CN109557529B - A Radar Target Detection Method Based on Statistical Modeling of Generalized Pareto Distribution Clutter - Google Patents

Abstract

The invention discloses a radar target detection method based on generalized Pareto distribution clutter statistical modeling, and aims to improve radar target detection performance. The technical scheme is that the maximum likelihood estimation of the logarithm is combined with the particle swarm optimization algorithm, a cost function is designed through the maximum likelihood estimation of the logarithm, and the particle swarm optimization algorithm is utilized to minimize the cost function so as to obtain the optimal parameter estimation result; and determining a detection threshold by using the optimal parameter estimation result to complete radar target detection. The invention can ensure the optimality of the solution of the parameter to be estimated in the maximum likelihood sense, improve the precision and the robustness of parameter estimation, improve the accuracy of the detection threshold and ensure the reliability of the radar target detection result.

Description

A Radar Target Detection Method Based on Statistical Modeling of Generalized Pareto Distribution Clutter

technical field

The invention belongs to the field of radar target detection, and relates to a radar target detection method for precise statistical modeling of generalized Pareto distribution by using particle swarm optimization algorithm.

Background technique

Statistical modeling of clutter data (that is, accurate fitting of the histogram of clutter data using a suitable distribution or function) is an important problem in the field of radar target detection. Combined function) and accurate acquisition of model parameters is an important basis for target detection algorithm design (such as detection threshold, false alarm probability, detection probability calculation, etc.). If there is a large error in the model parameter estimation, the accurate detection threshold cannot be obtained, resulting in a decrease in the detection probability and an increase in the false alarm probability.

Commonly used clutter data fitting distribution models include Rayleigh distribution model, Weibull distribution model, K distribution model, Gamma distribution model, etc. These distribution models have their own scope of application. For example, the Rayleigh distribution model is often used to describe low-resolution uniform clutter data, and the K distribution model is often used to describe high-resolution sea clutter data under low grazing angle conditions.

The generalized Pareto distribution model is an important statistical distribution model named after the Italian economist Vilfredo Pareto. The model was initially applied in the fields of economics, physics, hydrology and seismology, and has been gradually applied to the statistical modeling of high-resolution radar clutter data, such as literature 1: G.V. Weinberg, "Assessing Pareto fit to high- resolution high-grazing-angle sea clutter", Electronics Letters, 2011, 47(8): 516-517 (G.V.Weinberg published a paper in the 8th issue of Volume 47 of "Electronic Letters" in 2011, the Chinese title is "Pareto The evaluation of the fitting ability of distribution to sea clutter with high incidence angle") shows that the generalized Pareto distribution model has good model matching and Statistical distribution fitting capabilities. It can be seen that the parameter estimation of the generalized Pareto distribution model is an important technology involved in solving the problem of target detection in the background of sea clutter under the special application conditions of large incident angles.

The generalized Pareto distribution model can be expressed as

In the formula, f(·) represents the probability density function, z is the independent variable (represents the amplitude of the clutter signal, and its value is greater than zero), k represents the shape parameter, σ(σ>0) represents the scale parameter, and exp(·) represents The exponential function with the natural constant e as the base, f(z|k,σ) represents the probability density function of the clutter signal amplitude z including unknown distribution parameters (k and σ).

Obviously, when k=0, the generalized Pareto distribution degenerates into an exponential distribution model, and the solution of the parameter σ is relatively simple, which is beyond the scope of the present invention.

In the case of k<0, the parameters of the generalized Pareto distribution model (namely k, σ) usually use the r-order central moment E(z ^r )(E(·) Indicates that the mathematical expectation is taken, r≥1 and is an integer) for estimation, such as the current common method of moment estimation (Method of Moments, MoM), probability weighted moment method (ProbabilityWeighted Moments Method, PWMM), likelihood moment method (Likelihood Moment Method , LMM) and so on. Document 2: P.de Zea Bermudez, Samuel Kotz, "Parameter estimation of the generalized Pareto distribution Parts I&II", Stat.Plann.Inference 140, 2010: 1353-1388 (P.deZea Bermudez et al. In the paper published in Volume 140 of Reasoning Journal, the Chinese title is "Model Parameter Estimation of Generalized Pareto Distribution"), which summarizes, analyzes and compares several parameter estimation methods of generalized Pareto distribution models, and points out that the maximum likelihood Estimation (Maximum Likelihood Estimation, MLE) is the most effective estimation method.

