A signal modulation recognition method based on feature fusion

Abstract

The present invention discloses a signal modulation recognition method based on feature fusion, comprising: obtaining a data set of modulated signal samples; preprocessing the modulated signal samples, including extracting a spectrum and constellation diagram of the modulated signal samples; normalizing the spectrum and constellation diagram, and concatenating the normalized spectrum and constellation diagrams to obtain a corresponding fused feature diagram, and dividing the fused feature diagram into a training set and a test set; establishing a deep neural network model, the deep neural network model including a CNN network, an LSTM network, an attention mechanism module, and a classifier connected in sequence; inputting the fused feature diagram in the training set into the deep neural network model for training to obtain a trained deep neural network model; inputting the fused feature diagram of the actual received signal into the trained deep neural network model for recognition to obtain a modulation recognition result. This method enhances feature diversity and has high recognition accuracy and generalization capability.

Description

Technical Field

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

[0002] With the continuous development and widespread application of wireless communication technology, radio modulation recognition technology is becoming increasingly important to ensure the efficient use of radio resources and the safe operation of wireless communication systems. However, traditional signal modulation recognition technology, which relies primarily on the technician's prior knowledge, is no longer able to adapt to the increasingly complex communication environment. Therefore, automatic modulation recognition technology has emerged.

[0003] Existing automatic modulation recognition methods are mainly divided into two categories: signal modulation recognition based on decision theory and signal modulation recognition based on features. The former mainly classifies by calculating the likelihood ratio of the signal. This method requires sufficient prior knowledge when selecting the likelihood function, and the recognition accuracy is greatly affected by the channel. Feature-based signal modulation recognition technology has rapidly become mainstream thanks to the development of machine learning. This method mainly extracts the characteristic quantity of the signal and trains the neural network in a supervised or unsupervised manner to achieve signal classification and recognition. The following two solutions are mainly used:

[0004] One approach is to design neural network architectures that leverage the differences between different modulation signal types. Examples include convolutional long short-term memory units, inception modules, residual networks, and graph convolutional neural networks (GCNs). However, as the number and complexity of candidate modulations increase, the adaptive features of a single network model are affected by factors such as noise, making the extracted features less rational and prone to classification errors.

[0005] The second approach is to explore effective feature representations as input to deep learning networks. For example, various features such as high-order cumulants, amplitude, and phase can be calculated from the IQ components as network inputs. This approach primarily extracts single features for recognition, but modulation recognition accuracy is limited under low signal-to-noise ratio conditions and places high demands on the selected features. Changes to the candidate modulation set often require new feature selection, and adjustments and improvements are necessary for different application scenarios. Summary of the Invention

[0006] The purpose of the present invention is to address the above problems and propose a signal modulation recognition method based on feature fusion, which enhances feature diversity and has high recognition accuracy and generalization ability.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] The present invention proposes a signal modulation recognition method based on feature fusion, comprising the following steps:

[0009] S1. Obtain a modulated signal sample data set;

[0010] S2. Preprocessing the modulated signal samples in the data set, the preprocessing includes extracting the spectrum and constellation diagram of the modulated signal samples;

[0011] S3, normalizing the spectrum graph and constellation graph, and concatenating the normalized spectrum graph and constellation graph of the same modulated signal sample to obtain a corresponding fused feature graph, and dividing the fused feature graph into a training set and a test set;

[0012] S4. Establish a deep neural network model. The deep neural network model includes a CNN network, an LSTM network, an attention mechanism module, and a classifier connected in sequence. The CNN network is used to extract signal spatial features, the LSTM network is used to extract signal temporal features, and the attention mechanism module is used to capture key spatiotemporal features.

[0013] S5. Input the fused feature map in the training set into the deep neural network model for training to obtain a trained deep neural network model;

[0014] S6. Input the fusion feature map of the actual received signal into the trained deep neural network model for recognition to obtain the modulation recognition result.

