CN115079103A - Multi-domain feature and LightGBM-based foil strip interference resisting method - Google Patents

Abstract

The invention belongs to the technical field of electronic countermeasure, data classification and signal identification, and particularly relates to an anti-foil strip interference method based on multi-domain features and LightGBM. The method comprises the following steps: preprocessing target echo data under foil strip interference of a horizontal polarization receiving channel of a broadband polarization radar to obtain a range profile sequence of the horizontal polarization receiving channel; judging the type of foil interference and segmenting a target to obtain a target distance image sequence and a foil interference distance image sequence; extracting and obtaining a characteristic vector from a time domain, a frequency domain and a polarization domain, and taking the characteristic vector obtained by each sequence as a sample; labeling the sample; carrying out data division on the sample with the label to obtain a training set and a testing set; and training and testing are carried out, and identification and countermeasure of foil strip interference are realized. The method has the advantages of good effect of resisting the interference of the shielding type foil strips, high classification accuracy, good robustness to sea clutter and high anti-interference efficiency.

Description

An anti-chaf interference method based on multi-domain features and LightGBM

technical field

The invention belongs to the technical fields of electronic countermeasures, data classification and signal identification, and in particular relates to an anti-chaf interference method based on multi-domain features and LightGBM.

Background technique

As one of the main methods of terminal guidance, radar active homing guidance has the advantages of long range, high detection accuracy, all-day and all-weather, etc., and has been widely used. However, the radar seeker also has the disadvantage of being easily affected by the electromagnetic environment of the battlefield because the active radar guidance needs to complete the functions of target detection, identification, positioning and tracking by emitting electromagnetic waves.

The chaff jamming is one of the most typical passive jammers against active radar seekers. The jamming scatters the radar electromagnetic waves by projecting a large number of chaff fibers, resulting in a strong return with clutter characteristics in the radar receiver. It can achieve the purpose of electromagnetic deception or suppression of radar, so as to effectively protect its own platform or other targets of its own, and play an extremely important role in electronic countermeasures. With the optimization of the chaff material, the progress of the manufacturing process and the improvement of the delivery equipment, the chaff interference has the advantages of simple manufacture, low cost, convenient use, easy access to wide-band characteristics, and the ability to interfere with different systems, different frequencies, and different directions at the same time. Radar and other advantages are one of the important threats to radar seekers. Therefore, how to improve the anti-chaf interference capability of the radar seeker is still a hot and difficult issue at home and abroad.

According to the projecting purpose of chaff bullets, chaff interference is mainly divided into shielded chaff interference and diluted or centroid chaff interference. At present, chaff anti-jamming methods mainly include: identification method based on signal characteristics, filter-based chaff anti-jamming method and composite guidance technology, but most of the chaff anti-jamming methods are aimed at diluted or centroid chaff interference ( Diluted chaff, centroid chaff, etc.), and the proposed feature is only a single transform domain.

The existing dilute chaff countermeasures based on polarization domain features first estimate the polarization scattering matrices of the target and the chaff false target, and then define and calculate the polarization scattering parameters of the target and the chaff false target respectively; then The absolute value of the correlation between co-polarized and cross-polarized channels is calculated. Finally, the classification and identification of chaff false targets, targets, aircraft and other radar targets are realized by support vector machine. The algorithm has a certain anti-interference effect, but the disadvantage is that it is difficult to obtain the polarization scattering parameters and cannot resist the interference of the shielded chaff.

In addition, there is a method for identifying the interference of chaff false targets based on the characteristics of the polarization domain. This method, firstly, defines and calculates the polarization ratio average value of the horizontal channel and the vertical channel polarization ratio of the real target and the chaff fake target respectively; then the average value of the smaller polarization ratio of multiple pulses is taken as the characteristic parameter ; Finally, the characteristic parameters of the true and false targets are compared, and the target with a larger value is judged as the target. Although the computational complexity of the features of this method is low, the environmental adaptability of the method is poor, and the stability of the feature parameters is poor when the sea clutter is large and the true and false target echoes are mixed.

There is also a method for distinguishing chaff false target jamming based on polarization domain features. Firstly, the average polarization ratio of the horizontal channel and the polarization ratio of the vertical channel of the real target and the chaff fake target are defined and calculated respectively; then the average value of the smaller polarization ratio of multiple pulses is taken as the characteristic parameter; finally, the The characteristic parameters of the true and false targets are compared, and the threshold is used for classification, and the target larger than the threshold is judged as the real target. Although the computational complexity of the features of this method is low, the environmental adaptability of the method is poor, and the stability of the feature parameters is poor when the sea clutter is large and the true and false target echoes are mixed.

In addition, the waveform entropy, correlation coefficient and scattering intensity ratio features are extracted based on the distance image, and the chaff interference is identified by the FCM algorithm and cluster analysis. Although this method has high accuracy, it does not take into account the overlapping of the distance images of false chaff targets and real targets, so it is not effective against masked chaff interference.

Therefore, based on the aforementioned prior art, it can be seen that the echo data of the shielded chaff interference is complex, difficult to classify according to the features of a single transform domain, and the recognition rate is low. Efficient identification of chaff interference is achieved.

SUMMARY OF THE INVENTION

The purpose of the present invention is to propose an anti-chaf interference method based on multi-domain features and LightGBM, aiming at the defects of complex echo data of the target under shielded chaff interference, difficulty in classifying according to the features of a single transform domain, and low recognition rate , the method not only has high accuracy, but also can resist a variety of chaff interference including diluted or centroid chaff interference and masked chaff interference.

