A communication signal automatic modulation recognition method based on multi-dimensional image information

Abstract

The application discloses a communication signal automatic modulation recognition method based on multi-dimensional image information, and the in-phase / quadrature data is converted into constellation data, differential constellation diagram data and amplitude / phase data after the input signal passes through an information fusion model, then the multi-dimensional data is subjected to kernel density estimation with adaptive bandwidth to obtain the fused multi-dimensional image data; then the image features are extracted through an image feature extraction model, the extracted features are converted from two-dimensional image features into one-dimensional feature vectors through a flattening layer, and the two full connection layers in the classifier are mapped into the prediction value of the modulation category of the input signal, so that the automatic modulation recognition of the communication signal is realized.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, and more specifically, relates to an automatic modulation and recognition method for communication signals based on multi-dimensional image information. Background Technology

[0002] With the rapid development of wireless communication technology, spectrum resources are becoming increasingly valuable. As a key technology in the field of communication, Automatic Modulation Recognition (AMR) aims to automatically and accurately identify received signals without needing to know the parameters of the communication system. AMR technology is particularly important in complex real-world communication environments, especially in non-cooperative systems, where it plays a crucial role in ensuring communication security, effective spectrum management, and preventing malicious attacks. Current AMR technologies mainly include decision theory-based methods and feature pattern recognition-based methods. Decision theory-based methods require prior information about the communication channel. While theoretically they can achieve better classification accuracy, this prior information is often unavailable in real-world non-cooperative systems. Traditional feature pattern recognition-based methods use statistical features for classification tasks, such as higher-order moments, cumulants, and time-frequency features. These methods are limited by the manual extraction of features, leading to the loss of depth information.

[0003] In recent years, with the development of artificial intelligence technology, deep learning methods have shown significant advantages in the field of automatic modulation recognition. Deep learning models such as Convolutional Neural Networks (CNN), Residual Neural Networks (ResNet), Long Short-Term Memory (LSTM), and Transformer Networks (TRN) can automatically learn signal features, reducing the workload of manual feature engineering. However, existing deep learning-based AMR techniques exhibit significantly reduced recognition performance in complex channel environments such as low signal-to-noise ratios and multipath fading. Furthermore, deep learning-based AMR algorithms are highly complex, making them difficult to implement in resource-constrained real-time systems. Therefore, developing lightweight automatic modulation recognition technologies adapted to low noise ratios has significant theoretical and practical value. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide an automatic modulation and recognition method for communication signals based on multi-dimensional image information. This method improves the recognition accuracy of modulation signals under low signal-to-noise ratio by using a lightweight network model based on residual networks and channel attention mechanisms.

[0005] To achieve the above-mentioned objectives, the present invention provides an automatic modulation and recognition method for communication signals based on multi-dimensional image information, characterized by comprising the following steps:

[0006] (1) Extract training data from datasets with known modulation types to form a training dataset;

[0007] (2) Construct a modulation recognition model based on residual network and channel attention mechanism;

[0008] (3) Use the training dataset to train a modulation recognition model based on residual network and channel attention mechanism;

[0009] (4) Use the trained modulation recognition model to identify the sampled data with unknown modulation type and obtain the modulation type of the sampled data.

[0010] The objective of this invention is achieved as follows:

[0011] This invention relates to an automatic modulation recognition method for communication signals based on multi-dimensional image information. The input signal is processed through an information fusion model, which transforms in-phase / orthogonal data into constellation data, differential constellation diagram data, and amplitude / phase data. Then, the multi-dimensional data undergoes adaptive bandwidth kernel density estimation to obtain fused multi-dimensional image data. Next, an image feature extraction model extracts image features. These extracted features are then flattened from two-dimensional image features into one-dimensional feature vectors. Finally, these vectors are mapped to predicted values ​​of the input signal modulation category through two fully connected layers in a classifier, thus achieving automatic modulation recognition of the communication signal.

[0012] Meanwhile, the automatic modulation and recognition method for communication signals based on multi-dimensional image information of the present invention also has the following beneficial effects:

[0013] (1) The present invention is based on a lightweight network model with residual network and channel attention mechanism, which reduces the number of network model parameters and simplifies the hierarchical structure of the network model, thereby improving the training efficiency of the network model. Compared with other methods, this method improves the recognition accuracy of modulated signals under low signal-to-noise ratio and improves the generalization of the model.

[0014] (2) In view of the limitations of signal representation under different data formats, this invention proposes a multi-dimensional image information data, which obtains multi-modal deep-level data information by fusing multi-dimensional data in the modulation domain and time domain.