[0015] Preferably, the CNN network includes a first convolutional layer, a first Dropout layer, a first BN layer, a second convolutional layer, a second Dropout layer and a second BN layer connected in sequence.

[0016] Preferably, the output end of the LSTM network is further provided with a third Dropout layer.

[0017] Preferably, the classifier includes a first fully connected layer, a Relu activation function, a fourth Dropout layer and a second fully connected layer connected in sequence.

[0018] Preferably, the coefficient of each Dropout layer is set to 0.1.

[0019] Preferably, the modulated signal samples in the data set are IQ signals.

[0020] Preferably, the dataset uses the RadioML2016.10a public dataset.

[0021] Preferably, the loss function of the deep neural network model adopts the cross entropy function with the l2 regularization term added, and the formula is as follows:

[0022]

[0023] In the formula, m is the number of modulation signal categories, y is the actual value, is the predicted value of the deep neural network model, is the l2 regularization term, λ is the regularization parameter, n is the number of iterations, w i is the model weight parameter of the i-th iteration.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: the method converts the modulated signal into image representations of a spectrum diagram and a constellation diagram respectively, extracts the spectrum diagram and constellation diagram of the signal, fuses and splices them as network input, realizes feature compensation, enhances feature diversity, utilizes CNN and LSTM to extract the spatiotemporal features of the signal, and adds an Attention model to extract important sequence features, with a high recognition accuracy. For example, when the signal-to-noise ratio is -2dB, the classification accuracy reaches 90.04%, and when the signal-to-noise ratio is 18dB, the classification accuracy is 97.65%; and by adding a dropout layer to the CNN network, LSTM network and classifier, it ensures that the fitting ability is improved while taking into account the generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Flow chart of the signal modulation recognition method based on feature fusion of the present invention;

[0026] Figure 2 Schematic diagram of the structure of the deep neural network model of the present invention;

[0027] Figure 3 This is a recognition accuracy graph of the signal modulation recognition method based on feature fusion of the present invention;

[0028] Figure 4 Schematic diagram of confusion matrix for different signal-to-noise ratios of the present invention;

[0029] Figure 5 This is a graph of recognition accuracy for different feature inputs under the deep neural network model of the present invention;

[0030] Figure 6 Input recognition accuracy graphs of different network models into the fusion feature graph of the present invention;

[0031] Figure 7 This is a comparison diagram of the present invention and the existing model. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0033] It should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein in the specification of this application are for the purpose of describing specific embodiments only and are not intended to limit this application.

[0034] To improve modulation recognition performance, a signal modulation recognition method based on the fusion of spectrogram and constellation diagram features is proposed. First, the modulated signal is converted into image representations of a spectrogram and a constellation diagram, respectively, and the fused representations are used as network input. A convolutional neural network (CNN) is then used to extract spatial features of the signal. The CNN-extracted features are then passed through a long short-term memory (LSTM) network to extract temporal features. Finally, an attention mechanism is used to capture key spatiotemporal features. Experimental results show that the fusion of spectrogram and constellation diagram input achieves feature compensation and enhances feature diversity, resulting in better recognition performance than existing signal modulation recognition methods.

[0035] like Figure 1-7 As shown, a signal modulation recognition method based on feature fusion includes the following steps:

[0036] S1. Obtain a modulated signal sample data set.

[0037] In one embodiment, the modulated signal samples in the data set are IQ signals.

[0038] In one embodiment, the dataset uses the RadioML 2016.10a public dataset.

[0039] S2. Preprocessing the modulated signal samples in the data set, where the preprocessing includes extracting the spectrum and constellation diagram of the modulated signal samples.

