Radio frequency electromagnetic interference open set identification method based on lightweight generator group

Abstract

The invention discloses a radio frequency electromagnetic interference open set identification method based on a lightweight generator group. The method comprises the steps that 1, a data set is constructed, a radio frequency electromagnetic interference data set is divided into known interference and unknown interference, the known interference is divided into a training set, a verification set and a test set, and the unknown interference serves as the test set; 2, applying the training set and the verification set of the known interference to unknown interference generation and model training to obtain an optimal model; 3, the optimal model serves as the classification model in the step 3, the recognition performance of the classification model is detected through the test set part of the known interference and the unknown interference, and the classification model is used for recognizing and classifying the known interference and effectively recognizing the unknown type interference at the same time. The method has the characteristics that the calculation overhead for generating the unknown interference is small, multiple types of interference are generated, and the unknown types of interference can be effectively identified.

Description

Technical Field

[0001] This invention belongs to the technical field of open set identification methods, specifically relating to an open set identification method for radio frequency electromagnetic interference based on a lightweight generator set. Background Technology

[0002] Large radio telescopes possess extremely high system sensitivity and observation bandwidth, making them highly sensitive to weak signals from space and easily susceptible to interference from other services. With the increasingly frequent use of spectrum resources, wireless communication technologies, including GPS and television broadcasting, are increasingly interfering with radio astronomy. The introduction of advanced electronic equipment, such as digital and intelligent products, into radio observatories further exacerbates the complexity of the electromagnetic environment. Existing technology proposes a method for detecting and identifying resident signals (publication number CN113992281A), but this method is limited to identifying known resident interference and does not extend to unknown interference.

[0003] In reality, with the increasing complexity of the electromagnetic environment, radio frequency interference faced by radio astronomy observations presents new challenges: on the one hand, the signal characteristics of traditional interference sources are constantly evolving with the iteration of communication technologies; on the other hand, unknown types of interference introduced by emerging electronic devices are constantly emerging. This dynamically changing interference environment fundamentally limits existing identification methods based on the closed-world assumption, making them unable to effectively cope with unknown types of interference signals that may exist in the observation data.

[0004] Therefore, it is urgent to establish an interference identification system with open set recognition capabilities. By constructing an open set feature space for interference signals, it is possible to accurately identify known interference types while reliably identifying unknown interferences that exceed the scope of prior knowledge, thereby providing a key guarantee for robust observations of radio telescopes in complex electromagnetic environments.

[0005] Current open set recognition algorithms are mainly divided into two categories: recognition algorithms based on discriminative models and generative models.

[0006] Open-set recognition methods based on discriminative models compare output logits with a threshold to reject unknown input images and classify known input images. However, these methods suffer from several significant drawbacks in practical applications. First, they heavily rely on a pre-set, fixed threshold, the selection of which often lacks theoretical guidance and requires repeated adjustments based on human experience, making them difficult to adapt to different datasets and task requirements. Second, since discriminative models are typically trained under the assumption of closed sets, their output logits have limited ability to distinguish between known and unknown classes, leading to misclassification in open-set scenarios. These limitations make traditional discriminative models unsuitable for meeting the interference recognition needs of radio astronomy in complex and dynamic electromagnetic environments.

[0007] Generative model-based open-set recognition methods utilize generative models to generate unknown samples or features, then train the network using these generated unknown samples and known samples together, effectively transforming the open-set recognition problem into a pseudo-closed-set recognition problem. While this method expands recognition capabilities by simulating unknown samples, it still has significant limitations in practical applications. Existing methods typically use only a single generator or mechanism to simulate unknown samples, and this simple generation approach struggles to adequately cover the complex interference patterns that may exist in real open environments. Due to the diverse nature of astronomical radio frequency interference signals, samples generated by a single generator often fail to accurately reflect the diversity of real unknown interference, resulting in limited effectiveness in reducing open-space risks. Summary of the Invention

[0008] In order to overcome the shortcomings of the existing technology, the present invention aims to provide a radio frequency electromagnetic interference open set identification method based on a lightweight generator set. This method has the characteristics of low computational overhead in generating unknown interference, generating a variety of types, and effectively identifying unknown types of interference.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for identifying open-set radio frequency electromagnetic interference based on a lightweight generator set includes the following steps; Step 1: Construct the dataset. Divide the radio frequency electromagnetic interference dataset into two parts: known interference and unknown interference. The known interference is further divided into training set, validation set, and test set, while the unknown interference is used as the test set. Step 2: Use the training and validation sets of known interferences to generate unknown interferences and train the model to obtain the optimal model; Step 3: The best model is used as the classification model. The recognition performance of the classification model is tested by the test set of known interference and unknown interference. The classification model is used to identify and classify known interference, and also to effectively identify unknown interference.

