Modulation recognition method for small sample signals based on multi-level relational metric network

Abstract

The present disclosure relates to a method for identifying modulation of small sample signals based on a multi-level relational metric network. The method comprises the following steps: establishing a network model of a multi-level relational metric network, wherein the network model comprises a multi-level feature embedding module and a multi-level relational metric module; inputting the small sample signal to be identified into the network model for reconstructed feature training and multi-level metric training, and finally obtaining the similarity score and corresponding weighting coefficient of each level of the small sample signal; and determining the final similarity score of all small sample signals by weighted summation. The method proposed in the present disclosure can improve the network's utilization rate of the feature vectors of each level of the sample, and at the same time, achieve signal data enhancement by reconstructing the sample feature vectors, thereby improving the generalization of network training and avoiding the occurrence of overfitting.

Description

Technical Field

[0001] The present disclosure relates to the technical field of small sample signal modulation recognition, and in particular to a small sample signal modulation recognition method based on a multi-level relational metric network. Background Art

[0002] Communication signal recognition holds significant research value in both military and civilian fields. However, facing the increasingly complex electromagnetic spectrum, traditional modulation recognition methods based on feature extraction and likelihood ratio hypothesis theory are unable to fully utilize sample characteristics, making them less applicable. With the recent development of deep learning technology, modulation recognition techniques based on deep learning have achieved excellent recognition results. However, this success often relies on thousands of training samples, and this insufficient sample size leads to serious fitting problems. In real-world communication environments, such a large number of training samples is often unavailable, hindering the widespread application of deep learning methods. Therefore, there is a need to address one or more of the issues identified in the aforementioned related technical solutions.

[0003] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the Invention

[0004] The purpose of the embodiment of the present disclosure is to provide a method for identifying modulation of a small sample communication signal based on a multi-level relational metric network, comprising the following steps:

[0005] Establishing a network model of a multi-level relationship measurement network, wherein the network model includes a multi-level feature embedding module and a multi-level relationship measurement module;

[0006] Inputting the small sample signal to be identified into the feature embedding module of each level respectively, performing reconstruction feature training, and obtaining the corresponding reconstructed feature vector of each level respectively;

[0007] Inputting the reconstructed feature vector corresponding to each level into the relationship measurement module of the corresponding level, performing multi-level measurement training, and obtaining the similarity score and corresponding weighting coefficient of the small sample signal at each level;

[0008] The final similarity scores of all the small sample signals are determined by weighted summation.

[0009] In an exemplary embodiment of the present disclosure, in the step of establishing a network model of a multi-level relationship measurement network, the feature embedding module is used It indicates that each level of the feature embedding module includes a plurality of residual blocks with convolution kernels connected in sequence, and a long short-term neural memory network connected to the last residual block.

[0010] In an exemplary embodiment of the present disclosure, in the step of establishing a network model of a multi-level relationship measurement network, the relationship measurement module is used It indicates that each level of the relationship measurement module includes two residual blocks with convolution kernels connected in sequence, and a fully connected network connected to the last residual block; the fully connected network includes a ReLU activation function.

[0011] In an exemplary embodiment of the present disclosure, in the step of inputting the small sample signal to be identified into the feature embedding module of each level, performing feature reconstruction training, and obtaining the reconstructed feature vector corresponding to each level,

[0012] The small sample signal includes a support set D S and query set D Q , the support set D S and the query set D Q The basic unit T constituting the small sample signal, T={D S , D Q}; the support set D S The query set D is composed of label signal samples. Q It is composed of the signal samples to be tested, and the support set D S and the query set D Q The label space of the signal samples in the same; if the support set D S It contains C types of signals, and each type of signal contains K signal samples, so the small sample signal is a C-way K-shot recognition signal.

[0013] In an exemplary embodiment of the present disclosure, in the step of inputting the small sample signal to be identified into the feature embedding module of each level, performing feature reconstruction training, and obtaining the reconstructed feature vector corresponding to each level,

[0014] For the C-way K-shot recognition signal, the support set D S and the query set D Q The signal samples in each level are input into the feature embedding module In each level, the feature embedding module The sample distribution characteristic statistics are output respectively, and the reconstructed characteristic vector corresponding to each level is generated respectively through a reparameterization method.