In fact, when the radar clutter has a large tail distribution, there will be a certain deviation between the statistical origin moment of the measured data and the theoretical origin moment of the clutter model of the corresponding order, and the deviation varies with the moment order (r) increase and the decrease of the number of samples of clutter data will increase, which will have a negative impact on the estimation method based on statistical moments. Maximum likelihood estimation has high estimation accuracy, but when the radar clutter distribution model (such as the generalized Pareto distribution model, that is, the expression of the probability density function) is relatively complex, it is difficult to obtain the analytical expression of the parameter estimation by this method; in In this case, although it can be passed as document 3: Tjalling J.Ypma, "Historical development of the Newton-Raphson method", SIAM Review, 1995, 37(4): 531-551 (Tjalling J.Ypma in 1995 in The Newton-Raphson numerical algorithm in the journal "Review of the American Society for Industrial and Applied Mathematics" published in the fourth issue of volume 37, the Chinese title is the historical development of the Newton-Raphson method), but this numerical method is easy to fall into the local Extreme values cannot obtain optimal parameter estimates.

Particle Swarm Optimization (PSO) is an important intelligent optimization algorithm, which has the characteristics of simple calculation, fast convergence speed, good stability, etc., and has the ability of global optimization. Through the PSO algorithm, it is expected to solve the problem of high-precision estimation of the parameters of complex models, thereby improving the performance of radar target detection. However, there is no public literature on how to use the PSO algorithm for high-precision estimation of Pareto clutter model parameters and high-performance target detection. .

Contents of the invention

The technical problem to be solved by the present invention is to provide a radar target detection method based on generalized Pareto distribution clutter statistical modeling to improve radar target detection performance.

The technical solution is to combine the logarithmic maximum likelihood (LML) estimation with the particle swarm optimization (PSO) algorithm, design the cost function through the logarithmic maximum likelihood estimation criterion, and use the particle swarm optimization algorithm to minimize The cost function is transformed to obtain the optimal parameter estimation result; the detection threshold is determined by using the optimal parameter estimation result to complete the radar target detection.

Specific steps are as follows:

The first step is to define and initialize the particle swarm. the way is:

1.1 Define particle swarm G:

Particle swarm is a set composed of a group of particles, where the particles are the parameters to be estimated, particle swarm can be defined as

G={p _i =(σ _i ,k _i ), v _i =(δσ _i ,δk _i ); i=1,2,...,I} (2)

In the formula, I is the number of particles in G, p _i is the position feature of the i-th particle in G, σ _i is the scale parameter of the i-th particle in G, and k _i is the shape parameter of the i-th particle in G ; v _i is the velocity characteristic of the i-th particle in G, δσ _i is the variation of the scale parameter σ _i of the i-th particle in G, and δk _i is the variation of the shape parameter ki of the _i -th particle in G.

1.2 Initialize the particle position feature:

Denote the N clutter data samples observed from the clutter area as z ₁ ,...,z _n ,...,z _N (1≤n≤N, n is an integer), then the N clutter data The sample mean and variance are and In the absence of prior information, the available and Denote the rough estimated values of parameters σ and k respectively, so the position feature of the i-th particle at time 0 is initialized as

In the formula, I is a positive integer, and the size of I is given by experience, generally 200-500 is appropriate; the superscript 0 indicates the initial state of the parameter (that is, time 0), is the position feature of the i-th particle in G at time 0, is the scale parameter of the i-th particle in G at time 0, is the shape parameter of the i-th particle in G at time 0; Represents a uniform distribution function, the lower limit of the interval is a, and the upper limit is b; represents the function from the uniform distribution Randomly select a number, that is, randomly select a number between the interval [a,b]. express is from the uniform distribution function Take random numbers in the middle (that is, in the interval Randomly pick a number between), is the lower limit, is the upper limit; express is from the uniform distribution function random number in is the lower limit, is the upper limit.

1.3 Initialize particle velocity features:

is the velocity characteristic of the i-th particle in G at time 0, is the scale parameter variation of the i-th particle in G at time 0, is the variation of the shape parameter of the i-th particle in G at time 0. express is from the uniform distribution function the number taken, express is from the uniform distribution function The number taken.

In the second step, let the iteration number variable t=0.