[0010] Step 1 specifically involves: Radio frequency electromagnetic interference data measured by Xinjiang Astronomical Observatory were selected, and interference data were extracted according to the business frequency band division and time-frequency diagrams were plotted. The service frequency bands were initially divided according to the L-band radio frequency allocation table, and the service frequency bands were refined according to the time-frequency diagram of the measured data. Clear interference signals that appeared multiple times in different azimuth angles were selected. A portion of the extracted service frequency bands is selected as known interference, and the other portion is selected as unknown samples. The signal-to-noise ratio of the above samples is -5dB. Unknown samples are not used in the training and validation of the network; they are used entirely in the testing phase.

[0011] Step 2 specifically involves: (1) Generation of unknown interference An unknown disturbance is generated using a generator set, which includes G1, G2, ... G... N There are N generators in total, used to generate unknown interference samples; The observatory's measured interference data, divided into known interference frequency bands, are used as input to each round of the generator. For each type of known interference C... i Each round of the generator generates corresponding unknown samples, and N rounds of the generator generate N types of unknown samples, thereby achieving an open space that covers as many potential unknown models as possible around the boundary of the known space.

[0012] As a subsequent block to the known samples, the generator receives the time-frequency plot from the known samples and generates unknown interference. This generator has only two layers of neural network, and each training round generates one type of unknown interference, forming a generator group, which enables the generation of multiple types of unknown interference in each training round. The diversity is gradually increased by limiting the difference between the newly generated and old unknown interference in each iteration; a set of generators is trained for each type of known samples.

[0013] The loss function of the generator is as follows: in The input sample has known interference. It is an unknown interference generated by the current generator. It is an unknown interference generated by a randomly selected old generator. Used to constrain the similarity between generated unknown disturbances and known input disturbances. Used to amplify the differences between newly generated and old unknown disturbances; (2) Model training The generator and the feature extraction module together constitute the open set recognition model. The known class samples and the unknown interference samples generated by the generator in each round are input into the feature extraction module to extract features. The specific operation steps are as follows: The input data first passes through the initial convolutional layer and pooling layer to initially extract features and reduce spatial dimensions; then, the data passes through the residual block layer to convert the feature map into a 512-dimensional feature vector, and finally through two fully connected layers. The first fully connected layer converts the 512-dimensional feature vector into a 128-dimensional feature vector, which is used to calculate the center feature vector of each class of samples and to draw the TSNE diagram. The second fully connected layer converts the 128-dimensional feature vector into a 3-dimensional feature vector and generates the class prediction probability value through the Sigmoid activation function.

[0014] The loss function for the open set recognition model is as follows: in The function is as follows: A composite loss function is used to jointly optimize the open set recognition model, where, It is the standard cross-entropy loss function, applied to the classification output of samples with known categories; its core purpose is to ensure the accuracy of the model's classification of known categories by minimizing the difference between the predicted probability distribution and the true label distribution, thus driving the model to learn the correct decision boundary for the known categories. It is the core triplet loss function, which operates on the model's embedded feature space; each The calculation of the term is based on the sampled triplet. , from anchor sample Positive samples (i.e., with anchor sample) (same category samples) and negative samples composition; This loss function is used to minimize ,in[.] + = max(0,·) represents the Hingle loss function, where m is the violation margin, requiring that the distance between negative pairs be at least a margin m greater than the distance between positive pairs. This forces a reduction in the distance between samples of the same class while increasing the distance between samples of different classes, thereby achieving a compact feature space within each class, separation between classes, and a wide interval between each class guaranteed by the margin m. The above loss function is used to constrain the model training direction, and the optimal model in each training process is saved.

[0015] Step 3 specifically involves: The optimal model obtained during the training process in step 2 is selected, and the features of the validation set samples are extracted using the optimal model. The feature space is used as the benchmark feature space for testing the classification module. The nearest neighbor distance method is selected, and the unknown interference detection is achieved by analyzing the sample distribution characteristics in the feature space.

