Signal modulation classification model and learning method based on sample recall incremental learning

Abstract

This solution discloses a signal modulation classification model and learning method based on sample recall incremental learning, including a memory recall module, a discrimination module and a classification network, and introduces a memory method based on class labels and sample labels. In the sample memory stage, the memory module will learn the samples and generate low-dimensional feature representations. In the sample recall stage, it is only necessary to input the class labels and sample labels into the recall module to accurately recall the corresponding samples. There is no need to save the sample data, only the trained recall module and sample label information need to be saved. By using the memory recall module framework as a warehouse for sample data, the dilemma of directly storing sample data and occupying a large amount of memory is overcome.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal recognition in wireless communications, and in particular relates to a signal modulation classification model based on sample recall incremental learning and a training method thereof. Background Art

[0002] In the field of communications technology, different modulation methods are required to facilitate wireless signal transmission, improve anti-interference capabilities, and meet diverse requirements. Radio signal modulation classification has been an ongoing topic of discussion and is widely used in both civilian and military fields. Traditional modulation classification methods mainly include maximum likelihood-based and feature-based methods. Automatic modulation classification based on likelihood-based methods is computationally complex and susceptible to the influence of wireless channel environments, making it difficult to deploy. Feature-based methods are considered a mapping relationship, and the rationality of feature selection strongly influences the design and performance of the identifier.

[0003] In recent years, with the exponential growth of computing power and the explosive growth of data volumes, machine learning and deep learning have experienced explosive growth in both industry and industry, and several outstanding neural network models have emerged. Deep learning, which learns the inherent patterns and representational features of sample data, has achieved remarkable results in natural language processing, computer vision, and bioinformatics. Modulation classification based on deep learning is also becoming mainstream.

[0004] Although modulation classification algorithms based on deep learning can achieve good classification results, the currently dominant machine learning paradigm relies on isolated learning, which prevents the learned knowledge from being retained for future learning. When learning a new task, without the input of previous data, the new model will often only recognize new signal samples, resulting in catastrophic forgetting of the old tasks. On the one hand, there is a desire to learn new knowledge, while on the other hand, it is necessary to prevent new input data from interfering with old knowledge. These two conflicting requirements constitute the stability-plasticity dilemma. Therefore, addressing catastrophic forgetting in incremental learning is currently the greatest challenge.

[0005] Incremental learning requires a balance between the performance of new and existing tasks, as well as the balance between computer storage resources and task performance. Implementing incremental learning within the frameworks of supervised classification, task-based incremental learning, and multi-head networks presents significant challenges. While incremental learning remains a hot research area and will likely continue to do so for a long time, future research should prioritize the exploration of biological systems and computational models. Summary of the Invention

[0006] The purpose of the present invention is to provide a radio modulation classification incremental learning method based on sample recall in order to solve the above problems.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A signal modulation classification model based on sample recall incremental learning includes a memory recall module, a discrimination module and a classification network. The memory recall module includes a memory module and a recall module spliced ​​in sequence. The memory module is used to memorize original samples, and the recall module is used to recall samples memorized by the memory module. In a class incremental training task, new radio signal samples and recalled samples recalled by the recall module are used as sample data of the model to train the model. The discrimination module is used to form an adversarial network with the memory recall module to train the memory recall module and the discrimination module. The classification network is used to output classification results based on its input.

[0009] In the above-mentioned signal modulation classification model based on sample recall incremental learning, in the first training task, the original samples only include the radio signal samples of this training task, and the memory recall module and the discrimination module are trained based on the radio signal samples of this training task; the classification network is trained based on the radio signal samples of this training task or based on the recall module trained on the recalled samples of the radio signal samples;

[0010] During the incremental training phase for tasks other than the first one, the original samples include the radio signal samples from the current training task and the recalled samples from the recall module. The memory recall module and the discrimination module are trained based on the radio signal samples and recalled samples from the current training task. The input to the classification network includes the recalled samples from the recall module and the radio signal samples from the current training task, that is, the recalled samples from the recall module for old samples and new samples, or the recalled samples from the recall module for old and new samples. During the testing or usage phase, the input to the classification network is the radio signal samples to be classified. The old samples are the radio signal samples from the previous training task, and the new samples are the radio signal samples from the current training task.

[0011] The samples memorized in the previous training task are recalled and restored by the recall module, and then memorized again using the memory module after this training. For the samples that have been memorized, the original samples refer to the samples restored by the recall module in this training task.

