Multi-stage control method for lower limb exoskeleton

Abstract

The invention discloses a multi-stage control method for a lower limb exoskeleton, is applied to the field of exoskeleton robots, and aims to solve the problem that the design control precision is reduced due to the uncertainty of a model in the prior art. According to the controller designed by the invention, in a high-level control layer, a human body training mode is determined by a motion intention of an operator so as to generate a reference gait track; in the middle control layer, a variable admittance controller is designed, and three training modes of human exoskeleton cooperative movement, namely a passive mode, an active mode and a passive-to-active mode, are planned. In a low-level control loop, a method with a radial basis function neural network estimation function and a fixed-time convergence controller with input dead zone compensation are provided so as to ensure that an exoskeleton joint position tracks an expected trajectory output by an admittance loop. In order to avoid the Zeno phenomenon of the designed controller, an event trigger mechanism is used to determine the execution time of sampling and transmitting signals. Finally, the effectiveness of the proposed control strategy is verified through simulation and experimental results.

Description

Technical Field

[0001] The present invention belongs to the field of exoskeleton robots, and in particular relates to a human exoskeleton control technology. Background Art

[0002] As a representative of multidisciplinary systems, robotics has become popular across a wide range of fields, driving economic development and technological advancement. Wearable robots are a typical example of high-load dynamic human-machine coupling systems. Exoskeleton robots, as a typical wearable robot, form a human-machine coupling system in combination with the human body. Combining the advantages of a robot's high mechanical strength and high-load dynamics with the human's environmental perception and task analysis capabilities, they possess enormous potential and broad application value in areas such as medical rehabilitation, military deployment, industrial production, and assistance for the elderly and the disabled. Control strategies are primarily categorized into operator-passive control and operator-cooperative control, depending on the operator's level of involvement. In practice, exoskeleton assistance should be adjusted based on movement intent and the operator's needs. Furthermore, exoskeleton robots should possess safety features and be able to assess the operator's wearable comfort.

[0003] For patient-passive control, exoskeletons are used for severely impaired hemiplegic patients without any motor abilities and follow a predefined trajectory. Therefore, a reliable position controller should be designed to ensure tracking performance. Furthermore, the effectiveness of the admittance controller depends on the performance of the low-level position controller. Therefore, the design of a high-performance position controller is crucial for lower-limb exoskeletons. However, whether it is an adaptive cooperative control strategy that ensures accurate joint torque estimation and position control, or an extended state observer (ESO) with backstepping iteration to compensate for unmeasured system states, model uncertainties, and unmodeled dynamics, these methods typically require precise model parameters. In fact, model uncertainties can reduce the control accuracy of the design. Therefore, finding an intelligent control algorithm to handle unknown model uncertainties is of great significance.

[0004] Traditional time-triggered control methods require high update frequencies and transmissions, which can lead to network congestion, low overall system performance and energy efficiency, and are difficult to address communication resource constraints.

[0005] Generally speaking, most control methods can make the tracking error converge to the origin within a finite time. However, in many practical engineering applications, the tracking error only converges to a small neighborhood within a finite time. Furthermore, in many practical systems, information such as the initial state is difficult to obtain. Therefore, it is urgent to find a controller that converges within a specified time interval. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention proposes a multi-level control method for a lower limb exoskeleton.

[0007] The technical solution adopted by the present invention is: a multi-level control method for a lower limb exoskeleton, based on a multi-level controller including: a high-level control layer, an intermediate admittance control layer and a low-level position control layer; the high-level control layer generates a reference gait trajectory based on the intention of the person wearing the lower limb exoskeleton; the intermediate admittance control layer obtains the output of the intermediate admittance control layer based on the reference gait trajectory generated by the high-level control layer, the lower limb exoskeleton joint position data, and the lower limb exoskeleton interaction torque; the low-level position control layer obtains a control torque with an input dead zone based on the output of the intermediate admittance control layer and RBFNN and event triggering, and the control torque with the input dead zone acts on the lower limb exoskeleton to control the movement of the lower limb exoskeleton.

