A deepemd-based few-shot automatic modulation mode recognition method

Abstract

The application discloses a few-lens automatic modulation mode recognition method based on DeepEMD, and the recognition method comprises the following steps: acquiring a signal to be recognized; and obtaining the modulation category of the signal to be recognized by using a trained DeepEMD network, wherein the trained DeepEMD network is obtained by training a training sample set, the DeepEMD network comprises a feature vector group extraction network, a feature vector group weight generation network and an EMD network, the feature vector group extraction network is used for extracting a feature vector group, the feature vector group weight generation network is used for generating the weight of the feature vector group, the EMD network is used for recognizing the modulation category of the signal according to the feature vector group and the corresponding weight, and the feature vector group comprises a plurality of feature vectors. The DeepEMD network adopted in the application fully utilizes the local information of the signal, meanwhile, the weight of each feature is calculated, so that the features with high correlation play a leading role in the recognition, and the noise interference is reduced. Meanwhile, ideal results can be obtained in the case of sample deficiency.

Description

Technical Field

[0001] This invention belongs to the field of signal modulation, specifically relating to a method for automatic modulation mode recognition with few lenses based on DeepEMD. Background Technology

[0002] Automatic modulation classification (AMC) is an intermediate process in signal detection and demodulation, designed to identify the modulation scheme of received wireless communication signals. It plays an important role in real life, such as in spectrum detection, cognitive radio, and attitude perception.

[0003] Traditional recognition methods can be broadly categorized into two types based on their principles: likelihood estimation-based methods and feature-based methods. However, with the increasing prevalence of digital modulation techniques, these two types are no longer sufficient for everyday needs.

[0004] With the development of deep learning, it has gradually become the preferred method for solving the AMC (Acceptable Classification) problem. This is mainly because modern communication systems use denser modulation schemes to achieve efficient, high-capacity data transmission, and deep neural networks have strong learning capabilities for complex data and can achieve high classification accuracy. Although deep learning has achieved some success in AMC tasks, it can be observed that different signal-to-noise ratios may cause signals to be embedded in a given metric space far from the same class. Deep neural networks can avoid this problem using fully supervised learning when there is a sufficient training set, but this is unavoidable when there are few samples. In addition, a single feature vector cannot represent the sample signal well and may lose some important local information. Local information can provide class discriminative and transferable information, which is an important aspect of achieving AMC. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a method for automatic modulation mode recognition with few lenses based on DeepEMD.

[0006] The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] A method for identifying automatic modulation modes with few lenses based on DeepEMD, the method comprising:

[0008] S1. Acquire the signal to be identified;

[0009] S2. Obtain the modulation category of the signal to be identified using the trained DeepEMD network. The trained DeepEMD network is trained using a training sample set. The DeepEMD network includes a feature vector extraction network, a feature vector weight generation network, and an EMD network. The feature vector extraction network is used to extract feature vector groups, the feature vector weight generation network is used to generate weights for the feature vector groups, and the EMD network is used to identify the modulation category of the signal based on the feature vector groups and their corresponding weights. The feature vector groups include several feature vectors.

[0010] Optionally, the feature vector extraction network includes two convolutional networks connected in parallel. Each convolutional network includes a first convolutional layer, a first ReLU activation layer, a first batch normalization layer, a first max pooling layer, a second convolutional layer, a second ReLU activation layer, a second batch normalization layer, a second max pooling layer, a third convolutional layer, a third ReLU activation layer, a third batch normalization layer, and a third max pooling layer connected in sequence.

[0011] The feature vector group weight generation network is used to select a larger value between the relevance score and zero, and normalize the larger value to obtain the weight of the i-th feature vector. The relevance score is calculated by taking the dot product of the i-th feature vector of one feature vector group and the average feature of another feature vector group.

[0012] The EMD network is used to calculate the cosine similarity between two feature vectors in two feature vector groups, and obtains a similarity score based on the cosine similarity and the corresponding weights.

[0013] Optionally, the training method for the DeepEMD network includes:

[0014] The feature vector extraction network is pre-trained to obtain a pre-trained feature vector extraction network.

[0015] The DeepEMD network is obtained by training the pre-trained feature vector extraction network, the feature vector weight generation network, and the EMD network.

