CN104215935B - A kind of radar shell target identification method based on decision-making Weighted Fusion - Google Patents

Abstract

The invention belongs to radar target identification method, relate to conventional low-resolution radar to shell classification target recognition methods, particularly relate to shell classification target multiple features fusion recognition methods.Present invention shell class target has the advantages that to determine flight path, positional information first by normal radar detection target, and the track characteristics such as the flying height of target, flight speed, trajectory tortuosity, vertical direction acceleration are calculated with this, utilize these track characteristics to use decision Tree algorithms to carry out Classification and Identification；Re-use the RCS value during the target flight that radar obtains, extract the features such as RCS serial mean, variance, maximum, utilize these RCS features to use algorithm of support vector machine to carry out Classification and Identification；Finally two kinds of recognition results are weighted Decision fusion, obtain final recognition result.The present invention improves the classification accuracy rate of the single grader of single features；And achieve with this and under conventional low-resolution radar, shell classification target effectively to be identified.

Description

A kind of radar shell target identification method based on decision-making Weighted Fusion

Technical field

The invention belongs to radar target identification method, relate to conventional low-resolution radar to shell classification target recognition methods, Particularly relate to shell classification target multiple features fusion recognition methods.

Background technology

In modern battlefield, rocket gun, cannon and mortar are currently used offensive weapons the most frequently.In war In, if this kind of shell target can be detected in time and identify, it is possible to implement effectively to intercept to it before shell is ignited, improve Air defense operation ability, reduces loss, grasps battlefield supremacy.And utilize Radar Technology unfriendly target is carried out detect identification be One of technological means the most effectively.

At present, the study hotspot of radar target identification method is concentrated mainly on high resolution radar identification aspect.High-resolution thunder The target echo reached can reflect the detailed information such as the shape of target, size and structure.These information can be as target recognition Validity feature.And active service radar mostly is conventional low-resolution radar, it is impossible to obtain the architectural feature of target, based on high resolution radar Target identification method is to low-resolution radar inapplicable.Therefore, from reality application, according to the difference of target to be identified, The target identification technology special for conventional low-resolution radar research is significant.

Summary of the invention

It is an object of the invention to based on active service routine low-resolution radar, it is provided that a kind of radar based on decision-making Weighted Fusion Shell target identification method.

The present invention has the feature determining flight path according to shell class target, first by normal radar detection target Positional information, and the track characteristics such as the flying height of target, flight speed, trajectory tortuosity, vertical direction acceleration are calculated with this, These track characteristics are utilized to use decision Tree algorithms to carry out Classification and Identification.Re-use the thunder during the target flight that radar obtains Reach scattering section (Radar Cross-Section, RCS) value, extract the features such as the average of RCS sequence, variance, maximum, profit Algorithm of support vector machine is used to carry out Classification and Identification by these RCS features.Finally two kinds of recognition results are weighted decision-making to melt Close, obtain final recognition result.

A kind of radar shell target identification method based on decision-making Weighted Fusion, specific as follows:

S1, obtained each target RCS sequence in flight course by radar detection and position sequence constitutes track sets, will Described track sets is randomly divided into training dataset, test data set and target data set to be identified, wherein, a RCS sequence Being made up of the RCS value of n sampling instant of continuous print, described position sequence is by distance R of n sampling instant of continuous print, azimuth A Constituting with angle of pitch E, 1≤n≤N, N are test sample sum, 10 ＜ N≤20；

S2, extracting the motion feature of target j of sampling instant i according to position sequence described in S1, described motion feature includes Target j flight speed V_i, target j flying height Z_i, target j vertical direction accelerationTarget j horizontal direction accelerationWith Target j trajectory tortuosityThe target j motion feature vector set X that described motion feature composition five is tieed up, wherein, j=1,2 ..., M, i=1,2 ..., N, M are target classification number；