[0008] Compared with the prior art, the present invention has the following beneficial effects:

[0009] 1. Traditional time-triggered control methods require high update frequencies and transmissions, which can lead to network congestion. To conserve communication resources, this paper proposes an event-triggered method to reduce update frequencies. The event-triggered mechanism (ETM) can significantly reduce unnecessary data transmission and computation, thereby improving overall system performance and energy efficiency and addressing communication resource constraints.

[0010] 2. This invention designs a fixed-time convergence control method with input dead-zone compensation. Generally, most control methods can converge the tracking error to the origin within a finite time. However, in many practical systems, obtaining information such as the initial state is difficult. The fixed-time control method proposed in this invention can address this issue. Furthermore, to compensate for the dead-zone constraint, this invention also proposes a robust adaptive radial basis function neural network (RBFNN)-based method to reduce its impact on the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a multi-level control strategy for human-exoskeleton collaborative motion;

[0012] Figure 2 It is the operator gait experiment and corresponding data processing;

[0013] Among them, (a) is a typical general gait of human daily movement, (b) is a human gait experiment used to measure 3D motion and plantar force, and (c) is the experimental data of LSTM and Opensim.

[0014] Figure 3 The joint trajectory of the operator is sampled and the torque is estimated through the LSTM network.

[0015] Figure 4 It is the evaluation metric of the LSTM network.

[0016] Figure 5 It is the hip joint position response in PM training mode.

[0017] Figure 6 It is the related control torque of input dead zone and the event triggering time interval of two joints in PM training mode.

[0018] Figure 7 It is the exoskeleton joint response in AM training mode.

[0019] Figure 8 It is the exoskeleton joint response in PAM training mode.

[0020] Figure 9 It is the IAE indicator of different training modes.

[0021] Figure 10 It is the exoskeleton joint response in PM training mode.

[0022] Figure 11 In PAM training mode, the input dead zone is the time interval between event triggers of two joints.

[0023] Figure 12 It is a two-degree-of-freedom lower limb exoskeleton platform.

[0024] Figure 13 It is the human-exoskeleton interaction torque in PM training mode.

[0025] Figure 14 In the AM training mode, the human-exoskeleton interaction torque and the operator's active torque are estimated through LSTM. DETAILED DESCRIPTION

[0026] To facilitate those skilled in the art to understand the technical content of the present invention, the present invention is further explained below with reference to the accompanying drawings.

[0027] The present invention proposes an event-triggered fixed-time fuzzy control method for uncertain nonlinear systems with predetermined performance. The method includes the following steps: first, establishing a two-degree-of-freedom lower-limb exoskeleton dynamic model with joint friction and input dead zone. Secondly, the design of a multi-level controller begins by defining the compliance modulation factor and time-varying parameters in the high-level control layer, which are used to adjust the admittance model parameters and change the influence of the admittance controller to achieve different mode changes. Then, the middle-level admittance control layer is designed to reduce the interaction torque during human-exoskeleton collaborative motion. Finally, a fixed-time convergence controller with input dead zone compensation is designed in the low-level position control layer. Furthermore, an event-triggered mechanism (ETM) is used in backstepping iterations to reduce communication resources and avoid the Zeno phenomenon in the designed controller. Finally, simulations and experiments are performed to demonstrate the proposed method.

[0028] like Figure 1 As shown in FIG, the control system based on the method of the present invention comprises four parts, wherein A is a high-level control layer, B is an intermediate admittance layer, C is a low-level position control layer, and D is a human exoskeleton collaborative motion platform. The implementation process of the method of the present invention comprises the following steps:

[0029] S1: Construct a two-degree-of-freedom lower limb exoskeleton dynamic model with joint friction and input dead zone. The dynamic model expression is as follows:

[0030]

[0031] Where q∈R 2 、 and are the exoskeleton joint position, velocity and acceleration respectively, and are all two-dimensional vectors (R 2 ), where the elements in the two-dimensional vector are real numbers; M(q), and G(q)∈R 2 They are the inertia matrix, Coriolis term and gravity term respectively, where the first two terms are 2×2 matrices and the gravity term is a two-dimensional vector. is the joint friction term, τ ext is the human-exoskeleton interaction torque, and D(τ)∈R 2 is the control torque with input dead zone, which is also a two-dimensional vector; D(τ) is the dead zone effect combination of all joint control inputs.

