Automatic modulation recognition method and device based on consistency constraint and feature prototype
By introducing consistency constraints and feature prototypes into the deep learning model, the accuracy of modulated signal recognition in low signal-to-noise ratio environments is improved, solving the problem of relying on expert features and thresholds in existing technologies, and achieving efficient signal classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2024-09-24
- Publication Date
- 2026-06-23
Smart Images
Figure CN119276667B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of automatic modulation recognition technology, specifically relating to an automatic modulation recognition method and apparatus based on consistency constraints and feature prototypes. Background Technology
[0002] With the rapid development of information technology, the requirements for wireless communication are becoming increasingly stringent, making the utilization and management of the wireless spectrum crucial. Full wireless spectrum sensing technology, based on software-defined radio, can monitor the wireless spectrum in real time and assess channel quality and signal strength. Full wireless spectrum sensing technology needs to be combined with related technical tasks such as identification and classification to efficiently monitor and manage the spectrum. Among these, modulation signal classification, as one of the key tasks of wireless communication systems, has become an important means of ensuring the quality and stability of wireless communication transmission. Furthermore, modulation signal classification can also be used in areas such as network security, spectrum allocation, and signal anti-interference in wireless communication.
[0003] However, real-world communication is subject to various complex interferences and channel fading, which severely impact signal structure and characteristics, posing significant challenges to signal modulation identification technology. To improve accuracy and reliability, expand applicability, and reduce computational complexity, researchers worldwide are actively exploring and researching to achieve more efficient, intelligent, and reliable automatic modulation identification technology in the future, meeting stringent requirements for communication quality and stability.
[0004] Initially, automatic modulation identification (EMI) primarily relied on manually selected signal features for classification based on decision boundaries. The identification accuracy was highly dependent on the chosen features, and manual feature selection was extremely complex. Channel noise and multipath fading significantly impacted identification efficiency. While subsequent machine learning methods offered advantages in classification efficiency and performance compared to traditional methods, their effectiveness still depended to some extent on expert experience. Deep learning methods, however, differ significantly. Deep learning networks, with their powerful feature extraction capabilities, can extract high-dimensional features from communication signal data, effectively achieving communication signal identification and classification tasks without relying on manually selected expert features. Experiments have demonstrated that deep learning-based modulation identification methods are more suitable for processing radio time-series sample data, achieving equivalent accuracy several times higher than traditional feature-based classifiers. This end-to-end training of raw data using neural networks effectively learns discriminative features and has achieved great success in modulation identification tasks.
[0005] While deep learning has made some progress in modulation pattern recognition, its performance in low signal-to-noise ratio (SNR) environments still has significant room for improvement. With the increasing complexity of the electromagnetic environment in wireless communication, received signals frequently encounter various types of interference, posing a challenge to the accuracy of signal processing. Particularly serious is the possibility that in some applications, adversaries may use intelligent interference sources to automatically lock onto and attack specific frequencies of our signals, greatly increasing the difficulty for receivers to distinguish between valid signals and interference in the frequency domain. Furthermore, adversaries may also employ strategies such as reducing transmitter power to decrease the likelihood of their signals being intercepted, resulting in extremely low SNR signals received. Given that modulation pattern recognition is a crucial step in the demodulation process, in-depth research and improvement of modulation pattern recognition capabilities under low SNR conditions are of paramount practical significance for ensuring communication security and enhancing information acquisition capabilities in specific scenarios. Summary of the Invention
[0006] To address the aforementioned problems in the prior art, this invention provides an automatic modulation recognition method and apparatus based on consistency constraints and feature prototypes. The technical problem to be solved by this invention is achieved through the following technical solution:
[0007] In a first aspect, the present invention provides an automatic modulation recognition method based on consistency constraints and feature prototypes, comprising:
[0008] Acquire the modulation signal to be identified;
[0009] The modulated signal to be identified is input into the trained automatic modulation recognition model, which extracts features from the high signal-to-noise ratio data and the low signal-to-noise ratio data in the modulated signal to be identified, outputs the recognition probability, and obtains the classification and recognition result of the modulated signal to be identified.