[0040] The signal's time-frequency distribution reflects the distribution of signal energy on the time-frequency plane and can be directly used for signal modulation identification. By converting the time-frequency spectrogram into a grayscale image, with pixel values ​​representing the energy at the corresponding time-frequency nodes, the radio signal modulation classification problem is transformed into an image recognition problem. The fundamental approach to converting signals into the frequency domain is the Fourier transform (FT). The FT transform transforms signals from the time domain to the frequency domain. Its advantage is that it converts difficult-to-process time domain signals into the simpler, more manageable frequency domain for analysis and processing, simplifying the computational complexity of signal processing. To facilitate computer calculations, signal functions must be defined on a discrete time series. Therefore, computers generally use discrete Fourier transform formulas for calculations, such as the fast Fourier transform, to reduce complexity and increase computational speed.

[0041] constellation Figure 1 It is generally called a two-dimensional representation of a modulated signal. The samples are plotted as scattered points on the complex plane (also called the IQ plane). Its horizontal coordinate (X-axis) and vertical coordinate (Y-axis) represent the I component and Q component, respectively. Each point in the diagram can contain four pieces of information. Specifically, the projection of the point on the X-axis defines the peak amplitude of the in-phase component, and the projection of the point on the Y-axis defines the peak amplitude of the orthogonal component. The length of the line (vector) connecting the point to the origin is the peak amplitude of the signal element (the combination of the X component and the Y component), and the angle between the line and the X-axis is the phase of the signal element. Due to its powerful perception ability, it is often used for automatic signal modulation type and order information analysis.

[0042] S3. Normalize the spectrum graph and constellation graph, and concatenate the normalized spectrum graph and constellation graph of the same modulated signal sample to obtain the corresponding fused feature graph, and divide the fused feature graph into a training set and a test set.

[0043] In order to fully explore the information contained in the data and make the extracted characteristic information complementary and diverse, this application preprocesses the modulated signal samples in the data set, extracts the constellation diagram and spectrum diagram of the signal, and then uses the serial fusion method to obtain the fusion information (fusion feature map). The splicing formula is as follows:

[0044] F fusion =[F1, F2]

[0045] Where, F1 represents the normalized spectrum diagram, F2 represents the normalized constellation diagram, and F fusion Represents a fused feature map. A fused feature map is a 28×28 two-dimensional grayscale image. Fusion of the spectrum map and constellation map can improve signal feature extraction and modulation signal recognition.

[0046] S4. Establish a deep neural network model. The deep neural network model includes a CNN network, an LSTM network, an attention mechanism module and a classifier connected in sequence. Among them, the CNN network is used to extract signal spatial features, the LSTM network is used to extract signal temporal features, and the attention mechanism module is used to capture key spatiotemporal features.

[0047] In one embodiment, the CNN network includes a first convolutional layer, a first Dropout layer, a first BN layer, a second convolutional layer, a second Dropout layer, and a second BN layer connected in sequence.

[0048] In one embodiment, the output end of the LSTM network is further provided with a third Dropout layer.

[0049] In one embodiment, the classifier includes a first fully connected layer, a Relu activation function, a fourth Dropout layer, and a second fully connected layer connected in sequence.

[0050] In one embodiment, the coefficient of each Dropout layer is set to 0.1.

[0051] like Figure 1 、 2 As shown, the spectrum diagram and constellation diagram of different modulated signals can represent the differences between the modulated signals, and the fusion of the spectrum diagram and constellation diagram features can make full use of the diversity and complementarity between the two features to achieve better signal modulation recognition effect. Therefore, this application converts the signal modulation recognition problem into an image recognition problem, and uses a convolutional neural network (CNN network) and a long short-term memory network (LSTM network) to extract the spatiotemporal features of the signal, and uses an attention mechanism to capture the key features extracted by the neural network, which has good noise resistance. Specifically:

[0052] First, a convolutional neural network performs convolution operations to obtain high-level features (signal spatial features) and inputs them into a long short-term memory network. The long short-term memory network then receives the tensor processed by the convolutional neural network and extracts temporal features. Finally, an attention mechanism module (ATT) calculates the weights of the spatiotemporal features of the CNN and LSTM networks to obtain a complete feature representation of the sequence. A classifier is used to predict the signal modulation type. For example, if the dataset contains 11 modulation signals, the number of nodes is set to 11, and the output is an 11-dimensional probability vector. The index corresponding to the maximum value is the corresponding classification result.

