CN106033189B - Flying robot's pose network response surface device - Google Patents
Flying robot's pose network response surface device Download PDFInfo
- Publication number
- CN106033189B CN106033189B CN201510113369.8A CN201510113369A CN106033189B CN 106033189 B CN106033189 B CN 106033189B CN 201510113369 A CN201510113369 A CN 201510113369A CN 106033189 B CN106033189 B CN 106033189B
- Authority
- CN
- China
- Prior art keywords
- neural network
- pose
- control
- response surface
- value
- 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.)
- Expired - Fee Related
Links
Abstract
The invention discloses a kind of flying robot's pose network response surface devices, the Pose Control especially suitable for more rotor flying robots.Controller includes: rolling optimization device, neural network prediction model, inverse neural network model, delay link, feedback element, addition comparator, subtraction comparator；Controller models flying robot's dynamic process using neural network, predicts；The initial value of iteration optimization algorithms is determined using inverse neural network；Local optimum is carried out using Newton-Raphson method；Controller also has feedback control function.Compared to modelling by mechanism, neural network model has higher precision, and the PREDICTIVE CONTROL based on this model has higher control performance；Using the method for objective function optimization, power consumption can be reduced to the greatest extent under the premise of guaranteeing control performance；Compared with PID control, control overshoot is small, and control effect is more steady；Compared with sliding formwork control, it will not lead to the problem of and tremble shake.
Description
Technical field
The present invention relates to the Pose Control devices of flying robot a kind of, are especially used for the pose of more rotor flying robots
Controller belongs to robotic technology field.
Background technique
Flying robot generally refers to be equipped with various sensors, the nolo flight with certain perception and intelligence
Device.There is the successful application case of some flying robots in recent years with the development of science and technology in succession, for example, plus
Alan's (Aeryon) laboratory development Scout of Waterloo of putting on airs is a kind of novel flying machine that offender can be enabled to quail
Device people, may replace monitoring camera；In Japan, flying robot helps peasant to manage rice field, monitors corn growing, sows agriculture
Medicine；In Africa, the people of wild animal is catched and killed in flying robot's tracking with video camera.Currently, every profession and trade is to flying robot
Demand it is larger, after having scholarly forecast 2015, the business application of flying robot explosively increases appearance, extensively
Application prospect also promotes the research of flying robot's technology, since current flying robot generally uses multi-rotor aerocraft,
Therefore hot spot is become to the control technology research of multi-rotor aerocraft；Multi-rotor aerocraft belongs to multi input, multi output, has by force
The nonlinear control system of coupling, lag etc., modelling by mechanism use Newton-Euller method mostly, and this model, which often exists, to be lost
With the higher problem of degree；Control to multi-rotor aerocraft, most commonly used at present is still PID control, due to PID control sheet
That control is realized according to the error generated in matter, therefore, it is difficult to error generate before just provide one for eliminate mistake
The control amount of difference, control precision be not high；There is document proposition to use sliding formwork control, but " trembles shake " present in sliding formwork control and ask
Topic is difficult to solve；Also there is document to propose PREDICTIVE CONTROL, but by using model, there are the higher problem of mismatch, control effects
Fruit is unsatisfactory.In addition, therefore, battery continues since flying robot is to be powered using battery, and battery is all heavier
Working time is always the bottleneck of flying robot's technology, how under the premise of guaranteeing to control precision, is farthest reduced
Power consumption is also larger problem present in control, and PID control and sliding formwork control are difficult to solve this problem.
In various intelligent control methods, it is based on prediction model that Model Predictive Control (MPC), which derives from industrial practice,
Finite time-domain optimal control.MPC is applied widely in recent years, but from the point of view of application, is mainly also limited to linear or quasi-
Linear process causes this phenomenon mainly due to expend very big cost in Nonlinear Mechanism modeling, and hardly results in standard
True model.
Artificial neural network can sufficiently approach complicated Nonlinear Mapping relationship, have study and adapt to uncertain system
Dynamic characteristic and stronger robustness and the characteristics of fault-tolerance, this make its become to nonlinear system establish prediction model and
One of key technology of optimal control.If artificial neural network is combined and can be suitable for predictive control algorithm
The network response surface of Control of Nonlinear Systems.
