A wireless signal recognition method using a dual system architecture

Abstract

The application provides a wireless signal recognition method adopting a double-system architecture, and belongs to the field of wireless communication.The method combines a data-driven deep learning system (system one) and a knowledge-driven traditional communication system (system two), and the two systems work in turn to realize mutual optimization and upgrading between the systems: the preliminary recognition result output by system one helps to reduce the calculation time of system two; the unknown class samples with labels output by system two help to enrich and strengthen the training data set of system one.In addition, system one combines a classification network, a generation network and a self-supervised network, can effectively learn the potential representation and features of data, better understand the internal structure and pattern of data, and improve the recognition performance of known samples and unknown samples.The closed set recognition performance of the double system is better than that of the mainstream recognizers CNN and ResNet, and the open set recognition performance is better than that of the SoftMax open set recognizer, so that the closed set / open set recognition of wireless signals with high accuracy and low delay is realized.

Description

Technical Field

[0001] This invention belongs to the field of signal recognition in wireless communication, and mainly involves the design of data-driven systems based on deep learning methods and the design of knowledge-driven systems based on traditional communication theories. Background Technology

[0002] Wireless signal identification (WSI) technology plays a crucial role in today's communication era, with wide applications in electronic reconnaissance, spectrum management, and intrusion detection. In recent years, the explosive growth of IoT devices in key application scenarios such as smart industry, smart cities, and smart cars has led to severe spectrum scarcity and security vulnerabilities. WSI technology mitigates these security vulnerabilities by identifying signal types, parameters, and fingerprint characteristics, and enables collaborative sharing of spectrum resources, significantly improving spectrum utilization.

[0003] Existing wireless signal recognition technologies can be broadly categorized into two types: traditional communication methods and deep learning methods. Deep learning methods utilize advanced neural networks, treating a large dataset of collected wireless signals as a single dataset. Through iterative training, model parameters are optimized, and signal features are automatically extracted to achieve wireless signal classification and recognition. This method requires minimal time investment to achieve effective signal recognition, and with the development of artificial intelligence, it has gradually become the mainstream approach for signal recognition. However, this method is overly reliant on data-driven approaches, suffers from poor generalization performance, and lacks interpretability; its recognition performance is severely limited by the size and distribution of the dataset. On the other hand, the essence of deep learning classification methods is selecting parameters with the highest matching degree from the candidate set, which is a "closed-set" recognition method and cannot achieve true blind signal recognition. Traditional communication methods, based on manually extracted expert features, can achieve both closed-set signal recognition using classifiers and open-set blind signal recognition using traditional communication theory and statistical analysis methods. This method can produce high-accuracy results and reliably identify signals under different electromagnetic environments. However, this method typically has extremely high computational complexity and limited recognition efficiency, making it difficult to apply in scenarios with high real-time requirements. Summary of the Invention

[0004] To address the problem of wireless signal identification under non-cooperative communication conditions, this invention aims to propose a joint deep learning method and a traditional communication method for signal identification, achieving high-precision, low-latency open-set and closed-set identification of wireless signals. This method consists of a data-driven system based on deep learning (System 1) and a knowledge-driven system based on traditional communication algorithms (System 2). System 1 includes a generator network, a classification network, and a self-supervised network to achieve preliminary identification of known classes and detection of unknown classes. System 2 includes a closed-set identification module and an open-set identification module to achieve accurate identification of known classes and parameter identification of unknown classes. The two systems work collaboratively, achieving mutual optimization and upgrades: System 2 performs parameter identification and label assignment for unknown samples, enhancing the identification and detection capabilities of System 1 by strengthening the dataset; as the performance of System 1 improves, the computational traversal range of System 2 is reduced, decreasing its identification time. Furthermore, the dual-system framework combines deep learning and traditional communication methods to achieve integrated detection and identification of unknown samples.

[0005] The technical solution of this invention is a wireless signal identification method employing a dual-system architecture, the method comprising:

[0006] Step 1: Generate datasets for training and testing;

[0007] Step 1.1: Generate dataset samples;

[0008] Define the wireless signal identification task, generate a parameter candidate set according to different tasks, generate wireless signal data samples using hardware and software collaborative tools, and assign them corresponding labels;

[0009] Step 1.2: Dataset partitioning;

[0010] Generate a certain number of samples, randomly shuffle the samples, and divide them into training set, validation set, and test set in a ratio of 8:1:1.

