High-precision anti-interference tracking control method for waist exoskeleton

Abstract

The invention discloses a high-precision anti-interference tracking control method for a waist exoskeleton, and belongs to the technical field of exoskeleton robot control. According to the method, a combined control strategy based on a radial basis function neural network and a universal nonlinear disturbance observer is provided for solving the problems that a traditional exoskeleton dynamic model is insufficient in precision and the trajectory tracking performance is affected by external disturbance. The real-time compensation of model errors is realized by constructing a parallel exoskeleton dynamic model framework and utilizing a radial basis function neural network to approach unmodeled dynamic and parameter uncertain components in a system model in an online manner; meanwhile, human-computer interaction interference and external environment disturbance generated in the exoskeleton wearing process are observed and dynamically counteracted through a universal nonlinear interference observer. According to the method, the control precision under the external interference of the exoskeleton is effectively improved, the safety and control robustness of rehabilitation training are improved, and a novel control normal form is provided for the wearable medical robot.

Description

Technical Field

[0001] The present invention relates to a high-precision disturbance rejection tracking control method for a waist exoskeleton, belonging to the technical field of exoskeleton robot control. Background Art

[0002] In the field of waist rehabilitation training, exoskeleton robots need to achieve sub-millimeter-level motion trajectory tracking accuracy, which poses extremely high requirements for the control system. The current technology mainly faces three major challenges:

[0003] First of all, there are significant difficulties in system modeling. The exoskeleton has complex multi-body dynamics characteristics, including time-varying structural parameters, non-linear joint friction, and dynamically changing load characteristics, resulting in obvious deviations between the theoretical model and the actual system. This model mismatch will seriously affect the control accuracy. Secondly, the working environment is full of uncertainties. The muscle force fluctuations, sudden spasm movements of patients during the rehabilitation process, and environmental disturbances will form a composite disturbance source. These disturbances have characteristics such as strong randomness and wide frequency band, and traditional control methods are difficult to effectively cope with. Moreover, the adaptability of existing control strategies is insufficient. The PID controller with fixed parameters cannot meet the needs of different stages of rehabilitation: high auxiliary torque is required at the initial stage to ensure safety, and the auxiliary needs to be reduced at the later stage to promote active recovery. This rigid control method not only affects the curative effect but also has potential safety hazards.

[0004] Currently, the control schemes adopted clinically have obvious limitations: the observer based on linear assumptions has inaccurate estimation of non-linear disturbances; the single feed-forward compensation has a lag in response to sudden disturbances. This results in the tracking error of existing systems often being above the millimeter level, making it difficult to meet the needs of fine rehabilitation. Summary of the Invention

[0005] In order to reduce the system tracking error and improve the accuracy of exoskeleton tracking control, the present invention provides a high-precision disturbance rejection tracking control method for a waist exoskeleton, and the technical solutions are as follows:

[0006] Step 1: Perform dynamic modeling on the parallel waist exoskeleton, and separate the uncertain part of the modeling and external disturbances;

[0007] Step 2: Based on the constructed dynamic model of the parallel waist exoskeleton, use a radial basis neural network to achieve local approximation of the uncertain part of the modeling;

[0008] Step 3: Based on the constructed dynamic model of the parallel waist exoskeleton, use a radial general non-linear disturbance observer to estimate the external disturbance force of the exoskeleton, and feedback for compensation.

[0009] Optionally, the dynamic model of the parallel waist exoskeleton is:

[0010]

[0011] In the formula, M q is the inertia matrix of the exoskeleton in the joint space, C q is the Coriolis force and centrifugal force matrix of the exoskeleton in the joint space, G q is the gravity term of the exoskeleton in the joint space, F qf is the friction force of the exoskeleton in the joint space, τ is the joint force, F qext is the external disturbance, are respectively the actual acceleration and actual velocity of the exoskeleton joint.

