Signal modulation identification method based on AGRUNN-DDFN

Abstract

The invention discloses a signal modulation identification method based on AGRUNN-DDFN, and the method comprises 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 signal one-dimensional data set and a signal two-dimensional data set; s3, constructing a signal identification network, wherein the signal identification network comprises a feature extraction network, a fusion layer, a connection layer and an output layer which are cascaded in series; s4, performing deep learning training on the signal identification network, and taking the trained signal identification network as a signal identification modulator; and S5, using the test signal data set as the input of a signal identification modulator, wherein the output of the signal identification modulator is a modulation type identification result. According to the method, the advantages of the attention gating cyclic convolutional network and the decoupling dynamic filtering convolutional network are integrated, and the signal feature extraction robustness and the signal identification accuracy are improved.

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 key issue in the field of non-partner communication signal recognition technology. How to extract subtle signal features for modulation type recognition is a challenging task. Current research is mainly divided into two categories: one is to use expert experience to select and extract subtle signal features; the other is to automatically extract subtle signal features based on artificial intelligence.

[0003] Signal feature extraction methods based on expert experience face the problem of not having a deep understanding of the nature of signal feature generation, 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), recurrent neural networks (RNN) and generative adversarial networks (GAN). However, there are still many technical problems. For example, the problem of efficient labeling of heterogeneous signal data, the problem of low data processing accuracy, single data representation form, insufficient signal feature expression and integrity in unstable electromagnetic environments. At the same time, there are also problems with network model and training generalization adaptation. Due to the constraints of factors such as the spatial invariance of the network convolution base layer, the difference in signal weight and the limited number of data samples, conventional neural network models face the problems of low multi-mode signal recognition accuracy, 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 signal modulation recognition, the present invention proposes a signal modulation recognition method based on AGRUN-DDFN, which combines the advantages of attention gated recurrent convolutional network for extracting temporal cumulative global features of signals, decoupled dynamic filtering convolutional network for extracting content-adaptive multi-dimensional local features and multi-domain feature fusion, so as to solve the above problems.

[0005] The present application discloses a signal modulation identification method based on AGRUN-DDFN, comprising the following steps: S1. Acquire a received signal data set by signal reception sampling, and divide the received signal data set into a training signal data set and a test signal data set; S2, preprocessing the received signal data set to obtain a signal one-dimensional data set and a signal two-dimensional data set; S3. Construct a signal recognition network, which includes a serially cascaded feature extraction network, a fusion layer, a connection layer, and an output layer. The feature extraction network includes two upper and lower branches of a decoupled dynamic filtering convolutional network and an attention gated recurrent convolutional network that are cascaded in parallel. S4, using the preprocessed training signal data set to perform deep learning training on the signal recognition network, and the trained signal recognition network is used as a signal recognition modulator; S5. Use the preprocessed 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.

[0006] Preferably, the preprocessing of the received signal data set comprises the following steps: Adaptively cancel the background noise of the received signal data set to generate a signal data set with a relatively flat signal spectrum In order to unify the signal data set component dimensions and the adverse effects of received modulation signal noise on signal recognition network numerical overflow and network convergence, the signal data set needs to be Normalization is performed to obtain a one-dimensional data set of signals , the signal normalization process is:

[0007] in, It means taking the mean value, Indicates the standard deviation; One-dimensional data of the signal Perform time-frequency transformation to generate a two-dimensional signal data set . For a one-dimensional data set of signals Time domain data , using short time Fourier transform (STFT: Short time Fourier Transform), that is, a length of The time domain signal , using the window function The signal Divide into multiple segments for Fourier transform processing. Window function Slide on the timeline to adjust the signal Perform segmented interception and then use the window function With signal Multiply them, and then perform DFT calculation on the result to get the time-frequency matrix, that is, the two-dimensional time-frequency diagram of the signal :

[0008] in, and is the time variable, is the frequency variable, The length is The sliding window function controls the length of the data involved in Fourier transform and the windowing type.

[0009] For two-dimensional time-frequency diagram Normalization is performed to obtain a two-dimensional signal data set .

[0010] Preferably, the upper branch of the feature extraction network is a decoupled dynamic filtering convolutional network, which includes several decoupled dynamic filtering convolutional blocks, each of which is connected to a pooling layer, and the last decoupled dynamic filtering convolutional block is connected to the fusion layer after passing through the pooling layer.