[0053] The CNN network is set up as a two-layer network, each consisting of a convolutional layer, a dropout layer, and a batch normalization layer connected in sequence. The convolutional layers are two-dimensional. Zero padding is applied to the input samples to maintain the image size before and after the convolution operation, allowing for full learning of the edge features of the samples. Both layers use the same parameters: a 3×3 convolution kernel size and a convolution stride of 2. A ReLU activation function is applied to the output of each convolutional layer. After two convolutional layers, the output feature size is 128×(7×7).

[0054] Batch Normalization (BN) is used after each convolution operation to normalize the data distribution to speed up the convergence of the network and improve the accuracy. (1) ,...x (l) ,...x (d) ), for its k-th dimension input x (k) The BN process can be expressed as:

[0055]

[0056] The input of the CNN network of this application is a two-dimensional grayscale image, so in the above formula, d = 2, x = (x (1) , x (2) ). E[x (k) ] represents the average value of the input neurons in a batch of data, Represents the standard deviation of the input values ​​of each neuron. To ensure that the distribution of the transformed data remains unchanged, two parameters γ and β are introduced. Then:

[0057]

[0058] When β (k) =E[x (k) ], When , the network can recover the learned feature distribution y (k) .

[0059] The Long Short-Term Memory (LSTM) network receives the upper-layer output signal and extracts its temporal features. LSTM addresses data dependency by setting up gate units and memory cells within the neural network. The number of neurons is set to 128, and the feature size of the LSTM output is 128×49. A CNN is used to extract spatial features of the signal, and then an LSTM is used to extract temporal features based on these spatial features. To prevent non-critical or excessive features from influencing the prediction results, an attention mechanism is introduced to extract the key features of the sequence.

[0060] The classifier includes a first fully connected layer, a Relu activation function, a fourth Dropout layer, and a second fully connected layer connected in sequence. The number of neurons in the first fully connected layer is 128, which receives the features processed by the upper layer. The output feature size is 128×1. The second fully connected layer is also called the output layer. It uses the Softmax activation function to meet the requirements of multi-classification. The signal category in this embodiment is 11. According to the structure of the network, a dropout layer is added to the output of the convolutional layer of the CNN network, the output of the LSTM network, and the output of the Relu activation function, and the coefficients are all set to 0.1. While ensuring the improvement of fitting ability, the generalization ability is also taken into account.

[0061] The Relu activation function is a common neural activation function with sparsity, which enables the sparse model to better mine relevant features and fit the training data. Relu is used in the convolutional layer to accelerate the convergence of the model and improve the training performance of the neural network.

[0062] The Softmax activation function is commonly used in multi-classification problems as the activation function for the network output layer. The Softmax function normalizes the output values, converting them into probabilities where the sum of all probabilities equals 1. In deep neural networks, the final fully connected layer uses the Softmax function for classification and recognition, ultimately outputting the probability of each signal.

[0063] S5. Input the fused feature map in the training set into the deep neural network model for training to obtain a trained deep neural network model.

[0064] In one embodiment, the loss function of the deep neural network model adopts the cross entropy function with an l2 regularization term added, and the formula is as follows:

[0065]

[0066] In the formula, m is the number of modulation signal categories, y is the actual value, is the predicted value of the deep neural network model, is the l2 regularization term, λ is the regularization parameter, n is the number of iterations, w i is the model weight parameter of the i-th iteration.

