CN104730509A - Radar detection method based on knowledge auxiliary permutation detection - Google Patents

Abstract

The invention discloses a radar detection method based on knowledge auxiliary permutation detection. The radar detection method comprises the following steps that (1) according to basic parameters of a set false alarm probability, a reference unit, a pulse number and the like, a sequence number of a threshold in all permutation statistical magnitudes in a descending order is calculated, further, all effective strategies are determined, and an effective strategy matrix is formed; (2) an effective reference unit is selected according to a to-be-detected unit and GIS information of a reference unit, and the effective reference unit and a sampling value of the to-be-detected unit form an effective data matrix; (3) a detection statistical magnitude is calculated according to data of the to-be-detected unit, a permutation calculation magnitude is calculated according to the effective strategy matrix, the threshold is determined accordingly, the detection statistical magnitude is compared with the threshold, and therefore a judgment is made. The radar detection method based on knowledge auxiliary permutation detection effectively improves the performance of an algorithm under a complex clutter environment, greatly widens the application range, effectively reduces the calculation magnitude, and increases the target detection speed.

Description

A kind of radar detecting method of knowledge based complement replacement inspection

Technical field

The invention belongs to Radar Targets'Detection field, relate to the radar detecting method of a kind of knowledge based complement replacement inspection.

Background technology

CFAR (CFAR) keeps the target detection technique of constant false-alarm probability under detecting and referring to the condition that a class constantly can change at noise level.Common CFAR detection algorithm can be divided into two classes: Parameteric CFAR and nonparametric CFAR.Parameteric CFAR algorithm supposes that the distribution pattern of clutter is known usually, only needs to estimate some unknown parameters, and design thresholding makes target detection have CFAR characteristic under the assumptions accordingly.Nonparametric CFAR algorithm does not need the distribution pattern knowing clutter, and it is the inspection policies under the weak hypothesis about ground unrest or clutter statistical characteristics with constant false-alarm probability.In the non-homogeneous clutter environment of complexity, these traditional CFAR detection techniques are difficult to obtain good performance usually.So an important directions of CFAR research is in recent years exactly, by means of knowledge assistance, the various prior imformation of abundant fusion (such as target travel information, weather information, road traffic map and geography information figure etc.), the CFAR algorithm of design self-adaptation complex clutter environment.The people such as A.De Maio utilize various possible prior imformation, devise the Parameteric CFAR detection algorithm that several knowledge based is auxiliary.When processing non-homogeneous clutter environment, the adaptive detection algorithm that these knowledge based are assisted because adding data screening process, and shows the performance being obviously better than traditional C FAR and detecting.

Current knowledge assistance CFAR algorithm mainly merges prior imformation formation on the basis that common Parameteric CFAR detects, thus need equally to suppose that the distribution pattern of clutter is known, this means that it is only applicable to the clutter type that a group distribution function has clear and definite analytic expression.In practical application, this, by being a serious unfavorable factor, causes the scope of application of knowledge assistance Parameteric CFAR algorithm greatly limited, because in most cases the distribution pattern of clutter is unknown or cannot use concrete function representation.

Summary of the invention

The technical problem to be solved in the present invention is: for traditional C FAR algorithm penalty in non-homogeneous clutter environment, the problem of novel knowledge assistance Parameteric CFAR algorithm scope of application critical constraints, the radar detecting method of a kind of knowledge based complement replacement inspection is proposed, the method, by detection perform excellent under the permutation test cascade of data selector and nonparametric is achieved complex clutter environment, has the scope of application more widely simultaneously.

The technical solution adopted for the present invention to solve the technical problems is: a kind of radar detecting method of knowledge based complement replacement inspection, and the method performing step is as follows:

Step (1), according to setting false-alarm probability and the basic parameter such as reference unit and pulse number, calculate thresholding in all displacement statistics by sequence number during descending arrangement, and then determine all available strategies and form effective strategy matrix thus;

Step (2), select effective reference unit according to the GIS information of unit to be detected and reference unit, and form effective data matrix by the sampled value of effective reference unit and unit to be detected;

Step (3), calculate detection statistic according to the data of unit to be detected, according to available strategy matrix computations displacement statistic and and then determine thresholding, detection statistic and thresholding are compared thus complete judgement.