[0012] In the above-mentioned signal modulation classification model based on sample recall incremental learning, the memory recall module includes an embedded latent vector layer embedded in the memory module and the recall module;

[0013] The memory module is used to convert the original samples into low-dimensional features for memorization. In the first training task, the memory module converts the radio signal samples of this training task into low-dimensional features for memorization. In subsequent training tasks, the memory module converts the radio signal samples of this training task and the radio signal samples recalled by the recall module, i.e., the recall samples, into low-dimensional features for memorization:

[0014] z i =Net1(x i ) (3)

[0015] Among them, x i represents the original sample of the i-th type signal, z i Represents the latent variable characteristics of the i-th type signal;

[0016] Under each training task, the label information of the original sample of the training task is saved; at the same time, the embedded latent vector layer dynamically stores the latent variable features of the memory module based on the label information of the sample. The label information of each sample includes the class label and sample labels

[0017] The recall module recalls the corresponding recall samples based on the latent variable features of the memory:

[0018]

[0019] The loss function used to train the recall module is:

[0020]

[0021] is the recall sample obtained by the recall module for the i-th type of signal, x i is the original sample of the i-th type signal.

[0022] In the above-mentioned signal modulation classification model based on sample recall incremental learning, the adversarial network composed of the discrimination module and the memory recall module includes the following loss function:

[0023]

[0024] l gen = -log(Net3(z)) (14)

[0025] The loss function (13) is used to update the discrimination module, and the loss function (14) is used to update the memory module to confuse the discrimination module;

[0026] The classification network includes a cross entropy loss function:

[0027]

[0028] Where ρ represents the batch size, is the predicted class label, is the true class label.

[0029] In the aforementioned signal modulation classification model based on sample recall incremental learning, the recall module is also used to perform sample recall after performing the following conversion on the input class label and sample label:

[0030] Use one-hot encoding scheme to class labels Encode the class label of the i-th sample Expressed as

[0031]

[0032] Label the samples Converted to a low-dimensional continuous value, the sample label of the i-th sample Expressed as

[0033]

[0034] In the new training task, the embedded latent vector layer generates the corresponding latent variable features based on the converted label information:

[0035] z i =CodeEmbedding(y 1,i ,y 2,i ) (10)

[0036] The recall module recalls the corresponding recall samples based on the generated latent variable features.

[0037] A signal modulation classification incremental learning method based on sample recall. In the Nth training task, the incremental learning method includes:

[0038] The radio signal samples of this training task are input into this model, and the radio signal samples have class labels and sample labels

[0039] The memory module memorizes the original samples of this training task, which include the radio signal samples of this training task and the recalled samples recalled by the recall module;

[0040] The recall module is based on the class labels saved in the N-1th training task Sample Label and the memory module recalls the memory content in the N-1th training task to obtain the recall sample;

[0041] Use the original samples to train the memory recall module, discrimination module, and classification network;

[0042] N is a natural number greater than 1;

[0043] The trained classification network is used to classify the radio signal samples to be classified.

[0044] When a new task arrives, the memory recall module is trained with the new radio signal samples and the recall samples of the old recall module saved last time to obtain a new memory recall module. Then, the new recall module is used to train the classification network with the recall samples of the old task and the new task, or the new recall module is used to train the classification network with the recall samples of the old task and the new task.

[0045] The trained memory module memorizes the original samples of this training again, and saves the currently trained new recall module, waiting for the next task.

[0046] In the above-mentioned incremental learning method for signal modulation classification based on sample recall, the memory process of the memory module is as follows:

[0047] z i =Net1(x i ) (3)

[0048] Among them, x i represents the original sample of the i-th type signal, z i Represents the latent variable characteristics of the i-th type signal;

[0049] Under each training task, save the label information of the original sample of the training task

[0050] When a new task arrives, the recall module recalls the sample as follows:

[0051]

[0052] in, express The class label of express The sample labels are input into the recall module, and the original sample set recalled is:

[0053]

[0054] The new radio signal sample dataset and the recalled sample dataset form a new dataset:

[0055]

[0056] Using the new dataset to train the classification network;

[0057] And the loss function for training the classification network is the cross entropy loss function:

[0058]

[0059] Where ρ represents the batch size, is the predicted class label, is the true class label;

[0060] The loss function for training the recall module is: in is the output of the recall module, x i It is a real sample.

