Communication modulation identification method, system and device based on complex number cyclic convolutional network, and medium

Abstract

The invention provides a communication modulation identification method, system and device based on a complex number cyclic convolution network and a medium, and belongs to the field of communication, and the method comprises the steps: constructing a basic unit of a network model through a complex number cyclic convolution unit and a complex number cross term, a complex number cross term performs cross term operation on the complex number domain network weight of the complex number cyclic convolution unit and the output time domain feature and complex number domain feature to obtain the real part and imaginary part features of the time domain complex signal; and connecting the four groups of basic units through a multi-scale feature fusion structure, and constructing a complex number cyclic convolutional network model for communication modulation identification. According to the method, the time domain and complex domain characteristics of the communication signals are extracted by utilizing the characteristics of cyclic expansion convolution extraction of the time sequence characteristics and no boundary effect and combining the complex characteristics of the communication signals, so that the classification and identification of the communication signal modulation modes are realized, the problem of the edge effect of the TCN network is solved, the modulation identification accuracy is improved, and the method is suitable for popularization and application. And a good basis is provided for subsequent signal demodulation analysis.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a communication modulation recognition method, system, device and medium based on a complex cyclic convolutional network. Background Art

[0002] With the development of communication technology, various types of communication radiation source signals are superimposed in the same time, space and even frequency domain, and it is necessary to distinguish and analyze them in order to understand the signal system. At the same time, in the field of military security, it is necessary to safely manage the large number of captured non-cooperative communication signals and understand their communication system, laws and intelligence information.

[0003] Modulation recognition technology plays a fundamental and key role in understanding the communication signal system, communication environment and other information. Communication signal modulation recognition has received extensive attention as the basis of space electromagnetic signal processing. In recent years, with the deepening of research, significant progress has been made in modulation recognition technology.

[0004] At present, modulation recognition methods are mainly divided into two categories: model-driven methods and data-driven methods. Among them, model-driven methods are further divided into two categories: likelihood ratio-based methods and feature extraction-based methods. Data-driven methods are the application of deep learning in the field of wireless communications.

[0005] In recent years, the emergence of radio frequency machine learning (RFML) methods that apply deep machine learning to radio frequency (RF) data research has greatly promoted the development of data-driven methods in spatial electromagnetic signal modulation recognition. Temporal Convolutional Networks (TCN) is a variant of Convolutional Neural Networks (CNN), and its main structure includes dilated causal convolutional layers and residual structures. Compared with traditional recurrent neural network (RNN) models, TCN can not only process time series data, but also has the parallelism and flexibility of CNN, and has a strong ability to extract time domain features of communication signals.

[0006] The commonly used deep learning (DL) modulation recognition methods are studied on the basis of RNN and CNN. However, the long short-term memory (LSTM) network, as a typical network structure in RNN, is prone to gradient disappearance after too many training times, and the CNN network lacks the ability to extract the time domain features of the signal. In order to utilize the powerful feature extraction capability of CNN, some scholars proposed that TCN use dilated convolution to extract time domain features, but the convolution kernel of the TCN network is completed by zero padding when extracting edge features, which leads to the generation of edge effects. At the same time, in the existing modulation recognition research, the communication modulation signal is split into real and imaginary parts for processing, which fails to reflect the signal amplitude and phase characteristics well, and the network structure is constructed as a real network according to the signal processing form, which fails to make full use of the complex domain characteristics of the signal, resulting in low communication signal modulation recognition accuracy. Summary of the invention

[0007] Aiming at the problem of edge effect caused by zero padding to complete the extraction of edge features during TCN network training, and the problem of how to improve the accuracy of communication signal modulation recognition, the present invention proposes to use the characteristics of cyclic dilated convolution to extract signal timing features without boundary effects, and discloses a communication modulation recognition method, system, device and medium based on a complex cyclic convolutional network in combination with the complex characteristics of communication signals. Through this scheme, time domain and complex domain features of communication signals can be extracted, and the feature distribution differences of signals with different modulation modes can be fully explored to realize the classification and recognition of communication signal modulation modes, thereby solving the problem of edge effect of TCN network and improving the accuracy of modulation recognition, providing a good foundation for subsequent signal demodulation analysis.

