CN102880188B - Aircraft modeling method based on maximum information-reliability online identification criterion - Google Patents

Abstract

The invention discloses an aircraft modeling method based on a maximum information-reliability online identification criterion. The technical problems that the proving correctness of aerodynamic models and parameters verified by a flight test is poor caused by a simplified model identification criterion are solved. The technical scheme includes that data lengths are considered and reliable parameters are introduced in a maximum information criterion, an online modeling criterion is corrected, the division of U-D of Rj and Rj+1 is evaluated for the measurement variance, and a scalar online model selection and verified discriminant is obtained. The method has the advantages that aerodynamic models and moment models of an aircraft can be established on line directly in accordance with flight test data, and the technical problems of aircraft high attack angle model establishment and verification according to flight tests caused by a simplified model determination criterion are solved.

Description

Modeling of Vehicle method based on maximum fault information-confidence level on-line identification criterion

Technical field

The present invention relates to a kind of Modeling of Vehicle method, particularly relate to a kind of Modeling of Vehicle method based on maximum fault information-confidence level on-line identification criterion.

Background technology

According to aircraft aerodynamic model and parameter, not only can determine the control stability of aircraft, also can be ground and provide correct mathematical model with aerial emulator; Wind tunnel experiment and the calculated results of checking aircraft aerodynamic parameter; For design and the improvement of aircraft control system provides master data; Identify the flight quality of sizing aircraft; The flight quality of research high performance airplane; Carry out crash analysis of aviation accident etc.; Set up exactly aircraft mathematical model problem with completely different by the theoretical method of the modellings by mechanism such as philosophy, theorem, mainly according to the input and output data of experiment gained, set up model, its main theoretical basis of the reform of Chinese economic structure is Nonlinear Systems Identification and nonlinear flight dynamics; When aircraft is done low incidence microvariations flight, flight force and moment mould can be got once item, i.e. Bryan model representation by platform labor series expansion.When Mach number, height one timing, this model is Linear Time Invariant model, and this model, because form simply and is always used till today, becomes the foundation stone of pneumatic mathematical model; Adopt this model, aerocraft system identification has been estimated the systematic parameter of known mathematical model with regard to having become; Modern combat aircraft, tactical missile doing that wartime requirement is more motor-driven, fault speed tailspin even, its angle of attack can be from tens degree, tens degree until spend 100 more, can not adopt linear model; The caused non-permanent downwash flow field in body-shedding vortex, separated whirlpool that Aircraft at High Angle of Attack forms, steady model can not be suitable for again.Research non-permanent, nonlinear aerodynamic model of aircraft under At High Angle of Attack become current aircraft development in the urgent need to problem.Yet, the identification complex of Nonlinear Aerodynamic, it is general Nonlinear Systems Identification problem, the funtcional relationship between input quantity and state is difficult to determine, need to carries out identification to model; The key of Model Distinguish is modeling criterion and optimization algorithm, and for given version, application modeling criterion is determined the optimum exponent number of model and select optimization model from candidate model; Because measured data contains noise, modeling criterion can not only be investigated the error of fitting size to available data, and considers other factors, otherwise will make model incorrect; Conventionally, modeling criterion should be able to make the model optimizing have following characteristics: 1. the model existing flying quality of matching well; 2. model is every an obvious physical significance; 3. model can be predicted the measured data under conditions of similarity; 4. under the suitable condition of performance, order is minimum; The most frequently used identification Method is method of gradual regression, and its principle is item by item the predictor that affects conspicuousness to be selected into, and the little factor of impact is rejected, and sets up the method for regression equation; This method is calculated simple, practical; But this method has two obvious shortcomings: the one, and choice criteria is determined by people, and does not provide the confidence level of result; The 2nd, the accumulation of error is large, easily leaks choosing and falsely drops; For this reason, people usually adopt maximum fault information criterion AIC method to the dummy vehicle identification problem of having relatively high expectations, but the method processing speed is slow, and signal to noise ratio (S/N ratio) is hour Model Distinguish low precision; Due under nonlinear situation; can only carry out numerical integration to aircraft nonlinear equation; carry out sensitivity matrix calculating and iterative computation; thereby make complexity and the calculated amount of calculating more much higher than Linear Estimation; also make the matching between model output and experimental data become more difficult simultaneously; for online Approach For Identification of Model Structure problem, existing AIC criterion does not directly provide result, and aerodynamic model and Verification that other short-cut method usually can cause flight test to provide are incorrect.

Summary of the invention

The aerodynamic model and the poor deficiency of Verification correctness that in order to overcome existing simplified model identification criterion, cause flight test to provide, the invention provides a kind of Modeling of Vehicle method based on maximum fault information-confidence level on-line identification criterion.The method is by analyzing the impact of data length and introducing confidence level parameter, maximum fault information criterion is revised, obtained new model on-line identification criterion, by new criterion, set up index modeling, directly can, for flight test line modeling and the modelling verification of aircraft, can avoid the technical matters according to flight test is set up and checking aircraft at high angle of attack model exists.

