A modulation waveform recognition method and system based on fusion features

Abstract

The present invention belongs to the field of modulation recognition technology, and in particular relates to a modulation waveform recognition method and system based on fusion features. The method comprises: constructing the instantaneous feature vector, high-order cumulant vector and constellation feature vector of the received IQ signal respectively, combining the IQ signal, inputting a pre-established and trained recognition model at the same time, and outputting the modulation mode of the IQ signal; the recognition model comprises four branches connected in parallel and then cascaded in sequence to form a connection layer, a fully connected layer and an output layer, wherein the first branch and the second branch are two convolution units with the same structure, the first branch is used to further extract the key features of the instantaneous feature vector, and the second branch is used to further extract the key features of the high-order cumulant vector; the third branch comprises a convolutional neural network with four layers of convolution units cascaded to extract the constellation features of the IQ signal; the fourth branch adopts a residual network to extract the original IQ signal features.

Description

Technical Field

[0001] The present invention belongs to the technical field of modulation recognition, and in particular relates to a modulation waveform recognition method and system based on fusion features. Background Art

[0002] Traditional feature extraction-based modulation waveform recognition methods require experts to design and extract effective features based on the specific communication environment to distinguish the types of modulated signals. Different feature extraction methods are required for different modulation modes and channel environments, which consumes a lot of manual time and effort. Furthermore, traditional feature extraction-based recognition methods are poorly robust to noise and interference. When the signal is interfered with by noise, recognition performance will drop sharply. Summary of the Invention

[0003] The purpose of the present invention is to overcome the defects of the prior art and propose a modulation waveform recognition method based on fusion features.

[0004] In order to achieve the above object, the present invention proposes a modulation waveform recognition method based on fusion features, comprising:

[0005] Construct the instantaneous feature vector, high-order cumulant vector and constellation feature vector of the received IQ signal respectively, combine the IQ signal, input the pre-established and trained recognition model at the same time, and output the modulation mode of the IQ signal;

[0006] The recognition model includes four branches connected in parallel and then cascaded in sequence, including a connection layer, a fully connected layer, and an output layer. The first branch and the second branch are two convolutional units with the same structure. The first branch is used to further extract the key features of the instantaneous feature vector, and the second branch is used to further extract the key features of the high-order cumulant vector. The third branch includes a convolutional neural network with four layers of convolutional units cascaded, which is used to extract the constellation diagram features of the IQ signal. The fourth branch uses a residual network to extract the original IQ signal features.

[0007] Preferably, the instantaneous characteristic vector f of the IQ signal transient for:

[0008] f transient =[σ ma ,σ aa ,σ iaa ,σ ap ,σ iap ,PSD,PRR,PAR], the shape is [1,8], where σ ma Characterizes the change of the instantaneous amplitude of the signal, σ aa Characterizes the change in the absolute value of the signal amplitude, σ iaa Characterizes the instability or fluctuation of the signal amplitude, σ apCharacterizes the change of the absolute value of the signal phase, σ iap It represents the standard deviation of the absolute value of the instantaneous phase at the iterative zero center. PSD represents the maximum value of the power spectral density of the normalized sampling signal. PRR is the peak-to-mean square ratio, which represents the maximum value of the power spectral density of the signal in the frequency domain. PAR is the peak-to-average ratio, which represents the ratio of the signal peak value to the mean value.

[0009] Preferably, the high-order cumulant vector f of the IQ signal HOC for:

[0010] f HOC =[C 20 ,C 21 ,C 40 ,C 41 ,C 42 ,C 60 ,C 63 ,C 80 ], the shape is [1, 8],

[0011] C 20 =M 20

[0012] C 21 =M 21

[0013]

[0014] C 41 =M 41 -3M 20 M 21

[0015]

[0016] M pq is the p-order mixing moment of the stationary random process x(n), satisfying the following formula:

[0017] M pq =E{[x(n) p-q ][x * (n)] q}

[0018] Among them, x * (n) is the conjugate signal of x(n), q represents the difference between the order of the mixing moment and the order p, and E represents the expectation.

