Modulation category identification method and related apparatus, electronic device, storage medium

Abstract

The application discloses a modulation category identification method and related device, electronic equipment and storage medium, wherein the modulation category identification method comprises: extracting a first feature map of a to-be-identified signal; wherein the first feature map comprises first sub-feature maps of a plurality of channels, and the first sub-feature maps of the plurality of channels have the same resolution; performing prediction on the first feature map based on a plurality of dimensions to obtain weight parameters of the first feature map in each dimension, and weighting the first feature map based on the weight parameters in each dimension to obtain a weighted feature map in each dimension; wherein the plurality of dimensions comprise at least one of a channel dimension and a spatial dimension; fusing at least the weighted feature map in each dimension to obtain a fused feature map; and classifying based on the fused feature map to obtain a modulation category of the to-be-identified signal. The above scheme can improve the accuracy and robustness of modulation category identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, in particular to a modulation category identification method and related device, electronic equipment and storage medium. BACKGROUND

[0002] Modulation category identification of a communication signal refers to a process of identifying and analyzing an obtained communication signal and obtaining a modulation category. Generally, the modulation category includes, but is not limited to, digital modulation modes such as ASK (Amplitude Shift Keying), PSK (Phase Shift Keying), FSK (Frequency Shift Keying), QAM (Quadrature Amplitude Modulation), GMSK (Gaussian Filtered Minimum Shift Keying), and analog modulation modes such as FM (Frequency Modulation) and AM (Amplitude Modulation).

[0003] Modulation category identification of a communication signal is an indispensable step for demodulation of a communication signal received by a receiver, and has important application value and prospect in many scenarios such as spectrum monitoring and communication countermeasures. It is found through research that existing identification methods are extremely dependent on prior information such as carrier frequency value, code rate and symbol timing, or extremely dependent on artificially designed feature parameters such as amplitude histogram, frequency histogram and differential phase histogram, thereby affecting the accuracy and robustness of modulation category identification, especially in non-ideal environments such as low signal-to-noise ratio, Doppler shift, multiple modulation categories and less prior information, the influence is particularly obvious. Therefore, how to improve the accuracy and robustness of modulation category identification has become a problem to be solved. SUMMARY

[0004] The technical problem solved by the present application is to provide a modulation category identification method and related device, electronic equipment and storage medium, which can improve the accuracy and robustness of modulation category identification.

[0005] To solve the above technical problems, the first aspect of the present application provides a modulation category identification method, comprising: extracting a first feature map of a signal to be identified; wherein the first feature map comprises a plurality of channel first sub-feature maps, and the plurality of channel first sub-feature maps have the same resolution; predicting based on the first feature map in a plurality of dimensions respectively to obtain weight parameters of the first feature map in each dimension, and weighting the first feature map based on the weight parameters of each dimension respectively to obtain a weighted feature map of each dimension; wherein the plurality of dimensions include at least one of a channel dimension and a spatial dimension; fusing based on at least the weighted feature map of each dimension to obtain a fused feature map; and classifying based on the fused feature map to obtain a modulation category of the signal to be identified.

[0006] To solve the above technical problems, the second aspect of the present application provides a modulation category identification device, comprising: a feature extraction module, a weight prediction module, a feature weighting module, a feature fusion module and a modulation classification module, the feature extraction module is used to extract a first feature map of a signal to be identified; wherein the first feature map comprises a plurality of channel first sub-feature maps, and the plurality of channel first sub-feature maps have the same resolution; the weight prediction module is used to predict based on the first feature map in a plurality of dimensions respectively to obtain weight parameters of the first feature map in each dimension, wherein the plurality of dimensions include at least one of a channel dimension and a spatial dimension; the feature weighting module is used to weight the first feature map based on the weight parameters of each dimension respectively to obtain a weighted feature map of each dimension; the feature fusion module is used to fuse based on at least the weighted feature map of each dimension to obtain a fused feature map; and the modulation classification module is used to classify based on the fused feature map to obtain a modulation category of the signal to be identified.

[0007] To solve the above technical problems, the third aspect of the present application provides an electronic device, comprising a memory and a processor coupled to each other, the memory stores program instructions, and the processor is used to execute the program instructions to realize the modulation category identification method of the first aspect.

[0008] To solve the above technical problems, the fourth aspect of the present application provides a computer readable storage medium, which stores program instructions capable of being executed by a processor, and the program instructions are used to realize the modulation category identification method of the first aspect.

[0009] The above scheme extracts a first feature map of the signal to be identified, which includes first sub-feature maps of several channels with the same resolution. Based on this, predictions are made in several dimensions based on the first feature map to obtain weighting parameters of the first feature map in each dimension. The first feature map is then weighted based on the weighting parameters of each dimension to obtain weighted feature maps of each dimension. The several dimensions include at least one of channel dimension and spatial dimension. Based on this, the weighted feature maps of at least each dimension are fused to obtain a fused feature map. The fused feature map is then used for classification to obtain the modulation category of the signal to be identified. On the one hand, since the modulation category identification process does not rely on prior information and manually designed feature parameters, it can reduce the impact of environmental factors such as low signal-to-noise ratio, Doppler frequency shift, multiple modulation categories, and limited prior information on modulation identification. On the other hand, by combining weighted feature maps of several dimensions such as channel dimension and spatial dimension for modulation classification, the feature information related to the modulation category in the signal to be identified can be extracted as fully as possible, which helps to adapt to as many modulation categories as possible. Therefore, it can improve the accuracy and robustness of modulation category recognition. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an embodiment of the modulation category identification method of this application;

[0011] Figure 2 This is a schematic diagram of the framework of an embodiment of the debugging category recognition model;

[0012] Figure 3 This is a schematic diagram of the framework of an embodiment of a weight prediction network;

[0013] Figure 4 This is a schematic diagram illustrating the effect of the modulation category identification method of this application under different signal-to-noise ratios;

[0014] Figure 5 This is a schematic diagram of the framework of an embodiment of the modulation category identification device of this application;

[0015] Figure 6 This is a schematic diagram of the framework of an embodiment of the electronic device of this application;

[0016] Figure 7 This is a schematic diagram of the frame of an embodiment of the receiver of this application;

[0017] Figure 8 This is a schematic diagram of the framework of an embodiment of the communication device of this application;

[0018] Figure 9 This is a schematic diagram of a framework of an embodiment of the computer-readable storage medium of this application. Detailed Implementation

[0019] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0020] In the following description, specific details such as particular system architectures, interfaces, and technologies are presented for illustrative purposes rather than for limiting purposes, in order to provide a thorough understanding of this application.

