A modulation recognition method and apparatus based on multimodal residual attention network

Abstract

This invention relates to a modulation recognition method and apparatus based on a multimodal residual attention network. The method includes: acquiring modulation signal data to be identified; performing energy normalization and amplitude / phase data calculation on the modulation signal data to be identified to obtain amplitude and phase data of the modulation signal to be identified; inputting the above data into an intelligent recognition network model to obtain the modulation mode of the modulation signal data to be identified; wherein the intelligent recognition network model includes: a multimodal feature fusion module, an ECA-based residual network module, a feature pyramid network module, and a classification module. This invention can identify a wide range of modulation types, has no limitation on signal length, solves the problems of low accuracy and poor noise resistance in traditional modulation recognition methods, and reduces the workload and uncertainty of manually designing feature parameters.

Description

Technical Field

[0001] This invention relates to intelligent identification technology for communication modulation modes in the field of wireless communication technology, and in particular to a modulation identification method and apparatus based on a multimodal residual attention network. Background Technology

[0002] In modern information technology, the electromagnetic field has an increasingly significant impact on social development, people's activities, and national security. Communication modulation signals are the primary carriers of information in the electromagnetic field, providing the foundation for wireless information exchange. In the field of military communications, communication modulation pattern identification technology is a crucial component of radio reconnaissance and electronic warfare, enabling the monitoring of enemy intelligence and the restriction of communications, thereby enhancing the combat effectiveness of the military. In the field of civilian communications, communication modulation pattern identification technology has been widely applied in radio management, helping to monitor unknown interference signals and illegal signals present during communication, thus improving the level of social modernization. It can be seen that communication modulation pattern identification has become a fundamental function in communication signal processing without prior knowledge or with limited prior knowledge.

[0003] With the development of modern information technology, the types of wireless communication devices are increasing, resulting in a wide variety of modulation styles. From the earliest analog modulation to digital modulation, multiple modulations are mixed together, leading to an increasingly complex electromagnetic environment and placing higher demands on modulation identification technology. Modulation type identification is of great significance in both military and civilian applications, and corresponding solutions need to be proposed.

[0004] Traditional modulation recognition techniques, after long-term development, can be broadly categorized into likelihood ratio recognition methods based on decision theory, statistical pattern recognition methods based on feature extraction, and knowledge-based methods. These methods have three main drawbacks: 1. They rely on expert experience for feature extraction and model building, making it difficult to complete quickly and with high quality when dealing with large amounts of data; 2. They struggle to adapt to complex and changing electromagnetic environments, requiring new processing schemes when noise, fading, and interference vary; 3. They are difficult to extract effective features when processing short burst signals, resulting in unsatisfactory recognition performance. Therefore, in today's complex and changing electromagnetic environment, developing efficient, adaptive, robust, and intelligent modulation recognition techniques is extremely important.

[0005] With the development of hardware and artificial intelligence algorithms, deep learning has demonstrated high adaptability and satisfactory results in many fields. Deep learning provides a new approach to solving modulation recognition. Compared with traditional methods, deep learning can automatically extract features from massive amounts of data, while also overcoming the problems of low recognition efficiency, low classification accuracy, poor robustness, and poor scalability that exist in traditional methods. Summary of the Invention

[0006] This invention provides a modulation recognition method based on a multimodal residual attention network to solve the problems of low accuracy and poor noise resistance of traditional modulation recognition methods.

[0007] The technical solution adopted by this invention to solve its technical problem is: to provide a modulation recognition method based on a multimodal residual attention network, comprising the following steps:

[0008] Acquire the modulation signal data to be identified;

[0009] The modulation signal data to be identified is processed by energy normalization and amplitude and phase data calculation to obtain the amplitude data and phase data of the modulation signal to be identified.

[0010] The energy-normalized modulation signal data to be identified, the amplitude data, and the phase data are input into the intelligent identification network model to obtain the modulation mode of the modulation signal data to be identified; wherein, the intelligent identification network model includes:

[0011] The multimodal feature fusion module is used to extract features from the energy-normalized modulation signal data to be identified, the amplitude data, and the phase data, and to fuse the extracted features together by splicing to obtain fused features;

[0012] The residual network module based on ECA is used to extract deep features from the fused features to obtain multi-level features;

[0013] The feature pyramid network module is used to perform global average pooling on each layer of the multi-level features, and then concatenate the processed results along the channel dimension to obtain multi-level fused features.