Further, the method adopts the structure of data selector cascade permutation test to realize, after the data of data selector to reference unit are screened, only retain the reference data that those are similar to element characteristics to be detected, and then utilize these reference datas to carry out object detection process by permutation test.

Further, the false-alarm probability according to setting in described step (1) and the basic parameter such as reference unit and pulse number, calculate thresholding in all displacement statistics by sequence number during descending arrangement, and then determine all available strategies and form effective strategy matrix thus, this step is preprocessing process, only need when carrying out target detection to multiple unit perform once and store acquired results, and perform once when need not detect each unit.

Further, this method propose the method for discrimination of concept about " available strategy " and " available strategy ", and the displacement statistic determination thresholding utilizing available strategy corresponding.

The present invention compared with prior art advantage is:

1) the GIS information by introducing priori is screened reference unit, selects the reference data with unit to be detected with similar characteristic, effectively improves the performance of algorithm under complex clutter environment;

2) the target detection stage adopts the permutation test process of nonparametric, makes algorithm of the present invention can be applied to the unknown of various clutter distribution pattern or with in the situation of concrete function representation, greatly cannot extend the scope of application;

3) in the permutation test stage, have employed a kind of brand-new implementation method, significantly reduce calculated amount, improve target detection speed.

Accompanying drawing explanation

Fig. 1 is observation data model conventional in Radar Targets'Detection;

Fig. 2 is the particular flow sheet of the radar detecting method of a kind of knowledge based complement replacement inspection of the present invention.

Embodiment

The present invention is introduced in detail below in conjunction with the drawings and the specific embodiments.

In the present invention, GIS information (GIS and Geographic Information System) is applied in Radar Targets'Detection, proposes the detection method of data selector cascade permutation test, effectively improve the detection perform in complex clutter environment; Meanwhile, devise a kind of brand-new permutation test concrete methods of realizing, greatly reduce the calculated amount of permutation test.

The radar detecting method of a kind of knowledge based complement replacement inspection of the present invention, its concrete implementing procedure as shown in Figure 2, specifically comprises following 3 steps:

1, according to false-alarm probability and the basic parameter such as reference unit and pulse number of setting, calculate thresholding in all displacement statistics by sequence number during descending arrangement, and then determine all available strategies and form effective strategy matrix thus.

Permutation test is a kind of typical nonparametric CFAR detection algorithm, makes it keep constant false alarm rate only to need the joint distribution function of clutter to meet " permutation invariance ", and without the need to knowing the concrete distribution pattern of clutter.Therefore, when about the prior imformation of clutter type less, permutation test has applicability widely, but permutation test also exists the problems such as calculated amount is excessive.

Consider pulsed radar observation model as shown in Figure 1: whole observation area comprises M reference unit of unit to be detected and symmetria bilateralis distribution thereof, and altogether carries out N independent scan.The sampling all observing units being carried out to i-th scanning is denoted as:

x _i＝(x _i,1,x _i,2,…,x _i,M/2,x _i,0,x _i,M/2+1,…,x _i,M) ^T(1)

Wherein, x _i0for the sampling of unit to be detected, x _i1, x _i2..., x _iMfor the sampling of reference unit, i=1,2 ..., N.So whole observation data can be denoted as the matrix that (M+1) × N ties up in order to easy, the row vector of matrix X is designated as respectively, r ₀=(x _1,0, x _2,0..., x _{n, 0}) for treating the sampling of detecting unit N scanning; r _j=(x _{1, j}, x _{2, j}..., x _n,j) be the sampling that a jth reference unit is scanned for N time, j=1,2 ..., M.

Definition H ₀be assumed to be unit place to be detected and there is not target, H ₁be assumed to be unit place to be detected and there is target.At H ₀under supposing, vector x _i, i=1,2 ..., each component of N is independent same distribution (IID); And at H ₁under supposing, only has x _i1, x _i2..., x _iMiID.By H ₀and H ₁lower stochastic variable x _i0probability density function be denoted as f respectively _0i(x _i0) and f _1i(x _i0), i=1,2 ..., N.Because the two kinds of lower reference unit place of hypothesis driftlessness all the time, so x _ijprobability density function be always f _0i(x _ij), i=1,2 ..., N and j=1,2 ..., M.Thus, the probability density function of matrix X is:

$H_0:f_x(x_1,x_2,...,x_N|H_0)=\underset{i=1}{\overset{N}{\mathrm{\Π}}}{f}_{0i}\left(x_{i0}\right)\underset{j=1}{\overset{M}{\mathrm{\Π}}}{f}_{0i}\left(x_{\mathrm{ij}}\right)---\left(2\right)$

$H_1:f_x(x_1,x_2,...,x_N|H_1)=\underset{i=1}{\overset{N}{\mathrm{\Π}}}{f}_{1i}\left(x_{i0}\right)\underset{j=1}{\overset{M}{\mathrm{\Π}}}{f}_{0i}\left(x_{\mathrm{ij}}\right)---\left(3\right)$

Formula (3) and formula (2) are done compare and take the logarithm, can obtain log-likelihood ratio (i.e. detection statistic) is:

$T\left(X\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\ln \left(\frac{f_{1i}\left(x_{i0}\right)}{f_{0i}\left(x_{i0}\right)}\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{a}_i\left(x_{i0}\right)---\left(4\right)$

Wherein, $a_i\left(x_{i0}\right)=\ln \left(\frac{f_{1i}\left(x_{i0}\right)}{f_{0i}\left(x_{i0}\right)}\right),i=\mathrm{1,2},...,N.$

From the often row x of matrix X _iin appoint get a sampled value and with the x in its alternate form (4) _i0(this corresponds to will in X with x _i0displacement), can obtain:

$T_k(x_1,x_2,...,x_N)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\ln \left(\frac{f_{1i}\left(x_{{\mathrm{ik}}_i}\right)}{f_{0i}\left(x_{{\mathrm{ik}}_i}\right)}\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{a}_i\left(x_{{\mathrm{ik}}_i}\right)---\left(5\right)$

Above formula is called displacement statistic.Wherein, k=(k ₁, k ₂..., k _n) and k _i∈ 0,1 ..., M}, i=1,2 ..., N, obviously, formula (4) be formula (5) k=(0,0 ..., 0) time special case.

For observing matrix X=(x ₁, x ₂..., x _n), different according to the value of k, altogether can obtain (M+1) ⁿindividual different T _k() value.By these T _k() according to the arrangement of descending order, and gets Q as thresholding (wherein, Q is by the false-alarm probability P set _fathe value determined).So permutation test can be expressed as:

At H ₀under supposing, vector x _i, i=1,2 ..., each component of N meets IID characteristic, then all (M+1) ⁿindividual T _k() is that equiprobability occurs, so false-alarm probability is:

$P_{\mathrm{fa}}=\frac{Q}{{(M+1)}^N}---\left(7\right)$

Visible, false-alarm probability P _fabe only the function of parameter M, N and Q, and have nothing to do with clutter distribution function, this just permutation test without the need to knowing that clutter distribution pattern can keep the reason of constant false alarm rate.

Can directly obtain according to formula (7):

Q＝P _fa×(M+1) ^N(8)

This is in traditional permutation test, the sequence number of thresholding T in all displacement statistics of descending arrangement.And in radar detecting method for the inspection of a kind of knowledge based complement replacement of the present invention, the data of original reference unit need first to be used further to permutation test after data selector screens.The reference unit number that tentation data selector switch retains after screening is M _s, then in follow-up permutation test, the sequence number of thresholding T in all displacement statistics should be:

Q _s＝P _fa×(M _s+1) ^N(9)

In the present invention, data selector can export (a M after screening sampled data _s+ 1) the valid data matrix X of × N dimension _s, the sampled value of the first behavior unit to be detected of this matrix, other behavior the sampled value of selected reference unit.By matrix X _seach element press a _i(x _ij)=ln (f _1i(x _ij)/f _0i(x _ij)) carry out process and obtain matrix A _s, then to A _seach column element carry out descending sort, can matrix be obtained obviously, also be (M _s+ 1) × N, specifically might as well be denoted as it:

${\tilde{A}}_s\stackrel{\mathrm{\Δ}}{=}\left[\begin{array}{cccc}{\tilde{a}}_{11}& {\tilde{a}}_{21}& ...& {\tilde{a}}_{N1}\\ {\tilde{a}}_{12}& {\tilde{a}}_{22}& ...& {\tilde{a}}_{N2}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ {\tilde{a}}_{1,M_s+1}& {\tilde{a}}_{2,M_s+1}& ...& a_{N,M_s+1}\end{array}\right]$

