Doppler radar ground clutter false target discrimination method based on multi-feature prediction

Abstract

The invention relates to the field of radar target detection, in particular to a Doppler radar ground clutter false target discrimination method based on multi-feature prediction. Firstly, effective feature parameters are extracted according to the difference between a target and clutter echo signals, several optimal classification features are obtained based on all 10 features in the first training, and after the optimal features are obtained in the second training, a database is extracted according to the optimal features and trained to obtain a k value. According to the method, the optimal feature is self-adaptive, k is also self-adaptive, the number of used features and the number of used group number comparison are self-adaptive, and the method can automatically optimize parameters and extract the optimal discrimination method according to the collected data, so that the capability of identifying weak targets and complex environment targets is improved.

Description

Technical Field

[0001] The invention relates to the field of radar target detection, and in particular to a Doppler radar ground clutter false target discrimination method based on multi-feature prediction. Background Art

[0002] The feature detection method is an important means to solve the problem of radar target detection. The feature information of the echo signal can be used to determine whether there is a target in the area. In the radar's detection airspace, both moving targets (cars, ships, birds, airplanes, etc.) and fixed targets (buildings, billboards, air conditioners, etc.) may reflect and form echoes.

[0003] In the radar echo, there is a temporal correlation between the historical frame data and the current frame data, and the historical data can provide rich prior information for the current data. However, in the prior art processing, the historical data is not associated with each other. For example, the Chinese invention patent "CN117607818A-A method for generating a clutter map update coefficient and a method for generating a clutter map" discloses the steps of calculating the optimal update times of the frequency domain channel and calculating the update coefficient of the clutter map corresponding to the optimal update times of the frequency domain channel. The present invention solves the problem that the current clutter map update speed and update effect cannot be taken into account at the same time through the optimized clutter map update coefficient, and is also applicable to radars of other working types and frequency bands. The invention has further improved the speed of establishing the clutter map. However, due to the lack of association with historical data, its weak target recognition ability has not been improved. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a Doppler radar ground clutter false target discrimination method based on multi-feature prediction. The present invention is equivalent to having two training processes. The first training is based on all 10 features to obtain the best classification features. The second training is to extract the database according to the best features after the best features are obtained and train to obtain the k value. The present invention applies locust optimization KNN technology to radar classification, improves the classification and recognition capabilities of radar-like methods, especially for weak target classification and recognition, with significant performance improvement.

[0005] The technical solution of the present invention is: a Doppler radar ground clutter false target discrimination method based on multi-feature prediction, comprising the following steps:

[0006] Step 1: Divide different resolution units according to distance and angle;

[0007] Step 2: Based on historical data, the false alarm and target data features are counted, false alarm and target marks are made, and a database is established. The data features constitute a data feature set;

[0008] Step 3: construct a feature standard library, extract historical data clutter and target sequences, extract data feature sets to form classification samples, and perform 0-1 standardization on the classification samples;

[0009] Step 4: Select the optimal feature subset based on the locust optimization algorithm; specifically, the following steps are included:

[0010] Step 41, the classified samples are divided into a training data set and a test set;

[0011] Step 42, initialize the maximum number of iterations Times, the population size size, the variable range, and the maximum and minimum parameters of the control parameters;

[0012] Step 43, initializing the position of the locust colony, calculating the initial individual fitness, and obtaining the optimal locust position and fitness;

[0013] Step 44, calculating the current best fitness;

[0014]

[0015] Where X i is the position of the first locust, X j is the position of the Jth locust, c is the decreasing coefficient, N is the number of locusts, T d is the best solution of locust position in d-dimensional space so far, s is social power,

[0016] in:

[0017] f is the attraction strength; l is the attraction length scale; f and l are set values, ub d and lb d are the upper and lower bounds of r of the social force s function; r is the distance between locusts;

[0018] Step 45, determine whether the number of iterations has been reached, if so, the program ends, and the current position of the global optimal solution is the optimal classification feature subset;

[0019] Step 46, updating the decreasing coefficient c;

[0020]

[0021] In this formula, iter is the current iteration number, c max is the maximum value of the decreasing coefficient, c min The decreasing coefficient is the minimum value;

[0022] Step 47, standardize the distance between locusts, and the normalization interval is [1, 4];

[0023] The distance between locusts is calculated as: dij =|x j -x i | is the distance r between the Ith locust and the Jth locust, and then the distance r is normalized;