[0015] In an exemplary embodiment of the present disclosure, the sample distribution characteristic statistics include the eigenvector mean and the eigenvector variance The formula for reconstructing the feature vector includes:

[0016]

[0017] Where x represents the input signal sample; Represents the reconstructed feature vector; ε represents the parameter randomly sampled from the Gaussian distribution; ⊙ represents the multiplication of the corresponding elements; v represents the vth level; θ represents the network parameter of the feature embedding module. In an exemplary embodiment of the present disclosure, the reconstructed feature vector corresponding to each level is input into the relationship measurement module of the corresponding level, multi-level measurement training is performed, and the similarity score and corresponding weighting coefficient of the small sample signal of each level are obtained respectively.

[0018] When the support set D S When the characteristic vector of the signal in the cth class at the υth level represents the mean of the characteristic vectors of all signal samples in this class, we get:

[0019]

[0020] Among them, x c represents the c-th type signal; x ci represents the i-th signal sample of the c-th type signal;

[0021] Then the output formula of the relationship measurement module includes:

[0022]

[0023] in, express and the signal sample x to be measured j The υ-th level eigenvector Similarity feature map output by the υ-1 level relationship measurement module The characteristic connection of φ represents the network parameter of the relationship measurement module; if υ=1, then

[0024] In an exemplary embodiment of the present disclosure, in the reconstruction feature training, a multi-category classification neural network is used to train a given training set D m-tr The features are embedded into the module It is trained as a multi-class signal classifier, and the multivariate cross entropy loss function l is used during the training process. CE , according to the regularization strategy, in the multivariate cross entropy loss function l CE A penalty term is introduced in the feature embedding module, that is, the network parameter θ satisfies,

[0025]

[0026] Among them, δ iRepresents the mean of the variance output by each level of feature embedding module; m represents the total number of signal samples to be tested; λ represents the hyperparameter for adjusting the regularization penalty term; ← represents updating θ according to the multivariate cross entropy loss function.

[0027] In an exemplary embodiment of the present disclosure, in the multi-level metric training, the relationship metric module The training strategy of Episode is adopted, the trained network parameters θ of the feature embedding module are fixed, and only the network parameters φ of the relationship measurement module are updated in the multi-level relationship measurement training. m-tr Randomly sample a C-way K-shot recognition signal and use the multivariate cross entropy loss function l CE Calculate the training loss, which is:

[0028]

[0029] Where N represents the total number of samples of each type of signal in the query set, y represents the label of the signal sample; r c，j represents the similarity score of small sample signals; y c Indicates the label of the c-th type signal; y j Indicates the label of the j-th type of signal; 1(y x =y j ) means when y c =y j The value is 1.

[0030] In an exemplary embodiment of the present disclosure, in the step of determining the final similarity scores of all the small sample signals by weighted summation,

[0031] The calculation formula of the final similarity score includes:

[0032]

[0033] in,

[0034] Represents the similarity score of the small sample signal; the similarity score of the small sample signal The calculation formula includes:

[0035]

[0036] Represents the weighting coefficient: the weighting coefficient The calculation formula includes:

[0037]

[0038] f1v represents a fully connected network for calculating similarity scores; Represents a fully connected network that calculates weighted coefficients.

[0039] The technical solution provided by the present disclosure may have the following beneficial effects:

[0040] In the disclosed embodiments, a small sample signal modulation recognition method based on a multi-level relational measurement network is proposed. The method reconstructs and trains each level of features through a feature embedding module and reconstructs the feature vector using a reparameterization technique, thereby avoiding the problem of overfitting that is prone to occur when using deep neural networks. The similarity of each level of features of the small signal sample is measured through a relational measurement module, and the final similarity score between the small sample signals is determined by weighted summation. This multi-level measurement method fully considers the differences between the features of signal samples at all levels, effectively improving the recognition ability of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the specification, serve to explain the principles of the present disclosure. Obviously, the drawings described below are only some embodiments of the present disclosure, and those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0042] Figure 1 A schematic diagram showing the steps of a small sample signal modulation identification method based on a multi-level relational metric network in an exemplary embodiment of the present disclosure is shown;

[0043] Figure 2 A schematic diagram showing the structure of a network model of a multi-level relationship measurement network in an exemplary embodiment of the present disclosure;

[0044] Figure 3a and Figure 3b A schematic structural diagram of a v-th level feature embedding module in an exemplary embodiment of the present disclosure is shown;

