A refined feature recognition method for radar emitter signals based on self-weighted structured feature selection

Abstract

The present invention relates to a refined feature recognition method for radar emitter signals based on self-weighted structured feature selection. Signal data emitted by radar devices of different models are collected, and the original data is subjected to signal data preprocessing and normalization processing. A refined feature recognition model for radar emitter signals based on self-weighted structured feature selection is constructed. The emitter signal data is used to perform running tests on the above model, and the obtained recognition results are output, so as to obtain the type of the recognized radar emitter signal. Recognition is based on unsupervised feature selection in machine learning, avoiding the construction of neural networks and the data training process in deep learning. The number of hyperparameters in the model is greatly reduced, avoiding the adjustment of hyperparameters, and making the model easy to implement in actual situations.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar emitter signal processing in electronic countermeasure, and particularly relates to a refined feature recognition method for radar emitter signals based on self-weighted structured unsupervised feature selection. Background Art

[0002] With the advent of the information age, mobile communication technologies, wireless network devices, and Internet of Things facilities have all achieved rapid development and received extensive attention. In recent years, with the widespread deployment of 4G technology, the commercialization of 5G technology, and the start of frontier research on 6G technology, more and more communication facilities have entered people's lives. As a result, the working mechanisms have become gradually more complex and variable, which requires more reliable and secure encryption measures. In modern electronic information warfare, more battlefield intelligence of the opponent, especially radar emitter signals, is intercepted through electronic reconnaissance and search, and refined feature information of the opponent's emitter signals and the model numbers of the affiliated devices are obtained through analysis, and then information on the platforms carrying the opponent's emitters is mastered, thereby providing strong support for front-line electronic countermeasure and electronic jamming. During the process of signal generation or signal capture, due to the hardware defects of the signal source device, some subtle features will be added to the emitter signals, and these features are unique. The refined feature recognition technology for radar emitter signals is a technology that accurately identifies the opponent's radar emitter devices by performing refined feature analysis on the captured radar signals of the opponent, and plays an important role in the analysis and control of the electromagnetic spectrum in the electronic battlefield and military field. Therefore, the radar emitter feature recognition technology has great strategic significance and occupies the core position of electronic warfare.

[0003] Traditional radar emitter signal recognition technologies mainly extract and analyze features manually. This method requires people to have prior knowledge of the obtained radar signals, and then manually analyze the features and match them with the existing individual emitter databases to identify the emitter signals. However, since manual feature analysis is affected by human factors, there are certain limitations in the characterization of emitter signal information, which will also cause problems in practical applications, mainly manifested in poor robustness and insufficient and incomplete characterization of features. In recent years, deep learning technologies have also demonstrated strong learning and training capabilities in many fields. Therefore, experts have also tried to introduce deep learning into the field of emitter signal recognition. The reason why deep learning can demonstrate strong feature information characterization capabilities is usually at the cost of sufficient training data. However, in practical application fields, especially in non-cooperative communication scenarios, it is very difficult to obtain sufficient data of a certain emitter signal for deep neural network training and learning.

[0004] In recent years, due to its high representation ability, machine learning methods have been introduced by experts into the field of radiation source signal recognition technology. In particular, unsupervised refined feature selection has received extensive attention and research because it does not require prior knowledge of the label information of radiation source signal data. Traditional unsupervised feature selection methods are based on an assumption that the weights of the refined data of each sample in the radiation source signal are the same and play the same role in feature representation. Then, based on this assumption, the original radiation source data is constructed into a graph, and the feature selection process is carried out simultaneously. However, these methods cannot well distinguish samples that contribute differently in the data, and the recognition accuracy of the radiation source is relatively low when the noise is large, resulting in unsatisfactory subsequent effects. Summary of the Invention

[0005] Technical Problems to be Solved

[0006] To avoid the deficiencies of the prior art, the present invention provides a refined feature recognition method for radar radiation source signals based on self-weighted structured feature selection, which has a short analysis time, a fast convergence speed, and high recognition accuracy and efficiency.

