Anti-data missing radio station individual identification method and system based on knowledge graph enhancement
By converting electromagnetic signals into structured knowledge graphs and using large language models for feature encoding, the problems of interference resistance and data loss in complex electromagnetic environments for radio station identification technology are solved, achieving high-precision and robust individual radio station identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2026-01-19
- Publication Date
- 2026-04-28
AI Technical Summary
Existing radio station identification technologies have poor anti-interference capabilities in complex electromagnetic environments, and their identification accuracy is low, especially when data is missing, making it difficult to effectively utilize the correlation information between signals.
A knowledge graph-based augmentation method is adopted to transform electromagnetic signals into structured knowledge graphs. By constructing temporal and attribute associations between signals, feature encoding is performed using a large language model (BERT) to achieve high-precision and robust individual radio station identification.
Even with 10% to 20% missing data, the recognition accuracy can still be maintained above 95%, which is significantly better than traditional methods. The recognition accuracy reaches 98.05%, solving the problem of confusion of multi-source heterogeneous signals in complex environments.
Smart Images
Figure CN121935702A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wireless communication monitoring and electronic countermeasures technology, and relates to a method and system for identifying radio stations with data loss resistance based on knowledge graph enhancement. Background Technology
[0002] As a critical information carrier in the electromagnetic spectrum environment, the identification capability of general-purpose radios directly determines the mission effectiveness in fields such as national defense electronic countermeasures, civilian spectrum supervision, and emergency communication support. In the modern complex electromagnetic environment, radio signals exhibit significant characteristics of multi-source heterogeneity, dynamic changes, and non-uniform distribution. Currently, identification technologies for general-purpose radios are mainly divided into two categories: methods based on statistical features and methods based on deep learning.
[0003] Early radio station identification methods largely relied on machine learning schemes based on statistical features. These methods typically used manually designed features such as signal modulation and pulse parameters, combined with communication rules and protocols within the communication network, to extract features from spectrum monitoring data in the time, frequency, and spatial domains. Commonly used classification algorithms included Support Vector Machines (SVM) and decision trees, which identified radio signals by constructing behavioral pattern classifiers. While these methods possess a degree of interpretability under specific rules, their core limitation lies in their over-reliance on standard communication radio feature data and human prior knowledge. Traditional algorithms have high requirements for data completeness and focus only on local data of the signal itself, failing to uncover potential correlations between different signals. In complex scenarios with multiple radio stations coexisting, the robustness of the algorithms is poor and prone to misjudgments if signal parameters are missing or overlapping.
[0004] In recent years, with the development of artificial intelligence technology, radio station recognition based on deep learning has gradually become a research hotspot. Models such as Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM) have been widely applied in this field. For example, the received radio station IQ signals can be directly used, or the signals can be visualized as two-dimensional images and input into the neural network to achieve end-to-end output from raw features to recognition results. Compared with traditional methods, deep learning solutions reduce reliance on cumbersome preprocessing steps and significantly improve the automation of feature extraction.
[0005] However, existing deep learning methods still face significant technical bottlenecks in practical applications. First, these methods often require a large amount of high-quality data with accurate labels. In real electromagnetic environments, limitations such as electronic interference, equipment noise, and terrain obstruction frequently lead to the "small sample size" problem, signal waveform distortion, or missing data sampling. Second, most current deep learning models still focus on the identification and classification of individual signal features, lacking effective mining of deep relationships between signals (such as temporal logic, behavioral patterns, and interactions). When some signal parameters (such as level and communication cycle) are missing, existing models struggle to compensate using contextual information, resulting in a sharp drop in recognition accuracy and failing to meet the requirements for high-precision and high-reliability recognition. Therefore, there is an urgent need for a radio station identification technology that can overcome the boundaries of single signal information and leverage relational knowledge to enhance resilience against missing data. Summary of the Invention
[0006] In view of this, the purpose of this invention is to provide a method and system for identifying radio stations with data loss resistance based on knowledge graph enhancement. This method and system are based on a general radio station identification framework of a communication knowledge graph association big language model. First, discrete electromagnetic signals are transformed into structured knowledge graphs. Semantic subgraphs are constructed by mining the temporal and attribute associations between signals. Then, the powerful contextual reasoning ability of the big language model (BERT) is used to transform the structured graphs into text descriptions for feature encoding, ultimately achieving high-precision and highly robust radio station identification.