[0032] The specific lower limb exoskeleton system in this embodiment can refer to the patent application with patent application number 202111332323.7. The lower limb exoskeleton system is an existing known technology and will not be introduced in detail in this invention.

[0033] The input dead zone model D of the i-th joint in D(τ)i (τ i ) is designed as follows:

[0034]

[0035] where h i,1 >0,h i,2 >0,g i,1 <0,g i,2 >0,τ i represents the input of the system, h i is the slope parameter, g i is the intercept parameter, h i,1 and h i,2 Indicates the slope in different input ranges, g i,1 and g i,2 is the threshold that defines the deadband range.

[0036] S2: The design of the multi-level controller is as follows:

[0037] The three-layer multi-level control strategy for human-exoskeleton collaborative motion is as follows: Figure 1 As shown, it includes: a high-level control layer, a middle-level admittance control layer, and a low-level position control layer. It has two typical tasks:

[0038] (1) Figure 2 As shown in , sufficient gait trajectories are generated according to different training modes of the operator, such as Figure 2 As shown in Figure A, the operator's training modes can include jogging and sprinting. In this embodiment, five teaching gait trajectories are collected for dynamic time alignment. These five trajectories contain multiple gait cycles, covering natural variations in walking (such as speed and amplitude differences). The gait cycle is obtained based on different sensor acquisition cycles. The 3D motion capture system in this embodiment specifically collects data using SEMG (surface electromyography) sensors, IMU sensors, and plantar pressure sensors.

[0039] (2) Ensure high exoskeleton joint tracking accuracy and low human-exoskeleton interaction torque, ensuring wearable comfort for the operator.

[0040] At the high-level control layer, the present invention defines a compliance modulation factor γ(t) and a time-varying parameter η(t). γ(t) is used to adjust the parameters of the admittance model and alter the influence of the admittance controller. Different modes of change can be achieved based on γ(t) and η(t). η(t) is used to adjust the reference trajectory of the admittance controller, ensuring a slow trajectory change during actual experiments and improving the safety of the exoskeleton system. Three different training modes are then determined by the operator's movement intention to generate reference gait trajectories.

[0041] PM is the passive mode, AM is the active mode, and PAM: initially the exoskeleton leads the movement, but as the training progresses, the system will gradually reduce the exoskeleton's assistance and increase the operator's active participation, thereby helping the operator gradually recover autonomous movement ability. Table 1 shows the corresponding modulation parameters γ(t) and η(t) in each mode such as PM, AM, and PAM. Since the gait data q sampled in the real-time environment is h With obvious time series, the operator active torque τ is estimated by LSTM h To evaluate the performance of wearable comfort operators.

[0042] Table 1 Modulation parameters in different training modes

[0043]

[0044] The gait data of four operators were collected by 3D motion capture system, including hip and knee joint angles and time series data such as Figure 4 As shown, Figure 3 As shown in Figure 1, the operator's active torque is estimated using an LSTM network, and the wearable comfort index J is calculated in combination with the interaction torque. The evaluation index J of the wearable comfort performance in human-exoskeleton collaborative motion is defined as follows:

[0045]

[0046] in is the LSTM estimate of the effective torque of the simulated partial operator. τ ext is the human-exoskeleton interaction torque, which is calculated by the spring-damper model in the simulation part or measured by the 3D force sensor in the experiment part, μ i (i=1, 2, 3) are designed weight parameters.