[0012] Optionally, the process of using a radial basis neural network to design the control law in step 2 includes:

[0013] Let the parameters of the nominal model of the exoskeleton be M q0 , C q0 , G q0 , F qf0 respectively. For the nominal model, the trajectory tracking control law based on a PD controller is obtained:

[0014]

[0015] where, K d represents the proportional gain matrix, K p represents the differential gain matrix;

[0016] Substitute the trajectory tracking control law formula into the dynamic model of the parallel waist exoskeleton to obtain:

[0017]

[0018] where, is the acceleration error, represents the velocity error, e q = q - q d represents the trajectory error, q d are respectively the desired acceleration, angular velocity and trajectory of the exoskeleton joint; q represents the actual trajectory of the exoskeleton joint;

[0019] Let:

[0020]

[0021] Then u is the uncertain part of the dynamic modeling, where, ΔM q = M q0 - M q , ΔC q = C q0 - C q , ΔG q = G q0 - G q, ΔF qf = F qf0 - F qf ;

[0022] The radial basis neural network is used to approximate u, and the radial basis neural network algorithm is as follows:

[0023]

[0024] where x is the network input, is the neural network output, is the estimated value of the optimal neural network weight, is the output of the Gaussian basis function:

[0025]

[0026] where c i is the center of the Gaussian function, and σ i is the variance of the Gaussian function;

[0027] Assume that the neural network output is continuous, and when the neural network weight is the optimal neural network weight W * there is an ideal output:

[0028]

[0029] Satisfy that for any sufficiently small positive real number ε, ‖ξ‖ ≤ ε, where ξ is the ideal approximation error, its bound is ξ 0 = sup‖ξ‖;

[0030] According to the above information, the control law based on the radial basis neural network is expressed as:

[0031] τ = Τ 1 + Τ 2

[0032] where,

[0033] Optionally, the estimation process of the external disturbance force in step 3 includes:

[0034] According to the parallel waist exoskeleton dynamics model, take x 1 = q, and its system state equation is expressed as:

[0035]

[0036] Let the estimation of the external disturbance be Then the estimation error of the external disturbance is:

[0037]

[0038] Since there is no prior knowledge to support the true external disturbance, without loss of generality, let its derivative be 0, that is:

[0039]

[0040] Let \(L(x)\in R\) i×i be the observation gain matrix, and the control law is designed as:

[0041]

[0042] Substituting the system state equation into the above control law, we get:

[0043]

[0044] where \(f\) 1 (x)=C q x 2 +G q +F qf , subtracting

[0045]

[0046] from both sides of the above equation and integrating both sides, we get:

[0047]

[0048] where

[0049] Let an intermediate variable be derived as:

[0050]

[0051] The estimation of the external disturbance is obtained from the above observer for \(z\):

[0052]

[0053] Introduce term, at this time \(\rho(x)\) is \(L(x)M\) q x 2 , then the final observer is designed as:

[0054]

[0055] Adjust \(L(x)\) to obtain the ideal control effect.

[0056] The second object of the present invention is to provide a waist exoskeleton disturbance rejection tracking control system, including:

[0057] A kinetic modeling module, configured to perform kinetic modeling on a parallel waist exoskeleton, and isolate the uncertain part of the modeling and external disturbances;

[0058] A radial basis neural network, configured to locally approximate the uncertain part of the modeling using a radial basis neural network based on the constructed kinetic model of the parallel waist exoskeleton;

[0059] A feedback compensation module, configured to estimate the external disturbance force of the exoskeleton using a radial general nonlinear disturbance observer based on the constructed kinetic model of the parallel waist exoskeleton, and perform feedback compensation.

[0060] Optionally, the kinetic model of the parallel waist exoskeleton is:

[0061]

[0062] where M q is the inertia matrix of the exoskeleton in the joint space, C q is the Coriolis force and centrifugal force matrix of the exoskeleton in the joint space, G q is the gravity term of the exoskeleton in the joint space, F qf is the friction force of the exoskeleton in the joint space, τ is the joint force, F qext is the external disturbance, are respectively the actual acceleration and actual velocity of the exoskeleton joint.