[0008] The purpose of the present invention is specifically achieved through the following technical solutions:

[0009] The present invention discloses a communication modulation recognition method based on a complex cyclic convolutional network, comprising:

[0010] A basic unit of the network model is constructed by a complex cyclic convolution unit Ccycle-conv and a complex cross term Ccross-term, wherein the complex cross term performs a cross term operation on the complex domain network weight of the complex cyclic convolution unit and the output time domain features and complex domain features to obtain the real part features and imaginary part features of the time domain complex signal;

[0011] The four groups of basic units are connected through a multi-scale feature fusion structure to construct a complex cyclic convolutional network model CCCN for communication modulation recognition.

[0012] Furthermore, the complex circular convolution unit Ccycle-conv is composed of a complex circular convolution real part Icycle-conv and a complex circular convolution imaginary part Qcycle-conv; the complex circular convolution real part Icycle-conv and the complex circular convolution imaginary part Qcycle-conv are both circular convolutions with no hidden layer, a dilation rate of 1, and a convolution kernel size of 3 based on a circular dilated convolution structure;

[0013] The real part of the input signal or characteristic I n-1 ′ Input the real part of the complex circular convolution unit Icycle-conv to obtain the real part feature I of the signal n ; The imaginary part Q of the input signal or characteristic n-1 ′ Input the complex circular convolution unit's complex circular convolution imaginary part Qcycle-conv, and get the imaginary part feature Q of the signal n .

[0014] Wherein, the complex cycle convolution unit Ccycle-conv is:

[0015]

[0016] Among them, y t represents the output of the circular dilated convolution in the complex circular convolution unit, k represents the circular convolution kernel from 0 to K-1, K represents the convolution kernel size, represents the convolution kernel of the lth layer, x represents the signal value at a certain moment, and t represents the signal sequence [x1,x2,...,x N ] at time t, and t∈{1,2,...,N}, N represents the signal sequence [x1,x2,...,x N ] length; d l It represents the dilation rate of the l-th layer of cyclic dilated convolution, and l represents the number of layers.

[0017] Wherein, the mapping source of the complex cross-term Ccross-term is:

[0018] s(t)=x(t)*h(t)+u(t)

[0019] =[x I (t)+jx Q (t)]*[h I (t)+jh Q (t)]+[u I (t)+ju Q (t)]

[0020] =[x I (t)*h I (t)-x Q (t)*hQ (t)+u I (t)]+j[x I (t)*h Q (t)+x Q (t)*h I (t)+u Q (t)]

[0021] Where s(t) represents the signal intercepted by the non-cooperative receiver, x(t) represents the time-domain complex modulated signal transmitted by the cooperative transmitter, h(t) represents the transfer function of the wireless channel of the complex linear time-invariant system, u(t) represents the additive Gaussian white noise in the spatial electromagnetic environment where the channel is located, * represents the convolution operation, · I (t) represents the real part of the complex number, Q (t) represents the imaginary part of the complex number, x I (t) represents the real value of the time domain complex modulated signal transmitted by the partner transmitter, x Q (t) represents the imaginary part of the time domain complex modulated signal transmitted by the partner transmitter, j represents the imaginary unit, and h I (t) represents the real part of the transfer function of the wireless channel of the complex linear time-invariant system, h Q (t) represents the imaginary part of the transfer function of the complex linear time-invariant system wireless channel, u I (t) represents the real value of the additive white Gaussian noise in the electromagnetic environment of the channel, u Q (t) represents the imaginary part of the additive white Gaussian noise in the spatial electromagnetic environment where the channel is located.

[0022] The cross term calculation method is as follows:

[0023] I n ′=I n w r-n -Q n w i-n

[0024] Q n ′=I n w i-n +Q n w r-n

[0025] Among them, I n ′, Q n ′ represents the output data of the nth complex cross term, that is, the real and imaginary characteristics of the output time domain complex signal, I n , Q n represents the input data of the nth complex cross term, that is, the real time feature and imaginary time feature output by the complex circular convolution unit, n∈{1,2,3,4} represents the sequence number; wr-n represents the network weight of the complex circular convolution real part Icycle-conv of the nth complex circular convolution unit, w i-n Represents the network weight of the complex circular convolution imaginary part Qcycle-conv of the nth complex circular convolution unit.