[0021] In this paper, the terms "system" and "network" are often used interchangeably. The term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, "many" in this paper means two or more.

[0022] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the modulation category identification method of this application. Specifically, it may include the following steps:

[0023] Step S11: Extract the first feature map of the signal to be identified.

[0024] In this embodiment of the disclosure, the first feature map includes a first sub-feature map with several channels, and the first sub-feature maps of the several channels have the same resolution. For ease of description, the number of channels of the first feature map can be denoted as C, and the resolution of the first sub-feature map of each channel can be denoted as H*W. Then the first feature map can be represented as U∈R H×W×C .

[0025] In one implementation scenario, the signal to be identified is a communication signal intercepted by the receiver. Specifically, it can include several sampling points in the time domain, such as intermediate frequency (IF) sampling points. Furthermore, each sampling point can include sampling time and signal amplitude. After connecting several sampling points in chronological order, the signal to be identified can also be represented on a two-dimensional coordinate system as a curve with time on the horizontal axis and amplitude on the vertical axis. The above description only represents a few possible ways of representing the signal to be identified in practical applications and is not limited here.

[0026] In one implementation scenario, to improve the efficiency of modulation recognition, a modulation recognition model can be pre-trained. This model can include a feature extraction network, which may include, but is not limited to, convolutional neural networks. Based on this, features can be extracted from the signal to be recognized using the feature extraction network to obtain a first feature map. Specifically, to improve the accuracy of the modulation recognition model, several sample signals can be pre-collected, and these signals can be labeled with their modulation categories. Based on this, the feature extraction network can be used to extract features from the sample signals to obtain sample feature maps. Then, prediction is performed based on these sample feature maps to obtain the predicted modulation category of the sample signal. Thus, the network parameters of the modulation recognition model can be adjusted based on the difference between the sample modulation category and the predicted modulation category. This allows the feature extraction network to extract the feature information related to the modulation category in the signal as accurately and fully as possible during training, thus improving the recognition performance of the modulation recognition model. It should be noted that the specific measurement methods for the above differences can be found in loss functions such as cross-entropy, and the specific adjustment process of the above parameters can be found in optimization methods such as gradient descent, which will not be elaborated here.

[0027] In one implementation scenario, unlike the aforementioned extraction methods, the first feature map can also be obtained by fusing the feature maps extracted at multiple scales after multi-scale feature extraction of the signal to be identified. Specifically, feature extraction can be performed on the signal to be identified at several scales to obtain second feature maps at several scales, and each scale's second feature map is extracted by a convolutional kernel of the corresponding scale. Based on this, the second feature maps at several scales can be fused to obtain the first feature map of the signal to be identified. This method, by fusing the second feature maps extracted at several scales to obtain the first feature map, can fully explore deep feature representations and help improve the accuracy of the first feature map.

[0028] In a specific implementation scenario, the "several dimensions" can include, but are not limited to, one, two, three, or even four dimensions, etc., and are not limited here. For example, to maximize the accuracy of the first feature map, the number of categories in the "several dimensions" can be increased; conversely, to minimize the computational load caused by multi-dimensional feature extraction, the number of categories in the "several dimensions" can be decreased. In particular, to achieve a balance between accuracy and computational load, the number of categories in the "several dimensions" can be set to a moderate level. For example, the "several dimensions" can include two dimensions, and the specific setting method is not limited here.

[0029] In a specific implementation scenario, each scale can correspond to a number of convolutional kernels (e.g., 1 kernel, 2 kernels, 3 kernels, 4 kernels, etc.). Kernel widths are the same but lengths differ for the same scale, while kernel widths differ for different scales. Taking two types of dimensions as an example, for the first dimension, it can correspond to convolutional kernels with the same width but different lengths, such as 1*1, 4*1, and 8*1. For the second dimension, it can correspond to convolutional kernels with the same width but different lengths, such as 1*2, 2*2, and 8*2. The convolutional kernels for the first and second dimensions have different widths. It should be noted that the above-mentioned configuration of convolutional kernels for several dimensions is merely one possible configuration in practical applications and does not limit the specific configuration of the convolutional kernels for each dimension.

[0030] In a specific implementation scenario, as mentioned earlier, each scale corresponds to several convolutional kernels. Kernel widths are the same but lengths differ for the same scale, while kernel widths differ for different scales. Based on this, for each scale, features can be extracted from the signal to be recognized using the corresponding convolutional kernels, resulting in several third sub-feature maps for that scale. These third sub-feature maps are then concatenated to obtain a second feature map for that scale. Furthermore, the second feature maps for several scales can be adjusted to the same resolution, and the adjusted second feature maps are then concatenated to obtain the first feature map. Taking two dimensions as an example, for the first dimension, feature extraction can be performed using convolutional kernels of the same width but different lengths (e.g., 1x1, 4x1, 8x1), resulting in three third sub-feature maps for that dimension. These third sub-feature maps are then concatenated to obtain the second feature map for that dimension. Similarly, for the second dimension, feature extraction can be performed using convolutional kernels of the same width but different lengths (e.g., 1x2, 2x2, 8x2), resulting in three third sub-feature maps for that dimension. These third sub-feature maps are then concatenated to obtain the second feature map for that dimension. Based on this, the second feature maps for both dimensions can be adjusted to the same resolution, and then concatenated to obtain the first feature map U∈R. H×W×CIn the above method, each scale corresponds to several convolutional kernels. Kernels at the same scale have the same width but different lengths, while kernels at different scales have different widths. For each scale, features are extracted from the signal to be recognized based on the convolutional kernels corresponding to that scale, resulting in several third sub-feature maps. These third sub-feature maps are then concatenated to obtain a second feature map for that scale. Furthermore, the second feature maps at several scales are adjusted to the same resolution, and then concatenated to obtain a first feature map. Therefore, within the same dimension, concatenating the third sub-feature maps extracted by convolutional kernels with the same width but different lengths allows for the fusion of feature information extracted by convolutional kernels with different receptive fields within the same dimension. This helps to further extract deeper feature representations of the signal to be recognized, thereby improving the accuracy of the first feature map.

