CN103963938A - Intelligent self-adaptive prediction control system based on slant-rudder ship anti-pitching device - Google Patents

Abstract

The invention relates to a method for controlling a slant-rudder ship anti-pitching device, in particular to an intelligent self-adaptive prediction control system based on the slant-rudder ship anti-pitching device. The intelligent self-adaptive prediction control system comprises a heaving and pitching reference information generator, an input reference trajectory generator, an information constraint system, a control quantity fusion information generator, a control quantity estimator, a co-state fusion information generator, a co-state sequence estimator, a co-state fusion information quantity inverse time sequence initial value estimator, a co-state inverse time sequence initial value estimator, a sensor system and a controller. The intelligent self-adaptive prediction control system is suitable for a system such as the slant-rudder ship anti-pitching device which has nonlinearity, coupling and time variation, and is high in control accuracy and good in robustness.

Description

A kind of based on oblique rudder ship antipitching device intelligent adaptive Predictive Control System

Technical field

What the present invention relates to is a kind of oblique rudder ship antipitching device control method.

Background technology

Oblique rudder ship antipitching device multi-variable system, due to non-linear, close coupling, time stickiness and the characteristic such as parameter is uncertain, make the design of its control system there is certain challenge.

Ship stabilization is one of important subject of Naval Architecture and Ocean Engineering with controlling.Rollstabilization aspect has obtained many achievements in research at present.Aspect boats and ships pitching stabilization, achievement in research is little, and they adopt half submerged body and fin the pitching stabilization such as to combine.Rudder stabilization has many advantages compared with stabilizer, antirolling tank.It can utilize existing rudder and maneuvering system on ship, only need increase controller, sensor a few components, not only economy but also practicality.Prerequisite is that boats and ships are to liking Twin Rudders or many rudders.

Document " Chen Hongli; Zhao Xiren; Ye Kui; Peng Xiuyan. the LQG method research of boats and ships pitching stabilization control. the journal .2004 of Harbin Engineering University; 25 (4): 407-411 " in ignored the coupling between rolling, pitching, utilize the pitching stabilization of ∏ type diving rudder, and be to carry out linear quadratic gaussian (LQG) for permanent equation under Longitudinal Movement of Ship typical case sea situation to control.The present invention adopts 45 ° of flap type rudders that tilt to install, and can realize the decoupling zero control of rolling, pitching, improves lift of rudder coefficient, and pitching subtracts and while shaking, do not produce driftage control torque.

Many scholars, by the research to rudder stabilization for many years, find that the uncertain problem of motion of ship model becomes one of key factor of left and right rudder stabilization technology application success or failure.

The present invention utilizes the tank experiments test result of strip theory according to hydrodynamic force coefficient under typical sea situation, set up the ship control intelligent adaptive model with ship course (angle), the speed of a ship or plane and sea condition real-time change, in conjunction with suppressing the impact of ship model uncertainty on pitching stabilization effect based on oblique rudder ship antipitching device information fusion predictive control with the robustness that improves controller.Thereby ensure the safe and reliable operation of oblique rudder ship antipitching device.

Summary of the invention

The object of the invention is to have serious coupling and non-linear for oblique rudder ship antipitching device, provide a kind of applicable varying duty based on tandem combination neural net model, can obtain the oblique rudder ship antipitching device self-adaptation control method of good controller performance.

The object of the present invention is achieved like this:

Based on oblique rudder ship antipitching device intelligent adaptive Predictive Control System, comprising hangs down swings with pitching reference information maker, input reference locus maker, information constrained system, controlling quantity fuse information maker, controlling quantity estimator, the fuse information maker of assisting state, association's status switch estimator, association's state fusion quantity of information against time series initial value estimator, association's state against time series initial value estimator, sensing system, controller.Input reference locus maker receive hang down swing with the data of pitching reference information maker and sensing system after obtain inputting reference locus, send to association's state against time series initial value estimator and assist status switch estimator parallel input reference locus; Controlling quantity fuse information maker receives controlled fuse information amount after the data of information constrained system, and the fuse information amount of control is passed to controlling quantity estimator; Association's state receives association's state fusion quantity of information against time series initial value estimator and obtain association's state against time series initial estimate amount after the data of time series initial value estimator, association's state is issued to association's status switch estimator against time series initial estimate amount, assist status switch estimator to receive estimated sequence between the inverse time that obtains assisting state after the data of fuse information maker of association state, controlling quantity estimator receives between inverse time of association's state and obtains optimal control amount sequence after estimated sequence data simultaneously; Last controller sends to control command the actuating unit steering wheel of boats and ships, realizes the pitching control of surface ship.

Oblique rudder ship antipitching device intelligent adaptive forecast Control Algorithm of the present invention can also comprise:

It is that boats and ships hang down to swinging with pitch angle reference information and offer definite value (expectation value) that described oblique rudder ship antipitching device vertical swings with pitching reference information maker.

The sensing system of described oblique rudder ship antipitching device refers to, under certain sea condition, gathers boats and ships course made good, the speed of a ship or plane, rudder angle, hangs down and swing and displacement pickup, accelerometer, the gyroscope of pitch angle, log, boat-carrying theodolite.

The input reference locus maker of described oblique rudder ship antipitching device is

${Y^{*}}_j={\mathrm{\Ω}}^{j-k}{Y}_k+{(1-\mathrm{\Ω})}^{j-k}{\stackrel{\‾}{R}}_k$

In formula, Y _khang down for the boats and ships of sensing system collection are actual swing, pitch angle, Y _jfor to Y _k(j-k) forward step prediction, swing and pitching reference information given value for hanging down, Ω ∈ (0,1) is softening coefficient, and Ω choose the overall balance needing between taking into account system response characteristic and robustness.

