Small sample signal automatic modulation identification method and device

Abstract

The invention relates to the technical field of signal modulation identification, and particularly discloses a small sample signal automatic modulation identification method and device. The method comprises the following steps: performing task extraction in a source domain training set to obtain a support set and a query set; constructing a signal category reasoning model, respectively inputting the samples of the support set and the samples of the query set into a multi-level embedding function, and respectively averaging the feature maps of each level of the support samples of each category to obtain the feature maps of each level of each support category; through a multi-level relation measurement module, calculating each level of similarity relation score between the query sample and the support category feature map, and carrying out weighted summation to obtain a final similarity score; reconstructing the feature map of the support sample through a decoder, and calculating to obtain a classification error and a reconstruction error; carrying out back propagation on errors, and updating model parameters; and performing signal identification on target domain data by using the trained model. The objective of the invention is to solve the small sample dilemma of a traditional deep learning method in an AMR problem.

Description

Technical Field

[0001] The present invention relates to the field of signal modulation recognition, and in particular to a method and device for automatic modulation recognition of small sample signals. Background Art

[0002] Automatic signal modulation recognition is an important technology in the field of cognitive communications. Its purpose is to determine the modulation mode of the signal based on the received radio signal. It is the basis for further processing of non-cooperative signals and an important prerequisite for achieving efficient spectrum perception, spectrum understanding and spectrum utilization in non-cooperative communication scenarios. It is also one of the key research topics in the field of wireless communications in recent years.

[0003] Traditional signal automatic modulation recognition methods can be divided into methods based on maximum likelihood theory and methods based on expert features. The method based on likelihood theory is to perform maximum likelihood classification on the statistical characteristics of the signal according to the discrimination criteria. The method based on expert features is to manually design features, transform the signal into a certain feature space, and then design a classifier for classification. These two methods usually require a lot of prior knowledge, and their recognition accuracy and versatility are usually low.

[0004] With the development of deep learning technology, researchers have gradually noticed its powerful feature extraction and expression capabilities, and applied it to the field of automatic modulation recognition, and achieved considerable research results. In 2016, O'SHEA et al. first proposed a modulation recognition method that uses a convolutional neural network (CNN) to directly process the original in-phase and quadrature (IQ) signal. Subsequently, network architectures such as recurrent neural networks, long short-term memory networks, and denoising autoencoders have also been gradually applied to the problem of automatic modulation recognition of signals, and their recognition accuracy has been continuously improved. However, the performance of deep learning-based methods is heavily dependent on a large amount of high-quality labeled data. In actual non-cooperative scenarios, it is impractical to obtain sufficient high-quality signal samples in advance. Therefore, it is of great significance to study how to solve the small sample problem faced in the field of signal modulation recognition. Summary of the invention

[0005] In view of the above problems, an object of the present invention is to provide a method for automatic modulation recognition of small sample signals.

[0006] The second object of the present invention is to provide an automatic modulation identification device for small sample signals.

[0007] The first technical solution adopted by the present invention is:

[0008] S1. Extract C-Way K-shot tasks in the source domain training set to obtain the support set and query set;

[0009] S2, constructing a signal category inference model, inputting the samples of the support set and the samples of the query set into the multi-level embedding function of the signal category inference model respectively, obtaining the feature graphs of the support samples at all levels and the feature graphs of the query samples at all levels, and averaging the feature graphs of the support samples at all levels of each category to obtain the feature graphs of each support category at all levels;

[0010] S3, calculating the similarity relationship scores between the feature graphs at all levels of the query sample and the feature graphs at all levels of the support categories through the multi-level relationship measurement module of the signal category inference model, and performing weighted summation on the similarity relationship scores at all levels to obtain the similarity score between each support category corresponding to the query sample, thereby determining the predicted label of the query sample;

[0011] S4, reconstructing the feature graph of the supporting sample through the decoder of the signal category inference model to obtain a reconstructed supporting sample;

[0012] S5. Based on the predicted label and the reconstructed support sample, calculate and obtain a classification error and a reconstruction error;

[0013] S6, back-propagating the classification error and the reconstruction error to update the signal category inference model parameters;

[0014] S7. Repeat S1-S6 to iteratively train the signal category inference model, and use the trained signal category inference model to perform signal recognition on the target domain data to obtain a signal recognition result.

[0015] Optionally, S2 includes:

[0016] S21, inputting each support sample in the support set into the multi-level embedding function to obtain feature maps of each level of the support sample;

[0017] S22, averaging the feature maps of each level of the support samples of the same category to obtain feature maps of each level of each support category;

[0018] S23, sending each query sample in the query set into a multi-level embedding function to obtain feature graphs of each level of the query sample.