[0007] Technical Solution

[0008] A refined feature recognition method for radar radiation source signals based on self-weighted structured feature selection, characterized by the following steps:

[0009] Step 1: Collect the signal data emitted by radar devices of different models, and at the same time intercept the pulse data in the signal to generate an individual recognition sample set of radar radiation source signals, that is, the original data;

[0010] Step 2: Perform signal data preprocessing on the original data obtained in Step 1;

[0011] Step 3: Perform normalization processing on the data that has undergone signal preprocessing in Step 2;

[0012] Step 4: Construct a refined feature recognition model for radar radiation source signals based on self-weighted structured feature selection;

[0013] The refined feature recognition model of the radar radiation source signal:

[0014]

[0015] Where is two-dimensional, where n is the number of samples, d is the dimension of the data, and the i-th sample of the data is represented by ; W is the learned orthogonal projection matrix; m is the dimension of the low-dimensional space, k is the row sparsity of the transformation matrix, is a diagonal matrix whose diagonal elements represent the weights of sample data, sij To solve the adjacency probability, γ and λ are regularization term parameters;

[0016] Step 5: Use the radiation source signal data to perform a running test on the model established in Step 4, input the recognition features into the classifier for classification, so as to obtain the type of the recognized radar radiation source signal.

[0017] A further technical solution of the present invention: The preprocessing described in Step 2 includes denoising processing and signal enhancement processing.

[0018] A further technical solution of the present invention: The model optimization process in Step 4 is as follows:

[0019] According to the relationship between the matrix two-norm and the trace operation The model is equivalent to:

[0020]

[0021] It can be seen that the model (14) contains three variables, namely the weight matrix Θ, the similarity matrix S, and the projection matrix W. For its solution problem, the present invention designs an iterative optimization algorithm. The specific process is as follows:

[0022] ① Initialization: Initialize the weight matrix where I d is the identity matrix of dimension d;

[0023] ② Solve Θ:

[0024] When the variables S and W are fixed, problem (14) is transformed into:

[0025]

[0026] According to the properties of the trace operation, it can be changed to:

[0027]

[0028] Apply the theorem to problem (16), and there is:

[0029]

[0030] Obviously, this is a quadratic problem, so the present invention uses the augmented Lagrangian multiplier method to solve this problem;

[0031] ③ Solve S:

[0032] When the variables Θ and W are fixed, problem (14) is changed to:

[0033]

[0034] Its Lagrangian function is as follows:

[0035]

[0036] where Ψ = {ψ i | i = 1, 2,..., n} and Δ = {δ ij | i, j = 1, 2,..., n} are Lagrange multipliers; according to the KKT conditions, we can obtain

[0037]

[0038] Also, because of the restriction of the obtained s ij is as follows:

[0039]

[0040] ④ Solve for W:

[0041] When the variables Θ and S are fixed, problem (14) becomes:

[0042]

[0043] We let A = βI - Θ T X T L S XΘ + λΘ, then solving problem (22) becomes solving the following problem:

[0044]

[0045] where β is large enough to ensure that A is a positive semi - definite matrix, so β is set to the largest eigenvalue of the matrix B = Θ T X T L S XΘ + λΘ; for the solution of problem (23), we divide it into two cases to solve respectively, namely rank(A) ≤ m and rank(A) > m; the optimization algorithm of the present invention uses different methods to solve them respectively.

[0046] ⑤ Repeat the above steps ②, ③ and ④ until the model function converges (the convergence condition is |obj t+1 +obj t | ≤ 10 -4 ), at this time, the W obtained in step ④ is the orthogonal projection matrix required by our feature analysis module, and then the first m features are selected according to the descending order of .

[0047] A further technical solution of the present invention: the classifier described in step 5 includes a linear classifier, a non - linear classifier and a deep learning neural network.

[0048] A computer system, characterized by comprising: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the above method.

[0049] A computer-readable storage medium, characterized by storing computer-executable instructions which, when executed, are used to implement the above method.

[0050] Beneficial effects

[0051] A refined feature recognition method for radar emitter signals based on self-weighted structured feature selection provided by the present invention solves the problem of applying the same weight to the role of refined features in traditional machine learning-based emitter signal recognition methods, simplifies the processing process of complex data, and can utilize the original information of emitter signals and the calculation of imposing row constraints on the transformation matrix to obtain more representative and characterizing signal features in the data, and finally obtains a reliable refined feature recognition effect of radar emitter signals.

[0052] Compared with the existing technologies, the present invention has the following beneficial effects:

[0053] First, it is easy to implement. Compared with the emitter recognition method based on deep learning, the present invention uses unsupervised feature selection in machine learning as the basis for recognition, avoiding the construction of neural networks and the data training process in deep learning. Most importantly, it greatly reduces the number of hyperparameters in the model, avoids the adjustment of hyperparameters, and makes the model easy to implement in actual situations.