[0024] Step 48, updating the position of the locust, calculating the updated locust fitness, and comparing it with the historical best fitness. If the fitness is better than the historical best, then update it, otherwise do not update it;

[0025] Step 49, update the number of iterations, and return to step 44;

[0026] Step 5: Use KNN classifier to construct the best classifier parameters; specifically, the following steps are included:

[0027] Step 51, the samples composed of the best classification feature subset are used as feature subset classification samples, and the feature subset classification samples are divided into a training data set and a test set;

[0028] Step 52, initialize the distance to the maximum value, and the distance adopts Manhattan distance;

[0029] The distance measurement method uses Manhattan distance, and the d-dimensional space point a(x 11 , x 12 , x 13 , …x 1n ) and b(x 21 x 22 x 23 …x 2n ) has a Manhattan distance of

[0030]

[0031] X 1i , X 2i are the d features of the I-th locust that are less than or equal to 10,

[0032] Step 53, calculating the distance dist between the unknown sample and each training sample;

[0033] Step 54, obtain the maximum distance maxdist among the current k nearest samples;

[0034] Step 55: If dist is less than maxdist, the training sample is used as a K-nearest neighbor sample;

[0035] Step 56: Repeat steps 2, 3, and 4 until the distances between all unknown samples and all training samples are calculated;

[0036] Step 57, counting the number of occurrences of each class label in the K-nearest neighbor samples;

[0037] Step 58, select the class label with the largest occurrence frequency as the best classifier parameter k;

[0038] Step 6 accumulates the radar sampling point data over time, calculates the best classification features of the echo signals of the same grid unit for 10 consecutive frames based on the locust optimization algorithm in step 4, adopts the best classifier parameter k in step 5, uses the KNN classifier for prediction and classification, and filters out clutter false alarms.

[0039] According to the Doppler radar ground clutter false target discrimination method based on multi-feature prediction as described above, it is characterized in that the data feature set includes 10 data features.

[0040] According to the Doppler radar ground clutter false target discrimination method based on multi-feature prediction as described above, it is characterized by: the 10 data features are average signal-to-clutter ratio, margin factor, relative energy peak height, relative Doppler spectrum variation coefficient, peak factor, Doppler relative velocity variance, echo distance, number of echo data sampling points, Doppler relative velocity mean, and relative energy concentration.

[0041] According to the Doppler radar ground clutter false target discrimination method based on multi-feature prediction as described above, it is characterized in that: in step 41, 90% of the classified samples are used as training data sets, and 10% are used as test sets.

[0042] According to the Doppler radar ground clutter false target discrimination method based on multi-feature prediction as described above, it is characterized in that: in step 51, 90% of the feature subset classification samples are used as a training data set, and 10% are used as a test set.

[0043] The beneficial effects of the present invention are: improving the ability to identify weak targets and targets in complex environments, especially being able to quickly identify weak targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A block diagram of the invented algorithm.

[0045] Figure 2 Block diagram of the optimization algorithm for Locust. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0047] like Figure 1 As shown, a Doppler radar ground clutter false target discrimination method based on multi-feature prediction of the present invention comprises the following steps:

[0048] Step 1: Divide the resolution units into different units according to distance and angle. First, divide the distance into M units (e.g., one distance unit covers 30m). Each distance unit is divided into N resolution units according to azimuth (e.g., one azimuth unit is 5 degrees). The number of the jth azimuth unit of the i-th distance unit is r. ij .

[0049] Step 2 is based on historical data, and 10 data features of false alarm and target are counted, and the mark of false alarm and target is made, and database is established, 10 data features constitute data feature set, but not all features are useful to target classification. Redundant and irrelevant features may reduce the performance of classifier, increase computational complexity. In optimization algorithm, optimization algorithm adaptively selects A features (A value is undetermined, i.e., in actual work, multiple data features can be selected arbitrarily according to the situation) that are used to distinguish target and clutter in real time, the features of different targets may be different, and optimization algorithm is started in time-sharing, and the feature number and feature data of selection all can change, and 10 features are counted respectively from time (relative energy peak height, peak factor etc.) frequency (relative Doppler spectrum variation coefficient etc.) domain, so that method of the present invention comprises time domain feature and frequency domain feature, thereby improves accurate recognition ability of the present invention.

