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

Abstract

This invention discloses a radio signal recognition method based on feature and IQ comparison pre-training, belonging to the field of radio signal recognition technology. It solves the problems of insufficient recognition accuracy in existing radio signal recognition methods due to data scarcity and a single data source. The method includes: extracting IQ sequence data and feature data for each radio signal sample to construct a sequence-feature pair dataset; using a parallel sequence branch network and feature branch network as pre-training models, and performing unsupervised comparative learning on the pre-trained models using the sequence-feature pair dataset; adding a classification layer after the trained sequence branch network and updating the sequence branch network; training the weight parameters of the classification layer using a small number of labeled samples to obtain a radio signal recognition model based on the sequence branch network; extracting IQ sequence data of newly received unlabeled radio signals and inputting this data into the radio signal recognition model to identify 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 comparison pre-training. Background Art

[0002] With the rapid development of wireless communication technology, radio signal recognition, as a crucial component of the communications field, is widely used in various fields, including communications, radar, and electronic warfare. Traditional signal recognition methods rely primarily on expert-designed features, analyzing the signal's instantaneous, statistical, and spectral characteristics. However, in complex electromagnetic environments and low signal-to-noise ratio conditions, traditional feature extraction-based signal recognition methods have shown certain limitations. The rise of deep learning technology has provided a new solution for radio signal recognition. Using deep neural networks, researchers can automatically learn feature representations from large amounts of data, significantly improving the accuracy and robustness of modulation recognition. However, deep learning methods still face challenges in practical applications. In particular, when labeled data is scarce or training samples are insufficient, models are prone to overfitting and have limited generalization capabilities.

[0003] As an unsupervised learning method, contrastive learning has demonstrated significant advantages in addressing data scarcity. Rather than relying on labeled data, contrastive learning constructs pairs of positive and negative samples, learning the similarities and differences between samples and optimizing the model's embedding space. This brings similar samples closer together in the feature space, while dissimilar samples move further apart, thereby generating high-quality feature representations. This mechanism offers new possibilities for few-sample modulation recognition. However, in complex radio environments, due to the limited expressive power of a single modal feature, the model struggles to fully capture the deep characteristics of the signal, resulting in reduced recognition performance under complex conditions such as low signal-to-noise ratios and multipath effects. Therefore, relying solely on a single data source (e.g., sequence data) for contrastive learning still has limitations. Summary of the Invention

[0004] In view of the above analysis, an embodiment of the present invention aims to provide a radio signal recognition method based on feature and IQ comparison pre-training, so as to solve the problems of existing radio signal recognition methods such as limited feature expression, insufficient recognition accuracy and poor generalization ability in complex environments caused by data scarcity and a single data source.

[0005] The present invention discloses a radio signal recognition method based on feature and IQ comparison pre-training, the method comprising:

[0006] Extract the IQ sequence data and feature data of each radio signal sample respectively and construct a sequence-feature pair dataset;

[0007] The parallel sequence branch network and feature branch network are used as pre-training models, and unsupervised comparative learning is performed on the pre-training models using the sequence-feature pair dataset to obtain a pre-training model that has passed the training.

[0008] A classification layer is added after the trained sequence branch network, and the sequence branch network is updated. The weight parameters of the classification layer are trained using a small number of labeled samples to obtain a radio signal recognition model based on the sequence branch network.

[0009] The IQ sequence data of the newly received unlabeled radio signal is extracted and input into the radio signal recognition model to identify the corresponding signal type.

[0010] On the basis of the above solution, the present invention also makes the following improvements:

[0011] Furthermore, the sequence-feature pair dataset is used to perform unsupervised comparative learning on the pre-trained model, performing:

[0012] Initialize the model parameters of the sequence branch network and feature branch network in the pre-trained model;

[0013] During each model training process, sequence-feature pairs of B radio signal samples are randomly selected from the sequence-feature pair dataset and input into the pre-training model. The model parameters of the sequence branch network and the feature branch network are optimized according to the output of the pre-training model to obtain a pre-trained model that has passed the training.