[0063] Optionally, the process of designing a control law using a radial basis neural network includes:

[0064] Let the parameters of the nominal model of the exoskeleton be M q0 , C q0 , G q0 , F qf0 . For the nominal model, a trajectory tracking control law based on a PD controller is obtained:

[0065]

[0066] where K d represents the proportional gain matrix, and K p represents the differential gain matrix;

[0067] Substitute the trajectory tracking control law formula into the kinetic model of the parallel waist exoskeleton to obtain:

[0068]

[0069] where, is the acceleration error, represents the velocity error, eq = q - q d represents the trajectory error, q d are the desired acceleration, angular velocity, and trajectory of the exoskeleton joint respectively; q represents the actual trajectory of the exoskeleton joint;

[0070] Let:

[0071]

[0072] Then u is the uncertain part of the dynamic modeling, where, ΔM q = M q0 - M q , ΔC q = C q0 - C q , ΔG q = G q0 - G q , ΔF qf = F qf0 - F qf ;

[0073] Use a radial basis neural network to approximate u, and the radial basis neural network algorithm is:

[0074]

[0075] where, x is the network input, is the neural network output, is the estimated value of the optimal neural network weight, is the output of the Gaussian basis function:

[0076]

[0077] where c i is the center of the Gaussian function, and σ i is the variance of the Gaussian function;

[0078] Assume that the neural network output is continuous, and when the neural network weight is the optimal neural network weight W * there is an ideal output:

[0079]

[0080] Satisfies that for any sufficiently small positive real number ε, there is ‖ξ‖ ≤ ε, where ξ is the ideal approximation error, and its bound is ξ 0 = sup‖ξ‖;

[0081] According to the above information, the control law based on the radial basis neural network is expressed as:

[0082] τ = Τ 1 + Τ 2

[0083] Wherein,

[0084] Optionally, the process of estimating the external disturbance force by the feedback compensation module includes:

[0085] According to the parallel waist exoskeleton dynamics model, let x 1 = q, Its system state equation is expressed as:

[0086]

[0087] Let the estimate of the external disturbance be Then the estimation error of the external disturbance is:

[0088]

[0089] Since there is no prior knowledge to support the true external disturbance, without loss of generality, let its derivative be 0, that is:

[0090]

[0091] Let L(x) ∈ R i×i be the observation gain matrix, and the control law is designed as:

[0092]

[0093] Substitute the system state equation into the above control law to get:

[0094]

[0095] Wherein, f 1 (x) = C q x 2 + G q + F qf , subtract

[0096]

[0097] Integrate both sides of the above equation to get:

[0098]

[0099] Wherein,

[0100] Let an intermediate variable Take the derivative of it to get:

[0101]

[0102] The estimation of the external disturbance is obtained from the above observer for z:

[0103]

[0104] Introduce term, where ρ(x) is L(x)M q x 2 , then the final observer is designed as:

[0105]

[0106] Adjust L(x) to obtain the ideal control effect.

[0107] The third object of the present invention is to provide a waist exoskeleton disturbance rejection tracking control device, including:

[0108] A sensor module for real-time acquisition of joint position, velocity, acceleration and external force signals;

[0109] A processor module that executes the waist exoskeleton disturbance rejection tracking control method described above and calculates the joint control torque;

[0110] A driving module for driving the exoskeleton joint movement according to the joint control torque.

[0111] The fourth object of the present invention is to provide a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the waist exoskeleton disturbance rejection tracking control method described above is implemented.

