Automatic modulation identification method based on multi-feature lightweight network

By adopting a multi-feature lightweight network in automatic modulation recognition, combined with deep separable convolution and gated recursive unit modules, the problem of increasing parameter quantity and calculation complexity in the existing technology is solved, and the resource-constrained environment is realized The efficient detection performance is improved, achieving a high recognition accuracy rate.

CN119071119AInactive Publication Date: 2024-12-03ZHENGZHOU UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411015127.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-12-03
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

While the existing automatic modulation identification network improves performance, the parameter quantity and calculation complexity increase, resulting in unusability in resource-constrained environments, and lightweight networks still need to improve detection performance.

Method used

Using multi-feature lightweight network, the convolution and gated recursive unit modules can be separated by deep separation, combined with batch normalization layer and Dropout layer, build a network for automatic modulation recognition, reducing the amount of parameters and calculation complexity, and improving the Detection performance.

Benefits of technology

With low parameter quantity and calculation complexity, better detection performance was achieved, reaching an identification accuracy of 62.92%, and 64.95% identification accuracy on the extended data set, significantly better than other networks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119071119A_ABST
    Figure CN119071119A_ABST
Patent Text Reader

Abstract

The invention provides an automatic modulation identification method based on a multi-feature lightweight network. The method comprises the following steps: acquiring I / Q signal samples from public radio machine learning data sets RML2016.10a and RML2016.10b; according to the obtained I / Q signal samples, instantaneous amplitude and instantaneous phase samples are extracted, and real part and imaginary part samples are obtained after FFT operation is carried out on the I / Q signals; constructing a multi-feature lightweight learning network for automatic modulation; dividing the signal samples into a training set, a verification set and a test set, and randomly selecting samples of each modulation mode under each signal-to-noise ratio for training; and verifying the trained network by using a test set, and checking the modulation recognition effect. According to the method, the parameter quantity and the calculation quantity of the network are relatively low, and meanwhile, relatively high modulation identification accuracy is obtained.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of military communication technology, and in particular to an automatic modulation recognition method based on a multi-feature lightweight network. Background Art

[0002] In wireless communication systems, the use of Automatic Modulation Recognition (AMR) is crucial for quickly identifying the modulation mode of the received signal. In the military field, AMR is used for electronic support, electronic reconnaissance, etc. In the civilian field, AMR is used for spectrum sensing, etc. Traditional AMR methods can be divided into two categories: Maximum Likelihood-Based (LB) methods and Feature-Based (FB) methods. The LB method can achieve the theoretically optimal recognition effect under known noise and channel parameters, but non-cooperative communications often lack prior information. The FB method classifies by manually extracting features. The classification effect depends largely on the effect of feature extraction, and professionals are required to extract features.

[0003] In recent years, deep learning (DL) has been applied in automatic modulation recognition and has better results than traditional AMR methods.

[0004] Although there are many types of AMR networks now, and they have good performance. However, while improving performance, many networks increase the number of parameters and computational complexity, and cannot be used in some resource-constrained environments. In addition, although some lightweight networks reduce computational complexity, their detection performance still needs to be improved. Therefore, this paper proposes a lightweight network based on multi-channel (A Multi-channel Deepthwise Separable Convolution and Gated Recurrent Unit Network, MSGNet), which achieves good detection performance with lower parameter amount and computational complexity. Summary of the invention

[0005] In view of the defects in the prior art, the present invention proposes an automatic modulation recognition method based on a multi-feature lightweight network, which achieves better detection performance with lower parameter quantity and computational complexity.

[0006] In a first aspect, the present invention provides an automatic modulation recognition method based on a multi-feature lightweight network, the method comprising:

[0007] S1: Get I / Q signal samples from the public radio machine learning datasets RML2016.10a and RML2016.10b;

[0008] S2: extracting instantaneous amplitude and instantaneous phase samples and real and imaginary samples obtained by performing FFT operation on the I / Q signal according to the acquired I / Q signal samples;

[0009] S3: Building a multi-feature lightweight learning network for automatic modulation;

[0010] S301: The network inputs data from three channels, the input data including I / Q data, A / P data and R / I data;

[0011] S302: The network uses deep separable convolution and GRU modules with small number of parameters and low computational complexity;

[0012] S303: The network adds BN layer and Dropout layer to prevent overfitting and speed up training;

[0013] S4: Divide the signal samples into a training set, a validation set, and a test set, and randomly select samples of each modulation mode at each signal-to-noise ratio for training;

[0014] S5: Use the test set to verify the trained network and check the modulation recognition effect.