From matrix each row in choose one-component respectively and obtain one group (wherein, m _i∈ 1,2 ..., M _s+ 1}, i=1,2 ..., N), a displacement statistic can be tried to achieve thus this process corresponds to vector again due to matrix each column element be descending sort, if so m=(m ₁, m ₂..., m _n) and n=(n ₁, n ₂..., n _n) meet n _i≤ m _ito i=1,2 ..., N sets up, then necessarily have so, for vector m, at least can find individually to be not less than statistic.If then scarcely at front Q _sthe row of individual maximum statistic, at this moment claim m to be " invalidation policy "; Otherwise, just may belong to front Q _sone of individual maximum statistic, claims m to be " available strategy ".Using all available strategies as a line, effective strategy matrix can be formed.Because statistic corresponding to any invalidation policy be not at front Q _sthe row of individual maximum statistic, so thresholding T is Q in displacement statistic corresponding to available strategy matrix _sthat large statistic.

Comprehensive above analysis, determine that effective strategy matrix specifically can be divided into two steps:

1. calculating formula Q _s=P _fa× (M _s+ 1) ⁿ.

2. to each policy vector m=(m ₁, m ₂..., m _n), m _i∈ 1,2 ..., M _s+ 1}, i=1 ..., N, if then this vector m is judged to available strategy.Using all available strategies as a line, form effective strategy matrix.

It should be noted that, Q _sand available strategy matrix only with parameter P _fa, M _srelevant with N, and have nothing to do with the concrete sampled value of radar observation unit.When carrying out target detection to multiple observing unit, due to parameter P _fa, M _simmobilize with N, so using the present invention the 1st step as preprocessing process, only can perform before detection starts once and event memory, and need not perform during each unit inspection.Such process can reduce calculated amount greatly.

2, select effective reference unit according to the GIS information of unit to be detected and reference unit, and form effective data matrix by the sampled value of effective reference unit and unit to be detected.

In the present invention, data selector needs the GIS information utilizing priori to screen reference unit.Here used GIS information, refers to grid observation area being divided into fixed size, then uses the matrix of the geographic entity (as land, ocean, the woods etc.) corresponding to each grid of numeral.Generally speaking, have the region of identical geographic entity, its noise performance is also close, this just in the present invention data selector utilize GIS to filter out the basis of even clutter unit.

Here adopt the fixed number selector switch than being easier to realize, also known as FKA data selector, this selector switch can select fixed value M from an initial M reference unit _sindividual reference unit.In order to easy, first consider GIS grid and distance by radar-localizer unit situation one to one.The GIS information of unit to be detected (CUT) and reference unit is denoted as Y respectively ₀and Y _j, j=1 ..., M, and defined function:

$D\left(j\right)=\left\{\begin{array}{c}0,Y_j\≠Y_0\\ 1,Y_j=Y_0\end{array}\right.---\left(10\right)$

If then effective reference unit is chosen as and meets D (j)=1 and the nearest M of distance CUT _sindividual reference unit; If then effective reference unit is chosen as all meet D (j)=1 reference unit and D (j)=0 middle distance CUT nearest individual reference unit.Using the data of unit to be detected as the first row, the data of each effective reference unit are also embarked on journey separately, form (a M _s+ 1) matrix of × N dimension is valid data matrix X _s.

In some cases, the resolution of GIS may be higher than the resolution of radar a lot, will comprise multiple GIS unit in such distance by radar-localizer unit.Now, data selector will become complicated a little.Assuming that always have L kind geography information type, numbering is respectively 1,2 ..., L.To each radar observation unit, statistics be wherein numbered l ∈ 1,2 ..., the GIS unit number of L}, can obtain a L n dimensional vector n v, it can be used as the geography information eigenvector of this radar cell.So, jth ∈ 1,2 ..., the geography information similarity of M} reference unit and unit to be detected can represent with following formula:

$\mathrm{SIM}=\frac{\⟨v_j,v_0\⟩}{\left|\right|v_j\left|\right|\left|\right|v_0\left|\right|}---\left(11\right)$

Wherein, v _jand v ₀be respectively the geography information eigenvector of a jth reference unit and unit to be detected, <> represents inner product operation, || || represent that vector asks mould.In this case, FKA data selector only need select the M that makes SIM maximum _sindividual reference unit.

