Modulation identification method and system for complex value convolution fusion type multi-stream space-time network

Abstract

The invention relates to the field of wireless communication, and particularly discloses a modulation identification method and system for a complex value convolution fusion type multi-stream space-time network. Comprising the following steps: dividing an obtained IQ sample signal into an I channel signal, a Q channel signal and two IQ channel signals, respectively carrying out spatial feature extraction through convolution to obtain spatial features of IQ data, adding multiplicative Gaussian noise between LSTM and Bi-LSTM, inputting the spatial features of the IQ signals into the multiplicative Gaussian noise to carry out time feature extraction to obtain a time sequence feature output result, and outputting the time sequence feature output result. And outputting the time sequence characteristic output result through full connection, and adding multiplicative Gaussian noise to obtain an identification result. The objective of the invention is to fully mine complex features, improve precision, eliminate overfitting of a model and increase robustness of the model.

Description

Technical Field

[0001] The present invention relates to the field of wireless communications, and in particular to a modulation recognition method and system for a complex-valued convolution fusion type multi-stream spatiotemporal network. Background Art

[0002] In the field of wireless communications, signal analysis and processing technology mainly involves a series of complex processes such as signal detection, identification and recovery under the condition that the channel and signal parameters are not completely known. As a key link in this series of processes, automatic modulation classification is not only crucial for signal demodulation and parameter extraction, but also plays a guiding role in selecting the correct technical parameters when interfering with communications. Especially in the actual environment of non-cooperative communication, the research on automatic modulation recognition is not only of theoretical importance, but also has significant practical application value.

[0003] In 2016, T.O'Shea et al. used CNN (convolutional neural network) network to directly process IQ signals to verify the potential of deep learning in automatic modulation recognition, and released the RML2016.10a dataset, which stimulated the research boom of DL in the field of automatic modulation recognition. Subsequently, various deep learning network models based on CNN (convolutional neural network), DNN (deep neural network), ResNet (deep residual network) and RNN (recurrent neural network) were widely used in automatic modulation recognition. In 2017, T.O'Shea et al. combined the ability of CNN (convolutional neural network) to process frequency features and the advantage of RNN (recurrent neural network) to capture timing information to create the CLDNN model, which successfully extracted key features and improved classification accuracy. Xu et al. innovatively extracted IQ signal features from the time domain and spatial domain, and used CNN (convolutional neural network) and LSTM (long short-term memory network) fusion network for processing, verifying that the network converges quickly and can accurately distinguish easily confused modulation modes such as 16QAM and 64QAM. In the same year, Xu et al. proposed a MCLDNN network based on spatiotemporal multi-channels. The network combines multi-channel input, spatial feature mapping, temporal feature extraction, and fully connected classification layers to enhance the model's understanding of complex signals and further improve classification accuracy. Soon after, Cui et al. proposed the IQCLNet network to further optimize the number, size, and depth of convolution kernels to more finely extract IQ timing and phase features, verifying the superiority of the network. The IQGMCL network proposed by Hu et al. improved the IQCLNet network by adjusting the size of the convolution kernel and adding Gaussian noise to the fully connected layer, while improving the accuracy and eliminating the overfitting of the model. J. Krzyston et al. introduced the concept of complex convolution and realized a process similar to two-dimensional real convolution through the constructed numerical convolution module. This method only requires a small modification of network parameters to improve recognition accuracy, especially under low signal-to-noise ratio conditions. It can be seen that the automatic modulation classification (automatic modulation recognition) technology based on deep learning has gone through a development process from preliminary signal processing to complex multi-network models. This progress solves the problems of traditional methods relying on manual experience and large computational complexity.

[0004] Although the above algorithms have achieved certain results in the field of automatic modulation classification, problems such as average network recognition rate and model overfitting still need further research. Summary of the invention

[0005] In view of the above problems, an object of the present invention is to provide a modulation recognition method of a complex-valued convolutional fusion multi-stream spatiotemporal network.

[0006] The second object of the present invention is to provide a modulation recognition system of a complex-valued convolution fusion type multi-stream spatiotemporal network.

[0007] The first technical solution adopted by the present invention is:

[0008] S1, dividing the acquired IQ sample signal into an I channel signal, a Q channel signal and two IQ channel signals, respectively extracting spatial features through different convolution methods to obtain spatial features of IQ data;

[0009] S2, adding multiplicative Gaussian noise between LSTM and Bi-LSTM, and inputting the spatial features of the IQ signal into them to extract the time features, so as to obtain the time series feature output result;

[0010] S3. The time series feature output result is fully connected and multiplicative Gaussian noise is added to obtain a recognition result.