[0015] The step S1 specifically includes:

[0016] The RML2016.10a dataset contains 11 modulation modes, including 8 digital modulation modes BPSK, QPSK, 8PSK, QAM16, QAM64, CPFSK, GFSK, PAM4 and 3 analog modulation modes AM-SSB, AM-DSB, WBFM. The signal-to-noise ratio range of each modulation mode is -20dB to 18dB. There are 1000 samples at each signal-to-noise ratio, and a total of 220,000 samples.

[0017] The RML2016.10b dataset is an extended version of the RML2016.10a dataset with a larger number of samples. It contains 10 modulation modes except AM-SSB. The signal-to-noise ratio range is also -20dB to 18dB. Each modulation mode has 6000 samples at each signal-to-noise ratio, and a total of 1,200,000 samples.

[0018] Each sample of the two data sets includes 128 sampling points, with sampling frequency deviation, center frequency offset, additive Gaussian white noise, etc.

[0019] The step S2 specifically includes:

[0020] Each sample in the data set includes 128 sampling points, and each sampling point r(n) is represented by two signals, I and Q:

[0021] r(n)=[r I (n),r Q (n)]

[0022] Among them, r I (n) represents the in-phase sample signal, r Q (n) represents the orthogonal sample signal;

[0023] In order to more comprehensively extract the hidden features of the I / Q signal, we extract the instantaneous amplitude features and the instantaneous phase features of the I / Q signal:

[0024]

[0025] Among them, A(n) represents the instantaneous amplitude, P(n) represents the instantaneous phase, and we use A / P samples to represent it; in addition, since I / Q and A / P are both time domain features, in order to further extract frequency domain features, we perform FFT operation on the baseband signal to extract the real part and imaginary part:

[0026]

[0027] Where F1 represents the real part of the sample, F Q Represents the imaginary part of the sample, which we represent using R / I samples.

[0028] The step S3 specifically includes:

[0029] The proposed network is a lightweight network that uses deep separable convolution and GRU with small number of parameters and low computational complexity. The deep separable convolution layer can reduce the number of parameters while extracting the spatial and channel correlation of the input data at the same time. Compared with LSTM, GRU has a simpler network structure, fewer parameters, faster calculation speed, and is comparable to LSTM in processing long-term dependency tasks. Then, we add BN layer and Dropout layer to prevent overfitting, and finally classify through the fully connected layer. The network has three channels of input data, including I / Q data, A / P data and R / I data. I / Q data and R / I data are first batch normalized through the BN layer, and then the spatial features are extracted through the 1*8 convolution kernel. and frequency domain features; the A / P data is divided into two paths, which are fused in the channel dimension after passing through a one-dimensional convolution layer, and then the time domain features are extracted through GRU; we use the BN layer to normalize the features extracted by GRU, and then further extract the spatial features through a 1*8 convolution layer; then, we fuse the features extracted from the three channels, and use a 5*5 convolution layer and a GRU layer containing 128 cells for further feature extraction; finally, two fully connected layers with 128 and 11 neurons respectively are used for classification, and Dropout is added after each fully connected layer to prevent overfitting. The activation function of the first fully connected layer uses selu, and the activation function of the second fully connected layer uses softmax. The output result of the second fully connected layer is the probability of each modulation type.

[0030] The step S4 specifically includes:

[0031] The ratio of the data set, validation set, and test set is set to 6:2:2, and samples of each modulation mode at each signal-to-noise ratio are randomly selected;

[0032] We use the cross entropy loss function to verify the debugging recognition classification loss, and the descent algorithm uses the Adam optimizer. During the training process, we introduced the early stopping training strategy, and the initial learning rate was set to 0.001. If the training loss of the validation set did not decrease in 5 epochs, the learning rate was multiplied by 0.5. If the training loss of the validation set did not decrease for 50 consecutive epochs, we stopped training and recorded the network weights when the training loss was the smallest. We set Dropout to 0.5 and batch_size to 400. All experimental modules are built in the Keras library, supported by NVIDIA CUDA and GeForce GTX 1650 GPU.

[0033] The step S5 specifically includes:

[0034] In the training of the RML2016.10a dataset, convergence was achieved around 110 rounds of training, and the maximum accuracy reached 62.92%; in the training of the RML2016.10b dataset, the convergence speed was faster than that of the RML2016.10a dataset, and it was close to stability at 90 rounds, with a maximum recognition accuracy of 64.95%.