[0019] Preferably, the constellation diagram of the IQ signal is constructed using the following method:

[0020] Project the in-phase and quadrature components of the modulated signal onto a two-dimensional image of shape [h, w] to construct a constellation diagram of the modulated signal, where i0 and i1 are the I-axis range, q0 and q1 are the Q-axis range, and h and w represent the resolution of the I-axis and Q-axis, respectively.

[0021] Initialize the two-dimensional grid point matrix P to a zero matrix and calculate the grid spacing Δg of the I and Q axes i and Δg q ;

[0022] Map the in-phase component I data and the orthogonal component Q data of the IQ data to a grid point (x, y) in the constellation image, where x and y represent the I-axis coordinate and the Q-axis coordinate respectively;

[0023] Add 1 to the pixel value of that grid point.

[0024] Preferably, the first branch and the second branch are two convolutional units with the same structure, and the convolutional unit includes a single convolutional module and a channel attention network SENet, wherein,

[0025] The single convolution module includes convolution calculation, batch normalization, nonlinear activation function and maximum pooling operation;

[0026] The processing process of the channel attention network SENet includes: reducing the features in each channel of the input data into a single real value by average pooling, and then using the full connection operation to compress the channel to reduce the amount of model calculation, and then restoring the number of channels through the ReLU layer and the fully connected layer to obtain the channel weight vector, and finally multiplying the original input data with the corresponding weight to output data with a shape of [10,3].

[0027] Preferably, the number of convolution kernels of the first three convolution units of the third branch is 20, the number of convolution kernels of the last convolution unit is 10, and the size of the convolution kernel is 3×3. In order to splice with the feature data of other branches, the shape of the output data is adjusted from [10, 2, 2] to [10, 4].

[0028] Preferably, the fourth branch includes two convolution units and two residual stacking units, the convolution kernel size of the two convolution units is 3, the number of output channels of the first convolution unit is 64, the number of output channels of the second convolution unit 2 is 10, and the residual stacking unit is formed by cascading the residual unit and the maximum pooling operation, and the output data shape is [10, SignalLength / 8], where SignalLength is the length of the signal to be identified.

[0029] Preferably, the connection layer is used to splice the output data of each branch to obtain feature fusion data with a shape of [10, SignalLength / 8+10];

[0030] The fully connected layer is used to complete the mapping from signal characteristics to modulation mode;

[0031] The output layer includes two layers of fully connected networks.

[0032] Preferably, the step of training the recognition model includes:

[0033] Establish a training set, input the recognition model, adjust the model parameters until the training requirements are met, and obtain a trained recognition model.

[0034] On the other hand, the present invention proposes a modulation waveform recognition system based on fusion features, comprising:

[0035] A feature extraction module is used to construct an instantaneous feature vector, a high-order cumulant vector and a constellation feature vector of the received IQ signal respectively;

[0036] The recognition module is used to combine the output of the feature extraction module with the IQ signal, input a pre-established and trained recognition model, and output the modulation mode of the IQ signal;

[0037] The recognition model includes four branches connected in parallel and then cascaded in sequence, including a connection layer, a fully connected layer, and an output layer. The first branch and the second branch are two convolutional units with the same structure. The first branch is used to further extract the key features of the instantaneous feature vector, and the second branch is used to further extract the key features of the high-order cumulant vector. The third branch includes a convolutional neural network with four layers of convolutional units cascaded, which is used to extract the constellation diagram features of the IQ signal. The fourth branch uses a residual network to extract the original IQ signal features.

[0038] Compared with the prior art, the advantages of the present invention are:

[0039] This paper proposes a method for modulated waveform recognition based on fused features. It employs a convolutional neural network to extract instantaneous features, high-order cumulants, constellation diagrams, and higher-level feature representations of the original signal. Utilizing a channel attention mechanism, the model selectively focuses on feature channels that are useful for the recognition task, thereby reducing the impact of redundant information on the recognition results. This method addresses the problems of low recognition accuracy and poor generalization in low signal-to-noise ratio environments, often encountered by traditional methods. It also boasts rapid algorithm convergence and requires minimal training time. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a network structure diagram of the modulation waveform recognition model based on feature fusion of the present invention. DETAILED DESCRIPTION

