A Signal Modulation Recognition Method Based on AGRUN-DDFN

Abstract

The present invention discloses a signal modulation recognition method based on AGRUN-DDFN, comprising the following steps: S1, obtaining a received signal data set and dividing the received signal data set into a training signal data set and a test signal data set; S2, performing preprocessing to obtain a one-dimensional signal data set and a two-dimensional signal data set; S3, constructing a signal recognition network, the signal recognition network including a feature extraction network, a fusion layer, a connection layer, and an output layer connected in series; S4, performing deep learning training on the signal recognition network, and using the signal recognition network after training completion as a signal recognition modulator; S5, using the test signal data set as the input of the signal recognition modulator, and the output of the signal recognition modulator is the modulation type recognition result. The present invention combines the advantages of the attention gated recurrent convolutional network and the decoupled dynamic filtering convolutional network, improving the robustness of signal feature extraction and the accuracy of signal recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of signal recognition, and in particular to a signal modulation recognition method based on AGRUN-DDFN. Background Art

[0002] Signal modulation recognition is a core and key issue in the field of non-cooperative communication signal recognition technology. How to extract the subtle features of signals for modulation type recognition is a challenging task. The current research mainly falls into two categories: one is to select and extract the subtle features of signals by using expert human experience; the other is to automatically extract the subtle features of signals based on artificial intelligence.

[0003] The signal feature extraction method based on expert experience faces the problems of insufficient understanding of the essence of signal features, the application of mathematical tools, and the recognition mechanism, and it is difficult to understand and extract the signal features of modern multi-mode communication systems, resulting in complex feature extraction and low accuracy. The more commonly used artificial intelligence network models are mainly convolutional neural networks (CNN: Convolutional Neural Networks), recurrent neural networks (RNN: Recurrent Neural Networks), and generative adversarial networks (GAN: Generative Adversarial Networks), etc. However, there are still many technical problems. For example, the problem of efficient annotation of heterogeneous signal data, and the problems of low data processing accuracy, single data representation form, insufficient signal feature expressiveness and integrity in an unstable electromagnetic environment. At the same time, there are also problems of network model and training generalization adaptation. Restricted by factors such as the space-invariant characteristics of the network convolutional layer, the difference in signal body weight, and the limited number of data samples, conventional neural network models face problems such as low accuracy in multi-mode signal recognition, unbalanced model training, and small sample recognition. Summary of the Invention

[0004] In order to solve the problems of complex feature extraction and low accuracy of signal modulation recognition in the existing technology, the present invention proposes a signal modulation recognition method based on AGRUN-DDFN, which combines the advantages of the attention gated recurrent convolutional network that can extract the time-accumulated global features of signals, the decoupled dynamic filtering convolutional network that can extract content-adaptive multi-dimensional local features and multi-domain feature fusion, etc., to solve the above problems.

[0005] The present application discloses a signal modulation recognition method based on AGRUN-DDFN, including the following steps:

[0006] S1. Obtain a received signal data set through signal reception sampling, and divide the received signal data set into a training signal data set and a test signal data set;

[0007] S2. Preprocess the received signal dataset to obtain a one-dimensional signal dataset and a two-dimensional signal dataset;

[0008] S3. Construct a signal recognition network, which includes a feature extraction network, a fusion layer, a connection layer, and an output layer in serial cascade. The feature extraction network includes two parallel branches in cascade, an upper branch and a lower branch, which are composed of a decoupled dynamic filtering convolutional network and an attention gating recurrent convolutional network;

[0009] S4. Use the preprocessed training signal dataset to perform deep learning training on the signal recognition network. After training, the signal recognition network is used as a signal recognition modulator;

[0010] S5. Use the preprocessed test signal dataset as the input of the signal recognition modulator. The output of the signal recognition modulator is the recognition result of the modulation type.