[0045] Figure 4 A schematic diagram showing the structure of a v-th level relationship measurement module in an exemplary embodiment of the present disclosure is shown;

[0046] Figure 5 A schematic diagram illustrating a feature vector generation process in an exemplary embodiment of the present disclosure;

[0047] Figure 6 An algorithm framework diagram of a multi-level relationship measurement network in an exemplary embodiment of the present disclosure is shown;

[0048] Figure 7A graph showing how the recognition accuracy of a 5-way K-shot recognition signal changes with the signal-to-noise ratio in a simulation experiment of an exemplary embodiment of the present disclosure;

[0049] Figure 8a and Figure 8b A graph showing how the accuracy of the algorithm in identifying a signal changes with the signal-to-noise ratio at different metric levels in a simulation experiment of an exemplary embodiment of the present disclosure;

[0050] Figure 9a and Figure 9b A comparison diagram showing recognition performance under different algorithm conditions in a simulation experiment of an exemplary embodiment of the present disclosure;

[0051] Figure 10a and Figure 10b A graph showing the performance comparison between different meta-learning methods and the algorithm of the present disclosure in a simulation experiment of an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0052] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be embodied in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0053] In addition, the accompanying drawings are merely schematic illustrations of the present disclosure and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0054] This example embodiment provides a method, such as Figure 1 As shown, the following steps may be included:

[0055] Step S101: establishing a network model of a multi-level relationship measurement network, wherein the network model includes a multi-level feature embedding module and a multi-level relationship measurement module;

[0056] Step S102: inputting the small sample signal to be identified into the feature embedding module at each level, performing reconstruction feature training, and obtaining the corresponding reconstructed feature vectors at each level;

[0057] Step S103: inputting the reconstructed feature vector corresponding to each level into the relationship measurement module of the corresponding level, performing multi-level measurement training, and obtaining the similarity score and corresponding weighting coefficient of the small sample signal at each level;

[0058] Step S104: determining the final similarity scores of all the small sample signals by weighted summation.

[0059] This method reconstructs and trains each level of features through a feature embedding module and reconstructs feature vectors using reparameterization technology, which avoids the problem of overfitting that is prone to occur when using deep neural networks; it measures the similarity of each level of features of small signal samples through a relational measurement module, and determines the final similarity score between small sample signals through weighted summation; this multi-level measurement method fully considers the differences between the features of signal samples at all levels, and effectively improves the recognition ability of the algorithm.

[0060] Below, each step of the above method in this exemplary embodiment will be described in more detail.

[0061] In step S101, Figure 2 As shown, it can be seen that the principle of the method proposed in this embodiment is as follows:

[0062] The biggest difference between relational networks and traditional metric learning models is that relational networks can measure the similarity between samples through neural networks. The relational metric network mainly consists of two modules: feature embedding module and relationship measurement modules For small sample recognition signals, first, the feature embedding module transforms the support set D in the small sample signal into S and query set D Q The signal samples in are mapped to the feature metric space f θ (x), then the eigenvector of the signal sample to be tested is compared with the support set D S The characteristic vectors of various signal samples are spliced ​​and then passed through the relationship measurement module Calculate the signal to be tested and the support set D S The similarity scores between different types of signals. Figure 2 The support set D in S Contains 5 types of signal samples x i , Z(f θ (x i ), f θ (x)), i = 1, 2, 3, 4, 5 represents the sample feature f of the signal to be tested θ (x) and the features f of the 5 known signal samples in the support set θ (x i ) The concatenated feature vector; the signal sample to be tested and the support set D SThe similarity score between known signals can be expressed as: i =g φ (Z(f θ (x i ), f θ (x))), x i ∈D s .

[0063] In terms of training strategy, the relational network adopts an Episode-based training strategy based on the meta-learning idea. Both the training and testing processes are based on the small sample recognition task T = {D S , D Q} is the basic unit, where D S 、D Q Denote the support set and query set respectively, D S It consists of a small number of labeled signal samples, D Q It is composed of the signal samples to be measured, D S and D Q The inner signal sample label space is the same, D m-tr If D S It contains C types of signals, and each type has K signal samples, then the small sample recognition task is called C-way K-shot recognition task.