Summary of the invention
The invention discloses a kind of flying robot's pose network response surface devices, especially suitable for more rotor flyings
The Pose Control of robot；Main purpose is the relatively Accurate Model for solving the problems, such as flying robot, and is guaranteeing control performance
Under the premise of reduce power consumption the problem of.
The technical scheme adopted by the invention is that: flying robot's pose network response surface device, comprising: roll excellent
Change device, neural network prediction model, inverse neural network model, nuRank delay link, nwIt is rank delay link, first-order lag link, anti-
Feedback link, addition comparator, subtraction comparator；Network response surface device is using neural network prediction model to flying machine
People's dynamic process is modeled, is predicted；The initial value of iteration optimization algorithms is determined using inverse neural network；Using Newton-Raphson
Algorithm optimizes；Controller has feedback control function.
The dynamic process of flying robot may be expressed as: with discrete nonlinear equation
w(k+1) = f(u(k),u(k-1),…,u(k-nu+1),w(k) ,w(k-1),…,w(k-nw+1))
Wherein: w (k+1), w (k), w (k-1) ..., w (k-nw+ 1) kth+1 is respectively indicated, k, k-1 ..., k-nw+ 1 moment
The pose parameter value of flying robot；u(k),u(k-1),…,u(k-nu+ 1) kth is respectively indicated, k-1 ..., k-nu+ 1 moment flew
The control magnitude of row robot；nw, nuThe order of delay is respectively indicated, these numerical value can be measured by the method for experiment；
Function f indicates a unknown Nonlinear Mapping relationship.
By Rp=[u(k),u(k-1),…,u(k-nu+1),w(k),w(k-1),…,w(k-nw+ 1)] as neural network
Input, the output by w (k+1) as neural network only need to be according to the numerical value training neural network measured above, so that it may determine
The weight of neural network a, so that it is determined that neural network enables the dynamic of the neural network flyby robot special
Property.
For the controller of flying robot, at the kth moment, u (k-1) ..., u (k-nu+1), w(k) ,w(k-
1),…,w(k-nwIt+1) is obviously all known constant, as long as utilizing the neural network at this point, provide a control amount u (k)
The pose parameter w (k+1) at next moment can be calculated, so, this trained neural network can serve as one
The prediction model of a flying robot's pose.
In order to realize the purpose for reducing power consumption under the premise of guaranteeing control performance, rolling optimization can be technically used
Method, the objective function of optimization determines are as follows:
J = (wm(k+1)–wr’(k+1))2+λu(k)2
In formula: wm(k+1) it indicates by the pose value at the k+1 moment of Neural Network model predictive, wr' (k+1) indicate it is calibrated
It is externally input afterwards to refer to pose value, (wm(k+1)–wr’(k+1))2Value it is smaller, control performance is better, u (k)2Value get over
Small, then the electric energy consumed is fewer, and λ is weight factor, for compromising between control performance and power consumption.Target letter
Several minimum values corresponds to the optimum control amount that power consumption can be reduced under the premise of guaranteeing control performance.
For the real-time of Guarantee control system, a kind of local value optimization algorithm, that is, ox is used to the optimizing of objective function
- the inferior algorithm of pressgang, iterative formula are as follows:
In formula: u (k)j+1,u(k)jRespectively indicate the control amount of jth+1 time and iteration j, last point of the right
Subrepresentation objective function J is to the first-order partial derivative of control amount u (k) in u (k)=u (k)jWhen value, denominator indicate objective function J
To the second-order partial differential coefficient of control amount u (k) in u (k)=u (k)jWhen value.
Known by iterative formula, iteration needs to set the i.e. u (k) of initial value0, ideal scheme is to choose to have most
The control amount of excellent control performance is initial value, the advantage of doing so is that can prevent algorithm from falling into other minimums, to guarantee
The control performance of system；For this purpose, can be realized using an inverse neural network, the inverse kinematics process of flying robot can be indicated
For discrete nonlinear equation:
u(k) = g(u(k-1),…,u(k-nu+1), w(k+1),w(k) ,w(k-1),…,w(k-nw+1))
Wherein: the inverse Nonlinear Mapping relationship of function g expression f；As long as at this time by Ri=[ u(k-1),…,u(k-nu+1),
wr’(k +1),w(k), w(k-1),…,w(k-nw+ 1)] as the input of neural network, by u (k) as the defeated of neural network
Out, wr' (k+1) be taken as experimental data w (k+1), it can use the data of experiment measurement still just to train inverse neural network, this
The output of sample neural network enables to wm(k+1) = wr' (k+1), so corresponding to optimal control performance.