[0011] Step 2: Design and train System 1. System 1 includes a generator network, a classification network, and a self-supervised network. The input data first passes through the generator network. The output data of the generator network is reconstructed from the input data and used as the input to the classification module. The output of the classification module passes through the self-supervised network to obtain the preliminary identification results of the samples.

[0012] Step 2.1: Design and training of the generative network;

[0013] The generative network employs a deep autoencoder / decoder architecture, consisting of an encoder network E(x), a decoder network G(x), and a discriminator network D. Here, x represents the original input sample. The encoder network E(x) comprises a convolutional network and ResNet encoding blocks, the decoder network G(x) comprises a decoder block and a deconvolutional network, and the discriminator network D is trained using a Wasserstein discrimination mechanism with gradient penalty. These three parts are jointly trained, with the optimization objective being to minimize the reconstruction error ||xG(E(x))||1. During training, the loss function L is continuously and alternately optimized. D and L G :

[0014]

[0015]

[0016] in, This is the interpolation gradient penalty term, where λ is the gradient penalty weight coefficient. This parameter needs to be adjusted experimentally to balance the influence of the original loss function and the gradient penalty term. This is the gradient operator, used to calculate the gradient value of a vector; For the original sample x and the generated sample The difference sample is calculated using the following formula:

[0017]

[0018] Where ε is a random number between 0 and 1, and when ε is close to 1, the difference sample Closer to the original sample x; when ε approaches 0, the difference sample... Closer to the generated sample During the training of generative networks Always between x and The function space of the discriminator D is constrained by the difference between the generator and the discriminator, which alleviates the fluctuations during training and promotes network convergence. P(D) improves the mode collapse problem in GAN training by penalizing the difference between the generator and the discriminator. The generator network can accurately reconstruct known class samples, but it cannot accurately reconstruct unknown class samples. This invention explores the similarity between the input sample and the reconstructed sample, and determines the high similarity sample as a known sample and the low similarity sample as an unknown sample.

[0019] Step 2.2: Design and training of classification networks and self-supervised networks;

[0020] The classification network, based on ResNet, addresses the performance degradation issue in deep neural network training. A self-supervised network acts as an aid to the classification network, randomly transforming the two-dimensional input data and predicting the transformation pattern using self-supervision, thus enabling the network to learn the intrinsic structure and latent features of the data at a deeper level. During training, the sample data is first augmented; the augmented data input to the classification network is... x is the original data, and t(x) is the output of x after random transformation, with the transformation exponent being r. The data is the reconstructed data of t(x) after passing through the generator network; the classification loss l is calculated by comparing the true label and the predicted label of x. c Treating r as the true label and using cross-entropy to calculate the self-supervised loss l ss The joint loss function of the classification network and the self-supervised network is l t =α1l c +α2l ss , where α1 and α2 are weight coefficients; making System 1 place greater emphasis on the classification task;

[0021] Step 3: Design System 2, which includes a closed set recognition module and an open set recognition module. If the sample detected by System 1 is a known sample, the closed set recognition module is used for further accurate recognition. If the sample detected by System 1 is an unknown sample, the original sample is input into the open set recognition module of System 2 for specific parameter recognition.

[0022] Step 4: Combine System 1 and System 2 and test their performance;

[0023] A dual-system wireless signal identification framework is constructed by combining a data-driven System 1 and a knowledge-driven System 2. System 1 is responsible for the preliminary classification of known samples and the detection of unknown samples, while System 2 is responsible for the precise classification of known samples and the parameter identification of unknown samples. Let the candidate set parameters of the known classes be [K1, K2, K3, K4, ..., K]. n ];