[0011] Preferably, the preprocessing of the received signal dataset includes the following steps:

[0012] Perform background noise adaptive cancellation on the received signal dataset to generate a signal dataset with a relatively flat signal spectrum , in order to unify the component dimensions of the signal dataset and eliminate the adverse effects of the noise of the received modulation signal on numerical overflow and network convergence of the signal recognition network, it is necessary to perform normalization processing on the signal dataset to obtain a one-dimensional signal dataset . The signal normalization processing is as follows:

[0013]

[0014] where, represents taking the mean value, represents taking the standard deviation;

[0015] Perform time-frequency transformation processing on the one-dimensional signal to generate a two-dimensional signal dataset . For the time-domain data of the one-dimensional signal dataset , the short-time Fourier transform (STFT: Short time Fourier Transform) is adopted, that is, a time-domain signal with a length of is divided into multiple segments by using a window function and then subjected to Fourier transform processing. The window function slides on the time axis, performs segmented interception on the signal , and then the window function is multiplied by the signal Multiply, and then perform DFT calculation on the multiplication result to obtain a time-frequency matrix, that is, the two-dimensional time-frequency diagram of the signal :

[0016]

[0017] Among them, and are time variables, is the frequency variable, is a sliding window function with a length of to control the data length and windowing type involved in the Fourier transform.

[0018] Perform normalization processing on the two-dimensional time-frequency diagram to obtain the two-dimensional data set of the signal .

[0019] Preferably, the upper branch of the feature extraction network is a decoupled dynamic filtering convolutional network, including several decoupled dynamic filtering convolutional blocks. After each decoupled dynamic filtering convolutional block, a pooling layer is connected. After the last decoupled dynamic filtering convolutional block passes through the pooling layer, it is connected to the fusion layer.

[0020] The lower branch of the feature extraction network is an attention gated recurrent convolutional network, including several attention gated recurrent convolutional blocks. After each attention gated recurrent convolutional block, a pooling layer is connected. After the last attention gated recurrent convolutional block passes through the pooling layer, it is connected to the fusion layer.

[0021] Preferably, the pooling method of the pooling layer is average pooling. By adjusting the parameters of each pooling operation, redundant kernel features are gradually removed for simplification to reduce network parameters and increase generalization ability.

[0022] Preferably, the input of the decoupled dynamic filtering convolutional network is the two-dimensional data set of the signal , and the input of the attention gated recurrent convolutional network is the one-dimensional data set of the signal .

[0023] Preferably, the decoupled dynamic filtering convolutional block includes a first convolutional layer, a decoupled dynamic filtering block, and a second convolutional layer connected in series in sequence. The decoupled dynamic filtering convolutional block uses a shortcut connection.

[0024] The data processing of the first convolutional layer includes 1×1 convolution operation, batch normalization processing and activation function mapping processing in sequence. The data processing of the decoupled dynamic filter block includes decoupled dynamic filtering, batch normalization processing and activation function mapping processing in sequence. The data processing of the second convolutional layer includes 1×1 convolution operation, batch normalization processing and activation function mapping processing in sequence. Among them, the role of normalization processing is to speed up the network convergence process and enhance the feature classification effect. The activation function of the activation function mapping processing adopts the ReLU function.

[0025] Preferably, the data processing of the decoupled dynamic filtering includes spatial filtering parameter construction, channel filtering parameter construction, spatial channel filtering parameter fusion and kernel parameter application in sequence;

[0026] The spatial filtering parameter construction includes the following steps:

[0027] Assume that the input data have channels, and its frequency direction size is , the time direction is , the input is expressed as , set the length of the desired spatial filter along the frequency direction to , the length along the time direction is set to , that is, expect to use The filter processes the input features, then an input channel is applied , the output channel is of Convolutional Layer Process it. , and obtain the spatial parameters ,Right now:

[0028]

[0029] right The length of each spatial position is The vector is normalized, that is:

[0030]

[0031]

[0032] in, and are learnable parameters, The frequency position is and the time position is The spatial filtering parameters output at ;

[0033] Will Copy along channel dimension to After regularization, the spatial filtering parameters are obtained ;

[0034] The construction of the channel filtering parameters includes the following steps:

[0035] For the input data take the mean of each channel to obtain , and then continuously use two convolutions to process and finally obtain the channel branch parameters , the two convolutions are respectively expressed as and , is a constant within the value range of interval, that is:

[0036]

[0037]

[0038] Among them, is the input data at channel , frequency position , and time position ;

[0039] Copy along the second dimension times, and then copy times along the third dimension to obtain , after regularization, the channel filtering parameters ;

[0040] The fusion of the spatial channel filtering parameters includes the following steps:

[0041] Multiply the obtained spatial filtering parameters and the channel filtering parameters point by point to obtain , and then disassemble the first dimension of into three new dimensions with sizes of , and to obtain the final ;

[0042] For each channel, each time, and each frequency position of the input data, there is a corresponding filtering parameter with a size of . This filtering parameter is dynamically generated according to the input and is used for convolution with the input signal, called the dynamic convolution kernel.