[0061] In the above-mentioned incremental learning method for signal modulation classification based on sample recall, the memory module stores the memorized latent variable features into the embedded latent code vector layer; the embedded latent code vector layer dynamically stores the latent variable features memorized by the memory module based on the label information of the sample; the embedded latent code vector layer is represented as: CodeEmbedding = {z1, z2, ..., z L}, where L represents the length of the latent variable;

[0062] The recall module outputs recall samples based on the identified latent variable features:

[0063]

[0064] In the above-mentioned incremental learning method for signal modulation classification based on sample recall, the recall module uses a one-hot encoding scheme to classify the class labels. Encode the class label of the i-th sample Expressed as

[0065]

[0066] The recall module labels the samples Converted to a low-dimensional continuous value, the sample label of the i-th sample Expressed as

[0067]

[0068] In the new training task, the embedded latent vector layer generates the corresponding latent variable features based on the converted label information:

[0069] z i =CodeEmbedding(y 1,i ,y 2,i ) (10)

[0070] Where i is the index of the sample, and the recall module outputs the recall sample based on the generated latent variable features;

[0071] Through the memory module, the radio signal sample data of each training task is converted into the aggregated posterior distribution q(z) of the potential representative

[0072]

[0073] where q(z) is the encoding distribution, p d (x) is the data distribution, q(z|x) is the posterior distribution;

[0074] The training process of the memory module and the discrimination module is as follows:

[0075] First, update the discrimination module, treat the prior distribution z′ as a real sample, and use the aggregated posterior distribution z as a false sample to update the discrimination module. The loss function is expressed as formula (13):

[0076]

[0077] Then the memory module is updated to confuse the discrimination module, and the loss function is expressed as formula (14):

[0078] l gen = -log(Net3(z)) (14).

[0079] In the above-mentioned incremental learning method for signal modulation classification based on sample recall, in the first training task, the classification network is trained based on the radio signal samples of this training task; the memory module memorizes the radio signal samples of this training task;

[0080] The method further comprises:

[0081] The memory module selectively memorizes samples whose signal-to-noise ratio is higher than a set threshold;

[0082] After the recall module recalls the sample, the data augmentation method of adding Gaussian white noise is used to generate the modulated signal sample in the low signal-to-noise ratio range: in is the signal after adding noise, x is the original signal, a is the noise factor, and w is Gaussian white noise.

[0083] The advantages of the present invention are:

[0084] 1. This solution introduces a memory method based on class labels and sample labels. During the sample memory stage, the memory module learns the samples and generates low-dimensional feature representations. During the sample recall stage, the corresponding samples can be accurately recalled by simply inputting the class labels and sample labels into the recall module. This solution does not require saving sample data, only the trained recall module and sample label information. By using the memory recall module framework as a repository for sample data, it overcomes the problem of directly storing sample data, which requires a large amount of memory.

[0085] 2. This invention proposes a learning approach that combines partial memorization and data augmentation for modulated signal datasets with varying signal-to-noise ratios. When a new task arrives, only high-SNR samples are memorized. When recalling old samples to train the classifier model, data augmentation is used to generate more low-SNR data. This enriched data diversity leads to higher classification performance than previous approaches. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 This is a diagram of the incremental learning framework for radio modulation classification based on sample recall of the present invention;

[0087] Figure 2 This is a structural diagram of the memory recall module of the present invention;

[0088] Figure 3 This is a structural diagram of the identification module of the present invention;

[0089] Figure 4 A structural diagram of the classification network of the present invention;

[0090] Figure 5 This is the classification result diagram of each method in the comparative experiment based on the dataset SignalDatat 1;

[0091] Figure 6 This is the classification result diagram of each method in the comparative experiment based on the dataset SignalData2;

[0092] Figure 7 This is the actual noise effect diagram;

[0093] Figure 8 This is a comparison chart of the classification performance of each method after data enhancement;

[0094] Figure 9 The following is a comparison chart of the classification performance of each method after data enhancement under different signal-to-noise ratios on the test set. DETAILED DESCRIPTION

[0095] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0096] In order to solve the catastrophic forgetting problem of radio signals in the process of incremental learning, this paper proposes a signal modulation classification model and learning method based on sample recall incremental learning. The overall framework is as follows: Figure 1 As shown, it mainly includes memory recall module, discrimination module and classification network.

[0097] In the memory recall module, the memory module Net1 is used to convert the original samples into low-dimensional features and store them in a separate layer - the embedded latent vector layer. By combining the label information saved from the last training task, the recall module Net2 can accurately recall the original samples.

[0098] The discriminative module Net3 is used to regularize the posterior distribution of the data using any prior distribution.

[0099] The classification network ζ() is used to predict the classification results based on sample information, and can also be used to evaluate the final recall effect of the sample.

[0100] When a new task arrives, the memory module Net1 will remember the recalled samples and the new task samples again, and store the low-dimensional features of the recalled samples and the new task samples in a separate layer. In the next task, by combining the label information, the recall module Net2 can accurately recall the remembered samples again.