The information constrained system of described oblique rudder ship antipitching device, comprises following 3 information representation formulas:

The hard constraint information representation formula being determined by system state equation is

${\stackrel{\‾}{X}}_j=\stackrel{\‾}{\mathrm{\Φ}}{\stackrel{\‾}{X}}_{j-1}+\stackrel{\‾}{\mathrm{\Ψ}}{u}_{j-1}+\stackrel{\‾}{\mathrm{\Γ}}{W}_{f,j-1}$

In formula for system state equation matrix of coefficient, W _{f, j-1}for zero-mean and covariance matrix are I _{2 × 1}white noise.

The soft-constraint information representation formula being determined by association's state estimation error is:

${\widehat{\stackrel{\‾}{X}}}_j={\stackrel{\‾}{X}}_j+E_j$

E in formula _jfor zero-mean and covariance matrix are white noise.

By requiring to control the as far as possible little soft-constraint information representation formula determining of energy be:

0＝u _j-1+n _j-1

N in formula _j-1for zero-mean and covariance matrix are R _j-1 ^-1white noise.

By the soft-constraint information representation formula that requires system output tracking input reference locus to determine

$Y_j^{*}=Y_j+{\stackrel{\‾}{M}}_j$

In formula for zero-mean and covariance matrix for white noise.

The controlling quantity fuse information maker of described oblique rudder ship antipitching device is

In formula covariance matrix.

Association's state fusion quantity of information of described oblique rudder ship antipitching device against time series initial value estimator is:

$P_{k+N}^{-1}={\stackrel{\‾}{D}}^T{Q}_{k+N}\stackrel{\‾}{D}+{\stackrel{\‾}{\mathrm{\Φ}}}^T{(I_{n\×n}/\mathrm{\λ}+\stackrel{\‾}{\mathrm{\Ψ}}{R}_{k+N}^{-1}{\stackrel{\‾}{\mathrm{\Ψ}}}^T)}^{-1}\stackrel{\‾}{\mathrm{\Φ}}$

In formula for systematic observation matrix, λ is enough little positive nonzero number, I _{n × n}for the identity matrix of n × n dimension, Q _jfor covariance matrix, V _jfor systematic survey noise covariance matrix.

Association's state of described oblique rudder ship antipitching device against time series initial value estimator is

${\widehat{\stackrel{\‾}{X}}}_{k+N}=P_{k+N}\{-{\stackrel{\‾}{\mathrm{\Φ}}}^T{(I_{n\×n}/\mathrm{\λ}+\stackrel{\‾}{\mathrm{\Ψ}}{R}_{k+N}^{-1}{\stackrel{\‾}{\mathrm{\Φ}}}^T)}^{-1}\stackrel{\‾}{\mathrm{\Ψ}}+{\stackrel{\‾}{D}}^T{Q}_{k+N}{Y}_{k+N}^{*}\}$

The fuse information maker of association's state of described oblique rudder ship antipitching device is

In formula,

Association's status switch estimator of described oblique rudder ship antipitching device is

${\widehat{\stackrel{\‾}{X}}}_j=P_j[{\stackrel{\‾}{\mathrm{\Φ}}}^T{O}_j^{-1}({\widehat{\stackrel{\‾}{X}}}_{j+1}-\stackrel{\‾}{\mathrm{\Ψ}})+{\stackrel{\‾}{D}}^T{Q}_j{Y}_j^{*}]$

J=k+1 in formula ..., k+N.

The controlling quantity estimator of described oblique rudder ship antipitching device is

Beneficial effect of the present invention is:

The present invention adopts 45 ° of flap type rudders that tilt to install, and has set up the hydrodynamic force nonlinear adaptive model of mind (tandem combination neural net model) according to the tank experiments test result of hydrodynamic force coefficient under typical sea situation based on strip theory.According to instrument on ship, can record in real time the course angle in variation, the speed of a ship or plane and sea condition, utilize hydrodynamic force self-adapting intelligent model to online calculate the hydrodynamic force coefficient of oblique rudder ship antipitching device, constantly update oblique rudder ship antipitching device model, then neural network adaptive model is combined with predictive control, because becoming institute when slow, procedure parameter causes that forecast model output error can revise timely, thereby improve the dynamic property of system, being applicable to oblique rudder ship antipitching device has non-linear like this, coupling, the system of time variation, control accuracy is high, robustness is good.

Brief description of the drawings

Fig. 1 is the structured flowchart of oblique rudder ship antipitching device intelligent adaptive information fusion Predictive Control System;

Fig. 2 is oblique rudder control chart;

Fig. 3 is generalized regression nerve networks constructional drawing;

Fig. 4 is optimal stopping rule optimized network flow chart;

Fig. 5 expanding system block diagram;

Fig. 6 is tandem combination neural net constructional drawing;

Fig. 7 is for controlling effect relative value.

Specific implementation method

Below in conjunction with accompanying drawing, the present invention is described further.

The present invention is to provide a kind of intelligent adaptive Predictive Control System of the oblique rudder ship antipitching device based on information fusion best guess.Under certain sea condition, sensing system gathers boats and ships course made good, the speed of a ship or plane, rudder angle, hangs down and swings and pitch angle, hangs down and swings and pitching reference information maker in conjunction with boats and ships, generates input reference locus; Quadratic Optimal Control performance index function is converted into the soft-constraint information to system keeps track input and control energy, the hard constraint information of coupling system equation of state, all information of fused controlling amount, obtain association's status switch under rolling optimization performance figure and the best guess of control sequence; Evaluate the control effect of controlling quantity according to controller performance critical for the evaluation, and revise current controlling quantity according to Actual Control Effect of Strong, and finally select the control inputs of optimum oblique rudder ship antipitching device, pass to controller and generate control command, drive steering wheel, realize the pitching stabilization control of surface ship.

Principle of work of the present invention is: core of the present invention is the design of oblique rudder ship antipitching device intelligent adaptive advanced control system.This system is the combination of oblique rudder ship antipitching device intelligent adaptive integrated model and information fusion predictive control strategy.