S3, count average RM of S1 one RCS sequence, variance RD and maximum R_maxConstitute three-dimensional RCS feature Vector set Y={y_j(i-n)| j=1,2 ..., M；I=1,2 ..., N}, wherein,FeatureThe most corresponding target j RCS average RM in (i-n) individual sample window_(i-n), variance RD_(i-n), maximum Rmax_(i-n), described (i-n) individual sample window is that continuous n sampling instant is constituted, DescribedDescribedDescribedK is the kth sampled data in (i-n) individual sample window；

S4, in target j motion feature vector set X described in S2 training data use classical decision tree algorithm (Classification And Regression Tree, CART) carries out classification based training, generates categorised decision tree (Decision Tree, DT)；

S5, utilize each motion feature vector sample to each target in test data set described in S1 of the decision tree DT described in S4 x_iCarrying out Classification and Identification, output recognition result is five n dimensional vector nsDescribed x_i For the motion feature vector of each target in ith sample moment, describedIn have and only a nonzero element, described non-zero Element number i.e. corresponds to the target class alias identified, and described nonzero element is equal to 1；

S6, utilize S4 gained categorised decision tree that test data set described in S1 is carried out target classification, add up described classification certainly The classification results confidence level of plan Tree ClassifierWherein,Presentation class decision-making It is classification confidence during target j that Tree Classifier provides recognition result, [*]^TRepresent transposition；

RCS feature vector set Y described in S7, use S3 constitutes the training sample set (y of support vector machine_i,z_i), to described training Sample set (y_i,z_iTraining data in) uses Nonlinear Support Vector Machines one to one to carry out classification based training, obtains optimum classifier FunctionWherein, the kernel function that described vector machine is selected isz_iFor sample vector collection y_iTarget designation, b^*For bias vector, δ is kernel functional parameter, A Lagrange multiplier, sgn (*) is sign function；

S8, utilize each RCS characteristic vector sample to target each in test data set of optimum classifier function g (y) described in S7 This Y carries out Classification and Identification, and output differentiates that result is

S9, according to C described in S6^treeDescribed in S8Count the classification of support vector machine classifier based on RCS feature Result confidence levelWherein,Represent support vector machine classifier and provide knowledge Other result is classification confidence during target j；

S10, to described in S8With the C described in S9^svmBeing weighted merging, the class weights obtaining certain sample dividesThe classification number finally determining sample to be identified isThat classification that in i.e. M weighting point, score value is maximum.

Further, n=10 described in S1.

Further, extracting the motion feature of each sampling instant target j described in S2, concrete grammar is as follows:

S21, according to distance R described in S1, azimuth A and the angle of pitch set up the target rectangular coordinate system of sampling instant i, $\begin{array}{c}{z}_i=R_i\·\sin \left(E_i\right)\\ x_i=R_i\·\cos \left(E_i\right)\·\sin \left(A_i\right)\\ y_i=R_i\·\cos \left(E_i\right)\·\cos \left(A_i\right)\end{array},$ Wherein, R_iFor the target range information of sampling instant i, A_iTarget bearing for sampling instant i Angle information, E_iTarget pitch angle information for sampling instant i；

S22, by adjacent two moment target j coordinate extract motion feature, described adjacent two moment are sampling instant i And sampling instantWherein, T is sampling time interval, v_iFor The instantaneous velocity of target j in the i-th moment,It is the horizontal velocity of the i-th moment target j,It is that the i-th moment target j is at Vertical Square Speed upwards,It is the trajectory slope of the i-th moment target j,It is the i-th moment target j acceleration in vertical direction,It it is the i-th moment target j acceleration in the horizontal direction；

S23, take each target 5 motion features in each moment and constitute the Target Motion Character vector set X={x of five dimensions_ji|j =1,2 ..., M；I=1,2 ..., N}, wherein,Spy for the jth target ith sample moment Levy vector, featureThe most corresponding target j is at the instantaneous velocity v in ith sample moment_i, flight height Degree z_i, vertical direction instantaneous accelerationHorizontal direction instantaneous accelerationBallistic trajectory curvature

Further, the classification results confidence level C of statistical classification decision tree classifier described in S6^treeComputational methods are:Wherein, z_iFor sample x_iTarget class number, num () represent statistics meet The sample size of conditional expression in bracket,The number of times correctly classified for jth class target sample,For all test specimens This is categorized as jth classification target number of times, N by mistake_jFor jth classification target sample number in test sample.