[0016] Optionally, the feature vector extraction network is pre-trained to obtain a pre-trained feature vector extraction network, including:

[0017] A first training sample is selected from the training sample set, and the first training sample is split into two paths and input into the feature vector group extraction network to obtain two feature vector groups to be trained.

[0018] The two sets of feature vectors to be trained are concatenated to obtain the concatenated set of feature vectors.

[0019] The concatenated feature vector group is input into the fully connected layer to obtain the score of the predicted class;

[0020] The softmax function is used to convert the predicted class score into the modulated class probability;

[0021] The probability of the modulation category is logarithmically calculated, and a first loss value is calculated using the first cross-entropy loss function. Backpropagation is then performed based on the first loss value to update the parameters of the feature vector extraction network until a first preset condition is met, thus obtaining the pre-trained feature vector extraction network.

[0022] Optionally, the training sample set includes a training support set and a training query set;

[0023] The DeepEMD network is obtained by training the pre-trained feature vector extraction network, the feature vector weight generation network, and the EMD network, including:

[0024] The training support set is split into two paths and input into the pre-trained feature vector extraction network to obtain the first feature vector group. The training query set is split into two paths and input into the pre-trained feature vector extraction network to obtain the second feature vector group.

[0025] The first feature vector group and the second feature vector group are input into the feature vector group weight generation network to obtain the weights of the feature vectors in the first feature vector and the weights of the feature vectors in the second feature vector.

[0026] Calculate the first cosine similarity between the feature vectors in the first feature vector group and the feature vectors in the second feature vector group;

[0027] Based on the weights of the feature vectors in the first feature vector group, the weights of the feature vectors in the second feature vector group, and the first cosine similarity, the first optimal match is obtained using OpenCV library functions;

[0028] A similarity score is obtained based on the first optimal match and the first cosine similarity.

[0029] Based on the similarity score, a second loss value is calculated using the second cross-entropy loss function. Backpropagation is then performed based on the second loss value to update the parameters of the DeepEMD network until a second preset condition is met, resulting in a trained DeepEMD network.

[0030] Optionally, the training method for the DeepEMD network further includes:

[0031] Obtain a test sample set, which includes a labeled test support set and a test query set to be identified;

[0032] The test support set and the test query set are input into the trained DeepEMD network to obtain the modulation type of the modulation signal.

[0033] Optionally, the test support set and the test query set are input into the trained DeepEMD network to obtain the modulation type of the modulated signal, including:

[0034] The test support set is input into the feature vector extraction network to obtain the third feature vector set;

[0035] The test query set is input into the feature vector extraction network to obtain the fourth feature vector set;

[0036] The third and fourth feature vector sets are input into the feature weight generation network to obtain the weights of the feature vectors in the third feature vector set and the weights of the feature vectors in the fourth feature vector set;

[0037] Calculate the second cosine similarity between the feature vectors in the third feature vector group and the feature vectors in the fourth feature vector group;

[0038] Based on the weights of the feature vectors in the third feature vector group, the weights of the feature vectors in the fourth feature vector group, and the second cosine similarity, the second optimal match is obtained using OpenCV library functions;

[0039] The final similarity score is obtained based on the second optimal match and the second cosine similarity.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] The DeepEMD network used in this invention fully utilizes local information of the signal and calculates weights for each feature, ensuring that highly relevant features play a dominant role in recognition and reducing noise interference. It can also achieve ideal results even with limited samples.

[0042] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating a method for automatic modulation mode recognition with few lenses based on DeepEMD, provided in an embodiment of the present invention.

[0044] Figure 2 This is a flowchart illustrating the implementation of a method for automatic modulation mode recognition with few lenses based on DeepEMD, as provided in an embodiment of the present invention.

[0045] Figure 3This is a diagram illustrating the training and testing framework of a DeepEMD network provided in an embodiment of the present invention.

[0046] Figure 4 This is a feature vector group extraction network structure diagram provided in an embodiment of the present invention;

[0047] Figure 5 This is a feature vector group weight generation network structure diagram provided in an embodiment of the present invention.