Feedback compensation part obtains error with actually detected value to realize using prediction pose value, and the error at k moment is w
(k)-wm(k), after correction are as follows: e=δ (w (k)-wm(k)), wherein δ is feedback correcting coefficient；The realization of feedback compensation passes through amendment
Reference input is realized, if k moment external reference input is wr(k+1), the then reference input after correcting are as follows: wr ’(k+1) =
wr (k+1)+ δ(w(k)-wm(k))。
To the order n of delayw, nu, the memory-aided method of implementation use benefit, such as to nuRank delay, uses
nuA storage unit stores kth, k-1 ..., k-n in kth sampling instant respectivelyuThe value of+1 sampling instant is adopted in kth+1
At the sample moment, updating storage unit is k+1 → k, k → k-1 ..., k-nu+2→k-nu+ 1, abandon kth-nuThe value at+1 moment.
In conclusion flying robot's network response surface device is completed by following steps:
Step 1: initialization, network response surface device needs to initialize following parameter: feedback compensation when system electrification
Coefficient δ, weight factor λ, terminal parameter α, maximum number of iterations ξ, delay link order nu 、nw, the number of iterations j=0, prolong
The value of slow link storage unit, wm(1)；.
Step 2: feedback compensation, network response surface device measures the pose parameter w (k) at k moment, according to wr ’(k+1)
= wr(k+1)+ δ(w(k)-wm(k)) to reference pose signal wr(k+1) it is corrected, the reference pose w after being correctedr ’
(k+1)。
Step 3: calculating the initial value of rolling optimization, the initial value u (k) of rolling optimization is calculated0, neural network prediction control
Device processed calls inverse neural network model, according to formula: u (k)0 = lwi×a(iwi×Ri+b1i)+ b2iCalculate the first of rolling optimization
Initial value u (k)0。
Step 4: carrying out rolling optimization according to Newton-Raphson iterative formula obtains optimum control amount,
(1) by formula:
It calculates u (k)j+1；
(2) if ︱ u (k)j+1 -u(k)j︱<α or j>ξ then turns the 5th step, otherwise, j+1 → j, u (k) is arrangedj+1→
u(k)j, turn (1).
Step 5: being controlled with optimum control amount obtained in the previous step, optimum control amount, u* (k)=(u (k) are calculatedj+1
+ u(k)j)/2, and control flying robot's controlled system is gone with u* (k), network response surface device calls nerve net
Network prediction model obtains the pose predicted value w of next sampling instantm(k+1), the storage unit values for updating delay link are u (k
+1)→u(k), u(k)→u(k-1),…, u(k-nu+2)→u(k-nu+ 1) the value u (k-n at earliest moment, is abandonedu+1)；w
(k)→w(k-1), w(k-1)→w(k-2),…, w(k-nw+2)→w(k -nw+ 1) the value w (k-n at earliest moment, is abandonedw+
1)。
The beneficial effects of the present invention are: neural network model has higher precision, to flying machine compared with modelling by mechanism
The dynamic process Approximation effect of device people is good, but also the PREDICTIVE CONTROL based on this model has higher control performance；It uses
The method of objective function optimization can reduce power consumption to the greatest extent under the premise of guaranteeing control performance；Compared with PID control, by
It can be controlled in advance in front of following error generates, therefore control overshoot is small, control effect is more steady；With sliding formwork control
It compares, will not lead to the problem of and tremble shake.
Detailed description of the invention
Fig. 1 is controller structure diagram of the invention, in figure: 1001, network response surface device；1002, rolling optimization
Device；1003, neural network prediction model；1004, inverse neural network model；1005,nuRank delay link；1006,nwRank ring retard
Section；1007, first-order lag link；1008, feedback element；1009, addition comparator；1010, subtraction comparator；1011, it flies
Robot controlled system.
Fig. 2 is neural network structure figure of the present invention, in figure: 2001, input vector；2002, input layer biases
Vector；2003, weight matrix is inputted；2004, accumulator；2005, neuron activation functions；2006, output layer bias vector；
2007, weight matrix is exported；2008, output vector.