[0047] In PM passive mode, the exoskeleton strictly follows

[0048] Track definition trajectory Figure 5 As shown, the interaction torque is low as Figure 13 As shown in the figure, the operator's active participation in this mode is 0, and the comfort level J depends only on the human-machine interaction torque. In the AM active mode, the exoskeleton completely follows the operator's intention. Figure 7 However, if the operator's muscle strength is insufficient, the interaction torque will increase as shown in Figure 14 As shown in Figure 2, J deteriorates and the wearability of the exoskeleton decreases. In the passive-active transition mode of PAM, the exoskeleton assistance is gradually reduced by dynamically adjusting γ(t) and η(t). Figure 8 shown.

[0049] S3: For the intermediate admittance layer, in order to reduce the interaction torque in the human-exoskeleton collaborative motion, the exoskeleton admittance model is constructed as a mass-spring-damper model, so that

[0050]

[0051] Where △q=q d -q r ,q d ,q r ∈R 2 are the output and input of the admittance loop; where q d and q r is a two-dimensional vector, q d and q r They represent the output and input of the admittance loop, both of which belong to the two-dimensional real space, namely R 2 represents a two-dimensional real space. M, B, and K are the inertia matrix, damping matrix, and stiffness matrix, respectively. 0 ≤ γ(t) ≤ 1 is the compliance modulation factor, as shown in Table 1.

[0052] If the human-exoskeleton torque τ is measured in a real-time environment ext and plans the reference joint positions q determined by the operator motion r , then the regulation of △q in the admittance loop is calculated by the double integral of the mass-spring-damper model formula.

[0053] At the same time, the time-varying parameter η(t) is intended to adjust the output trajectory q of the admittance loop d , making

[0054]

[0055] where q h is the operator’s motion trajectory, 0≤η(t)≤1, as shown in Table 1.

[0056] S4: For the low-level position control layer, backstepping iteration: the system state error of the exoskeleton system is defined as:

[0057]

[0058] where z = [z1, z2] T ∈R 2 is the position tracking error, e=[e1,e2] T ∈R 2 is the velocity error, q d =[q d1 ,q d2 ] T ∈R 2 is the exoskeleton expected value output by the admittance loop, α=[α1,α2]T ∈R 2 is a dummy control variable.

[0059] Define a cascaded Lyapunov function as follows:

[0060]

[0061] According to the Lyapunov theorem, in order to make the system stable, the designed Lyapunov function must be positive definite, so according to the state error z=[z1,z2] defined in the backstepping iteration T ∈R 2 and velocity error e=[e1,e2] T ∈R 2 Design Lyapunov function.

[0062] Then, the derivative of V1 is:

[0063]

[0064] At the same time, the dummy control variable α i Designed to:

[0065]

[0066] where b 1i 、c 1i >0 is a design parameter, sgn(·) is a sign function. Substituting the above formula into It turns out that:

[0067]

[0068] Secondly, the derivative of V2 is:

[0069]

[0070] in and z = [z1,z2] T , M, C, G, are the inertia matrices in the exoskeleton dynamics modeling (M(q)∈R 2×2 ), Coriolis matrix Gravity term (G(q)∈R 2 ).

[0071] RBFNN consists of an input layer, a hidden layer, and an output layer. Each neuron in the hidden layer uses a radial basis function (such as a Gaussian function) as an activation function to output the input layer. The output layer is a linear weighted combination of the hidden layer outputs. RBFNN is used to approximate the unknown model function F. i (X i ), as shown below:

[0072] F i (X i )=W i *T Φ i (X i )+ε i (X i ), i=1,2

[0073] where ε i (X i ) is the approximation error, so There is no actual physical meaning here, it is only used to limit the error to converge to any number greater than 0, that is, X i =[x1,...,x n ] T ∈R n is the input vector, Φ i (X i )=[Φ i,1 (X i ),...,Φ i,n (X i )] T ∈R N (N>1) is the radial basis function vector, N is the node number, and W i =[W i,1 ,...,W i,N ] T ∈R N is the ideal weight vector.

