A modulation recognition method based on transfer meta-learning

Abstract

The present disclosure relates to a modulation recognition method based on transfer meta-learning. The method comprises: pre-training network parameters of a deep neural network through all signal samples in a training set to obtain network parameters of the pre-trained deep neural network; wherein the network parameters of the deep neural network include feature extraction parameters and classifier parameters; optimizing the network parameters obtained by pre-training through different small sample recognition tasks in meta-training; wherein different small sample recognition tasks are randomly extracted from all signal samples in the training set; fine-tuning the optimized network parameters through a new small sample recognition task in meta-testing, so that the fine-tuned network parameters recognize the modulation style of the to-be-tested signal in the new small sample recognition task. According to the present disclosure, the fine-tuned network parameters can improve the recognition accuracy of the modulation style of the to-be-tested signal when recognizing the modulation style of the to-be-tested signal in the new small sample recognition task.

Description

Technical Field

[0001] This disclosure relates to the field of signal processing technology, and in particular to a modulation recognition method based on transfer meta-learning. Background Technology

[0002] Modulation pattern recognition of communication signals is a crucial step between signal detection and demodulation, and it has significant research value in fields such as information reconnaissance, electronic warfare, and electromagnetic spectrum monitoring. Traditional communication signal modulation recognition involves analyzing and extracting features from the target signal, and then designing a specific classifier based on these features to classify the signal. However, with the rapid development of modern communication technology, traditional methods are no longer suitable for modulation recognition tasks in complex electromagnetic environments.

[0003] In related technologies, deep learning techniques are used to solve the modulation recognition problem. One approach involves directly learning the IQ sequence samples of the signal using a Convolutional Neural Network (CNN), achieving the recognition of 11 types of modulation signals. Alternatively, ResNet-50 and Inception V2 networks are used to extract features from the signal constellation density matrix, and combined with a pre-noise filtering network to achieve accurate recognition of low signal-to-noise ratio (SNR) digital modulation signals. Another approach cascades a Bidirectional Long Short-Term Memory (BiLSTM) network with a CNN, achieving a recognition rate of 90% for 11 types of modulation signals when the SNR exceeds 4dB. A third approach employs a hybrid parallel network, where a CNN and a gated recurrent unit extract the spatial and temporal features of the signal samples respectively, and three attention mechanisms are used to weight and combine different features, highlighting the more discriminative parts of the extracted features, resulting in an accuracy improvement of at least 5% compared to a single network.

[0004] While deep learning-based modulation recognition methods have achieved significant recognition results, their success often requires at least several thousand labeled training samples. Insufficient data samples can lead to overfitting in the network. However, with the continuous expansion of practical applications, the labeled signal samples provided by communication signal environments are far from meeting the requirements of these deep learning-based modulation recognition methods, which greatly limits the application of deep learning technology in the field of modulation recognition.

[0005] Therefore, it is necessary to provide a new technical solution to improve one or more of the problems existing in the above solutions.

[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] The purpose of this disclosure is to provide a modulation recognition method based on transfer meta-learning, thereby overcoming, at least to some extent, one or more problems caused by the limitations and defects of related technologies.

[0008] A modulation recognition method based on transfer meta-learning, according to an embodiment of this disclosure, includes:

[0009] The network parameters of the deep neural network are pre-trained using all signal samples in the training set to obtain the pre-trained network parameters of the deep neural network; wherein, the network parameters of the deep neural network include feature extraction parameters and classifier parameters;

[0010] The network parameters obtained from pre-training are optimized by using different few-shot recognition tasks in meta-training; wherein, the different few-shot recognition tasks are randomly selected from all signal samples in the training set.

[0011] The optimized network parameters are fine-tuned by a new few-shot recognition task in the meta-test, so that the fine-tuned network parameters can recognize the modulation pattern of the signal to be tested in the new few-shot recognition task.

[0012] In embodiments of this disclosure, the step of pre-training the network parameters of the deep neural network using all signal samples in the training set to obtain the pre-trained network parameters of the deep neural network includes:

[0013] The feature extraction parameters and the classifier parameters are randomly initialized, and then updated and optimized using the training set recognition loss function.

[0014]

[0015] Where D represents the training set, α represents the learning rate during pre-training, θ represents the feature extraction parameters, θ represents the classifier parameters, θ′ represents the updated and optimized classifier parameters, and L D ([θ;θ]) represents the recognition loss function of the training set. This represents the gradient of the training loss on the training set.

[0016] In embodiments of this disclosure, the training set recognition loss can be expressed as:

[0017]

[0018] Where x represents the signal sample, y represents the signal label corresponding to x, and L D ([θ;θ]) represents the recognition loss of the training set, l(f [θ；θ] (x),y) represent the cross-entropy loss function, f[ θ；θ ] represents a deep neural network model.