Oblique rudder ship antipitching device intelligent adaptive Predictive Control System is the output in future of the multistep prediction device controlled system under following reference locus effect based on intelligent adaptive integrated model, according to rolling optimization criterion, combining information fusion forecasting control policy, controlled Performance Evaluating Indexes, evaluate the control effect of controlling quantity according to controller performance critical for the evaluation, and revise current controlling quantity according to Actual Control Effect of Strong, finally select the control inputs of optimum oblique rudder ship antipitching device.As shown in Figure 1, oblique rudder ship antipitching device intelligent adaptive advanced control method of the present invention, is mainly made up of the link such as multistep prediction device and information fusion predictive control based on intelligent adaptive integrated model.In Fig. 1, for the j step predictor of state, wherein 1≤j≤N+1, N is prediction step; Y ^*(k+j) be reference locus, tandem combination neural net and oblique rudder ship antipitching device composition intelligent adaptive integrated model.

The oblique rudder of first step utilization is realized the decoupling zero control in ship rolling, pitching and course

The installation at 45 ° of rudder and horizontal surface, when roll damping (Fig. 2 (a)), left rudder guide margin is upwards time, and right standard rudder guide margin is downward, and the pitchmoment opposite direction that left and right rudder causes, cancels out each other; Pitching subtracts (Fig. 2 (b)) while shaking, left and right rudder guide margin in the same way, only produce making a concerted effort vertically upward, the component that left and right rudder produces is in the horizontal direction cancelled out each other, thus, rolling control and pitching control have realized automatically decouple. and simultaneously pitching subtracts and while shaking, does not produce driftage control torque.

The foundation of the second step Longitudinal Movement of Ship differential equation

Set up the Longitudinal Movement of Ship differential equation based on strip theory

Document [Chen Hongli. based on the longitudinal multivariate Random Control Method research of π type rudder ship [D]. Harbin: Harbin Engineering University's doctorate paper, 2004.12] utilize strip theory, according to Newton's second law, set up boats and ships under irregular wave action and hung down and swing and coupled motions (Longitudinal Movement of Ship) differential equation of pitching:

$\begin{array}{c}(a_{33}+\mathrm{\Δ})\stackrel{\·\·}{z}+b_{33}\stackrel{\·}{z}+c_{33}z+a_{35}\stackrel{\·\·}{\mathrm{\θ}}+b_{35}\stackrel{\·}{\mathrm{\θ}}+c_{35}\mathrm{\θ}=F_R+F_{w3}\\ a_{53}\stackrel{\·\·}{z}+b_{53}\stackrel{\·}{z}+c_{53}z+(I_5+a_{55})\stackrel{\·\·}{\mathrm{\θ}}+b_{55}\stackrel{\·}{\mathrm{\θ}}+c_{55}\mathrm{\θ}=F_R{X}_R+M_{w3}\end{array}---\left(1\right)$

In formula:

Z---represent that the vertical of hull swings, m

θ---represent the pitching of hull, rad

Pitching moment of inertia I ₅, kg.m ²

Displacement Δ, kg

F _r---represent lift of rudder, kg.m/s ²

X _r---represent the fore-and-aft distance of lift of rudder center to hull center of gravity, m

F _w3represent the vertical exciting force, kg.m/s of swinging of wave ²

M _w5represent wave pitching disturbance torque, kg.m ²/ s ²

A ₃₃---represent added mass, kg,

A ₃₅---represent moment of mass, kg.m,

A ₅₃---represent moment of mass, kg.m,

A ₅₅---represent rotor inertia, kg.m ²,

B ₃₃---represent dumping force coefficient,

B ₃₅---represent dumping force coefficient, $b_{35}=V{\∫}_{H_s}^{H_b}{A}_{33}^{\′}\mathrm{dx}-{\∫}_{H_s}^{H_b}{N}_{22}^{\′}\mathrm{xdx}$

C ₃₃---represent recuperability coefficient,

C ₃₅---represent recuperability coefficient, $c_{35}=V{\∫}_{H_s}^{H_b}{N}_{33}^{\′}\mathrm{dx}-\mathrm{\ρg}{\∫}_{H_s}^{H_b}{B}_{22}^{*\′}\mathrm{xdx}$

B ₅₃---represent damping moment coefficient,

B ₅₅---represent damping moment coefficient,

C ₅₃---represent righting moment coefficient,

C ₅₅---represent righting moment coefficient, $c_{55}=\mathrm{\ρg}{\∫}_{H_s}^{H_b}{B^{*}x}^2\mathrm{dx}-V{\∫}_{H_s}^{H_b}{N}_{22}^{\′}\mathrm{xdx}-V^2{\∫}_{H_S}^{H_b}{A}_{33}^{\′}\mathrm{dx}$

B ^*---represent the local cross-sectional width of boats and ships, m

V---represent speed of the ship in meters per second, kn

H _b---represent the coordinate of bow, m

H _s---represent the coordinate of stern, m

A ' ₃₃---represent the vertical added mass in cross section,

N ' ₂₂---represent the local heave damping coefficient of boats and ships,

---can be calculated by Griem collection of illustrative plates.

Wherein a _ij, b _ij, c _ij(i, j=3 or 5)---be referred to as hydrodynamic force coefficient, wherein c _ijcan be by calculating and being constant.A _ij, b _ijuse w _ijrepresent, obtain by formula (2):

$w_{\mathrm{ij}}=\left(\underset{k=1}{\overset{30}{\mathrm{\Σ}}}\sqrt{2s\left({\mathrm{\ω}}_k\right){\mathrm{d\ω}}_k}\right|w_{\mathrm{ijk}}\left|\right)/\underset{k=1}{\overset{30}{\mathrm{\Σ}}}\sqrt{2s\left({\mathrm{\ω}}_k\right){\mathrm{d\ω}}_k}---\left(2\right)$

According to strip theory, utilization is calculated to obtain w in the data of the regular wave disturbance sink experiment measuring of the different speed of a ship or plane, course and different frequency _ijk.