[0024] In System 1, the augmented data is detected as either a known sample or an unknown sample by the classification network; if the sample is detected as a known sample, System 1 outputs the confidence score [b1,b2,b3,b4,……,b] for each parameter in the candidate set. n ], where b i This represents the confidence level corresponding to the parameters of the i-th candidate set, with a value between 0 and 1; System 1 combines the original samples and the candidate set [K] with confidence levels higher than the threshold. t1 ,K t2 ,K t3Both are input into the closed set recognition module of System 2 for precise recognition; if the sample is detected as an unknown sample, System 1 directly inputs the original sample into the open set recognition module of System 2 for specific parameter recognition;

[0025] In System 2, if the input sample is a known class sample, the closed set recognition module sequentially detects the candidate set [K] provided by System 1. t1 ,K t2 ,K t3 The system identifies the parameters of each sample and then uses them to obtain more accurate recognition results. If the input sample is an unknown sample, the system relies on the open set recognition module to identify the specific parameters of the sample. After the system completes the identification of the parameters of the unknown sample, it assigns a label to the sample and incorporates it into the training dataset of the system, thereby strengthening the dataset and improving the performance of the system.

[0026] Furthermore, in step 2.2, α1 = 0.8 and α2 = 0.2.

[0027] Furthermore, in System 2, the specific method in the closed set identification module is as follows:

[0028] First, define the candidate set of encoding parameters [K1, K2, K3, K4, ..., K]. n Then, based on the encoded sequence and the encoder, the check equation is calculated. Based on the check equation, the corresponding metric value [d1, d2, d3, d4, ..., d] is calculated by traversing the parameters of each candidate set. n ], where d i Let be the metric value corresponding to the parameters of the i-th candidate set, and the parameter corresponding to the maximum metric value is the final recognition result;

[0029] The specific method in the open set recognition module is as follows:

[0030] The open-set identification method for Polar codes in the field of channel coding identification is as follows: The open-set identification algorithm for Polar codes needs to traverse from the minimum code length to the maximum code length. Each traversal requires calculating the codeword matrix X and the Kronecker matrix D under the current code length. Then, based on X and D, the dual vector h of different coding positions is obtained. The positions of information bits and frozen bits are determined by comparing a specific statistic γ and a threshold value Λ. Finally, after the traversal is completed, the code length value is obtained based on the special relationship between the Polar code rate and the code length.

[0031] In summary, this invention has the following beneficial effects for wireless signal identification tasks:

[0032] 1. This invention addresses the problem of wireless signal identification by proposing a dual-system identification framework for the first time, which can effectively solve the problem of identifying closed-set and open-set wireless signals;

[0033] 2. This invention achieves low-latency, high-precision identification of wireless signals through the sequential collaborative operation of two systems, and enhances the diversity of sample data by labeling unknown samples, thereby enabling mutual optimization and upgrading between systems.

[0034] 3. This invention combines the advantages of traditional communication methods and deep learning methods. It uses deep learning methods to quickly detect unknown samples and uses traditional communication methods to identify the specific parameters of unknown samples, thus achieving the integration of unknown sample detection and parameter identification.

[0035] 4. This invention enhances the representation of data samples by introducing generative networks and self-supervised networks, and learns the deep intrinsic structure of data. Its closed-set recognition performance surpasses that of the current advanced ResNet recognition network and other typical neural networks, and its open-set recognition performance surpasses that of the SoftMax open-set recognizer. Attached Figure Description

[0036] Figure 1 A flowchart illustrating the design process of the dual-system framework;

[0037] Figure 2 This is a schematic diagram of the dual-system framework working mode;

[0038] Figure 3 This is a schematic diagram of the internal network structure of System 1;

[0039] Figure 4 This is a detailed structural diagram of the dual-system framework. Detailed Implementation

[0040] To make the objectives, technical methods, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0041] This invention addresses the problem of wireless signal identification in non-cooperative communication by proposing a general dual-system framework. This framework consists of two subsystems: System 1 comprises a classification network, a generator network, and a self-supervised network, enabling preliminary identification of known samples and detection of unknown samples; System 2 comprises a closed-set identification module and an open-set identification module, enabling accurate identification of known samples and identification of unknown samples.

