Radar interference detection and analysis method based on time-frequency transformation

Abstract

The invention discloses a radar interference detection and analysis method based on time-frequency transformation, and belongs to the technical field of radar interference detection.The method comprises the steps that firstly, rough detection is completed through the characteristics of energy and entropy, and a rough interference area is screened out; extracting frequency domain skewness and kurtosis to construct a radar interference detection model to finish fine detection; and finally, the trained SVM model is used to realize final detection and evaluation. Experimental results show that the model realizes better interference detection performance; compared with a traditional detection method, the method does not need a large number of interference echo samples, multiple features such as energy, entropy, skewness and kurtosis can be merged into the same vector, the optimal weight and the nonlinear boundary are automatically learned, and complementarity between the features is effectively utilized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar jamming detection, and particularly relates to a radar jamming detection analysis method based on time-frequency transformation. BACKGROUND

[0002] With the development of radar technology, a series of research results have emerged in the field of radar jamming detection in recent years. Generally speaking, the jamming detection method of radar can be mainly divided into two categories: traditional method and deep learning method. In the public paper "Radar Active Jamming Recognition Based on Time-frequency Feature Extraction", the generation mechanism of several jamming signals is studied, the jamming model is established, the feature differences of five typical jamming signals are analyzed, the Choi-Williams distribution is used as the description method of jamming signals, the obtained jamming signal time-frequency two-dimensional distribution image is given, the effective time-frequency domain feature parameter set is extracted, the feature set is reduced through simulation comparison, and finally the support vector machine SVM is used to recognize the radar active jamming. However, this method has large amount of calculation, which is not conducive to engineering implementation, and how to extract more efficient feature parameters needs further research. In the public paper "A Radar Main Lobe Jamming Detection Method Based on Time-frequency Space Feature", a radar main lobe jamming detection method based on time-frequency space feature is proposed. This method analyzes and verifies the differences in time-space distribution and time-frequency distribution characteristics of jamming and target signals as feature quantities, and respectively uses time-space analysis and time-frequency analysis methods to detect and verify different types of jamming such as typical long-distance natural scene, suppression jamming and deception jamming. However, this algorithm can only detect main lobe jamming, and is difficult to adapt to side lobe jamming, and the algorithm can only detect active jamming, and is difficult to realize classification and recognition.

[0003] The existing method has the following disadvantages: 1. The performance of the traditional jamming detection method based on feature extraction mainly depends on the discrimination ability of artificial features, the application range is limited, and the recognition accuracy is poor.

[0004] 2. The jamming detection method based on neural network needs a large number of training samples, the training cost is high, and the robustness is poor. SUMMARY

[0005] The present application aims at: in view of the problem that the interference signal and the radar echo signal overlap in multiple parameter domains, so that the radar is difficult to identify whether it is interfered, and thus cannot adopt effective methods for interference detection and suppression, the present application provides a radar interference detection analysis method based on time-frequency transformation, which finds the characteristic parameters capable of effectively representing the radar received signal from the perspective of time-frequency analysis, and detects and identifies the radar interference signal.

[0006] The technical scheme adopted by the present application is as follows: A radar interference detection analysis method based on time-frequency transformation, the method comprising: establishing a time domain model, and performing time-frequency joint analysis on the radar signal through the time domain model; coarse detection, establishing an energy and entropy value feature detector, performing energy difference analysis on the signal processed by the time-frequency model, and performing entropy value calculation after the analysis to obtain a preliminary candidate interference region; fine detection, extracting frequency domain high value statistical features from the candidate unit screened out by the coarse detection; establishing a support vector machine model, combining the coarse detection and the fine detection, constructing a training set to train the support vector machine model, and performing final detection of the interference region through the support vector machine model.

[0007] Further, the time domain model in step 1 is specifically as follows: the time-frequency domain is represented by using a smoothing WVD transformation, as follows: (1) The parameters in the formula are and respectively windowed, as follows: (2).

[0008] Further, the energy difference analysis in the coarse detection specifically comprises: the spectrum energy of each frame of the radar signal after SPWVD smoothing is added up to obtain the total energy on the time sequence, then the mean value and the mean square deviation are calculated and normalized, an energy threshold is set, and the time points higher than the threshold are identified as possible abnormal energy to roughly screen out the interference region.

