Radio signal identification method based on feature and IQ comparison pre-training

Abstract

The invention discloses a radio signal identification method based on feature and IQ comparison pre-training, belongs to the technical field of radio signal identification, and solves the problems of insufficient identification precision and the like caused by data scarcity and single data source in the existing radio signal identification method. The method comprises the following steps: extracting IQ sequence data and feature data of each radio signal sample, and constructing a sequence-feature pair data set; taking the parallel sequence branch network and the feature branch network as a pre-training model, and performing unsupervised comparative learning on the pre-training model by using the sequence-feature pair data set; adding a classification layer after the trained sequence branch network, and updating the sequence branch network; training weight parameters of the classification layer by using a small number of labeled samples to obtain a radio signal identification model based on the sequence branch network; and extracting the IQ sequence data of the newly received unlabeled radio signal, inputting the IQ sequence data into the radio signal identification model, and identifying the corresponding signal type.

Description

Technical Field

[0001] The present invention relates to the technical field of radio signal recognition, and in particular to a radio signal recognition method based on feature and IQ contrast pre-training. Background Art

[0002] With the rapid development of wireless communication technology, radio signal recognition, as a crucial part in the field of communication, is widely applied in multiple fields such as communication, radar, and electronic warfare. Traditional signal recognition methods mainly rely on features designed by experts, and analyze instantaneous features, statistical features, spectral features, etc. of signals. However, in the face of complex electromagnetic environments and low signal-to-noise ratio conditions, traditional signal recognition methods based on feature extraction show certain limitations. The rise of deep learning technology provides a new solution for radio signal recognition. Through deep neural networks, researchers can automatically learn feature representations from a large amount of data, greatly improving the accuracy and robustness of modulation recognition. However, deep learning methods still face some challenges in practical applications, especially when the labeled data is scarce or the training samples are insufficient, the model is prone to overfitting and has limited generalization ability.

[0003] As an unsupervised learning method, contrastive learning has shown significant advantages in dealing with the problem of scarce data. Contrastive learning does not rely on labeled data, but constructs positive sample pairs and negative sample pairs to learn the similarities and differences between samples, optimize the embedding space of the model, make similar samples closer in the feature space, and dissimilar samples farther away, thereby generating high-quality feature representations. This mechanism provides new possibilities for few-shot modulation recognition. However, in a complex radio environment, due to the limited expressive ability of single-modal features, it is difficult for the model to fully capture the deep characteristics in the signal, resulting in a decline in recognition performance under complex conditions such as low signal-to-noise ratio and multipath effects. Therefore, there are still limitations in relying solely on a single data source (such as only relying on sequence data) for contrastive learning. Summary of the Invention

[0004] In view of the above analysis, the embodiments of the present invention aim to provide a radio signal recognition method based on feature and IQ contrast pre-training to solve the problems of limited feature expression, insufficient recognition accuracy, and poor generalization ability in complex environments caused by scarce data and single data sources in existing radio signal recognition methods.

[0005] The present invention discloses a radio signal recognition method based on feature and IQ contrast pre-training, and the method includes: Extract the IQ sequence data and feature data of each radio signal sample respectively, and construct a sequence-feature pair data set; Taking the parallel sequence branch network and feature branch network as the pre-trained model, and using the sequence-feature pair dataset to perform unsupervised contrastive learning on the pre-trained model to obtain a pre-trained model that passes the training; Add a classification layer after the sequence branch network that passes the training, and update the sequence branch network; use a small number of labeled samples to train the weight parameters of the classification layer to obtain a radio signal recognition model based on the sequence branch network; Extract the IQ sequence data of the newly received unlabeled radio signal and input it into the radio signal recognition model to identify the corresponding signal type.

[0006] Based on the above solution, the present invention also makes the following improvements: Furthermore, when using the sequence-feature pair dataset to perform unsupervised contrastive learning on the pre-trained model, the following is executed: Initialize the model parameters of the sequence branch network and feature branch network in the pre-trained model; In each model training process, randomly extract B sequence-feature pairs of radio signal samples from the sequence-feature pair dataset and input them into the pre-trained model, and optimize the model parameters of the sequence branch network and feature branch network according to the output of the pre-trained model to obtain a pre-trained model that passes the training.