[0048] Figure 6 This is an EMD network structure diagram provided in an embodiment of the present invention. Detailed Implementation

[0049] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0050] Example 1

[0051] Currently, automatic modulation classification is mainly treated as a multi-classification problem. Convolutional neural networks obtain feature vector groups through convolutional layers, activation layers, batch normalization layers, and pooling layers, then convert them into one-dimensional vectors through fully connected layers. The category corresponding to the maximum value output by the softmax function is used as the predicted category. Maqsood Hussain Shah, in his paper "An effective approach for low-complexity maximum likelihood based automatic modulation classification of STBC-MIMO systems," proposed a low-complexity likelihood-based method that uses zero-forced equalization to modify the typical average likelihood ratio test (ALRT) function. Liu Qiaoping et al., in their paper "Automatic Modulation Recognition Based on Deep Neural Networks," proposed generating IQ datasets with 20 signal-to-noise ratios and 8 modulation types based on GNU Radio, and using the trained deep neural network model RESNET for classification testing. This method has a simple structure but requires a large number of samples. Zhang Tingting et al., in their paper "Research on Automatic Modulation Recognition Method Based on Deep Residual Networks," proposed using a recurrent neural network structure to represent electromagnetic signal sequences and established a recognition method based on a multi-layer bidirectional GRU network structure. Cui Tianshu et al., in their reference "Automatic Modulation Recognition Algorithm for Satellite Signals Using Convolutional Neural Networks," proposed using convolutional networks to first extract in-phase orthogonal correlation features of the signal, then extract time-domain features, and finally extract the mean of features from each channel for classification. However, this method does not fully consider the influence of local signals on modulation recognition. Shi Yunhao et al., in their reference "A Small-Sample Modulation Recognition Method Based on Ensemble Learning and Feature Dimensionality Reduction," proposed a small-sample modulation classification model based on ensemble learning and feature dimensionality reduction. This method first constructs a feature set by integrating manually extracted features with features automatically extracted by deep learning. Then, a feature selection algorithm is designed to optimize the feature set and generate an efficient feature subset. Finally, a high-performance classifier with fast convergence is used to distinguish the signals. However, this method requires manually designed features and a large number of samples for training. Existing technologies struggle to achieve good results with limited samples and cannot fully utilize local information in the signal.

[0052] The purpose of this invention is to provide a small-sample automatic modulation scheme identification method based on DeepEMD, thereby improving the accuracy of small-sample automatic modulation scheme identification by fully mining local information. Please see below. Figure 1 , Figure 2 and Figure 3 , Figure 1 This is a flowchart illustrating a method for automatic modulation mode recognition with few lenses based on DeepEMD, provided in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the implementation of a method for automatic modulation mode recognition with few lenses based on DeepEMD, as provided in an embodiment of the present invention. Figure 3 This is a training and testing framework diagram of a DeepEMD network provided in an embodiment of the present invention. The method for automatic modulation mode recognition with few shots based on DeepEMD includes:

[0053] S1. Acquire the signal to be identified;

[0054] S2. The modulation category of the signal to be identified is obtained using the trained DeepEMD network. The trained DeepEMD network is trained using a training sample set. The DeepEMD network includes a feature vector extraction network, a feature vector weight generation network, and an EMD network. The feature vector extraction network is used to extract feature vector groups, the feature vector weight generation network is used to generate the weights of the feature vector groups, and the EMD (earth mover's distances) network is used to identify the modulation category of the signal based on the feature vector groups and their corresponding weights. The feature vector groups include several feature vectors.

[0055] In other words, this embodiment can pre-set a sample set with labels (marking the category of the signal), then input the signal to be identified and the labeled sample set into the trained DeepEMD network. The trained DeepEMD network outputs a similarity score, and then selects the label of the sample set corresponding to the highest similarity score as the modulation category of the signal to be identified, thereby identifying the modulation category of the signal to be identified.

[0056] In this embodiment, a training sample set T is first provided. train and test sample set T test C×k samples are selected from the training sample set according to each category as the training support set S1, and the remaining samples are... One sample is used as the training query set Q1; similarly, the test sample set is divided into a labeled test support set S2 and a query set to be identified Q2, and the test sample set is a set of small sample modulated signals to be identified.

[0057] In this context, the modulation categories of the signals in the test sample set are different from those in the training sample set; where C represents the number of modulation categories in the training sample set; and k is the number of samples of each modulation category in the training support set. The total number of samples in the training sample set belonging to class k.

[0058] For example, WBFM, BPSK, CPFSK, PAM4, 16QAM, 64QAM, and QPSK from the RML2016.10a dataset are used as the training sample set T. trainDuring training, C classes are randomly selected from the training sample set, and k signal samples from each class are selected as the training support set S1. The remaining samples are... One signal is used as the training query set Q1.

