Communication signal modulation recognition method and device based on multi-scale feature fusion network

Abstract

The application provides a modulation recognition method and device based on a multi-scale feature fusion network, reads an input signal, and pre-processes the input signal to obtain a baseband signal; signal feature vectors are obtained by performing feature extraction on the baseband signal; the signal feature vectors are input into a trained multi-scale feature fusion network model to select the real type of the input signal. The scale feature fusion network has three channels, and multiple scale signal feature maps can be obtained through the three channels to improve the signal modulation recognition accuracy. Moreover, the application considers the expression ability and size of the model, has smaller memory requirements, shorter training time, and is easier to adjust and deploy. The application extracts five kinds of signal features of the input signal as the input of the network model, and compared with directly inputting the time domain signal, the signal recognition accuracy is obviously improved, and the application has stronger anti-noise ability and is not easily affected by noise.

Description

Technical Field

[0001] This invention belongs to the field of signal recognition technology, specifically relating to a communication signal modulation recognition method and apparatus based on a multi-scale feature fusion network. Background Technology

[0002] With the development of information technologies such as radio communication, satellite navigation, network communication, and software-defined radio, signal recognition is playing an increasingly important role in military applications, and deep learning algorithms have been widely applied in the field of signal recognition.

[0003] Existing technology describes a technical solution for signal recognition using deep learning networks. In this solution, a VGG16 convolutional neural network is used for signal recognition. This VGG16 convolutional neural network consists of 13 convolutional layers and 3 fully connected layers. The large number of parameters and deep layers give the VGG network a strong fitting ability, but its disadvantages are also obvious: long training time, which is not conducive to parameter tuning; large memory consumption, which is not conducive to deployment; and the model is prone to overfitting when the training dataset is small. Furthermore, when using the VGG16 convolutional neural network to recognize time-domain signal data, the network's recognition results are easily affected by noise. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a communication signal modulation recognition method and apparatus based on a multi-scale feature fusion network. The technical problem to be solved by this invention is achieved through the following technical solution:

[0005] In a first aspect, the present invention provides a communication signal modulation recognition method based on a multi-scale feature fusion network, comprising:

[0006] S100: Read the input signal and preprocess the input signal to obtain the baseband signal;

[0007] S200, feature extraction is performed on the baseband signal to obtain a signal feature vector;

[0008] S300, the signal feature vector is input into the trained multi-scale feature fusion network model, so that the multi-scale feature fusion network model uses convolutional kernels of different sizes in three channels to process the signal features to obtain feature maps of different scales, uses a fusion layer to fuse the feature maps of different scales output by the three channels, uses a fully connected layer to process the fusion result, and then uses an output layer to classify the probability that the input signal belongs to a different type.

[0009] S400, select the type with the highest probability as the true type of the input signal.

[0010] Secondly, the present invention provides a communication signal modulation identification device based on a multi-scale feature fusion network, comprising:

[0011] The reading module is configured to read the input signal and preprocess the input signal to obtain the baseband signal;

[0012] The extraction module is configured to extract features from the baseband signal to obtain a signal feature vector;

[0013] The recognition module is configured to input the signal feature vector into a trained multi-scale feature fusion network model, so that the multi-scale feature fusion network model processes the signal features using convolutional kernels of different sizes in three channels to obtain feature maps of different scales, uses a fusion layer to fuse the feature maps of different scales output from the three channels, uses a fully connected layer to process the fusion result, and then uses an output layer to classify the probability that the input signal belongs to a different type.

[0014] The selection module is configured to select the type with the highest probability as the true type of the input signal.

[0015] Beneficial effects:

[0016] This invention provides a modulation recognition method and apparatus based on a multi-scale feature fusion network. The method involves reading an input signal and preprocessing it to obtain a baseband signal; extracting features from the baseband signal to obtain a signal feature vector; and inputting the signal feature vector into a trained multi-scale feature fusion network model to select the true type of the input signal. The multi-scale feature fusion network of this invention has three channels, enabling the generation of signal feature maps from multiple scales, thereby improving the accuracy of signal modulation recognition. It also balances the expressive power and size of the model, requiring less memory, having a shorter training time, and being easier to deploy and debug. This invention extracts five signal features from the input signal as input to the network model. Compared to directly inputting the time-domain signal, this significantly improves signal recognition accuracy, has stronger noise resistance, and is less susceptible to noise.