[0031] In a specific implementation scenario, to improve the efficiency of modulation recognition, a modulation recognition model can be pre-trained, and this model can include a feature extraction network. Unlike the aforementioned methods, in this embodiment, the feature extraction network can include several extraction sub-networks corresponding to different scales, and a fusion sub-network connected to each extraction sub-network. As mentioned earlier, each scale-corresponding extraction sub-network includes several convolutional kernels; convolutional kernels at the same scale have the same width but different lengths, while convolutional kernels at different scales have different widths. Please refer to [reference needed]. Figure 2 , Figure 2 This is a schematic diagram of the framework of an embodiment of the debugging category recognition model. For example... Figure 2 As shown, for the signal to be identified, before performing feature extraction at several scales, it can first be passed through a two-dimensional convolutional layer (such as one that can contain...). Figure 2 The output feature map F∈R is obtained by using a 3*1 convolution kernel. H0×W0×C0 Where H0 and W0 represent the height and width of the output feature map, and C0 represents the number of channels in the output feature map. The output feature map can then be fed into the extraction sub-networks corresponding to each scale, such as... Figure 2As shown, the output feature maps can be fed into the left and right extraction sub-networks respectively. For the extraction sub-networks corresponding to each scale, the feature F can be processed first using a 1*1 convolutional kernel. Then, for the left extraction sub-network, several third sub-feature maps corresponding to that dimension can be extracted using convolutional kernels of different lengths but the same width, such as 4*1 and 8*1 convolutional kernels. These are then concatenated at a fine-grained level to obtain the second feature map corresponding to that dimension. For the right extraction sub-network, several third sub-feature maps corresponding to that dimension can be obtained using convolutional kernels of the same width but different lengths, such as 1*2, 2*2, and 8*2. These are then concatenated at a fine-grained level to obtain the second feature map corresponding to that dimension. In addition, the fusion sub-network can include convolutional kernels (such as...) used to adjust the resolution of the second feature maps corresponding to each dimension. Figure 2 By fusing 1x2 convolutional kernels in the fusion subnetwork, and then processing with 1x1 convolutional kernels at a coarse-grained level, a multi-scale deep representation of the signal to be identified across different receptive fields can be obtained, i.e., the first feature map U∈R. H×W×C Furthermore, as can be seen from this process, the feature extraction method in this embodiment does not require any prior information, nor does it require manual feature analysis of the signal. Instead, the deep signal representation is obtained automatically through data mining by the neural network. It should be noted that... Figure 2 The feature extraction network shown is merely one possible network structure in practical applications and does not limit the network structure of the feature extraction network. In the above method, the first feature map is extracted by the feature extraction network, which includes several extraction sub-networks corresponding to different scales, and fusion sub-networks connected to each extraction sub-network. Each extraction sub-network corresponding to each scale includes several convolutional kernels. The convolutional kernels corresponding to the same scale have the same width but different lengths, while the convolutional kernels corresponding to different scales have different widths. Therefore, while simplifying the network structure of the feature extraction network as much as possible, it can fully extract the deep feature representation of the signal to be identified, thereby improving the accuracy of the first feature map.

[0032] Step S12: Based on the first feature map, make predictions in several dimensions to obtain the weight parameters of the first feature map in each dimension, and then weight the first feature map based on the weight parameters of each dimension to obtain the weighted feature map of each dimension.

[0033] In this embodiment of the disclosure, the plurality of dimensions includes at least one of a channel dimension and a spatial dimension. For example, the plurality of dimensions may include a channel dimension, or it may include a spatial dimension, or it may include both a channel dimension and a spatial dimension; this is not limited here. It should be noted that, in order to maximize the accuracy of modulation category recognition, the plurality of dimensions can be as numerous as possible; for example, the plurality of dimensions may be set to include both a channel dimension and a spatial dimension; this is not limited here. Furthermore, the channel dimension refers to the image channel dimension of the first feature map, that is, the image channel is used as the basic unit of weighting, while the spatial dimension refers to the pixel position dimension of the first feature map, that is, the pixel position is used as the basic unit of weighting.

[0034] In one implementation scenario, when there are several dimensions, including the channel dimension, prediction can be made based on the first feature map in the channel dimension to obtain the first weights of several channels. The weight parameters of the first feature map in the channel dimension include the first weights of several channels. Based on this, the first sub-feature maps of the same channel can be weighted according to the first weights of each channel to obtain a weighted feature map in the channel dimension. This method, when there are several dimensions, including the channel dimension, predicts the first weights of each channel using the first feature map, and then weights the first sub-feature maps accordingly to obtain the weighted feature map in the channel dimension. Therefore, it can enhance features important for modulation category identification and suppress features irrelevant to modulation category identification in the channel dimension, which helps improve the accuracy of subsequent modulation category identification.

[0035] In a specific implementation scenario, during the prediction of the first weight, the first sub-feature maps of several channels can be globally pooled (e.g., global max pooling, global average pooling, etc.) to obtain a global feature map. This global feature map includes the first feature value of each channel. Prediction is then performed based on the global feature map to obtain the first weight of each channel. For example, for the first feature map U∈R... H×W×C In this regard, the first sub-feature map of each channel can be represented as:

[0036] U = [u1, u2, ... u] i ...u c ]u i ∈R H×W ……(1)

[0037] In the above formula (1), u i This represents the i-th first sub-feature map. Based on this, global average pooling can be performed on the first sub-feature maps of each channel to obtain the first eigenvalue of the corresponding channel. Combining the first eigenvalues ​​of each channel yields the global feature map Z∈R. 1×1×C For example, the k-th first feature value z in the global feature map Z. kIt can be represented as:

[0038]

[0039] In a specific implementation scenario, as mentioned earlier, to improve the efficiency of modulation category recognition, a modulation recognition model can be pre-trained, and this model may include a first prediction network. See [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram of the framework of an embodiment of a weight prediction network. Figure 3 As shown, the weight prediction network may specifically include a first prediction network, which may include a global pooling layer (e.g., a global average pooling layer, a global max pooling layer, etc.) and a fully connected layer. The global pooling layer is used to perform global pooling on the first sub-feature maps of several channels respectively to obtain a global feature map, and the global feature map includes the first feature value of each channel. The fully connected layer is used to make predictions based on the global feature map to obtain the first weight of each channel. For example, as shown... Figure 3 As shown, the first prediction network can include two fully connected layers, and each of the two fully connected layers can be followed by an activation layer. For example, the activation layer following the first fully connected layer can be a ReLU (Rectified Linear Unit), and the activation layer following the second fully connected layer can be a sigmoid layer. In this case, the channel-dimensional weight parameter M can be expressed as:

[0040]

[0041] In the above formula (3), Z represents the global feature map, W1 and W2 represent the network weights of the first fully connected layer and the second fully connected layer, respectively, and δ1 and δ2 represent the activation functions after the first and second fully connected layers, respectively. It should be noted that the first weight of each channel contained in the channel dimension weight parameter M represents the importance of each channel. For example, the larger the first weight of a channel, the higher the importance of that channel; conversely, the smaller the first weight of a channel, the lower the importance of that channel. In the above method, the channel dimension weight parameter is predicted by the first prediction network. The first prediction network includes a global pooling layer and a fully connected layer. The global pooling layer is used to perform global pooling on the first sub-feature maps of several channels to obtain a global feature map, and the global feature map includes the first feature value of each channel. The fully connected layer is used to make predictions based on the global feature map to obtain the first weight of each channel. Therefore, the accuracy of weight prediction can be improved while simplifying the first prediction network as much as possible.

[0042] In a specific implementation scenario, after obtaining the channel-dimensional weight parameter M, the channel-dimensional weight parameter M can be weighted correspondingly with the first feature map U along the channel dimension to obtain the channel-dimensional weighted feature map. It should be noted that weighting along the channel dimension does not change the number of channels or the feature map resolution of the first feature map.

[0043] In one implementation scenario, when there are several dimensions, including a spatial dimension, prediction can be made based on the first feature map in the spatial dimension to obtain the second weights for each pixel position. The weight parameters obtained from the first feature map in the spatial dimension include the second weights for each pixel position. Based on this, the second sub-feature maps for the same pixel position can be weighted according to the second weights for each pixel position to obtain a weighted feature map in the spatial dimension. In this method, when there are several dimensions, including a spatial dimension, the second weights for each pixel position are predicted using the first feature map, and the second sub-feature maps for each pixel position are weighted accordingly to obtain the weighted feature map in the spatial dimension. Therefore, this method can enhance features important for modulation category identification and suppress features irrelevant to modulation category identification in the spatial dimension, thus helping to improve the accuracy of subsequent modulation category identification.

[0044] In a specific implementation scenario, the pixel values ​​at the same pixel position in the first sub-feature maps of several channels can form a second sub-feature map corresponding to that pixel position. That is, for the first feature map U∈R... H×W×C In this case, H×W second sub-feature maps of size 1×C can be formed. For ease of description, the second sub-feature map at each pixel position can be represented as:

[0045] U = [u 1,1 ,u 1,2 ...u i,j ...u H,W ]u i ∈R 1×C ……(4)

[0046] In the above formula (4), u i,jThis represents the second sub-feature map at pixel position (i, j). Based on this, the second sub-feature map at each pixel position can be dimensionality-reduced to obtain the second feature value for each pixel position. In other words, the second sub-feature map of size 1×C at each pixel position can be reduced to a one-dimensional second feature value. Based on this, normalization can be performed on the second feature value at each pixel position to obtain the second weight for each pixel position. This method, which performs dimensionality reduction on the second sub-feature map at each pixel position to obtain the second feature value for each pixel position, and then normalizes based on the second feature value to obtain the second weight for each pixel position, simplifies the prediction of the second weight while maximizing the accuracy of the second weight.

[0047] In a specific implementation scenario, as mentioned earlier, to improve the efficiency of modulation category recognition, a modulation recognition model can be pre-trained, and this model may include a second prediction network. See [link to relevant documentation]. Figure 3 As shown in the weight prediction network, the weight prediction network may include a second prediction network, and the second prediction network may include a one-dimensional convolutional layer (e.g., a 1*1 convolutional kernel) and a normalization layer (e.g., a sigmoid). The one-dimensional convolutional layer is used to reduce the dimensionality of the second sub-feature map of each pixel position to obtain the second feature value of each pixel position. The normalization layer is used to normalize based on the second feature value of each pixel position to obtain the second weight of each pixel position. For example, the second sub-feature map of each pixel position shown in the above formula (4) can be input into a one-dimensional convolutional layer such as a 1*1 to obtain the second feature values ​​of H*W pixel positions. Then, through a normalization layer such as a sigmoid, the weight parameter Q∈R of the spatial dimension can be obtained. H×W By weighting the first feature map U along its corresponding dimensions in space, a weighted feature map along the spatial dimensions can be obtained. It should be noted that weighting along the spatial dimensions does not change the number of channels or the resolution of the first feature map.

[0048] Step S13: At least based on the weighted feature maps of each dimension, perform fusion to obtain a fused feature map.

[0049] In one implementation scenario, the weighted feature maps of each dimension can be fused by adding corresponding points (i.e., pixel positions) to obtain a fused feature map. In other words, for the weighted feature maps of each dimension, the pixel values ​​at the same pixel position in these weighted feature maps can be added together to obtain the fused feature map.

[0050] In one implementation scenario, please refer to the following: Figure 2 ,like Figure 2As shown, to reduce information loss, a fused feature map can be obtained by fusing the first feature map and the weighted feature maps of each dimension. Specifically, the corresponding points (i.e., pixel positions) of the first feature map and the weighted feature maps of each dimension can be added together to obtain the fused feature map. This method, which fuses the first feature map and the weighted feature maps of each dimension to obtain the fused feature map, further references the first feature map during the feature map fusion process, thus minimizing information loss and improving the accuracy of subsequent modulation category recognition.

[0051] Step S14: Classify the signal based on the fused feature map to obtain the modulation category of the signal to be identified.

[0052] In one implementation scenario, to improve classification accuracy, sample signals of different modulation categories can be collected in advance, and signal feature maps of each sample signal can be extracted. For example, feature extraction and weighting can be performed using the feature extraction network and weight prediction network in the aforementioned modulation recognition model to obtain sample feature maps of the sample signals. For details, please refer to the aforementioned extraction process of the fused feature map, which will not be repeated here. Based on this, for sample signals belonging to the same modulation category, the sample feature maps of these sample signals can be averaged to obtain the signal feature map of that modulation category. Thus, during the classification process, the feature distance between the signal feature map of each modulation category and the fused feature map of the signal to be identified can be obtained separately, and the modulation category corresponding to the smallest feature distance can be taken as the modulation category of the signal to be identified.