In order to achieve the above object, the present invention adopts the following technical solutions:

The method for resisting chaff interference includes the following steps:

S1. The preprocessing module preprocesses the target echo data under the chaff interference of the horizontally polarized receiving channel of the broadband polarized radar to obtain the range image sequence of the horizontally polarized receiving channel;

The preprocessing includes pulse compression, target distance image alignment and normalization operations;

S2, the target segmentation module performs chaff interference type judgment and target segmentation on the range image sequence of the horizontally polarized receiving channel obtained in S1, and obtains the target range image sequence and the chaff interference range image sequence;

S2 specifically includes the following sub-steps:

S21. Calculate the average power of the range image sequence of the horizontally polarized receiving channel, and set a threshold to determine the type of chaff interference. If the power is less than the threshold, it is a diluted or centroid type of chaff interference, otherwise, it is a shielded type of chaff interference;

S22. When the type of the chaff interference is diluting or centroid chaff interference, detect the range image sequence of the horizontally polarized receiving channel through CA-CFAR, and perform target segmentation on the over-detected signal. Segmentation, specifically, taking the midpoint of the range image of each over-detected target as the center and taking n range resolution units as the range image sequence of the target, and obtaining the target range image sequence and the chaff interference range image sequence; when the chaff When the interference type is masked chaff interference, if the target distance is unknown, divide a target every n range resolution units of the range image, otherwise the range image sequence is centered on the range resolution unit of the known target. The range resolution unit is used as the target range image sequence, and n distance resolution units are taken as the chaff interference range image sequence in the remaining areas without target information, and the target range image sequence and the chaff interference range image sequence are obtained;

S3. The feature extraction module extracts the range image concentration, energy ratio, energy mean, energy variance, waveform width and average similarity from the time domain, respectively, from the target range image sequence and the chaff interference range image sequence obtained in S2, from the frequency domain. Extract spectrum broadening, frequency domain skewness, frequency domain kurtosis, power spectrum envelope fluctuation and power spectrum Shannon entropy, extract polarization ratio mean, polarization ratio variance and polarization similarity from polarization domain, and combine distance images Concentration, energy ratio, energy mean, energy variance, waveform width, average similarity, spectrum broadening, frequency domain skewness, frequency domain kurtosis, power spectrum envelope fluctuation, power spectrum Shannon entropy, polarization ratio mean, The polarization ratio variance and polarization similarity are used to obtain the eigenvectors corresponding to all target distance image sequences and chaff interference distance image sequences, and the eigenvectors obtained from each sequence are used as a sample to obtain all samples;

Specifically, it includes the following sub-steps:

S31, extracting the distance image concentration, energy ratio, energy mean, energy variance, waveform width and average similarity from the time domain, specifically:

S311, extracting a range image of a pulse of target and chaff interference from the target range image sequence and the chaff interference range image sequence obtained in S2;

S312, according to the range image of a pulse interfered by the target and chaff obtained in S311, extract the range image concentration, energy ratio, energy mean, energy variance, waveform width and average similarity from the time domain;

The extraction of the range image concentration from the time domain is specifically: sorting the amplitudes of all the range resolution units of the range image of a pulse interfered by the target and the chaff, and based on the range resolution unit amplitude sequence sorted from large to small Calculate the ratio of its upper quartile to its peak value to get the distance image concentration;

The extraction of the energy ratio from the time domain is specifically: sorting the amplitudes of all the range resolution cells of the range image of a pulse interfered by the target and the chaff, and calculating based on the range resolution cell magnitude sequence sorted from large to small. The ratio of the sum of the amplitudes of the first m distance resolution units to the sum of the amplitudes of all n distance resolution units, the energy ratio is obtained, and m is less than n;

The extracting the average energy value from the time domain is specifically: according to the radar cross section of the chaff cloud is generally larger than the radar cross section of the target, calculating the average value of the amplitudes of the range images of the two to obtain the average energy value of the two;

The extraction of the energy variance from the time domain is specifically as follows: according to the difference in the distribution of the distance images of the target and the chaff cloud, the variance of the amplitudes of the distance images of the two is calculated to obtain the energy variance of the two;

The extraction of the waveform width from the time domain is specifically: according to the difference in the length of the target and the chaff cloud in the radial direction of the radar, calculate the distance occupied when the main peak amplitude of the range image of a pulse interfered by the target and the chaff is attenuated by 90% The product of the number of resolution units and the distance resolution is the waveform width;

The extracting the average similarity from the time domain is specifically: calculating the average similarity of two adjacent distance images according to the geometric distribution of the target and the chaff cloud and the difference in volume transformation during motion;

S32, extracting spectrum broadening, frequency-domain skewness, frequency-domain kurtosis, power spectrum envelope fluctuation degree and power spectrum Shannon entropy from the frequency domain;

S321, performing fast Fourier transform on the range image of a pulse of the target and chaff interference obtained in S311 to obtain the frequency spectrum of the target and chaff interference;

S322, according to the spectrum of the target and chaff interference obtained in S321, extract spectrum broadening, frequency domain skewness, and frequency domain kurtosis from the frequency domain;

The extracting spectrum broadening from the frequency domain is specifically as follows: according to the difference in the motion characteristics of the target and the chaff cloud, calculating the spectrum widths of the two to obtain the spectrum widths of the two;

The extracting the frequency domain skewness from the frequency domain is specifically: calculating the skewness coefficient of the frequency spectrum of the target and the chaff cloud according to the difference in the spectral shapes of the target and the chaff cloud to obtain the frequency domain skewness;

The extracting the frequency domain kurtosis from the frequency domain is specifically: calculating the kurtosis coefficient of the frequency spectrum of the target and the chaff cloud according to the difference in the spectral shape of the target and the chaff cloud to obtain the frequency domain kurtosis;

S323, calculating the target and chaff interference spectrum obtained in S321 to obtain the target and chaff interference power spectrum;

S324, extract the envelope fluctuation degree of the power spectrum and the Shannon entropy of the power spectrum from the frequency domain according to the power spectrum of the target and the chaff interference;

The extraction of the power spectrum envelope fluctuation degree from the frequency domain is specifically: according to the difference in the degree of change of the power spectrum envelope amplitude of the target and the chaff cloud, calculating the envelope fluctuation degree of the power spectrum of the two, and obtaining the power spectrum envelope network fluctuation;

The Shannon entropy of the power spectrum extracted from the frequency domain is specifically: the normalized power value of each frequency resolution unit of the power spectrum of the target and the chaff cloud is used to replace the probability value in the entropy function, and the power spectra of the two are calculated respectively. Shannon entropy;

S33. Extract the polarization ratio mean, polarization ratio variance and polarization similarity from the polarization domain;

S331, extracting the co-polarization distance image and cross-polarization distance image of the target and the chaff interference from the target range image sequence and the chaff interference range image sequence obtained in S2, and extracting the polarization similarity from the polarization domain;