[0019] Optionally, the S3 includes:

[0020] S31, concatenating the first-level query sample feature graph and the first-level support category feature graph, and performing calculations using the first-level relationship submodule in the multi-level relationship measurement module to obtain a first-level similarity feature graph and a first-level similarity relationship score;

[0021] S32, concatenating the previous level similarity feature graph, the query feature graph at the same level, and the support category feature graph at the same level as the input of the next level relationship measurement module, until the similarity relationship scores between the query sample feature graphs at all levels and the support category feature graphs at all levels are calculated;

[0022] S33, performing weighted summation on the similarity relationship scores between the feature graphs of each level of the query sample and the feature graphs of each level of the support category to obtain a similarity relationship score between the query sample and each support category, and determining a predicted label for the query sample according to the highest similarity relationship score.

[0023] Optional,

[0024] The v-th level similarity feature graph expression is as follows:

[0025]

[0026] The first-level similarity feature graph expression is as follows:

[0027]

[0028] In the formula, is the feature map of the v-th level support sample, is the feature map of the query sample at level v, is the similarity feature map output by the v-1th level relation module. θ and They are multi-level embedding function and multi-level relationship measurement module respectively.

[0029] Optional,

[0030] The similarity score expression between the query sample and each supporting category is as follows:

[0031]

[0032] in,

[0033]

[0034] The similarity feature graph relationship score of level v is as follows:

[0035]

[0036] In the formula, α v is a fully connected layer, C is a constant, is the scalar attention weight.

[0037] The second technical solution adopted by the present invention is: a small sample signal automatic modulation recognition device, comprising: a task extraction module, used to perform C-Way K-shot task extraction in a source domain training set to obtain a support set and a query set;

[0038] A feature graph generation module is used to construct a signal category inference model, input the samples of the support set and the samples of the query set into the multi-level embedding function of the signal category inference model respectively, obtain the feature graphs of the support samples at all levels and the feature graphs of the query samples at all levels, and respectively average the feature graphs of the support samples at all levels of each category to obtain the feature graphs of each support category at all levels;

[0039] A multi-level relationship measurement module, used to calculate the similarity relationship scores between the feature graphs of each level of the query sample and the feature graphs of each level of the support category through the multi-level relationship measurement module of the signal category inference model, and perform weighted summation of the similarity relationship scores at each level to obtain the similarity score between each support category corresponding to the query sample, thereby determining the predicted label of the query sample;

[0040] A reconstruction module, used to reconstruct the feature graph of the support sample through the decoder of the signal category inference model to obtain a reconstructed support sample;

[0041] An error calculation module, used to calculate a classification error and a reconstruction error based on the predicted label and the reconstructed support sample;

[0042] A back propagation module, used for back propagating the classification error and the reconstruction error to update the signal category inference model parameters;

[0043] The model training module is used to repeatedly iteratively train the signal category inference model, and use the trained signal category inference model to perform signal recognition on the target domain data to obtain a signal recognition result.

[0044] Optionally, the feature map generation module includes:

[0045] A support feature map module, used for inputting each support sample in the support set into the multi-level embedding function to obtain feature maps of each level of the support sample;

[0046] A mean calculation module is used to calculate the mean of the feature maps of each level of the support samples of the same category to obtain the feature maps of each level of each support category;

[0047] The query feature graph module is used to send each query sample in the query set into a multi-level embedding function to obtain feature graphs of each level of the query sample.

[0048] Optionally, the multi-level relationship measurement module includes:

[0049] A first-level similarity score calculation module is used to concatenate the first-level query sample feature graph and the first-level support category feature graph, and use the first-level relationship submodule in the multi-level relationship measurement module to calculate to obtain a first-level similarity feature graph and a first-level similarity relationship score;

[0050] A similarity score calculation module at each level, used to concatenate the similarity feature graph at the previous level, the query feature graph at the same level, and the support category feature map at the same level as the input of the next level relationship measurement module, until the similarity relationship scores between the feature graphs at each level of the query sample and the feature graphs at each level of the support category are calculated;

[0051] The weighted summation module is used to perform weighted summation on the similarity relationship scores between the feature graphs of each level of the query sample and the feature graphs of each level of the support category to obtain the similarity relationship score between the query sample and each support category, and determine the predicted label of the query sample according to the highest similarity relationship score.

[0052] Optional,

[0053] The v-level similarity feature graph expression is as follows:

[0054]

[0055] The first-level similarity feature graph expression is as follows:

[0056]

[0057] In the formula, is the v-th level support feature map, is the v-th level query feature map, is the similarity feature map.

[0058] Optional,

[0059] The similarity score expression between the query sample and each supporting category is as follows:

[0060]

[0061] in,

[0062]

[0063] The similarity feature graph relationship score of level v is as follows:

[0064]

[0065] In the formula, α v is a fully connected layer, C is a constant, is the scalar attention weight.