[0053] In one implementation scenario, unlike the aforementioned classification methods, to further adapt to as many modulation categories as possible, the modulation recognition model can further include a category classification network. After obtaining the fused feature map, the category classification network can be used to classify and predict the fused feature map, obtaining the probability values ​​of the signal to be identified belonging to each modulation category. The modulation category corresponding to the highest probability value can then be used as the modulation category of the signal to be identified. Please refer to further details. Figure 2 The category classification network can include a global pooling layer, a fully connected layer, and a normalization layer. The global pooling layer performs global pooling on the fused feature map to obtain a pooled feature map. The fully connected layer processes the pooled feature map to obtain values ​​representing the probability of various modulation categories. The normalization layer normalizes these values ​​to obtain the probability values ​​of the signal to be identified belonging to each modulation category. For example, for a fused feature map of size H*W*C, it can be compressed into a pooled feature map of size 1*(H*W*C) through global pooling, then processed by several (e.g., three) fully connected layers, and finally normalized by a function such as softmax to obtain the probability values ​​of the signal to be identified belonging to each modulation category. Thus, the modulation recognition model can be used to complete the end-to-end identification from the signal sampling point to the modulation category.

[0054] In one implementation scenario, an experimental platform was built for the embodiments of this disclosure through actual testing, as shown in Table 1. Table 1 is a configuration table of an embodiment of the experimental platform. Based on this, test data was collected from communication signals of 24 different modulation categories, specifically as follows: OOK (On-Off Keying), ASK4, ASK8, BPSK (Binary Phase Shift Keying), QPSK (Quadrature Phase Shift Keying), PSK8, PSK16, PSK32, APSK16 (Amplitude Phase Shift Keying), APSK32, APSK64, APSK128, QAM16, QAM32, QAM64, QAM128, QAM256, AMSSBWC, AMSSBSC, AMDSBWC, AMDSBSC, FM, GMSK, and OQPS (offset-QPSK). The specific meanings of the above modulation categories can be found in the relevant technical details of communication modulation, which will not be elaborated here. Furthermore, the test data is stored alternately in I / Q sampling and 32-bit floating-point format. To verify the recognition performance under different signal-to-noise ratios (SNRs), data at 26 different SNRs are prepared for each modulation category (data in batches of 2dB from -20dB to 30dB). Each modulation category contains 4096 data points at each SNR, with each data point containing 2048 sampling points, for a total of 2,555,904 data points. During the experiment, 85% of the data can be randomly selected as the training set, and the remaining 15% as the test set.

[0055] Table 1 Configuration table of an embodiment of the test platform

[0056]

[0057] Furthermore, to verify the recognition effect of the embodiments of this disclosure, decision trees, and SVM (Support Vector Machine) and ResNet-based deep residual network schemes, which are currently leading in the field of communication modulation type recognition, were constructed as baseline systems for comparative analysis. The latter two schemes extract time-frequency distribution and higher-order cumulative quantity features, respectively. The experimental results of the embodiments of this disclosure and the above three schemes are shown in Table 2, which is a comparison table of the embodiments of this disclosure with other schemes.

[0058] Table 2 Comparison of the embodiments of this disclosure with other embodiments.

[0059]

[0060] As shown in Table 2, at a signal-to-noise ratio of 10dB, compared to feature extraction-based pattern recognition schemes, regardless of whether the backend classifier uses traditional machine learning classifiers such as SVM or deep neural network-based classifiers, the recognition accuracy of the scheme in this disclosure is significantly improved, with a relative improvement of over 22%. Please refer to... Figure 4 , Figure 4 This is a schematic diagram illustrating the effect of the modulation category identification method of this application under different signal-to-noise ratios. For example... Figure 4 As shown, the solution of this disclosure can adapt to different signal-to-noise ratios, and its recognition rate is greater than 93.6% when the signal-to-noise ratio is greater than 10dB. Based on the above test results, it can be found that the solution of this disclosure can effectively characterize the deep discriminative features of communication signals under different modulation categories, and can adapt to different transmission environments and support the recognition of multiple modulation categories, showing a significant improvement in recognition performance compared with existing technical solutions.

[0061] The above scheme extracts a first feature map of the signal to be identified, which includes first sub-feature maps of several channels with the same resolution. Based on this, predictions are made in several dimensions based on the first feature map to obtain weighting parameters of the first feature map in each dimension. The first feature map is then weighted based on the weighting parameters of each dimension to obtain weighted feature maps of each dimension. The several dimensions include at least one of channel dimension and spatial dimension. Based on this, the weighted feature maps of at least each dimension are fused to obtain a fused feature map. The fused feature map is then used for classification to obtain the modulation category of the signal to be identified. On the one hand, since the modulation category identification process does not rely on prior information and manually designed feature parameters, it can reduce the impact of environmental factors such as low signal-to-noise ratio, Doppler frequency shift, multiple modulation categories, and limited prior information on modulation identification. On the other hand, by combining weighted feature maps of several dimensions such as channel dimension and spatial dimension for modulation classification, the feature information related to the modulation category in the signal to be identified can be extracted as fully as possible, which helps to adapt to as many modulation categories as possible. Therefore, it can improve the accuracy and robustness of modulation category recognition.

[0062] Please see Figure 5 , Figure 5This is a schematic diagram of a framework of an embodiment of the modulation category identification device 50 of this application. The modulation category identification device 50 includes: a feature extraction module 51, a weight prediction module 52, a feature weighting module 53, a feature fusion module 54, and a modulation classification module 55. The feature extraction module 51 is used to extract a first feature map of the signal to be identified; wherein, the first feature map includes first sub-feature maps of several channels, and the first sub-feature maps of several channels have the same resolution; the weight prediction module 52 is used to predict based on the first feature map in several dimensions to obtain the weight parameters of the first feature map in each dimension, wherein, the several dimensions include at least one of channel dimension and spatial dimension; the feature weighting module 53 is used to weight the first feature map based on the weight parameters of each dimension to obtain a weighted feature map of each dimension; the feature fusion module 54 is used to fuse based on at least the weighted feature maps of each dimension to obtain a fused feature map; the modulation classification module 55 is used to classify based on the fused feature map to obtain the modulation category of the signal to be identified.