[0064] Given a training set D during the algorithm implementation m-tr With the test set D m-te , where D m-tr Contains a large number of labeled signal samples, D m-tr With D m-te The sample label spaces of D are disjoint. m-tr The signal samples in the D m-tr In the process of randomly sampling multiple small sample recognition tasks, the tasks are input into the relational network in sequence, and the network model is trained to learn a highly efficient metric function, so that when the network faces a signal recognition task that has never been seen in the training process during the test phase, it can also accurately recognize the signal to be tested. Specifically, for the C-way K-shot recognition task, during each training Episode iteration, the network model first randomly selects C types of signals in the training set, randomly extracts K samples from each type of signal to form a support set, and then randomly extracts samples from the remaining samples of each type to form a query set. The training loss on the Episode is calculated based on the recognition results, and the network parameters are updated through the reverse gradient propagation algorithm, that is,

[0065]

[0066] Where θ represents the network parameters of the feature embedding module; φ represents the network parameters of the relationship measurement module; Loss(·) represents the loss function; y represents the label of the signal sample x in the query set, Represents the predicted label.

[0067] The above is the principle of this embodiment. Specifically, in this embodiment, To represent the feature embedding module; To represent the relationship measurement module.

[0068] As shown in Figure 3, convolutional neural networks have been widely used for signal feature extraction, but it is often difficult to use a deeper structure, resulting in insufficient signal features. The skip connection structure of the residual network can effectively avoid the gradient vanishing problem caused by too deep a network layer. Here, a feature embedding module is constructed based on the principle of the residual network. Each level of feature embedding module includes multiple residual blocks with convolution kernels connected in sequence. At the same time, considering the timing characteristics of the in-phase orthogonal sequence of the modulated signal, the post-feature embedding module at each level A long short-term memory network LSTM is cascaded after the last residual block, so that the temporal characteristics of the signal samples can be learned at the same time.

[0069] exist Figure 3a The specific structure of the residual block is given in

[15] : it contains three convolutional layers, the input and output of the convolutional block use residual connections, and the final output is obtained through a maximum pooling layer; Figure 3b Shown is an LSTM cascaded after the last residual block, which also includes a Flatten layer as a transition.

[0070] like Figure 4 As shown, each level of feature embedding module Each corresponds to a relationship measurement module Each level of relationship measurement module consists of two residual blocks with convolution kernels connected in sequence, and a fully connected network connected to the last residual block. The fully connected network contains a ReLU activation function, which can be used to obtain a feature similarity map as the input of the next level of relationship measurement module. Each level of feature similarity map is input into the two fully connected layers f1. v and The Sigmoid activation function is used to calculate the similarity score and weighting coefficient.

[0071] In step S102, the small sample signal is represented by the support set D S and query set D Q Composition, support set D S and query set D Q The basic unit T that constitutes the small sample signal, T={DS , D Q}; Support set D S The query set D consists of label signal samples. Q It is composed of the signal samples to be tested, and the support set D S and the query set D Q The label space of the signal samples in is the same. If the support set D S It contains C types of signals, and each type of signal contains K signal samples, so the small sample signal is a C-way K-shot recognition signal.

[0072] like Figure 5 As shown, for the C-way K-shot recognition signal, the support set D S and query set D Q The signal samples in each level are input into the feature embedding module Each level of feature embedding module The sample distribution characteristic statistics are output respectively, and the reconstructed characteristic vector is generated by a reparameterization method.

[0073] Here, the sample distribution characteristic statistics include the eigenvector mean and the eigenvector variance The calculation formula for reconstructing the eigenvector is:

[0074]

[0075] Where x represents the input signal sample; Represents the reconstructed feature vector; ε represents the parameter randomly sampled from the Gaussian distribution; ⊙ represents the multiplication of corresponding elements; υ represents the υth level; θ represents the network parameters of the feature embedding module.

[0076] Since the feature vectors obtained by each level of feature embedding module are random, the effect of data enhancement can be achieved, thereby effectively avoiding the overfitting problem during network training.

[0077] In step S103, when the support set D S When the characteristic vector of the c-th type signal at the υ level represents the mean of the characteristic vectors of all signal samples of this type,

[0078]

[0079] Among them, x c represents the c-th type signal; x ci represents the i-th signal sample of the c-th type signal, then the output formula of the relationship measurement module includes:

[0080]

[0081] in, express and the signal sample x to be measured j The υ-th level eigenvector Similarity feature map output by the υ-1 level relationship measurement module The characteristic connection of φ represents the network parameter of the relationship measurement module; if υ=1, then

[0082] It should be noted here that if Figure 6 As shown,

[0083] In the reconstruction feature training, a multi-category classification neural network is used to m-tr Feature embedding module It is trained as a multi-class signal classifier, and the multivariate cross entropy loss function l is used during the training process. CE , according to the regularization strategy, in the multivariate cross entropy loss function l CE A penalty term is introduced in the embedding module network parameter θ,

[0084]

[0085] Among them, δ i Represents the mean of the variance output by each level of feature embedding module; m represents the total number of signal samples to be tested; λ represents the hyperparameter for adjusting the regularization penalty term; ← represents updating θ according to the multivariate cross entropy loss function.

