A design method of neural network integral sliding mode controller for flexible joint manipulator based on disturbance observer
By designing a neural network integral sliding mode controller based on perturbation observer in a flexible joint robot arm, combining the RBF neural network and integral sliding mode surface, the problems of complex design and output constraints of the flexible joint robot arm control scheme are solved, and high Accuracy trajectory tracking and robust control.
Patent Information
- Application Number
- CN202210548816.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-20
- Publication Date
- 2024-11-26
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The design of the control scheme of flexible joint robotic arms is complex. The existing control methods rely on multiple derivatization to increase system instability and calculation burden, making it difficult to effectively deal with time-varying external interference and output constraint problems.
Design a flexible joint robotic arm neural network integral sliding mode controller based on perturbation observer, combining RBF neural network and integral sliding mode surface to reduce dependence on dynamic model parameters, reduce the number of virtual control laws and derivative times, through The new perturbation observer estimates the uncertainty of the system, enabling output constraints and high-precision tracking.
It realizes high-precision trajectory tracking of flexible joint robot arms under model-free conditions, reduces system tracking error and steady-state error, alleviates the calculation explosion problem in inversion control, and improves control accuracy and robustness.
Smart Images

Figure CN114952835B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of automatic control, and in particular relates to a design method of a neural network integral sliding mode controller of a flexible joint mechanical arm based on a disturbance observer for trajectory tracking control of the flexible joint mechanical arm. Background Art
[0002] In recent decades, due to its wide application in aerospace, deep-sea exploration, industrial manufacturing, medical and other fields, the research on flexible joint manipulators has attracted more and more attention. Compared with traditional rigid manipulators, flexible joint manipulators have the advantages of light weight, small size, low energy consumption, and high load-to-weight ratio. In addition, since the joints of the manipulator are flexible, this greatly improves the safety of the manipulator in actual use. When the flexible joint collides, the flexible joint can effectively alleviate the collision. Although the flexible joint manipulator has great advantages in practical applications, the design of its control scheme also has great difficulties. Many researchers are working hard to design a more ideal control scheme to manipulate the flexible joint manipulator so that it can have better performance in various operations.
[0003] It is worth noting that the control scheme design of flexible joint manipulators is not mature compared with rigid joint manipulators. It is well known that flexible joint manipulators are complex dynamic systems, in which there are many parameters to be designed in the controller. In addition, some parameters are not easy to obtain under certain conditions. Therefore, people are considering designing a method that has lower requirements for manipulator parameters in feedback control. In recent years, the application of neural networks in learning ability mapping and parallel processing has attracted the attention of researchers. It can reduce the dependence on the parameters of the manipulator dynamic system in controller design, thereby realizing model-free control. Although the application of radial basis neural networks in flexible manipulator control has made great progress, in most previous radial basis neural network control studies, the weight estimation update law of the neural network is heavily dependent on the tracking error and instantaneous estimation data, and the convergence of the system estimation error in the neural network control cannot be achieved. In addition, radial basis neural networks cannot cope well with time-varying external disturbances.
[0004] In most trajectory tracking control literature, disturbance observer is a commonly used technique for dealing with time-varying external disturbance problems, which can achieve finite time convergence. In addition, this technique can also compensate for the approximation error of the radial basis neural network and make up for the shortcoming that the radial basis neural network estimation error cannot converge asymptotically. Therefore, a "neural network + disturbance observer" joint control law can be designed.
[0005] Generally, in the design of control methods for flexible joint manipulators, the most common control technologies are fourth-order backstepping control and fourth-order dynamic surface control. Both methods require the designer to perform multiple derivatives of state variables to obtain virtual control laws. Multiple derivatives will increase the instability of the system and the possibility of singularities when the computer is solving. Therefore, we can consider finding a way to reduce the number of virtual control laws and the number of derivatives of virtual control laws in controller design, thereby reducing the possibility of singularities and alleviating the "computational explosion" problem in backstepping control.
[0006] In addition, sliding mode control is a highly robust control method with perfect transient performance. For the controller of the flexible manipulator, choosing a suitable sliding mode surface will significantly improve its performance. This will also be one of the topics explored in this article.