[0007] Furthermore, the pre-trained model implements each model training according to the following process: Randomly extract B sequence-feature pairs of radio signal samples from the sequence-feature pair dataset; input the IQ sequence data and feature data in the B sequence-feature pairs of radio signal samples into the sequence branch network and feature branch network respectively for feature extraction to obtain the corresponding sequence feature vectors and signal feature vectors; Construct a cosine similarity matrix according to the sequence feature vectors and signal feature vectors of B radio signal samples respectively to calculate the total loss of the pre-trained model; Optimize the model parameters of the sequence branch network and feature branch network according to the total loss of the pre-trained model until the model training end condition is met, stop training, and obtain a pre-trained model that passes the training.

[0008] Furthermore, when calculating the total loss of the pre-trained model, the following is executed: Construct the sequence feature matrix and signal feature matrix of the current training batch according to the sequence feature vectors and signal feature vectors of B radio signal samples respectively; Calculate the total loss of the pre-trained model according to the cosine similarity matrix between the sequence feature matrix and the signal feature matrix.

[0009] Furthermore, when calculating the total loss of the pre-trained model, the following is also executed: Calculate the cross-entropy losses of the sequence branch network and the feature branch network according to the cosine similarity matrix between the sequence feature matrix and the signal feature matrix; Calculate the total loss of the pre-trained model according to the cross-entropy losses of the sequence branch network and the feature branch network.

[0010] Furthermore, optimize the model parameters of the sequence branch network and the feature branch network according to the total loss of the pre-trained model, and execute: Take half of the total loss of the pre-trained model as the branch loss gradients of the sequence branch network and the feature branch network respectively, and optimize the model parameters of the sequence branch network and the feature branch network respectively.

[0011] Furthermore, the cosine similarity matrix of the sequence feature matrix relative to the signal feature matrix is expressed as: (1) where and represent the sequence feature matrix and the signal feature matrix respectively; represents the scaling factor matrix; The dimension of is ; The cosine similarity matrix of the signal feature matrix relative to the sequence feature matrix is expressed as: Furthermore, the cross-entropy losses and of the sequence branch network and the feature branch network are respectively expressed as: (3) (4) where the sequence label and represent calculating the cross-entropy loss between and ;

[0012] Furthermore, to obtain the radio signal recognition model based on the sequence branch network, execute: Load the sequence branch network that has passed the training, and freeze the model parameters of the sequence branch network; Add a classification layer after the sequence branch network that has passed the training to obtain the updated sequence branch network, and the number of nodes in the classification layer is the same as the number of signal type labels to be recognized; Use a small number of labeled samples to input the updated sequence branch network, and train the weight parameters of the classification layer to obtain the radio signal recognition model based on the sequence branch network.

[0013] Furthermore, the feature data includes instantaneous features, statistical features, and spectral features.

[0014] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: The present invention proposes a radio signal recognition method based on feature and IQ contrast pre-training, which has the following advantages: (1) By combining contrastive learning with multi-modal learning, the present invention can give full play to the advantages of IQ sequence data and artificial features, improve the recognition ability of radio signals in the case of scarce data, and break through the limitations of traditional single data source methods. Especially in the case of scarce data, it can significantly improve the recognition ability of radio signals, and at the same time extract deep features from a large number of unlabeled samples, providing basic support for signal recognition.

[0015] (2) Through the collaborative learning of artificial features and IQ sequence data, the present invention effectively assists the sequence data with feature data, enabling the sequence data to more accurately capture the subtle changes of the signal under the guidance of artificial features; at the same time, by fusing features of different modalities, the model can more comprehensively characterize the signal characteristics from multiple dimensions, thereby improving the discrimination ability of the sequence branch for different signal types in complex environments and enhancing the model's perception ability of signal subtle changes. It effectively improves the generalization performance of the model in complex environments. Especially in the case of low signal-to-noise ratio and complex channel conditions, the robustness and discrimination ability are significantly enhanced.

[0016] (3) In the fine-tuning stage, the present invention only retains the sequence branch network, and in the inference stage, only IQ sequence data is used for recognition. It no longer depends on the feature branch network, greatly reducing the model complexity and the demand for computing resources, while reducing the dependence on labeled data and improving the recognition accuracy in small sample scenarios.

[0017] (4) The present invention provides a new idea for radio signal recognition in scenarios with scarce labeled data, providing strong support for related research and practice, especially applicable to scenarios where data acquisition is difficult or the labeling cost is high.