[0112] The beneficial effects of the present invention are:

[0113] Through the innovative composite control architecture, the present invention significantly improves the control performance and safety of the waist rehabilitation exoskeleton system. Specifically, it has the following outstanding advantages:

[0114] In terms of intelligent adaptability, the present invention uses a radial basis neural network to perform real-time online learning and compensate for the unmodeled parts in the system dynamics, including complex factors such as time-varying parameters and non-linear friction, effectively solving the problem of accuracy degradation caused by model mismatch in traditional control methods.

[0115] In terms of anti-interference ability, the present invention designs a general non-linear disturbance observer to dynamically estimate and compensate for external disturbances such as patient muscle force fluctuations and sudden spasm movements, achieving fast response and precise suppression of composite disturbances.

[0116] In terms of precision control, through the collaborative work of the radial basis neural network and the general non-linear disturbance observer, combined with the basic PD control, the present invention forms a "feedforward-feedback-compensation" closed-loop control structure, which improves the system trajectory tracking precision to the sub-millimeter level.

[0117] In terms of safety guarantee, the adaptive control strategy can intelligently adjust the assistance intensity according to the patient's rehabilitation stage, providing sufficient protection in the initial stage of training and promoting active recovery in the later stage, ensuring the safety and effectiveness of the entire rehabilitation process.

[0118] The present invention effectively solves the problems of poor model adaptability and weak anti-interference ability in the prior art, effectively improves the control accuracy, can provide a more accurate, safe and personalized rehabilitation training experience for patients, and has significant clinical application value. Brief Description of the Drawings

[0119] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0120] Figure 1 is a flowchart of the high-precision disturbance rejection tracking control method for the lumbar exoskeleton of the present invention.

[0121] Figure 2 is a schematic diagram of the implementation of the high-precision disturbance rejection tracking control method for the lumbar exoskeleton provided in Embodiment 1 of the present invention.

[0122] Figure 3 is a comparison chart of errors between the traditional single radial basis control method provided in Embodiment 1 of the present invention and the method of the present invention under different external disturbances. Detailed Embodiments

[0123] To make the objectives, technical solutions and advantages of the present invention clearer, the following will further describe the embodiments of the present invention in detail with reference to the drawings.

[0124] Embodiment 1:

[0125] This embodiment provides a parallel lumbar exoskeleton composite control method based on the collaboration of a radial basis neural network and a general non-linear disturbance observer. As Figure 1 and Figure 2 shown, it includes the following steps:

[0126] Step 1: Perform dynamic modeling on the parallel lumbar exoskeleton, and separate the uncertain part and external disturbance of the modeling.

[0127] Step 2: Based on the established dynamic model of the parallel waist exoskeleton, use a radial basis neural network to achieve local approximation of the uncertain part of the modeling;

[0128] Step 3: Based on the established dynamic model of the parallel waist exoskeleton, use a radial general nonlinear disturbance observer to estimate the external disturbance force of the exoskeleton and feedback for compensation.

[0129] Among them, the dynamic model of the parallel waist exoskeleton in its joint space can be expressed as:

[0130]

[0131] In the formula, M q is the inertia matrix of the exoskeleton in the joint space, C q is the Coriolis force and centrifugal force matrix of the exoskeleton in the joint space, G q is the gravity term of the exoskeleton in the joint space, F qf is the friction force of the exoskeleton in the joint space, τ is the joint force, F qext is the external disturbance, are the actual acceleration and actual velocity of the exoskeleton joint respectively.

[0132] Let the parameters of the nominal model of the exoskeleton be M q0 , C q0 , G q0 , F qf0 respectively. Then, for the nominal model, the trajectory tracking control law based on the PD controller can be written as:

[0133]

[0134] Among them, K d represents the proportional gain matrix, and K p represents the differential gain matrix.

[0135] Substitute the trajectory tracking control law formula (2) into formula (1), and we can get:

[0136]

[0137] Among them, is the acceleration error, represents the velocity error, e q =q - q d represents the trajectory error, q d are the expected acceleration, angular velocity and trajectory of the exoskeleton joint respectively; q represents the actual trajectory of the exoskeleton joint.