[0042] The dual-system structure proposed in this invention operates sequentially and collaboratively. Its purpose is to achieve high-accuracy, low-time-consumption wireless signal identification and to enable mutual optimization and upgrades between the systems. Its operating mode is as follows: Figure 2As shown, System 1 can automatically extract the depth features of wireless signals and output the detection results of known and unknown classes of samples. If a sample is determined to be a known sample by System 1, a preliminary identification result of the known class is output, which includes the confidence level of each candidate parameter in the candidate set. Multiple candidate parameters with high confidence levels, along with the original sample, are input into the closed-set identification module of System 2 for precise identification. Therefore, the closed-set identification module does not need to traverse and calculate the entire candidate set; it only needs to further check the set of candidate parameters with high confidence provided by System 1, significantly reducing the time consumption of the identification task. If a sample is determined to be an unknown sample by System 1, the original sample is directly input into the open-set identification module of System 2 for specific parameter identification. After completing the parameter identification of the unknown sample, the open-set identification module assigns a label to the unknown sample and uses it as a known sample for the next round of training of System 1, thereby achieving dataset reinforcement and improving the identification and detection capabilities of System 1. As the identification and detection performance of System 1 continues to improve, the provided candidate parameter set becomes more accurate, further narrowing the computational traversal range of System 2, and the identification time of System 2 will continue to decrease.

[0043] The specific steps for constructing the dual-system framework for wireless signal identification according to this invention are as follows, and its implementation process is as follows: Figure 1 As shown. To facilitate the description of specific implementation details, channel coding parameter identification, one of the important tasks in the field of wireless communication signal identification, is taken as an example.

[0044] Step 1: Generate datasets for training and testing.

[0045] Step 1.1: Dataset Sample Generation. Channel coding is performed on the information sequence of length k to obtain a coded sequence c of length n. This coded sequence is then modulated, resulting in data x. If the channel is set to additive white Gaussian noise, the received sequence y is...

[0046] y = x + w

[0047] Where y is a standard sample sequence, x is the modulated coded data, and w is additive noise, each element of which has a mean of 0 and a variance of σ. 2 Gaussian random variables.

[0048] Step 1.2: Dataset Partitioning. Generate a certain number of samples, and randomly select 80% as the training set, 10% as the validation set, and 10% as the test set. The training set is used to optimize network parameters until the network converges; the validation set is used to adjust network hyperparameters, including batch size, number of training iterations, learning rate, optimizer type, etc.; the test set is used to test network performance. For closed-set recognition tests, the test samples consist only of samples from known classes; for open-set recognition tests, the test samples consist of samples from both known and unknown classes.

[0049] Step 2: Design and train the generative network, classification network, and self-supervised network in System 1. The specific structure of System 1 is as follows: Figure 3 As shown.

[0050] Step 2.1: Generate network training design and training.

[0051] The generative network employs a deep autoencoder / decoder architecture, primarily consisting of three parts: an encoder network E(x) mapping data to the feature space, a decoder network G(x) mapping data back from the feature space, and a discriminator network D distinguishing between fake and real reconstructed data. Here, x represents the original input sample. The encoder network E(x) consists of one convolutional layer and six ResNet residual blocks, the decoder network G(x) consists of six decoder blocks and two deconvolutional layers, and the discriminator network D is trained using a Wasserstein discrimination mechanism with gradient penalty. These three parts are jointly trained, with the optimization objective being to minimize the reconstruction error ||xG(E(x))||1. During training, the loss function L is continuously and alternately optimized. D and L G :

[0052]

[0053]

[0054] in, This is the interpolation gradient penalty term, where λ is the gradient penalty weight coefficient, to ensure a balance between the loss function and the gradient penalty; For the original sample x and the generated sample The difference sample is calculated using the following formula:

[0055]

[0056] In the formula, ε is a random number between 0 and 1, such that... Always between x and This effectively constrains the output range of the discriminator D, reduces training instability, and helps the network converge.

[0057] Ideally, a generative network can accurately reconstruct samples of known classes, meaning the input sample and the reconstructed sample of the known class have high similarity. Conversely, during training, because the generative network does not extract features from samples of unknown classes, it cannot accurately reconstruct them, meaning the input sample and the reconstructed sample of the unknown class have low similarity. This invention explores the similarity between the input sample and the reconstructed sample, classifying highly similar samples as known samples and low-similarity samples as unknown samples.