[0011] The lower branch of the feature extraction network is an attention gated recurrent convolutional network, which includes several attention gated recurrent convolutional blocks, each of which is connected to a pooling layer, and the last attention gated recurrent convolutional block is connected to the fusion layer after passing through the pooling layer.

[0012] Preferably, the pooling method of the pooling layer is average pooling, and by adjusting the parameters of each pooling operation, redundant core features are gradually removed and simplified to reduce network parameters and increase generalization ability.

[0013] Preferably, the input of the decoupled dynamic filtering convolutional network is a two-dimensional data set of signals , the input of the attention gated recurrent convolutional network is a one-dimensional data set of signals .

[0014] Preferably, the decoupled dynamic filtering convolution block comprises a first convolution layer, a decoupled dynamic filtering block and a second convolution layer which are serially cascaded in sequence, and the decoupled dynamic filtering convolution block adopts a shortcut connection.

[0015] 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.

[0016] 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; The spatial filtering parameter construction includes the following steps: Assume that the input data have channels, and their 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:

[0017] right The length of each spatial position is The vector is normalized, that is:

[0018]

[0019] in, and are learnable parameters, The frequency position is and the time position is The spatial filtering parameters output at ; Will Copy along channel dimension Arrival , after regularization, the spatial filtering parameters are obtained ; The channel filter parameter construction includes the following steps: For input data The average value of each channel is , and then use two Convolution pair Processing is performed to finally obtain the channel branch parameters ,two Convolution is represented as and , The value range is A constant in the interval, that is:

[0020]

[0021] in, For the channel The frequency position is , time position is Input data at Will Copy along the second dimension times, and then replicate along the third dimension. Get , after regularization, the channel filter parameters are obtained ; The spatial channel filter parameter fusion includes the following steps: The obtained spatial filtering parameters and channel filter parameters Multiply point by point to get , and then The first dimension of is decomposed into , and The three new dimensions give the final ; For each channel, each time and each frequency position of the input data, All have corresponding sizes The filter parameters are dynamically generated according to the input and used for convolution with the input signal, which is called a dynamic convolution kernel.

[0022] The kernel parameter application includes the following steps: 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:

[0023] 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; Preferably, the attention gated recurrent convolution block comprises a gated recurrent unit module and an attention transformation module which are serially cascaded in sequence.

[0024] 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:

[0025]

[0026]

[0027] 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.

[0028] Preferably, the fusion layer will dimensional local eigenvectors and The local feature vector of the dimension is expanded by concatenating the head and tail The activation function used in the connection layer is the ReLU function, and the output layer uses the Softmax function for multi-classification output.

[0029] Beneficial effects of the present invention: (1) The present invention avoids the drawbacks of complexity and low accuracy of manually extracted features in traditional signal recognition feature extraction methods, and directly uses the original received signal to distinguish the modulation type, which is more universal and comprehensive.

[0030] (2) Compared with existing artificial intelligence algorithms, the present invention fully exploits the content-adaptive time-frequency domain local features and time-accumulated “key and significant” global features of the signal, and utilizes multi-domain fusion of global and local features to greatly improve the robustness of signal feature extraction and the accuracy of signal recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1A flowchart of a signal modulation identification method based on AGRUN-DDFN according to an embodiment of the present invention; Figure 2 A schematic diagram of the network structure of a decoupled dynamic filtering convolution block according to an embodiment of the present invention; Figure 3 A schematic diagram of a network structure of a decoupled dynamic filter according to an embodiment of the present invention; Figure 4 A schematic diagram of the network structure of an attention gated recurrent convolution block according to an embodiment of the present invention; Figure 5 A structural block diagram of signal modulation identification according to an embodiment of the present invention; FIG6 (a) is a schematic diagram of a confusion matrix for modulated signal recognition when the signal-to-noise ratio is 2 according to an embodiment of the present invention; FIG6 (b) is a schematic diagram of a confusion matrix for modulated signal recognition when the signal-to-noise ratio is 4 according to an embodiment of the present invention; FIG6 (c) is a schematic diagram of a confusion matrix for modulated signal recognition when the signal-to-noise ratio is 6 according to an embodiment of the present invention; FIG6( d ) is a schematic diagram of a confusion matrix for modulated signal recognition when the signal-to-noise ratio is 8 according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and examples.

[0033] The embodiment of the present application discloses a signal modulation identification method based on AGRUN-DDFN, and the identification process is as follows: Figure 1 As shown, the following steps are included: 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.