[0066] Beneficial effects of the above technical solution:

[0067] The present invention performs C-Way in the source domain training set K-shot tasks are extracted to obtain a support set and a query set, and a signal category inference model is constructed. The samples of the support set and the samples of the query set are respectively input into the multi-level embedding function of the signal category inference model to obtain feature graphs of the support samples at all levels and feature graphs of the query samples at all levels, and the feature graphs of the support samples at all levels of each category are averaged to obtain feature graphs of each support category at all levels. The similarity relationship scores between the feature graphs of the query samples at all levels and the feature graphs of the support categories at all levels are calculated through the multi-level relationship measurement module of the signal category inference model, and the similarity relationship scores at all levels are weighted and summed to obtain the similarity scores between each support category corresponding to the query sample, and then the predicted label of the query sample is determined. The feature graph of the support sample is reconstructed through the decoder of the signal category inference model to obtain a reconstructed support sample. Based on the predicted label and the reconstructed support sample, a classification error and a reconstruction error are calculated, and the classification error and the reconstruction error are back-propagated to update the parameters of the signal category inference model, and the signal category inference model is repeatedly iteratively trained. The trained signal category inference model is used to perform signal recognition on the target domain data to obtain a signal recognition result. The construction of the MCRN-CR framework solves the small sample dilemma faced by traditional deep learning methods in the AMR problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A flowchart of a method for automatic modulation recognition of a small sample signal provided by the present invention;

[0069] Figure 2 A schematic diagram of the structure of a small sample signal automatic modulation identification device provided by the present invention;

[0070] Figure 3 The overall framework diagram of the MCRN-CR proposed in the present invention;

[0071] Figure 4 This is the encoder structure diagram;

[0072] Figure 5 It is the structure diagram of the decoder;

[0073] Figure 6 Comparison chart of recognition accuracy of different models under all signal-to-noise ratios. DETAILED DESCRIPTION

[0074] The following detailed description of the embodiments of the present invention is further described in detail in conjunction with the accompanying drawings and examples. The detailed description of the following embodiments and the accompanying drawings are used to exemplarily illustrate the principles of the present invention, but cannot be used to limit the scope of the present invention, that is, the present invention is not limited to the preferred embodiments described, and the scope of the present invention is defined by the claims.

[0075] In the description of the present invention, it should be noted that, unless otherwise specified, “plurality” means two or more than two; the terms “first”, “second”, etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance; for ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0076] Embodiment 1

[0077] An embodiment of the present invention provides a method for automatic modulation recognition of a small sample signal, such as Figure 1 and Figure 3 As shown, S1. C-Way K-shot task extraction is performed in the source domain training set to obtain a support set and a query set; S2. A signal category inference model is constructed, and samples of the support set and samples of the query set are respectively input into the multi-level embedding function of the signal category inference model to obtain feature graphs at all levels of the support samples and feature graphs at all levels of the query samples, and the feature graphs at all levels of the support samples of each category are averaged to obtain feature graphs at all levels of each support category; S3. The similarity relationship scores between the feature graphs at all levels of the query samples and the feature graphs at all levels of the support categories are calculated through the multi-level relationship measurement module of the signal category inference model, and the similarity relationship scores at all levels are weighted and summed to obtain The similarity score between each supporting category corresponding to the query sample is used to determine the predicted label of the query sample; S4, the feature map of the supporting sample is reconstructed through the decoder of the signal category inference model to obtain the reconstructed supporting sample; S5, based on the predicted label and the reconstructed supporting sample, the classification error and the reconstruction error are calculated; S6, the classification error and the reconstruction error are back-propagated to update the signal category inference model parameters; S7, repeat S1-S6 to iteratively train the signal category inference model, and use the trained signal category inference model to perform signal recognition on the target domain data to obtain the signal recognition result.

[0078] In order to solve the small sample problem in the field of signal modulation recognition, this paper proposes a multi-level comparison based class reconstruction relationship network (MCRN-CR), whose framework structure is as follows: Figure 3MCRN-CR consists of two parts: class reconstruction and multi-level comparative relation network. The class reconstruction part is used to generate low-dimensional potential representation of input samples as the embedding of support samples and query samples; the multi-level comparative relation network is an improved version of the relation network, which is used to achieve recognition tasks under small sample conditions.

[0079] The class reconstruction part includes two parts: the encoder and the decoder. The encoder generates the potential feature expression zs of the supporting sample xs, and the decoder further reconstructs the supporting sample xs. s After training, the smaller the reconstruction error, the better zs can represent the sample features. Since complex neural networks can process data containing real and imaginary parts and provide richer and more diverse feature expression capabilities than real neural networks, the present invention uses complex convolution as the basic unit of the encoder and decoder. Figure 4-Figure 5 As shown, the structure of the encoder and decoder includes five layers of complex convolution.