The technical solution adopted for the present invention to solve the technical problems is: a kind of Modeling of Vehicle method based on maximum fault information-confidence level on-line identification criterion, is characterized in comprising the following steps:

The state equation of the aircraft candidate family that step 1, flight test are to be determined is

$\stackrel{.}{x}\left(t\right)=f\left\{f_0\right[x\left(t\right),{\mathrm{\Ω}}_0],f_1[x\left(t\right),{\mathrm{\θ}}_1],...,f_q[{\mathrm{\θ}}_q,x\left(t\right)],t\}---\left(1\right)$

Observation equation is

$\left\{\begin{array}{c}y\left(t\right)=g[x\left(t\right),\mathrm{\Ω},t]=g\left\{g_0\right[x\left(t\right),{\mathrm{\Ω}}_0],g_1[x\left(t\right),{\mathrm{\θ}}_1],...,g_q[{\mathrm{\θ}}_q,x\left(t\right)],t\}\\ z\left(t_k\right)=y\left(t_k\right)+v\left(k\right)\end{array}\right.---\left(2\right)$

(1), in (2) formula, x (t) is n dimension state vector; Y (t) is m dimension observation vector; f{f ₀[x (t), Ω ₀], f ₁[x (t), θ ₁] ..., f _q[θ _q, x (t)], t}, g{g ₀[x (t), Ω ₀], g ₁[x (t), θ ₁] ..., g _q[θ _q, x (t)], t} is the known model structure function to be determined of expression formula, f ₀[x (t), Ω ₀], g ₀[x (t), Ω ₀] be the model that must be selected into according to physical concept, f _i[x (t), θ _i], g _i[x (t), θ _i] (i=1,2 ..., q) be candidate family, z (t _k) be at t _kconstantly to y (t _k) measured value; Ω is the parameter vector of unknown dimension, Ω ₀parameter vector for known dimension; V (k), for measuring noise, supposes that variance is R _kzero-mean white Gaussian noise; f _i[x (t), θ _i], g _i[x (t), θ _i] (i=1,2 ..., q) whether in model, occur and Ω ₀, θ _i(i=1,2 ..., value q) needs identification, and q is known candidate family number;

Conventionally higher to the model structure accuracy requirement of aircraft, the present invention provides following maximum fault information-confidence level criterion:

$\left|\frac{\ln L_2[x\left(t\right),{\mathrm{\Ω}}_2,t]-\ln L_1[x\left(t\right),{\mathrm{\Ω}}_1,t]}{\ln L_1[x\left(t\right),{\mathrm{\Ω}}_1,t]}\right|>\mathrm{\δ},---\left(3\right)$

In formula, L ₁[x (t), Ω ₁, t], L ₂[x (t), Ω ₂, t] and for getting different parameter vector Ω ₁and Ω ₂maximum likelihood function, δ is given arithmetic number, represent confidence level parameter,

$\left\{\begin{array}{c}\ln L_1[x\left(t\right),{\mathrm{\Ω}}_1,t]=-\frac{\mathrm{mN}}{2}\ln \left(\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{\stackrel{\‾}{\mathrm{\η}}}_1^T\left(k\right)R_1^{-1}\left(k\right){\stackrel{\‾}{\mathrm{\η}}}_1\left(k\right)\right)-\frac{\mathrm{mN}}{2}[\ln \left(\frac{2\mathrm{\π}}{\mathrm{mN}}\right)+1]-\frac{1}{2}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}\ln \left|R_1\left(k\right)\right|\\ \ln L_2[x\left(t\right),{\mathrm{\Ω}}_2,t]=-\frac{\mathrm{mN}}{2}\ln \left(\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{\stackrel{\‾}{\mathrm{\η}}}_2^T\left(k\right)R_2^{-1}\left(k\right){\stackrel{\‾}{\mathrm{\η}}}_2\left(k\right)\right)-\frac{\mathrm{mN}}{2}[\ln \left(\frac{2\mathrm{\π}}{\mathrm{mN}}\right)+1]-\frac{1}{2}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}\ln \left|R_2\left(k\right)\right|\end{array}\right.---\left(4\right)$

$,{\stackrel{\‾}{\mathrm{\η}}}_1\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\Ω}}_1,t_k],$ ${\stackrel{\‾}{\mathrm{\η}}}_2\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\Ω}}_2,t_k],$ $R_1=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{\stackrel{\‾}{\mathrm{\η}}}_1\left(k\right){\stackrel{\‾}{\mathrm{\η}}}_1^T\left(k\right),$ Ω ₁and Ω ₂for different parameter vectors, N is current data length, and ln is natural logarithm symbol;

Step 2, according to supposition f ₀[x (t), Ω ₀], g ₀[x (t), Ω ₀], Ω ₀=Ω ₀by optimization algorithm, be selected into model, and calculated by following algorithm iteration:

Make j=0,1,2 ..., q, supposes f _j[x (t), θ _j], g _j[x (t), θ _j], Ω _jbe selected into model, selected in such a way other candidate family:

Ask (4) formula maximum value, iterative computation:

$\mathrm{\Δ}{\mathrm{\Ω}}_j=A_j^{-1}{b}_j---\left(5\right)$

And

$R_j=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{\stackrel{\‾}{\mathrm{\η}}}_j\left(k\right){\stackrel{\‾}{\mathrm{\η}}}_j^T\left(k\right),$ ${\stackrel{\‾}{\mathrm{\η}}}_j\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\Ω}}_j,t_k]---\left(6\right)$

(5), in (6) formula: $\mathrm{\Δ}{\mathrm{\Ω}}_j={\mathrm{\Ω}}_j-{\widehat{\mathrm{\Ω}}}_j,$ $b_j=\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{\left(\frac{\∂y}{\∂{\mathrm{\Ω}}_j^T}\right)}^T{R}_j^{-1}[z\left(t_k\right)-y\left(t_k\right)],$

$A_j=\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{\left(\frac{\∂y}{\∂{\mathrm{\Ω}}_j^T}\right)}^T{R}_j^{-1}\frac{\∂y}{\∂{\mathrm{\Ω}}_j^T}=B_j^T{P}_j^{-1}{B}_j,$ $B_j^T=[{\left(\frac{\∂y\left(t_1\right)}{\∂{\mathrm{\Ω}}_j^T}\right)}^T,{\left(\frac{\∂y\left(t_2\right)}{\∂{\mathrm{\Ω}}_j^T}\right)}^T,\·\·\·,{\left(\frac{\∂y\left(t_N\right)}{\∂{\mathrm{\Ω}}_j^T}\right)}^T]$

$P_j^{-1}=\mathrm{diag}[R_j^{-1},R_j^{-1},\·\·\·R_j^{-1}],$

If ${\mathrm{\Ω}}_{j+1}=\left[\begin{array}{c}{\mathrm{\Ω}}_j\\ {\mathrm{\θ}}_{j+1}\end{array}\right],$ And N+1 secondary data is come, θ _j+1be selected into or rejecting and verification condition are: when

$\frac{\ln \left|R_{\mathrm{Nj}}\left(k\right)\right|-\ln \left|R_{N(j+1)}\left(k\right)\right|}{|m(\ln 2\mathrm{\&pi;}+1)+\ln |R_{\mathrm{Nj}}\left(k\right)\left|\right|}>\mathrm{\&delta;}$ And $|-\frac{m}{2}[1+\ln \left(2\mathrm{\&pi;}\right)]+\ln \frac{{\left|R_N\left(k\right)\right|}^N}{{\left|R_{N+1}\left(k\right)\right|}^{N+1}}|<\mathrm{\&epsiv;}---\left(7\right)$

Time, θ _j+1, f _j+1[x (t), θ _j+1], g _j+1[x (t), θ _j+1] be selected into model, and ${\mathrm{\&Omega;}}_{j+1}=\left[\begin{array}{c}{\mathrm{\&Omega;}}_j\\ {\mathrm{\&theta;}}_{j+1}\end{array}\right];$ Otherwise reject f _j+1[x (t), θ _j+1], g _j+1[x (t), θ _j+1] candidate item, and Ω _j+1=Ω _j;

(7) in formula: ε is given positive number, $R_{\mathrm{Nj}}=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j^T\left(k\right),$ $R_{N(j+1)}=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}^T\left(k\right),$

$R_{(N+1)j}=\frac{1}{N+1}\underset{k=1}{\overset{N+1}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j^T\left(k\right),$ $R_{(N+1)(j+1)}=\frac{1}{N+1}\underset{k=1}{\overset{N+1}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}^T\left(k\right),$

${\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\&Omega;}}_j,t_k],$ ${\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\&Omega;}}_{j+1},t_k];$

Step 3, because the dimension m of aircraft measurement vector y is larger, adopt Gram-Schmidt orthogonalization method to R _nj, R _{(N+1) j}, R _{n (j+1)}and R _{(N+1) (j+1)}carry out U-D decomposition, R _nj, R _{(N+1) j}, R _{n (j+1)}and R _{(N+1) (j+1)}u-D be respectively: $R_{\mathrm{Nj}}=U_{\mathrm{RNj}}{D}_{\mathrm{RNj}}{U}_{\mathrm{RNj}}^T,$ $R_{(N+1)j}=U_{R(N+1)j}{D}_{R(N+1)j}{U}_{R(N+1)j}^T,$ $R_{N(j+1)}=U_{\mathrm{RN}(j+1)}{D}_{\mathrm{RN}(j+1)}{U}_{\mathrm{RN}(j+1)}^T,$ $R_{(N+1)(j+1)}=U_{R(N+1)(j+1)}{D}_{R(N+1)(j+1)}{U}_{R(N+1)(j+1)}^T,$ In formula, U _rNj, U _{r (N+1) j}, U _{rN (j+1)}, U _{r (N+1) (j+1)}for unit upper triangular matrix; D _rNj=diag[d _rNj(1), d _rNj(2) ..., d _rNj(m)], D _{rN (j+1)}=diag[d _{rN (j+1)}(1), d _{rN (j+1)}(2) ..., d _{rN (j+1)}(m)], D _{r (N+1) j}=diag[d _{r (N+1) j}(1), d _{r (N+1) j}(2) ..., d _{r (N+1) j}(m)], D _{r (N+1) (j+1)}=diag[d _{r (N+1) (j+1)}(1), d _{r (N+1) (j+1)}(2) ..., d _{r (N+1) (j+1)}(m)]; Diag is diagonal angle symbol;