[0034] According to the Nyquist sampling theorem, the signal is received and sampled, and redundancy should be minimized while ensuring the acquisition of complete signal information. The data is stored in the hdf5 format to generate a complete received signal data set. The received signal data set is divided into a training signal data set and a test signal data set.

[0035] S2. Preprocess the received signal data set to obtain a one-dimensional signal data set and a two-dimensional signal data set.

[0036] The broadband spectrum estimation method based on the Welch periodogram method is used to adaptively cancel the background noise of the received signal data set to generate a signal data set with a relatively flat signal spectrum. In order to unify the signal data set component dimensions and the adverse effects of received modulation signal noise on signal recognition network numerical overflow and network convergence, the signal data set needs to be Normalization is performed to obtain a one-dimensional data set of signals , the signal normalization process is:

[0037] in, It means taking the mean value, Indicates the standard deviation; One-dimensional data of the signal Perform time-frequency transformation to generate a two-dimensional signal data set Specifically, for a one-dimensional data set of signals Time domain data , using short time Fourier transform (STFT: Short time Fourier Transform), that is, a length of The time domain signal , using the window function The signal Divide into multiple segments for Fourier transform processing. Window function Slide on the timeline to adjust the signal Perform segmented interception and then use the window function With signal Multiply them, and perform DFT calculation on the result to get the time-frequency matrix, that is, the two-dimensional time-frequency diagram of the signal :

[0038] in, and is the time variable, is the frequency variable, The length is The sliding window function controls the length of the data involved in Fourier transform and the windowing type.

[0039] For two-dimensional time-frequency diagram Normalization is performed to obtain a two-dimensional signal data set , Normalization Method and Signal Dataset The normalization process is the same.

[0040] S3. Build a signal recognition network. The signal recognition network includes a serially cascaded feature extraction network, a fusion layer, a connection layer, and an output layer. The feature extraction network includes the upper and lower branches of the parallel cascaded decoupled dynamic filtering convolutional networks (DDFN) and the attention gated recurrent unit networks (AGRUN).

[0041] Among them, the upper branch of the feature extraction network is a decoupled dynamic filtering convolutional network, which is used to extract adaptive local features of the multi-dimensional content of the signal, and the input is a two-dimensional data set of the signal , the content-adaptive time-frequency domain local features of the output 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 a one-dimensional data set of the signal. , the temporal accumulation of the output signal is a key global feature.

[0042] The fusion layer of this embodiment will dimensional local eigenvectors and The local feature vector of the dimension is expanded by concatenating the head and tail dimensional fusion feature vector to achieve the "diversity" of rich features. Its mathematical expression is: .

[0043] The connection layer of this embodiment is The highly abstract features are integrated to reduce redundant features, and several "good" features are selected. The activation function used in the connection layer is the ReLU function.

[0044] The output layer of this embodiment uses the Softmax function for multi-classification output, generating a length of The output vector of each value in the output vector represents the probability value of all samples corresponding to the neuron input , the expression of Softmax function is:

[0045] in, Output the classification probability value for each class, For the The output of a neuron, is the number of categories.

[0046] The decoupled dynamic filtering convolution network of this embodiment includes a plurality of decoupled dynamic filtering convolution blocks, and the decoupled dynamic filtering convolution blocks are used to input a two-dimensional signal data set. Perform several decoupled dynamic filter convolutions. Each decoupled dynamic filter convolution block is connected to a pooling layer, and the last decoupled dynamic filter convolution block is connected to the fusion layer after the pooling layer. The network structures of the several decoupled dynamic filter convolution blocks used are the same. Only the network parameters of each decoupled dynamic filter convolution block are adjusted to realize the signal from the signal two-dimensional data set. The signal adaptive low-level features are gradually extracted to recognizable high-level features. The pooling method used in the pooling layer is average pooling. By adjusting the parameters of each pooling operation, the redundant core features are gradually removed and simplified to reduce network parameters and increase generalization ability.

[0047] The network structure of the decoupled dynamic filter convolution block is as follows Figure 2 As shown, it includes a first convolutional layer, a decoupled dynamic filter block, and a second convolutional layer that are serially cascaded in sequence. 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 uses the ReLU function. The decoupled dynamic filter convolution block adopts a shortcut connection, that is, the input data skips the first convolutional layer and the decoupled dynamic filter block, and after the 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.

