A radar communication modulation recognition method based on hierarchical classification

Abstract

The present invention discloses a radar communication modulation recognition method based on hierarchical classification, comprising the following steps: S1, acquiring modulation signal data; S2, normalizing the modulation signal data; S3, performing wavelet filtering on the normalized data; S4, performing feature extraction on the filtered data; S5, using the extracted features as input to a support vector machine, and constructing a hierarchical modulation recognition classifier based on the support vector machine. The present invention utilizes wavelet filtering to reduce the impact of the noise environment on the characteristics of the received signal, and employs time domain statistical features to mitigate frequency bias problems caused by carrier frequency estimation deviations. By hierarchically classifying the signal through a support vector machine trained with corresponding features, the accuracy and stability of signal recognition are improved, and good robustness can be achieved under noise and frequency bias conditions.

Description

Technical Field

[0001] The present invention belongs to the field of signal modulation recognition, and is particularly suitable for recognition of radar and communication modulation signals in complex electromagnetic environments. Background Art

[0002] Due to the rapid development of electronic technology, electronic warfare (EW) is becoming a critical component of modern warfare, exerting a significant impact on the battlefield. The development of EW-related technologies is of great significance to the future security and development of the nation. Electronic reconnaissance is a key component of EW and a prerequisite for its organization and implementation. Modern combat deployments of EW require the identification and tracking of intercepted signals on the battlefield to effectively interpret the signal information. Determining the signal modulation form is a prerequisite for extracting the parameters required for signal analysis, providing decisive intelligence guidance for determining precise EW mission targets and jamming methods. This makes signal modulation recognition a crucial research direction in electronic reconnaissance systems within complex electromagnetic environments.

[0003] After years of development and research, modulation recognition algorithms can now be broadly categorized into two main categories: hypothesis testing algorithms based on decision theory and pattern recognition algorithms based on feature extraction. Due to the non-cooperative nature of reconnaissance missions, pattern recognition algorithms based on feature extraction are more effective. These methods primarily include those based on instantaneous features, those based on time-frequency analysis and wavelet transforms, those based on spectral analysis, modulation recognition algorithms based on constellation geometry, and those based on cyclostationary theory.

[0004] Feature extraction-based pattern recognition algorithms consist of three parts: feature extraction, feature selection, and pattern recognition. Feature extraction extracts typical features from signals of different modulation types as a basis for recognition. Its goal is to focus on patterns with distinct categorical differences and minimize the dataset size, thereby improving recognition efficiency and reducing computational complexity. Currently, the features used in feature extraction-based pattern recognition methods primarily include time-domain and frequency-domain features, transient information, high-order cumulants, cyclic spectra, time-frequency information, and constellation diagrams. Designing a classifier based on a small number of features not only improves overall classifier performance when the sample size is small but also simplifies the feature acquisition process in most cases, reducing the cost of the recognition system. After feature extraction, a classifier is designed based on the obtained features. Typical pattern recognition classifiers include decision tree (DT) classifiers, support vector machine (SVM) classifiers, and neural network classifiers. Ultimately, the classifier identifies the category of the modulated signal based on the features.

[0005] The paper "D. Saharia, M.R. Boruah, N.K. Pathak and N. Sarma, "An Ensemble-based Modulation Recognition using Feature Extraction," 2021 International Conference on Intelligent Technologies (CONIT), Hubli, India, 2021, pp. 1-6" proposes a modulation recognition method based on random forests. This method uses a hierarchical structure to determine the threshold of sample features at each node, ultimately obtaining the recognition result. While the decision tree structure is simple and low in complexity, its single threshold-based determination makes it less robust to noise. A misclassification at one node can affect the next node, and the method is less resilient to noise and frequency deviation.

[0006] The paper "D. Shuli, L. Zhipeng and Z. Linfeng," AModulation Recognition Algorithm based on Cyclic Spectrum and SVM Classification," presented at the 2020 IEEE 4th Information Technology, Networking, Electronic and Automation Control Conference (ITNEC), Chongqing, China, 2020, pp. 2123-2127, proposes a modulation recognition method based on a support vector machine. This method selects the second-order cyclic spectrum characteristics of the signal and the statistical moment characteristics of the signal amplitude and implements it using a support vector machine classifier. However, this method only recognizes digitally modulated signals and cannot meet the requirements for recognizing various communication radar modulation signals in complex electromagnetic environments.