[0086] In the multi-level metric training, the relational metric module The training strategy of Episode is adopted, and the network parameters θ of the trained feature embedding module are fixed. In the multi-level relationship metric training, only the network parameters φ of the relationship metric module are updated. m-tr Randomly sample a C-way K-shot recognition signal and use the multivariate cross entropy loss function l CE Calculate the training loss, which is:

[0087]

[0088] Where N represents the total number of samples of each type of signal in the query set, y represents the label of the signal sample; r c，j represents the similarity score of small sample signals; y c Indicates the label of the c-th type signal; y j Indicates the label of the j-th type of signal; 1(y c =y j ) means when y c =y jThe value of φ is 1. Here, the network parameter φ of the relationship measurement module is updated through multiple Episode training.

[0089] In step S104, the calculation formula of the final similarity score is as follows:

[0090]

[0091] in,

[0092] Represents the similarity score of small sample signals; Similarity score of small sample signals The calculation formula includes:

[0093]

[0094] Represents a weighting coefficient; the weighting coefficient The calculation formula includes:

[0095]

[0096] f1 v represents a fully connected network for calculating similarity scores; Represents a fully connected network that calculates weighted coefficients.

[0097] Formula (7) is the similarity score of the small sample signal obtained in step S103; Formula (8) is the corresponding weighting coefficient obtained in step S103.

[0098] In order to verify the effectiveness of the method proposed in this embodiment, the following simulation experiment was carried out.

[0099] The signal samples required for this simulation experiment are obtained by sampling from the RadioML2018.01 modulated signal public dataset. The dataset contains 24 types of modulated signals, including 19 types of digital modulated signals and 5 types of analog modulated signals. The signal-to-noise ratio distribution range of each type of signal is -20dB to 30dB, with a distribution interval of 2dB. The signal data format is [1024,2], that is, the I and Q data of the signal. In this simulation experiment, signal samples with a signal-to-noise ratio ranging from -10dB to 20dB in the dataset are selected for experimental simulation. The algorithm of this embodiment belongs to a meta-learning algorithm. The signal sample label space required for its training and testing phases does not intersect. Different types of signals are selected for training, verification, and testing respectively. The specific modulation styles of each dataset are shown in Table 1 below:

[0100] Table 1: Modulation patterns of the signal dataset

[0101]

[0102] The algorithm model in this application is built using the PyTorch neural network architecture in Python. The feature embedding module learns and updates network parameters using the Adam optimization algorithm. The algorithm presents a learning rate set to 0.01, and the training update iteration is 100 epochs. The batch size is 256, and the learning rate is reduced to 50% of the original after 20 epochs. The relationship measurement module optimizes network parameters using the SGD algorithm, with an initial learning rate set to 0.001. The algorithm updates and iterates for 200 epochs. Each epoch randomly samples 200 small sample recognition tasks from the training set. After each epoch, the recognition ability of the current model is verified by using a batch of recognition tasks from the validation set. The algorithm selects the model with the highest recognition accuracy on the validation set as the final network model. When testing the algorithm's recognition performance, 100 test recognition tasks are randomly sampled from the test set under each test condition, and the average recognition accuracy of all test tasks is used as the final recognition accuracy of the model under that condition.

[0103] Analysis of algorithm recognition performance under different signal sample size conditions:

[0104] The algorithm proposed in this application is designed for the problem of low-sample modulation recognition. To verify its recognition performance under small sample conditions, simulation experiments are conducted under different signal sample sizes. The metric level of the algorithm's network model is set to 4. First, different strategies are used on the training set to train the network parameters of the feature embedding module and the relationship measurement module respectively. Then, the network recognition performance is tested by sampling multiple different small sample recognition tasks from the test set.