The invention has the beneficial effects as follows:

By analyze trajectory class target flight path feature, be extracted the flight speed during target flight, highly, 5 motion features such as vertical direction acceleration, horizontal direction acceleration and trajectory tortuosity are as classification foundation, and use decision tree Method carries out Classification and Identification；Consider the unbalanced caused grader recognition result statistics of different classes of target sample quantity by mistake Difference, devises more reasonably grader classification results confidence calculations method；Take full advantage of different characteristic and different classification The classification results that device produces has the feature of complementarity, uses the mode of weighting to achieve the decision-making of different characteristic difference grader Merge, improve the classification accuracy rate of the single grader of single features；And achieve under conventional low-resolution radar big gun with this Play classification target effectively to identify.Emulation data are identified test, and correct recognition rata reaches 90%.

Accompanying drawing explanation

Fig. 1 is decision-making Weighted Fusion radar shell target identification method flow chart.

Detailed description of the invention

Below in conjunction with embodiment and accompanying drawing, describe technical scheme in detail.

As it is shown in figure 1, use shell rigid body six degree of freedom trajectory model, Runge Kutta algorithm is used to solve equation, to compeling to hit The trajectory track of big gun, rocket gun and howitzer carries out simulation calculation.During calculating, use the fight of STK software emulation Machine and the flight path of helicopter, every class target all produces 100 tracks by the way of being randomly provided initial transmissions parameter Sequence.Use FEKO computed in software mortar, rocket gun, howitzer, fighter plane and helicopter five class target (that is, target classification Number M=5) RCS value, every every track of classification target and RCS value are all according to the specific works parameter of radar of some model Add-on system error.With reference to the work schedule of the radar of some model described, emulation obtains every classification target each flight rail Each sampling instant t in mark_iRCS value Rcs_iWith each sampling instant t_iTarget position information: distance R_i, azimuth A_i, bow Elevation angle E_i。

Randomly choose from 100 track sets of every classification target 30 as training dataset, 30 as test numbers According to collection, remaining 40 as data set to be identified.

According to the motion feature of target location sequential extraction procedures each sampling instant target, build target in the ith sample moment Rectangular coordinate system: $\begin{array}{c}{z}_i=R_i\·\sin \left(E_i\right)\\ x_i=R_i\·\cos \left(E_i\right)\·\sin \left(A_i\right)\\ y_i=R_i\·\cos \left(E_i\right)\·\cos \left(A_i\right)\end{array},$ Motion feature is extracted by the coordinates of targets in the most adjacent two momentWherein, T is sampling time interval, v_iIt it is the target wink in the i-th moment Shi Sudu,It is the horizontal velocity of the i-th moment target,It is the i-th moment target speed in vertical direction,When being i-th Carve target trajectory slope,It is the i-th moment target acceleration in vertical direction,Be the i-th moment target in the horizontal direction On acceleration.Take each target 5 motion features in each moment and constitute five D feature vectors collection X={x_ji| j=1,2 ..., M； I=1,2 ..., N}, M are target classification number, and N is test sample sum,For jth target i-th The characteristic vector of individual sampling instant, feature The most corresponding target j is in the wink in ith sample moment Shi Suduv_i, flying height z_i, vertical direction instantaneous accelerationHorizontal direction instantaneous accelerationBallistic trajectory curvature

The RCS value of often continuous n=10 sampling instant (in one section of sample window) is constituted a RCS sample sequence, statistics RCS feature in each sequence:

Wherein, RM_i、RD_i、Rmax_iIt is respectively certain target ith sample

RCS average, variance and maximum in window.

