A modulation mode identification method and system of a communication signal with interference

Abstract

The application discloses a modulation mode identification method and system of a communication signal with interference, and belongs to the technical field of communication signal modulation mode identification under deep learning. The application solves the problem of low accuracy of existing methods in identifying modulation modes under the condition that communication signals are interfered. The application uses the most suitable network structure to extract features from IQ signals, amplitude-phase AP signals and signal information in high-order cumulants, and then uses the fusion results of the three features as the input of a subsequent model to identify the signal modulation mode. Even in the presence of interference signals, the application can effectively identify the signal modulation type, and can solve the problem of low identification accuracy of various commonly used neural networks under the influence of signal interference. The application can be applied to modulation mode identification of communication signals with interference.

Description

Technical Field

[0001] This invention belongs to the field of communication signal modulation mode recognition technology under deep learning, specifically relating to a method and system for identifying the modulation mode of communication signals with interference. Background Technology

[0002] Identifying the modulation scheme of communication signals is a common problem in non-cooperative communication environments, where a third party attempts to access the communication system of two communicating parties without authorization. In this process, the third party must not interfere with the normal communication between the two parties, and must receive, estimate, and demodulate the transmitted signal without any prior information. Signal modulation scheme identification is a crucial step in the demodulation process of unknown signals, determining whether subsequent work can proceed smoothly. Traditional signal modulation scheme identification methods mainly employ maximum likelihood hypothesis testing based on decision theory and pattern recognition methods based on feature extraction. The drawbacks of these two methods are limited generalization ability and high computational complexity. With the development of deep learning technology, it has gradually been applied to signal modulation scheme identification. However, current research on deep learning-based signal modulation scheme identification mainly focuses on improving the identification performance of network models under low signal-to-noise ratio conditions. The datasets used in these studies do not consider the influence of interference signals during simulation. Under datasets containing interference, the identification performance of these networks is significantly reduced, failing to effectively identify the modulation scheme of communication signals.

[0003] In summary, existing methods have low accuracy in identifying modulation patterns when communication signals are subject to interference. Therefore, how to identify the modulation patterns of signals affected by interference is a pressing problem to be solved in the field of deep learning-based communication signal modulation pattern identification. Summary of the Invention

[0004] This invention addresses the problem of low accuracy in modulation pattern identification using existing methods when communication signals are subject to interference. It proposes a method and system for identifying modulation patterns in communication signals with interference.

[0005] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0006] According to one aspect of the present invention, a method for identifying the modulation scheme of interfering communication signals is provided, the method specifically comprising the following steps:

[0007] Step 1: Perform amplitude-phase conversion on the interfering IQ communication signal to obtain the amplitude-phase AP signal, which includes an amplitude signal and a phase signal;

[0008] Step 2: Denoise and extract initial features from the interfering IQ communication signal, and denoise and extract initial features from the amplitude-phase AP signal;

[0009] Step 3: Process the initial characteristics of the IQ communication signal to obtain the final characteristics of the IQ communication signal;

[0010] Step 4: Process the initial characteristics of the amplitude-phase AP signal to obtain the final characteristics of the amplitude-phase AP signal;

[0011] Step 5: Calculate the higher-order cumulants of the IQ communication signal, and then extract the features of the IQ communication signal based on the higher-order cumulants;

[0012] Step 6: Fuse the IQ communication signal features extracted in Step 3, the amplitude-phase AP signal features extracted in Step 4, and the IQ communication signal features extracted in Step 5 to obtain fused features;

[0013] Step 7: Pass the fusion features obtained in Step 6 through the channel attention module and the temporal attention module in sequence; then merge all dimensions of the feature map output by the temporal attention module, and the merged result is one-dimensional data.

[0014] The merged result is then passed sequentially through the eighth fully connected layer, the fifteenth ReLU activation function layer, the ninth fully connected layer, the sixteenth ReLU activation function layer, and the tenth fully connected layer;

[0015] The classification result is determined based on the output of the tenth fully connected layer: the output of the tenth fully connected layer is a one-dimensional vector, and the modulation mode corresponding to the index of the point with the largest value in the one-dimensional vector is the recognition result.

[0016] Furthermore, the amplitude-phase conversion of the interfering IQ communication signal to obtain the amplitude-phase AP signal specifically involves:

[0017]

[0018]

[0019] in, Indicating the amplitude signal of the first Data from each point, Indicates the first phase signal Data from each point, Indicates the first signal in the I-channel Data from each point, Indicates the first Q-channel signal Data from each point, This represents the arctangent function.

[0020] Furthermore, the specific process of step two is as follows:

[0021] The interfering IQ communication signal is converted into a matrix form, and then the converted matrix is ​​passed through the first convolutional layer. The output of the first convolutional layer is then used as the input of the second convolutional layer.

[0022] Take the absolute value of each element in the feature map output by the second convolutional layer, and then perform average pooling on the absolute value corresponding to each channel in the feature map to obtain the average value corresponding to each channel.

