Wireless signal automatic modulation identification method based on CGAF and residual error identification network model

Abstract

The invention discloses a wireless signal automatic modulation identification method based on a CGAF and a residual error identification network model, and the method comprises the steps: firstly, enabling the I, Q and I / Q one-dimensional time features of a wireless signal to be mapped to a two-dimensional space image domain through a complex number field Gramb angle field algorithm; therefore, the technical bottlenecks of high time domain feature similarity and low feature space separability of a modulation mode are effectively solved. On the basis, a channel segmentation residual neural network is designed as a classifier, feature extraction is performed through channel segmentation and an attention mechanism, and information is simplified by using a residual module, so that the classification precision of a modulation mode and the system robustness are remarkably improved. According to the method provided by the invention, the accuracy and robustness of modulation recognition are remarkably improved in a complex wireless channel environment, the problems of single feature and insufficient model learning ability of a traditional method are solved, and reliable technical support is provided for application of automatic modulation recognition in the fields of cognitive radio, military reconnaissance and the like.

Description

Technical Field

[0001] The present invention relates to the field of communication signal processing, and is mainly aimed at the problem of automatic classification of modulated signals with highly similar time domain features in complex electromagnetic environments. Specifically, it is a method for automatic modulation recognition of wireless signals based on CGAF and residual recognition network models. Background Art

[0002] Automatic modulation recognition provides an important technical means for electromagnetic spectrum situational awareness in military applications such as satellite communication monitoring and tactical data link analysis. However, due to complex factors such as wireless channel interference, the time domain characteristics of specific modulation signals are highly similar, making them difficult to distinguish.

[0003] Automatic modulation identification (AMRI) is a fundamental branch of cognitive radio and military reconnaissance, aiming to identify signal modulation types affected by time-varying channels. Traditional AMI methods can be broadly categorized into two main categories: likelihood-based (LB) methods and feature-based (FB) methods. LB methods construct likelihood functions between the received signal and candidate modulation types and perform classification based on the maximum likelihood criterion. Their theoretical classification error can reach the Cramér-Rao lower bound. However, the computational complexity of LB methods is limited by the high-dimensional joint probability density estimation problem and their strong dependence on prior knowledge such as channel state information, significantly limiting their practicality in dynamic, time-varying channels. In contrast, FB methods employ a two-stage architecture of "feature extraction + classifier": first, discriminative signal features are extracted using an expert-designed feature extractor, and then classification is performed based on machine learning models (such as support vector machines and random forests). Although FB methods reduce computational complexity through feature dimensionality reduction, their performance is limited by the empirical and scenario-specific nature of manual feature design, making them difficult to effectively capture the time-frequency nonlinear coupling characteristics of modulation features under complex channel conditions.

[0004] In recent years, the rapid development of deep learning has driven research on data-driven automatic modulation recognition methods. Compared with traditional automatic modulation recognition methods, deep learning can automatically learn the feature representation of modulated signals, reduce reliance on prior knowledge, and demonstrate strong robustness and generalization capabilities in complex environments. However, current deep learning-based automatic modulation recognition methods still face two technical bottlenecks: First, the current mainstream datasets only provide the in-phase / quadrature (I / Q) time-domain sampling sequence of the baseband signal and lack multi-domain joint features, resulting in insufficient feature space separability; second, the mainstream neural network architecture has a significantly better perception of multi-dimensional features than one-dimensional time series signals, and the sparse time-domain structure of the I / Q sequence further exacerbates the difficulty of model learning. These problems jointly restrict the practical application of deep learning-based automatic modulation recognition methods in complex interference scenarios. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides a method for automatic modulation recognition of wireless signals based on CGAF and residual recognition network model. This method maps a one-dimensional time signal sequence into a two-dimensional joint feature map through the CGAF (complex domain Gram angular field) algorithm, realizes the time-frequency domain union, and combines the channel segmentation residual network to perform multi-level feature fusion and classification. This method not only overcomes the defects of the existing automatic modulation recognition method based on deep learning in the single time domain feature expression and the insufficient perception ability of the neural network for one-dimensional signals, but also significantly improves the recognition accuracy and robustness of the system in complex wireless channel environments.

[0006] The technical solution of the present invention to solve the above technical problems is to design a method for automatic modulation identification of wireless signals based on CGAF and residual recognition network model, characterized in that the method specifically includes the following steps:

[0007] S1: Obtain training data

[0008] Select signal samples of I / Q timing signals containing K modulation schemes as the original data set, where the K modulation schemes include digital modulation scheme types and analog modulation scheme types; divide no less than 60% of the signal samples in the original data set into a training set, and the rest into a validation set, and both the training set and the validation set contain signal samples of the above K modulation schemes;

[0009] S2: Data preprocessing

[0010] The sample signals in the training set and the validation set are processed respectively, and the I channel signal and Q channel signal in the time domain of each sample signal are extracted respectively, and the corresponding IQ channel signal is obtained, where IQ = I + jQ, where j is a complex unit; then the I channel signal and the Q channel signal in the time domain are fast Fourier transformed respectively to obtain the I channel signal and the Q channel signal in the complex domain respectively;

[0011] Afterwards, the complex domain Gram angle field is used to perform feature mapping on the complex domain I channel signal, Q channel signal, and IQ channel signal respectively to obtain the mapping images of the corresponding channels. The mapping images of each channel are then connected to the same feature space according to the channel dimension to obtain the global mapping image of the sample signal.