[0011] Optionally, the S1 includes:

[0012] S11: dividing the IQ sample signal into the I channel signal, the Q channel signal, a first IQ channel signal and a second IQ channel signal;

[0013] S12 causes the I channel signal to pass through a plurality of layers of convolution kernels to perform convolution operations to obtain a first spatial feature;

[0014] S13 causes the Q channel signal to pass through a plurality of layers of convolution kernels to perform convolution operation to obtain a second spatial feature;

[0015] S14: extracting features of the first IQ channel signal in a three-step convolution manner to obtain a third spatial feature;

[0016] S15: extracting features of the second IQ channel signal by complex convolution to obtain a fourth spatial feature;

[0017] S16 combines and filters the first spatial feature, the second spatial feature, the third spatial feature and the fourth spatial feature to obtain the spatial feature of the IQ data.

[0018] Optionally, the S15 includes:

[0019] S151, performing a zero padding operation on the second IQ channel signal;

[0020] S152, performing a convolution operation on the zero-padded complex data and the complex convolution kernel to obtain a two-dimensional matrix;

[0021] S153, reconstructing and arranging the two-dimensional matrix to obtain a reconstructed matrix;

[0022] S154. Perform a linear transformation based on the reconstruction matrix to obtain the fourth spatial feature.

[0023] Optionally, S2 includes:

[0024] S21. Processing the spatial features of the IQ data by using LSTM to obtain a first time series feature;

[0025] S22, the first time series feature is input into Bi-LSTM through multiplicative Gaussian noise;

[0026] S23. Use Bi-LSTM to perform data processing on the first time series feature after multiplicative Gaussian noise to obtain the time series feature output result.

[0027] Optionally, the S3 includes:

[0028] S31, outputting the time series feature output result through full connection to obtain a full connection output result;

[0029] S32. The fully connected output result is subjected to multiplicative Gaussian noise to obtain a recognition result.

[0030] The second technical solution adopted by the present invention is:

[0031] A spatial feature extraction module is used to divide the acquired IQ sample signal into an I channel signal, a Q channel signal and two IQ channel signals, and perform spatial feature extraction respectively through different convolution methods to obtain the spatial features of the IQ data;

[0032] A time feature extraction module adds multiplicative Gaussian noise between LSTM and Bi-LSTM, and inputs the spatial features of the IQ signal into it to extract the time features, thereby obtaining a time series feature output result;

[0033] The output module outputs the time series feature results through full connection and adds multiplicative Gaussian noise to obtain the recognition result.

[0034] Optionally, the spatial feature extraction module includes:

[0035] A sample division module, dividing the IQ sample signal into the I channel signal, the Q channel signal, a first IQ channel signal and a second IQ channel signal;

[0036] A first convolution module, which performs convolution operation on the I channel signal through a plurality of layers of convolution kernels to obtain a first spatial feature;

[0037] A second convolution module performs convolution operation on the Q channel signal through a plurality of convolution kernel layers to obtain a second spatial feature;

[0038] A third convolution module performs feature extraction on the first IQ channel signal according to a three-step convolution method to obtain a third spatial feature;

[0039] a fourth convolution module, performing feature extraction on the second IQ channel signal by means of complex convolution to obtain a fourth spatial feature;

[0040] A merging module merges and filters the first spatial feature, the second spatial feature, the third spatial feature and the fourth spatial feature to obtain the spatial feature of the IQ data.

[0041] Optionally, the fourth convolution module includes:

[0042] a zero-padding module, performing a zero-padding operation on the second IQ channel signal;

[0043] The complex convolution module performs convolution operation on the zero-padded complex data and the complex convolution kernel to obtain a two-dimensional matrix;

[0044] A matrix reconstruction module is used to reconstruct and arrange the two-dimensional matrix to obtain a reconstructed matrix;

[0045] A linear transformation module performs a linear transformation based on the reconstruction matrix to obtain the fourth spatial feature.

[0046] Optionally, the time feature extraction module includes:

[0047] A data processing module, using LSTM to process the spatial features of the IQ data to obtain a first time series feature;

[0048] In the first noise module, the first time series feature is input into the Bi-LSTM through multiplicative Gaussian noise;

[0049] The time series feature output module uses Bi-LSTM to process the first time series feature after multiplicative Gaussian noise to obtain the time series feature output result.