[0014] Furthermore, the pre-trained model implements each model training according to the following process:

[0015] 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 sequence-feature pairs of the B radio signal samples into the sequence branch network and the feature branch network for feature extraction, respectively, to obtain the corresponding sequence feature vector and signal feature vector;

[0016] Construct a cosine similarity matrix based on the sequence feature vectors and signal feature vectors of B radio signal samples to calculate the total loss of the pre-trained model;

[0017] The model parameters of the sequence branch network and the feature branch network are optimized according to the total loss of the pre-trained model until the model training end conditions are met, and the training is stopped to obtain a pre-trained model that has passed the training.

[0018] Furthermore, to calculate the total loss of the pre-trained model, execute:

[0019] Construct the sequence feature matrix and signal feature matrix of the current training batch based on the sequence feature vectors and signal feature vectors of B radio signal samples respectively;

[0020] The total loss of the pre-trained model is calculated based on the cosine similarity matrix between the sequence feature matrix and the signal feature matrix.

[0021] Furthermore, to calculate the total loss of the pre-trained model, we also execute:

[0022] According to the cosine similarity matrix between the sequence feature matrix and the signal feature matrix, the cross entropy loss of the sequence branch network and the feature branch network is calculated;

[0023] The total loss of the pre-trained model is calculated based on the cross entropy loss of the sequence branch network and the feature branch network.

[0024] Furthermore, the model parameters of the sequence branch network and feature branch network are optimized based on the total loss of the pre-trained model. Execute:

[0025] Half of the total loss of the pre-trained model is used as the branch loss gradient of the sequence branch network and the feature branch network, respectively, to optimize the model parameters of the sequence branch network and the feature branch network.

[0026] Furthermore, the cosine similarity matrix of the sequence feature matrix relative to the signal feature matrix Expressed as:

[0027] (1)

[0028] in, 、 Represent the sequence feature matrix and signal feature matrix respectively; represents the scaling factor matrix; The dimension is ;

[0029] Cosine similarity matrix of signal feature matrix relative to sequence feature matrix Expressed as:

[0030] (2)

[0031] Furthermore, the cross entropy loss of the sequence branch network and the feature branch network 、 Respectively expressed as:

[0032] (3)

[0033] (4)

[0034] Among them, the sequence tag , Representative Calculation and The cross entropy loss between them.

[0035] Furthermore, the radio signal recognition model based on the sequence branch network is obtained by executing:

[0036] Load the trained sequence branch network and freeze the model parameters of the sequence branch network;

[0037] A classification layer is added after the trained sequence branch network to obtain an updated sequence branch network. The number of nodes in the classification layer is consistent with the number of signal type labels to be identified.

[0038] A small amount of labeled samples is used to input the updated sequence branch network to train the weight parameters of the classification layer, and a radio signal recognition model based on the sequence branch network is obtained.

[0039] Furthermore, the characteristic data includes instantaneous characteristics, statistical characteristics and spectral characteristics.

[0040] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0041] The present invention proposes a radio signal recognition method based on feature and IQ comparison pre-training, which has the following advantages:

[0042] (1) By combining contrastive learning with multimodal learning, this invention can fully leverage the advantages of IQ sequence data and artificial features, improve the ability to recognize radio signals in data-scarce conditions, and break through the limitations of traditional single-data-source methods. In particular, it can significantly improve the ability to recognize radio signals in data-scarce conditions, while also extracting deep features from a large number of unlabeled samples, providing fundamental support for signal recognition.

[0043] (2) This invention uses the collaborative learning of artificial features and IQ sequence data to effectively assist the sequence data with the feature data, enabling the sequence data to more accurately capture subtle changes in the signal under the guidance of artificial features. At the same time, by fusing features from different modalities, the model can more comprehensively characterize signal characteristics from multiple dimensions, thereby improving the sequence branch's ability to distinguish different signal types in complex environments and enhancing the model's ability to perceive subtle changes in the signal. This effectively improves the model's generalization performance in complex environments, especially under low signal-to-noise ratio and complex channel conditions, significantly enhancing its robustness and distinguishing ability.