[0080] S1. Extract C-Way K-shot tasks in the source domain training set to obtain the support set and query set;

[0081] Task extraction is C-way K-shot task extraction.

[0082] It should be noted that C-way K-shot task extraction refers to randomly extracting C categories in the training set, extracting K samples from each category to form a support set, and then extracting Q samples as a query set. The category labels of the support set are known and are used to construct the embedding benchmark of category C in the metric space, and the query set is the sample to be predicted.

[0083] It is further noted that in a task extraction, if K>1, the embedding function needs to perform an average pooling operation on all supporting sample feature maps of the same category when calculating the support category feature map of each level, so as to ensure that only one feature map is generated for each category. The class-level feature map is combined with the feature map of each query sample, so that the number of relationship scores for a single query sample is always C in the 1-shot or k-shot setting.

[0084] S2, constructing a signal category inference model, inputting the samples of the support set and the samples of the query set into the multi-level embedding function of the signal category inference model respectively, obtaining the feature graphs of the support samples at all levels and the feature graphs of the query samples at all levels, and averaging the feature graphs of the support samples at all levels of each category to obtain the feature graphs of each support category at all levels;

[0085] It should be noted that the feature map includes a shallow feature map and a deep feature map.

[0086] S21, inputting each support sample in the support set into the multi-level embedding function to obtain feature maps of each level of the support sample;

[0087] S22, averaging the feature maps of each level of the support samples of the same category to obtain feature maps of each level of each support category;

[0088] S23, sending each query sample in the query set into a multi-level embedding function to obtain feature graphs of each level of the query sample.

[0089] S3, calculating the similarity relationship scores between the feature graphs at all levels of the query sample and the feature graphs at all levels of the support categories through the multi-level relationship measurement module of the signal category inference model, and performing weighted summation on the similarity relationship scores at all levels to obtain the similarity score between each support category corresponding to the query sample, thereby determining the predicted label of the query sample;

[0090] If you want to accurately measure and match query samples and supporting categories, you need to learn nonlinear distances at different feature levels on both deep and shallow abstract features and comprehensively calculate the relationship score.

[0091] Among them, the relationship measurement module is used to calculate the nonlinear distance between the query sample and the supporting category, and obtain the relationship score between each sample and different categories, which is used to further determine the category of the query sample. The specific structure of the relationship measurement module is shown in Table 1:

[0092] Table 1 Hierarchical structure of relationship modules

[0093]

[0094] S31, concatenating the first-level query sample feature graph and the first-level support category feature graph, and performing calculations using the first-level relationship submodule in the multi-level relationship measurement module to obtain a first-level similarity feature graph and a first-level similarity relationship score;

[0095] S32, concatenating the previous level similarity feature graph, the query feature graph at the same level, and the support category feature graph at the same level as the input of the next level relationship measurement module, until the similarity relationship scores between the query sample feature graphs at all levels and the support category feature graphs at all levels are calculated;

[0096] S33, performing weighted summation on the similarity relationship scores between the feature graphs of each level of the query sample and the feature graphs of each level of the support category to obtain a similarity relationship score between the query sample and each support category, and determining a predicted label for the query sample according to the highest similarity relationship score.

[0097] Specifically, the support samples and query samples are respectively embedded into the function to generate the v-th level feature map and Then concatenate it into Then it is sent to the corresponding v-th level relation module for comparison.

[0098] At level v-1, the relation module outputs sample x i and x j Similarity feature map Then the v-th level relation module takes both the embedding output of the v-th level support and query sample as input, and the similarity feature map of the v-1-th level relation module as input.

[0099] The v-level similarity feature graph expression is as follows:

[0100]

[0101] For the first-level relationship module, since it has no previous input, the first-level similarity feature graph expression is set as follows:

[0102]

[0103] In the formula, is the feature map of the v-th level support sample, is the feature map of the query sample at level v, is the similarity feature map output by the v-1th level relation module. θ and They are multi-level embedding function and multi-level relationship measurement module respectively.

[0104] Assuming that q(.) represents average pooling and full connection operations, the similarity (relationship) score of the feature graph of the support sample and the query sample output by each relationship module at level v, that is, the similarity feature graph relationship score at level v, is as follows:

[0105]

[0106] Further query sample x j And each support category y c The similarity score expression is as follows:

[0107]

[0108] in,

[0109]

[0110] In the formula, α v is a fully connected layer, C is a constant, is the scalar attention weight.