[0050] Optionally, the output module includes:

[0051] A fully connected output module, which outputs the time series feature output result through full connection to obtain a fully connected output result;

[0052] In the second noise module, the fully connected output result is subjected to multiplicative Gaussian noise to obtain a recognition result.

[0053] Beneficial effects of the above technical solution:

[0054] The present invention introduces a complex convolution module as a branch of the IQ data stream input, which retains the phase information of the signal and the intrinsic correlation between the real part and the imaginary part to the maximum extent, and fully exploits the complex features. The network adopts a space-time dual module combination. The spatial feature extraction module contains multiple parallel processing streams to fully exploit the IQ signal characteristics from multiple dimensions. The combination of space-time modules helps to capture the evolution characteristics of the signal in time series. Multiplicative Gaussian noise is added between LSTM and Bi-LSTM, and between the time feature extraction module and the fully connected output module to eliminate overfitting of the model while improving the accuracy and increase the robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A flowchart of a modulation recognition method for a complex-valued convolution fusion type multi-stream spatiotemporal network provided by the present invention;

[0056] Figure 2 A schematic diagram of the structure of a modulation recognition system of a complex-valued convolution fusion type multi-stream spatiotemporal network provided by the present invention;

[0057] Figure 3 Schematic diagram of complex convolution flow of the present invention;

[0058] Figure 4 It is a schematic diagram of the spatial feature module;

[0059] Figure 5 It is a schematic diagram of the time feature extraction module;

[0060] Figure 6 It is a schematic diagram of a complex-valued convolutional fusion multi-stream spatiotemporal network;

[0061] Figure 7 Schematic diagram of the comparison between the network of the present invention and the existing network on the RML 2016.10a dataset;

[0062] Figure 8 This is a schematic diagram showing the accuracy of network training of the present invention;

[0063] Fig. 9 Schematic diagram of network training loss performance 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] Embodiment 1

[0067] An embodiment of the present invention provides a modulation recognition method of a complex-valued convolution fusion multi-stream spatiotemporal network, such as Figure 1 As shown, a modulation recognition method of a complex-valued convolution fusion multi-stream spatiotemporal network is characterized by comprising: S1, dividing the acquired IQ sample signal into an I channel signal, a Q channel signal and two IQ channel signals, respectively extracting spatial features through different convolution methods to obtain the spatial features of the IQ data; S2, adding multiplicative Gaussian noise between LSTM and Bi-LSTM, and inputting the spatial features of the IQ signal into them for time feature extraction to obtain a time series feature output result; S3, outputting the time series feature output result through full connection and adding multiplicative Gaussian noise to obtain a recognition result. It aims to fully tap complex features, improve accuracy, eliminate overfitting of the model, and increase the robustness of the model.

[0068] Since the IQ signal is a complex expression of two in-phase sequences and orthogonal sequences, previous studies often input the real and imaginary parts as separate channels into the isolated information extraction features based on the neural network for this complex data, thereby ignoring the inherent physical correlation between them. The complex convolution module can process IQ data more comprehensively, effectively capture the richer features between IQ signals, further explore their internal relationships, and help the model better understand the complex characteristics of the signal. The idea of ​​complex convolution, that is, two-dimensional real convolution and linear transformation are equivalent to one-dimensional complex convolution, allows the neural network to calculate complex convolution in the operation of the real deep learning paradigm, so that the complex signal still retains the relationship between the real and imaginary parts after the real convolution operation.

[0069] S1, dividing the acquired IQ sample signal into an I channel signal, a Q channel signal and two IQ channel signals, respectively extracting spatial features through different convolution methods to obtain spatial features of IQ data;

[0070] S11: dividing the IQ sample signal into the I channel signal, the Q channel signal, a first IQ channel signal and a second IQ channel signal;

[0071] S12 causes the I channel signal to pass through a plurality of layers of convolution kernels to perform convolution operations to obtain a first spatial feature;

[0072] S13 causes the Q channel signal to pass through a plurality of layers of convolution kernels to perform convolution operation to obtain a second spatial feature;

[0073] S14: extracting features of the first IQ channel signal in a three-step convolution manner to obtain a third spatial feature;

[0074] S15: extracting features of the second IQ channel signal by complex convolution to obtain a fourth spatial feature;