[0054] Second, it has a fast convergence speed. Aiming at the deficiencies of the traditional method that requires relying on manual experience to extract refined features of emitter signals and requires category information, the present invention uses an unsupervised model to recognize emitter signals. The model is used to select representative feature vectors, and then clustering analysis is performed based on these selected feature vectors to complete the task of emitter signal recognition, improving the convergence speed, reducing the influence of external factors, and having high robustness.

[0055] Third, it has a high recognition accuracy. The present invention changes the disadvantage of the same role of different features in the data in the past, adopts a self-weighted method, and enables the model to automatically learn the different contribution degrees between refined features. At the same time, the present invention incorporates the idea of maximizing information entropy into the model and imposes The row sparse constraint is adopted, and finally, by combining the construction process of similar graphs and the feature selection process, the discriminability and representational ability of the selected features in unsupervised feature selection can be greatly improved, thereby enhancing the accuracy of radar emitter signal feature recognition and having certain engineering practical value.

[0056] Fourthly, the refined feature recognition technology of radar emitter signals mainly refers to the refined feature analysis and measurement of the received radar electromagnetic signals, and then determining the individual to which the radar emitter that sends the signal belongs according to the internal structure of the signal, and then it can be associated with the individual platform of the emitter, which has important military strategic significance. Brief Description of the Drawings

[0057] The drawings are only for the purpose of illustrating specific embodiments and are not considered to be a limitation of the present invention. Throughout the drawings, the same reference numerals represent the same components.

[0058] Figure 1 is a flowchart of the refined feature recognition system of radar emitter signals of the present invention;

[0059] Figure 2 is a core idea diagram of an unsupervised feature selection algorithm based on self-weighted structure. Detailed Embodiments

[0060] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0061] As Figure 1 shown, the present invention provides a refined feature recognition method for radar emitter signals based on self-weighted structured feature selection, and its specific implementation process is as follows:

[0062] 1. Obtain original data, perform signal data preprocessing and normalization processing

[0063] In reality, the electromagnetic signals received from radar radiation source devices are usually "unclean", and the signals are more or less "polluted" during the processes of acquisition, propagation, and conversion. Especially in military electronic warfare, the captured electromagnetic signals of the opponent are more vulnerable to external factors, resulting in a large number of noise characteristics and singular values in the signal data. Therefore, it is particularly important to perform noise reduction processing on the signal data. The traditional signal denoising method is Fourier transform, but it has a drawback that when using Fourier transform for analysis, its constructed functions are periodic sine waves or cosine waves, which have certain limitations. In signal transmission, if the spectra of the signal and the noise are not in the same frequency band, researchers often use the method of filter filtering to remove the noise. At the same time, wavelet transform can perform more subtle processing on the signal through transformation, and can characterize certain features of the signal, meeting the requirements of analyzing the signal in the time domain and frequency domain. There are usually two types of threshold denoising: soft threshold denoising and hard threshold denoising. The difference lies in the different ways of processing the threshold, and thus different preprocessing results are obtained. The fuzzy maximum value method for denoising can handle the situation where the signal is not entirely continuous, discontinuous at a certain point, or the derivative of a certain order is discontinuous, achieving the purpose of noise reduction. Then we perform signal enhancement and normalization processing on the denoised data to obtain matrix data of the radar radiation source signal that is more conducive to subsequent feature analysis.

[0064] 2. Feature Analysis

[0065] After preprocessing the captured radiation source signals, it enters the important part of the present invention and the innovation of the present invention, namely the feature selection module. Unsupervised learning is applied to the present invention because it can perform a certain degree of learning on data without the need for data annotation. Assume the matrix data obtained from the first step is two-dimensional, where n is the number of samples and d is the dimension of the data. The i-th sample of the data is represented by . The present invention proposes a refined feature recognition method for radiation source signals based on machine learning. The core of this method is double-sparse unsupervised feature selection based on self-weighting and row constraints, as shown in Figure 2 . The specific implementation process is as follows:

[0066] [1] Construct a projection clustering model for adaptive neighbors

[0067] Graph theory is generally considered a tool for describing the local manifold structure in data. When data points are considered as the vertices of graph G, there will always be a similarity matrix corresponding to the adjacency matrix of G. We use the following model to solve the adjacency probability s ij :

[0068]

[0069] where \(W\) is an orthonormal projection matrix obtained by learning.

[0070] [2] Unsupervised Feature Selection Based on Unsupervised Self-Weighting and Row Constrained Double Sparsity

[0071] The objective function of the traditional unsupervised feature selection method for structured graph optimization is:

[0072]

[0073] where \(c\) is the number of data categories, and \(\gamma\) and \(\lambda\) are regularization term parameters.