S (ω) is random sea wave spectrum, adopts 15 ^ththe two-parameter spectrum of ITTC, wherein H _1/3there is adopted wave height, T ₁on average spend null cycle, they are by different sea condition values.

Utilize formula (2) calculate hydrodynamic force coefficient be with sea condition, the speed of a ship or plane and course change but with the typical sea situation of frequency-independent under hydrodynamic force coefficient a _ij, b _ij.In order in real time, accurately to control Ship Motion Attitude, need set up a motion of ship 3 D intelligent model based on the speed of a ship or plane, sea condition and course angle real-time change, the hydrodynamic force coefficient in the equation of motion is continually varying nonlinear function in three dimensional space.Approaching quickly and accurately these Nonlinear Curved is to set up the key of model of mind.

2. the structure of generalized regression nerve networks and principle

Utilize radial basis neuron and linear neuron can set up generalized regression nerve networks (Generalized RegressionNeural Network-GRNN), the network of this form is through being usually used in approximation of function.GRNN structure as shown in Figure 3.This network comprises two-layer, and interlayer (hidden layer) is basic unit radially, and output layer is linear layer.

GRNN shown in Fig. 3 has Q group input vector, and the element number of every group of vector is m, and interlayer has n radial basis neuron.

For fear of the phenomenon of crossing study, improve training speed and the generalization ability of GRNN, utilize optimal stopping rule to be optimized network.

Fig. 4 flow chart, wherein parameter when utilizing optimal stopping rule: training set T, checking collection C, current validation error Errornew, last time validation error Erroroed, weight w _ij, counting machine count, fault-tolerant thresholding S, frequency of training zone bit x0, frequency of training k, training error E, error cost function ε, cycle number R, initial weight the weights variation delta w trying to achieve by gradient descent method _ij.

3. the foundation of Longitudinal Movement of Ship hydrodynamic force model of mind

Longitudinal boats and ships hydrodynamic force coefficient is pressed to sea condition, speed of a ship or plane refinement grouping, adopt generalized regression nerve networks to set up the tandem combination neural net model of hydrodynamic force coefficient.

The 3rd step, based on oblique rudder ship antipitching device information fusion predictive control.

Oblique rudder ship pitching stabilization discrete state spatial model based on wave disturbance modeling

Utilize Longitudinal Movement of Ship differential equation (1), set up equation of state and observational equation discretization and (get sampling period T _s=0.2 second), utilize hydrodynamic force model of mind to obtain hydrodynamic force coefficient or 5, substitution equation of state is tried to achieve the eigenwert of system, judges the stability of system, can control, controllability.

As follows based on oblique rudder ship pitching stabilization discrete time state equation

X _k＝ΦX _k-1+Ψu _k-1+ΓW _k-1(3)

Y _k＝DX _k+V _k(4)

In formula: $X={\left[\begin{array}{cccc}{x}_1& x_2& x_3& x_4\end{array}\right]}^T={\left[\begin{array}{cccc}z& \stackrel{\·}{z}& \mathrm{\θ}& \stackrel{\·}{\mathrm{\θ}}\end{array}\right]}^T,$ u＝F _R，Y＝[x ₁x ₃] ^T， $D=\left(\begin{array}{cccc}1& 0& 0& 0\\ 0& 0& 1& 0\end{array}\right),$ V is two-dimensional measurement noise, W ^t=[F _w3m _w5] be the disturbance of random seaway to boats and ships, regard system interference as, subscript k is the value moment after relevant variable discretization.

The spectral density of wave disturbance (comprising disturbance force and distrubing moment) is a narrow-band spectrum, can be considered to a kind of coloured noise with Gaussian distribution, (output signal is at this moment no longer real wave disturbance to regard wave disturbance power and distrubing moment spectrum the output of the formed filter that white noise drives as, but the two has identical statistical property, be all zero-mean, the coloured noise of identical covariance and same spectrum density), as shown in Figure 5.Wherein, w ₃(t) be wave disturbance power forming filter G ₃(s) input white noise signal, w ₅(t) be wave disturbance moment forming filter G ₅(s) input white noise signal, their averages are zero, variance is 1; Y _f(t) be the coloured noise with wave disturbance with identical statistical property, it is used for " substituting " original wave disturbance W (t),

Wave is chosen as follows to the formed filter of the longitudinal disturbance force of boats and ships and distrubing moment, (in order to facilitate the exponent number of ship control system research and design and reduction institute ships control system, the order of G (s) should not be obtained too high):

$G_i\left(s\right)=\frac{y_i\left(s\right)}{w_i\left(s\right)}=\frac{b_{i1}s}{s^2+a_{i1}+a_{i2}},i=\mathrm{3,5}---\left(5\right)$

Utilize wave disturbance spectrum, obtain a by least square method of recursion _i1, a _i2, b _i1, i=3,5.