[0138] Let:

[0139]

[0140] Then \(u\) is the uncertain part of the dynamic modeling, where \(\Delta M\) q \( = M\) q0 \(-M\) q , \(\Delta C\) q \( = C\) q0 \(-C\) q , \(\Delta G\) q \( = G\) q0 \(-G\) q , \(\Delta F\) qf \( = F\) qf0 \(-F\) qf ;

[0141] If the error between the actual model and the nominal model is not considered, that is, \(\Delta M\) q , \(\Delta C\) q , \(\Delta G\) q , \(\Delta F\) qf are all 0, then the above error equation can be transformed into:

[0142]

[0143] The above formula shows that trajectory tracking can be completed by a PD controller. However, it is difficult to achieve precise control of trajectory tracking only by a PD controller, and \(u\) is difficult to measure in actual engineering and needs to be further estimated and compensated.

[0144] Use a radial basis function neural network (RBFNN) to approximate \(u\). The radial basis function neural network algorithm is:

[0145]

[0146] where \(x\) is the network input, is the neural network output, is the estimated value of the optimal neural network weight, is the output of the Gaussian basis function:

[0147]

[0148] where \(c\) i is the center of the Gaussian function and \(\sigma\) i is the variance of the Gaussian function.

[0149] Assume that the neural network output is continuous, and when the neural network weight is the optimal neural network weight \(W\) * there is an ideal output:

[0150]

[0151] satisfies that for any sufficiently small positive real number \(\epsilon\), \(\|\xi\| \leq \epsilon\), where \(\xi\) is the ideal approximation error. Its boundary is ξ 0 = sup‖ξ‖. Based on the above information, the control law based on RBF can be designed as:

[0152] τ = Τ 1 + Τ 2 (9)

[0153] where

[0154] Substituting Equation (9) into Equation (1), the error state equation can be obtained as:

[0155]

[0156] Take The above equation can be written as:

[0157]

[0158] where the PD control matrix The coefficient matrix B = (0I) T , and I is the identity matrix.

[0159] Let Make the following transformation:

[0160]

[0161] In the formula, Then Equation (11) can be written as:

[0162]

[0163] Define the Lyapunov function as:

[0164]

[0165] where γ > 0, is the trace of the matrix . Since the real part of the eigenvalues of matrix A is negative, there exist positive definite matrices P and Q that satisfy the Lyapunov equation PA + A T P = -Q.

[0166] Taking the derivative of both sides of Equation (14) gives:

[0167]

[0168] where Then the above equation can be transformed into:

[0169]

[0170] The adaptive law can be taken as Let be 0, Equation (16) can be transformed into:

[0171]

[0172] Let λ Pmax be the largest eigenvalue of matrix P, and λ Qmin be the smallest eigenvalue of matrix Q, then:

[0173]

[0174] That is, when , there is L y ≥ 0, and when and only when , That is, when , According to the LaSalle invariance principle, the closed-loop system is asymptotically stable.

[0175] The above has completed the design and derivation of the control algorithm based on the radial basis neural network. Next, a general nonlinear disturbance observer will be designed to further compensate for the external disturbance force.

[0176] From Equation (1), taking x 1 = q, its system state equation can be written as:

[0177]

[0178] At this time, represents the actual velocity, x 1 represents the actual trajectory, represents the actual acceleration, τ is the control variable, F qext is the external disturbance. Let the estimation of the external disturbance be Then the estimation error of the external disturbance is:

[0179]

[0180] Since there is no prior knowledge to support the true external disturbance, without loss of generality, assume its derivative is 0, that is:

[0181]

[0182] Let L(x) ∈ R i×i , be the observation gain matrix, and the control law can be designed as:

[0183]

[0184] Substituting Equation (19) into the above control law, we can get:

[0185]