[0014] The classification module is used to classify the modulation mode of the modulation signal data to be identified based on the multi-level fusion features.

[0015] The multimodal feature fusion module includes:

[0016] The first feature extraction unit, composed of two-dimensional convolutional blocks, is used to extract features from the modulated signal data to be identified after energy normalization, and obtain the first feature.

[0017] The second feature extraction unit, composed of one-dimensional convolutional blocks, is used to extract features from the amplitude data to obtain the second feature.

[0018] The third feature extraction unit, composed of one-dimensional convolutional blocks, is used to extract features from the phase data to obtain the third feature.

[0019] The splicing unit is used to splice the first feature, the second feature, and the third feature to obtain the fused feature.

[0020] The ECA-based residual network module includes multiple stacked ECA residual blocks. Each ECA residual block extracts features through three one-dimensional convolutional blocks, then calculates the average value of each feature channel through global pooling, performs convolution processing on the feature channels, and calculates attention weights using the Sigmoid function to achieve weighted features.

[0021] The classification module includes:

[0022] A fully connected layer is used to map the multi-level fused features to the dimension of the number of categories;

[0023] The Softmax activation layer is used to convert the output of the fully connected layer into a probability distribution through the Softmax activation function.

[0024] The intelligent recognition network model is trained in the following way:

[0025] Collect modulated signal data with modulation type labels;

[0026] Energy normalization and amplitude-phase data calculation are performed on all modulated signal data to obtain the amplitude and phase data of each modulated signal data.

[0027] Training and test datasets were constructed based on energy-normalized modulated signal data, amplitude data, and phase data.

[0028] The intelligent recognition network model is trained using the training dataset. During training, a fixed length of data is randomly truncated from each data sample in the training dataset. During testing, data samples from the test dataset of arbitrary length are used for testing.

[0029] The loss function of the intelligent recognition network model is: FL(p k )=-α k (1-p k ) γ log(p k ), where FL(p k α is the loss function of the intelligent recognition network model. k p represents the category weight coefficient. k γ represents the probability of the correct category predicted by the intelligent recognition network model, and γ is the aggregation parameter.

[0030] The technical solution adopted by this invention to solve its technical problem is: to provide a modulation recognition device based on a multimodal residual attention network, comprising:

[0031] The acquisition module is used to acquire the modulation signal data to be identified;

[0032] The processing module is used to perform energy normalization and amplitude and phase data calculation on the modulation signal data to be identified, so as to obtain the amplitude data and phase data of the modulation signal to be identified.

[0033] An identification module is used to input the energy-normalized modulation signal data to be identified, the amplitude data, and the phase data into an intelligent identification network model to obtain the modulation mode of the modulation signal data to be identified; wherein, the intelligent identification network model includes:

[0034] The multimodal feature fusion module is used to extract features from the energy-normalized modulation signal data to be identified, the amplitude data, and the phase data, and to fuse the extracted features together by splicing to obtain fused features;

[0035] The residual network module based on ECA is used to extract deep features from the fused features to obtain multi-level features;

[0036] The feature pyramid network module is used to perform global average pooling on each layer of the multi-level features, and then concatenate the processed results along the channel dimension to obtain multi-level fused features.

[0037] The classification module is used to classify the modulation mode of the modulation signal data to be identified based on the multi-level fusion features.

[0038] The technical solution adopted by the present invention to solve its technical problem is: to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-mentioned modulation recognition method based on a multimodal residual attention network.

[0039] The technical solution adopted by the present invention to solve its technical problem is: to provide a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the above-mentioned modulation recognition method based on a multimodal residual attention network.