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the modulation recognition process based on a multi-scale feature fusion network provided by this invention;

[0019] Figure 2 This is a flowchart of the communication signal modulation recognition process based on a multi-scale feature fusion network.

[0020] Figure 3 Diagram of the multi-scale feature fusion network structure;

[0021] Figure 4 The graph shows the change in the accuracy of neural network modulation recognition based on feature vectors as a function of the number of iterations.

[0022] Figure 5 The accuracy of signal recognition under different signal-to-noise ratios for neural network modulation recognition based on feature vectors is shown in the figure.

[0023] Figure 6 The graph shows the accuracy results for recognition based on feature vectors and based on time-domain signal modulation. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0025] Firstly, combining Figure 1 and Figure 2 This invention provides a communication signal modulation recognition method based on a multi-scale feature fusion network, comprising:

[0026] S100: Read the input signal and preprocess the input signal to obtain the baseband signal;

[0027] S200, feature extraction is performed on the baseband signal to obtain a signal feature vector;

[0028] The signal feature vector includes spectral line features, squared spectral line features, envelope features, entropy features, and higher-order cumulant features.

[0029] The signal feature vector of this invention includes spectral line features (t pl ,t c ,t pr ), spectral characteristics of square spectrum (t) 2pl ,t 2c ,t 2pr ), higher-order cumulant feature F e Signal envelope feature R, information entropy feature H x These features are combined into a one-dimensional feature vector V, which is then used as the input to the neural network.

[0030] V = [t] pl ,t c ,t pr ,t 2pl ,t 2c ,t 2pr ,F e ,R,H x ];

[0031] This invention can modulate and identify six types of signals: BPSK, QPSK, GMSK, 16QAM, PCM / FM, and USB.

[0032] S300, the signal feature vector is input into the trained multi-scale feature fusion network model, so that the multi-scale feature fusion network model uses convolutional kernels of different sizes in three channels to process the signal features to obtain feature maps of different scales, uses a fusion layer to fuse the feature maps of different scales output by the three channels, uses a fully connected layer to process the fusion result, and then uses an output layer to classify the probability that the input signal belongs to a different type.

[0033] S400, select the type with the highest probability as the true type of the input signal.

[0034] The actual types of the input signals include: BPSK, QPSK, GMSK, 16QAM, PCM / FM, and USB.

[0035] The multi-scale feature fusion network model of this invention uses convolutional kernels with receptive fields of different sizes to scan feature vectors to obtain multi-scale signal feature maps. These feature maps of different modalities are then fused and used to identify the signal modulation mode. This can more accurately characterize signal feature information, enhance the expressive power of the network model, and improve the accuracy of signal recognition.

[0036] Existing deep learning-based signal modulation recognition techniques have many parameters, deep layers, long training time, are not conducive to parameter tuning, consume a lot of memory, are not conducive to deployment, and are prone to overfitting when the training dataset is small. Compared with existing technologies, the modulation recognition method based on multi-scale feature fusion network proposed in this invention takes into account the expressive power and size of the model. Without reducing the recognition rate, it optimizes the network structure, reduces training time, reduces memory requirements, and makes the network easier to deploy.

[0037] Existing deep learning-based signal modulation recognition techniques use time-domain signal data as network input. Compared to existing techniques, the modulation recognition method based on a multi-scale feature fusion network proposed in this invention extracts time-domain signal feature vectors as model input, including five signal features: spectral line features, squared spectral line features, higher-order cumulant features, signal envelope features, and information entropy features. This method performs modulation recognition on six modulation signals: BPSK, QPSK, GMSK, 16QAM, PCM / FM, and USB, improving the model's recognition rate and enhancing its noise resistance.

[0038] In one specific embodiment, reference is made to... Figure 3The multi-scale feature fusion network model of this invention includes three channels, each channel comprising four convolutional layers and two pooling layers. The four convolutional layers and two pooling layers are connected in a sequential manner: two convolutional layers are connected to one pooling layer. The number of convolutional kernels in the convolutional layers are set to 64 and 128, respectively. The multidimensional result after convolution is flattened to one dimension through a flattening layer. The outputs of the three channels are connected to one fusion layer, and the output of the fusion layer is connected to three sequentially connected fully connected layers. The last fully connected layer is the output layer. In the three channels, the kernel size of the convolutional layers in the first channel is 1×1, the kernel size of the convolutional layers in the second channel is 3×3, and the kernel size of the convolutional layers in the third channel is 5×5.