[0023] The vector composed of the average values ​​of each channel is passed through the first fully connected layer. The output of the first fully connected layer is then used as the input of the first batch normalization layer. Finally, the output of the first batch normalization layer is used as the input of the first ReLU activation function layer.

[0024] The output of the first ReLU activation function layer is used as the input of the second fully connected layer, and then the output of the second fully connected layer is passed through the Sigmoid activation function.

[0025] The output of the Sigmoid activation function is multiplied by a vector composed of the average values ​​of each channel. Each element in the multiplication result represents the threshold of each channel.

[0026] The corresponding channels in the output feature map of the second convolutional layer are subjected to soft thresholding using the channel thresholds to obtain the soft thresholding result of the output feature map of the second convolutional layer.

[0027] The soft thresholding result of the output feature map of the second convolutional layer is then added to the output of the first convolutional layer, and the result is used as the initial feature of the extracted IQ communication signal with interference.

[0028] Furthermore, the specific process of step three is as follows:

[0029] The initial features of the IQ communication signal are used as the inputs to the third convolutional layer and the first causal dilated convolutional layer, respectively, and the output of the first causal dilated convolutional layer is used as the input to the second ReLU activation function layer.

[0030] The output of the second ReLU activation function layer is then used as the input of the second causal dilated convolutional layer. The output of the second causal dilated convolutional layer is added to the output of the third convolutional layer, and the result of the addition is used as the input of the third ReLU activation function layer.

[0031] The output of the third ReLU activation function layer is used as the input of the fourth convolutional layer and the third causal dilation convolutional layer, respectively. Then, the output of the third causal dilation convolutional layer is used as the input of the fourth ReLU activation function layer.

[0032] The output of the fourth ReLU activation function layer is used as the input of the fourth causal dilated convolutional layer. The output of the fourth causal dilated convolutional layer is added to the output of the fourth convolutional layer, and the result of the addition is used as the input of the fifth ReLU activation function layer.

[0033] The output of the fifth ReLU activation function layer is used as the input of the fifth convolutional layer and the fifth causal dilation convolutional layer, respectively. Then, the output of the fifth causal dilation convolutional layer is passed as the input of the sixth ReLU activation function layer.

[0034] The output of the sixth ReLU activation function layer is used as the input of the sixth causal dilated convolutional layer. The output of the sixth causal dilated convolutional layer is added to the output of the fifth convolutional layer, and the result of the addition is used as the input of the seventh ReLU activation function layer.

[0035] The output of the seventh ReLU activation function layer is used as the input of the sixth and seventh convolutional layers, respectively. Then the output of the seventh convolutional layer is passed through the second batch normalization layer and the eighth ReLU activation function layer in sequence.

[0036] The output of the eighth ReLU activation function layer is used as the input of the eighth convolutional layer, and the output of the eighth convolutional layer is used as the input of the third batch normalization layer.

[0037] Then, the output of the sixth convolutional layer is added to the output of the third batch normalization layer, and the result is used as the input of the ninth ReLU activation function layer. The output of the ninth ReLU activation function layer is used as the input of the ninth and tenth convolutional layers, respectively.

[0038] The output of the tenth convolutional layer is passed sequentially through the fourth batch normalization layer and the tenth ReLU activation function layer, and the output of the tenth ReLU activation function layer is used as the input of the eleventh convolutional layer.

[0039] The output of the eleventh convolutional layer is used as the input of the fifth batch normalization layer. The outputs of the ninth convolutional layer and the fifth batch normalization layer are added together, and the result is used as the input of the eleventh ReLU activation function layer. The output of the eleventh ReLU activation function layer is fused in the channel dimension and the width dimension, and the fused result is used as the final feature of the IQ communication signal.

[0040] Furthermore, the specific process of step four is as follows:

[0041] The channel dimension and width dimension of the initial features of the amplitude-phase AP signal are merged, the two dimensions of the merged result are transposed, and the transposed result is passed through the third fully connected layer.

[0042] Then, position encoding is added to the output of the third fully connected layer. The result after adding position encoding is passed through the first encoder, the second encoder, and the third encoder in sequence (the encoders in this invention are all Transformer encoders). The output of the third encoder is used as the input of the fourth fully connected layer.

[0043] The two dimensions of the output of the fourth fully connected layer are transposed, and the transposed result is used as the input of the twelfth convolutional layer. The output of the twelfth convolutional layer is used as the final feature of the amplitude-phase AP signal.