[0067] The accuracy evaluation model is evaluated by identification. The accuracy calculation formula is:

[0068]

[0069] Where n is the number of samples, is the predicted label of the i-th sample, y iis the true label of the i-th sample. The label is a one-hot vector. For example, if there are 11 types of modulation signals, 11 categories are required. Each modulation signal corresponds to a label of length 11, and only one can be 1; the others are 0. For example, the label corresponding to the first modulation signal is [10000000000]. The higher the accuracy, the better the recognition effect of this article.

[0070] S6. Input the fusion feature map of the actual received signal into the trained deep neural network model for recognition to obtain the modulation recognition result.

[0071] The following is a detailed description through specific embodiments.

[0072] This example uses the RadioML2016.10a dataset, which is a commonly used dataset sample in current modulation recognition research. It contains 11 modulation signal types, including three analog modulation signals: double-sideband amplitude modulation (AM-SDB), single-sideband amplitude modulation (AM-SSB), and wideband frequency modulation (WBFM); and eight digital modulation signals: octal phase shift keying (8PSK), BPSK, continuous phase frequency shift keying (CPFSK), Gaussian frequency shift keying (GFSK), four-pulse amplitude modulation (PAM4), hexadecimal quadrature amplitude modulation (16QAM), and sixty-four-quadrature amplitude modulation (64QAM). Modulation (64QAM) and Quadrature Phase Shift Keying (QPSK). The ratio of the training set to the test set is 7:3. The specific parameters of the dataset are shown in Table 1:

[0073] Table 1 RadioML2016.10a dataset parameters

[0074]

[0075] The optimization algorithm is the Adam algorithm, which is an improvement of the stochastic gradient descent algorithm and a combination of the Momentum algorithm and the RMSprop algorithm. The cross entropy function with the l2 regularization term is used to calculate the neural network loss value as the network performance indicator. Its expression is:

[0076]

[0077] In the formula, m is the number of modulation signal categories, y is the actual value, is the predicted value of the deep neural network model, is the l2 regularization term, which mainly reduces the complexity of the network and prevents overfitting. λ is the regularization parameter, n is the number of iterations, and w i is the model weight parameter of the i-th iteration, and the network weight is reversely adjusted through the optimization algorithm.

[0078] like Figure 3 As shown in the figure, the recognition accuracy (precision) gradually improves with the increase of the signal-to-noise ratio. When the signal-to-noise ratio is between -20dB and 10dB, the recognition accuracy is below 35%. When the signal-to-noise ratio is between -10dB and -2dB, the signal recognition accuracy increases rapidly, reaching 90.04% at -2dB. When the signal-to-noise ratio is between 0dB and 8dB, the recognition accuracy fluctuates slightly, and then stabilizes at 8dB. When the signal-to-noise ratio reaches 18dB, the recognition accuracy reaches the highest level of 97.65%.

[0079] like Figure 4 As shown in Figure 2, the confusion matrix is ​​used to observe the recognition of each category. The horizontal axis of the confusion matrix is ​​the predicted label of the sample, and the vertical axis is the true label of the sample. Figure 4 The confusion matrix under different signal-to-noise ratios is shown. A signal-to-noise ratio of -2dB shows good results, indicating that the proposed method has good recognition accuracy at low signal-to-noise ratios. However, confusion issues arise between WBFM and AM-SSB, as well as between WBFM and AM-DSB. This is because they are both analog modulations, and the signals are generated using the same audio source signal, with silent segments, making some of their features more difficult to distinguish. When the signal-to-noise ratio is 18dB, the confusion situation is significantly improved because the proposed spatial and temporal features can more accurately capture the periodic trends corresponding to the modulation type at high signal-to-noise ratios.

[0080] Compared to single features, combining multiple features as network input can better leverage feature compensation, enhance feature diversity, and thus improve recognition accuracy. Furthermore, different feature combinations can significantly impact recognition results. The experiment maintained the same deep neural network model, using both single and multi-feature combinations as input features to verify the impact of different input features on recognition results.