Fig. 3 is quadrotor flying robot pose schematic diagram.
Fig. 4 is controller work flow diagram.
Specific embodiment
Embodiment 1: illustrate specific real-time mode, quadrotor flying robot position for selection quadrotor flying robot
For appearance as shown in figure 3, input control quantity is the speed control signal of four rotor motors, output quantity is robot pose parameter,
In, the control signal at k moment are as follows: u (k)=[u1(k),u2(k),u3(k),u4(k)], current quadrotor control signal is generally adopted
With duty ratio, i.e. u1(k),u2(k),u3(k),u4(k) duty ratio of the control square wave of four motors, pose parameter are respectively indicated
For w (k)=[ψ (k), φ (k), θ (k)], ψ (k), φ (k), θ (k) is respectively the yaw angle at k moment, roll angle, pitch angle.
It should be noted that since neural network is that the general mapping relations of one kind approach device, so proposed by the present invention
Method is suitble to various flying robots, the difference is that element number and physics dimension that dominant vector u (k) includes are different；
Although this illustrates for sentencing quadrotor flying robot, versatility of the invention is had no effect on.
Control system architecture as shown in Figure 1, in dotted line frame 1001 be network response surface device, specifically include:
1002, rolling optimization device；1003, neural network prediction model；1004, inverse neural network model；1005,nuRank delay link；
1006、nwRank delay link；1007, first-order lag link；1008, feedback element；1009, addition comparator；1010, subtraction ratio
Compared with device；In Fig. 1,1011 be flying robot's controlled system.
The pose data of experiment measurement quadrotor flying robot: 800 are randomly generated within the scope of the permissible value of control amount
A data, u1(1),u2(1),u3(1),u4(1), u1(2),u2(2),u3(2),u4(2),…,u1(200),u2(200),u3
(200),u4(200), these control amounts are inputted in corresponding sampling instant respectively, measures corresponding output pose parameter: ψ
(2),φ(2),θ(2), ψ(3) ,φ(3),θ(3),…, ψ(201) ,φ(201),θ(201)。
According to the data of the above experiment measurement, delay order n is determined with the method for parameter identificationu,nw。
Neural network prediction model is established as shown in Fig. 2, the input vector (2001) of neural network is Rp= [u(k), u
(k-1),…, u(k-nu+1), w(k), w(k-1),…, w(k-nw+1)]=[u1(k),u2(k),u3(k),u4(k), u1
(k-1),u2(k-1),u3(k-1),u4(k-1),…, u1(k-nu+1),u2(k-nu+1),u3(k-nu+1),u4(k-nu+1),ψ
(k),φ(k),θ(k),ψ(k-1),φ(k-1),θ(k-1),…, ψ(k-nw+1),φ(k-nw+1),θ(k-nw+ 1) it], inputs
The number of vector (2001) is 4 nu +3nw；Output vector (2008) is the prediction pose w at k+1 momentm(k+1) = [ψm (k+
1),φm (k+1),θm(k+1)]；Neuron activation functions (2005) use sigmoid function: a (t)=1/ (1+e-t)；Input
Layer bias vector (2002) b1m, input weight matrix (2003) iwm, output layer bias vector (2006) b2m, output weight square
Battle array (2007) lwmIt can be obtained according to the data of the above experiment measurement by neural network training method；Neural network prediction
The relational expression of model are as follows: wm(k+1) = lwm×a(iwm×Rp+b1m)+ b2m。
The structure of inverse neural network model is also as shown in Fig. 2, the input vector (2001) of network is Ri= [ wr’(k+1),
u(k-1),…, u(k-nu+1), w(k), w(k-1),…, w(k-nw+1)]=[ψ(k+1),φ(k+1),θ(k+1), u1
(k-1),u2(k-1),u3(k-1),u4(k-1),…, u1(k-nu+1),u2(k-nu+1),u3(k-nu+1),u4(k-nu+1), ψ
(k),φ(k),θ(k),ψ(k-1),φ(k-1),θ(k-1),…, ψ(k-nw+1),φ(k-nw+1),θ(k-nw+ 1) it], inputs
The number of vector (2001) is 4 nu +3nw-1；Output vector (2008) is control amount u (k)=[u at k moment1(k),u2
(k),u3(k),u4(k)]；Neuron activation functions (2005) still use sigmoid function: a (t)=1/ (1+e-t)；Input layer
Bias vector (2002) b1i, input weight matrix (2003) iwi, output layer bias vector (2006) b2i, output weight matrix
(2007) lwiIt can be obtained according to the data of the above experiment measurement by neural network training method；Inverse neural network prediction
The relational expression of model are as follows: u (k)=lwi×a(iwi×Ri+b1i)+ b2i。
The implementation of delay link (1005,1006,1007) uses the memory-aided method of benefit, such as to nuRank ring retard
It saves (1005), uses nuA storage unit stores kth, k-1 ..., k-n in kth sampling instant respectivelyu+ 1 sampling instant
Value, in+1 sampling instant of kth, updating storage unit is k+1 → k, k → k-1 ..., k-nu+2→k-nu+ 1, abandon kth-nu+
The value at 1 moment.