[0010] The trained automatic modulation recognition model is obtained by training the initial automatic modulation recognition model with data of a preset category; the data of the preset category includes modulation signals.
[0011] Secondly, the present invention also includes an automatic modulation identification device based on consistency constraints and feature prototypes, comprising:
[0012] The data acquisition module is used to acquire the modulation signal to be identified;
[0013] The data processing module is used to input the modulated signal to be identified into the trained automatic modulation recognition model, extract features of high signal-to-noise ratio data in the modulated signal to be identified, extract features of low signal-to-noise ratio data in the modulated signal to be identified, output the recognition probability, and obtain the classification recognition result of the modulated signal to be identified.
[0014] The trained automatic modulation recognition model is obtained by training the initial automatic modulation recognition model with data of a preset category; the data of the preset category includes modulation signals.
[0015] The beneficial effects of this invention are:
[0016] The present invention provides an automatic modulation recognition method and apparatus based on consistency constraints and feature prototypes, which can better cope with low signal-to-noise ratio environments, no longer rely on expert experience, and will not cause the signal features to be excessively suppressed or lost. The model of the present invention does not generate additional computational overhead after deployment, and does not require manual setting of threshold hyperparameters.
[0017] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0018] Figure 1 This is a flowchart of an automatic modulation recognition method based on consistency constraints and feature prototypes provided in an embodiment of the present invention;
[0019] Figure 2 This is a schematic diagram of an automatic modulation recognition model provided in an embodiment of the present invention;
[0020] Figure 3 This is a schematic diagram of the simulation experiment results provided in the embodiments of the present invention. Detailed Implementation
[0021] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.
[0022] Among the existing technologies, feature extraction-based methods rely on expert features and perform poorly in low signal-to-noise ratio environments; soft-threshold deep learning-based methods may lead to excessive suppression or loss of signal features, and the network's generalization ability is limited; wavelet denoising-based methods require additional wavelet denoising preprocessing of the data, increasing the computational complexity of the algorithm, and are extremely sensitive to the choice of threshold.
[0023] In view of this, the present invention provides an automatic modulation recognition method based on consistency constraints and feature prototypes. Based on deep learning, the method calculates feature prototypes of high signal-to-noise ratio data and applies consistency constraints to the features extracted from low signal-to-noise ratio data and the feature prototypes of high signal-to-noise ratio data, enabling the model to autonomously learn noise-resistant features and improve the recognition performance of the deep learning model in high-noise scenarios.
[0024] Please see Figure 1 , Figure 1 This is a flowchart of an automatic modulation recognition method based on consistency constraints and feature prototypes provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of an automatic modulation recognition model provided in an embodiment of the present invention. The present invention provides an automatic modulation recognition method based on consistency constraints and feature prototypes, comprising:
[0025] S101. Obtain the modulation signal to be identified.
[0026] S102. Input the modulation signal to be identified into the trained automatic modulation recognition model, extract features of high signal-to-noise ratio data in the modulation signal to be identified, extract features of low signal-to-noise ratio data in the modulation signal to be identified, output the recognition probability, and obtain the classification recognition result of the modulation signal to be identified.
[0027] The trained automatic modulation recognition model is obtained by training the initial automatic modulation recognition model with data of a preset category; the data of the preset category includes modulation signals.