The third step is to obtain the cost function of the particle at the tth iteration. the way is:

3.1 Transform the complex exponential distribution of the generalized Pareto distribution model (the situation when k<0) into the logarithmic likelihood function shown in formula (9), namely

In the formula, ln( ) represents the logarithmic function based on the natural constant e, and L(z|k,σ) represents the likelihood function of the clutter signal amplitude z including unknown distribution parameters (k and σ).

It can be seen that the maximum likelihood estimation of parameters k and σ can be obtained by solving the partial derivatives of formula (9), at this time,

Obviously, the explicit expressions about the parameters k and σ cannot be obtained by formula (10) and formula (11), that is, the partial The derivative result is set to zero to obtain the solution of the parameters k and σ, and the present invention uses the particle swarm optimization algorithm to obtain the numerical solution, and a cost function needs to be constructed.

3.2 Construct the cost function shown in formula (12) according to formulas (10) and (11)

In the formula, |·| is the symbol of the absolute value. By minimizing the cost function, T(σ,k) can be continuously approached to 0. This process is equivalent to making the solutions of formula (10) and formula (11) reach the optimum.

3.3 Find the fitness value corresponding to the I particle in G through the formula (12), the method is

Substituting the position characteristics of the I particle in G in the t-th iteration in turn into the formula (12), the cost function value of the I particle in the t-th iteration is obtained as the fitness value of the I particle in the t-th iteration. For the i-th particle in G, the specific method is to use its position feature in the t-th iteration (which is Parameter pair) into formula (12), calculate And let the particle fitness value of the i-th particle in the t-th iteration be

The fitness value of the tth group of I particles obtained by the tth iteration of I particles is expressed as

The fourth step is to calculate the individual extremum pbest ^t and the global extremum gbest ^t of the particle swarm in the tth iteration. the way is:

4.1 According to Find the particle corresponding to the minimum particle fitness value in the t-th iteration, i.e.

In the formula, means to find first The smallest value in , and find the particle corresponding to this smallest value (expressed as a parameter pair (σ′,k′)). (assuming that the serial number of the minimum value is i, the particle found is which is

4.2 According to Find the particle with the smallest fitness value in all iterations from 0 to t times, and use it as the global extremum gbest ^t of the first t iterations, that is,

In the formula, (σ ^* , k ^* ) represents the particle corresponding to the minimum fitness value of the first t iterations.

In the fifth step, let t=t+1.

In the sixth step, according to the obtained individual extremum pbest ^t and global extremum gbest ^t , update the position and velocity characteristics of I particles in G, the method is:

6.1 Let i=1.

6.2 The position feature and velocity feature of the i-th particle are updated according to formula (15) and formula (16):

Among them, w ^t-1 ＝0.9-0.5·(t-1)/t _max is the inertia weight factor for t-1 times, and t _max is the maximum number of iterations (a positive integer, usually 100-200); rand Represents a uniformly distributed random number between [0,1]; c ₁ and c ₂ are learning factors (or acceleration constants), which are positive real numbers and are usually taken as 2.

6.3 Let i=i+1.

6.4 Determine whether i≤I is true, if true, go to 6.2; otherwise, it means that the position and velocity characteristics of I particles in G have been updated, and go to the seventh step.

The seventh step is to judge whether t is equal to the maximum number of iterations t _max , and if so, take the global extremum gbest ^t , that is, (σ ^* , k ^* ) as the final estimation result of the parameter pair (σ, k), and execute the eighth step step; otherwise, go to step three.

The eighth step is to use (σ ^* , k ^* ) for radar target detection, the method is:

8.1 Use (σ ^* , k ^* ) to reconstruct the probability density function corresponding to the generalized Pareto distribution model which is

In the formula, Represents the approximate function about the independent variable z obtained by (σ ^* , k ^* ) reconstruction.

8.2 Given the target detection false alarm rate P _f (P _f usually ranges from 10 ^-4 to 10 ^-2 ), calculate the detection threshold th according to formula (18)

In the formula, the integral formula can be solved by the Newton-Cotes (Newton-Cotes) numerical integral formula. For the specific solution, please refer to literature 3: Ye Qixiao, Shen Yonghuan, etc., "Handbook of Practical Mathematics (2nd Edition)", Science Press, In 2006, 719-720.