[0007] To achieve the above objectives, the present invention provides the following technical solution: A knowledge graph-based method for identifying radio stations with resistance to missing data, comprising the following steps: S1. The collected raw electromagnetic signals are cleaned, missing values are imputed, and normalized to obtain a standardized signal tensor; S2. Construct a radio signal knowledge graph based on the standardized signal tensor. The knowledge graph includes signal entity nodes, radio entity nodes, and time entity nodes, and constructs attribute relationships and temporal relationships between entities. S3. Using the signal to be identified as the central node, extract its K-hop neighborhood association subgraph in the knowledge graph, and convert the association subgraph into a text sequence described in natural language; S4. Input the text sequence into a pre-trained large language model for semantic feature encoding, and output the radio station category of the signal to be identified, thereby realizing individual radio station identification.
[0008] Furthermore, step S1 specifically includes: S11. Collect raw radio data and clean the collected raw electromagnetic signals (including data from multiple sensors and multiple time windows); S12. Set a physical threshold to remove abnormal samples, and use the K-nearest neighbor algorithm to interpolate and fill in the missing features; S13. Construct a globally consistent time axis and normalize each feature to build a standardized signal tensor. Assuming the original signal feature vector The normalization formula is as follows: in The characteristic mean, The standard deviation is given; the final output is a standard structured data sequence containing frequency, bandwidth, amplitude, occurrence time, duration, and station ID.
[0009] Furthermore, step S2 specifically includes: S21. Convert discrete signals into knowledge graphs ,in For a collection of entities, For a set of relations; the system first instantiates three types of entity nodes: signal nodes ( ), radio nodes ( ) and time nodes ( ); S22. System constructs multidimensional semantic relationships: In addition to basic attribute relationships, dynamic temporal relationships are constructed; the system groups signals by radio station ID and sorts them by time, and for adjacent signals... and Establish bidirectional time-series edges: ; S23. To explicitly quantize the transmission rhythm, the system further calculates the time difference between adjacent signals. Explicitly encode it as an edge attribute of the emitter node. The calculated microsecond intervals are written into the graph as relation attributes to generate triples. .
[0010] Finally, all data is serialized and stored in a graph database, forming an electromagnetic knowledge network with complete temporal logic.
[0011] Furthermore, in step S3, the signal to be identified is used as the central node to extract its... K-hop neighborhood related subgraph Subsequently, a template mechanism is used to transform the subgraph. Convert to natural language description text This mechanism maps structured triples to natural language sentences, and integrates numerical features with topological structures into sequence data that can be understood by a large language model.
[0012] Furthermore, in step S4, the bidirectional encoder representation network BERT is constructed and fine-tuned to achieve accurate recognition, generating text... After WordPiece word segmentation, it is converted into a vector representation E through an embedding layer: The model is set to a maximum input length of 512, and uses a multi-layer self-attention mechanism to capture long-distance dependencies for the input sequence. The formula for calculating attention is: in These are the query, key, and value matrices, respectively.
[0013] Furthermore, in step S4, during the training phase, extract... Tag hidden state vector The data is mapped to the radio station category space through a fully connected layer; during training, the cross-entropy loss function is used for end-to-end fine-tuning, and the objective function is defined as: in The total number of samples, This is a real radio station tag. The model predicts probabilities for the signal; by learning the “contextual semantics” of the signal, the model can deeply understand the contextual semantics and behavioral patterns of radio signals, thus maintaining extremely high recognition robustness even when some data is missing.
[0014] Furthermore, in step S3, the value of K in the K-hop neighborhood association subgraph is 2.
[0015] The present invention also provides a knowledge graph-based augmentation-based system for identifying radio stations that is resistant to missing data.
[0016] The beneficial effects of this invention are as follows: Compared with existing radio station identification technologies, this invention has significant advantages.
[0017] First, this solution demonstrates extremely strong robustness against missing data. Traditional deep learning or statistical methods have extremely high requirements for data integrity; once signal parameters are missing, recognition performance drops sharply. However, this invention introduces a large language model to process sequential text features and leverages BERT's powerful context completion and inference capabilities. Even when the missing data ratio reaches 10% to 20%, the recognition accuracy remains above 95%, significantly outperforming traditional methods.
[0018] Secondly, this solution innovatively mines deep-level correlations between signals by constructing a radio signal knowledge graph. Existing technologies often focus on local features of the signal itself (such as modulation scheme), neglecting the temporal and behavioral relationships between signals. This invention utilizes the associated subgraphs of the graph structure as features, integrating temporal adjacency relationships and time interval relationships, effectively solving the confusion problem of multi-source heterogeneous signals in complex environments. Finally, this solution achieves high recognition accuracy. Experimental results show that, with the dual support of "knowledge graph association + large language model," this method achieves an accuracy of 98.05% on general radio station recognition tasks, far exceeding mainstream algorithms such as Support Vector Machine (67.82%) and LSTM (92.14%).
[0019] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a diagram of the general radio station identification framework based on a knowledge graph-related large language model of the present invention; Figure 2 A rendering of the knowledge graph construction for radio signals; Figure 3 The result of extracting data from the association graph. Detailed Implementation
[0021] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.