Transfer function is above converted to state space form,

${\stackrel{\·}{m}}_i=A_i{m}_i+C_i{w}_i---\left(6\right)$

y _i＝H _im _i(7)

In formula: $A_i=\left[\begin{array}{cc}0& 1\\ -a_{i2}& -a_{i1}\end{array}\right],C_i=\left[\begin{array}{c}0\\ 1\end{array}\right],$ H _i＝[0 b _i1]，i＝3,5。

By i=3,5 are associated with

$\begin{array}{c}\stackrel{\·}{M}=A_{f}M+C_f{W}_f\\ Y_f=H_{f}M\end{array}---\left(8\right)$

In formula: M ^t=[m ₃m ₅], $W_f^T=\left[\begin{array}{cc}{w}_3& w_5\end{array}\right],Y_f^T=\left[\begin{array}{cc}{y}_3& y_5\end{array}\right],$ A _f＝diag[A ₃Α ₅]，

Fig. 5 is equivalent to the series connection (being called expanding system) of original system and formed filter, and after expansion, the disturbance of new system is all white noise, i.e. its input is dimensional Gaussian white noise sequence, and output remains the observed reading of state z and θ, and expanding system can be expressed as

$\begin{array}{c}\stackrel{\·}{\stackrel{\‾}{X}}=\stackrel{\‾}{A}\stackrel{\‾}{X}+\stackrel{\‾}{B}u+\stackrel{\‾}{C}{W}_f\\ Y=\stackrel{\‾}{H}\·\stackrel{\‾}{X}+V\end{array}---\left(9\right)$

In formula: $\stackrel{\‾}{X}=\left[\begin{array}{c}X\\ M\end{array}\right],\stackrel{\‾}{A}=\left[\begin{array}{cc}A& {\mathrm{CH}}_f\\ 0& A_f\end{array}\right],\stackrel{\‾}{B}=\left[\begin{array}{c}B\\ 0\end{array}\right],\stackrel{\‾}{C}=\left[\begin{array}{c}0\\ C_f\end{array}\right],\stackrel{\‾}{H}=\left[\begin{array}{cc}H& 0\end{array}\right],$ Q _wff＝diag[1 1]。

Equation after discretization is

${\stackrel{\‾}{X}}_k=\stackrel{\‾}{\mathrm{\Φ}}{\stackrel{\‾}{X}}_{k-1}+\stackrel{\‾}{\mathrm{\Ψ}}{u}_{k-1}+\stackrel{\‾}{\mathrm{\Γ}}{W}_{f,k-1}---\left(10\right)$

$Y_k=\stackrel{\‾}{D}{\stackrel{\‾}{X}}_k+V_k---\left(11\right)$

Oblique rudder ship pitching stabilization information fusion predictive control

If objective function is:

$J=E[\underset{j=k+1}{\overset{k+N}{\mathrm{\Σ}}}{(Y_j-{Y^{*}}_j)}^2+\underset{j=k+1}{\overset{k+N_u}{\mathrm{\Σ}}}{\left({\mathrm{Ru}}_{j-1}\right)}^2]---\left(12\right)$

Wherein, N is prediction step, N _ufor step size, R is for controlling weighting factor, Y _jfor to Y _k(j-k) forward step prediction, Y ^* _jfor the input reference locus after filter softening, be taken as::

${Y^{*}}_j={\mathrm{\Ω}}^{j-k}{Y}_k+{(1-\mathrm{\Ω})}^{j-k}{\stackrel{\‾}{R}}_k---\left(13\right)$

In formula, swing and pitching reference information given value for hanging down, Ω ∈ (0,1) is softening coefficient, and Ω choose the overall balance needing between taking into account system response characteristic and robustness.

The object of information fusion control is by merging all information about controlling quantity, estimating optimal control amount sequence, is the information fusion SEQUENTIAL ALGORITHM that solves optimal control amount sequence below.

About controlling quantity, comprise following 3 information representation formulas:

The hard constraint information representation formula being determined by system state equation is

${\stackrel{\‾}{X}}_j=\stackrel{\‾}{\mathrm{\Φ}}{\stackrel{\‾}{X}}_{j-1}+\stackrel{\‾}{\mathrm{\Ψ}}{u}_{j-1}+\stackrel{\‾}{\mathrm{\Γ}}{W}_{f,j-1}---\left(14\right)$

2) the soft-constraint information representation formula being determined by association's state estimation error is:

${\widehat{\stackrel{\‾}{X}}}_j={\stackrel{\‾}{X}}_j+E_j---\left(15\right)$

3) by requiring to control the as far as possible little soft-constraint information representation formula determining of energy be:

0＝u _j-1+n _j-1(16)

E in formula _jfor zero-mean and covariance matrix are white noise, n _j-1for zero-mean and covariance matrix are R _j-1 ^-1white noise.

Simultaneous formula (14) and (15) can obtain

${\widehat{\stackrel{\‾}{X}}}_j=\stackrel{\‾}{\mathrm{\Φ}}{\stackrel{\‾}{X}}_{j-1}+\stackrel{\‾}{\mathrm{\Ψ}}{u}_{j-1}+\stackrel{\‾}{\mathrm{\Γ}}{W}_{f,j-1}+E_j---\left(17\right)$

Formula (17) is further converted to

${\widehat{\stackrel{\‾}{X}}}_j=\stackrel{\‾}{\mathrm{\Φ}}{\stackrel{\‾}{X}}_{j-1}+\stackrel{\‾}{\mathrm{\Ψ}}{u}_{j-1}+{\mathrm{\ω}}_j---\left(18\right)$

ω in formula _jfor zero-mean and covariance matrix are white noise

Controlling quantity fuse information maker is

Controlling quantity estimator is

About association's state against time series , comprise altogether 4 information representation formulas:, in information representation formula (14)～(16) and performance index function formula (12) by the soft-constraint information representation formula that requires system output tracking input reference locus to determine

$T_j^{*}=Y_j+{\stackrel{\‾}{M}}_j---\left(21\right)$

In formula for zero-mean and covariance matrix are white noise.

Simultaneous information representation formula (16), (18), are estimated as about the information fusion of association's status switch in formula (20)

${\widehat{\stackrel{\‾}{X}}}_{j+1}-{\mathrm{\ω}}_{j+1}=\stackrel{\‾}{\mathrm{\Φ}}{\stackrel{\‾}{X}}_j+\stackrel{\‾}{\mathrm{\Ψ}}-\stackrel{\‾}{\mathrm{\Ψ}}{n}_j---\left(22\right)$

?