[0028] Specifically, in this embodiment, before training the initial automatic modulation recognition model, the following steps are also included:
[0029] Construct an automatic modulation recognition model; among which,
[0030] The automatic modulation recognition model includes a feature extractor and a classifier. The feature extractor includes a first convolutional block, a second convolutional block, a third convolutional block, a fourth convolutional block, a fifth convolutional block, a first temporal layer, a second temporal layer, a first fully connected layer, and a second fully connected layer. The classifier includes a classification layer. The input of the first convolutional block receives a sine wave and a cosine wave of the signal to be processed. The input of the second convolutional block receives a cosine wave of the signal to be processed. The input of the third convolutional block receives a sine wave of the signal to be processed. The input of the fourth convolutional block receives a first signal, which is the signal processed by the second convolutional block and the signal processed by the third convolutional block. The connection signal is as follows: the input of the fifth convolutional block receives the second signal, which is the connection signal between the signal processed by the first convolutional block and the signal processed by the fourth convolutional block; the input of the first timing layer receives the signal processed by the fifth convolutional block; the input of the second timing layer receives the signal processed by the first timing layer; the input of the first fully connected layer receives the signal processed by the second timing layer; the input of the second fully connected layer receives the signal processed by the first fully connected layer; the input of the classification layer receives the signal processed by the second fully connected layer; and the output of the classification layer outputs the prediction score of the signal to be processed.
[0031] In this embodiment, the first, second, third, fourth, and fifth convolutional blocks have the same structure and all include a convolutional layer, a batch normalization layer, and a RuLU activation layer. The first convolutional block has 50 convolutional kernels, a kernel size of 2*8, a stride of 1, and 0 padding in its convolutional layer. The second convolutional block has 50 convolutional kernels, a kernel size of 1*8, a stride of 1, and 0 padding in its convolutional layer. The third convolutional block has 50 convolutional kernels, a kernel size of 1*8, a stride of 1, and 0 padding in its convolutional layer. The fourth convolutional block has 50 convolutional kernels, a kernel size of 1*8, a stride of 1, and 0 padding in its convolutional layer. The fifth convolutional block has 100 convolutional kernels, a kernel size of 2*5, a stride of 1, and 0 padding in its convolutional layer.
[0032] The hidden layer dimension of the first time series layer is 128, and the hidden layer dimension of the second time series layer is 128;
[0033] The output dimension of the first fully connected layer is 128, and the output dimension of the second fully connected layer is 11.
[0034] In this embodiment, data of a preset category is used as the training dataset to train the initial automatic modulation recognition model, including:
[0035] Obtain the training dataset, which includes multiple training samples, each of which contains a modulated signal;
[0036] Label the types of training samples in the training dataset to obtain the true labels of the types of training samples;
[0037] The training dataset and the true labels of the types of training samples in the training dataset are input into the initial automatic modulation recognition model for iterative training until the preset conditions are met, resulting in a well-trained automatic modulation recognition model.
[0038] In this embodiment, the training dataset and the true labels of the types of training samples in the training dataset are input into the initial automatic modulation recognition model for iterative training, including:
[0039] For the first iteration of training, the training dataset is input into the automatic modulation recognition model. Training samples with a signal-to-noise ratio greater than or equal to 0dB and correctly identified in the training dataset are processed by the feature extractor to extract features and obtain the first feature vector of the training process. Based on the first feature vector of each class of training samples, the feature prototype of each class of training samples is obtained.
[0040] In the second iteration of training, the training dataset is input into the automatic modulation recognition model. Training samples with a signal-to-noise ratio less than 0dB or greater than -10dB and correctly identified in the training dataset are processed by the feature extractor to obtain the second feature vector of the training process. Using the feature prototypes of each class of training samples in the first iteration of training, a consistency constraint is applied to the second feature vector of the training process to update the parameters of the current automatic modulation recognition model. Using the updated parameters of the automatic modulation recognition model, the feature prototypes of each class of training samples are recalculated to update the feature prototypes of each class of training samples.
[0041] For the third iteration of training, the feature prototypes of each class of training samples updated in the previous iteration are used to apply a consistency constraint to the second feature vector of the third iteration of training until the preset number of training iterations is completed.
[0042] It should be noted that the l-th iteration process requires using the feature prototypes of each class of training samples obtained in the (l-1)-th iteration process to apply a consistency constraint to the second feature vector in the l-th iteration process.
[0043] It should be noted that correct identification is specifically represented by the maximum probability score output by the signal sample after it is input into the model, which matches the true label of the sample.
[0044] In this embodiment, the preset condition includes minimizing the loss function during a finite number of training iterations. Optionally, during training, the optimizer uses Adam, and the training iterations are 200 epochs.