The extraction of the polarization similarity from the polarization domain is specifically: calculating the difference between the co-polarization and the cross-polarization distance image of the target and the chaff according to the co-polarization distance image and the cross-polarization distance image of the interference between the target and the chaff. Similarity, get polarization similarity;

S332, obtaining the polarization ratio distance image of the target and the chaff interference according to the co-polarization distance image and the cross-polarization distance image of the target and the chaff interference obtained in S331;

S333. According to the polarization ratio distance image of the target and chaff interference obtained in S331, extract the polarization ratio mean value and the polarization ratio variance sum from the polarization domain;

The extracting the polarization ratio mean value from the polarization domain is specifically: calculating the mean value of the polarization ratio range image of the interference of the target and the chaff to obtain the polarization ratio mean value;

The extracting the polarization ratio variance from the polarization domain is specifically: calculating the variance of the polarization ratio range image of the interference between the target and the chaff to obtain the polarization ratio variance;

S34. Combined distance image concentration, energy ratio, energy mean, energy variance, waveform width, average similarity, spectrum broadening, frequency domain skewness, frequency domain kurtosis, power spectrum envelope fluctuation, power spectrum Shannon entropy, The mean polarization ratio, polarization ratio variance, and polarization similarity are used to obtain the eigenvectors.

S4, the label labeling module labels all the samples obtained in S3 to obtain labeled samples;

The labeling label is specifically: setting the feature vector label of the target range image sequence as 1, and the feature vector label of the chaff interference range image sequence as 0, to obtain a labeled sample;

S5. The data division module divides the labeled samples obtained in S4 into data according to a certain proportion to obtain a training set and a test set;

S6. The LightGBM training module uses the training set obtained in S5 and the grid search method to obtain the LightGBM classifier with the optimal hyperparameters, and then inputs the training set into the LightGBM classifier with the optimal hyperparameters for training, and obtains the lowest classification cross entropy. LightGBM classifier;

S6 specifically includes the following sub-steps:

S61. Use the training set obtained in S5 and the grid search method to obtain the LightGBM classifier with optimal hyperparameters;

S62. Use the training set obtained in S5, the growth-by-leaf algorithm with depth limit and the sparse feature optimization algorithm based on histogram to train the LightGBM classifier with the optimal hyperparameters, and obtain the LightGBM classifier with the lowest classification cross entropy ;

S7, LightGBM classification module inputs the test set obtained in S5 into the LightGBM classifier with the lowest classification cross entropy obtained in S6 for classification, so as to realize the identification and confrontation of chaff interference;

So far, from S1 to S7, an anti-chaf interference method based on multi-domain features and LightGBM has been completed.

beneficial effect

Compared with the existing classification method, an anti-chaf interference method based on multi-domain features and LightGBM according to the present invention has the following beneficial effects:

1. The anti-chaff interference system benefits from the target segmentation module, the feature extraction module and the LightGBM training module. It can not only resist the diluted or centroid-type chaff interference, but also the masked chaff interference, and has a high classification accuracy. , the confrontation effect is stable;

2. The anti-chaff interference method benefits from the fact that the feature extraction module in S3 adopts the features of multiple transform domains, including time domain, frequency domain and polarization domain, so as to avoid inapplicability due to the influence of a single transform domain feature by the environment. Strong defects, good robustness to sea clutter, and strong environmental adaptability;

3. The anti-chaff interference method benefits from the lightweight and efficient LightGBM classifier used in the LightGBM training module, which avoids the defects of the anti-chaff interference method that is classified according to the threshold, which is greatly affected by the environment and difficult to determine. LightGBM classification The binary tree structure and feature segmentation points of the classifier can be learned adaptively. The training time of the LightGBM classifier is much shorter than the training time of the support vector machine used in other inventive methods, and the anti-interference efficiency is high.

Description of drawings

Fig. 1 is the composition and connection relation of a kind of anti-chaf jamming system based on multi-domain feature and LightGBM of the present invention;

2 is a fast time-slow time diagram of a preprocessed full-range image sequence in an anti-chaff interference method based on multi-domain features and LightGBM of the present invention;

Fig. 3 is the flow chart of feature extraction in a kind of anti-chaf interference method based on multi-domain feature and LightGBM of the present invention;

4 is a system block diagram of signal processing when an anti-chaff jamming system based on multi-domain features and LightGBM of the present invention is specifically implemented;

Fig. 5 is a kind of anti-chaff interference method based on multi-domain feature and LightGBM of the present invention, the confrontation result on the dilution or centroid chaff interference of different sea conditions;

FIG. 6 shows the confrontation results of the masked chaff interference with different interference signal ratios when an anti-chaff interference method based on multi-domain features and LightGBM is specifically implemented.

Detailed ways

The method for anti-chaf interference based on multi-domain features and LightGBM according to the present invention will be further described and described in detail below with reference to the accompanying drawings and embodiments.

Example 1

This embodiment describes the process of using the method of the present invention to perform anti-jamming on the target echo data under the background of chaff interference received by the broadband polarized radar. The broadband polarized radar transmits electromagnetic waves in a period of 90 pulses of horizontally polarized waves and then 70 vertically polarized waves. The data set is the target echo data received by the radar horizontal polarization channel under three backgrounds. The first background is the background without sea clutter dilution or chaff interference. There are 7 groups of data in this background, and the number of pulses in each group is 3200, including 1600 pulses in the diffusion stage and 1600 pulses in the stabilization stage. The second background is the background of different sea conditions diluted or centroid chaff interference background. The 6 groups of signal-to-noise ratios (SCR) under this background are 14.03dB, 9.58dB, 6.84dB, 4.97dB, 3.57dB, 2.49dB respectively. Data, the number of pulses in each group is 12800, including 1600 pulses in the growth phase and 11200 pulses in the stationary phase. The third background is the shielded chaff interference background with different S/R under low sea conditions. The S/R of the 10 groups of this background are -20dB, -13.98dB, -6.02dB, -3.10 using the simulated S/R respectively. dB, -0.92dB, 0dB, 6.02dB, 9.54dB, 12.04dB, 13.98dB, the signal-to-noise ratio is 14.03dB, the number of pulses is 12800, including 1600 pulses in the growth period and 11200 pulses in the stable period. One pulse data dimension is 1x2048, one pulse data contains both target and chaff interference, and the target class is 2.

