Small sample increment automatic modulation identification method based on confrontation margin and correction prototype

Abstract

The invention provides a small sample increment automatic modulation recognition method based on confrontation margin and a modified prototype, which comprises the following steps: optimizing margin for a recognition model in a basic training stage, so that the recognition model can give consideration to adaptability and stability of new and old categories during incremental learning, and meanwhile, the recognition capability of the new category is enhanced; and in the incremental learning stage, the class prototype of the new class is dynamically adjusted, so that the deviation caused by insufficient samples is reduced, and the representation accuracy of the new class is improved. The invention aims to solve the challenge of how to balance the stability of the known modulation type and the learning efficiency of the new modulation type in small sample incremental learning in the existing method, so that the generalization ability and accuracy of the model are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of communication signal classification, and particularly relates to a few-shot incremental automatic modulation recognition method based on adversarial margin and corrected prototypes. Background Art

[0002] With the rapid development of wireless communication technology, its application scenarios have become increasingly rich, covering multiple fields such as civil and public safety. Automatic modulation recognition (AMR), as a key technology among them, plays a crucial role in ensuring the normal operation and security of communication systems.

[0003] In the civil field, AMR helps to achieve efficient spectrum utilization and management. By accurately identifying the modulation modes of communication signals, spectrum resources can be better allocated, communication efficiency can be improved, and the growing communication needs of people can be met. In the field of public safety, such as in applications like cognitive radio and reconnaissance, electronic countermeasures, and spectrum sensing, AMR can timely detect abnormal signals, prevent unauthorized device intrusion, reduce the risks of malicious attacks and data leakage, and thus ensure the reliable operation of the Internet of Things system.

[0004] The emergence of deep learning technology has brought new development opportunities for AMR. Deep learning networks, such as convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and Transformers, can effectively utilize the spatial and temporal correlations contained in signals by virtue of their powerful feature representation capabilities. These networks can automatically learn the features of signals without the need for complex manual feature extraction, thereby improving the accuracy of automatic modulation recognition.

[0005] However, most traditional AMR methods are based on the closed-set scenario assumption, that is, it is considered that all possible modulation types are known during the training stage, and the training data contains samples of these modulation types. This assumption has great limitations in the actual complex communication environment. In the actual communication environment, new modulation types often appear after the training is completed, and the number of their samples is extremely limited, which makes the model based on the closed-set assumption unable to handle this situation.

[0006] To address the limitations of closed-set recognition methods, the Few-Shot Incremental Automatic Modulation Recognition (FSI-AMR) technology has emerged. Few-Shot Incremental Learning (FSIL) is a technology that enables a model to maintain its learning ability for new classes and prevent forgetting of existing classes even with only a very small number of new samples. In automatic modulation recognition, this means that the model can quickly adapt and accurately identify the modulation type when continuously receiving new and limited modulation signals. Different from traditional methods, few-shot incremental learning can solve the problem of continuously emerging new modulation types in practical applications while retaining and consolidating the information of the old classes that have been learned. However, few-shot incremental automatic modulation recognition faces many challenges. The most prominent challenges are catastrophic forgetting and the adaptability to new classes. In few-shot incremental automatic modulation recognition, catastrophic forgetting refers to the situation where the model may forget the knowledge of existing classes when learning new classes, resulting in a decrease in the recognition accuracy of old modulation types. This problem is particularly prominent in the incremental learning process because the model needs to continuously learn new modulation types while ensuring that its classification ability for old classes is not affected. In addition, how to efficiently process a very small number of new samples is also a huge challenge. In practical applications, the samples of new modulation types are often extremely limited, making it difficult for traditional deep learning methods to effectively learn the features of new classes. Deep neural networks usually require a large amount of labeled data for effective training, but in the incremental learning scenario, the model must be effectively trained and generalized with only a few samples. How to extract sufficient class features from a small number of samples and ensure the adaptability of the model to new classes is the key problem faced by the few-shot incremental automatic modulation recognition technology.

[0007] The prior art provides the following technical solutions:

[0008] The literature [K. Zhu, Y. Cao, W. Zhai, J. Cheng, and Z.-J. Zha, "Self-promoted prototype refinement for few-shot class-incremental learning," in Proceedings of the IEEE / CVF conference on computer vision and pattern recognition, 2021, pp. 6801–6810] uses the prototype learning method to generate prototypes of new classes through a small number of new samples and classify them by measuring the similarity between the interference features and the prototypes. This method has been widely used in few-shot class incremental learning (FSCIL). By generating class prototypes of new classes, it can perform effective incremental learning with extremely few samples.

