Automatic modulation recognition method based on binary neural network

Abstract

The application discloses an automatic modulation recognition method based on a binary neural network and relates to the technical fields of deep learning and wireless communication. The method comprises the following steps: a binary neural network-based automatic modulation recognition network model BMR-Net is built; the BMR-Net is trained by using an automatic modulation recognition training set, the neural network weight obtained through training is stored, and the trained BMR-Net model is saved as an inference model; wireless communication signals are collected; the collected signals are subjected to I / Q preprocessing; the preprocessed wireless communication signals are input into the inference model for processing; and the inference model outputs a predicted signal modulation recognition type. The method can improve the modulation recognition accuracy from 45.62% to 84.36%, can significantly reduce the requirements for computing and storage resources, and can improve the inference speed of the automatic modulation recognition model based on deep learning.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and wireless communication technology, and in particular to an automatic modulation recognition method based on a binary neural network. Background Technology

[0002] Modulation identification of wireless communication signals plays a crucial role in the field of cognitive radio, with wide applications in areas such as electromagnetic spectrum interference detection, cognitive radio fault detection, dynamic spectrum acquisition, spectrum regulation, and countermeasures. Before the rise of deep learning methods, academia and industry relied on pattern recognition and hypothesis testing for communication signal modulation identification. However, these methods have the following drawbacks: 1) over-reliance on prior knowledge, resulting in a large workload for manual intervention; 2) weak generalization ability, making it difficult to adapt to different identification scenarios; and 3) low identification accuracy, failing to meet application requirements in complex scenarios.

[0003] Given the significant success of deep learning technology in fields such as computer vision and natural language processing, the industry has begun to apply deep learning techniques for automatic modulation recognition in recent years. The paper "Convolutional radiomodulation recognition networks (Engineering Applications of Neural Networks), August 2016" proposed an automatic modulation recognition method based on convolutional neural networks and released the widely used Radio ML series datasets. The paper "An Improved Neural Network Pruning Technology for Automatic Modulation Classification in Edge Devices" (IEEE Transactions on Vehicular Technology, Vol. 69, No. 5, May 2020) designed a lightweight VT-CNN2 model for automatic modulation recognition. The paper "Research on Automatic Modulation Recognition Based on Deep Learning" (Computer Science, Vol. 49, No. 5, May 2021) summarized the work on automatic modulation recognition based on deep learning technology. It can be seen that current deep learning technology has been widely applied in the field of automatic modulation recognition, and its recognition accuracy and generalization ability are superior to traditional methods.

[0004] However, current deep learning-based automatic modulation recognition methods have very high computational and storage requirements. Deploying these models on resource-constrained wireless communication devices such as radios presents two problems: first, simply migrating the massive deep learning models to embedded devices leads to insufficient computational and storage resources; second, even if the deep learning models are loaded onto embedded devices, their enormous computational demands result in excessively long computation times, failing to meet the requirements of fast signal processing in wireless communication. Therefore, designing a lightweight deep learning model with low resource requirements and fast inference speed to complete automatic modulation recognition tasks on resource-constrained platforms is an essential requirement for the practical application of automatic modulation recognition in wireless communication.

[0005] Binary neural networks (BNNs) are a lightweight neural network structure that transforms both the input and weight values ​​of a neural network model into -1 and 1. This converts computationally intensive convolution operations into XOR and bitcount operations, reducing computational load by 58 times and storage requirements by 16 times. Therefore, BNNs show promise for automatic modulation recognition in resource-constrained scenarios. However, directly applying current BNN models for automatic modulation recognition still faces two challenges: firstly, training convergence is difficult; secondly, the model's recognition accuracy is low. Therefore, designing an automatic modulation recognition method based on BNNs with fast convergence and high recognition accuracy is a pressing issue. Summary of the Invention

[0006] The technical problem to be solved by the present invention is how to provide an automatic modulation recognition method based on binary neural networks that can significantly reduce the computational and storage resource requirements of deep learning models used for automatic modulation recognition and improve the inference speed of deep learning models on embedded devices such as wireless communication radios.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an automatic modulation recognition method based on a binary neural network, comprising the following steps:

[0008] S101: Construct an automatic modulation recognition network model BMR-Net based on a binary neural network;

[0009] S102: Train the BMR-Net network model using the automatic modulation recognition training set. The training method adopts the gradient descent backpropagation method based on mini-batch training until the network model converges. Store the neural network weights obtained from the training and save the trained BMR-Net model as an inference model.