[0009] Further, the entropy value calculation is specifically represented by information entropy, the probability of the occurrence of a signal in an information source is defined as and the information entropy H is as follows: (3) The energy difference analysis is combined with entropy calculation to complete the rough detection, and a preliminary candidate interference region is obtained.

[0010] Further, the frequency domain high value statistical features in the fine detection include skewness and kurtosis. The skewness is specifically as follows: (4) In the formula, is a variance, is a mean value, is a cumulant, and when the distribution curve data is closer to symmetry, the skewness value is closer to zero; The kurtosis is specifically as follows: (5) The interference is discriminated through the skewness and the kurtosis.

[0011] Further, the features obtained through the rough detection and the fine detection are combined into a feature vector, a support vector machine model (SVM) is used for binary classification, and an accurate interference region is obtained.

[0012] Further, the method further includes detection evaluation, and the evaluation is specifically performed through the following formula: (6).

[0013] As described above, due to the adoption of the above technical solutions, the present application has the following beneficial effects: The radar interference detection analysis method based on time-frequency transformation first completes rough detection through the characteristics of energy and entropy, filters out a general interference region; then extracts frequency domain skewness and kurtosis, constructs a radar interference detection model to complete fine detection; finally, a trained SVM model is used to realize final detection and evaluation; the experimental results show that the model realizes good interference detection performance; compared with the traditional detection method, the present application does not need a large number of interference echo samples, can integrate multiple features such as energy, entropy, skewness and kurtosis into the same vector, automatically learns the optimal weight and nonlinear boundary, and effectively utilizes the complementarity between the features. The required samples of the training set are not many, the prediction speed is fast, and the generalization and robustness of the model are high. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 FIG. 1 is a flowchart of a radar interference detection analysis method based on time-frequency transformation of the present application; Figure 2 FIG. 2 is a SPWVD time-frequency distribution diagram in the method of the present application; Figure 3A schematic diagram of a principle of an SVM soft segmentation classification model in the method of the application; Figure 4 A schematic diagram of a fine detection result in the method of the application; Figure 5 A schematic diagram of an SVM detection result in the method of the application. DETAILED DESCRIPTION

[0015] The application will be described in further detail below with reference to the drawings.

[0016] In order to make the objectives, technical solutions and advantages of the application clearer, the application will be further described below with reference to the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application.

[0017] Example 1 This example provides a radar interference detection and analysis method based on time-frequency transformation, as shown in Figure 1 The method is implemented according to the following steps: First, a time-domain model of a radar receiving signal is established, and then a time-frequency joint analysis is performed on the radar receiving signal, and Wigner-Ville distribution (WVD) is used as a description method of the radar receiving signal. There are many methods for representing the time-frequency domain of a signal, and since WVD has good resolution and is the best time-frequency distribution in terms of time-frequency concentration, WVD is used in this example. However, since WVD has cross terms, in order to ensure its resolution while excluding the influence of cross terms, as shown in Figure 2 Smooth WVD (SPWVD) transformation is used, specifically as follows: (1) The parameters and in the formula are windowed, as follows: (2).

[0018] An energy and entropy feature detector is designed to realize coarse detection by using the distribution difference between the energy of an interference signal and the energy of a target echo in a time-frequency plane, and to quickly screen out abnormal units on a time-frequency grid.

[0019] Since the energy of the interference signal is usually concentrated or abnormally distributed in the time-frequency plane, the interference or target signal may exist in the place with high energy or abnormal distribution. The energy of each frame of the signal after SPWVD smoothing is added to obtain the total energy in the time sequence. Then the mean and the mean square deviation are calculated for normalization processing, and a suitable energy threshold is set. The time points higher than the threshold are considered to have abnormal energy, i.e. interference or target. In this way, the region where the interference exists can be roughly screened out, and other irrelevant regions are excluded.