[0018] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent specification, and some advantages can be made obvious from the specification, or understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained through the content specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings are only for the purpose of showing specific embodiments, and are not considered as a limitation of the present invention. Throughout the drawings, the same reference signs represent the same components; Figure 1Flowchart of the radio signal recognition method based on feature and IQ contrast pre-training provided by the embodiments of the present invention. Detailed implementation manners

[0020] The following will specifically describe the preferred embodiments of the present invention with reference to the accompanying drawings. The accompanying drawings form a part of this application and are used together with the embodiments of the present invention to explain the principles of the present invention, rather than to limit the scope of the present invention.

[0021] A specific embodiment of the present invention discloses a radio signal recognition method based on feature and IQ contrast pre-training. The flowchart is as Figure 1 shown.

[0022] Step S1: Extract the IQ sequence data and feature data of each radio signal sample respectively, and construct a sequence-feature pair dataset.

[0023] The following will explain the specific implementation process of step S1 as follows.

[0024] Step S11: Select a matching radio signal sample dataset according to the radio signal recognition requirements.

[0025] In the specific implementation process, the radio signal recognition requirements can be determined according to the signal types included in the radio signal to be recognized, so as to ensure that the selected radio signal sample dataset covers all signal types of the radio signal to be recognized.

[0026] Step S12: Extract the IQ sequence data of each radio signal sample in the radio signal sample dataset respectively.

[0027] Specifically, the IQ sequence data refers to the complex representation form of the radio signal, where I represents the in-phase component and Q represents the quadrature component.

[0028] In this embodiment, the radio signal can be expressed as: (1) where represents the signal of the th sampling point in the radio signal, represents the th sampling point in the radio signal, is the signal length representing the radio signal; , , respectively represent the amplitude, channel impulse response, and additive white Gaussian noise of the th sampling point in the radio signal sample, Represents the carrier frequency offset caused by the Doppler effect or the clock error between the transmitter and the receiver. Represents the random phase deviation. In the specific implementation process, the additive white Gaussian noise can be selected as the additive white Gaussian noise with a mean of zero and a variance of .

[0029] Extract the IQ components from the radio signal, which is expressed as: (2) (3) Among them, and respectively represent the in-phase component and the quadrature component of the th sampling point in the radio signal. and respectively extract the real part and the imaginary part of . After that, stack the IQ components to form the IQ sequence data with a dimension of : : (4) Step S13: Extract the feature data of each radio signal sample in the radio signal sample dataset respectively.

[0030] In this embodiment, the feature data includes instantaneous features, statistical features, and spectral features, which are specifically described as follows.

[0031] (1) Instantaneous features The instantaneous features are based on the instantaneous amplitude, phase, and frequency of the radio signal, and capture the changes at specific time points. The instantaneous features of the radio signal include: instantaneous amplitude, instantaneous phase, instantaneous frequency, standard deviation of the absolute value of the instantaneous amplitude, standard deviation of the absolute value of the instantaneous frequency, peak value of the instantaneous amplitude, and kurtosis of the instantaneous frequency.

[0032] (2) Statistical features The statistical features describe the statistical attributes of the radio signal and reflect the distribution characteristics of the signal over time. The statistical features of the radio signal include higher-order mixed moments and cumulants. Specifically, the cumulants can be defined through the joint cumulant function.

[0033] The statistical features can reflect the overall statistical characteristics of the signal in the time domain and reveal the changing trend of its distribution. For example, the higher-order mixed moments can be used to describe the skewness and kurtosis of the signal amplitude distribution, while the cumulants can more deeply characterize the statistical structure of the signal, especially in characterizing its characteristics deviating from the Gaussian distribution. These features provide an effective basis for the fine analysis and category discrimination of the modulation signal, thereby enhancing the accuracy and stability of the recognition method.

[0034] (3) Spectrum characteristics The spectrum characteristics can reflect the spectral characteristics of radio signals by extracting the frequency-domain information of radio signals. The spectrum characteristics of radio signals mainly include the maximum spectral power density, spectral symmetry, and local maxima of the spectrum.

[0035] Common spectrum characteristics such as the maximum power spectral density, spectral symmetry, and local maxima can effectively reflect the frequency-domain characteristics of signals and help distinguish different modulation types. For example, certain modulation methods have specific spectral distribution patterns. By extracting these spectrum characteristics, the corresponding patterns can be identified in the frequency domain, thereby improving the accuracy of modulation recognition.

