CN114326405B - Neural network backstepping control method based on error training - Google Patents
Neural network backstepping control method based on error training Download PDFInfo
- Publication number
- CN114326405B CN114326405B CN202111669359.4A CN202111669359A CN114326405B CN 114326405 B CN114326405 B CN 114326405B CN 202111669359 A CN202111669359 A CN 202111669359A CN 114326405 B CN114326405 B CN 114326405B
- Authority
- CN
- China
- Prior art keywords
- neural network
- error
- function
- training
- nonlinear
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Landscapes
- Feedback Control In General (AREA)
Abstract
A neural network backstepping control method based on error training solves the problems that the existing neural network backstepping control method is low in convergence speed and the neural network can not accurately estimate unmodeled dynamics, so that system tracking errors are causedThe method has a large problem, and belongs to the field of neural network backstepping control methods of nonlinear systems. The invention comprises the following steps: s1, establishing a nonlinear n-order system state space model containing unmodeled dynamics, wherein a state variable is [ x ] 1 ,...,x n ] T (ii) a S2, determining an error variable z 1 And z i ,z 1 =x 1 ‑y d ,z i =x i ‑α i‑1 Wherein α is i‑1 Representing a virtual control function; s3, establishing an error z i The input of the differential estimator is z i Output is as Is composed of(ii) is estimated; s4, obtained by S3Calculating the estimation error of the current radial basis function neural network, and performing gradient descent training on the weight of the neural network based on the estimation error to obtainS5 according to alpha n 、A control input signal for the non-linear system is calculated.
Description
Technical Field
The invention belongs to the field of a neural network backstepping control method of a nonlinear system.
Background
The basic principle of the method is that the neural network can approximate any unknown function with a certain error to estimate the unmodeled dynamics in the system, and the estimated value of the neural network is fed back to the nonlinear system through negative feedback in the design process of the backstepping method, so that the disturbance of the unmodeled dynamics to the system is reduced. Most of the current strategies employ net back-stepping control with final consistent bounding and sigma-adjustment-based neural network weight update strategies. The final consistency and the bounding are defined as stability in infinite time, and the system designed by the method has a low convergence speed and is not suitable for some control systems with strict convergence time of the system state. In addition, the method based on σ adjustment has the effect of reducing the lyapunov function, rather than directly reducing the approximation error of the neural network, and thus has a relatively limited ability to compensate for unmodeled dynamic quantities.
Disclosure of Invention
The invention provides a neural network backstepping control method based on error training, aiming at the problems that the existing neural network backstepping control method is low in convergence speed, and the neural network cannot accurately estimate unmodeled dynamics, so that the system tracking error is larger.
The invention discloses a neural network backstepping control method based on error training, which comprises the following steps:
s1, establishing a nonlinear n-order system state space model containing unmodeled dynamics, wherein a given target signal is y d The state variable is [ x ] 1 ,...,x n ] T ;
S2, determining an error variable z 1 And z i ,z 1 =x 1 -y d ,z i =x i -α i-1 Wherein α is i-1 Representing a virtual control function;
wherein i =1, \8230;, n,denotes the intermediate variable, g i Is a non-linear n-th order system intrinsic parameter, g 0 =z 0 =0,μ i1 、μ i2 、ε i Are all constant and are positive numbers, W i T S i (X i ) Representing a radial basis neural network, W i =[w i1 ,...,w iN ] T Is a weight matrix of the neural network, S i (X i ) Representing the basis function vector, S, of a neural network i (X i ) Each of the basis functions in (1) is a gaussian function having the same center and fixed width, X i =[x 1 ,...,x i ]The input vector is an input vector of the neural network, and N is the number of nodes of the neural network;
s3, establishing an error z i The input of the differential estimator is z i Output isIs->(ii) an estimate of (d);
s4, obtained by S3Calculating the estimation error of the current radial basis function neural network, performing gradient descent training on the weight of the neural network based on the estimation error, and obtaining->
Preferably, in S3, the differential estimator is:
where Ψ (y) = (1-e) -by )/(1+e -by ),λ, η, b are constants and positive numbers, ω i Ψ (y) is used to estimate the sign function sgn (y), - @, for the state quantity of the differential estimator>Is z i In the first derivative of (D), in conjunction with a signal from a signal pickup device>Is omega i ξ denotes the estimation error of the differential estimator, b @>Is the upper limit of ξ.