[0019] In embodiments of this disclosure, the optimization of the pre-trained network parameters through different few-shot recognition tasks in meta-training includes the step of randomly selecting the different few-shot recognition tasks from all signal samples in the training set, comprising:

[0020] The feature extraction parameters in the pre-trained deep neural network are frozen, and learnable scaling and translation parameters are introduced to scale and translate the feature extraction parameters.

[0021] In the embodiments of this disclosure, during the scaling and translation operation, scaling and translation parameters [Ψ1,Ψ2] are set. When the input is a signal sample x, the output of the deep neural network can be expressed as:

[0022]

[0023] Where W represents the weights of the deep neural network, and b represents the bias of the deep neural network.

[0024] In embodiments of this disclosure, the optimization of the pre-trained network parameters through different few-shot recognition tasks in meta-training; wherein the step of randomly selecting the different few-shot recognition tasks from all signal samples in the training set includes:

[0025] New classifier parameters are set according to different few sample recognition tasks, and the new classifier parameters are initialized to obtain the initialized new classifier parameters.

[0026] In embodiments of this disclosure, the optimization of the pre-trained network parameters through different few-shot recognition tasks in meta-training; wherein the step of randomly selecting the different few-shot recognition tasks from all signal samples in the training set includes:

[0027] The initialization parameters of the new classifier are optimized by using the recognition loss optimization function generated in the meta-training to obtain the optimal initialization parameters of the new classifier.

[0028] In the embodiments of this disclosure, the initialization parameters of the optimal new classifier are:

[0029]

[0030] The initial values ​​of the scaling and translation parameters Ψ1 and Ψ2 are set to 1 and 0, respectively. Let θ0 represent the recognition loss optimization function generated during meta-training, θ″0 represent the initialization parameters of the new classifier, θ″0 represent the optimal initialization parameters of the new classifier, and β represent the first learning rate during meta-training.

[0031] In the embodiments of this disclosure, the scaling and translation parameters are updated using the recognition loss optimization function generated in the meta-test to obtain the updated and optimized scaling and translation parameters:

[0032]

[0033] in, Let p(T) represent the recognition loss optimization function generated in the meta-test, and let T represent the task distribution. i T represents the identification task of the i-th small sample drawn from the task distribution. i ~p(T), Ψ i Ψ represents the scaling and translation parameters corresponding to the i-th small sample recognition task. i ′ represents the updated and optimized scaling and translation parameters corresponding to the i-th small sample recognition task, where i is a positive integer, and γ represents the second learning rate in meta-training.

[0034] In the embodiments of this disclosure, the initialization parameters of the new classifier are updated using the recognition loss optimization function generated in the meta-test to obtain the updated and optimized initialization parameters of the new classifier:

[0035]

[0036] Where θ0′ represents the initialization parameters of the new classifier after the update and optimization.

[0037] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects:

[0038] In one embodiment of this disclosure, the network parameters of the deep neural network are pre-trained using all signal samples in the training set through the above method. The network parameters obtained from the pre-training are optimized through different few-shot recognition tasks in the meta-training. The optimized network parameters are then fine-tuned through new few-shot recognition tasks in the meta-testing, so that the fine-tuned network parameters can quickly adapt to the new few-shot recognition tasks. Furthermore, when recognizing the modulation pattern of the signal to be tested in the new few-shot recognition task, the accuracy of the recognition of the modulation pattern of the signal to be tested can be improved based on the fine-tuned network parameters.

[0039] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0040] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0041] Figure 1 A flowchart illustrating a modulation recognition method based on transfer meta-learning in an exemplary embodiment of this disclosure is shown schematically.

[0042] Figure 2 This schematic diagram illustrates the overall block diagram of the modulation recognition method based on transfer meta-learning in an exemplary embodiment of the present disclosure;

[0043] Figure 3a This schematic diagram illustrates an n-level residual block of the ResNet feature extraction network in an exemplary embodiment of the present disclosure.

[0044] Figure 3b This schematic diagram illustrates the ResNet network structure of the ResNet feature extraction network in an exemplary embodiment of the present disclosure.