[0035] It can be seen from the above technical scheme that the present invention provides an automatic modulation recognition method based on a multi-feature lightweight network. First, the network adopts a lightweight module: a deep separable convolution and a gated recurrent unit, wherein the deep separable convolution layer can achieve the spatial and channel correlation of the input data while reducing the amount of parameters, while the GRU is simpler in network structure, has fewer parameters, and is faster in calculation than the LSTM, and is comparable to the LSTM in processing long-term dependency tasks. Secondly, the module inputs separated amplitude / phase samples, in-phase / orthogonal phase samples from three input data channels, and real / imaginary samples obtained by Fourier transforming the baseband signal, and obtains the signal space, time and frequency characteristics. Finally, in order to speed up the training speed of the network and prevent overfitting, a batch normalization layer and a Dropout algorithm are added to the network. The simulation results on the benchmark data set show that the network has low parameter and calculation amounts and has good modulation recognition performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 It is the depth convolution process in depth-separable convolution;

[0038] Figure 2 It is the point-by-point convolution process in the depth-wise separable convolution;

[0039] Figure 3 This is the GRU implementation block diagram;

[0040] Figure 4 It is the MSGNet network;

[0041] Figure 5is the training loss and average recognition accuracy of the training set and validation set, where (a) is the average recognition accuracy of the training set and validation set on the RML2016.10a dataset; (b) is the training loss of the training set and validation set on the RML2016.10a dataset; (c) is the average recognition accuracy of the training set and validation set on the RML2016.10b dataset; (d) is the training loss of the training set and validation set on the RML2016.10b dataset;

[0042] Figure 6 is the recognition accuracy of various networks on the RML2016.10a dataset;

[0043] Figure 7 is the recognition accuracy of various networks on the RML2016.10b dataset;

[0044] Figure 8 is the confusion matrix of the network at 0dB and 8dB;

[0045] Fig. 9 It is a flow chart of an automatic modulation recognition method based on a multi-feature lightweight network provided by the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] like Fig. 9 As shown, an automatic modulation recognition method based on a multi-feature lightweight network is provided in one embodiment of the present invention, and the method comprises the following steps:

[0048] S1: Get I / Q signal samples from the public radio machine learning datasets RML2016.10a and RML2016.10b;

[0049] S2: extracting instantaneous amplitude and instantaneous phase samples and real and imaginary samples obtained by performing FFT operation on the I / Q signal according to the acquired I / Q signal samples;

[0050] S3: Building a multi-feature lightweight learning network for automatic modulation;

[0051] S301: The network inputs data from three channels, the input data including I / Q data, A / P data and R / I data;

[0052] S302: The network uses deep separable convolution and GRU modules with small number of parameters and low computational complexity;

[0053] S303: The network adds BN layer and Dropout layer to prevent overfitting and speed up training;

[0054] S4: Divide the signal samples into a training set, a validation set, and a test set, and randomly select samples of each modulation mode at each signal-to-noise ratio for training;

[0055] S5: Use the test set to verify the trained network and check the modulation recognition effect.

[0056] In this embodiment, the specific process of step S1 is as follows:

[0057] The RML2016.10a dataset contains 11 modulation modes, including 8 digital modulation modes BPSK, QPSK, 8PSK, QAM16, QAM64, CPFSK, GFSK, PAM4 and 3 analog modulation modes AM-SSB, AM-DSB, WBFM. The signal-to-noise ratio range of each modulation mode is -20dB to 18dB. There are 1000 samples at each signal-to-noise ratio, and a total of 220,000 samples.

[0058] The RML2016.10b dataset is an extended version of the RML2016.10a dataset with a larger number of samples. It contains 10 modulation modes except AM-SSB. The signal-to-noise ratio range is also -20dB to 18dB. Each modulation mode has 6000 samples at each signal-to-noise ratio, and a total of 1,200,000 samples.

[0059] Each sample of the two data sets includes 128 sampling points, with sampling frequency deviation, center frequency offset, additive Gaussian white noise, etc.

[0060] In this embodiment, the specific process of step S2 is as follows:

[0061] S21: Obtain I / Q signal samples.

[0062] Each sample in the data set includes 128 sampling points, and each sampling point r(n) is represented by two signals, I and Q:

[0063] r(n)=[r I (n),r Q (n)] (1)

[0064] Among them, r I (n) represents the in-phase sample signal, r Q (n) represents the orthogonal sample signal.