[0044] (3) The present invention retains only the sequence branch network during the fine-tuning phase and uses only IQ sequence data for recognition during the inference phase. This eliminates reliance on the feature branch network, significantly reducing model complexity and computing resource requirements. It also reduces reliance on labeled data and improves recognition accuracy in small sample scenarios.

[0045] (4) This invention provides a new approach to radio signal recognition in scenarios where labeled data is scarce, and provides strong support for related research and practice. It is particularly suitable for scenarios where data acquisition is difficult or the labeling cost is high.

[0046] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following description, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. Throughout the drawings, the same reference symbols denote the same components.

[0048] Figure 1 This is a flowchart of a radio signal recognition method based on feature and IQ comparison pre-training provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0050] A specific embodiment of the present invention discloses a radio signal recognition method based on feature and IQ comparison pre-training, the flow chart is as follows: Figure 1 shown.

[0051] Step S1: extract the IQ sequence data and feature data of each radio signal sample respectively, and construct a sequence-feature pair data set.

[0052] The specific implementation process of step S1 is described below.

[0053] Step S11: selecting a matching radio signal sample data set according to radio signal recognition requirements.

[0054] During the specific implementation process, the identification requirements of the radio signal can be determined according to the signal types contained in the radio signal to be identified, so as to ensure that the selected radio signal sample data set covers all signal types of the radio signal to be identified.

[0055] Step S12: extracting IQ sequence data of each radio signal sample in the radio signal sample data set respectively.

[0056] Specifically, IQ sequence data refers to a complex number representation of a radio signal, where I represents an in-phase component (In-phase) and Q represents a quadrature component (Quadrature).

[0057] In this embodiment, the radio signal can be expressed as:

[0058] (1)

[0059] in, Indicates the first The signal of the sampling point, Indicates the first sampling points, To represent the signal length of the radio signal; 、 、 Represents the first The amplitude of the sampling points, the channel impulse response, the additive white Gaussian noise, Indicates the carrier frequency offset caused by the Doppler effect or the clock error between the transmitter and the receiver. In the specific implementation process, the additive Gaussian white noise can be selected with a mean of zero and a variance of Additive white Gaussian noise.

[0060] Extract the IQ components from the radio signal, expressed as:

[0061] (2)

[0062] (3)

[0063] in, 、 Represents the first The in-phase component and quadrature component of the sampling points, and Extract separately The real and imaginary parts of . Afterwards, the IQ components are stacked to form a dimension of IQ sequence data :

[0064] (4)

[0065] Step S13: extracting feature data of each radio signal sample in the radio signal sample data set.

[0066] In this embodiment, the characteristic data includes instantaneous characteristics, statistical characteristics, and spectral characteristics, which are described in detail as follows.

[0067] (1) Instantaneous characteristics

[0068] Instantaneous features capture changes at a specific point in time based on the instantaneous amplitude, phase, and frequency of a radio signal. These features 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.

[0069] (2) Statistical characteristics

[0070] Statistical characteristics describe the statistical properties of radio signals and reflect their temporal distribution. They include high-order mixing moments and cumulants. Specifically, cumulants can be defined using a joint cumulant function.

[0071] Statistical features can reflect the overall statistical properties of a signal in the time domain and reveal the changing trends of its distribution. For example, high-order mixing moments can be used to describe the skewness and kurtosis of the signal's amplitude distribution, while cumulants can provide a deeper understanding of the signal's statistical structure, particularly its deviations from a Gaussian distribution. These features provide an effective basis for detailed analysis and classification of modulated signals, thereby enhancing the accuracy and stability of recognition methods.