[0015] (3) To address the impact of signal-to-noise ratio on signal quality and modulation recognition accuracy, an adaptive bandwidth kernel density estimation method is used to estimate the density of signals under different modes. The density estimate is used to replace the estimated data, preserving the spatial network features of the original image data while reducing the interference of signal-to-noise ratio on feature information.

[0016] (4) Based on residual network and channel attention mechanism, an image feature extraction model is proposed. By reducing the kernel size of the convolutional layer, removing some pooling layers, and reducing the number of channels in each residual layer, a lightweight residual network is constructed, which improves the training efficiency of the model and reduces the model complexity. After each residual block, a channel attention mechanism SE module is introduced. Based on the feature value of each channel, the contribution of each channel to the total feature is calculated. The channel weights are assigned according to the contribution. The channel weights are multiplied by the features to obtain the features with added attention mechanism, thereby enhancing the feature information.

[0017] (4) By introducing L2 parameters during training, the loss function is constrained to prevent the trained model from overfitting and improve the model's generalization ability. Attached Figure Description

[0018] Figure 1 This is a flowchart of the automatic modulation and recognition method for communication signals based on multi-dimensional image information according to the present invention;

[0019] Figure 2 This is a structural diagram of a modulation recognition model based on residual networks and channel attention mechanisms;

[0020] Figure 3 This is the confusion matrix diagram tested on the RML2018.01A dataset used in this embodiment, where (a) is the confusion matrix diagram at 6dB signal-to-noise ratio, (b) is the confusion matrix diagram at 0dB signal-to-noise ratio, and (c) is the confusion matrix diagram at -6dB signal-to-noise ratio.

[0021] Figure 4 This is a simulation comparison chart of the recognition accuracy of the present invention and other methods. Detailed Implementation

[0022] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0023] Example

[0024] Figure 1 This is a flowchart of the automatic modulation and recognition method for communication signals based on multi-dimensional image information according to the present invention.

[0025] In this embodiment, as Figure 1 As shown, the present invention provides an automatic modulation and recognition method for communication signals based on multi-dimensional image information, comprising the following steps:

[0026] (1) Extract training data from datasets with known modulation types to form a training dataset;

[0027] The dataset containing 24 modulation types was stratified and extracted to form the training dataset X = {x1, x2, ..., x...} i ,…,x N}, where each set of training data x i For in-phase or quadrature data, it is represented as: This represents orthogonal I / Q data, where n is the sampling point number in the orthogonal data, and N is the number of training data sets.

[0028] In this embodiment, Python is used to process the RML2018.01A dataset. Binary data is read from the data file and converted into in-phase / quadrature complex signal format to obtain the RML2018.01A dataset, as shown in Table 1.

[0029]

[0030] Table 1

[0031] Based on the signal-to-noise ratio (SNR) from -14dB to 30dB, the dataset was divided into 2dB intervals. Under each SNR, 2048 samples were extracted stratified according to 24 modulation categories. Each sample included 1024 sampling points, resulting in 49152 data samples under each SNR. A total of 23 SNRs were used to obtain the training dataset, as shown in Table 2.

[0032]

[0033] Table 2

[0034] (2) Construct a modulation recognition model based on residual network and channel attention mechanism;

[0035] In this embodiment, as Figure 2 As shown, the modulation recognition model based on residual network and channel attention mechanism includes an information fusion model, a feature extraction model, and a classifier model;

[0036] Among them, such as Figure 2 As shown, the information fusion model includes: a sequentially connected multidimensional data extraction layer, a two-dimensional image mapping layer, a kernel density estimation layer, and an image fusion layer; the multidimensional data extraction layer further includes a parallelly connected autocorrelation function layer, a constellation diagram mapping layer, and an amplitude-phase response layer.

[0037] like Figure 2As shown, the feature extraction model includes: convolutional layers and four stacked residual networks; each residual network consists of a residual block and a channel attention mechanism (SE) module connected sequentially; the residual block includes two convolutional layers, two batch normalization layers, and a ReLU layer, and is used to extract multi-channel image features; the SE module includes two convolutional layers, one pooling layer, one ReLU layer, and one Sigmad layer. The SE module assigns channel weights according to the contribution of each channel to the feature, which is used to enhance effective features and suppress useless features;

[0038] like Figure 2 As shown, the classifier model includes: a flattening layer, two fully connected layers, a batch normalization layer, and a ReLU layer; the input to the classifier is the multi-channel image features output by the image feature extraction model, and the classifier completes the prediction of the modulation category.

[0039] (3) Training a modulation recognition model based on residual network and channel attention mechanism;

[0040] (3.1) Extract h = 256 sets of training data from the training dataset as a batch input for this round of training and input them into the modulation recognition model;

[0041] (3.2) Obtain each set of training data x through an information fusion model. i Multidimensional image data;

[0042] (3.2.1) will Converted into constellation data in the modulation domain (I i [n],Q i [n]):

[0043]

[0044] Among them, I i [n] and Q i [n] represents the in-phase and quadrature components of the constellation data, respectively, and T represents the sampling period of the signal;

[0045] (3.2.2) Extracting difference constellation diagram data using autocorrelation function

[0046]

[0047] in, and These represent the real and imaginary components of the differential constellation diagram data, respectively, and k represents the conjugate signal. The difference step size, j represents the imaginary part;

[0048] (3.2.3) will Converted to amplitude / phase data in the time domain (A i [n],P i [n]):

[0049]

[0050] Among them, A i [n] and P i [n] represents the training data x. i The amplitude and phase;

[0051] (3.2.4) Extracting constellation diagram data using adaptive bandwidth kernel density estimation (I) i [n],Q i [n]), Differential constellation data and amplitude / phase data (A i [n],P i The kernel density estimation distribution of [n]) is:

[0052]

[0053] in, Representative (I) i [n],Q i [n]), and (A) i [n],P i Any set of data in [n]) express The kernel density estimation function, h I and h Q They are and The bandwidth of kernel density estimation is given by S, where S is the total number of sampling points, and K(a,b) represents the kernel function for data (a,b), using a Gaussian kernel function. The bandwidth and Gaussian kernel of kernel density estimation are obtained from the following formula:

[0054]

[0055] Where, σ l This represents the standard deviation of the data on the l-axis.

[0056] (3.2.5) Merge the multi-dimensional kernel density estimates into multi-dimensional image data according to the channel direction:

[0057]

[0058] (3.3) Extract feature data from multi-dimensional image data using a feature extraction model;

[0059] (3.3.1) The input to the feature extraction model is multi-dimensional image data. Its dimensions are denoted as [H,W,C]=[64,64,3]; first, Channel expansion is performed using a convolutional layer with dimensions k*k = 3*3, number of channels C1 = 32, stride s = 1, and padding p = 1. After another convolution operation, the initial features output are... Its dimensions are [H] F W F C F = [64, 64, 32], the size satisfies:

[0060]

[0061] C F =C1=32,

[0062] (3.3.2) Initial characteristics Deeper feature data is further extracted using a residual network based on a channel attention mechanism;

[0063] This residual network consists of four stacked residual layers. Each residual layer includes a residual structure and a SE module. The residual structure consists of a sequentially connected convolutional layer + a batch normalization layer + a ReLU layer + a convolutional layer + a batch normalization layer.

[0064] Initial features The input is fed into the first residual network based on the channel attention mechanism, and feature data is obtained through the first residual block. Its dimensions are Features output by residual blocks As input to the first SE module, each channel is sized as The feature data is represented as 1*1 data, and after learning the representation through two convolutional layers, the final output has a dimension of 1*1. The channel weights R1 are then used to obtain the output features with increased channel attention:

[0065]

[0066] in, This is the feature data output by the first residual network, and its dimension is...

[0067] (3.3.3) Feature data The input is used as the input to the next residual network, and the operation process of the first residual network is repeated, and so on, until the feature data is obtained after the fourth residual network operation. Its dimensions are

[0068] (3.4) Predict the type of modulation signal using a classifier model;

[0069] The input to the classifier model is feature data. Each channel size is determined by flattening. The feature matrix flattened into a length is The eigenvectors are obtained, thus yielding the flattened eigenvector matrix F. flatten Its size is The flattened vector matrix undergoes an equal-dimensional feature transformation via a first fully connected layer. This first fully connected layer has 256 neurons in both its input and output layers. It is connected to a batch normalization layer and then uses a ReLU activation layer. The second fully connected layer has 256 neurons in its input layer and 24 neurons in its output layer. These neurons in the second layer use a Softmax activation layer. After being learned and represented by these two fully connected layers, the feature vector matrix yields the predicted probabilities (P1, P2, ..., P...) for 24 modulation types. j ,…,P 24 The final predicted modulation type is:

[0070]

[0071] in, This represents the probability that the i-th training data belongs to the j-th modulation type;

[0072] In this embodiment, after the above processing, the corresponding modulation type can be predicted for all h groups of training data in this batch.