In sum, determine that effective data matrix specifically can be divided into following three steps:

1. the geography information similarity between each reference unit and unit to be detected is determined according to formula (10) or formula (11);

2. from all reference units, M is selected _sthe highest unit of unit geography information similarity individual and to be detected is as effective reference unit (if similarity is identical, then the reference unit that chosen distance unit to be detected is nearer);

3. using the data of unit to be detected as the first row, the data of each effective reference unit are also embarked on journey separately, form (a M _s+ 1) matrix of × N dimension is valid data matrix X _s.

3, calculate detection statistic according to the data of unit to be detected, according to available strategy matrix computations displacement statistic and and then determine thresholding, detection statistic and thresholding are compared thus complete judgement.

Valid data matrix X is obtained by data selector _safterwards, just need to carry out target detection with permutation test.When adopting direct computing method to carry out permutation test, need to calculate all (M _s+ 1) ⁿindividual displacement statistic T _k(), more therefrom select Q _sindividual maximal value is as thresholding.For slightly bigger M _sand N, this will be very huge calculated amount.Consider this point, devise a kind of brand-new permutation test implementation method in the present invention, the method only needs to calculate displacement statistic corresponding to all available strategies, thus also known as work " available strategy " method.

Assuming that FKA data selector selects is the m altogether _sindividual reference unit, then effectively observation data matrix can be expressed as:

$X_s\stackrel{\mathrm{\&Delta;}}{=}\left[\begin{array}{c}{r}_0\\ r_{k_1}\\ r_{k_2}\\ .\\ .\\ .\\ r_{k_{M_s}}\end{array}\right]=\left[\begin{array}{cccc}{x}_{10}& x_{20}& ...& x_{N0}\\ x_{1k_1}& x_{2k_1}& ...& x_{N{k}_1}\\ x_{1k_2}& x_{2k_2}& ...& x_{N{k}_2}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ x_{1k_{M_s}}& x_{2k_{M_s}}& ...& x_{N{k}_{M_s}}\end{array}\right]$

Wherein, the sampling of detecting unit N scanning is treated in the first behavior, and remaining row is the sampling to N the scanning of effective reference unit.Q is tried to achieve in the present invention the 1st step _swith under the precondition of available strategy matrix, the detailed process of this implementation method is as follows:

1. according to matrix X _scalculate new matrix A _s

$A_s\stackrel{\mathrm{\&Delta;}}{=}\left[\begin{array}{cccc}{a}_{10}& a_{20}& ...& a_{N0}\\ a_{1k_1}& a_{2k_1}& ...& a_{N{k}_1}\\ a_{1k_2}& a_{2k_2}& ...& a_{N{k}_2}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ a_{1k_{M_s}}& a_{2k_{M_s}}& ...& a_{N{k}_{M_s}}\end{array}\right]$

Wherein, a _ij=a _i(x _ij)=ln (f _1i(x _ij)/f _0i(x _ij)), i=1,2 ..., N and in addition, under Gaussian Clutter background and fixed target situation, quadratic detection can be reduced to or linear detection a _ij=| x _ij|.

2. to matrix A _sthe first row element summation, obtain detection statistic

3. by matrix A _seach column element by descending sort, obtain new matrix

${\tilde{A}}_s\stackrel{\mathrm{\&Delta;}}{=}\left[\begin{array}{cccc}{\tilde{a}}_{11}& {\tilde{a}}_{21}& ...& {\tilde{a}}_{N1}\\ {\tilde{a}}_{12}& {\tilde{a}}_{22}& ...& {\tilde{a}}_{N2}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ {\tilde{a}}_{1,M_s+1}& {\tilde{a}}_{2,M_s+1}& ...& a_{N,M_s+1}\end{array}\right]$

Wherein, ${\tilde{a}}_{i1}\&GreaterEqual;{\tilde{a}}_{i2}\&GreaterEqual;...\&GreaterEqual;{\tilde{a}}_{i,M_s+1},i=\mathrm{1,2},...,N.$

4. according to matrix calculate the displacement statistic that in available strategy matrix, each available strategy is corresponding, then by its descending sort, and select Q _sindividual as thresholding T.

5. compare detection statistic T (X) and thresholding T, make judgement.