[0105] In this simulation experiment, the task is "5-way K-shot" recognition, that is, the support set of each task consists of 5 types of modulated signals, and each type of signal includes K labeled samples. The query set uses N = 20 signal samples for each type of signal. The algorithm tests the change of average recognition accuracy with signal-to-noise ratio under different values ​​of K, such as Figure 7 shown.

[0106] from Figure 7 It can be seen that the recognition accuracy of the test recognition task "5-way K-shot" under different conditions gradually increases with the signal-to-noise ratio, and tends to be stable when the signal-to-noise ratio is greater than 10dB. Figure 7It can be seen that when the signal sample size shot value is 1, the algorithm can achieve good recognition capabilities. When the signal-to-noise ratio is 20dB, the recognition accuracy reaches 81.27%, which fully demonstrates the feasibility of the algorithm proposed in this application under low-sample conditions. The algorithm recognition rate gradually increases with the increase of sample size. When the sample size shot values ​​are 5, 10, 20, 50 and 100, respectively, compared with 1 shot, the recognition accuracy increases by 7.95%, 10.08%, 10.93%, 11.82% and 11.94%.

[0107] The feature embedding module, through training, adapts to the sample distribution characteristics of the entire dataset and can fully extract the features of different signal samples within the dataset. Using an episode-based strategy, the relationship measurement module learns an efficient feature measurement function, enabling accurate recognition of new types of signals sampled from the test set. It was also found that when the sample size exceeds 5 shots, the algorithm's recognition performance essentially reaches saturation, and subsequent increases in sample size gradually mitigate the improvement in the algorithm's recognition rate.

[0108] Analysis of algorithm recognition performance under different metric levels:

[0109] This application uses a multi-level metric to determine the similarity of samples. In order to verify the impact of different metric levels on the recognition performance of the algorithm, this simulation experiment sets 3, 4, 5 and 6 levels of metric respectively. The "5-way 1-shot" and "5-way 5-shot" tasks are simulated under different metric levels. The algorithm's recognition accuracy changes with the signal-to-noise ratio. Figure 8a and Figure 8b shown.

[0110] The network parameter counts, training time for the relationship measurement module, and the "5-way 5-shot" person recognition accuracy at 20dB for different metric levels are shown in Table 2 below:

[0111] Table 2: Comparison of network parameter volume and training time performance

[0112]

[0113] from Figure 8a and Figure 8bAs can be seen from the results, the algorithm's recognition accuracy gradually improves with the increase in the number of metric levels. At 20 dB, the "5-way 5-shot" recognition accuracy in a 6-level relational metric network test reaches 97.5%. Compared to the 3-, 4-, and 5-level relational metric networks, the test recognition accuracy increases by 10.01%, 5.68%, and 2.62%, respectively. In the proposed algorithm, the feature embedding module adds an L8TM network layer to the residual network, which can simultaneously learn the temporal characteristics of the signal samples and improve the representativeness of the proposed features.

[0114] As can be seen from Table 2, as the number of metric levels increases, the number of network parameters also increases rapidly, which further improves the network's performance and ability, enabling it to learn more detailed and more segmented features of sample signals. However, this also increases the difficulty of network training, resulting in longer and longer training times.

[0115] Analysis of algorithm improvement performance:

[0116] This application improves the relational network algorithm, uses a multi-level measurement method to determine the similarity between samples, and realizes data enhancement through feature reconstruction. In order to verify the effectiveness of the algorithm improvement proposed in this application, this simulation experiment conducted relevant comparative experiments. In the experiment, the relational network was used as the "Baseline", and multi-level measurement and feature reconstruction operations were added to this basic network, and simulation analysis was performed. All experiments used a 4-level measurement network model, and all modulated signal data sets are shown in Table 1; the simulation experiment results at different signal-to-noise ratio levels are shown in Table 1. Figure 9a and Figure 9b shown.