The maximum fault information criterion of modelling verification is write as: when

$\left\{\begin{array}{c}\frac{\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}[\ln d_{\mathrm{RNj}}\left(i\right)-\ln d_{\mathrm{RN}(j+1)}\left(i\right)]}{|m(\ln 2\mathrm{\&pi;}+1)+\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\ln d_{\mathrm{RNj}}\left(i\right)|}>\mathrm{\&delta;}\\ |-\frac{m}{2}[1+\ln \left(2\mathrm{\&pi;}\right)]+N\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\ln d_{\mathrm{RNj}}\left(i\right)-(N+1)\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\ln d_{R(N+1)j}\left(i\right)|<\mathrm{\&epsiv;}\end{array}\right.---\left(8\right)$

During establishment, θ _i+1, f _j+1[x (t), θ _j+1], g _j+1[x (t), θ _j+1] to be selected into model correct, and ${\mathrm{\&Omega;}}_{j+1}=\left[\begin{array}{c}{\mathrm{\&Omega;}}_j\\ {\mathrm{\&theta;}}_{j+1}\end{array}\right];$ Otherwise reject f _j+1[x (t), θ _j+1], g _j+1[x (t), θ _j+1] candidate item, and Ω _j+1=Ω _j.

The invention has the beneficial effects as follows: due to by considering data length in maximum fault information criterion and introducing confidence level parameter, revised line modeling criterion; To measuring variance, estimate R _jand R _j+1u-D decompose, the online Model Selection of scalar and checking discriminant have been obtained, be convenient to directly according to test flight data, set up online aircraft aerodynamic force, moment model, avoid simplified model criterion to cause, avoided the technical matters according to flight test is set up and checking aircraft at high angle of attack model exists.

Below in conjunction with embodiment, the present invention is elaborated.

Embodiment

The Modeling of Vehicle method concrete steps that the present invention is based on maximum fault information-confidence level on-line identification criterion are as follows:

1, many aircraft are less than 60 conventional candidate family forms while spending at the angle of attack and are:

$\stackrel{.}{x}\left(t\right)=\mathrm{\&Phi;}\left({\mathrm{\&Omega;}}_0\right)f_0\left[x\left(t\right)\right]+{\mathrm{\&theta;}}_1{f}_1\left[x\left(t\right)\right]+...+{\mathrm{\&theta;}}_q{f}_q\left[x\left(t\right)\right]---\left(1\right)$

$\left\{\begin{array}{c}y\left(t\right)=g[x\left(t\right),\mathrm{\&Omega;}]=\mathrm{\&Psi;}\left({\mathrm{\&Omega;}}_0\right)g_0\left[x\left(t\right)\right]+{\mathrm{\&theta;}}_1{g}_1\left[x\left(t\right)\right]+...+g_q[{\mathrm{\&theta;}}_q,x\left(t\right)]\\ z\left(t_k\right)=y\left(t_k\right)+v\left(k\right)\end{array}\right.---\left(2\right)$

(1), in (2) formula,, x (t) is n dimension state vector; Y (t) is m dimension observation vector, Φ (Ω ₀) f ₀[x (t)], Ψ (Ω ₀) g ₀[x (t)] model for being selected into according to physical concept, θ _if _i[x (t)], θ _ig _i[x (t)] (i=1,2 ..., q) be candidate family, z (t _k) be at t _kconstantly to y (t _k) measured value; Ω is the parameter vector of unknown dimension, Ω ₀parameter vector for known dimension; V (k), for measuring noise, supposes that variance is R _kzero-mean white Gaussian noise; θ _if _i[x (t)], θ _ig _i[x (t)] (i=1,2 ..., q) whether in model, occur and Ω ₀, θ _i(i=1,2 ..., value q) needs identification., q is known candidate family number;

Conventionally higher to the model structure accuracy requirement of aircraft, the present invention provides following maximum fault information-confidence level criterion:

$\left|\frac{\ln L_2[x\left(t\right),{\mathrm{\&Omega;}}_2,t]-\ln L_1[x\left(t\right),{\mathrm{\&Omega;}}_1,t]}{\ln L_1[x\left(t\right),{\mathrm{\&Omega;}}_1,t]}\right|>\mathrm{\&delta;},---\left(3\right)$

In formula, L ₁[x (t), Ω ₁, t], L ₂[x (t), Ω ₂, t] and for getting different parameter vector Ω ₁and Ω ₂maximum likelihood function, δ is given arithmetic number, represent confidence level parameter,