[0101] Specifically, the following describes the various modules and processes involved in this solution:

[0102] Memory Recall Module

[0103] During the memorization process, the IQ components of the radio signal are used as the input of the network and are expressed as

[0104] I=Re{x(t)},Q=Im{x(t)} (1)

[0105] Where I is the real part of the signal and Q is the imaginary part of the signal, which can be further expressed as

[0106] I=[I1,I2,L,I l ],Q=[Q1,Q2,L,Q l ] (2)

[0107] Where l represents the length of the signal.

[0108] The memory module Net1 converts the signal into a low-dimensional feature representation:

[0109] z i =Net1(x i ) (3)

[0110] where x i and zi Represent the original sample and latent variable features of the i-th type of signal respectively. For the samples of this training task, the original sample is the new radio signal sample of this training task. For the samples of the previous training task, they are the samples that have been memorized. The original sample refers to the sample restored by the recall module in this training task. Through this process, the original data is converted into the aggregated posterior distribution q(z) of the potential representative

[0111]

[0112] where q(z) is the encoding distribution, p d (x) is the data distribution and q(z|x) is the posterior distribution.

[0113] The characteristic of this solution is that it introduces the class label-based and sample labels Memory method, class label Used to represent different modulation categories, sample labels Used to represent each signal sample. These two types of labels are represented as:

[0114]

[0115] Where G and K represent the number of categories and the number of samples respectively.

[0116] Furthermore, this scheme uses a one-hot encoding scheme to class labels Encoding to enable the neural network to better process data, the class label of the i-th sample Expressed as

[0117]

[0118] In addition, this solution introduces the Word2Vec method into the neural network, and converts the sample label into a low-dimensional continuous value by embedding, which can avoid the problem that the vector dimension is too large due to too many samples and too many K, which is not conducive to calculation. Expressed as:

[0119]

[0120] Furthermore, this scheme uses the CodeEmbedding layer (embedded latent vector layer) to store low-dimensional features in the form of a dictionary so that the sample accurately responds to its two labels.

[0121] The CodeEmbedding layer is represented as: CodeEmbedding = {z1, z2, ..., z L}(8), where L represents the length of the latent variable. During the training phase, the CodeEmbeddi layer dynamically updates the low-dimensional features based on the label information.

[0122] The recall module Net2 is a sample recovery model. In the training state of the network, the mean square error is used as the loss function to minimize the reconstruction error between the input and output. The loss function is:

[0123]

[0124] in Is the output of the recall module. In the testing phase or the use phase, it is only necessary to store and The code is sent to the recall module, which automatically converts the input label information into y1 and y2. The CodeEmbedding layer accurately responds to the converted label information and generates the corresponding latent variable features:

[0125] z i =CodeEmbedding(y 1,i ,y 2,i ) (10)

[0126] Where i is the index of the sample, and the recall module recalls the sample based on the identified latent variable features:

[0127]

[0128] Specifically, the memory recall module provided by this solution is as follows: Figure 2 As shown in the figure. The network consists of a convolutional layer, a maximum pooling layer, a fully connected layer, a deconvolution layer, and a CodeEmbedding layer. "Conv" represents the convolutional layer, "Maxpool" represents the maximum pooling layer to compress the size of the input signal, "fc" represents the fully connected layer, "Deconv" represents the deconvolution layer to restore the size of the signal, and "CodeEmbedding" represents the embedded latent code vector layer. " / 2" represents a convolution window with a step of 2 to reduce the height and width of the output, and "2 / " represents a deconvolution window with a step of 2 to increase the height and width of the output. In the sample storage stage, the embedded latent code vector layer dynamically stores the feature vector of the sample according to the label information. In the sample recall stage, the corresponding original signal sample is restored through the input label information.

[0129] Identification module

[0130] In the discrimination module, the prior distribution is assumed to be a normal distribution p(z), which can be expressed as

[0131]

[0132] where σ2 Denotes the variance and μ denotes the mean. Using the adversarial training criterion, the prior distribution p(z) and the aggregated posterior distribution q(z) are matched and regularized, forcing the memory module to transform the data distribution into the desired prior distribution.

[0133] Adversarial training is essentially similar to the theory of GAN to deceive a real-time updated discriminator, but unlike traditional GAN, the input of the discrimination module is not a real sample or a fake sample, but a posterior distribution and a normal prior distribution after data transformation.

[0134] The first step is to update the discrimination module and the prior distribution in the discrimination module Net3 The output of is considered as a true sample, while the output of the posterior distribution z in the discrimination module is considered as a false sample. The loss function is:

[0135]

[0136] The second step is to update the memory module to confuse the discrimination module. The loss function is:

[0137] l gen = -log(Net3(z)) (14)

[0138] The entire adversarial network is trained by alternately repeating the first and second steps. When the entire adversarial network training is completed, the network maps the prior distribution of the data to the posterior distribution. The memory module can better remember the samples, and the recall module will be a deep generative model.