[0186] Among them, f 1 (x) = C q x 2 + G q + F qf , but in practice it is difficult to obtain. It is necessary to eliminate the terms included in the above formula to simplify the observer. Subtract

[0187]

[0188] from both sides of Equation (23) at the same time. Integrating both sides of the above formula gives:

[0189]

[0190] Among them,

[0191] Let an intermediate variable Taking the derivative of it gives:

[0192]

[0193] The estimation of external disturbance can be obtained from the above observer for z:

[0194]

[0195] In order to obtain better estimation performance and make its application more convenient, according to the UNDO design method of Feilong Zhang et al., it is necessary to introduce the term on the basis of (26). At this time, ρ(x) is L(x)M q x 2 , then the final observer can be designed as:

[0196]

[0197] In application, only need to adjust L(x) multiple times to obtain better control effect.

[0198] This embodiment significantly improves the control performance and safety of the waist rehabilitation exoskeleton system through an innovative composite control architecture. Specifically, it has the following outstanding advantages:

[0199] In terms of intelligent adaptability, in this embodiment, a radial basis neural network is adopted to learn and compensate the unmodeled part in the system dynamics in real time online, including complex factors such as time-varying parameters and nonlinear friction, effectively solving the problem of reduced accuracy caused by model mismatch in traditional control methods; in terms of anti-interference ability, in this embodiment, a general nonlinear disturbance observer is designed to dynamically estimate and compensate external disturbances such as the muscle strength fluctuation of the patient and sudden spasm movements, achieving fast response and precise suppression of composite disturbances; in terms of precision control accuracy, in this embodiment, through the collaborative work of a radial basis neural network and a general nonlinear disturbance observer, combined with basic PD control, a closed-loop control structure of "feedforward-feedback-compensation" is formed, improving the system trajectory tracking accuracy to the sub-millimeter level; in terms of safety guarantee, the adaptive control strategy of this embodiment can intelligently adjust the assistance strength according to the patient's rehabilitation stage, providing sufficient protection in the initial stage of training and promoting active recovery in the later stage, ensuring the safety and effectiveness of the entire rehabilitation process.

[0200] Combined with Figure 3 , the present invention compares the error between a single radial basis neural network controller and the combined method of the present invention by applying different external disturbances to the exoskeleton. The external disturbances are set in two groups. One group is a fixed external disturbance F ext0 = [10N 10N 10N 1N·m 1N·m 1N·m] T , and the other group is a high-frequency external disturbance F ext1 = [10sin(3t)N 20sin(3t)N 30sin(3t)N sin(3t)N·m 2sin(3t)N·m 3sin(3t)N·m] T . The results show that under different external disturbances, the tracking error fluctuations of each branch chain of the high-precision anti-interference tracking control of the present invention are all smaller than those under the single radial basis neural network tracking control. It can be seen that the general nonlinear disturbance observer helps to further improve the control accuracy of the exoskeleton on the basis of the radial basis neural network trajectory tracking control. The data shows that for the fixed external disturbance F ext0 , the control accuracy of each branch chain of the tracking control of the present invention has an average increase of 31.80% compared with the single radial basis neural network method. For the high-frequency external disturbance F ext1 , the control accuracy of each branch chain of the tracking control of the present invention has an average increase of 42.84% compared with the single radial basis neural network control method.

[0201] Embodiment 2:

[0202] This embodiment provides a parallel waist exoskeleton control system for implementing the method described in Embodiment 1, including:

[0203] The dynamics modeling module is configured to perform dynamics modeling on the parallel waist exoskeleton, and separate the uncertain part of the modeling and external disturbances;

[0204] The radial basis neural network is configured to locally approximate the uncertain part of the modeling by using the radial basis neural network based on the established dynamics model of the parallel waist exoskeleton;

[0205] The feedback compensation module is configured to estimate the external disturbance force of the exoskeleton by using the radial general nonlinear disturbance observer based on the established dynamics model of the parallel waist exoskeleton, and feedback for compensation.