[0020] After the energy threshold screening, the entropy value is calculated. Generally, the region with low entropy value, high kurtosis and high effective number of pulses is the target signal, and the region with high entropy value and low effective number of pulses is the interference signal. The main information entropy, i.e. Shannon entropy, is calculated in this embodiment. The information entropy is an effective index for quantifying the uncertainty of the system state or signal. The index can be deeply combined with various signal processing methods to extract the characteristics of the signal from different transform domains. In the radar system, the received signal is composed of target echo and noise. Due to the existence of random noise, the specific form of the received signal is often different even if the same signal is received twice, showing inherent randomness. The information entropy provides an effective mathematical tool to accurately describe the uncertainty degree of the received signal.

[0021] The signal The probability of occurrence in an information source is defined as The mathematical expression of the information entropy H is: (3) The combination of energy statistics and entropy analysis can complete the coarse detection to obtain the preliminary candidate region, providing the time-frequency range for subsequent fine detection. The advantage of coarse detection is fast and global scanning, without the need to calculate complex skewness or kurtosis, and the efficiency is high, but false positives may occur, which need to be further screened by fine detection.

[0022] Further fine detection is performed on the candidate units screened out by coarse detection, and the skewness and kurtosis of the extracted high-order statistical characteristics in the frequency domain are used for further judgment, as follows: The essence is to change each candidate time-frequency unit into a stable feature vector to realize the subsequent detection. After completing the coarse detection based on the energy and entropy characteristics, the candidate units where the interference signals may exist are marked in the time-frequency diagram. However, since the coarse detection only depends on the energy threshold and the entropy characteristics, the result often has a certain false alarm. Therefore, it is necessary to further extract the high-order statistical characteristics with higher discrimination in the candidate units to realize accurate identification. Therefore, the skewness and kurtosis of the frequency domain statistical characteristics are introduced in the fine detection stage to represent the narrowband interference signal by analyzing the asymmetry and sharpness of the local spectrum distribution.

[0023] Firstly, the candidate unit after rough detection is extracted, and the output of the rough detection stage is a set of suspicious time-frequency unit collection. Set the index set of these units on the time-frequency plane as: (7) In the formula, f represents the frequency index, t represents the time index, for each candidate time frame , the corresponding one-dimensional spectrum slice is taken as the object of further analysis, which can convert the two-dimensional time-frequency detection problem into a one-dimensional spectrum statistical feature extraction problem.

[0024] For the current electronic warfare core used deception jamming, it accurately replicates radar signals through digital radio memory (DRFM) and other technologies, and transmits after delay and frequency shift, thus generating false targets difficult to distinguish on radar. And deception jamming is almost narrowband jamming, so this embodiment mainly analyzes narrowband jamming.

[0025] The spectral energy of narrowband jamming is usually concentrated in a narrow frequency band range, which shows obvious peak characteristics in the frequency domain. This kind of interference can often be equivalent to the superposition of several single frequency signals. Since the narrowband jamming signal has local prominent characteristics in the frequency domain, while random noise or normal target echo is often uniformly distributed in the frequency domain, in order to enhance the interference characteristics, local smoothing of the spectrum is needed. The specific method is: for each candidate time frame , a time window is selected around it , the spectral mean of all frames in the window is calculated, and the frequency average spectrum is obtained as follows: (8) After local smoothing, not only can the influence of random noise be reduced, but also the steady-state characteristics of the interference component can be highlighted, making the subsequent high-order features more robust and improving the robustness of detection.

[0026] After obtaining the frequency average spectrum, its probability distribution characteristics can be calculated. First, the average spectrum amplitude is normalized to a probability distribution, and then the frequency spectrum statistical characteristics are calculated based on the probability distribution.

[0027] Skewness is a measure used in statistics to measure the direction and degree of data distribution skewness, which is a numerical characteristic of the degree of asymmetry of statistical data distribution. It describes the skewness or asymmetry of data distribution to some extent, that is, the shape of the data distribution around the mean. It can be defined as the third-order standardized moment of the sample, and the specific definition is: (4) In the formula, is variance, is mean, is cumulant, when the distribution curve data is closer to symmetry, its skewness value will be closer to zero.