Figure 1 shows the composition and connection relationship of the anti-chaf interference system based on multi-domain features and LightGBM; as can be seen from Figure 1, the classification system includes a preprocessing module, a target segmentation module, a feature extraction module, and a label labeling module. , a data division module, a LightGBM training module and a LightGBM classification module; and the preprocessing module is connected to the target segmentation module, the target segmentation module is connected to the feature extraction module, the feature extraction module is connected to the label labeling module, and the label labeling module is connected to the data division module. The data division module LightGBM training module is connected, and the LightGBM training module is connected with the LightGBM classification module.

S1. The preprocessing module preprocesses the target echo data under the chaff interference of the horizontally polarized receiving channel of the broadband polarized radar to obtain the range image sequence of the horizontally polarized receiving channel;

The preprocessing includes pulse compression, target distance image alignment and normalization operations;

FIG. 2 is a fast time-slow time diagram of the range image sequence of the preprocessed horizontally polarized receiving channel in the anti-chaff interference method based on multi-domain features and LightGBM. It can be seen from Figure 2 that the range image sequence contains information in three dimensions: time domain, frequency domain and polarization domain. The data can clearly reflect the difference between chaff interference and target in time domain, frequency domain and polarization domain. .

S2, the target segmentation module performs chaff interference type judgment and target segmentation on the range image sequence of the horizontally polarized receiving channel obtained in S1, and obtains the target range image sequence and the chaff interference range image sequence;

S2 specifically includes the following sub-steps:

S21. Calculate the average power of the range image sequence of the horizontally polarized receiving channel, and set a threshold to determine the type of chaff interference. If the power is less than the threshold, it is a diluted or centroid type of chaff interference, otherwise, it is a shielded type of chaff interference;

S22. When the type of the chaff interference is diluting or centroid chaff interference, detect the range image sequence of the horizontally polarized receiving channel through CA-CFAR, and perform target segmentation on the over-detected signal. Segmentation, specifically, taking the midpoint of the range image of each over-detected target as the center and taking n (n=320) range resolution units as the range image sequence of the target, to obtain the target range image sequence and the chaff interference range image sequence; When the type of the chaff interference is masked chaff interference, if the target distance is unknown, a target is divided every n range resolution units of the range image, otherwise the range image sequence is divided into the range resolution unit of the known target. Take n range resolution units as the target range image sequence for the center, and take n range resolution units as the chaff interference range image sequence in the remaining areas without target information, and obtain the target range image sequence and the chaff interference range image sequence;

S3. The feature extraction module extracts the range image concentration, energy ratio, energy mean, energy variance, waveform width and average similarity from the time domain, respectively, from the target range image sequence and the chaff interference range image sequence obtained in S2, from the frequency domain. Extract spectrum broadening, frequency domain skewness, frequency domain kurtosis, power spectrum envelope fluctuation and power spectrum Shannon entropy, extract polarization ratio mean, polarization ratio variance and polarization similarity from polarization domain, and combine distance images Concentration, energy ratio, energy mean, energy variance, waveform width, average similarity, spectrum broadening, frequency domain skewness, frequency domain kurtosis, power spectrum envelope fluctuation, power spectrum Shannon entropy, polarization ratio mean, The polarization ratio variance and polarization similarity are used to obtain the eigenvectors corresponding to all target distance image sequences and chaff interference distance image sequences, and the eigenvectors obtained from each sequence are used as a sample to obtain all samples;

Fig. 3 is the flow chart of feature extraction in the described anti-chaf interference method based on multi-domain feature and LightGBM;

Specifically, it includes the following sub-steps:

S31, extracting the distance image concentration, energy ratio, energy mean, energy variance, waveform width and average similarity from the time domain, specifically:

S311 , extract the range image s=[s(1),s(2),...,s(j),... ,s(n)];

S312, according to the range image of a pulse interfered by the target and chaff obtained in S311, extract the range image concentration, energy ratio, energy mean, energy variance, waveform width and average similarity from the time domain;

The extraction of the range image concentration from the time domain is specifically: sorting the amplitudes of all the range resolution units of the range image of a pulse interfered by the target and the chaff, and based on the range resolution unit amplitude sequence sorted from large to small o=[o(1),o(2),…,o(j),…,o(n)], calculate the ratio of its upper quartile Q ₃ to its peak value o(1), and get the distance image concentration

其中n为距离分辨单元个数，且m小于n；The extraction of the energy ratio from the time domain is specifically: sorting the amplitudes of all the range resolution cells of the range image of a pulse interfered by the target and the chaff, and calculating based on the range resolution cell magnitude sequence sorted from large to small. The ratio of the sum of the amplitudes of the first m distance resolution units to the sum of the amplitudes of all n distance resolution units to obtain the energy ratio

where n is the number of distance resolution units, and m is less than n;

The extraction of the energy mean value from the time domain is specifically as follows: according to the radar cross section of the chaff cloud is generally larger than that of the target, calculating the mean value of the range image amplitude of a pulse interfered by the target and the chaff to obtain the energy mean value

The extraction of the energy variance from the time domain is specifically as follows: according to the difference in the range image distribution of the target and the chaff cloud, the variance of the amplitude of the range image of a pulse interfered by the target and the chaff cloud is calculated to obtain the energy variance

The extraction of the waveform width from the time domain is specifically: according to the difference in the length of the target and the chaff cloud in the radial direction of the radar, calculate the distance occupied when the main peak amplitude of the range image of a pulse interfered by the target and the chaff is attenuated by 90% The product of the number of resolution units and the distance resolution is the waveform width W=|k ₂ -k ₁ |×Δr, where k ₁ and k ₂ are subscripts at 90% attenuation of the main peak, and Δr is the distance resolution;

The extraction of the average similarity from the time domain is specifically: calculating the average of two adjacent distance images s(j) and s′(j) according to the geometric distribution of the target and the chaff cloud and the difference in volume transformation during motion similarity

S32, extracting spectrum broadening, frequency-domain skewness, frequency-domain kurtosis, power spectrum envelope fluctuation degree and power spectrum Shannon entropy from the frequency domain;

S321, performing fast Fourier transform on the range image of a pulse of the target and chaff interference obtained in S311 to obtain the frequency spectrum W(f) of the target and chaff interference;