[0072] (3) Spectral characteristics

[0073] Spectral characteristics can reflect the spectral characteristics of radio signals by extracting frequency domain information of radio signals. Spectral characteristics of radio signals mainly include maximum spectral power density, spectral symmetry, and local maxima of the spectrum.

[0074] Common spectral features such as maximum power spectral density, spectral symmetry, and local maxima can effectively reflect the frequency domain characteristics of a signal and help distinguish different modulation types. For example, some modulation methods have specific spectral distribution patterns. By extracting these spectral features, the corresponding patterns can be identified in the frequency domain, thereby improving the accuracy of modulation recognition.

[0075] In this embodiment, the instantaneous characteristics, statistical characteristics, and spectral characteristics of the radio can all be calculated using existing methods, which will not be described in detail here. Instantaneous characteristics, statistical characteristics, and spectral characteristics describe different characteristics of radio signals from the time domain and frequency domain, respectively. By comprehensively using these feature information, the characteristics of the radio signal can be more comprehensively captured, the accuracy and robustness of modulation recognition can be improved, and sequence branches can be effectively helped to significantly improve recognition performance. By combining these traditional features with deep learning technology, this embodiment can effectively improve recognition performance, especially signal recognition in complex and dynamic channel environments.

[0076] After extracting the feature data of each radio signal sample, the feature data of the radio signal sample may be represented as a 1*K-dimensional vector, where K represents the total number of all features in the feature data.

[0077] Step S14: constructing the IQ sequence data and feature data of each radio signal sample into a sequence-feature pair; summarizing the sequence-feature pairs of all radio signal samples in the radio signal sample dataset to construct a sequence-feature pair dataset.

[0078] Preferably, during the specific implementation process, a data integrity check can be performed on the sequence-feature pair of each radio signal sample in the radio signal sample data set, and radio signal samples that fail the data integrity check (such as missing data) are eliminated. The sequence-feature pairs of the remaining radio signal samples are aggregated to construct a sequence-feature pair data set, thereby ensuring the integrity and validity of the sequence-feature pair data set.

[0079] Sequence-feature pair dataset Expressed as:

[0080] (5)

[0081] in, 、 Represents the first IQ sequence data and feature data of radio signal samples, is the total number of valid radio signal samples.

[0082] Step S2: Use the parallel sequence branch network and feature branch network as pre-training models, and use the sequence-feature pair dataset to perform unsupervised comparative learning on the pre-training models to obtain a trained pre-training model.

[0083] The sequence branch network and the feature branch network are used to extract features from the IQ sequence data and the feature data, respectively. Preferably, in this embodiment, the sequence branch network can be implemented using a feedforward neural network with a fully connected layer as the output layer.

[0084] Exemplarily, the sequence branch network can be implemented using a ResNeXt model. This embodiment provides a specific structure of a ResNeXt model, and the structural parameters are shown in Table 1.

[0085] Table 1 Structural parameters of the ResNeXt model

[0086]

[0087] The ResNeXt model in Table 1 consists of 7 layers, including a convolutional layer, four residual blocks (denoted as ResX1, ResX2, ResX3 and ResX4 respectively), a pooling layer, and finally a (Generally, M is less than L) A fully connected layer (denoted as Fc) with 1 hidden node 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. For example, the number of groups of group convolution can be Setting it to 32 means that the grouped convolution contains 32 groups.

[0088] In this embodiment, a sequence branch network is used to extract sequence feature vectors of IQ sequence data. The implementation process can be expressed as follows:

[0089] (6)

[0090] in, represents the sequence feature vector extracted by the sequence branch network, , Indicates the first character vector in the sequence The characteristics of the dimensions, represents the dimension of the sequence feature vector, represents the parameters of the sequence branching network, Represents the mapping between the input and output of a sequence-branching network.