Can find out, no matter be the available strategy method proposed in direct computing method or the present invention, main calculated amount is all displacement statistic T _kthe calculating of ().Therefore, can using the mark of the displacement statistic number of required calculating as calculated amount size.For available strategy method, need to calculate statistic corresponding to all available strategies, its number is:

$N_{\mathrm{val}}=\mathrm{Pr}(\underset{i=1}{\overset{N}{\mathrm{\&Pi;}}}{m}_i\&le;Q_s)\&times;{(M_s+1)}^N---\left(12\right)$

Wherein, m _iget 1,2 equiprobably ..., M _s+ 1, i=1,2 ..., N, represent probability.And direct computing method needs to calculate all statistics, its number is:

N _dir＝(M _s+1) ^N(13)

Formula (12) and formula (13) ask ratio, can obtain:

$\mathrm{\&gamma;}=\frac{N_{\mathrm{val}}}{N_{\mathrm{dir}}}=\mathrm{Pr}(\underset{i=1}{\overset{N}{\mathrm{\&Pi;}}}{m}_i\&le;Q_s)---\left(14\right)$

Above formula can be similar to the ratio of the calculated amount regarding these two kinds of implementation methods as.Its value is less, shows that available strategy method efficiency is higher.Again because γ≤1 perseverance is set up, so available strategy method scarcely can be poorer than direct computing method.In fact, at Q _swhen less, the calculated amount of available strategy method will much smaller than direct computing method.

Method below by emulation is verified the present invention.The data that experiment adopts are: N before intercepting from Canadian Mcmaster IPIX radar actual measurement file 13 _r=102 range units, 220 ° to 304 ° are total to N _a=1834 orientation to data.In object detection field, IPIX radar data disclosed in Mcmaster university is one of the conventional data for evaluation algorithms performance.The platform configuration that experiment adopts is: the CPU of Intel Core i7-3770,4GB internal memory.The optimum configurations that experiment adopts is: false-alarm probability P _fa=10 ^-6, reference window length M=12, multiple scanning times N=7, reference unit number M selected by data selector _s=9.Adopt the radar detecting method of knowledge based complement replacement inspection in traditional permutation test method and the present invention to carry out false-alarm detection to observation area respectively, the former false-alarm number is 638, and the false-alarm number of the latter is 99.Then, each orientation of observation data in, random selecting range gate insertion point target, and carry out target detection by these two kinds of detection methods respectively.Result shows, in the present invention, the radar detecting method of knowledge based complement replacement inspection has the advantage of about 3dB than traditional permutation test method.Finally, adopt direct computing method and available strategy method as the concrete methods of realizing of replacing check portion in the radar detecting method that the present invention is based on knowledge assistance permutation test and dividing respectively, 100 observing units are detected.Time needed for direct computing method is 2018s, and the time needed for available strategy method is only 0.26s.

The techniques well known related in the present invention does not elaborate.

Claims (4)

1. a radar detecting method for knowledge based complement replacement inspection, is characterized in that performing step is as follows:

Step (1), according to setting false-alarm probability and the basic parameter such as reference unit and pulse number, calculate thresholding in all displacement statistics by sequence number during descending arrangement, and then determine all available strategies and form effective strategy matrix thus;

Step (2), select effective reference unit according to the GIS information of unit to be detected and reference unit, and form effective data matrix by the sampled value of effective reference unit and unit to be detected;

Step (3), calculate detection statistic according to the data of unit to be detected, according to available strategy matrix computations displacement statistic and and then determine thresholding, detection statistic and thresholding are compared thus complete judgement.

2. the radar detecting method of knowledge based complement replacement inspection according to claim 1, it is characterized in that: the method adopts the structure of data selector cascade permutation test to realize, after the data of data selector to reference unit are screened, only retain the reference data that those are similar to element characteristics to be detected, and then utilize these reference datas to carry out object detection process by permutation test.

3. the radar detecting method of knowledge based complement replacement inspection according to claim 1, it is characterized in that: the false-alarm probability according to setting in described step (1) and the basic parameter such as reference unit and pulse number, calculate thresholding in all displacement statistics by sequence number during descending arrangement, and then determine all available strategies and form effective strategy matrix thus, this step is preprocessing process, only need when carrying out target detection to multiple unit perform once and store acquired results, and perform once when need not detect each unit.

4. the radar detecting method of knowledge based complement replacement inspection according to claim 1, it is characterized in that: the method for discrimination this method proposing concept about " available strategy " and " available strategy ", and the displacement statistic determination thresholding utilizing available strategy corresponding.