[0007] It is worth noting that there is a lot of literature on dealing with input saturation, but not enough attention has been paid to output constraints. In practical applications, such as healthcare tasks and wearable exoskeleton manipulators, output constraints are very necessary to avoid harm to operators during use. This will also be one of the contents of this paper. Summary of the invention
[0008] In view of the above problems, the present invention proposes a design method of a neural network integral sliding mode controller for a flexible joint manipulator based on a disturbance observer, which is suitable for trajectory tracking control of a flexible joint manipulator affected by model uncertainty and external disturbances. A "neural network + disturbance observer" joint control method is proposed. This method reduces the dependence on the parameters of the dynamic model of the flexible joint manipulator in controller design, overcomes the shortcoming that the neural network estimation error cannot converge asymptotically, alleviates the "computational explosion" problem in the inverse control design, and applies the integral sliding surface to improve the estimation accuracy and reduce the steady-state error.
[0009] The present invention proposes a design method for a neural network integral sliding mode controller for a flexible joint manipulator based on a disturbance observer. The specific design scheme is as follows:
[0010] Step 1, establish a dynamic model of an n-DOF flexible joint robotic arm;
[0011] Step 2, divide the model in step 1 into two subsystems, and design an integral sliding surface for each subsystem;
[0012] Step 3, using RBF neural network to approximate the unknown dynamic parameters inside the model;
[0013] Step 4, using the new disturbance observer to estimate the lumped uncertainty consisting of the estimation error of the RBF neural network and the external disturbance of the model;
[0014] Step 5, using the obstacle Lyapunov function to design a disturbance observer-based flexible joint manipulator neural network integral sliding mode controller with position output constraint performance, to achieve model-free control of the manipulator.
[0015] Furthermore, the specific steps of establishing the dynamic model of the n-DOF flexible joint mechanical arm in step 1 are as follows:
[0016]
[0017]
[0018] In the formula, Respectively represent the angular position, angular velocity and angular acceleration of the robot arm link side and the motor shaft side. M(q) is the symmetric positive definite inertia matrix, is the centrifugal force and Coriolis force matrix, G(q) is the gravity vector, is the friction torque, τ is the control input, τ d is the unknown time-varying external disturbance, J m represents the positive definite diagonal matrix of motor moment of inertia, and K represents the positive definite diagonal matrix representing the spring stiffness.
[0019] Furthermore, the specific steps of step 2 are as follows: first, the dynamic model of the flexible joint manipulator is decomposed into two subsystems: the connecting rod side subsystem
[0020]
[0021] and motor side subsystem
[0022]
[0023] In these two subsystems, we can see that τ has a significant effect on q m The impact is direct, m The influence of τ on q is direct, while the influence of τ on q is indirect, so we define an intermediate virtual control quantity q md , it is q m The ideal trajectory is as long as τ can drive q m Tracking md , then at this time in q m Driven by q, it can track the ideal trajectory of the end of the upper robot arm q d .
[0024] From this we define the error signal as follows:
[0025] e=qq d (45)
[0026] em =q m -q md (46) Next, we design the integral sliding surface for these two error signals
[0027]
[0028]
[0029] Where, gains k1 and k2 are positive constants.
[0030] Furthermore, the specific steps of step 3 are as follows: first, the principle of RBFNN neural network is introduced. RBFNN can approximate any nonlinear function, and its mathematical expression is:
[0031] f(x)=W T h(x)+ε (49)
[0032] Where W is the ideal weight matrix, h(x) is the Gaussian basis function vector, and ε is the bounded estimation error of the neural network, whose value satisfies the inequality |ε|≤ε N (ε N is an upper bound of ε).
[0033] Then, we used the above neural network principle to analyze the dynamics model of the flexible joint robot arm. G(q) matrix is estimated.
[0034]
[0035]
[0036] In the formula, G ij (q) respectively represent The i-th row and j-th column element in the G(q) matrix. t represents the total number of nodes in the neural network. Therefore, the matrix G(q) can be expressed as
[0037]
[0038]
[0039] In the formula It is a new operator symbol defined for the convenience of expression. The specific expression of its operation rules is as follows:
[0040]
[0041]
[0042] Matrix W CBy sub-matrix W Cij (i=1,...,n;j=1,...,n); matrix W G By sub-matrix W Gij (i=1,...,n;j=1); matrix h C (q), h G (q) are respectively composed of sub-matrices and h Gij (q)(i=1,...,n;j=1); Estimation error matrix ε G (q) is composed of submatrices ε Cij (i=1,...,n; j=1,...,n) and ε Gij (i=1,...,n; j=1); wherein i represents the row where the submatrix is located, and j represents the column where the submatrix is located.