[0048] The network structure of decoupled dynamic filtering is as follows Figure 3 As shown, the data processing includes spatial filter parameter construction, channel filter parameter construction, spatial channel filter parameter fusion and kernel parameter application in sequence, and decoupled dynamic filtering is performed on the input data.

[0049] The spatial filtering parameter construction includes the following steps: Assume that the input data have channels, and their frequency direction size is , the time direction is , the input is expressed as ,in is a real variable; the length of the desired spatial filter along the frequency direction is set 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:

[0050] Then, yes The length of each spatial position is The vector is normalized, that is:

[0051]

[0052] in, and are learnable parameters, The frequency position is and the time position is The spatial filter parameters output at .

[0053] Finally, Copy along channel dimension Arrival , after regularization, the spatial filtering parameters are obtained .

[0054] The channel filter parameter construction includes the following steps: First, the input data The average value of each channel is , and then use two Convolution pair After processing, the channel branch parameters are finally obtained. ,two Convolution is represented as and , The value range is A constant in the interval, that is:

[0055]

[0056] in, For the channel The frequency position is , time position is The input data at .

[0057] Will Copy along the second dimension times, and then replicate along the third dimension. Get , after regularization, the channel filter parameters are obtained .

[0058] The spatial channel filter parameter fusion includes the following steps: The obtained spatial filtering parameters and channel filter parameters Multiply point by point to get , and then The first dimension of is decomposed into , and The three new dimensions give the final .

[0059] For each channel, each time and each frequency position of the input data, All have corresponding sizes The filter parameters are dynamically generated according to the input and used for convolution with the input signal, which is called a dynamic convolution kernel.

[0060] The application of kernel parameters, that is, the application of kernel parameters for decoupling filtering, includes the following steps: 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:

[0061] 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 The input data.

[0062] The attention gated recurrent convolution network of this embodiment includes several attention gated recurrent convolution blocks. The attention gated recurrent convolution blocks perform one-dimensional data of the input signal on the input signal. Perform several attention gated recurrent convolutions, connect each attention gated recurrent convolution block with a pooling layer, and connect the last attention gated recurrent convolution block with the fusion layer after the pooling layer. The network structures of the several attention gated recurrent convolution blocks used are the same, and only the network parameters of each attention gated recurrent convolution block are adjusted to achieve the one-dimensional data of the signal. The low-level features related to the time dimension are gradually extracted to recognizable high-level features. The pooling method of the pooling layer used is average pooling.

[0063] The network structure of the attention gated recurrent convolutional block is as follows Figure 4 As shown, it includes a gated recurrent unit module (GRU) and an attention transformation module that are serially cascaded 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.

[0064] The gated recurrent unit module in this embodiment solves the long-term dependency and historical forgetting problems in feature extraction of traditional networks. The network training formula is as follows:

[0065]

[0066]

[0067]

[0068] in, represents the update gate of the gated recurrent unit module, Represents the reset gate of the gated recurrent unit module, and the subscript is used to identify the moment, i.e. Indicates the degree to which the state information of the previous moment is brought into the current state. Indicates the degree to which the current state ignores the state of the previous moment. for The weight matrix of the gate, for The weight matrix of the gate, is the weight matrix of the output state, for Input data at time, for The candidate state at the moment, for The output status at the moment, for Activation function, used to activate the control gate, for Activation function, used to activate candidate states.

[0069] The attention transformation module used in this embodiment can mine the intrinsic 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. The network training formula is as follows:

[0070]

[0071]

[0072] 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.

[0073] S4. Use the preprocessed training signal data set to perform deep learning training on the signal recognition network, and the trained signal recognition network is used as a signal recognition modulator.

[0074] S5. Use the preprocessed 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.

[0075] In a specific embodiment, the present application is verified by data simulation. The received signal data set adopts the RML2018.01 modulated signal data set. Assume that the signal data set has 2555904 items, including 24 modulation modes such as BPSK, QPSK, 8PSK, FM, GMSK, QAM16, QAM64, etc., and each modulation mode has 106496 modulation signals. Assume that each modulation of the received signal data set contains 26 signal-to-noise ratios, and each signal-to-noise ratio contains 4096 modulation signals. The size of each modulated signal is (2,1024), where 2 corresponds to I / Q two-way modulation signals and 1024 corresponds to 1024 sampling points. In this embodiment, it is assumed that 90% of the received signal data set is used as a training signal data set for training and learning the signal recognition network, and 10% of the received signal data set is used as a test signal data set for testing the performance of the signal recognition network.