[0028] Kurtosis, also known as kurtosis coefficient, is a statistical quantity used to describe the sharpness or flatness of the probability density distribution curve. Intuitively, kurtosis reflects the height of the peak of the distribution curve at its mean value, i.e. the sharpness of the peak, and also reflects the thickness of the tail. The general method for calculating the kurtosis of a random variable is the ratio of the fourth central moment to the square of the variance. The specific calculation formula is: (5) Kurtosis includes normal distribution, thick tail, and thin tail. The kurtosis value of normal distribution is 3. If the kurtosis of a certain distribution is greater than 3, it means that the distribution has a thick tail and the distribution curve is relatively sharp. If its kurtosis is less than 3, it means that the distribution has a thin tail and the distribution curve is relatively flat. According to the above characteristics, kurtosis can be used to describe the shape of data distribution, which helps to understand the kurtosis characteristics of data and also can identify whether there are outliers or long-tail distributions in the data set.

[0029] When there is narrowband interference, the spectral distribution shows a clear single-peak protrusion at a certain frequency, which makes the absolute value of skewness significantly increase, and the distribution curve becomes asymmetric. At the same time, the spectral distribution near the interference point is more sharp than random noise, which leads to a significant increase in kurtosis. If the spectrum is composed of only noise or target echo, the spectral distribution is approximately smooth and symmetric, the skewness is close to zero, and the kurtosis is close to the theoretical value 3 of normal distribution. Therefore, skewness and kurtosis can be used as important discriminant indicators for narrowband interference detection.

[0030] Combining coarse detection and fine detection, construct a training set to train the support vector machine (SVM) model, and use SVM for binary classification, which includes: Combine the features obtained after coarse detection and fine detection into a feature vector, and use SVM for binary classification, i.e. interference / non-interference classification. The SVM algorithm training samples mainly come from Sentinel-1 SAR satellite images, and the data is downloaded from the official website for training.

[0031] SVM is a high-efficiency machine learning method, its main idea is to find an optimal separation plane in the feature space, so as to classify the samples. The goal of this algorithm is to maximize the separation distance between different categories of data points, so as to determine the decision boundary. It is worth noting that only a few key data points, called support vectors, ultimately determine the position of the separation boundary.

[0032] Compared with other classification algorithms, SVM has strong generalization ability, small storage overhead when processing high-dimensional data, and is easy to deploy in engineering. With the help of kernel function, SVM can flexibly handle linear and nonlinear classification problems. Its mathematical essence is to solve the hyperplane that maximizes the classification interval under the constraint condition.

[0033] For example, in a two-dimensional plane, the decision boundary assumption is a hyperplane, and the interval distance reflects the degree of separation between the two classes of data. According to whether the samples are allowed to exist near the boundary, SVM can be divided into "hard interval" and "soft interval" two forms. Since real data is often not completely linearly separable, and there may be individual abnormal samples, the soft interval classification method is usually used, that is, a small number of samples are allowed to fall into the interval area or even be misclassified, and the classification model can be referred to as shown in Figure 3 .

[0034] However, in most cases, it is difficult to find a hyperplane to perfectly separate the data set, and the samples cannot be linearly separated in low-dimensional space. Therefore, the data needs to be upgraded to map it to high-dimensional space, and the separation is realized in high-dimensional space.

[0035] The kernel function used in SVM is also called a feature transformation function, which can map the original features to a high-dimensional space while avoiding direct high-dimensional operations. Through the kernel trick, the inner product operation in the high-dimensional feature space is converted into a function calculation in the original low-dimensional space. This way not only saves computing resources, but also makes the support vector machine more efficient in handling nonlinear classification problems. Assuming the mapping function is , it is easy to get that the kernel function is a function of variables and , and has nothing to do with the mapping function itself.

[0036] Kernel functions can generally be divided into linear kernel functions, Gaussian kernel functions, and sigmoid kernel functions. In this invention, the number of selected features is only two, but the number of samples is large. Therefore, considering comprehensively, the Gaussian kernel function is selected as the kernel function of SVM. The expression of the Gaussian kernel function is as follows: (9) In the formula, is the parameter of the Gaussian kernel function, which controls the decay speed of the similarity between samples. The smaller the value is, the faster the similarity decreases. Set the jamming signal ratio JSR to 2dB, and the change range of SNR or JNR to-5 to 5dB with an interval of 1dB. The accuracy on the validation set is 98%, so the model can be considered to have good performance.