[0073] (3.5) Calculate the loss function value L after this round of training:

[0074]

[0075] Where L represents the loss function value after regularization, L0 represents the loss function of the model before regularization, λ is the regularization coefficient, and w τ p represents the weight parameters in the τth iteration. i ρ represents the probability value of the true modulation type corresponding to the i-th group of training data, and ρ represents the current iteration round;

[0076] (3.6) Determine whether the loss function value L has converged or the number of training iterations has reached the set maximum value. If so, the training ends and the trained modulation recognition model is obtained; otherwise, proceed to step (3.7).

[0077] (3.7) Update the weight parameters of the modulation recognition model;

[0078] Differentiating the loss function yields the weight parameters w for the current ρ-th iteration.ρ :

[0079]

[0080] Update the weight parameters as follows:

[0081]

[0082] Where η is the weighting adjustment coefficient;

[0083] (3.8) Initialize the modulation recognition model according to the updated weight parameters, and then return to step (3.1) for the next round of training;

[0084] (4) Sample data x = (r) of unknown modulation type I [n],r Q [n]) is input into the trained modulation recognition model to predict the modulation type corresponding to the sampled data.

[0085] Example Simulation

[0086] In this embodiment, the PyTorch library was used in Python, and the model was built, trained, and tested on an RTX 4090 GPU server. A simplified version of the RML2018.01A dataset was used as the test dataset. The training epochs were set to 150, and the batch size to 256. The Adam optimizer was used, with a learning rate of 0.001, a regularization coefficient of 0.0001, and a dropout rate of 0.4. The cross-entropy loss function was employed. The confusion matrix of the experiment is shown below. Figure 3 As shown, the horizontal axis represents the predicted label, and the vertical axis represents the actual label; where (a) represents the classification of 24 modulation categories at a signal-to-noise ratio of 6dB, (b) represents the classification at a signal-to-noise ratio of 0dB, and (c) represents the classification at a signal-to-noise ratio of -6dB; Figure 3 It can be seen that at a signal-to-noise ratio of -6dB, the classification accuracy of all modulation categories is higher than 0.5. At a signal-to-noise ratio of 6dB, the confusion levels of AM modulation categories, 32PSK, and 16PSK are significantly improved. Figure 3 The classification results show that the network model of this invention effectively improves the overall classification accuracy of modulation categories; the classification accuracy of this experiment compared with existing methods is as follows: Figure 4As shown, under a signal-to-noise ratio (SNR) greater than 6 dB, the classification accuracy of the method of this invention is among the highest of all experimental methods. Except for the LSTM model, the classification accuracy of the method of this invention is higher than the other four. Under an SNR less than 2 dB, the advantages of the method of this invention are significantly demonstrated. As the SNR decreases, at -14 dB, the classification accuracy of the model network of this invention reaches 0.55, while the classification accuracy of other methods is only around 0.15. This shows that the invention has high performance under a wide range of SNR conditions and has significant advantages under low SNR conditions.

[0087] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for automatic modulation and recognition of communication signals based on multi-dimensional image information, characterized in that, Includes the following steps: (1) Extract training data from datasets with known modulation types to form a training dataset; For containing The datasets of different modulation types are hierarchically extracted to form the training dataset. Each set of training data For in-phase or quadrature data, it is represented as: , express Two orthogonal data streams The sampling point number in the orthogonal data. The number of training data sets; (2) Construct a modulation recognition model based on residual network and channel attention mechanism; (3) Train a modulation recognition model based on residual network and channel attention mechanism using the training dataset; (3.1) Extracting from the training dataset The training data is used as a batch of input for this round of training and is fed into the modulation recognition model; (3.2) Obtain each set of training data through an information fusion model. Multidimensional image data; (3.2.1) will Converted into constellation data in the modulation domain : ； in, and These represent the in-phase and quadrature components of constellation data, respectively. The sampling period of the signal; (3.2.2) Extracting difference constellation diagram data using autocorrelation function : ； in, and These represent the real and imaginary components of the difference constellation diagram data, respectively. Represents conjugate signal The difference step size, Indicates the imaginary part; (3.2.3) will Convert to amplitude / phase data in the time domain : ； in, and They represent the training data respectively. The amplitude and phase; (3.2.4) Extracting constellation diagram data using adaptive bandwidth kernel density estimation Differential constellation data and amplitude / phase data The kernel density estimation distribution: ； in, represent , and Any set of data in express The kernel density estimation function, and They are and Bandwidth of kernel density estimation This is the total number of sampling points. Representing data The kernel function for kernel density estimation is a Gaussian kernel function; the bandwidth and Gaussian kernel for kernel density estimation are obtained by the following formula: ； ； ； in, express Standard deviation of data on the road; (3.2.5) Merge the multi-dimensional kernel density estimates into multi-dimensional image data according to the channel direction: ； (3.3) Extract feature data from multi-dimensional image data using a feature extraction model; (3.3.1) The input to the feature extraction model is multi-dimensional image data. Its dimensions are denoted as First, Channel expansion is performed using a convolutional layer with a size of [size missing]. The number of channels is stride is Fill with After another convolution operation, the initial features output are: Its size is The size meets the following requirements: ； ； ； (3.3.2) Initial characteristics Deeper feature data is further extracted using a residual network based on a channel attention mechanism; This residual network consists of four stacked residual layers. Each residual layer includes a residual structure and a SE module. The residual structure consists of a sequentially connected convolutional layer + a batch normalization layer + a ReLU layer + a convolutional layer + a batch normalization layer. Initial features The input is fed into the first residual network based on the channel attention mechanism, and feature data is obtained through the first residual block. Its size is Features output by the residual block As input to the first SE module, each channel is sized as The feature data is represented as 1*1 data, and after learning the representation through two convolutional layers, the final output has a dimension of 1*1. Channel weights Finally, the output features with increased channel attention are obtained: ； in, This is the feature data output by the first residual network, and its dimension is... ; (3.3.3) Feature data The input is used as the input to the next residual network, and the operation process of the first residual network is repeated, and so on, until the feature data is obtained after the fourth residual network operation. Its size is ; (3.4) Predict the type of modulation signal using a classifier model; The input to the classifier model is feature data. The size of each channel is determined by flattening the layers. The feature matrix flattened into a length is The eigenvectors are obtained, thus yielding the flattened eigenvector matrix. Its size is The flattened vector matrix undergoes an equal-dimensional feature transformation through a first fully connected layer. This first fully connected layer has E neurons in its input layer and E neurons in its output layer. It is connected to a batch normalization layer and then uses a ReLU activation layer. The second fully connected layer has E neurons in its input layer and M neurons in its output layer. The neurons in this second layer use a Softmax activation layer. After being learned and represented by these two fully connected layers, the feature vector matrix yields the predicted probabilities for M modulation types. The final predicted modulation type is: ； in, Indicates the first The training data of the group belongs to the first The probability of a modulation type; (3.5) Calculate the loss function value after this round of training. : ； ； in, This represents the loss function value after regularization. This represents the loss function of the model before regularization is applied. It is the added regularization coefficient. Indicates the first Weight parameters during round iteration Indicates the first The probability value of the true modulation type corresponding to the training data set. Indicates the current iteration round; (3.6) Determine the value of the loss function If the training has converged or the number of training iterations has reached the set maximum value, the training ends and the trained modulation recognition model is obtained; otherwise, proceed to step (3.7). (3.7) Update the weight parameters of the modulation recognition model; Taking the derivative of the loss function, we obtain the current... Weight parameters during round iteration : ； Update the weight parameters as follows: ； in, This is the weighting adjustment coefficient; (3.8) Initialize the modulation recognition model according to the updated weight parameters, and then return to step (3.1) for the next round of training; (4) Use the trained modulation recognition model to identify the sampled data with unknown modulation type and obtain the modulation type of the sampled data.

2. The automatic modulation and recognition method for communication signals based on multi-dimensional image information according to claim 1, characterized in that, The modulation recognition model based on residual networks and channel attention mechanism includes an information fusion model, a feature extraction model, and a classifier model; The information fusion model includes: a sequentially connected multidimensional data extraction layer, a two-dimensional image mapping layer, a kernel density estimation layer, and an image fusion layer; the multidimensional data extraction layer further includes a parallelly connected autocorrelation function layer, a constellation diagram mapping layer, and an amplitude-phase response layer. The feature extraction model includes: convolutional layers and four stacked residual networks; each residual network consists of a residual block and a channel attention mechanism (SE) module connected sequentially; wherein, the residual block includes two convolutional layers, two batch normalization layers and one ReLU layer, and the residual block is used to extract multi-channel image features; the SE module includes two convolutional layers, one pooling layer, one ReLU layer and one Sigmoid layer, and the SE module assigns channel weights according to the contribution of each channel to the feature, which is used to enhance effective features and suppress useless features; The classifier model includes: a flattening layer, two fully connected layers, a batch normalization layer, and a ReLU layer; the classifier input is the multi-channel image features output by the image feature extraction model, and the classifier completes the prediction of the modulation category.

Images