[0009] The literature [D.-W. Zhou, F.-Y. Wang, H.-J. Ye, L. Ma, S. Pu, and D.-C. Zhan, "Forward compatible few-shot class-incremental learning," in Proceedings of the IEEE / CVF conference on computer vision and pattern recognition, 2022, pp. 9046–9056] uses the feature subspace optimization method to strive to preset or learn a highly scalable feature space for future new classes in order to better support the learning of unknown classes. The literature proposes a forward compatible training method to reserve space for multiple incremental prototypes to enhance the model's adaptability to new classes.

[0010] The literature [B.Yang, M.Lin, Y.Zhang, B.Liu, X.Liang, R.Ji, and Q.Ye, “Dynamic support network for few-shot class incremental learning,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.45, no.3, pp.2945–2951, 2022] uses a dynamic network structure update method to automatically adjust the network architecture according to the input data features, providing higher flexibility and reducing the risk of overfitting. This method helps the model effectively learn new categories in incremental learning while maintaining stability for old categories by dynamically adding or compressing network nodes.

[0011] The literature [H.Liu, L.Gu, Z.Chi, Y.Wang, Y.Yu, J.Chen, and J.Tang, “Few-shot class-incremental learning via entropy-regularized data-free replay,” in European Conference on Computer Vision. Springer, 2022, pp.146–162] uses a data replay method to mitigate the catastrophic forgetting problem by using episodic memory to replay samples of old tasks. It includes two ways: generative replay and direct replay. Among them, generative replay uses a generative model to fit the data distribution of old tasks, while direct replay directly uses typical samples of old tasks for replay.

[0012] However, the defects of the above existing technologies are as follows:

[0013] Although the prototype learning method can generate prototypes of new categories, when dealing with complex relationships between categories, especially when there are fuzzy boundaries between categories, the accuracy of the prototypes is still problematic. The update of class prototypes depends on a small number of samples, which may lead to biases in the new class prototypes and affect the model's ability to recognize modulation signals of new categories.

[0014] The feature subspace optimization method provides better scalability for the model, but it usually assumes sufficient prior knowledge to preset the feature space. In practical applications, especially when unknown modulation types appear, the expansion ability of the feature space may not be as expected. In addition, this method may not be able to operate effectively in the case of insufficient samples.

[0015] The dynamic network structure update method has high flexibility, but the process of adjusting the network structure will increase the computational overhead. Moreover, in some applications, the dynamic changes in the network may lead to an overly complex model, thereby affecting the training and inference efficiency.

[0016] The data replay method avoids catastrophic forgetting by replaying samples of old tasks. However, these methods are usually designed for 2D image data, and when applied to 1D communication signals, there may be problems of incompatibility and misadaptation. In addition, the pseudo-samples generated by generative replay may not be able to fully simulate the complex characteristics of communication signals, resulting in a large difference between the generated signal samples and the real communication signals, affecting the generalization ability of the model, and thus reducing the accuracy of classifying communication signals. Summary of the Invention

[0017] To solve the above problems existing in the prior art, the present invention provides a few-shot incremental automatic modulation recognition method based on adversarial margin and corrected prototypes. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0018] A few-shot incremental automatic modulation recognition method based on adversarial margin and corrected prototypes includes:

[0019] S100, obtaining a first modulation signal of multiple known modulation categories and a second modulation signal of a modulation category different from that of the first modulation signal;

[0020] S200, in the basic training stage, using all the first modulation signals to train a preset recognition model to obtain a pre-trained recognition model;

[0021] S300, in the incremental learning stage, freezing the parameters of the pre-trained recognition model for the adjustment signals of the basic type, and using the second modulation signal and the first modulation signal to perform additional training on the pre-trained model with frozen parameters to obtain a corrected recognition model;

[0022] S400, using the corrected recognition model to classify and recognize the modulation signal to be recognized obtained from the communication system to obtain the category to which the modulation signal to be recognized belongs.