Take each target 3 RCS statistical values in each sample window and constitute three-dimensional feature vector set Y={y_ji| j=1,2 ..., M；I=1,2 ..., N-10},Wherein, featureCorresponding target j is adopted in i-th respectively RCS average RM in sample window_i, variance RD_i, maximum Rmax_i。

Classical CART decision Tree algorithms is used to carry out classification based training the motion feature vector set X of the training data obtained, Generate categorised decision tree DT.

Utilize the decision tree DT each motion feature vector sample x to each target in test data set_iCarry out Classification and Identification, Output recognition result is five n dimensional vector ns:These 5 results have and only has One nonzero element, this nonzero element numbering i.e. corresponds to the target class alias identified, and described nonzero element is equal to 1.

The classification results confidence level of statistical decision Tree ClassifierWhereinGeneration It is classification confidence during target j that table sort device provides recognition result.Now, C^treeIt is five dimensional vectors.

Owing to there were bigger difference 5 class target respective flight time (landing from being transmitted into), cause putting at every classification carrying The characteristic vector length got, the i.e. size of feature samples collection have the biggest difference.Therefore, at statistical sorter to certain classification target Needing during result confidence level to consider such target and other classification target sample number differences, concrete mode is as follows:

Wherein, z_iFor sample x_iTarget class number, num () represent System

Meter meets the sample size of conditional expression in bracket,The number of times correctly classified for jth class target sample,It is categorized as jth classification target number of times, N by mistake for all test samples_jFor jth classification target sample number, N in test sample For test sample sum.

Training sample set (the y of support vector machine is constituted with the RCS feature vector set Y of the training data obtained_i,z_i), z_iFor Sample y_iTarget designation, use OTO (one to one) Nonlinear Support Vector Machines carry out classification based training,

Selection kernel function:

Solve optimization formula:

Obtain optimal solution: α^*=(α₁ ^*α₂ ^*,...,α_l ^*)^T

Optimum biasing:

Thus optimum classifier function can be obtained be:

Utilize optimum classifier function g (y) each RCS characteristic vector sample y to target each in test data set_iCarry out point Class identification, output differentiates that result isBy formula

The classification results adding up support vector machine classifier based on RCS feature is put Letter

DegreeWhereinRepresent grader to be given when recognition result is target j Classification confidence.

For certain sample sequence to be identified, extract the motion feature vector set of target sample to be identified:

Wherein, n is the number of samples of target to be identified.

For certain sample sequence to be identified, extract the RCS feature vector set of target sample sequence to be identified:

Utilize the decision tree DT the obtained motion feature vector p to each sample to be identified_iCarrying out Classification and Identification, output is sentenced It is not 5 classification target recognition results:5 results have and only one non- Neutral element, this nonzero element numbering i.e. corresponds to the target class alias identified, and described nonzero element is equal to 1.

Utilize optimum classifier function g (y) the obtained RCS characteristic vector r to each sample to be identified_iCarry out classification to know Not, output differentiates that result is

It is weighted two recognition results merging, obtains 5 class weights of certain sample i are divided:

Finally, determine that the classification number k of sample i to be identified is those classifications that in 5 weightings point, score value is maximum:

Employing emulates the correct of the normal radar Ballistic Target recognition methods based on Decision fusion of the data verification present invention Discrimination.Emulation creates the totally 5 classification target flight including mortar, rocket gun, howitzer, helicopter, fighter plane Data, every class target all comprises each 40 of the flight path being randomly provided emission parameter, adds up for discrimination.To 5 classifications Mark is respectively adopted 3 kinds of methods and is identified statistics: is used alone RCS feature and is supported vector machine identification, is used alone track Feature carries out decision tree identification, uses the Decision fusion method of the present invention to be identified, and recognition result is added up in table one.

Table one

From table one, its average recognition rate of traditional decision-tree being used alone track characteristic reaches 88.5%, individually makes Being 87.6% by its average recognition rate of support vector machine method of RCS feature, the discrimination that two kinds of methods obtain is essentially identical.Make The average recognition rate obtained by the Decision fusion method of the present invention reaches 92.3%, has had bright relative to first two method discrimination Aobvious raising.