[0043] The kernel parameter application includes the following steps:

[0044] Apply dynamic convolution kernel parameters For input data Filtering is performed, and the obtained filtering output result is recorded as , and its training formula is:

[0045]

[0046] in, express In the channel ( ), the frequency position is ( ), time position is ( ) The size is The position of the filter is The filter parameters at ; Indicates that the channel is The frequency position is , time position is Input data;

[0047] Preferably, the attention gated recurrent convolution block comprises a gated recurrent unit module and an attention transformation module which are serially cascaded in sequence.

[0048] Preferably, the attention transformation module is used to explore the intrinsic relationship between the feature vectors extracted by the gated recurrent unit module, that is, to extract more critical features and enhance the effectiveness of feature extraction. The network training formula of the attention transformation module is as follows:

[0049]

[0050]

[0051]

[0052] in, , is the feature vector output by the gated recurrent unit module, is the length of the feature vector, is the number of eigenvectors, , is the attention feature vector, is the number of attention feature vectors, and is the feature weight, for Activation function, and is the activated feature vector, and is a set of learnable attention transformation parameters.

[0053] Preferably, the fusion layer concatenates the - dimensional local feature vector and the - dimensional local feature vector head - to - tail to expand into the - dimensional fusion feature vector. The activation function used in the connection layer is the ReLU function, and the output layer uses the Softmax function for multi - classification output.

[0054] Advantages of the present invention:

[0055] (1) The present invention avoids the disadvantages of low complexity and low accuracy in artificial feature extraction in traditional signal recognition feature extraction methods, directly uses the original received signal for modulation type discrimination, and is more general and comprehensive.

[0056] (2) Compared with existing artificial intelligence algorithms, the present invention fully excavates the content - adaptive time - frequency domain local features and time - accumulated "key significant" global features of the signal, and uses the multi - domain fusion of global features and local features to greatly improve the robustness of signal feature extraction and the accuracy of signal recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is the flowchart of the signal modulation recognition method based on AGRUN - DDFN according to the embodiment of the present invention;

[0058] Figure 2 is the schematic diagram of the network structure of the decoupled dynamic filtering convolution block according to the embodiment of the present invention;

[0059] Figure 3 is the schematic diagram of the network structure of the decoupled dynamic filtering according to the embodiment of the present invention;

[0060] Figure 4 is the schematic diagram of the network structure of the attention gating recurrent convolution block according to the embodiment of the present invention;

[0061] Figure 5 is the structural block diagram of signal modulation recognition according to the embodiment of the present invention;

[0062] Figure 6(a) is the schematic diagram of the confusion matrix of modulation signal recognition with a signal - to - noise ratio of 2 according to the embodiment of the present invention;

[0063] Figure 6(b) is the schematic diagram of the confusion matrix of modulation signal recognition with a signal - to - noise ratio of 4 according to the embodiment of the present invention;

[0064] Figure 6(c) is the schematic diagram of the confusion matrix of modulation signal recognition with a signal - to - noise ratio of 6 according to the embodiment of the present invention;

[0065] Figure 6(d) is a schematic diagram of the confusion matrix for modulating signal recognition when the signal-to-noise ratio of the embodiment of the present invention is 8. Detailed implementation manners

[0066] To make the purpose, technical solutions and advantages of the present application clearer and more understandable, the following takes examples with reference to the accompanying drawings and further elaborates on the present application in detail.

[0067] The embodiment of the present application discloses a signal modulation recognition method based on AGRUN-DDFN, and the recognition processing flow is as Figure 1 shown, including the following steps:

[0068] S1. Obtain a received signal dataset through signal reception and sampling, and divide the received signal dataset into a training signal dataset and a test signal dataset.

[0069] According to the Nyquist sampling theorem, receive and sample the signal. While ensuring the acquisition of complete signal information, redundancy should be minimized as much as possible, and it is stored in the hdf5 format to generate a complete received signal dataset. Divide the received signal dataset into a training signal dataset and a test signal dataset.

[0070] S2. Preprocess the received signal dataset to obtain a one-dimensional signal dataset and a two-dimensional signal dataset.