[0045] Figure 4 This schematic diagram illustrates the format of scaling and translation parameters in an exemplary embodiment of the present disclosure;

[0046] Figure 5a This illustration schematically shows the performance variation curves of different methods when the training sample size N=50 in an exemplary embodiment of the present disclosure;

[0047] Figure 5b This illustration schematically shows the performance variation curves of different methods when the training sample size N = 1000 in an exemplary embodiment of this disclosure;

[0048] Figure 6a This illustration schematically shows a global variation of the identification performance of different methods with sample size in exemplary embodiments of this disclosure;

[0049] Figure 6b This illustration schematically shows a local variation of the identification performance of different methods with sample size in exemplary embodiments of the present disclosure;

[0050] Figure 7a This illustration schematically shows the training accuracy of the MAML method when using ResNet in an exemplary embodiment of this disclosure;

[0051] Figure 7b This illustration schematically shows a diagram of the training loss of the MAML method when using ResNet in an exemplary embodiment of this disclosure;

[0052] Figure 8a This schematic diagram illustrates the training accuracy of the transfer meta-learning method in an exemplary embodiment of the present disclosure.

[0053] Figure 8b The diagram illustrates a training loss diagram using a transfer meta-learning method in an exemplary embodiment of this disclosure. Detailed Implementation

[0054] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary 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.

[0055] This example implementation provides a modulation recognition method based on transfer meta-learning. (Reference) Figure 1 As shown, the method may include:

[0056] Step S101: Pre-train the network parameters of the deep neural network using all signal samples in the training set to obtain the pre-trained network parameters of the deep neural network; wherein, the network parameters of the deep neural network include feature extraction parameters and classifier parameters.

[0057] Step S102: Optimize the network parameters obtained from pre-training by using different few-shot recognition tasks in meta-training; wherein, the different few-shot recognition tasks are randomly selected from all signal samples in the training set.

[0058] Step S103: Fine-tune the optimized network parameters through the new few-shot recognition task in the meta-test, so that the fine-tuned network parameters can recognize the modulation pattern of the signal to be tested in the new few-shot recognition task.

[0059] The above method pre-trains the network parameters of the deep neural network using all signal samples in the training set, optimizes the pre-trained network parameters using different few-shot recognition tasks in meta-training, and fine-tunes the optimized network parameters using new few-shot recognition tasks in meta-testing. This allows the fine-tuned network parameters to quickly adapt to new few-shot recognition tasks, and improves the recognition accuracy of the modulation pattern of the signal under test when recognizing the modulation pattern of the signal under test in new few-shot recognition tasks based on the fine-tuned network parameters.

[0060] Below, we will refer to Figures 1 to 2 The steps of the method described above in this example embodiment will be explained in more detail.

[0061] In step S101, the network parameters of the deep neural network are pre-trained using all signal samples in the training set to obtain the pre-trained network parameters of the deep neural network; wherein, the network parameters of the deep neural network include feature extraction parameters and classifier parameters.

[0062] Specifically, the modulation recognition method using transfer meta-learning includes two stages: a transfer pre-training stage and a meta-learning stage. The first stage is the transfer pre-training stage, which involves pre-training a deep neural network (DNN) using all signal samples in the training set. This yields the network parameters of the pre-trained DNN, including feature extraction parameters and classifier parameters. The classifier parameters are set for all signal samples. The deep neural network, structurally speaking, can be divided into feature extraction and classification parts, each composed of convolutional layers and fully connected layers, respectively. This embodiment designs the feature extraction network of the modulation recognition method based on a shallow convolutional neural network (ConvNet) and a deep residual network (ResNet). For example... Figure 3a and Figure 3b As shown, the ConvNet in this disclosure consists of 5 convolutional layers. Each convolutional layer consists of a (3×1) kernel convolution, a batch normalization layer, a ReLU activation function, and a (2×1) kernel max pooling layer. The number of convolutional kernels is set to 16, 32, 64, 128, and 256 respectively. The ResNet consists of multi-level residual blocks. Each residual block contains 3 (3×1) convolutions. The (1×1) convolutions ensure that the input and output formats are consistent, achieving skip-layer connections. Finally, a (2×1) max pooling layer is cascaded. The residual structure can avoid the gradient vanishing problem caused by excessive network depth.

[0063] In step S102, the network parameters obtained from pre-training are optimized by different few-shot recognition tasks in meta-training; wherein, the different few-shot recognition tasks are randomly selected from all signal samples in the training set.

[0064] Specifically, after the transfer learning pre-training stage, the meta-learning stage begins. A small number of signal samples are randomly selected from all signal samples in the training set to form different small sample recognition tasks. Each small sample recognition task has the same number of signal sample categories. The meta-learning stage is further divided into meta-training and meta-testing processes. During the meta-training process, the network parameters obtained from pre-training are optimized through different small sample recognition tasks to obtain optimized network parameters. These optimized network parameters are then applied in subsequent recognition of test signal samples.