${\widehat{\stackrel{\‾}{X}}}_j=\stackrel{\‾}{\mathrm{\Φ}}{\stackrel{\‾}{X}}_j+\stackrel{\‾}{\mathrm{\Ψ}}+{\stackrel{\‾}{V}}_j---\left(23\right)$

In formula, for zero-mean and covariance matrix are O _jwhite noise,, and meet

Secondly, simultaneous output equation formula (11) and information representation formula (21) obtain

In formula for zero-mean and covariance matrix are for Q _j ^-1white noise.

About association's state information representation formula comprise formula (23) and (25), association state fuse information maker be

Assist status switch estimator to be

${\widehat{\stackrel{\‾}{X}}}_j=P_j[{\stackrel{\‾}{\mathrm{\Φ}}}^T{O}_j^{-1}({\widehat{\stackrel{\‾}{X}}}_{j+1}-\stackrel{\‾}{\mathrm{\Ψ}})+{\stackrel{\‾}{D}}^T{Q}_j{Y}_j^{*}]---\left(27\right)$

J=k+1 in formula ..., k+N.

Association's status information after following N step and input reference locus information are all unknown, assist State fusion estimation assist state about association's state quantity of information be now about association's state fuse information amount be but, due to positive semidefinite matrix often, therefore, nonsingular for guarantee information amount, get (k+N+1) moment association's state quantity of information wherein λ is enough little positive nonzero number, I _{n × n}for the identity matrix of n × n dimension, now, association's state fusion quantity of information against time series initial value estimator is

$P_{k+N}^{-1}={\stackrel{\‾}{D}}^T{Q}_{k+N}\stackrel{\‾}{D}+{\stackrel{\‾}{\mathrm{\Φ}}}^T{(I_{n\×n}/\mathrm{\λ}+\stackrel{\‾}{\mathrm{\Ψ}}{R}_{k+N}^{-1})}^{-1}\stackrel{\‾}{\mathrm{\Φ}}---\left(28\right)$

Association's state against time series initial value estimator is

${\widehat{\stackrel{\‾}{X}}}_{k+N}=P_{k+N}\{-{\stackrel{\‾}{\mathrm{\Φ}}}^T{(I_{n\×n}/\mathrm{\λ}+\stackrel{\‾}{\mathrm{\Ψ}}{R}_{k+N}^{-1}{\stackrel{\‾}{\mathrm{\Ψ}}}^T)}^{-1}\stackrel{\‾}{\mathrm{\Ψ}}+{\stackrel{\‾}{D}}^T{Q}_{k+N}{Y}_{k+N}^{*}\}---\left(29\right)$

1 adopts generalized regression nerve networks to set up the tandem combination neural net model of hydrodynamic force coefficient.

Before carrying out the training of neural network, first utilize a neural network to find out the erroneous point of observed data, train the functional relation of a neuron network simulation observed data, thereby utilize neural network output valve to substitute erroneous point, data are revised, ensured that the Coefficient Space of model approaches with actual.Because each input variable exists certain difference on the order of magnitude, thereby must do normalized to every one dimension input data.

Taking 442T surface craft as example, the longitudinal hydrodynamic force coefficient data of boats and ships that actual experiment is measured, 3 grades, 4 grades, 5 grades of corresponding typical sea condition grades respectively, typical case's speed of a ship or plane 6 saves, 12 joints, 18 save, 24 joints, typical case's course angle is from 0 °～180 °, get a sub-value, 13 data points altogether, altogether 156 (=3 × 4 × 13) group data every 15 °.

By the observation that after the actual checking computations to conventional method for normalizing and processing, data distribute, finally determine and choose the basic formula of one group of formula below as normalization method and renormalization: $\left\{\begin{array}{c}x=0.1+0.8(q-q_{\min })/(q_{\max }-q_{\min })\\ y=q_{\min }+(q_{\max }-q_{\min })(x-0.1)/0.8\end{array}\right.$

Wherein, q _max, q _minrepresent respectively maxim and minimum value in sample, q represents the actual value of sample, x, and y represents respectively the sample after sample and the renormalization after normalization method.

By 156 groups of data refinement groupings, with hydrodynamic force coefficient a ₃₃for example, shown in Fig. 6, divide the modeling of three grades of composition tandem array configurations, GRNNi_j_k in figure, i represents i level, and j represents Hai Qing, and k represents the speed of a ship or plane.

This network divides three grades to form tandem form by 10 GRNN, wherein the first order is made up of 4 GRNN, and its input variable is 1 neuron (course angle), and output variable is all also 1 (fixing sea condition, the fixing speed of a ship or plane, the corresponding hydrodynamic force coefficient of course angle); The second stage is made up of 4 GRNN, and its input variable is 2 neurons, and output variable is 2 neurons or 1 neuron (input and output are all the corresponding hydrodynamic force coefficients of course angle); The third stage is made up of 2 GRNN, and its input variable is 2 neurons, and output variable is 1 neuron (being all the corresponding hydrodynamic force coefficients of course angle).

Test samples is got respectively every kind of sea condition (3 grades, 4 grades, 5 grades), when every kind of speed of a ship or plane (6 joints, 12 joints, 18 joints, 24 joints) 15 °, 120 °, 165 ° totally 36 groups, other situation (every kind of sea condition, when every kind of speed of a ship or plane 0 °, 30 °, 45 °, 60 °, 75 °, 90 °, 105 °, 135 °, 150 °, 180 °) is totally 120 groups of training samples.

Network is through training, inspection, and finally tandem combination neural net institute established model maximum relative error is 0.95%.

While applying optimal stopping rule to GRNN modeling, we are provided with a fault-tolerant thresholding S, prevent from becoming large when certain or some checking samples (signal) have temporarily caused validation error because of burst error, think that network has entered overfitting by mistake, in this enforcement, fault-tolerant thresholding S gets 3.

The intelligent adaptive predictive control of oblique rudder ship antipitching device

In enforcement, (20) formula state estimation value obtain according to expansion Kalman filtering method.Kalman filtering recurrence equation is shown in formula (30)～formula (35).