[0043] Furthermore, the specific steps of step 4 are to introduce the new disturbance observer introduced. Consider the following nonlinear system
[0044]
[0045] where f1(x), is a nonlinear function, u is the system input, and Δ(t) is the lumped uncertainty of the system.
[0046] For the above system, the new disturbance observer can be designed as follows:
[0047]
[0048]
[0049] Among them, K o is a positive definite diagonal matrix representing the gain, It represents the estimation of the system lumped uncertainty Δ(t). The expression of the lumped uncertainty of the flexible joint robot arm will be introduced in the next step.
[0050] Furthermore, the specific steps of step 5 are as follows: the first step is to design a controller based on the dynamic model of the flexible joint manipulator. (For ease of description, in the following, we use symbols M, C, G, and F to represent the matrix M(q), G(q), ).
[0051] First choose the following barrier Lyapunov function
[0052]
[0053] Taking the derivative of V1, we have
[0054]
[0055] In the formula, e i Represents the i-th element of vector e. k ai is the design parameter in the obstacle Lyapunov function. By designing this parameter, the error value e i Limited to the interval [-k ai ;k ai ]middle.
[0056] Next, we design the following Lyapunov function
[0057]
[0058] Taking the derivative of V2, we have
[0059]
[0060] in
[0061]
[0062] Using equations (3) and (5), the term (A) in equation (20) can be transformed into
[0063]
[0064] Then, we define (21) = -K1s, and we can get the virtual control law q md The following expression
[0065]
[0066] Where K1 is a positive definite diagonal matrix representing the control gain.
[0067] Next, using equations (4) and (6), term (B) in equation (20) can be transformed into
[0068]
[0069] Then, we define (23) = -K2s m , we can get the actual control law τ
[0070]
[0071] Where K2 is a positive definite diagonal matrix representing the control gain.
[0072] The second step is to design a neural network integral sliding mode controller for the flexible joint manipulator based on disturbance observer. First, the matrix M(q) can be written as
[0073] M(q)=M no +ΔM (66)
[0074] Where M no represents the known nominal value of the matrix M(q), and ΔM represents the unknown uncertainty of the matrix M(q). According to formulas (10), (11), and (25), the dynamic equations (1) and (2) of the flexible manipulator can be rewritten as
[0075]
[0076]
[0077] In the formula, represents the lumped uncertainty of this model. For this lumped uncertainty, we use the new disturbance observer mentioned above to estimate it. Transforming equation (1) into equation (14)
[0078]
[0079] M in the formula no (q) -1 τ D (t) corresponds to Δ(t) in equation (14), M no (q) -1 Corresponding to f(x1) in equation (14), K(q m -q) corresponds to u in equation (14), Corresponding to equation (14) Corresponding to equation (14)
[0080] Therefore, this new disturbance observer can be designed as
[0081]
[0082]
[0083]
[0084] In the formula, Denotes the lumped uncertainty τ D (t) is an estimated value.
[0085] Since in practical applications, we cannot get the exact weight matrix W C , W G , but only their valuations can be obtained So we define
[0086]
[0087]
[0088] Therefore, according to the transformed dynamic equations (26) and (27), equations (28) and (29) are used to correct the equations in (26): Replace with For τ in (26) D (t) by replacing it, we can obtain the virtual control law q in the neural network integral sliding mode controller of the flexible joint manipulator based on disturbance observer md The expression is
[0089] The adaptive update rate of the neural network weights is as follows
[0090]
[0091]
[0092] In the formula, the sub-matrix is a matrix The submatrix of , i represents the row of the submatrix, j represents the column of the submatrix; Γ Cij ,Γ Gi1 is a positive definite diagonal matrix; η C , η G is a very small positive real number; r i Represents vectors The jth (j=1,...,n) and ith (i=1,...,n) elements of r, vector r is defined as the following expression
[0093]
[0094]
[0095] Where Λ is a positive definite diagonal matrix to be designed.
[0096] Since in actual situations, the state variables involved in the control law is unmeasurable, therefore, we apply a first-order filter to observe and estimate these two state variables.
[0097] For estimation The design of the first-order filter is as follows
[0098]
[0099]
[0100] In the formula, ζ1 is a positive constant, is the output of the filter. We use As a state variable The estimated value of .