[0071] Adopt a broadband spectrum estimation method based on the Welch periodogram method to adaptively cancel the background noise of the received signal dataset, generating a signal dataset with a relatively flat signal spectrum . To unify the component dimensions of the signal dataset and the adverse effects of the received modulation signal noise on numerical overflow and network convergence of the signal recognition network, it is necessary to normalize the signal dataset to obtain a one-dimensional signal dataset , and the signal normalization process is:

[0072]

[0073] where, represents taking the mean value, represents taking the standard deviation;

[0074] Perform time-frequency transformation processing on the one-dimensional signal to generate a two-dimensional signal dataset . Specifically, for the time-domain data of the one-dimensional signal dataset , the short-time Fourier transform (STFT: Short time FourierTransform) is adopted, that is, a time-domain signal with a length of is used with a window function Divide the signal into multiple segments for Fourier transform processing. The window function slides on the time axis, segments the signal by intercepting, and then multiplies the window function with the signal The result of the multiplication is subjected to DFT calculation to obtain a time-frequency matrix, that is, the two-dimensional time-frequency diagram of the signal :

[0075]

[0076] Among them, and are time variables, is the frequency variable, is a sliding window function with a length of which controls the data length participating in the Fourier transform and the windowing type.

[0077] Perform normalization processing on the two-dimensional time-frequency diagram to obtain the two-dimensional data set of the signal , and the normalization processing method is the same as that of the signal data set .

[0078] S3. Construct a signal recognition network. The signal recognition network includes a feature extraction network, a fusion layer, a connection layer, and an output layer in serial cascade. The feature extraction network includes two parallel upper and lower branches of a decoupled dynamic filtering convolutional network (Decoupled Dynamic Filtering Networks, DDFN) and an attention gated recurrent convolutional network (Attention Gated Recurrent Unit Networks, AGRUN) in parallel cascade.

[0079] Among them, the upper branch of the feature extraction network is a decoupled dynamic filtering convolutional network, which is used to extract the adaptive local features of the multi-dimensional content of the signal. The input is the two-dimensional data set of the signal , and the output is the content-adaptive time-frequency domain local features of the signal. The lower branch of the feature extraction network is an attention gated recurrent convolutional network, which is used to extract the time-accumulated significant key global features of the signal. The input signal is the one-dimensional data set of the signal , and the output is the time-accumulated significant key global features of the signal.

[0080] The fusion layer in this embodiment concatenates the -dimensional local feature vector and the -dimensional local feature vector head-to-tail to expand into a -dimensional fusion feature vector, achieving the effect of enriching the "diversity" of features. Its mathematical expression is .

[0081] The connection layer of this embodiment integrates the three-dimensional highly abstract features to reduce redundant features, selects several "good" features, and the activation function adopted by the connection layer is the ReLU function.

[0082] The output layer of this embodiment uses the Softmax function for multi-classification output, generating an output vector with a length of Each value of the output vector represents the probability value of the corresponding input of the neuron for all samples , and the expression of the Softmax function is:

[0083]

[0084] where is the probability value of each output classification, is the output of the neuron, and

[0085] The decoupled dynamic filtering convolutional network of this embodiment includes several decoupled dynamic filtering convolutional blocks, and the decoupled dynamic filtering convolutional blocks perform several times of decoupled dynamic filtering convolution on the input signal two-dimensional data set . After each decoupled dynamic filtering convolutional block, there is a pooling layer connected. After the last decoupled dynamic filtering convolutional block passes through the pooling layer, it is connected to the fusion layer. The network structures of the several decoupled dynamic filtering convolutional blocks adopted are the same, and only the network parameters of each decoupled dynamic convolutional block are adjusted to realize the gradual extraction of the signal's self-adaptive low-level features to recognizable high-level features from the signal two-dimensional data set . The pooling method adopted by the pooling layer is average pooling. By adjusting the parameters of each pooling operation, redundant kernel features are gradually removed and refined to reduce network parameters and increase generalization ability.

[0086] The network structure of the decoupled dynamic filtering convolutional block is as shown in Figure 2As shown in the figure, it includes a first convolutional layer, a decoupled dynamic filtering block, and a second convolutional layer that are serially cascaded in sequence. The data processing of the first convolutional layer sequentially includes 1×1 convolutional operation, batch normalization processing, and activation function mapping processing. The data processing of the decoupled dynamic filtering block sequentially includes decoupled dynamic filtering, batch normalization processing, and activation function mapping processing. The data processing of the second convolutional layer sequentially includes 1×1 convolutional operation, batch normalization processing, and activation function mapping processing. Among them, the role of the normalization processing is to accelerate the network convergence process and enhance the feature classification effect. The activation functions used in the activation function mapping processing are all ReLU functions. The decoupled dynamic filtering convolutional block adopts a shortcut connection, that is, the input data skips the first convolutional layer and the decoupled dynamic filtering block, and after performing residual summation with the output data of the batch normalization processing of the second convolutional layer, it is connected to the activation function mapping of the second convolutional layer to prevent overfitting and improve the iteration speed.