S322, according to the spectrum W(f) of the target and chaff interference obtained in S321, extract spectrum broadening, frequency domain skewness, and frequency domain kurtosis from the frequency domain;

The extracting spectrum broadening from the frequency domain is specifically as follows: according to the difference in the motion characteristics of the target and the chaff cloud, calculating the spectrum widths of the two to obtain the spectrum widths W _f =|f ₂ -f ₁ |×Δf , where f ₁ and f ₂ are the subscripts at 90% attenuation of the main peak of the spectrum, and Δf is the frequency resolution;

其中E(W(f))为频谱的均值，D(W(f))为频谱的标准差；The extraction of the frequency domain skewness from the frequency domain is specifically as follows: according to the difference in the spectral shapes of the target and the chaff cloud, calculating the skewness coefficients of the frequency spectra of the two to obtain the frequency domain skewness

where E(W(f)) is the mean of the spectrum, and D(W(f)) is the standard deviation of the spectrum;

The extraction of the frequency domain kurtosis from the frequency domain is specifically: according to the difference in the spectral shape of the target and the chaff cloud, calculating the kurtosis coefficient of the frequency spectrum of the two to obtain the frequency domain kurtosis

其中N为频谱的点数；S323. Calculate the target and the chaff interference spectrum obtained in S321 to obtain the power spectrum of the target and the chaff interference

where N is the number of points in the spectrum;

S324, extract the envelope fluctuation degree of the power spectrum and the Shannon entropy of the power spectrum from the frequency domain according to the power spectrum of the target and the chaff interference;

其中E(S(f))为功率谱的均值，D(S(f))为功率谱的标准差；The extraction of the power spectrum envelope fluctuation degree from the frequency domain is specifically: according to the difference in the degree of change of the power spectrum envelope amplitude of the target and the chaff cloud, calculating the envelope fluctuation degree of the power spectrum of the two, and obtaining the power spectrum envelope network fluctuation

where E(S(f)) is the mean of the power spectrum, and D(S(f)) is the standard deviation of the power spectrum;

The Shannon entropy of the power spectrum extracted from the frequency domain is specifically: the normalized power value of each frequency resolution unit of the power spectrum of the target and the chaff cloud is used to replace the probability value in the entropy function, and the power spectra of the two are calculated respectively. Shannon entropy

S33. Extract the polarization ratio mean, polarization ratio variance and polarization similarity from the polarization domain;

S331 , extract the co-polarized distance image _shh and cross-polarized distance image s _hv of the target and chaff interference from the target distance image sequence and chaff interference distance image sequence obtained in S2 , and extract the polarization similarity from the polarization domain ;

其中E(s_hh(j))为共极化距离像的均值，E(s_hv(j))为交叉极化距离像的均值；The extraction of the polarization similarity from the polarization domain is specifically: calculating the co-polarization and cross-polarization of the target and the chaff according to the co-polarization distance image _shh and the cross-polarization distance image s _hv of the interference between the target and the chaff The similarity of the distance image is obtained to obtain the polarization similarity

where E(s _hh (j)) is the mean value of the co-polarized range image, and E(s _hv (j)) is the mean value of the cross-polarized range image;

S332, according to the co-polarization distance image s _hh and cross-polarization distance image s _hv of the target and chaff interference obtained in S331, obtain the polarization ratio distance image of the target and chaff interference

S333. According to the polarization ratio distance image of the target and chaff interference obtained in S331, extract the polarization ratio mean value and the polarization ratio variance sum from the polarization domain;

The extracting the polarization ratio mean value from the polarization domain is specifically: calculating the mean value of the polarization ratio range image of the interference of the target and the chaff to obtain the polarization ratio mean value

The extracting the polarization ratio variance from the polarization domain is specifically: calculating the variance of the polarization ratio range profile of the target and the chaff interference to obtain the polarization ratio variance

S34. Combined distance image concentration, energy ratio, energy mean, energy variance, waveform width, average similarity, spectrum broadening, frequency domain skewness, frequency domain kurtosis, power spectrum envelope fluctuation, power spectrum Shannon entropy, The mean polarization ratio, polarization ratio variance, and polarization similarity are used to obtain the eigenvectors.

S4, the label labeling module labels all the samples obtained in S3 to obtain labeled samples;

The labeling label is specifically: setting the feature vector label of the target range image sequence as 1, and the feature vector label of the chaff interference range image sequence as 0, to obtain a labeled sample;

S5. The data division module divides the labeled samples obtained in S4 into data according to a certain proportion to obtain a training set and a test set;

S6. The LightGBM training module uses the training set obtained in S5 and the grid search method to obtain the LightGBM classifier with the optimal hyperparameters, and then inputs the training set into the LightGBM classifier with the optimal hyperparameters for training, and obtains the lowest classification cross entropy. LightGBM classifier;

S6 specifically includes the following sub-steps:

S61. Use the training set obtained in S5 and the grid search method to obtain the LightGBM classifier with optimal hyperparameters;

S62. Use the training set obtained in S5, the growth-by-leaf algorithm with depth limit and the sparse feature optimization algorithm based on histogram to train the LightGBM classifier with the optimal hyperparameters, and obtain the LightGBM classifier with the lowest classification cross entropy ;

The S7 and LightGBM classification modules input the test set obtained in S5 into the LightGBM classifier with the lowest classification cross entropy obtained in S6 for classification, so as to realize the identification and confrontation of chaff interference.

So far, from S1 to S7, an anti-chaf interference method based on multi-domain features and LightGBM has been completed.

FIG. 4 is a system block diagram of signal processing when the anti-chaff jamming system based on multi-domain features and LightGBM is specifically implemented.

In the specific implementation, after preprocessing the echo data of the broadband polarized radar target of the three backgrounds, feature extraction is performed on the chaff data and target data of each pulse, and finally the feature vector is obtained, which is the input of the LightGBM identifier. sample. The first background: no sea clutter dilution or centroid chaff interference background, target data and chaff interference data each have 22400 samples. The second background: the background of different sea conditions diluted or centroid type chaff interference There are 12,800 samples each for the target data and the chaff interference data. The third background: the background of the shielded chaff interference with different S/S ratios in low sea conditions. The S/S ratios of the 10 groups of the backgrounds are -20dB, -13.98dB, -6.02dB, -3.10 using the simulated S/S ratios respectively. dB, -0.92dB, 0dB, 6.02dB, 9.54dB, 12.04dB, 13.98dB, the signal-to-noise ratio is 14.03dB, and the target data and chaff data each have 12800 samples.