[0036] In this embodiment, the instantaneous characteristics, statistical characteristics, and spectrum characteristics of the radio can all be calculated using existing methods, which will not be elaborated here. The instantaneous characteristics, statistical characteristics, and spectrum characteristics respectively describe different characteristics of radio signals from the time domain and frequency domain. By comprehensively using these characteristic information, the characteristics of radio signals can be captured more comprehensively, improving the accuracy and robustness of modulation recognition, and effectively helping the sequence branch to significantly improve the recognition performance. In this embodiment, by combining these traditional characteristics with deep learning technology, the recognition performance can be effectively improved, especially for signal recognition in complex and dynamic channel environments.

[0037] After extracting the characteristic data of each radio signal sample, the characteristic data of the radio signal sample can be represented as a 1*K-dimensional vector, where K represents the total number of all characteristics in the characteristic data.

[0038] Step S14: Construct the sequence-characteristic pairs from the IQ sequence data and characteristic data of each radio signal sample; summarize the sequence-characteristic pairs of all radio signal samples in the radio signal sample dataset to construct a sequence-characteristic pair dataset.

[0039] Preferably, in the specific implementation process, data integrity checks can also be performed on the sequence-characteristic pairs of each radio signal sample in the radio signal sample dataset, and the radio signal samples that fail the data integrity check (such as data missing) are excluded. Summarize the sequence-characteristic pairs of the remaining radio signal samples to construct a sequence-characteristic pair dataset, thereby ensuring the integrity and effectiveness of the sequence-characteristic pair dataset.

[0040] Sequence-characteristic pair dataset It is expressed as: (5) Wherein, , respectively represent the IQ sequence data and feature data of radio signal samples, is the total number of valid radio signal samples.

[0041] Step S2: Use the parallel sequence branch network and feature branch network as the pre-trained model, and perform unsupervised contrast learning on the pre-trained model using the sequence-feature pair dataset to obtain the pre-trained model that passes the training.

[0042] The sequence branch network and the feature branch network are respectively used to extract features from the IQ sequence data and the feature data. Preferably, in this embodiment, the sequence branch network can be implemented by a feed-forward neural network with a fully connected layer as the output layer.

[0043] Exemplarily, the sequence branch network can be implemented using the ResNeXt model. This embodiment provides a specific structure of the ResNeXt model, and the structure parameters are shown in Table 1.

[0044] Table 1 Structure parameters of the ResNeXt model

[0045] The ResNeXt model in Table 1 consists of 7 layers, including one convolutional layer, four residual blocks (denoted as ResX1, ResX2, ResX3, and ResX4), one pooling layer, and finally a fully connected layer (denoted as Fc) with (Generally, M is less than L) hidden nodes, which is used to output the sequence feature vector of the IQ sequence data. The four residual blocks are repeated 3 times, 4 times, 6 times, and 3 times respectively. Exemplarily, in this embodiment, the number of groups of grouped convolutions can be set to 32, which means that the grouped convolution contains 32 groups.

[0046] In this embodiment, using the sequence branch network to extract the sequence feature vector of the IQ sequence data, its implementation process can be expressed as: (6) where, represents the sequence feature vector extracted by the sequence branch network, , represents the feature of the th dimension in the sequence feature vector, represents the dimension of the sequence feature vector, represents the parameters of the sequence branch network, represents the mapping between the input and output of the sequence branch network.

[0047] Preferably, the feature branch network can be implemented using a CNN model. In this embodiment, the signal feature vector of the feature data is extracted by the feature branch network, and its implementation process can be expressed as: (7) Wherein, represents the signal feature vector extracted by the feature branch network, , represents the -th dimension feature in the signal feature vector, represents the parameters of the feature branch network, represents the mapping between the input and output of the feature branch network.

[0048] Next, the training process of the pre-trained model is specifically described as follows.

[0049] Step S21: Initialize the model parameters of the sequence branch network and the feature branch network in the pre-trained model.

[0050] Step S22: In each model training process, randomly extract B sequence-feature pairs of radio signal samples from the sequence-feature pair dataset and input them into the pre-trained model. Optimize the model parameters of the sequence branch network and the feature branch network according to the output of the pre-trained model to obtain a pre-trained model that passes the training.