[0074] What the present invention proposes is an unsupervised feature selection based on self-weighted structure. On the basis of the above-mentioned previous work, considering that the model without adding self-weighting constraints cannot distinguish the contribution degrees of different refined features to the data, is vulnerable to the influence of noise in the data, and this method cannot well solve the subspace sparsity problem. To solve the above problems, the present invention embeds the construction process of the adaptive neighbor graph into the feature selection process, and at the same time introduces the ideas of self-weighting and maximizing information entropy, and integrates the subspace row sparsity \(l\) 2,0 constraint, and designs an objective function to make the selected data features more representative. Therefore, the mathematical model proposed by the present invention is:

[0075]

[0076] where \(m\) is the dimension of the low-dimensional space, \(k\) is the row sparsity of the transformation matrix, is a diagonal matrix whose diagonal elements represent the weights of the sample data.

[0077] [3] Model Optimization

[0078] According to the relationship between the matrix two-norm and the trace operation Equation (13) is equivalent to:

[0079]

[0080] It can be seen that the model (14) contains three variables, namely the weight matrix \(\Theta\), the similarity matrix \(S\) and the projection matrix \(W\). For its solution problem, the present invention designs an iterative optimization algorithm. The specific process is as follows:

[0081] ① Initialization: Initialize the weight matrix where \(I\) d is the identity matrix of dimension \(d\).

[0082] ② Solve for \(\Theta\):

[0083] When the variables \(S\) and \(W\) are fixed, problem (14) is transformed into:

[0084]

[0085] According to the properties of the trace operation, it can be changed to:

[0086]

[0087] Apply the theorem to problem (16), and we have:

[0088]

[0089] Obviously, this is a quadratic problem. Therefore, the present invention uses the augmented Lagrangian multiplier method (ALM) to solve this problem.

[0090] ③ Solve for S:

[0091] When the variables Θ and W are fixed, problem (14) is changed to:

[0092]

[0093] Its Lagrangian function is:

[0094]

[0095] where Ψ = {ψ i | i = 1, 2,..., n} and Δ = {δ ij | i, j = 1, 2,..., n} are Lagrange multipliers. According to the KKT conditions, we can obtain

[0096]

[0097] Also, because of the restriction of the obtained s ij is:

[0098]

[0099] ④ Solve for W:

[0100] When the variables Θ and S are fixed, problem (14) becomes:

[0101]

[0102] We let A = βI - Θ T X T L S XΘ + λΘ, then solving problem (22) becomes solving the following problem:

[0103]

[0104] where β is large enough to ensure that A is a positive semi - definite matrix, so set β as matrix B = Θ T X T L S the largest eigenvalue of XΘ + λΘ. For solving problem (23), we divide it into two cases to solve respectively, which are rank(A) ≤ m and rank(A) > m. The optimization algorithm of the present invention adopts different methods to solve them respectively.

[0105] ⑤ Repeat the above steps ②, ③ and ④ until the model function converges (the convergence condition is |obj t+1 +obj t | ≤ 10 -4 ), at this time, the W obtained in step ④ is the orthogonal projection matrix required by our feature analysis module, and then select the top m features according to the descending order.

[0106] 4. Radiation source identification and classification

[0107] In the above step 3, a set of detailed analysis and feature selection subsets of the refined features of radar emitter signals with strong representation ability can be obtained. Then, based on the learned feature subsets, researchers can use the feasible and effective classification algorithms in this module to perform refined feature recognition on radar emitter signals, and then obtain their categories. Finally, the classification results obtained through the process of the present invention are output for subsequent work. At the same time, the selection of the classification algorithm used in this module is also very important. Common classifiers can be divided into two types: linear classifiers and non-linear classifiers. Linear classifiers mainly include linear discriminant analysis LDA, logistic regression, and support vector machine SVM, etc. Non-linear classifiers mainly include naive Bayes classification, K-nearest neighbor KNN, and decision tree, etc. The Bayesian classifier is a general term for a class of classification algorithms. Based on Bayes' theorem, through the prior probability of an object, the posterior probability, that is, the probability that the object belongs to a certain class, is calculated using Bayes' formula, and the class with the maximum posterior probability is selected as the class to which the object belongs. SVM is a kernel function-based method. It maps feature vectors to a high-dimensional space through certain kernel functions and then establishes a linear discriminant function. The optimal solution is, in a sense, to maximize the distance between the feature vectors closest to the separation surface in the two classes and the separation surface. K-nearest neighbor classification first stores all training samples, and then analyzes (including methods such as voting and calculating weighted sums) the K nearest neighbors around a new sample, and then labels the new sample as the class with the highest frequency among the K nearest neighbors. The decision tree is a non-parametric estimation method. When the variable under study is a qualitative variable, the established decision tree is called a classification tree, and when it is a quantitative variable, it is called a regression tree. In addition, due to the simplicity of the decision tree, it has become the cornerstone of some more useful algorithms. At the same time, the feature matrix of the individual signal of the radar emitter obtained in the above step 3 can also be input into a deep learning neural network for classification recognition, so as to realize the individual recognition of radar emitters. Therefore, in practical applications, we adopt different classifiers according to the different internal structures of the data to achieve better classification results.