[0117] from Figure 9a and Figure 9bAs can be seen from the results, compared to the basic relational network, the multi-level metric and feature reconstruction operations both improve the recognition performance of the proposed algorithm to a certain extent. For example, when the test signal-to-noise ratio (SNR) is 20dB, the network model using multi-level metric and feature reconstruction improves the recognition accuracy of the "5-way 1-shot" task by 5.92% and 5.05%, respectively, and the recognition accuracy of the "5-way 5-shot" task by 4.41% and 4.07%, respectively. The multi-level metric measures the similarity between all levels of sample features, while also considering the differences between local and global features. During training, the algorithm learns weights corresponding to the similarity between feature vectors at all levels of the sample, adaptively increasing the weights corresponding to the level metric results that have a greater impact on the final recognition result. This mitigates recognition errors caused by the lack of representativeness of a single feature and improves the algorithm's recognition ability. Feature reconstruction draws on the idea of ​​variational autoencoders. The algorithm learns the distribution corresponding to the sample feature vectors and uses this distribution to reconstruct the current feature vector. Since the current feature vector is randomly reconstructed each time, it achieves data augmentation and avoids overfitting during network training. From the simulation results, we can see that multi-level measurement and multi-level reconstruction can effectively improve the recognition performance of the algorithm.

[0118] Analysis of the recognition performance of the algorithm under different small sample signal learning:

[0119] This simulation experiment selects 5 different down-sample learning algorithms for experimental comparison with the algorithm proposed in this application. These 5 algorithms are: relational network, twin network, prototype network, matching network, and MAML algorithm. In the comparative experiment, the network model of the algorithm proposed in this application is set as a relational metric network with 4-level metrics. Due to performance limitations of the MAML algorithm, a deeper neural network cannot be used. Therefore, a 5-layer convolutional neural network will be used as the feature extraction network of MAML in this simulation experiment, and the other comparative learning algorithms use "ResNet34" as the feature extraction network of the network model. The data sets used in the simulation experiments all use the modulated signals shown in Table 1. In the "5-way 1-shot" and "5-way 5-shot" tasks, the recognition accuracy of the centralized algorithm changes with the signal-to-noise ratio as shown below. Figure 10a and 10b shown.

[0120] from Figure 10a and 10bIt can be seen that the algorithm proposed in this application can achieve higher recognition accuracy compared with other meta-learning algorithms. The MAML algorithm has poor recognition performance compared to other algorithms. The MAML algorithm is a meta-learning algorithm for learning initialization. It finds an optimal initialization parameter for the network through learning multiple tasks, so that the network can quickly adapt when facing the recognition task of new categories of signals. However, the MAML algorithm needs to fine-tune the network when facing new categories of signal recognition tasks. Due to the small number of samples, it is difficult to use a network with a large number of parameters, which limits the further improvement of the network recognition performance; the other comparison algorithms are all traditional metric learning algorithms. The circular network and the twin network both use Euclidean distance to measure the similarity between samples. The matching network measures the similarity between samples by cosine distance. The algorithm proposed in this application uses a nonlinear neural network to measure the similarity between samples, which can achieve a more accurate representation of the distance between different signal samples compared to other algorithms; in addition, this application also adopts multi-level measurement and realizes data enhancement by reconstructing sample feature vectors, which further improves the recognition performance of the algorithm proposed in this application. However, the algorithm proposed in this application requires more network parameters to be trained than other algorithms, which also increases the difficulty of the algorithm proposed in this application.

[0121] In summary, the simulation experimental results show that this application is a further improvement to the problems existing in the small sample recognition algorithm based on metric learning. This method reconstructs and trains each level of features through the feature embedding module, and reconstructs the feature vector using reparameterization technology, which avoids the problem of overfitting that is prone to occur when using deep neural networks; the similarity of each level of features of the small signal sample is measured through the relational measurement module, and the final similarity score between the small sample signals is determined by weighted summation; this multi-level measurement method fully considers the differences between the features of signal samples at all levels, and effectively improves the recognition ability of the algorithm.

[0122] It should be noted that although several units of the system for action execution are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiment of the present disclosure, the features and functions of two or more units described above can be concretized in one unit. Conversely, the features and functions of a unit described above can be further divided into multiple units for concretization. Some or all of the units can be selected according to actual needs to achieve the purpose of the disclosed solution. Those of ordinary skill in the art can understand and implement it without paying creative work.

[0123] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the appended claims.