[0026] Furthermore, the method of connecting the four groups of basic units through a multi-scale feature fusion structure includes:

[0027] The first feature extracted through one layer of basic units and the second feature extracted through two layers of basic units are fused in the real and imaginary parts through the concatenate function to obtain a first fused feature, and the first fused feature is extracted through one layer of basic units to obtain a fourth feature;

[0028] The third feature extracted by the three-layer basic unit and the fourth feature are fused through the concatenate function to obtain the second fused feature;

[0029] The second fusion feature is fully connected by the Dense layer;

[0030] The Softmax layer is used to realize the classification and recognition of signal modulation methods.

[0031] The present invention also provides a communication modulation recognition system based on a complex cyclic convolutional network, comprising:

[0032] A basic unit construction module is used to construct a basic unit of a network model through a complex cyclic convolution unit Ccycle-conv and a complex cross term Ccross-term, wherein the complex cross term performs a cross term operation on the complex domain network weights of the complex cyclic convolution unit and the output time domain features and complex domain features to obtain real and imaginary features of a time domain complex signal;

[0033] The CCCN building module is used to connect four groups of basic units through a multi-scale feature fusion structure to construct a complex cyclic convolutional network model CCCN for communication modulation recognition.

[0034] The present invention also provides a communication modulation identification device based on a complex cyclic convolutional network, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the communication modulation identification method based on a complex cyclic convolutional network when executing the computer program.

[0035] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the communication modulation identification method based on a complex cyclic convolutional network.

[0036] Beneficial effects:

[0037] 1. The present invention proposes a method, system, device and medium for communication modulation recognition based on a complex cyclic convolutional network, which uses a complex cyclic convolution unit and a complex cross term to extract the time domain and complex domain features of the communication signal, and uses a multi-scale feature fusion structure to deeply mine the signal features, thereby improving the recognition rate of the communication signal modulation mode;

[0038] 2. The complex cyclic convolutional network model CCCN disclosed in the present invention has both the powerful feature extraction capability of CNN and the ability of LSTM to extract time series features, which solves the edge effect caused by zero padding to complete the extraction of edge features during TCN network training;

[0039] 3. Combining the complex characteristics of communication modulation signals, Ccycle-conv and Ccross-term are designed to extract the time domain and complex domain features of the signal, which enhances the network's learning ability for communication signals;

[0040] 4. Utilize the fusion of low-dimensional features and high-dimensional features to fully extract communication signal features and thus improve the accuracy of signal modulation recognition.

[0041] 5. The complex cyclic convolutional network model CCCN disclosed in the present invention fully considers the complex nature of the communication system and fully learns the complex domain and time domain characteristics of the communication modulation signal;

[0042] 6. The scheme proposed in the present invention has theoretical value for complex electromagnetic signal processing methods and has positive reference significance for improving the accuracy of communication modulation signal recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The present invention is further described in detail below based on the accompanying drawings and embodiments.

[0044] Figure 1 It is a structure and principle diagram of a complex circular convolution unit and a complex cross term provided by the present invention;

[0045] Figure 2 It is a schematic diagram of the cyclic dilated convolution provided by the present invention;

[0046] Figure 3 It is a CCCN flow chart provided by the present invention;

[0047] Figure 4 It is a schematic diagram of the comparison curve of network layer recognition accuracy provided by the present invention;

[0048] Figure 5 It is a schematic diagram of the comparison curve of recognition accuracy with and without feature fusion provided by the present invention;

[0049] Figure 6It is a schematic diagram of the comparison curve of the feature fusion structure recognition accuracy provided by the present invention;

[0050] Figure 7 It is a schematic diagram of the comparison curve of the recognition accuracy of the Base model basic network structure provided by the present invention. DETAILED DESCRIPTION

[0051] Embodiment 1 of the present invention provides a communication modulation recognition method based on a complex cyclic convolutional network, comprising:

[0052] The basic unit of the network model is constructed by the complex cycle convolution unit Ccycle-conv and the complex cross term Ccross-term, such as Figure 1 As shown, in the basic unit, the complex cross term performs cross term operations on the complex domain network weights of the complex circular convolution unit and the output time domain features and complex domain features to obtain the real part features and imaginary part features of the time domain complex signal;

[0053] In order to avoid the edge effect in the process of extracting the time domain features of the communication signal and to extract the complex domain features of the signal, a set of Ccycle-conv and Ccross-term is used as the basic unit of the network model.

[0054] In order to fully explore the sample distribution characteristics of communication signals with different modulation modes, four groups of basic units are connected through a multi-scale feature fusion structure to construct a complex cyclic convolutional network model CCCN for communication modulation recognition.

[0055] The complex circular convolution unit Ccycle-conv is composed of a complex circular convolution real part Icycle-conv and a complex circular convolution imaginary part Qcycle-conv; the complex circular convolution real part Icycle-conv and the complex circular convolution imaginary part Qcycle-conv are both based on Figure 2 The circular dilated convolution structure shown is based on a circular convolution with no hidden layer, a dilation rate of 1, and a convolution kernel size of 3;

[0056] Figure 1 The Ccycle-conv-n shown represents the nth complex cycle convolution unit used in CCCN, and Ccross-term-n represents the nth complex cross term used in CCCN; the real part of the input signal or feature I n-1 ′ Input the real part of the complex circular convolution unit Icycle-conv to obtain the real part feature I of the signal n ; The imaginary part Q of the input signal or characteristic n-1 ′ Input the complex circular convolution unit's complex circular convolution imaginary part Qcycle-conv, and get the imaginary part feature Q of the signal n .

[0057] Wherein, the complex cycle convolution unit Ccycle-conv is:

[0058]

[0059] Among them, y t represents the output of the circular dilated convolution in the complex circular convolution unit, k represents the circular convolution kernel from 0 to K-1, K represents the convolution kernel size (convolution kernel size), represents the convolution kernel of the lth layer, x represents the signal value at a certain moment, and t represents the signal sequence [x1,x2,...,x N ] at time t, and t∈{1,2,...,N}, N represents the signal sequence [x1,x2,...,x N ] length; d l It represents the dilation rate of the l-th layer of cyclic dilated convolution, and l represents the number of layers.

[0060] Wherein, the mapping source of the complex cross-term Ccross-term is:

[0061] s(t)=x(t)*h(t)+u(t)

[0062] =[x I (t)+jx Q (t)]*[h I (t)+jh Q (t)]+[u I (t)+ju Q (t)]

[0063] =[x I (t)*h I (t)-x Q (t)*h Q (t)+u I (t)]+j[x I (t)*h Q (t)+x Q (t)*h I (t)+u Q (t)]

[0064] Where s(t) represents the signal intercepted by the non-cooperative receiver, x(t) represents the time-domain complex modulated signal transmitted by the cooperative transmitter, h(t) represents the transfer function of the wireless channel of the complex linear time-invariant system, u(t) represents the additive Gaussian white noise in the spatial electromagnetic environment where the channel is located, * represents the convolution operation, · I (t) represents the real part of the complex number, Q (t) represents the imaginary part of the complex number, x I(t) represents the real value of the time domain complex modulated signal transmitted by the partner transmitter, x Q (t) represents the imaginary part of the time domain complex modulated signal transmitted by the partner transmitter, j represents the imaginary unit, and h I (t) represents the real part of the transfer function of the wireless channel of the complex linear time-invariant system, h Q (t) represents the imaginary part of the transfer function of the complex linear time-invariant system wireless channel, u I (t) represents the real value of the additive white Gaussian noise in the electromagnetic environment of the channel, u Q (t) represents the imaginary part of the additive white Gaussian noise in the spatial electromagnetic environment where the channel is located.

[0065] The nth complex cross-term Ccross-term-n performs the cross-term operation shown in the following formula on the network weights and output features of the nth complex cyclic convolution unit Ccycle-conv-n to obtain the real part and the imaginary part of the complex signal output by the quasi-complex communication system, wherein the cross-term operation method is:

[0066] I n ′=I n w r-n -Q n w i-n

[0067] Q n ′=I n w i-n +Q n w r-n

[0068] Among them, I n ′, Q n ′ represents the output data of the nth complex cross term, that is, the real and imaginary characteristics of the output time domain complex signal, I n , Q n represents the input data of the nth complex cross term, that is, the time domain real part feature and time domain imaginary part feature output by the complex circular convolution unit, n∈{1,2,3,4} represents the sequence number of the complex cross term and the complex circular convolution unit; w r-n represents the network weight of the complex circular convolution real part Icycle-conv of the nth complex circular convolution unit, w i-n Represents the network weight of the complex circular convolution imaginary part Qcycle-conv of the nth complex circular convolution unit.

[0069] Furthermore, the method of connecting the four groups of basic units through a multi-scale feature fusion structure includes:

[0070] The first feature extracted through one layer of basic units and the second feature extracted through two layers of basic units are fused in the real and imaginary parts through the concatenate function to obtain a first fused feature, and the first fused feature is extracted through one layer of basic units to obtain a fourth feature;

[0071] The third feature extracted by the three-layer basic unit is directly fused with the fourth feature through the concatenate function to obtain the second fused feature;

[0072] The second fusion feature is fully connected by the Dense layer;

[0073] The Softmax layer is used to realize the classification and recognition of signal modulation methods.

[0074] like Figure 3 As shown in the figure, a group of Ccycle-conv and Ccross-term is a basic unit of a network model. CCCN fuses the lower dimensional features I1′ and Q1′ extracted by one layer of basic units with the higher dimensional features I2′ and Q2′ extracted by two layers of basic units in the real and imaginary parts through concatenate functions to obtain the first fused features I′ and Q′, and directly fuses the third features I3′ and Q3′ extracted by three layers of basic units with the first fused features I′ and Q′ through the fourth features I4′ and Q4′ extracted by the basic units through concatenate functions to obtain the second fused features, and then fully connects the second fused features by the Dense layer, and finally uses the Softmax layer to realize the classification and recognition of signal modulation methods. The network structure parameter settings of CCCN are shown in Table 1.

[0075] Table 1

[0076]

[0077] The present invention also provides a communication modulation recognition system based on a complex cyclic convolutional network, comprising:

[0078] A basic unit construction module is used to construct a basic unit of a network model through a complex cyclic convolution unit Ccycle-conv and a complex cross term Ccross-term, wherein the complex cross term performs a cross term operation on the complex domain network weights of the complex cyclic convolution unit and the output time domain features and complex domain features to obtain real and imaginary features of a time domain complex signal;

[0079] The CCCN building module is used to connect four groups of basic units through a multi-scale feature fusion structure to construct a complex cyclic convolutional network model CCCN for communication modulation recognition.

[0080] The present invention also provides a communication modulation identification device based on a complex cyclic convolutional network, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the communication modulation identification method based on a complex cyclic convolutional network when executing the computer program.

[0081] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the communication modulation identification method based on a complex cyclic convolutional network.

[0082] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0083] In order to facilitate those skilled in the art to understand the present invention, specific examples are now provided to illustrate the technical solutions of the present invention in detail.

[0084] Take the modulation recognition basic data set shown in Table 2a and the recognition network training set and test set shown in Table 2b as examples:

[0085] Table 2a

[0086]

[0087] After the data set in Table 2a is denoised by F-CNRN, modulation recognition is performed on the data set in Table 2b:

[0088] Table 2b

[0089]

[0090] Based on the Meta Multi-Task Network (MMTN) framework, rapid experimental verification is carried out on small sample datasets.

[0091] Example 1 Communication modulation recognition method based on complex cyclic convolutional network

[0092] The influence of the number of convolution layers on the network performance in the complex cyclic convolutional network model CCCN is verified. The parameters of the meta-multitask framework structure are kept unchanged. The modulation recognition accuracy is calculated when the number of Ccycle-conv and Ccross-term layers is 1, 2, 3, 4, and 5. The effectiveness of the CCCN network layer setting is illustrated by comparing the recognition rates under different numbers of layers. Figure 4 And as shown in Table 3.