[0016] The specific steps are as follows: First, the nearest neighbor distance between each sample in the test set and the sample in the validation set is calculated in the feature space. Based on this characteristic, the nearest neighbor distances of all test samples are sorted. The decision threshold is automatically determined by detecting abrupt changes in the sorted distance sequence. These abrupt changes reflect the natural boundary between known and unknown categories in the distance distribution. Finally, samples with nearest neighbor distances greater than the threshold are judged as unknown interference, while samples with distances less than the threshold are classified into the known category to which their nearest neighbor belongs.

[0017] A radio frequency electromagnetic interference open set identification system based on a lightweight generator set, the system includes an unknown interference generator set, a feature extractor and a classification module; Among them, the unknown interference generator group generates diverse unknown samples through a progressive interference simulation strategy, effectively expanding the coverage of the open space; the feature extractor and classification module are based on the ResNet network and achieve feature representation of radio frequency electromagnetic interference data signals through deep residual learning. The feature extractor and classification module achieve end-to-end optimization through a collaborative training mechanism: the unknown samples generated by the generator group and the real known samples together constitute the training set, driving the feature extractor to learn sensitive discriminative features for unknown interference while maintaining high classification accuracy for known interference.

[0018] The beneficial effects of this invention are: This invention proposes an adaptive thresholding mechanism based on the geometric properties of feature space: leveraging the inherent properties of "compactness within classes and separation between classes," it automatically detects statistical abrupt changes in the sorted array by analyzing the nearest neighbor distance distribution of known samples to determine the optimal threshold. This method requires no manual intervention, is mathematically interpretable, and its dynamic adaptability significantly improves the robustness of the model in open environments. Furthermore, it requires only one efficient sorting to achieve the decision, resulting in extremely low computational cost.

[0019] This invention innovatively designs a lightweight generator set that collaboratively generates a wider range of diverse unknown interference samples around the known class decision boundary, reducing open space risks. The generator set has a lightweight overall structure, controlling computational overhead while ensuring generation diversity; its simulation process simultaneously optimizes the feature space—enhancing both the intra-class compactness of known class samples and strengthening the inter-class separability between known classes and between known and unknown classes, thus compressing the uncertainty of the decision boundary from the source. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall technical process of the present invention.

[0021] Figure 2 This is a diagram of the generator model structure.

[0022] Figure 3 This is a structural diagram of the feature extraction module.

[0023] Figure 4 This is a schematic diagram of the confusion matrix for the test set. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings.

[0025] The present invention discloses an open-set radio frequency electromagnetic interference identification system based on a lightweight generator set, as shown in the attached figure. Figure 1As shown, it mainly consists of three parts: an unknown interference generator group, a feature extractor, and a classification module. Among them, the generator group generates diverse unknown samples through a progressive interference simulation strategy, effectively expanding the coverage of the open space. The ResNet network, as the core module for feature extraction and classification, realizes the feature representation of the interference signal through deep residual learning. The two modules achieve end-to-end optimization through a collaborative training mechanism: the unknown samples generated by the generator group and the real known samples together constitute the training set, driving the feature extraction network to learn sensitive discriminative features for unknown interference while maintaining high classification accuracy for known interference.

[0026] A method for identifying open-set radio frequency electromagnetic interference based on a lightweight generator set includes the following steps; Step 1: Building the dataset The technique used in this invention is interference data measured by the Xinjiang Astronomical Observatory. Interference data is extracted based on the service frequency bands and time-frequency diagrams are plotted. The extracted service frequency bands are 1310-1412MHz, 1429-1448MHz, 1666-1700MHz, and 1746-1765MHz. The first three categories are considered known interference, and the last category is considered unknown. The signal-to-noise ratio of these samples is -5dB. The number of samples in each of the three known categories is 192, which are divided into training, validation, and test sets in a 3:3:1 ratio. The number of unknown samples is 32, which are not used in the training and validation processes and are used entirely for the testing phase.

[0027] Step 2: Generation of unknown interference and model training (1) Generation of unknown interference An unknown disturbance is generated using a generator set. The generator model structure diagram is shown in the attached figure. Figure 1 , Figure 2 As shown, where G1, G2...G N There are N generators used to generate unknown interference samples; C0, C 1、 C2 consists of three types of known samples, corresponding to observatory-measured interference data in the 1310-1412MHz, 1429-1448MHz, and 1666-1700MHz service frequency bands, respectively. These samples serve as input to the generator in each round. For each type of known interference C... i Each round of the generator can generate corresponding unknown samples, and N rounds of the generator can generate N types of unknown samples, thereby achieving an open space that covers as many potential unknown models as possible around the boundary of the known space.