[0111] S4, reconstructing the feature graph of the supporting sample through the decoder of the signal category inference model to obtain a reconstructed supporting sample;

[0112] S5. Based on the predicted label and the reconstructed support sample, calculate and obtain a classification error and a reconstruction error;

[0113] During the training process, the class reconstruction part first generates support samples x s The low-dimensional latent representation z s On the one hand, s The similarity calculation between the participating and query samples, on the other hand, needs to be further s Feed into the decoder to generate reconstructed support samples In order to eliminate the influence of noise, the reconstruction error L is further calculated. re When , the high signal-to-noise ratio samples of the same category as the support samples are As a reconstruction goal,

[0114]

[0115] Then, the classifier in the MCRN part predicts the category label of the query sample and calculates the cross entropy classification error with the true label:

[0116]

[0117] In the formula, y is the query sample, is the predicted label of the query sample.

[0118] The class reconstruction part helps the encoder as the embedding function to extract more recognizable sample features and increase the difference between embeddings of different categories, thus helping to improve the signal recognition performance under small sample conditions. re and L ce Together they constitute the model error, and their coefficients are λ re and λ ce .

[0119] L=λ ce L ce +λ re L re

[0120] S6, back-propagating the classification error and the reconstruction error to update the signal category inference model parameters;

[0121] S7. Repeat S1-S6 to iteratively train the signal category inference model, and use the trained signal category inference model to perform signal recognition on the target domain data to obtain a signal recognition result.

[0122] During training, each iteration will extract C-way K-shot tasks from the training set, and obtain the support set of each training task. and queryset Wherein m=K*C, n=Q*C.

[0123] Then, according to S3, the relationship score r between each query sample xj and each support category yc is calculated c, j , and further determine the predicted label of the query sample Secondly, the hidden layer feature z of the supporting sample calculated by the encoder is s It is further fed into the decoder to reconstruct the support samples Then, the classification error and reconstruction error are calculated according to the predicted label and the reconstructed sample, and finally the error is back-propagated to update the encoder and decoder parameters of the class reconstruction part and the parameters of the relationship module.

[0124]

[0125] This embodiment is a method for automatic modulation recognition of small sample signals, which obtains a support set and a query set by obtaining a small sample signal set and performing C-way K-shot task extraction from the small sample signal set; the encoder calculates each support category in the support set and each query sample in the query set to obtain a support feature map and a query feature map; the relationship measurement module calculates the support feature map and the query feature map to obtain a similarity feature map and a relationship score to determine the predicted label of the query sample; the decoder calculates the feature map to obtain a reconstructed support sample; based on the predicted label and the reconstructed support sample, the classification error and the reconstruction error are calculated; the classification error and the reconstruction error are back-propagated to update the parameters of the encoder, the decoder and the relationship measurement module. Constructing the MCRN-CR framework solves the small sample dilemma faced by traditional deep learning methods in the AMR problem.

[0126] Embodiment 2

[0127] An embodiment of the present invention provides a small sample signal automatic modulation identification device 200, such as Figure 2 As shown, it includes: a task extraction module 201, which is used to extract C-Way K-shot tasks in the source domain training set to obtain a support set and a query set;

[0128] The feature graph generation module 202 constructs a signal category inference model, inputs the samples of the support set and the samples of the query set into the multi-level embedding function of the signal category inference model respectively, obtains the feature graphs of the support samples at all levels and the feature graphs of the query samples at all levels, and respectively averages the feature graphs of the support samples at all levels of each category to obtain the feature graphs of each support category at all levels;

[0129] The multi-level relationship measurement module 203 is used to calculate the similarity relationship scores between the feature graphs of the query sample at each level and the feature graphs of the support categories at each level through the multi-level relationship measurement module of the signal category inference model, and perform weighted summation of the similarity relationship scores at each level to obtain the similarity score between each support category corresponding to the query sample, thereby determining the predicted label of the query sample;

[0130] A reconstruction module 204, configured to reconstruct the feature graph of the support sample through a decoder of the signal category inference model to obtain a reconstructed support sample;

[0131] An error calculation module 205 is used to calculate a classification error and a reconstruction error based on the predicted label and the reconstructed support sample;

[0132] A back propagation module 206, used for back propagating the classification error and the reconstruction error to update the signal category inference model parameters;

[0133] The model training module 207 is used to repeatedly iteratively train the signal category inference model, and use the trained signal category inference model to perform signal recognition on the target domain data to obtain a signal recognition result.

[0134] Furthermore, the feature map generation module 202 includes:

[0135] A support feature map module 2021 is used to input each support sample in the support set into the multi-level embedding function to obtain feature maps of each level of the support sample;

[0136] A mean calculation module 2022 is used to calculate the mean of the feature maps of each level of the support samples of the same category to obtain the feature maps of each level of each support category;

[0137] The query feature graph module 2023 is used to send each query sample in the query set into a multi-level embedding function to obtain feature graphs of each level of the query sample.

[0138] Furthermore, the multi-level relationship measurement module 203 includes:

[0139] The first-level similarity score calculation module 2031 is used to concatenate the first-level query sample feature graph and the first-level support category feature graph, and use the first-level relationship submodule in the multi-level relationship measurement module to calculate to obtain a first-level similarity feature graph and a first-level similarity relationship score;

[0140] The similarity score calculation module 2032 at each level is used to concatenate the similarity feature graph at the previous level, the query feature graph at the same level, and the support category feature map at the same level as the input of the next level relationship measurement module until the similarity relationship scores between the feature graphs at each level of the query sample and the feature graphs at each level of the support category are calculated;

[0141] The weighted summation module 2033 is used to perform weighted summation on the similarity relationship scores between the feature graphs of each level of the query sample and the feature graphs of each level of the support category to obtain the similarity relationship score between the query sample and each support category, and determine the predicted label of the query sample according to the highest similarity relationship score.

[0142] Furthermore, it also includes:

[0143] The v-th level similarity feature graph expression is as follows:

[0144]

[0145] The first-level similarity feature graph expression is as follows:

[0146]

[0147] In the formula, is the feature map of the v-th level support sample, is the feature map of the query sample at level v, is the similarity feature map output by the v-1th level relation module. θ and They are multi-level embedding function and multi-level relationship measurement module respectively.

[0148] Furthermore, it also includes:

[0149] The similarity score expression between the query sample and each supporting category is as follows:

[0150]

[0151] in,

[0152]

[0153] The similarity feature graph relationship score of level v is as follows:

[0154]

[0155] In the formula, α v is a fully connected layer, C is a constant, is the scalar attention weight.

[0156] The present embodiment is a small sample signal automatic modulation recognition device, a task extraction module, which is used to perform C-Way K-shot task extraction in the source domain training set to obtain a support set and a query set; a feature graph generation module, which constructs a signal category inference model, and inputs the samples of the support set and the samples of the query set into the multi-level embedding function of the signal category inference model respectively to obtain the feature graphs of each level of the support sample and the feature graphs of each level of the query sample, and respectively averages the feature graphs of each level of the support sample of each category to obtain the feature graphs of each level of each support category; a multi-level relationship measurement module, which is used to calculate the similarity relationship scores between the feature graphs of each level of the query sample and the feature graphs of each level of the support category through the multi-level relationship measurement module of the signal category inference model, and weighted sum the similarity relationship scores at each level to obtain the query. The similarity score between each supporting category corresponding to the sample is used to determine the predicted label of the query sample; the reconstruction module is used to reconstruct the feature map of the supporting sample through the decoder of the signal category inference model to obtain the reconstructed supporting sample; the error calculation module is used to calculate the classification error and reconstruction error based on the predicted label and the reconstructed supporting sample, the back propagation module is used to back propagate the classification error and reconstruction error, and update the signal category inference model parameters; the model training module is used to repeatedly iteratively train the signal category inference model, and use the trained signal category inference model to perform signal recognition on the target domain data to obtain the signal recognition result. The construction of the MCRN-CR framework solves the small sample dilemma faced by traditional deep learning methods in the AMR problem.

[0157] Embodiment 3

[0158] The dataset and experimental settings are as follows:

[0159] To verify and evaluate the effectiveness of the proposed method, this paper conducts experiments on the RML2018.01A modulated signal dataset. This dataset consists of 24 modulated signals. Each signal sample contains I\Q two-way data in the format of [1024,2]. The signal-to-noise ratio range is -20dB to 30dB, and the interval is 2dB. A single category has 4096 samples at each signal-to-noise ratio, totaling 2555904 data.

[0160] First, the data set is divided, with 14 modulation categories as known source domain training sets and the other 10 as target domain data sets. Since the original data set is large in size and redundant for the experiments in this article, only 1,000 samples are taken for each category at a single signal-to-noise ratio. In a small sample scenario, the existing training samples in the target domain are usually scarce. Therefore, in order to meet this setting, a subset of the target domain data set is divided as a known small sample data set, and the number is 2% of the number of target domain samples, that is, 20 samples are taken for a single modulation method at each signal-to-noise ratio. During each test, the support set is extracted from this subset. The remaining part is the test set, and query samples are extracted from it during each test. The specific division is shown in Table 2.

[0161] The hardware environment of the experiment is: Intel(R)Core(TM)i7-10700k CPU@3.8GHz, NVIDIAGeForceGTX3090 GPU; the software environment is python3.8, pytorch deep learning framework.