[0037] The test set is used to verify and evaluate the detection performance of the detection model, as follows: In addition, the generalization ability of the model can be verified, and the detection accuracy is The performance of the detection model is evaluated. The expression is as follows: (6) As shown in Figure 4 and Figure 5 verified, 96.5% can be achieved; the test results show that the model can well fit the training data, has good generalization ability, and shows good detection performance.

[0038] In the traditional interference detection method, when the eigenvalue is used for interference detection, one-dimensional threshold judgment is usually performed, it is difficult to simultaneously utilize multiple complementary information, and a large amount of non-interference and interference echo data need to be used for feature extraction and comparison experiments, and then a detection threshold is set, which not only needs a large amount of interfered echo data but also increases the labor cost. In the method of the present application, through the combination of coarse detection and fine detection, multiple features such as energy, entropy, skewness and kurtosis can be integrated into the same vector, the optimal weight and nonlinear boundary are automatically learned, and the complementarity between features is effectively utilized. Moreover, the generalization of SVM is relatively strong, the required samples of the training set are not many, and the prediction speed is fast enough.

[0039] In this paper, specific examples are applied to explain the principles and implementation modes of the present application, and the above examples are only used to help understand the method of the present application and its core idea. It should be pointed out that for ordinary skilled persons in the technical field, some improvements and modifications can be made to the present application without departing from the principles of the present application, and these improvements and modifications also fall within the protection scope of the claims of the present application.

Claims

1. A method of radar jamming detection analysis based on time-frequency transform, characterized in that, The method comprises: A time domain model is established, and radar signals are analyzed in time-frequency domain by the time domain model; Coarse detection is performed by establishing an energy and entropy value feature detector, and energy difference analysis is performed on the signals processed by the time-frequency model, and entropy value calculation is performed after the analysis to obtain a preliminary candidate interference region; Fine detection is performed by extracting frequency domain high value statistical features from the candidate units screened by the coarse detection; A support vector machine model is established, a training set is constructed by combining the coarse detection and the fine detection to train the support vector machine model, and final detection of the interference region is performed by the support vector machine model.

2. The radar jamming detection analysis method based on time-frequency transform according to claim 1, characterized in that, The time domain model in step 1 is specifically as follows: A smooth WVD transform is used to represent the time-frequency domain, and the formula is as follows: （1） wherein is a time delay, t and f are time and frequency variables, respectively; The parameters in the formula are and are windowed as follows: （2） In the formula, is a frequency smoothing window function, and g(u) is a time smoothing window function, which are respectively used for smoothing the time-frequency distribution in the frequency domain and in the time domain.

3. The radar jamming detection and analysis method based on time-frequency transform according to claim 1, characterized in that, The energy difference analysis in the coarse detection specifically comprises: The energy of each frame of the radar signals after the SPWVD smoothing is summed to obtain the total energy on the time sequence, and then the mean value and the mean square deviation are calculated and normalized, an energy threshold is set, and the time points higher than the threshold are determined to possibly exist abnormal energy, and the interference region is roughly screened out.

4. The radar jamming detection analysis method based on time-frequency transform according to claim 3, characterized in that, The entropy value calculation is specifically represented by information entropy, and the signal The probability of occurrence in an information source is defined as The information entropy H is as follows: （3） The energy difference analysis and the entropy calculation are combined to complete the coarse detection, and a preliminary candidate interference region is obtained.

5. The method of claim 1, wherein, The frequency domain high value statistical features in the fine detection comprise skewness and kurtosis. The skewness is specifically as follows: （4） In the formula, is the variance, is the mean, is the skewness, and the closer the distribution curve data is to symmetry, the closer its skewness value will be to zero; The kurtosis is specifically as follows: （5） The skewness and the kurtosis are used for interference discrimination.

6. The method of radar jamming detection analysis based on time-frequency transform according to claim 1, characterized in that, The features obtained by the coarse detection and the fine detection are combined into a feature vector, a support vector machine model SVM is used for binary classification, and an accurate interference region is obtained.

7. The method of claim 1, wherein, The method further comprises detection evaluation, and the evaluation is specifically performed by the following formula: （6） In the formula, TP represents true cases, TN represents true negative cases, FP represents false positive cases, FN represents false negative cases.

Images