[0059] The 8PSK, AM-DSB, and AM-SSB samples from the RML2016.10a dataset were used as the test sample set T. test During testing, C classes are randomly selected from the test sample set, and k samples from each class are selected as the test support set S2. The remaining samples are... One sample is used as the test query set Q2.

[0060] In this embodiment, the provided DeepEMD network includes a feature vector group extraction network, a feature vector group weight generation network, and an EMD network.

[0061] For details, please see Figure 4 The feature vector extraction network consists of two parallel convolutional networks. Each convolutional network consists of a first convolutional layer, a first ReLU activation layer, a first batch normalization layer, a first max pooling layer, a second convolutional layer, a second ReLU activation layer, a second batch normalization layer, a second max pooling layer, a third convolutional layer, a third ReLU activation layer, a third batch normalization layer, and a third max pooling layer, connected in sequence.

[0062] The three convolutional kernels are all 2×2 in size and have a stride of 3, with a total of 16 kernels in the convolutional layer; the max pooling layers are all 2×2 in size and have a stride of 2.

[0063] For details, please see Figure 5 The feature vector group weight generation network is used to select a larger value between the relevance score and zero, and normalize the larger value to obtain the weight of the i-th feature vector. The relevance score is calculated by taking the dot product of the i-th feature vector of one feature vector group and the average feature of another feature vector group. The average feature is the sum of all feature vectors in the feature vector group divided by the size of the feature vector.

[0064] The feature vector group weight generation network uses a cross-reference mechanism to obtain the weights of each feature vector. Specifically, the cross-reference mechanism uses the dot product of the feature vectors of one feature vector group and the average feature of another feature vector group to obtain the relevance score. Assuming u... i v j If the feature vectors are from two feature vector groups respectively, then the formula for calculating the relevance score is:

[0065]

[0066] Among them, s i It is the eigenvector ui The corresponding initial weights, max is to ensure the weight values ​​are non-negative, H is the length of the feature vector group, W is the height of the feature vector group, therefore HW is the size of the feature vector group, after calculating s i Afterwards, regarding s i Normalization yields the eigenvector u i The final weight This ensures that both parties are matched with the same total weight. The calculation formula is:

[0067]

[0068] Among them, s j It is the feature vector v j The corresponding initial weights. Similarly, u i With v j By exchanging the positions, the feature vector v can be obtained. j The final weight

[0069] For details, please see Figure 6 The EMD network is used to calculate the cosine similarity between two feature vectors in two feature vector groups, and obtains a similarity score based on the cosine similarity and the corresponding weights.

[0070] The EMD network first calculates the cosine similarity of each feature vector in two feature vector sets. Specifically, it calculates the cosine similarity of each feature vector u in a feature vector set. i With each feature vector v in another feature vector group j The cosine similarity is calculated using the following formula:

[0071]

[0072] Among them, c ij Let T be the cosine similarity and T be the transpose.

[0073] After calculating the cosine similarity, the EMD is calculated by combining the weights of the feature vectors. Here, the OpenCV library function is used. The OpenCV library function solves the linear programming problem through an improved simplex algorithm, which is fast and accurate.

[0074] In one specific embodiment, the training method for the DeepEMD network includes:

[0075] Step 1: Pre-train the feature vector extraction network to obtain the pre-trained feature vector extraction network.

[0076] Specifically, the training sample set is divided into two paths, which are then fed into the feature vector extraction network. The outputs of the two paths are then concatenated to form the overall signal features. The specific steps are as follows: First, three classes are randomly selected from the training sample set, with 10,000 sample signals selected from each class. Each sample signal is then sequentially fed into the feature vector extraction network and mapped into the feature space. The mapping function is f(x). i ).

[0077] Step 1.1: Select the first training sample from the training sample set, and input the first training sample into the feature vector extraction network in two ways to obtain two feature vector groups to be trained.

[0078] First, set the number of iterations to t, where t ≥ 1000, and initialize the number of iterations to t = 0.

[0079] Step 1.2: Concatenate the two sets of feature vectors to be trained to obtain the concatenated feature vector set.

[0080] Step 1.3: Input the concatenated feature vector group into the fully connected layer to obtain the score of the predicted class.