[0041] This paper proposes a modulation waveform recognition network based on fusion features, such as Figure 1 As shown in the figure, both branches 1 and 2 use CNNs instead of traditional methods such as decision trees to further extract key signal features. Due to the small size of the input data for instantaneous features and high-order cumulants, the number of effective features that can be extracted by the neural network is limited. Therefore, the number of convolutional units in the instantaneous and high-order cumulant feature learning module is set to 1. Branch 3 uses a convolutional neural network composed of four cascaded layers of convolutional units to extract signal constellation features. To fully capture the effective signal features, branch 4 uses a residual neural network to extract raw signal features. The connection layer fuses the feature data from different branches and outputs it to the fully connected classification layer. The final output is the modulation mode of the signal to be identified. Furthermore, a channel attention network (SENet) is cascaded after the convolutional units to adaptively learn the weights of feature channels to better adapt to different datasets and application scenarios, thereby improving the model's generalization ability.

[0042] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0043] Example 1

[0044] An embodiment of the present invention proposes a modulation waveform recognition method based on feature fusion.

[0045] The modulation waveform recognition method based on feature fusion first constructs the instantaneous feature vector, high-order cumulant vector, and constellation feature vector of the modulation signal to be identified. These feature vectors and the original IQ (In-Phase and Quadrature, IQ) signal are then input into the recognition network. The network is trained on training samples, and the parameters that achieve the highest recognition accuracy in the validation set are selected as the final model parameters. Finally, the model's recognition performance under different signal-to-noise ratio conditions is verified on a test set, confirming the effectiveness of the proposed method.

[0046] The instantaneous characteristic f of the received signal transient It can be expressed as:

[0047] f transient =[σ ma ,σ aa ,σ iaa ,σ ap ,σ iap ,PSD,PRR,PAR] (1)

[0048] The shape is [1,9], where:

[0049] (1)σ ma Characterizes the change of the instantaneous amplitude of the signal. The mathematical expression is:

[0050]

[0051] Where N represents the number of sampling points, A cn (i) represents the zero-centered normalized instantaneous amplitude.

[0052] (2) Standard deviation σ of the absolute value of the zero-centered normalized instantaneous amplitude aa :

[0053] σ aa Characterizes the change in the absolute value of the signal amplitude. The mathematical expression is:

[0054]

[0055] (3) The standard deviation σ of the absolute value of the iterative zero-centered normalized instantaneous amplitude iaa :

[0056] σ iaa Characterizes the instability or fluctuation of signal amplitude, which is relatively large compared to σ aa It can better distinguish high-order amplitude modulation signals. The mathematical expression is:

[0057]

[0058] (4) Standard deviation σ of the absolute value of the instantaneous phase at the zero center ap :

[0059] σ ap Characterizes the change of the absolute value of the signal phase. The mathematical expression is:

[0060]

[0061] Where φ zc (i) is the zero-centered instantaneous phase, φ(i) represents the fold-free phase sequence, which is the sum of the instantaneous phase θ(i) and the phase sequence correction factor C(i). The expressions of θ(i) and C(i) are shown in Equation (6).

[0062]

[0063] (5) The standard deviation σ of the absolute value of the instantaneous phase at the iterative zero center iap :

[0064]

[0065] Where, φ' zc (i) is the iterative zero-centered instantaneous phase.

[0066] (6) Maximum power spectral density (PSD) of the normalized sampling signal:

[0067] PSD is the maximum value of the power spectral density of the signal in the frequency domain, which represents the main frequency characteristics of the signal. The mathematical expression is:

[0068]

[0069] Where x(i) represents the received signal.

[0070] (7) Peak-to-Revenue Ratio (PRR):

[0071] PRR is an indicator used to measure the ratio of the peak value to the root mean square value of a signal. Its physical meaning is to measure the ratio of the instantaneous peak value in the signal to the overall energy. The mathematical expression is:

[0072]

[0073] (8) Peak-to-average ratio (PAR):

[0074] PAR is an indicator used to measure the ratio of the peak value to the mean value of a signal. Its physical meaning is to measure the ratio of the instantaneous peak value of the signal to the mean value of the signal. The calculation formula is:

[0075]

[0076] The p-order mixing moment M of a stationary random process x(n) pq The mathematical expression is shown in formula (13):

[0077] M pq =E{[x(n) p-q ][x * (n)] q} (13)

[0078] Among them, x * (n) is the conjugate signal of x(n). The high-order cumulant of the modulated signal is shown in Equation (14).