[0075] S151, performing a zero padding operation on the second IQ channel signal;

[0076] S152, performing a convolution operation on the zero-padded complex data and the complex convolution kernel to obtain a two-dimensional matrix;

[0077] Assume that the two-dimensional IQ data is expressed as:

[0078] Z (t) =I (t) cosω c QUR (t) sinω c t

[0079] When input into a deep learning model, IQ data is expressed in the form of a 2×N matrix:

[0080]

[0081] The expression of the complex convolution kernel is introduced as follows:

[0082] V m =T m +jH m ,T m ,

[0083] Its matrix expression is as follows:

[0084]

[0085] For the complex input data Z n With the complex convolution kernel V m Performing convolution operation, we get the following result:

[0086] Z n *V m =(I n +jQ n )*(T m +jH m )

[0087] Z n *V m =(I n*T m -Q n *H m )+j(I n *H m +Q n *T m )

[0088] S153, reconstructing and arranging the two-dimensional matrix to obtain a reconstructed matrix;

[0089] S154. Perform a linear transformation based on the reconstruction matrix to obtain the fourth spatial feature.

[0090] Z n ZeroPadding operation is performed on the top and bottom of the matrix.

[0091]

[0092] With V m Perform convolution operation and obtain the expression of two-dimensional matrix form X as follows:

[0093] X=Z n *V m =[I*TQ*HI*H+Q*T]

[0094] Z n Carry out Z n *V m The XDL expression for convolution as a real number is:

[0095]

[0096] When XDL does It is equivalent to X during linear transformation, realizing the process in which one-dimensional complex convolution is equivalent to two-dimensional real convolution after linear transformation.

[0097] S16 combines and filters the first spatial feature, the second spatial feature, the third spatial feature and the fourth spatial feature to obtain the spatial feature of the IQ data.

[0098] like Figure 4 As shown in the figure, the IQ sample signal size is 128×2, and the signal is divided into I channel, Q channel and IQ channel as the data source for spatial feature extraction. The I channel and Q channel are respectively subjected to 3 layers of small convolution kernels as independent channels and then connected in parallel to further extract features as the first and second branches of the spatial feature extraction module. By referring to the VGG small convolution, smaller convolution kernels are used to increase the network depth. The use of multiple nonlinear layers can effectively improve the performance of the model and the feature extraction capability, and learn more complex feature representations.

[0099] The IQ paths use the three-step convolution method of (2, 1), (1, 3), and (1, 3) to extract features as the third branch of the spatial feature extraction module.

[0100] Extracting the correlation features between IQ signals reduces the dimension of the data on the one hand, and reduces the frequency domain changes of the data in subsequent processing on the other hand;

[0101] The IQ two-way is padded with zero according to (1, 1) and then convolved with the convolution kernel (2, 3) and then linearly transformed to realize the IQ complex convolution as the fourth branch of the spatial feature extraction module.

[0102] like Figure 5 As shown in the figure, the introduction of the complex convolution module can process IQ data more comprehensively, effectively capture richer features between IQ signals, further explore their internal relationships, and help the model better understand the complex features of the signal. Finally, the signals of the four branches are further filtered with a (9, 5) convolution kernel to extract features and input into the time feature extraction module.

[0103] S2, adding multiplicative Gaussian noise between LSTM and Bi-LSTM, and inputting the spatial features of the IQ signal into them to extract the time features, so as to obtain the time series feature output result;

[0104] S21. Processing the spatial features of the IQ data by using LSTM to obtain a first time series feature;

[0105] S22, the first time series feature is input into Bi-LSTM through multiplicative Gaussian noise;

[0106] S23. Use Bi-LSTM to perform data processing on the first time series feature after multiplicative Gaussian noise to obtain the time series feature output result.

[0107] like Figure 5 As shown in the figure, the spatial feature extraction module fully mines the intrinsic relationship of IQ data through multi-stream branches, while the temporal feature extraction module utilizes the advantages of LSTM and Bi-LSTM networks in extracting time series features, and fully mines the effective information of the previous and next time nodes through the hidden layer to improve the recognition accuracy.

[0108] After passing through the spatial feature extraction module, the data input format is (124, 100), and the data enters the LSTM network forget gate to filter out valid information. The Sigmoid function filters out the unimportant information of the previous moment and this moment. The tanh function operation result selects the important information and transmits it to the next moment. The output data format is (124, 128). The Bi-LSTM network can perceive the "context" information more effectively through forward input and backward input, and dig deeper into the time series features.