[0162] Table 2 Dataset division

[0163]

[0164] The comparison of recognition performance with other networks is as follows:

[0165] In order to verify the effectiveness of the proposed method under small sample conditions and the superiority of recognition performance, we compared the recognition performance of MCRN-CR with other networks under the 5-way 5-shot setting. The compared models include the recent small sample signal modulation recognition methods, AMCRN, AMR-CapsNet, and the small sample learning methods in the image field, prototype network (PN) and relation network (RN). At the same time, in order to prove that the proposed method can solve the small sample dilemma faced by traditional deep learning methods, a classifier is added to the structure of the encoder to form an automatic modulation recognition model based on a complex neural network, AMR-CVNN, which is used as a comparison item. In order to ensure the fairness of the experiment, AMR-CapsNet and AMCRN maintain the same settings as the original text; the laboratory settings of PN and RN are consistent with MCRN-CR. AMR-CVNN adopts the idea of ​​transfer learning to first pre-train on the source domain dataset, and then fine-tune the weight parameters on the target domain dataset.

[0166] Figure 6The recognition accuracy of all the comparison models under all signal-to-noise ratios on a given dataset is shown. It can be seen from the figure that when the signal-to-noise ratio is greater than 0dB, the proposed method MCRN-CR has the best recognition performance. When the signal-to-noise ratio is lower than 0dB, the recognition performance of all metric-based meta-learning methods is not much different, but the recognition performance of MCRN-CR is much higher than that of AMR-CapsNet and AMR-CVNN.

[0167] Table 3 shows more specific recognition performance statistics of each comparison model. The highest recognition accuracy of MCRN-CR reached 89.25%. Under all signal-to-noise ratios, the average recognition accuracy of MCRN-CR reached 65.98%, and the average recognition accuracy reached 87.41% when SNR>0dB. Compared with AMR-CVNN, the average recognition accuracy of MCRN-CR below 0dB increased by nearly 10%, the average recognition accuracy above 0dB increased by 5.86%, and the average recognition accuracy at all signal-to-noise ratios increased by 8.04%. Therefore, it can be proved that the MCRN-CR proposed in this paper can effectively solve the small sample dilemma faced by traditional deep learning methods in AMR problems.

[0168] Table 3 Comparison of average recognition accuracy of different methods

[0169] Model -20:2:-2dB 0:2:30dB -20:2:30dB Highest accuracy MCRN-CR 31.69％ 87.41％ 65.98％ 89.25％ AMR-CapsNet 19.22％ 44.42％ 34.73％ 46.40％ AMCRN 32.13％ 83.09％ 63.49％ 86.25％ PN 30.93％ 83.00％ 62.98％ 85.41％ RN 30.72％ 84.08％ 63.56％ 86.68％ AMR-CVNN 20.16％ 81.55％ 57.94％ 85.04％

[0170] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for automatic modulation identification of small sample signals, characterized in that: include: S1. Extract C-Way K-shot tasks in the source domain training set to obtain the support set and query set; S2, constructing a signal category inference model, inputting the samples of the support set and the samples of the query set into the multi-level embedding function of the signal category inference model respectively, obtaining the feature graphs of the support samples at all levels and the feature graphs of the query samples at all levels, and averaging the feature graphs of the support samples at all levels of each category to obtain the feature graphs of each support category at all levels; S3, calculating the similarity relationship scores between the feature graphs at all levels of the query sample and the feature graphs at all levels of the support categories through the multi-level relationship measurement module of the signal category inference model, and performing weighted summation on the similarity relationship scores at all levels to obtain the similarity score between each support category corresponding to the query sample, thereby determining the predicted label of the query sample; S4, reconstructing the feature graph of the supporting sample through the decoder of the signal category inference model to obtain a reconstructed supporting sample; S5. Based on the predicted label and the reconstructed support sample, calculate and obtain a classification error and a reconstruction error; S6, back-propagating the classification error and the reconstruction error to update the signal category inference model parameters; S7. Repeat S1-S6 to iteratively train the signal category inference model, and use the trained signal category inference model to perform signal recognition on the target domain data to obtain a signal recognition result.

2. The method for automatic modulation identification of small sample signals according to claim 1, characterized in that: The S2 includes: S21, inputting each support sample in the support set into the multi-level embedding function to obtain feature maps of each level of the support sample; S22, averaging the feature maps of each level of the support samples of the same category to obtain feature maps of each level of each support category; S23, sending each query sample in the query set into a multi-level embedding function to obtain feature graphs of each level of the query sample.