[0079]

[0080] The high-order cumulative characteristic f of the received signal HOC It can be expressed as:

[0081] f HOC =[C 20 ,C 21 ,C 40 ,C 41 ,C 42 ,C 60 ,C 63 ,C 80 ] (15)

[0082] The shape is [1,8].

[0083] Project the in-phase component and quadrature component of the modulated signal to a shape of [p i ,qi ], the modulation signal constellation is constructed, and the pseudo code is shown in Table 1. Where i0 and i1 are the I axis range, q0 and q1 are the Q axis range, and p i and p q Represents the resolution of the I axis and Q axis respectively. First, initialize the two-dimensional grid point matrix P to a zero matrix and calculate the grid spacing Δg of the two axes i , Δg q ; Secondly, the input I data and Q data are mapped to a grid point (x, y) in the constellation image, where x and y represent the I-axis coordinate and the Q-axis coordinate respectively; finally, the pixel value of the grid point is increased by 1.

[0084] Table 1 Constellation map mapping algorithm

[0085]

[0086]

[0087] IQ signal instantaneous characteristics f transient The shape is [1,9]. The instantaneous feature learning module consists of a single convolutional module and a channel attention network (SENet). The convolutional module includes convolution calculations, batch normalization, nonlinear activation functions, and maximum pooling operations. SENet first reduces the features within each channel of the input data to a single real value through average pooling. Secondly, a fully connected operation is used to compress C channels into C / r channels, reducing the model's computational complexity. Then, a ReLU layer and a fully connected layer with C channels are used to restore the number of channels and obtain the channel weight vector. Finally, the original input data is multiplied by the corresponding weights to produce the output data. As can be seen, the input and output data of SENet have the same shape.

[0088] High-order cumulant characteristic data f HOC The shape is [1,8]. The high-order cumulant feature learning module further extracts the effective features between the high-order cumulants of the signal. The network structure is the same as that of the instantaneous feature learning module. hoc After the high-order cumulant feature learning module, the shape becomes [10,3].

[0089] The core concept of the constellation feature learning module is to extract richer features by increasing the network depth. Increasing the network width, compared to increasing the number of network layers, may introduce more computational overhead. Furthermore, when the network width is too large, it can easily lead to feature duplication, increasing redundant computation. Therefore, the constellation feature extraction module consists of four cascaded convolutional units. The first three convolutional units have 20 convolution kernels, while the last has 10, with a kernel size of 3×3. To facilitate concatenation with feature data from other branches, the output data is reshaped from [10, 2, 2] to [10, 4].

[0090] The IQ feature learning module consists of two convolutional units and two residual stacking units. Convolutional unit 1 has 64 output channels, while convolutional unit 2 has 10 output channels, and the convolution kernel size is 3 for both. The residual stacking unit is composed of a cascade of residual units and a maximum pooling operation. By introducing residual connections, the IQ feature extraction module provides a shortcut within each residual unit, allowing gradients to more easily reach the underlying layers, helping to alleviate the vanishing gradient problem. Furthermore, the skip connections in the residual units help reduce the correlation between parameters, improve parameter utilization efficiency, and reduce the difficulty of network optimization.

[0091] The output data of the four branches have shapes of [10, 3], [10, 3], [10, 4], and [10, SignalLength / 8], respectively. The connection layer concatenates the output data of each branch to generate feature fusion data of the shape [10, SignalLength / 8+10]. The fused features are input to a fully connected classification module to map signal features to modulation types. The output layer contains two fully connected layers, with 20 hidden units and the number of modulation signal types in the dataset, respectively.

[0092] Example 2

[0093] The present invention proposes a modulation waveform recognition system based on fusion features, which is implemented based on the method of Example 1. The system includes:

[0094] A feature extraction module is used to construct an instantaneous feature vector, a high-order cumulant vector and a constellation feature vector of the received IQ signal respectively;

[0095] The recognition module is used to combine the output of the feature extraction module with the IQ signal, input a pre-established and trained recognition model, and output the modulation mode of the IQ signal;

[0096] The recognition model includes four branches connected in parallel and then cascaded in sequence, including a connection layer, a fully connected layer, and an output layer. The first branch and the second branch are two convolutional units with the same structure. The first branch is used to further extract the key features of the instantaneous feature vector, and the second branch is used to further extract the key features of the high-order cumulant vector. The third branch includes a convolutional neural network with four layers of convolutional units cascaded, which is used to extract the constellation diagram features of the IQ signal. The fourth branch uses a residual network to extract the original IQ signal features.