[0058] Step 2.2: Design and training of classification networks and self-supervised networks.

[0059] The classification network is based on the ResNet network, which consists of six residual blocks, each composed of two convolutional layers and a cross-layer connection. The cross-layer connection directly adds the input features to the output, allowing gradients to propagate directly to earlier layers during forward propagation, thus avoiding information loss and addressing the performance degradation problem in deep neural networks. The input layer has a dimension of 1×N, where N is the length of the sample data, and the output layer has a dimension of M, where M is the number of candidate sets. The final activation function of the classification network is softmax, while the activation functions of other layers are ReLU. A self-supervised network acts as an aid to the classification network, randomly transforming the two-dimensional input data and predicting the transformation patterns using a self-supervised approach, prompting the network to learn the intrinsic structure and latent features of the data at a deeper level. A total of 14 random transformations are used, each obtained by flipping and rotating the two-dimensional data.

[0060] During training, the sample data is first augmented:

[0061]

[0062] Where x is the original data, This is the reconstructed data after passing through a generative network. z represents x and... The merged augmented data is then input into the classification network to obtain the classification loss l. c If we introduce self-supervised loss l ss Then, a transformation needs to be randomly selected from the available transformation set, and the transformed data is used as the augmented input:

[0063]

[0064] Where t(x) is the output of the original data after random transformation, and the transformation exponent is r. Let t(x) be the reconstructed data after passing through the generator network. Treat r as the true label and calculate l using cross-entropy. ss The joint training loss function l of classification networks and self-supervised networks t for:

[0065] l t =α1l c +α2l ss

[0066] Where α1 and α2 are the weights of the classification network and the self-supervised network, respectively. In System 1, in order to give more importance to the classification task, α1 = 0.8 and α2 = 0.2 are chosen.

[0067] Step 3: Design the closed set identification algorithm based on traditional communication algorithms in System 2.

[0068] The closed-set identification algorithm in System 2 is based on traditional communication theory and requires the design of specific signal identification algorithms for specific identification tasks. Taking channel coding identification as an example, the traditional closed-set identification algorithm needs to first set a candidate set of coding parameters [K1, K2, K3, K4, ..., K]. n Then, based on the encoded sequence and the encoder, the check equation is calculated. Based on the check equation, the corresponding metric value [d1, d2, d3, d4, ..., d] is calculated by traversing the parameters of each candidate set. n ], where d i Let be the metric value corresponding to the parameters of the i-th candidate set, and the parameter corresponding to the maximum metric value is the final recognition result. i Calculation methods typically include least squares optimization, log-likelihood ratio optimization, and cosine similarity method.

[0069] Step 4: Design the open set recognition algorithm based on traditional communication algorithms in System 2.

[0070] Traditional open-set recognition algorithms also require specific recognition methods designed for specific recognition tasks. A significant difference from closed-set recognition algorithms is that open-set recognition does not have any candidate sets; it identifies various parameters of the signal solely based on the signal sample data itself. Therefore, open-set recognition algorithms typically have extremely high complexity. Taking the open-set recognition task of Polar codes in the field of channel coding recognition as an example, the goal of this task is to identify Polar code parameters, including code length and frozen bit positions. The open-set recognition algorithm for Polar codes needs to traverse from the minimum code length to the maximum code length. Each traversal requires calculating the codeword matrix X and the Kronecker matrix D for the current code length. Then, based on X and D, the dual vector h for different coding positions is derived. The positions of information bits and frozen bits are determined by comparing a specific statistic γ and a threshold value Λ. Finally, after the traversal is complete, the code length value is obtained based on the special relationship between the Polar code rate and code length. The entire open-set recognition algorithm involves a large number of traversal calculations and matrix operations, resulting in high complexity and time consumption.

[0071] Step 5: Combine System 1 and System 2 and test their performance. The overall system architecture after combination is as follows: Figure 4 As shown.