[0206] Some steps in the embodiments of the present invention can be implemented by software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk, etc.

[0207] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A waist exoskeleton anti-disturbance tracking control method, characterized in that: The method comprises: Step 1: Dynamically model the parallel waist exoskeleton and separate the uncertain parts of its modeling and external disturbances; Step 2: Based on the constructed parallel waist exoskeleton dynamics model, a radial basis neural network is used to achieve local approximation of the uncertain part of the modeling; Step 3: Based on the established parallel waist exoskeleton dynamic model, a radial general nonlinear disturbance observer is used to estimate the external disturbance force of the exoskeleton and provide feedback for compensation.

2. The anti-disturbance tracking control method for waist exoskeleton according to claim 1, characterized in that: The parallel waist exoskeleton dynamics model is: Where M q is the inertia matrix of the exoskeleton in the joint space, C q is the Coriolis force and centrifugal force matrix of the exoskeleton in the joint space, G q is the gravity term of the exoskeleton in the joint space, F qf is the friction force of the exoskeleton in the joint space, τ is the joint force, F qext For external interference, are the actual acceleration and actual velocity of the exoskeleton joints, respectively.

3. The anti-disturbance tracking control method for waist exoskeleton according to claim 2, characterized in that: The process of using the radial basis neural network to design the control law in step 2 includes: Assume that the parameters of the exoskeleton nominal model are M q0 , C q0 , G q0 、F qf0 , for the nominal model, the trajectory tracking control law based on the PD controller is obtained: Among them, K d represents the proportional gain matrix, K p represents the differential gain matrix; Substituting the trajectory tracking control law formula into the parallel waist exoskeleton dynamics model, we get: in, is the acceleration error, represents the speed error, e q =qq d represents the trajectory error, q d are the expected acceleration, angular velocity and trajectory of the exoskeleton joints, respectively; q represents the actual trajectory of the exoskeleton joints; set up: Then u is the uncertain part of the dynamic modeling, where ΔM q =M q0 -M q , ΔC q =C q0 -C q , ΔG q =G q0 -G q , ΔF qf =F qf0 -F qf ; Use radial basis neural network to approximate u. The radial basis neural network algorithm is: Among them, x is the network input, is the neural network output, is the estimate of the optimal neural network weights, is the output of the Gaussian basis function: where c i is the center of the Gaussian function, σ i is the variance of the Gaussian function; Assume that the neural network outputs is continuous, and when the neural network weight is the optimal neural network weight W * The ideal output is: It satisfies that for any sufficiently small positive real number ε, ‖ξ‖≤ε, ξ is the ideal approximation error, Its bound is ξ0=sup‖ξ‖; According to the above information, the control law based on radial basis neural network is expressed as: τ=Τ1+Τ2 in, 4. The anti-disturbance tracking control method for waist exoskeleton according to claim 3, characterized in that: The process of estimating the external disturbance force in step 3 includes: According to the parallel waist exoskeleton dynamics model, x1=q, The system state equation is expressed as: Assume that the estimate of external disturbance is Then the estimation error of external disturbance is: Since there is no prior knowledge to support the real external interference, without loss of generality, we set its derivative to 0, that is: Let L(x)∈R i×i is the observed gain matrix, and the designed control law is: Substituting the system state equation into the above control law, we get: Where f1(x) = C q x2+G q +F qf , subtract both sides of the above equation Integrating both sides of the above equation gives: in, Set an intermediate variable Taking the derivative of it, we get: The estimate of the external disturbance is obtained from the above observer for z: Introduction term, then ρ(x) is L(x)M q x2, then the final observer design is: The ideal control effect can be obtained by adjusting L(x).