[0065] In step S103, the optimized network parameters are fine-tuned by a new few-shot recognition task in the meta-test, so that the fine-tuned network parameters can recognize the modulation pattern of the signal to be tested in the new few-shot recognition task.

[0066] Specifically, during the meta-testing process in the meta-learning phase, the optimized network parameters are fine-tuned. This fine-tuned network parameters improve the recognition accuracy of test samples when identifying the modulation patterns of the test signals in the new few-shot recognition task. The number of signal sample categories in the new few-shot recognition task is the same as the number of signal sample categories in the original few-shot recognition task.

[0067] In one embodiment, the step of pre-training the deep neural network parameters using all signal samples in the training set to obtain the pre-trained deep neural network parameters includes:

[0068] The feature extraction parameters and the classifier parameters are randomly initialized, and then updated and optimized using the training set recognition loss function.

[0069]

[0070] Where D represents the training set, α represents the learning rate during pre-training, θ represents the feature extraction parameters, θ represents the classifier parameters, θ′ represents the updated and optimized classifier parameters, and L D ([θ;θ]) represents the recognition loss function of the training set. This represents the gradient of the training loss on the training set.

[0071] Specifically, in the transfer learning pre-training phase, the network parameters of the deep neural network (DNN) are optimized using all signal samples in the training set. First, the feature extraction parameters θ and classifier parameters θ of the DNN are randomly initialized. Then, gradient descent is used to optimize the network parameters [θ; θ], that is, the DNN parameters and classifier parameters are updated and optimized using the training set recognition loss function, resulting in updated and optimized feature extraction parameters and classifier parameters. The updated and optimized feature extraction parameters are then used in the meta-learning phase. The network parameters [θ; θ] of the deep neural network include the feature extraction parameters θ and the classifier parameters θ.

[0072] In one embodiment, the training set identification loss function can be expressed as:

[0073]

[0074] Where x represents the signal sample, y represents the signal label corresponding to x, and L D ([θ;θ]) represents the recognition loss of the training set, l(f [θ；θ] (x),y) represent the cross-entropy loss function, f[ θ；θ ] represents a deep neural network model.

[0075] Specifically, the feature extraction parameters and classifier parameters are updated and optimized using the training set recognition loss function shown in the above formula (2).

[0076] In one embodiment, the optimization of the pre-trained network parameters through different few-shot recognition tasks in meta-training; wherein the step of randomly selecting the different few-shot recognition tasks from all signal samples in the training set includes:

[0077] The feature extraction parameters in the pre-trained deep neural network are frozen, and learnable scaling and translation parameters are introduced to scale and translate the feature extraction parameters.

[0078] Specifically, since the number of signal samples in the training set is not equal to the number of signal categories to be identified in the few-shot recognition task (e.g., the pre-trained classifier is a 10-class (10-way) signal classifier, while the meta-learning process requires a 5-class (5-way) signal classifier), only the feature extraction parameters of the deep neural network (DNN) are retained during the meta-learning stage. The feature extraction parameters θ in the pre-trained DNN are frozen during the meta-learning stage, and a new classifier is set for the few-shot recognition task. These new classifier parameters are referred to as the new classifier parameters. Furthermore, learnable scaling and translation parameters are introduced to scale and translate the pre-trained feature extraction parameters, thus transferring these parameters to the meta-learning process of the few-shot recognition task. The few-shot recognition task includes training data T. tr and test data T te .

[0079] In one embodiment, during the scaling and translation operation, scaling and translation parameters [Ψ1, Ψ2] are set. When the input is a signal sample x, the output of the deep neural network can be represented as:

[0080]

[0081] Where W represents the weights of the deep neural network, and b represents the bias of the deep neural network.

[0082] Specifically, in the meta-learning stage, when scaling and translating the pre-trained feature extraction parameters, learnable scaling and translating parameters [Ψ1,Ψ2] are set, and when the input is a signal sample x, the output of the deep neural network is obtained.

[0083] In one embodiment, the optimization of the pre-trained network parameters through different few-shot recognition tasks in meta-training; wherein the step of randomly selecting the different few-shot recognition tasks from all signal samples in the training set includes:

[0084] New classifier parameters are set according to different few-sample recognition tasks, and the new classifier parameters are initialized to obtain new classifier initialization parameters.

[0085] Specifically, before meta-training, for the few-sample recognition task, new classifier parameters are randomly set and initialized to obtain the new classifier initialization parameters.

[0086] In one embodiment, the optimization of the pre-trained network parameters through different few-shot recognition tasks in meta-training; wherein the step of randomly selecting the different few-shot recognition tasks from all signal samples in the training set includes:

[0087] The initialization parameters of the new classifier are optimized by using the training recognition loss optimization function generated in the meta-training to obtain the optimal initialization parameters of the new classifier.