[0063] The above-described scheme, on the one hand, reduces the impact of environmental factors such as low signal-to-noise ratio, Doppler shift, multiple modulation categories, and limited prior information on modulation category identification because it does not rely on prior information or manually designed feature parameters during the modulation category identification process. On the other hand, by combining weighted feature maps of several dimensions such as channel dimension and spatial dimension for modulation classification, it can extract as much feature information related to the modulation category as possible from the signal to be identified, which helps to adapt to as many modulation categories as possible. Therefore, it can improve the accuracy and robustness of modulation category identification.

[0064] In some disclosed embodiments, when the multiple dimensions include the channel dimension, the weight prediction module 52 includes a first prediction submodule, which is used to predict based on the first feature map in the channel dimension to obtain the first weights of the multiple channels; wherein, the weight parameters of the first feature map in the channel dimension include the first weights of the multiple channels; the weight prediction module 52 includes a first weighting submodule, which is used to weight the first sub-feature map of the same channel based on the first weight of each channel to obtain a weighted feature map in the channel dimension.

[0065] Therefore, in the case of several dimensions including the channel dimension, the first weight of each channel is predicted by the first feature map, and the first sub-feature map is weighted accordingly to obtain the weighted feature map of the channel dimension. Thus, it is possible to enhance the features that are important for modulation category identification and suppress the features that are irrelevant to modulation category identification in the channel dimension, which helps to improve the accuracy of subsequent modulation category identification.

[0066] In some disclosed embodiments, the first prediction submodule includes a global pooling unit, which is used to perform global pooling on the first sub-feature maps of several channels respectively to obtain a global feature map; wherein, the global feature map includes the first feature value of each channel; the first prediction submodule includes a weight prediction unit, which is used to make predictions based on the global feature map to obtain the first weight of each channel.

[0067] Therefore, by performing global pooling on the first sub-feature maps of several channels to obtain a global feature map, and the global feature map includes the first feature value of each channel, and then making predictions based on the global feature map to obtain the first weight of each channel, the accuracy of the first weight can be improved while simplifying the prediction of the first weight.

[0068] In some disclosed embodiments, the channel dimension weight parameters are predicted by a first prediction network, which includes a global pooling layer and a fully connected layer. The global pooling layer is used to perform global pooling on the first sub-feature maps of several channels to obtain a global feature map, and the global feature map includes the first feature value of each channel. The fully connected layer is used to make predictions based on the global feature map to obtain the first weight of each channel.

[0069] Therefore, the channel dimension weight parameters are predicted by the first prediction network, which includes a global pooling layer and a fully connected layer. The global pooling layer is used to perform global pooling on the first sub-feature maps of several channels to obtain a global feature map, and the global feature map includes the first feature value of each channel. The fully connected layer is used to make predictions based on the global feature map to obtain the first weight of each channel. Thus, the accuracy of weight prediction can be improved while simplifying the first prediction network as much as possible.

[0070] In some disclosed embodiments, the pixel values ​​at the same pixel position of the first sub-feature map of several channels constitute the second sub-feature map. When the several dimensions include the spatial dimension, the weight prediction module 52 includes a second prediction sub-module, which is used to predict based on the first feature map in the spatial dimension to obtain the second weight of each pixel position; wherein, the weight parameters obtained by the first feature map in the spatial dimension include the second weight of each pixel position; the weight prediction module 52 includes a second weighting sub-module, which is used to weight the second sub-feature map at the same pixel position based on the second weight of each pixel position to obtain a weighted feature map in the spatial dimension.

[0071] Therefore, in the case of several dimensions including the spatial dimension, the second weight of each pixel position is predicted by the first feature map, and the second sub-feature map of each pixel position is weighted accordingly to obtain the weighted feature map of the spatial dimension. Thus, it is possible to enhance the features that are important for modulation category identification and suppress the features that are irrelevant to modulation category identification in the spatial dimension, which helps to improve the accuracy of subsequent modulation category identification.

[0072] In some disclosed embodiments, the second prediction submodule includes a feature dimensionality reduction unit, which is used to perform feature dimensionality reduction on the second sub-feature map of each pixel position to obtain the second feature value of each pixel position; the second prediction submodule includes a normalization unit, which is used to normalize based on the second feature value of each pixel position to obtain the second weight of each pixel position.

[0073] Therefore, by performing feature dimensionality reduction on the second sub-feature map of each pixel location to obtain the second feature value of each pixel location, and then normalizing based on the second feature value of each pixel location to obtain the second weight of each pixel location, the accuracy of the second weight can be improved as much as possible while simplifying the prediction of the second weight.

[0074] In some disclosed embodiments, the weight parameters of the spatial dimension are predicted by a second prediction network, which includes a one-dimensional convolutional layer and a normalization layer. The one-dimensional convolutional layer is used to reduce the spatial dimension of the second sub-feature map at each pixel location to obtain the second feature value at each pixel location. The normalization layer is used to normalize based on the second feature value at each pixel location to obtain the second weight at each pixel location.

[0075] Therefore, the spatial dimension weight parameters are predicted by the second prediction network, which includes a one-dimensional convolutional layer and a normalization layer. The one-dimensional convolutional layer is used to reduce the spatial dimension of the second sub-feature map at each pixel position to obtain the second feature value at each pixel position. The normalization layer is used to normalize based on the second feature value at each pixel position to obtain the second weight at each pixel position. Thus, the accuracy of weight prediction can be improved while simplifying the second prediction network as much as possible.

[0076] In some disclosed embodiments, the feature extraction module 51 includes an extraction submodule for performing feature extraction at several scales based on the signal to be identified, to obtain second feature maps at several scales; wherein the second feature maps at each scale are extracted by convolution kernels at the corresponding scales; the feature extraction module 51 includes a fusion submodule for fusing the second feature maps at several scales to obtain a first feature map of the signal to be identified.

[0077] Therefore, by fusing the second feature maps extracted at several scales to obtain the first feature map, we can fully explore the deep feature representation and help improve the accuracy of the first feature map.