[0010] S103: Acquire the wireless communication signal to be identified;

[0011] S104: Perform I / Q preprocessing on the acquired signal;

[0012] S105: Input the wireless communication signal to be identified into the inference model described in step S102;

[0013] S106: Output predicted signal modulation identification type.

[0014] A further technical solution is that the network model BMR-Net includes: one 2D convolutional layer (Conv2D), one normalization layer (BN), six residual blocks, four max pooling layers, three fully connected layers, two dropout layers, and a softmax layer. The output of the 2D convolutional layer (Conv2D) is connected to the input of the normalization layer (BN). The output of the normalization layer (BN) is connected to the input of the max pooling layer (Maxpooling-1). The output of the max pooling layer (Maxpooling-1) is connected to the input of the residual block (basic block-1). The output of the residual block (basic block-1) is connected to the input of the residual block (basic block-2). The output of the residual block (basic block-2) is connected to the input of the max pooling layer (Maxpooling-2). The output of the max pooling layer (Maxpooling-2) is connected to the input of the residual block (basic block-3). The output of the residual block (basic block-3) is connected to the input of the residual block (basic block-4). The output of block-4 is connected to the input of the max-pooling layer Maxpooling-3; the output of Maxpooling-3 is connected to the input of the residual block basic block-5; the output of basic block-5 is connected to the input of the residual block basic block-6; and the output of basic block-6 is connected to the input of the residual block basic block-6. The output of block-6 is connected to the input of the max pooling layer Maxpooling-4. The output of Maxpooling-4 is connected to the input of the Flatten layer. The output of the Flatten layer is connected to the input of the fully connected layer Fc-1. The output of the fully connected layer Fc-1 is connected to the input of the Dropout-1 layer. The output of the Dropout-1 layer is connected to the input of the fully connected layer Fc-2. The output of the fully connected layer Fc-2 is connected to the input of the Dropout-2 layer. The output of the Dropout-2 layer is connected to the input of the fully connected layer Fc-3. The output of the fully connected layer Fc-3 is connected to the input of the Softmax layer. The output of the Softmax layer is the output of the network model BMR-Net.

[0015] The beneficial effects of the above technical solution are as follows: the method constructs a residual block-based neural network model, BMR-Net, which propagates information from the previous layer layer by layer, thus compensating to some extent for the information loss caused by binarization of network parameters. Furthermore, by designing a second-order approximation of the binarization function during backpropagation, the problems of difficult training and low recognition accuracy in automatic modulation recognition based on binary neural networks are solved. Experiments show that compared to directly binarizing the automatic modulation recognition neural network based on AlexNet, the method of this invention improves the modulation recognition accuracy from 45.62% to 84.36%. At the same time, it significantly reduces the computational load and storage requirements of deep learning-based automatic modulation recognition methods, thus enabling the deployment of automatic modulation recognition algorithms in embedded devices and demonstrating good practical application value. Attached Figure Description

[0016] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0017] Figure 1 This is a flowchart of the method described in the embodiments of the present invention;

[0018] Figure 2 This is a diagram of the overall network structure of BMR-Net in the method described in this embodiment of the invention;

[0019] Figure 3 This is a flowchart of the processing of the basic residual block in the method described in the embodiments of the present invention;

[0020] Figure 4 It directly binarizes the signal modulation method recognition accuracy and confusion matrix based on the AlexNet-based automatic modulation recognition neural network;

[0021] Figure 5 The method described in the embodiments of the present invention is used to identify the accuracy and confusion matrix of the signal modulation mode;

[0022] Figure 6 It refers to the recognition accuracy of each modulation method under different signal-to-noise ratio conditions;

[0023] Figure 7 It describes the change in modulation scheme recognition accuracy during training as the number of iterations increases. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0026] like Figure 1 As shown in the figure, an embodiment of the present invention discloses an automatic modulation recognition method based on a binary neural network, the method comprising the following steps:

[0027] S101: Construct an automatic modulation recognition network model BMR-Net based on a binary neural network. The structure of the BMR-Net network model is as follows: Figure 2 As shown, it mainly consists of one 2D convolutional layer (Conv2D), one normalization layer (BN), six residual blocks (basicblock-1, basic block-2, basic block-3, basic block-4, basic block-5, basic block-6), four max pooling layers (Maxpooling-1, Maxpooling-2, Maxpooling-3, Maxpooling-4), three fully connected layers (Fc-1, Fc-2, Fc-3), and two dropout layers (Dropout-1, Dropout-2). The last layer of BMR-Net is a Softmax layer. The output dimension of the Softmax layer is equal to the number of modulation types to be identified. The input of BMR-Net is I / Q data, and the output is the modulation type corresponding to the input signal.