5. A waist exoskeleton anti-disturbance tracking control system, characterized in that: The system comprises: A dynamic modeling module, configured to perform dynamic modeling on the parallel waist exoskeleton and separate the uncertain parts of its modeling and external disturbances; A radial basis function neural network is configured to realize local approximation of the uncertain part of modeling based on the constructed parallel waist exoskeleton dynamics model; The feedback compensation module is configured to estimate the external disturbance force of the exoskeleton based on the established parallel waist exoskeleton dynamic model using a radial general nonlinear disturbance observer and provide feedback for compensation.

6. The waist exoskeleton anti-disturbance tracking control system according to claim 5, characterized in that: The parallel waist exoskeleton dynamics model is: Where M q is the inertia matrix of the exoskeleton in the joint space, C q is the Coriolis force and centrifugal force matrix of the exoskeleton in the joint space, G q is the gravity term of the exoskeleton in the joint space, F qf is the friction force of the exoskeleton in the joint space, τ is the joint force, F qext For external interference, are the actual acceleration and actual velocity of the exoskeleton joints, respectively.

7. The waist exoskeleton anti-disturbance tracking control system according to claim 6, characterized in that: The process of designing a control law using a radial basis function neural network includes: Assume that the parameters of the exoskeleton nominal model are M q0 , C q0 , G q0 、F qf0 , for the nominal model, the trajectory tracking control law based on the PD controller is obtained: Among them, K d represents the proportional gain matrix, K p represents the differential gain matrix; Substituting the trajectory tracking control law formula into the parallel waist exoskeleton dynamics model, we get: in, is the acceleration error, represents the speed error, e q =qq d represents the trajectory error, q d are the expected acceleration, angular velocity and trajectory of the exoskeleton joints, respectively; q represents the actual trajectory of the exoskeleton joints; set up: Then u is the uncertain part of the dynamic modeling, where ΔM q =M q0 -M q , ΔC q =C q0 -C q , ΔG q =G q0 -G q , ΔF qf =F qf0 -F qf ; Use radial basis neural network to approximate u. The radial basis neural network algorithm is: Among them, x is the network input, is the neural network output, is the estimate of the optimal neural network weights, is the output of the Gaussian basis function: where c i is the center of the Gaussian function, σ i is the variance of the Gaussian function; Assume that the neural network outputs is continuous, and when the neural network weight is the optimal neural network weight W * The ideal output is: It satisfies that for any sufficiently small positive real number ε, ‖ξ‖≤ε, ξ is the ideal approximation error, The bound of ξ is ξ0 = sup‖ξ‖; According to the above information, the control law based on radial basis neural network is expressed as: τ=Τ1+Τ2 in, 8. The waist exoskeleton anti-disturbance tracking control system according to claim 7, characterized in that: The process of estimating the external disturbance force by the feedback compensation module includes: According to the parallel waist exoskeleton dynamics model, x1=q, The system state equation is expressed as: Assume that the estimate of external disturbance is Then the estimation error of external disturbance is: Since there is no prior knowledge to support the real external interference, without loss of generality, we set its derivative to 0, that is: Let L(x)∈R i×i is the observed gain matrix, and the designed control law is: Substituting the system state equation into the above control law, we get: Where f1(x) = C q x2+G q +F qf , subtract both sides of the above equation Integrating both sides of the above equation gives: in, Set an intermediate variable Taking the derivative of it, we get: The estimate of the external disturbance is obtained from the above observer for z: Introduction term, then ρ(x) is L(x)M q x2, then the final observer design is: The ideal control effect can be obtained by adjusting L(x).

9. A waist exoskeleton anti-disturbance tracking control device, characterized in that: include: Sensor module, used to collect joint position, velocity, acceleration and external force signals in real time; A processor module, executing the anti-disturbance tracking control method for the waist exoskeleton as described in any one of claims 1 to 4, and calculating the joint control torque; A driving module is used to drive the exoskeleton joint to move according to the joint control torque.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the anti-disturbance tracking control method for the waist exoskeleton as described in any one of claims 1 to 4 is implemented.

Images