[0091] Preferably, the feature branch network can be implemented using a CNN model. In this embodiment, the feature branch network is used to extract the signal feature vector of the feature data, and its implementation process can be expressed as:

[0092] (7)

[0093] in, represents the signal feature vector extracted by the feature branch network, , Indicates the first The characteristics of the dimensions, represents the parameters of the feature branch network, Represents the mapping between the input and output of the feature branch network.

[0094] Below, the training process of the pre-training model is described in detail.

[0095] Step S21: Initialize the model parameters of the sequence branch network and the feature branch network in the pre-trained model.

[0096] Step S22: During each model training process, B sequence-feature pairs of radio signal samples are randomly selected from the sequence-feature pair dataset and input into the pre-training model. The model parameters of the sequence branch network and the feature branch network are optimized according to the output of the pre-training model to obtain a pre-trained model that has passed the training.

[0097] Specifically, the pre-trained model implements each model training according to the following process.

[0098] 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 from these B radio signal sample sequence-feature pairs into the sequence branch network and 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 session, i.e., the training batch size. For example, 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 speed up model training. Of course, other appropriate values ​​can also be selected depending on the specific situation.

[0099] Specifically, the sequence feature vector and signal feature vector corresponding to the sequence-feature pair of the b-th radio signal sample are expressed as 、 , b takes 1 to B. Combining the representation of the feature vector extracted by the previous sequence branch network and feature branch network, , ,in, 、 They represent the sequence feature vector corresponding to the bth radio signal sample, the Features of dimensions.

[0100] Step S222: constructing a cosine similarity matrix based on the sequence feature vectors and signal feature vectors of the B radio signal samples to calculate the total loss of the pre-trained model.

[0101] Specifically, the sequence feature matrix and signal feature matrix of the current training batch are constructed according to the sequence feature vectors and signal feature vectors of B radio signal samples respectively; and the total loss of the pre-training model is calculated according to the cosine similarity matrix between the sequence feature matrix and the signal feature matrix.

[0102] Sequence feature matrix , signal feature matrix .

[0103] Therefore, the cosine similarity matrix of the sequence feature matrix relative to the signal feature matrix is Expressed as:

[0104] (8)

[0105] in, represents the scaling factor matrix; The dimension is .

[0106] Specifically,

[0107]

[0108] As we know, 、 Where, where b ranges from 1 to B, the sequence feature vector and signal feature vector represent the same radio signal sample, and thus constitute a positive sample pair. Sequence feature vectors and signal feature vectors with different subscripts correspond to different radio signal samples, and thus constitute a negative sample pair. During unsupervised contrastive learning optimization, it is necessary to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs, thereby improving the model's ability to understand and distinguish signal and sequence features.

[0109] Cosine similarity matrix of signal feature matrix relative to sequence feature matrix Expressed as:

[0110] (9)

[0111] Then, the cross entropy loss of the sequence branch network and the feature branch network is calculated 、 :

[0112] (10)

[0113] (11)

[0114] Among them, the sequence tag , Representative Calculation and The cross entropy loss between them.

[0115] Specifically, assuming , It represents the similarity between the i-th sequence feature vector and the j-th signal feature vector.

[0116] right Each row of computes a probability distribution separately:

[0117] (12)

[0118] Take the negative log-probability of the positive sample pairs and average them to calculate the cross entropy loss of the sequence branch network , expressed as:

[0119] (13)

[0120] Cross-entropy loss for feature-branch networks The calculation process of is similar to that of , which will not be described here.

[0121] At this point, the total loss of the pre-trained model for:

[0122] (14)

[0123] 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 has passed the training.

[0124] Specifically, half of the total loss of the pre-trained model is used as the branch loss gradient of the sequence branch network and the feature branch network, respectively, to optimize the model parameters of the sequence branch network and the feature branch network, and then jump to the next model training; until the end condition of the model training is met, the training is stopped and a pre-trained model that has passed the training is obtained.

[0125] It should be noted that formula (12) is used as the optimization function of the pre-training model. The symmetrical design of the branch losses of the sequence branch network and the feature branch network can make the pre-training model more robust in modeling the feature matching relationship of two different modalities.

