Non-cooperative communication signal modulation identification method based on multi-scale feature fusion extraction

Abstract

The invention relates to a non-cooperative communication signal modulation identification method based on multi-scale feature fusion extraction, which belongs to the technical field of cognitive radio communication, and comprises the following steps: generating time domain signals of communication signals in different modulation patterns, and preprocessing to obtain a plurality of baseband IQ signals as sample data; forming a communication signal data set with the corresponding modulation pattern as a sample label; initializing a multi-scale feature fusion extraction network model, and training the multi-scale feature fusion extraction network model by using the modulation signal data set to obtain a trained multi-scale feature fusion extraction network model; and obtaining a time domain signal of a non-cooperative communication signal in real time, performing preprocessing, inputting the trained multi-scale feature fusion extraction network model, and obtaining a recognition result of a corresponding time domain signal modulation pattern. Accurate and efficient identification of the non-cooperative communication signal modulation pattern is realized.

Description

Technical Field

[0001] The present invention relates to the field of cognitive radio communication technology, and in particular to a non-cooperative communication signal modulation recognition method based on multi-scale feature fusion extraction. Background Art

[0002] Non-cooperative communication signals refer to the communication method in which an unauthorized third party accesses the cooperative communication system without affecting the normal communication between the two parties in cooperative communication. In non-cooperative communication scenarios, the receiving party usually cannot obtain complete information about the transmitter, such as the precise carrier frequency, symbol timing, modulation style, etc. Therefore, modulation recognition of non-cooperative communication signals is a key technology in fields such as electronic reconnaissance and wireless spectrum management. It requires the reconnaissance system to be able to identify the modulation style of the signal without prior information, relying only on the received communication signal itself.

[0003] Modulation recognition of non-cooperative communication signals is an important prerequisite for demodulation, decoding and information extraction. It is also an important part of communication signal reconnaissance. It is widely used in intelligence reconnaissance, electronic countermeasures and other fields, and has important application value and research significance.

[0004] In recent years, with the rise of deep learning technology, it has shown strong advantages in data processing capabilities. More and more researchers are trying to use different deep learning methods to study signal modulation style recognition. In traditional networks, the main network structures are based on CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network). The spatial features of the signal sequence are extracted based on the CNN network, and the temporal relationship in the signal sequence is extracted based on the RNN network. However, both extract features locally in the signal and cannot achieve global information interaction, resulting in the inability to further improve the accuracy of communication signal modulation style recognition. In recent years, the Transformer-based network structure has broken through the limitations of CNN and RNN networks due to its excellent global information extraction ability and achieved global information (long-distance) interaction. The Transformer model mainly uses the long-distance correlation in the signal through the attention mechanism to extract the global features of the signal, but it lacks the ability to extract local features, and the modulation recognition effect cannot be further improved. Neural network structure design is an important method that combines expert knowledge and data-driven. The ideal network structure can have both efficiency and performance. The problem with existing technologies is how to improve the existing deep learning models and combine the advantages of different networks to improve the accuracy of modulation style recognition.

[0005] Traditional deep learning network models show good feature extraction capabilities and high recognition accuracy under high signal-to-noise ratios, but are insufficient in extracting signal features under low signal-to-noise ratios, and the recognition accuracy is not high. In addition, in actual application scenarios, it is impossible to deploy too large a network model and high requirements are placed on the rapid response of the model. The traditional deep neural network model has a complex network structure and it is difficult to strike a balance between recognition accuracy and computational complexity, which limits its application in actual scenarios.

[0006] Therefore, how to achieve higher modulation pattern recognition accuracy under low signal-to-noise ratio with less computational cost and improve communication reconnaissance capability is a technical problem that technicians in this field need to solve. Summary of the invention

[0007] In view of the above analysis, an embodiment of the present invention aims to provide a non-cooperative communication signal modulation recognition method based on multi-scale feature fusion extraction, so as to solve the problems that in the existing method, the traditional deep learning model is difficult to effectively extract signal features in a low signal-to-noise ratio environment, resulting in low modulation recognition accuracy, and in practical applications, there are technical problems such as too large model, too much calculation and slow response speed.

[0008] The present invention provides a non-cooperative communication signal modulation recognition method based on multi-scale feature fusion extraction, comprising the following steps:

[0009] Generate time domain signals of communication signals under different modulation modes and perform preprocessing to obtain multiple baseband IQ signals as sample data, and form a communication signal data set with the corresponding modulation modes as sample labels;

[0010] Initializing a multi-scale feature fusion extraction network model, and using the modulated signal data set to train the multi-scale feature fusion extraction network model to obtain a trained multi-scale feature fusion extraction network model;

[0011] The time domain signal of the non-cooperative communication signal is acquired in real time and preprocessed, and the trained multi-scale feature fusion extraction network model is input to obtain the recognition result of the corresponding time domain signal modulation style.

[0012] Furthermore, the time domain signal is preprocessed, including:

[0013] Converting the time domain signal into a baseband signal by orthogonal down conversion;

[0014] Filtering the baseband signal to remove noise from the baseband signal to obtain a filtered baseband signal;

[0015] Performing orthogonal demodulation on the filtered baseband signal to obtain a corresponding baseband IQ signal;

[0016] Each baseband IQ signal is sliced ​​to obtain a plurality of baseband IQ signal slices, and each baseband IQ signal slice is used as sample data.

[0017] Furthermore, the multi-scale feature fusion extraction network model includes an input layer, a multi-scale convolution module MSC, a CNN feature fusion module, a global feature extraction module GFE, a classification module and an output layer;

[0018] The input layer is used to receive the baseband IQ signal slice data;

[0019] The multi-scale convolution module MSC uses three parallel group convolution modules GCNN1, GCNN2 and GCNN3 with different convolution kernels to respectively extract the first, second and third local features of the baseband IQ signal slice, and concatenates the first, second and third local features to obtain multi-scale local features of the signal;

[0020] The CNN feature fusion layer sequentially fuses the multi-scale local features of the signal using three one-dimensional convolutions to obtain fusion features of the signal;

[0021] The global feature extraction module GFE performs linear transformation and self-attention calculation on the fusion features of the signal based on a multi-head attention mechanism to obtain the global features of the signal;

[0022] The classification module is used to classify the global features of the signal and output the predicted probability of each modulation pattern;

[0023] The output layer is used to output the recognition result of the modulation pattern with the maximum predicted probability corresponding to the baseband IQ signal.

[0024] Further, in a group convolution module: the baseband IQ signal is input into a two-dimensional convolution conv2d for feature extraction, and then Relu activation is performed to obtain a two-dimensional convolution local feature;

[0025] Inputting the two-dimensional convolution local features into one-dimensional group convolutions Groupconv1, Groupconv2 and Groupconv3 in sequence for feature extraction to obtain the local features of the baseband IQ signal;

[0026] Among them, the groups G of the one-dimensional group convolutions Groupconvl, Groupconv2 and Groupconv3 are different, and Relu activation is performed after each group convolution.

[0027] Furthermore, the fusion features of the signal are linearly transformed and self-attention calculated to obtain the global features of the signal, including:

[0028] Performing a linear transformation on the fused features to generate corresponding query vector Q, key vector K and value vector V;

[0029] Calculating a self-attention matrix mapping using the correlation between the query vector Q and the key vector K;

[0030] Weight the self-attention matrix mapping with the value vector V and scale it by a scaling factor to obtain an output matrix Attention(Q, K, V);

[0031] The query vector Q, key vector K and value vector V are projected independently, each vector is projected h times, and the output of each projection is spliced ​​to obtain the global feature output of the signal.

[0032] Furthermore, the classification module includes an average pooling layer, a fully connected layer and a softmax layer;

[0033] The average pooling layer is used to further reduce the dimension of the global features of the signal and retain the main features;

[0034] The fully connected layer is used to map the global features after average pooling to the category space;

[0035] The softmax layer is used to convert the output of the fully connected layer into a probability value, where each category corresponds to a probability value.

[0036] Furthermore, the training process of the multi-scale feature fusion extraction network model includes:

[0037] Dividing the communication signal data set into a training set, a validation set and a test set according to the modulation mode and the signal-to-noise ratio;

[0038] Set the batch size, total number of training times, and initial learning rate;

[0039] The training set is used to perform model training to obtain a loss function of the training and a model of the training; the validation set is used to validate the model of the training to obtain a loss function of the validation set;

[0040] Compare the loss function of the validation set with the minimum loss function of the trained set. If the loss function of the validation set is less than the minimum loss function of the trained set, save the model of the training set and update the minimum loss function of the trained set to the loss function value of the training set to start the next training. Otherwise, proceed to the next training set directly.

[0041] If the loss function of each validation set in m trainings is greater than the minimum loss function of the trained set, the learning rate is decayed to n times the original value and training continues; otherwise, the original learning rate is used for training; where m and n are preset values;

[0042] If the number of training times reaches the total number of training times, a trained model is obtained;

[0043] The trained model is tested using the test set to obtain the accuracy of the test set. If the accuracy of the test set reaches a preset threshold, the trained model is used as the trained multi-scale feature fusion extraction network model; otherwise, the model is retrained.

[0044] Furthermore, the loss function is a cross entropy loss function, as shown below:

[0045]

[0046] Among them, r i is the i-th sample data, P(y=c|r i ) is the probability of predicting the c-th modulation pattern, y i,c is the sample label, W is the model parameter, N is the total number of training samples, and C is the total number of modulation style categories.

[0047] Furthermore, the convolution kernels of the group convolution modules GCNN1, GCNN2 and GCNN3 are 3, 7 and 11 respectively.

[0048] Furthermore, the modulation pattern includes 8 digital modulation patterns and 3 analog modulation patterns;

[0049] The eight digital modulation modes include BPSK, QPSK, 8PSK, PAM4, QAM64, AM-DSB, AM-SSB, and CPFSK;

[0050] The three analog modulation styles are GFSK, WBFM, and QAM16.

[0051] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0052] 1. The present invention proposes a network model based on multi-scale feature fusion extraction for non-cooperative communication signal modulation pattern recognition. The local features of the signal at different scales are extracted by convolutional layers based on different scales, and the global features of the signal are extracted by using a multi-head self-attention mechanism to obtain the complete features of the signal. This enhances the model's characterization of the signal modulation pattern under low signal-to-noise ratio and improves the accuracy of modulation recognition.

[0053] 2. The present invention optimizes the parameters and computational complexity of the network model by replacing ordinary convolution with group convolution. At the same time, the extracted complete features are further optimized through average pooling operation, making the model more lightweight, reducing the memory required for model deployment and speeding up the model reasoning and recognition time, so that the model can adapt to more application scenarios.

[0054] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0056] Figure 1 This is a flow chart of a non-cooperative communication signal modulation identification method based on multi-scale feature fusion extraction in an embodiment of the present invention;

[0057] Figure 2 It is a schematic diagram of the structure of a multi-scale feature fusion extraction network model in an embodiment of the present invention;

[0058] Figure 3 Schematic diagram of the structure of a multi-scale convolution module MSC in an embodiment of the present invention;

[0059] Figure 4 This is a schematic diagram of the structure of the global feature extraction module GFE in an embodiment of the present invention;

[0060] Figure 5 Schematic diagram of the principle of feature extraction of a multi-scale convolution module in an embodiment of the present invention;

[0061] Figure 6 Schematic diagram of the group convolution principle in an embodiment of the present invention;

[0062] Figure 7 It is a comparison chart of experimental data in the embodiment of the present invention. DETAILED DESCRIPTION

[0063] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0064] The research significance of the present invention lies in: by preprocessing the original signal, improving the existing deep learning model, mining the characteristics of different modulations of the signal from the original signal, and optimizing the network parameters and computational complexity, the ability to characterize the signal under low signal-to-noise ratio is enhanced, the accuracy of modulation pattern recognition is improved, and the parameters and computational complexity required by the model are balanced.

[0065] A specific embodiment of the present invention discloses a non-cooperative communication signal modulation recognition method based on multi-scale feature fusion extraction, such as Figure 1 As shown, the following steps are included:

[0066] Step S1, generating time domain signals of communication signals under different modulation modes and preprocessing to obtain multiple baseband IQ signals as sample data, and forming a communication signal data set with corresponding modulation modes as sample labels;

[0067] Step S2, initializing a multi-scale feature fusion extraction network model, and using the modulated signal data set to train the multi-scale feature fusion extraction network model to obtain a trained multi-scale feature fusion extraction network model;

[0068] Step S3: acquiring the time domain signal of the non-cooperative communication signal in real time and preprocessing it, inputting it into the trained multi-scale feature fusion extraction network model, and obtaining the recognition result of the corresponding time domain signal modulation pattern.

[0069] Step S1, specifically.

[0070] In order to run efficiently on devices with limited hardware resources, we choose to directly use the time domain signals of communication signals for training and recognition of deep learning models to reduce memory usage and computing costs.

[0071] The present invention simulates and generates time domain signals of communication signals under different modulation modes, and preprocesses the time domain signals, including:

[0072] Converting the time domain signal into a baseband signal by orthogonal down conversion;

[0073] Filtering the baseband signal to remove noise from the baseband signal to obtain a filtered baseband signal;

[0074] The filtered baseband signal is orthogonally demodulated to obtain a corresponding baseband IQ signal; each baseband IQ signal is sliced ​​to obtain a plurality of baseband IQ signal slices, and each baseband IQ signal slice is used as sample data.

[0075] Perform filtering to remove noise from the baseband signal and improve signal quality;

[0076] Each baseband IQ signal is sliced, and each sample is sliced ​​into a uniform size to facilitate model processing, and the sliced ​​baseband IQ signal is used as sample data.

[0077] Exemplarily, each sample slice size is an IQ time domain signal in a 2*128 format. 2*128 represents 2 channels, each channel including 128 sampling points.

[0078] The present invention directly uses the IQ time domain signal instead of performing frequency domain or time domain transformation on the signal.

[0079] Exemplarily, the modulation modes in the present invention include 8 digital modulation modes and 3 analog modulation modes;

[0080] The eight digital modulation modes include BPSK, QPSK, 8PSK, PAM4, QAM64, AM-DSB, AM-SSB, and CPFSK;

[0081] The three analog modulation styles are GFSK, WBFM, and QAM16.

[0082] The modulation style type of each sample is manually labeled, and the modulation style serves as the sample label of the sample data.

[0083] The sample data and the corresponding sample labels constitute the communication signal dataset.

[0084] The function of step S1 is to generate and preprocess the baseband IQ signal of the communication signal, slice it into sample data in a unified format, and provide a modulation style label for each sample to construct a communication signal dataset for deep learning model training.

[0085] Step S2 is divided into steps S21-S22.

[0086] Step S21, initializing the multi-scale feature fusion extraction network model.

[0087] like Figure 2 As shown, the multi-scale feature fusion extraction network model includes an input layer, a multi-scale convolution module MSC, a CNN feature fusion module, a global feature extraction module GFE, a classification module and an output layer;

[0088] The input layer is used to receive the baseband IQ signal slice data;

[0089] The multi-scale convolution module MSC uses three parallel group convolution modules GCNN1, GCNN2 and GCNN3 with different convolution kernels to respectively extract the first, second and third local features of the baseband IQ signal slice, and concatenates the first, second and third local features to obtain multi-scale local features of the signal;

[0090] The CNN feature fusion layer sequentially fuses the multi-scale local features of the signal using three one-dimensional convolutions to obtain fusion features of the signal;

[0091] The global feature extraction module GFE performs linear transformation and self-attention calculation on the fusion features of the signal based on a multi-head attention mechanism to obtain the global features of the signal;

[0092] The classification module is used to classify the global features of the signal and output the predicted probability of each modulation pattern;

[0093] The output layer is used to output the recognition result of the modulation pattern with the maximum predicted probability corresponding to the baseband IQ signal.

[0094] The multi-scale convolution module MSC (Multi-Scale Convolution) is used to extract local features of the signal at different scales, such as Figure 3 shown. Figure 3 and Figure 5 The following figure shows the schematic diagram of signal feature extraction using convolution kernels of different sizes. Signals with different modulation styles exhibit different characteristics in different frequency ranges. In order to effectively identify signals with different modulation styles, different convolution kernels are used to extract features from the signal to obtain feature information at different scales.

[0095] Input X is the baseband IQ signal slice sample data, and inputs three parallel group convolution modules GCNN1, GCNN2 and GCNN3 with different convolution kernels;

[0096] The convolution kernels of the group convolution modules GCNN1, GCNN2 and GCNN3 are 3, 7 and 11 respectively.

[0097] The local features of the signal are extracted by using three convolution kernels of different sizes.

[0098] In a group convolution module:

[0099] The baseband IQ signal is input into a two-dimensional convolution conv2d for feature extraction, and then Relu activation is performed to obtain a two-dimensional convolution local feature;

[0100] Inputting the two-dimensional convolution local features into one-dimensional group convolutions Groupconv1, Groupconv2 and Groupconv3 in sequence for feature extraction to obtain the local features of the baseband IQ signal;

[0101] Among them, the groups G of the one-dimensional group convolutions Groupconvl, Groupconv2 and Groupconv3 are different, and Relu activation is performed after each group convolution.

[0102] Each group convolution module extracts the local feature operation process, as shown in formula (1):

[0103]

[0104] Among them, X is the input feature map, Z(p,q) represents the value of the pth and qth positions of the output feature map, W is the convolution kernel, b is the bias, and m and n are the row and column indices of the convolution kernel.

[0105] The result of each Relu activation is shown in formula (2):

[0106] Relu(x)=max(0,x) Formula (2)

[0107] If the input value x is greater than or equal to 0, output x; if x is less than 0, output 0.

[0108] The Relu activation function simplifies the calculation of gradients in back propagation during network training, while alleviating the gradient vanishing and gradient exploding phenomena during training.

[0109] (1) For the group convolution module GCNN1:

[0110] Two-dimensional convolution conv2d(2, 3, 48), the convolution kernel is 2*3, 48 channels;

[0111] One-dimensional convolution Groupconv1 (48*48*1*3, G=4), the convolution kernel is 1*3, and the group G is 4;

[0112] One-dimensional convolution Groupconv2 (48*48*1*3, G=8), the convolution kernel is 1*3, and the group G is 8;

[0113] One-dimensional convolution Groupconv3 (48*48*1*3, G=16), the convolution kernel is 1*3, and the group G is 16;

[0114] (2) For the group convolution module GCNN2:

[0115] Two-dimensional convolution conv2d(2, 7, 48), the convolution kernel is 2*7, 48 channels;

[0116] One-dimensional convolution Groupconv1 (48*48*1*7, G=4), the convolution kernel is 1*7, and the group G is 4;

[0117] One-dimensional convolution Groupconv2 (48*48*1*7, G=8), the convolution kernel is 1*7, and the group G is 8;

[0118] One-dimensional convolution Groupconv3 (48*48*1*7, G=16), the convolution kernel is 1*7, and the group G is 16;

[0119] (3) For the group convolution module GCNN3:

[0120] Two-dimensional convolution conv2d(2, 11, 48), the convolution kernel is 2*11, 48 channels;

[0121] One-dimensional convolution Groupconv1 (48*48*1*11, G=4), the convolution kernel is 1*11, and the group G is 4;

[0122] One-dimensional convolution Groupconv2 (48*48*1*11, G=8), the convolution kernel is 1*11, and the group G is 8;

[0123] One-dimensional convolution Groupconv3 (48*48*1*11, G=16), the convolution kernel is 1*11, and the group G is 16;

[0124] The multi-scale convolution module MSC performs two-dimensional convolution on the signal with kernel sizes of 2*3, 2*7, and 2*11 to extract the IQ path information of the signal; then, through three layers of one-dimensional convolution, the convolution kernels are 1*3, 1*7, and 1*11 respectively, and the number of channels of all convolution layers is set to 48, to obtain local feature maps of the signal at three different scales.

[0125] In the step of extracting multi-scale features, a large convolution kernel is introduced to extract local features, which increases the parameters and computational complexity of the model. Group convolution is used to replace the resources occupied by the ordinary convolution optimization model. The schematic diagram of the principle is shown in the figure. Figure 6 As shown in the figure, the principle diagrams of ordinary convolution, convolution with group number G=2, and convolution with group number G=4 are shown respectively. Group convolution can reduce the number of parameters and the amount of computation required in convolution. The number of parameters and the amount of computation required for group convolution are 1 / G of that for ordinary convolution.

[0126] In the multi-scale convolution module MSC, GCNN1, GCNN2 and GCNN3 use group convolution to extract local features. The number of groups G of Groupconv1, Groupconv2 and Groupconv3 in GCNN1, GCNN2 and GCNN3 are 4, 8 and 16 respectively, which reduces the parameters and amount of computation required for each convolution layer operation for extracting local features to 1 / 4, 1 / 8 and 1 / 16 of the original ordinary convolution respectively.

[0127] By increasing the number of groups G layer by layer, the computational complexity and the number of parameters are gradually reduced, while ensuring that a high feature representation capability is maintained in shallower feature extraction and gradually reducing the computational burden in deeper layers.

[0128] The first local feature output by the group convolution module GCNN1 is F msc,3 ;

[0129] The second local feature output by the group convolution module GCNN2 is F msc,7 ;

[0130] The third local feature output by the group convolution module GCNN3 is F msc,11 .

[0131] The first, second and third local features are concatenated by Concat to obtain the multi-scale local feature F msc , as shown in formula (3):

[0132] F msc =Concat(F msc,3 ,F msc,7 ,F msc,11 ) Formula (3)

[0133] Among them, F msc,3 、F msc,7 、F msc,11 The local features are obtained by the group convolution modules GCNN1, GCNN2 and GCNN3 with convolution kernels of sizes 3, 7 and 11 respectively.

[0134] For multi-scale local features F msc Through three layers of convolution layers Conv1(1,3,64), Conv2(1,3,64), and Conv3(1,3,64), we can fully extract the relationship between features at different scales and extract the fusion features of the signal. The number of output channels of each convolution layer is set to 64.

[0135] The fused features are input into the global feature module GFE (Global Feature Extraction), so that the model has the function of extracting the global features of the signal, ignoring irrelevant information on the signal as a whole, focusing on key information related to signal modulation, and improving the recognition accuracy.

[0136] The fusion features of the signal are linearly transformed and self-attention calculated to obtain the global features of the signal, including:

[0137] Performing a linear transformation on the fused features to generate corresponding query vector Q, key vector K and value vector V;

[0138] Calculating a self-attention matrix mapping using the correlation between the query vector Q and the key vector K;

[0139] Weight the self-attention matrix mapping with the value vector V and scale it by a scaling factor to obtain an output matrix Attention(Q, K, V);

[0140] The query vector Q, key vector K and value vector V are projected independently, each vector is projected h times, and the output of each projection is spliced ​​to obtain the global feature output of the signal.

[0141] The global feature extraction module GFE is implemented based on the multi-head attention mechanism. First, different linear transformations are performed on the fused features, as shown in formula (4):

[0142]

[0143] Among them, F represents the fusion feature of the input signal, W q , W k and W v is a learnable parameter, a weight matrix for linear transformation.

[0144] By using the correlation between the query vector Q and the key vector K, a self-attention matrix mapping containing global information is calculated Among them, K T represents the transpose of K.

[0145] Then, the self-attention matrix mapping is weighted with the corresponding value vector V to obtain the output matrix Attention(Q,K,V), as shown in formula (5):

[0146]

[0147] in, is the scaling factor.

[0148] The query vector Q, key vector K, and value vector V are projected independently, each vector is projected h times, each time using different parameters. The outputs of these projections are connected together, as shown in formula (6):

[0149] head j =Attention(Q j ,K j ,V j ) Formula (6)

[0150] Among them, j = 1, 2, ..., h.

[0151] By splicing multiple heads, we can get the global characteristics of the signal, as shown in formula (7):

[0152] MultiHead=Concat(head 1 ,...,head h ) Formula (7)

[0153] Among them, MultiHead is the global feature of the signal.

[0154] The global feature MultiHead is input into the classification module for modulation style classification and identification to obtain the signal modulation style.

[0155] The classification module includes an average pooling layer, a fully connected layer and a softmax layer;

[0156] The average pooling layer is used to further reduce the dimension of the global features of the signal and retain the main features;

[0157] The fully connected layer is used to map the global features after average pooling to the category space;

[0158] The softmax layer is used to convert the output of the fully connected layer into a probability value, where each category corresponds to a probability value.

[0159] For the global features of the signal, the feature dimension is first reduced through the average pooling operation while retaining the main features. The main purpose is to reduce the computational complexity and number of parameters required by the model while preventing the model from overfitting.

[0160] The average pooled features are input into the fully connected layer, and the network is able to better process the features of different modulation styles of the input signal through a combination of linear transformation and nonlinear activation function. Finally, the probability of each modulation style is output through the softmax layer.

[0161] Step S22: Use the modulated signal data set to train the multi-scale feature fusion extraction network model to obtain a trained multi-scale feature fusion extraction network model.

[0162] The training process of the multi-scale feature fusion extraction network model includes:

[0163] Dividing the communication signal data set into a training set, a validation set and a test set according to the modulation mode and the signal-to-noise ratio;

[0164] Set the batch size, total number of training times, and initial learning rate;

[0165] The training set is used to perform model training to obtain a loss function of the training and a model of the training; the validation set is used to validate the model of the training to obtain a loss function of the validation set;

[0166] Compare the loss function of the validation set with the minimum loss function of the trained set. If the loss function of the validation set is less than the minimum loss function of the trained set, save the model of the training set and update the minimum loss function of the trained set to the loss function value of the training set to start the next training. Otherwise, proceed to the next training set directly.

[0167] If the loss function of each validation set in m trainings is greater than the minimum loss function of the trained set, the learning rate is decayed to n times the original value and training continues; otherwise, the original learning rate is used for training; where m and n are preset values;

[0168] If the number of training times reaches the total number of training times, a trained model is obtained;

[0169] The trained model is tested using the test set to obtain the accuracy of the test set. If the accuracy of the test set reaches a preset threshold, the trained model is used as the trained multi-scale feature fusion extraction network model; otherwise, the model is retrained.

[0170] Exemplarily, the communication signal data set is divided into a training set (60%), a validation set (20%) and a test set (20%) according to different modulation modes and signal-to-noise ratios;

[0171] During initialization, the batch size is set to 256, the initial learning rate is 0.001, and the total number of training times is 200;

[0172] m is 10 rounds and n is 0.4 times.

[0173] The Adam optimizer and the cross entropy loss function are used to train the function. The loss function is the cross entropy loss function, as shown in formula (8):

[0174]

[0175] Among them, r i is the i-th sample data, P(y=c|r i ) is the probability of predicting the c-th modulation pattern, y i,c is the sample label, W is the model parameter, N is the total number of training samples, and C is the total number of modulation style categories.

[0176] The modulation style is predicted using the finally trained network parameters, as shown in formula (9):

[0177]

[0178] Compared with the existing modulation pattern recognition model based on deep learning, the present invention mainly extracts the local features of the signal at different scales, and extracts the global features of the signal through a multi-head self-attention mechanism, fully exploits the signal features and recognizes the modulation pattern of the signal. At the same time, compared with the existing deep learning modulation pattern recognition model, the present invention optimizes the memory and the amount of calculation required by the model through group convolution, so as to reduce the calculation cost. Compared with the existing deep learning modulation recognition model, the network model proposed in the present invention shows a higher recognition rate and better noise resistance.

[0179] Step S3: acquiring the time domain signal of the non-cooperative communication signal in real time and preprocessing it, inputting it into the trained multi-scale feature fusion extraction network model, and obtaining the recognition result of the corresponding time domain signal modulation pattern.

[0180] The effect of the present invention can be further illustrated by the following simulation:

[0181] In order to demonstrate the feasibility of the method proposed by the multi-scale feature fusion extraction network model MFEE (Multi-scale Feature Extraction and Fusion Network) of the present invention, Figure 7 As shown in the figure, the simulation experiment is compared with the CNN network model (CNN), LSTM network model (LSTM), CNN and LSTM hybrid model (MCLDNN, PET-CGDNN) and the network based on Transformer structure (MCformer). The effects of each model under different signal-to-noise ratios are shown in the figure. Figure 7 shown.

[0182] In the embodiment, under the condition of high signal-to-noise ratio greater than 2dB, the average recognition rate of MFEE is 8.38% to 11.61% higher than that of other models, indicating that in a high signal-to-noise ratio environment, MFEE can maintain a high recognition accuracy and performs best among all compared models.

[0183] In the signal-to-noise ratio range of -6dB to 0dB, due to the low signal-to-noise ratio, the recognition accuracy of MFEE has decreased, but the average recognition rate of MFEE is still 4.73% to 9.09% higher than other models, with significant performance improvement, indicating that MFEE has better noise resistance and recognition performance under low signal-to-noise ratio conditions, reflecting the robustness of MFEE to noise.

[0184] The method of the present invention has the advantage of using high recognition accuracy compared with the existing signal modulation recognition models in the prior art by designing a feature extraction backbone network structure; at the same time, the method of the present invention has significant improvements over other network models under low signal-to-noise ratios, demonstrating that the MFEE model is robust to noise.

[0185] In summary, a non-cooperative communication signal modulation recognition method based on multi-scale feature fusion extraction according to an embodiment of the present invention has the following beneficial effects:

[0186] 1. The present invention proposes a network model based on multi-scale feature fusion extraction for non-cooperative communication signal modulation pattern recognition. The local features of the signal at different scales are extracted by convolutional layers based on different scales, and the global features of the signal are extracted by using a multi-head self-attention mechanism to obtain the complete features of the signal. This enhances the model's characterization of the signal modulation pattern under low signal-to-noise ratio and improves the accuracy of modulation recognition.

[0187] 2. The present invention optimizes the parameters and computational complexity of the network model by replacing ordinary convolution with group convolution. At the same time, the extracted complete features are further optimized through average pooling operation, making the model more lightweight, reducing the memory required for model deployment and speeding up the model reasoning and recognition time, so that the model can adapt to more application scenarios.

[0188] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A non-cooperative communication signal modulation recognition method based on multi-scale feature fusion extraction, characterized in that: The steps include: Generate time domain signals of communication signals under different modulation modes and perform preprocessing to obtain multiple baseband IQ signals as sample data, and form a communication signal data set with the corresponding modulation modes as sample labels; Initializing a multi-scale feature fusion extraction network model, and using the modulated signal data set to train the multi-scale feature fusion extraction network model to obtain a trained multi-scale feature fusion extraction network model; The time domain signal of the non-cooperative communication signal is acquired in real time and preprocessed, and the trained multi-scale feature fusion extraction network model is input to obtain the recognition result of the corresponding time domain signal modulation style.

2. The method according to claim 1, characterized in that: Preprocess the time domain signal, including: Converting the time domain signal into a baseband signal by orthogonal down conversion; Filtering the baseband signal to remove noise from the baseband signal to obtain a filtered baseband signal; Performing orthogonal demodulation on the filtered baseband signal to obtain a corresponding baseband IQ signal; Each baseband IQ signal is sliced ​​to obtain a plurality of baseband IQ signal slices, and each baseband IQ signal slice is used as sample data.

3. The method according to claim 1, characterized in that: The multi-scale feature fusion extraction network model includes an input layer, a multi-scale convolution module MSC, a CNN feature fusion module, a global feature extraction module GFE, a classification module and an output layer; The input layer is used to receive the baseband IQ signal slice data; The multi-scale convolution module MSC uses three parallel group convolution modules GCNN1, GCNN2 and GCNN3 with different convolution kernels to respectively extract the first, second and third local features of the baseband IQ signal slice, and concatenates the first, second and third local features to obtain multi-scale local features of the signal; The CNN feature fusion layer sequentially fuses the multi-scale local features of the signal using three one-dimensional convolutions to obtain fusion features of the signal; The global feature extraction module GFE performs linear transformation and self-attention calculation on the fusion features of the signal based on a multi-head attention mechanism to obtain the global features of the signal; The classification module is used to classify the global features of the signal and output the predicted probability of each modulation pattern; The output layer is used to output the recognition result of the modulation pattern with the maximum predicted probability corresponding to the baseband IQ signal.

4. The method according to claim 3, characterized in that: In a group convolution module: the baseband IQ signal is input into a two-dimensional convolution conv2d for feature extraction, and then Relu activation is performed to obtain a two-dimensional convolution local feature; Inputting the two-dimensional convolution local features into one-dimensional group convolutions Groupconv1, Groupconv2 and Groupconv3 in sequence for feature extraction to obtain the local features of the baseband IQ signal; Among them, the groups G of the one-dimensional group convolutions Groupconvl, Groupconv2 and Groupconv3 are different, and Relu activation is performed after each group convolution.

5. The method according to claim 3, characterized in that: The fusion features of the signal are linearly transformed and self-attention calculated to obtain the global features of the signal, including: Performing a linear transformation on the fused features to generate corresponding query vector Q, key vector K and value vector V; Calculating a self-attention matrix mapping using the correlation between the query vector Q and the key vector K; Weight the self-attention matrix mapping with the value vector V and scale it by a scaling factor to obtain an output matrix Attention(Q, K, V); The query vector Q, key vector K and value vector V are projected independently, each vector is projected h times, and the output of each projection is spliced ​​to obtain the global feature output of the signal.

6. The method according to claim 3, characterized in that: The classification module includes an average pooling layer, a fully connected layer and a softmax layer; The average pooling layer is used to further reduce the dimension of the global features of the signal and retain the main features; The fully connected layer is used to map the global features after average pooling to the category space; The softmax layer is used to convert the output of the fully connected layer into a probability value, where each category corresponds to a probability value.

7. The method according to claim 1, characterized in that: The training process of the multi-scale feature fusion extraction network model includes: Dividing the communication signal data set into a training set, a validation set and a test set according to the modulation mode and the signal-to-noise ratio; Set the batch size, total number of training times, and initial learning rate; The training set is used to perform model training to obtain a loss function of the training and a model of the training; the validation set is used to validate the model of the training to obtain a loss function of the validation set; Compare the loss function of the validation set with the minimum loss function of the trained set. If the loss function of the validation set is less than the minimum loss function of the trained set, save the model of the training set and update the minimum loss function of the trained set to the loss function value of the training set to start the next training. Otherwise, proceed to the next training set directly. If the loss function of each validation set in m trainings is greater than the minimum loss function of the trained set, the learning rate is decayed to n times the original value and training continues; otherwise, the original learning rate is used for training; where m and n are preset values; If the number of training times reaches the total number of training times, a trained model is obtained; The trained model is tested using the test set to obtain the accuracy of the test set. If the accuracy of the test set reaches a preset threshold, the trained model is used as the trained multi-scale feature fusion extraction network model; otherwise, the model is retrained.

8. The method according to claim 7, characterized in that: The loss function is a cross entropy loss function, as shown below: Among them, r i is the i-th sample data, P(y=c|r i ) is the probability of predicting the c-th modulation pattern, y i,c is the sample label, W is the model parameter, N is the total number of training samples, and C is the total number of modulation style categories.

9. The method according to claim 3, characterized in that: The convolution kernels of the group convolution modules GCNN1, GCNN2 and GCNN3 are 3, 7 and 11 respectively.

10. The method according to any one of claims 1 to 9, characterized in that: The modulation modes include 8 digital modulation modes and 3 analog modulation modes; The eight digital modulation modes include BPSK, QPSK, 8PSK, PAM4, QAM64, AM-DSB, AM-SSB, and CPFSK; The three analog modulation styles are GFSK, WBFM, and QAM16.

Images