoskeleton can thereby update the assistance torque curve in real time according to different motion information of the human body at different moments. Compared with the prior art, the assistance curve provided by this method for the exoskeleton has higher real-time performance and is more in line with the dynamics of human joints.

[0037] The control method of the present invention uses different models for assistance according to the support phase and the swing phase, can solve the problem that the ground reaction force during daily walking is difficult to predict, and can obtain the joint torque only with easily measured parameters. Description of the Drawings

[0038] Figure 1 It is a schematic structural diagram of a seven-link model of the human lower limb.

[0039] Figure 2 It is an architecture diagram of a multi-layer classification-regression fusion neural network algorithm.

[0040] Figure 3 It is a control block diagram of a force controller.

[0041] Figure 4 It is a schematic diagram of an exoskeleton assistance module. Detailed Embodiments

[0042] The present invention will be further explained below in conjunction with embodiments and the accompanying drawings, but this is not intended to limit the protection scope of the present application.

[0043] A hip exoskeleton assist control method of the present invention includes the following steps:

[0044] Step 1: Starting from joint dynamics, establish a Newton-Euler joint torque model for the human lower limb.

[0045] Step 2: Collect the kinematic and dynamic data of the human lower limb and normalize them. Construct a multi-layer classification-regression fusion neural network. Use the human motion information as the input of the multi-layer classification-regression fusion neural network, and the hip joint torque as the output of the multi-layer classification-regression fusion neural network. Train the multi-layer classification-regression fusion neural network to establish a neural network joint torque model for predicting the human hip joint torque.

[0046] Step 3: When the wearer performs a walking motion, the exoskeleton provides corresponding assist torques in real time according to the wearer's motion state.

[0047] The present invention can provide corresponding assistance to the hip joint in real time based on the motion information of the wearer at a certain moment, improving the assistance efficiency and reducing the physical energy consumption of the wearer.

[0048] Further, specifically included in step S1 are:

[0049] a Calculate the physiological parameters of the human lower limb segments. Obtain the body data based on human statistics, including the lengths, masses, and center of gravity positions of various body parts. The calculation formulas are as follows:

[0050]

[0051]

[0052]

[0053] Among them, H is the height, L1 is the thigh length, L2 is the calf length, L3 is the foot length, C1 is the distance from the center of gravity of the thigh to the hip joint, C2 is the distance from the center of gravity of the calf to the knee joint, C3 is the distance from the center of gravity of the foot to the ankle joint, M is the body weight, m1 is the thigh mass, m2 is the calf mass, and m3 is the foot mass.

[0054] b Simplify the human lower limb into a seven-link model, as Figure 1 shown, and calculate the transformation matrix between each joint and the adjacent joint.

[0055] c Based on the physiological parameters of the human lower limb segments and the transformation matrix between joints obtained in steps a and b, establish a Newton-Euler joint torque model for the human lower limb, and calculate the joint forces of each joint of the human lower limb during walking:

[0056]

[0057]

[0058]

[0059] In the formula, are respectively the rotation matrices from the origin of the human body base coordinate to the hip joint, knee joint, and ankle joint; f0 is the ground reaction force on the sole of the foot, which is 0 during the swing phase; f1, f2, and f3 are respectively the joint forces of the hip joint, knee joint, and ankle joint; a1, a2, and a3 are respectively the centroid accelerations of the thigh, calf, and foot; and g is the acceleration due to gravity.

[0060] Finally, the joint torques of each joint are calculated, and the formula is as follows:

[0061]

[0062]

[0063]

[0064] Finally, the calculation formula for the hip joint is

[0065]

[0066] In the formula, M1, M2, and M3 are respectively the joint torques of the hip joint, knee joint, and ankle joint; r1, r2, and r3 are respectively the centroid coordinates of the thigh, calf, and foot; c1, c2, and c3 are respectively the joint coordinates of the hip joint, knee joint, and ankle joint, ω1, ω2, and ω3 are respectively the joint angular velocities of the hip joint, knee joint, and ankle joint, α1, α2, and α3 are respectively the joint angular accelerations of the hip joint, knee joint, and ankle joint, and I1, I2, and I3 are respectively the moments of inertia of the thigh, calf, and foot about their respective centroids;

[0067] The calculation formula for the hip joint is the Newton-Euler joint torque model.