Claims (4)

1. a radar shell target identification method based on decision-making Weighted Fusion, it is characterised in that comprise the steps:

S1, obtained each target RCS sequence in flight course by radar detection and position sequence constitutes track sets, by described Track sets is randomly divided into training dataset, test data set and target data set to be identified, and wherein, a RCS sequence is by even The RCS value composition of n continuous sampling instant, described position sequence by distance R of n sampling instant of continuous print, azimuth A with bow Elevation angle E is constituted, and 1≤n≤N, N are test sample sum, 10 ＜ N≤20；

S2, extracting the motion feature of target j of sampling instant i according to position sequence described in S1, described motion feature includes target j Flight speed V_i, target j flying height Z_i, target j vertical direction accelerationTarget j horizontal direction accelerationAnd target J trajectory tortuosity θ_i ¹, by described motion feature constitute five dimension target j motion feature vector set X, wherein, j=1,2 ..., M, i= 1,2 ..., N, M are target classification number；

S3, count average RM of S1 one RCS sequence, variance RD and maximum R_maxConstitute three-dimensional RCS characteristic vector Collection Y={y_j(i-n)| j=1,2 ..., M；I=1,2 ..., N}, wherein, $y_{j(i-n)}={\[y_{j(i-n)}^1,y_{j(i-n)}^2,y_{j(i-n)}^3\]}^T,$ FeatureThe most corresponding target j RCS average RM in (i-n) individual sample window_(i-n), variance RD_(i-n), maximum Rmax_(i-n), described (i-n) individual sample window is that continuous n sampling instant is constituted, DescribedDescribed ${\mathrm{RD}}_{(i-n)}=\frac{{\mathrm{\Σ}}_{k=0}^{k=n-1}{({\mathrm{Rcs}}_{(i-n)+k}-{\mathrm{RM}}_{(i-n)})}^2}{n},$ Described ${\mathrm{Rmax}}_{(i-n)}=\underset{k=0}{\overset{n-1}{\max }}\left({\mathrm{Rcs}}_{(i-n)+k}\right),$ K is the kth sampled data in (i-n) individual sample window；

S4, in target j motion feature vector set X described in S2 training data use classical decision tree algorithm (Classification And Regression Tree, CART) carries out classification based training, generates categorised decision tree (Decision Tree, DT)；

S5, utilize each motion feature vector sample x to each target in test data set described in S1 of the decision tree DT described in S4_iEnter Row Classification and Identification, output recognition result is five n dimensional vector nsDescribed x_iFor The motion feature vector of each target in ith sample moment, describedIn have and only a nonzero element, described non-zero entry Element numbering i.e. corresponds to the target class alias identified, and described nonzero element is equal to 1；

S6, utilize S4 gained categorised decision tree that test data set described in S1 is carried out target classification, add up described categorised decision tree The classification results confidence level of grader $C^{tree}={\[c_1^{tree},c_2^{tree},c_3^{tree},...,c_j^{tree},...,c_M^{tree}\]}^T,$ Wherein,Presentation class decision tree It is classification confidence during target j that grader provides recognition result, [*]^TRepresent transposition；

RCS feature vector set Y described in S7, use S3 constitutes the training sample set (y of support vector machine_i,z_i), to described training sample Collection (y_i,z_iTraining data in) uses Nonlinear Support Vector Machines one to one to carry out classification based training, obtains optimum classifier function $g\left(y\right)=sgn\{\left({\mathrm{\ω}}^{*}y\right)+b^{*}\}=sgn\{\underset{j=1}{\overset{l}{\Σ}}{a_j}^{*}{z}_{j}K(y_j,y)+b^{*}\},$ Wherein, the kernel function that described vector machine is selected isz_iFor sample vector collection y_iTarget designation, b^*For bias vector, δ is kernel functional parameter, A Lagrange multiplier, sgn (*) is sign function；