[0023] Advantageous Effects:

[0024] 1. The recognition model used in this application can distinguish different types of modulation signals by introducing a transfer classifier and a discriminative classifier. The transfer classifier uses a negative margin to enhance the adaptability of new classes, while the discriminative classifier uses a positive margin to strengthen the discrimination ability for old classes. Through adversarial learning, the recognition model can avoid catastrophic forgetting while learning new classes and maintain the recognition ability for old classes. This application effectively balances the model's rapid adaptation to new classes and memory of old classes, enabling it to maintain good classification performance during incremental learning.

[0025] 2. This application aims to solve the problem of inaccurate class prototype update in incremental learning. By weighting samples, the class prototypes of new classes are corrected to reduce the negative impact of a small number of samples on prototype update. This application calculates the similarity between each sample and the class prototype and uses similarity weighting to calculate the class prototype, thereby enhancing the model's recognition ability in the case of small samples, ensuring that the class prototypes of new classes are more accurate, effectively improving the recognition accuracy in the incremental learning stage, and maintaining good adaptability to new and old classes.

[0026] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a schematic flowchart of a small-sample incremental automatic modulation recognition method based on adversarial margin and corrected prototype provided by the present invention;

[0028] Figure 2 is a schematic diagram of the process in the basic training stage provided by the present invention;

[0029] Figure 3 is a schematic diagram of the process in the incremental learning stage provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] The present invention will be further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0031] The object of the present invention is to enable the model to balance the adaptability and stability of new and old classes during incremental learning by optimizing the margin through adversarial margin-driven training, while enhancing the recognition ability of new classes; and the corrected prototype learning reduces the deviation caused by insufficient samples by dynamically adjusting the class prototypes of new classes, improving the representation accuracy of new classes. This method aims to address the challenge of how to balance the stability of known modulation types and the learning efficiency of new modulation types in small-sample incremental learning, thereby improving the generalization ability and accuracy of the model.

[0032] Such as Figure 1As shown in the figure, the present invention provides a few-shot incremental automatic modulation recognition method based on adversarial margin and corrected prototypes, including:

[0033] S100, obtaining first modulation signals of multiple known modulation categories and second modulation signals of modulation categories different from those of the first modulation signals;

[0034] It should be noted that the second modulation signals are few-shot samples, and the quantity is far less than that of the first modulation signals.

[0035] S200, in the basic training stage, using all the first modulation signals to train a preset recognition model to obtain a pre-trained recognition model; wherein, the recognition model includes a feature extractor, a transfer classifier, and a discriminative classifier connected in sequence.

[0036] S300, in the incremental learning stage, freezing the parameters of the pre-trained recognition model for the adjustment signals of the basic types, and using the second modulation signals and the first modulation signals to perform additional training on the pre-trained model with frozen parameters to obtain a corrected recognition model;

[0037] S400, using the corrected recognition model to classify and recognize the modulation signals to be recognized obtained from the communication system to obtain the category to which the modulation signals to be recognized belong.

[0038] In a specific implementation manner of the present invention, referring to Figure 2 , S200 includes:

[0039] S210, in the basic training stage, inputting the first modulation signals into the feature extractor for feature extraction to obtain the feature embeddings of the first modulation signals;

[0040] S220, inputting the feature embeddings of the first modulation signals into the transfer classifier to extract shared features to obtain feature embeddings;

[0041] S230, inputting the feature embeddings output by the transfer classifier into the discriminative classifier to obtain the predicted category to which the first modulation signals belong;

[0042] S240, according to the modulation categories and predicted categories of the first modulation signals, calculating the loss value of the total loss function;

[0043] Referring to Figure 2 , the input is K types of known modulation type signals, and a 1D residual neural network (a variant of ResNet18) is used for feature extraction. The input signal is an IQ signal (2x1024 array), which is processed through multiple convolutional layers and residual blocks to extract signal features. After feature extraction, the model generates feature embeddings and performs classification through fully connected layers. Two classifiers are used: a transfer classifier and a discriminative classifier a transfer classifier uses a negative margin to help the model extract shared features during the training of the base classes, thereby enhancing the transfer ability to new classes. The discriminative classifier uses a positive margin to ensure the high discriminative ability of the model for the base classes, thereby avoiding overfitting and forgetting. Based on f θ (x) performs the AMC task. Then, f θ (x) generates the feature embedding of the discriminative classifier through the fully connected layer g(·). Therefore, Based on g(f (x)) performs the AMC task. θ (x)) performs the AMC task.

