A signal modulation recognition method based on complex neural network

Abstract

The present invention relates to the field of radio communication technology, and specifically discloses a signal modulation recognition method based on a complex neural network, comprising: step S110, obtaining a training sample; step S120, dividing the signal in the training sample into a training set and a verification set according to a preset ratio; step S130, building a signal modulation recognition model; step S140, using the training set to train the signal modulation recognition model, and saving the model parameters obtained after each round of training; step S150, using the verification set to verify the signal modulation recognition model applied to each model parameter, and selecting the model parameter with the highest recognition accuracy on the verification set as the optimal model parameter; step S140, inputting a communication signal of a modulation mode to be identified into the signal modulation recognition model applying the optimal model parameter, outputting a predicted category probability, and setting the category label corresponding to the maximum value of the predicted category probability as the final recognized modulation category.

Description

Technical Field

[0001] The present invention relates to the field of radio communication technology, and in particular to a signal modulation recognition method based on complex neural network. Background Art

[0002] In the era of Industry 5.0, automatic modulation recognition in industrial equipment is necessary for intelligence, aiming to help ensure the normal operation of communication systems, information security and effective use of spectrum resources. How to effectively and automatically identify modulation recognition of signals has a wide range of application value in the field of communications. At present, automatic modulation recognition technology based on deep learning has become the focus of research. Although there are relatively rich literatures on the application of deep learning in communications, most of them regard the complex modulated signals of the input model, namely the in-phase component and the orthogonal component, as two independent channels, which ignores the intrinsic connection between them and fails to maintain the amplitude and phase information in the signal, which will reduce the interpretability of the model and lead to weakened generalization performance of the model.

[0003] A complex neural network is an artificial neural network whose input and network parameters are complex numbers. It can directly process the phase and amplitude information of the signal and improve the data expression ability. It is a more suitable model for processing wave-related information than real neural networks. It provides a new solution for processing problems with complex characteristics or calculations in the complex domain.

[0004] At present, research on complex neural networks in signal processing has been carried out, but there is still a lack of complex neural networks applied in the field of signal modulation recognition. In addition, with the introduction of more and more high-precision models, the computing power requirements for recognition equipment are gradually increasing. However, model training takes a long time and the computational complexity is very high. High-precision and large-scale recognition models cannot be well applied in practical scenarios. Especially in the case of limited hardware resources in industrial scenarios, how to enhance the practicality of automatic modulation recognition methods based on complex neural networks deployed on industrial equipment, improve generalization, and reduce model complexity are problems that need to be solved at present. Summary of the invention

[0005] In view of the above problems, the purpose of the present invention is to provide a signal modulation recognition method based on complex neural network, which can enhance the practicality of signal modulation recognition deployed on industrial equipment, improve robustness and reduce model complexity.

[0006] The present invention provides a signal modulation recognition method based on complex neural network, comprising:

[0007] Step S110, obtaining a training sample; the training sample is a communication modulation signal that has been marked with a modulation category label;

[0008] Step S120, using a non-repetitive random sampling method to divide the signals in the training sample into a training set and a validation set according to a preset ratio;

[0009] Step S130, constructing a signal modulation recognition model; the signal modulation recognition model includes a data enhancement module, a complex convolution layer, a complex batch normalization layer, a complex activation function, a complex depth-separable convolution module, a complex channel attention mechanism, a complex global average pooling layer, and a complex fully connected layer;

[0010] Step S140, based on the adaptive moment estimation optimization algorithm and the cross entropy loss function, the signal modulation recognition model is trained using the training set, and the model parameters obtained after each round of training are saved;

[0011] Step S150, using the verification set to verify the signal modulation recognition model applied to each of the model parameters, and selecting the model parameter with the highest recognition accuracy on the verification set as the optimal model parameter;

[0012] Step S160, inputting the communication signal of the modulation mode to be identified into the signal modulation identification model applying the optimal model parameters, outputting the predicted category probability, and setting the category label corresponding to the maximum value of the predicted category probability as the final identified modulation category.

[0013] In a possible implementation, step 120 includes:

[0014] Randomly selecting a plurality of non-repeating integers according to a preset ratio for the signals in the training samples as index values ​​of the training set or the validation set;

[0015] According to the index value, the corresponding training sample is extracted from the training samples and divided into the training set or the validation set according to a preset ratio.

[0016] In a possible implementation, step S130 includes:

[0017] Step S131, inputting the first data into the data enhancement module to obtain the second data; the data format of the first data is N×2×M, and the data format of the second data is 4N×2×M;

[0018] Step S132, inputting the second data into a first complex convolutional layer, a first complex batch normalization layer and a complex activation function to obtain a first feature; the data format of the first feature is 4N×16×32;

[0019] Step S133, inputting the first feature into a second complex convolutional layer and a second complex batch normalization layer to extract features to obtain a second feature; the data format of the second feature is 4N×32×61;

[0020] Step S134, inputting the second feature into the first complex depthwise separable convolution module and the second complex depthwise separable convolution module to extract features to obtain a third feature; the data format of the third feature is 4N×128×16;

[0021] Step S135, inputting the third feature into the complex channel attention mechanism to obtain a weighted fourth feature; the data format of the fourth feature is 4N×128×16;

[0022] Step S136, inputting the fourth feature into a complex global average pooling layer to obtain an average feature of each channel; the data format of the average feature is 4N×128×1;

[0023] Step S137, input the average feature into the complex fully connected layer and output the predicted category probability; the data format of the predicted category probability is N×N 调制方式 ;

[0024] Among them, N 调制方式 is the number of modulation types in the training samples, M is the number of samples on the co-directional or orthogonal components, and N is an integer.

[0025] In a possible implementation, step S131 includes:

[0026] The data enhancement module expands the data by rotation, and the formula is as follows:

[0027]

[0028] Where x represents the first input data. represents the expanded second data, θ represents the rotation angle, R() represents the real part operation, corresponding to the in-phase component, I() represents the imaginary part operation, corresponding to the orthogonal component, and the second data The number of samples is 4 times that of the first data x.

[0029] In a possible implementation, step S132 includes:

[0030] The second data is input into the first complex convolution layer, the first complex batch normalization layer and the complex activation function according to the following formula to obtain the first feature F1:

[0031]

[0032] Among them, CConv1() represents the first complex convolution layer operation, CBN1() represents the first complex batch normalization layer operation, and CReLU() represents the complex activation function operation. Indicates the expanded second data.

[0033] In a possible implementation, step S133 includes:

[0034] The first feature is input into the second complex convolution layer and the second complex batch normalization layer for feature extraction to obtain the second feature F2 according to the following formula:

[0035] F2=CBN2(CConv2(F1))

[0036] Among them, CConv2() represents the second complex convolution layer operation, CBN2() represents the second complex batch normalization operation, and F1 is the first feature.

[0037] In a possible implementation, step S134 includes:

[0038] The second feature is input into the first complex depthwise separable convolution module and the second complex depthwise separable convolution module to extract features to obtain the third feature F3 according to the following formula:

[0039] F3=Block2(Block1(F2))

[0040] Among them, Block1() represents the first complex depth-wise separable convolution module operation, Block2() represents the second complex depth-wise separable convolution module operation, and F2 is the second feature.

[0041] In a possible implementation, step S135 includes:

[0042] The third feature is input into the complex channel attention mechanism according to the following formula to obtain the fourth feature F4 after weighted processing:

[0043] F4=F3×Sigmoid(CConv”(CReLU(CConv’(GAP(R(F3)))+iGAP(I(F3))))))

[0044] Among them, GAP() is a real global average pooling operation, R() and I() represent the operations of obtaining the real part and the imaginary part respectively, i is the imaginary unit, CConv'() is the complex convolution layer one operation, CConv"() is the complex convolution layer two operation, CReLU() represents the complex activation function operation, Sigmoid() is the activation function, and F3 is the third feature.

[0045] In a possible implementation, step S136 includes:

[0046] The fourth feature is input into the complex global average pooling layer according to the following formula to obtain the average feature F5 of each channel:

[0047] F5=CGAP(F4)=GAP(R(F4))+iGAP(I(F4))

[0048] Among them, GAP() is a real global average pooling operation, R() and I() represent the operations of obtaining the real part and the imaginary part respectively, i is the imaginary unit, F4 is the fourth feature, and CGAP() is a complex global average pooling operation;

[0049] The step S137 includes:

[0050] The average feature is input into the complex fully connected layer according to the following formula and the predicted category probability P is output:

[0051] P=CFC(F5)=FC(R(F5))+iFC(I(F5))

[0052] Among them, FC() is a real full connection operation, R() and I() represent the operations of obtaining the real part and the imaginary part respectively; i is the imaginary unit, F5 is the average feature, and CFC() is a complex full connection operation.

[0053] In a possible implementation, step S140 includes:

[0054] Inputting the signal in the training set into the signal modulation recognition model to obtain a predicted probability, and taking the category label corresponding to the predicted probability with the largest value as the predicted category label;

[0055] Inputting the real category label of the signal and the predicted category label into a cross entropy loss function to obtain a loss value, and updating the network parameters of the signal modulation recognition model according to the loss value;

[0056] Using the adaptive moment estimation optimization algorithm to train the signal modulation recognition model, and setting an initial value of the learning rate;

[0057] When the value of the cross entropy loss function does not decrease within a preset number of iterations, multiplying the learning rate by a preset multiple and updating;

[0058] When the learning rate reaches a preset value, the training is stopped, and the training of the signal modulation recognition model through the constructed training set is completed.

[0059] The signal modulation recognition method based on complex neural network provided by the present invention, the complex channel attention mechanism proposed by the present invention is an improvement of the channel attention mechanism module, so that it can realize complex operations to be applicable to complex neural networks. The present invention is suitable for automatic modulation recognition of signals under the condition that stable prior information cannot be obtained, and enhances the practicality of the automatic modulation recognition method based on complex neural network deployed on industrial equipment, and has the characteristics of fewer model parameters, low model complexity and good generalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 A schematic diagram of a flow chart of a signal modulation identification method provided by an embodiment of the present invention;

[0061] Figure 2 A structural diagram of a signal modulation identification model provided by an embodiment of the present invention;

[0062] Figure 3 A schematic diagram of the structure of a complex depthwise separable convolution module provided by an embodiment of the present invention;

[0063] Figure 4 A schematic diagram of the structure of the complex channel attention mechanism provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The following detailed description of the embodiments of the present invention is further described in detail in conjunction with the accompanying drawings and examples. The detailed description of the following embodiments and the accompanying drawings are used to exemplarily illustrate the principles of the present invention, but cannot be used to limit the scope of the present invention, that is, the present invention is not limited to the preferred embodiments described, and the scope of the present invention is defined by the claims.

[0065] In the description of the present invention, it should be noted that, unless otherwise specified, “plurality” means two or more than two; the terms “first”, “second”, etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance; for ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0066] Figure 1 A flow chart of a signal modulation identification method provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the present invention provides a signal modulation recognition method based on complex neural network, comprising:

[0067] Step S110, obtaining training samples;

[0068] In a possible implementation, the training sample is a communication modulation signal that has been labeled with a modulation category label.

[0069] Step S120, using a non-repetitive random sampling method to divide the signals in the training sample into a training set and a validation set according to a preset ratio;

[0070] In a possible implementation, a plurality of non-repeating integers are randomly selected according to a preset ratio from signals in the training samples as index values ​​of the training set or the validation set;

[0071] According to the index value, the corresponding training sample is extracted from the training sample and divided into a training set or a validation set according to a preset ratio.

[0072] In one example, the random.choice function in the numpy library is used in the computer programming language python to randomly select N signals in the training sample according to a preset ratio. 训 / 验 A non-repeated integer is used as the index value of the training set or validation set.

[0073] The calculation formula for the index value of the training set or validation set is as follows:

[0074]

[0075] Among them, N 训 / 验 Indicates the number of samples in the training set or validation set, r 训 Represents the preset ratio of the training set, r 验 Represents the preset proportion of the validation set, N 训练样本 Indicates the number of training samples.

[0076] The number of modulation types in the training samples is N 调制方式 , the data format of each sample is 2×M, where M is the number of samples in the same direction or orthogonal component. Therefore, the sample size of the training set after division is N 训 ×2×M, the sample size of the validation set is N 验 ×2×M.

[0077] Step S130, constructing a signal modulation recognition model;

[0078] In one possible implementation, the signal modulation recognition model includes a data augmentation module, a complex convolution layer, a complex batch normalization layer, a complex activation function, a complex depthwise separable convolution module, a complex channel attention mechanism, a complex global average pooling layer, and a complex fully connected layer; Figure 2 A structural diagram of a signal modulation identification model provided in an embodiment of the present invention.

[0079] In a possible implementation, step S130 includes:

[0080] Step S131, inputting the first data into a data enhancement module to obtain second data; the data format of the first data is N×2×M, and the data format of the second data is 4N×2×M;

[0081] Step S132, inputting the second data into the first complex convolution layer, the first complex batch normalization layer and the complex activation function to obtain the first feature;

[0082] The data format of the first feature is 4N×16×32. The first complex convolution layer has 16 filters and a kernel size of 1×3.

[0083] Step S133, inputting the first feature into the second complex convolution layer and the second complex batch normalization layer to extract the feature to obtain the second feature;

[0084] Among them, the data format of the second feature is 4N×32×61, and the second complex convolution layer has 32 filters and a kernel size of 1×3.

[0085] Step S134, inputting the second feature into the first complex depthwise separable convolution module and the second complex depthwise separable convolution module to extract features to obtain a third feature;

[0086] Among them, the data format of the third feature is 4N×128×16, the first complex depth-wise separable convolution module has 64 filters and a kernel size of 1×3, and the second complex depth-wise separable convolution module has 128 filters and a kernel size of 1×3.

[0087] Step S135, inputting the third feature into the complex channel attention mechanism to obtain a weighted fourth feature;

[0088] Among them, the data format of the fourth feature is 4N×128×16.

[0089] Step S136, inputting the fourth feature into the complex global average pooling layer to obtain the average feature of each channel;

[0090] The data format of the average feature is 4N×128×1.

[0091] Step S137, input the average feature into the complex fully connected layer and output the predicted category probability;

[0092] Among them, the data format of the predicted category probability is N×N 调制方式 , the number of hidden units in the complex fully connected layer is N 调制方式 .

[0093] Among them, N 调制方式 is the number of modulation types in the training samples, M is the number of samples on the co-directional or orthogonal components, and N is an integer.

[0094] In a possible implementation, step S131 includes:

[0095] The data enhancement module expands the data by rotation. The formula is as follows:

[0096]

[0097] Where x represents the first input data. represents the expanded second data, θ represents the rotation angle, R() represents the real part operation, corresponding to the in-phase component, I() represents the imaginary part operation, corresponding to the orthogonal component, and the second data The number of samples is 4 times that of the first data x.

[0098] In one possible implementation, the complex convolution kernel in the complex convolution layer is K c =K r +jK i , for data Perform complex convolution operation CConv() to extract feature F CConv As shown below:

[0099]

[0100] Where * indicates the convolution operation.

[0101] The output of the previous network layer is used as the complex input h and input into the complex batch normalization layer for complex batch normalization operation CBN() to extract the feature F CBN The following formula:

[0102]

[0103] Where γ and β are scaling parameters and translation parameters in the complex domain, respectively. γ is a matrix containing four parameters similar to the covariance matrix. β is a learnable complex parameter. The dimensions are the same, is the normalized representation of the complex input h, which also corresponds to the feature F extracted by the previous complex convolutional layer. cconv .

[0104] The gamma matrix looks like this: The γ in the scaling parameter γ rr and γ ii Initialized to γ ir and γ ri The real and imaginary parts of the complex offset β are initialized to 0.

[0105] Normalization of complex input h The calculation formula is as follows:

[0106]

[0107] in, is the square root of the inverse matrix of the input covariance matrix V, It obeys the standard complex distribution with mean 0, covariance 1, and pseudo-covariance 0. The formula of the covariance matrix V is as follows:

[0108]

[0109] Cov(,) is the covariance operator, R(·) and I(·) represent the operations of obtaining the real and imaginary parts, respectively.

[0110] Step S132 includes: inputting the output of the previous layer of network as the complex input z into the complex activation function layer to perform a complex activation operation CReLU (·) to extract the feature F CReLU , as shown in the following formula:

[0111] F CReLU =CReLU(z)=ReLU(R(z))+iReLU(I(z))

[0112]

[0113] Among them, ReLU(·) is the real activation function ReLU operation, R(·) and I(·) represent the operations of obtaining the real part and the imaginary part respectively, and i is the imaginary unit.

[0114] Therefore, step S132 inputs the second data into the first complex convolution layer, the first complex batch normalization layer and the complex activation function according to the following formula to obtain the first feature F1:

[0115]

[0116] Among them, CConv1() represents the first complex convolution layer operation, CBN1() represents the first complex batch normalization layer operation, and CReLU() represents the complex activation function operation. Indicates the expanded second data.

[0117] In a possible implementation, step S133 includes:

[0118] The first feature is input into the second complex convolution layer and the second complex batch normalization layer for feature extraction to obtain the second feature F2 according to the following formula:

[0119] F2=CBN2(CConv2(F1))

[0120] Among them, CConv2() represents the second complex convolution layer operation, CBN2() represents the second complex batch normalization operation, and F1 is the first feature.

[0121] In one possible implementation, the complex depthwise separable convolution module is a module with a residual connection function consisting of a complex convolution layer, a complex depthwise separable convolution layer, a complex batch normalization layer, and a complex maximum pooling layer, such as Figure 3 As shown, Figure 3 A schematic diagram of the structure of a complex depthwise separable convolution module provided by an embodiment of the present invention. Wherein, Cin represents the input channel of the convolution layer, and its value a is equal to the output channel of the previous layer network; Cout represents the output channel of the convolution layer, and its value b is equal to the number of filters;

[0122] Therefore, step S134 inputs the second feature into the first complex depthwise separable convolution module and the second complex depthwise separable convolution module to extract the feature to obtain the third feature F3 according to the following formula:

[0123] F3=Block2(Block1(F2))

[0124] Among them, Block1() represents the first complex depth-wise separable convolution module operation, Block2() represents the second complex depth-wise separable convolution module operation, and F2 is the second feature.

[0125] In one possible implementation, Figure 4 A schematic diagram of the structure of the complex channel attention mechanism provided by an embodiment of the present invention is shown in FIG. Figure 4 As shown in Figure 1, the complex channel attention mechanism is composed of complex global average pooling, complex convolutional layers, complex activation functions and sigmoid activation functions.

[0126] Step S135 includes: inputting the third feature into the complex channel attention mechanism according to the following formula to obtain a weighted fourth feature F4:

[0127] F4=F3×Sigmoid(CConv”(CReLU(CConv’(GAP(R(F3)))+iGAP(I(F3))))))

[0128] Among them, GAP() is a real global average pooling operation, R() and I() represent the operations of obtaining the real part and the imaginary part respectively, i is the imaginary unit, CConv'() is the complex convolution layer one operation, CConv"() is the complex convolution layer two operation, CReLU() represents the complex activation function operation, Sigmoid() is the activation function, and F3 is the third feature.

[0129] In a possible implementation, step S136 includes:

[0130] The fourth feature is input into the complex global average pooling layer according to the following formula to obtain the average feature F5 of each channel:

[0131] F5=CGAP(F4)=GAP(R(F4))+iGAP(I(F4))

[0132] Among them, GAP() is a real global average pooling operation, R() and I() represent the operations of obtaining the real part and the imaginary part respectively; i is the imaginary unit, F4 is the fourth feature, and CGAP() is a complex global average pooling operation.

[0133] In a possible implementation, step S137 includes:

[0134] The average feature is input into the complex fully connected layer according to the following formula and the predicted category probability P is output:

[0135] P=CFC(F5)=FC(R(F5))+iFC(I(F5))

[0136] Among them, FC() is a real full connection operation, R() and I() represent the operations of obtaining the real part and the imaginary part respectively; i is the imaginary unit, F5 is the average feature, and CFC() is a complex full connection operation.

[0137] Step S140, based on the adaptive moment estimation optimization algorithm and the cross entropy loss function, the signal modulation recognition model is trained using the training set, and the model parameters obtained after each round of training are saved;

[0138] In a possible implementation, step S140 includes:

[0139] The signal in the training set is input into the signal modulation recognition model to obtain the predicted probability, and the category label corresponding to the predicted probability with the largest value is used as the predicted category label; where the predicted probability P 训 The data format is N 训 ×N 调制方式 .

[0140] The real category label and predicted category label of the signal are input into the cross entropy loss function to obtain the loss value, and the network parameters of the signal modulation recognition model are updated according to the loss value;

[0141] Among them, the calculation method of the cross entropy loss function is as follows:

[0142]

[0143] Among them, Loss represents the value of the cross entropy loss function, L 真实k is the true category label of signal k; L 预测k is the predicted category label of signal k, and n is the number of signals.

[0144] The signal modulation recognition model is trained using the adaptive moment estimation optimization algorithm and the initial value of the learning rate is set;

[0145] When the value of the cross entropy loss function does not decrease within the preset number of iterations, the learning rate is multiplied by the preset multiple and updated;

[0146] When the learning rate reaches a preset value, the training is stopped, and the training of the signal modulation recognition model using the constructed training set is completed.

[0147] Among them, the adaptive moment estimation optimization algorithm is implemented by using the adaptive moment estimation optimizer created by the torch.optim.adam function in Python, which is used to dynamically adjust the learning rate of each parameter so that the lightweight modulation recognition model of communication signals based on complex neural networks can obtain better convergence performance.

[0148] Step S150, using the validation set to validate the signal modulation recognition model applied to each model parameter, and selecting the model parameter with the highest recognition accuracy on the validation set as the optimal model parameter;

[0149] Step S160, input the communication signal of the modulation mode to be identified into the signal modulation identification model using the optimal model parameters, output the predicted category probability, and set the category label corresponding to the maximum value of the predicted category probability as the final identified modulation category.

[0150] The following is a description with reference to specific embodiments.

[0151] The experiment was verified on the IQ signal dataset RML2016.10a generated by GNU Radio software platform simulation, as shown in Table 1. Table 1 shows the details of the dataset RML2016.10a. The hardware and software configurations used in the experiment are shown in Table 2; some model training hyperparameter settings of the experiment are shown in Table 3. The initial value of the learning rate is 0.001. When the validation loss does not decrease within 10 iterations, it is multiplied by a coefficient of 0.5; other hyperparameters are default values.

[0152] Table 1 Dataset information

[0153]

[0154] Table 2 Experimental hardware and software settings

[0155]

[0156] Table 3 Model training hyperparameter settings

[0157]

[0158] The accuracy is used in the experiment to measure the performance of the generalization ability of the network model. The recognition accuracy can be calculated as:

[0159]

[0160] In this experiment, the number of training sample modulation types is 11, and the data format of each sample is 2 × 128. Therefore, after dividing the training set according to the ratio of 6:2:2, the sample size is 132000 × 2 × 128, the sample size of the validation set is 44000 × 2 × 128, and the sample size of the test set is 44000 × 2 × 128.

[0161] Then, a lightweight modulation recognition model of communication signals based on complex neural networks is constructed:

[0162] The training samples obtained by division are input into the constructed modulation recognition model, and the predicted probability of the training samples is output. The real category label and predicted probability of the signal are input into the cross entropy loss function, the loss value is calculated, and the network parameters are updated based on the loss value; and the validation samples obtained by division are used to verify the model after each round of training; after 200 rounds of iterative training, the model parameters with the highest recognition accuracy in the validation samples are selected as the final model parameters, and finally the test samples are input into the final model to obtain the model recognition accuracy results.

[0163] In order to verify the generalization and effectiveness of the signal modulation recognition model constructed by the present invention, referred to as the CVDSCNNSE model, the current classic automatic modulation recognition models are selected to provide a benchmark comparison, including MC, MCLDNN, PET-CGDNN and CVDSCNN.

[0164] Table 4 shows the recognition accuracy of the CVDSCNNSE model and other benchmark models on different data sets, as well as the FLOPs and parameter count of the model. Among them, the average accuracy is expressed as the average recognition accuracy of all modulation styles under all signal-to-noise ratios, that is, -20 to 18dB, and the recognition accuracy under the highest signal-to-noise ratio is expressed as the recognition accuracy obtained when the signal-to-noise ratio is 20dB; FLOPs is expressed as the number of floating-point calculations. FLOPs and parameter count are two important indicators for measuring model complexity. Among them, the smaller the value of FLOPs, the less computing resources and time the model requires, and the lower the complexity of the model; the parameter count is the number of parameters that need to be learned in the model. The smaller the parameter count of the model, the less storage and video memory it needs to occupy when training the model.

[0165] Table 4

[0166]

[0167] As shown in Table 4, on the RML2016.10a dataset, compared with other benchmark models, the CVDSCNNSE model has the highest average accuracy of 61.85%, and the model has the lowest number of parameters, which is only 0.03M. Compared with models with the same recognition accuracy, the CVDSCNNSE model has the smallest FLOPs value, which is 190.15M. This shows that the model parameters used in the method proposed in the present invention are small and the model complexity is low. And its good performance on the RML2016.01a dataset shows that the model has good generalization.

[0168] The signal modulation recognition method based on complex neural network provided by the present invention has the following effects:

[0169] (1) The complex channel attention mechanism proposed in the present invention can be better applied to complex neural networks. By calculating the correlation between different channels, it can promote information exchange and integration between channels, recalibrate the feature response of each channel, and improve the feature representation ability. It also has the advantages of fewer parameters and lighter calculation.

[0170] (2) The signal modulation recognition model provided by the present invention has the characteristics of low computational complexity, few model parameters, and good generalization, which is conducive to deployment on computing-constrained devices.

[0171] (3) The present invention can realize automatic extraction of characteristic information of communication signals for modulation mode identification without the need for receiving signal parameters and channel prior information.

[0172] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A signal modulation recognition method based on complex neural network, characterized in that: include: Step S110, obtaining a training sample; the training sample is a communication modulation signal that has been marked with a modulation category label; Step S120, using a non-repetitive random sampling method to divide the signals in the training sample into a training set and a validation set according to a preset ratio; Step S130, constructing a signal modulation recognition model, wherein the signal modulation recognition model includes a data enhancement module, a complex convolution layer, a complex batch normalization layer, a complex activation function, a complex depthwise separable convolution module, a complex channel attention mechanism, a complex global average pooling layer, and a complex fully connected layer; Step S140, based on the adaptive moment estimation optimization algorithm and the cross entropy loss function, the signal modulation recognition model is trained using the training set, and the model parameters obtained after each round of training are saved; Step S150, using the verification set to verify the signal modulation recognition model applied to each of the model parameters, and selecting the model parameter with the highest recognition accuracy on the verification set as the optimal model parameter; Step S160, inputting the communication signal of the modulation mode to be identified into the signal modulation identification model applying the optimal model parameters, outputting the predicted category probability, and taking the category label corresponding to the maximum value of the predicted category probability as the final identified modulation category; The step S130 includes: Step S131, inputting the first data into the data enhancement module to obtain the second data; the data format of the first data is N×2×M, and the data format of the second data is 4N×2×M; Step S132, inputting the second data into a first complex convolutional layer, a first complex batch normalization layer and a complex activation function to obtain a first feature; the data format of the first feature is 4N×16×32; Step S133, inputting the first feature into a second complex convolutional layer and a second complex batch normalization layer to extract features to obtain a second feature; the data format of the second feature is 4N×32×61; Step S134, inputting the second feature into the first complex depthwise separable convolution module and the second complex depthwise separable convolution module to extract features to obtain a third feature; the data format of the third feature is 4N×128×16; Step S135, inputting the third feature into the complex channel attention mechanism to obtain a weighted fourth feature; the data format of the fourth feature is 4N×128×16; Step S136, inputting the fourth feature into a complex global average pooling layer to obtain an average feature of each channel; the data format of the average feature is 4N×128×1; Step S137, input the average feature into the complex fully connected layer and output the predicted category probability; the data format of the predicted category probability is N×N 调制方式 ; Among them, N 调制方式 is the number of modulation types in the training samples, M is the number of samples on the co-directional or orthogonal components, and N is an integer.

2. The signal modulation identification method according to claim 1, characterized in that: The step 120 includes: Randomly selecting a plurality of non-repeating integers according to a preset ratio for the signals in the training samples as index values ​​of the training set or the validation set; According to the index value, the corresponding training sample is extracted from the training samples and divided into the training set or the validation set according to a preset ratio.

3. The signal modulation identification method according to claim 1, characterized in that: The step S131 includes: The data enhancement module expands the data by rotation, and the formula is as follows: Where x represents the first input data. represents the expanded second data, θ represents the rotation angle, R() represents the real part operation, corresponding to the in-phase component, I() represents the imaginary part operation, corresponding to the orthogonal component, and the second data The number of samples is 4 times that of the first data x.

4. The signal modulation identification method according to claim 1, characterized in that: The step S132 includes: The second data is input into the first complex convolution layer, the first complex batch normalization layer and the complex activation function according to the following formula to obtain the first feature F1: Among them, CConv1() represents the first complex convolution layer operation, CBN1() represents the first complex batch normalization layer operation, and CReLU() represents the complex activation function operation. Indicates the expanded second data.

5. The signal modulation identification method according to claim 1, characterized in that: The step S133 includes: The first feature is input into the second complex convolution layer and the second complex batch normalization layer for feature extraction to obtain the second feature F2 according to the following formula: F2=CBN2(CConv2(F1)) Among them, CConv2() represents the second complex convolution layer operation, CBN2() represents the second complex batch normalization operation, and F1 is the first feature.

6. The signal modulation identification method according to claim 1, characterized in that: The step S134 includes: The second feature is input into the first complex depthwise separable convolution module and the second complex depthwise separable convolution module to extract features to obtain the third feature F3 according to the following formula: F3=Block2(Block1(F2)) Among them, Block1() represents the first complex depth-wise separable convolution module operation, Block2() represents the second complex depth-wise separable convolution module operation, and F2 is the second feature.

7. The signal modulation identification method according to claim 1, characterized in that: The step S135 comprises: The third feature is input into the complex channel attention mechanism according to the following formula to obtain the fourth feature F4 after weighted processing: F4=F3×Sigmoid(CConv”(CReLU(CConv'(GAP(R(F3))+iGAP(I(F3))))))wherein, GAP() is a real global average pooling operation, R() and I() represent the operations of obtaining the real part and the imaginary part respectively, i is an imaginary unit, CConv'() is a complex convolution layer one operation, CConv"() is a complex convolution layer two operation, CReLU() represents a complex activation function operation, Sigmoid() is the activation function, and F3 is the third feature.

8. The signal modulation identification method according to claim 1, characterized in that: The step S136 includes: The fourth feature is input into the complex global average pooling layer according to the following formula to obtain the average feature F5 of each channel: F5=CGAP(F4)=GAP(R(F4))+iGAP(I(F4)) Wherein, GAP() is a real global average pooling operation, R() and I() represent operations for obtaining the real part and the imaginary part, respectively; i is an imaginary unit, F4 is the fourth feature, and CGAP() is a complex global average pooling operation; The step S137 includes: The average feature is input into the complex fully connected layer according to the following formula and the predicted category probability P is output: P=CFC(F5)=FC(R(F5))+iFC(I(F5)) Among them, CFC() is a real full connection operation, R() and I() represent the operations of obtaining the real part and the imaginary part respectively; i is the imaginary unit, F5 is the average feature, and CGAP() is a complex full connection operation.

9. The signal modulation identification method according to claim 1, characterized in that: The step S140 includes: Inputting the signal in the training set into the signal modulation recognition model to obtain a predicted probability, and taking the category label corresponding to the predicted probability with the largest value as the predicted category label; Inputting the real category label of the signal and the predicted category label into a cross entropy loss function to obtain a loss value, and updating the network parameters of the signal modulation recognition model according to the loss value; Using the adaptive moment estimation optimization algorithm to train the signal modulation recognition model, and setting an initial value of the learning rate; When the value of the cross entropy loss function does not decrease within a preset number of iterations, the learning rate is multiplied by a preset multiple and updated; when the learning rate reaches a preset value, the training is stopped, and the training of the signal modulation recognition model through the constructed training set is completed.

Images