[0097] Innovation:

[0098] To address the challenges of traditional feature extraction-based modulation waveform recognition methods, such as complex feature extraction and noise sensitivity, which hinder high-precision recognition of multiple modulation modes for communication satellites, a new modulation waveform recognition method based on fused features is proposed. A convolutional neural network is used to extract instantaneous features, high-order cumulants, constellation diagrams, and higher-level feature representations of the original signal. A channel-attention mechanism is employed to selectively focus on feature channels that are useful for the recognition task, thereby reducing the impact of redundant information on the recognition results. Experimental results demonstrate that the modulation waveform recognition method based on fused features outperforms the other four feature types, with rapid convergence and short training time. To further validate the effectiveness of the proposed method, this paper compares it with traditional machine learning and deep learning algorithms on a publicly available dataset. The results demonstrate that the proposed method achieves a maximum recognition accuracy of 91%, an improvement of nearly 16% over traditional machine learning algorithms. Compared with comparable deep learning methods, the proposed method reduces the number of parameters by approximately 77% and improves efficiency by 8.4%.

[0099] The recognition performance of the proposed method is significantly better than that of the traditional machine learning algorithm. The average recognition accuracy of the best machine learning method is nearly 16% lower than that of the proposed method.

[0100] Compared with deep learning methods, the proposed method has a slightly higher recognition accuracy than the MACLDNN algorithm, but the number of model parameters is reduced by nearly 77%, and the training time is shortened by 5.27 seconds. This shows that the proposed recognition method achieves good results in both recognition accuracy and recognition efficiency.

[0101] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and are intended to be encompassed by the claims of the present invention.

Claims

1. A modulation waveform recognition method based on fusion features, comprising: Construct the instantaneous feature vector, high-order cumulant vector and constellation feature vector of the received IQ signal respectively, combine the IQ signal, input the pre-established and trained recognition model at the same time, and output the modulation mode of the IQ signal; The recognition model includes four branches connected in parallel and then cascaded in sequence to form a connection layer, a fully connected layer and an output layer, wherein the first branch and the second branch are two convolution units with the same structure, and the convolution unit includes a single convolution module and a channel attention network SENet. The processing process of the channel attention network SENet includes: reducing the features in each channel of the input data into a single real value by average pooling, and then using a fully connected operation to compress the channel to reduce the amount of model calculation, and then restoring the number of channels through the ReLU layer and the fully connected layer to obtain the channel weight vector, and finally multiplying the original input data with the corresponding weight to output data with a shape of [10,3]; the first branch is used to further extract the key features of the instantaneous feature vector, and the second branch is used to further extract the key features of the high-order cumulant vector; the third branch includes a convolutional neural network with four layers of convolution units cascaded to extract the constellation diagram features of the IQ signal; the fourth branch uses a residual network to extract the original IQ signal features.

2. The modulation waveform recognition method based on fusion features according to claim 1 is characterized in that: The instantaneous eigenvector f of the IQ signal transient for: f transient =[σ ma ,σ aa ,σ iaa ,σ ap ,σ iap ,PSD,PRR,PAR], the shape is [1,8], where σ ma Characterizes the change of the instantaneous amplitude of the signal, σ aa Characterizes the change in the absolute value of the signal amplitude, σ iaa Characterizes the instability or fluctuation of the signal amplitude, σ ap Characterizes the change in the absolute value of the signal phase, σ iap It represents the standard deviation of the absolute value of the instantaneous phase at the iterative zero center. PSD represents the maximum value of the power spectral density of the normalized sampling signal. PRR is the peak-to-mean square ratio, which represents the maximum value of the power spectral density of the signal in the frequency domain. PAR is the peak-to-average ratio, which represents the ratio of the signal peak value to the mean value.