[0109] S3. The time series feature output result is fully connected and multiplicative Gaussian noise is added to obtain a recognition result.

[0110] S31, outputting the time series feature output result through full connection to obtain a full connection output result;

[0111] S32. The fully connected output result is subjected to multiplicative Gaussian noise to obtain a recognition result.

[0112] In the process of neural network training, in order to alleviate the overfitting problem of the training model, noise is generally added to eliminate the influence of isolated features on the overall result. Common noises include additive noise and multiplicative noise. Multiplicative Gaussian noise is normally distributed, and the probability density expression is:

[0113]

[0114] Where σ represents the standard deviation of the noise level, σ 2 is its variance, and 1 is the average value of the noise level. Suppose the original input signal expression is S(t), then the expression of the output signal after introducing multiplicative Gaussian noise is:

[0115] S′(t)=S(t)×f(x)

[0116] In the time series feature module, although the introduction of LSTM and Bi-LSTM networks has certain advantages in extracting time series data, there will also be problems such as excessive parameters and overfitting in training. Therefore, multiplicative Gaussian noise is added between LSTM and Bi-LSTM and between the time feature extraction module and the fully connected output module. Gaussian noise can eliminate the overfitting of the model and also has a certain role in improving the accuracy and generalization ability of the model.

[0117] like Figure 7 , Figure 8 , Fig. 9 As shown in Table 1, a modulation recognition method and system of a CNN convolution and complex-valued convolution fusion type multi-stream spatiotemporal network has high recognition accuracy, and the network is compared with the existing advanced modulation signal recognition network, namely CNN, CLDNN, ResNET, LSTM, PET-CGDNN, MCLDNN, IQGMCL network, and the comparison experiment is carried out on the RML 2016.10a data set. From the comparison results, it can be seen that the present invention has a higher recognition effect.

[0118] Table 1 Network recognition accuracy indicators

[0119]

[0120] This embodiment is a modulation recognition method of a complex-valued convolution fusion multi-stream spatiotemporal network, including dividing the acquired IQ sample signal into an I channel signal, a Q channel signal and two IQ channel signals, respectively extracting spatial features through different convolution methods to obtain the spatial features of the IQ data, adding multiplicative Gaussian noise between LSTM and Bi-LSTM, and inputting the spatial features of the IQ signal into them for time feature extraction to obtain a time series feature output result, outputting the time series feature output result through full connection and adding multiplicative Gaussian noise to obtain a recognition result. A complex convolution module is introduced as a branch of the IQ data stream input to maximize the retention of the signal's phase information and the intrinsic correlation between the real part and the imaginary part, and fully exploit the complex features. The network adopts a combination of space-time dual modules, and the spatial feature extraction module contains multiple parallel processing streams to fully exploit the IQ signal features from multiple dimensions. The combination of space-time modules helps capture the evolution characteristics of the signal in time series. Multiplicative Gaussian noise is added between LSTM and Bi-LSTM, and between the time feature extraction module and the fully connected output module to improve the accuracy while eliminating the overfitting of the model and increasing the robustness of the model.

[0121] Embodiment 2

[0122] An embodiment of the present invention provides a modulation recognition system 200 of a complex-valued convolution fusion multi-stream spatiotemporal network, such as Figure 2 As shown, including:

[0123] The spatial feature extraction module 201 is used to divide the acquired IQ sample signal into an I channel signal, a Q channel signal and two IQ channel signals and perform spatial feature extraction respectively through different convolution methods to obtain the spatial features of the IQ data;

[0124] A time feature extraction module 202 is used to add multiplicative Gaussian noise between LSTM and Bi-LSTM, and input the spatial features of the IQ signal into the module to perform time feature extraction, thereby obtaining a time series feature output result;

[0125] The output module 203 is used to output the time series feature output result through full connection and add multiplicative Gaussian noise to obtain a recognition result.