[0081] Specifically, the first training sample is, for example, a modulated signal of shape 2×128. It is split into two paths and sent to the feature vector extraction network, resulting in two 8×16 feature vector groups. Finally, the two feature vector groups are concatenated to obtain a 16×16 feature vector group, which is then sent to the fully connected layer to obtain a 1×256 feature vector. The mapping function is defined as:

[0082]

[0083] In this process, each training sample in the training sample set will obtain a mapping representation after passing through the feature vector extraction network and the fully connected layer. Simultaneously predict the signal modulation type.

[0084] Step 1.4: Use the softmax function to convert the predicted class score into the modulated class probability.

[0085] Step 1.5: Take the logarithm of the probability of the modulation category and calculate the first loss value using the first cross-entropy loss function. Then, backpropagate based on the first loss value to update the parameters of the feature vector extraction network until the first preset condition is met, thus obtaining the pre-trained feature vector extraction network.

[0086] Specifically, suppose the score of the predicted class of the i-th signal sample obtained through the fully connected layer is z. i First, use the softmax function to predict the class score z. i The probability of converting to the corresponding modulation category of the i-th signal sample is given by the following formula:

[0087]

[0088] Then, taking the logarithm of probability p to amplify the probability difference between predicted samples, the first cross-entropy loss function is calculated; let the loss function be defined as:

[0089]

[0090] Where, q i Defined as the true category of signal modulation mode, and C as the total number of categories in the training sample set. Then, the loss function is backpropagated to update the parameters of the feature vector extraction network until the training iterations are reached, resulting in a pre-trained feature vector extraction network.

[0091] Step 2: Train the DeepEMD network using the pre-trained feature vector extraction network, feature vector weight generation network, and EMD network to obtain the trained DeepEMD network.

[0092] Specifically, the training sample set T train The dataset is divided into a support set and a query set, which are then fed into a feature vector extraction network to obtain the feature vector set of the support set samples (i.e., the first feature vector set) and the feature vector set of the query samples (i.e., the second feature vector set). The predicted signal modulation category is then obtained through a feature weight generation network and an EMD network. The DeepEMD network is then trained to obtain the trained DeepEMD network.

[0093] Step 2.1: Divide the training support set into two inputs and feed them into the pre-trained feature vector extraction network to obtain the first feature vector set. Divide the training query set into two inputs and feed them into the pre-trained feature vector extraction network to obtain the second feature vector set.

[0094] Set the number of iterations to t, where t ≥ 1000, and initialize the number of iterations t = 0. In the training sample set T... train In the training support set S1, k sample signals from each class are randomly selected from C classes, and the remaining signal samples are used as the training query set Q1. The training support set and the training query set are then sent to the feature vector extraction network in two separate paths to obtain the feature vector group of the support set sample class and the feature vector group of the sample to be queried.

[0095] The specific steps are as follows:

[0096] First, the sample signals x from the training support set are... i The data is split into two streams and fed into the feature vector extraction network to obtain the feature mapping representation. Secondly, the C-type modulation signal mapping representation contained in the support set S1 is trained. The class prototype of a class-C modulated signal in the feature space is calculated using the following formula:

[0097]

[0098] Where, n k S represents the class prototype of the k-th type of modulated signal in the spatial feature space; k x represents the set of modulated signals of class k in the training support set; i S represents k The i-th modulated signal, y i For x i Corresponding tag; |S k |for S k The total number of modulated signals; f() is the mapping function of the spatial feature extraction network; f(x) i ) is the sample signal x i Mapping representation within the feature space.

[0099] Then the sample signals x in the training query set will be used. j The inputs are split into two paths and fed into the feature vector extraction network to obtain the feature mapping representation.

[0100] Step 2.2: Input the first feature vector group and the second feature vector group into the feature vector group weight generation network to obtain the weights of the feature vectors in the first feature vector and the weights of the feature vectors in the second feature vector.

[0101] Step 2.3: Calculate the first cosine similarity between the eigenvectors in the first eigenvector group and the eigenvectors in the second eigenvector group.

[0102] Specifically, the feature vector μ of the first feature vector group in the training support set is calculated. i The feature vector ν of the second feature vector group in the training query set j The cosine similarity is calculated using the following formula:

[0103]

[0104] Step 2.4: Based on the weights of the feature vectors in the first feature vector group, the weights of the feature vectors in the second feature vector group, and the first cosine similarity, the first optimal match is obtained using OpenCV library functions, that is, the cost required to transform the first feature vector group into the second feature vector group is minimized.