[0101] For estimation The design of the first-order filter is as follows:
[0102]
[0103]
[0104] In the formula, ζ2 is a positive constant, is the output of the filter. We use As a state variable The estimated value of .
[0105] Therefore, substitute equation (34) into equation (24) and use The state variables in the formula are By replacing, the actual control law of the neural network integral sliding mode controller of the flexible joint manipulator based on disturbance observer can be obtained as follows:
[0106]
[0107] The above technical solution achieves the following beneficial effects:
[0108] (1) A joint control law of "RBF neural network + disturbance observer" is proposed. The lumped uncertainty composed of the estimation error of the RBF neural network and the external disturbance is observed again by the disturbance observer, which can greatly reduce the tracking error of the system.
[0109] (2) The neural network is used to estimate each element in the uncertainty matrix of the flexible joint robot arm dynamics model, which improves the estimation accuracy of the neural network and makes it easier to combine the neural network technology with other advanced control technologies.
[0110] (3) The use of integral sliding surface improves the error convergence speed and reduces the steady-state error.
[0111] (4) A new inversion design method was used to reduce the order of the inversion design and effectively alleviate the "computational explosion" problem in inversion control. BRIEF DESCRIPTION OF THE DRAWINGS
[0112] Figure 1 It is a structural block diagram of the controller in the present invention;
[0113] Figure 2Schematic diagram of the physical model of a two-link flexible joint mechanical arm in an embodiment of the present invention;
[0114] Figure 3 is a schematic diagram of the joint position tracking of a robotic arm in an embodiment of the present invention;
[0115] Figure 4 Schematic diagram of the tracking error of the joint position of the robot arm in an embodiment of the present invention;
[0116] Figure 5 Schematic diagram of robot arm joint velocity tracking error in an embodiment of the present invention;
[0117] Figure 6 Schematic diagram of the joint torque of the robot arm in an embodiment of the present invention. DETAILED DESCRIPTION
[0118] The present invention is further explained below in conjunction with specific embodiments. In order to better illustrate the present invention, the proposed controller is verified by using MATLAB numerical simulation. Figure 1 is the structural block diagram of the controller. Figure 2 is a schematic diagram of the physical model of a two-link flexible joint robotic arm. The results are shown in Figures 3 to 6 The specific steps of the controller design method are as follows:
[0119] Step 1: Establish the dynamic model of the n-DOF flexible joint robot arm. The specific steps are as follows:
[0120]
[0121]
[0122] In the formula, Respectively represent the angular position, angular velocity and angular acceleration of the robot arm link side and the motor shaft side. M(q) is the symmetric positive definite inertia matrix, is the centrifugal force and Coriolis force matrix, G(q) is the gravity vector, is the friction torque, τ is the control input, τ d is the unknown time-varying external disturbance, J m represents the positive definite diagonal matrix of motor moment of inertia, and K represents the positive definite diagonal matrix representing the spring stiffness.
[0123] Step 2, first we define the error signal as follows:
[0124] e=qq d (3)
[0125] e m =q m -q md (4) Next, we design the integral sliding surface for these two error signals
[0126]
[0127]
[0128] Where, gains k1 and k2 are positive constants.
[0129] Step 3, first introduce the principle of RBFNN neural network. RBFNN can approximate any nonlinear function, and its mathematical expression is:
[0130] f(x)=W T h(x)+ε (85)
[0131] Where W is the ideal weight matrix, h(x) is the Gaussian basis function vector, and ε is the bounded estimation error of the neural network, whose value satisfies the inequality |ε|≤ε N (ε N is an upper bound of ε).
[0132] Then, we used the above neural network principle to analyze the dynamics model of the flexible joint robot arm. G(q) matrix is estimated.
[0133]
[0134]
[0135] In the formula, G ij (q) respectively represent The i-th row and j-th column element in the G(q) matrix. t represents the total number of nodes in the neural network. Therefore, the matrix G(q) can be expressed as
[0136]
[0137]
[0138] In the formula It is a new operator symbol defined for the convenience of expression. The specific expression of its operation rules is as follows:
[0139]
[0140]
[0141] Matrix W C By sub-matrix W Cij (i=1,...,n;j=1,...,n); matrix W G By sub-matrix W Gij(i=1,...,n;j=1); matrix h C (q), h G (q) are respectively composed of sub-matrices and h Gij (q)(i=1,...,n;j=1); Estimation error matrix ε G (q) is composed of submatrices ε Cij (i=1,...,n; j=1,...,n) and ε Gij (i=1,...,n; j=1); wherein i represents the row where the submatrix is located, and j represents the column where the submatrix is located.