[0139] The identification module Net3 provided by this solution is as follows Figure 3 The network consists of four fully connected layers, three ReLU layers, and a Sigmoid layer. The network determines whether the input is true or false based on the input distribution, and the Sigmoid layer presents the judgment result in the form of probability.

[0140] Classification Network

[0141] The classification network is the main network used to implement radio signal classification, and also uses the IQ of the radio signal as the input of the network. In the first training task, the memory recall module and the discrimination module are trained based on the radio signal samples of this training task. After the training is completed, the trained recall module can be used to train the classification network with the recalled samples of the radio signal samples, or the classification network can be directly trained with the original samples of the radio signal samples. In order to better evaluate the memory recall effect of the memory recall module proposed in this solution, this embodiment uses the former. In general scenarios, the latter can be preferred. The trained memory module will then memorize the radio signal samples of this training task, and save the trained memory recall module and the label information of this training task for the next training task. During each training task, the memory module memorizes the radio signal samples of the current training task and the recalled samples recalled by the recall module. The recall module recalls the original samples based on the information memorized by the memory module from the previous training task and the saved class labels and sample labels. The classification network is trained based on the radio signal samples of the current training task and the recalled recalled samples, i.e., the recalled radio signal samples, thereby achieving simultaneous training of new and old tasks, so as not to forget the old tasks when training new tasks. Of course, the classification network can also be used to evaluate the recall effect of the recall module. As previously mentioned, in order to better evaluate the memory recall effect of the memory recall module proposed in this solution, this embodiment uses the recall module to train the classification network on the recalled samples of old and new samples.

[0142] The classification network provided by this solution is as follows Figure 4 As shown in the figure, it includes a convolutional layer, two residual blocks, an average pooling layer, a maximum pooling layer, a fully connected layer and a softmax layer. Among them, "Maxpool" represents the maximum pooling layer, "Avgpool" represents the average pooling layer, "fc" represents the fully connected layer, "Residual-block1-X" represents the residual block 1, and "Resideal-block2-X" represents the residual block 2. The convolutional layer added by the residual block 2 is used to make the output dimension consistent. " / 2" represents a convolution window with a stride of 2 to reduce the height and width of the output. Finally, the parameters are converted into probability distributions through the softmax layer. Suppose there is a set of labeled datasets for training In addition to the radio signal samples of the current classification task, this training set may also contain recall samples of previous tasks. This scheme uses cross entropy as the training loss function to determine the closeness between the actual output and the expected output:

[0143]

[0144] Where ρ is the batch size. The closer the true value is to the predicted value, the smaller the loss function value.

[0145] In the incremental learning process: after training the previous task, save the label information And the trained recall module Net2 for the next incremental learning. When a new task arrives, the saved label information is input into the recall module Net2, and the original sample is recalled through the recall module Net2

[0146]

[0147] in express The true class label, express The true sample label of , the original sample set recalled is

[0148]

[0149] The new task dataset and the recalled sample dataset form a new dataset, which is expressed as

[0150]

[0151] The dataset will be used to train new and old tasks and will be memorized for use in the next task.

[0152] In practical communication environments, interference from communication devices and the external environment can cause the final transmitted modulated signal to be in a non-ideal state, resulting in modulated signals with varying signal-to-noise ratios at the receiver. For modulated signals with low signal-to-noise ratios, the memory cost is prohibitive. To address this, this solution proposes a partial memory sample selection strategy that only memorizes modulated signal samples with signal-to-noise ratios above a set threshold.

[0153] On this basis, this scheme increases the generalization ability of the classification network through data augmentation after memorizing a small amount of high signal-to-noise ratio data, so as to minimize the burden of memory space for memorized samples. After recalling the samples, the data augmentation method of adding Gaussian white noise is used to generate modulated signal samples in the low signal-to-noise ratio range. The process can be described as follows: in is the signal after adding noise, x is the original signal, a is the noise factor, and w is Gaussian white noise. Data augmentation solves the problem of imbalanced signal-to-noise ratios in samples, improves the generalization ability of the classification model, enables the classifier to recognize signals with low signal-to-noise ratios, and improves the robustness of the classification model.

[0154] To verify the effectiveness of this solution, this embodiment uses the radio signal datasets SingalDatat 1 and SingalData2. The datasets are divided into different task sequences, imitating the incremental process. The impact of the number of incremental categories and the number of sequences on the incremental results is considered, and the effect is evaluated on the same convolutional network classifier. The experimental results show that this solution can effectively solve the catastrophic forgetting problem of reinforcement learning. Some experimental data and experimental procedures are as follows:

[0155] SingalData1 contains 110,000 signal samples, including 11 types of modulation signals: BPSK, QPSK, 8PSK, 16QAM, 64QAM, PAM4, GFSK, CPFSK, B-FM, DSB-AM, and SSB-AM. Each modulation signal has 10,000 samples and a uniform 30dB signal-to-noise ratio. Each signal sample is 1024 bytes long.