$\left\{\begin{array}{c}\ln L_1[x\left(t\right),{\mathrm{\&Omega;}}_1,t]=-\frac{\mathrm{mN}}{2}\ln \left(\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_1^T\left(k\right)R_1^{-1}\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_1\left(k\right)\right)-\frac{\mathrm{mN}}{2}[\ln \left(\frac{2\mathrm{\&pi;}}{\mathrm{mN}}\right)+1]-\frac{1}{2}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}\ln \left|R_1\left(k\right)\right|\\ \ln L_2[x\left(t\right),{\mathrm{\&Omega;}}_2,t]=-\frac{\mathrm{mN}}{2}\ln \left(\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_2^T\left(k\right)R_2^{-1}\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_2\left(k\right)\right)-\frac{\mathrm{mN}}{2}[\ln \left(\frac{2\mathrm{\&pi;}}{\mathrm{mN}}\right)+1]-\frac{1}{2}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}\ln \left|R_2\left(k\right)\right|\end{array}\right.---\left(4\right)$

$,{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_1\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\&Omega;}}_1,t_k],$ ${\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_2\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\&Omega;}}_2,t_k],$ $R_1=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_1\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_1^T\left(k\right),$ Ω ₁and Ω ₂for different parameter vectors, N is current data length, and ln is natural logarithm symbol;

2, according to supposition f ₀[x (t), Ω ₀], g ₀[x (t), Ω ₀], Ω ₀=Ω ₀by optimization algorithm, be selected into model, and calculated by following algorithm iteration:

Make j=0,1,2 ..., q, supposes f _j[x (t), θ _j], g _j[x (t), θ _j], Ω _jbe selected into model, selected in such a way other candidate family:

Ask (4) formula maximum value, iterative computation:

$\mathrm{\&Delta;}{\mathrm{\&Omega;}}_j=A_j^{-1}{b}_j---\left(5\right)$

And

$R_j=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j^T\left(k\right),$ ${\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\&Omega;}}_j,t_k]---\left(6\right)$

(5), in (6) formula: $\mathrm{\&Delta;}{\mathrm{\&Omega;}}_j={\mathrm{\&Omega;}}_j-{\widehat{\mathrm{\&Omega;}}}_j,$ $b_j=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(\frac{\&PartialD;y}{\&PartialD;{\mathrm{\&Omega;}}_j^T}\right)}^T{R}_j^{-1}[z\left(t_k\right)-y\left(t_k\right)],$

$A_j=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(\frac{\&PartialD;y}{\&PartialD;{\mathrm{\&Omega;}}_j^T}\right)}^T{R}_j^{-1}\frac{\&PartialD;y}{\&PartialD;{\mathrm{\&Omega;}}_j^T}=B_j^T{P}_j^{-1}{B}_j,$ $B_j^T=[{\left(\frac{\&PartialD;y\left(t_1\right)}{\&PartialD;{\mathrm{\&Omega;}}_j^T}\right)}^T,{\left(\frac{\&PartialD;y\left(t_2\right)}{\&PartialD;{\mathrm{\&Omega;}}_j^T}\right)}^T,\&CenterDot;\&CenterDot;\&CenterDot;,{\left(\frac{\&PartialD;y\left(t_N\right)}{\&PartialD;{\mathrm{\&Omega;}}_j^T}\right)}^T]$

$P_j^{-1}=\mathrm{diag}[R_j^{-1},R_j^{-1},\&CenterDot;\&CenterDot;\&CenterDot;R_j^{-1}],$

If ${\mathrm{\&Omega;}}_{j+1}=\left[\begin{array}{c}{\mathrm{\&Omega;}}_j\\ {\mathrm{\&theta;}}_{j+1}\end{array}\right],$ And N+1 secondary data is come, θ _j+1be selected into or rejecting and verification condition are: when

$\frac{\ln \left|R_{\mathrm{Nj}}\left(k\right)\right|-\ln \left|R_{N(j+1)}\left(k\right)\right|}{|m(\ln 2\mathrm{\&pi;}+1)+\ln |R_{\mathrm{Nj}}\left(k\right)\left|\right|}>\mathrm{\&delta;}$ And $|-\frac{m}{2}[1+\ln \left(2\mathrm{\&pi;}\right)]+\ln \frac{{\left|R_N\left(k\right)\right|}^N}{{\left|R_{N+1}\left(k\right)\right|}^{N+1}}|<\mathrm{\&epsiv;}---\left(7\right)$

Time, θ _i+1, θ _jf _j[x (t)], θ _jg _j[x (t)] is selected into model, and ${\mathrm{\&Omega;}}_{j+1}=\left[\begin{array}{c}{\mathrm{\&Omega;}}_j\\ {\mathrm{\&theta;}}_{j+1}\end{array}\right];$ Otherwise reject θ _jf _j[x (t)], θ _jg _j[x (t)] candidate item, and Ω _j+1=Ω _j;

(7) in formula: ε is given positive number, $R_{\mathrm{Nj}}=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j^T\left(k\right),$ $R_{N(j+1)}=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}^T\left(k\right),$