[0105] Specifically, the weights of the feature vectors in the first feature vector group, the weights of the feature vectors in the second feature vector group, and the first cosine similarity are substituted into the OpenCV library function to obtain the first optimal match.

[0106] Step 2.5: Obtain the similarity score based on the first optimal match and the first cosine similarity.

[0107] Specifically, the similarity score is calculated using the following formula:

[0108]

[0109] Where H is the length of the feature vector group, W is the height of the feature vector group, and HW represents the size of the feature vector group.

[0110] Step 2.6: Based on the similarity score, calculate the second loss value using the second cross-entropy loss function, and perform backpropagation based on the second loss value to update the parameters of the DeepEMD network until the second preset condition is met, such as reaching a preset number of times, to obtain the trained DeepEMD network.

[0111] In this embodiment, after obtaining the trained DeepEMD network, the trained DeepEMD network is also tested using a test set, including:

[0112] Step 3: Obtain the test sample set. The test sample set includes a labeled test support set and a test query set to be identified. The test support set S2 is a labeled small sample modulated signal, and the query set Q2 is a small sample modulated signal to be identified.

[0113] Step 4: Input the test support set and test query set into the trained DeepEMD network to obtain the modulation type of the modulated signal.

[0114] Step 4.1: Input the test support set into the feature vector extraction network to obtain the third feature vector set.

[0115] Specifically, from the test sample set T test We select the labeled test support set S2 and the test query set Q2 to be queried. We then send the test support set S2 into the feature vector extraction network to obtain the third feature vector set, whose mapping expression is as follows:

[0116]

[0117] Where, n k S represents the class prototype of the k-th type of modulated signal in the spatial feature space; k This indicates that the test supports the k-th type of modulation signal set; x i S represents k The i-th modulated signal, y i For x i Corresponding tag; |S k |for S kThe total number of modulated signals; f() is the mapping function of the spatial feature extraction network; f(x) i ) is the sample signal x i Mapping representation within the feature space.

[0118] Step 4.2: Input the test query set into the feature vector extraction network to obtain the fourth feature vector group, which is the feature vector group of the sample to be queried.

[0119] Specifically, the modulated signals of the small samples to be identified in the test query set are sent to the feature vector extraction network to obtain the fourth feature vector group f(x) corresponding to each sample signal. j ).

[0120] Step 4.3: Input the third and fourth feature vector sets into the feature vector set weight generation network to obtain the weights of the feature vectors in the third feature vector set. The weights of the eigenvectors in the fourth eigenvector group

[0121] Step 4.4: Calculate the second cosine similarity between the eigenvectors in the third eigenvector group and the eigenvectors in the fourth eigenvector group.

[0122] Step 4.5: Based on the weights of the feature vectors in the third feature vector group, the weights of the feature vectors in the fourth feature vector group, and the second cosine similarity, the second optimal match is obtained using OpenCV library functions.

[0123] Step 4.6: Obtain the final similarity score based on the second optimal matching and the second cosine similarity. Compare the final similarity scores of the sample signals to be queried in the test query set with the final similarity scores of each sample signal in the test support set, and select the one with the highest final similarity score as the category of the predicted modulation signal.

[0124] The DeepEMD network used in this invention fully utilizes local information of the signal and calculates weights for each feature, ensuring that highly relevant features play a dominant role in recognition and reducing noise interference. It can also achieve ideal results even with limited samples.

[0125] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0126] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0127] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In this specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. While certain measures are described in different embodiments, this does not mean that these measures cannot be combined to produce good results.