[0126] Specifically, the Adam optimizer can be used to optimize the model parameters of the sequence branch network and the feature branch network. Preferably, the end condition of the model training can be: reaching a preset number of training times and the total loss of the pre-trained model converges.

[0127] For example, 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. In a certain training process, after 10 training rounds, the learning rate can decay to 80% of the original value.

[0128] This step introduces contrastive learning for radio signal recognition. Using pairs of positive samples (same signal) and negative samples (different signals), the model learns effective feature representations, bringing similar signals closer together in feature space and dissimilar signals farther apart. This mechanism enables the model to learn and extract deep signal features from a large number of unlabeled samples, providing foundational model support for subsequent signal recognition tasks.

[0129] Step S3: Add a classification layer after the trained sequence branch network and update the sequence branch network; use a small amount 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.

[0130] It should be noted that the feature branch network needs to extract a large amount of feature data and has high computational complexity. In this embodiment, the feature data is extracted by the feature branch network, and its function is to perform comparative learning with the sequence branch network in the comparative learning phase, so that the sequence branch network learns a more robust representation in the training phase, and finally relies only on the lightweight sequence branch network to complete the recognition task in the inference phase. Therefore, after obtaining the sequence branch network with high feature extraction accuracy through step S2, only IQ sequence data can be used in the prediction phase, reducing the need for feature engineering. That is, due to the auxiliary role of the feature branch in the comparative learning phase, the prediction phase can achieve efficient and high-performance prediction results.

[0131] In this step, the trained sequence branching network is loaded and its model parameters are frozen to preserve the knowledge accumulated during the pre-training phase. A classification layer is added after the trained sequence branching network to obtain an updated sequence branching network. The number of nodes in the classification layer matches the number of signal type labels to be identified (i.e., the number of signal types in the labeled samples) to adapt to the recognition task. A small number of labeled samples (i.e., radio signal samples labeled with signal type labels) are fed into the updated sequence branching network to train the weight parameters of the classification layer, thereby obtaining a radio signal recognition model based on the sequence branching network.

[0132] When training the weight parameters of the classification layer, a lower learning rate can be used. This method ensures the stability of the model weights during fine-tuning and avoids drastic changes in the weights. This method improves the model's performance on specific tasks while retaining the general features learned during pre-training.

[0133] 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 identified. For example, 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 special signals, etc., and their corresponding signal type tags can be 0, 1, 2, 3, 4, 5, and 6 respectively.

[0134] It should also be noted that the signal type of the radio signals in the radio signal sample data set collected in step S1 should match the signal type of the radio signal sample annotated with the signal type label here.

[0135] Step S4: extracting the IQ sequence data of the newly received unlabeled radio signal and inputting it into the radio signal recognition model to identify the corresponding signal type.

[0136] In summary, the present invention proposes a radio signal recognition method based on feature and IQ contrast pre-training. This method utilizes the unsupervised learning mechanism of contrastive learning and multimodal feature technology to significantly improve the model's recognition performance in complex channel environments and small sample size scenarios. This method compares IQ sequence data with artificially designed feature data for received radio signal samples, capturing signal features from multiple dimensions. The feature data effectively supplements the IQ sequence data, thereby enhancing the model's generalization capability. During pre-training, a large amount of unlabeled IQ sequence data and feature data are first used for contrastive learning. This enables the model to automatically learn the similarities and differences between signals in an unsupervised environment, thereby enhancing its feature representation and discrimination capabilities when faced with new signals. During the fine-tuning phase, the sequence branch network is further trained using a small number of labeled samples to ensure its ability to effectively identify specific signal types. Finally, during the prediction phase, signal recognition is performed solely using the sequence branch network, without relying on the feature branch network. This not only simplifies the model structure and reduces computational complexity, but also reduces the computing resource requirements, enabling efficient operation in resource-constrained environments.