[0072] A dual-system wireless signal identification framework is constructed by combining a data-driven System 1 and a knowledge-driven System 2. System 1 is responsible for the initial classification of known samples and the detection of unknown samples, while System 2 is responsible for the precise classification of known samples and the parameter identification of unknown samples. Let the candidate set parameters of the known classes be [K1, K2, K3, K4, ..., K]. n The specific workflow is as follows:

[0073] In System 1, sample data is input into a generator network for data augmentation. The augmented data is then passed through a classification network and detected as either a known or unknown sample. If a sample is detected as a known sample, System 1 outputs a preliminary identification result for that known sample. This result includes the confidence score [b1, b2, b3, b4, ..., b] for each parameter in the candidate set. n ], where b i represents the confidence level corresponding to the i-th candidate set parameter, with a value between 0 and 1. b i The higher the value, the higher the confidence level, and the more reliable the classification result. Assuming the three parameters with the highest confidence are [K...] t1 ,K t2 ,K t3 Then System 1 will combine the original samples and the high-confidence candidate set [K] t1 ,K t2 ,K t3 Both samples are input into the closed-set recognition module of System 2 for precise recognition; if the sample is detected as an unknown sample, System 1 directly inputs the original sample into the open-set recognition module of System 2 for specific parameter recognition.

[0074] In System 2, if the input sample is a known class sample, the closed set identification module sequentially detects the high-confidence candidate set [K] provided by System 1. t1 ,K t2 ,K t3 The various parameters in System 2 are used to arrive at a more accurate recognition result. Compared with traditional closed-set recognition algorithms, the closed-set recognition algorithm in System 2 does not need to traverse and calculate the entire candidate set; it only needs to further check the high-confidence candidate set provided by System 1, which significantly reduces the recognition time. If the input sample is an unknown sample, the open-set recognition module identifies the specific parameters of the sample. For channel coding, the parameters of the sample include code length, code rate, generator polynomial, etc. After System 2 completes the identification of unknown sample parameters, it assigns a label to the sample and incorporates it into the training dataset of System 1, thereby strengthening the dataset and improving the performance of System 1. It is worth noting that in traditional methods of wireless signal recognition, due to the difference in conditions between fully blind recognition and semi-blind recognition (with the existence of a candidate set), open-set recognition has higher computational complexity and lower recognition accuracy than closed-set recognition.

[0075] This invention proposes a general dual-system framework for closed-set and open-set identification of wireless signals. The proposed dual-system framework consists of two subsystems: System 1 comprises a classification network, a generator network, and a self-supervised network based on deep learning methods, used for preliminary classification of known samples and identification of unknown samples; System 2 comprises a closed-set identification module and an open-set identification module based on traditional communication methods, used for precise classification of known samples and identification of specific parameters of unknown samples, respectively. The proposed dual-system framework effectively improves the identification accuracy of known samples, reduces identification time, accurately detects unknown samples and identifies their specific parameters, and achieves optimization and upgrades between the systems through sequential collaborative operation.