[0087] The network structure of the decoupled dynamic filtering is as Figure 3 shown, and its data processing sequentially includes spatial filtering parameter construction, channel filtering parameter construction, spatial-channel filtering parameter fusion, and kernel parameter application to perform decoupled dynamic filtering on the input data.

[0088] The spatial filtering parameter construction includes the following steps:

[0089] Assume the input data has channels, its size in the frequency direction is , and its size in the time direction is , and the input is represented as , where is a real variable; set the length of the expected generated spatial filter along the frequency direction to , and the length along the time direction to , that is, expect to use a filter of to process the input features, then apply a convolutional layer with an input channel of and an output channel of to process it, , and obtain the spatial parameter , that is:

[0090]

[0091] Then, normalize the vector of length at each spatial position of , that is:

[0092]

[0093] ​

[0094] Among them, and are learnable parameters, is the spatial filtering parameter output at the frequency position of and the time position of At the place.

[0095] Finally, is copied along the channel dimension times to After regularization, the spatial filtering parameter is obtained.

[0096] The construction of the channel filtering parameter includes the following steps:

[0097] First, the mean value of each channel of the input data is taken to obtain , and then two convolutions are continuously used to process , and finally the channel branch parameter is obtained. The two convolutions are respectively expressed as and . is a constant within the value range of interval, that is:

[0098]

[0099]

[0100] Among them, is the input data at the channel of , the frequency position of , and the time position of At the place.

[0101] Will is copied along the second dimension times, and then copied along the third dimension times to obtain . After regularization, the channel filtering parameter is obtained.

[0102] The fusion of the spatial channel filtering parameter includes the following steps:

[0103] The obtained spatial filtering parameter and the channel filtering parameter are multiplied point by point to obtain , and then the first dimension of is disassembled into sizes of , and The three new dimensions obtain the final .

[0104] For each channel, each time, and each frequency position of the input data, all have corresponding filtering parameters of size . The filtering parameters are dynamically generated according to the input and are used to convolve with the input signal, which is called a dynamic convolution kernel.

[0105] The application of kernel parameters, that is, the decoupled filtering using kernel parameters, includes the following steps:

[0106] Apply the dynamic convolution kernel parameters to the input data for filtering, and denote the obtained filtered output result as , and its training formula is:

[0107]

[0108] where, represents the filter parameter at position ( ) in the filter of size ( ) at channel ( ) and frequency position ; represents the input data at channel , frequency position , and time position .

[0109] The attention gating recurrent convolutional network of this embodiment includes several attention gating recurrent convolutional blocks. The attention gating recurrent convolutional blocks perform several attention gating recurrent convolutions on the one-dimensional signal data of the input. A pooling layer is connected after each attention gating recurrent convolutional block. The last attention gating recurrent convolutional block is connected to the fusion layer after passing through the pooling layer. The network structures of the several attention gating recurrent convolutional blocks adopted are the same, and only the network parameters of each attention gating recurrent convolutional block are adjusted to gradually extract the low-level features related to the front and back in the time dimension from the one-dimensional signal data to recognizable high-level features. The pooling method of the adopted pooling layer is average pooling.

[0110] The network structure of the attention gating recurrent convolutional block is as Figure 4 ​As shown, it includes a gated recurrent unit module (GRU) and an attention transformation module connected in series in sequence. The gated recurrent unit module is used to extract the temporal cumulative global features of the signal, and the attention transformation module is used to extract the significant key temporal cumulative global features of the signal.

[0111] The gated recurrent unit module in this embodiment solves the problems of long-term dependence and historical forgetting in feature extraction by traditional networks. Its network training formula is as follows:

[0112]

[0113]

[0114]

[0115]

[0116] Among them, represents the update gate of the gated recurrent unit module, represents the reset gate of the gated recurrent unit module. The subscript is used to identify the moment, that is, represents the degree to which the state information of the previous moment is brought into the current state, represents the degree to which the current state ignores the state of the previous moment, is the weight matrix of the gate, is the weight matrix of the gate, is the weight matrix of the output state, is the input data at the moment of is the candidate state at the moment of is the output state at the moment of is the activation function, used to activate the control gate, is the activation function, used to activate the candidate state.