8.3 Target detection, the method is:

8.3.1 Obtain J observation data through real-time radar observation, J≥1, let j=1;

8.3.2 Judging whether the jth observation data y _j actually acquired by the radar has a target, the method is:

If y _j ≥ th, output the conclusion of "the jth observation data has a target"; if y _j < th, output the conclusion of "the jth observation data has no target";

8.3.3 Determine whether j is less than J, if it is satisfied, set j=j+1, and go to 8.3.2; otherwise, it means that the radar real-time observation data processing is completed and the target detection is completed.

Beneficial effects of the present invention:

1. the 3rd step of the present invention constructs the estimating formula (formula (9)) of Pareto distribution model parameter by logarithm maximum likelihood function, guarantees that the optimality that parameter to be estimated is solved under maximum likelihood meaning;

2. The fourth step of the present invention utilizes the particle swarm optimization algorithm to have the ability of global optimization, overcomes the sensitivity to the initial value and is easy to converge on the local extremum when the traditional numerical method (such as Newton-Raphson, etc.) solves the model parameter estimation problem It can improve the accuracy and robustness of parameter estimation and ensure the reliability of radar target detection results.

3. The present invention combines the PSO algorithm with the maximum likelihood parameter estimation method of the generalized Pareto distribution model parameters, improves the accuracy of the detection threshold th, and then improves the detection performance of the radar target.

Description of drawings

Fig. 1 is the overall flow chart of the present invention.

Detailed ways

Embodiments of the present invention will be described below in conjunction with drawings and experiments.

Fig. 1 shows the overall flow chart of the present invention, and the specific experimental process will be described below by taking the number of clutter samples N=100 as an example. On this basis, different experimental conditions are set (such as changing the number of clutter samples N and the model shape parameter k, etc.) to investigate the detection performance of the present invention under different experimental conditions.

The first step is to define and initialize the particle swarm. Obtain N=100 clutter data samples from the clutter area, set the particle swarm size I=500, the particle swarm is defined as G={p _i ＝(σ _i ,k _i ),v _i ＝(δσ _i ,δk _i ); i=1,2,…,I} The initial position component of the particle swarm Respectively and Interval random sampling, initial velocity component unified in Random sampling among them, where i=1, 2, ..., I; the maximum number of iterations is t _max =200.

In the second step, let t=0.

The third step is to calculate the particle fitness value in the tth (t≥0) iteration. According to the cost function T(σ,k) in the formula (12), the fitness value corresponding to the I particle in G is calculated, and the fitness value of all I particles in the t-th iteration of the I particle is obtained, and the i-th particle Particle fitness value The fitness value of the tth group of I particles obtained by the tth iteration of I particles is expressed as

The fourth step is to use formula (13) and formula (14) to find the individual extremum pbest ^t and the global extremum gbest ^t of the particle swarm during t iterations, gbest ^t = (σ ^* , k ^* );

In the fifth step, let t=t+1.

The sixth step is to update the velocity and position components of one particle according to formula (15) and formula (16).

The seventh step is to judge whether t reaches the maximum number of iterations t _max . If it is satisfied, output (σ ^* , k ^* ) as the parameter estimation result, and go to the eighth step; otherwise, return to the third step.

The eighth step, radar target detection. First construct the probability density function according to formula (17) On this basis, use the formula (18) to determine the detection threshold th, and finally obtain J observation data (J = 500 in the experiment) through real-time radar observation, and carry out effective analysis of the J observation data according to steps 8.3.2 to 8.3.3. There is no target judgment to complete the target detection.

In order to illustrate the estimation performance of the present invention to the parameters of the generalized Pareto distribution model under different shape parameters and different clutter sample numbers, three typical parameter pairs (σ, k) (corresponding to three groups of different experiments) were selected in the embodiment. Simulation. The parameter estimation accuracy is evaluated by the mean error (Mean Error, ME) and the root mean square error (Root of Mean Square Error, RMSE), the expressions of which are:

In the formula, Indicates the estimate of the parameter υ (υ stands for scale parameter or shape parameter) obtained in the mth simulation, and M indicates the number of Monte Carlo simulations.