[0142] Step 4, introduce the new disturbance observer. Consider the following nonlinear system
[0143]
[0144] where f1(x), is a nonlinear function, u is the system input, and Δ(t) is the lumped uncertainty of the system.
[0145] For the above system, the new disturbance observer can be designed as follows:
[0146]
[0147]
[0148] Among them, K o is a positive definite diagonal matrix representing the gain, It represents the estimation of the system lumped uncertainty Δ(t). The expression of the lumped uncertainty of the flexible joint robot arm will be introduced in the next step.
[0149] Step 5: The first step is to design a controller based on the dynamic model of the flexible joint manipulator. (For ease of description, in the following, we use the symbols M, C, G, and F to represent the matrix M(q). G(q), ).
[0150] First choose the following barrier Lyapunov function
[0151]
[0152] Taking the derivative of V1, we have
[0153]
[0154] In the formula, e iRepresents the i-th element of vector e. k ai is the design parameter in the obstacle Lyapunov function. By designing this parameter, the error value e i Limited to the interval [-k ai ;k ai ]middle.
[0155] Next, we design the following Lyapunov function
[0156]
[0157] Taking the derivative of V2, we have
[0158]
[0159] in
[0160]
[0161] Using equations (3) and (5), the term (A) in equation (20) can be transformed into
[0162]
[0163] Then, we define (21) = -K1s, and we can get the virtual control law q md The following expression
[0164]
[0165] Where K1 is a positive definite diagonal matrix representing the control gain.
[0166] Next, using equations (4) and (6), term (B) in equation (20) can be transformed into
[0167]
[0168] Then, we define (23) = -K2s m , we can get the actual control law τ
[0169]
[0170] Where K2 is a positive definite diagonal matrix representing the control gain.
[0171] The second step is to design a neural network integral sliding mode controller for the flexible joint manipulator based on disturbance observer. First, the matrix M(q) can be written as
[0172] M(q)=M no +ΔM (94)
[0173] Where M norepresents the known nominal value of the matrix M(q), and ΔM represents the unknown uncertainty of the matrix M(q). According to formulas (10), (11), and (25), the dynamic equations (1) and (2) of the flexible manipulator can be rewritten as
[0174]
[0175]
[0176] In the formula, represents the lumped uncertainty of this model. For this lumped uncertainty, we use the new disturbance observer mentioned above to estimate it. Transforming equation (1) into equation (14)
[0177]
[0178] M in the formula no (q) -1 τ D (t) corresponds to Δ(t) in equation (14), M no (q) -1 Corresponding to f(x1) in equation (14), K(q m -q) corresponds to u in equation (14), Corresponding to equation (14) Corresponding to equation (14)
[0179] Therefore, this new disturbance observer can be designed as
[0180]
[0181]
[0182]
[0183] In the formula, Denotes the lumped uncertainty τ D (t) is an estimated value.
[0184] Since in practical applications, we cannot get the exact weight matrix W C , W G , but only their valuations can be obtained So we define
[0185]
[0186]
[0187] Therefore, according to the transformed dynamic equations (26) and (27), equations (28) and (29) are used to correct the equations in (26): Replace with For τ in (26) D (t) by replacing it, we can obtain the virtual control law q in the neural network integral sliding mode controller of the flexible joint manipulator based on disturbance observer md The expression is
[0188] The adaptive update rate of the neural network weights is as follows
[0189]
[0190]
[0191] In the formula, the sub-matrix is a matrix The submatrix of , i represents the row of the submatrix, j represents the column of the submatrix; Γ Cij ,Γ Gi1 is a positive definite diagonal matrix; η C , η G is a very small positive real number; r i Represents vectors The jth (j=1,...,n) and ith (i=1,...,n) elements of r, vector r is defined as the following expression
[0192]
[0193]
[0194] Where Λ is a positive definite diagonal matrix to be designed.
[0195] Since in actual situations, the state variables involved in the control law is unmeasurable, therefore, we apply a first-order filter to observe and estimate these two state variables.
[0196] For estimation The design of the first-order filter is as follows
[0197]
[0198]
[0199] In the formula, ζ1 is a positive constant, is the output of the filter. We use As a state variable The estimated value of .