Claims

1. A wireless signal identification method employing a dual-system architecture, the method comprising: Step 1: Generate datasets for training and testing; Step 1.1: Generate dataset samples; Define the wireless signal identification task, generate a parameter candidate set according to different tasks, generate wireless signal data samples using hardware and software collaborative tools, and assign them corresponding labels; Step 1.2: Dataset partitioning; Generate a certain number of samples and divide the samples into training set, validation set and test set; Step 2: Design and train System 1. System 1 includes a generator network, a classification network, and a self-supervised network. The input data first passes through the generator network. The output data of the generator network is reconstructed from the input data and used as the input of the classification module. The output of the classification module passes through the self-supervised network to obtain the preliminary identification results of the samples. Step 2.1: Design and training of the generative network; The generator network employs a deep autoencoder / decoder architecture, consisting of an encoder network E(x), a decoder network G(x), and a discriminator network D. Here, x represents the original input sample. The encoder network E(x) comprises a convolutional network and ResNet encoding blocks, the decoder network G(x) comprises a decoder block and a deconvolutional network, and the discriminator network D is trained using a Wasserstein discrimination mechanism with gradient penalty. These three parts are jointly trained, with the optimization objective being to minimize the reconstruction error ||xG(E(x))||1. During training, the loss function L is continuously and alternately optimized. D and L G : in, This is the interpolation gradient penalty term, where λ is the gradient penalty weight coefficient. This parameter needs to be adjusted experimentally to balance the influence of the original loss function and the gradient penalty term. This is the gradient operator, used to calculate the gradient value of a vector; For the original sample x and the generated sample The difference sample is calculated using the following formula: Where ε is a random number between 0 and 1, and when ε is close to 1, the difference sample Closer to the original sample x; when ε approaches 0, the difference sample... Closer to the generated sample During the training of generative networks Always between x and This constrains the function space of the discriminator D, mitigating fluctuations during training and promoting network convergence. Step 2.2: Design and training of classification networks and self-supervised networks; First, the sample data is augmented. The augmented data input for the classification network is... x is the original data, and t(x) is the output of x after random transformation, with the transformation exponent being r. The data is the reconstructed data of t(x) after passing through the generator network; the classification loss l is calculated by comparing the true label and the predicted label of x. c Treating r as the true label and using cross-entropy to calculate the self-supervised loss l ss The joint loss function of the classification network and the self-supervised network is l t =α1l c +α2l ss , where α1 and α2 are weighting coefficients; Step 3: Design System 2, which includes a closed set recognition module and an open set recognition module. If the sample detected by System 1 is a known sample, the closed set recognition module is used for further accurate recognition. If the sample detected by System 1 is an unknown sample, the original sample is input into the open set recognition module of System 2 for specific parameter recognition. Step 4: Combine System 1 and System 2 and test their performance; A dual-system wireless signal identification framework is constructed by combining a data-driven System 1 and a knowledge-driven System 2. System 1 is responsible for the preliminary classification of known samples and the detection of unknown samples, while System 2 is responsible for the precise classification of known samples and the parameter identification of unknown samples. Let the candidate set parameters of the known classes be [K1, K2, K3, K4, ..., K]. n ]; In System 1, the augmented data is detected as either a known sample or an unknown sample by the classification network; if the sample is detected as a known sample, System 1 outputs the confidence score [b1,b2,b3,b4,……,b] for each parameter in the candidate set. n ], where b i This represents the confidence level corresponding to the parameters of the i-th candidate set, with a value between 0 and 1; System 1 combines the original samples and the candidate set [K] with confidence levels higher than the threshold. t1 ,K t2 ,K t3 Both are input into the closed set recognition module of System 2 for precise recognition; if the sample is detected as an unknown sample, System 1 directly inputs the original sample into the open set recognition module of System 2 for specific parameter recognition; In System 2, if the input sample is a known class sample, the closed set recognition module detects the candidate set [K] provided by System 1 one by one. t1 ,K t2 ,K t3 The system identifies the parameters in the dataset and then uses them to obtain more accurate recognition results. If the input sample is an unknown sample, the system relies on the open set recognition module to identify the specific parameters of the sample. After the system completes the identification of the parameters of the unknown sample, it assigns a label to the sample and then incorporates it into the training dataset of the system.

2. The wireless signal identification method employing a dual-system architecture as described in claim 1, characterized in that, In step 2.2, α1 = 0.8 and α2 = 0.

2.

3. The wireless signal identification method employing a dual-system architecture as described in claim 1, characterized in that, In System 2, the specific method in the closed set identification module is as follows: First, define the candidate set of encoding parameters [K1, K2, K3, K4, ..., K]. n Then, based on the encoded sequence and the encoder, the check equation is calculated. Based on the check equation, the corresponding metric value [d1, d2, d3, d4, ..., d] is calculated by traversing the parameters of each candidate set. n ], where d i Let be the metric value corresponding to the parameters of the i-th candidate set, and the parameter corresponding to the maximum metric value is the final recognition result; The specific method in the open set recognition module is as follows: The open-set identification method for Polar codes in the field of channel coding identification is as follows: The open-set identification algorithm for Polar codes needs to traverse from the minimum code length to the maximum code length. Each traversal requires calculating the codeword matrix X and the Kronecker matrix D under the current code length. Then, based on X and D, the dual vector h of different coding positions is obtained. The positions of information bits and frozen bits are determined by comparing a specific statistic γ and a threshold value Λ. Finally, after the traversal is completed, the code length value is obtained based on the special relationship between the Polar code rate and the code length.

Images