[0065] S22: Get A / P signal samples

[0066] In order to more comprehensively extract the hidden features of the I / Q signal, we extract the instantaneous amplitude features and the instantaneous phase features of the I / Q signal:

[0067]

[0068] Among them, A(n) represents the instantaneous amplitude, P(n) represents the instantaneous phase, and we use A / P samples to represent it.

[0069] S23: Obtain R / I signal samples.

[0070] In addition, since I / Q and A / P are both time domain features, in order to further extract frequency domain features, we perform FFT operation on the baseband signal to extract the real and imaginary parts:

[0071]

[0072] Where F1 represents the real part of the sample, F Q Represents the imaginary part of the sample, which we represent using R / I samples.

[0073] In this embodiment, the specific process of step S3 is as follows:

[0074] The proposed network is a lightweight network that uses deep separable convolution and GRU with small number of parameters and low computational complexity. The deep separable convolution layer can reduce the number of parameters and extract the spatial and channel correlation of the input data at the same time. Compared with LSTM, GRU has a simpler network structure, fewer parameters, faster calculation speed, and comparable effect to LSTM when processing long-term dependency tasks. Then, we add BN layer and Dropout layer to prevent overfitting, and finally classify through the fully connected layer. The network has three channels of input data, including I / Q data, A / P data and R / I data. Among them, I / Q data and R / I data are first batch normalized by BN layer, and then spatial features and frequency domain features are extracted by 1*8 convolution kernel. A / P data is divided into two paths, which are fused in the channel dimension after passing through one-dimensional convolution layer, and then time domain features are extracted by GRU. We use BN layer to normalize the features extracted by GRU, and then further extract spatial features through a 1*8 convolution layer. Next, we fuse the features extracted from the three channels and use them for further feature extraction through a 5*5 convolutional layer and a GRU layer containing 128 cells; finally, we use two fully connected layers with 128 and 11 neurons respectively for classification, and add Dropout after each fully connected layer to prevent overfitting. The activation function of the first fully connected layer is selu, and the activation function of the second fully connected layer is softmax. The output result of the second fully connected layer is the probability of each modulation type.

[0075] Figure 1 It is the deep convolution process in the depthwise separable convolution. The deep convolution stage performs convolution operations on each channel of the input tensor independently with a single channel as the unit. When the input tensor has C input channels, the deep convolution operation will use C different convolution kernels to process each channel separately and generate C intermediate feature maps.

[0076] Figure 2 It is the point-by-point convolution process in the depth-wise separable convolution. The point-by-point convolution stage uses a 1x1 convolution kernel to perform point-by-point dot product operations on the intermediate feature map generated in the depth convolution stage. This 1x1 convolution kernel can reduce the number of channels of the feature map from C to a smaller value to generate the final output feature map.

[0077] Figure 3 This is the implementation process of GRU. The GRU structure consists of a reset gate, an update gate, a candidate hidden state, and a hidden state. In the figure, x t Indicates the current input information, h t-1 is the previous hidden state, r t represents the reset gate, z tis defined as the update gate, we set the tanh function, means the candidate hidden state, h t It is the hidden state passed to the next moment.

[0078] The specific implementation process is: According to the formula, we can get the reset gate r t The result is that the formula is to calculate the update gate z t The process of candidate hidden states According to the formula, t , z t and We can finally get the hidden state h at the next moment according to the formula t , the state update formula is:

[0079] r t =σ(W xr x t +W hr h t-1 +b r ) (4)

[0080] z t =σ(W xz x t +W hz h t-1 +b z ) (5)

[0081]

[0082] Figure 4 It is the MSGNet network framework. The network inputs data from three channels. The input data includes I / Q data, A / P data and R / I data, which are obtained by step S2, and then pass through a series of sub-modules. Finally, the output result of the second fully connected layer is the probability of each modulation type.

[0083] In this embodiment, step S4 specifically includes:

[0084] The ratio of the data set, validation set, and test set is set to 6:2:2, and samples of each modulation mode at each signal-to-noise ratio are randomly selected;

[0085] We use the cross entropy loss function to verify the debugging recognition classification loss. The descent algorithm uses the Adam optimizer. During the training process, we introduce the early stopping training strategy and set the initial learning rate to 0.001. If the training loss of the validation set does not decrease in 5 epochs, the learning rate is multiplied by 0.5. If the training loss of the validation set does not decrease for 50 consecutive epochs, we stop training and record the network weights when the training loss is the smallest. We set Dropout to 0.5 and batch_size to 400. All experimental modules are built in the Keras library and supported by NVIDIA CUDA and GeForce GTX 1650 GPU.

[0086] In this embodiment, step S5 specifically includes:

[0087] In the training of the RML2016.10a dataset, convergence was achieved around 110 rounds of training, and the maximum accuracy reached 62.92%; in the training of the RML2016.10b dataset, the convergence speed was faster than that of the RML2016.10a dataset, and it was close to stability at 90 rounds, with a maximum recognition accuracy of 64.95%.

[0088] Figure 5 are the training set and validation set losses and recognition accuracy. Figure (a) and (b) are the training results of the RML2016.10a dataset. We can see that convergence is achieved around 110 rounds of training, and the maximum accuracy reaches 62.92%. Figures (c) and (d) are the training results of the RML2016.10b dataset. Compared with the RML2016.10a dataset, the convergence speed is faster, and it is close to stability after 90 rounds, and the maximum recognition accuracy is 64.95%.

[0089] Figure 6 This is the recognition accuracy of various networks in the RML2016.10a dataset. It can be seen that the network we proposed is significantly better than other networks above -8dB, with the highest accuracy of 93.4% at 14dB. The average recognition accuracy of MSGNet in 0-18dB reaches 92.418%, which is 1%-11% higher than other networks.

[0090] Figure 7 It is the recognition accuracy of various networks under the RML2016.10b dataset, which is to verify the robustness of the network. The results are shown in the figure. It can be seen that the accuracy of MSGNet is still better than other networks.

[0091] Figure 8It is the confusion matrix of MSGNet network at 0dB and 8dB. As shown in the figure, the recognition effect of most modulation modes is good. However, WBFM and AM-DSB still have great confusion. This is because both are generated by sampling analog audio signals, in which there are silent periods.

[0092] Table I shows the accuracy of each modulation type at 0 dB in the dataset RML2016.10a. We can see that the MSGNet network achieves the maximum recognition accuracy under most modulation types, including BPSK, CPFSK, GFSK, 4PAM, 16-QAM, 64-QAM, and QPSK. Among them, there is a significant improvement in QPSK recognition, which is 5% higher than the IC-AMCNet network with the highest recognition accuracy among other networks.

[0093]

[0094] Table II is to verify the role of each module of the MSGNet network. We conducted an ablation experiment on the RML2016.10a dataset; the sub-networks are MSGNet-A, MSGNet-B, MSGNet-C, and MSGNet-D. The figure shows the accuracy of each network under various signal-to-noise ratios, and it can be seen that the effect of the MSGNet network is better than other networks.

[0095]

[0096] Table III compares the complexity of the MSGNet network with other networks. We compare them in terms of parameter quantity, computational complexity, 0-18dB average recognition accuracy, and average recognition accuracy, as shown in the table. It can be seen that the MSGNet network is only slightly higher than DAE and PET-CGDNN in terms of parameter quantity and computational complexity, but its average recognition accuracy is significantly higher than theirs. Compared with the MCLDNN network with the highest recognition rate, the MSGNet network has significantly lower parameter quantity and computational complexity, and the average recognition accuracy of our proposed network is 1.1% higher than that of the MCLDNN network.

[0097]

[0098] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. An automatic modulation recognition method based on a multi-feature lightweight network, characterized in that: The method comprises: S1: Get I / Q signal samples from the public radio machine learning datasets RML2016.10a and RML2016.10b; S2: extracting instantaneous amplitude and instantaneous phase samples and real and imaginary samples obtained by performing FFT operation on the I / Q signal according to the acquired I / Q signal samples; S3: Building a multi-feature lightweight learning network for automatic modulation; S301: The network inputs data from three channels, the input data including I / Q data, A / P data and R / I data; S302: The network uses deep separable convolution and GRU modules with small number of parameters and low computational complexity; S303: The network adds BN layer and Dropout layer to prevent overfitting and speed up training; S4: Divide the signal samples into a training set, a validation set, and a test set, and randomly select samples of each modulation mode at each signal-to-noise ratio for training; S5: Use the test set to verify the trained network and check the modulation recognition effect.

2. The automatic modulation recognition method based on a multi-feature lightweight network according to claim 1 is characterized in that: The step S1 specifically includes: The RML2016.10a dataset contains 11 modulation modes, including 8 digital modulation modes BPSK, QPSK, 8PSK, QAM16, QAM64, CPFSK, GFSK, PAM4 and 3 analog modulation modes AM-SSB, AM-DSB, WBFM. The signal-to-noise ratio range of each modulation mode is -20dB to 18dB. There are 1000 samples at each signal-to-noise ratio, and a total of 220,000 samples. The RML2016.10b dataset is an extended version of the RML2016.10a dataset with a larger number of samples. It contains 10 modulation modes except AM-SSB. The signal-to-noise ratio range is also -20dB to 18dB. Each modulation mode has 6000 samples at each signal-to-noise ratio, and a total of 1,200,000 samples. Each sample of the two data sets includes 128 sampling points, with sampling frequency deviation, center frequency offset, additive Gaussian white noise, etc.

3. The automatic modulation recognition method based on a multi-feature lightweight network according to claim 1 is characterized in that: The step S2 specifically includes: Each sample in the data set includes 128 sampling points, and each sampling point r(n) is represented by two signals, I and Q: r(n)=[r I (n),r Q (n)] Among them, r I (n) represents the in-phase sample signal, r Q (n) represents the orthogonal sample signal; In order to more comprehensively extract the hidden features of the I / Q signal, we extract the instantaneous amplitude features and the instantaneous phase features of the I / Q signal: Among them, A(n) represents the instantaneous amplitude, P(n) represents the instantaneous phase, and we use A / P samples to represent it; in addition, since I / Q and A / P are both time domain features, in order to further extract frequency domain features, we perform FFT operation on the baseband signal to extract the real part and imaginary part: Where F1 represents the real part of the sample, F Q Represents the imaginary part of the sample, which we represent using R / I samples.

4. The automatic modulation recognition method based on a multi-feature lightweight network according to claim 1 is characterized in that: The step S3 specifically includes: The proposed network is a lightweight network that uses deep separable convolution and GRU with small number of parameters and low computational complexity. The deep separable convolution layer can reduce the number of parameters while extracting the spatial and channel correlation of the input data at the same time. Compared with LSTM, GRU has a simpler network structure, fewer parameters, faster calculation speed, and is comparable to LSTM in processing long-term dependency tasks. Then, we add BN layer and Dropout layer to prevent overfitting, and finally classify through the fully connected layer. The network has three channels of input data, including I / Q data, A / P data and R / I data. I / Q data and R / I data are first batch normalized through the BN layer, and then the spatial features are extracted through the 1*8 convolution kernel. and frequency domain features; the A / P data is divided into two paths, which are fused in the channel dimension after passing through a one-dimensional convolution layer, and then the time domain features are extracted through GRU; we use the BN layer to normalize the features extracted by GRU, and then further extract the spatial features through a 1*8 convolution layer; then, we fuse the features extracted from the three channels, and use a 5*5 convolution layer and a GRU layer containing 128 cells for further feature extraction; finally, two fully connected layers with 128 and 11 neurons respectively are used for classification, and Dropout is added after each fully connected layer to prevent overfitting. The activation function of the first fully connected layer uses selu, and the activation function of the second fully connected layer uses softmax. The output result of the second fully connected layer is the probability of each modulation type.

5. The automatic modulation recognition method based on a multi-feature lightweight network according to claim 1 is characterized in that: The step S4 specifically includes: The ratio of the data set, validation set, and test set is set to 6:2:2, and samples of each modulation mode at each signal-to-noise ratio are randomly selected; We use the cross entropy loss function to verify the debugging recognition classification loss. The descent algorithm uses the Adam optimizer. During the training process, we introduce the early stopping training strategy and set the initial learning rate to 0.

001. If the training loss of the validation set does not decrease in 5 epochs, the learning rate is multiplied by 0.

5. If the training loss of the validation set does not decrease for 50 consecutive epochs, we stop training and record the network weights when the training loss is the smallest. We set Dropout to 0.5 and batch_size to 400. All experimental modules are built in the Keras library and supported by NVIDIA CUDA and GeForce GTX1650 GPU.

6. The automatic modulation recognition method based on a multi-feature lightweight network according to claim 1 is characterized in that: The step S5 specifically includes: In the training of the RML2016a dataset, convergence was achieved around 110 rounds of training, and the maximum accuracy reached 62.92%; in the training of the RML2016b dataset, the convergence speed was faster than that of the RML2016a dataset, and it was close to stability at 90 rounds, with a maximum recognition accuracy of 64.95%.

Citation Information

Patent Citations

  • Automatic modulation classification method based on deep learning network fusion

    CN115563468A

  • Lightweight automatic modulation identification method based on multi-channel fusion

    CN118337576A