[0156] SingalData2 contains 156,000 samples, including 12 modulation types: BPSK, QPSK, 8PSK, OQPSK, 2FSK, 4FSK, 8FSK, 16QAM, 32QAM, 64QAM, 4PAM, and 8PAM. Each modulation type has 13,000 samples and 26 levels of signal-to-noise ratio (SNR) ranging from -20 to 30 dB in 2 dB increments. Each signal sample is 512 bytes long.

[0157] This experiment compares this approach with four different baseline methods: fine-tuning, feature extraction, iCaRL, and joint training. The same learning rate strategy and parameters were used for all methods. All experiments were conducted in PyTorch. The model was trained on a computer with an Intel Core i7-9750H @ 2.60GHz CPU, 32GB of RAM, and a GeForce RTX2080 GPU.

[0158] Fine-turning involves loading a previously trained model for an old task and adding a new fully-connected layer corresponding to the new task. This newly added fully-connected layer is then connected to the fully-connected layer from the old task into a single output. After fine-tuning, the optimized network model is used to classify signals from both the old and new tasks.

[0159] The feature extraction method, Feat.Extraction, is similar to fine-tuning. It loads a previously trained model for the old task and adds new fully-connected layers corresponding to the new task. These added fully-connected layers are concatenated with the fully-connected layers of the old task to form a single output. Unlike fine-tuning, when training a new network using training data for the new task, the parameters of all layers except the newly added fully-connected layers are frozen. The new task passes through the old model and benefits from the complex features learned by the network. Aside from the fully-connected layers, the remaining network structure can be considered a feature extractor.

[0160] Joint Training: When training the current task, the entire network is trained together with data from previous tasks. The advantage of joint training is that it achieves the best results among all classification results and resists catastrophic forgetting of old tasks, but it requires higher learning costs and storage resources.

[0161] iCaRL includes a fixed-size sample storage space. After completing an incremental task, the trained model for the current task is saved. The number of samples per class that can be stored is calculated based on the sample storage space size. The average characteristics of each class are calculated, and samples close to the average characteristics of each class are stored. This allows some new data to be added to the sample storage space and some old data to be removed, ensuring that the stored samples are representative. When a new task arrives, the classification network is trained using all data from the current new class and representative old data from the sample storage space. Knowledge distillation is performed between the old and new models to ensure that the characteristics of the old task are retained as much as possible when learning the new task.

[0162] SignalDatat1:

[0163] This experiment compares the effects of four baselines and the proposed method in the incremental process. The experiment divides the signal dataset SignalDatat1 into different recognition sequences. The task sequences are shown in Table 2. The classification results of each method are shown in Table 2. Figure 5 As shown. Figure 5As can be seen, fine-tuning and feature extraction can lead to catastrophic forgetting, while iCaRL can slightly alleviate catastrophic forgetting, while our method (ours) and joint training achieve competitive results. Although a smaller learning rate is used for old nodes in fine-tuning, small changes in parameters can also undermine the network's ability to handle old tasks, leading to catastrophic forgetting. In the incremental learning paradigm of feature extraction, the ability to learn a new task depends on the characteristics of the previous model. This method only learns feature representations of new classes, resulting in a bias towards new classes and catastrophic forgetting. The joint training paradigm utilizes all new and old data to thoroughly train a new network from a new starting point, so it can maximize the representation of the characteristics of all signals and is often used as an upper limit for incremental learning. The sample recall method proposed in our scheme, however, incurs less cost than joint training while achieving similar performance advantages.

[0164] Table 2 Information of each task sequence

[0165] Task SignalDatat1 1 BPSK, QPSK, 8PSK 2 16QAM, 64QAM 3 4PAM,GFSK 4 CPFSK,B-FM 5 DSB-AM,SSB-AM

[0166] SignalDatat2:

[0167] 1. Partial Memory

[0168] In order to evaluate the effect of sample selection on learning performance, this experiment selected a suitable memory sample range on SignalData2 with different SNR ranges, including 10dB and above, 0dB and above, and 10dB and above. The signal dataset is divided into different recognition sequences, and the task sequences are shown in Table 3. The classification performance of each incremental process is compared according to the task sequence in Table 3. The classification performance of each incremental stage is shown in Table 4. It can be seen that when the SNR threshold is 0dB, the classification accuracy of each incremental step is higher than that of the other two SNR cases. This means that if the SNR is too high, the performance will be affected due to the limited number of samples. However, if the SNR is too low, the performance will be affected by the low SNR samples. Based on the storage cost and the final performance, samples with a SNR of 0dB or above are finally selected as the stored data.