[0044] Furthermore, the specific process of step five is as follows:

[0045] Step 51: Calculate the IQ communication signal First-order mixing moment:

[0046]

[0047] in, Indicates the IQ communication signal at the 1st Data from each point, This represents the cumulative multiplication count of the conjugate operation, * represents the conjugate operation, and E represents calculating the mean. express First-order mixing moment;

[0048]

[0049] In the formula, Represents the imaginary unit;

[0050] Step 52: Calculate the higher-order mixing moments using the method described in Step 51. , , , , , , and Then, eight higher-order cumulants were calculated. , , , , , , and ;

[0051] Step 53, Utilize , , , , , , , , , , , , , , and The real and imaginary parts are combined to form a vector of length 32;

[0052] The combined vector is passed sequentially through the fifth fully connected layer, the sixth batch normalization layer, the twelfth ReLU activation function layer, the sixth fully connected layer, the seventh batch normalization layer, the thirteenth ReLU activation function layer, the seventh fully connected layer, the eighth batch normalization layer, and the fourteenth ReLU activation function layer. Then, the output of the seventh fully connected layer is subjected to dimensional transformation, and the dimensional transformation result is used as the IQ communication signal feature extracted based on higher-order cumulants.

[0053] Furthermore, the specific process of step five two is as follows:

[0054]

[0055]

[0056]

[0057]

[0058]

[0059]

[0060]

[0061] .

[0062] Furthermore, in the vector formed by combining the elements in step five, the first to eighth elements are respectively: , , , , , , , The real part of the , the 9th to the 16th elements are respectively: , , , , , , , The imaginary part, the 17th to the 24th elements are respectively: , , , , , , and The real part of the , the 25th to the 32nd elements are respectively: , , , , , , and The imaginary part.

[0063] According to another aspect of the present invention, a modulation scheme identification system for interfering communication signals is provided. The system includes a signal preprocessing module, a first soft-threshold denoising module, a second soft-threshold denoising module, an IQ signal feature extraction module, an amplitude-phase signal feature extraction module, a higher-order cumulant conversion and feature extraction module, a feature fusion module, a dual attention mechanism unit, and a classification module, wherein:

[0064] The signal preprocessing module is used to perform amplitude-phase conversion on the interfering IQ communication signal to obtain the amplitude-phase AP signal;

[0065] The first soft threshold denoising module is used to process the interfering IQ communication signal and extract the initial features of the IQ communication signal.

[0066] The second soft threshold denoising module is used to process the amplitude-phase AP signal and extract the initial features of the amplitude-phase AP signal;

[0067] The IQ signal feature extraction module is used to process the initial features of the IQ communication signal and extract the final features of the IQ communication signal.

[0068] The amplitude-phase signal feature extraction module is used to process the initial features of the amplitude-phase AP signal and extract the final features of the amplitude-phase AP signal.

[0069] The higher-order cumulant conversion and feature extraction module is used to process the IQ communication signal with interference to obtain IQ communication signal features based on higher-order cumulant extraction.

[0070] The feature fusion module is used to fuse the final features of the IQ communication signal, the final features of the amplitude-phase AP signal, and the features of the IQ communication signal extracted based on the higher-order cumulant to obtain fused features;

[0071] The dual attention mechanism unit is used to process the fused features, passing the fused features sequentially through the channel attention module and the temporal attention module;

[0072] The classification module is used to classify based on the output of the dual attention mechanism unit to obtain the modulation mode recognition result.

[0073] The beneficial effects of this invention are:

[0074] This invention combines the signal information from IQ signals, amplitude-phase AP signals, and higher-order cumulants, using these three signal information expressed in different ways as subsequent inputs. A neural network model is designed to extract useful information to the greatest extent. The designed neural network model is used for targeted feature extraction, which can effectively identify the signal modulation type even in the presence of interference signals.

[0075] Experimental results show that, when the interference signal types include single-tone interference, multi-tone interference, broadband interference, narrowband interference, and time-domain pulse interference, and the signal-to-interference ratio (SIR) is greater than or equal to -8dB, the method of this invention can achieve an accuracy of at least 70% in identifying the modulation mode of the received signal when the SIR is 4dB, and can achieve an accuracy of nearly 80% when the SIR is higher. In contrast, the accuracy of traditional deep learning network models is only about 60% when the SIR is also -8dB. Attached Figure Description

[0076] Figure 1 This is a flowchart of a modulation scheme identification method for interference-laden communication signals according to the present invention;

[0077] Figure 2 This is a flowchart of the soft threshold denoising module;

[0078] In the diagram, the numbers in parentheses for the 2D convolutional layers represent, in order, the number of output channels, the kernel width, and the kernel length; the numbers in the boxes of the fully connected layers represent the length of the output data.

[0079] Figure 3 This is a flowchart of the IQ signal feature extraction module;

[0080] In the diagram, the numbers in parentheses for the 2D convolutional layer represent the number of output channels, kernel width, and kernel length, respectively; ReLU indicates that the calculation is performed using the ReLU nonlinear transformation formula; the numbers in parentheses for the causal dilated convolutional layer represent the number of output channels, kernel width, and kernel length, respectively, and b represents the dilation coefficient. Setting the dilation coefficients of the causal dilated convolutional layer to 1, 2, and 4 can change the receptive field of the convolutional kernel.

[0081] Figure 4 This is a flowchart of the amplitude and phase signal feature extraction module;

[0082] In the diagram, the numbers in the fully connected layer boxes represent the length of the output data; the positional encoding part uses learnable positional encoding; in the Transformer encoder box, h represents the number of attention heads, and d represents the data length of the intermediate layer of the feedforward network; the numbers in parentheses in the one-dimensional convolutional layer mean, in order, the number of output channels and the length of the convolutional kernel.

[0083] Figure 5 This is a flowchart of the high-order cumulant conversion and feature extraction module;

[0084] In the diagram, the numbers in the boxes of fully connected layers represent the length of the output data;

[0085] Figure 6 This is a structural diagram of the neural network (IAC-Net) proposed in this invention;

[0086] In the diagram, the numbers inside the boxes of fully connected layers represent the length of the output data;

[0087] Figure 7 This is a comparison chart of the recognition accuracy of the IAC-Net proposed in this invention and traditional neural networks under a 4dB signal-to-noise ratio environment. Detailed Implementation

[0088] The neural network designed in this invention can simultaneously train or predict by inputting multiple signals at once. Each input signal is processed in the same way within the neural network, and the following process applies to each input signal.

[0089] Specific implementation method one: Combining Figure 1 This embodiment describes a method for identifying the modulation scheme of interfering communication signals. The method specifically includes the following steps:

[0090] Step 1: Perform amplitude-phase conversion on the interfering IQ communication signal to obtain an amplitude-phase (AP) signal, which includes an amplitude signal and a phase signal.

[0091] The IQ signal consists of mutually orthogonal I-channel and Q-channel signals. The storage format of the IQ signal is a two-dimensional floating-point sequence with width n and length m. An effective implementation is n=2, m=128, representing an I-channel signal with 128 sampling points and a Q-channel signal with 128 sampling points; specifically:

[0092]

[0093]

[0094] in, Indicating the amplitude signal of the first Data from each point, Indicates the first phase signal Data from each point, Indicates the first signal in the I-channel Data from each point, Indicates the first Q-channel signal Data from each point;

[0095] Step 2: Denoise and extract initial features from the interfering IQ communication signal, and denoise and extract initial features from the amplitude-phase AP signal;

[0096] Step two can suppress the influence of Gaussian white noise in the signal. This invention processes interfering IQ communication signals and amplitude-phase AP signals in the same way. For interfering IQ communication signals, the IQ communication signals are stored in matrix form; that is, the data of each point in the I-channel signal is stored in the first row of the matrix, and the data of each point in the Q-channel signal is stored in the second row of the matrix. For amplitude-phase AP signals, the data of each point in the A-channel signal is stored in the first row of the matrix, and the data of each point in the P-channel signal is stored in the second row of the matrix. The following explanation uses an interfering IQ communication signal as an example. Figure 2 As shown, specifically:

[0097] The interfering IQ communication signal is converted into a matrix form, and then the converted matrix is ​​passed through the first convolutional layer. The output of the first convolutional layer is then used as the input of the second convolutional layer.

[0098] Take the absolute value of each element in the feature map output by the second convolutional layer, and then perform average pooling on the absolute value corresponding to each channel in the feature map to obtain the average value corresponding to each channel.

[0099] The vector composed of the average values ​​of each channel is passed through the first fully connected layer. The output of the first fully connected layer is then used as the input of the first batch normalization layer. Finally, the output of the first batch normalization layer is used as the input of the first ReLU activation function layer.

[0100] The output of the first ReLU activation function layer is used as the input of the second fully connected layer, and then the output of the second fully connected layer is passed through the Sigmoid activation function.

[0101] The output of the Sigmoid activation function is multiplied by a vector composed of the average values ​​of each channel. Each element in the multiplication result represents the threshold of each channel.

[0102] The corresponding channels in the output feature map of the second convolutional layer are soft-thresholded using the channel threshold to obtain the soft-thresholding result of the output feature map of the second convolutional layer. The weaker features in the convolution result are removed by soft-thresholding to suppress the influence of noise.

[0103] Taking any one channel as an example, specifically:

[0104] The threshold of the current channel is denoted as For any element in the current channel of the output feature map of the second convolutional layer ,right The soft threshold processing method is as follows:

[0105]

[0106] in, Indicates to The results of soft thresholding;

[0107] Similarly, after performing soft thresholding on each element of the current channel, the soft thresholding result of the feature map of the current channel is obtained.

[0108] The soft thresholding result of the output feature map of the second convolutional layer is then added to the output of the first convolutional layer to prevent useful features from being completely eliminated during the thresholding process and to ensure that the useful information contained in the features is not lost. The sum is used as the initial feature of the extracted IQ communication signal with interference.

[0109] Step 3: Process the initial characteristics of the IQ communication signal to obtain the final characteristics of the IQ communication signal;

[0110] like Figure 3 As shown, specifically:

[0111] The initial features of the IQ communication signal are used as the inputs to the third convolutional layer and the first causal dilated convolutional layer, respectively, and the output of the first causal dilated convolutional layer is used as the input to the second ReLU activation function layer.

[0112] The output of the second ReLU activation function layer is then used as the input of the second causal dilated convolutional layer. The output of the second causal dilated convolutional layer is added to the output of the third convolutional layer, and the result of the addition is used as the input of the third ReLU activation function layer.