[0128] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for automatic modulation mode recognition with few lenses based on DeepEMD, characterized in that, The identification method includes: S1. Acquire the signal to be identified; S2. Obtain the modulation category of the signal to be identified using the trained DeepEMD network, wherein the trained DeepEMD network is trained using a training sample set, and the DeepEMD network includes a feature vector group extraction network, a feature vector group weight generation network, and an EMD network. The feature vector group extraction network is used to extract feature vector groups, the feature vector group weight generation network is used to generate weights for the feature vector groups, and the EMD network is used to identify the modulation category of the signal based on the feature vector groups and the corresponding weights. The feature vector groups include several feature vectors. The training method for the DeepEMD network includes: The feature vector extraction network is pre-trained to obtain a pre-trained feature vector extraction network. The pre-trained feature vector extraction network, the feature vector weight generation network, and the EMD network are used to train the DeepEMD network to obtain the trained DeepEMD network. The training sample set includes a training support set and a training query set; The DeepEMD network is obtained by training the pre-trained feature vector extraction network, the feature vector weight generation network, and the EMD network, including: The training support set is split into two paths and input into the pre-trained feature vector extraction network to obtain the first feature vector group. The training query set is split into two paths and input into the pre-trained feature vector extraction network to obtain the second feature vector group. The first feature vector group and the second feature vector group are input into the feature vector group weight generation network to obtain the weights of the feature vectors in the first feature vector and the weights of the feature vectors in the second feature vector. Calculate the first cosine similarity between the feature vectors in the first feature vector group and the feature vectors in the second feature vector group; Based on the weights of the feature vectors in the first feature vector group, the weights of the feature vectors in the second feature vector group, and the first cosine similarity, the first optimal match is obtained using OpenCV library functions; A similarity score is obtained based on the first optimal match and the first cosine similarity. Based on the similarity score, a second loss value is calculated using the second cross-entropy loss function. Backpropagation is then performed based on the second loss value to update the parameters of the DeepEMD network until a second preset condition is met, resulting in a trained DeepEMD network.

2. The method for identifying automatic modulation modes with few lenses according to claim 1, characterized in that, The feature vector extraction network includes two convolutional networks connected in parallel. Each convolutional network includes a first convolutional layer, a first ReLU activation layer, a first batch normalization layer, a first max pooling layer, a second convolutional layer, a second ReLU activation layer, a second batch normalization layer, a second max pooling layer, a third convolutional layer, a third ReLU activation layer, a third batch normalization layer, and a third max pooling layer connected in sequence. The feature vector group weight generation network is used to select a larger value between the relevance score and zero, and normalize the larger value to obtain the weight of the i-th feature vector. The relevance score is calculated by taking the dot product of the i-th feature vector of one feature vector group and the average feature of another feature vector group. The EMD network is used to calculate the cosine similarity between two feature vectors in two feature vector groups, and obtains a similarity score based on the cosine similarity and the corresponding weights.

3. The method for identifying automatic modulation modes with few lenses according to claim 2, characterized in that, The feature vector extraction network is pre-trained to obtain a pre-trained feature vector extraction network, including: A first training sample is selected from the training sample set, and the first training sample is divided into two paths and input into the feature vector group extraction network to obtain two feature vector groups to be trained. The two sets of feature vectors to be trained are concatenated to obtain the concatenated set of feature vectors. The concatenated feature vector group is input into the fully connected layer to obtain the score of the predicted class; The softmax function is used to convert the predicted class score into the modulated class probability; The probability of the modulation category is logarithmically calculated, and a first loss value is calculated using the first cross-entropy loss function. Backpropagation is then performed based on the first loss value to update the parameters of the feature vector extraction network until a first preset condition is met, thus obtaining the pre-trained feature vector extraction network.

4. The method for identifying automatic modulation modes with few lenses according to claim 3, characterized in that, The training method for the DeepEMD network also includes: Obtain a test sample set, which includes a labeled test support set and a test query set to be identified; The test support set and the test query set are input into the trained DeepEMD network to obtain the modulation type of the modulation signal.

5. The method for identifying automatic modulation modes with few lenses according to claim 4, characterized in that, The test support set and the test query set are input into the trained DeepEMD network to obtain the modulation type of the modulated signal, including: The test support set is input into the feature vector extraction network to obtain the third feature vector set; The test query set is input into the feature vector extraction network to obtain the fourth feature vector set; The third and fourth feature vector sets are input into the feature weight generation network to obtain the weights of the feature vectors in the third feature vector set and the weights of the feature vectors in the fourth feature vector set; Calculate the second cosine similarity between the feature vectors in the third feature vector group and the feature vectors in the fourth feature vector group; Based on the weights of the feature vectors in the third feature vector group, the weights of the feature vectors in the fourth feature vector group, and the second cosine similarity, the second optimal match is obtained using OpenCV library functions; The final similarity score is obtained based on the second optimal match and the second cosine similarity.

Images