Feedback fraction is realized by subtraction comparator (1010), feedback element (1008), addition comparator (1009), according to figure
1, feedback quantity are as follows: δ (w (k)-wm(k)), wherein δ is feedback correcting coefficient, the reference input after feedback compensation are as follows: wr ’(k
+1) = wr (k+1)+ δ(w(k)-wm(k)), feedback compensation is achieved.
The objective function form of rolling optimization device (1002) are as follows: J=(wr ’(k+1)- wm(k+1))2+λu(k) 2,
In, λ is weight factor, is optimized using Newton-Raphson numerical value Local Optimization Algorithm, iterative formula are as follows:
Wherein, u (k)j+1,u(k)jRespectively indicate the control amount of jth+1 time and iteration j, last point of the right
Subrepresentation objective function J is to the first-order partial derivative of control amount u (k) in u (k)=u (k)jWhen value, denominator indicate objective function J
To the second-order partial differential coefficient of control amount u (k) in u (k)=u (k)jWhen value.
The initial value u (k) of iteration0It is determined by inverse neural network, i.e. u (k)0 = lwi×a(iwi×Ri+b1i)+ b2i, repeatedly
For termination condition are as follows: ︱ u (k)j+1 -u(k)j︱ < α, α are a lesser normal number；In addition, for the real-time for guaranteeing algorithm,
Maximum number of iterations ξ is set, if the number of iterations is higher than the value, terminates iterative process；Optimum control amount is by public affairs after the completion of iteration
Formula u* (k)=(u (k)j+1 + u(k)j)/2 obtain.
The working principle of network response surface device (1001): illustrating by taking the k sample moment as an example, neural network prediction control
Device (1001) processed measures the pose parameter w (k) at k moment first, according to wr ’(k+1) = wr(k+1)+ δ(w(k)-wm(k)) right
With reference to pose signal wr(k+1) it is corrected, the reference pose w after being correctedr ’(k+1)；Secondly, network response surface
Device (1001) calls inverse neural network model (1004), according to formula: u (k)0 = lwi×a(iwi×Ri+b1i)+ b2iCalculate rolling
The initial value u (k) of dynamic optimization0, optimum control amount u* (k) is obtained according to Newton-Raphson iterative formula, goes to control with u* (k)
Flying robot's controlled system (1011)；Finally, updating the storage unit values of delay link, network response surface device
(1001) neural network prediction model (1003) are called to obtain the pose predicted value w of next sampling instantm(k+1), in order to
Next sampling instant is used for feedback compensation.
Further, as shown in figure 4, the detailed work steps of network response surface device (1001) are as follows:
Step 1: initialization, network response surface device (1001) needs to initialize following parameter when system electrification: anti-
Present the order n of correction coefficient δ, weight factor λ, terminal parameter α, maximum number of iterations ξ, delay linku 、nw, the number of iterations j=
0, the value of delay link storage unit, wm(1)。
Step 2: feedback compensation, network response surface device (1001) measures the pose parameter w (k) at k moment first,
According to wr ’(k+1) = wr(k+1)+ δ(w(k)-wm(k)) to reference pose signal wr(k+1) it is corrected, after obtaining correction
Reference pose wr ’(k+1)。
Step 3: calculating the initial value of rolling optimization:
Calculate the initial value u (k) of rolling optimization0, the inverse neural network model of network response surface device (1001) calling
(1004), according to formula: u (k)0 = lwi×a(iwi×Ri+b1i)+ b2iCalculate the initial value u (k) of rolling optimization0。
Step 4: carrying out rolling optimization according to Newton-Raphson iterative formula obtains optimum control amount:
(1) by formula:
It calculates u (k)j+1；
(2) if ︱ u (k)j+1 -u(k)j︱<α or j>ξ then turns the 5th step, otherwise, j+1 → j, u (k) is arrangedj+1→
u(k)j, turn (1).
