Method for automatically identifying signal modulation mode of non-cooperative communication system

Abstract

The invention discloses an automatic identification method for a signal modulation mode of a non-cooperative communication system, and relates to the technical field of information perception and identification. The method comprises the following steps: performing time-frequency analysis on existing modulation signal I / Q channel data to obtain model input data; the processed data are input into a deep supervision architecture based on adaptive auxiliary loss, the network architecture comprises a deep supervision network model, and parameter optimization is carried out through an adaptive loss function, so that common supervision of deep and shallow layer features of the network is realized; meanwhile, the problem of class imbalance in a real data set and the problem of possible noise labels are considered, and a group of optimal model parameters are obtained through optimization of a certain number of iterations; and reasoning the trained model to obtain a modulation mode to which the input signal belongs. According to the method, feature information of different dimensions can be fully learned, the problem of category imbalance in a data set and the possible problem of noise labels are effectively solved, and the model recognition precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of information perception and recognition, and in particular to a method for automatically recognizing a signal modulation mode of a non-cooperative communication system. Background Art

[0002] Automatic identification of communication system signal modulation is one of the core functions of non-cooperative communication reconnaissance systems. With the development of wireless communication technology, the modulation types of signals have become more complex and diverse, and the signals are easily interfered during transmission, resulting in the original features not being obvious. Traditional signal modulation recognition methods generally use multiple feature extraction methods for joint analysis, which need to be adjusted according to different environments. The processing flow is complicated and the accuracy is low. Existing deep learning methods tend to ignore the class imbalance problem in the training samples in the process of modeling a large amount of data, resulting in low recognition accuracy of the few sample categories. Therefore, how to use a large amount of data for modeling and realize high-accuracy automatic recognition of signals with different modulation modes is of great practical significance. Summary of the invention

[0003] In view of this, the present invention proposes a method for automatically identifying the signal modulation mode of a non-cooperative communication system. The method can fully learn feature information of different dimensions, effectively solve the problem of category imbalance in the data set and the problem of possible noise labels, and improve the model recognition accuracy.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0005] A method for automatically identifying a signal modulation mode of a non-cooperative communication system comprises the following steps:

[0006] Step 1, obtaining a training sample set, wherein the training sample set includes a plurality of known IQ signals, and there are a total of C modulation modes of each IQ signal in the training sample set;

[0007] Step 2, preprocessing the IQ signal in the training sample set to obtain a time-frequency diagram corresponding to the IQ signal;

[0008] Step 3, construct a training neural network model, the input of the training neural network model is the time-frequency diagram corresponding to the IQ signal, the output is the probability of the IQ signal corresponding to C modulation modes, the actual modulation mode of the IQ signal itself is the true value label, and the adaptive auxiliary loss function is used to calculate the total loss value. According to the total loss value, the training neural network model is updated by the stochastic gradient descent method and the error back propagation method until the total loss value converges, the current training neural network model is saved, and some of its structures are selected as the recognition neural network model;

[0009] Step 4: pre-process the IQ signal of the modulation mode to be identified, input the time-frequency diagram corresponding to the IQ signal into the recognition neural network model, output the probability values ​​corresponding to C modulation modes respectively, record the modulation mode corresponding to the maximum probability value as the modulation mode of the current IQ signal, and complete the automatic recognition of the modulation mode of the non-cooperative communication system signal.

[0010] Furthermore, in step 1, the modulation modes of the IQ signals include 16qam, 2ask, 2fsk, 2psk, 32qam, 4ask, 4fsk, 4psk, 64qam, 8fsk and 8psk, and C=11.

[0011] Furthermore, the specific method of preprocessing in step 2 and step 4 is:

[0012] The IQ signal is preprocessed by short-time Fourier transform method:

[0013]

[0014] Where x(t) is the IQ signal, w(t-τ) indicates windowing of the signal, τ is the time window size, t is time, ω is the angular frequency, and j is the imaginary unit. S(ω,t) is a two-dimensional matrix, each column of which represents the result sequence after Fourier transform of the obtained segments after windowing the signal at different positions. S(ω,t) is the time-frequency diagram corresponding to the IQ signal.