[0007] Although various modulation recognition schemes have been proposed in existing literature, traditional modulation recognition algorithms are usually designed to identify specific modulation types. The features extracted and classifiers designed to identify specific modulation types have poor stability when facing the recognition of more signal types. Summary of the Invention

[0008] The present invention aims to overcome the shortcomings of the prior art by providing a radar communication modulation recognition method based on hierarchical classification. During feature extraction, the present invention utilizes wavelet filtering to filter received signals in a noisy environment, mitigating the impact of noise on signal characteristics. A support vector machine trained with the corresponding features performs hierarchical classification of the signals in a decision-tree-like structure, improving the accuracy and stability of signal recognition. This method enables the recognition of various radar intra-pulse modulation and communication signal analog and digital modulation methods under noise and frequency offset conditions.

[0009] The object of the present invention is to achieve the following technical solution: a radar communication modulation recognition method based on hierarchical classification, comprising the following steps:

[0010] S1. Acquire modulation signal data; signal modulation modes include: radar signal pulse modulation CW, LFM, Barker, Costas, Barker / LFM, Barker / Costas; communication signal analog modulation DSB, LSB, USB, FM and digital modulation BASK, 4ASK, BFSK, BPSK, QPSK, 16QAM signals;

[0011] S2, normalizing the modulated signal data;

[0012] S3, performing wavelet filtering on the normalized data;

[0013] S4, extracting features from the filtered data;

[0014] S5. Using the extracted features as the input of a support vector machine, a hierarchical modulation recognition classifier is constructed based on the support vector machine; determining the kernel function of the support vector machine and the characteristic parameters of the classification, and training the classifier.

[0015] The beneficial effects of the present invention are as follows: the radar communication signal modulation recognition method based on hierarchical classification of the present invention utilizes the obtained signal waveform and combines the time and frequency domain characteristics of the signal to establish a modulation recognition model based on multi-dimensional features. During feature extraction, wavelet filtering is used to filter the received signal under the influence of the noise environment to reduce the impact of noise on the signal characteristics; the signal is hierarchically classified into a decision tree structure by a support vector machine trained with corresponding features to improve the accuracy and stability of signal recognition, and realize the recognition of various radar intra-pulse modulation and communication signal analog and digital modulation methods under noise and frequency bias conditions. The present invention completes the effective recognition of the modulation types of various radar signals and communication signals, which not only reduces the complexity of the classification structure, but also improves the accuracy and stability of signal recognition, and has important application prospects in the field of radar and communication modulation recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1Flowchart of the radar communication signal modulation identification scheme provided by the present invention;

[0017] Figure 2 A schematic diagram comparing wavelet denoising results provided by an embodiment of the present invention;

[0018] Figure 3 A schematic diagram of the structure of a hierarchical modulation recognition classifier provided by an embodiment of the present invention;

[0019] Figure 4 A diagram showing the modulation signal recognition results provided by an embodiment of the present invention;

[0020] Figure 5 This is a graph showing the accuracy of signal recognition under different noise conditions according to an embodiment of the present invention;

[0021] Figure 6 This is a graph showing the signal recognition accuracy under different frequency offsets according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to solve the problem of comprehensive identification of radar and communication radiation sources in complex electromagnetic environments and meet the demand for stable signal recognition of signals of various modulation types, the present invention uses a support vector machine classifier to perform wavelet denoising and multidimensional feature extraction on the received signal, and outputs the recognition results through a hierarchical support vector machine classification structure.

[0023] In the present invention, wavelet analysis is used to filter the received signal under the influence of a noisy environment to reduce the impact of noise on signal characteristics; time domain statistical features are used to reduce the frequency offset problem caused by carrier frequency estimation deviation and the like; at the same time, a support vector machine trained with corresponding features is used to perform hierarchical classification of the signal in a decision tree structure to improve the accuracy and stability of signal recognition, realize the recognition of various radar intra-pulse modulation and communication signal analog and digital modulation methods, and can recognize signals under noise and frequency offset conditions.

[0024] In order to facilitate the description of the present invention, the following terms are explained herein.