[0045] Specifically, the expression for the loss function L1 in the first iteration of training is:
[0046]
[0047] Among them, Y label Let X be the training sample. label The corresponding predicted score, Indicates training sample X label The corresponding real tags.
[0048] Specifically, the expression for the loss function L2 in the second iteration of training is:
[0049]
[0050] Among them, f i This represents the feature prototype of the training sample. Let F(·) represent a training sample with a signal-to-noise ratio less than 0dB and a true label of class i, where n represents the total number of training sample classes, i = 1, ..., n, and F(·) represents the feature extractor function.
[0051] Specifically, the feature prototype f of the i-th type of training sample i The expression is:
[0052]
[0053] Where Mean(·) represents the mean function, F represents the training sample of class i in the training dataset that has a signal-to-noise ratio greater than 0 dB and is correctly identified, and F(·) represents the feature extractor function.
[0054] In summary, the automatic modulation recognition method based on consistency constraints and feature prototypes provided by this invention acquires the feature prototypes of each class of training samples in each round of training during the training process. It then uses the updated feature prototypes of each class of samples in this round to apply consistency constraints to the feature vectors extracted in the next round of training. In a finite number of iterations, this minimizes the loss function, determines the parameters of the automatic modulation recognition model, and constructs the trained automatic modulation recognition model. This method allows the model to autonomously learn the noise-resistant features of the signal during training, improving the model's ability to recognize low signal-to-noise ratio data without adding a denoising module before the model.
[0055] In an optional embodiment of the present invention, the effectiveness of the automatic modulation recognition method based on consistency constraints and feature prototypes provided in the above embodiment is verified by simulation experiments, specifically as follows:
[0056] I. Simulation Conditions
[0057] The hardware platform for the simulation experiment in this embodiment is: GPU is NVIDIA GeForce GTX 1060, and video memory size is 6GB.
[0058] The software platform for the simulation experiment in this embodiment is Windows 10.
[0059] II. Simulation Content
[0060] To verify the effectiveness of the algorithm provided in the above embodiments, an automatic modulation recognition simulation experiment was conducted.
[0061] In this embodiment, the RML2016.10b dataset was used for training. It contains 10 different types of modulation signals, including 8 digital modulation signals (8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, QPSK) and 2 analog modulation signals (AM-DSB, WBFM). Each modulation signal contains 20 signal-to-noise ratios (SNR) of -20:2:18dB, with 6000 data points per SNR for each modulation signal, totaling 1,200,000 data points. 60% of the data was selected as the training set, 20% as the validation set, and 20% as the test set.
[0062] III. Simulation Results
[0063] Simulation results are as follows Figure 3 As shown, the yellow line represents the accuracy of the baseline model at a given signal-to-noise ratio (SNR), while the blue line represents the accuracy of the model after applying the method provided in the above embodiments of the present invention. In the low to medium SNR region, the model accuracy is improved by 1% to 4%, and the overall accuracy increases from 62.04% to 63.06%, indicating that the method proposed in this invention has a significant improvement in processing low SNR data.
[0064] Based on the same inventive concept, this invention also provides an automatic modulation identification device based on consistency constraints and feature prototypes, used to implement the automatic modulation identification method based on consistency constraints and feature prototypes provided in the above embodiments of this invention. Embodiments of the method are described above and will not be repeated here. The device includes:
[0065] The data acquisition module is used to acquire the modulation signal to be identified;
[0066] The data processing module is used to input the modulated signal to be identified into the trained automatic modulation recognition model, extract features of high signal-to-noise ratio data in the modulated signal to be identified, extract features of low signal-to-noise ratio data in the modulated signal to be identified, output the recognition probability, and obtain the classification recognition result of the modulated signal to be identified.
[0067] The trained automatic modulation recognition model is obtained by training the initial automatic modulation recognition model with data of a preset category; the data of the preset category includes modulation signals.
[0068] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.
[0069] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.