Considering that the scale parameter σ has little influence on the estimation results, but the shape parameter k has a greater influence on the generalized Pareto distribution, and the smaller the value of k, the larger the tail of the generalized Pareto distribution. The present invention mainly analyzes and evaluates the estimation effect of described method to shape parameter, sets σ=1 for this reason, and k=-0.3,-0.2,-0.1 (promptly the value of k is taken as-0.3 respectively,-0.2,-0.1 , carry out three groups of different experiments), M=500, the generalized Pareto clutter model target detection method (LML-PSO) based on particle swarm optimization of the present invention is tested, and compare it with traditional methods (such as MoM, PWMM, LMM etc. ) performance difference between. The experiment was carried out on a general-purpose computer platform and realized by using Matlab software. The obtained parameters (σ, k) were tested under the conditions of different clutter data samples (N=25, 50, 100, 500) and different estimation methods (MoM, PWMM, LMM and LML-PSO). The following simulation results are shown in Table 1-Table 3. Among them, Table 1 to Table 3 respectively correspond to the mean error (ME) and root mean square error (RMSE) under the conditions of k=-0.3, k=-0.2 and k=-0.1.

It can be seen from Tables 1 to 3 that the average error and root mean square error of the LML-PSO of the present invention are significantly smaller than those of the other three methods.

Table 1

Table 2

table 3

In order to further illustrate the influence of the estimation result of shape parameter k on target detection, taking k=-0.3 as an example, under the application condition of false alarm probability P _f =0.01, if there is no parameter estimation error, formula (18) can be used to solve The real target detection threshold is th=9.898. When there are different degrees of shape parameter estimation errors (that is, both ME and RMSE are not 0), the corresponding target detection thresholds are shown in Table 4. The error range in the table is -0.12 to 0.12 (the error interval is 0.002), That is, it corresponds to an error range of ±40% (-0.3×40%～0.3×40%) of the real parameter. When the estimation error is -0.12, the target detection threshold is th=14.023; when the estimation error is 0.12, the target detection threshold is th=7.149.

estimation error -0.120 -0.118 -0.116 -0.114 -0.112 -0.110 -0.108 -0.106 -0.104 -0.102 -0.100 -0.098 detection threshold 14.023 13.939 13.856 13.774 13.692 13.61 13.529 13.449 13.369 13.29 13.212 13.134 estimation error -0.096 -0.094 -0.092 -0.090 -0.088 -0.086 -0.084 -0.082 -0.080 -0.078 -0.076 -0.074 detection threshold 13.056 12.979 12.903 12.827 12.751 12.676 12.602 12.528 12.455 12.382 12.31 12.238 estimation error -0.072 -0.070 -0.068 -0.066 -0.064 -0.062 -0.060 -0.058 -0.056 -0.054 -0.052 -0.050 detection threshold 12.166 12.096 12.025 11.955 11.886 11.817 11.749 11.681 11.613 11.546 11.479 11.413 estimation error -0.048 -0.046 -0.044 -0.042 -0.040 -0.038 -0.036 -0.034 -0.032 -0.030 -0.028 -0.026 detection threshold 11.347 11.282 11.217 11.153 11.089 11.026 10.963 10.900 10.838 10.776 10.715 10.654 estimation error -0.024 -0.022 -0.020 -0.018 -0.016 -0.014 -0.012 -0.010 -0.008 -0.006 -0.004 -0.002 detection threshold 10.593 10.533 10.473 10.414 10.355 10.296 10.238 10.181 10.123 10.066 10.01 9.954 estimation error 0.002 0.004 0.006 0.008 0.010 0.012 0.014 0.016 0.018 0.020 0.022 0.024 detection threshold 9.842 9.787 9.733 9.678 9.624 9.571 9.518 9.465 9.412 9.360 9.308 9.257 estimation error 0.026 0.028 0.030 0.032 0.034 0.036 0.038 0.040 0.042 0.044 0.046 0.048 detection threshold 9.206 9.155 9.104 9.054 9.005 8.955 8.906 8.857 8.809 8.761 8.713 8.665 estimation error 0.050 0.052 0.054 0.056 0.058 0.060 0.062 0.064 0.066 0.068 0.070 0.072 detection threshold 8.618 8.571 8.525 8.478 8.433 8.387 8.342 8.297 8.252 8.207 8.163 8.119 estimation error 0.074 0.076 0.078 0.080 0.082 0.084 0.086 0.088 0.090 0.092 0.094 0.096 detection threshold 8.076 8.033 7.990 7.947 7.904 7.862 7.820 7.779 7.738 7.696 7.656 7.615 estimation error 0.098 0.100 0.102 0.104 0.106 0.108 0.110 0.112 0.114 0.116 0.118 0.120 detection threshold 7.575 7.535 7.495 7.456 7.417 7.378 7.339 7.300 7.262 7.224 7.187 7.149