[0040] Beneficial effects

[0041] By adopting the above-mentioned technical solution, this invention has the following advantages and positive effects compared with the prior art: The intelligent recognition network model in this invention enhances the feature extraction capability of communication modulation signals through a multimodal feature fusion module, improving the recognition performance when the feature differences between different modulation types are small. Simultaneously, this intelligent recognition network model uses multi-level features, thus better adapting to signals of different SPS. Furthermore, this intelligent recognition network model uses an ECA-based residual network module as its backbone; the ECA mechanism can filter features effective for modulation recognition, thereby improving the performance of the intelligent recognition network model. The intelligent recognition network model of this invention can achieve the recognition of many modulation types without limitation on signal length, solving the problems of low accuracy and poor noise resistance in traditional modulation recognition methods, while reducing the workload and uncertainty of manually designing feature parameters. Attached Figure Description

[0042] Figure 1 This is a flowchart of the modulation recognition method based on a multimodal residual attention network according to the first embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the intelligent recognition network model in the first embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of the structure of a one-dimensional convolutional block and a two-dimensional convolutional block in the first embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of the structure of the ECA residual block in the first embodiment of the present invention;

[0046] Figure 5 This is a graph showing the change in accuracy of the training and test sets with the number of iterations in the first embodiment of the present invention;

[0047] Figure 6 This is the confusion matrix diagram of the simulation test set in the first embodiment of the present invention;

[0048] Figure 7 This is a graph showing the variation of the simulation test set with signal-to-noise ratio accuracy in the first embodiment of the present invention;

[0049] Figure 8 This is a graph showing the change in accuracy of the simulation test set as a function of the input signal length in the first embodiment of the present invention;

[0050] Figure 9 This is a confusion matrix diagram of the actual dataset in the first embodiment of the present invention. Detailed Implementation

[0051] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0052] The first embodiment of the present invention relates to a modulation recognition method based on a multimodal residual attention network, such as... Figure 1 As shown, it includes the following steps:

[0053] Step 1: Obtain the modulation signal data to be identified.

[0054] Step 2: Perform energy normalization and amplitude and phase data calculation on the modulation signal data to be identified to obtain the amplitude data and phase data of the modulation signal to be identified.

[0055] Step 3: Input the energy-normalized modulation signal data to be identified, the amplitude data, and the phase data into the intelligent identification network model to obtain the modulation mode of the modulation signal data to be identified.

[0056] like Figure 2 As shown, the intelligent recognition network model in this embodiment includes a multimodal feature fusion module, a residual network module based on ECA (Efficient Channel Attention Network), a Feature Pyramid Networks (FPN) module, and a classification module. The output of this intelligent recognition network model can be expressed as:

[0057] Y = S(G mean (f(X)))

[0058] Where f(·) represents the features extracted by the multimodal feature fusion module and the ECA-based residual network module, and G mean (·) represents the feature pyramid network module, and S(·) represents the classification module.

[0059] For f(·), the output can be written as:

[0060] f(X) = [F1, ..., F M ]

[0061] Where M represents the number of outputs. f(·) is based on a CNN, which features parameter sharing and local connections. f(·) also exhibits locality and translation invariance, meaning that each output F... m It only depends on the local input, that is:

[0062] Fm =f([x n ,…,x n+k ])

[0063] Where k is the size of the feature receptive field. Meanwhile, in any time segment [x]... n ,…,x n+k Since they are independent and identically distributed, and have the same statistical properties, we have:

[0064] mean(F i ) = mean(F j )

[0065] Where i and j represent arbitrary times. G mean (·) This mainly involves global average pooling of the time-sampled features F, which can be understood as estimating the mean of features F. Therefore, the model can adapt to data of different lengths; the longer the data, the better G becomes. mean The more accurate the estimate of mean(F) by f(X) is, the better the modulation recognition performance will be.

[0066] In this embodiment, the multimodal feature fusion module is used to extract features from the energy-normalized modulation signal data to be identified, the amplitude data, and the phase data, and then fuse the extracted features together to obtain fused features.