[0070] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. An automatic modulation recognition method based on consistency constraints and feature prototypes, characterized in that, include: Acquire the modulation signal to be identified; The modulated signal to be identified is input into a trained automatic modulation recognition model to extract features of high signal-to-noise ratio data and low signal-to-noise ratio data in the modulated signal to be identified, output the recognition probability, and obtain the classification and recognition result of the modulated signal to be identified. The trained automatic modulation recognition model is obtained by training an initial automatic modulation recognition model using data of a preset category as the training dataset. The automatic modulation recognition model includes a feature extractor and a classifier; the data of the preset category includes modulation signals; training the initial automatic modulation recognition model using data of the preset category as the training dataset includes: Obtain a training dataset, which includes multiple training samples, each of which includes a modulated signal; The types of training samples in the training dataset are labeled to obtain the true labels of the types of the training samples. The training dataset and the true labels of the types of training samples in the training dataset are input into the initial automatic modulation recognition model for iterative training until the preset conditions are met, and the trained automatic modulation recognition model is obtained. The training dataset and the true labels of the types of training samples in the training dataset are input into the initial automatic modulation recognition model for iterative training, including: For the first iteration of training, the training dataset is input into the automatic modulation recognition model. The training samples in the training dataset with a signal-to-noise ratio greater than or equal to 0dB and correctly identified are subjected to feature extraction by the feature extractor to obtain the first feature vector of the training process. Based on the first feature vector of each class of training samples, the feature prototype of each class of training samples is obtained. For the second iteration of training, the training dataset is input into the automatic modulation recognition model. Training samples in the training dataset with a signal-to-noise ratio less than 0dB or greater than -10dB and correctly identified are processed by the feature extractor to extract features, resulting in a second feature vector for the training process. Using the feature prototypes of each class of training samples from the first iteration, a consistency constraint is applied to the second feature vector of the second iteration to update the parameters of the current automatic modulation recognition model. Using the updated parameters of the automatic modulation recognition model, the feature prototypes of each class of training samples are recalculated to update the feature prototypes of each class of training samples. For the third iteration of training, the feature prototypes of each class of training samples updated in the previous iteration are used to apply a consistency constraint to the second feature vector of the third iteration of training until the preset number of training iterations is completed.
2. The automatic modulation recognition method based on consistency constraints and feature prototypes according to claim 1, characterized in that, Before training the initial automatic modulation recognition model, the following steps are also included: Construct an automatic modulation recognition model; among which, The feature extractor includes a first convolutional block, a second convolutional block, a third convolutional block, a fourth convolutional block, a fifth convolutional block, a first temporal layer, a second temporal layer, a first fully connected layer, and a second fully connected layer. The classifier includes a classification layer. The input of the first convolutional block receives a sine wave and a cosine wave of the signal to be processed. The input of the second convolutional block receives a cosine wave of the signal to be processed. The input of the third convolutional block receives a sine wave of the signal to be processed. The input of the fourth convolutional block receives a first signal, which is a concatenation signal between the signal processed by the second convolutional block and the signal processed by the third convolutional block. The input of the fifth convolutional block... The first time-series layer receives a second signal, which is a concatenation signal between the signal processed by the first convolutional block and the signal processed by the fourth convolutional block. The input of the first time-series layer receives the signal processed by the fifth convolutional block. The input of the second time-series layer receives the signal processed by the first time-series layer. The input of the first fully connected layer receives the signal processed by the second time-series layer. The input of the second fully connected layer receives the signal processed by the first fully connected layer. The input of the classification layer receives the signal processed by the second fully connected layer. The output of the classification layer outputs the prediction score of the signal to be processed.