S8, utilize optimum classifier function g (y) described in S7 that each RCS characteristic vector sample Y of target each in test data set is entered Row Classification and Identification, output differentiates that result is $f_i^{svm}=\[f_{i1}^{svm},f_{i2}^{svm},f_{i3}^{svm},...,f_{ij}^{svm},...,f_{iN}^{svm}\];$

S9, according to C described in S6^treeDescribed in S8Count the classification results of support vector machine classifier based on RCS feature Confidence level $C^{svm}={\[c_1^{svm},c_2^{svm},c_3^{svm},...,c_j^{svm},...,c_M^{svm}\]}^T,$ Wherein,Represent support vector machine classifier and provide identification knot Fruit is classification confidence during target j；

S10, to described in S8With the C described in S9^svmBeing weighted merging, the class weights obtaining certain sample divides $F_i=\left[\begin{array}{c}{F}_i^1\\ F_i^2\\ F_i^3\\ ...\\ F_i^j\\ ...\\ F_i^M\end{array}\right]=C^{tree}\·f_i^{tree}+C^{svm}\·f_i^{svm}=\left[\begin{array}{c}{c}_1^{tree}\·f_{i1}^{tree}+c_1^{svm}\·f_{i1}^{svm}\\ c_2^{tree}\·f_{i2}^{tree}+c_2^{svm}\·f_{i2}^{svm}\\ c_3^{tree}\·f_{i3}^{tree}+c_3^{svm}\·f_{i3}^{svm}\\ ...\\ c_j^{tree}\·f_{ij}^{tree}+c_j^{svm}\·f_{ij}^{svm}\\ ...\\ c_M^{tree}\·f_{iM}^{tree}+c_M^{svm}\·f_{iM}^{svm}\end{array}\right],$ The classification number finally determining sample to be identified isThat classification that in i.e. M weighting point, score value is maximum.

A kind of radar shell target identification method based on decision-making Weighted Fusion the most according to claim 1, its feature exists In: n=10 described in S1.

A kind of radar shell target identification method based on decision-making Weighted Fusion the most according to claim 1, its feature exists In: extracting the motion feature of each sampling instant target j described in S2, concrete grammar is as follows:

S21, according to distance R described in S1, azimuth A and the angle of pitch set up the target rectangular coordinate system of sampling instant i, $\begin{array}{c}{z}_i=R_i\·\sin \left(E_i\right)\\ x_i=R_i\·\cos \left(E_i\right)\·\sin \left(A_i\right)\\ y_i=R_i\·\cos \left(E_i\right)\·\cos \left(A_i\right)\end{array},$ Wherein, R_iFor the target range information of sampling instant i, A_iTarget bearing for sampling instant i Angle information, E_iTarget pitch angle information for sampling instant i；

S22, being extracted motion feature by the target j coordinate in adjacent two moment, described adjacent two moment are sampling instant i and adopting Sample moment i+1:Wherein, T is sampling time interval, v_iWhen being i-th The instantaneous velocity of target j carved,It is the horizontal velocity of the i-th moment target j,Be the i-th moment target j in vertical direction Speed, θ_i ²It is the trajectory slope of the i-th moment target j,It is the i-th moment target j acceleration in vertical direction,For I-th moment target j acceleration in the horizontal direction；

S23, take each target 5 motion features in each moment and constitute the Target Motion Character vector set X={x of five dimensions_ji| j=1, 2,…,M；I=1,2 ..., N}, wherein,Feature for the jth target ith sample moment Vector, featureThe most corresponding target j is at the instantaneous velocity v in ith sample moment_i, flight height Degree z_i, vertical direction instantaneous accelerationHorizontal direction instantaneous accelerationBallistic trajectory curvature θ_i ³。

A kind of radar shell target identification method based on decision-making Weighted Fusion the most according to claim 1, its feature exists In: the classification results confidence level C of statistical classification decision tree classifier described in S6^treeComputational methods are:Wherein, z_iFor sample x_iTarget class number, num () represent statistics meet includes The sample size of conditional expression in number,The number of times correctly classified for jth class target sample,For all test samples It is categorized as jth classification target number of times, N by mistake_jFor jth classification target sample number in test sample.