[0044] The total loss function includes a transfer learning loss and a discriminative learning loss which is expressed by the formula:

[0045]

[0046] In the formula, is the total loss function, is the transfer learning loss, is the discriminative learning loss, x i is the input modulated signal, y i is the true modulation class label of x i , and represent the linear classification weights of the transfer classifier i and the discriminative classifier corresponding to the y class, f(x i ) is the feature embedding extracted from x i by the feature extractor, θ represents the trainable parameters of the recognition model, f θ (x i ) represents the feature embedding calculated through the trainable parameters, τ is the temperature parameter used to adjust the smoothness of the softmax function, m neg and m pos are hyperparameters representing the negative margin and the positive margin, λ is the weight parameter controlling the relative importance of the two parts in the transfer learning loss and the discriminative learning loss, and g(·) represents the fully connected layer.

[0047] S250, updates the trainable parameters of the feature extractor, the transfer classifier, and the discriminative classifier using the loss value;

[0048] S260. Repeat S210 to S250 until the basic training cutoff condition is reached to obtain a pre-trained recognition model.

[0049] In a specific embodiment of the present invention, referring to Figure 3 , S300 includes:

[0050] S310. In the incremental learning stage, freeze the trainable parameters of the feature extractor and the transfer classifier in the pre-trained recognition model.

[0051] S320. First step, use the first modulation signal to pass through the pre-trained recognition model with frozen trainable parameters, calculate the first modulation prototype of the first modulation signals belonging to the same type, and use it as the coefficient of the learnable parameters of the decision classifier; second step, use the second modulation signal to pass through the pre-trained recognition model with frozen trainable parameters, calculate the second modulation prototype of the second modulation signals belonging to the same type, and use it as the coefficient of the learnable parameters of the decision classifier; third step, re-select the second modulation signal to perform incremental training on the decision classifier in the pre-trained recognition model with frozen trainable parameters to obtain a corrected recognition model.

[0052] In a specific embodiment of the present invention, the first step in S320 includes:

[0053] S321a. Input the first modulation signal into the feature extractor with frozen trainable parameters to obtain the feature embedding of the input first modulation signal, and obtain the first feature representation through the fully connected layer in the transfer classifier.

[0054] S322a. Use the first feature representation to calculate the first modulation prototype of the first feature representations belonging to the same type.

[0055] S323a. Assign the first modulation prototype to the coefficient of the trainable parameters of the decision classifier.

[0056] Referring to Figure 2, in the incremental stage, the model faces new modulation types (there is little data for these new categories, usually few-shot). At this time, the training set consists of the previous base categories and the new incremental categories, with a total of K' kinds of signals. To avoid catastrophic forgetting, during incremental training, the model freezes the feature extraction network and classifier for the base categories (i.e., freezes the parameters for the base categories) and only adjusts the part for the new categories. To address the problem of few samples for the new categories, Rectified Prototypical Learning is used. For each new category, the model calculates its class prototype and rectifies it to make the prototype more accurate and reduce the influence of abnormal samples. The influence of typical samples is enhanced through weighting, and the influence of noise samples is reduced.

[0057] Calculating the class prototype: In the incremental stage, when samples of the new category arrive, first generate the feature embedding of each sample through the network. Then, calculate the prototype of each category, usually the mean of the feature embeddings of all samples in that category. That is, the first modulation prototype in S322a is expressed by the formula:

[0058]

[0059] where μ t,k represents the class prototype of category k in the t-th incremental learning stage, N t,k represents the number of samples of category k in the t-th incremental learning stage, x t,n represents the input data of the n-th sample in the t-th incremental learning stage, represents the feature embedding extracted from x by the feature extractor t,n with parameters θ t being the learnable parameters trained in the t-th incremental learning stage for generating feature representations, denotes the feature representation output by the fully connected layer of the transfer classifier.

[0060] The second step in S320 includes:

[0061] S321b, input the second modulation signal into the feature extractor with frozen trainable parameters to obtain the feature embedding of the input second modulation signal, and obtain the second feature representation through the fully connected layer in the transfer classifier;

[0062] S322b, use the second feature representation and through weighted calculation, obtain the second modulation prototype of the second modulation signals belonging to the same type;

[0063] S323c, taking the second modulation prototype as the coefficient of the trainable parameter of the decision classifier.