[0200] For estimation The design of the first-order filter is as follows:
[0201]
[0202]
[0203] In the formula, ζ2 is a positive constant, is the output of the filter. We use As a state variable The estimated value of .
[0204] Therefore, substitute equation (34) into equation (24) and use The state variables in the formula are By replacing, the actual control law of the neural network integral sliding mode controller of the flexible joint manipulator based on disturbance observer can be obtained as follows:
[0205]
[0206] In the MATLAB 2019a environment, the present invention uses simulink and two-joint flexible joint robot model parameters to simulate and verify a flexible joint robot neural network integral sliding mode controller based on a disturbance observer designed by the present invention (indicated by the letter group "DNISM" in the coordinate diagram below) and compare it with some other control algorithms, such as "flexible joint robot control using linear sliding surface" (indicated by the letter group "LSS" in the coordinate diagram below), "flexible joint robot control using only neural network" (indicated by the letter group "SNNC" in the coordinate diagram below), and "fourth-order dynamic surface flexible joint robot control" (indicated by the letter group "DSC" in the coordinate diagram below):
[0207] (1) The simulation parameters are as follows
[0208]
[0209]
[0210]
[0211]
[0212] Where q1 and q2 represent the angular positions of the two joints in the dynamic model of the flexible joint manipulator. Represent the angular velocities of the two joints respectively. The nominal values of the parameters of the dynamic model system are The gravitational acceleration g is 9.8 kg / N. The initial simulation values of each state variable are selected as q(0) = [0.02, 0.02] T , q m (0) = [0.001, 0.001] T ,
[0213] Ideal trajectory q 1d ,q 2d Select The external time-varying disturbance is set to Considering the existence of system uncertainty in reality, we set the actual values of the dynamic system parameters of the two-joint flexible joint robot arm to fluctuate 20% upward from their nominal values, that is, The control gain of the controller is set to The parameters of the integral sliding surface are set to k1 = 35, k2 = 20; the parameters in the barrier Lyapunov function are set to k a1 =0.1, k a2 = 0.1. As for the neural network, we use an 11-node RBF neural network to And predict each element in the G(q) matrix. The center of the neural network Gaussian function is evenly distributed in the interval [-0.3, 0.3], and the width of the Gaussian function is set to b Cij =10(i=1,...,n; j=1,...,n), b Gij =10(i=1,...,n; j=1). The parameter of the neural network weight adaptive law is set to Γ Cij =100,Γ Gi1 =100,η Cij =0.0001, η Gi1 =0.0001, The parameters of the new disturbance observer are set as The parameters of the first-order filter are set to ζ1=0.0001, ζ2=0.0001.
[0214] Result description:
[0215] Figure 3 It is a schematic diagram of the simulation of the position tracking of the two joints of the robot arm. It can be seen from the figure that the two flexible robot arm joints in the present invention can track the desired trajectory in a very short time, which reflects the advantage of the fast tracking of the present invention.
[0216] Figure 4It is a schematic diagram of the simulation of the tracking error of the two joint angular positions of the robot arm. It can be seen from the figure that the steady-state error of the angular position of the two robot arms in the present invention is smaller than the tracking error of other control methods, and the error curve is smoother, which reflects the advantage of high tracking accuracy of the present invention.
[0217] Figure 5 It is a schematic diagram of the simulation of the tracking error of the angular velocity of the two joints of the robot arm. It can be seen from the figure that the angular velocity steady-state error of the two robot arms in the present invention is smaller than the tracking error of other control methods, and the error curve is smoother, which reflects the advantage of high tracking accuracy of the present invention.
[0218] Figure 6 It is a simulation diagram of the input torque of the two joints of the robot. By comparing with the input torque of the “fourth-order dynamic surface flexible joint robot control method”, it can be seen that the control input curves of the two joints in the present invention are smooth and continuous, and the “computational explosion” problem and singularity problem existing in the inversion control are well avoided. However, due to the multiple derivations of the virtual control law in the “fourth-order dynamic surface flexible joint robot control method”, the torque curve produces a large jump, and the “computational explosion” and singularity problems still exist.
[0219] In summary, the control scheme designed in the present invention can achieve high-precision tracking of the desired trajectory in a short time without the need for an accurate model of the flexible joint robot arm, and also exhibits strong robustness against interference, and can achieve asymptotic stabilization of the tracking error.
[0220] The above specific implementation cases are only for the convenience of people in this research field to understand the present invention, but the present invention is not only applicable to the situations in the cases. For ordinary technicians in this technical field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.