3. The method for automatic modulation identification of small sample signals according to claim 2, characterized in that: The S3 includes: S31, concatenating the first-level query sample feature graph and the first-level support category feature graph, and performing calculations using the first-level relationship submodule in the multi-level relationship measurement module to obtain a first-level similarity feature graph and a first-level similarity relationship score; S32, concatenating the previous level similarity feature graph, the query feature graph at the same level, and the support category feature graph at the same level as the input of the next level relationship measurement module, until the similarity relationship scores between the query sample feature graphs at all levels and the support category feature graphs at all levels are calculated; S33, performing weighted summation on the similarity relationship scores between the feature graphs of each level of the query sample and the feature graphs of each level of the support category to obtain a similarity relationship score between the query sample and each support category, and determining a predicted label for the query sample according to the highest similarity relationship score.

4. The method for automatic modulation identification of small sample signals according to claim 3, characterized in that: The v-th level similarity feature graph expression is as follows: The first-level similarity feature graph expression is as follows: In the formula, is the feature map of the v-th level support sample, is the feature map of the query sample at level v, is the similarity feature map output by the v-1th level relation module. θ and They are multi-level embedding function and multi-level relationship measurement module respectively.

5. The method for automatic modulation identification of small sample signals according to claim 3, characterized in that: The similarity score expression between the query sample and each supporting category is as follows: in, The similarity feature graph relationship score of level v is as follows: In the formula, α v is a fully connected layer, C represents the number of support categories, is the scalar attention weight.

6. A small sample signal automatic modulation identification device, characterized in that: include: The task extraction module is used to extract C-Way K-shot tasks from the source domain training set to obtain the support set and query set; A feature graph generation module is used to construct a signal category inference model, input the samples of the support set and the samples of the query set into the multi-level embedding function of the signal category inference model respectively, obtain the feature graphs of the support samples at all levels and the feature graphs of the query samples at all levels, and respectively average the feature graphs of the support samples at all levels of each category to obtain the feature graphs of each support category at all levels; A multi-level relationship measurement module, used to calculate the similarity relationship scores between the feature graphs of each level of the query sample and the feature graphs of each level of the support category through the multi-level relationship measurement module of the signal category inference model, and perform weighted summation of the similarity relationship scores at each level to obtain the similarity score between each support category corresponding to the query sample, thereby determining the predicted label of the query sample; A reconstruction module, used to reconstruct the feature graph of the support sample through the decoder of the signal category inference model to obtain a reconstructed support sample; An error calculation module, used to calculate a classification error and a reconstruction error based on the predicted label and the reconstructed support sample; A back propagation module, used for back propagating the classification error and the reconstruction error to update the signal category inference model parameters; The model training module is used to repeatedly iteratively train the signal category inference model, and use the trained signal category inference model to perform signal recognition on the target domain data to obtain a signal recognition result.

7. The small sample signal automatic modulation identification device according to claim 6, characterized in that: The feature map generation module comprises: A support feature map module, used for inputting each support sample in the support set into the multi-level embedding function to obtain feature maps of each level of the support sample; A mean calculation module is used to calculate the mean of the feature maps of each level of the support samples of the same category to obtain the feature maps of each level of each support category; The query feature graph module is used to send each query sample in the query set into a multi-level embedding function to obtain feature graphs of each level of the query sample.

8. The small sample signal automatic modulation identification device according to claim 6, characterized in that: The multi-level relationship measurement module includes: A first-level similarity score calculation module is used to concatenate the first-level query sample feature graph and the first-level support category feature graph, and use the first-level relationship submodule in the multi-level relationship measurement module to calculate to obtain a first-level similarity feature graph and a first-level similarity relationship score; A similarity score calculation module at each level, used to concatenate the similarity feature graph at the previous level, the query feature graph at the same level, and the support category feature map at the same level as the input of the next level relationship measurement module, until the similarity relationship scores between the feature graphs at each level of the query sample and the feature graphs at each level of the support category are calculated; The weighted summation module is used to perform weighted summation on the similarity relationship scores between the feature graphs of each level of the query sample and the feature graphs of each level of the support category to obtain the similarity relationship score between the query sample and each support category, and determine the predicted label of the query sample according to the highest similarity relationship score.

9. The automatic modulation identification device for small sample signals according to claim 8, characterized in that: The v-th level similarity feature graph expression is as follows: The first-level similarity feature graph expression is as follows: In the formula, is the feature map of the v-th level support sample, is the feature map of the query sample at level v, is the similarity feature map output by the v-1th level relation module. θ and They are multi-level embedding function and multi-level relationship measurement module respectively.

10. The automatic modulation identification device for small sample signals according to claim 8, characterized in that: The similarity score expression between the query sample and each supporting category is as follows: in, The similarity feature graph relationship score of level v is as follows: In the formula, α v is a fully connected layer, C is a constant, is the scalar attention weight.

Images