[0028] The output of the 2D convolutional layer Conv2D is connected to the input of the normalization layer BN. The output of the normalization layer BN is connected to the input of the max pooling layer Maxpooling-1. The output of the max pooling layer Maxpooling-1 is connected to the input of the residual block basic block-1. The output of the residual block basic block-1 is connected to the input of the residual block basic block-2. The output of the residual block basic block-2 is connected to the input of the max pooling layer Maxpooling-2. The output of the max pooling layer Maxpooling-2 is connected to the input of the residual block basic block-3. The output of the residual block basic block-3 is connected to the input of the residual block basic block-4. The output of the residual block basic block-4 is connected to the input of the max pooling layer Maxpooling-3. The output of the max pooling layer Maxpooling-3 is connected to the input of the residual block basic block-5. The output of the residual block basic block-5 is connected to the input of the residual block basic block-5. The input of block-6 is connected to the network layer. The output of the residual block basic block-6 is connected to the input of the max pooling layer Maxpooling-4. The output of the max pooling layer Maxpooling-4 is connected to the input of the Flatten layer. The output of the Flatten layer is connected to the input of the fully connected layer Fc-1. The output of the fully connected layer Fc-1 is connected to the input of the Dropout-1 layer. The output of the Dropout-1 layer is connected to the input of the fully connected layer Fc-2. The output of the fully connected layer Fc-2 is connected to the input of the Dropout-2 layer. The output of the Dropout-2 layer is connected to the input of the fully connected layer Fc-3. The output of the fully connected layer Fc-3 is connected to the input of the Softmax layer. The output of the Softmax layer is the output of the BMR-Net network model. Table 1 shows the network structure of BMR-Net in tabular form.

[0029] Table 1: BMR-Net Network Structure

[0030]

[0031] The six basic residual blocks have similar structures, differing only in their input and output dimensions (specific dimensions are shown in Table 1). Their basic structures are as follows: Figure 3 As shown;

[0032] Taking the first residual block (basic block-1) as an example, its input is represented as x. The basic block-1 consists of the following components:

[0033] 1) Binarization layer (Sign): Binarizes the input of the residual block to -1 or 1. The calculation method is as follows:

[0034]

[0035] Where, x b The input is after binarization;

[0036] 2) Binary convolutional layer: x b With the binarized weight W b Performing binary convolution, since x b and W b Both are -1 or 1. The convolution operation is as follows: First, x... b With W b If we perform a XOR operation and then count the number of -1s and 1s respectively as m and n, then the formula for calculating binary convolution is:

[0037] a = x b *W b =nm

[0038] Where a is the output of the binary convolutional layer, and * represents a two-dimensional convolution operation;

[0039] 3) Batch Normalization (BN) layer:

[0040]

[0041] Where b is the output of the normalization layer. Let σ(a) be the mean of a, and σ(a) be the standard deviation of a.

[0042] 4) If a step connection adds x and b, then the output of basic block-1 is represented as O = x + b;

[0043] With this design, each residual block retains the original input information of the previous layer to a certain extent, which makes up for the information loss caused by input and weight binarization to a certain extent.

[0044] S102: Train the BMR-Net model using the automatic modulation recognition training set: The training method adopts the gradient descent backpropagation method based on mini-batch training until the network model converges, store the neural network weights obtained from the training, and save the trained BMR-Net model as an inference model.

[0045] Since the binarization function is non-differentiable, its gradient cannot be directly calculated when the gradient is propagated to the binarization layer during training. To address this issue, traditional methods typically use a `clip()` function as an approximation of the original binarization function during backpropagation. However, the derivative of the `clip()` function differs significantly from the derivative of the binarization function. This mismatch between forward inference and the backpropagation binarization layer affects the network's recognition accuracy. This invention uses a second-order approximation of the binarization function when the gradient is propagated to the binarization layer. The specific implementation method is as follows:

[0046] Standard binarization functions are used during model inference.