Claims
1. A design method for a neural network integral sliding mode controller for a flexible joint manipulator based on a disturbance observer, characterized in that: The steps include: Step 1, establish a dynamic model of an n-DOF flexible joint robotic arm; Step 2, divide the model in step 1 into two subsystems, and design an integral sliding surface for each subsystem; Step 3, using RBF neural network to approximate the unknown dynamic parameters inside the model; Step 4, using the disturbance observer to estimate the lumped uncertainty consisting of the estimation error of the RBF neural network and the external disturbance of the model; Step 5, using the obstacle Lyapunov function to design a neural network integral sliding mode controller for a flexible joint manipulator based on a disturbance observer with position output constraint performance, to achieve model-free control of the manipulator; The specific steps of establishing the dynamic model of the n-DOF flexible joint manipulator in step 1 are as follows: In the formula, q, They represent the angular position, angular velocity and angular acceleration of the robot arm link side, respectively, and q m , They represent the angular position, angular velocity and angular acceleration of the motor shaft side respectively; M(q) is the symmetric positive definite inertia matrix, is the centrifugal force and Coriolis force matrix, G(q) is the gravity vector, is the friction torque, τ is the control input, τ d (t) is the unknown time-varying external disturbance, J m represents the positive definite diagonal matrix of motor moment of inertia, and K represents the positive definite diagonal matrix of spring stiffness; The specific steps of step 2 are as follows: First, the dynamic model of the flexible joint manipulator is decomposed into two subsystems: the connecting rod side subsystem and motor side subsystem In these two subsystems, τ has an influence on q m The impact is direct, m The influence of τ on q is direct, while the influence of τ on q is indirect, so an intermediate virtual control quantity q is defined md , it is q m The ideal trajectory is as long as τ can drive q m Tracking md , then at this time in q m Driven by q, it can track the ideal trajectory of the end of the upper robot arm q d ; The error signal is thus defined as follows: e=q-q d (3) e m =q m -q md (4) Next, we design the integral sliding surface for these two error signals: In the formula, gains k1 and k2 are positive constants; The specific steps of step 3 are: firstly, introduce the principle of RBF neural network; RBF approximates any nonlinear function, and its mathematical expression is: f(x)=W T h(x)+ε (7) Where W is the ideal weight matrix, h(x) is the Gaussian basis function vector, and ε is the bounded estimation error of the neural network, whose value satisfies the inequality |ε|≤ε N , ε N is an upper bound of ε; Then, the above neural network principle is used to analyze the dynamics model of the flexible joint robot arm. G(q) matrix is estimated; In the formula, G ij (q) respectively represent The i-th row and j-th column element in the G(q) matrix; t represents the total number of nodes in the neural network; therefore, the matrix G(q) is expressed as In the formula It is a new operator symbol defined for the convenience of expression. Its operation rules are specifically expressed as follows: Matrix W C By sub-matrix W Cij (i=1,...,n;j=1,...,n); matrix W G By sub-matrix W Gij (i=1,...,n;j=1); matrix h C (q), h G (q) are respectively composed of sub-matrices (i=1,...,n;j=1,...,n) and h Gij (q)(i=1,...,n;j=1); Estimation error matrix ε G (q) is composed of submatrices ε Cij (i=1,...,n; j=1,...,n) and ε Gij (i=1,...,n; j=1); wherein i represents the row where the submatrix is located, and j represents the column where the submatrix is located.
2. The method for designing a neural network integral sliding mode controller for a flexible joint manipulator based on a disturbance observer according to claim 1 is characterized in that: The specific steps of step 4 are as follows: first, the disturbance observer introduced is introduced; consider the following nonlinear system where f1(x), is a nonlinear function, u is the system input, and Δ(t) is the lumped uncertainty of the system; For the above system, the disturbance observer is designed as follows: Among them, K o is a positive definite diagonal matrix representing the gain, Represents an estimate of the lumped uncertainty Δ(t).
Citation Information
Patent Citations
Adaptive terminal sliding mode control method
CN109564406A
Mechanical arm self-adaptive trajectory tracking control method
CN112180729A
Flexible joint control method based on sliding mode observer
CN112643670A
Trajectory tracking control method based on adaptive neural network high-order dynamic sliding mode
CN114147713A
Adaptive terminal sliding mode control method
WO2018023201A1