Preferably, in S5, the control input signal u of the nonlinear system:
preferably, in S1, the nonlinear n-order system state space model including unmodeled dynamics is:
wherein f is i (. Cndot.) is an unknown nonlinear continuous function representing unmodeled dynamics of the system, u represents the control input signal of the nonlinear system, and y represents the nonlinear lineAn output of the sexual system;
preferably, in S2, the method for acquiring the virtual control function includes:
using the obtained error z i Designing a Lyapunov function V:
the derivative values are:
and selecting a virtual control function according to the derivative of the Lyapunov function.
Preferably, in S1, S is i (X i ) The kth element s of k (X i ) Represents:
where exp represents an exponential function, b k Representing the k-th basis function width, C, of the neural network k Is the k-th basis function center vector, C, of the neural network k =[c 1,k ,...,c n,k ] T Wherein c is i,k Is a constant.
Preferably, in S4, the compound obtained in S3 is usedCalculating the estimation error of the current radial basis function neural network, and performing gradient descent training on the weights of the neural network based on the estimation error, wherein the weight W is i The update rate is specifically expressed as:
W n the update rate is specifically expressed as:
wherein, the first and the second end of the pipe are connected with each other,represents a weight W i First derivative of (a), r is the learning rate, r>0;/>And/or>Are respectively in>And/or>Is estimated.
Preferably, the nonlinear n-order system is a motor rotation angle system.
The method has the advantages that by utilizing a fixed time bounded stability theory and an updating strategy based on error training neural network weight, compared with the traditional adaptive neural network control method, the method has the advantages that the system has shorter convergence time and smaller tracking error.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a graph of the motor turn angle response with time on the abscissa, under the method of the present invention;
FIG. 3 is a diagram of a dynamic curve of the neural network estimation and system unmodeled under the method of the present invention, with the ordinate being the estimation value of the neural network and the abscissa being time;
FIG. 4 is a graph of neural network weights plotted on the ordinate versus time for neural network weights under the method of the present invention;
FIG. 5 is a graph of the neural network estimation error in the method of the present invention, with the ordinate being the neural network estimation error and the abscissa being time;
fig. 6 is a graph of a motor input torque signal plotted on the ordinate versus time on the abscissa, in accordance with the method of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
It should be noted that the embodiments and features of the embodiments may be combined with each other without conflict.
The invention is further described with reference to the following drawings and specific examples, which are not intended to be limiting.
The neural network backstepping control method based on error training of the embodiment comprises the following steps:
step one, establishing a nonlinear n-order system state space model containing unmodeled dynamics comprises the following steps:
wherein f is i () is an unknown nonlinear continuous function representing unmodeled dynamics of the system, u represents the control input signal of the nonlinear system, and y represents the output of the nonlinear system; given a target signal of y d The state variable is [ x ] 1 ,...,x n ] T ;
Step two, determining an error variable z 1 And z i ,z 1 =x 1 -y d ,z i =x i -α i-1 Wherein α is i-1 Representing a virtual control function;
the virtual control function of the present embodiment is designed according to a fixed-time bounded stability theory, and the virtual control function designed in the present embodiment is:
wherein i =1, \8230;, n,denotes the intermediate variable, g i G, a non-linear n-th order system intrinsic parameter 0 =z 0 =0,μ i1 、μ i2 、ε i Are all constant and positive, are>Representing a radial basis neural network, W i =[w i1 ,...,w iN ] T Is a weight matrix of the neural network, S i (X i ) Representing the basis function vector, S, of a neural network i (X i ) Each of the basis functions in (a) is a gaussian function having the same center and fixed width, X i =[x 1 ,...,x i ]The input vector of the neural network is defined, and N is the number of nodes of the neural network;
in this embodiment, the process of designing the virtual control function is as follows: designing a Lyapunov function V by using an error variable; first derivative of time of Lyapunov functionAnd selecting a virtual control function according to the first derivative of the Lyapunov function.
Step three, establishing error z i The input of the differential estimator is z i Output isIs->(ii) an estimate of (d);
step four, utilizing the product obtained in step threeCalculating the estimation error of the current radial basis function neural network, performing gradient descent training on the weight of the neural network based on the estimation error, and obtaining->
Step five, according to alpha n 、And determining a fixed time neural network backstepping controller to calculate a control input signal of the nonlinear system.
The fixed time neural network backstepping controller of the embodiment is as follows:
wherein u is a control input signal of the nonlinear system; mu.s n1 ,μ n2 ,ε n Is constant and is a positive number, W n =[w n1 ,...,w nN ] T Is a weight matrix of the neural network, S n (X n ) Representing the basis function vector, S, of a neural network n (X n ) Each of which is a gaussian function having the same center and fixed width,X n =[x 1 ,...,x n ]Is the input vector of the neural network, and N is the number of the nodes of the neural network.