[0050] The 10 data features of the present invention include average signal-to-noise ratio, margin factor which is the ratio of signal peak value to root mean square amplitude, relative energy peak height, relative Doppler spectrum variation coefficient, peak factor (margin factor is the ratio of signal peak value to root mean square amplitude. Similar to the peak factor, the root mean square amplitude and the root mean square value (effective value) correspond to each other, and the formula of the root mean square is the arithmetic square root of the average value of the sum of signal squares, and the root mean square amplitude is the square of the average value of the arithmetic square roots.), Doppler relative velocity variance, echo distance, number of echo data sampling points, Doppler relative velocity mean, and relative energy concentration parameter. The parameters of the present invention are known in the art, and descriptions of some parameters are given below.

[0051] Parameter 1, average signal-to-noise ratio: Take n frames of data and calculate r using the formula ij The clutter amplitude of the unit

[0052] PowA=E(r ij )+δ(r ij )

[0053] Among them, E is the mean operation, δ is the standard deviation operation,

[0054]

[0055] Where X(m) is r ij The radar echo amplitude of the unit.

[0056] Parameter 2, margin factor is the ratio of the signal peak value to the root mean square amplitude, and the calculation formula is as follows

[0057]

[0058] Parameter 3: Relative energy peak height

[0059]

[0060] Parameter 4: Relative Doppler spectrum variation coefficient

[0061]

[0062] Where X m is the echo data amplitude of each distance unit, R is the number of surrounding units, generally 5, CV(r) refers to the Doppler spectrum variation coefficient of the 5 nearby units, CV(n) refers to the Doppler spectrum variation coefficient of the current unit, and RCV(n) refers to the relative Doppler spectrum variation coefficient.

[0063] Parameter 5, peak factor: The peak factor is the ratio of the signal peak to the effective value (RMS, root mean square value), which represents the extreme degree of the peak in the waveform.

[0064] Step 3: Feature standard library construction

[0065] Extract the historical data clutter and target sequence, extract 10 features of the data, form classification samples, and perform 0-1 standardization on the classification samples. Each set of features has a known corresponding label. The classification sample is the standard library, the classification standard of the KNN algorithm. The data features calculated from the newly extracted unlabeled data are k-nearest calculated with the data from the standard library. In the locust optimization algorithm, the sample composed of 10 features, 90% of the classification samples are used as training data sets, and 10% are used as test sets to select the best classification feature subset. In KNN, similarly, the samples composed of the best classification feature subset (less than or equal to 10, KNN training samples) are used as feature subset classification samples, 90% of the feature subset classification samples are used as training data sets, and 10% are used as test sets to determine the best k value.

[0066] Step 4: Select the optimal feature subset based on the Grasshopper Optimization Algorithm (GOA algorithm).

[0067] The KNN classifier of the locust optimization algorithm is constructed, and the locust optimization algorithm is used to select the best classification feature subset. The best feature subset is a subset of 10 features. According to the minimum classification error standard, the classification error no longer changes after 100 iterations. The feature set corresponding to this error is the best classification feature subset. As the best classification standard, the classification error is controllable. The classification error is related to the size of the feature database and can be as low as 0.02%. In radar, the classification features of strong targets and weak targets are different. The locust optimization algorithm is required to select the best classification features. The number of input parameters changes over time and is not fixed. This is a characteristic of machine learning algorithms. The best classification features are selected adaptively. Locust optimization is an intelligent optimization algorithm with high search efficiency and fast convergence speed. The algorithm's special adaptive mechanism can well balance the global and local search processes and has good optimization accuracy.

[0068] Step 41: 90% of the classified samples are used as training data sets and 10% as test sets

[0069] Step 42: Initialization: Initialize the maximum number of iterations Times, population size size, variable range, maximum and minimum values ​​of control parameters and other parameters.

[0070] If L is the maximum number of iterations, we take 100, and iter is the current number of iterations;

[0071] In this example, the number of iterations is set to 100, the population size is 10-30, the maximum value of the update parameter c is 1, and the minimum value is 0.00004.

[0072] Step 43, initialize the position (feature subset) of the locust population, calculate the initial individual fitness, and obtain the optimal locust position and fitness. The locust population is N, and the position of each locust is X = (x1, x2, x3, ..., xd), d < = 10; d is randomly selected, and the fitness function is F, which is the classification accuracy;

[0073] Step 44, calculating the current best fitness;

[0074]

[0075] Where X i is the position of the first locust, X j is the position of the Jth locust, c is the decreasing coefficient, c max The maximum value is 1, c min The minimum value is 0.00004, N is the number of locusts, T d is the best solution for the locusts' positions in d-dimensional space so far, s is the social force, which indicates mutual repulsion when it is negative and mutual attraction when it is positive.