3. The automatic modulation recognition method based on consistency constraints and feature prototypes according to claim 2, characterized in that, The first, second, third, fourth, and fifth convolutional blocks have the same structure and all include a convolutional layer, a batch normalization layer, and a RuLU activation layer. The first convolutional block has 50 convolutional kernels with a kernel size of 2*8, a stride of 1, and 0 padding. The second convolutional block has 50 convolutional kernels with a kernel size of 1*8, a stride of 1, and 0 padding. The third convolutional block has 50 convolutional kernels with a kernel size of 1*8, a stride of 1, and 0 padding. The fourth convolutional block has 50 convolutional kernels with a kernel size of 1*8, a stride of 1, and 0 padding. The fifth convolutional block has 100 convolutional kernels with a kernel size of 2*5, a stride of 1, and 0 padding. The hidden layer dimension of the first time sequence layer is 128, and the hidden layer dimension of the second time sequence layer is 128; The output dimension of the first fully connected layer is 128, and the output dimension of the second fully connected layer is 11.
4. The automatic modulation recognition method based on consistency constraints and feature prototypes according to claim 1, characterized in that, The preset conditions include minimizing the loss function within a preset number of training iterations.
5. The automatic modulation recognition method based on consistency constraints and feature prototypes according to claim 4, characterized in that, The loss function in the first iteration of training The expression is: ； in, Represented as training samples The corresponding predicted score, Indicates training samples The corresponding real tags.
6. The automatic modulation recognition method based on consistency constraints and feature prototypes according to claim 4, characterized in that, The loss function in the second iteration of training The expression is: ； in, This represents the feature prototype of the training sample. This indicates that the signal-to-noise ratio is less than 0dB and the true label is the first one. Training samples of the class, This represents the total number of training sample categories. , This represents the feature extractor function.
7. The automatic modulation recognition method based on consistency constraints and feature prototypes according to claim 6, characterized in that, The first Feature prototype of training samples The expression is: ； in, Represents the mean function, This indicates that the first correctly identified data in the training dataset has a signal-to-noise ratio greater than 0 dB. Training samples of the class, This represents the feature extractor function.
8. An automatic modulation recognition device based on consistency constraints and feature prototypes, characterized in that, include: The data acquisition module is used to acquire the modulation signal to be identified; The data processing module is used to input the modulation signal to be identified into the trained automatic modulation recognition model, extract features of high signal-to-noise ratio data in the modulation signal to be identified, extract features of low signal-to-noise ratio data in the modulation signal to be identified, output the recognition probability, and obtain the classification recognition result of the modulation signal to be identified. The trained automatic modulation recognition model is obtained by training an initial automatic modulation recognition model using data of a preset category as the training dataset; the automatic modulation recognition model includes a feature extractor and a classifier; the data of the preset category includes modulation signals; training the initial automatic modulation recognition model using data of the preset category as the training dataset includes: Obtain a training dataset, which includes multiple training samples, each of which includes a modulated signal; The types of training samples in the training dataset are labeled to obtain the true labels of the types of the training samples. The training dataset and the true labels of the types of training samples in the training dataset are input into the initial automatic modulation recognition model for iterative training until the preset conditions are met, and the trained automatic modulation recognition model is obtained. The training dataset and the true labels of the types of training samples in the training dataset are input into the initial automatic modulation recognition model for iterative training, including: For the first iteration of training, the training dataset is input into the automatic modulation recognition model. The training samples in the training dataset with a signal-to-noise ratio greater than or equal to 0dB and correctly identified are subjected to feature extraction by the feature extractor to obtain the first feature vector of the training process. Based on the first feature vector of each class of training samples, the feature prototype of each class of training samples is obtained. For the second iteration of training, the training dataset is input into the automatic modulation recognition model. Training samples in the training dataset with a signal-to-noise ratio less than 0dB or greater than -10dB and correctly identified are processed by the feature extractor to extract features, resulting in a second feature vector for the training process. Using the feature prototypes of each class of training samples from the first iteration, a consistency constraint is applied to the second feature vector of the second iteration to update the parameters of the current automatic modulation recognition model. Using the updated parameters of the automatic modulation recognition model, the feature prototypes of each class of training samples are recalculated to update the feature prototypes of each class of training samples. For the third iteration of training, the feature prototypes of each class of training samples updated in the previous iteration are used to apply a consistency constraint to the second feature vector of the third iteration of training until the preset number of training iterations is completed.