[0039] This invention designs a multi-scale feature fusion network that employs early fusion, constructing a 3-channel, 7-layer convolutional network (only convolutional and fully connected layers are calculated). Each channel of the network has 4 convolutional layers, 2 pooling layers, 1 fusion layer, and 3 fully connected layers. The kernel sizes of the convolutional layers in the three channels are 1×1, 3×3, and 5×5, respectively. Five signal features—spectral line features, squared spectral line features, envelope features, entropy features, and higher-order cumulant features—are used to form a feature vector as the input to the network model. The convolutional and fully connected layers use ReLU as the activation function, the output layer uses softmax as the activation function, and the cross-entropy function is used as the loss function. The different kernel sizes in each channel's convolutional layers result in different receptive field sizes, leading to the output of feature maps at different scales. These feature maps are fused in the fusion layer, then passed through the fully connected layer, and finally classified in the output layer, outputting six different probability values ​​corresponding to six different signal types.

[0040] The convolutional layer (Conv) is the core of a convolutional neural network. Each convolutional layer can contain multiple convolutional kernels. These kernels slide across the input data, traversing every position in the input data from left to right and top to bottom. At each position, the kernel performs a convolution operation with a local region of the input data, producing an output value. The kernel then aggregates all the output values ​​to obtain an output feature map, which contains the feature information of the input data.

[0041] Pooling is a method of replacing the values ​​in a specific region with a specific value. Generally, there are two types of pooling: max pooling and mean pooling. Max pooling uses the maximum value in the region to replace all values ​​in that region; it is a linear function. Mean pooling, on the other hand, uses the average value in the region to replace all values; it is a non-linear function. Both max pooling and mean pooling can effectively capture image features, thus performing downsampling. Downsampling can significantly reduce image size and is not sensitive to image translation and rotation.

[0042] Fully connected layers, also known as dense layers, are the most commonly used neural network layers in deep learning. Each neuron in a fully connected layer is connected to all neurons in the previous layer, hence the name "fully connected network." In a fully connected layer, each neuron has its own weights and biases, which need to be trained to minimize the loss function. The output of a fully connected layer is a vector that can be used as the input to the next layer. Fully connected layers are commonly used for classification problems. The last fully connected layer in a neural network often uses the softmax function as its activation function, outputting the probability distribution for each class to perform classification.

[0043] The convolutional and fully connected layers of this invention use ReLU as the activation function:

[0044] ReLU(x) = max(0,x);

[0045] The output layer uses softmax as the activation function:

[0046]

[0047] And use the cross-entropy function as the loss function:

[0048]

[0049] y i p represents the true label of category i. i This represents the model's predicted value for i, k is the number of categories, and N is the total number of samples.

[0050] In one specific embodiment, S100 of the present invention includes:

[0051] S110, read the input signal;

[0052] S120, The input signal is transformed into a complex signal using the Hilbert transform;

[0053] S130 performs spectrum shifting on the transformed input signal to convert it into a baseband signal.

[0054] In one specific embodiment, the trained multi-scale feature fusion network model is pre-trained using a training set, the generation process of which includes:

[0055] Choose from the following six modulation signal types: BPSK, QPSK, GMSK, 16QAM, PCM / FM, and USB.

[0056] 500 signal samples of each type of modulation signal are generated at 5dB intervals between -10 and 20dB, with each signal sample having 5000 sampling points, for a total of 21000 sampling samples.

[0057] 60% of the sampled data is randomly selected as the training set, and the remainder is used as the validation set.

[0058] The accuracy curves of the training and validation sets in this invention show a rapid overall increase with significant local fluctuations in the first 50 iterations, then tend to flatten out after 50 iterations, ultimately achieving an accuracy of 96.67% for the training set and 95.71% for the validation set. Figure 4 As shown. The modulation recognition results of the multi-scale feature fusion network model of this invention on six types of signals under different signal-to-noise ratios in the validation set are as follows. Figure 5 As shown.

[0059] In one specific embodiment, the training process of the multi-scale feature fusion network model includes:

[0060] The sampled samples in the training set are input into a preset multi-scale feature fusion network model, so that the multi-scale feature fusion network model uses convolutional kernels of different sizes in three channels to process the features of the sampled samples to obtain feature maps of different scales. The fusion layer fuses the feature maps of different scales output by the three channels, the fully connected layer processes the fusion result, and then the output layer classifies the sampled samples to obtain the probability that the sampled samples belong to different types.