[0064] For new classes, since there are only a few samples, directly calculating the mean prototype may be affected by abnormal samples. To reduce the influence of noisy samples, the modified prototype learning method weights the similarity of each sample.

[0065] In a specific embodiment of the present invention, S322b includes:

[0066] S322b1, calculating the similarity between the input second modulation signal and other modulation signals using the second feature representation;

[0067] By calculating the similarity between each sample and other samples of the same class (Euclidean distance based on feature embedding) to determine its "typicality" (the more typical a sample is, the more it is considered to represent the class), which is expressed by the formula:

[0068]

[0069] In the formula, g(f θ (x i )) represents the second feature representation output by the fully connected layer of the transfer classifier for f θ (x i ), g(f θ (x j )) represents the second feature representation output by the fully connected layer of the transfer classifier for f θ (x j ), f θ (x i ) represents the feature embedding extracted from the input second modulation signal x by the feature extractor i , f θ (x j ) represents the feature embedding extracted from the input second modulation signal x by the feature extractor j , and N represents the number of other modulation signals;

[0070] S322b2, calculating the contribution of the input second modulation signal to the modulation type using the similarity to obtain a weighting coefficient;

[0071] Using these similarity scores to weight the contribution of each sample to the calculation of the class prototype to obtain a weighting coefficient which is expressed as:

[0072]

[0073] S322b3, calculate the second modulation prototype of the second modulation signals belonging to the same type by using the weighting coefficient;

[0074] Use the weighted similarity value to update the class prototype to obtain a more robust prototype representation, expressed as:

[0075]

[0076] In a specific embodiment of the present invention, the third step in S320 includes:

[0077] S321c, reselect the second modulation signal and input it into the feature extractor with frozen trainable parameters to obtain the feature embedding of the second modulation signal input, and obtain the predicted category to which the second modulation signal belongs through the transfer classifier and the discriminative classifier;

[0078] S322c, calculate the loss value of the cross-entropy loss function by using the predicted category to which the second modulation signal belongs in S327 and the true label category of the reselected second modulation signal, and update the learnable parameters of the discriminative classifier by using the loss value;

[0079] S323c, repeat S321c - S322c until the incremental training cut-off condition is reached to obtain the corrected recognition model.

[0080] In a specific embodiment of the present invention, S400 includes:

[0081] Input the modulation signal to be recognized obtained from the communication system into the corrected recognition model, so that the corrected recognition model selects the modulation category corresponding to the modulation prototype with the minimum distance among the distances between the modulation signal to be recognized and the known modulation prototypes as the predicted category to which the modulation signal to be recognized belongs.

[0082] In the application stage, input the modulation signal x to be recognized t into the corrected recognition model in step 2 to obtain the feature embedding calculate its minimum distance to all known modulation prototypes μ t,k denoted as dis t :

[0083]

[0084] Select the modulation type to which the modulation signal belongs when the minimum distance is selected as the category of the modulation signal to be recognized.

[0085] Illustrate the advantages of this application by comparing with the prior art.

[0086] Compared with Document 1 in the prior art, the method of the present invention can be regarded as an upgrade of the prototype learning method. Traditional prototype learning methods generate class prototypes and use these prototypes for classification. However, in dealing with few-shot or incremental learning scenarios, they are easily affected by the limitations of the number of samples and inaccurate updates of class prototypes. Compared with the prototype learning method, the present invention not only generates prototypes of new classes but also effectively balances the flexibility and stability of the model by introducing an adversarial margin-driven method. Traditional prototype learning methods may be affected by few samples, resulting in inaccurate updates of class prototypes, thus affecting the learning effect of new classes. Through adversarial margin guidance, the present invention can reduce catastrophic forgetting, ensure the improvement of adaptability to new classes and memory ability to old classes during the incremental learning process, and thus enhance the overall recognition effect.

[0087] Compared with Document 2 in the prior art, the present invention optimizes the class prototype representation by modifying the prototype learning method, avoiding the limitations of relying on a preset or virtual space. Traditional feature subspace optimization methods reserve space for virtual incremental prototypes, but these methods may not be able to fully adapt to the dynamic changes in the real world and the emergence of unknown classes. The present invention adjusts the class prototype by weighting samples, ensuring that the class prototypes of new classes are more accurate, and at the same time avoiding over-reliance on virtual prototypes to enhance the scalability of the model.