[0047] During model training, assuming the vector representation to be propagated from the binarized layer in the residual block to the previous layer is y, then during backpropagation, what is actually propagated is the second-order approximation σ(y) of y, which is calculated as follows:

[0048]

[0049] In this embodiment, the publicly available RadioML 2016a automatic modulation recognition dataset is used to train the aforementioned binary neural network. The modulation schemes include 11 digital and analog signals: BPSK, 8PSK, QPSK, CPFSK, GFSK, PAM4, QAM16, QAM64, AM-DSB, AM-SSB, and WB-FM. The signal-to-noise ratio range of the signals is -20 to 18, the signal length is 128, and the number of samples is 2.2 × 10⁻⁶. 5 During training, 60% of the RadioML 2016.10a dataset was used as the training set, 20% as the validation set, and 20% as the test set. The network model was built using Python and the Keras deep learning framework. Training was performed using an NVIDIA GeForce RTX 3090 graphics card and the accompanying CUDA software. The loss function was cross-entropy loss. Batch gradient descent was employed, with each batch of training data consisting of 1024 data points and 200 iterations. The Adam algorithm was used for optimization, adjusting and optimizing the network hyperparameters until the network accuracy no longer improved.

[0050] S103: Acquire the wireless communication signal to be identified: The signal received by the wireless communication receiver is r(t), which can be expressed as: r(t)=s(t)*c+n(t), where s(t) is the original signal after modulation, c is the channel loss caused by various reasons, and n(t) is additive white Gaussian noise;

[0051] S104: I / Q preprocessing of the acquired signal: The in-phase component I(j) and quadrature component Q(j) of the acquired signal are used as inputs for feature extraction, with a dimension of 128×2. Before identifying the signal modulation scheme, the I / Q data are preprocessed by normalization: First, the mean μ of I(j) and Q(j) is calculated respectively. I μ Q :

[0052]

[0053]

[0054] Then calculate the standard deviation of the two sets of data separately:

[0055]

[0056]

[0057] Finally, the normalized I / Q data is calculated using the following formula:

[0058]

[0059]

[0060] S105: Input the wireless communication signal to be identified into the inference model described in S101: Input the preprocessed I / Q data described in step S104. and Input the inference model described in S101.

[0061] S106: Output prediction signal modulation identification type: and The inference model performs forward computation layer by layer, and finally outputs the predicted signal modulation identification type.

[0062] Experimental verification:

[0063] Figure 4 The accuracy and confusion matrix of signal modulation scheme recognition were demonstrated by directly binarizing the parameters of the AlexNet-based automatic modulation recognition neural network. Figure 5 It uses the signal modulation method of BMR-Net to identify accuracy and confusion matrix.

[0064] Compared to directly binarizing the AlexNet-based automatic modulation recognition neural network, BMR-Net exhibits good normalization in the confusion matrix and improves the modulation recognition accuracy from 45.62% to 84.36%. Figure 6This represents the recognition accuracy of BMR-Net under each modulation scheme and different signal-to-noise ratio (SNR) conditions. The recognition accuracy is lower for QAM16, QAM64, and WBFM modulation schemes. Figure 6 This indicates that QAM16 and QAM64 modulation schemes are easily confused. However, when SNR ≥ 4dB, the recognition accuracy for most modulation schemes can reach over 90%. Figure 7 This describes the change in modulation scheme recognition accuracy of BMR-Net during training as the number of iterations increases. The model accuracy improves rapidly between 0 and 80 iterations, and converges around 150 iterations.

[0065] The automatic modulation recognition method based on binary neural networks described in this invention effectively solves the problems of difficult training convergence and low model recognition accuracy. At the same time, it significantly reduces the computational load and storage requirements of deep learning-based automatic modulation recognition methods. Therefore, the automatic modulation recognition algorithm can be well deployed in embedded devices and has good practical application value.