[0068] Furthermore, step S2 specifically includes:

[0069] a Obtain the angular velocities and angular accelerations of the hip joint, knee joint, and ankle joint, the centroid accelerations of the thigh, calf, and foot, the rotation angles of the hip joint, knee joint, and ankle joint in the sagittal plane, coronal plane, and horizontal plane, the joint coordinates of the hip joint, knee joint, and ankle joint, the centroid coordinates of the thigh, calf, and foot, and the joint torques of each joint, and establish a kinematic parameter dataset for normal human walking; and normalize these data.

[0070] b Based on the prediction effects of existing neural networks in different gait phases during a gait cycle, the classification layer of the multi-layer classification-regression fusion neural network algorithm uses the SVM (Support Vector Machine) neural network algorithm. Taking the joint angles of the human lower limbs as the feature vector input, the gait phases of the human body during the stance phase are divided into four phases: initial contact, loading response, mid-stance, and terminal stance, denoted as k1, k2, k3, and k4. The regression layer uses a parallel neural network jointly trained by NTM (Neural Turing Machine) and RNN (Recurrent Neural Network). The normalized kinematic parameter data is combined into a feature vector and input into this parallel neural network. Each vector outputs the predicted value of the hip joint moment of the human body after passing through the regression layer. After the regression layer is trained, the output value of the regression layer will be input into the fusion layer. The role of the fusion layer is to calculate the weight ratio between different regression layer results under different divided phases. The backpropagation algorithm is used to calculate the weights in this task. Its architecture is as Figure 2 shown.

[0071] Finally, the predicted value of the hip joint moment of the multi-layer classification-regression fusion neural network algorithm satisfies the following formula:

[0072] M = X(k)M N + Y(k)M R

[0073] where M N is the predicted value of the hip joint moment by the NTM neural network, M R is the predicted value of the hip joint moment by the RNN neural network, X(k) is the weight ratio of the NTM neural network prediction value, Y(k) is the weight ratio of the RNN neural network prediction value, and the weight values are all related to the gait phases divided by the SVM algorithm.

[0074] Using the kinematic parameter dataset during normal human walking to train the second neural network to obtain a neural network joint moment model for predicting the joint moment of the hip joint; taking the kinematic parameters in the kinematic parameter dataset as feature values and the hip joint moment during human walking as the target value for neural network training.

[0075] Furthermore, step S3 specifically includes:

[0076] When walking with a wearable exoskeleton, the wearer's motion state is divided into a stance phase and a swing phase, and the judgment basis is the plantar pressure of the human body during walking, with a threshold of 0. That is, when the plantar pressure is greater than 0, the assist controller determines that the wearer enters the stance phase and adjusts the assist mode to the neural network joint torque model assist. When the plantar pressure is equal to 0, the assist controller determines that the wearer enters the swing phase and adjusts the assist mode to the Newton-Euler joint torque model assist. The wearable IMU sensor captures the motion information of the human body during walking in real time and determines the human motion state based on the human motion information, so that the assist controller continuously adjusts the assist mode and the motor output torque.

[0077] In the present invention, the human motion state is divided into a stance phase and a swing phase. In this embodiment, the judgment method is an XGBoost machine learning classification model pre-constructed with the rotation angles of three joints in the sagittal plane, coronal plane, and horizontal plane as inputs and the gait phase as the output. When the output of the XGBoost machine learning classification model is the stance phase, the assist controller determines that the wearer enters the stance phase and adjusts the assist mode to the neural network joint torque model assist. When the output of the XGBoost machine learning classification model is the swing phase, the assist controller determines that the wearer enters the swing phase and adjusts the assist mode to the Newton-Euler joint torque model assist.