[0051] Specifically, the pre-trained model implements each model training according to the following process.

[0052] Step S221: Randomly extract B sequence-feature pairs of radio signal samples from the sequence-feature pair dataset; input the IQ sequence data and feature data in the B sequence-feature pairs of radio signal samples into the sequence branch network and the feature branch network respectively to obtain the corresponding sequence feature vector and signal feature vector. B represents the number of samples input into the model during each training, that is, the size of the training batch. Exemplarily, if the sequence-feature pair dataset contains tens of thousands of sequence-feature pairs of radio signal samples, B can be 128 or 256 to accelerate the model training speed. Of course, other appropriate values can also be selected according to the specific situation.

[0053] Specifically, the sequence feature vector and the signal feature vector corresponding to the sequence-feature pair of the b-th radio signal sample are respectively expressed as 、 , where b ranges from 1 to B. Combining the representation methods of the feature vectors extracted by the sequence branch network and the feature branch network before, , , wherein, 、 respectively represent the Features of one dimension

[0054] Step S222: Construct a cosine similarity matrix respectively according to the sequence feature vectors and signal feature vectors of B radio signal samples, so as to calculate the total loss of the pre-trained model.

[0055] Specifically, construct a sequence feature matrix and a signal feature matrix of the current training batch respectively according to the sequence feature vectors and signal feature vectors of B radio signal samples; calculate the total loss of the pre-trained model according to the cosine similarity matrix between the sequence feature matrix and the signal feature matrix.

[0056] Sequence feature matrix , signal feature matrix .

[0057] Therefore, the cosine similarity matrix of the sequence feature matrix relative to the signal feature matrix is expressed as:[[]] (8) Wherein,[[]] represents a scaling factor matrix;[[]] The dimension of[[]] .

[0058] Specifically,[[]]

[0059] Since it can be known that[[]] , are the sequence feature vector and signal feature vector of the same radio signal sample, b ranges from 1 to B, therefore, it is a positive sample pair; the sequence feature vectors and signal feature vectors with different subscripts correspond to different radio signal samples, and they are negative sample pairs. In the process of optimizing unsupervised contrast learning, it is necessary to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs, so as to improve the model's understanding and discrimination ability of signal features and sequence features.

[0060] The cosine similarity matrix of the signal feature matrix relative to the sequence feature matrix is expressed as:[[]] (9) Subsequently, calculate the cross-entropy losses , of the sequence branch network and the feature branch network:[[]] (10) (11) Wherein, the sequence label , represents the calculation of and The cross - entropy loss between

[0061] Specifically, assume that , represents the similarity between the i - th sequence feature vector and the j - th signal feature vector.

[0062] For each row of , calculate the probability distribution respectively: (12) Take the negative logarithm probability of the positive sample pairs, and calculate the average value to calculate the cross - entropy loss of the sequence branch network , which is expressed as: (13) The cross - entropy loss of the feature branch network has a similar calculation process and will not be elaborated here.

[0063] At this time, the total loss of the pre - trained model is: (14) Step S223: Optimize the model parameters of the sequence branch network and the feature branch network according to the total loss of the pre - trained model until the model training end condition is met, stop training, and obtain a pre - trained model that passes the training.

[0064] Specifically, take half of the total loss of the pre - trained model as the branch loss gradients of the sequence branch network and the feature branch network respectively, optimize the model parameters of the sequence branch network and the feature branch network respectively, and jump to the next model training; until the model training end condition is met, stop training, and obtain a pre - trained model that passes the training.

[0065] It should be noted that formula (12) is used as the optimization function of the pre - trained model, and the symmetric design of the branch losses of the sequence branch network and the feature branch network can make the pre - trained model more robust in modeling the feature matching relationship of the two different modalities.

[0066] Specifically, the Adam optimizer can be used to optimize the model parameters of the sequence branch network and the feature branch network. Preferably, the model training end condition can be: reaching the preset number of training times, and the total loss of the pre - trained model converges.

[0067] Exemplarily, during the pre - training process, the initial learning rate of the sequence branch network and the feature branch network can be set to 0.001, and the maximum number of model training times can be set to 50. During a certain training process, after 10 training rounds, the learning rate can be decayed to 80% of the original value.