[0169] Table 3 Information of each task sequence

[0170] Task SignalDatat2 1 BPSK, QPSK, 8PSK 2 OQPSK, 2FSK, 4FSK 3 8FSK, 16QAM, 32QAM 4 64QAM, 4PAM, 8PAM

[0171] Table 4 Task accuracy at different signal-to-noise ratio thresholds

[0172] Task -10dB 0dB 10dB 1 73.34 74.21 67.27 2 61.99 69.86 61.68 3 32.61 64.20 59.54 4 38.30 58.34 56.43

[0173] This experiment compares the performance of four baselines and the proposed method in the incremental process on the dataset SignalData2. The classification results of each method are as follows: Figure 6 As shown. Figure 6 As can be seen, both the fine-tuning incremental learning paradigm and the feature extraction incremental learning paradigm still suffer from catastrophic forgetting. The iCaRL incremental learning paradigm, by introducing a distillation loss function and some old samples, can slightly mitigate catastrophic forgetting. The final classification performance of our approach is close to that of joint training. Overall, our approach effectively combats catastrophic forgetting during incremental learning on this dataset.

[0174] 2. Data Augmentation

[0175] For the dataset SignalData2 with various signal-to-noise ratio modulated signals, this scheme proposes a method of storing less parts and data enhancement, which only needs to store some high signal-to-noise ratio signals and generate low signal-to-noise ratio signal data by adding noise and power normalization to the recalled signal samples. The actual noise addition effect is as follows Figure 7 As shown, the left side is the I sequence, the right side is the Q sequence, the blue line is the IQ channel sequence of the original signal, and the red line is the IQ channel sequence after noise addition.

[0176] Using the 10dB and above dataset as memory samples, the final classification performance is demonstrated through data enhancement based on recall. The incremental tasks are still completed according to the task sequence in Table 3. The experimental results are as follows: Figure 8 As shown in the figure, we can see that the classification performance has been significantly improved after data augmentation, and the classification performance of the first and second increments is slightly better than that of joint training, which shows that the data set after data augmentation has greatly enhanced the generalization ability of the model.

[0177] In order to better demonstrate the effect of data augmentation, this scheme compares the classification accuracy of different signal-to-noise ratios on the test set with the final classification models of different incremental methods. Figure 9 As shown by Figure 9 As can be seen, after data augmentation, the recognition rate for low signal-to-noise ratios improves significantly, approaching the effect of joint training for all signal-to-noise ratios. Because iCaRL only retains a subset of characteristic samples, it achieves a modest improvement in recognition rate for high signal-to-noise ratios compared to fine-tuning and feature extraction.

[0178] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope of the appended claims.

Claims

1. A signal modulation classification model based on sample recall incremental learning, characterized in that: The system comprises a memory recall module, a discrimination module and a classification network. The memory recall module comprises a memory module and a recall module spliced ​​in sequence. The memory module is used to memorize original samples, and the recall module is used to recall samples memorized by the memory module. In the class incremental training task, the new radio signal samples and the recalled samples recalled by the recall module are used as sample data of the model to train the model. The discrimination module is used to form an adversarial network with the memory recall module to train the memory recall module and the discrimination module. The classification network is used to output classification results based on its input. The memory recall module includes an embedded latent code vector layer embedded in the memory module and the recall module; The memory module is used to convert the original sample into low-dimensional features for memory: z i =Net1(x i ) (3) Among them, x i represents the original sample of the i-th type signal, z i Represents the latent variable characteristics of the i-th type signal; In each training task, the label information of the original sample of the training task is saved; at the same time, the embedded latent code vector layer dynamically stores the latent variable features of the memory module based on the label information of the sample CodeEmbedding={z1,z2,...,z L }, the label information of each sample includes the class label and sample labels The recall module recalls the corresponding recall samples based on the latent variable features of the memory: The loss function used to train the recall module is: is the recall sample obtained by the recall module for the i-th type of signal, x i is the original sample of the i-th type signal.

2. The signal modulation classification model based on sample recall incremental learning according to claim 1, characterized in that In the first training task, the original samples only include the radio signal samples of this training task, and the memory recall module and the discrimination module are trained based on the radio signal samples of this training task; The classification network is trained based on the radio signal samples of this training task or is trained based on the recall module on the recall samples of the radio signal samples; In the incremental training phase of a non-first task, the original samples include the radio signal samples of the current training task and the recalled samples recalled by the recall module; the input of the classification network includes the recalled samples recalled by the recall module and the radio signal samples of the current training task or includes the recalled samples of the recall module for old samples and new samples. In the testing or use phase, the input of the classification network is the radio signal samples to be classified.