Step 5: being controlled with optimum control amount obtained in the previous step, optimum control amount, u* (k)=(u (k) are calculatedj+1
+ u(k)j)/2, and control flying robot's controlled system (1011), network response surface device are removed with u* (k)
(1001) neural network prediction model (1003) are called to obtain the pose predicted value w of next sampling instantm(k+1), update is prolonged
The storage unit values of slow link are u (k+1) → u (k), u (k) → u (k-1) ..., u (k-nu+2)→u(k -nu+ 1) it, abandons
Value u (k-the n at earliest momentu+1)；w(k)→w(k-1), w(k-1)→w(k-2),…, w(k-nw+2)→w(k –nw+ 1),
Abandon the value w (k-n at earliest momentw+1)。
Claims (4)
1. flying robot's pose network response surface device, comprising: rolling optimization device, neural network prediction model, inverse mind
Through network model, nuRank delay link, nwRank delay link, first-order lag link, feedback element, addition comparator, subtraction compare
Device；Network response surface device models flying robot's dynamic process using neural network prediction model, predicts；Make
The initial value of iteration optimization algorithms is determined with inverse neural network；It is optimized using Newton-Raphson method；Neural network prediction control
Device processed has feedback control function, and the objective function of rolling optimization device is determined as J=(wr ’(k+1)- wm(k+1))2+λu(k) 2,
Wherein, λ is weight factor, wr ’(k+1)、wm(k+1) be respectively+1 sampling instant of kth correction after with reference to pose and prediction bits
Appearance, pose parameter are w (k+1)=[ψ (k+1), φ (k+1), θ (k+1)], and ψ (k+1), φ (k+1), θ (k+1) is respectively k+1
The yaw angle at moment, roll angle, pitch angle, u (k) are the control amount of kth sampling instant；It is carried out using Newton-Raphson method
Local optimum, iterative formula are as follows:
In formula: u (k)j+1, u(k)jRespectively indicate the control amount of jth+1 time and iteration j, last molecule table of the right
Show objective function J to the first-order partial derivative of control amount u (k) in u (k)=u (k)jWhen value, denominator indicate J pairs of objective function
The second-order partial differential coefficient of control amount u (k) is in u (k)=u (k)jWhen value；Stopping criterion for iteration are as follows: ︱ u (k)j+1 -u(k)j︱ < α,
α is terminal parameter；For the real-time for guaranteeing algorithm, maximum number of iterations ξ is set, if the number of iterations is higher than the value, termination changes
For process；Optimum control amount presses formula u* (k)=(u (k) after the completion of iterationj+1 + u(k)j)/2 obtain；Neural network is pre-
It surveys model and inverse neural network model includes input vector, input layer bias vector, input weight matrix, accumulator, neuron
Activation primitive, output layer bias vector, output weight matrix, output vector；Neuron activation functions use sigmoid function: a
(t)=1/(1+e-t)；The input vector of neural network prediction model is Rp= [u(k), u(k-1),…, u(k-nu+1), w
(k), w(k-1),…, w(k-nw+ 1)], wherein nw, nuThe order of delay is respectively indicated, the number of input vector is 4 nu +
3nw；Output vector is the prediction pose w at k+1 momentm(k+1)；The input vector of inverse neural network model is Ri= [ wr’(k+
1), u(k-1),…, u(k-nu+1), w(k), w(k-1),…, w(k-nw+ 1)], the number of input vector is 4 nu +
3nw-1；Output vector is the control amount u (k) at k moment.