[0015] Further, the training neural network model in step 3 includes a first convolutional layer, a second convolutional layer, a first maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a second maximum pooling layer, a fifth convolutional layer to a seventh convolutional layer, a third maximum pooling layer, an eighth convolutional layer to a tenth convolutional layer, a fourth maximum pooling layer, an eleventh convolutional layer to a thirteenth convolutional layer, a fifth maximum pooling layer and a first fully connected layer connected in sequence;

[0016] Among them, the structures of the first convolutional layer to the thirteenth convolutional layer are the same, all including one 3×3 convolutional layer, and the output results of the 3×3 convolutional layer are batch normalized and input into the ReLU activation function to obtain the corresponding output results; the windows of the first maximum pooling layer to the fifth maximum pooling layer are all 3×3; the first fully connected layer is a two-layer structure, the first layer includes 512 neurons, and the second layer includes 11 neurons;

[0017] In addition, the training neural network model also includes the fourteenth to eighteenth convolutional layers, the fourteenth convolutional layer has the same structure as the fifteenth convolutional layer, both including two parallel 3×3 convolutional layers, and each 3×3 convolutional layer is connected to a 1×1 convolutional layer, and the outputs of the two 1×1 convolutional layers are dot-producted and input into a 1×1 convolutional layer, and then sequentially connected to an adaptive average pooling layer and a fully connected layer; the output results of the ReLU activation function in the first and second convolutional layers correspond to the inputs of the two parallel 3×3 convolutional layers in the fourteenth convolutional layer; the output results of the ReLU activation function in the third and fourth convolutional layers correspond to the inputs of the two parallel 3×3 convolutional layers in the fifteenth convolutional layer;

[0018] The sixteenth to eighteenth convolutional layers have the same structure, including three parallel 3×3 convolutional layers, and each 3×3 convolutional layer is connected to a 1×1 convolutional layer, and the outputs of the three 1×1 convolutional layers are dot-producted and input into a 1×1 convolutional layer, and then sequentially connected to an adaptive average pooling layer and a fully connected layer; the output results of the ReLU activation function in the fifth to seventh convolutional layers correspond to the inputs of the three parallel 3×3 convolutional layers in the sixteenth convolutional layer; the output results of the Relu activation function in the eighth to tenth convolutional layers correspond to the inputs of the three parallel 3×3 convolutional layers in the seventeenth convolutional layer; the output results of the ReLU activation function in the eleventh to thirteenth convolutional layers correspond to the inputs of the three parallel 3×3 convolutional layers in the eighteenth convolutional layer;

[0019] The fully connected layer structures in the fourteenth to eighteenth convolutional layers are the same, both of which are two-layer structures, the first layer includes 21 neurons, and the second layer includes 11 neurons; the time-frequency map corresponding to the IQ signal is converted into a corresponding feature map through each convolutional layer, and the adaptive average pooling layer in the fourteenth to eighteenth convolutional layers is used to adjust the size of the feature map to 7×7.

[0020] Furthermore, in step 3, the adaptive auxiliary loss function is used to calculate the total loss value. According to the total loss value, the specific method of updating the training neural network model by using the stochastic gradient descent method and the error back propagation method is as follows:

[0021]

[0022] α=1 / (1+e -A );

[0023] β=1 / (1+e -B );