[0028] As a follow-up block to the known samples, the generator receives the time-frequency graph from the known samples to generate unknown interference. This generator has only two neural network layers, making it smaller in size compared to common generative models such as GANs, and significantly reducing computational overhead when generating unknown interference. Furthermore, unlike common generative models that can only generate one type of unknown interference per model and have a small simulated open space, this generator can generate one type of unknown interference in each training iteration, forming a generator group that can generate multiple types of unknown interference in each training iteration. To avoid the similarity of the simulated open space generated by each generator when training generators independently, a set of generation patterns is trained progressively. Diversity is gradually increased by limiting the difference between new and old generated unknown interference in each iteration. A set of generators is trained for each type of known samples, constructing as many unknown interference samples as possible based on the known samples, thereby covering as much of the potential unknown model's open space as possible around the boundary of the known space.

[0029] The generator loss function is as follows: in The input sample has known interference. It is an unknown interference generated by the current generator. It is an unknown disturbance generated by a randomly selected old generator. Used to constrain the similarity between generated unknown disturbances and known input disturbances. Used to amplify the differences between newly generated and old unknown disturbances.

[0030] (2) Model training The generator and the feature extraction module together constitute the open set recognition model. The three known classes of samples and the unknown interference samples generated by the generator in each round are input into the feature extraction module to extract features. Feature extraction module as attached Figure 3As shown. The specific process of the feature extraction module is as follows: The input data first passes through an initial convolutional layer (Conv2d, kernel size 7×7, stride 2, padding 3) for preliminary feature extraction, and then passes through a max pooling layer (Maxpool, pooling kernel size 3×3, stride 2, padding 1) to reduce the spatial dimension; then, the data passes through four sets of residual blocks for deep feature enhancement. The first three sets of residual blocks each contain three consecutive 3×3 convolutions (stride 1, padding 1) and one 3×3 convolution for reducing the spatial dimension (stride 2, padding 1). The fourth set of residual blocks completes the final feature extraction through three consecutive 3×3 convolutions (stride 1, padding 1). The output feature map is converted into a 512-dimensional feature vector by global average pooling (avgpool); to meet the needs of subsequent analysis and prediction, the 512-dimensional feature vector is first converted into a 128-dimensional feature vector by the first fully connected layer (FC (512, 128)) to calculate the center feature vector of each class of samples and to draw TSNE. Visualization diagram. The second fully connected layer (FC (128, 3)) reduces the dimensionality to a 3D feature vector, and finally, the Sigmoid activation function is used to generate the class prediction probability value.

[0031] The loss function for the open set recognition model is as follows: in The function is as follows: This method employs a composite loss function to jointly optimize the open set recognition model. Among these, It is the standard cross-entropy loss function, applied to the classification output of samples with known categories. Its core purpose is to ensure the accuracy of the model's classification of known categories by minimizing the difference between the predicted probability distribution and the true label distribution, thereby driving the model to learn the correct decision boundary for the known categories. It is the core triplet loss function, which operates on the model's embedded feature space. Each The calculation of the term is based on the sampled triplet. , from anchor sample Positive samples (i.e., with anchor sample) (same category samples) and negative samples Composition. The loss function aims to minimize ,in[.] += max(0,·) represents the Hingle loss function, where m is the violation margin. It requires that the distance between negative pairs be at least m greater than the distance between positive pairs, thereby forcibly reducing the distance between samples of the same class while increasing the distance between different classes. This achieves a compact feature space within classes, separation between classes, and a wide margin guaranteed by the margin m between each class. This structure ensures that during the testing phase, samples belonging to known classes will fall into their corresponding compact clusters and be accurately classified, while truly unknown samples are more likely to fall into unknown regions far from all known clusters, thus being effectively identified as unknown and rejected.