[0088] Compared with Document 3 in the prior art, the adversarial margin-driven method and the modified prototype learning method of the present invention avoid the computational burden and overfitting problems caused by frequent adjustments of the network architecture. Although dynamic network structure updates provide flexibility, they may lead to increased computational overhead and are difficult to deploy in real-time applications. In contrast, the present invention can effectively handle incremental learning tasks and reduce the computational complexity without significantly adjusting the network structure by optimizing the prototype representation and using the adversarial margin method.

[0089] Compared with Document 4 in the prior art, the present invention does not rely on the generated or directly replayed old task samples, but reduces catastrophic forgetting and more effectively responds to the rapid learning of new classes through adversarial margins and modified prototype learning. Although data replay methods can alleviate the forgetting problem, they often rely on generated pseudo-data or old samples, which may not accurately represent new modulation types. The present invention ensures the balance between new classes and old classes through a more reasonable prototype update mechanism, avoiding the potential problems of generating pseudo-samples.

[0090] It should be noted that the terms "first" and "second" in the present invention are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality of" means two or more, unless otherwise specifically defined.

[0091] Although the present application has been described in conjunction with various embodiments herein, however, in the process of implementing the claimed present application, those skilled in the art can understand and achieve other variations of the disclosed embodiments by viewing the accompanying drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "one" does not exclude a plurality of cases.

[0092] The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A small sample increment automatic modulation recognition method based on adversarial margin and modified prototype, characterized in that: include: S100, obtaining a plurality of first modulation signals of known modulation types and a second modulation signal of a modulation type different from the first modulation signal; S200, in a basic training stage, using all first modulated signals to train a preset recognition model to obtain a pre-trained recognition model; S300, in the incremental learning stage, freezing the parameters of the adjustment signal for the basic type in the pre-trained recognition model, and performing additional training on the pre-trained model of the frozen parameters using the second modulation signal and the first modulation signal to obtain a modified recognition model; S400, using the modified recognition model to classify and recognize the modulated signal to be recognized obtained from the communication system, to obtain the category to which the modulated signal to be recognized belongs.

2. The small sample increment automatic modulation recognition method based on adversarial margin and modified prototype according to claim 1 is characterized in that: The recognition model includes a feature extractor, a transfer classifier and a discriminative classifier connected in sequence.

3. The small sample increment automatic modulation recognition method based on adversarial margin and modified prototype according to claim 2 is characterized in that: S200 includes: S210, in a basic training stage, inputting the first modulated signal into the feature extractor to perform feature extraction, thereby obtaining feature embedding of the first modulated signal; S220, inputting the feature embedding of the first modulated signal into the transferability classifier to extract shared features to obtain feature embedding; S230, embedding the feature output by the transfer classifier into the discriminative classifier to obtain the predicted category of the first modulated signal; S240, calculating a loss value of a total loss function according to a modulation category and a predicted category of the first modulated signal; S250, using the loss value to update the trainable parameters of the feature extractor, the transfer classifier, and the discriminative classifier; S260, repeat S210 to S250 until the basic training cutoff condition is reached to obtain a pre-training recognition model.

4. The small sample increment automatic modulation recognition method based on adversarial margin and modified prototype according to claim 3 is characterized in that: The total loss function includes transferable learning loss and discriminative learning loss, which is expressed as: In the formula, is the total loss function, is the transferable learning loss, is the discriminative learning loss, x i is the input modulation signal, y i For x i The true modulation class label, and Indicates that i Class-corresponding transfer classifier and discriminative classifier The linear classification weight, f(x i ) is the feature extractor for x i The extracted features are embedded, θ represents the trainable parameters of the recognition model, f θ (x i ) represents the feature embedding calculated by the trainable parameters, τ is the temperature parameter used to adjust the smoothness of the softmax function, and m neg and m pos is a hyperparameter, representing the negative margin and negative margin, λ is a weight parameter that controls the relative importance of the two parts in transferable learning loss and discriminative learning loss, and g(·) represents a fully connected layer.