$R_{(N+1)j}=\frac{1}{N+1}\underset{k=1}{\overset{N+1}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j^T\left(k\right),$ $R_{(N+1)(j+1)}=\frac{1}{N+1}\underset{k=1}{\overset{N+1}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}^T\left(k\right),$

${\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\&Omega;}}_j,t_k],$ ${\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\&Omega;}}_{j+1},t_k];$

3, the dimension m of aircraft measurement vector y is larger conventionally, adopts Gram-Schmidt orthogonalization method to R _nj, R _{(N+1) j}, R _{n (j+1)}and R _{(N+1) (j+1)}carry out U-D decomposition, R _nj, R _{(N+1) j}, R _{n (j+1)}and R _{(N+1) (j+1)}u-D be respectively: $R_{\mathrm{Nj}}=U_{\mathrm{RNj}}{D}_{\mathrm{RNj}}{U}_{\mathrm{RNj}}^T,$ $R_{(N+1)j}=U_{R(N+1)j}{D}_{R(N+1)j}{U}_{R(N+1)j}^T,$ $R_{N(j+1)}=U_{\mathrm{RN}(j+1)}{D}_{\mathrm{RN}(j+1)}{U}_{\mathrm{RN}(j+1)}^T,$ $R_{(N+1)(j+1)}=U_{R(N+1)(j+1)}{D}_{R(N+1)(j+1)}{U}_{R(N+1)(j+1)}^T,$ In formula, U _rNj, U _{r (N+1) j}, U _{rN (j+1)}, U _{r (N+1) (j+1)}for unit upper triangular matrix; D _rNj=diag[d _rNj(1), d _rNj(2) ..., d _rNj(m)], D _{rN (j+1)}=diag[d _{rN (j+1)}(1), d _{rN (j+1)}(2) ..., d _{rN (j+1)}(m)], D _{r (N+1) j}=diag[d _{r (N+1) j}(1), d _{r (N+1) j}(2) ..., d _{r (N+1) j}(m)], D _{r (N+1) (j+1)}=diag[d _{r (N+1) (j+1)}(1), d _{r (N+1) (j+1)}(2) ..., d _{r (N+1) (j+1)}(m)]; Diag is diagonal angle symbol;

The maximum fault information criterion of modelling verification can be write as: when

$\left\{\begin{array}{c}\frac{\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}[\ln d_{\mathrm{RNj}}\left(i\right)-\ln d_{\mathrm{RN}(j+1)}\left(i\right)]}{|m(\ln 2\mathrm{\&pi;}+1)+\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\ln d_{\mathrm{RNj}}\left(i\right)|}>\mathrm{\&delta;}\\ |-\frac{m}{2}[1+\ln \left(2\mathrm{\&pi;}\right)]+N\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\ln d_{\mathrm{RNj}}\left(i\right)-(N+1)\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\ln d_{R(N+1)j}\left(i\right)|<\mathrm{\&epsiv;}\end{array}\right.---\left(8\right)$

During establishment, θ _j+1, θ _jf _j[x (t)], θ _jg _jit is correct that [x (t)] is selected into model, and ${\mathrm{\&Omega;}}_{j+1}=\left[\begin{array}{c}{\mathrm{\&Omega;}}_j\\ {\mathrm{\&theta;}}_{j+1}\end{array}\right];$ Otherwise reject θ _jf _j[x (t)], θ _jg _j[x (t)] candidate item, and Ω _j+1=Ω _j.

Claims (1)

1. the Modeling of Vehicle method based on maximum fault information-confidence level on-line identification criterion, is characterized in that comprising the following steps:

The state equation of the aircraft candidate family that step 1, flight test are to be determined is

$\stackrel{\&CenterDot;}{x}\left(t\right)=f\left\{f_0\right[x\left(t\right),{\mathrm{\&Omega;}}_0],f_1[x\left(t\right),{\mathrm{\&theta;}}_1],...,f_q[{\mathrm{\&theta;}}_q,x\left(t\right)],t\}---\left(1\right)$

Observation equation is

$\left\{\begin{array}{c}y\left(t\right)=g[x\left(t\right),\mathrm{\&Omega;},t]=g\left\{g_0\right[x\left(t\right),{\mathrm{\&Omega;}}_0],g_1[x\left(t\right),{\mathrm{\&theta;}}_1],...,g_q[{\mathrm{\&theta;}}_q,x\left(t\right)],t\}\\ z\left(t_k\right)=y\left(t_k\right)+v\left(k\right)\end{array}\right.---\left(2\right)$