Table 4

As can be seen from the table:

1) Overall, the estimation errors (or absolute values of errors) of various methods decrease with the increase of the number of clutter data samples (N), indicating that the larger the number of clutter data samples, the higher the accuracy of parameter estimation, and the estimation performance the better;

2) In three groups of different experiments, under the same condition of the number of clutter data samples, the method of the present invention has the smallest average error and root mean square error compared with other three estimation methods, indicating that it has consistent Higher parameter estimation accuracy;

3) Through comparison, it can be seen that even when the number of clutter data samples is small (such as N=50), the LML-PSO method still has a high estimation accuracy, which shows that it has good adaptability to small samples.

4) In contrast, the LML-PSO method has the smallest parameter estimation error, indicating that the calculated target detection threshold is the closest to the theoretical value of the detection threshold, and correspondingly the detection performance is also the best.

In summary, compared with the traditional method, the method of the present invention has obvious advantages in accuracy and performance for parameter estimation of the generalized Pareto distribution model, and can well adapt to changes in model parameters, and can effectively improve the performance of long-tail distribution clutter background. Radar target detection performance.

While the invention has been described in detail with reference to the foregoing embodiments, it is to be understood that the invention is not limited to the disclosed embodiments. Various changes in form and details will occur to those skilled in the art. The invention encompasses modifications within the spirit and scope of the appended claims.

Claims (7)