[0067] The multimodal feature fusion module in this embodiment includes: a first feature extraction unit, composed of two-dimensional convolutional blocks, used to extract features from the energy-normalized modulation signal data to be identified, to obtain a first feature; a second feature extraction unit, composed of one-dimensional convolutional blocks, used to extract features from the amplitude data, to obtain a second feature; a third feature extraction unit, composed of one-dimensional convolutional blocks, used to extract features from the phase data, to obtain a third feature; and a splicing unit, used to splice the first feature, the second feature, and the third feature to obtain a fused feature.

[0068] Figure 3 The diagram shows the specific structures of two-dimensional and one-dimensional convolutional blocks. The main structures include convolutional network layers, batch normalization layers, and ReLU activation layers. The multimodal feature fusion module in this embodiment enhances the feature extraction capability of communication modulation signals and improves recognition performance when feature differences between different modulation types are small.

[0069] The ECA-based residual network module in this embodiment is used for deep feature extraction of the fused features, resulting in multi-level features. This ECA-based residual network module includes four stacked ECA residual blocks. Figure 4This is a structural diagram of the ECA residual block. In the ECA residual block, features are extracted through three one-dimensional convolutional blocks, then the average value of each feature channel is calculated through global pooling, convolution is performed on the feature channels, and the attention weights are calculated using the Sigmoid function to achieve weighted feature processing.

[0070] This implementation uses multi-level features to better adapt to signals of different SPS. In addition, by using an ECA-based residual network module as the backbone, the ECA mechanism can filter features that are effective for modulation identification, thereby improving the performance of the intelligent identification network model.

[0071] In this embodiment, the feature pyramid network module performs global average pooling on each layer of the multi-level features and concatenates the processed results along the channel dimension to obtain multi-level fused features. This feature pyramid network module performs global average pooling on the four feature maps extracted from each ECA residual block and then concatenates them along the channel dimension to achieve the effect of multi-level feature fusion. The global average pooling used ensures that features from different levels can be fused together and is also a necessary means to adapt to signals of different lengths during inference.

[0072] The classification module in this embodiment is used to classify the modulation pattern of the modulation signal data to be identified based on the multi-level fusion features. This classification module includes: a fully connected layer, used to map the multi-level fusion features to the dimension of the number of categories; and a Softmax activation layer, used to convert the output of the fully connected layer into a probability distribution using the Softmax activation function. Assuming the input feature of this classification module is x, the probability p of the k-th category obtained after calculation by the fully connected layer and the Softmax activation layer... k It can be calculated using the following formula:

[0073]

[0074] Where w is the weight parameter in the fully connected layer, b is the bias parameter in the fully connected layer, and M is the total number of features.

[0075] The training process of the intelligent recognition network model in this embodiment is as follows:

[0076] Data preparation:

[0077] In this implementation, a large amount of labeled simulation data was generated based on MATLAB, containing nineteen modulation types: {"CW":0,"AM":1,"AM-DSB":2,"AM-SSB":3,"FM":4,"ASK":5,"BPSK":6,"QPSK":7,"OQPSK":8,"PI_4QPSK":9,"8PSK":10,"APSK16":11,"APSK32":12,"16QAM":13,"32QAM":14,"64QAM":15,"MSK":16,"2FSK":17,"4FSK":18}. This includes five analog modulation types and fourteen digital modulation types. To better reflect real signals, the simulation also considered the effects of different signal-to-noise ratios, SPS, frequency offsets, and modulation parameters. The specific parameter information contained in this dataset is shown in Table 1.

[0078] Table 1 Dataset Parameter Information Table

[0079]

[0080]

[0081] Each data sample is 16384 in length. There are a total of 2,280,000 simulation data samples in this study. 80% of the samples, totaling 18,240,000, are divided into the training set, and 20% of the samples, totaling 456,000, are divided into the test set.

[0082] Data processing:

[0083] Energy normalization and amplitude / phase data calculation are performed on all data, assuming each data sample Energy-normalized data S i It is obtained from the following formula:

[0084]

[0085] Amplitude data A i ′ and phase data We obtain it from the following formula:

[0086]

[0087] A i ′=A i / A mean

[0088] Model training:

[0089] The specific parameter settings for model training are shown in Table 2.