2. flying robot's pose network response surface device according to claim 1, it is characterised in that: nuRank delay
Link, nwRank delay link, first-order lag link use benefit memory-aided method when implementation, to nuRank delay link, makes
Use nuA storage unit stores kth, k-1 ..., k-n in kth sampling instant respectivelyuThe value of+1 sampling instant, in kth+1
Sampling instant, updating storage unit is k+1 → k, k → k-1 ..., k-nu+2→k-nu+ 1, abandon kth-nuThe value at+1 moment.
3. flying robot's pose network response surface device according to claim 1, it is characterised in that: feedback loop
Section, addition comparator, subtraction comparator constitute the feedback fraction of network response surface device, feedback quantity are as follows: δ (w (k)-wm
(k)), wherein δ is feedback correcting coefficient, the reference input after feedback compensation are as follows: wr ’(k+1) = wr (k+1)+ δ(w
(k)-wm(k)), wr(k+1) pose parameter before being expressed as k+1 time correction.
4. flying robot's pose network response surface device according to claim 1, it is characterised in that: work step
Are as follows:
Step 1: initialization, network response surface device needs to initialize following parameter: feedback correcting coefficient when system electrification
δ, weight factor λ, terminal parameter α, maximum number of iterations ξ, delay link order nu、nw, the number of iterations j=0, delay link
The value w of storage unitm(1)；
Step 2: feedback compensation, network response surface device measures the pose parameter w (k) at k moment, according to wr ’(k+1) =
wr(k+1)+ δ(w(k)-wm(k)) to reference pose signal wr(k+1) it is corrected, the reference pose w after being correctedr ’(k+
1)；
Step 3: calculating the initial value of rolling optimization:
Calculate the initial value u (k) of rolling optimization0, network response surface device calls inverse neural network model, according to formula: u
(k)0 = lwi×a(iwi×Ri+b1i)+ b2iCalculate the initial value u (k) of rolling optimization0, wherein lwiIndicate output weight square
Battle array, iwiIndicate input weight matrix, b1iIndicate input layer bias vector, b2iIndicate output layer bias vector；
Step 4: carrying out rolling optimization according to Newton-Raphson iterative formula obtains optimum control amount:
(1) by formula:
It calculates u (k)j+1；
(2) if ︱ u (k)j+1 -u(k)j︱<α or j>ξ then turns the 5th step, otherwise, j+1 → j, u (k) is arrangedj+1→u(k)j, turn
(1)；
Step 5: being controlled with optimum control amount obtained in the previous step, optimum control amount, u* (k)=(u (k) are calculatedj+1 +
u(k)j)/2, and control flying robot's controlled system is gone with u* (k), network response surface device calls neural network
Prediction model obtains the pose predicted value w of next sampling instantm(k+1), the storage unit values for updating delay link are u (k+
1)→u(k), u(k)→u(k-1),…, u(k-nu+2)→u(k-nu+ 1) the value u (k-n at earliest moment, is abandonedu+1)；w(k)
→w(k-1), w(k-1)→w(k-2),…, w(k-nw+2)→w(k -nw+ 1) the value w (k-n at earliest moment, is abandonedw+1)。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510113369.8A CN106033189B (en) | 2015-03-16 | 2015-03-16 | Flying robot's pose network response surface device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510113369.8A CN106033189B (en) | 2015-03-16 | 2015-03-16 | Flying robot's pose network response surface device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106033189A CN106033189A (en) | 2016-10-19 |
| CN106033189B true CN106033189B (en) | 2019-02-22 |
Family
ID=57150723
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510113369.8A Expired - Fee Related CN106033189B (en) | 2015-03-16 | 2015-03-16 | Flying robot's pose network response surface device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106033189B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106054598A (en) * | 2016-05-05 | 2016-10-26 | 安徽农业大学 | Robot adaptive steering single neuron PID control method |
| CN108280043B (en) * | 2018-01-29 | 2021-11-23 | 北京润科通用技术有限公司 | Method and system for quickly predicting flight trajectory |
| CN111210877B (en) * | 2020-01-03 | 2023-11-24 | 山东天岳先进科技股份有限公司 | Method and device for deducing physical parameters |
| CN111399375A (en) * | 2020-03-04 | 2020-07-10 | 重庆邮电大学 | Neural network prediction controller based on nonlinear system |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101763036A (en) * | 2009-12-29 | 2010-06-30 | 江苏大学 | Lysine fermentation process feeding prediction control system and method based on fuzzy neural network |