[0093] Table 3

[0094]

[0095]

[0096] pass Figure 4 It can be seen from the recognition rate curve that when the signal-to-noise ratio is -6 to 18dB, the recognition accuracy is significantly lower than that of other cases when the network is one layer. Combined with Table 3, it can be seen that the one-layer network does not fully extract the features of small sample data, so the recognition accuracy is the lowest at -20 to -2dB, 0 to 18dB and -20 to 18dB. As can be seen from Table 3, the recognition accuracy is the highest when the number of network layers is 3, and the overall recognition accuracy reaches 76.18%. In the case of 1 to 3 layers, the recognition rate increases with the increase of the number of layers. After the number of layers is greater than 3, although the recognition rate of the 5-layer network is 0.52% higher than that of the 4-layer network at -20 to -2dB, it decreases by 0.73% at 0 to 18dB, and decreases by 0.1% overall. The recognition rate of the 3 to 5-layer network shows an overall downward trend, indicating that as the number of network layers increases, the extraction of small sample data features appears to be over-learned. Therefore, when the number of Ccycle-conv and Ccross-term layers is 3, the learning effect of small sample communication modulation signal data is the best.

[0097] Example 2 Communication modulation recognition method based on complex cyclic convolutional network

[0098] The influence of the multi-scale feature fusion structure on the network performance in the complex recurrent convolutional network model CCCN is verified. The parameters of the meta-multi-task framework structure are kept unchanged, and the modulation recognition accuracy with and without the multi-scale feature fusion structure is calculated respectively. The network without multi-scale feature fusion is a 3-layer network structure; the network with multi-scale feature fusion is also a 3-layer network structure, the difference is that the input of the third layer network is a fusion of the lower dimensional features extracted by the 1-layer network and the higher dimensional features extracted by the 2-layer network. By comparing the recognition rates in the two cases, the effectiveness of the multi-scale feature fusion designed by the CCCN network is illustrated. The recognition rate comparison with and without the feature fusion structure in the 3-layer network is shown as follows. Figure 5 And as shown in Table 4.

[0099] Table 4

[0100]

[0101] Depend on Figure 5 It can be seen from the accuracy curve combined with Table 4 that the recognition accuracy of the multi-scale feature fusion structure is 0.75%, 0.8% and 0.77% higher than that of the non-feature fusion structure in the three cases of SNR -20~-2dB, 0~18dB and -20~18dB, respectively, indicating that the multi-scale feature fusion structure can enhance the expression ability of small sample data features and thus improve the accuracy of small sample modulation recognition.

[0102] The effectiveness of the multi-scale feature fusion structure design in CCCN is verified: the parameters of the meta-multi-task framework structure are kept unchanged, and the modulation recognition accuracy under the three multi-scale feature fusion structures is calculated. The first structure is that the lower dimensional features extracted by the 1-layer network are fused with the higher dimensional features extracted by the 2-layer network and then input into the third-layer network; the second structure is the CCCN structure, the lower dimensional features extracted by the 1-layer network and the higher dimensional features extracted by the 2-layer network are fused and input into the complex circular convolution unit of the basic unit, and the obtained output features are then fused with the high dimensional features extracted by the 3-layer network; the third structure is to perform complex circular convolution on the final fused features of the second structure for feature extraction. The effectiveness of the multi-scale feature fusion structure of the CCCN network is illustrated by comparing the recognition rates of the three structures. The recognition rates of different feature fusion structures are compared as shown below. Figure 6 And as shown in Table 5.

[0103] Table 5

[0104]

[0105] from Figure 6 It can be seen from the recognition rate curve that when the SNR is -20 to -2dB, the recognition accuracy of the multi-scale feature fusion structure of CCCN represented by the red line is significantly higher than that of the other two fusion structures. It can be seen from Table 5 that although the recognition accuracy of CCCN is 82.88% at 0 to 18dB, which is 0.95% lower than fusion1, because the CCCN structure adds the fusion of the features extracted by the 3-layer network compared to the fusion1 structure, and it can be seen from Table 5 that the 3-layer network has a more sufficient learning of the -20 to -2dB signal, CCCN has a higher recognition accuracy at -20 to -2dB, and the Acc reaches 77.35% at -20 to 18dB, which is significantly higher than the other two structures. Fusion3 performs one more convolution operation than CCCN, resulting in excessive learning of small sample data features, which causes the recognition accuracy to decline. The experimental results show that the multi-scale feature fusion structure in CCCN proposed in the example of the present invention can achieve the best learning effect on the features of small sample communication modulation signals.