[0068] This step uses a contrastive learning method for radio signal recognition. By using positive sample pairs (the same type of signal) and negative sample pairs (different signals), the model can learn effective feature representations, making similar signals closer in the feature space and different signals farther apart. This mechanism enables the model to learn and extract the deep features of signals from a large number of unlabeled samples, providing basic model support for subsequent signal recognition tasks.

[0069] Step S3: Add a classification layer after the sequence branch network that has passed training, and update the sequence branch network; use a small number of labeled samples to train the weight parameters of the classification layer to obtain a radio signal recognition model based on the sequence branch network.

[0070] It should be noted that the feature branch network needs to extract a large amount of feature data, resulting in high computational complexity. In this embodiment, the feature data is extracted by the feature branch network, and its role is to perform contrastive learning with the sequence branch network in the contrastive learning stage, enabling the sequence branch network to learn more robust representations during the training stage and finally relying only on the lightweight sequence branch network to complete the recognition task in the inference stage. Therefore, after obtaining a sequence branch network with high feature extraction accuracy through step S2, only IQ sequence data can be used in the prediction stage, reducing the need for feature engineering. That is, due to the auxiliary role of the feature branch in the contrastive learning stage, efficient and high-performance prediction results can be achieved in the prediction stage.

[0071] In this step, load the sequence branch network that has passed training and freeze the model parameters of the sequence branch network to retain the knowledge accumulated in the pre-training stage; add a classification layer after the sequence branch network that has passed training to obtain an updated sequence branch network. The number of nodes in the classification layer is the same as the number of signal type labels to be recognized (i.e., the number of signal types in the labeled samples) to adapt to the recognition task. Use a small number of labeled samples (i.e., radio signal samples with a small number of labeled signal type labels) to input the updated sequence branch network and train the weight parameters of the classification layer, thereby obtaining a radio signal recognition model based on the sequence branch network.

[0072] During the process of training the weight parameters of the classification layer, a relatively low learning rate can be used for training. This method ensures the weight stability of the model during fine-tuning, avoids drastic changes in the weights, and thus enhances the performance of the model on specific tasks while retaining the general features learned during pre-training.

[0073] It should be noted that in this embodiment, the radio signal samples labeled with signal type tags cover radio signals of various signal types to be recognized. Exemplarily, common radio signal types include communication signals, broadcast signals, navigation and positioning signals, radar and detection signals, Internet of Things and low-power signals, emergency and dedicated signals, etc., and their corresponding signal type tags can be 0, 1, 2, 3, 4, 5, 6 respectively.

[0074] It should also be noted that the signal type of the radio signals in the radio signal sample dataset collected in step S1 should match the signal type of the radio signal samples labeled with signal type tags here.

[0075] Step S4: Extract the IQ sequence data of the newly received unlabeled radio signal and input it into the radio signal recognition model to identify the corresponding signal type.

[0076] In summary, the present invention proposes a radio signal recognition method based on feature and IQ contrast pre-training. By using the unsupervised learning mechanism of contrastive learning and multi-modal feature technology, the recognition performance of the model in complex channel environments and small sample scenarios is significantly improved. This method captures signal features from multiple dimensions by performing contrastive learning on the IQ sequence data and the manually designed feature data of the received radio signal samples, and uses the feature data to effectively assist the IQ sequence data, thereby enhancing the generalization ability of the model. During the pre-training process, first, a large amount of unlabeled IQ sequence data and feature data are used for contrastive learning, enabling the model to automatically learn the similarities and differences between signals in an unsupervised environment, so as to have stronger feature representation and discrimination capabilities when facing new signals. In the fine-tuning stage, the sequence branch network is further trained with a small number of labeled samples to ensure that it can effectively identify specific signal types. Finally, in the prediction stage, only the sequence branch network is used for signal recognition without relying on the feature branch network. This not only simplifies the model structure, reduces the computational complexity, but also reduces the demand for computing resources, enabling it to operate efficiently in resource-constrained environments.

[0077] Those skilled in the art can understand that all or part of the processes of implementing the above embodiment methods can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a disk, an optical disk, a read-only memory, or a random access memory, etc.