5. The small sample increment automatic modulation recognition method based on adversarial margin and modified prototype according to claim 2 is characterized in that: S300 includes: S310, freezing the trainable parameters of the feature extractor and the transfer classifier in the pre-trained recognition model in the incremental learning stage; S320, the first step is to use the first modulated signal to calculate the first modulation prototype of the first modulated signal belonging to the same type through the pre-trained recognition model with frozen trainable parameters, and use it as the coefficient of the learnable parameter of the decision classifier; the second step is to use the second modulated signal to calculate the second modulation prototype of the second modulated signal belonging to the same type through the pre-trained recognition model with frozen trainable parameters, and use it as the coefficient of the learnable parameter of the decision classifier; the third step is to reselect the second modulated signal to perform incremental training on the decision classifier in the pre-trained recognition model with frozen trainable parameters to obtain a revised recognition model.

6. The small sample increment automatic modulation recognition method based on adversarial margin and modified prototype according to claim 5 is characterized in that: The first step in S320 includes: S321a, inputting the first modulated signal into a feature extractor with frozen trainable parameters to obtain a feature embedding of the first modulated signal, and obtaining a first feature representation through a fully connected layer in a transfer classifier; S322a, using the first feature representation to calculate a first modulation prototype of a first feature representation belonging to the same type; S323a, assigning the first modulation prototype to the coefficient of the trainable parameter of the decision classifier; The second step in S320 includes: S321b, inputting the second modulated signal into a feature extractor with frozen trainable parameters to obtain a feature embedding of the second modulated signal, and obtaining a second feature representation through a fully connected layer in a transfer classifier; S322b, using the second characteristic representation and performing weighted calculation to obtain a second modulation prototype of a second modulation signal of the same type; S323c: Use the second modulation prototype as a coefficient of a trainable parameter of the decision classifier.

7. The small sample increment automatic modulation recognition method based on adversarial margin and modified prototype according to claim 6 is characterized in that: The first modulation prototype in S322a is expressed by the formula: In the formula, μ t,k represents the class prototype of category k in the tth incremental learning stage, N t,k represents the number of samples of category k in the tth incremental learning phase, x t,n represents the input data of the nth sample in the tth incremental learning phase, Represents the feature extractor From x t,n The feature embedding extracted from t is a learnable parameter trained in the incremental learning phase t to generate feature representations, express Feature representation of the output of the fully connected layer of the transfer classifier.

8. The small sample increment automatic modulation recognition method based on adversarial margin and modified prototype according to claim 6, characterized in that: S322b includes: S322b1, using the second feature representation to calculate the similarity between the input second modulation signal and other modulation signals, expressed as: In the formula, g(f θ (x i )) means f θ (x i ) is the second feature representation output by the fully connected layer of the transfer classifier, g(f θ (x j )) means f θ (x j ) The second feature representation output by the fully connected layer of the transfer classifier, f θ (x i ) indicates that the feature extractor The second modulated signal x is input i The feature embedding extracted from θ (x j ) indicates that the feature extractor The second modulated signal x is input j The feature embedding extracted from , N represents the number of other modulated signals; S322b2, using the similarity to calculate the contribution of the input second modulation signal to the modulation type, to obtain a weighting coefficient, expressed as: S322b3, using the weighted coefficient to calculate a second modulation prototype of a second modulation signal of the same type, expressed as:

9. The small sample increment automatic modulation recognition method based on adversarial margin and modified prototype according to claim 5, characterized in that: The third step in S320 includes: S321c, reselecting the second modulated signal and inputting it into the feature extractor of the frozen trainable parameters to obtain feature embedding of the second modulated signal, and obtaining the predicted category of the second modulated signal through the transfer classifier and the decision classifier; S322c, using the predicted category of S327 and the real label category of the re-selected second modulated signal, calculating the loss value of the cross entropy loss function, and using the loss value to update the learnable parameters of the decision classifier; S323c, repeat S321c-S322c until the incremental training cutoff condition is reached to obtain a revised recognition model.

10. The small sample increment automatic modulation recognition method based on adversarial margin and modified prototype according to claim 1, characterized in that: S400 includes: The modulation signal to be identified obtained from the communication system is input into the modified recognition model so that the modified recognition model selects the modulation category corresponding to the modulation prototype with the smallest distance between the modulation signal to be identified and the known modulation prototype as the predicted category to which the modulation signal to be identified belongs.

Images