[0078] The assist torque curves in the existing hip exoskeleton assist control methods are all pre-designed before the required assist actions occur and do not calculate the joint torque required by the exoskeleton in real time according to the real-time kinematic information of the wearer during a gait cycle. In the present invention, the motion information of the wearer at a certain moment can be used as the input of the neural network joint torque model or the Newton-Euler joint torque model to directly calculate the joint torque required by the wearer at the current moment, so as to provide corresponding assistance to the hip joint in real time, improve the assist efficiency, and reduce the physical energy consumption of the wearer.

[0079] Embodiment

[0080] This embodiment proposes a hip exoskeleton walking assist control method, including:

[0081] S1 Establish a human lower limb joint dynamics model: Calculate the physiological parameters of the human lower limb segments. According to the physiological parameters of the human lower limb, simplify the human lower limb into a seven-link model, calculate the transformation matrix between each joint and the adjacent joint, and then establish a human lower limb joint Newton-Euler recursive inverse dynamics model based on the calculated physiological parameters of the human lower limb segments and the joint transformation matrix.

[0082] Obtain body data based on human statistics, including the lengths, masses, and center of gravity positions of various body parts. The calculation formulas are as follows:

[0083]

[0084]

[0085]

[0086] Among them, H is height, L1 is thigh length, L2 is calf length, L3 is foot length, C1 is the distance from the center of gravity of the thigh to the hip joint, C2 is the distance from the center of gravity of the calf to the knee joint, C3 is the distance from the center of gravity of the foot to the ankle joint, M is body weight, m1 is thigh mass, m2 is calf mass, and m3 is foot mass. The lower limbs of the human body are simplified into a seven-link model, and the transformation matrix between each joint and the adjacent joint is calculated. Taking the pelvis as the body base coordinate of the human body, the formula is as follows:

[0087]

[0088] In the formula is the rotation matrix from the origin of the body base coordinate of the human body to the hip joint. α1, γ1, and β1 are the rotation angles of the hip joint in the sagittal plane, coronal plane, and horizontal plane, respectively. Similarly, the rotation matrix from the hip joint to the knee joint can be obtained and the rotation matrix from the knee joint to the ankle joint Next, calculate the rotation matrix from the origin of the body base coordinate of the human body to the knee joint

[0089]

[0090] The rotation matrix from the origin of the body base coordinate of the human body to the ankle joint

[0091]

[0092] Subsequently, calculate the joint forces of each joint of the lower limbs of the human body during walking:

[0093]

[0094]

[0095]

[0096] In the formula, f0 is the ground reaction force on the sole of the foot, which is 0 during the swing phase. f1, f2, and f3 are the joint forces of the hip joint, knee joint, and ankle joint, respectively. a1, a2, and a3 are the centroid accelerations of the thigh, calf, and foot, respectively. g is the acceleration due to gravity.

[0097] Finally, calculate the expression of the joint torque of each joint. The formula is as follows:

[0098]

[0099]

[0100]

[0101] The final hip joint calculation formula is

[0102]

[0103] In the formula, M1, M2, and M3 are the joint torques of the hip joint, knee joint, and ankle joint respectively; r1, r2, and r3 are the centroid coordinates of the thigh, calf, and foot respectively; c1, c2, and c3 are the joint coordinates of the hip joint, knee joint, and ankle joint respectively, ω1, ω2, and ω3 are the joint angular velocities of the hip joint, knee joint, and ankle joint respectively, α1, α2, and α3 are the joint angular accelerations of the hip joint, knee joint, and ankle joint respectively, and I1, I2, and I3 are the moments of inertia of the thigh, calf, and foot about their respective centroids.