In this embodiment, the vector of the basis function of the neural network is selected from Gaussian function S i (X i ) The kth element s of k (X i ) Represents:
wherein exp represents an exponential function, b k Representing the k-th basis function width, C, of the neural network k Is the k-th basis function center vector, C, of the neural network k =[c 1,k ,...,c n,k ] T Wherein c is i,k I = 1.. And n is a constant.
In the step of the present embodiment, the compound obtained in the third step is usedCalculating the estimation error of the current radial basis function neural network, and performing gradient descent training on the weight of the neural network based on the estimation error, wherein the weight W is i The update rate is specifically expressed as:
W n the update rate is specifically expressed as:
wherein the content of the first and second substances,represents a weight W i First derivative of (a), r is the learning rate, r>0;/>And/or>Are respectively in>And/or>Is estimated.
The differential estimator in step three of the present embodiment is a differential estimator with fixed time convergence, and specifically includes:
where Ψ (y) = (1-e) -by )/(1+e -by ),λ, η, b are constants and are positive numbers, ω i Ψ (y) is used to estimate the sign function sgn (y), - @, for the state quantity of the differential estimator>Is z i Is first derivative of->Is omega i ξ denotes the estimation error of the differential estimator, b @>Is the upper limit of ξ.
The fixed-time bounded stability of the present embodiment is demonstrated:
substituting (2) and (3) into coordinate transformation comprises:
For any continuous function f (X), there is an ideal network weight W * So thatLet f be the function to be evaluated of the neural network if the weight of the neural network is based on a gradient descent algorithm>Update then has->Therefore, when we update the network weights according to (5) and (6), we can get:
from the inequality of mean, have
For x i ∈R,i=1,...,n,ι∈[0,1]Of (| x) 1 |+...+|x n |) ι ≤(|x 1 |) ι +...+(|x n |) ι And (| x) 1 |+...+|x n |) 2 ≤[(|x 1 |) 2 +...+(|x n |) 2 ]N according to (14), (15) and (16) are
Wherein mu 1 =min{μ 11 ,...,μ n1 },μ 2 =min{μ 12 ,...,μ n2 }/n,According to the fixed time bounded stability theory, the system meets the condition
WhereinIs a constant, the tracking error of the system will be at a fixed time T s Converge into a small neighborhood around the origin, T s The values of (A) are:
after the syndrome is confirmed.
Example 1:
consider the following motor rotation angle system:
wherein x is 1 (unit rad) is the output angle, x, of the motor angle system 2 (unit rad/s) is the angular velocity of the motor. j (unit N.m) 2 ) The system input u (in N · m) is the input torque for the system moment of inertia.
Taking the initial value of the system as x 1 (0)=0,x 2 (0) =0, constant j =20. To demonstrate the simulation, assume that the system unmodeled dynamics areThis function cannot be used directly to design the system control input u. System target signal set to y d (t)=0.1。
The parameters in the virtual control function (4) and the system control input function (5) are taken as mu 11 =15,μ 12 =20,ε 1 =0.4,μ 21 =150,μ 22 =3,ε 2 And =0.01. (6) The medium neural network comprises 12 nodes, the elements of the central vector of the basis function are all 0, the widths are all 0.5, and the initial values of the weights are all 0. The parameters in the differential estimators (7), (8) are λ =10, η =10, b =1. The parameters in the error-training based update strategy (11) are taken as r =0.0002. The system sampling interval time is 0.001 seconds.
The control is carried out in the manner of fig. 1, and fig. 2 shows a motor rotation angle response curve under the action of the method of the invention; FIG. 3 is a graph of an estimate for a neural network under the method of the present invention; FIG. 4 is a graph illustrating neural network weight update under the method of the present invention; FIG. 5 is a graph of the neural network estimated error under the method of the present invention; FIG. 6 is a graph of motor input torque in accordance with the method of the present invention.
Conclusion one: from fig. 2, it can be obtained that the convergence rate of the system is faster and the tracking error of the system is smaller under the method of the present invention.
And a second conclusion: 3-5, the neural network can adjust the network weight according to the error better under the method of the present invention, thereby estimating the unmodeled dynamics in the system more accurately, and leading the system to have smaller tracking error.