[0137] Those skilled in the art will appreciate that all or part of the process steps of the above-described embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, such as a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0138] The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A radio signal recognition method based on feature and IQ comparison pre-training, characterized in that: The radio signal identification method comprises: Extract the IQ sequence data and feature data of each radio signal sample respectively and construct a sequence-feature pair dataset; The parallel sequence branch network and feature branch network are used as pre-training models, and unsupervised comparative learning is performed on the pre-training models using the sequence-feature pair dataset to obtain a pre-training model that has passed the training. The trained sequence branch network is loaded and its model parameters are frozen. A classification layer is added after the trained sequence branch network to obtain an updated sequence branch network, where the number of nodes in the classification layer is consistent with the number of signal type labels to be identified. The updated sequence branch network is input with IQ sequences of a small number of radio signal samples labeled with signal type labels to train the weight parameters of the classification layer, thereby obtaining a radio signal recognition model based on the sequence branch network. Extracting the IQ sequence data of the newly received unlabeled radio signal and inputting it into the radio signal recognition model to identify the corresponding signal type; The unsupervised contrastive learning of the pre-trained model using the sequence-feature pair dataset is performed: Initialize the model parameters of the sequence branch network and feature branch network in the pre-trained model; During each model training process, the IQ sequence data and feature data from the sequence-feature pairs of B radio signal samples randomly selected from the sequence-feature pair dataset are input into the sequence branch network and feature branch network respectively for feature extraction to obtain the corresponding sequence feature vectors and signal feature vectors. Based on the sequence feature vectors and signal feature vectors of the B radio signal samples, a cosine similarity matrix is ​​constructed to calculate the total loss of the pre-trained model. The model parameters of the sequence branch network and the feature branch network are optimized according to the total loss of the pre-trained model until the model training end conditions are met, and the training is stopped to obtain a pre-trained model that has passed the training.

2. The radio signal recognition method based on feature and IQ comparison pre-training according to claim 1, characterized in that: To calculate the total loss of the pre-trained model, execute: Construct the sequence feature matrix and signal feature matrix of the current training batch based on the sequence feature vectors and signal feature vectors of B radio signal samples respectively; The total loss of the pre-trained model is calculated based on the cosine similarity matrix between the sequence feature matrix and the signal feature matrix.

3. The radio signal recognition method based on feature and IQ comparison pre-training according to claim 2, characterized in that: Calculate the total loss of the pre-trained model and also execute: According to the cosine similarity matrix between the sequence feature matrix and the signal feature matrix, the cross entropy loss of the sequence branch network and the feature branch network is calculated; The total loss of the pre-trained model is calculated based on the cross entropy loss of the sequence branch network and the feature branch network.

4. The radio signal recognition method based on feature and IQ comparison pre-training according to claim 3, characterized in that: Optimize the model parameters of the sequence branch network and feature branch network based on the total loss of the pre-trained model and execute: Half of the total loss of the pre-trained model is used as the branch loss gradient of the sequence branch network and the feature branch network, respectively, to optimize the model parameters of the sequence branch network and the feature branch network.

5. The radio signal recognition method based on feature and IQ comparison pre-training according to claim 4, characterized in that: Cosine similarity matrix of sequence feature matrix relative to signal feature matrix Expressed as: in, 、 Represent the sequence feature matrix and signal feature matrix respectively; represents the scaling factor matrix; The dimension is ; Cosine similarity matrix of signal feature matrix relative to sequence feature matrix Expressed as: 。 6. The radio signal recognition method based on feature and IQ comparison pre-training according to claim 5, characterized in that: Cross entropy loss of sequence branch network and feature branch network 、 Respectively expressed as: Among them, the sequence tag , Representative Calculation and The cross entropy loss between them.

7. The radio signal recognition method based on feature and IQ comparison pre-training according to claim 6, characterized in that: The characteristic data includes instantaneous characteristics, statistical characteristics and spectrum characteristics.

Images