[0088] Specifically, based on the new classifier initialization parameters obtained above, the recognition loss function generated during the meta-training process is used to optimize them, resulting in the optimal new classifier initialization parameters.

[0089] In one embodiment, the initialization parameters of the optimal new classifier are:

[0090]

[0091] The initial values ​​of the scaling and translation parameters Ψ1 and Ψ2 are set to 1 and 0, respectively. Let θ0 represent the training recognition loss optimization function generated in meta-training, θ″0 represent the initialization parameters of the new classifier, θ″0 represent the optimal initialization parameters of the new classifier, and β represent the first learning rate in meta-training.

[0092] Specifically, as shown in formula (4), the initialization parameters of the optimal new classifier are calculated. The initial values ​​of the scaling and translation parameters Ψ1 and Ψ2 are set to 1 and 0 respectively, and the training data T in the meta-training is used. tr The optimal initialization parameters for the new classifier are obtained by optimizing the training loss function and the first learning rate β in the meta-training. Among these, the two learnable scaling and translation parameters Ψ1 and Ψ2 have different structures depending on the structure of the feature extraction parameters, such as... Figure 4 As shown, when the feature extraction weight parameter format is C×N k×k1×k2, then the corresponding scaling and translation parameters Ψ1 and Ψ2 are in the format C×N respectively. k ×1×1 and 1×N k ×1×1. Compared to general MAML methods, the modulation recognition method based on transfer meta-learning disclosed in this invention only needs to update the scaling and translation parameters Ψ1 and Ψ2 during the meta-learning stage, reducing the number of parameters required for training deep neural networks and lowering the training difficulty of modulation recognition methods using deep neural networks under small sample conditions. In one embodiment, the scaling and translation parameters are updated using the test recognition loss optimization function generated in the meta-training, resulting in the updated and optimized scaling and translation parameters:

[0093]

[0094] in, Let p(T) represent the test recognition loss optimization function generated in the meta-test, and let T represent the task distribution. i T represents the identification task of the i-th small sample drawn from the task distribution. i ~p(T), Ψ i Ψ represents the scaling and translation parameters corresponding to the i-th small sample recognition task. i ′ represents the updated and optimized scaling and translation parameters corresponding to the i-th small sample recognition task, where i is a positive integer, and γ represents the second learning rate in meta-training.

[0095] Specifically, through the test data T in meta-learning and meta-training. te The test recognition loss optimization function generated in the process, the second learning rate γ in the meta-training, and the scaling and translation parameters are substituted into formula (5) to obtain the updated and optimized scaling and translation parameters.

[0096] In one embodiment, the initialization parameters of the new classifier are updated using the recognition loss optimization function generated in the meta-test, resulting in the updated and optimized initialization parameters of the new classifier:

[0097]

[0098] Where θ0′ represents the initialization parameters of the new classifier after the update and optimization.

[0099] Specifically, through the test data T in meta-learning and meta-training. te The test recognition loss optimization function generated in the process, the second learning rate γ in the meta-training, the classifier initialization parameter θ0 and the scaling and translation parameter are substituted into formula (6) to obtain the updated and optimized new classifier initialization parameters.

[0100] In this modulation recognition method based on transfer meta-learning, multiple different few-shot recognition tasks are learned during the meta-training process to find updated and optimized scaling and translation parameters Ψ1 and Ψ2, as well as new classifier initialization parameters θ0. During the meta-testing process, the deep neural network fine-tunes the new classifier initialization parameters by learning a small number of labeled signal samples in the new few-shot recognition task, thereby achieving accurate recognition of the test signal in the new few-shot recognition task.

[0101] The embodiments of this disclosure will be further illustrated below through simulation experiments.

[0102] 1. Experimental Dataset and Parameter Settings

[0103] RadioML2018.01 is a publicly available modulation recognition dataset proposed by Deepsig. This dataset contains 24 classes of modulation signals. The experiment selected modulation signals with a signal-to-noise ratio (SNR) ranging from -10dB to 20dB, with an SNR distribution interval of 2dB. The dataset required for the simulation experiment in this section was constructed by sampling 1000 samples from each SNR point of each signal class. The signal data format is [1024,2], representing I and Q-path components with a sequence length of 1024. In the experiment, the training set, validation set, and test set contain 10, 7, and 7 classes of modulation signals, respectively. The specific modulation patterns of each dataset are shown in Table 1. During the meta-learning process, a large number of different N-way K-shot recognition tasks are randomly sampled from these datasets to implement the modulation recognition method based on transfer meta-learning in this embodiment.