[0024] Among them, for the first fully connected layer and the fully connected layers in the fourteenth to eighteenth convolutional layers, the 11 neurons in the second layer output C feature space vectors and calculate their corresponding adaptive auxiliary loss function values ​​respectively: Specifically, the C feature space vectors are converted into C probability values ​​p(n|S) through the Softmax function, where p(n|S) represents the probability that the current IQ signal corresponds to the nth modulation mode, n=1,2,3,...,C; q(n|S) represents the true value distribution of the current IQ signal, if the nth modulation mode is the true modulation mode of the current IQ signal itself, then q(n|S)=1, otherwise q(n|S)=0; k n is the number of IQ signal samples belonging to the nth modulation mode in the training sample set; A and B are hyperparameters; the adaptive auxiliary loss function values ​​l corresponding to the first fully connected layer and the fully connected layers from the 14th to the 18th convolutional layers are SAAL The sum is calculated to get the total loss value. According to the total loss value, the parameters and hyperparameters A and B in the training neural network model are updated using the stochastic gradient descent method and the error back propagation method.

[0025] Furthermore, the specific method of selecting some of the structures as the identification neural network model in step 3 is:

[0026] For the training neural network model saved after the training, the first convolution layer, the second convolution layer, the first maximum pooling layer, the third convolution layer, the fourth convolution layer, the second maximum pooling layer, the fifth convolution layer to the seventh convolution layer, the third maximum pooling layer, the eighth convolution layer to the tenth convolution layer, the fourth maximum pooling layer, the eleventh convolution layer to the thirteenth convolution layer, the fifth maximum pooling layer and the first fully connected layer connected in sequence are selected as the recognition neural network model;

[0027] The input of the recognition neural network model is the time-frequency diagram corresponding to the IQ signal, and the output is C feature space vectors, which are converted into C probability values ​​through the Softmax function.

[0028] Due to the adoption of the above technical solution, the beneficial effects of the present invention compared with the prior art are:

[0029] 1. The present invention adopts a deep learning method to construct a specific model, which can fully model a large amount of data, and is an end-to-end approach that can realize automatic identification of signal modulation methods.

[0030] 2. The present invention performs time-frequency analysis on the original signal and, combined with the network architecture of DSN-SAAL, can fully learn feature information of different dimensions, effectively solve the problem of category imbalance in the data set and the problem of possible noise labels, and improve the model recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is an overall flow chart of a method for automatically identifying a signal modulation mode in a non-cooperative communication system according to an embodiment of the present invention.

[0032] Figure 2 Schematic diagram of the DSN-SAAL network architecture in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0034] A method for automatically identifying a signal modulation mode in a non-cooperative communication system, such as Figure 1 As shown, the following steps are included:

[0035] Step 1, obtaining a training sample set, wherein the training sample set includes a plurality of known IQ signals, and there are a total of C modulation modes of each IQ signal in the training sample set;

[0036] Step 2, preprocessing the IQ signal in the training sample set to obtain a time-frequency diagram corresponding to the IQ signal;

[0037] Step 3, construct a training neural network model, the input of the training neural network model is the time-frequency diagram corresponding to the IQ signal, the output is the probability of the IQ signal corresponding to C modulation modes, the actual modulation mode of the IQ signal itself is the true value label, and the adaptive auxiliary loss function is used to calculate the total loss value. According to the total loss value, the training neural network model is updated by the stochastic gradient descent method and the error back propagation method until the total loss value converges, the current training neural network model is saved, and some of its structures are selected as the recognition neural network model;

[0038] Step 4: pre-process the IQ signal of the modulation mode to be identified, input the time-frequency diagram corresponding to the IQ signal into the recognition neural network model, output the probability values ​​corresponding to C modulation modes respectively, record the modulation mode corresponding to the maximum probability value as the modulation mode of the current IQ signal, and complete the automatic recognition of the modulation mode of the non-cooperative communication system signal.

[0039] Furthermore, in step 1, the modulation modes of the IQ signals include 16qam, 2ask, 2fsk, 2psk, 32qam, 4ask, 4fsk, 4psk, 64qam, 8fsk and 8psk, and C=11.

[0040] Specifically, in this embodiment, a software radio platform (GNU Radio) is used to collect 242,000 IQ data of 11 modulation signals at a signal-to-noise ratio of 5dB-15dB as original input signals, specifically including 16qam, 2ask, 2fsk, 2psk, 32qam, 4ask, 4fsk, 4psk, 64qam, 8fsk, and 8psk.

[0041] Furthermore, the specific method of preprocessing in step 2 and step 4 is:

[0042] The IQ signal is preprocessed by short-time Fourier transform method:

[0043]

[0044] Among them, x(t) is the IQ signal, w(t-τ) represents the windowing of the signal, τ is the time window size, t represents time, ω represents the angular frequency, and j is the imaginary unit; S(ω,t) is a two-dimensional matrix, each column represents the result sequence after the signal is windowed at different positions and the obtained segments are Fourier transformed; S(ω,t) is the time-frequency diagram corresponding to the IQ signal. By performing time-frequency analysis using the short-time Fourier transform method, we can explore the differences in the time-frequency characteristics of signals under different modulation methods.

[0045] Specifically, in this embodiment, the IQ signal is 1×2×1024, where 2 represents the IQ channel, 1024 is the signal sampling length, the short-time Fourier transform sets the window type to a rectangular window, and the window length is 128, and the obtained 1×128×17 data is input into the training neural network model;

[0046] Furthermore, if Figure 2 As shown, the training neural network model in step 3 includes the first convolution layer, the second convolution layer, the first maximum pooling layer, the third convolution layer, the fourth convolution layer, the second maximum pooling layer, the fifth convolution layer to the seventh convolution layer, the third maximum pooling layer, the eighth convolution layer to the tenth convolution layer, the fourth maximum pooling layer, the eleventh convolution layer to the thirteenth convolution layer, the fifth maximum pooling layer and the first fully connected layer connected in sequence;

[0047] Among them, the structures of the first convolutional layer to the thirteenth convolutional layer are the same, all including one 3×3 convolutional layer, and the output results of the 3×3 convolutional layer are batch normalized and input into the ReLU activation function to obtain the corresponding output results; the windows of the first maximum pooling layer to the fifth maximum pooling layer are all 3×3; the first fully connected layer is a two-layer structure, the first layer includes 512 neurons, and the second layer includes 11 neurons;

[0048] In addition, the training neural network model also includes the fourteenth to eighteenth convolutional layers, the fourteenth convolutional layer has the same structure as the fifteenth convolutional layer, both including two parallel 3×3 convolutional layers, and each 3×3 convolutional layer is connected to a 1×1 convolutional layer, and the outputs of the two 1×1 convolutional layers are dot-producted and input into a 1×1 convolutional layer, and then sequentially connected to an adaptive average pooling layer and a fully connected layer; the output results of the ReLU activation function in the first and second convolutional layers correspond to the inputs of the two parallel 3×3 convolutional layers in the fourteenth convolutional layer; the output results of the ReLU activation function in the third and fourth convolutional layers correspond to the inputs of the two parallel 3×3 convolutional layers in the fifteenth convolutional layer;

[0049] The sixteenth to eighteenth convolutional layers have the same structure, including three parallel 3×3 convolutional layers, and each 3×3 convolutional layer is connected to a 1×1 convolutional layer, and the outputs of the three 1×1 convolutional layers are dot-producted and input into a 1×1 convolutional layer, and then sequentially connected to an adaptive average pooling layer and a fully connected layer; the output results of the ReLU activation function in the fifth to seventh convolutional layers correspond to the inputs of the three parallel 3×3 convolutional layers in the sixteenth convolutional layer; the output results of the Relu activation function in the eighth to tenth convolutional layers correspond to the inputs of the three parallel 3×3 convolutional layers in the seventeenth convolutional layer; the output results of the ReLU activation function in the eleventh to thirteenth convolutional layers correspond to the inputs of the three parallel 3×3 convolutional layers in the eighteenth convolutional layer;

[0050] The fully connected layer structures in the fourteenth to eighteenth convolutional layers are the same, both of which are two-layer structures, the first layer includes 21 neurons, and the second layer includes 11 neurons; the time-frequency map corresponding to the IQ signal is converted into a corresponding feature map through each convolutional layer, and the adaptive average pooling layer in the fourteenth to eighteenth convolutional layers is used to adjust the size of the feature map to 7×7.

[0051] Specifically, the training neural network model in this embodiment is a deep supervised network with a self-adaptive auxiliary loss (DSN-SAAL), and parameter optimization is performed through an adaptive loss function; for the training of the DSN-SAAL architecture, a stochastic gradient descent with a momentum of 0.9 and a weight decay of 5e-4 is used as the optimizer, and the learning rate is initialized to 0.001, and decays to 10% of the original every 40 generations (a total of 50 generations of training);

[0052] Furthermore, in step 3, the adaptive auxiliary loss function is used to calculate the total loss value. According to the total loss value, the specific method of updating the training neural network model by using the stochastic gradient descent method and the error back propagation method is as follows:

[0053]

[0054] α=1 / (1+e -A );

[0055] β=1 / (1+e -B );

[0056] Among them, for the first fully connected layer and the fully connected layers in the fourteenth to eighteenth convolutional layers, the 11 neurons in the second layer output C feature space vectors and calculate their corresponding adaptive auxiliary loss function values ​​respectively: Specifically, the C feature space vectors are converted into C probability values ​​p(n|S) through the Softmax function, where p(n|S) represents the probability that the current IQ signal corresponds to the nth modulation mode, n=1,2,3,...,C; q(n|S) represents the true value distribution of the current IQ signal, if the nth modulation mode is the true modulation mode of the current IQ signal itself, then q(n|S)=1, otherwise q(n|S)=0; k n is the number of IQ signal samples belonging to the nth modulation mode in the training sample set; A and B are hyperparameters; the adaptive auxiliary loss function values ​​l corresponding to the first fully connected layer and the fully connected layers from the 14th to the 18th convolutional layers are SAAL The sum is calculated to get the total loss value. According to the total loss value, the parameters and hyperparameters A and B in the training neural network model are updated using the stochastic gradient descent method and the error back propagation method.

[0057] Furthermore, the specific method of selecting some of the structures as the identification neural network model in step 3 is:

[0058] For the training neural network model saved after the training, the first convolution layer, the second convolution layer, the first maximum pooling layer, the third convolution layer, the fourth convolution layer, the second maximum pooling layer, the fifth convolution layer to the seventh convolution layer, the third maximum pooling layer, the eighth convolution layer to the tenth convolution layer, the fourth maximum pooling layer, the eleventh convolution layer to the thirteenth convolution layer, the fifth maximum pooling layer and the first fully connected layer connected in sequence are selected as the recognition neural network model;

[0059] The input of the recognition neural network model is the time-frequency diagram corresponding to the IQ signal, and the output is C feature space vectors, which are converted into C probability values ​​through the Softmax function.

[0060] Specifically, in this embodiment, precision (PRE) is used as the evaluation standard, and the detailed scoring algorithm is as follows:

[0061]

[0062] Among them, TP is a true sample, FP is a false sample, and the class value is obtained by the above formula. The specific experimental details are as follows:

[0063] For the obtained recognition neural network model, reasoning was performed on an additional set of 242,000 test data. The corresponding PRE mean and variance results were obtained after 10 experiments as shown in Table 1:

[0064] Table 1 Experimental results

[0065] Model PRE STFT+DSN-SAAL 97.36±0.02

[0066] It can be seen from the experimental results that the accuracy of this embodiment has reached more than 97%. The combination of the short-time Fourier transform time-frequency analysis method and the DSN-SAAL model architecture has good generalization ability and can effectively identify the modulation mode of unknown signals.

[0067] In summary, the present invention can fully learn feature information of different dimensions, effectively solve the problem of category imbalance in the data set and the problem of possible noise labels, and improve the model recognition accuracy.

[0068] Those skilled in the art will appreciate that the embodiments described are intended to help readers understand the principles of the present invention, and should be understood that the scope of protection of the present invention is not limited to the embodiments described. For those skilled in the art, the present invention may have various changes and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for automatically identifying a signal modulation mode in a non-cooperative communication system, characterized in that: The following steps are involved: Step 1, obtaining a training sample set, wherein the training sample set includes a plurality of known IQ signals, and there are a total of C modulation modes of each IQ signal in the training sample set; Step 2, preprocessing the IQ signal in the training sample set to obtain a time-frequency diagram corresponding to the IQ signal; Step 3, construct a training neural network model, the input of the training neural network model is the time-frequency diagram corresponding to the IQ signal, the output is the probability of the IQ signal corresponding to C modulation modes, the actual modulation mode of the IQ signal itself is the true value label, and the adaptive auxiliary loss function is used to calculate the total loss value. According to the total loss value, the training neural network model is updated by the stochastic gradient descent method and the error back propagation method until the total loss value converges, the current training neural network model is saved, and some of its structures are selected as the recognition neural network model; Step 4: pre-process the IQ signal of the modulation mode to be identified, input the time-frequency diagram corresponding to the IQ signal into the recognition neural network model, output the probability values ​​corresponding to C modulation modes respectively, record the modulation mode corresponding to the maximum probability value as the modulation mode of the current IQ signal, and complete the automatic recognition of the modulation mode of the non-cooperative communication system signal.

2. The method for automatically identifying a signal modulation mode of a non-cooperative communication system according to claim 1, characterized in that: In step 1, the modulation modes of the IQ signals include 16qam, 2ask, 2fsk, 2psk, 32qam, 4ask, 4fsk, 4psk, 64qam, 8fsk and 8psk, and C=11.

3. The method for automatically identifying a signal modulation mode of a non-cooperative communication system according to claim 2, characterized in that: The specific method of preprocessing in step 2 and step 4 is: The IQ signal is preprocessed by short-time Fourier transform method: Where x(t) is the IQ signal, w(t-τ) indicates windowing of the signal, τ is the time window size, t is time, ω is the angular frequency, and j is the imaginary unit. S(ω,t) is a two-dimensional matrix, each column of which represents the result sequence after Fourier transform of the obtained segments after windowing the signal at different positions. S(ω,t) is the time-frequency diagram corresponding to the IQ signal.