[0108] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should all be covered within the protection scope of the present invention.

Claims

1. A refined feature recognition method for radar emitter signals based on self-weighted structured feature selection, characterized in that The steps are as follows: Step 1: Collect the signal data emitted by radar devices of different models, and at the same time intercept the pulse data in the signals to generate a radar emitter signal individual recognition sample set, that is, the original data; Step 2: Perform signal data preprocessing on the original data obtained in Step 1; Step 3: Perform normalization processing on the data that has undergone signal preprocessing in Step 2; Step 4: Construct a refined feature recognition model for radar emitter signals based on self-weighted structured feature selection; The refined feature recognition model for radar emitter signals: Among them is two-dimensional, where n is the number of samples, d is the dimension of the data, and the i-th sample of the data is represented by ; W is the learned orthogonal projection matrix; m is the dimension of the low-dimensional space, k is the row sparsity of the transformation matrix, is a diagonal matrix whose diagonal elements represent the weights of the sample data, s ij is the adjacency probability, and γ and λ are regularization term parameters; Step 5: Use the emitter signal data to perform a running test on the model established in Step 4, input the recognition features into the classifier for classification, so as to obtain the type of the recognized radar emitter signal.

2. The refined feature recognition method for radar emitter signals based on self-weighted structured feature selection according to claim 1, wherein: The preprocessing described in Step 2 includes denoising processing and signal enhancement processing.

3. A refined feature recognition method for radar emitter signals based on self-weighted structured feature selection according to claim 1, characterized in that: The model optimization process in Step 4 is as follows: According to the relationship between the matrix 2-norm and the trace operation The model is equivalent to: It can be seen that problem (1) contains three variables, namely the weight matrix Θ, the similarity matrix S, and the projection matrix W. For its solution problem, an iterative optimization algorithm is designed; the specific process is as follows: ① Initialization: Initialize the weight matrix where I d is the identity matrix of dimension d; ② Solve Θ: When the variables S and W are fixed, problem (1) is transformed into: According to the properties of the trace operation, it can be changed to: Apply the theorem to problem (3), we get: Obviously, formula (4) is a quadratic problem, so the augmented Lagrangian multiplier method is used to solve this problem; ③ Solve S: When the variables Θ and W are fixed, problem (1) is changed to: Its Lagrangian function is: where Ψ = {ψ i | i = 1, 2,..., n} and Δ = {δ ij | i, j = 1, 2,..., n} are Lagrange multipliers; according to the KKT conditions, we can obtain Also because there is restriction, so the obtained s ij is: ④ Solve W: When the variables Θ and S are fixed, problem (1) becomes: Let \(A = \beta I-\Theta\) T X T L S If \(X\Theta+\lambda\Theta\), then solving problem (9) becomes solving the following problem: where β is large enough to ensure that A is a positive semi - definite matrix, so β is set to matrix B = Θ T X T L S the largest eigenvalue of XΘ + λΘ; for the solution of problem (10), it is divided into two cases for separate solutions, namely rank(A) ≤ m and rank(A) > m; ⑤ Repeat the above steps ②, ③, and ④ until the model function converges. The convergence condition is |obj t+1 +obj t | ≤ 10 -4 . At this time, the W obtained in step ④ is the orthogonal projection matrix required by the feature analysis module. Then, select the top m features according to the descending order.

4. A refined feature recognition method for radar emitter signals based on self-weighted structured feature selection according to claim 1, characterized in that: The classifier described in Step 5 includes a linear classifier, a non-linear classifier, and a deep learning neural network.

5. A computer system, characterized in that Including: One or more processors, a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in claim 1.

6. A computer-readable storage medium, characterized in that Stored with computer-executable instructions, the instructions are used to implement the method described in claim 1 when executed.

Images