[0104] Table 1 Sample Dataset

[0105]

[0106] The simulation experiment used the PyTorch neural network architecture under Python to build the network model. The experiment was conducted on a server with Windows 7 system, 32GB of memory, and an NVIDIA P4000 graphics card.

[0107] In the pre-training phase, the network parameters were optimized using the Gradient Descent (SGD) algorithm with an initial learning rate of 0.01. Training was conducted for 100 epochs, with the learning rate reduced to 50% every 20 epochs. The model with the highest validation recognition rate was selected for the meta-learning phase. In the meta-training phase, the Adam modulation recognition method was optimized, with inner and outer loop learning rates set to 0.001 and 0.0001, respectively. The model with the highest validation recognition rate was selected as the final training model. In the meta-testing phase, 600 test recognition tasks were sampled from the test set to test the network model's recognition accuracy. The average recognition accuracy of all test tasks was used to characterize the final recognition performance of the modulation recognition method. In each recognition task, 15 signal samples from each signal class were selected as test samples.

[0108] 2. Performance Analysis of Transfer Meta-Learning Methods

[0109] (1) Method performance analysis

[0110] The transfer meta-learning method can solve the modulation recognition problem under conditions of a very small number of labeled signal samples. To verify the recognition performance of the modulation recognition method, simulation experiments were conducted under different sample sizes. To further demonstrate the performance advantage of the transfer meta-learning method, several different modulation recognition methods were selected, namely CNN, Transfer Learning (TL), and Data Augmentation (DA). The performance differences of all modulation recognition methods under different sample sizes were compared and analyzed. In this disclosure, the feature extraction network of the transfer meta-learning method is set as a ResNet network. To ensure the reliability of the experimental results, according to the principle of single variable experiment, the dataset used for all comparison methods in the transfer meta-learning method is the test set shown in Table 1. The network parameters of the comparison methods were trained by sampling a portion of the signal samples in this dataset, and the remaining samples were used to test the recognition performance of the algorithms. The transfer meta-learning method in this disclosure uses the dataset in Table 1 for simulation experiments. Figure 5a and Figure 5b The diagram illustrates the transfer meta-learning method of this disclosure for the "5-way 5-shot" and "5-way 1-shot" test tasks, as well as the comparison method, with the test recognition accuracy as a function of signal-to-noise ratio when the training sample size N for each type of signal is 50 and 100, respectively.

[0111] from Figure 6a and Figure 6bAs can be seen, when the training sample size N of the comparison methods is 50, the transfer meta-learning method of this disclosure achieves the best recognition performance for both the "5-way 1-shot" and "5-way 5-shot" test tasks compared to other comparison methods. When N is 100, the transfer meta-learning method of this disclosure still outperforms other comparison methods in the recognition performance of the "5-way 5-shot" task when the signal-to-noise ratio of the test signal is greater than 0dB. Experimental results demonstrate that the transfer meta-learning method of this disclosure can still achieve a high recognition accuracy when there are only a few training samples of the test signal. When the sample size used is much smaller than other comparison methods, the transfer meta-learning method of this disclosure can achieve better recognition performance. The main reason is that the transfer meta-learning method of this disclosure adopts a meta-learning strategy to optimize the network model. That is, during the training phase, the model network parameters are optimized by learning a large number of different small sample recognition tasks. Through this cross-task training method, the network model has the ability to "learn to learn" and can quickly adapt to new tasks. Unlike meta-learning strategies, the contrastive method directly learns the mapping relationship between signal samples and their corresponding labels. The trained network can directly identify the modulation pattern of new test signal samples. However, for the network model to master this ability, it needs to learn a large number of different signal samples. Insufficient training samples can lead to a decrease in recognition performance. Experimental results show that the transfer meta-learning method in this disclosure maintains high recognition accuracy even when there are only one labeled sample per signal class. Compared to the three contrastive methods, the transfer meta-learning method in this disclosure has a significant performance advantage. When the signal-to-noise ratio of the test signal is 20dB, the global and local changes in recognition accuracy of different methods with the amount of training samples are shown in the following figures. Figure 6a and Figure 6b As shown. From Figure 6a and Figure 6b As can be seen, the transfer meta-learning method in this disclosure requires only 20 samples to achieve the best recognition performance, which is far less than the sample size required by other comparative methods. When the training sample size of the signal to be tested is only a few, the recognition performance of the transfer meta-learning method in this disclosure is significantly better than that of the comparative methods. For example, when the training sample size of each type of signal is 20, compared with the three modulation recognition methods based on CNN, TL, and DA, the recognition accuracy of the transfer meta-learning method in this disclosure is improved by 79.66%, 72.42%, and 78.05%, respectively.

[0112] (2) Method Comparison Ablation Experiment

[0113] To further improve the recognition accuracy of the network model when the number of labeled signal samples is only a few, the transfer meta-learning method in this disclosure adds a transfer pre-training process based on the Model-Agnostic Meta-Learning (MAML) method, and sets trainable and updatable scaling and translation parameters [Ψ1,Ψ2] during the meta-learning process. To verify the effectiveness of these improvements, ablation experiments were conducted on the dataset shown in Table 1. When the signal-to-noise ratio of the test signal is 20dB, the experimental results for the 5-way modulation signal recognition task are shown in Table 2. The network that has undergone transfer pre-training will freeze the network parameters of the feature extraction part during the meta-learning stage.

[0114] Table 2 Comparison of Methods in Ablation Experiments

[0115]

[0116] As shown in Table 2, the improvements to the MAML method in this disclosure have all further increased the network's recognition accuracy. Analysis of the data in the table reveals that when the MAML method uses a deeper ResNet as the feature extraction network, the network's recognition accuracy actually decreases. This is because the MAML method has only a very small number of training samples for each recognition task, such as... Figure 7a , Figure 7b As shown in Figure 1, when using a deep feature extraction network, directly updating the entire network parameters makes it difficult to train the network to converge. The network's recognition accuracy and loss fluctuate drastically throughout the meta-training process, preventing the network from achieving a stable learning effect and leading to a significant decline in recognition performance. The transfer learning process further improves the network's feature extraction capability for signal samples with the same distribution, alleviating the learning pressure on the network during the meta-learning stage. Freezing the network parameters of the feature extraction part during meta-learning, i.e., freezing the feature extraction parameters, greatly reduces the number of network parameters that the network needs to learn and update, thus lowering the difficulty of network training. Figure 8a , Figure 8b As shown, the transfer meta-learning method, when using a deeper ResNet feature extraction network, allows the network to converge to its optimal state after 60 training iterations. Furthermore, freezing the network parameters of the feature extraction part during the meta-learning stage allows the training process to focus more on learning the parameters of the classification part, enhancing the sensitivity of the classification network to the extracted signal features. On the other hand, setting learnable scaling and translation parameters [Ψ1,Ψ2] to transfer the network parameters of the feature extraction part, and obtaining the optimal [Ψ1,Ψ2] through meta-training, helps the feature extraction network adapt to the recognition task of new signal classes as quickly as possible during the meta-testing stage, further improving the network's recognition performance.

[0117] (3) Performance comparison of different meta-learning methods

[0118] To further verify the performance advantages of the transfer meta-learning method, several other meta-learning algorithms were selected for comparative experiments, including Prototypical Network (PN), Relation Network (RN), Matching Network (MN), Siamese Network (SN), and Meta-leaner LSTM. These meta-learning methods were used to solve the modulation recognition problem proposed in this disclosure when there are only a few labeled signal samples. All methods were simulated on the dataset shown in Table 1. When the signal-to-noise ratio of the test signal is 20dB, the test results of the above methods on the 5-way 5-shot and 5-way 1-shot recognition tasks are shown in Table 3. As can be seen from the data in Table 3, the transfer meta-learning method of this disclosure achieves better recognition results compared with other methods.

[0119] Table 3 Performance Comparison of Different Meta-Learning Methods

[0120]

[0121]

[0122] The transfer meta-learning method disclosed in this paper has significant performance advantages, achieving optimal recognition results in both 5-way 5-shot and 5-way 1-shot recognition tasks. Among the five meta-learning methods compared, PN, RN, MN, and SN all belong to metric-based meta-learning methods. These methods learn a suitable feature metric space through multiple training tasks. When facing a new class of signal recognition task, they do not need to update network parameters; they only need to map the labeled signal samples and the test signal samples to the trained feature metric space, calculate the distance between samples using a certain distance metric function, and find the nearest neighbor class to determine the recognition result. Due to the limitation of sample size, metric-based meta-learning methods find it difficult to learn a high-efficiency feature metric space. Compared with the transfer meta-learning method in this paper, the recognition accuracy of these four metric-based meta-learning methods on the 5-way 5-shot recognition task decreased by 8.61%, 7.98%, 18.48%, and 33.16%, respectively. The Meta-leaner LSTM method trains an LSTM meta-network to learn the update rules for the base network parameters. When faced with a new type of signal recognition task, the network parameters can be updated more accurately under the guidance of the LSTM meta-network to quickly adapt to the new task. However, the training loss of the base network and the gradient of the meta-network parameters both depend on the parameters of the meta-network, resulting in high computational complexity. Compared with the transfer meta-learning method disclosed in this paper, the Meta-leaner LSTM method reduces the recognition accuracy by 31.16% on the 5-way 5-shot recognition task. The test recognition rate of the transfer meta-learning method disclosed in this paper is also much higher than other meta-learning methods on the 5-way 1-shot recognition task.

[0123] This disclosure presents a transfer meta-learning method to address modulation recognition problems with only a few labeled signal samples. It reduces the training difficulty of the MAML method using deep feature extraction networks through transfer pre-training, improving the network's ability to extract signal features. To reduce the number of parameters required for training when facing new signal recognition tasks, the network parameters for feature extraction are frozen during the meta-learning stage, and scaling and translation parameters are introduced to transfer them to the learning process for new tasks. During the meta-testing stage, the network fine-tunes the classification parameters starting from the optimal initial weights obtained through meta-training. This allows the network to quickly adapt to new signal recognition tasks with only a small number of signal samples. Comparative experiments further verify the feasibility of this transfer meta-learning method and its performance advantages over other methods under conditions of extremely small sample sizes.

[0124] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.

Claims

1. A modulation recognition method based on transfer meta-learning, characterized in that, The method comprises: Pre-training network parameters of a deep neural network through all signal samples in a training set, to obtain network parameters of the deep neural network after pre-training; wherein the network parameters of the deep neural network comprise feature extraction parameters and classifier parameters; Optimizing the network parameters obtained by pre-training through different small sample identification tasks in meta-training; wherein the different small sample identification tasks are respectively obtained by random sampling from all signal samples in the training set; Fine-tuning the optimized network parameters through a new small sample identification task in meta-testing, so that the fine-tuned network parameters identify the modulation pattern of a signal to be tested in the new small sample identification task; The step of optimizing the network parameters obtained by pre-training through different small sample identification tasks in meta-training, wherein the different small sample identification tasks are respectively obtained by random sampling from all signal samples in the training set, comprises: Freezing the feature extraction parameters in the deep neural network obtained by pre-training, and introducing learnable scaling and translation parameters to perform scaling and translation operations on the feature extraction parameters; The scaling and translation operation is performed by setting scaling and translation parameters When the input is a signal sample The deep neural network outputs can be represented as: （3） wherein, denote weights of the deep neural network, denote biases of the deep neural network; Setting new classifier parameters according to the different small sample identification tasks, and initializing the new classifier parameters to obtain new classifier initialization parameters; Setting new classifier parameters according to the different small sample identification tasks, and initializing the new classifier parameters to obtain new classifier initialization parameters; The optimal new classifier initialization parameters are: （4） wherein the scaling translation parameter and are set to 1 and 0, respectively, as initial values, denotes a recognition loss optimization function generated in meta-training, denotes a new classifier initialization parameter, denotes an optimal new classifier initialization parameter, denotes a first learning rate in meta-training, denotes a feature extraction parameter, denotes a classifier parameter, denotes a deep neural network model; Updating the scaling and translation parameters through an identification loss optimization function generated in the meta-testing, to obtain updated and optimized scaling and translation parameters: （5） wherein, represents an identification loss optimization function generated in meta testing, represents a task distribution, represents an i-th small sample identification task extracted from the task distribution, i.e. , represents a scaling and translation parameter corresponding to the i-th small sample identification task, represents an updated optimization scaling and translation parameter corresponding to the i-th small sample identification task, i is a positive integer, represents a second learning rate in meta training; Updating the new classifier initialization parameters through an identification loss optimization function generated in the meta-testing, to obtain updated and optimized new classifier initialization parameters: （6） wherein, denotes the updated optimized new classifier initialization parameters.

2. The modulation identification method based on transfer meta-learning according to claim 1, characterized in that, The step of pre-training network parameters of a deep neural network through all signal samples in a training set, to obtain network parameters of the deep neural network after pre-training, wherein the network parameters of the deep neural network comprise feature extraction parameters and classifier parameters, comprises: Randomly initializing the feature extraction parameters and the classifier parameters, and updating and optimizing the feature extraction parameters and the classifier parameters through a training set identification loss function: （1） wherein, denotes the training set, denotes the learning rate in pre-training, denotes the feature extraction parameters, denotes the classifier parameters, denotes the updated optimized classifier parameters, denotes the updated optimized classifier parameters, denotes the training set identification loss function, denotes the gradient of the training set training loss.

3. The method of claim 2, wherein the method further comprises: The training set identification loss function can be expressed as: （2） wherein, denotes a signal sample, denotes a corresponding signal label, denotes a training set recognition loss, denotes a cross-entropy loss function, denotes a deep neural network model.

Images