[0081] like Figure 5 As shown in the figure, for a single input feature, when the signal-to-noise ratio (SNR) is 18dB, the recognition accuracy of the constellation diagram is 93.51%, and the recognition accuracy of the spectrum diagram is 84.16%, demonstrating that the constellation diagram's feature representation is superior in the deep neural network model designed in this paper. For multiple features, when the SNR is 18dB, the recognition accuracy of the power spectrum + constellation diagram is 92.51%, and the recognition accuracy of the spectrum diagram + constellation diagram is 97.65%, indicating that the feature combination selected in this paper achieves better feature compensation and higher recognition accuracy. Furthermore, the recognition accuracy of multi-feature input is generally better than that of single-feature input.

[0082] Different network model parameters will also have a great impact on the recognition results. The experiment fixed the network input to spectrum + constellation diagram; by setting the network model and parameters, the recognition effect of different networks was verified. Figure 6 As shown in the data, when the signal-to-noise ratio (SNR) is 18dB, the CNN's overall accuracy is relatively low, reaching 81.24%. In comparison, the LSTM's accuracy improves, reaching 88.73%. The attention-based long short-term memory (Att-LSTM) network surpasses the previous two models, achieving a recognition accuracy of 91.24%. Our proposed method (Att-CNN-LSTM) combines the advantages of these models, achieving a recognition accuracy of 97.65%. This demonstrates the effectiveness of our method based on extensive experimental results. Specifically, the Att-CNN-LSTM is a signal modulation recognition method proposed in this paper that fuses spectrogram and constellation features. This deep neural network model utilizes a joint recognition framework of a convolutional neural network (CNN) based on an attention mechanism and a long short-term memory (LSTM).

[0083] In order to verify the superiority of the signal modulation recognition model in this paper, the proposed Att-CNN-LSTM model is compared with existing signal recognition models, including the CLDNN model, DAE-LSTM model, NACF model and TFA-SCNN model. Among them, CLDNN is a recognition model based on convolutional long short-term deep neural network, reference: NEWest, TO'Shea. Deep architectures for modulation recognition [C]. Baltimore, America: 2017 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN), 2017: 1-6. DAE-LSTM is a recognition model based on LSTM denoising auto-encoder, reference: Z.Ke and H.Vikalo, Real-Time Radio Technology and Modulation Classification via an LSTM Auto-Encoder [J]. IEEE Transactions on Wireless Communications, 2022, 21(1): 370-382. NACF ​​represents a novel attention cooperative model, which includes recurrently connected CNN (CCNN), bidirectional RNN (BRNN), a fusion strategy based on global average and maximum pooling (GAMP), and auxiliary classification GAN (ACGAN). Reference: Chen S, Zhang Y, He Z, et al. A Novel Attention Cooperative Framework for Automatic Modulation Recognition [J]. IEEE Access, 2020, 8: 15673-15686. TFA-SCNN is a joint recognition model of time-frequency attention mechanism and convolutional neural network based on spectrum graph. Reference: Lin S, Zeng Y, Gong Y. Learning of Time-Frequency Attention Mechanism for Automatic Modulation Recognition [J]. IEEE Wireless Communications Letters, 2022, 11 (4): 707-711. In order to ensure the reliability of the experiment, all experiments were carried out in the same experimental environment and repeated 10 times. The average value of the 10 experiments was taken as the final experimental result. The experimental results are shown in the figure. Figure 7 shown.