$\begin{array}{c}\widehat{\stackrel{\‾}{X}}\left(0\right)=E\left[\stackrel{\‾}{X}\left(0\right)\right]={\mathrm{\μ}}_0\\ P\left(0\right)=\mathrm{Cov}[\stackrel{\‾}{X}\left(0\right)-{\mathrm{\μ}}_0,\stackrel{\‾}{X}\left(0\right)-{\mathrm{\μ}}_0]\end{array}---\left(30\right)$

${\widehat{\stackrel{\‾}{X}}}_{k/k-1}=\stackrel{\‾}{\mathrm{\Φ}}{\widehat{\stackrel{\‾}{X}}}_{k-1}+{\stackrel{\‾}{\mathrm{\Ψ}}u}_{k-1}---\left(31\right)$

$P_{k/k-1}=\stackrel{\‾}{\mathrm{\Φ}}{P}_{k-1}{\stackrel{\‾}{\mathrm{\Φ}}}^T+\stackrel{\‾}{\mathrm{\Γ}}{Q}_{\mathrm{wff}}{\stackrel{\‾}{\mathrm{\Γ}}}^T---\left(32\right)$

$K_k=P_{k/k-1}{\stackrel{\‾}{D}}^T{[\stackrel{\‾}{D}{P}_{k/k-1}{\stackrel{\‾}{D}}^T+Q_w]}^{-1}---\left(33\right)$

${\widehat{\stackrel{\‾}{X}}}_k={\widehat{\stackrel{\‾}{X}}}_{k/k-1}+K_k[Y_k-\stackrel{\‾}{D}{\widehat{\stackrel{\‾}{X}}}_{k/k-1}]---\left(34\right)$

$\begin{array}{c}{P}_k=[I-K_k\stackrel{\‾}{D}]P_{k/k-1}{[I-K_k\stackrel{\‾}{D}]}^T\\ +K_k{Q}_w{K_k}^T\end{array}---\left(35\right)$

Initial value is got measure noise V _kcovariance matrix

Q _vv＝diag[20.3×10 ^-42.25×10 ^-6]

Operational configuration parameter (Hai Qing, the speed of a ship or plane, course angle) information, as the input of tandem combination neural net, is carried out identification oblique rudder ship antipitching device intelligent adaptive model by tandem combination neural net.

Based on the intelligent adaptive integrated model prediction device of oblique rudder ship antipitching device and tandem combination neural net composition, by merging all information about controlling quantity, provide optimal control sequence by controlling quantity estimator.

When working control, the fusion that at every turn only solves current k moment control law is estimated only by controlling quantity against last component of seasonal effect in time series add system, the key step of algorithm is:

Put k=0, system brings into operation;

Initial parameter is set: softening coefficient Ω, prediction step number N, maximum iteration time d, stops moment K;

The given signal of input system weight matrix Q is set _j, R _k, R _j, j=k+1, k+2 ..., k+N, is asked by kalman filtering recurrence equation

Put iterations i=1;

The fusion that solves association's status switch according to the contrary time orientation in formula (26)～(29) is estimated and quantity of information

The fusion that solves k moment controlling quantity according to formula (19) and (20) is estimated

If iterations i >=d, performs step 8); Otherwise i=i+1, returns and asks step 5);

If meet k >=K, system is out of service, otherwise k=k+1 returns to step 3).

In the present invention, get N=3, N _u=2, λ=10 ^-4, softening coefficient Ω=0.2, K=800, R=10 ^-2, Q=diag[3,900].

Table 1 is the statistics (4 grades of speed of a ship or plane 18kn course angle 120o of sea condition) of the control effect relative value of 1000 second time period, and pitching stabilization control effect reaches 74%, and maximum rudder angle is 11o.

The present embodiment shows, because sea condition is complicated and changeable, the impact of the factors such as while wind-engaging, wave and stream, the hydrodynamic force coefficient of oblique rudder ship changes constantly, the present invention adopts 45 ° of flap type rudders that tilt to install, and has set up the hydrodynamic force nonlinear adaptive model of mind (tandem combination neural net model) according to the tank experiments test result of hydrodynamic force coefficient under typical sea situation based on strip theory.According to instrument on ship, can record in real time course angle, the speed of a ship or plane and sea condition in variation, utilize hydrodynamic force self-adapting intelligent model to online calculate the hydrodynamic force coefficient of oblique rudder ship antipitching device, constantly update oblique rudder ship antipitching device model, then neural network adaptive model is combined with predictive control, cause that because procedure parameter becomes institute when slow forecast model output error can revise timely, thereby improve the dynamic property of system.Result shows that this routine control method has good parameter robustness.

Claims (10)

1. one kind based on oblique rudder ship antipitching device intelligent adaptive Predictive Control System, comprising hangs down swings and pitching reference information maker (1), input reference locus maker (2), information constrained system (3), controlling quantity fuse information maker (4), controlling quantity estimator (5), the fuse information maker (6) of association's state, association's status switch estimator (7), association's state fusion quantity of information is against time series initial value estimator (8), association's state is against time series initial value estimator (9), sensing system (10), controller (11), it is characterized in that: input reference locus maker (2) receive hang down swing with the data of pitching reference information maker (1) and sensing system (10) after obtain inputting reference locus, send to association's state against time series initial value estimator (9) and assist status switch estimator (7) parallel input reference locus, controlling quantity fuse information maker (4) receives controlled fuse information amount after the data of information constrained system (3), and the fuse information amount of control is passed to controlling quantity estimator (5), association's state receives association's state fusion quantity of information against time series initial value estimator (9) and obtain association's state against time series initial estimate amount after the data of time series initial value estimator (8), association's state is issued to association's status switch estimator (7) against time series initial estimate amount, assist status switch estimator (7) to receive estimated sequence between the inverse time that obtains assisting state after the data of fuse information maker (6) of association state simultaneously, controlling quantity estimator (5) receives between inverse time of association's state and obtains optimal control amount sequence after estimated sequence data, pass to controller (11) and generate control command, last controller (11) sends to control command the actuating unit steering wheel of boats and ships, realizes the pitching stabilization control of surface ship.