[0104] S2 collects the kinematic and kinetic data of the human lower limb and normalizes them. The motion information of the human body is used as the input of the neural network, and the hip joint torque is used as the output of the second neural network. The second neural network is trained to establish a neural network joint torque model for predicting the human hip joint torque: the second neural network is a multi-layer classification-regression fusion neural network, including: a classification layer, a regression layer, and a fusion layer;

[0105] The classification layer uses the SVM (Support Vector Machine) neural network algorithm, with the joint angles of each joint of the human lower limb as the input of the feature vector. The gait phase of the human body during the stance phase is divided into four phases: initial contact, loading response, mid-stance, and terminal stance, and the four phases are denoted as k1, k2, k3, and k4. The regression layer uses a parallel neural network jointly trained by NTM (Neural Turing Machine) and RNN (Recurrent Neural Network). The normalized kinematic parameter data are combined into a feature vector and input into this parallel neural network. Each vector outputs the predicted value of the hip joint torque of the human body after passing through the regression layer. After the regression layer is trained, the output value of the regression layer will enter the fusion layer to calculate the weight ratio between different results of the regression layer under different divided phases.

[0106] Referring to the kinetic model established in step S1, specific human kinematic parameters are extracted as feature values, and these kinematic parameter data and the joint torque data of the human body are collected and normalized. The normalization formula is as follows:

[0107]

[0108] Where x0 is the original data, x is the normalized data, x min is the minimum value in the original data, x max is the maximum value in the original data.

[0109] The predicted hip joint moment of the final multi-layer classification-regression fusion neural network algorithm satisfies the following formula:

[0110] M = X(k)M N + Y(k)M R

[0111] Where M N is the hip joint moment value predicted by the NTM, and M R is the hip joint moment value predicted by the RNN. X(k) is the weight ratio of the NTM prediction value, and Y(k) is the weight ratio of the RNN prediction value. The weight values are all related to the gait phases divided by the SVM algorithm.

[0112] The fusion layer includes four fusion layers, which calculate the weights of different phases of the stance phase respectively, and obtain the predicted hip joint moment value M according to the above formula.

[0113] When the wearer performs walking motion, the exoskeleton provides corresponding assistive moments in real time according to the wearer's motion state:

[0114] The wearable IMU sensor captures the motion information of the human body during walking in real time, judges the human motion state based on the human motion information, divides the wearer's motion state into the stance phase and the swing phase, and the assistive controller thus continuously adjusts the assistive mode and the motor output moment.

[0115] When the wearer is in the stance phase, the ground reaction force required for joint moment calculation cannot be measured, so the moment assistive curve generated by the neural network joint moment model is used for assistance.

[0116] When the wearer is in the stance phase, the ground reaction force is 0, and the moment assistive curve generated by the Newton-Euler joint moment model can be used for assistance.

[0117] Taking the human motion information captured by the wearable inertial sensor as the input, the target assistive moment is output by the lower limb joint moment model in two motion states of stance phase assistance and swing phase assistance, and then the target human-machine interaction force is calculated. The formula is as follows:

[0118]

[0119] Where M is the target assistive moment output by the lower limb joint moment model, L is the distance between the hip joint and the exoskeleton assistive point. θ is the angle between the wearer's thigh and the tension belt, which is related to the wearer's gait phase.

[0120] In this embodiment, the hip exoskeleton adopted is a front-back pull flexible exoskeleton. It is attached near the knee joint of the wearer's thigh. The rear tension belt and the front tension belt are respectively connected to the power-assisted points on the inner and outer sides of the attachment. One end of each tension belt is connected to the power-assisted point, and the other end is connected to the motor fixed on the waist. Each tension belt is controlled by one motor. The motor is connected to the motor driver, and both the motor driver and the motor are installed on the waist to achieve the control of hip joint assistance. A force sensor is set at the position where the tension belt is connected to the attachment. When the outside is stressed, the angle θ is the angle between the front tension belt and the thigh. When the inside is stressed, the angle θ is the angle between the rear tension belt and the thigh. The installation schematic diagram of the components of the exoskeleton near the knee joint is as Figure 4 shown.