The first experiment: train the LightGBM classifier with the training set without the background of sea clutter dilution or centroid chaff interference, and test the adversarial results of the trained LightGBM classifier on the test set. For dilution or centroid chaff interference, identification is confrontation. Therefore, the indicators are commonly used indicators for evaluating the recognition effect of classifiers: accuracy rate, precision rate, recall rate, and F1 score. In order to evaluate the ability of radar target detection, the false alarm rate is also used as an evaluation indicator. Table 1 shows the adversarial results of the trained LightGBM classifier on the test set without the background of sea clutter dilution or centroid chaff interference.

Table 1 The countermeasures against sea clutter-free or centroid chaff interference

Evaluation Metrics mean variance Accuracy 99.98% 1.399e-08 accuracy 99.99% 7.8859e-09 recall 99.99% 7.8859e-09 F1 Score 99.98% 1.4006e-08 false alarm rate 0.07% 4.4176e-08

All indicators except the false alarm rate in Table 1 are all higher than 99%, indicating that an anti-chaf interference method based on multi-domain features and LightGBM of the present invention is effective for diluted or centroid chaff in the background of no sea clutter. The anti-interference effect is very good; the variance of each index in Table 1 is less than 0.0001, indicating that the anti-chaf interference method based on multi-domain features and LightGBM of the present invention has good stability.

The second experiment: Train the LightGBM classifier with the training set with different sea-flavoured or centroid-style chaff interference backgrounds, and test the confrontation results of the trained LightGBM classifier on the test set.

FIG. 5 is the result of confrontation on the dilution or centroid chaff interference under different sea conditions when an anti-chaff interference method based on multi-domain features and LightGBM of the present invention is specifically implemented. It can be seen from Figure 5 that the adversarial results of the trained LightGBM classifier on the test set change with SCR, and the recall rate of the LightGBM classifier on the test data set gradually decreases as the power of sea clutter increases. , the false alarm probability of the LightGBM classifier increases with the decrease of SCR, indicating that sea clutter has a certain impact on the recognition effect of the model; all indicators (except the false alarm rate) of the LightGBM classifier on the test data set are in More than 99%, the classification accuracy rate is high, that is to say, an anti-chaf interference method based on multi-domain features and LightGBM of the present invention can well identify and resist the dilution or centroid type of chaff interference under different sea conditions; The recognition accuracy under the highest sea conditions is less than 1 percentage point lower than the recognition accuracy without sea clutter data, indicating that an anti-chaf interference method based on multi-domain features and LightGBM of the present invention is suitable for the sea clutter background. Has better robustness.

The third experiment: Use the training set of the masked chaff interference background with different interference signal ratios to train the LightGBM classifier, and test the training of the LightGBM classifier on the test set of the masked chaff interference background corresponding to the interference signal ratio. result.

FIG. 6 shows the confrontation results of the masked chaff interference with different interference signal ratios when an anti-chaff interference method based on multi-domain features and LightGBM is specifically implemented. For the shielded chaff interference, the mixed echo is used to identify whether there is a target in the range image of the fixed distance resolution unit, that is, the target is detected and roughly ranged. It can be seen from Fig. 6 that the adversarial results of the trained LightGBM classifier on the test set change with the S/S ratio. The larger the S/S ratio is, the recognition effect decreases slightly, but the decrease is less than 1%, indicating that the present invention An anti-chafing method based on multi-domain features and LightGBM has good anti-chafing effect on masked chaff.

The fourth experiment: the LightGBM classifiers were respectively trained using the training sets extracted based on the method of the present invention from three backgrounds, and the confrontation results of the trained LightGBM classifiers on the test sets were tested. The invention patent "Method for Recognition of Millimeter-Wave Fuze Chaff Interference Based on Extraction of Distance Image Features" is referred to as the existing method 1. According to the existing method 1, the waveform entropy, correlation coefficient and scattering intensity ratio features are extracted to form a training set, and three backgrounds are used. The training set respectively trains the LightGBM classifier, and tests the adversarial results of the trained LightGBM classifier on the test set. The method of the paper "Multi-feature target anti-jamming detection technology based on support vector machine (SVM)" is referred to as the existing method 2. According to the existing method 2, the mean value of echo scattered energy, waveform fluctuation characteristics, effective echo waveform width, The scatter center dimension and the target distribution feature constitute the training set, and the SVMs are trained separately using the training sets of three backgrounds, and the adversarial results of the trained SVMs on the test set are tested. The adversarial results of the two methods are shown in Table 2.

Table 2 Adversarial results of three methods in three contexts

In Table 2, the identification of the chaff interference and the target under the three backgrounds of the method of the present invention are all higher than 99%, indicating that the anti-chaf interference method based on multi-domain features and LightGBM of the present invention is effective for various types of chaff. The anti-interference effect is very good; the recognition effect of using the features of the existing method 1 and the LightGBM classifier is also 92%, but both are lower than the method of the present invention, especially the recognition accuracy under the third background is higher than that of the present invention The method is 7.42% lower than the method of the present invention, indicating that the recognition effect of this method on the interference of the masked chaff is slightly worse than that of the method of the present invention; the existing method 2 has the best recognition effect on the interference of the chaff in the first background and the second background among the three methods. The accuracy rate is the lowest, with an accuracy rate of 57.90% in the third background, indicating that the method has almost no ability to identify the masked chaff interference.

The fifth experiment: train LightGBM classifier and SVM using the training set in three backgrounds, and record the time required for training and the accuracy of the trained LightGBM classifier and SVM on the test set in three backgrounds. The results recorded during training time are shown in Table 3. The results of the accuracy record are shown in Table 4.