[0076] in:

[0077] f is the attraction strength; l is the attraction length scale; f and l are set values. In this embodiment, l=1.2, f=0.8, ub d and lb d is the upper and lower bounds of r of the social force s function; r is the distance between locusts. In this embodiment, ub d and lb d Take them as 5 and 1 respectively.

[0078] Step 45, determine whether the number of iterations has been reached. If reached, the program ends, and the current position of the global optimal solution is the best classification feature subset; the best classification feature subset is a subset of the 10 features.

[0079] Step 46, updating the decreasing coefficient c;

[0080]

[0081] In this formula, iter is the current iteration number, c max is the maximum value of the decreasing coefficient, c min The decreasing coefficient is the minimum value;

[0082] Step 47, standardize the distance between locusts, and the normalization interval is [1, 4];

[0083] The distance between locusts is calculated as: d ij =|x j -x i | is the distance r between the Ith locust and the Jth locust, and then the distance r is normalized.

[0084] Step 48, updating the position of the locust, calculating the updated locust fitness, and comparing it with the historical best fitness. If the fitness is better than the historical best, then update it, otherwise do not update it;

[0085] Step 49, update the number of iterations, and return to step 44;

[0086] Step 5: KNN classifier construction

[0087] The best classifier parameter k is obtained based on historical data. According to the selected best feature subset, the data corresponding to the effective features are used as historical data. Each set of data has a corresponding label. Generally, the cross-validation method is used to select the k value that best suits the data. In cross-validation, the historical data set is divided into several subsets. The model is repeatedly trained on these subsets, and the k value that gives the model the highest validation accuracy is selected. The specific steps include the following.

[0088] Step 51: Samples composed of the best classification feature subset (less than or equal to 10, KNN training samples) are used as feature subset classification samples, 90% of the feature subset classification samples are used as training data sets, and 10% are used as test sets to obtain the optimal k value.

[0089] Step 52, initialize the distance to the maximum value, and the distance adopts Manhattan distance;

[0090] The distance measurement method uses Manhattan distance, and the d-dimensional space point a(x 11 , x 12 , x 13 , …x 1n ) and b(x 21 x 22 x 23 …x 2n ) is (d is random, n is the maximum value of d)

[0091]

[0092] X 1i , X 2i are the d features of the I-th locust that are less than or equal to 10,

[0093] Step 53, calculating the distance dist between the unknown sample and each training sample;

[0094] Step 54, obtain the maximum distance maxdist among the current k nearest samples;

[0095] Step 55: If dist is less than maxdist, the training sample is used as a K-nearest neighbor sample;

[0096] Step 56: Repeat steps 2, 3, and 4 until the distances between all unknown samples and all training samples are calculated;

[0097] Step 57, counting the number of occurrences of each class label in the K-nearest neighbor samples;

[0098] Step 58: Select the class label with the highest frequency as the class label of the unknown sample, that is, the class label with the highest frequency is the best classifier parameter k.

[0099] The optimal classifier parameter k is the number of k groups of unlabeled new data that are closest to the training data. According to the principle of minority obeys majority, the label corresponding to the majority data is the new data label.

[0100] Step 6 accumulates the radar sampling point data over time. The sampling points of one frame of radar data are limited and have no obvious features. The effective features of the echo signal can be obtained by using the accumulation method. The best classification features calculated based on the locust optimization algorithm in step 4 for echo signals of the same grid unit for 10 consecutive frames are calculated. The best classifier parameter k in step 5 is used to perform prediction classification using the KNN classifier to filter out clutter false alarms. The best features of the present invention are adaptive, and k is also adaptive. The number of features and the number of groups used for comparison are all adaptive. The method of the present invention can automatically optimize parameters according to the collected data and extract the optimal discrimination method, thereby improving the ability to identify weak targets and targets in complex environments, and can quickly identify weak targets, such as the recognition probability of weak targets with an average signal-to-noise ratio of -30dB and a relative energy peak height of -40dB is 80%.

[0101] The present invention starts from the point of view of point trace recognition, establishes a ground clutter false alarm training set, first extracts effective feature parameters based on the difference between the target and the clutter echo signal, and only the distance information of the target can be extracted using the feature clutter map method. For small targets, it is also necessary to extract the target angle and speed information. The present invention can identify weak targets and targets in complex environments, and can quickly identify targets.