[0113] The output of the third ReLU activation function layer is used as the input of the fourth convolutional layer and the third causal dilation convolutional layer, respectively. Then, the output of the third causal dilation convolutional layer is used as the input of the fourth ReLU activation function layer.

[0114] The output of the fourth ReLU activation function layer is used as the input of the fourth causal dilated convolutional layer. The output of the fourth causal dilated convolutional layer is added to the output of the fourth convolutional layer, and the result of the addition is used as the input of the fifth ReLU activation function layer.

[0115] The output of the fifth ReLU activation function layer is used as the input of the fifth convolutional layer and the fifth causal dilation convolutional layer, respectively. Then, the output of the fifth causal dilation convolutional layer is passed as the input of the sixth ReLU activation function layer.

[0116] The output of the sixth ReLU activation function layer is used as the input of the sixth causal dilated convolutional layer. The output of the sixth causal dilated convolutional layer is added to the output of the fifth convolutional layer, and the result of the addition is used as the input of the seventh ReLU activation function layer.

[0117] The output of the seventh ReLU activation function layer is used as the input of the sixth and seventh convolutional layers, respectively. Then the output of the seventh convolutional layer is passed through the second batch normalization layer and the eighth ReLU activation function layer in sequence.

[0118] The output of the eighth ReLU activation function layer is used as the input of the eighth convolutional layer, and the output of the eighth convolutional layer is used as the input of the third batch normalization layer.

[0119] Then, the output of the sixth convolutional layer is added to the output of the third batch normalization layer, and the result is used as the input of the ninth ReLU activation function layer. The output of the ninth ReLU activation function layer is used as the input of the ninth and tenth convolutional layers, respectively.

[0120] The output of the tenth convolutional layer is passed sequentially through the fourth batch normalization layer and the tenth ReLU activation function layer, and the output of the tenth ReLU activation function layer is used as the input of the eleventh convolutional layer.

[0121] The output of the eleventh convolutional layer is used as the input of the fifth batch normalization layer. The outputs of the ninth convolutional layer and the fifth batch normalization layer are added together, and the result is used as the input of the eleventh ReLU activation function layer. The output of the eleventh ReLU activation function layer is fused in the channel dimension and the width dimension, and the fused result is used as the final feature of the IQ communication signal.

[0122] Step 4: Process the initial characteristics of the amplitude-phase AP signal to obtain the final characteristics of the amplitude-phase AP signal;

[0123] like Figure 4As shown, the initial features of the AP branch are first transformed into a shape suitable for the Transformer input through feature mapping. Then, preprocessing before entering the Transformer module is completed through positional encoding. Next, feature extraction is performed by the Transformer encoder part, which has 8 attention heads and a feedforward network intermediate layer of length 256. Finally, transposition and splitting are used to change the feature shape to adapt it to the subsequent feature fusion process. Specifically:

[0124] The channel dimension and width dimension of the initial features of the amplitude-phase AP signal are merged, the two dimensions of the merged result are transposed, and the transposed result is passed through the third fully connected layer.

[0125] Then, position encoding is added to the output of the third fully connected layer. The result after adding position encoding is passed through the first encoder, the second encoder, and the third encoder in sequence (the encoders in this invention are all Transformer encoders). The output of the third encoder is used as the input of the fourth fully connected layer.

[0126] The two dimensions of the output of the fourth fully connected layer are transposed, and the transposed result is used as the input of the twelfth convolutional layer. The output of the twelfth convolutional layer is used as the final feature of the amplitude-phase AP signal.

[0127] Step 5: Calculate the higher-order cumulants of the IQ communication signal, and then extract the features of the IQ communication signal based on the higher-order cumulants;

[0128] like Figure 5 As shown, specifically:

[0129] Step 51: Calculate the IQ communication signal First-order mixing moment:

[0130]

[0131] in, Indicates the IQ communication signal at the 1st Data from each point, This represents the cumulative multiplication count of the conjugate operation, * represents the conjugate operation, and E represents calculating the mean. express First-order mixing moment;

[0132]

[0133] In the formula, Represents the imaginary unit;

[0134] Step 52: Calculate the higher-order mixing moments using the method described in Step 51. , , , , , , and Eight higher-order cumulative quantities , , , , , , and These are artificial features commonly used in traditional signal modulation recognition, proven to effectively reflect higher-order information of signals, and suitable for feature extraction. Therefore, eight higher-order cumulants are then calculated. , , , , , , and :

[0135]

[0136]

[0137]

[0138]

[0139]

[0140]

[0141]

[0142]

[0143] Step 53, Utilize , , , , , , , , , , , , , , and The real and imaginary parts are combined to form a vector of length 32;

[0144] In the combined vector, the first to eighth elements are respectively: , , , , , , , The real part of the , the 9th to the 16th elements are respectively: , , , , , , , The imaginary part, the 17th to the 24th elements are respectively: , , , , , , and The real part of the , the 25th to the 32nd elements are respectively: , , , , , , and The imaginary part;

[0145] The combined vector is passed sequentially through the fifth fully connected layer, the sixth batch normalization layer, the twelfth ReLU activation function layer, the sixth fully connected layer, the seventh batch normalization layer, the thirteenth ReLU activation function layer, the seventh fully connected layer, the eighth batch normalization layer, and the fourteenth ReLU activation function layer. The output of the seventh fully connected layer is then subjected to dimensional transformation to adapt it to the subsequent feature fusion. The dimensional transformation result is used as the IQ communication signal feature extracted based on the higher-order cumulant.

[0146] Step 6: Fuse the IQ communication signal features extracted in Step 3, the amplitude-phase AP signal features extracted in Step 4, and the IQ communication signal features extracted in Step 5 (i.e., splice the three features along the channel dimension) to obtain the fused features;

[0147] Step 7: Pass the fused features obtained in Step 6 through the channel attention module and the temporal attention module in sequence. This can further enhance the proportion of useful features in the feature data block and improve the network's recognition performance. Then, merge all dimensions of the feature map output by the temporal attention module. The merged result is one-dimensional data.

[0148] The merged result is then passed sequentially through the eighth fully connected layer, the fifteenth ReLU activation function layer, the ninth fully connected layer, the sixteenth ReLU activation function layer, and the tenth fully connected layer;

[0149] The classification result is determined based on the output of the tenth fully connected layer: the output of the tenth fully connected layer is a one-dimensional vector, the length of which is the total number of signal modulation mode categories, and the modulation mode corresponding to the position index of the point with the largest value in the one-dimensional vector is the recognition result.

[0150] Specific Implementation Method Two: Combining Figure 6 This embodiment describes a modulation scheme identification system for interfering communication signals. The system includes a signal preprocessing module, a first soft-threshold denoising module, a second soft-threshold denoising module, an IQ signal feature extraction module, an amplitude-phase signal feature extraction module, a higher-order cumulant conversion and feature extraction module, a feature fusion module, a dual attention mechanism unit, and a classification module. The entire neural network model of this invention is named IAC-Net (IQ-Amplitude-Cumulant Multi-stream Fusion Network), wherein:

[0151] The signal preprocessing module is used to perform amplitude-phase conversion on the interfering IQ communication signal to obtain the amplitude-phase AP signal;

[0152] The first soft threshold denoising module is used to process the interfering IQ communication signal and extract the initial features of the IQ communication signal.

[0153] The second soft threshold denoising module is used to process the amplitude-phase AP signal and extract the initial features of the amplitude-phase AP signal;

[0154] The IQ signal feature extraction module is used to process the initial features of the IQ communication signal and extract the final features of the IQ communication signal.

[0155] The amplitude-phase signal feature extraction module is used to process the initial features of the amplitude-phase AP signal and extract the final features of the amplitude-phase AP signal.

[0156] The higher-order cumulant conversion and feature extraction module is used to process the IQ communication signal with interference to obtain IQ communication signal features based on higher-order cumulant extraction.

[0157] The feature fusion module is used to fuse the final features of the IQ communication signal, the final features of the amplitude-phase AP signal, and the features of the IQ communication signal extracted based on the higher-order cumulant to obtain fused features;

[0158] The dual attention mechanism unit is used to process the fused features, passing the fused features sequentially through the channel attention module and the temporal attention module;

[0159] The classification module is used to classify based on the output of the dual attention mechanism unit to obtain the modulation mode recognition result.

[0160] Experimental Section

[0161] In the method of this invention, by using three branches to extract the features of the IQ signal to be identified, the features of the amplitude and phase signals, and the feature information based on the higher-order cumulants, the signal features extracted from multiple angles can complement each other, which improves the signal modulation mode recognition capability of the neural network under the influence of interference and ensures that it can perform modulation recognition work normally in the interference environment.

[0162] Figure 7 This is a comparison chart of the recognition performance of this invention with other traditional neural networks under a 4dB signal-to-noise ratio (SNR) environment. The dataset used was generated by Gnuradio software, with an SNR range of -20dB to 20dB and a step size of 4dB. It includes eight digital modulation schemes: 8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, and QPSK, with a sampling rate of 200kHz. The channel environment consists of additive white Gaussian noise, Rician-type small-scale fading, center frequency offset (CFO), and sample rate offset (SRO). Figure 7 As shown, the method of this invention (IAC-Net) is significantly superior to traditional neural networks in almost all cases. It overcomes the problem that pure convolutional networks (such as CNN and ResNet networks in the figure) focus on the response of interference and useful signals in interference environments, while ignoring the high signal-to-interference ratio (SIR) caused by the useful signal itself, resulting in poor recognition performance. Moreover, with an SIR of 4dB, it can achieve an accuracy of 70% at an SIR of -8dB, and an accuracy of nearly 80% at even higher SIRs, thus realizing effective recognition of signal modulation methods in interference environments.

[0163] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for identifying the modulation scheme of interfering communication signals, characterized in that, The method specifically includes the following steps: Step 1: Perform amplitude-phase conversion on the interfering IQ communication signal to obtain the amplitude-phase AP signal, which includes an amplitude signal and a phase signal; Step 2: Denoise and extract initial features from the interfering IQ communication signal, and denoise and extract initial features from the amplitude-phase AP signal; Step 3: Process the initial characteristics of the IQ communication signal to obtain the final characteristics of the IQ communication signal; Step 4: Process the initial characteristics of the amplitude-phase AP signal to obtain the final characteristics of the amplitude-phase AP signal; Step 5: Calculate the higher-order cumulants of the IQ communication signal, and then extract the features of the IQ communication signal based on the higher-order cumulants; Step 6: Fuse the IQ communication signal features extracted in Step 3, the amplitude-phase AP signal features extracted in Step 4, and the IQ communication signal features extracted in Step 5 to obtain fused features; Step 7: Pass the fusion features obtained in Step 6 through the channel attention module and the temporal attention module in sequence; then merge all dimensions of the feature map output by the temporal attention module, and the merged result is one-dimensional data. The merged result is then passed sequentially through the eighth fully connected layer, the fifteenth ReLU activation function layer, the ninth fully connected layer, the sixteenth ReLU activation function layer, and the tenth fully connected layer; The classification result is determined based on the output of the tenth fully connected layer: the output of the tenth fully connected layer is a one-dimensional vector, and the modulation mode corresponding to the index of the point with the largest value in the one-dimensional vector is the recognition result.

2. The method for identifying the modulation scheme of a communication signal with interference according to claim 1, characterized in that, The amplitude-phase conversion of the interfering IQ communication signal to obtain the amplitude-phase AP signal is specifically performed as follows: in, Indicating the amplitude signal of the first Data from each point, Indicates the first phase signal Data from each point, Indicates the first signal in the I-channel Data from each point, Indicates the first Q-channel signal Data from each point.

3. The method for identifying the modulation scheme of a communication signal with interference according to claim 2, characterized in that, The specific process of step two is as follows: The interfering IQ communication signal is converted into a matrix form, and then the converted matrix is ​​passed through the first convolutional layer. The output of the first convolutional layer is then used as the input of the second convolutional layer. Take the absolute value of each element in the feature map output by the second convolutional layer, and then perform average pooling on the absolute value corresponding to each channel in the feature map to obtain the average value corresponding to each channel. The vector composed of the average values ​​of each channel is passed through the first fully connected layer. The output of the first fully connected layer is then used as the input of the first batch normalization layer. Finally, the output of the first batch normalization layer is used as the input of the first ReLU activation function layer. The output of the first ReLU activation function layer is used as the input of the second fully connected layer, and then the output of the second fully connected layer is passed through the Sigmoid activation function. The output of the Sigmoid activation function is multiplied by a vector composed of the average values ​​of each channel. Each element in the multiplication result represents the threshold of each channel. The corresponding channels in the output feature map of the second convolutional layer are subjected to soft thresholding using the channel thresholds to obtain the soft thresholding result of the output feature map of the second convolutional layer. The soft thresholding result of the output feature map of the second convolutional layer is then added to the output of the first convolutional layer, and the result is used as the initial feature of the extracted IQ communication signal with interference.

4. The modulation scheme identification method for interference-laden communication signals according to claim 3, characterized in that, The specific process of step three is as follows: The initial features of the IQ communication signal are used as the inputs to the third convolutional layer and the first causal dilated convolutional layer, respectively, and the output of the first causal dilated convolutional layer is used as the input to the second ReLU activation function layer. The output of the second ReLU activation function layer is then used as the input of the second causal dilated convolutional layer. The output of the second causal dilated convolutional layer is added to the output of the third convolutional layer, and the result of the addition is used as the input of the third ReLU activation function layer. The output of the third ReLU activation function layer is used as the input of the fourth convolutional layer and the third causal dilation convolutional layer, respectively. Then, the output of the third causal dilation convolutional layer is used as the input of the fourth ReLU activation function layer. The output of the fourth ReLU activation function layer is used as the input of the fourth causal dilated convolutional layer. The output of the fourth causal dilated convolutional layer is added to the output of the fourth convolutional layer, and the result of the addition is used as the input of the fifth ReLU activation function layer. The output of the fifth ReLU activation function layer is used as the input of the fifth convolutional layer and the fifth causal dilation convolutional layer, respectively. Then, the output of the fifth causal dilation convolutional layer is passed as the input of the sixth ReLU activation function layer. The output of the sixth ReLU activation function layer is used as the input of the sixth causal dilated convolutional layer. The output of the sixth causal dilated convolutional layer is added to the output of the fifth convolutional layer, and the result of the addition is used as the input of the seventh ReLU activation function layer. The output of the seventh ReLU activation function layer is used as the input of the sixth and seventh convolutional layers, respectively. Then the output of the seventh convolutional layer is passed through the second batch normalization layer and the eighth ReLU activation function layer in sequence. The output of the eighth ReLU activation function layer is used as the input of the eighth convolutional layer, and the output of the eighth convolutional layer is used as the input of the third batch normalization layer. Then, the output of the sixth convolutional layer is added to the output of the third batch normalization layer, and the result is used as the input of the ninth ReLU activation function layer. The output of the ninth ReLU activation function layer is used as the input of the ninth and tenth convolutional layers, respectively. The output of the tenth convolutional layer is passed sequentially through the fourth batch normalization layer and the tenth ReLU activation function layer, and the output of the tenth ReLU activation function layer is used as the input of the eleventh convolutional layer. The output of the eleventh convolutional layer is used as the input of the fifth batch normalization layer. The outputs of the ninth convolutional layer and the fifth batch normalization layer are added together, and the result is used as the input of the eleventh ReLU activation function layer. The output of the eleventh ReLU activation function layer is fused in the channel dimension and the width dimension, and the fused result is used as the final feature of the IQ communication signal.

5. The modulation scheme identification method for interference-laden communication signals according to claim 4, characterized in that, The specific process of step four is as follows: The channel dimension and width dimension of the initial features of the amplitude-phase AP signal are merged, the two dimensions of the merged result are transposed, and the transposed result is passed through the third fully connected layer. Then, position encoding is added to the output of the third fully connected layer. The result after adding position encoding is passed through the first encoder, the second encoder, and the third encoder in sequence (the encoders in this invention are all Transformer encoders). The output of the third encoder is used as the input of the fourth fully connected layer. The two dimensions of the output of the fourth fully connected layer are transposed, and the transposed result is used as the input of the twelfth convolutional layer. The output of the twelfth convolutional layer is used as the final feature of the amplitude-phase AP signal.

6. The modulation scheme identification method for interference-laden communication signals according to claim 5, characterized in that, The specific process of step five is as follows: Step 51: Calculate the IQ communication signal First-order mixing moment: in, Indicates the IQ communication signal at the 1st Data from each point, This represents the cumulative multiplication count of the conjugate operation, * represents the conjugate operation, and E represents calculating the mean. express First-order mixing moment; In the formula, Represents the imaginary unit; Step 52: Calculate the higher-order mixing moments using the method described in Step 51. , , , , , , and Then, eight higher-order cumulants were calculated. , , , , , , and ; Step 53, Utilize , , , , , , , , , , , , , , and The real and imaginary parts are combined to form a vector of length 32; The combined vector is passed sequentially through the fifth fully connected layer, the sixth batch normalization layer, the twelfth ReLU activation function layer, the sixth fully connected layer, the seventh batch normalization layer, the thirteenth ReLU activation function layer, the seventh fully connected layer, the eighth batch normalization layer, and the fourteenth ReLU activation function layer. Then, the output of the seventh fully connected layer is subjected to dimensional transformation, and the dimensional transformation result is used as the IQ communication signal feature extracted based on higher-order cumulants.

7. The modulation scheme identification method for interference-laden communication signals according to claim 6, characterized in that, The specific process of step 52 is as follows: 。 8. The modulation scheme identification method for interference-laden communication signals according to claim 7, characterized in that, In the vector formed by combining the elements in step five, the first to eighth elements are respectively: , , , , , , , The real part of the , the 9th to the 16th elements are respectively: , , , , , , , The imaginary part, the 17th to the 24th elements are respectively: , , , , , , and The real part of the , the 25th to the 32nd elements are respectively: , , , , , , and The imaginary part.

9. A modulation scheme identification system for communication signals with interference, characterized in that, The system includes a signal preprocessing module, a first soft-threshold denoising module, a second soft-threshold denoising module, an IQ signal feature extraction module, an amplitude-phase signal feature extraction module, a higher-order cumulant conversion and feature extraction module, a feature fusion module, a dual attention mechanism unit, and a classification module, wherein: The signal preprocessing module is used to perform amplitude-phase conversion on the interfering IQ communication signal to obtain the amplitude-phase AP signal; The first soft threshold denoising module is used to process the interfering IQ communication signal and extract the initial features of the IQ communication signal. The second soft threshold denoising module is used to process the amplitude-phase AP signal and extract the initial features of the amplitude-phase AP signal; The IQ signal feature extraction module is used to process the initial features of the IQ communication signal and extract the final features of the IQ communication signal. The amplitude-phase signal feature extraction module is used to process the initial features of the amplitude-phase AP signal and extract the final features of the amplitude-phase AP signal. The higher-order cumulant conversion and feature extraction module is used to process the IQ communication signal with interference to obtain IQ communication signal features based on higher-order cumulant extraction. The feature fusion module is used to fuse the final features of the IQ communication signal, the final features of the amplitude-phase AP signal, and the features of the IQ communication signal extracted based on the higher-order cumulant to obtain fused features; The dual attention mechanism unit is used to process the fused features, passing the fused features sequentially through the channel attention module and the temporal attention module; The classification module is used to classify based on the output of the dual attention mechanism unit to obtain the modulation mode recognition result.

Images