[0126] Further, the spatial feature extraction module 201 includes:

[0127] A sample division module 2011, configured to divide the IQ sample signal into the I channel signal, the Q channel signal, a first IQ channel signal and a second IQ channel signal;

[0128] A first convolution module 2012 is used to perform a convolution operation on the I channel signal through a plurality of layers of convolution kernels to obtain a first spatial feature;

[0129] A second convolution module 2013 is used to perform a convolution operation on the Q channel signal through a plurality of layers of convolution kernels to obtain a second spatial feature;

[0130] A third convolution module 2014 is used to extract features of the first IQ channel signal in a three-step convolution manner to obtain a third spatial feature;

[0131] A fourth convolution module 2015 is used to extract features of the second IQ channel signal by complex convolution to obtain a fourth spatial feature;

[0132] The merging module 2016 is used to merge and filter the first spatial feature, the second spatial feature, the third spatial feature and the fourth spatial feature to obtain the spatial feature of the IQ data.

[0133] Further, the fourth convolution module 2015 includes:

[0134] A zero-padding module 20151, configured to perform a zero-padding operation on the second IQ channel signal;

[0135] The complex convolution module 20152 is used to perform a convolution operation on the zero-padded complex data and the complex convolution kernel to obtain a two-dimensional matrix;

[0136] A matrix reconstruction module 20153 is used to reconstruct and arrange the two-dimensional matrix to obtain a reconstructed matrix;

[0137] The linear transformation module 20154 is used to perform a linear transformation based on the reconstruction matrix to obtain the fourth spatial feature.

[0138] Further, the time feature extraction module 202 includes:

[0139] A data processing module 2021 is used to process the spatial features of the IQ data using LSTM to obtain a first time series feature;

[0140] A first noise module 2022, used for inputting the first time series feature into the Bi-LSTM via multiplicative Gaussian noise;

[0141] The time series feature output module 2023 is used to use Bi-LSTM to perform data processing on the first time series feature after multiplicative Gaussian noise to obtain the time series feature output result.

[0142] Further, the output module 203 includes:

[0143] A fully connected output module 2031 is used to output the time series feature output result through a fully connected output to obtain a fully connected output result;

[0144] The second noise module 2032 is used to subject the fully connected output result to multiplicative Gaussian noise to obtain a recognition result.

[0145] This embodiment is a modulation recognition system of a complex-valued convolution fusion multi-stream spatiotemporal network. The spatial feature extraction module is used to divide the acquired IQ sample signal into an I channel signal, a Q channel signal and two IQ channel signals, respectively, and extract spatial features through different convolution methods to obtain the spatial features of the IQ data; the time feature extraction module adds multiplicative Gaussian noise between LSTM and Bi-LSTM, and inputs the spatial features of the IQ signal into it for time feature extraction to obtain the time series feature output result; the output module outputs the time series feature output result through full connection and adds multiplicative Gaussian noise to obtain the recognition result. The complex convolution module is introduced as a branch of the IQ data stream input to maximize the retention of the signal's phase information and the intrinsic correlation between the real part and the imaginary part, and fully explore the complex features. The network adopts a combination of space-time dual modules. The spatial feature extraction module contains multiple parallel processing streams to fully explore the IQ signal features from multiple dimensions. The combination of space-time modules helps capture the evolution characteristics of the signal in time series. Multiplicative Gaussian noise is added between LSTM and Bi-LSTM, and between the time feature extraction module and the fully connected output module to improve the accuracy while eliminating the overfitting of the model and increasing the robustness of the model.

[0146] 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 modulation recognition method of a complex-valued convolutional fusion multi-stream spatiotemporal network, characterized in that: include: S1, dividing the acquired IQ sample signal into an I channel signal, a Q channel signal and two IQ channel signals, respectively extracting spatial features through different convolution methods to obtain spatial features of IQ data; S2, adding multiplicative Gaussian noise between LSTM and Bi-LSTM, and inputting the spatial features of the IQ signal into them to extract the time features, so as to obtain the time series feature output result; S3. The time series feature output result is fully connected and multiplicative Gaussian noise is added to obtain a recognition result.

2. The modulation recognition method of the complex-valued convolution fusion multi-stream spatiotemporal network according to claim 1 is characterized in that: The S1 includes: S11: dividing the IQ sample signal into the I channel signal, the Q channel signal, a first IQ channel signal and a second IQ channel signal; S12 causes the I channel signal to pass through a plurality of layers of convolution kernels to perform convolution operations to obtain a first spatial feature; S13 causes the Q channel signal to pass through a plurality of layers of convolution kernels to perform convolution operation to obtain a second spatial feature; S14: extracting features of the first IQ channel signal in a three-step convolution manner to obtain a third spatial feature; S15: extracting features of the second IQ channel signal by complex convolution to obtain a fourth spatial feature; S16 combines and filters the first spatial feature, the second spatial feature, the third spatial feature and the fourth spatial feature to obtain the spatial feature of the IQ data.