[0090] Table 2 Specific parameters for model training

[0091]

[0092] The Adam (Adaptive Moment Estimation) optimizer is an adaptive optimization algorithm that adjusts the learning rate based on historical gradient information. It's a variant of gradient descent used to update the weights of a neural network. Warmup uses a small learning rate at the beginning of training, and then gradually increases the pre-set learning rate after the system has stabilized. Cosine decay is a commonly used learning rate adjustment strategy; the learning rate initially increases slightly and then decreases linearly, while the decrease is achieved using a cosine function.

[0093] Batch size is the amount of data input to the model during each backpropagation, Epoch is the number of iterations across all data. Input shape is the size of the model's input during training.

[0094] In this embodiment, the loss function of the intelligent recognition network model is Focal loss. Focal loss is a modification of the standard cross-entropy loss. It reduces the loss weight of easily classified samples and increases the loss weight of difficult-to-classify samples by introducing an adjustment factor. It is calculated as follows:

[0095] FL(p k )=-α k (1-p k ) γ log(p k )

[0096] Where, p k α represents the probability of the correct category predicted by the intelligent recognition network model. k γ is the class weight coefficient, used to handle the problem of imbalanced class numbers, and γ is the aggregation parameter used to adjust the weights of easy and difficult samples. After calculating the loss, backpropagation is used to update the model parameters to train the model.

[0097] Figure 5 It is a curve showing the change in accuracy on the training and validation sets with the number of iterations. Figure 5 As can be seen, the accuracy of the training set continues to rise with the increase in the number of iterations, but the accuracy of the validation set has stabilized, indicating that convergence has been achieved. The experimental results of the simulation dataset are shown in Table 3.

[0098] Table 3. Experimental Results of Simulation Dataset

[0099] Training accuracy Test accuracy 99.4％ 99.1％

[0100] The confusion matrix of the simulation test set is shown below. Figure 6 The overall accuracy is high, with less confusion between different modulation types. The accuracy versus signal-to-noise ratio curve is shown in the figure. Figure 7 As can be seen, the accuracy increases with the increase in signal-to-noise ratio (SNR), achieving approximately 92% accuracy at an SNR of 3dB, indicating that it adapts well to low SNR conditions. The model can use data of arbitrary length as input; by changing the length of the input signal, the accuracy curve as a function of the input signal length can be obtained. Figure 8 As the length of the input signal data increases, the accuracy initially increases rapidly, then grows more slowly. When the input data length is greater than 1024, the accuracy is greater than 92%, and the model can adapt well to short burst signals.

[0101] Furthermore, to better illustrate the advantages of the present invention, actual collected data was used for verification. The parameters of the actual collected dataset are shown in Table 4.

[0102] Table 4 Parameter Table of Actual Data Set

[0103]

[0104]

[0105] The parameter settings for the actual data and the simulation data differ to verify the model's generalization performance. The main differences are as follows: 1) The sampling rate and symbol rate differ to verify that the model can adapt to signals with different symbol rates in real-world conditions; 2) The SPS differs: the simulation data uses integers between 5 and 15 for the SPS, while the actual data uses 7.5 and 12.5, values ​​outside the simulation data's sampling values; 3) The modulation parameters for each modulation scheme (including FM, AM, ASK, and FSK) are also selected from values ​​outside the simulation data during data acquisition, demonstrating the generalization ability of the modulation parameters; 4) The signal-to-noise ratio (SNR) of the measured data is related to the real channel; 5) The frequency offset is also related to the real channel and is randomly adjusted during data acquisition.

[0106] The model trained using simulation data was tested on real-world data with an input length of 4096. The results are shown in Table 5.

[0107] Table 5. Test Results of Actual Data Collection

[0108] Accuracy of actual collected data test 97.5％

[0109] See confusion matrix Figure 9 Compared to the simulation data, the results showed a decline, but the accuracy rate was still high, proving that the method of this implementation method has good adaptability.

[0110] It is easy to see that the intelligent recognition network model of the present invention can realize the recognition of many modulation types, without limiting the length of the signal, solving the problems of low accuracy and poor noise resistance of traditional modulation recognition methods, while reducing the workload and uncertainty of manually designing feature parameters.