[0012] After processing each sample signal in the training set and the validation set into a corresponding global mapping image, a training image set and a validation image set are obtained respectively;

[0013] S3: Building a recognition network model

[0014] The recognition network model includes a PCN module, an FD-Unit module, a first AE-Unit module, a first Res-Unit module, a second AE-Unit module, a second Res-Unit module, a third AE-Unit module, a third Res-Unit module, a GRU module, an Avgpool layer, an FC layer, and a Softmax layer;

[0015] The PCN module performs Z normalization on the eigenvalues ​​of each channel of the input image, and then concatenates the results of the Z normalization of the eigenvalues ​​of each channel in channel order to obtain the output of the PCN module;

[0016] The output of the PCN module is the input of the FD-Unit module; the FD-Unit module includes a channel segmentation operation, three convolution branch networks and a channel splicing operation. The features input to the FD-Unit module first undergo a channel segmentation operation to evenly divide the output features of the PCN module into three intermediate features according to the channel dimension, and then the three intermediate features are respectively input to three parallel convolution branch networks; the outputs obtained by the three parallel convolution branch networks are then subjected to a channel splicing operation to obtain the output of the FD-Unit module; the first convolution branch in the three parallel convolution branch networks consists of a 3×3 convolution kernel depthwise separable convolution module and a 1×1 convolution kernel generalized convolution module. The first convolution branch is composed of two convolution modules connected in series, the output of the previous module is used as the input of the next module, and the output of the last module is the output of the convolution branch; the second convolution branch is composed of a 3×3 convolution kernel ordinary convolution module and a 1×1 convolution kernel depthwise separable convolution module connected in series, the output of the previous module is used as the input of the next module, and the output of the last module is the output of the convolution branch; the third convolution branch is composed of a 3×3 convolution kernel depthwise separable convolution module and two 1×1 convolution kernel ordinary convolution modules connected in series, where the 3×3 convolution kernel depthwise separable convolution module is set between the two 1×1 convolution kernel ordinary convolution modules, and the output of the last 1×1 convolution kernel ordinary convolution module is the output of the convolution branch;

[0017] The first AE-Unit module takes the output of the FD-Unit module as input; the first AE-Unit module includes a channel splitting operation, two convolution branches and a channel splicing operation; the features input to the first AE-Unit module first undergo a channel splitting operation to divide the input features into two intermediate features, and then the two intermediate features are respectively input into two parallel convolution branch networks. The outputs obtained by the two parallel convolution branch networks are then subjected to a channel splicing operation to obtain the output of the AE-Unit module; the first convolution branch in the two parallel convolution branch networks is composed of a 3×3 convolution kernel, a depth-separable convolution module with a stride of 2, a channel attention module, and a 1×1 convolution kernel ordinary convolution module in series. The output of the previous module is used as the input of the next module, and the output of the last module is the output of the convolution branch; the second convolution branch is symmetrical with the first convolution branch, and its setting is exactly the same as the first convolution branch;

[0018] The first Res-Unit module takes the output of the first AE-Unit module as input. The first Res-Unit module consists of two convolution branches, the main and side, and a channel splicing operation. The features input to the first Res-Unit module are respectively input into the two convolution branches. The results of the two convolution branches are subjected to a channel splicing operation to obtain the output of the first Res-Unit module. The main convolution branch is composed of a 1×1 convolution kernel ordinary convolution module, a 3×3 convolution kernel grouped convolution module, and a 1×1 convolution kernel ordinary convolution module connected in series. The output of the previous module is used as the input of the next module, and the output of the last module is the output of the convolution branch. The side convolution branch consists of only a 3×3 convolution kernel depth-separable convolution module, and its output is the output of the side convolution branch of the first Res-Unit module. After processing by the first Res-Unit module, the feature size remains unchanged, and the feature channel is doubled.

[0019] The basic structures of the first AE-Unit, the second AE-Unit module, and the third AE-Unit module are the same, and the network parameters are not shared; the basic structures of the first Res-Unit module, the second Res-Unit module, and the third Res-Unit module are the same, and the network parameters are not shared; after each AE-Unit module, the number of feature channels doubles and the size of the feature map is halved; after each Res-Unit module, the number of feature channels doubles and the size of the feature map remains unchanged;

[0020] The feature map output by the third Res-Unit module undergoes a flattening operation to flatten the spatial dimension of the feature map into one dimension, and this dimension is used as the sequence length, that is, the time step, and the channel dimension is used as the feature dimension. The obtained output is used as the input of the GRU module;

[0021] The output of the GRU module is processed in turn by the average pooling layer, the fully connected layer, and the Softmax layer to obtain the probability distribution of each modulation type of the signal. The modulation type corresponding to the maximum probability is the modulation type of the signal;

[0022] S4: Training the recognition network model

[0023] In the process of training the recognition network model, the trainable parameters of the recognition network model are first initialized; the convolutional layer weights are assigned using the He initialization method; the bias term is initialized to 0; the learning rate is initially set to 0.001, and the weight decay strategy is used to prevent overfitting, and the regularization coefficient is set to e -4 ;

[0024] The input of the recognition network model training is the global mapping image F of the sample signal CGAF , the corresponding label is the modulation type to which the sample signal belongs, expressed as a one-hot encoding; the recognition network model outputs the predicted probability of each category of the sample signal on the K types of modulation modes; the training loss of a batch is calculated using the cross entropy loss function to measure the prediction value The difference between the true label y is expressed as:

[0025]

[0026] Where B represents the number of samples in a batch, y i，c is the actual label of sample i in category c, is the model's predicted probability of sample i for category c;

[0027] According to the training loss of a batch, the Adam optimizer is used for parameter optimization to update the trainable parameters of the recognition network model once, completing the training of a batch of sample signals; the network parameters when the previous batch of sample signals is trained are used as the initial parameters for the next batch of training, and this process is repeated continuously until the last batch of sample signals in the training image set is trained, completing a round of training; after each round of training, the sample signals in the verification image set are input into the current network model, and the loss value of the network model in this round on the verification set is calculated; the network parameters when the previous round of training is completed are used as the initial parameters for the next round of training, and the training process is repeated for a round. When the early stopping mechanism condition set according to the loss value on the verification set is reached or the training round reaches the preset maximum value, the training of the recognition network model is completed; the set of network parameters with the smallest loss value on the verification set is selected and saved to obtain the optimal recognition network model;

[0028] S5: Automatic modulation recognition

[0029] Obtain a wireless signal to be identified. First, convert the wireless signal into an I / Q timing signal form through IQ modulation. Then, according to the data preprocessing method in S2, obtain a global mapping image of the wireless signal. Then, input the global mapping image into the optimal recognition network model in S4 to obtain the probability distribution of the K-type modulation mode of the wireless signal. The modulation mode type corresponding to the maximum probability is the modulation mode type of the wireless signal.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention is based on a wireless signal automatic modulation recognition method based on CGAF and residual recognition network models. This method first maps the I, Q, I / Q one-dimensional time features of the wireless signal to the two-dimensional spatial image domain through the complex domain Gram angular field algorithm, thereby effectively solving the technical bottleneck of high similarity of time domain features and low feature space separability of the modulation mode. On this basis, a channel segmentation residual neural network is designed as a classifier, feature extraction is performed through channel segmentation and attention mechanism, and the residual module is used to simplify information, thereby significantly improving the classification accuracy and system robustness of the modulation mode. The method proposed in the present invention performs well when processing modulation signal types with high time domain feature similarity (such as GFSK and CPFSK, AM-SSB and AM-SSB). And under high signal-to-noise ratio, the highest recognition accuracy reaches 93.88%, which is better than the existing automatic modulation recognition method. Comparative experimental results show that the method proposed in this invention significantly improves the accuracy and robustness of modulation recognition in complex wireless channel environments, overcomes the problems of traditional methods in terms of single features and insufficient model learning ability, and provides reliable technical support for the application of automatic modulation recognition in fields such as cognitive radio and military reconnaissance. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a structural diagram of the recognition network model of an embodiment of the wireless signal automatic modulation recognition method based on CGAF and residual recognition network model of the present invention.

[0032] Figure 2 It is a schematic diagram of the structure and principle of the FD-Unit module of the recognition network model of an embodiment of the wireless signal automatic modulation recognition method based on CGAF and residual recognition network model of the present invention.

[0033] Figure 3 It is a schematic diagram of the structure and principle of the first AE-Unit module of the recognition network model of an embodiment of the wireless signal automatic modulation recognition method based on CGAF and residual recognition network model of the present invention.

[0034] Figure 4It is a schematic diagram of the structure and principle of the first Res-Unit module of the recognition network model of an embodiment of the wireless signal automatic modulation recognition method based on CGAF and residual recognition network model of the present invention.

[0035] Figure 5 This is a schematic diagram of the structure and principle of the GRU module of the recognition network model of an embodiment of the wireless signal automatic modulation recognition method based on CGAF and residual recognition network model of the present invention.

[0036] Figure 6 This is a flow chart of a method for performing feature mapping of I / Q timing signals using a complex domain Gram angle field to obtain corresponding mapping images in an embodiment of the present invention's method for automatic modulation recognition of wireless signals based on CGAF and residual recognition network models.

[0037] Figure 7 This is the recognition accuracy curve of the wireless signal automatic modulation recognition method based on CGAF and residual recognition network model of the present invention and the other five automatic modulation recognition methods under different signal-to-noise ratios. DETAILED DESCRIPTION

[0038] The specific embodiments of the present invention are given below. The specific embodiments are only used to further illustrate the present invention and do not limit the scope of protection of the claims of the present invention.

[0039] The present invention provides a method for automatic modulation recognition of wireless signals based on CGAF and residual recognition network model, which specifically includes the following steps:

[0040] S1: Obtain training data

[0041] Signal samples of I / Q timing signals containing K modulation schemes are selected as the original data set, where the K modulation schemes include digital modulation scheme types and all analog modulation scheme types; no less than 60% of the signal samples in the original data set are divided into a training set, and the rest are a validation set, and both the training set and the validation set contain signal samples of the above-mentioned K modulation schemes.

[0042] This example uses the public dataset RML2016.10a, which contains 220,000 signal samples from eight digital modulation schemes and three analog modulation schemes. Each signal sample includes a signal sequence and a corresponding modulation type label. The eight digital modulation schemes are 8PSK, BPSK, CPFSK, GFSK, 4-PAM, 16-QAM, 64-QAM, and QPSK; the three analog modulation schemes are AM-DSB, AM-SSB, and WBFM. The signal-to-noise ratio ranges from -20dB to 18dB, with a 2dB increment.

[0043] 80% of the public RML2016.10a dataset was used as the original dataset. 75% of the signal samples in the original dataset were divided into the training set and 25% into the validation set. The remaining 20% ​​of the public RML2016.10a dataset was used as the test set. Each of the training, validation, and test sets contained signal samples for the eight digital modulation schemes and three analog modulation schemes mentioned above.

[0044] S2: Data preprocessing

[0045] The RML2016.10a dataset is stored as I / Q time series signals with a dimension of 2×128, where 2 represents the I channel and Q channel, and 128 represents the number of sampling points for each signal. Common representations are as follows:

[0046] r(n)=A(n)e j(ωn+θ) x(n)+σ(n),n=0,1,…,N-1

[0047]

[0048] Where r(n) represents the modulated signal preprocessed by the transmitter, σ(n) is additive Gaussian noise, A(n) represents the channel gain, ω and θ represent the frequency offset and phase offset, respectively, and j is a complex unit. N = 128. x(n) represents the original signal sequence, whose real and imaginary parts constitute the I-channel and Q-channel signals, respectively. The IQ-channel signal is a complex signal constructed from the I-channel and Q-channel signals.

[0049] The sample signals in the training set, validation set, and test set are processed separately, and the time domain I channel signal and Q channel signal of each sample signal are extracted. The corresponding IQ channel signal is obtained, and IQ = I + jQ. The time domain I channel signal and Q channel signal are then transformed by fast Fourier transform (FFT) to obtain the complex domain I channel signal and Q channel signal.

[0050] Fast Fourier Transform (FFT) can convert time domain signals into complex domain signals. The formula is as follows:

[0051]

[0052] Where X[k] represents the result after FFT transformation, T[n] represents the extracted I channel signal or Q channel signal in the time domain; N is the number of sampling points of the signal.

[0053] Afterwards, the complex domain Gram Angular Field (CGAF) is used to perform feature mapping on the I-channel signal, Q-channel signal and IQ-channel signal in the complex domain to obtain the mapping images of the corresponding channels. Then, the mapping images of each channel are connected to the same feature space according to the dimension of the channel to obtain the global mapping image of the sample signal.

[0054] Taking the IQ channel signal as an example, the process of using the complex domain Gram angle field to perform feature mapping to obtain the mapping image of the corresponding channel is explained:

[0055] Assume that the IQ channel signal of a sample signal is X=[X1, X2, ..., X n ,…,X N-1 ], element X n is a complex point, which can be expressed in terms of magnitude and phase: where |X n | is the modulus of the element point, is the phase of the element point, j is a complex unit, θ n For element X n The phase angle of . First, the complex signal needs to be scaled by the following formula:

[0056]

[0057] Then, based on the time domain mapping and the inner product formula in the complex domain, the following new mapping method in the complex domain is adopted:

[0058] CGAF[i, j]=sin(arg[(X′ i ·X′ j ) * ])

[0059] X′ i , X′ j is the element X in the complex signal X i 、X j The corresponding eigenvalue obtained after scaling using the above formula, X′ i ·X′ j Indicates the calculation of the inner product of the two, (·) * Indicates that the inner product result is in complex conjugate form, arg[] is the function of taking the phase angle, and CGAF[i, j] represents the element X in the IQ channel signal X of the sample signal. i 、X j The pixel value of the pixel coordinate [i, j] of the mapped CGAF image; map each element in the IQ channel signal X to obtain the mapped image IQ of the IQ channel signal X CGAF ;

[0060] Using the same method, the complex domain I channel signal and Q channel signal are feature mapped using the complex domain Gram angle field to obtain the corresponding mapping image I of the I channel signal. CGAF , the mapping image Q of the Q channel signal CGAF ; Then map the image I CGAF , Mapping Image IQ CGAF , mapping image Q CGAF Connect to the same feature space according to the dimension of the channel to obtain the global mapping image F of the sample signal CGAF .

[0061] F CGAF =Concat(I CGAF , IQ CGAF , Q CGAF )

[0062] After processing each sample signal in the training set, validation set, and test set into a corresponding global mapping image, the training image set, validation image set, and test image set are obtained respectively.

[0063] S3: Building a recognition network model

[0064] The recognition network model is a residual recognition network model, specifically including a PCN module, an FD-Unit module, a first AE-Unit module, a first Res-Unit module, a second AE-Unit module, a second Res-Unit module, a third AE-Unit module, a third Res-Unit module, a GRU (gated recurrent unit) module, an Avgpool layer, an FC layer, and a Softmax layer.

[0065] The PCN module performs Z-normalization processing on the eigenvalues ​​of each channel of the input image, and then concatenates the results of the Z-normalization processing of the eigenvalues ​​of each channel in channel order to obtain the output of the PCN module.

[0066] The specific working principle of the PCN module is as follows: Assume that the image samples x in a given batch i (Its shape is C×H×W, i.e. it contains C channels, height is h, and width is W), and for each channel there is the following normalization formula:

[0067]

[0068] where x i，c represents the eigenvalue of the cth channel of the i-th sample, μ i，c is the mean of the channel eigenvalues, σ i，c is the standard deviation of the channel eigenvalue, x i，c，h，wrepresents the pixel value of the i-th sample in the batch at the c-th channel, h-th row, and w-th column. In addition, in order to avoid division by 0, when σ i，c = 0, then let σ i，c = 1. At the same time, after preprocessing the samples of each channel, they are combined according to the original channel order to obtain the output of the PCN module M is the number of samples in a batch, C is the image sample x i The input image is processed by the PCN module to obtain an output feature with a dimension of [3, 128, 128]. That is, the dimension of the output feature is consistent with the input image.

[0069] The output of the PCN module is the input of the FD-Unit module. The FD-Unit module includes a channel splitting operation (i.e. Figure 1 S operation in ), three convolutional branch networks and a channel splicing operation (i.e. Figure 1 The features input to the FD-Unit module first undergo a channel splitting operation to evenly split the output features of the PCN module into three intermediate features according to the channel dimension. That is, the input of [3, 128, 128] is split into three intermediate features of [1, 128, 128]. These three intermediate features are then input into three parallel convolutional branch networks respectively. The outputs of the three parallel convolutional branch networks then undergo a channel splicing operation to obtain the output of the FD-Unit module, whose output dimension is [24, 128, 128]. The first convolution branch in the three parallel convolutional branch networks consists of a 3×3 convolutional kernel depthwise separable convolution module and a 1×1 convolutional kernel ordinary convolution module in series. The output of the first module serves as the input of the second module, and the output of the last module serves as the output of the convolution branch. The second convolution branch consists of a 3×3 convolutional kernel ordinary convolution module and a 1×1 convolutional kernel depthwise separable convolution module in series. The output of the first module serves as the input of the second module, and the output of the last module serves as the output of the convolution branch. The third convolution branch consists of a 3×3 convolution kernel depthwise separable convolution module and two 1×1 convolution kernel ordinary convolution modules in series, where the 3×3 convolution kernel depthwise separable convolution module is set between the two 1×1 convolution kernel ordinary convolution modules, and the output of the last 1×1 convolution kernel ordinary convolution module is the output of this convolution branch.

[0070] The first AE-Unit module takes the output of the FD-Unit module as input. It consists of a channel splitting operation, two convolution branches, and a channel concatenation operation. The features input to the first AE-Unit module first undergo a channel splitting operation, splitting the input features of [24, 128, 128] into two intermediate features of [12, 128, 128]. These two intermediate features are then fed into two parallel convolutional branch networks. The outputs of these two parallel convolutional branch networks then undergo a channel concatenation operation to produce the output of the AE-Unit module, with dimensions of [48, 56, 56]. The first convolutional branch in the two parallel convolutional branch networks consists of a 3×3 convolution kernel, a depthwise separable convolution module with a stride of 2, a channel attention module (SEblock), and a 1×1 convolution kernel, all connected in series. The output of the first module serves as the input to the second module, and the output of the last module serves as the output of the convolution branch. The second convolution branch is symmetrical to the first, and its configuration is identical to that of the first. The channel attention mechanism of the channel attention module calculates feature weights through global average pooling to achieve adaptive feature enhancement.

[0071] The first Res-Unit module takes the output of the first AE-Unit module as input and consists of two convolutional branches, a main convolutional branch and a side convolutional branch, followed by a channel concatenation operation. The features input to the first Res-Unit module are fed into the two convolutional branches, and the results of the two convolutional branches are concatenated to form the output of the first Res-Unit module. The main convolutional branch consists of a 1×1 normal convolutional module, a 3×3 grouped convolutional module, and a 1×1 normal convolutional module, connected in series. The output of each module serves as the input to the next module, and the output of the last module serves as the output of that convolutional branch. The side convolutional branch consists of a single 3×3 depthwise separable convolutional module, the output of which serves as the output of the side convolutional branch of the first Res-Unit module. After processing by the first Res-Unit module, the feature size remains unchanged, but the number of feature channels is doubled, i.e., [96, 56, 56].

[0072] The basic structures of the first AE-Unit, the second AE-Unit module, and the third AE-Unit module are the same, and the network parameters are not shared. The basic structures of the first Res-Unit module, the second Res-Unit module, and the third Res-Unit module are the same, and the network parameters are not shared. Figure 1 The network structure diagram is analogous to this. After each AE-Unit module, the number of feature channels doubles and the size of the feature map is halved. After each Res-Unit module, the number of feature channels doubles and the size of the feature map remains unchanged.

[0073] The feature map output by the third Res-Unit module is flattened (i.e. Figure 1 The F operation in (Flattening) flattens the spatial dimensions (height and width) of the feature map into one dimension, and uses this dimension as the sequence length, i.e., the time step, and the channel dimension as the feature dimension. The output is used as the input of the GRU module. After processing by the third Res-Unit module, the feature dimension of the feature map obtained is [384, 14, 14]. First, the height and width dimensions are merged to obtain a two-dimensional matrix of [384, 196], where 196 = 14 × 14. Then it is serialized, with the channel dimension (384) as the feature dimension and the dimension after spatial flattening (196) as the sequence length. Therefore, the final data shape input to the GRU module is [384, 196].

[0074] The GRU module implements temporal dynamic modeling through a gating mechanism. Its process can be divided into four steps: First, the input feature contains the input vector x of the current time step. t and the hidden state h at the previous time step t-1 , where t represents the input time step, and the two are used as initial input. Then they are processed in parallel by the dual gate branch: In the update gate branch, x t and h t-1 After the weight matrix W Z 、U Z Linear transformation and bias b z Superposition, and then activation through the Sigmoid (σ) function to generate the updated gating signal z t =σ(W Z x t +U Z h t-1 +b z ), used to control the retention ratio of historical information; in the symmetric reset gate branch, the independent parameter W r 、U r and b r Generate reset gate signal r t =σ(W r x t +U r h t-1 +b r ), determines the degree of forgetting of the historical state. Then, the candidate hidden state Generated by reset gate control: linear transformation W h x t and the reset history state U h (r t ⊙h t-1 ) and add bias b h, activated by Tanh Where W h and U h Process x separately t and h t-1 The linear matrix, ⊙ represents element-by-element multiplication, and the formula is expressed as A=[a ij ],B=[b ij ],A⊙B=[a ij b ij ]. Finally, update the goalkeeper's historical state h t-1 With candidate status Fusion by weight, output current hidden state Complete the iterative update of the timing status.

[0075] The output of the GRU module is processed in sequence by the average pooling layer (Avgpool), the fully connected layer (FC), and the Softmax layer to obtain the probability distribution of each modulation type of the signal. The modulation type corresponding to the maximum probability is the modulation type of the signal.

[0076] S4: Training the recognition network model

[0077] In the process of training the recognition network model, the trainable parameters of the recognition network model are first initialized. The He initialization method is used to assign values ​​to the convolutional layer weights to adapt to the characteristics of the RELU activation function and improve the convergence speed of the network; the bias term is initialized to 0. Its learning rate is initially set to 0.001, and the weight decay (L2 regularization) strategy is used to prevent overfitting. The regularization coefficient is set to e -4 .

[0078] The input of the recognition network model training is the global mapping image F of the sample signal CGAF , whose shape is [3, 128, 128], and the corresponding label is the modulation type of the sample signal, expressed as one-hot encoding. The recognition network model outputs the predicted probability of each category of the sample signal on K types of modulation (11 in RML2016.10a). The training loss of a batch is calculated using the cross entropy loss function to measure the prediction value. The difference between the true label y is expressed as:

[0079]

[0080] Where B represents the number of samples in a batch, y i，c is the actual label of sample i in category c, is the model's predicted probability of sample i for category c.

[0081] Based on the training loss of a batch, parameter optimization uses the Adam optimizer to update the trainable parameters of the recognition network model, completing the training of a batch of sample signals. The network parameters at the completion of the previous batch of sample signals are used as the initial parameters for the next batch of training, and this process is repeated until the last batch of sample signals in the training image set is trained, completing a training round. After each training round, the sample signals in the validation image set are input into the current network model, and the loss of the network model on the validation set is calculated. The network parameters at the completion of the previous training round are used as the initial parameters for the next training round, and the training process is repeated. The recognition network model training is completed when the early stopping condition set based on the loss value on the validation set is met or the number of training rounds reaches a preset maximum. The set of network parameters with the lowest loss value on the validation set is selected and saved to obtain the optimal recognition network model.

[0082] The early stopping mechanism automatically terminates training when the validation set loss does not decrease for 10 consecutive rounds.

[0083] S5: Automatic modulation recognition

[0084] Obtain a wireless signal to be identified. First, convert the wireless signal into an I / Q timing signal form through IQ modulation. Then, according to the data preprocessing method in S2, obtain a global mapping image of the wireless signal. Then, input the global mapping image into the optimal recognition network model in S4 to obtain the probability distribution of the K-type modulation mode of the wireless signal. The modulation mode type corresponding to the maximum probability is the modulation mode type of the wireless signal.

[0085] To validate the model's performance, the original dataset, excluding the training and validation sets, was divided into a test set containing signal samples from the eight digital modulation schemes and three analog modulation schemes described above. A series of experiments were conducted using the optimal recognition network model. The results demonstrated that the proposed method performed well when dealing with easily confused modulation types (such as GFSK and CPFSK, and AM-SSB and AM-SSB). Furthermore, the highest recognition accuracy reached 93.88% at high signal-to-noise ratios.

[0086] Table 1 shows the highest and average recognition accuracy rates for several algorithms, comparing the proposed method with several state-of-the-art automatic modulation recognition methods, including TFA-SCNN, GAF-MAE, FC-MLP, ResNeXt-GRU, and iFormer. Each method was trained on the same training and validation sets, and recognition performance data was calculated on the same test set.

[0087] Table 1 Comparison of recognition effects between the method of the present invention and other methods

[0088]

[0089]

[0090] Figure 7 The recognition accuracy (acc) curves of each model under different signal-to-noise ratios (dB) are displayed. The results show that the recognition performance of the method proposed in the present invention is significantly better than other models. At 18dB, the recognition accuracy rate reaches 93.88%. In addition, the method performs well when processing signal modulation types whose time domain characteristics are easily confused (such as CPFSK and GFSK, AM-DSB and AM-SSB). Among them, the highest recognition rate for AM-DSB and AM-SSB can reach 100%, and the recognition rates for CPFSK and GFSK reach 97% and 96% respectively. This shows that the method proposed in the present invention has excellent performance. The research provides a more advanced solution for automatic modulation recognition, which is a significant improvement.

[0091] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made in accordance with the spirit of the present invention are intended to be covered by the scope of protection of the present invention.

[0092] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A wireless signal automatic modulation recognition method based on CGAF and residual recognition network model, characterized in that: The method specifically comprises the following steps: S1: Obtain training data Select signal samples of I / Q timing signals containing K modulation schemes as the original data set, where the K modulation schemes include digital modulation scheme types and analog modulation scheme types; divide no less than 60% of the signal samples in the original data set into a training set, and the rest into a validation set, and both the training set and the validation set contain signal samples of the above K modulation schemes; S2: Data preprocessing The sample signals in the training set and the validation set are processed respectively, and the I channel signal and Q channel signal in the time domain of each sample signal are extracted respectively, and the corresponding IQ channel signal is obtained, where IQ = I + jQ, where j is a complex unit; then the I channel signal and the Q channel signal in the time domain are fast Fourier transformed respectively to obtain the I channel signal and the Q channel signal in the complex domain respectively; Afterwards, the complex domain Gram angle field is used to perform feature mapping on the complex domain I channel signal, Q channel signal, and IQ channel signal respectively to obtain the mapping images of the corresponding channels. The mapping images of each channel are then connected to the same feature space according to the channel dimension to obtain the global mapping image of the sample signal. After processing each sample signal in the training set and the validation set into the corresponding global mapping image, the training image set and the validation image set are obtained respectively; S3: Building a recognition network model The recognition network model includes a PCN module, an FD-Unit module, a first AE-Unit module, a first Res-Unit module, a second AE-Unit module, a second Res-Unit module, a third AE-Unit module, a third Res-Unit module, a GRU module, an Avgpool layer, an FC layer, and a Softmax layer; The PCN module performs Z normalization on the eigenvalues ​​of each channel of the input image, and then concatenates the results of the Z normalization of the eigenvalues ​​of each channel in channel order to obtain the output of the PCN module; The output of the PCN module is the input of the FD-Unit module; The FD-Unit module includes a channel segmentation operation, three convolution branch networks and a channel splicing operation. The features input to the FD-Unit module first undergo a channel segmentation operation to evenly divide the output features of the PCN module into three intermediate features according to the channel dimension, and then the three intermediate features are respectively input to three parallel convolution branch networks; the outputs obtained by the three parallel convolution branch networks are then subjected to a channel splicing operation to obtain the output of the FD-Unit module; the first convolution branch in the three parallel convolution branch networks is composed of a 3×3 convolution kernel depth-separable convolution module and a 1×1 convolution kernel ordinary convolution module in series, and the former The output of one module is used as the input of the next module, and the output of the last module is the output of the convolution branch. The second convolution branch is composed of a 3×3 convolution kernel ordinary convolution module and a 1×1 convolution kernel depthwise separable convolution module in series. The output of the previous module is used as the input of the next module, and the output of the last module is the output of the convolution branch. The third convolution branch is composed of a 3×3 convolution kernel depthwise separable convolution module and two 1×1 convolution kernel ordinary convolution modules in series, where the 3×3 convolution kernel depthwise separable convolution module is set between the two 1×1 convolution kernel ordinary convolution modules, and the output of the last 1×1 convolution kernel ordinary convolution module is the output of the convolution branch. The first AE-Unit module takes the output of the FD-Unit module as input; the first AE-Unit module includes a channel splitting operation, two convolution branches and a channel splicing operation; the features input to the first AE-Unit module first undergo a channel splitting operation to divide the input features into two intermediate features, and then the two intermediate features are respectively input into two parallel convolution branch networks. The outputs obtained by the two parallel convolution branch networks are then subjected to a channel splicing operation to obtain the output of the AE-Unit module; the first convolution branch in the two parallel convolution branch networks is composed of a 3×3 convolution kernel, a depth-separable convolution module with a stride of 2, a channel attention module, and a 1×1 convolution kernel ordinary convolution module in series. The output of the previous module is used as the input of the next module, and the output of the last module is the output of the convolution branch; the second convolution branch is symmetrical with the first convolution branch, and its setting is exactly the same as the first convolution branch; The first Res-Unit module takes the output of the first AE-Unit module as input. The first Res-Unit module consists of two convolution branches, the main and side, and a channel splicing operation. The features input to the first Res-Unit module are respectively input into the two convolution branches. The results of the two convolution branches are subjected to a channel splicing operation to obtain the output of the first Res-Unit module. The main convolution branch is composed of a 1×1 convolution kernel ordinary convolution module, a 3×3 convolution kernel grouped convolution module, and a 1×1 convolution kernel ordinary convolution module connected in series. The output of the previous module is used as the input of the next module, and the output of the last module is the output of the convolution branch. The side convolution branch consists of only a 3×3 convolution kernel depth-separable convolution module, and its output is the output of the side convolution branch of the first Res-Unit module. After processing by the first Res-Unit module, the feature size remains unchanged, and the feature channel is doubled. The basic structures of the first AE-Unit, the second AE-Unit module, and the third AE-Unit module are the same, and the network parameters are not shared; the basic structures of the first Res-Unit module, the second Res-Unit module, and the third Res-Unit module are the same, and the network parameters are not shared; after each AE-Unit module, the number of feature channels doubles and the size of the feature map is halved; after each Res-Unit module, the number of feature channels doubles and the size of the feature map remains unchanged; The feature map output by the third Res-Unit module undergoes a flattening operation to flatten the spatial dimension of the feature map into one dimension, and this dimension is used as the sequence length, that is, the time step, and the channel dimension is used as the feature dimension. The obtained output is used as the input of the GRU module; The output of the GRU module is processed in turn by the average pooling layer, the fully connected layer, and the Softmax layer to obtain the probability distribution of each modulation type of the signal. The modulation type corresponding to the maximum probability is the modulation type of the signal; S4: Training the recognition network model In the process of training the recognition network model, the trainable parameters of the recognition network model are first initialized; the convolutional layer weights are assigned using the He initialization method; the bias term is initialized to 0; the learning rate is initially set to 0.001, and the weight decay strategy is used to prevent overfitting, and the regularization coefficient is set to e -4 ; The input of the recognition network model training is the global mapping image F of the sample signal CGAF , the corresponding label is the modulation type to which the sample signal belongs, expressed as a one-hot encoding; the recognition network model outputs the predicted probability of each category of the sample signal on the K types of modulation modes; the training loss of a batch is calculated using the cross entropy loss function to measure the prediction value The difference between the true label y is expressed as: Where B represents the number of samples in a batch, y i,c is the actual label of sample i in category c, is the model's predicted probability of sample i for category c; Based on the training loss of a batch, the Adam optimizer is used for parameter optimization to update the trainable parameters of the recognition network model once, completing the training of a batch of sample signals. The network parameters at the completion of the training of the previous batch of sample signals are used as the initial parameters for the training of the next batch, and this process is repeated until the last batch of sample signals in the training image set is trained, completing a round of training. After each round of training, the sample signals in the validation image set are input into the current network model, and the loss value of the network model in this round on the validation set is calculated; the network parameters at the completion of the previous round of training are used as the initial parameters for the next round of training, and the training process is repeated. When the early stopping mechanism conditions set according to the loss value on the validation set are met or the training round reaches the preset maximum value, the training of the recognition network model is completed; the set of network parameters with the smallest loss value on the validation set is selected and saved to obtain the optimal recognition network model; S5: Automatic modulation recognition Obtain a wireless signal to be identified. First, convert the wireless signal into an I / Q timing signal form through IQ modulation. Then, according to the data preprocessing method in S2, obtain a global mapping image of the wireless signal. Then, input the global mapping image into the optimal recognition network model in S4 to obtain the probability distribution of the K-type modulation mode of the wireless signal. The modulation mode type corresponding to the maximum probability is the modulation mode type of the wireless signal.

2. The method for automatic modulation identification of wireless signals based on CGAF and residual recognition network model according to claim 1, characterized in that: In S1, the K modulation schemes include 8 digital modulation schemes and 3 analog modulation schemes.

3. The method for automatic modulation identification of wireless signals based on CGAF and residual recognition network model according to claim 2, characterized in that: The 8 digital modulation schemes are 8PSK, BPSK, CPFSK, GFSK, 4-PAM, 16-QAM, 64-QAM, and QPSK; the 3 analog modulation schemes are AM-DSB, AM-SSB, and WBFM.

4. The method for automatic modulation identification of wireless signals based on CGAF and residual recognition network model according to claim 1, characterized in that: In S1, the original dataset is 80% of the signal samples in the public dataset RML2016.10a.

5. The method for automatic modulation identification of wireless signals based on CGAF and residual recognition network model according to claim 1, characterized in that: In S2, the formula for converting the time domain signal into the complex domain signal using fast Fourier transform is as follows: Where X[k] represents the result after FFT transformation, T[n] represents the extracted I channel signal or Q channel signal in the time domain; N is the number of sampling points of the signal.

6. The method for automatic modulation identification of wireless signals based on CGAF and residual recognition network model according to claim 1, characterized in that: In S2, taking the IQ channel signal as an example, the process of using the complex domain Gram angle field to perform feature mapping to obtain the mapping image of the corresponding channel is explained: Assume that the IQ channel signal of a sample signal is X=[X1,X2,…,X n ,…,X N-1 ], element X n is a complex point, which can be expressed in terms of magnitude and phase: where |X n | is the modulus of the element point, is the phase of the element point, j is a complex unit, θ n For element X n The phase angle of ; first you need to scale the complex signal by the following formula: Then, based on the time domain mapping and the inner product formula in the complex domain, the following new mapping method in the complex domain is adopted: CGAF[i,j]=sin(arg[(X′ j ·X′ j ) * ]) X′ i , X′ j is the element X in the complex signal X i 、X j The corresponding eigenvalue obtained after scaling using the above formula, X′ i ·X′ j Indicates the calculation of the inner product of the two, (·) * Indicates that the inner product result is in complex conjugate form, arg[] is the function of taking the phase angle, and CGAF[i,j] represents the element X in the IQ channel signal X of the sample signal. i 、X j The pixel value of the pixel coordinate [i, j] of the mapped CGAF image; map each element in the IQ channel signal X to obtain the mapped image IQ of the IQ channel signal X CGAF .

7. The method for automatic modulation identification of wireless signals based on CGAF and residual recognition network model according to claim 1, characterized in that: In S3, the specific working principle of the PCN module is as follows: Assume that the image samples x in a given batch i , its shape is C×H×W, that is, it contains C channels, height is H, width is W, and for each channel there is the following normalization formula: where x i,c represents the eigenvalue of the cth channel of the i-th sample, μ i,c is the mean of the channel eigenvalues, σ i,c is the standard deviation of the channel eigenvalue, x i,c,h,w Represents the pixel value of the i-th sample in the batch at the c-th channel, h-th row, and w-th column position; in addition, in order to avoid division by 0, when σ i,c = 0, then let σ i,c =1; At the same time, after preprocessing the samples of each channel, they are combined according to the original channel order to obtain the output of the PCN module M is the number of samples in a batch, C is the image sample x i The number of channels in .

8. The method for automatic modulation identification of wireless signals based on CGAF and residual recognition network model according to claim 1, characterized in that: In S3, the GRU module implements temporal dynamic modeling through a gating mechanism. The process can be divided into four steps: First, the input feature contains the input vector x of the current time step. t and the hidden state h at the previous time step t-1 , where t represents the input time step, and the two are used as initial inputs; then processed in parallel by the dual-gating branch: in the update gate branch, x t and h t-1 After the weight matrix W Z 、U Z Linear transformation and bias b z Superposition, and then through the Sigmoid() function σ activation to generate the updated gating signal z t =σ(W Z x t +U Z h t-1 +b z ), used to control the retention ratio of historical information; in the symmetric reset gate branch, the independent parameter W r 、U r and b r Generate reset gate signal r t =σ(W r x t +U r h t-1 +b r ), determines the degree of forgetting of the historical state; then, the candidate hidden state Generated by reset gate control: linear transformation W h x t and the reset history state U h (r t ⊙h t-1 ) and add bias b h , activated by Tanh Where W h and U h Process x separately t and h t-1 The linear matrix, ⊙ represents element-by-element multiplication, and the formula is expressed as A=[a ij ],B=[b ij ],A⊙B=[a ij b ij ]; Finally, update the goalkeeper's historical status h t-1 With candidate status Fusion by weight, output current hidden state Complete the iterative update of the timing status.

9. The method for automatic modulation identification of wireless signals based on CGAF and residual recognition network model according to claim 1, characterized in that: In S4, the early stopping mechanism is to automatically terminate the training when the validation set loss does not decrease for 10 consecutive rounds.

Images