[0061] The loss value of the sampled sample is calculated based on the probability that the sampled sample belongs to a different type and the cross-entropy loss function. If the loss value is greater than the threshold, the weights of the multi-scale feature fusion network model are adjusted.

[0062] The process of training the multi-scale feature fusion network model with the sampled samples in the training set is repeated until the required number of iterations is reached;

[0063] The multi-scale feature fusion network model that has reached the required number of iterations is validated using the validation set, and the multi-scale feature fusion network model that passes the validation is determined as the successfully trained multi-scale feature fusion network model.

[0064] In one specific embodiment, before inputting the sampled samples from the training set into a preset multi-scale feature fusion network model, the communication signal modulation recognition method based on the multi-scale feature fusion network further includes:

[0065] The proportion of inactive neurons in the multi-scale feature fusion network model is set to 0; edge padding is set to "same"; the optimizer is set to the Adam optimization method; the learning rate is set to 0.001; the loss function is set to the cross-entropy loss function; the number of iterations is set to 200; and the validation frequency is set to 1.

[0066] Through repeated experiments, this invention has obtained better training parameters for the multi-scale feature fusion network as shown above, which can improve the training efficiency of the training model.

[0067] This invention extracts time-domain signal feature vectors as input to a multi-scale feature fusion network model. It extracts five signal features: spectral line features, squared spectral line features, envelope features, entropy features, and higher-order cumulant features, forming a feature vector that serves as input to the network model. Compared to directly inputting the time-domain signal, this significantly improves signal recognition accuracy, enhances noise resistance, and reduces susceptibility to noise. Figure 6 As shown.

[0068] Secondly, the present invention provides a communication signal modulation identification device based on a multi-scale feature fusion network, comprising:

[0069] The reading module is configured to read the input signal and preprocess the input signal to obtain the baseband signal;

[0070] The extraction module is configured to extract features from the baseband signal to obtain a signal feature vector;

[0071] The recognition module is configured to input the signal feature vector into a trained multi-scale feature fusion network model, so that the multi-scale feature fusion network model processes the signal features using convolutional kernels of different sizes in three channels to obtain feature maps of different scales, uses a fusion layer to fuse the feature maps of different scales output from the three channels, uses a fully connected layer to process the fusion result, and then uses an output layer to classify the probability that the input signal belongs to a different type.

[0072] The selection module is configured to select the type with the highest probability as the true type of the input signal.

[0073] This invention provides a modulation recognition method and apparatus based on a multi-scale feature fusion network. The method involves reading the input signal and preprocessing it to obtain a baseband signal; extracting features from the baseband signal to obtain a signal feature vector; and inputting the signal feature vector into a trained multi-scale feature fusion network model to select the true type of the input signal. The multi-scale feature fusion network of this invention has three channels, which can obtain signal feature maps at multiple scales, thereby improving the accuracy of signal modulation recognition. Furthermore, this invention balances the expressive power and size of the model, requiring less memory, having a shorter training time, and being easier to debug and deploy. This invention extracts five signal features from the input signal as input to the network model. Compared to directly inputting the time-domain signal, this significantly improves signal recognition accuracy, has stronger noise resistance, and is less affected by noise.

[0074] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0075] Although this application has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0076] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A communication signal modulation recognition method based on a multi-scale feature fusion network, characterized in that, include: S100: Read the input signal and preprocess the input signal to obtain the baseband signal; S200, feature extraction is performed on the baseband signal to obtain a signal feature vector; the signal feature vector includes spectral line features, square spectral line features, envelope features, entropy features, and higher-order cumulant features; S300, the signal feature vector is input into the trained multi-scale feature fusion network model, so that the multi-scale feature fusion network model uses convolutional kernels of different sizes in three channels to process the signal features to obtain feature maps of different scales, uses a fusion layer to fuse the feature maps of different scales output by the three channels, uses a fully connected layer to process the fusion result, and then uses an output layer to classify the probability that the input signal belongs to a different type. S400, Select the type with the highest probability as the true type of the input signal; S100 includes: S110, read the input signal; S120, The input signal is transformed into a complex signal using the Hilbert transform; S130, the transformed input signal is spectrum shifted to convert the transformed input signal into a baseband signal; The multi-scale feature fusion network model includes three channels, each channel including four convolutional layers and two pooling layers. The four convolutional layers and two pooling layers are connected in such a way that two convolutional layers are connected in sequence to one pooling layer. The output of each channel is connected to one fusion layer, and the output of the fusion layer is connected to three fully connected layers in sequence. The last fully connected layer is the output layer. Of the three channels, the kernel size of the convolutional layer in the first channel is... The kernel size of the convolutional layer in the second channel is The kernel size of the convolutional layer in the second channel is .