[0076] In the signal preprocessing of this embodiment, it is assumed that the time-frequency transform STFT parameters are: the number of samples is 256, the overlap length is 127, the window function type is Hamming, the window function length is 128, the FFT length is 128, and the output signal two-dimensional data matrix dimension is 128×128.

[0077] Use Figure 5 The 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 constructed with the following settings: The input data dimension of the upper branch decoupled dynamic filter convolution network is 16×128×128, which is obtained by copying the signal two-dimensional data 16 times. It contains 3 decoupled dynamic filter convolution blocks. The network structure of each decoupled dynamic filter convolution block is the same, and the network structure parameter configuration is dynamically adjustable. The channel dimension, frequency domain dimension and time domain dimension of the first, second and third decoupled dynamic filter convolution blocks are 16×128×128, 16×32×32 and 64×8×8 respectively. The upper branch decoupled dynamic filter convolution network contains 3 pooling layers. The first pooling layer contains 16 features, the data dimension is 64×64, and the average value pooling is used for extraction, and the extraction ratio is 2; the second pooling layer contains 32 features, the data dimension is 16×16, and the average value pooling is used for extraction, and the extraction ratio is 2; the third pooling layer contains 64 features, the data dimension is 1×1, and the average value pooling is used for extraction, and the extraction ratio is 8. The number of batch normalization processing samples of the decoupled dynamic filter convolution block of the upper branch decoupled dynamic filter convolution network is 512.

[0078] The lower branch attention gated recurrent convolution network contains 3 attention gated recurrent convolution blocks. The network structure of each attention gated recurrent convolution block is the same, and the network structure parameter configuration is dynamically adjustable. The filter data dimensions of the first, second, and third attention gated recurrent convolution blocks are 16×1024, 32×256, and 64×32, respectively. The lower branch attention gated recurrent convolution network contains 3 pooling layers. The first pooling layer contains 16 features, the data dimension is 1×512, and the average value pooling is used for extraction, and the extraction ratio is 2; the second pooling layer contains 32 features, the data dimension is 1×128, and the average value pooling is used for extraction, and the extraction ratio is 2; the third pooling layer contains 128 features, the data dimension is 1×1, and the average value pooling is used for extraction, and the extraction ratio is 32.

[0079] 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 24 modulation type signal recognition results.

[0080] The recognition rates of different modulation types under different signal-to-noise ratios were tested. Figures 6 (a)-6 (d) show the confusion matrix for the recognition of 24 modulation signals under different noise scenarios with signal-to-noise ratios of 2, 4, 6, and 8. It can be seen intuitively from the figure that when the signal-to-noise ratio is 8dB, most modulation types have a high recognition rate, and only the recognition of 128QAM, 128APSK, and 64QAM is poor. This is because these signals are similar and difficult to recognize. When the signal-to-noise ratio drops to 6dB, it can be found that the recognition rates of most modulation types are still at a high level, but as the signal-to-noise ratio further decreases (that is, the noise increases), the diagonal of the confusion matrix becomes more and more blurred, and the recognition rates of various modulation methods decrease.

[0081] The prior art has difficulty in demodulating signals for most high-order modulation types in signal data sets even when the signal-to-noise ratio is 8dB. The signal recognition network constructed in the present application can perform non-cooperative recognition of most types of modulated signals under an 8dB signal-to-noise ratio. The reason is that the signal recognition network combines the advantages of the AGRUN network that can extract global features of signal time accumulation, the DDFN network that can extract content-adaptive multi-dimensional local features and multi-domain feature fusion, and fully exploits the time and content correlation of the modulated signal, solving the problems of weak network generalization ability and decreased accuracy in recognizing multi-mode modulated signals due to irrelevant content of convolution operations and small convolution perception field of view. This reflects the recognition robustness and generalization ability of the signal recognition network of the present application.

[0082] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A signal modulation identification method based on AGRUN-DDFN, characterized in that: The following steps are involved: S1. Acquire a received signal data set by signal reception sampling, and divide the received signal data set into a training signal data set and a test signal data set; S2, preprocessing the received signal data set to obtain a signal one-dimensional data set and a signal two-dimensional data set; S3. Construct a signal recognition network, which includes a serially cascaded feature extraction network, a fusion layer, a connection layer, and an output layer. The feature extraction network includes two upper and lower branches of a decoupled dynamic filtering convolutional network and an attention gated recurrent convolutional network that are cascaded in parallel. S4, using the preprocessed training signal data set to perform deep learning training on the signal recognition network, and the trained signal recognition network is used as a signal recognition modulator; S5. Use the preprocessed 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.