1. a radar target detection method based on generalized Pareto distribution clutter statistical modeling, is characterized in that comprising the following steps: The first step is to define and initialize the particle swarm by: 1.1 Define particle swarm G: Particle swarm is a set composed of a group of particles, that is, parameters to be estimated. Particle swarm is defined as G={p _i =(σ _i ,k _i ), v _i =(δσ _i ,δk _i ); i=1,2,...,I} (2) In the formula, I is the number of particles in G, p _i is the position feature of the i-th particle in G, σ _i is the scale parameter of the i-th particle in G, and k _i is the shape parameter of the i-th particle in G ; v _i is the velocity characteristic of the i-th particle in G, δσ _i is the variation of the scale parameter σ _i of the i-th particle in G, and δk _i is the variation of the shape parameter k _i of the i-th particle in G; 1.2 Initialize the particle position feature: Denote the N clutter data samples observed from the clutter area as z ₁ ,...,z _n ,...,z _N , 1≤n≤N, n is an integer, then the N clutter data samples The mean and variance of are respectively and In the absence of prior information, the and Represent the rough estimated values of the parameters σ and k, respectively, k represents the shape parameter, σ represents the scale parameter, σ>0, the position feature of the i-th particle at time 0 is initialized as In the formula, I is a positive integer; the superscript 0 indicates the initial state of the parameter, that is, time 0, is the position feature of the i-th particle in G at time 0, is the scale parameter of the i-th particle in G at time 0, is the shape parameter of the i-th particle in G at time 0; Represents a uniform distribution function, the lower limit of the interval is a, and the upper limit is b; " ” means from the uniform distribution function Randomly select a number, that is, randomly select a number between the interval [a,b]; express is from the uniform distribution function random number in the interval, that is, in the interval Take a random number between is the lower limit, is the upper limit; express is from the uniform distribution function random number in is the lower limit, is the upper limit; 1.3 Initialize particle velocity features: is the velocity characteristic of the i-th particle in G at time 0, is the scale parameter variation of the i-th particle in G at time 0, is the shape parameter variation of the i-th particle in G at time 0, express is from the uniform distribution function the number taken, express is from the uniform distribution function the number taken; In the second step, let the iteration number variable t=0; The third step is to obtain the cost function of the particle at the t-th iteration, the method is: 3.1 Transform the complex exponential distribution of the generalized Pareto distribution model into the logarithmic likelihood function shown in formula (9), k≠0, that is In the formula, ln( ) represents the logarithmic function with the natural constant e as the base, and L(z|k,σ) represents the likelihood function of the clutter signal amplitude z including the unknown distribution parameters k and σ; The maximum likelihood estimation of parameters k and σ is obtained by solving the partial derivative of formula (9), at this time, we have 3.2 Construct the cost function shown in formula (12) according to formulas (10) and (11): In the formula, |·| is the sign of the absolute value. By minimizing the cost function, T(σ,k) is continuously approaching 0. This process is equivalent to making the solutions of formula (10) and formula (11) reach optimal; 3.3 Calculate the fitness value corresponding to the I particle in G through the formula (12). The fitness value of the tth group of the I particle obtained by the t iteration of the I particle is expressed as The fourth step is to calculate the individual extremum pbest ^t and the global extremum gbest ^t of the particle swarm in the t-th iteration, the method is: 4.1 According to Find the particle corresponding to the minimum particle fitness value in the t-th iteration, i.e. In the formula, means to find first The minimum value in , and find the particle corresponding to this minimum value, expressed as a parameter pair (σ′,k′); 4.2 According to Find the particle with the smallest fitness value in all iterations from 0 to t times, and use it as the global extremum gbest ^t of the first t iterations, that is, (σ ^* , k ^* ) is the particle corresponding to the minimum fitness value of the previous t times; The fifth step, let t=t+1; In the sixth step, according to the obtained individual extremum pbest ^t and global extremum gbest ^t , update the position and velocity characteristics of I particles in G; The seventh step is to judge whether t is equal to the maximum number of iterations t _max , if it is satisfied, take the global extremum gbest ^t , that is, (σ ^* , k ^* ) as the final estimation result of the parameter pair (σ, k), and execute the eighth step ; Otherwise, go to the third step; The eighth step is to use (σ ^* , k ^* ) for radar target detection, the method is: 8.1 Use (σ ^* , k ^* ) to reconstruct the probability density function corresponding to the generalized Pareto distribution model which is In the formula, Represents the approximate function about the independent variable z obtained by (σ ^* , k ^* ) reconstruction; 8.2 Given the target detection false alarm rate P _f , calculate the detection threshold th according to the formula (18) 8.3 Target detection, the method is: 8.3.1 Obtain J observation data through real-time radar observation, J≥1, let j=1; 8.3.2 Judging whether the jth observation data y _j actually acquired by the radar has a target, the method is: If y _j ≥ th, output the conclusion of "the jth observation data has a target"; if y _j < th, output the conclusion of "the jth observation data has no target"; 8.3.3 Determine whether j is less than J, if it is satisfied, set j=j+1, and go to 8.3.2; otherwise, it means that the radar real-time observation data processing is completed and the target detection is completed. 2. the radar target detection method based on generalized Pareto distribution clutter statistical modeling as claimed in claim 1, is characterized in that described I gets 200～500. 3. the radar target detection method based on generalized Pareto distribution clutter statistical modeling as claimed in claim 1, it is characterized in that the method for asking for the adaptability value corresponding to I particle in G described in 3.3 steps is: sequentially the first In the t iteration, the position characteristics of the I particle in G are substituted into formula (12), and the cost function value of the I particle in the t iteration is obtained as the fitness value of the I particle in the t iteration; for the i particle in G Particles, the specific method is to feature its position in the tth iteration which is Substituting the parameter pairs into formula (12), calculate And let the particle fitness value of the i-th particle in the t-th iteration be 4. the radar target detection method based on generalized Pareto distribution clutter statistical modeling as claimed in claim 1, it is characterized in that the method for the position characteristic and the speed characteristic of I particle in the described update G of the 6th step is: 6.1 let i=1; 6.2 The position feature and velocity feature of the i-th particle are updated according to formula (15) and formula (16): Among them, w ^t-1 ＝0.9-0.5·(t-1)/t _max is the inertia weight factor for t-1 times, t _max is the maximum number of iterations, which is a positive integer; rand represents the value between [0,1] Uniformly distributed random numbers between; c ₁ and c ₂ are learning factors, which are positive integers; 6.3 Let i=i+1; 6.4 Determine whether i≤I is true, if it is true, go to 6.2; otherwise, it means that the position characteristics and velocity characteristics of I particles in G have been updated, and end. 5. The radar target detection method based on generalized Pareto distribution clutter statistical modeling as claimed in claim 4, characterized in that said t _max is 100-200; c ₁ and c ₂ are both 2. 6 . The radar target detection method based on generalized Pareto distribution clutter statistical modeling according to claim 1 , wherein the P _f is 10 ⁻⁴ to 10 ⁻² . 7. the radar target detection method based on generalized Pareto distribution clutter statistical modeling as claimed in claim 1, it is characterized in that described formula (18) asks detection threshold th by Newton-Cotes namely Newton-Cotes numerical integral formula to solve.