[0106] Example 3 Communication modulation recognition method based on complex cyclic convolutional network:

[0107] Comparison experiments with other networks:

[0108] To verify the effectiveness of CCCN: Keep the meta-multitask framework structure parameters unchanged, replace CCCN with CNN, LSTM, and CLDNN to achieve the ablation of CCCN, and conduct small sample communication signal modulation recognition experiments under different basic networks used in the Base model to further verify the effectiveness of the complex recurrent convolutional network proposed in this invention. The replacement network structure parameters used are shown in Table 6.

[0109] Table 6

[0110]

[0111]

[0112] Three typical modulation recognition networks are used to replace CCCN and the recognition accuracy is compared. Figure 7 And as shown in Table 7.

[0113] Table 7

[0114]

[0115] from Figure 7 It can be clearly seen from the recognition rate curve that when the basic network is CCCN, the classification and recognition accuracy of small sample communication modulation signals is the highest, and the overall recognition rate reaches 77.35%, which is 7.89%, 12.62% and 17.89% higher than CNN, LSTM and CLDNN respectively. The recognition accuracy is the lowest when the basic network is LSTM. This is because although LSTM can extract the time domain features of communication modulation signals, the feature learning ability of LSTM is not as good as CLDNN and CNN when the signal data is a small sample. When the basic network is CLDNN, the recognition accuracy is higher than LSTM and lower than CNN. This is because CLDNN has several more layers of one-dimensional convolutional units Conv1D than LSTM, which enhances the feature extraction ability of the network, while CLDNN has several more layers of LSTM than CNN, resulting in gradient explosion in the feature extraction of small sample data, which ultimately reduces the recognition rate. The experimental results further illustrate the effectiveness of the basic network CCCN proposed in the present invention.

[0116] Beneficial effects of the embodiments:

[0117] 1. The present invention proposes a method, system, device and medium for communication modulation recognition based on a complex cyclic convolutional network, which uses a complex cyclic convolution unit and a complex cross term to extract the time domain and complex domain features of the communication signal, and uses a multi-scale feature fusion structure to deeply mine the signal features, thereby improving the recognition rate of the communication signal modulation mode;

[0118] 2. The complex cyclic convolutional network model CCCN disclosed in the present invention has both the powerful feature extraction capability of CNN and the ability of LSTM to extract time series features, which solves the edge effect caused by zero padding to complete the extraction of edge features during TCN network training;

[0119] 3. Combining the complex characteristics of communication modulation signals, Ccycle-conv and Ccross-term are designed to extract the time domain and complex domain features of the signal, which enhances the network's learning ability for communication signals;

[0120] 4. Utilize the fusion of low-dimensional features and high-dimensional features to fully extract communication signal features and thus improve the accuracy of signal modulation recognition.

[0121] 5. The complex cyclic convolutional network model CCCN disclosed in the present invention fully considers the complex nature of the communication system and fully learns the complex domain and time domain characteristics of the communication modulation signal;

[0122] 6. The scheme proposed in the present invention has theoretical value for complex electromagnetic signal processing methods and has positive reference significance for improving the accuracy of communication modulation signal recognition.

[0123] 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 communication modulation recognition method based on a complex cyclic convolutional network, characterized in that: include: A basic unit of the network model is constructed by a complex circular convolution unit and a complex cross term, wherein the complex cross term performs a cross term operation on the complex domain network weight of the complex circular convolution unit and the output time domain features and complex domain features to obtain the real part features and imaginary part features of the time domain complex signal; The four groups of basic units are connected through a multi-scale feature fusion structure to construct a complex recurrent convolutional network model for communication modulation recognition.

2. A communication modulation recognition method based on a complex cyclic convolutional network as claimed in claim 1, characterized in that: The complex circular convolution unit is composed of a complex circular convolution real part and a complex circular convolution imaginary part; the complex circular convolution real part and the complex circular convolution imaginary part are circular convolutions without hidden layers, with a dilation rate of 1 and a convolution kernel size of 3 based on a circular dilated convolution structure; The real part of the input signal or feature is input into the real part of the complex circular convolution unit to obtain the real part of the signal; the imaginary part of the input signal or feature is input into the imaginary part of the complex circular convolution unit to obtain the imaginary part of the signal.