3. The modulation identification method of the complex-valued convolution fusion multi-stream spatiotemporal network according to claim 2 is characterized in that: The S15 includes: S151, performing a zero padding operation on the second IQ channel signal; S152, performing a convolution operation on the zero-padded complex data and the complex convolution kernel to obtain a two-dimensional matrix; S153, reconstructing and arranging the two-dimensional matrix to obtain a reconstructed matrix; S154. Perform a linear transformation based on the reconstruction matrix to obtain the fourth spatial feature.

4. The modulation identification method of the complex-valued convolution fusion multi-stream spatiotemporal network according to claim 1 is characterized in that: The S2 includes: S21. Processing the spatial features of the IQ data by using LSTM to obtain a first time series feature; S22, the first time series feature is input into Bi-LSTM through multiplicative Gaussian noise; S23. Use Bi-LSTM to process the first time series feature after multiplicative Gaussian noise to obtain the time series feature output result.

5. The modulation recognition method of the complex-valued convolution fusion multi-stream spatiotemporal network according to claim 1 is characterized in that: The S3 includes: S31, outputting the time series feature output result through full connection to obtain a full connection output result; S32. The fully connected output result is subjected to multiplicative Gaussian noise to obtain the recognition result.

6. A modulation recognition system of a complex-valued convolutional fusion multi-stream spatiotemporal network, characterized in that: include: A spatial feature extraction module is used to divide the acquired IQ sample signal into an I channel signal, a Q channel signal and two IQ channel signals, and perform spatial feature extraction respectively through different convolution methods to obtain the spatial features of the IQ data; A time feature extraction module, used for adding multiplicative Gaussian noise between LSTM and Bi-LSTM, and inputting the spatial features of the IQ signal into the module for time feature extraction to obtain a time series feature output result; The output module is used to output the time series feature output result through full connection and add multiplicative Gaussian noise to obtain the recognition result.

7. The modulation identification system of complex-valued convolution fusion multi-stream spatiotemporal network according to claim 6 is characterized in that: The spatial feature extraction module comprises: A sample division module, used for dividing the IQ sample signal into the I channel signal, the Q channel signal, a first IQ channel signal and a second IQ channel signal; A first convolution module, used to perform a convolution operation on the I channel signal through a plurality of layers of convolution kernels to obtain a first spatial feature; A second convolution module, used for performing a convolution operation on the Q channel signal through a plurality of layers of convolution kernels to obtain a second spatial feature; A third convolution module, used to perform feature extraction on the first IQ channel signal in a three-step convolution manner to obtain a third spatial feature; a fourth convolution module, configured to perform feature extraction on the second IQ channel signal by means of complex convolution to obtain a fourth spatial feature; A merging module is used to merge and filter the first spatial feature, the second spatial feature, the third spatial feature and the fourth spatial feature to obtain the spatial feature of the IQ data.

8. The modulation identification system of complex-valued convolution fusion multi-stream spatiotemporal network according to claim 7 is characterized in that: The fourth convolution module comprises: A zero-padding module, used for performing a zero-padding operation on the second IQ channel signal; The complex convolution module is used to perform convolution operation on the zero-padded complex data and the complex convolution kernel to obtain a two-dimensional matrix; A matrix reconstruction module, used for reconstructing and arranging the two-dimensional matrix to obtain a reconstructed matrix; A linear transformation module is used to perform a linear transformation based on the reconstruction matrix to obtain the fourth spatial feature.

9. The modulation identification system of complex-valued convolution fusion multi-stream spatiotemporal network according to claim 6 is characterized in that: The time feature extraction module comprises: A data processing module, used to process the spatial features of the IQ data using LSTM to obtain a first time series feature; A first noise module, used for inputting the first time series feature into the Bi-LSTM via multiplicative Gaussian noise; The time series feature output module is used to use Bi-LSTM to perform data processing on the first time series feature after multiplicative Gaussian noise to obtain the time series feature output result.

10. The modulation identification system of complex-valued convolution fusion multi-stream spatiotemporal network according to claim 6, characterized in that: The output module comprises: A fully connected output module, used for outputting the time series feature output result through full connection to obtain a fully connected output result; The second noise module is used to pass multiplicative Gaussian noise on the fully connected output result to obtain the recognition result.

Images