[0074] and Among them C i ∈R n ,ω i ∈R 1 (i=1,…,N) are Gaussian functions Φ i (X) Center and width. W i * represents the ideal weight matrix (unknown), W i Represents the actual estimated weight matrix, the goal is to approximate W i * ,f(X i ) represents the nonlinear uncertainty of the i-th joint in the exoskeleton system, which is approximated by a radial basis function neural network.

[0075] In the model used in this embodiment, Φ is defined i (X i )=[Φ i,1 (X i ),Φ i,2 (X i)] T , where Φ i (X i ) is the hidden layer output vector of the radial basis function neural network (RBFNN), which represents the input X i The nonlinear mapping result, Φ i,j (X i )(j=1,2) the output of the j-th radial basis function neural network (RBFNN), The ideal weight values of the j-th radial basis function neural network (RBFNN).

[0076] According to Young's inequality, we have

[0077]

[0078] where θ i =||W i * || 2 is an unknown constant, a i >0. and e i F i Substitution get:

[0079]

[0080] An event-triggered mechanism (ETM) is designed for use in the low-level position control layer. It can dynamically adjust the update frequency of the control signal and reduce the computation and communication burden. To reduce the use of communication resources, the event-triggered mechanism is designed as follows:

[0081]

[0082] in is a temporary control variable, τ i (t) is the actual control variable with ETM, m i (t) = ω i (t)-τ i (t) is the measurement error, δ i >0 is used to control the smoothness of the trigger function, 0<ρ i <1 is the proportional coefficient for adjusting the trigger threshold, s i >0 is a fixed offset of the trigger threshold, t i,k is the event trigger moment, k∈Z + , t i,1 is the initial moment. Those skilled in the art will know that i (t) is a continuous function, ω i (t i,k ) is a discrete function.

[0083] Dead zone parameter estimation: Consider the input dead zone model D of the servo motor mentioned above. i (τ i ), unconstrained control variables The design is as follows:

[0084]

[0085] in is the control variable with input deadband, and the matrix superscript T represents the transpose of the matrix.

[0086] In practice, the two input deadband parameters h i and g i is unknown. Therefore, an adaptive law is used to estimate the parameter matrix H i =[H i,1 ,H i,2 ] T , where H i,1 =(1 / h i ) and H i,2 =-(g i / h i ).

[0087] If the constraint control variable It was designed later, as mentioned in the formula designed by ETM above. Obtained by the following formula

[0088]

[0089] in It is H i Parameter estimation. According to the formula designed by ETM, the actual control variable u with ETM i Conclusion

[0090]

[0091] where |γ i,1 (t)|≤1 and |γ i,2 (t)|≤1 are two time-varying parameters. Therefore, ω in the ETM design formula is i Substitute (t) into τ i In, we get

[0092]

[0093] By using the generalized Young's inequality and the hyperbolic tangent inequality, we can get the value of e i τ i Simplified as follows:

[0094]

[0095] Then, e i τ i Substitution satisfy

[0096]

[0097] Taking into account Weight estimation error and dead zone parameter The estimation error (the input of RBFNN is a combination of system states, specifically: q and The actual position and velocity of the exoskeleton joints (2D vectors, corresponding to the hip and knee joints), q d,i , The desired joint positions and velocities (outputs from the mid-level admittance controller); the output of the RBFNN is F i (X i ), which is used to approximate the unknown linear terms in the system dynamics. The total cascade Lyapunov function V3 is defined as:

[0098]

[0099] in θ i is the square norm of the ideal weight of RBFNN, is θ i The estimated value of is obtained through the subsequent adaptive law formula, is the estimation error, reflecting the accuracy of RBFNN weight approximation, H i is the real vector of dead zone parameters, H i The estimated value of is obtained through the adaptive law formula in the following text. is the dead zone parameter estimation error, K i is a positive definite matrix, r i is a constant.