Claims

1. An automatic modulation recognition method based on a binary neural network, characterized in that... Includes the following steps: S101: Construct an automatic modulation recognition network model BMR-Net based on a binary neural network; S102: Train the BMR-Net network model using the automatic modulation recognition training set. The training method adopts the gradient descent backpropagation method based on mini-batch training until the network model converges. Store the neural network weights obtained from the training and save the trained BMR-Net model as an inference model. S103: Acquire the wireless communication signal to be identified; S104: Perform I / Q preprocessing on the acquired signal; S105: Input the wireless communication signal to be identified into the inference model described in step S102; S106: Output predicted signal modulation identification type; The network model BMR-Net includes: one 2D convolutional layer (Conv2D), one normalization layer (BN), six residual blocks, four max pooling layers, three fully connected layers, two dropout layers, and a softmax layer. The output of the 2D convolutional layer (Conv2D) is connected to the input of the normalization layer (BN). The output of the normalization layer (BN) is connected to the input of the max pooling layer (Maxpooling-1). The output of the max pooling layer (Maxpooling-1) is connected to the input of the residual block (basic block-1). The output of the residual block (basic block-1) is connected to the input of the residual block (basic block-2). The output of the residual block (basic block-2) is connected to the input of the max pooling layer (Maxpooling-2). The output of the max pooling layer (Maxpooling-2) is connected to the input of the residual block (basic block-3). The output of the residual block (basic block-3) is connected to the input of the residual block (basic block-4). The output of block-4 is connected to the input of the max-pooling layer Maxpooling-3; the output of Maxpooling-3 is connected to the input of the residual block basic block-5; the output of basic block-5 is connected to the input of the residual block basic block-6; and the output of basic block-6 is connected to the input of the residual block basic block-6. The output of block-6 is connected to the input of the max pooling layer Maxpooling-4. The output of Maxpooling-4 is connected to the input of the Flatten layer. The output of the Flatten layer is connected to the input of the fully connected layer Fc-1. The output of the fully connected layer Fc-1 is connected to the input of the Dropout-1 layer. The output of the Dropout-1 layer is connected to the input of the fully connected layer Fc-2. The output of the fully connected layer Fc-2 is connected to the input of the Dropout-2 layer. The output of the Dropout-2 layer is connected to the input of the fully connected layer Fc-3. The output of the fully connected layer Fc-3 is connected to the input of the Softmax layer. The output of the Softmax layer is the output of the network model BMR-Net. The six basic residual blocks have similar structures. Basic residual block-1, whose input is represented as x, consists of the following: 1) Binarization layer Sign: Binarizes the input of the residual block to -1 or 1. The calculation method is as follows: ； in, The input is after binarization; 2) Binary convolutional layer: Weights with binarization Performing binary convolution operations, because and Both are -1 or 1. The convolution operation is as follows: First, ... and If we perform a XOR operation and then count the number of -1s and 1s respectively as m and n, then the formula for calculating binary convolution is: ； Where 'a' is the output of the binary convolutional layer. Represents a two-dimensional convolution operation; 3) Normalized layer BN: ； Where b is the output of the normalization layer. Let a be the mean of a. Let a be the standard deviation of a; 4) A step connector connects x and... Adding them together, the output of basic block-1 is represented as: ; With this design, each residual block retains the original input information of the previous layer to a certain extent; During model training, assuming that the vector representation to be propagated from the binarization layer in the residual block to the previous layer is y, then during backpropagation, what is actually propagated is a second-order approximation of y. The calculation method is as follows: 。 2. The automatic modulation recognition method based on a binary neural network as described in claim 1, characterized in that: The network model BMR-Net takes I and Q signals as inputs and outputs the modulation type corresponding to the input signal.

3. The automatic modulation recognition method based on a binary neural network as described in claim 1, characterized in that, In step S102: During the training process, 60% of the RadioML 2016.10a dataset is used as the training set, 20% as the validation set, and 20% as the test set.

4. The automatic modulation recognition method based on a binary neural network as described in claim 1, characterized in that, The network model was built using the Python language and the Keras deep learning framework.

5. The automatic modulation recognition method based on a binary neural network as described in claim 1, characterized in that, The training process used an NVIDIA GeForce RTX 3090 graphics card and its accompanying software CUDA, with cross-entropy loss as the loss function. Batch gradient descent was used during training, with each batch of training data being 1024 and the number of iterations set to 200. The Adam algorithm was used for optimization, adjusting and optimizing the network hyperparameters until the network accuracy could no longer be improved.

6. The automatic modulation recognition method based on a binary neural network as described in claim 1, characterized in that: In step S103: The signal received by the wireless communication receiver is r(t), which is represented as: ； Where s(t) is the original signal after modulation, c is the channel loss caused by various reasons, and n(t) is additive white Gaussian noise.

7. The automatic modulation recognition method based on a binary neural network as described in claim 1, characterized in that, In step S104: The in-phase component I(j) and quadrature component Q(j) of the acquired signal are used as inputs for feature extraction, with a dimension of 128. Before identifying the signal modulation mode, the I / Q data are preprocessed by normalization. First, the mean values ​​of I(j) and Q(j) are calculated respectively. , : ； Then calculate the standard deviation of the two sets of data separately: ； Finally, the normalized I / Q data is calculated using the following formula: 。

Images