[0025] Term 1: Hierarchical classification

[0026] A decision tree is a tree-like classification structure based on decision theory, which classifies data hierarchically from root to leaf. The root is the input of a set of feature parameters for the sample to be classified. Each layer of the decision tree has multiple nodes, and each node has a classification rule based on the predicted value of the feature. When the data meets the classification rule, it is divided into the corresponding node in the next layer of the decision tree for classification. The final classification results in multiple leaves of the decision tree, each corresponding to a known, specified category. The hierarchical classification recognition algorithm used in this paper refers to the structure of the decision tree. To address the problem that decision tree classification is susceptible to noise, the nodes are replaced with support vector machine classifiers.

[0027] The technical solution of the present invention is further described below with reference to the accompanying drawings.

[0028] like Figure 1 As shown, a radar communication modulation recognition method based on hierarchical classification of the present invention comprises the following steps:

[0029] S1. Acquire modulated signal data; the modulated signal is a time domain signal. The signal modulation methods used in this embodiment include: radar signal pulse modulation CW, LFM, Barker, Costas, Barker / LFM, Barker / Costas; communication signal analog modulation DSB, LSB, USB, FM and digital modulation BASK, 4ASK, BFSK, BPSK, QPSK, 16QAM signals.

[0030] S2. Normalize the modulated signal data. To reduce the influence of noise and other factors on the recognition results, normalize the received signal x(t) by the maximum absolute value of the signal, i.e. x(t) / ||x(t)|| max In this embodiment, the normalized data set is divided into a training set and a test set in a ratio of 3:1, which are used to train the model and test the recognition effect of the model respectively.

[0031] S3. Perform wavelet filtering on the normalized data to reduce the impact of noise on the signal. Wavelet filtering achieves multi-scale decomposition and reconstruction of the signal by selecting appropriate wavelet basis functions and scale parameters. The series decomposition and reconstruction formula of wavelet filtering is:

[0032]

[0033] Where J is the number of decomposition series, W j,k is the jth level, kth wavelet coefficient, ψ j,k(t) is the j-th level and k-th wavelet function, which is the version of the mother wavelet ψ(t) at a specific scale j and translation k. As can be seen from formula (1), by weighted summation of wavelet functions at different scales and positions, the original signal can be reconstructed or filtered. By setting a threshold, certain frequency components of the signal can be selectively retained or removed, thereby achieving the purpose of denoising and filtering. The threshold denoising method adopted is soft threshold denoising based on the minimum mean square error criterion, and the Daubechies wavelet basis function is used as the mother wavelet for wavelet transform, and the wavelet order is 4. The Daubechies wavelet has good regularity, that is, the smoothing error introduced by the wavelet as a sparse basis is not easily detected, making the signal reconstruction process relatively smooth. The time domain contrast of the signal after wavelet filtering is as follows: Figure 2 As shown in Figure 2, the result obtained by extracting the time domain features of the signal is less affected by noise.

[0034] S4. Perform feature extraction on the filtered data. Feature extraction is a method of projecting high-dimensional data into a low-dimensional space to represent samples using features. It is often used when the original data has a high dimension and a large number of samples.

[0035] The extracted features include:

[0036] (1) Duty factor: The duty factor of a signal is expressed as the pulse time divided by the signal period. Let the simulation signals be x1(t) and x2(t), and the signal sampling length be N. Follow the steps below to solve the duty factor: First, select one of the signals x i (t), i = 1, 2, and calculate the amplitude mean μ; use the mean to cut off the part of the signal below μ, and repeat the operation four times; calculate the amplitude mean again, and count the number m of points in the signal with amplitude greater than the mean; calculate the duty factor a = 1-m / N; a reflects the difference in waveforms generated by radar signal pulse modulation and communication signal modulation. The duty factor of the radar signal pulse modulation signal is significantly smaller than that of the communication modulation signal.

[0037] (2) The maximum value of the normalized zero-centered instantaneous amplitude spectrum; the maximum value of the normalized zero-centered instantaneous amplitude spectrum is a statistical characteristic quantity based on the instantaneous amplitude, and its mathematical expression is:

[0038]

[0039] Among them, N is the number of sampling points, A cn (i) is the normalized zero-centered instantaneous amplitude; γ maxThis characteristic measures the degree of change in instantaneous amplitude and reflects the fluctuation of the envelope. It can be used to distinguish between modulation modes with stable and unstable envelope values. Radar modulation signals experience significant envelope fluctuations due to pulse modulation. Frequency modulation (FM) and FSK (including BFSK) exhibit no envelope fluctuations, while ASK (including BASK and 4ASK) exhibits subtle but noticeable envelope fluctuations.

[0040] (3) The first-order absolute origin moment of the normalized zero-centered instantaneous amplitude:

[0041]

[0042] Among them, A cn (i) = A n (i) -1; absEnv also reflects the changes in instantaneous amplitude and signal envelope. The ASK amplitude modulation signal and radar modulation signal reflected by this characteristic parameter have large envelope changes, while frequency modulation has no envelope fluctuations.

[0043] (4) Standard deviation of the instantaneous amplitude of the normalized zero-centered non-weak signal segment:

[0044]

[0045] Among them, σ da Characterizes the change of the instantaneous amplitude of the signal; σ da The signal characteristics reflected by the characteristic parameter are similar to those of the first two characteristic parameters.

[0046] (5) Signal envelope stability: The signal envelope stability is based on the time domain waveform characteristics obtained from the filtered signal, and its mathematical expression is:

[0047]

[0048] A s Reflects the changes in instantaneous amplitude and signal envelope; A s The larger the value, the greater the amplitude change, which can show the signal amplitude characteristics of the pulse radar signal pulse modulation method. The amplitude change of analog amplitude modulation is large, resulting in envelope fluctuations. Frequency modulation has no envelope fluctuations. The radar signal pulse modulation A s Smaller.

[0049] Since the frequency offset has little effect on the signal amplitude feature, the amplitude feature should be used in the reclassification. The feature extraction is performed on the normalized signal obtained in step S2. max 、absEnv、σ da Based on , the following features are added:

[0050] (6) The first-order absolute central moment of the phase of the zero-centered non-weak signal segment:

[0051]

[0052] Among them, θ(i) represents the phase of the strong signal; absPhase reflects the change of the phase. QPSK signals are obviously different from other signals, and this feature is less affected by the signal-to-noise ratio.

[0053] (7) The first-order absolute central moment of the absolute value of the phase of the zero-centered non-weak signal segment:

[0054]

[0055] absPhase2 also reflects the phase change, and can distinguish amplitude-modulated DSB, ASK, and BPSK signals that contain phase information.

[0056] (8) Standard deviation of the instantaneous phase nonlinear component of the zero-centered non-weak signal segment:

[0057]

[0058] σ dp Characterizes the change in the instantaneous phase of the signal. It can be used to distinguish modulations that contain direct phase information (such as BPSK, LSB, USB) from modulations that do not contain direct phase information (such as AM, BASK).

[0059] (9) The first-order absolute origin moment of the normalized instantaneous frequency of the zero-centered non-weak signal segment:

[0060]

[0061] in, f(i) is the instantaneous frequency of the signal. When the signal is analog modulated, f(i) is the modulating signal frequency; when the signal is digitally modulated, f(i) is the symbol rate. Under ideal noise-free conditions, absFreq is non-zero for BFSK and FM signals. On the other hand, for BPSK signals, since f(i) is always zero, absFreq is also zero.

[0062] (10) Standard deviation of the instantaneous frequency of the normalized zero-centered non-weak signal segment:

[0063]

[0064] σ af Characterizes the change of the instantaneous frequency of the signal.

[0065] (11) Spectral symmetry:

[0066]

[0067] Among them, SR , S L The sum of the moduli of the Fourier transform of the left and right halves of the signal, respectively. absSymm measures the symmetry of a signal about the carrier frequency. This characteristic parameter can be used to separate USB and LSB signals from other signals.

[0068] S5, using the extracted features as the input of the support vector machine, constructing a hierarchical modulation recognition classifier based on the support vector machine; determining the kernel function of the support vector machine and the characteristic parameters of the classification, and training the classifier; the hierarchical modulation recognition classifier of the present invention includes four support vector machines SVM1, SVM2, SVM3, and SVM4, the input of SVM1 is the characteristic signal of the modulation signal of S4, and the output is connected to SVM2, SVM3, and SVM4 respectively, as shown in FIG. Figure 3 As shown in Figure 2, each support vector machine is trained separately using the corresponding training set and feature parameters.

[0069] The training method of support vector machine is:

[0070] S51. Support vector machines are a data-oriented classification algorithm whose goal is to determine a hyperplane that separates different data. First, a nonlinear transformation is performed to project the input vector into a high-dimensional space. Then, a search is performed in this new space to ultimately find an optimal classification surface that accurately separates the two classes. An optimal classification surface means that it can correctly distinguish between the two classes of samples while maintaining the maximum classification distance between them.

[0071] Let H be the classification surface, and planes H1 and H2 are parallel to H. The distance between H1 and H2 is called the classification margin, and they pass through sample points of the corresponding category while minimizing the distance from H. The mathematical formula for the classification surface is x·w+b=0, where x is the eigenvector of the input data, representing a point in feature space; w is the normal vector; and b is the bias term used to determine the position of the classification surface. The classification surface is normalized so that the sample set for which the classification samples are linearly separable satisfies the condition:

[0072] y i [(w·x i )+b]-1≥0, i=1,..,n (12)

[0073] where y i Represents each eigenvector x of the i-th i The corresponding label or output. During the SVM training process, the goal is to find the optimal w and b that can maximize the separation of data points of different categories while keeping the interval as large as possible. According to the formula, the classification interval is 2 / ||w||. Finding the maximum classification interval is to find ||w|| 2 Minimum value. That is, formula (12) is satisfied and the classification surface is let ||w|| 2When / 2 reaches the minimum value, the optimal classification surface is obtained, and the support vector is the data sample point located on the interval boundaries H1 and H2.

[0074] Finding the maximum classification margin, one of the core principles of SVM, is a way to control its generalization ability. VC dimension is a concept used to measure the complexity of a classification model, describing the maximum number of data points that the model can separate. According to VC dimension theory, assuming that the samples are distributed within a sphere with a diameter of 2R in a high-dimensional space of dimension N, then the VC dimension of a high-dimensional plane set f(x,w,b)=sgn{(w·x)+b} that satisfies ||w||≤A satisfies the following condition:

[0075] h≤min([R 2 A 2 ],N)+1 (13)

[0076] Where h is the VC dimension of the model, A represents the size of the hypothesis space; R 2 A 2 It represents the product of the complexity of the hypothesis space and the square of the size of the hypothesis space, and is used to measure the overall complexity of the entire hypothesis space. The formula provides a theoretical limit for the model. Finding the minimum upper limit of the VC dimension can give the minimum ||w|| 2 ; Using the Lagrangian optimization method to process (13), equations (14) and (15) are the conditions required for solution:

[0077]

[0078] T i ≥0, i=1,...,n (15)

[0079] Solve the maximum value of formula (16):

[0080]

[0081] Where T i is a Lagrange multiplier, corresponding to each sample; this is a quadratic function optimization problem with only one solution, subject to conditional constraints. It is easy to prove that the solution T corresponding to the support vector sample i It is not zero and generally only accounts for a small part of the solution.

[0082] The optimal classification function expression is finally obtained:

[0083]

[0084] Only the samples corresponding to the support vectors are summed in formula (17). The classification threshold b can be obtained by taking the median of any support vector or the support vector samples corresponding to two categories. * ;

[0085] For the nonlinear separability condition, a relaxation term a is added in (12) i ≥0, becomes:

[0086] y i [(w·x)+b]-1+a i ≥0, i=1,...,n (18)

[0087] At this point, we need to comprehensively consider how to maximize the classification interval while maintaining the accuracy of sample classification. The minimum value is obtained by solving the generalized optimal classification surface. Among them, C>0 is the penalty factor, which is responsible for adjusting the penalty degree corresponding to the wrong classification and the correct classification. Changing its value will affect the classification accuracy. The Lagrangian optimization method is used to perform a dual transformation on the generalized optimal classification surface problem. The basic principle is similar to the linear case, except that condition (15) becomes:

[0088] 0≤T i ≤C, i=1,..,n (19)

[0089] According to the conclusion of Equation (13), for linear functions in N-dimensional space, their VC dimension will be significantly reduced under the condition of ||w||≤A, ensuring the stability of generalization and obtaining a set of functions with lower VC dimension in non-low-dimensional space. In addition, after Lagrangian optimization of the original problem, the algorithm complexity no longer increases with the increase of spatial dimension, but increases with the increase of support vectors. This allows us to effectively handle and solve high-dimensional problems.

[0090] In order to solve the optimal classification surface of nonlinear classification problems, the nonlinear problem can be transformed into a linear problem through mapping, thereby solving the nonlinear problem in a high-dimensional space. This transformation method is complex and difficult to implement directly. However, it can be obtained from the dual transformation formula that only the inner product operation is actually required in the transformed space, because the optimization function and the classification function only include the direct operation of the inner product in the formula (x i ·x j ), so that nonlinear problems can be solved directly using functions in the original dimension without having to solve complex dimensional transformation methods. According to the relevant theories of functional analysis, to correspond to the inner product in a certain high-dimensional space, the kernel function K(x i ,x j ) must meet Mercer conditions.

[0091] Use appropriate inner product function K(x i ,x j ) to construct the optimal classification surface, transforming the nonlinear classification problem into a linear classification problem without increasing the amount of computation. At this point, the objective function will be rewritten as:

[0092]

[0093] The corresponding classification function also becomes:

[0094]

[0095] S52, using radial basis function as the SVM classifier kernel function, the radial basis function is:

[0096]

[0097] The classification decision function corresponding to SVM is

[0098]

[0099] Where 1 / σ 2 The radial basis function (RBF) is a kernel function that is used for classification. This is because the dimensionality of the high-dimensional space corresponding to the RBF can be infinite, and any finite number of samples projected into the high-dimensional space by the inner product corresponding to the RBF is linearly separable.

[0100] The choice of penalty factor C and kernel function parameter σ has a significant impact on the generalization and performance of the support vector machine. The value of C represents the tolerance for misclassification; larger values ​​indicate lower tolerance. The value of σ represents the correlation between support vectors; larger values ​​indicate stronger correlation. Excessively large or low values ​​of C and σ will result in reduced recognition accuracy. Therefore, choosing the right C and σ is crucial for ensuring the performance of the SVM classifier.

[0101] C and σ can be set empirically or through cross-validation. This method uses cross-validation to train the SVM. A subset of the features extracted from the generated simulation samples is selected for SVM training. The remaining subset is used to test the model generated from the training set and determine the recognition accuracy. This process is repeated several times, and the average recognition rate under this parameter setting is calculated. The recognition rates under different parameter settings are then compared. Ultimately, the optimal performance of the generated classification model is determined, and the appropriate kernel function parameter settings are determined. In this embodiment, the parameters of the classifier kernel function selected are C = 2000 and σ = 0.9.

[0102] S53, the modulation recognition of multiple signals realized by the present invention requires a multi-classification classifier, and support vector machines are generally used for binary classification problems, so it is necessary to improve the traditional binary classification SVM to realize SVM multi-class classification. Multi-classifiers can be obtained by combining binary classification support vector machines. Commonly used promotion methods include one-to-one (OAO) method and one-to-many (OAA) method. Under the one-to-one (OAO) method, for N classification problems, it is necessary to train and construct N(N-1) / 2 SVM classifiers, select two types of samples from N classes without repetition and use their features to form a training set to train each SVM classifier, and input the sample to be classified into the multi-classifier when classifying the test set. Each SVM will determine its category, and finally, all SVM classification results are combined. The category that is determined the most times is the output category of the multi-classifier for this sample. The Libsvm support vector machine classification tool based on the Matlab platform used by the present invention to process the multi-classification problem uses the OAO classification method.

[0103] S54. Use the training data obtained in step S4 as the input of the support vector machine constructed in step S5. Use features (1)-(5) extracted in step S4 to train SVM1. The initial signal set can be divided into three categories. The first category includes DSB, LSB, USB, BASK, 4ASK, BPSK, QPSK, and 16QAM; the second category includes FM and BFSK; and the third category includes CW, LFM, Barker, Costas, Barker / LFM, and Barker / Costas. Then, use features (6)-(11) extracted in step S4 to train SVM2, SVM3, and SVM4, respectively, for the corresponding three types of signal data sets.

[0104] Table 1 shows the communication modulation signal dataset used for training and testing the classifiers in this embodiment. It includes 10 baseband signals with different modulation types and signal bandwidths, including DSB, LSB, USB, FM, 2ASK, 4ASK, 2FSK, 2PSK, QPSK, and 16QAM. The bandwidth ranges from 1 MHz to 60 MHz, with an interval of 0.3 MHz. The sampling frequency is 250 MHz. Table 2 shows the radar modulation signal dataset used for training and testing the classifiers in this embodiment. It includes 6 baseband signals with different modulation types and signal bandwidths, including CW, LFM, Barker, Costas, Barker / LFM, and Barker / Costas. The bandwidth ranges from 1 MHz to 60 MHz, with an interval of 0.3 MHz. The sampling frequency is 250 MHz. The pulse width ranges from 0.8 to 20 µs, with an interval of 4.8 µs.

[0105] Gaussian white noise was added to all simulated signals, with a signal-to-noise ratio ranging from 0 to 16 dB in 2 dB increments. The frequency offset ranged from 0 to 20 kHz in 5 kHz increments. The generated modulated signals were independent of each other, with each segment consisting of 8192 samples. Training and test samples were selected from the generated signals. The training set contained 36,000 samples, and the test set contained 12,000 samples. All data were generated using MATLAB 2018a.

[0106] When SNR=8dB, the overall sorting success rate of the method proposed in the embodiment of the present invention can reach 99.42%. The confusion matrix is ​​as follows: Figure 4 shown. Figure 5 The signal recognition accuracy diagram under different noise conditions when the bias is zero, Figure 6 This is a graph showing the signal recognition accuracy under different biases when the noise is 8dB.

[0107] Table 1 Communication modulation signal dataset parameter correspondence table

[0108]

[0109] Table 2 Radar modulation signal dataset parameter correspondence table

[0110]

[0111] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A radar communication modulation recognition method based on hierarchical classification, characterized in that: The following steps are involved: S1. Obtain modulated signal data; Signal modulation methods include: radar signal pulse modulation CW, LFM, Barker, Costas, Barker / LFM, Barker / Costas; communication signal analog modulation DSB, LSB, USB, FM and digital modulation BASK, 4ASK, BFSK, BPSK, QPSK, 16QAM signals; S2, normalizing the modulated signal data; S3, performing wavelet filtering on the normalized data, wherein the wavelet filtering realizes multi-scale decomposition and reconstruction of the signal by selecting wavelet basis functions and scale parameters; S4. Extract features from the filtered data; the extracted features include: (1) duty cycle; (2) maximum value of the normalized zero-centered instantaneous amplitude spectrum; (3) first-order absolute origin moment of the normalized zero-centered instantaneous amplitude; (4) standard deviation of the normalized zero-centered non-weak signal segment instantaneous amplitude; (5) signal envelope stationarity; (6) first-order absolute central moment of the zero-centered non-weak signal segment phase; (7) first-order absolute central moment of the absolute value of the zero-centered non-weak signal segment phase; (8) standard deviation of the nonlinear component of the zero-centered non-weak signal segment instantaneous phase; (9) first-order absolute origin moment of the normalized instantaneous frequency of the zero-centered non-weak signal segment; (10) standard deviation of the normalized zero-centered non-weak signal segment instantaneous frequency; (11) spectrum symmetry; S5. Using the extracted features as input to a support vector machine, and constructing a hierarchical modulation recognition classifier based on the support vector machine; determining a kernel function of the support vector machine and characteristic parameters of the classification, and training the classifier; The hierarchical modulation recognition classifier includes four support vector machines SVM1, SVM2, SVM3, and SVM4. The input of SVM1 is the characteristic signal of the modulation signal of S4, and the output is connected to SVM2, SVM3, and SVM4 respectively; The features (1)-(5) extracted in step S4 are used to train SVM1, and the initial signal set is divided into three categories. The first category includes DSB, LSB, USB, BASK, 4ASK, BPSK, QPSK, and 16QAM; the second category includes FM and BFSK; and the third category includes CW, LFM, Barker, Costas, Barker / LFM, and Barker / Costas. The features (6)-(11) extracted in step S4 are used to train SVM2, SVM3, and SVM4, respectively.

Images