| CN101825871A (en) * | 2010-04-09 | 2010-09-08 | 哈尔滨工程大学 | Intelligent adaptive control method for heave and pitch device for oblique rudder ship |
| CN101963784A (en) * | 2010-09-07 | 2011-02-02 | 江苏科技大学 | Marine dynamic positioning method based on support vector machine and generalized predictive control algorithm |
| CN104238366A (en) * | 2014-09-29 | 2014-12-24 | 中国科学院自动化研究所 | Predictive control method and device of piezoelectric ceramic actuator on basis of neural network |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7362892B2 (en) * | 2003-07-02 | 2008-04-22 | Lockheed Martin Corporation | Self-optimizing classifier |
-
2015
- 2015-03-16 CN CN201510113369.8A patent/CN106033189B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101763036A (en) * | 2009-12-29 | 2010-06-30 | 江苏大学 | Lysine fermentation process feeding prediction control system and method based on fuzzy neural network |
| CN101825871A (en) * | 2010-04-09 | 2010-09-08 | 哈尔滨工程大学 | Intelligent adaptive control method for heave and pitch device for oblique rudder ship |
| CN101963784A (en) * | 2010-09-07 | 2011-02-02 | 江苏科技大学 | Marine dynamic positioning method based on support vector machine and generalized predictive control algorithm |
| CN104238366A (en) * | 2014-09-29 | 2014-12-24 | 中国科学院自动化研究所 | Predictive control method and device of piezoelectric ceramic actuator on basis of neural network |
Non-Patent Citations (1)
| Title |
|---|
| 基于RBF_ARX模型的预测控制在四旋翼飞行器控制系统中的应用;罗倩兰;《中国优秀硕士学位论文全文数据库》;20121231;41-49 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106033189A (en) | 2016-10-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Peng et al. | Robust adaptive formation control of underactuated autonomous surface vehicles with uncertain dynamics | |
| CN106033189B (en) | Flying robot's pose network response surface device | |
| Cheng et al. | Dynamic properties of Elman and modified Elman neural network | |
| Luo et al. | A laguerre neural network-based ADP learning scheme with its application to tracking control in the Internet of Things | |
| Fan et al. | Convergence of online gradient method for feedforward neural networks with smoothing L1/2 regularization penalty | |
| Song et al. | Robust backpropagation training algorithm for multilayered neural tracking controller | |
| CN103235620A (en) | Greenhouse environment intelligent control method based on global variable prediction model | |
| Fang et al. | A course control system of unmanned surface vehicle (USV) using back-propagation neural network (BPNN) and artificial bee colony (ABC) algorithm | |
| CN116661307A (en) | Nonlinear system actuator fault PPB-SIADP fault-tolerant control method | |
| Hsu et al. | Recurrent fuzzy-neural approach for nonlinear control using dynamic structure learning scheme | |
| Perng et al. | FOPID controller optimization based on SIWPSO-RBFNN algorithm for fractional-order time delay systems | |
| Belhaj Salah et al. | A greenhouse modeling and control using deep neural networks | |
| CN107490962B (en) | Data-driven optimal control method for servo system | |
| de Jesus Rubio et al. | Comparison of three proposed controls to accelerate the growth of the crop | |
| Bakshi | Model reference adaptive control of quadrotor UAVs: A neural network perspective | |
| Jiang et al. | Research on manipulator trajectory tracking with model approximation RBF neural network adaptive control | |
| Zhou et al. | An improved EMD online learning-based model for gold market forecasting | |
| Yang et al. | PID controller based on the artificial neural network | |
| CN107065522B (en) | Fuzzy slow state feedback H-infinity control method for nonlinear switching double-time scale system | |
| Sun et al. | Unmanned aerial vehicles control study using deep deterministic policy gradient | |
| Sitompul | Adaptive neural networks for nonlinear dynamic systems identification | |
| Hendzel et al. | Neural network identifier of a four-wheeled mobile robot subject to wheel slip | |
| Jiangming et al. | Self-Tuning PID controller based on improved BP neural network | |
| Katal et al. | Design of Optimal QFT Controller and Prefilter for Buck Converter Using Metaheuristic Algorithms | |
| Liu | On a method of single neural PID feedback compensation control |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190222 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |