Automatic Signal Modulation Recognition Method

Abstract

The present invention discloses a method for automatic signal modulation recognition, comprising the following steps: step 1, obtaining a target signal, wherein the target signal is a signal to be modulated and identified; step 2, obtaining I / Q dual-channel data and A / P dual-channel data of the target signal, separating the I / Q dual-channel data and the A / P dual-channel data respectively to obtain independent I single-channel data, Q single-channel data, A single-channel data, and P single-channel data; step 3, building a symmetrical multi-channel deep learning system, constructing a data set, and training the built symmetrical multi-channel deep learning system to obtain a trained symmetrical multi-channel deep learning system; step 4, inputting a new signal to be modulated and identified into the trained symmetrical multi-channel deep learning system to determine the signal modulation recognition mode of the target signal. The present invention improves the recognition accuracy of the automatic modulation recognition technology based on the deep learning method, and has high recognition performance in distinguishing modulation modes.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic modulation recognition, and more particularly to a method for automatic signal modulation recognition. Background Art

[0002] Monitoring complex communication environments has become a critical issue, particularly identifying the modulation patterns of communication signals. Automatic Modulation Recognition (AMR) technology aims to quickly and reliably identify the modulation patterns of communication signals. This technology has a wide range of applications, particularly in non-cooperative communication scenarios.

[0003] Generally speaking, AMR technology can be divided into two traditional approaches: likelihood-based methods and feature learning-based methods. However, both approaches rely heavily on prior knowledge. In recent years, with the development of intelligent algorithms and computing hardware, a third category of AMR technology, based on deep learning, has demonstrated promising performance and application prospects.

[0004] In the past, several deep learning frameworks have been directly applied to AMR technology. For example, using convolutional neural networks (CNNs) to directly learn the signal's I / Q sequence has achieved better results than expert-based feature learning in certain aspects. Furthermore, recurrent neural networks (RNNs), which can fully learn temporal features, have also been widely used, and an improved version of the signal, the long short-term memory (LSTM) model, has been introduced. While these frameworks have achieved promising results, they have significant limitations, as they can only learn a single aspect of the signal's spatiotemporal characteristics. Based on this concept, the convolutional long short-term deep neural network (CLDNN) was proposed. Further research has led to the exploration of connections between different signal representations, such as I / Q channels, amplitude and phase, and even separate channel features. Summary of the Invention

[0005] Technical problem: In view of the insufficiency of existing AMR technology in learning signal features, the present invention provides a method for automatic signal modulation recognition, which fully learns various features of the signal and improves the performance and accuracy of signal modulation mode recognition.

[0006] Technical solution: In order to achieve the above-mentioned purpose of the invention, the technical solution provided by the present invention is: a signal automatic modulation recognition method, including two multi-channel feature mapping modules and an overall feature network classifier module.

[0007] The two multi-channel feature mapping modules, such as Figure 1 As shown, they are symmetrical to each other and are divided into Part A module and Part B module.

[0008] The Part A module includes a three-channel input, convolutional layers CONV1, CONV2, CONV3, first LSTM layers LSTM1 and LSTM2, a second LSTM layer LSTM5, and two identical Gaussian Dropout layers. The three-channel input of the Part A module respectively inputs input signal I / Q multi-channel data, input signal A (amplitude) single-channel data, and input signal P (phase) single-channel data. The three-channel input is connected to the input end of the convolutional layer, the output end of the I / Q multi-channel data is connected to the input end of the CONV1 layer, the output end of the A single-channel data is connected to the input end of the CONV2 layer, and the output end of the P single-channel data is connected to the input end of the CONV3 layer. The output of the convolutional layer is connected to the input end of the first LSTM layer after merging, the output of the CONV1 layer and the output of the CONV2 layer are connected to the input end of the LSTM1 layer after merging, and the output of the CONV1 layer and the output of the CONV3 layer are connected to the input end of the LSTM2 layer after merging. The outputs of the first LSTM layer are connected to GaussianDropout layers. The output of the LSTM1 layer is connected to the input of the GaussianDropout layer, and the output of the LSTM2 layer is connected to the input of another identical GaussianDropout layer. The outputs of the two GaussianDropout layers are combined and connected to the input of the second LSTM layer, LSTM5. The output of the LSTM5 layer serves as the output of the Part A module.

[0009] The Part B module also includes a three-channel input, convolutional layers CONV6, CONV4, CONV5, first LSTM layers LSTM3 and LSTM4, a second LSTM layer LSTM6, and two identical Gaussian Dropout layers. The three-channel input of the Part B module respectively inputs input signal A / P multi-channel data, input signal I single-channel data, and input signal Q single-channel data. The three-channel inputs are respectively connected to the input end of the convolutional layer, the output end of the A / P multi-channel data is connected to the input end of the CONV6 layer, the output end of the I single-channel data is connected to the input end of the CONV4 layer, and the output end of the Q single-channel data is connected to the input end of the CONV5 layer. The output of the convolutional layer is merged and connected to the input end of the first LSTM layer, the output of the CONV6 layer and the output of the CONV4 layer are merged and connected to the input end of the LSTM3 layer, and the output of the CONV6 layer and the output of the CONV5 layer are merged and connected to the input end of the LSTM4 layer. The output of the first LSTM layer is connected to a GaussianDropout layer, the output of the LSTM3 layer is connected to the input of a GaussianDropout layer, and the output of the LSTM4 layer is connected to the input of another GaussianDropout layer. The outputs of the two GaussianDropout layers are combined and connected to the input of the second LSTM layer, the LSTM6 layer. The output of the LSTM6 layer serves as the output of the Part B module.

[0010] The overall feature network classifier module, such as Figure 1The figure shows the Part C module. The Part C module includes a convolutional layer CONV7, an LSTM layer LSTM7, a Gaussian Dropout layer, two fully connected layers Denset1 and Denset2, two identical Dropout layers, and a Softmax output layer. The Part C module combines the outputs of the Part A module and the Part B module as input, which is connected to the input of the Convolutional layer CONV7. The output of the CONV7 layer is connected to the input of the LSTM layer LSTM7. The output of the LSTM7 layer is connected to the input of the Gaussian Dropout layer. The output of the Gaussian Dropout layer is connected to the input of the fully connected layer Denset1. The output of the Denset1 layer is connected to the input of the first Dropout layer. The output of the first Dropout layer is connected to the input of the fully connected layer Denset2. The output of the Denset2 layer is connected to the input of the second Dropout layer. The output of the second Dropout layer is connected to the input of the Softmax layer. The output of the Softmax layer serves as the output of the Part C module and the output of the symmetric multi-channel deep learning system.

[0011] Furthermore, the three-channel input of the multi-channel feature mapping module is a combination of any one multi-channel data and any two single-channel data among the input signal I / Q multi-channel data, the input signal A single-channel data, the input signal P single-channel data, the input signal A / P multi-channel data, the input signal I single-channel data, and the input signal Q single-channel data.

[0012] Furthermore, in the multi-channel feature mapping module, the convolution layers CONV1 and CONV6 are 2D convolution layers, and the convolution layers CONV2, CONV3, CONV4, and CONV5 are 1D convolution layers.

[0013] Furthermore, in the multi-channel feature mapping module, the convolutional layers CONV1 and CONV6 each have 50 filters of size 4*8, and the convolutional layers CONV2, CONV3, CONV4, and CONV5 each have 50 filters of size 2.

[0014] Furthermore, in the multi-channel feature mapping module, the first LSTM layers LSTM1, LSTM2, LSTM3, and LSTM4 each have 64 units, and the second LSTM layers LSTM5 and LSTM6 each have 128 units.

[0015] Furthermore, the convolution layer CONV7 in the overall feature network classifier module is a 1D convolution layer with 50 filters of size 8.

[0016] Furthermore, the LSTM layer LSTM7 in the overall feature network classifier module has 128 units.

[0017] Furthermore, in the overall feature network classifier module, the fully connected layer Denset1 has 256 neurons and the fully connected layer Denset2 has 128 neurons.

[0018] Furthermore, the GaussianDropout layers are all the same, and the GaussianDropout rate is 0.5.

[0019] Furthermore, the Dropout layers are all the same, and the Dropout rate is 0.5.

[0020] Beneficial effects: The automatic signal modulation recognition method of the present invention includes two multi-channel feature mapping modules and an overall feature network classifier module. The multi-channel feature mapping module uses multi-form inputs of the modulated signal, including I / Q sequence, amplitude phase, etc., to capture the connection between different forms of signal representation. Different from the traditional cascade method, the present invention learns the spatial and temporal characteristics of the modulated signal through the cross-hierarchical structure of CNN and LSTM, and seeks the hidden layer connection between different forms of the modulated signal through the reasonable use of the GaussianDropout layer, further expanding the possibility of learning the network, avoiding overfitting, and finally deepening the mapping relationship with the output. The present invention fully learns the characteristics of signals with different modulation modes, has high performance in distinguishing modulation modes, and improves the recognition accuracy of the automatic modulation recognition technology based on deep learning methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a structural diagram of the present invention;

[0022] Figure 2 (a) is the comparison of the overall confusion matrix of the present invention and GRU2;

[0023] Figure 2 (b) is a comparison of the overall confusion matrix of the present invention and PET-CGDNN;

[0024] Figure 2 (c) is the overall confusion matrix diagram of the present invention;

[0025] Figure 3 Comparison of the training and testing accuracy of the present invention and other deep learning systems;

[0026] Figure 4 Comparison of training and testing accuracy for different GaussianDropout layer settings in the present invention;

[0027] Table 1 compares the recognition accuracy of each modulation mode between the present invention and other deep learning systems at a signal-to-noise ratio of 14dB. DETAILED DESCRIPTION

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying 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 making creative efforts shall fall within the scope of protection of the present invention.

[0029] The application scenario of the present invention mainly refers to a single-input, single-output (SISO) communication system, in which the received signal can be expressed as:

[0030] r(t)=s(t)*h(t)+n(t) (1)

[0031] Where s(t) is the modulated signal sent by the transmitter, h(t) is the channel impulse response, and n(t) is the channel noise. After sampling, the transmitted signal r(t) becomes a discrete time signal r(n). r(n) can be obtained from the orthogonal sequence signal r I (n) and r Q (n) indicates.

[0032] Through the orthogonal sequence, the signal amplitude A(n) and phase P(n) expressions can be derived.

[0033]

[0034] like Figure 1 As shown, the signal automatic modulation recognition method includes two multi-channel feature mapping modules and an overall feature network classifier module.

[0035] like Figure 1 As shown, the two multi-channel feature mapping modules are symmetrical to each other and are divided into a Part A module and a Part B module.

[0036] like Figure 1As shown, the Part A module includes a three-channel input, convolutional layers CONV1, CONV2, and CONV3, first LSTM layers LSTM1 and LSTM2, a second LSTM layer LSTM5, and two identical Gaussian Dropout layers. The three-channel inputs of the Part A module respectively input the input signal I / Q multi-channel data, the input signal A (amplitude) single-channel data, and the input signal P (phase) single-channel data. The three-channel inputs are respectively connected to the input of the convolutional layer, the output of the I / Q multi-channel data is connected to the input of the CONV1 layer, the output of the A single-channel data is connected to the input of the CONV2 layer, and the output of the P single-channel data is connected to the input of the CONV3 layer. The output of the convolutional layer is merged and connected to the input of the first LSTM layer, the output of the CONV1 layer and the output of the CONV2 layer are merged and connected to the input of the LSTM1 layer, and the output of the CONV1 layer and the output of the CONV3 layer are merged and connected to the input of the LSTM2 layer. The outputs of the first LSTM layer are connected to GaussianDropout layers. The output of the LSTM1 layer is connected to the input of the GaussianDropout layer, and the output of the LSTM2 layer is connected to the input of another identical GaussianDropout layer. The outputs of the two GaussianDropout layers are combined and connected to the input of the second LSTM layer, LSTM5. The output of the LSTM5 layer serves as the output of the Part A module.

[0037] like Figure 1As shown, the Part B module also includes a three-channel input, convolutional layers CONV6, CONV4, CONV5, first LSTM layers LSTM3 and LSTM4, second LSTM layer LSTM6, and two identical Gaussian Dropout layers. The three-channel input of the Part B module respectively inputs input signal A / P multi-channel data, input signal I single-channel data, and input signal Q single-channel data. The three-channel inputs are respectively connected to the input end of the convolutional layer, the output end of the A / P multi-channel data is connected to the input end of the CONV6 layer, the output end of the I single-channel data is connected to the input end of the CONV4 layer, and the output end of the Q single-channel data is connected to the input end of the CONV5 layer. The output of the convolutional layer is merged and connected to the input end of the first LSTM layer, the output of the CONV6 layer and the output of the CONV4 layer are merged and connected to the input end of the LSTM3 layer, and the output of the CONV6 layer and the output of the CONV5 layer are merged and connected to the input end of the LSTM4 layer. The output of the first LSTM layer is connected to a GaussianDropout layer, the output of the LSTM3 layer is connected to the input of a GaussianDropout layer, and the output of the LSTM4 layer is connected to the input of another GaussianDropout layer. The outputs of the two GaussianDropout layers are combined and connected to the input of the second LSTM layer, the LSTM6 layer. The output of the LSTM6 layer serves as the output of the Part B module.

[0038] like Figure 1As shown, the overall feature network classifier module is Part C. Part C includes a convolutional layer CONV7, an LSTM layer LSTM7, a Gaussian Dropout layer, two fully connected layers Denset1 and Denset2, two identical Dropout layers, and a Softmax output layer. Part C combines the outputs of the Part A module and the Part B module as input, which is connected to the input of the convolutional layer CONV7. The output of the CONV7 layer is connected to the input of the LSTM layer LSTM7. The output of the LSTM7 layer is connected to the input of the Gaussian Dropout layer. The output of the Gaussian Dropout layer is connected to the input of the fully connected layer Denset1. The output of the Denset1 layer is connected to the input of the first Dropout layer. The output of the first Dropout layer is connected to the input of the fully connected layer Denset2. The output of the Denset2 layer is connected to the input of the second Dropout layer. The output of the second Dropout layer is connected to the input of the Softmax layer. The output of the Softmax layer is the output of the Part C module and the output of the symmetric multi-channel deep learning system.

[0039] The experiment of the present invention is based on the open source data set RML2016.10a, which includes a total of 11 modulation signal types, namely WBFM, AM-DSB, AM-SSB, BPSK, CPFSK, GFSK, 4-PAM, 16-QAM, 64-QAM, QPSK, and 8PSK, with a total of 220,000 modulation signals. These modulation signals are evenly distributed between the signal-to-noise ratio (SNR) of -20dB and 18dB, with each SNR step size of 2dB, and a total of 20 SNR conditions. The number of signals of each modulation type in each SNR condition is 1000, and the dimension of each modulation signal is (2, 128), where 2 represents that each signal has I / Q dual-channel data, and 128 represents that each channel has 128 sampling points.

[0040] In order to meet the input requirements of this deep learning system, the A / P dual-channel data of the modulated signal is first derived from the I / Q dual-channel data of the modulated signal in the dataset according to formulas (3) and (4), and then normalized accordingly. The I / Q dual-channel and A / P dual-channel are separated in sequence to form independent I single-channel data, Q single-channel data, A single-channel data, and P single-channel data of dimensions (1, 128), and I / Q dual-channel data and A / P dual-channel data of dimensions (2, 128), which are then submitted to the multi-channel feature mapping module as input. The dataset is divided into training set, validation set, and test set in a ratio of 6:1:3.

[0041] This paper sets up two experiments to evaluate the recognition performance of the symmetric multi-channel deep learning system. All experiments are based on Nvidia GPU acceleration and use the TensorFlow framework.

[0042] In the first experiment, four improved deep learning systems, CNN2, GRU2, CLDNN2, and PET-CGDNN, were selected as the first control group. These four models will use RML2016.10a as the benchmark data set. The recognition performance of this model is evaluated by comparing the recognition gap between the present invention and these four deep learning systems for each modulation mode. Taking into account that under normal circumstances, high-order digital debugging 64-QAM and 16-QAM require a channel signal-to-noise ratio greater than 13dB for normal communication under the 5G communication standard, the recognition performance of each deep learning system is mainly compared when the channel signal-to-noise ratio is 14dB. GRU2 and PET-CGDNN, two deep learning systems with higher overall recognition rates under a signal-to-noise ratio of 14dB, were further selected as the second control group. By analyzing the overall confusion matrix and overall test accuracy of the present invention and these two systems, the recognition performance and learning depth of the present invention are further evaluated.

[0043] In the second experiment, we explored the rationality of the GaussianDropout layer setting. We selected two control groups: no GaussianDropout layer, a GaussianDropout rate of 0.3, and a GaussianDropout rate of 0.8. We evaluated the GaussianDropout layer setting by comparing the overall test accuracy.

[0044] As shown in Table 1, under good communication conditions, the present invention achieves excellent recognition capabilities. Among all results, the difference with CNN2 is the largest, with an overall accuracy difference of nearly 8.545%. Considering that the features learned by CNN2 are relatively simple, focusing solely on the spatial characteristics of the received signal, the poor recognition performance is reasonable. We found that the architecture of the present invention can fully learn the various characteristics of the received signal, improving the accuracy of modulation mode recognition.

[0045] Table 1 Comparison of recognition accuracy of various modulation modes under a signal-to-noise ratio of 14dB

[0046]

[0047]

[0048] like Figure 2The following are the total confusion matrices of GRU2, PET-CGDNN, and the present invention. The elements on the diagonal represent the total recognition rate of each modulation at all signal-to-noise ratios, while the other elements represent the confusion rate of the two modulations. As can be seen from the figure, the recognition rates of 16-QAM and 64-QAM are: (a) 48 and 60, (b) 54 and 62, and (c) 56 and 64, respectively. The present invention has higher performance in 16-QAM and 64-QAM recognition.

[0049] like Figure 3 The figure shows the overall test accuracy of GRU2, PET-CGDNN, and the present invention as iterations increase. The trend shows that the present invention learns the optimal signal features as early as the 55th iteration, achieving the highest overall test accuracy, which is nearly 1.5% higher than the best accuracy of PET-CGDNN. At the end of the curve, it is clear that, with the exception of the present invention, PEF-CGDNN and GRU2 struggle to maintain stability, with their effective accuracy decreasing as iterations increase. Generally speaking, the present invention can achieve better performance more stably and earlier.

[0050] like Figure 4 As shown, the overall test accuracy changes under different GaussianDropout layer settings. The GaussianDropout layer setting has a very important impact on the recognition performance of the present invention. When there is no GaussianDropout layer or the GaussianDropout rate is not 0.5, the best overall test accuracy of the present invention drops by nearly 8%. From the details, without the GaussianDropout layer, the accuracy performance is the worst. The GaussianDropout layer can avoid overfitting by increasing the possibility of input combinations and obtains the maximum possibility when the rate is equal to 0.5. In summary, an appropriate GaussianDropout layer is indispensable.

[0051] Any details not described in detail herein are generally known to those skilled in the art. The preferred embodiments of the present invention have been described in detail above. It should be understood that a person skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solution that a person skilled in the art can arrive at through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A method for automatic signal modulation recognition, characterized in that: The following steps are involved: Step 1: Acquire a target signal, where the target signal is a signal to be modulated and identified; Step 2: Acquire I / Q dual-channel data and A / P dual-channel data of the target signal, separate the I / Q dual-channel data and the A / P dual-channel data to obtain independent I single-channel data, Q single-channel data, A single-channel data, and P single-channel data; Step 3: Build a symmetric multi-channel deep learning system and construct a data set. The data set includes I / Q dual-channel data, A / P dual-channel data, I single-channel data, Q single-channel data, A single-channel data, and P single-channel data. The data set is divided into a training set, a validation set, and a test set in a ratio of 6:1:

3. The built symmetric multi-channel deep learning system is trained to obtain a trained symmetric multi-channel deep learning system. Step 4: Input a new signal to be modulated and identified into the trained symmetric multi-channel deep learning system to determine the modulation recognition mode of the new signal to be modulated and identified; In step 3, the symmetric multi-channel deep learning system includes two multi-channel feature mapping modules and an overall feature network classifier module; The two multi-channel feature mapping modules are defined as Part A module and Part B module respectively: The Part A module includes a three-channel input, a convolutional layer CONV1, a convolutional layer CONV2, a convolutional layer CONV3, the first LSTM layer LSTM1 and LSTM2, the second LSTM layer LSTM5, and two Gaussian Dropout layers; The three-channel input of the Part A module inputs I / Q dual-channel data, A single-channel data, and P single-channel data respectively; The output end of the I / Q dual-channel data is connected to the input end of the convolutional layer CONV1, the output end of the A single-channel data is connected to the input end of the convolutional layer CONV2, and the output end of the P single-channel data is connected to the input end of the convolutional layer CONV3; The output of the convolutional layer CONV1 and the output of the convolutional layer CONV2 are combined and connected to the input end of the LSTM1, and the output of the convolutional layer CONV1 and the output of the convolutional layer CONV3 are combined and connected to the input end of the LSTM2; The output end of the LSTM1 is connected to one of the GaussianDropout layer input ends, and the output of the LSTM2 is connected to the other GaussianDropout layer input end; The outputs of the two GaussianDropout layers are combined and connected to the input of the LSTM5, and the output of the LSTM5 is the output of the Part A module; The Part B module includes a three-channel input, a convolutional layer CONV6, a convolutional layer CONV4, a convolutional layer CONV5, the first LSTM layer LSTM3 and LSTM4, the second LSTM layer LSTM6, and two Gaussian Dropout layers; The three-channel input of the Part B module respectively inputs A / P multi-channel data, I single-channel data, and Q single-channel data; The output end of the A / P multi-channel data is connected to the input end of the convolution layer CONV6, the output end of the I single-channel data is connected to the input end of the convolution layer CONV4, and the output end of the Q single-channel data is connected to the input end of the convolution layer CONV5; The output of the convolutional layer CONV6 and the output of the convolutional layer CONV4 are combined and connected to the input end of the LSTM3, and the output of the convolutional layer CONV6 and the output of the convolutional layer CONV5 are combined and connected to the input end of the LSTM4; The output of the LSTM3 is connected to one of the GaussianDropout layer inputs, and the output of the LSTM4 is connected to the other GaussianDropout layer input. The outputs of the two GaussianDropout layers are combined and connected to the input of LSTM6; The output end of the LSTM6 is the output end of the Part B module; The overall feature network classifier module is defined as a Part C module; the Part C module includes a convolutional layer CONV7, a third LSTM layer LSTM7, a Gaussian Dropout layer, a fully connected layer Denset1, a fully connected layer Denset2, two identical Dropout layers, and a Softmax output layer; The Part C module combines the output of the Part A module and the output of the Part B module as input and connects it to the input end of the convolutional layer CONV7; The output of the convolutional layer CONV7 is connected to the input of LSTM7, the output of the LSTM7 is connected to the input of the GaussianDropout layer, the output of the GaussianDropout layer is connected to the input of the fully connected layer Denset1, the output of the fully connected layer Denset1 is connected to the input of the first Dropout layer, the output of the first Dropout layer is connected to the input of the fully connected layer Denset2, the output of the fully connected layer Denset2 is connected to the input of the second Dropout layer, the output of the second Dropout layer is connected to the input of the Softmax output layer, and the output of the Softmax output layer is the output of the Part C module, that is, the output of the symmetric multi-channel deep learning system.

2. The method for automatic signal modulation recognition according to claim 1, wherein: For the Part A module and the Part B module, the inputs of each of the three channels are any one multi-channel data and any two single-channel data, where the multi-channel data includes: I / Q multi-channel data, A / P multi-channel data, and the single-channel data includes A single-channel data, P single-channel data, I single-channel data and Q single-channel data.

3. The method for automatic signal modulation recognition according to claim 1, wherein: In the multi-channel feature mapping module, the convolution layer CONV1 and the convolution layer CONV6 are 2D convolution layers, and the convolution layer CONV2, the convolution layer CONV3, the convolution layer CONV4, and the convolution layer CONV5 are 1D convolution layers.

4. The method for automatic signal modulation recognition according to claim 1, wherein: In the multi-channel feature mapping module, the convolutional layer CONV1 and the convolutional layer CONV6 each have 50 filters of size 4*8, and the convolutional layer CONV2, the convolutional layer CONV3, the convolutional layer CONV4, and the convolutional layer CONV5 each have 50 filters of size 2.

5. The method for automatic signal modulation recognition according to claim 1, wherein: In the multi-channel feature mapping module, LSTM1, LSTM2, LSTM3, and LSTM4 each have 64 units, and LSTM5 and LSTM6 each have 128 units.

6. The method for automatic signal modulation recognition according to claim 1, wherein: The convolution layer CONV7 in the overall feature network classifier module is a 1D convolution layer with 50 filters of size 8.

7. The method for automatic signal modulation recognition according to claim 1, wherein: The LSTM7 in the overall feature network classifier module has 128 units.

8. The method for automatic signal modulation recognition according to claim 1, wherein: In the overall feature network classifier module, the fully connected layer Denset1 has 256 neurons, and the fully connected layer Denset2 has 128 neurons.

9. The method for automatic signal modulation recognition according to claim 1, wherein: The GaussianDropout layers are all the same, and the GaussianDropout rate is 0.

5.

10. The method for automatic signal modulation recognition according to claim 1, wherein: The Dropout layers are all the same, and the Dropout rate is 0.5.

Images