3. The signal modulation classification model based on sample recall incremental learning according to claim 1, characterized in that The adversarial network composed of the discrimination module and the memory recall module includes the following loss functions: L gen =-log(Net3(z)) (14) The loss function (13) is used to update the discrimination module, and the loss function (14) is used to update the memory module; The classification network includes a cross entropy loss function: Where ρ represents the batch size, is the predicted class label, is the true class label.

4. The signal modulation classification model based on sample recall incremental learning according to claim 3, characterized in that Recall module class labels for input and sample labels Perform sample recall after the following transformation: Use one-hot encoding scheme to class labels Encode the class label of the i-th sample Expressed as Label the samples Converted to a low-dimensional continuous value, the sample label of the i-th sample Expressed as In the new training task, the embedded latent vector layer generates the corresponding latent variable features based on the converted label information z i =CodeEmbedding(y 1,i ,the 2,i ) (10) The recall module recalls the corresponding recall samples based on the generated latent variable features.

5. A signal modulation classification incremental learning method based on sample recall, characterized in that: In the Nth training task, the incremental learning method includes: The radio signal samples of this training task are input into this model, and the radio signal samples have class labels and sample labels The memory module memorizes the original samples of this training task, which include the radio signal samples of this training task and the recalled samples recalled by the recall module; The recall module is based on the class labels saved in the N-1th training task Sample Label and recalling the memory content of the memory module to obtain the recall sample; Use the original samples of this training to train the memory recall module and the discrimination module; use the radio signal samples and recall samples of this training task to train the classification network, or use the recall module to train the classification network with the recall samples of old samples and new samples; N is a natural number greater than 1; The trained classification network is used to classify the radio signal samples to be classified.

6. The incremental learning method for signal modulation classification based on sample recall according to claim 5, characterized in that The memory process of the memory module is as follows: z i =Net1(x i ) (3) Among them, x i represents the original sample of the i-th type signal, z i Represents the latent variable characteristics of the i-th type signal; Under each training task, save the label information of the original sample of the training task When a new task arrives, the recall module recalls the sample as follows: in, express The class label of express The sample labels are input into the recall module, and the original sample set recalled is: The new radio signal sample dataset and the recalled sample dataset form a new dataset: Using the new dataset to train the classification network; And the loss function for training the classification network is the cross entropy loss function: Where ρ represents the batch size, is the predicted class label, is the true class label; The loss function for training the recall module is: in is the output of the recall module, x i It is a real sample.

7. The incremental learning method for signal modulation classification based on sample recall according to claim 6, characterized in that: The memory module stores the latent variable features of the memory into the embedded latent code vector layer; the embedded latent code vector layer dynamically stores the latent variable features of the memory module based on the label information of the sample, which is expressed as: CodeEmbedding = {z1, z2, ..., z L }, where L represents the length of the latent variable; The recall module outputs recall samples based on the latent variable features of memory:

8. The incremental learning method for signal modulation classification based on sample recall according to claim 7, characterized in that: The recall module uses a one-hot encoding scheme to encode class labels Encode the class label of the i-th sample Expressed as The recall module labels the samples Converted to a low-dimensional continuous value, the sample label of the i-th sample Expressed as In the new training task, the embedded latent vector layer generates the corresponding latent variable features based on the converted label information: Where i is the index of the sample, and the recall module outputs the recall sample based on the generated latent variable features; Through the memory module, the radio signal sample data of each training task is converted into the aggregated posterior distribution q(z) of the potential representative where q(z) is the encoding distribution, p d (x) is the data distribution, q(z|x) is the posterior distribution; The training process of the memory module and the discrimination module is as follows: First, update the discrimination module, treat the prior distribution z′ as a real sample, and use the aggregated posterior distribution z as a false sample to update the discrimination module. The loss function is expressed as formula (13): Then the memory module is updated to confuse the discrimination module, and the loss function is expressed as formula (14): L gen =-log(Net3(z)) (14)。 9. The signal modulation classification incremental learning method based on sample recall according to any one of claims 5 to 8, characterized in that: In the first training task, the classification network is trained based on the radio signal samples of this training task; The memory module memorizes the radio signal samples of this training task; The method further comprises: The memory module selectively memorizes samples whose signal-to-noise ratio is higher than a set threshold; After the recall module recalls the sample, the data augmentation method of adding Gaussian white noise is used to generate the modulated signal sample in the low signal-to-noise ratio range: in is the signal after adding noise, x is the original signal, a is the noise factor, and w is Gaussian white noise.

Images