2. The communication signal modulation recognition method based on a multi-scale feature fusion network according to claim 1, characterized in that, The actual types of the input signals include: BPSK, QPSK, GMSK, 16QAM, PCM / FM, and USB.

3. The communication signal modulation recognition method based on a multi-scale feature fusion network according to claim 1, characterized in that, The multi-scale feature fusion network model that has been trained is obtained by pre-training using a training set, the generation process of which includes: Choose from the following six modulation signal types: BPSK, QPSK, GMSK, 16QAM, PCM / FM, and USB. 500 signal samples of each type of modulation signal are generated at 5dB intervals between -10 and 20dB, with each signal sample having 5000 sampling points, for a total of 21000 sampling points, and each sampling point being a sample. 60% of the sampled data is randomly selected as the training set, and the remainder is used as the validation set.

4. The communication signal modulation recognition method based on a multi-scale feature fusion network according to claim 3, characterized in that, The training process of the multi-scale feature fusion network model includes: The sampled samples in the training set are input into a preset multi-scale feature fusion network model, so that the multi-scale feature fusion network model uses convolutional kernels of different sizes in three channels to process the features of the sampled samples to obtain feature maps of different scales. The fusion layer fuses the feature maps of different scales output by the three channels, the fully connected layer processes the fusion result, and then the output layer classifies the sampled samples to obtain the probability that the sampled samples belong to different types. The loss value of the sampled sample is calculated based on the probability that the sampled sample belongs to a different type and the cross-entropy loss function. If the loss value is greater than the threshold, the weights of the multi-scale feature fusion network model are adjusted. The process of training the multi-scale feature fusion network model with the sampled samples in the training set is repeated until the required number of iterations is reached; The multi-scale feature fusion network model that has reached the required number of iterations is validated using the validation set, and the multi-scale feature fusion network model that passes the validation is determined as the successfully trained multi-scale feature fusion network model.

5. The communication signal modulation recognition method based on a multi-scale feature fusion network according to claim 4, characterized in that, Before inputting the sampled samples from the training set into the preset multi-scale feature fusion network model, the communication signal modulation recognition method based on the multi-scale feature fusion network further includes: The proportion of inactive neurons in the multi-scale feature fusion network model is set to 0; edge padding is set to "same"; the optimizer is set to the Adam optimization method; the learning rate is set to 0.001; the loss function is set to the cross-entropy loss function; the number of iterations is set to 200; and the validation frequency is set to 1.

6. A communication signal modulation recognition device based on a multi-scale feature fusion network, characterized in that, include: The reading module is configured to read the input signal and preprocess the input signal to obtain the baseband signal; The extraction module is configured to extract features from the baseband signal to obtain a signal feature vector; the signal feature vector includes spectral line features, square spectral line features, envelope features, entropy features, and higher-order cumulant features; The recognition module is configured to input the signal feature vector into a trained multi-scale feature fusion network model, so that the multi-scale feature fusion network model processes the signal features using convolutional kernels of different sizes in three channels to obtain feature maps of different scales, uses a fusion layer to fuse the feature maps of different scales output from the three channels, uses a fully connected layer to process the fusion result, and then uses an output layer to classify the probability that the input signal belongs to a different type. The selection module is configured to select the type with the highest probability as the true type of the input signal; The process of reading the input signal and preprocessing the input signal to obtain the baseband signal includes: S110, read the input signal; S120, The input signal is transformed into a complex signal using the Hilbert transform; S130, the transformed input signal is spectrum shifted to convert the transformed input signal into a baseband signal; The multi-scale feature fusion network model includes three channels, each channel including four convolutional layers and two pooling layers. The four convolutional layers and two pooling layers are connected in such a way that two convolutional layers are connected in sequence to one pooling layer. The output of each channel is connected to one fusion layer, and the output of the fusion layer is connected to three fully connected layers in sequence. The last fully connected layer is the output layer. Of the three channels, the kernel size of the convolutional layer in the first channel is... The kernel size of the convolutional layer in the second channel is The kernel size of the convolutional layer in the second channel is .

Images