[0117] The attention transformation module adopted in this embodiment can explore the internal relationship between the feature vectors extracted by the gated recurrent unit module, that is, extract more critical features and enhance the effectiveness of feature extraction. Its network training formula is as follows:

[0118]

[0119]

[0120]

[0121] Among them, , is the feature vector output by the gated recurrent unit module, is the length of the feature vector, is the number of feature vectors, , is the attention feature vector, is the number of attention feature vectors, and are the feature weighting weights, is the activation function, and are the activated feature vectors, and are a set of learnable attention transformation parameters.

[0122] S4. Use the preprocessed training signal dataset to perform deep learning training on the signal recognition network. The signal recognition network after training is used as the signal recognition modulator.

[0123] S5. Use the preprocessed test signal dataset as the input of the signal recognition modulator. The output of the signal recognition modulator is the modulation type recognition result.

[0124] In a specific embodiment, the present application is verified through data simulation. The received signal dataset uses the RML2018.01 modulation signal dataset. It is assumed that the signal dataset has 2,555,904 pieces, including a total of 24 modulation methods such as BPSK, QPSK, 8PSK, FM, GMSK, QAM16, and QAM64. Each modulation method has 106,496 modulation signals. It is assumed that each modulation in the received signal dataset contains 26 signal-to-noise ratios, and each signal-to-noise ratio contains 4,096 modulation signals. The size of each modulation signal is (2, 1024), where 2 corresponds to the I / Q two-channel modulation signal and 1024 corresponds to 1024 sampling points. In this embodiment, it is assumed that 90% of the received signal dataset is used as the training signal dataset for training and learning the signal recognition network, and 10% of the received signal dataset is used as the test signal dataset for testing the performance of the signal recognition network.

[0125] In the signal preprocessing of this embodiment, it is assumed that the STFT parameters of the time-frequency transformation are: the number of samples is 256, the overlapping length is 127, the window function type is Hamming, the window function length is 128, the FFT length is 128, and the two-dimensional data matrix dimension of the output signal is 128×128.

[0126] Adopt as Figure 5The neural network model shown is used as a signal recognition network, and the upper / lower branch feature extraction network, fusion layer, connection layer, and output layer are configured as follows:

[0127] The input data dimension of the upper branch decoupled dynamic filtering convolutional network is 16×128×128, which is obtained by replicating the two-dimensional signal data 16 times. It contains 3 decoupled dynamic filtering convolutional blocks. The network structures of each decoupled dynamic filtering convolutional block are the same, and the network structure parameter configurations are dynamically adjustable. The channel dimensions, frequency domain dimensions, and time domain dimensions of the 1st, 2nd, and 3rd decoupled dynamic filtering convolutional blocks are 16×128×128, 16×32×32, and 64×8×8 respectively. The upper branch decoupled dynamic filtering convolutional network contains 3 pooling layers. The 1st pooling layer contains 16 features, with a data dimension of 64×64, and uses average pooling for extraction, with an extraction ratio of 2; the 2nd pooling layer contains 32 features, with a data dimension of 16×16, and uses average pooling for extraction, with an extraction ratio of 2; the 3rd pooling layer contains 64 features, with a data dimension of 1×1, and uses average pooling for extraction, with an extraction ratio of 8. The number of samples for batch normalization in the decoupled dynamic filtering convolutional blocks of the upper branch decoupled dynamic filtering convolutional network is 512.

[0128] The lower branch attention gated recurrent convolutional network contains 3 attention gated recurrent convolutional blocks. The network structures of each attention gated recurrent convolutional block are the same, and the network structure parameter configurations are dynamically adjustable. The filtered data dimensions of the 1st, 2nd, and 3rd attention gated recurrent convolutional blocks are 16×1024, 32×256, and 64×32 respectively. The lower branch attention gated recurrent convolutional network contains 3 pooling layers. The 1st pooling layer contains 16 features, with a data dimension of 1×512, and uses average pooling for extraction, with an extraction ratio of 2; the 2nd pooling layer contains 32 features, with a data dimension of 1×128, and uses average pooling for extraction, with an extraction ratio of 2; the 3rd pooling layer contains 128 features, with a data dimension of 1×1, and uses average pooling for extraction, with an extraction ratio of 32.