Table 3. Training time (seconds) of LightGBM and SVM under three backgrounds

Table 4 Accuracy of LightGBM and SVM in three backgrounds

It can be seen from Table 3 that the training time of LightGBM is much shorter than the training time of SVM, indicating that the anti-chaff interference method based on multi-domain features and LightGBM of the present invention has a fast training time, and the binary tree structure and feature segmentation point of the LightGBM classifier can automatically Adapt to learning, high anti-interference efficiency. It can be seen from Table 4 that the accuracy of the LightGBM classifier in the first two contexts is similar to that of the SVM classifier, both greater than 99%, but the accuracy of the LightGBM classifier in the third context is much greater than that of the SVM classifier The accuracy of the shading chaff interference is better using the LightGBM classifier.

The above descriptions are only the preferred embodiments of the present invention, and the present invention should not be limited to the contents disclosed in the embodiments and the accompanying drawings. All equivalents or modifications accomplished without departing from the disclosed spirit of the present invention fall into the protection scope of the present invention.

Claims (7)

1. an anti-chaf interference system based on multi-domain features and LightGBM, is characterized in that: comprising a preprocessing module, a target segmentation module, a feature extraction module, a label labeling module, a data division module, a LightGBM training module and a LightGBM classification module; The preprocessing module is connected to the target segmentation module, the target segmentation module is connected to the feature extraction module, the feature extraction module is connected to the label labeling module, the label labeling module is connected to the data division module, the data division module is connected to the LightGBM training module, and the LightGBM training module is connected to the LightGBM classification module. modules are connected; The preprocessing module first performs pulse compression on the signals collected by the broadband polarization radar and then performs distance alignment to obtain the range image sequence of the horizontally polarized receiving channel; The target segmentation module determines the type of chaff interference from the range image sequence of the horizontally polarized receiving channel, and divides the determined shielded chaff interference into targets at equal intervals, and divides the determined dilute or The centroid chaff jamming performs target segmentation according to CA-CFAR, and obtains the target range image sequence and the chaff jamming range image sequence; The feature extraction module extracts the range profile concentration, energy ratio, energy mean, energy variance, waveform width, average similarity, spectrum broadening, frequency domain skewness, frequency domain based on the target range profile sequence and the chaff interference range profile sequence. Kurtosis, power spectrum envelope fluctuation, power spectrum Shannon entropy, polarization ratio mean, polarization ratio variance and polarization similarity, combine all features to get all eigenvectors; The label labeling module takes each feature vector as a sample, and labels it to obtain a labeled sample; The data division module divides the data of the labeled samples according to a certain proportion to obtain a training set and a test set; The LightGBM training module uses the training set and the grid search method to obtain the optimal hyperparameters of the LightGBM classifier, and then inputs the training set into the LightGBM classifier with the optimal hyperparameters for training to obtain the LightGBM classifier with the lowest classification cross entropy ; The LightGBM classification module test set inputs the LightGBM classifier with the lowest classification cross-entropy for classification, and outputs a classification result, where the classification result is 1 or 0, and its specific meaning is whether it is a real target or not. 2. an anti-chaf interference method based on multi-domain feature and LightGBM, is characterized in that: comprise the steps: S1. The preprocessing module preprocesses the target echo data under the chaff interference of the horizontally polarized receiving channel of the broadband polarized radar to obtain the range image sequence of the horizontally polarized receiving channel; S2, the target segmentation module performs chaff interference type judgment and target segmentation on the range image sequence of the horizontally polarized receiving channel obtained in S1, and obtains the target range image sequence and the chaff interference range image sequence; S3. The feature extraction module extracts the range image concentration, energy ratio, energy mean, energy variance, waveform width and average similarity from the time domain, respectively, from the target range image sequence and the chaff interference range image sequence obtained in S2, from the frequency domain. Extract spectrum broadening, frequency domain skewness, frequency domain kurtosis, power spectrum envelope fluctuation and power spectrum Shannon entropy, extract polarization ratio mean, polarization ratio variance and polarization similarity from polarization domain, and combine distance images Concentration, energy ratio, energy mean, energy variance, waveform width, average similarity, spectrum broadening, frequency domain skewness, frequency domain kurtosis, power spectrum envelope fluctuation, power spectrum Shannon entropy, polarization ratio mean, The polarization ratio variance and polarization similarity are used to obtain the eigenvectors corresponding to all target distance image sequences and chaff interference distance image sequences, and the eigenvectors obtained from each sequence are used as a sample to obtain all samples; S4, the label labeling module labels all the samples obtained in S3 to obtain labeled samples; S5. The data division module divides the labeled samples obtained in S4 into data according to a certain proportion to obtain a training set and a test set; S6. The LightGBM training module uses the training set obtained in S5 and the grid search method to obtain the LightGBM classifier with the optimal hyperparameters, and then inputs the training set into the LightGBM classifier with the optimal hyperparameters for training, and obtains the lowest classification cross entropy. LightGBM classifier; The S7 and LightGBM classification modules input the test set obtained in S5 into the LightGBM classifier with the lowest classification cross entropy obtained in S6 for classification, so as to realize the identification and confrontation of chaff interference. 3 . The method for anti-chafing interference according to claim 2 , wherein the preprocessing in S1 includes pulse compression, target range image alignment and normalization operations. 4 . 4. The method for resisting chaff interference as claimed in claim 2, characterized in that: S2 specifically comprises the following sub-steps: S21. Calculate the average power of the range image sequence of the horizontally polarized receiving channel, and set a threshold to determine the type of chaff interference. If the power is less than the threshold, it is a diluted or centroid type of chaff interference, otherwise, it is a shielded type of chaff interference; S22. When the type of the chaff interference is diluting or centroid chaff interference, detect the range image sequence of the horizontally polarized receiving channel through CA-CFAR, and perform target segmentation on the over-detected signal. Segmentation, specifically, taking the midpoint of the range image of each over-detected target as the center and taking n range resolution units as the range image sequence of the target, and obtaining the target range image sequence and the chaff interference range image sequence; when the chaff When the interference type is masked chaff interference, if the target distance is unknown, divide a target every n range resolution units of the range image, otherwise the range image sequence is centered on the range resolution unit of the known target. The range resolution unit is used as the target range profile sequence, and n range resolution units are taken as the chaff interference range profile sequence in the remaining areas without target information, and the target range profile sequence and the chaff interference range profile sequence are obtained. 