Modulation classification method based on complex-valued cumulative quantity neural network

Abstract

The application provides a modulation classification method based on a complex-valued cumulant neural network, which comprises the following steps: step A, setting the structure of the complex-valued cumulant neural network; step B, using a cumulant artificial feature extraction layer to respectively calculate the second-order cumulant complex-valued signal and the fourth-order cumulant complex-valued signal of the input intermediate frequency signal r, to obtain the second-order cumulant complex-valued signal C 20 、 21 and the fourth-order cumulant complex-valued signal C 40 、 41 、 42 ; step C, horizontally splicing the input intermediate frequency signal r, the second-order cumulant complex-valued signal C 20 、 21 , and the fourth-order cumulant complex-valued signal C 40 、 41 、 42 as column vectors respectively, and merging them into a cumulant artificial feature matrix used for modulation classification; step D, inputting the cumulant artificial feature matrix into a complex-valued deep feature extraction layer to obtain deep features; and step E, inputting the deep features into a feature classification layer to perform modulation classification and output a signal modulation result. Through complex-valued deep feature extraction, the application retains the phase information in the cumulant complex-valued signal which is beneficial to modulation classification, and realizes the improvement of the signal modulation classification accuracy.

Description

Technical Field

[0001] This application relates to the field of electronic reconnaissance technology, and in particular to a modulation classification method based on a complex-valued cumulant neural network. Background Technology

[0002] Modulation classification techniques are widely used in electronic reconnaissance and cognitive radio systems, enabling non-cooperative signal demodulation between electronic reconnaissance receivers and cognitive radio receivers. Traditional modulation classification methods are mostly based on maximum likelihood classification, which performs poorly in unknown environments.

[0003] In recent years, modulation classification methods based on artificial or deep features have gradually emerged due to their versatility across different application environments. Commonly used artificial features include time-domain signals, time-frequency diagrams, constellation diagrams, and higher-order cumulant eigenvalues, while deep features refer to signal features extracted by neural networks. Among these, time-domain signals are poorly adaptable to low signal-to-noise ratio environments; time-frequency diagrams are only suitable for frequency-modulated signals; constellation diagrams are only suitable for amplitude-phase modulated signals and require high-quality preprocessing for signal symbol synchronization; while higher-order cumulants have the advantage of low noise sensitivity, but in modulation classification applications, they are generally used in the form of eigenvalues, resulting in limited information and poor performance when classifying higher-order modulation types; neural networks are generally used to process real-valued signals, and for ease of use, the complex-valued signal to be classified is split into two real-valued signals, one with a real part and the other with an imaginary part, thus losing some phase information and hindering the improvement of modulation classification accuracy. Summary of the Invention

[0004] This application provides a modulation classification method based on a complex-valued cumulant neural network, which can be used to solve the technical problem of low modulation classification accuracy.

[0005] This application provides a modulation classification method based on a complex-valued cumulant neural network, the method comprising:

[0006] Step A: Define the structure of the complex-valued cumulant neural network; the complex-valued cumulant neural network includes a cumulant artificial feature extraction layer, a complex-valued deep feature extraction layer, and a feature classification layer;

[0007] Step B involves using a cumulant artificial feature extraction layer to calculate the second-order and fourth-order cumulant complex values ​​for the input intermediate frequency signal r, respectively, to obtain the second-order cumulant complex value signal C. 20 C 21 With the fourth-order cumulant complex-valued signal C 40 C 41 C 42 Among them, C 20 Let C be a second-order self-accumulator complex signal of r. 21 For r and r *The second-order cross-cumulant complex signal, C 40 Let C be a fourth-order self-accumulator complex signal of type r. 41 Let r, r, r and r * The fourth-order cross-cumulant complex signal, C 42 For r, r, r * With r * The fourth-order cross-cumulant complex signal, r * This represents the conjugate signal of the input intermediate frequency signal r;

[0008] Step C involves combining the input intermediate frequency signal r with the second-order cumulant complex-valued signal C. 20 C 21 Fourth-order cumulant complex-valued signal C 40 C 41 C 42 The columns are horizontally concatenated and merged into a cumulative artificial feature matrix F used for modulation classification. C ;

[0009] Step D, accumulating the artificial feature matrix F C Input the complex-valued deep feature extraction layer to obtain the deep features F D ;

[0010] Step E, transfer the depth features F D The input feature classification layer performs modulation classification and outputs the signal modulation result.

[0011] Further, step A, setting the structure of the complex-valued cumulant neural network; including:

[0012] Among them, the cumulative artificial feature extraction layer is used to calculate the cumulative complex value signal of the intermediate frequency signal r, and combine the cumulative complex value signal into the cumulative artificial feature matrix;

[0013] The complex-valued deep feature extraction layer consists of seven concatenated complex convolution-max pooling modules; each complex convolution-max pooling module group consists of one complex convolution module and one concatenated max pooling module.

[0014] The feature classification layer is configured with 7 complex convolution-max pooling modules to concatenate and expand the resulting features, and then uses a fully connected layer to classify M modulation types.

[0015] The feature classification layer consists of multiple fully connected modules.

[0016] Furthermore, the cumulative quantity complex value signal is a signal composed of corresponding cumulative quantity characteristic values ​​under multiple time delay conditions of the signal, and each signal value in the cumulative quantity signal is a complex value.

[0017] Furthermore, the second-order cumulative complex value signal C in the cumulative artificial feature extraction layer... 20 C 21 With the fourth-order cumulant complex-valued signal C 40 C 41 C 42 The calculation method is as follows:

[0018]

[0019]

[0020] Where Cum(·) represents the cumulative characteristic value calculation function, r (k) This represents the signal obtained by delaying the input intermediate frequency signal r by k sampling points, where L is the number of sampling points of the input intermediate frequency signal r. The delay of each cumulative complex signal increases from -L to L in steps of 1.

[0021] The Cum(·) cumulative characteristic value calculation function is expressed as follows:

[0022]

[0023] The cumulative amount of variable x is represented as Cum[x1,x2,…,x] k ], which is c x (I), where k is the order of the cumulant, and the variable x includes k signals x1, x2, ..., xk. k Let I = {1,2,…,k} be the subscript index set of x, and let l be the moment of order l. Set I has q groups. p Let I be the p-th group, ∑ be the summation operation, ∪ be the union operation, ! be the factorial operation, Π be the product operation, and E(·) be the expectation operation.

[0024] The method provided in this application combines low-noise-sensitive artificial features of cumulants with deep features to construct a complex-valued cumulant neural network. It uses a cumulant artificial feature extraction layer to extract the low-noise-sensitive complex-valued cumulant signal as the basis for artificial features, and a complex-valued deep feature extraction layer to further extract the deep features of the complex-valued cumulant signal, thereby achieving accurate modulation type classification. Compared with existing technologies, the most significant advantage of this invention is that by extracting the complex-valued cumulant signal, the dimension of cumulant features is expanded, preserving the low noise sensitivity of cumulant features while avoiding the deficiency of low information content when cumulants are used in feature value form; simultaneously, through complex-valued deep feature extraction, phase information in the complex-valued cumulant signal that is beneficial for modulation classification is preserved, thus improving the accuracy of signal modulation classification. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the modulation classification process of the present invention;

[0026] Figure 2 This is a schematic diagram of the complex-valued cumulant neural network model structure provided in an embodiment of the present invention;

[0027] Figure 3 This is a comparison chart of the classification accuracy of the embodiments of the present invention with other typical modulation classification methods. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0029] The embodiments of this application will now be described in conjunction with the accompanying drawings.

[0030] The method provided in this application combines low-noise-sensitive artificial features of cumulants with deep features to construct a complex-valued cumulant neural network. To avoid the limitation of limited information content when cumulants are used in the form of feature values, the method expands the cumulant feature dimension by using a cumulant artificial feature extraction layer within the complex-valued cumulant neural network to extract complex-valued cumulant signals, which also possess low noise sensitivity, as the basis for artificial features. Simultaneously, considering the positive role of phase information in modulation classification, a complex-valued deep feature extraction layer is used to retain the phase information in the complex-valued cumulant signals, further extracting and classifying the deep features of the complex-valued cumulant signals, effectively improving the accuracy of signal modulation classification.

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0032] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0033] See Figure 1 ,like Figure 1 The diagram shown is a schematic of the modulation classification process.

[0034] Step A: Define the structure of the complex-valued cumulant neural network; the complex-valued cumulant neural network includes a cumulant manual feature extraction layer, a complex-valued deep feature extraction layer, and a feature classification layer;

[0035] Among them, the cumulative artificial feature extraction layer is used to calculate the cumulative complex value signal of the intermediate frequency signal r, and combine the cumulative complex value signal into the cumulative artificial feature matrix;

[0036] The complex-valued deep feature extraction layer consists of seven concatenated complex convolution-max pooling modules; each complex convolution-max pooling module group consists of one complex convolution module and one concatenated max pooling module.

[0037] The feature classification layer is configured with 7 complex convolution-max pooling modules to concatenate and expand the resulting features, and then uses a fully connected layer to classify M modulation types.

[0038] The following explains the differences between complex-valued neural networks and ordinary neural networks, specifically the differences between complex-valued deep feature extraction layers and ordinary deep feature extraction layers.

[0039] Ordinary neural networks are built on real numbers; neuron states, connection weights, and activation functions are all in real-valued form. In contrast, complex-valued neural networks use complex-valued forms for neuron states, connection weights, and activation functions. Therefore, complex-valued neural networks can utilize the phase information of their input complex-valued signals, giving them a significant advantage in processing complex-valued data, such as modulation classification signal data.

[0040] The complex convolution operation formula in the complex convolution module of the complex value deep feature extraction layer is as follows:

[0041] W c *s c =(A c *x c -B c *y c )+i(B c *x c +A c *y c )

[0042] Among them, W c W represents the complex convolution kernel. c =A c +iB c A c B represents the real part of the complex convolution kernel. c s represents the imaginary part of the complex convolution kernel. c s represents the input signal of the complex convolution module. c =x c +iy c x c y represents the real part of the input signal. c The input signal is represented by the imaginary part, * indicates convolution, and i represents the imaginary unit.

[0043] Step B: Using the cumulant artificial feature extraction layer, calculate the second-order and fourth-order cumulant complex values ​​for the input intermediate frequency signal r, respectively, to obtain the second-order cumulant complex value signal C.20 C 21 With the fourth-order cumulant complex-valued signal C 40 C 41 C 42 Based on the theory of signal cumulants and combined with a large number of experimental results, we chose to simultaneously value the second-order cumulant complex signal C. 20 C 21 With the fourth-order cumulant complex-valued signal C 40 C 41 C 42 The cumulative artificial feature extraction layer used in complex-valued cumulative neural networks can effectively improve the modulation classification accuracy.

[0044] The cumulative quantity artificial feature extraction layer uses a second-order cumulative quantity complex-valued signal C. 20 C 21 With the fourth-order cumulant complex-valued signal C 40 C 41 C 42 , where C 20 Let C be a second-order self-accumulator complex signal of r. 21 For r and r * The second-order cross-cumulant complex signal, C 40 Let C be a fourth-order self-accumulator complex signal of type r. 41 Let r, r, r and r * The fourth-order cross-cumulant complex signal, C 42 For r, r, r * With r * The fourth-order cross-cumulant complex signal, r * This represents the conjugate signal of the input intermediate frequency signal r;

[0045] The cumulative quantity complex value signal is a signal composed of corresponding cumulative quantity characteristic values ​​under multiple time delay conditions of the signal, and each signal value in the cumulative quantity signal is a complex value.

[0046] The second-order cumulative complex value signal C in the cumulative artificial feature extraction layer 20 C 21 With the fourth-order cumulant complex-valued signal C 40 C 41 C 42 The calculation method is as follows:

[0047]

[0048]

[0049] Where Cum(·) represents the cumulative characteristic value calculation function, r (k)This represents the signal obtained by delaying the input intermediate frequency signal r by k sampling points, where L is the number of sampling points of the input intermediate frequency signal r. The delay of each cumulative complex signal increases from -L to L in steps of 1.

[0050] The Cum(·) cumulative characteristic value calculation function is expressed as follows:

[0051]

[0052] The cumulative amount of variable x is represented as Cum[x1,x2,…,x] k ], which is c x (I), where k is the order of the cumulant, and the variable x includes k signals x1, x2, ..., xk. k Let I = {1,2,…,k} be the subscript index set of x, and let l be the moment of order l. Set I has q groups. p Let I be the p-th group, ∑ be the summation operation, ∪ be the union operation, ! be the factorial operation, Π be the product operation, and E(·) be the expectation operation.

[0053] Step C: Combine the input intermediate frequency signal r with the second-order cumulant complex value signal C. 20 C 21 Fourth-order cumulant complex-valued signal C 40 C 41 C 42 The columns are horizontally concatenated and merged into a cumulative artificial feature matrix F used for modulation classification. C ;

[0054] Step D: Accumulate the artificial feature matrix F C Input the complex-valued deep feature extraction layer to obtain the deep features F D ;

[0055] The complex-valued deep feature extraction layer includes multiple complex convolution modules and pooling modules.

[0056] Step E: Transfer the depth features F D The input feature classification layer performs modulation classification and outputs the signal modulation result.

[0057] The feature classification layer includes multiple fully connected modules.

[0058] The beneficial effects of this invention are as follows: This invention provides a modulation classification method based on a complex-valued cumulant neural network. It combines low-noise-sensitive artificial features of cumulants with deep features to construct a complex-valued cumulant neural network. An artificial feature extraction layer extracts complex-valued cumulant signals with low noise sensitivity as the basis for artificial features, and a complex-valued deep feature extraction layer further extracts the deep features of the complex-valued cumulant signals to achieve accurate modulation type classification. Compared with existing technologies, the most significant advantage of this invention is that by extracting complex-valued cumulant signals, the dimension of cumulant features is expanded. While retaining the low noise sensitivity of cumulant features, it avoids the deficiency of low information content when cumulants are used in the form of feature values. Simultaneously, through complex-valued deep feature extraction, phase information in the complex-valued cumulant signals that is beneficial for modulation classification is retained, thereby improving the accuracy of signal modulation classification.

[0059] See Figure 2 ,like Figure 2 The diagram shown is a schematic diagram of the complex-valued cumulant neural network model structure provided in the embodiment. Figure 2 In this diagram, CvConv represents a complex-valued convolutional module, the number after CvConv indicates the number of convolutional kernels, and the multiplication after @ indicates the size of the convolutional kernel; MaxPool represents a max-pooling module, the multiplication after @ indicates the size of the pooling kernel; Dense represents a fully connected layer, the number representing the number of neurons; Dropout represents a Dropout layer, the number representing the Dropout probability. The complex-valued cumulant neural network has a cumulant artificial feature extraction layer, a complex-valued deep feature extraction layer, and a feature classification layer. The cumulant artificial feature extraction layer first calculates the cumulant complex-valued signal of the intermediate frequency signal r and combines it into a cumulant artificial feature matrix; the complex-valued deep feature extraction layer consists of 7 complex convolution-max-pooling module groups connected in series, each group consisting of one complex convolutional module connected to one max-pooling module; the feature classification layer, after the 7 complex convolution-max-pooling module groups, concatenates the resulting features and expands them, then uses a fully connected layer to classify M modulation types.

[0060] The effectiveness of this invention can be further demonstrated through the following simulation experiments.

[0061] 1. Simulation conditions:

[0062] The simulation experiments of this invention were implemented using the Python programming language and were based on the publicly available RadioML 2018.01A dataset. This dataset includes Rayleigh fading channel, carrier frequency estimation bias, symbol rate estimation bias, delay spread, and thermal noise in signal generation and acquisition. Furthermore, the simulation experiments of this invention utilized 10 amplitude-phase modulation types from the RadioML 2018.01A dataset, with each sample being an intermediate frequency signal and having 1024 sampling points.

[0063] 2. Simulation content:

[0064] The simulation experiment of this invention compares the modulation and classification performance of the method of this invention with that of four existing common methods, and the results are as follows: Figure 3 As shown. Figure 3 The curves marked with * represent the changes in classification accuracy of the method of the present invention; Figure 3 The curve marked with a circle in the middle represents the change curve of classification accuracy of the method based on the real / imaginary two components of the intermediate frequency signal + convolutional neural network (I / Q+CNN); Figure 3 The curves marked with squares in the middle represent the changes in classification accuracy of the method based on time-frequency amplitude spectrum + convolutional neural network (STFT+CNN); Figure 3 The curves marked with triangles in the middle represent the changes in classification accuracy of the constellation diagram + convolutional neural network (CD+CNN) method; Figure 3 The curve marked with × represents the change in classification accuracy of the method based on the real / imaginary components of the intermediate frequency signal + constellation diagram + convolutional neural network (I / Q+CD+CNN). Figure 3 The horizontal axis represents the signal-to-noise ratio (SNR) in decibels (dB), and the vertical axis represents the classification accuracy in percentages (%).

[0065] 3. Simulation Result Analysis:

[0066] from Figure 3 As can be seen, the classification accuracy curves obtained by the method of this invention are above the classification accuracy curves of the other four common methods, indicating that the modulation classification performance of the method of this invention is superior to the other four common methods. Simulation results demonstrate that this invention can effectively improve the accuracy of signal modulation classification.

[0067] The phase of a signal is a key piece of information for distinguishing its modulation classification. To better utilize signal phase information for neural network-based modulation classification, complex-valued neural networks are needed. While the advantage of complex-valued neural networks in modulation classification applications lies in utilizing the phase information of the input complex signal, this phase information is easily contaminated by noise, thus weakening its advantages. The calculation process of cumulant complex-valued signals involves integration or summation of the signal, which is equivalent to smoothing the signal. Smoothing reduces the influence of high-frequency noise, compensating for the weakening of the advantages of complex-valued neural networks in modulation classification applications. Therefore, cumulant complex-valued signals are a feature signal with low noise sensitivity.

[0068] The innovations of this invention are: 1. Combining artificial features of cumulant complex-valued signals with complex-valued deep features to construct a complex-valued cumulant neural network; 2. Using cumulant complex-valued signals to overcome the disadvantage that phase information is easily contaminated by noise when directly using intermediate frequency signals as input to the complex-valued deep feature extraction layer; 3. Cumulant complex-valued signals expand the dimension of cumulant features, while retaining the low noise sensitivity of cumulant features and avoiding the deficiency of small information content when cumulant is used in the form of feature values.

[0069] This invention provides a modulation classification method based on a complex-valued cumulant neural network. It combines low-noise-sensitive artificial features of cumulants with deep features to construct a complex-valued cumulant neural network. An artificial feature extraction layer extracts low-noise-sensitive complex-valued cumulant signals as the basis for artificial features, and a complex-valued deep feature extraction layer further extracts the deep features of the complex-valued cumulant signals to achieve accurate modulation type classification. Compared with existing technologies, the most significant advantage of this invention is that by extracting complex-valued cumulant signals, the dimension of cumulant features is expanded, preserving the low noise sensitivity of cumulant features while avoiding the limitation of low information content when cumulants are used in feature value form. Simultaneously, through complex-valued deep feature extraction, phase information in the complex-valued cumulant signals that is beneficial for modulation classification is preserved, thereby improving the accuracy of signal modulation classification.

[0070] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0071] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0072] The embodiments described above do not constitute a limitation on the scope of protection of this application.

Claims

1. A modulation classification method based on a complex-valued cumulant neural network, characterized in that, The method includes: Step A: Define the structure of the complex-valued cumulant neural network; the complex-valued cumulant neural network includes a cumulant manual feature extraction layer, a complex-valued deep feature extraction layer, and a feature classification layer; Step B involves using a cumulant artificial feature extraction layer to calculate the second-order and fourth-order cumulant complex values ​​for the input intermediate frequency signal r, resulting in the second-order cumulant complex value signal C. 20 C 21 With the fourth-order cumulant complex-valued signal C 40 C 41 C 42 Among them, C 20 Let C be a second-order self-accumulator complex signal of r. 21 For r and r * The second-order cross-cumulant complex signal, C 40 Let C be a fourth-order self-accumulator complex signal of type r. 41 Let r, r, r and r * The fourth-order cross-cumulant complex signal, C 42 For r, r, r * With r * The fourth-order cross-cumulant complex signal, r * This represents the conjugate signal of the input intermediate frequency signal r; Step C involves combining the input intermediate frequency signal r with the second-order cumulant complex-valued signal C. 20 C 21 Fourth-order cumulant complex-valued signal C 40 C 41 C 42 The columns are horizontally concatenated and merged into a cumulative artificial feature matrix F used for modulation classification. C ; Step D, accumulating the artificial feature matrix F C Input the complex-valued deep feature extraction layer to obtain the deep features F D ; Step E, transfer the depth features F D The input feature classification layer performs modulation classification and outputs the signal modulation result.

2. The method according to claim 1, characterized in that, Step A, setting the structure of the complex-valued cumulant neural network; including: The cumulative artificial feature extraction layer is used to calculate the cumulative complex value signal of the intermediate frequency signal r, and combine the cumulative complex value signal into a cumulative artificial feature matrix; The complex-valued deep feature extraction layer consists of seven concatenated complex convolution-max pooling modules; each complex convolution-max pooling module group consists of one complex convolution module and one concatenated max pooling module. The feature classification layer is configured with 7 complex convolution-max pooling modules to concatenate and expand the resulting features, and then uses a fully connected layer to classify M modulation types. The feature classification layer consists of multiple fully connected modules.

3. The method according to claim 1, characterized in that, The cumulative quantity complex value signal is a signal composed of corresponding cumulative quantity characteristic values ​​under multiple time delay conditions of the signal, and each signal value in the cumulative quantity signal is a complex value.

4. The method according to claim 1, characterized in that, The second-order cumulative complex value signal C in the cumulative artificial feature extraction layer 20 C 21 With the fourth-order cumulant complex-valued signal C 40 C 41 C 42 The calculation method is as follows: Where Cum(·) represents the cumulative characteristic value calculation function, r (k) This represents the signal obtained by delaying the input intermediate frequency signal r by k sampling points, where L is the number of sampling points of the input intermediate frequency signal r. The delay of each cumulative complex signal increases from -L to L in steps of 1. The Cum(·) cumulative characteristic value calculation function is expressed as follows: The cumulative amount of variable x is represented as Cum[x1,x2,…,x] k ], which is c x (I), where k is the order of the cumulant, and the variable x includes k signals x1, x2, ..., xk. k Let I = {1,2,…,k} be the subscript index set of x, and let l be the moment of order l. Set I has q groups. p Let I be the p-th group, ∑ be the summation operation, ∪ be the union operation, ! be the factorial operation, Π be the product operation, and E(·) be the expectation operation.

Images