Claims

1. A Doppler radar ground clutter false target discrimination method based on multi-feature prediction, comprising the following steps: Step 1: Divide different resolution units according to distance and angle; Step 2: Based on historical data, the false alarm and target data features are counted, false alarm and target marks are made, and a database is established. The data features constitute a data feature set; Step 3: construct a feature standard library, extract historical data clutter and target sequences, extract data feature sets to form classification samples, and perform 0-1 standardization on the classification samples; Step 4: Select the optimal feature subset based on the locust optimization algorithm; The specific steps include: Step 41, the classified samples are divided into a training data set and a test set; Step 42, initialize the maximum number of iterations Times, the population size size, the variable range, and the maximum and minimum parameters of the control parameters; Step 43, initializing the position of the locust colony, calculating the initial individual fitness, and obtaining the optimal locust position and fitness; Step 44, calculating the current best fitness; Where X I is the position of the first locust, X J is the position of the Jth locust, c is the decreasing coefficient, N is the number of locusts, T d is the best solution of locust position in d-dimensional space so far, s is social power, in: f is the attraction strength; l is the attraction length scale; f and l are set values, ub d and lb d are the upper and lower bounds of r of the social force s function; r is the distance between locusts; Step 45, determine whether the number of iterations has been reached, if so, the program ends, and the current position of the global optimal solution is the optimal classification feature subset; Step 46, updating the decreasing coefficient c; In this formula, iter is the current iteration number, c max is the maximum value of the decreasing coefficient, c min The decreasing coefficient is the minimum value; Step 47, standardize the distance between locusts, and the normalization interval is [1, 4]; The distance between locusts is calculated as: d IJ =|x I -x J | is the distance r between the Ith locust and the Jth locust, and then the distance r is normalized; Step 48, updating the position of the locust, calculating the updated locust fitness, and comparing it with the historical best fitness. If the fitness is better than the historical best, then update it, otherwise do not update it; Step 49, update the number of iterations, and return to step 44; Step 5: Use KNN classifier to construct the best classifier parameters; specifically, the following steps are included: Step 51, the samples composed of the best classification feature subset are used as feature subset classification samples, and the feature subset classification samples are divided into a training data set and a test set; Step 52, initialize the distance to the maximum value, and the distance adopts Manhattan distance; The distance measurement method uses Manhattan distance, and the d-dimensional space point a(x 11 , x 12 , x 13 , …x 1n ) and b(x 21 x 22 x 23 …x 2n ) has a Manhattan distance of X 1i , X 2i are the d features of the I-th locust that are less than or equal to 10, Step 53, calculating the distance dist between the unknown sample and each training sample; Step 54, obtain the maximum distance maxdist among the current k nearest samples; Step 55: If dist is less than maxdist, the training sample is used as a K-nearest neighbor sample; Step 56: Repeat steps 2, 3, and 4 until the distances between all unknown samples and all training samples are calculated; Step 57, counting the number of occurrences of each class label in the K-nearest neighbor samples; Step 58, select the class label with the largest occurrence frequency as the best classifier parameter k; Step 6 accumulates the radar sampling point data over time, calculates the best classification features of the echo signals of the same grid unit for 10 consecutive frames based on the locust optimization algorithm in step 4, adopts the best classifier parameter k in step 5, uses the KNN classifier for prediction and classification, and filters out clutter false alarms.

2. The Doppler radar ground clutter false target discrimination method based on multi-feature prediction according to claim 1, characterized in that: The data feature set includes 10 data features.

3. The Doppler radar ground clutter false target discrimination method based on multi-feature prediction according to claim 2 is characterized in that: The 10 data features are average signal-to-noise ratio, margin factor, relative energy peak height, relative Doppler spectrum variation coefficient, peak factor, Doppler relative velocity variance, echo distance, number of echo data sampling points, Doppler relative velocity mean, and relative energy concentration.

4. A Doppler radar ground clutter false target discrimination method based on multi-feature prediction according to claim 1 or 2, characterized in that: In step 41, 90% of the classified samples are used as a training data set, and 10% are used as a test set.

5. The Doppler radar ground clutter false target identification method based on multi-feature prediction according to claim 1 or 2, characterized in that: In step 51, 90% of the feature subset classification samples are used as a training data set, and 10% are used as a test set.

Images