[0129] The fusion layer contains 64 + 128 = 192 neuron nodes, the connection layer contains 64 neuron nodes, and the output layer uses a Softmax fully connected layer to output the recognition results of 24 modulation type signals.

[0130] Performance tests were conducted on the recognition rates of different modulation types at different signal-to-noise ratios. Figures 6(a)-6(d) show the confusion matrices for the recognition of 24 modulation signals in different noise scenarios with signal-to-noise ratios of 2, 4, 6, and 8. It can be directly seen from the figures that at a signal-to-noise ratio of 8 dB, most modulation types have relatively high recognition rates, and only the recognition of 128QAM, 128APSK, and 64QAM has relatively poor results because these signals are relatively similar, making it difficult to recognize. When the signal-to-noise ratio drops to 6 dB, it can be found that the recognition rates of most modulation types are still at a relatively high level. However, as the signal-to-noise ratio further decreases (i.e., the noise increases), the diagonal of the confusion matrix becomes increasingly blurred, and the recognition rates of various modulation methods all decrease.

[0131] In the prior art, for most high-order modulation types in a signal data set, even when the signal-to-noise ratio is 8 dB, it is difficult for cooperative communication to demodulate the signals. The signal recognition network constructed in this application can non-cooperatively recognize most modulation signal types at a signal-to-noise ratio of 8 dB. The reason is that the signal recognition network integrates the advantages of the AGRUN network that can extract the time-accumulated global features of signals, the DDFN network that can extract content-adaptive multi-dimensional local features, and multi-domain feature fusion, etc., fully exploiting the time and content correlations of modulation signals, and solving problems such as weak network generalization ability and decreased recognition accuracy for multi-mode modulation signals caused by content-irrelevant convolution operations and relatively small convolution perception fields. This reflects the recognition robustness and generalization ability of the signal recognition network of this application.

[0132] The above has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and all these changes and improvements fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A signal modulation recognition method based on AGRUN-DDFN, characterized in that, It includes the following steps: S1. Obtain a received signal dataset through signal reception and sampling, and divide the received signal dataset into a training signal dataset and a test signal dataset; S2. Preprocess the received signal dataset to obtain a one-dimensional signal dataset and a two-dimensional signal dataset; S3. Construct a signal recognition network. The signal recognition network includes a feature extraction network, a fusion layer, a connection layer, and an output layer connected in series. The feature extraction network includes upper and lower branches connected in parallel by a decoupled dynamic filtering convolutional network and an attention gated recurrent convolutional network. The upper branch of the feature extraction network is a decoupled dynamic filtering convolutional network, which includes several decoupled dynamic filtering convolutional blocks. A pooling layer is connected after each decoupled dynamic filtering convolutional block. After the last decoupled dynamic filtering convolutional block passes through the pooling layer, it is connected to the fusion layer. The lower branch of the feature extraction network is an attention gated recurrent convolutional network, which includes several attention gated recurrent convolutional blocks. A pooling layer is connected after each attention gated recurrent convolutional block. After the last attention gated recurrent convolutional block passes through the pooling layer, it is connected to the fusion layer; The decoupled dynamic filtering convolutional block includes a first convolutional layer, a decoupled dynamic filtering block, and a second convolutional layer connected in series in sequence. The decoupled dynamic filtering convolutional block uses a shortcut connection. The data processing of the first convolutional layer and the second convolutional layer both sequentially include 1×1 convolutional operation, batch normalization processing, and activation function mapping processing. The data processing of the decoupled dynamic filtering block sequentially includes decoupled dynamic filtering, batch normalization processing, and activation function mapping processing; The attention gated recurrent convolutional block includes a gated recurrent unit module and an attention transformation module connected in series in sequence. The attention transformation module is used to explore the internal relationship between the feature vectors extracted by the gated recurrent unit module. The network training formula of the attention transformation module is as follows: Among them, is the feature vector output by the gated recurrent unit module, N f is the length of the feature vector, B is the number of feature vectors, s q is the attention feature vector, S is the number of attention feature vectors, [α 1i , …, α Si ∈ R 1×S and α qi are the feature weighting weights, g(·) is the tansig activation function, δ qi and δ ql are the activated feature vectors, and are a set of learnable attention transformation parameters; S4. Use the preprocessed training signal dataset to perform deep learning training on the signal recognition network. The signal recognition network after training is used as a signal recognition modulator; S5. Use the preprocessed test signal dataset as the input of the signal recognition modulator. The output of the signal recognition modulator is the modulation type recognition result.