[0084] Depend on Figure 7 As can be seen, the overall recognition accuracy of the five models improves with increasing signal-to-noise ratio (SNR). This is because as the SNR increases, noise decreases, interference gradually decreases, and the effective features contained in the signal increase, leading to a continuous increase in the model's recognition accuracy. Between the SNR range of -20dB and -2dB, the recognition performance of the four control models is similar and all fall short of our Att-CNN-LSTM model. In particular, at a -2dB SNR, our proposed model achieves a recognition accuracy of 90.04% for modulated signals, while the CLDNN, DAE-LSTM, NACF, and TFA-SCNN models achieve no more than 84% at the same SNR. Therefore, our algorithm exhibits high recognition accuracy at low SNRs. Our model also achieves high recognition accuracy at high SNRs. When the SNR is 18dB, the proposed Att-CNN-LSTM algorithm achieves a recognition accuracy of 97.65%. At the same signal-to-noise ratio (SNR), the recognition accuracy of the CLDNN model, DAE-LSTM model, NACF model, and TFA-SCNN model was 84.16%, 92.75%, 94.10%, and 92.18%, respectively. Our model achieved improvements of 13.49%, 4.9%, 3.55%, and 5.47%, respectively, compared to these four models. In summary, our Att-CNN-LSTM model achieved higher recognition accuracy than other popular models, both at low and high SNRs, demonstrating the superiority of our model.

[0085] The signal modulation recognition method proposed in this application extracts the signal's spectrum diagram and constellation diagram respectively, fuses and splices them as network input, realizes feature compensation, enhances feature diversity, uses CNN and LSTM to extract the signal's spatiotemporal features, and adds an Attention model to extract important sequence features. It has a high recognition accuracy. For example, when the signal-to-noise ratio is -2dB, the classification accuracy reaches 90.04%, and when the signal-to-noise ratio is 18dB, the classification accuracy is 97.65%. By adding a dropout layer to the CNN network, LSTM network and classifier, it ensures that the fitting ability is improved while taking into account the generalization ability.

[0086] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] The above-described embodiments merely represent specific and detailed examples of the present application and should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A signal modulation recognition method based on feature fusion, characterized by: The signal modulation recognition method based on feature fusion comprises the following steps: S1. Obtain a modulated signal sample data set; S2. Preprocessing the modulated signal samples in the data set, wherein the preprocessing includes extracting a spectrum diagram and a constellation diagram of the modulated signal samples; S3, normalizing the spectrum graph and constellation graph, and concatenating the normalized spectrum graph and constellation graph of the same modulated signal sample to obtain a corresponding fused feature graph, and dividing the fused feature graph into a training set and a test set; S4. Establish a deep neural network model, wherein the deep neural network model includes a CNN network, an LSTM network, an attention mechanism module, and a classifier connected in sequence, wherein the CNN network is used to extract signal spatial features, the LSTM network is used to extract signal temporal features, and the attention mechanism module is used to capture key spatiotemporal features; The CNN network includes a first convolutional layer, a first Dropout layer, a first BN layer, a second convolutional layer, a second Dropout layer and a second BN layer connected in sequence; The loss function of the deep neural network model adopts the cross entropy function with the l2 regularization term added, and the formula is as follows: In the formula, m is the number of modulation signal categories, y is the actual value, is the predicted value of the deep neural network model, is the l2 regularization term, λ is the regularization parameter, n is the number of iterations, ω i is the model weight parameter of the i-th iteration; S5. Inputting the fused feature map in the training set into the deep neural network model for training to obtain a trained deep neural network model; S6. Input the fusion feature map of the actual received signal into the trained deep neural network model for recognition to obtain the modulation recognition result.

2. The signal modulation recognition method based on feature fusion according to claim 1, wherein: The output end of the LSTM network is further provided with a third Dropout layer.

3. The signal modulation recognition method based on feature fusion according to claim 1, wherein: The classifier includes a first fully connected layer, a Relu activation function, a fourth Dropout layer and a second fully connected layer which are connected in sequence.

4. The signal modulation identification method based on feature fusion according to any one of claims 1 to 3, characterized in that: The coefficient of each Dropout layer is set to 0.

1.

5. The signal modulation identification method based on feature fusion according to claim 1, wherein: The modulated signal samples in the data set are IQ signals.

6. The signal modulation recognition method based on feature fusion according to claim 5, characterized in that: The dataset uses the RadioML2016.10a public dataset.

Images