[0121] Taking the target human-machine interaction force F designed in two motion states of support-phase assistance and swing-phase assistance as the input, a force controller is designed. A force sensor is added at the power-assisted point of the exoskeleton. The target human-machine interaction force F a and the actual human-machine interaction force F measured by the force sensor a are subtracted to obtain the force deviation value F b . The calculation formula is F e =F e -F a . The force deviation value F b is input into the PID controller for calculation to obtain the target current value I required by the motor driver e , and it is subtracted from the detected current feedback I t to obtain the actual current value I required by the motor driver s . The motor driver drives the motor of the hip exoskeleton to move according to the current value I r and drives the attachment to apply an auxiliary force to the wearer, so as to achieve the target human-machine interaction force F t designed in two motion states of support-phase assistance and swing-phase assistance, and provide the target torque M output by the joint torque model for the wearer's hip joint. a The control block diagram of the force controller is as

[0122] shown. Based on the human motion information captured by the wearable inertial sensor, combined with the constructed human lower limb joint dynamics model, the target assistance torque in the current motion state is output, and the target human-machine interaction force F Figure 3 is calculated through the human-machine interaction force calculation formula. Taking the target human-machine interaction force F a as a positive value and the actual human-machine interaction force F measured by the force sensor a as a negative value, they are added (∑) to obtain the force deviation value F b . The force deviation value F e is input into the PID controller for calculation to obtain the target current value I required by the motor driver e ​t and subtract it from the detected current feedback I s to obtain the actual current value I required by the motor driver r Input I r into the motor driver, and the motor driver drives the motor of the hip exoskeleton according to the current value I t to perform motion and provide the target assistance torque in the current motion state for the wearer.

[0123] Those not described in the present invention are applicable to the prior art.

Claims

1. A hip exoskeleton assist control method, characterized in that The control method includes the following: Obtain the rotation angles of the hip joint, knee joint, and ankle joint in the sagittal plane, coronal plane, and horizontal plane during human movement, and label the rotation angles as the stance phase or swing phase to establish a motion state data set; Obtain the angular velocity and angular acceleration of the hip joint, knee joint, and ankle joint, the centroid acceleration of the thigh, calf, and foot, the rotation angles of the hip joint, knee joint, and ankle joint in the sagittal plane, coronal plane, and horizontal plane, the joint coordinates of the hip joint, knee joint, and ankle joint, the centroid coordinates of the thigh, calf, and foot, and the joint torques of each joint to establish a kinematic parameter data set during normal human walking; Train a first neural network using the motion state data set to identify whether the human motion state is the stance phase or the swing phase; Train a second neural network using the kinematic parameter data set during normal human walking to obtain a neural network joint torque model for predicting the joint torque of the hip joint; The wearer wears a hip exoskeleton and obtains in real time the rotation angles of the hip joint, knee joint, and ankle joint in the sagittal plane, coronal plane, and horizontal plane, and uses the trained first neural network to determine whether the wearer is in the stance phase or the swing phase during normal human walking; If it is the stance phase, use the neural network joint torque model to input the real-time kinematic parameters during normal human walking to obtain the joint torques of each joint for hip joint assistance; If it is the swing phase, provide the corresponding hip joint assistance torque in real time according to the Newton-Euler joint torque model; The Newton-Euler joint torque model is: In the formula, M1 is the joint torque of the hip joint; r1, r2, r3 are the centroid coordinates of the thigh, calf, and foot respectively; c1, c2, c3 are the joint coordinates of the hip joint, knee joint, and ankle joint respectively, ω1, ω2, ω3 are the joint angular velocities of the hip joint, knee joint, and ankle joint respectively, α1, α2, α3 are the joint angular accelerations of the hip joint, knee joint, and ankle joint respectively, I1, I2, I3 are the moments of inertia of the thigh, calf, and foot about their respective centroids; f1, f2, f3 are the joint forces of the hip joint, knee joint, and ankle joint respectively, and the expressions are: wherein are respectively the rotation matrices from the origin of the human body base coordinate system to the hip joint, knee joint, and ankle joint; f0 is the ground reaction force, which is 0 during the swing phase; a1, a2, and a3 are respectively the centroid accelerations of the thigh, calf, and foot; and g is the acceleration due to gravity.