2. The signal modulation recognition method based on AGRUN-DDFN according to claim 1, wherein The preprocessing of the received signal dataset includes the following steps: Perform background noise adaptive cancellation on the received signal data set to generate a signal data set For the signal data set Perform normalization processing to obtain the one-dimensional signal data set x. The calculation formula for signal normalization processing is as follows: Among them, E[·] represents taking the mean, and std[·] represents taking the standard deviation; Perform short-time Fourier transform on the one-dimensional signal dataset x to obtain a two-dimensional time-frequency diagram where n and m are time variables, k is a frequency variable, is a sliding window function of length N; For the two-dimensional time-frequency diagram perform normalization processing to obtain the two-dimensional signal data set S x .

3. The signal modulation recognition method based on AGRUN-DDFN according to claim 2, wherein, The pooling method of the pooling layer is average pooling.

4. The signal modulation recognition method based on AGRUN-DDFN according to claim 3, characterized in that The input of the decoupled dynamic filtering convolutional network is the two-dimensional signal dataset S x , and the input of the attention gated recurrent convolutional network is the one-dimensional signal dataset x.

5. The signal modulation recognition method based on AGRUN-DDFN according to claim 4, wherein The data processing of the decoupled dynamic filtering sequentially includes spatial filtering parameter construction, channel filtering parameter construction, spatial-channel filtering parameter fusion, and kernel parameter application; The spatial filtering parameter construction includes the following steps: Assume that the input data X has C channels, with a size of H in the frequency direction and a size of W in the time direction, and the input is represented as X ∈ R C ×H×W , set the length of the desired spatial filter along the frequency direction to K and the length along the time direction to L, that is, it is desired to process the input features using a K×L filter, then apply a 1×1 convolutional layer with C input channels and KL output channels to process it, obtaining the spatially parameterized z (SP) ∈ R KL×H×W , that is: For z (SP) normalize the vectors of length KL×1×1 at each spatial position, i.e.: where α and γ are learnable parameters, and z (SP) [h, w] are the spatial filtering parameters output at the frequency position h and the time position w; Copy z (SP) C times along the channel dimension to obtain z SP . After regularization, the spatial filtering parameter z SP ∈R C×KL×H×W ; The channel filtering parameter construction includes the following steps: The mean value of each channel of the input data X is taken to obtain X GAP ∈R C×1×1 , and then two consecutive 1×1 convolutions are used to process X GAP , and finally the channel branch parameter z (CH) ∈R CKL×1×1 is obtained. The two 1×1 convolutions are respectively denoted as and ε is a constant with a value range in the interval (0, 1), that is: Among them, X[c, h, w] is the input data at the channel c, frequency position h, and time position w; Duplicate z (CH) H times along the second dimension and then W times along the third dimension to obtain z CH . After regularization, obtain the channel filtering parameter z CH ∈R CKL×H×W ; The spatial-channel filtering parameter fusion includes the following steps: The obtained spatial filtering parameter z SP and the channel filtering parameter z CH are multiplied point by point to obtain z (DDF) ∈R CKL×H×W Then, the first dimension of z (DDF) is decomposed into three new dimensions of size C, K, and L to obtain the final z DDF ∈R C×K×L×H×W ; The kernel parameter includes the following steps: Apply the dynamic convolution kernel parameter z DDF Filter the input data X, and denote the resulting filtered output as Y ∈ R C ×H×W , and its training formula is: where z DDF [c, k, l, h, w] represents z DDF the filter parameter at position (k, l) in a filter of size K×L at channel c (1 << c << C), frequency position h (1 << h << H), and time position w (1 << w << W); X[c, h - k, w - l] represents the input data at channel c, frequency position h - k, and time position w - l.

6. The signal modulation recognition method based on AGRUN-DDFN according to claim 5, wherein, The fusion layer concatenates the Q1-dimensional local feature vector and the Q2-dimensional local feature vector head to tail to expand into a Q1+Q2-dimensional fusion feature vector. The activation function used by the connection layer is the ReLU function, and the output layer uses the Softmax function for multi-classification output.

Images