[0100] Then, the derivative of V3 yields

[0101]

[0102] For fixed-time convergence controller design: the actual control variable τ mentioned in the ETM design of this embodiment i (t) involves three terms, namely the constraint control variable ζ i , and two adaptive laws and Constrained control variable ζ i First, consider designing the fixed-time convergence term. and They are respectively the high-order error term and the sign function term, which are designed to ensure the tracking error e i Converges to a zero neighborhood in constant time, parameter b 2i , c 2i and p are used to adjust the convergence speed and boundary; secondly, the design of the neural network compensation term is considered. The uncertainty of the model is estimated by RBFNN; finally, the compensation items such as external interference suppression are considered and the design The term cancels out the Lyapunov function item to eliminate the impact on the system.

[0103] Adaptive Law In the design The term is used to adjust the weight parameters according to the tracking error and the neural network output to compensate for the model uncertainty. In addition, the damping term is introduced. Ensure that the estimate is bounded, Accelerate parameter convergence and suppress oscillations for high-order damping terms. Adaptive law Design of -e i K i Q i The term is the dead zone compensation term, which compensates for the nonlinear effect of the input dead zone, and The term is a linear damping term to prevent parameter divergence. is a high-order nonlinear damping term that enhances parameter convergence and suppresses chattering. The designed constraint control variable ζ i And two adaptive laws and As shown below:

[0104]

[0105] in is the speed error, b 2i and c 2i is a positive design parameter used to adjust the rate and robustness of fixed-time convergence, p is an exponential parameter to ensure fixed-time stability, and a i is the adjustment parameter of RBFNN approximation error, is the square norm of the RBFNN weight estimation error, Φ i is the basis function vector of RBFNN, τ ext is the human-computer interaction torque. i is the adaptive gain, λ 1i and λ 2i is the damping coefficient, used to prevent parameter drift, ||Φ i || 2 It is the norm of the RBFNN basis function, reflecting the uncertainty of the model. iis the estimated matrix of input dead zone parameters, K i is a positive definite matrix, adjusting the adaptive rate, Q i is the control variable correlation vector, η 1i and η 2i is the damping term coefficient used to enhance robustness.

[0106] Through relevant theoretical derivation, all states and estimation errors z i , e i , (i=1,...,n)∈Z converges to the following residual set:

[0107]

[0108] Where Z represents an integer set, δ is the upper bound of the disturbance, l and θ are controller parameters, and p is the exponential parameter. and are all bounded, because e i 、z i ,q d 、H i 、 and At the same time, the virtual control quantity α and the actual control input are bounded due to their relatively bounded parameters. In addition, according to the error definition and event triggering mechanism, ω i and u i is bounded. Therefore, all closed-loop signals are bounded.

[0109] Figures 5 to 11 The simulation results reflect the joint response, position response, torque and other related data of the exoskeleton in various modes such as PM, AM, and PAM, which indirectly confirms that the two-degree-of-freedom exoskeleton system can track the desired trajectory within a fixed time, and at the same time make the tracking error converge to meet the predetermined performance conditions. It has good control accuracy, meets the safety performance of the exoskeleton robot and the indicators for evaluating the wearable comfort of the operator, and effectively saves communication resources.

[0110] The mechanical structure of the two-degree-of-freedom exoskeleton plate used in the experiment is as follows Figure 10 As shown. The hip joint is fixed on a stainless steel bracket to ensure the stability of the prototype. In order to effectively perform human-exoskeleton collaborative motion tasks and ensure the operator's wearable comfort, the exoskeleton system is connected to the human body through an elastic belt. In order to ensure the safety of the operator, the exoskeleton constrains a reasonable range of motion by designing mechanical hard limits. According to the range of motion of the human lower limb joints, the hard limit of the hip joint is set to [-20°, 120°] and the knee joint is [-140°, 10°]. The two-degree-of-freedom exoskeleton platform consists of hardware and software, such as Figure 10 As shown in Figure 2, a core NI controller (cRIO-9035NI) is used to process sensor signals and send control commands. A PC (Thinkpad-T490, i5-10210U CPU, 1.60GHz) is used to provide the operator with a visual interface for training the exoskeleton. The exoskeleton is driven by two servo motors (GDM1-100N2 / 120N2) and actuators (Elmo-G-SOLHOR15 / 100EE). Two absolute encoders (INC-4-150 and INC-3-125) are used to measure the angles of the two joints, and four 3-D force sensors are used to measure the torque of the human-exoskeleton coupling. The software is based on MATLAB and LABVIEW. MATLAB is used to design and compile the control algorithm and generate a dynamic link library (DLL) to be embedded in LABVIEW for building the control block diagram, data display, and preprocessing.