(1), in (2) formula, x (t) is n dimension state vector; Y (t) is m dimension observation vector; f{f ₀[x (t), Ω ₀], f ₁[x (t), θ ₁] ..., f _q[θ _q, x (t)], t}, g{g ₀[x (t), Ω ₀], g ₁[x (t), θ ₁] ..., g _q[θ _q, x (t)], t} is the known model structure function to be determined of expression formula, f ₀[x (t), Ω ₀], g ₀[x (t), Ω ₀] be the model that must be selected into according to physical concept, f _i[x (t), θ _i], g _i[x (t), θ _i], i=1,2 ..., q is candidate family, z (t _k) be at t _kconstantly to y (t _k) measured value; Ω is the parameter vector of unknown dimension, Ω ₀parameter vector for known dimension; V (k), for measuring noise, supposes that variance is R _kzero-mean white Gaussian noise; f _i[x (t), θ _i], g _i[x (t), θ _i], i=1,2 ..., whether q occurs and Ω in model ₀, θ _i, i=1,2 ..., the value of q needs identification, and q is known candidate family number;

Conventionally higher to the model structure accuracy requirement of aircraft, provide following maximum fault information-confidence level criterion:

$\left|\frac{\ln L_2[x\left(t\right),{\mathrm{\&Omega;}}_2,t]-\ln L_1[x\left(t\right),{\mathrm{\&Omega;}}_1,t]}{\ln L_1[x\left(t\right),{\mathrm{\&Omega;}}_1,t]}\right|>\mathrm{\&delta;},---\left(3\right)$

In formula, L ₁[x (t), Ω ₁, t], L ₂[x (t), Ω ₂, t] and for getting different parameter vector Ω ₁and Ω ₂maximum likelihood function, δ is given arithmetic number, represent confidence level parameter,

$\left\{\begin{array}{c}\ln L_1[x\left(t\right),{\mathrm{\&Omega;}}_1,t]=-\frac{\mathrm{mN}}{2}\ln \left(\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_1^T\left(k\right)R_1^{-1}\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_1\left(k\right)\right)-\frac{\mathrm{mN}}{2}[\ln \left(\frac{2\mathrm{\&pi;}}{\mathrm{mN}}\right)+1]-\frac{1}{2}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}\ln \left|R_1\left(k\right)\right|\\ \ln L_2[x\left(t\right),{\mathrm{\&Omega;}}_2,t]=-\frac{\mathrm{mN}}{2}\ln \left(\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_2^T\left(k\right)R_2^{-1}\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_2\left(k\right)\right)-\frac{\mathrm{mN}}{2}[\ln \left(\frac{2\mathrm{\&pi;}}{\mathrm{mN}}\right)+1]-\frac{1}{2}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}\ln \left|R_2\left(k\right)\right|\end{array}\right.---\left(4\right)$ ， ${\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_1\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\&Omega;}}_1,t_k],{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_2\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\&Omega;}}_2,t_k],R_1=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_1\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_1^T\left(k\right),$ Ω ₁and Ω ₂for different parameter vectors, N is current data length, and ln is natural logarithm symbol;

Step 2, according to supposition f ₀[x (t), Ω ₀], g ₀[x (t), Ω ₀], Ω ₀=Ω ₀by optimization algorithm, be selected into model, and calculated by following algorithm iteration:

Make j=0,1,2 ..., q, supposes f _j[x (t), θ _j], g _j[x (t), θ _j], Ω _jbe selected into model, selected in such a way other candidate family:

Ask (4) formula maximum value, iterative computation:

$\mathrm{\&Delta;}{\mathrm{\&Omega;}}_j=A_j^{-1}{b}_j---\left(5\right)$

And

$R_j=\frac{1}{N}\underset{k=1}{\mathrm{\&Sigma;}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j^T\left(k\right),{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\&Omega;}}_j,t_k]---\left(6\right)$

(5), in (6) formula: $\mathrm{\&Delta;}{\mathrm{\&Omega;}}_j={\mathrm{\&Omega;}}_j-{\widehat{\mathrm{\&Omega;}}}_j,b_j=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(\frac{\&PartialD;y}{{\&PartialD;\mathrm{\&Omega;}}_j^T}\right)}^T{R}_j^{-1}[z\left(t_k\right)-y\left(t_k\right)],$

$A_j=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(\frac{\&PartialD;y}{\&PartialD;{\mathrm{\&Omega;}}_j^T}\right)}^T{R}_j^{-1}\frac{\&PartialD;y}{\&PartialD;{\mathrm{\&Omega;}}_j^T}=B_j^T{P}_j^{-1}{B}_j,B_j^T=[{\left(\frac{\&PartialD;y\left(t_1\right)}{\&PartialD;{\mathrm{\&Omega;}}_j^T}\right)}^T,{\left(\frac{\&PartialD;y\left(t_2\right)}{\&PartialD;{\mathrm{\&Omega;}}_j^T}\right)}^T,\&CenterDot;\&CenterDot;\&CenterDot;,{\left(\frac{\&PartialD;y\left(t_N\right)}{\&PartialD;{\mathrm{\&Omega;}}_j^T}\right)}^T]$

$P_j^{-1}=\mathrm{diag}[R_j^{-1},R_j^{-1},\&CenterDot;\&CenterDot;\&CenterDot;R_j^{-1}],$

If ${\mathrm{\&Omega;}}_{j+1}=\left[\begin{array}{c}{\mathrm{\&Omega;}}_j\\ {\mathrm{\&theta;}}_{j+1}\end{array}\right],$ And N+1 secondary data is come, θ _j+1be selected into or rejecting and verification condition are: when