2. The signal modulation identification method based on AGRUN-DDFN according to claim 1, characterized in that: The preprocessing of the received signal data set comprises the following steps: Perform background noise adaptive cancellation on the received signal data set to generate a signal data set ; For the signal dataset Normalization is performed to obtain a one-dimensional data set of signals , the signal normalization calculation formula is as follows: in, It means taking the mean value, Indicates the standard deviation; For a one-dimensional data set of signals Using short-time Fourier transform, we can get a two-dimensional time-frequency diagram : in, and is the time variable, is the frequency variable, The length is Sliding window function of For two-dimensional time-frequency diagram Normalization is performed to obtain a two-dimensional signal data set .

3. The signal modulation identification method based on AGRUN-DDFN according to claim 2 is characterized in that: The upper branch of the feature extraction network is a decoupled dynamic filter convolution network, which includes a plurality of decoupled dynamic filter convolution blocks, each of which is connected to a pooling layer, and the last decoupled dynamic filter convolution block is connected to the fusion layer after passing through the pooling layer; The lower branch of the feature extraction network is an attention gated recurrent convolutional network, which includes several attention gated recurrent convolutional blocks, each of which is connected to a pooling layer, and the last attention gated recurrent convolutional block is connected to the fusion layer after passing through the pooling layer.

4. The signal modulation identification method based on AGRUN-DDFN according to claim 3 is characterized in that: The pooling method of the pooling layer is average pooling.

5. The signal modulation identification method based on AGRUN-DDFN according to claim 4 is characterized in that: The input of the decoupled dynamic filtering convolutional network is a two-dimensional data set of signals , the input of the attention gated recurrent convolutional network is a one-dimensional data set of signals .

6. The signal modulation identification method based on AGRUN-DDFN according to claim 5, characterized in that: The decoupled dynamic filtering convolution block comprises a first convolution layer, a decoupled dynamic filtering block and a second convolution layer which are serially cascaded in sequence, and the decoupled dynamic filtering convolution block adopts a shortcut connection; The data processing of the first convolutional layer and the second convolutional layer both include 1×1 convolution operation, batch normalization processing and activation function mapping processing in sequence; The data processing of the decoupled dynamic filtering block includes decoupled dynamic filtering, batch normalization processing and activation function mapping processing in sequence.

7. The signal modulation identification method based on AGRUN-DDFN according to claim 6, characterized in that: 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; The spatial filtering parameter construction includes the following steps: Assume that the input data have channels, and their 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: right The length of each spatial position is The vector is normalized, that is: in, and are learnable parameters, The frequency position is and the time position is The spatial filtering parameters output at ; Will Copy along channel dimension Arrival , after regularization, we get the spatial filtering parameters ; The channel filter parameter construction includes the following steps: For input data The average value of each channel is , and then use two Convolution pair After processing, the channel branch parameters are finally obtained. ,two Convolution is represented as and , The value range is A constant in the interval, that is: in, For the channel The frequency position is , time position is Input data at Will Copy along the second dimension times, and then replicate along the third dimension. Get , after regularization, the channel filter parameters are obtained ; The spatial channel filter parameter fusion includes the following steps: The obtained spatial filtering parameters and channel filter parameters Multiply point by point to get , and then The first dimension of is decomposed into , and The three new dimensions give the final ; The kernel parameters include the following steps: 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: 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 The input data.

8. The signal modulation identification method based on AGRUN-DDFN according to claim 7, characterized in that: The attention gated recurrent convolution block includes a gated recurrent unit module and an attention transformation module which are serially cascaded in sequence.

9. The signal modulation identification method based on AGRUN-DDFN according to claim 8, characterized in that: The attention transformation module is used to mine the intrinsic relationship between the feature vectors extracted by the gated recurrent unit module. The network training formula of the attention transformation module is as follows: 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.

10. The signal modulation identification method based on AGRUN-DDFN according to claim 9, characterized in that: The fusion layer will dimensional local eigenvectors and The local feature vector of the dimension is expanded by concatenating the head and tail The activation function used in the connection layer is the ReLU function, and the output layer uses the Softmax function for multi-classification output.

Images