2. The hip exoskeleton assistance control method according to claim 1, wherein The second neural network is a multi-layer classification-regression fusion neural network, and the multi-layer classification-regression fusion neural network includes: a classification layer, a regression layer, and a fusion layer; The classification layer uses the SVM neural network algorithm. The SVM neural network algorithm takes the rotation angles of the hip joint, knee joint, and ankle joint in the sagittal plane, coronal plane, and horizontal plane during human movement as the input of the feature vector, divides the gait phase of the human body in the stance phase into four phases: initial contact, loading response period, mid-stance phase, and terminal stance phase, and records the four phases as k1, k2, k3, k4; The regression layer uses a parallel neural network trained synchronously by the Neural Turing Machine NTM and the Recurrent Neural Network RNN. The normalized kinematic parameter data combinations are combined into a feature vector and input into this parallel neural network, and each feature vector outputs the torque prediction value of the human hip joint after passing through the regression layer; After the regression layer is trained, the output value of the regression layer will be input into the fusion layer. The role of the fusion layer is to calculate the weight ratio between the results of different regression layers under different divided phases. The backpropagation algorithm is used to calculate the weights in this task; The final hip joint moment prediction value M of the multi-layer classification-regression fusion neural network satisfies the following formula: M = X(k)M N + Y(k)M R where M N is the hip joint moment value predicted by NTM, and M R is the hip joint moment value predicted by RNN. X(k) is the weight ratio of the NTM prediction value output by the fusion layer, Y(k) is the weight ratio of the RNN prediction value output by the fusion layer. The weight values are all related to the gait phase divided by the SVM neural network algorithm, and k is the phase; The multi-layer classification-regression fusion neural network is trained using the kinematic parameter dataset during normal human walking to obtain a neural network joint moment model for predicting the joint moment of the hip joint.

3. The hip exoskeleton assist control method according to claim 1, characterized in that The process of obtaining the kinematic parameter dataset during normal human walking is as follows: A human gait model is established in the VICON infrared optical motion capture system, and the kinematic parameter data during normal human walking is collected using infrared cameras. At the same time, the joint moment of the hip joint during normal human walking is collected in combination with a force platform.

4. The hip exoskeleton assist control method according to claim 1, wherein The hip exoskeleton includes a force sensor provided at the exoskeleton assistance point, a binding near the knee joint, a tension belt connecting the binding to the waist motor, a motor driver for driving the waist motor, a PID controller, and a wearable inertial sensor for capturing human motion information; Taking the human motion information captured by the wearable inertial sensor as input, and outputting the respective target assist torques in two motion states of the support phase assist and the swing phase assist, thereby calculating the target human-machine interaction force F according to the following formula a , Among them, M is the target assistance moment, which is obtained by the output of the neural network joint moment model or the Newton-Euler joint moment model, L is the distance between the hip joint and the exoskeleton assistance point; θ is the angle between the wearer's thigh and the tension belt of the hip exoskeleton; The target human-machine interaction force F designed in two motion states of support-phase assistance and swing-phase assistance a is used as the input to design a force controller: a force sensor is added at the exoskeleton assistance point, and the target human-machine interaction force F a is subtracted from the actual human-machine interaction force F measured by the force sensor b to obtain a force deviation value F e , and the calculation formula is F e = F a - F b ; the force deviation value F e is input into the PID controller for calculation to obtain the target current value I required by the motor driver t , and the detected current feedback I s is subtracted from it to obtain the actual current value I required by the motor driver r . The motor driver drives the motor of the hip exoskeleton to move according to the target current value I t and drives the binding to apply an auxiliary force to the wearer, thereby realizing the target human-machine interaction force F designed in two motion states of support-phase assistance and swing-phase assistance a to provide the required target torque M for the wearer's hip joint.

Images