[0078] In some disclosed embodiments, each scale corresponds to several convolutional kernels. The convolutional kernels corresponding to the same scale have the same width but different lengths, while the convolutional kernels corresponding to different scales have different widths. The extraction submodule includes a scale extraction unit, which is used to extract features from the signal to be recognized based on several convolutional kernels corresponding to each scale to obtain several third sub-feature maps corresponding to the scale. The extraction submodule includes a scale fusion unit, which is used to stitch together several third sub-feature maps corresponding to the scale to obtain a second feature map of the scale. The fusion submodule includes a resolution adjustment unit, which is used to adjust the second feature maps of several scales to the same resolution. The fusion submodule includes a feature map stitching unit, which is used to stitch together the adjusted second feature maps to obtain a first feature map.

[0079] Therefore, in the same dimension, the third sub-feature maps extracted by convolutional kernels with the same width but different lengths can be stitched together to fuse the feature information extracted by convolutional kernels with different receptive fields in the same dimension. This can help to further extract the deep feature representation of the signal to be identified, thereby helping to improve the accuracy of the first feature map.

[0080] In some disclosed embodiments, the first feature map is extracted by a feature extraction network, which includes several extraction sub-networks corresponding to different scales, and a fusion sub-network connected to each extraction sub-network. Each extraction sub-network corresponding to a different scale includes several convolutional kernels. The convolutional kernels corresponding to the same scale have the same width but different lengths, while the convolutional kernels corresponding to different scales have different widths.

[0081] Therefore, the first feature map is extracted by a feature extraction network, which includes several extraction sub-networks corresponding to different scales, and a fusion sub-network connected to each extraction sub-network. Each extraction sub-network at each scale includes several convolutional kernels. The convolutional kernels at the same scale have the same width but different lengths, while the convolutional kernels at different scales have different widths. Thus, the deep feature representation of the signal to be identified can be fully extracted while simplifying the network structure of the feature extraction network as much as possible, thereby improving the accuracy of the first feature map.

[0082] In some disclosed embodiments, the feature fusion module 54 is specifically used to fuse the first feature map and the weighted feature maps of each dimension to obtain a fused feature map.

[0083] Therefore, by fusing the first feature map and the weighted feature maps of each dimension, a fused feature map is obtained. Thus, the first feature map is further referenced during the feature map fusion process, which can minimize the loss of information and help improve the accuracy of subsequent modulation category recognition.

[0084] Please see Figure 6 , Figure 6This is a schematic diagram of an embodiment of the electronic device 60 of this application. The electronic device 60 includes a memory 61 and a processor 62 coupled to each other. The memory 61 stores program instructions, and the processor 62 is used to execute the program instructions to implement the steps in any of the above-described modulation category identification method embodiments.

[0085] Specifically, processor 62 controls itself and memory 61 to implement the steps in any of the modulation category identification method embodiments described above. Processor 62 can also be referred to as a CPU (Central Processing Unit). Processor 62 may be an integrated circuit chip with signal processing capabilities. Processor 62 can also be a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor. Furthermore, processor 62 can be implemented using integrated circuit chips.

[0086] The above-described scheme, on the one hand, reduces the impact of environmental factors such as low signal-to-noise ratio, Doppler shift, multiple modulation categories, and limited prior information on modulation category identification because it does not rely on prior information or manually designed feature parameters during the modulation category identification process. On the other hand, by combining weighted feature maps of several dimensions such as channel dimension and spatial dimension for modulation classification, it can extract as much feature information related to the modulation category as possible from the signal to be identified, which helps to adapt to as many modulation categories as possible. Therefore, it can improve the accuracy and robustness of modulation category identification.

[0087] Please see Figure 7 , Figure 7 This is a schematic diagram of a frame of an embodiment of the receiver 70 of this application. The receiver 70 includes a radio frequency circuit 71, an identification circuit 72, and a demodulation circuit 73 connected in sequence. The radio frequency circuit 71 is used to receive a signal to be identified. The identification circuit 72 is an electronic device as described in any of the above-described electronic device embodiments, used to identify the modulation category of the signal to be identified. The demodulation circuit 73 is used to demodulate the signal to be identified based on the identified modulation category. In addition, by way of example, besides the above-described circuits, the receiver 70 may further include, but is not limited to, analog-to-digital conversion circuits, gain circuits, etc., which are not limited here.

[0088] The above-described scheme, on the one hand, reduces the impact of environmental factors such as low signal-to-noise ratio, Doppler shift, multiple modulation categories, and limited prior information on modulation category identification because it does not rely on prior information or manually designed feature parameters during the modulation category identification process. On the other hand, by combining weighted feature maps of several dimensions such as channel dimension and spatial dimension for modulation classification, it can extract as much feature information related to the modulation category as possible from the signal to be identified, which helps to adapt to as many modulation categories as possible. Therefore, it can improve the accuracy and robustness of modulation category identification.

[0089] Please see Figure 8 , Figure 8 This is a schematic diagram of a framework of an embodiment of the communication device 80 of this application. The communication device 80 includes: an antenna 81, a duplexer 82, a receiver 83, and a transmitter 84. The antenna 81 is connected to the antenna connection terminal 821 of the duplexer 82, the receiver 83 is connected to the signal output terminal 822 of the duplexer 82, and the transmitter 84 is connected to the signal input terminal 823 of the duplexer 82. The receiver 83 is the receiver in any of the above-disclosed receiver embodiments. It should be noted that the antenna 81 is used to convert received electromagnetic waves in free space into electrical signals and transmit the electrical signals to the duplexer 82, so that the duplexer 82 transmits the electrical signals to the receiver 83. The antenna 81 is also used to convert the electrical signals emitted by the transmitter 84 into electromagnetic waves and radiate them into free space. The electrical signals emitted by the transmitter 84 are transmitted to the antenna 81 via the duplexer 82. It should be noted that the communication device 80 may include, but is not limited to: mobile phones, tablet computers, walkie-talkies, etc., and is not limited here.

[0090] The above-described scheme, on the one hand, reduces the impact of environmental factors such as low signal-to-noise ratio, Doppler shift, multiple modulation categories, and limited prior information on modulation category identification because it does not rely on prior information or manually designed feature parameters during the modulation category identification process. On the other hand, by combining weighted feature maps of several dimensions such as channel dimension and spatial dimension for modulation classification, it can extract as much feature information related to the modulation category as possible from the signal to be identified, which helps to adapt to as many modulation categories as possible. Therefore, it can improve the accuracy and robustness of modulation category identification.