Although the invention herein has been described with reference to particular embodiments, it is to be understood that these embodiments are merely illustrative of the principles and applications of the present invention. It is therefore to be understood that numerous modifications may be made to the illustrative embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that features described in different dependent claims and herein may be combined in ways different from those described in the original claims. It is also to be understood that features described in connection with individual embodiments may be used in other described embodiments.
Claims (5)
1. A neural network backstepping control method based on error training is characterized by comprising the following steps:
s1, establishing a nonlinear n-order system state space model containing unmodeled dynamics, wherein a given target signal is y d The state variable is [ x ] 1 ,...,x n ] T ;
The nonlinear n-order system state space model is a motor corner system model, and the motor corner system model is as follows:
x 1 is the output angle, x, of the motor angle system 2 The angular velocity of the motor is shown, j is the rotational inertia of the system, and u is the input torque; f (x) 1 ,x 2 ) For unmodeled dynamics of the system, y represents the output of the motor rotation angle system;
s2, determining an error variable z 1 And z i ,z 1 =x 1 -y d ,z i =x i -α i-1 Wherein α is i-1 Representing a virtual control function;
wherein, i =1, \8230, n,denotes the intermediate variable, g i G, a non-linear n-th order system intrinsic parameter 0 =z 0 =0,μ i1 、μ i2 、ε i Are all constant and are positive numbers, W i T S i (X i ) Representing a radial basis neural network, W i =[w i1 ,...,w iN ] T Is a weight matrix of the neural network, S i (X i ) Representing the basis function vector, S, of a neural network i (X i ) Each of the basis functions in (1) is a gaussian function having the same center and fixed width, X i =[x 1 ,...,x i ]The input vector of the neural network is defined, and N is the number of nodes of the neural network;
s3, establishing an error z i The input of the differential estimator is z i Output isIs->(ii) is estimated;
the differential estimator is:
where Ψ (y) = (1-e) -by )/(1+e -by ),λ, η, b are constants and are positive numbers, ω i Ψ (y) is used to estimate the sign function sgn (y), or->Is z i Is first derivative of->Is omega i ξ denotes the estimation error of the differential estimator, b @>An upper limit of ξ;
s4, obtained by S3Calculating estimation error of the current radial basis function neural network, performing gradient descent training on the weight of the neural network based on the estimation error to obtain->
Weight W i The update rate is specifically expressed as:
W n the update rate is specifically expressed as:
wherein it is present>Represents a weight W i First derivative of r i To the learning rate, r i >0;/>And &>Are respectively based on>And/or>(ii) an estimate of (d);
Control input signal u of nonlinear system:
2. the neural network backstepping control method based on error training as claimed in claim 1, wherein in S2, the method for obtaining the virtual control function comprises:
using the obtained error z i Designing a Lyapunov function V:
the derivative values are:
and selecting a virtual control function according to the derivative of the Lyapunov function.
3. The neural network backstepping control method based on error training as claimed in claim 1, wherein in S1, the S i (X i ) The kth element s of k (X i ) Represents:
wherein exp represents an exponential function, b k Representing the k-th basis function width, C, of the neural network k Is the k-th basis function center vector, C, of the neural network k =[c 1,k ,...,c n,k ] T Wherein c is i,k Is a constant.
4. A computer-readable storage device, in which a computer program is stored, which computer program, when being executed, carries out the method according to any one of claims 1 to 3.