[0078] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A radio signal recognition method based on feature and IQ contrast pre-training, characterized in that, The radio signal recognition method includes: Extract the IQ sequence data and feature data of each radio signal sample respectively, and construct a sequence-feature pair data set; Use the parallel sequence branch network and feature branch network as the pre-trained model, and perform unsupervised contrast learning on the pre-trained model using the sequence-feature pair data set to obtain a pre-trained model that passes the training; Add a classification layer after the sequence branch network that passes the training, and update the sequence branch network; use a small number of labeled samples to train the weight parameters of the classification layer to obtain a radio signal recognition model based on the sequence branch network; Extract the IQ sequence data of the newly received unlabeled radio signal and input it into the radio signal recognition model to identify the corresponding signal type.

2. The radio signal recognition method based on feature and IQ contrast pre-training according to claim 1, wherein The unsupervised contrast learning of the pre-trained model using the sequence-feature pair data set is performed as follows: Initialize the model parameters of the sequence branch network and feature branch network in the pre-trained model; In each model training process, randomly extract B sequence-feature pairs of radio signal samples from the sequence-feature pair data set and input them into the pre-trained model, and optimize the model parameters of the sequence branch network and feature branch network according to the output of the pre-trained model to obtain a pre-trained model that passes the training.

3. The radio signal recognition method based on feature and IQ contrast pre-training according to claim 2, wherein The pre-trained model implements each model training according to the following process: Randomly extract B sequence-feature pairs of radio signal samples from the sequence-feature pair data set; input the IQ sequence data and feature data in the B sequence-feature pairs of radio signal samples into the sequence branch network and feature branch network respectively for feature extraction to obtain the corresponding sequence feature vectors and signal feature vectors; Construct a cosine similarity matrix according to the sequence feature vectors and signal feature vectors of B radio signal samples respectively to calculate the total loss of the pre-trained model; Optimize the model parameters of the sequence branch network and feature branch network according to the total loss of the pre-trained model until the model training end condition is met, stop training, and obtain a pre-trained model that passes the training.

4. The radio signal recognition method based on feature and IQ contrast pre-training according to claim 3, characterized in that To calculate the total loss of the pre-trained model, perform: Construct a sequence feature matrix and a signal feature matrix for the current training batch according to the sequence feature vectors and signal feature vectors of B radio signal samples respectively; Calculate the total loss of the pre-trained model according to the cosine similarity matrix between the sequence feature matrix and the signal feature matrix.

5. The radio signal recognition method based on feature and IQ contrast pre-training according to claim 4, characterized in that To calculate the total loss of the pre-trained model, also perform: Calculate the cross-entropy loss of the sequence branch network and the feature branch network according to the cosine similarity matrix between the sequence feature matrix and the signal feature matrix; Calculate the total loss of the pre-trained model according to the cross-entropy loss of the sequence branch network and the feature branch network.

6. The radio signal recognition method based on feature and IQ contrast pre-training according to claim 5, characterized in that, To optimize the model parameters of the sequence branch network and the feature branch network according to the total loss of the pre-trained model, perform: Take half of the total loss of the pre-trained model as the branch loss gradients of the sequence branch network and the feature branch network respectively, and optimize the model parameters of the sequence branch network and the feature branch network respectively.

7. According to the radio signal recognition method based on feature and IQ contrast pre-training described in claim 5, characterized in that Cosine similarity matrix of the sequence feature matrix relative to the signal feature matrix Expressed as: Among them, and represent a sequence feature matrix and a signal feature matrix, respectively; represents a scaling factor matrix; The dimension of is Cosine similarity matrix of the signal feature matrix relative to the sequence feature matrix It is expressed as: 。 8. The radio signal recognition method based on feature and IQ contrast pre-training according to claim 7, wherein Cross-entropy loss of the sequence branch network and the feature branch network , which are respectively expressed as: Among them, the sequence tag , represents the calculation and of the cross-entropy loss between them.

9. The radio signal recognition method based on feature and IQ contrast pre-training according to any one of claims 1-8, characterized in that, To obtain the radio signal recognition model based on the sequence branch network, perform: Load the sequence branch network that has passed training and freeze the model parameters of the sequence branch network; Add a classification layer after the sequence branch network that has passed training to obtain an updated sequence branch network, where the number of nodes in the classification layer is the same as the number of signal type labels to be recognized; Use a small number of labeled samples to input the updated sequence branch network, train the weight parameters of the classification layer, and obtain a radio signal recognition model based on the sequence branch network.

10. The radio signal recognition method based on feature and IQ contrast pre-training according to claim 9, wherein, The feature data includes instantaneous features, statistical features, and spectral features.

Images