[0091] Please see Figure 9 , Figure 9 This is a schematic diagram of a framework of an embodiment of the computer-readable storage medium 90 of this application. The computer-readable storage medium 90 stores program instructions 91 that can be executed by a processor. The program instructions 91 are used to implement the steps in any of the above-described embodiments of the modulation category identification method.

[0092] The above-described scheme, on the one hand, reduces the impact of environmental factors such as low signal-to-noise ratio, Doppler shift, multiple modulation categories, and limited prior information on modulation category identification because it does not rely on prior information or manually designed feature parameters during the modulation category identification process. On the other hand, by combining weighted feature maps of several dimensions such as channel dimension and spatial dimension for modulation classification, it can extract as much feature information related to the modulation category as possible from the signal to be identified, which helps to adapt to as many modulation categories as possible. Therefore, it can improve the accuracy and robustness of modulation category identification.

[0093] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0094] The description of the various embodiments above tends to emphasize the differences between the various embodiments. The similarities or similarities between them can be referred to, and for the sake of brevity, they will not be repeated here.

[0095] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus implementations described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0096] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0097] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0098] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A modulation category identification method, characterized in that, include: For each scale, features are extracted from the signal to be identified based on several convolutional kernels corresponding to the scale, resulting in several third sub-feature maps corresponding to the scale. These third sub-feature maps are then concatenated to obtain a second feature map for the scale. Each scale corresponds to several convolutional kernels. Convolutional kernels with the same scale have the same width but different lengths, while convolutional kernels with different scales have different widths. The second feature maps of several scales are adjusted to the same resolution, and the adjusted second feature maps are stitched together to obtain a first feature map; wherein, the first feature map includes a first sub-feature map of several channels, and the first sub-feature maps of several channels have the same resolution; Based on the first feature map, prediction is performed in the channel dimension to obtain the first weight of the plurality of channels, and based on the first feature map, prediction is performed in the spatial dimension to obtain the second weight of each pixel position; wherein, the pixel values ​​of the first sub-feature maps of the plurality of channels at the same pixel position constitute the second sub-feature map. The first sub-feature maps of the same channel are weighted based on the first weight of each channel to obtain the weighted feature map of the channel dimension, and the second sub-feature maps of the same pixel position are weighted based on the second weight of each pixel position to obtain the weighted feature map of the spatial dimension. The fused feature map is obtained by fusing the weighted feature maps based at least on the channel dimension and the spatial dimension. Based on the fused feature map, the modulation category of the signal to be identified is obtained.

2. The method according to claim 1, characterized in that, The step of predicting the first weights of the plurality of channels based on the first feature map in the channel dimension includes: The first sub-feature maps of the plurality of channels are each subjected to global pooling to obtain a global feature map; wherein, the global feature map includes the first feature value of each of the channels; Based on the global feature map, prediction is performed to obtain the first weight of each channel.

3. The method according to claim 1, characterized in that, The weight parameters of the channel dimension are predicted by a first prediction network, which includes a global pooling layer and a fully connected layer. The global pooling layer is used to perform global pooling on the first sub-feature maps of the plurality of channels respectively to obtain a global feature map, and the global feature map includes the first feature value of each of the channels. The fully connected layer is used to make predictions based on the global feature map to obtain the first weight of each of the channels.

4. The method according to claim 1, characterized in that, The step of predicting based on the first feature map in the spatial dimension to obtain the second weights for each pixel location includes: The second sub-feature map of each pixel location is subjected to feature dimensionality reduction to obtain the second feature value of each pixel location; The second weight of each pixel position is obtained by normalizing the second feature value of each pixel position.

5. The method according to claim 1, characterized in that, The weight parameters of the spatial dimension are predicted by a second prediction network, which includes a one-dimensional convolutional layer and a normalization layer. The one-dimensional convolutional layer is used to perform feature dimensionality reduction on the second sub-feature map of each pixel position to obtain the second feature value of each pixel position. The normalization layer is used to normalize based on the second feature value of each pixel position to obtain the second weight of each pixel position.

6. The method according to claim 1, characterized in that, The first feature map is extracted by a feature extraction network, which includes extraction sub-networks corresponding to the plurality of scales, and fusion sub-networks connected to each of the extraction sub-networks respectively; The extraction subnetwork corresponding to each scale includes several convolutional kernels. The convolutional kernels corresponding to the same scale have the same width but different lengths, while the convolutional kernels corresponding to different scales have different widths.

7. The method according to claim 1, characterized in that, The fusion of weighted feature maps based at least on each of the aforementioned dimensions to obtain a fused feature map includes: The fused feature map is obtained by fusing the first feature map and the weighted feature maps of each dimension.

8. A modulation category identification device, characterized in that, include: The extraction submodule is used to extract features from the signal to be identified based on several convolutional kernels corresponding to each scale, to obtain several third sub-feature maps corresponding to the scale, and to concatenate the several third sub-feature maps corresponding to the scale to obtain the second feature map of the scale; wherein, each scale corresponds to several convolutional kernels, the convolutional kernels corresponding to the same scale have the same width but different lengths, and the convolutional kernels corresponding to different scales have different widths. The fusion submodule is used to adjust the second feature maps of several scales to the same resolution and stitch the adjusted second feature maps together to obtain a first feature map; wherein, the first feature map includes a first sub-feature map of several channels, and the first sub-feature maps of several channels have the same resolution; The weight prediction module is used to predict the first weight of the plurality of channels based on the first feature map in the channel dimension, and to predict the second weight of each pixel position based on the first feature map in the spatial dimension; wherein, the pixel values ​​of the first sub-feature maps of the plurality of channels at the same pixel position constitute the second sub-feature map. The feature weighting module is used to weight the first sub-feature maps of the same channel based on the first weight of each channel to obtain the weighted feature map of the channel dimension, and to weight the second sub-feature maps of the same pixel position based on the second weight of each pixel position to obtain the weighted feature map of the spatial dimension. The feature fusion module is used to fuse the weighted feature maps based at least on the channel dimension and the spatial dimension to obtain a fused feature map; The modulation classification module is used to classify the signal to be identified based on the fused feature map to obtain the modulation category.

9. An electronic device, characterized in that, The method includes a memory and a processor coupled to each other, wherein the memory stores program instructions and the processor executes the program instructions to implement the modulation category identification method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The device stores program instructions that can be executed by a processor, the program instructions being used to implement the modulation category identification method according to any one of claims 1 to 7.

Images