3. The modulation waveform recognition method based on fusion features according to claim 1 is characterized in that: The high-order cumulant vector f of the IQ signal HOC is: f HOC =[C 20 ,C 21 ,C 40 ,C 41 ,C 42 ,C 60 ,C 63 ,C 80 ], the shape is [1, 8], C 20 =M 20 C 21 =M 21 C 41 =M 41 -3M 20 M 21 M pq is the p-order mixing moment of the stationary random process x(n), satisfying the following formula: M pq =E{[x(n) p-q ][x * (n)] q } Among them, x * (n) is the conjugate signal of x(n), q represents the difference between the order of the mixing moment and the order p, and E represents the expectation.

4. The modulation waveform recognition method based on fusion features according to claim 1 is characterized in that: The constellation diagram of the IQ signal is constructed using the following method: Project the in-phase and quadrature components of the modulated signal onto a two-dimensional image of shape [h, w] to construct a constellation diagram of the modulated signal, where i0 and i1 are the I-axis range, q0 and q1 are the Q-axis range, and h and w represent the resolution of the I-axis and Q-axis, respectively. Initialize the two-dimensional grid point matrix P to a zero matrix and calculate the grid spacing Δg of the I and Q axes i and Δg q ; Map the in-phase component I data and the orthogonal component Q data of the IQ data to a grid point (x, y) in the constellation image, where x and y represent the I-axis coordinate and the Q-axis coordinate respectively; Add 1 to the pixel value of that grid point.

5. The modulation waveform recognition method based on fusion features according to claim 1 is characterized in that: The single convolution module includes convolution calculation, batch normalization, nonlinear activation function and maximum pooling operation.

6. The modulation waveform recognition method based on fusion features according to claim 1 is characterized in that: The number of convolution kernels of the first three convolution units of the third branch is 20, and the number of convolution kernels of the last convolution unit is 10. The size of the convolution kernel is 3×3. In order to splice the feature data with other branches, the shape of the output data is adjusted from [10, 2, 2] to [10, 4].

7. The modulation waveform recognition method based on fusion features according to claim 1 is characterized in that: The fourth branch includes two convolution units and two residual stacking units. The convolution kernel size of the two convolution units is 3, the number of output channels of the first convolution unit is 64, and the number of output channels of the second convolution unit is 10. The residual stacking unit is formed by cascading the residual unit and the maximum pooling operation. The output data shape is [10, SignalLength / 8], where SignalLength is the length of the signal to be identified.

8. The modulation waveform recognition method based on fusion features according to claim 1 is characterized in that: The connection layer is used to splice the output data of each branch to obtain feature fusion data with a shape of [10, SignalLength / 8+10]; The fully connected layer is used to complete the mapping from signal characteristics to modulation mode; The output layer includes two layers of fully connected networks.

9. The modulation waveform recognition method based on fusion features according to claim 1, characterized in that: Includes the training steps for the recognition model, including: Establish a training set, input the recognition model, adjust the model parameters until the training requirements are met, and obtain a trained recognition model.

10. A modulation waveform recognition system based on fusion features, characterized in that: include: A feature extraction module is used to construct an instantaneous feature vector, a high-order cumulant vector and a constellation feature vector of the received IQ signal respectively; The recognition module is used to combine the output of the feature extraction module with the IQ signal and input it into a pre-established and trained recognition model. It uses the channel attention mechanism to make the recognition model selectively focus on the feature channels that are useful for the recognition task, thereby reducing the impact of redundant information on the recognition results and outputting the modulation mode of the IQ signal; The recognition model includes four branches connected in parallel and then cascaded in sequence to form a connection layer, a fully connected layer and an output layer, wherein the first branch and the second branch are two convolution units with the same structure, and the convolution unit includes a single convolution module and a channel attention network SENet. The processing process of the channel attention network SENet includes: reducing the features in each channel of the input data into a single real value by average pooling, and then using a fully connected operation to compress the channel to reduce the amount of model calculation, and then restoring the number of channels through the ReLU layer and the fully connected layer to obtain the channel weight vector, and finally multiplying the original input data with the corresponding weight to output data with a shape of [10,3]; the first branch is used to further extract the key features of the instantaneous feature vector, and the second branch is used to further extract the key features of the high-order cumulant vector; the third branch includes a convolutional neural network with four layers of convolution units cascaded to extract the constellation diagram features of the IQ signal; the fourth branch uses a residual network to extract the original IQ signal features.

Images