[0111] A second embodiment of the present invention relates to a modulation recognition device based on a multimodal residual attention network, comprising:

[0112] The acquisition module is used to acquire the modulation signal data to be identified;

[0113] The processing module is used to perform energy normalization and amplitude and phase data calculation on the modulation signal data to be identified, so as to obtain the amplitude data and phase data of the modulation signal to be identified.

[0114] An identification module is used to input the energy-normalized modulation signal data to be identified, the amplitude data, and the phase data into an intelligent identification network model to obtain the modulation mode of the modulation signal data to be identified; wherein, the intelligent identification network model includes:

[0115] The multimodal feature fusion module is used to extract features from the energy-normalized modulation signal data to be identified, the amplitude data, and the phase data, and to fuse the extracted features together by splicing to obtain fused features;

[0116] The residual network module based on ECA is used to extract deep features from the fused features to obtain multi-level features;

[0117] The feature pyramid network module is used to perform global average pooling on each layer of the multi-level features, and then concatenate the processed results along the channel dimension to obtain multi-level fused features.

[0118] The classification module is used to classify the modulation mode of the modulation signal data to be identified based on the multi-level fusion features.

[0119] The multimodal feature fusion module includes:

[0120] The first feature extraction unit, composed of two-dimensional convolutional blocks, is used to extract features from the modulated signal data to be identified after energy normalization, and obtain the first feature.

[0121] The second feature extraction unit, composed of one-dimensional convolutional blocks, is used to extract features from the amplitude data to obtain the second feature.

[0122] The third feature extraction unit, composed of one-dimensional convolutional blocks, is used to extract features from the phase data to obtain the third feature.

[0123] The splicing unit is used to splice the first feature, the second feature, and the third feature to obtain the fused feature.

[0124] The ECA-based residual network module includes multiple stacked ECA residual blocks. Each ECA residual block extracts features through three one-dimensional convolutional blocks, then calculates the average value of each feature channel through global pooling, performs convolution processing on the feature channels, and calculates attention weights using the Sigmoid function to achieve weighted features.

[0125] The classification module includes:

[0126] A fully connected layer is used to map the multi-level fused features to the dimension of the number of categories;

[0127] The Softmax activation layer is used to convert the output of the fully connected layer into a probability distribution through the Softmax activation function.

[0128] The intelligent recognition network model is trained in the following way:

[0129] Collect modulated signal data with modulation type labels;

[0130] Energy normalization and amplitude-phase data calculation are performed on all modulated signal data to obtain the amplitude and phase data of each modulated signal data.

[0131] Training and test datasets were constructed based on energy-normalized modulated signal data, amplitude data, and phase data.

[0132] The intelligent recognition network model is trained using the training dataset. During training, a fixed length of data is randomly truncated from each data sample in the training dataset. During testing, data samples from the test dataset of arbitrary length are used for testing.

[0133] The loss function of the intelligent recognition network model is: FL(p k )=-α k (1-p k ) γ log(p k ), where FL(p k α is the loss function of the intelligent recognition network model. k p represents the category weight coefficient. k γ represents the probability of the correct category predicted by the intelligent recognition network model, and γ is the aggregation parameter.

[0134] The third embodiment of the present invention relates to an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the modulation recognition method based on a multimodal residual attention network of the first embodiment.

[0135] The fourth embodiment of the present invention relates to a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the modulation recognition method based on a multimodal residual attention network of the first embodiment.