[0032] Step 3: Known / Unknown Interference Classification Module The optimal model selected during training is used to extract features from the validation set samples. This feature space is then used as the baseline feature space for testing the classification module. Since a well-trained feature space is compact within classes and separate between classes, the nearest neighbors of known class samples are usually from the same class and are relatively close. However, unknown class samples, not having participated in training, will have feature representations that significantly deviate from the feature distribution of known classes, causing them to fall far from known clusters in the feature space, resulting in significantly larger nearest neighbor distance values. Therefore, this invention proposes a nearest neighbor distance method to detect unknown interference by analyzing the sample distribution characteristics in the feature space. The core idea of ​​this method is as follows: First, calculate the nearest neighbor distance between each sample in the test set and the validation set samples in the feature space. Based on this characteristic, sort the nearest neighbor distances of all test samples. Automatically determine the decision threshold by detecting abrupt changes in the sorted distance sequence—this abrupt change reflects the natural boundary between known and unknown classes in the distance distribution. Finally, samples with nearest neighbor distances greater than the threshold are judged as unknown interference, while samples with distances less than the threshold are classified into the known class to which their nearest neighbors belong. This method does not require a preset threshold and can adaptively identify unknown interference while maintaining high recognition accuracy for known categories.

[0033] The classification results of the model are presented in the form of a confusion matrix, as shown in the attached figure. Figure 4 As shown in the figure, the horizontal axis represents the predicted label, the vertical axis represents the true label, and the diagonal elements represent the correctly classified results. The test set has four classes: labels 0, 1, and 2 represent the three known classes, and label 3 represents the unknown class. Each class has 32 samples. As shown in the figure, the classification accuracy for both the known and unknown classes is 100%.

[0034] To address the shortcomings of traditional open set recognition methods in simulating open space, this invention proposes a lightweight generator set. This network simulates various disturbances around known disturbance boundaries, more broadly simulating open space and reducing open space risk. Simultaneously, it enables the network to learn known samples more comprehensively, enhancing classification capabilities between known classes and between known and unknown classes, as well as the compactness within known classes. Furthermore, the lightweight design of this generator network reduces computational burden and improves adaptability to different scenarios. The lightweight generator set's open space simulation mechanism employs a set of progressive lightweight generators to simulate and generate various types of unknown disturbances around the known class decision boundaries, resulting in a wider coverage of unknown open space samples or features, effectively reducing open space risk. Simultaneously, this mechanism promotes feature space learning, enhances the intra-class compactness and inter-class separability of known class samples, better matches adaptive threshold selection methods, and thus improves the accuracy of unknown and known class recognition.

[0035] To address the shortcomings of traditional open set identification methods that rely on manually set thresholds, this invention proposes an adaptive threshold selection method. This method, based on the geometric characteristics of the feature space (compactness within classes and separation between classes), automatically determines the optimal threshold by analyzing the statistical distribution of nearest neighbor distances, eliminating the need for manual intervention. Theoretically, it utilizes the differences in distance distributions between known and unknown classes to achieve objective decision-making by detecting abrupt changes. In practical applications, only a single sorting operation is required to complete the threshold decision, significantly reducing computational complexity. This data-driven approach improves model adaptability, ensures decision interpretability, and provides a more reliable and efficient open set identification scheme: Based on the inherent geometric characteristics of the feature space—compactness within classes and separation between classes—this method automatically determines the optimal threshold by analyzing the statistical distribution of nearest neighbor distances, requiring no manual intervention. Specifically, it utilizes the inherent differences in distance distributions to detect abrupt changes in the distance sorting array for objective decision-making, achieving high computational efficiency with only one sorting operation, and improving the model's adaptability and decision interpretability in dynamic open environments.

Claims

1. A radio frequency electromagnetic interference open set identification method based on a lightweight generator set, characterized in that, Comprise the following steps: Step 1: build a dataset, divide the radio frequency electromagnetic interference dataset into two parts of known interference and unknown interference, divide the known interference into training set, validation set and test set, and the unknown interference as test set; Step 2: use the training set and validation set of known interference to generate unknown interference and model training, and get the best model; Step 3: the best model is used as a classification model, and the test set of known interference and unknown interference are used to detect the identification performance of the classification model, and the classification model is used to identify and classify known interference and unknown interference.

2. The method of claim 1, wherein, The step 1 is specifically: Select the measured radio frequency electromagnetic interference data, extract the interference data according to the business frequency band division and draw the time-frequency graph; According to the L-band radio frequency division table, the business frequency band is preliminarily divided, and the business frequency band is refined according to the measured data time-frequency graph, and the interference signal which appears clearly and in different azimuth angles is selected; Select a part of the extracted business frequency band as known class interference, and the other part as unknown class sample.