2. one according to claim 1, based on oblique rudder ship antipitching device intelligent adaptive Predictive Control System, is characterized in that: it is that boats and ships hang down to swinging with pitch angle reference information and offer definite value that described oblique rudder ship antipitching device vertical swings with pitching reference information maker; The sensing system of described oblique rudder ship antipitching device refers to, gathers boats and ships course made good, the speed of a ship or plane, rudder angle, hangs down and swing and displacement pickup, the gyroscope of pitch angle, log, boat-carrying theodolite.

3. one according to claim 1, based on oblique rudder ship antipitching device intelligent adaptive Predictive Control System, is characterized in that: the input reference locus maker of described oblique rudder ship antipitching device is:

${Y^{*}}_j={\mathrm{\Ω}}^{j-k}{Y}_k+{(1-\mathrm{\Ω})}^{j-k}{\stackrel{\‾}{R}}_k$

In formula, Y _khang down for the boats and ships of sensing system collection are actual swing, pitch angle, Y _jfor to Y _k(j-k) forward step prediction, swing and pitching reference information given value for hanging down, Ω ∈ (0,1) is softening coefficient, and Ω choose the overall balance needing between taking into account system response characteristic and robustness.

4. one according to claim 1, based on oblique rudder ship antipitching device intelligent adaptive Predictive Control System, is characterized in that: the information constrained system of described oblique rudder ship antipitching device, comprises following information representation formula:

The hard constraint information representation formula being determined by system state equation is:

${\stackrel{\‾}{X}}_j=\stackrel{\‾}{\mathrm{\Φ}}{\stackrel{\‾}{X}}_{j-1}+\stackrel{\‾}{\mathrm{\Ψ}}{u}_{j-1}+\stackrel{\‾}{\mathrm{\Γ}}{W}_{f,j-1}$

In formula for system state equation matrix of coefficient, W _{f, j-1}for zero-mean and covariance matrix are I _{2 × 1}white noise;

The soft-constraint information representation formula being determined by association's state estimation error is:

${\widehat{\stackrel{\‾}{X}}}_j={\stackrel{\‾}{X}}_j+E_j$

E in formula _jfor zero-mean and covariance matrix are white noise;

By requiring to control the as far as possible little soft-constraint information representation formula determining of energy be:

0＝u _j-1+n _j-1

N in formula _j-1for zero-mean and covariance matrix are R _j-1 ^-1white noise;

Soft-constraint information representation formula by requiring system output tracking input reference locus to determine:

$Y_j^{*}=Y_j+{\stackrel{\‾}{M}}_j$

In formula for zero-mean and covariance matrix for white noise.

5. one according to claim 1, based on oblique rudder ship antipitching device intelligent adaptive Predictive Control System, is characterized in that: the controlling quantity fuse information maker of described oblique rudder ship antipitching device is:

In formula covariance matrix.

6. one according to claim 1, based on oblique rudder ship antipitching device intelligent adaptive Predictive Control System, is characterized in that: association's state fusion quantity of information of described oblique rudder ship antipitching device against time series initial value estimator is:

$P_{k+N}^{-1}={\stackrel{\‾}{D}}^T{Q}_{k+N}\stackrel{\‾}{D}+{\stackrel{\‾}{\mathrm{\Φ}}}^T{(I_{n\×n}/\mathrm{\λ}+\stackrel{\‾}{\mathrm{\Ψ}}{R}_{k+N}^{-1}{\stackrel{\‾}{\mathrm{\Ψ}}}^T)}^{-1}\stackrel{\‾}{\mathrm{\Φ}}$

In formula for systematic observation matrix, λ is enough little positive nonzero number, I _{n × n}for the identity matrix of n × n dimension, Q _jfor covariance matrix, V _jfor systematic survey noise covariance matrix.

7. one according to claim 1, based on oblique rudder ship antipitching device intelligent adaptive Predictive Control System, is characterized in that: association's state of described oblique rudder ship antipitching device against time series initial value estimator is:

${\widehat{\stackrel{\‾}{X}}}_{k+N}=P_{k+N}\{-{\stackrel{\‾}{\mathrm{\Φ}}}^T{(I_{n\×n}/\mathrm{\λ}+\stackrel{\‾}{\mathrm{\Ψ}}{R}_{k+N}^{-1}{\stackrel{\‾}{\mathrm{\Φ}}}^T)}^{-1}\stackrel{\‾}{\mathrm{\Ψ}}+{\stackrel{\‾}{D}}^T{Q}_{k+N}{Y}_{k+N}^{*}\}.$

8. one according to claim 1, based on oblique rudder ship antipitching device intelligent adaptive Predictive Control System, is characterized in that: the fuse information maker of association's state of described oblique rudder ship antipitching device is

In formula,

9. one according to claim 1, based on oblique rudder ship antipitching device intelligent adaptive Predictive Control System, is characterized in that: association's status switch estimator of described oblique rudder ship antipitching device is

${\widehat{\stackrel{\‾}{X}}}_j=P_j[{\stackrel{\‾}{\mathrm{\Φ}}}^T{O}_j^{-1}({\widehat{\stackrel{\‾}{X}}}_{j+1}-\stackrel{\‾}{\mathrm{\Ψ}})+{\stackrel{\‾}{D}}^T{Q}_j{Y}_j^{*}]$

J=k+1 in formula ..., k+N.

10. one according to claim 1, based on oblique rudder ship antipitching device intelligent adaptive Predictive Control System, is characterized in that: the controlling quantity estimator of described oblique rudder ship antipitching device is