[0136] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0137] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0138] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction methods implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0139] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0140] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A modulation recognition method based on a multimodal residual attention network, characterized in that, Includes the following steps: Acquire the modulation signal data to be identified; The modulation signal data to be identified is processed by energy normalization and amplitude and phase data calculation to obtain the amplitude data and phase data of the modulation signal to be identified. The energy-normalized modulation signal data to be identified, the amplitude data, and the phase data are input into the intelligent identification network model to obtain the modulation mode of the modulation signal data to be identified; wherein, the intelligent identification network model includes: The multimodal feature fusion module is used to extract features from the energy-normalized modulation signal data to be identified, the amplitude data, and the phase data, and to fuse the extracted features together by splicing to obtain fused features; The ECA-based residual network module is used to extract deep features from the fused features to obtain multi-level features. The ECA-based residual network module includes multiple stacked ECA residual blocks. The ECA residual blocks extract features through three one-dimensional convolutional blocks, then calculate the average value of each feature channel through global pooling, perform convolution processing on the feature channels, and use the Sigmoid function to calculate the attention weights to achieve feature weighting. The feature pyramid network module is used to perform global average pooling on each layer of the multi-level features, and then concatenate the processed results along the channel dimension to obtain multi-level fused features. The classification module is used to classify the modulation mode of the modulation signal data to be identified based on the multi-level fusion features.

2. The modulation recognition method based on a multimodal residual attention network according to claim 1, characterized in that, The multimodal feature fusion module includes: The first feature extraction unit, composed of two-dimensional convolutional blocks, is used to extract features from the modulated signal data to be identified after energy normalization, and obtain the first feature. The second feature extraction unit, composed of one-dimensional convolutional blocks, is used to extract features from the amplitude data to obtain the second feature. The third feature extraction unit, composed of one-dimensional convolutional blocks, is used to extract features from the phase data to obtain the third feature. The splicing unit is used to splice the first feature, the second feature, and the third feature to obtain the fused feature.

3. The modulation recognition method based on a multimodal residual attention network according to claim 1, characterized in that, The classification module includes: A fully connected layer is used to map the multi-level fused features to the dimension of the number of categories; The Softmax activation layer is used to convert the output of the fully connected layer into a probability distribution through the Softmax activation function.

4. The modulation recognition method based on a multimodal residual attention network according to claim 1, characterized in that, The intelligent recognition network model is trained in the following way: Collect modulated signal data with modulation type labels; Energy normalization and amplitude-phase data calculation are performed on all modulated signal data to obtain the amplitude and phase data of each modulated signal data. Training and test datasets were constructed based on energy-normalized modulated signal data, amplitude data, and phase data. The intelligent recognition network model is trained using the training dataset. During training, a fixed length of data is randomly truncated from each data sample in the training dataset. During testing, data samples from the test dataset of arbitrary length are used for testing.

5. The modulation recognition method based on a multimodal residual attention network according to claim 1, characterized in that, The loss function of the intelligent recognition network model is: FL(p k )=-α k (1-p k ) γ log(p k ), where FL(p k α is the loss function of the intelligent recognition network model. k p represents the category weight coefficient. k γ represents the probability of the correct category predicted by the intelligent recognition network model, and γ is the aggregation parameter.

6. A modulation recognition device based on a multimodal residual attention network, characterized in that, include: The acquisition module is used to acquire the modulation signal data to be identified; The processing module is used to perform energy normalization and amplitude and phase data calculation on the modulation signal data to be identified, so as to obtain the amplitude data and phase data of the modulation signal to be identified. An identification module is used to input the energy-normalized modulation signal data to be identified, the amplitude data, and the phase data into an intelligent identification network model to obtain the modulation mode of the modulation signal data to be identified; wherein, the intelligent identification network model includes: The multimodal feature fusion module is used to extract features from the energy-normalized modulation signal data to be identified, the amplitude data, and the phase data, and to fuse the extracted features together by splicing to obtain fused features; The ECA-based residual network module is used to extract deep features from the fused features to obtain multi-level features. The ECA-based residual network module includes multiple stacked ECA residual blocks. The ECA residual blocks extract features through three one-dimensional convolutional blocks, then calculate the average value of each feature channel through global pooling, perform convolution processing on the feature channels, and use the Sigmoid function to calculate the attention weights to achieve feature weighting. The feature pyramid network module is used to perform global average pooling on each layer of the multi-level features, and then concatenate the processed results along the channel dimension to obtain multi-level fused features. The classification module is used to classify the modulation mode of the modulation signal data to be identified based on the multi-level fusion features.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the modulation recognition method based on a multimodal residual attention network as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the modulation recognition method based on a multimodal residual attention network as described in any one of claims 1-5.

Images