Claims

1. A small sample signal modulation recognition method based on a multi-level relational metric network, characterized in that: The following steps are involved: Establishing a network model of a multi-level relationship measurement network, wherein the network model includes a multi-level feature embedding module and a multi-level relationship measurement module; Inputting the small sample signal to be identified into the feature embedding module of each level respectively, performing reconstruction feature training, and obtaining the corresponding reconstructed feature vector of each level respectively; The small sample signal includes the support set and queryset , the support set and the queryset The basic unit T constituting the small sample signal, The support set The query set consists of label signal samples. It is composed of the signal samples to be tested, and the support set and the queryset The label space of the signal samples in the same; if the support set It contains C types of signals, and each type of signal contains K signal samples, so the small sample signal is a C-way K-shot recognition signal; For the C-way K-shot recognition signal, the support set and the queryset The signal samples in each level are input into the feature embedding module In each level, the feature embedding module Outputting sample distribution characteristic statistics respectively, and generating the reconstructed characteristic vector corresponding to each level respectively through a reparameterization method; The sample distribution characteristic statistics include the eigenvector mean and the eigenvector variance ; The formula for reconstructing the feature vector includes: in, represents the input signal sample; represents the reconstructed feature vector; represents the parameters of random sampling from a Gaussian distribution; Indicates the multiplication of corresponding elements; Indicates the class; Represents the network parameters of the feature embedding module; Inputting the reconstructed feature vector corresponding to each level into the relationship measurement module of the corresponding level, performing multi-level measurement training, and obtaining the similarity score and corresponding weighting coefficient of the small sample signal at each level; The final similarity scores of all the small sample signals are determined by weighted summation.

2. The small sample signal modulation recognition method based on a multi-level relational metric network according to claim 1 is characterized in that: In the step of establishing a network model of a multi-level relationship measurement network, The feature embedding module is used It indicates that each level of the feature embedding module includes a plurality of residual blocks with convolution kernels connected in sequence, and a long short-term neural memory network connected to the last residual block.

3. The small sample signal modulation recognition method based on a multi-level relational metric network according to claim 1 is characterized in that: In the step of establishing a network model of a multi-level relationship measurement network, The relationship measurement module is used It indicates that each level of the relationship measurement module includes two residual blocks with convolution kernels connected in sequence, and a fully connected network connected to the last residual block; the fully connected network includes a ReLU activation function.

4. The small sample signal modulation recognition method based on a multi-level relational metric network according to claim 1 is characterized in that: In the step of inputting the reconstructed feature vector corresponding to each level into the relationship measurement module of the corresponding level, performing multi-level measurement training, and obtaining the similarity score and corresponding weighting coefficient of the small sample signal at each level, When the support set Neidi The signal described in the When the eigenvector of the class represents the mean of the eigenvectors of all signal samples of this class, we get: in, Indicates the Class signal; Indicates the Class I signal signal samples; Then the output formula of the relationship measurement module includes: in, express and the signal sample to be measured No. Level eigenvector and Similarity feature map output by the level-level relationship measurement module Feature connection; Represents the network parameters of the relationship measurement module; if ,but .

5. The small sample signal modulation recognition method based on a multi-level relational metric network according to claim 4 is characterized in that: In the reconstruction feature training, a multi-category classification neural network is used to The features are embedded into the module Trained as a multi-class signal classifier, using the multivariate cross entropy loss function , according to the regularization strategy, in the multivariate cross entropy loss function A penalty term is introduced in the feature embedding module network parameter satisfy, in, Represents the mean of the variance output by each level of feature embedding module; Indicates the total amount of signal samples to be tested; represents the hyperparameter for adjusting the regularization penalty term; Indicates updating according to the multivariate cross entropy loss function .

6. The small sample signal modulation recognition method based on a multi-level relational metric network according to claim 5, characterized in that: In the multi-level metric training, the relationship metric module Adopt Episode's training strategy to embed the trained features into the module network parameters Fixed, and only the network parameters of the relationship metric module are updated in the multi-level relationship metric training , in each Episode, from the training set Randomly sample a C-way K-shot recognition signal and use the multivariate cross entropy loss function Calculate the training loss, which is: Where N represents the total number of samples of each type of signal in the query set, Labels representing signal samples; Represents the similarity score of small sample signals; Indicates the c the label of the class signal; Indicates the j the label of the class signal; Indicates when = The value is 1.

7. The small sample signal modulation recognition method based on a multi-level relational metric network according to claim 6, characterized in that: In the step of determining the final similarity scores of all the small sample signals by weighted summation, The calculation formula of the final similarity score includes: in, Represents the similarity score of the small sample signal; the similarity score of the small sample signal The calculation formula includes: Represents the weighting coefficient: the weighting coefficient The calculation formula includes: represents a fully connected network for calculating similarity scores; Represents a fully connected network that calculates weighted coefficients.

Images