$\frac{\ln \left|R_{\mathrm{Nj}}\left(k\right)\right|-\ln \left|R_{N(j+1)}\left(k\right)\right|}{|m(\ln 2\mathrm{\&pi;}+1)+\ln |R_{\mathrm{Nj}}\left(k\right)\left|\right|}>\mathrm{\&delta;}$ And $|-\frac{m}{2}[1+\ln \left(2\mathrm{\&pi;}\right)]+\ln \frac{{\left|R_N\left(k\right)\right|}^N}{{\left|R_{N+1}\left(k\right)\right|}^{N+1}}|<\mathrm{\&epsiv;}---\left(7\right)$

Time, θ _j+1, f _j+1[x (t), θ _j+1], g _j+1[x (t), θ _j+1] be selected into model, and ${\mathrm{\&Omega;}}_{j+1}=\left[\begin{array}{c}{\mathrm{\&Omega;}}_j\\ {\mathrm{\&theta;}}_{j+1}\end{array}\right];$ Otherwise reject f _j+1[x (t), θ _j+1], g _j+1[x (t), θ _j+1] candidate item, and Ω _j+1=Ω _j;

(7) in formula: ε is given positive number, $R_{\mathrm{Nj}}=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j^T\left(k\right),R_{N(j+1)}=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}^T\left(k\right),$

$R_{(N+1)j}=\frac{1}{N+1}\underset{k=1}{\overset{N+1}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j^T\left(k\right),R_{(N+1)(j+1)}=\frac{1}{N+1}\underset{k=1}{\overset{N+1}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}\left(k\right){\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}^T\left(k\right),$

${\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_j\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\&Omega;}}_j,t_k],{\stackrel{\&OverBar;}{\mathrm{\&eta;}}}_{j+1}\left(k\right)=z\left(t_k\right)-g[x\left(t_k\right),{\mathrm{\&Omega;}}_{j+1},t_k];$

Step 3, because the dimension m of aircraft measurement vector y is larger, adopt Gram-Schmidt orthogonalization method to R _nj, R _{(N+1) j}, R _{n (j+1)}and R _{(N+1) (j+1)}carry out U-D decomposition, R _nj, R _{(N+1) j}, R _{n (j+1)}and R _{(N+1) (j+1)}u-D be respectively: $R_{\mathrm{Nj}}=U_{\mathrm{RNj}}{D}_{\mathrm{RNj}}{U}_{\mathrm{RNj}}^T,R_{(N+1)j}=U_{R(N+1)j}{D}_{R(N+1)j}{U}_{R(N+1)j}^T,$ $R_{N(j+1)}=U_{\mathrm{RN}(j+1)}{D}_{\mathrm{RN}(j+1)}{U}_{\mathrm{RN}(j+1)}^T,R_{(N+1)(j+1)}=U_{R(N+1)(j+1)}{D}_{R(N+1)(j+1)}{U}_{R(N+1)(j+1)}^T,$

In formula, U _rNj, U _{r (N+1) j}, U _{rN (j+1)}, U _{r (N+1) (j+1)}for unit upper triangular matrix;

D _RNj＝diag[d _RNj(1),d _RNj(2),…,d _RNj(m)]，D _RN(j+1)＝diag[d _RN(j+1)(1),d _RN(j+1)(2),…,d _RN(j+1)(m)]，

D _R(N+1)j＝diag[d _R(N+1)j(1),d _R(N+1)j(2),…,d _R(N+1)j(m)]，

D _{r (N+1) (j+1)}=diag[d _{r (N+1) (j+1)}(1), d _{r (N+1) (j+1)}(2) ..., d _{r (N+1) (j+1)}(m)]; Diag is diagonal angle symbol;

The maximum fault information criterion of modelling verification is write as: when

$\left\{\begin{array}{c}\frac{\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}[\ln d_{\mathrm{RNj}}\left(i\right)-\ln d_{\mathrm{RN}(j+1)}\left(i\right)]}{|m(\ln 2\mathrm{\&pi;}+1)+\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\ln d_{\mathrm{RNj}}\left(i\right)|}>\mathrm{\&delta;}\\ |-\frac{m}{2}[1+\ln \left(2\mathrm{\&pi;}\right)]+N\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\ln d_{\mathrm{RNj}}\left(i\right)-(N+1)\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\ln d_{R(N+1)j}\left(i\right)|<\mathrm{\&epsiv;}\end{array}\right.---\left(8\right)$

During establishment, θ _j+1, f _j+1[x (t), θ _j+1], g _j+1[x (t), θ _j+1] to be selected into model correct, and ${\mathrm{\&Omega;}}_{j+1}=\left[\begin{array}{c}{\mathrm{\&Omega;}}_j\\ {\mathrm{\&theta;}}_{j+1}\end{array}\right];$ Otherwise reject f _j+1[x (t), θ _j+1], g _j+1[x (t), θ _j+1] candidate item, and Ω _j+1=Ω _j.