[0111] Those skilled in the art will appreciate that the embodiments described herein are intended to aid the reader in understanding the principles of the present invention, and it should be understood that the scope of the present invention is not limited to such specific descriptions and embodiments. Various modifications and variations are readily apparent to those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims.

Claims

1. A multi-level control method for a lower limb exoskeleton, characterized in that: The multi-level controller it is based on includes: a high-level control layer, an intermediate admittance control layer and a low-level position control layer; the high-level control layer generates a reference gait trajectory based on the intention of the person wearing the lower limb exoskeleton; the intermediate admittance control layer obtains the output of the intermediate admittance control layer based on the reference gait trajectory generated by the high-level control layer, the lower limb exoskeleton joint position data, and the lower limb exoskeleton interaction torque; the low-level position control layer obtains a control variable with an input dead zone based on the output of the intermediate admittance control layer based on RBFNN and event triggering, and the control variable with an input dead zone acts on the lower limb exoskeleton to control the movement of the lower limb exoskeleton.

2. A multi-level control method for a lower limb exoskeleton according to claim 1, characterized in that: The advanced control layer generates a reference gait trajectory based on three training modes. Specifically, the advanced control layer defines a flexible modulation factor γ(t) and a time-varying parameter η(t). The values of γ(t) and η(t) corresponding to the first training mode are: γ(t) = 0 and η(t) = 0, the values of γ(t) and η(t) corresponding to the second training mode are: 0<γ(t)≤1 and η(t) = 1, and the values of γ(t) and η(t) corresponding to the third training mode are: 0<γ(t)≤1 and 0<η(t)<1.

3. A multi-level control method for lower limb exoskeleton according to claim 2, characterized in that: The intermediate admittance control layer is constructed as a mass-spring-damper model, which is expressed as follows: Where Δq(t)=q d (t)-q r (t),q d (t) and q r (t) represents the output and input of the admittance loop, M a 、B a and K a They are inertia, damping and stiffness respectively.

4. A multi-level control method for a lower limb exoskeleton according to claim 3, characterized in that: q d The calculation formula for (t) is: Among them, q h is the operator's motion trajectory.

5. The multi-level control method for lower limb exoskeleton according to claim 4, characterized in that: The control variable output by the low-level position control layer is expressed as τ i (t), τ i (t) includes three terms: constraint control variable ζ i , and two adaptive laws and in is the velocity error, q represents the actual position of the skeletal joint, and b 2i and c 2i is a positive definite design parameter, p is an exponential parameter, a i is the adjustment parameter of RBFNN approximation error, is the square norm of the RBFNN weight estimation error, Φ i is the basis function vector of RBFNN, τ ext is the human-computer interaction torque, r i is the adaptive gain, λ 1i and λ 2i is the damping term coefficient, ||Φ i || 2 is the norm of the RBFNN basis function, H i is the real vector of dead zone parameters, H i The estimated value of , the superscript T indicates the transpose, K i is a positive definite matrix, Q i is the control variable correlation vector, η 1i and η 2i is the damping term coefficient used to enhance robustness.

6. The multi-level control method for lower limb exoskeleton according to claim 5, characterized in that: The update frequency of the control variables output by the low-level position control layer is: in is a temporary control variable, τ i (t) is the actual control variable with event triggering, m i (t) = ω i (t)-τ i (t) is the measurement error, δ i Used to control the smoothness of the trigger function, δ i >0,ρ i Is the proportional coefficient for adjusting the trigger threshold, 0<ρ i <1,s i is the fixed offset of the trigger threshold, s i >0,t i,k is the event trigger moment, k is an integer.

Images