4. The method for automatically identifying a signal modulation mode of a non-cooperative communication system according to claim 3, characterized in that: The training neural network model in step 3 includes a first convolutional layer, a second convolutional layer, a first maximum pooling layer, a third convolutional layer, a fourth convolutional layer, a second maximum pooling layer, a fifth convolutional layer to a seventh convolutional layer, a third maximum pooling layer, an eighth convolutional layer to a tenth convolutional layer, a fourth maximum pooling layer, an eleventh convolutional layer to a thirteenth convolutional layer, a fifth maximum pooling layer and a first fully connected layer connected in sequence; Among them, the structures of the first convolutional layer to the thirteenth convolutional layer are the same, all including one 3×3 convolutional layer, and the output results of the 3×3 convolutional layer are batch normalized and input into the ReLU activation function to obtain the corresponding output results; the windows of the first maximum pooling layer to the fifth maximum pooling layer are all 3×3; the first fully connected layer is a two-layer structure, the first layer includes 512 neurons, and the second layer includes 11 neurons; In addition, the training neural network model also includes the fourteenth to eighteenth convolutional layers, the fourteenth convolutional layer has the same structure as the fifteenth convolutional layer, both including two parallel 3×3 convolutional layers, and each 3×3 convolutional layer is connected to a 1×1 convolutional layer, and the outputs of the two 1×1 convolutional layers are dot-producted and input into a 1×1 convolutional layer, and then sequentially connected to an adaptive average pooling layer and a fully connected layer; the output results of the ReLU activation function in the first and second convolutional layers correspond to the inputs of the two parallel 3×3 convolutional layers in the fourteenth convolutional layer; the output results of the ReLU activation function in the third and fourth convolutional layers correspond to the inputs of the two parallel 3×3 convolutional layers in the fifteenth convolutional layer; The sixteenth to eighteenth convolutional layers have the same structure, including three parallel 3×3 convolutional layers, and each 3×3 convolutional layer is connected to a 1×1 convolutional layer, and the outputs of the three 1×1 convolutional layers are dot-producted and input into a 1×1 convolutional layer, and then sequentially connected to an adaptive average pooling layer and a fully connected layer; the output results of the ReLU activation function in the fifth to seventh convolutional layers correspond to the inputs of the three parallel 3×3 convolutional layers in the sixteenth convolutional layer; the output results of the ReLU activation function in the eighth to tenth convolutional layers correspond to the inputs of the three parallel 3×3 convolutional layers in the seventeenth convolutional layer; the output results of the ReLU activation function in the eleventh to thirteenth convolutional layers correspond to the inputs of the three parallel 3×3 convolutional layers in the eighteenth convolutional layer; The fully connected layer structures in the fourteenth to eighteenth convolutional layers are the same, both of which are two-layer structures, the first layer includes 21 neurons, and the second layer includes 11 neurons; the time-frequency map corresponding to the IQ signal is converted into a corresponding feature map through each convolutional layer, and the adaptive average pooling layer in the fourteenth to eighteenth convolutional layers is used to adjust the size of the feature map to 7×7.

5. The method for automatically identifying a signal modulation mode of a non-cooperative communication system according to claim 4, characterized in that: In step 3, the adaptive auxiliary loss function is used to calculate the total loss value. According to the total loss value, the specific method of updating the training neural network model by using the stochastic gradient descent method and the error back propagation method is as follows: α＝1 / (1+e -A )； B=1 / (1+e -B ); Among them, for the first fully connected layer and the fully connected layers in the fourteenth to eighteenth convolutional layers, the 11 neurons in the second layer output C feature space vectors and calculate their corresponding adaptive auxiliary loss function values ​​respectively: Specifically, the C feature space vectors are converted into C probability values ​​p(n|S) through the Softmax function, where p(n|S) represents the probability that the current IQ signal corresponds to the nth modulation mode, n=1,2,3,...,C; q(n|S) represents the true value distribution of the current IQ signal, if the nth modulation mode is the true modulation mode of the current IQ signal itself, then q(n|S)=1, otherwise q(n|S)=0; k n is the number of IQ signal samples belonging to the nth modulation mode in the training sample set; A and B are hyperparameters; the adaptive auxiliary loss function values ​​l corresponding to the first fully connected layer and the fully connected layers from the 14th to the 18th convolutional layers are SAAL The sum is calculated to get the total loss value. According to the total loss value, the parameters and hyperparameters A and B in the training neural network model are updated using the stochastic gradient descent method and the error back propagation method.

6. The method for automatically identifying a signal modulation mode in a non-cooperative communication system according to claim 5, characterized in that: The specific method of selecting some of the structures in step 3 as the identification neural network model is: For the training neural network model saved after the training, the first convolution layer, the second convolution layer, the first maximum pooling layer, the third convolution layer, the fourth convolution layer, the second maximum pooling layer, the fifth convolution layer to the seventh convolution layer, the third maximum pooling layer, the eighth convolution layer to the tenth convolution layer, the fourth maximum pooling layer, the eleventh convolution layer to the thirteenth convolution layer, the fifth maximum pooling layer and the first fully connected layer connected in sequence are selected as the recognition neural network model; The input of the recognition neural network model is the time-frequency diagram corresponding to the IQ signal, and the output is C feature space vectors, which are converted into C probability values ​​through the Softmax function.

Images