5. The method for resisting chaff interference as claimed in claim 1, characterized in that: one of the feature vectors in S3 is specifically obtained through the following sub-steps: S31, extracting the distance image concentration, energy ratio, energy mean, energy variance, waveform width and average similarity from the time domain, specifically: S311, extracting a range image of a pulse of target and chaff interference from the target range image sequence and the chaff interference range image sequence obtained in S2; S312, according to the range image of a pulse interfered by the target and chaff obtained in S311, extract the range image concentration, energy ratio, energy mean, energy variance, waveform width and average similarity from the time domain; The extraction of the range image concentration from the time domain is specifically: sorting the amplitudes of all the range resolution units of the range image of a pulse interfered by the target and the chaff, and based on the range resolution unit amplitude sequence sorted from large to small Calculate the ratio of its upper quartile to its peak value to get the distance image concentration; The extraction of the energy ratio from the time domain is specifically: sorting the amplitudes of all the range resolution cells of the range image of a pulse interfered by the target and the chaff, and calculating based on the range resolution cell magnitude sequence sorted from large to small. The ratio of the sum of the amplitudes of the first m distance resolution units to the sum of the amplitudes of all n distance resolution units, the energy ratio is obtained, and m is less than n; The extracting the energy mean value from the time domain is specifically: according to the radar scattering cross section of the chaff cloud is generally larger than the radar scattering cross section of the target, calculate the mean value of the range image amplitude of a pulse interfered by the target and the chaff, and obtain the energy mean value; The extraction of the energy variance from the time domain is specifically as follows: according to the difference in the range image distribution of the target and the chaff cloud, the variance of the amplitude of the range image of a pulse interfered by the target and the chaff is calculated to obtain the energy variance; The extraction of the waveform width from the time domain is specifically: according to the difference in the length of the target and the chaff cloud in the radial direction of the radar, calculate the distance occupied when the main peak amplitude of the range image of a pulse interfered by the target and the chaff is attenuated by 90% The product of the number of resolution units and the distance resolution is the waveform width; The extraction of the average similarity from the time domain is specifically: calculating the average similarity of two adjacent distance images according to the geometric distribution of the target and the chaff cloud and the difference in volume transformation during motion. S32, extracting spectrum broadening, frequency-domain skewness, frequency-domain kurtosis, power spectrum envelope fluctuation degree and power spectrum Shannon entropy from the frequency domain; S321, performing fast Fourier transform on the range image of a pulse of the target and chaff interference obtained in S311 to obtain the frequency spectrum of the target and chaff interference; S322, according to the spectrum of the target and chaff interference obtained in S321, extract spectrum broadening, frequency domain skewness, and frequency domain kurtosis from the frequency domain; The extracting spectrum broadening from the frequency domain is specifically as follows: according to the difference in the motion characteristics of the target and the chaff cloud, calculating the spectrum widths of the two to obtain the spectrum widths of the two; The extracting the frequency domain skewness from the frequency domain is specifically: calculating the skewness coefficient of the frequency spectrum of the target and the chaff cloud according to the difference in the spectral shapes of the target and the chaff cloud to obtain the frequency domain skewness; The extracting the frequency domain kurtosis from the frequency domain is specifically: calculating the kurtosis coefficient of the frequency spectrum of the target and the chaff cloud according to the difference in the spectral shape of the target and the chaff cloud to obtain the frequency domain kurtosis; S323, calculating the target and chaff interference spectrum obtained in S321 to obtain the target and chaff interference power spectrum; S324, extract the envelope fluctuation degree of the power spectrum and the Shannon entropy of the power spectrum from the frequency domain according to the power spectrum of the target and the chaff interference; The extraction of the power spectrum envelope fluctuation degree from the frequency domain is specifically: according to the difference in the degree of change of the power spectrum envelope amplitude of the target and the chaff cloud, calculating the envelope fluctuation degree of the power spectrum of the two, and obtaining the power spectrum envelope network fluctuation; The Shannon entropy of the power spectrum extracted from the frequency domain is specifically: the normalized power value of each frequency resolution unit of the power spectrum of the target and the chaff cloud is used to replace the probability value in the entropy function, and the power spectra of the two are calculated respectively. Shannon entropy; S33. Extract the polarization ratio mean, polarization ratio variance and polarization similarity from the polarization domain; S331, extracting the co-polarization distance image and cross-polarization distance image of the target and the chaff interference from the target range image sequence and the chaff interference range image sequence obtained in S2, and extracting the polarization similarity from the polarization domain; The extraction of the polarization similarity from the polarization domain is specifically: calculating the difference between the co-polarization and the cross-polarization distance image of the target and the chaff according to the co-polarization distance image and the cross-polarization distance image of the interference between the target and the chaff. Similarity, get polarization similarity; S332, obtaining the polarization ratio distance image of the target and the chaff interference according to the co-polarization distance image and the cross-polarization distance image of the target and the chaff interference obtained in S331; S333. According to the polarization ratio distance image of the target and chaff interference obtained in S331, extract the polarization ratio mean value and the polarization ratio variance sum from the polarization domain; The extracting the polarization ratio mean value from the polarization domain is specifically: calculating the mean value of the polarization ratio range image of the interference of the target and the chaff to obtain the polarization ratio mean value; The extracting the polarization ratio variance from the polarization domain is specifically: calculating the variance of the polarization ratio range image of the interference between the target and the chaff to obtain the polarization ratio variance; S34. Combined distance image concentration, energy ratio, energy mean, energy variance, waveform width, average similarity, spectrum broadening, frequency domain skewness, frequency domain kurtosis, power spectrum envelope fluctuation, power spectrum Shannon entropy, The mean polarization ratio, polarization ratio variance, and polarization similarity are used to obtain the eigenvectors. 6 . The method for anti-chaf interference according to claim 1 , wherein the labeling in S4 is specifically: labeling the feature vector of the target range image sequence as 1, and the chaff interference range image sequence is 1. 7 . The feature vector label is 0, and the labeled sample is obtained. 7. The method for resisting chaff interference as claimed in claim 1, wherein S6 specifically comprises the following sub-steps: S61. Use the training set obtained in S5 and the grid search method to obtain the LightGBM classifier with optimal hyperparameters; S62. Use the training set obtained in S5, the growth-by-leaf algorithm with depth limit and the sparse feature optimization algorithm based on histogram to train the LightGBM classifier with the optimal hyperparameters, and obtain the LightGBM classifier with the lowest classification cross entropy .

Images