3. The method of claim 2, wherein the method is based on a lightweight generator set. The step 2 is specifically: (1) Unknown interference is generated by a generator set, which includes G1, G2, …, G N N generators in total; The astronomical observatory measured interference data of the service frequency band divided into known interference is taken as the input of each round generator, and for each type of known interference C i Each round generator generates a corresponding unknown sample, and N round generators generate N types of unknown samples; The generator receives the time-frequency graph from the known sample to generate unknown interference; (2) the generator and the feature extraction module constitute an open set recognition model, and the known class interference sample and the unknown interference sample generated by each round of generator are used as input data to input the feature extraction module for feature extraction.

4. The method of claim 3, wherein the method is based on a lightweight generator set. In the step (1), the generator is a two-layer neural network, and the generator generates a kind of unknown interference in each training, which constitutes a generator group, and realizes the generation of multiple unknown interferences in each training; the diversity is gradually increased by limiting the difference between new and old generated unknown interferences in each iteration; a group of generators is trained for each known sample.

5. The method of claim 4, wherein the method is based on a lightweight generator set. The loss function of the generator is as follows: wherein is the input sample of known interference, is the unknown interference generated by the current generator, is the unknown interference generated by a randomly selected old generator, is used to constrain the similarity between the generated unknown interference and the known input interference, is used to enlarge the difference between the new and old generated unknown interference.

6. The method of claim 5, wherein the method is based on a lightweight generator set. The specific operation steps of the step (2) are as follows: The input data first passes through the initial convolution layer and the pooling layer, preliminarily extracts the features and reduces the spatial dimension; then, the data is converted into a 512-dimensional feature vector through the residual block layer, and finally, through two fully connected layers, the first fully connected layer converts the 512-dimensional feature vector into a 128-dimensional feature vector, which is used for subsequent calculation of the center feature vector of each class sample and drawing TSNE graph, and the second fully connected layer converts the 128-dimensional feature vector into a 3-dimensional feature vector, and generates the class prediction probability value through the Sigmoid activation function.

7. The method of claim 6, wherein the method is based on a lightweight generator set. The loss function of the open set recognition model is as follows: wherein The function is as follows: A composite loss function is adopted to jointly optimize the open-set recognition model, wherein, is a standard cross-entropy loss function acting on the classification output of known-class samples; is a triplet loss function acting on the embedding feature space of the model; each is calculated based on a sampled triplet , which consists of an anchor sample , a positive sample (i.e., a sample of the same class as the anchor sample ), and a negative sample ; The loss function is used to minimize where [.] + = max(0, ·) represents a hinge loss function, m is a margin of violation, requiring the distance of negative pairs to be at least one margin m larger than the distance of positive pairs, which forces the distance between samples of the same class to be reduced, while the distance between different classes is enlarged, achieving a compact within-class and a large margin between classes in the feature space, and a wide gap between classes is guaranteed by the margin m. The above loss function is used to constrain the model training direction, and the optimal model in each training process is saved.

8. The method of claim 7, wherein the method is based on a lightweight generator set. The step 3 is specifically: Select the optimal model obtained in the training process of step 2, extract the features of the validation set samples using the optimal model, and use the feature space as the reference feature space for the test classification module; select the nearest neighbor distance method to realize the unknown interference detection by analyzing the sample distribution characteristics in the feature space.

9. The method of claim 8, wherein the method is based on a lightweight generator set. The specific operation is: First, the distance between each sample in the test set and the nearest neighbor of the validation set samples in the feature space is calculated. Based on this characteristic, the nearest neighbor distances of all test samples are sorted, and the decision threshold is automatically determined by detecting the mutation point in the distance sequence after sorting, which reflects the natural demarcation of known and unknown categories in the distance distribution. Finally, the samples with a nearest neighbor distance greater than the threshold are determined as unknown interference, and the samples with a distance less than the threshold are classified as the known category to which the nearest neighbor belongs.

10. A radio frequency electromagnetic interference open set identification system based on a lightweight generator set for implementing the method of any one of claims 1-9, characterized in that, The system comprises an unknown interference generator group, a feature extractor, and a classification module. The unknown interference generator group generates diversified unknown samples through a progressive interference simulation strategy, effectively expanding the coverage of the open space. The feature extractor and the classification module are based on the ResNet network and realize feature representation of radio frequency electromagnetic interference data signals through deep residual learning. The feature extractor and the classification module realize end-to-end optimization through a cooperative training mechanism: The unknown samples generated by the unknown interference generator group and the real known samples jointly constitute the training set, which drives the feature extractor to learn sensitive discriminative features for unknown interference while maintaining high classification accuracy for known interference.

Images