3. A communication modulation recognition method based on a complex cyclic convolutional network as claimed in claim 1 or 2, characterized in that: The complex circular convolution unit is: Among them, y t represents the output of the circular dilated convolution in the complex circular convolution unit, k represents the circular convolution kernel from 0 to K-1, K represents the convolution kernel size, represents the convolution kernel of the lth layer, x represents the signal value at a certain moment, and t represents the signal sequence [x1,x2,...,x N ] at time t, and t∈{1,2,...,N}, N represents the signal sequence [x1,x2,...,x N ] length; d l It represents the dilation rate of the l-th layer of cyclic dilated convolution, and l represents the number of layers.

4. A communication modulation recognition method based on a complex cyclic convolutional network as claimed in claim 1, characterized in that: The mapping source of the complex cross term is: s(t)=x(t)*h(t)+u(t) =[x I (t)+jx Q (t)]*[h I (t)+jh Q (t)]+[u I (t)+ju Q (t)] =[x I (t)*h I (t)-x Q (t)*h Q (t)+u I (t)]+j[x I (t)*h Q (t)+x Q (t)*h I (t)+u Q (t)] Where s(t) represents the signal intercepted by the non-cooperative receiver, x(t) represents the time-domain complex modulated signal transmitted by the cooperative transmitter, h(t) represents the transfer function of the wireless channel of the complex linear time-invariant system, u(t) represents the additive Gaussian white noise in the spatial electromagnetic environment where the channel is located, * represents the convolution operation, · I (t) represents the real part of the complex number, Q (t) represents the imaginary part of the complex number, x I (t) represents the real value of the time domain complex modulated signal transmitted by the partner transmitter, x Q (t) represents the imaginary part of the time domain complex modulated signal transmitted by the partner transmitter, j represents the imaginary unit, and h I (t) represents the real part of the transfer function of the wireless channel of the complex linear time-invariant system, h Q (t) represents the imaginary part of the transfer function of the complex linear time-invariant system wireless channel, u I (t) represents the real value of the additive white Gaussian noise in the electromagnetic environment of the channel, u Q (t) represents the imaginary part of the additive white Gaussian noise in the spatial electromagnetic environment where the channel is located.

5. A communication modulation recognition method based on a complex cyclic convolutional network as claimed in claim 1, characterized in that: The method of the cross term operation is: I n ′=I n w r-n -Q n w i-n Q n ′=I n w i-n +Q n w r-n Among them, I n ′, Q n ′ represents the output data of the nth complex cross term, that is, the real and imaginary characteristics of the output time domain complex signal, I n , Q n represents the input data of the nth complex cross term, that is, the real time feature and imaginary time feature output by the complex circular convolution unit, n∈{1,2,3,4} represents the sequence number; w r-n represents the network weight of the complex circular convolution real part Icycle-conv of the nth complex circular convolution unit, w i-n Represents the network weight of the complex circular convolution imaginary part Qcycle-conv of the nth complex circular convolution unit.

6. A communication modulation recognition method based on a complex cyclic convolutional network as claimed in claim 1, characterized in that: The method of connecting four groups of basic units through a multi-scale feature fusion structure includes: The first feature extracted through one layer of basic units and the second feature extracted through two layers of basic units are fused in the real and imaginary parts through the concatenate function to obtain a first fused feature, and the first fused feature is extracted through one layer of basic units to obtain a fourth feature; The third feature extracted by the three-layer basic unit and the fourth feature are fused through the concatenate function to obtain the second fused feature; The second fusion feature is fully connected by the Dense layer; The Softmax layer is used to realize the classification and recognition of signal modulation methods.

7. A communication modulation recognition system based on a complex cyclic convolutional network, characterized in that: include: A basic unit construction module is used to construct a basic unit of a network model through a complex circular convolution unit and a complex cross term, wherein the complex cross term performs a cross term operation on the complex domain network weight of the complex circular convolution unit and the output time domain features and complex domain features to obtain the real part features and imaginary part features of the time domain complex signal; The CCCN building module is used to connect four groups of basic units through a multi-scale feature fusion structure to construct a complex cyclic convolutional network model for communication modulation recognition.

8. A communication modulation recognition device based on a complex cyclic convolutional network, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method according to any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images