5. An error-training-based neural network back-stepping control system, comprising a storage device, a processor, and a computer program stored in the storage device and executable on the processor, wherein the processor executes the computer program to implement the method of any one of claims 1 to 3.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111669359.4A CN114326405B (en) | 2021-12-30 | 2021-12-30 | Neural network backstepping control method based on error training |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111669359.4A CN114326405B (en) | 2021-12-30 | 2021-12-30 | Neural network backstepping control method based on error training |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114326405A CN114326405A (en) | 2022-04-12 |
CN114326405B true CN114326405B (en) | 2023-04-07 |
Family
ID=81020454
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111669359.4A Active CN114326405B (en) | 2021-12-30 | 2021-12-30 | Neural network backstepping control method based on error training |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114326405B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114815618B (en) * | 2022-04-29 | 2023-05-16 | 哈尔滨工业大学 | Adaptive neural network tracking control method based on dynamic gain |
CN115952731B (en) * | 2022-12-20 | 2024-01-16 | 哈尔滨工业大学 | Active vibration control method, device and equipment for wind turbine blade |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111176122A (en) * | 2020-02-11 | 2020-05-19 | 哈尔滨工程大学 | Underwater robot parameter self-adaptive backstepping control method based on double BP neural network Q learning technology |
CN112904726A (en) * | 2021-01-20 | 2021-06-04 | 哈尔滨工业大学 | Neural network backstepping control method based on error reconstruction weight updating |
Family Cites Families (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102385316B (en) * | 2011-09-16 | 2013-03-20 | 哈尔滨工程大学 | Deepening controlling method of underactuated automatic underwater vehicle based on neural network back stepping method |
CN103336433B (en) * | 2013-04-25 | 2016-10-19 | 常州大学 | Mixed self-adapting Predictive Control System based on Backstepping and forecast Control Algorithm thereof |
CN103885386A (en) * | 2014-04-08 | 2014-06-25 | 北京工业大学 | Gray model thermal error data processing method based on Kalman filtering |
CN107479384B (en) * | 2017-09-05 | 2019-10-22 | 西北工业大学 | The non-backstepping control method of hypersonic aircraft neural network Hybrid Learning |
CN107544256B (en) * | 2017-10-17 | 2020-02-14 | 西北工业大学 | Underwater robot sliding mode control method based on self-adaptive backstepping method |
CN110022137B (en) * | 2019-02-15 | 2022-11-01 | 华侨大学 | Simple complementary fusion filtering and differential estimation method |
CN110405757B (en) * | 2019-07-06 | 2021-01-01 | 大国重器自动化设备(山东)股份有限公司 | Intelligent robot based on neural network |
CN110829933B (en) * | 2019-11-11 | 2021-09-10 | 南京理工大学 | Neural network output feedback self-adaptive robust control method based on transmitting platform |
-
2021
- 2021-12-30 CN CN202111669359.4A patent/CN114326405B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111176122A (en) * | 2020-02-11 | 2020-05-19 | 哈尔滨工程大学 | Underwater robot parameter self-adaptive backstepping control method based on double BP neural network Q learning technology |
CN112904726A (en) * | 2021-01-20 | 2021-06-04 | 哈尔滨工业大学 | Neural network backstepping control method based on error reconstruction weight updating |
Also Published As
Publication number | Publication date |
---|---|
CN114326405A (en) | 2022-04-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN114326405B (en) | Neural network backstepping control method based on error training | |
Zhang et al. | Observer-based terminal sliding mode control of non-affine nonlinear systems: Finite-time approach | |
CN106249599B (en) | Neural network prediction-based networked control system fault detection method | |
CN106774379B (en) | Intelligent supercoiled strong robust attitude control method | |
CN107422741B (en) | Learning-based cluster flight distributed attitude tracking control method for preserving preset performance | |
CN110647042A (en) | Robot robust learning prediction control method based on data driving | |
CN107592048A (en) | The adaptive chaos control method of fractional order brushless DC motor system | |
CN111459051A (en) | Discrete terminal sliding mode model-free control method with disturbance observer | |
CN108489015B (en) | Air conditioning system temperature control method based on pole allocation and Pade approximation | |
CN111965976B (en) | Robot joint sliding mode control method and system based on neural network observer | |
CN104950677A (en) | Mechanical arm system saturation compensation control method based on back-stepping sliding mode control | |
CN106970636B (en) | Control system and control method for controlling speed of aircraft | |
CN110471768B (en) | FastPCA-ARIMA-based load prediction method | |
CN113589689B (en) | Sliding mode controller design method based on multi-parameter self-adaptive neural network | |
CN112904726B (en) | Neural network backstepping control method based on error reconstruction weight updating | |
CN102749843B (en) | Self-adaptive feedback protection dynamic surface controller structure and design method | |
CN115981162A (en) | Sliding mode control trajectory tracking method of robot system based on novel disturbance observer | |
Zhang et al. | Adaptive neural network control for a class of interconnected pure-feedback time-delay nonlinear systems with full-state constraints and unknown measurement sensitivities | |
CN115180461B (en) | Tension data driving control method for new energy automobile compressor motor production equipment | |
CN112346342B (en) | Single-network self-adaptive evaluation design method of non-affine dynamic system | |
CN110824922A (en) | Smith estimation compensation method based on six-order B-spline wavelet neural network | |
CN114063458B (en) | Preset performance control method of non-triangular structure system independent of initial conditions | |
CN115857329A (en) | Tethered spacecraft complex attitude stability control method based on event triggering | |
CN115016273A (en) | Predefined time stability control method and system for single-link mechanical arm | |
CN113459083A (en) | Self-adaptive fixed time control method and system for mechanical arm under event trigger |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |