A non-cooperative communication signal modulation recognition method based on multi-scale feature fusion extraction

Abstract

The present application relates to a kind of non-cooperative communication signal modulation identification method based on multiscale feature fusion extraction, belong to cognitive radio communication technical field, including generating the time domain signal of communication signal under different modulation style and being preprocessed to obtain multiple baseband IQ signal as sample data, with corresponding modulation style as sample label composition communication signal dataset;Initialization multiscale feature fusion extraction network model, utilize the modulation signal dataset to train the multiscale feature fusion extraction network model, obtain trained multiscale feature fusion extraction network model;Real-time acquisition non-cooperative communication signal time domain signal and pre-processing, input the trained multiscale feature fusion extraction network model, obtain the identification result of corresponding time domain signal modulation style.Accurate and efficient identification of non-cooperative communication signal modulation style is realized.

Description

A method for non-cooperative communication signal modulation recognition based on multi-scale feature fusion extraction Technical Field

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

[0002] Non-cooperative communication signals refer to communication methods where an unauthorized third party accesses a cooperative communication system without affecting the normal communication between the cooperating parties. In non-cooperative communication scenarios, the receiver typically cannot obtain complete information about the transmitter, such as the precise carrier frequency, symbol timing, and modulation pattern. Therefore, modulation identification 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 pattern of a signal based solely on the received communication signal without prior information.

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

[0004] In recent years, with the rise of deep learning technology, it has demonstrated powerful advantages in data processing capabilities. More and more researchers are attempting to use different deep learning methods for signal modulation pattern recognition. Traditional networks mainly employ CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network) architectures. CNNs extract spatial features from signal sequences, while RNNs extract temporal relationships. However, both methods extract features locally, failing to achieve global information interaction, thus limiting the accuracy of modulation pattern recognition. In recent years, Transformer-based network structures have overcome the limitations of CNNs and RNNs due to their excellent global information extraction capabilities, enabling global (long-distance) information interaction. Transformer models primarily utilize attention mechanisms to extract global features from signals based on long-distance correlations, but they lack the ability to extract local features, further hindering modulation pattern recognition. Neural network architecture design is a crucial method combining expert knowledge and data-driven approaches; an ideal network structure can balance efficiency and performance. The current challenge lies in improving existing deep learning models and combining the advantages of different networks to enhance the accuracy of modulation pattern recognition.

[0005] Traditional deep learning network models exhibit good feature extraction capabilities and high recognition accuracy at high signal-to-noise ratios (SNRs). However, they are insufficient in extracting signal features at low SNRs, resulting in low recognition accuracy. Furthermore, in practical applications, it is impossible to deploy excessively large network models, and rapid model response is crucial. The complex network structure of traditional deep neural network models makes it difficult to balance recognition accuracy and computational complexity, thus limiting their application in real-world 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 capabilities is a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0007] Based on the above analysis, the present invention aims to provide a non-cooperative communication signal modulation recognition method based on multi-scale feature fusion extraction, in order to solve the technical problems of existing methods in low signal-to-noise ratio environments, where traditional deep learning models are difficult to effectively extract signal features, resulting in low modulation recognition accuracy, and in practical applications, the model is too large, the computational load is too heavy, and the response speed is slow.

[0008] This invention provides a method for identifying modulation of non-cooperative communication signals based on multi-scale feature fusion extraction, comprising the following steps:

[0009] The time-domain signals of communication signals under different modulation styles are generated and preprocessed to obtain multiple baseband IQ signals as sample data, which are then combined with the corresponding modulation styles as sample labels to form a communication signal dataset.

[0010] Initialize the multi-scale feature fusion extraction network model, and train the multi-scale feature fusion extraction network model using the modulation signal dataset to obtain the 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. It is then input into the trained multi-scale feature fusion extraction network model to obtain the recognition result of the corresponding time-domain signal modulation pattern.

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

[0013] The time-domain signal is converted into a baseband signal by orthogonal downconversion;

[0014] The baseband signal is filtered to remove noise and obtain a filtered baseband signal.

[0015] The filtered baseband signal is quadrature demodulated to obtain the corresponding baseband IQ signal;

[0016] Each baseband IQ signal is sliced ​​to obtain multiple 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 convolutional 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 extract the first, second and third local features of the baseband IQ signal slice, respectively, and splices the first, second and third local features to obtain the multi-scale local features of the signal.

[0020] The CNN feature fusion layer sequentially uses three one-dimensional convolutions to fuse the multi-scale local features of the signal to obtain the fused features of the signal.

[0021] The Global Feature Extraction (GFE) module performs linear transformation and self-attention calculation on the fused 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 identification result of the modulation pattern with the highest predicted probability corresponding to the baseband IQ signal.

[0024] Furthermore, 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 local features of the two-dimensional convolution;

[0025] The local features of the two-dimensional convolution are sequentially input into one-dimensional group convolutions Groupconv1, Groupconv2, and Groupconv3 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 Groupconv1, Groupconv2 and Groupconv3 are different, and ReLU activation is performed after each group convolution.

[0027] Furthermore, the fused features of the signal are subjected to linear transformation and self-attention calculation to obtain the global features of the signal, including:

[0028] The fused features are linearly transformed to generate the corresponding query vector Q, key vector K, and value vector V;

[0029] The self-attention matrix mapping is calculated using the correlation between the query vector Q and the key vector K;

[0030] The self-attention matrix is ​​mapped and weighted with the value vector V, and then scaled using a scaling factor to obtain the output matrix Attention(Q,K,V);

[0031] The query vector Q, key vector K, and value vector V are projected independently, with each vector being projected h times. The outputs of each projection are then concatenated 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 dimensionality of the global features of the signal while retaining 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 probability values, with one probability value corresponding to each category.

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

[0037] The communication signal dataset is divided into a training set, a validation set, and a test set according to the modulation pattern and signal-to-noise ratio.

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

[0039] The model is trained using the training set to obtain the loss function and the model trained in this training session. The model trained in this training session is then validated using the validation set to obtain the loss function of the validation set.

[0040] The loss function of the current validation set is compared with the smallest loss function that has been trained. If the loss function of the current validation set is less than the smallest loss function that has been trained, the model trained in the current iteration is saved, and the smallest loss function that has been trained is updated to the loss function value of the current iteration, and the next training iteration begins; otherwise, the next training iteration is performed directly.

[0041] If the loss function of the validation set is larger than the smallest loss function trained in each of the m training iterations, then the learning rate is decayed to n times the original value and training continues; otherwise, the original learning rate is used for training. Here, m and n are preset values.

[0042] If the number of training iterations reaches the total number of training iterations, a well-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 a trained multi-scale feature fusion extraction network model; otherwise, the model is retrained.

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

[0045]

[0046] Where, r i For the i-th sample data, P(y=c|r i y is the probability of predicting the c-th modulation pattern. i,c denoted as sample label, W as model parameter, N as total number of training samples, and C as total number of modulation pattern classifications.

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

[0048] Furthermore, the modulation styles include eight digital modulation styles and three analog modulation styles;

[0049] The eight digital modulation styles 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. This invention proposes a network model based on multi-scale feature fusion extraction for non-cooperative communication signal modulation pattern recognition. It extracts local features of the signal at different scales by using convolutional layers based on different scales, and uses a multi-head self-attention mechanism to extract global features of the signal, thereby obtaining complete features of the signal. This enhances the model's representation of signal modulation patterns under low signal-to-noise ratio and improves the accuracy of modulation recognition.

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

[0054] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0055] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

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

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

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

[0059] Figure 4 is a schematic diagram of the Global Feature Extraction Module (GFE) in an embodiment of the present invention;

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

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

[0062] Figure 7 is a comparison chart of experimental data in the embodiments of the present invention. Detailed Implementation

[0063] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form 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 intended to limit the scope of the present invention.

[0064] The significance of this invention lies in: by preprocessing the original signal, improving the existing deep learning model, mining the features between different modulations of the signal from the original signal, optimizing the network parameters and computational load, thereby enhancing the ability to represent the signal under low signal-to-noise ratio, improving the accuracy of modulation pattern recognition, and balancing the parameters and computational load required by the model.

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

[0066] Step S1: Generate time-domain signals of communication signals under different modulation styles and preprocess them to obtain multiple baseband IQ signals as sample data, and use them with the corresponding modulation styles as sample labels to form a communication signal dataset.

[0067] Step S2: Initialize the multi-scale feature fusion extraction network model, and train the multi-scale feature fusion extraction network model using the modulation signal dataset to obtain the trained multi-scale feature fusion extraction network model;

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

[0069] Step S1, specifically.

[0070] To run efficiently on devices with limited hardware resources, we chose to use the time-domain signals of the communication signals directly for training and recognizing deep learning models, thereby reducing memory usage and computational costs.

[0071] This invention simulates and generates time-domain signals of communication signals under different modulation styles, and preprocesses the time-domain signals, including:

[0072] The time-domain signal is converted into a baseband signal by orthogonal downconversion;

[0073] The baseband signal is filtered to remove noise and obtain a filtered baseband signal.

[0074] The filtered baseband signal is quadrature demodulated to obtain the corresponding baseband IQ signal; each baseband IQ signal is sliced ​​to obtain multiple baseband IQ signal slices, and each baseband IQ signal slice is used as sample data.

[0075] Filtering is performed to remove noise from the baseband signal and improve signal quality;

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

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

[0078] This invention directly utilizes the IQ time-domain signal, rather than the data obtained after frequency-domain or time-domain transformation of the signal.

[0079] For example, the modulation styles described in this invention include 8 digital modulation styles and 3 analog modulation styles;

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

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

[0082] For each sample, its modulation style type is manually labeled, and the modulation style serves as the sample label for the sample data.

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

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

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

[0086] Step S21: Initialize the multi-scale feature fusion extraction network model.

[0087] As shown in Figure 2, the multi-scale feature fusion extraction network model includes an input layer, a multi-scale convolutional 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 extract the first, second and third local features of the baseband IQ signal slice, respectively, and splices the first, second and third local features to obtain the multi-scale local features of the signal.

[0090] The CNN feature fusion layer sequentially uses three one-dimensional convolutions to fuse the multi-scale local features of the signal to obtain the fused features of the signal.

[0091] The Global Feature Extraction (GFE) module performs linear transformation and self-attention calculation on the fused 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 identification result of the modulation pattern with the highest predicted probability corresponding to the baseband IQ signal.

[0094] Local features at different scales of the signal are extracted using a multi-scale convolution (MSC) module, as shown in Figure 3. Figures 3 and 5 illustrate the feature extraction of the signal using convolution kernels of different sizes. Signals with different modulation patterns exhibit different characteristics in different frequency ranges. To effectively identify signals with different modulation patterns, feature extraction is performed on the signal using different convolution kernels to obtain feature information at different scales.

[0095] Input X is baseband IQ signal slice sample data, and input three parallel group convolutional modules with different convolutional kernels: GCNN1, GCNN2 and GCNN3.

[0096] The group convolutional modules GCNN1, GCNN2, and GCNN3 have 3, 7, and 11 convolutional kernels, respectively.

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

[0098] In a group convolutional 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 local features of the two-dimensional convolution.

[0100] The local features of the two-dimensional convolution are sequentially input into one-dimensional group convolutions Groupconv1, Groupconv2, and Groupconv3 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 Groupconv1, Groupconv2 and Groupconv3 are different, and ReLU activation is performed after each group convolution.

[0102] The local feature extraction operation process for each group of convolutional modules is shown in formula (1):

[0103]

[0104] Where X is the input feature map, Z(p,q) represents the value at position p and q 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 Equation (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 gradient calculation during backpropagation in the training process of a network, while also mitigating the vanishing and exploding gradient phenomena during training.

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

[0110] A 2D convolution conv2d(2, 3, 48) has a 2*3 kernel with 48 channels.

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

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

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

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

[0115] A 2D convolution conv2d(2, 7, 48) has a kernel of 2*7 and 48 channels.

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

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

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

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

[0120] A 2D convolution conv2d(2, 11, 48) has a kernel of 2*11 and 48 channels.

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

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

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

[0124] The multi-scale convolution module MSC extracts the IQ path information of the signal by performing two-dimensional convolution with kernel sizes of 2*3, 2*7, and 2*11 respectively. Then, through three layers of one-dimensional convolution with kernel sizes of 1*3, 1*7, and 1*11 respectively, and the number of channels in all convolutional layers is set to 48, local feature maps of the signal at three different scales are obtained.

[0125] In the multi-scale feature extraction step, a large convolution kernel is introduced to extract local features, increasing the model's parameters and computational load. Group convolution is used to replace ordinary convolution to optimize the model's resource usage. The principle diagram is shown in Figure 6, which shows the principle diagrams of ordinary convolution, convolution with G=2 groups, and convolution with G=4 groups. Group convolution can reduce the number of parameters and computational load required in convolution. The number of parameters and computational load required by group convolution is 1 / G of that of ordinary convolution.

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

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

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

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

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

[0131] By concatenating the first, second, and third local features using Concat, we 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 These are local features obtained by the group convolutional modules GCNN1, GCNN2, and GCNN3, respectively, using convolutional kernels of sizes 3, 7, and 11.

[0134] For multi-scale local features F msc By using three convolutional layers Conv1(1,3,64), Conv2(1,3,64), and Conv3(1,3,64), the relationships between features at different scales can be fully extracted, and the fusion features of the signal can be extracted. The number of output channels for each convolutional layer is set to 64.

[0135] By inputting the fused features into the global feature extraction (GFE) module, the model is able to extract global features of the signal, ignoring irrelevant information in the overall signal and focusing on key information related to signal modulation, thereby improving the recognition accuracy.

[0136] The fused features of the signal are subjected to linear transformation and self-attention calculation to obtain the global features of the signal, including:

[0137] The fused features are linearly transformed to generate the corresponding query vector Q, key vector K, and value vector V;

[0138] The self-attention matrix mapping is calculated using the correlation between the query vector Q and the key vector K;

[0139] The self-attention matrix is ​​mapped and weighted with the value vector V, and then scaled using a scaling factor to obtain the output matrix Attention(Q,K,V);

[0140] The query vector Q, key vector K, and value vector V are projected independently, with each vector being projected h times. The outputs of each projection are then concatenated to obtain the global feature output of the signal.

[0141] The Global Feature Extraction (GFE) module is implemented based on a multi-head attention mechanism. First, it performs different linear transformations on the fused features, as shown in Equation (4):

[0142]

[0143] Where F represents the fusion characteristics of the input signal, W q W k and W v These are learnable parameters, and the weight matrix is ​​used for linear transformation.

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

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

[0146]

[0147] in, This is the scaling factor.

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

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

[0150] Where j = 1, 2, ..., h.

[0151] By splicing the multiple signals together, the global characteristics of the signal are obtained, as shown in formula (7):

[0152] MultiHead=Concat(head1,...,head h ) Formula (7)

[0153] MultiHead is a global feature of the signal.

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

[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 dimensionality of the global features of the signal while retaining 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 probability values, with one probability value corresponding to each category.

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

[0160] The features after average pooling are input into the fully connected layer. Through a combination of linear transformation and nonlinear activation functions, the network can better handle the features between different modulation patterns of the input signal. Finally, the probability of each modulation pattern is output through a softmax layer.

[0161] Step S22: Train the multi-scale feature fusion extraction network model using the modulation signal dataset to obtain the trained multi-scale feature fusion extraction network model.

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

[0163] The communication signal dataset is divided into a training set, a validation set, and a test set according to the modulation pattern and signal-to-noise ratio.

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

[0165] The model is trained using the training set to obtain the loss function and the model trained in this training session. The model trained in this training session is then validated using the validation set to obtain the loss function of the validation set.

[0166] The loss function of the current validation set is compared with the smallest loss function that has been trained. If the loss function of the current validation set is less than the smallest loss function that has been trained, the model trained in the current iteration is saved, and the smallest loss function that has been trained is updated to the loss function value of the current iteration, and the next training iteration begins; otherwise, the next training iteration is performed directly.

[0167] If the loss function of the validation set is larger than the smallest loss function trained in each of the m training iterations, then the learning rate is decayed to n times the original value and training continues; otherwise, the original learning rate is used for training. Here, m and n are preset values.

[0168] If the number of training iterations reaches the total number of training iterations, a well-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 a trained multi-scale feature fusion extraction network model; otherwise, the model is retrained.

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

[0171] During initialization, the batch size was set to 256, the initial learning rate was 0.001, and the total number of training iterations was 200.

[0172] m represents 10 rounds, and n represents 0.4 times.

[0173] The function is trained using the Adam optimizer and the cross-entropy loss function, as shown in Equation (8):

[0174]

[0175] Where, r i For the i-th sample data, P(y=c|r i y is the probability of predicting the c-th modulation pattern. i,c denoted as sample label, W as model parameter, N as total number of training samples, and C as total number of modulation pattern classifications.

[0176] The modulation pattern is predicted using the final trained network parameters, as shown in Equation (9):

[0177]

[0178] Compared to existing deep learning-based modulation pattern recognition models, this invention primarily extracts local features at different scales of the signal and uses a multi-head self-attention mechanism to extract global features, thus fully mining signal characteristics and recognizing the signal modulation pattern. Furthermore, compared to existing deep learning modulation pattern recognition models, this invention optimizes the memory and computational requirements of the model through group convolution, reducing computational costs. The network model proposed in this invention demonstrates higher recognition rates and better noise resistance compared to existing deep learning modulation pattern recognition models.

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

[0180] The effects of this invention can be further illustrated by the following simulations:

[0181] To demonstrate the feasibility of the proposed method of the Multi-scale Feature Extraction and Fusion Network (MFEE) model in this invention, as shown in Figure 7, simulation experiments were conducted to compare the results of each model with CNN network model, LSTM network model, CNN and LSTM hybrid model (MCLDNN, PET-CGDNN), and the Transformer-based network (MCformer). The performance of each model under different signal-to-noise ratios is shown in Figure 7.

[0182] In the examples, under high signal-to-noise ratio conditions with a signal-to-noise ratio greater than 2dB, the average recognition rate of MFEE was 8.38% to 11.61% higher than that of other models, indicating that MFEE can maintain high recognition accuracy under high signal-to-noise ratio conditions and performs best among all comparison models.

[0183] Within the signal-to-noise ratio range of -6dB to 0dB, the recognition accuracy of MFEE decreased due to the low signal-to-noise ratio. However, the average recognition rate of MFEE was still 4.73% to 9.09% higher than that of other models, showing a significant performance improvement. This indicates that MFEE has better noise immunity and recognition performance under low signal-to-noise ratio conditions, demonstrating that MFEE is robust to noise.

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

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

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

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

[0188] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying modulation of non-cooperative communication signals based on multi-scale feature fusion extraction, characterized in that, The process includes the following steps: generating time-domain signals of communication signals under different modulation styles and preprocessing them to obtain multiple baseband IQ signals as sample data, which are then combined with the corresponding modulation styles as sample labels to form a communication signal dataset; initializing a multi-scale feature fusion extraction network model, and training the multi-scale feature fusion extraction network model using the communication signal dataset to obtain a trained multi-scale feature fusion extraction network model; the multi-scale feature fusion extraction network model includes an input layer, a multi-scale convolutional 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 baseband IQ signal slice data; the multi-scale convolutional module MSC uses three parallel group convolutional modules GCNN1, GCNN2, and GCNN3 with different convolutional kernels to extract the first, second, and third local features of the baseband IQ signal slices, and concatenates the first, second, and third local features to obtain the multi-scale local features of the signal; wherein, the group convolutional module GCNN1... The convolutional kernels of GCNN2 and GCNN3 are 3, 7, and 11, respectively, and the group convolutional operations within each group convolutional module are performed with an increasing number of groups G layer by layer. The CNN feature fusion module sequentially uses three one-dimensional convolutions to fuse the multi-scale local features of the signal to obtain the fused features of the signal. The global feature extraction module GFE performs linear transformation and self-attention calculation on the fused 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 highest predicted probability corresponding to the baseband IQ signal. The time-domain signal of the non-cooperative communication signal is acquired in real time and preprocessed, and then input into the trained multi-scale feature fusion extraction network model to obtain the recognition result of the corresponding time-domain signal modulation pattern. The non-cooperative communication signal is a third-party communication signal acquired by the receiver under the condition that it cannot obtain any prior information about the transmitted signal.

2. The method according to claim 1, characterized in that, The time-domain signal is preprocessed, including: converting the time-domain signal into a baseband signal through orthogonal downconversion; filtering the baseband signal to remove noise and obtain a filtered baseband signal; orthogonally demodulating the filtered baseband signal to obtain the corresponding baseband IQ signal; and slicing each baseband IQ signal to obtain multiple baseband IQ signal slices, with each baseband IQ signal slice serving as sample data.

3. The method according to claim 1, 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 local features of the two-dimensional convolution; the local features of the two-dimensional convolution are sequentially input into one-dimensional group convolutions Groupconv1, Groupconv2, and Groupconv3 for feature extraction to obtain local features of the baseband IQ signal; wherein the groups G of the one-dimensional group convolutions Groupconv1, Groupconv2, and Groupconv3 are different, and ReLU activation is performed after each group convolution.

4. The method according to claim 1, characterized in that, The process of performing linear transformation and self-attention calculation on the fused features of the signal to obtain the global features of the signal includes: 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 key vector K; and weighting the self-attention matrix mapping with the value vector V and scaling it using a scaling factor to obtain an output matrix. ; for the query vector Key vector Sum value vector Perform independent projections on each vector. The signal is projected multiple times, and the outputs of each projection are spliced ​​together to obtain the global feature output of the signal.

5. The method according to claim 1, 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 dimensionality of the global features of the signal while retaining 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 probability values, with one probability value corresponding to each category.

6. 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 dataset into a training set, a validation set, and a test set according to the modulation pattern and signal-to-noise ratio; setting the batch size, total number of training iterations, and initial learning rate; training the model using the training set to obtain the loss function and the model trained in this iteration; validating the model trained in this iteration using the validation set to obtain the loss function of the validation set; comparing the loss function of the validation set with the smallest loss function trained so far; if the loss function of the validation set is less than the smallest loss function trained so far, then saving the model trained in this iteration and updating the smallest loss function trained so far. The new loss function value is used for the current iteration, and the next training iteration begins; otherwise, the next training iteration proceeds directly. If the loss function of the validation set is larger than the smallest loss function trained in each of the m training iterations, the learning rate is decayed to n times the original value, and training continues; otherwise, the original learning rate is used for training. Here, m and n are preset values. If the total number of training iterations is reached, 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 a trained multi-scale feature fusion extraction network model; otherwise, model training is restarted.

7. The method according to claim 6, characterized in that, The loss function is the cross-entropy loss function, as shown below: in, For the i-th sample data, To predict the first The probability of a modulation pattern. For sample labels, For model parameters, The total number of training samples, This represents the total number of modulation style categories.

8. The method according to any one of claims 1-7, characterized in that, The modulation styles include 8 digital modulation styles and 3 analog modulation styles; the 8 digital modulation styles include BPSK, QPSK, 8PSK, PAM4, QAM64, AM-DSB, AM-SSB, and CPFSK; the 3 analog modulation styles are GFSK, WBFM, and QAM16.

Images