[0022] This invention addresses the problems of existing methods failing to effectively utilize inter-signal correlation information and having poor anti-interference capabilities. It proposes a general radio station identification framework based on a large language model associated with a communication knowledge graph. Figure 1This is a diagram of the general radio station identification framework based on a knowledge graph association big language model, as described in this invention. The method first transforms discrete electromagnetic signals into a structured knowledge graph. It then constructs a semantic subgraph by mining the temporal and attribute associations between signals. Finally, leveraging the powerful contextual reasoning capabilities of the Big Language Model (BERT), the structured graph is transformed into a text description for feature encoding, ultimately achieving high-precision and robust radio station identification.
[0023] Specifically, the method provided by this invention includes the following steps: Step 1: Multidimensional Preprocessing and Standardization of Radio Data The system first cleans the acquired raw electromagnetic signals (including data from multiple sensors and multiple time windows). Physical thresholds are set to remove outlier samples (such as frequencies). or amplitude The K-Nearest Neighbors (KNN) algorithm is then used to interpolate and fill in missing features. Subsequently, a globally consistent time axis is constructed, and features such as frequency, bandwidth, and amplitude are normalized to build a standardized signal tensor. Assume the original signal feature vector The normalization formula is as follows: in The characteristic mean, The standard deviation is given. The final output is a standard structured data sequence containing frequency, bandwidth, amplitude, occurrence time, duration, and station ID.
[0024] Step 2: Construction of Radio Signal Knowledge Graph and Generation of Temporal Chains This phase aims to transform discrete signals into knowledge graphs. ,in For a collection of entities, This is a set of relations. The system first instantiates three types of entity nodes: three types of nodes: signal nodes ( ), radio nodes ( ) and time nodes ( ).
[0025] Next, the system constructs multidimensional semantic relationships. In addition to basic attribute relationships (such as...) Besides, the core of this invention lies in constructing dynamic timing relationships. The system groups signals by radio station ID and sorts them by time; for adjacent signals... and Establish bidirectional time-series edges: To explicitly quantize the transmission rhythm, the system further calculates the time difference between adjacent signals. Explicitly encode it as an edge attribute of the emitter node. The calculated microsecond intervals are written into the graph as relation attributes to generate triples. .
[0026] Ultimately, all data is serialized and stored in a graph database, forming an electromagnetic knowledge network with complete temporal logic. Figure 2 A rendering of the knowledge graph construction for radio signals.
[0027] Step 3: Extraction of related subgraphs and conversion of text and images To uncover deep correlations, the algorithm uses the signal to be identified as the central node and extracts its... K-hop neighborhood related subgraph (Experiments show) (For best results). Then, a template mechanism is used to transform the subgraph. Convert to natural language description text This mechanism maps structured triples to natural language sentences, for example, transforming subgraphs into... Then a text example is generated. This transformation integrates numerical features with topological structure into sequence data that can be understood by a large language model. Figure 3 The result of extracting data from the association graph.
[0028] Step 4: Semantic Feature Encoding and Classification Based on BERT Model This invention constructs and fine-tunes a bidirectional encoder representation network (BERT) to achieve accurate recognition. The generated text... After WordPiece word segmentation, it is converted into a vector representation E through an embedding layer: The model has a maximum input length of 512 and utilizes a multi-layer self-attention mechanism to capture long-distance dependencies. For the input sequence... The formula for calculating attention is: in These are the query, key, and value matrices, respectively. During the training phase, the following are extracted: Tag hidden state vector The data is mapped to the radio station category space through a fully connected layer. During training, end-to-end fine-tuning is performed using the cross-entropy loss function, and the objective function is defined as: in The total number of samples, This is a real radio station tag. The model predicts probabilities for the signal. By learning the "contextual semantics" of the signal, the model can deeply understand the contextual semantics and behavioral patterns of radio signals, thus maintaining extremely high recognition robustness even when some data is missing.
[0029] The universal radio station identification method and system developed in this invention can be widely applied in the fields of wireless communication monitoring and electronic warfare. Specifically, in defense electronic warfare scenarios, this technology can assist reconnaissance systems in quickly and accurately identifying enemy radio station models and tactical intentions, grasping the battlefield communication situation in real time, and providing a basis for electronic jamming and tactical decision-making. In the field of civilian spectrum regulation, this technology can be used to efficiently identify illegal radio stations or fake base stations, maintaining normal electromagnetic spectrum order. In emergency communication support situations, especially in disaster relief sites where the electromagnetic environment is harsh due to earthquakes, floods, etc., this technology can quickly locate legitimate radio stations using limited and potentially missing signal data, ensuring the accurate transmission of rescue instructions.
[0030] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications should be covered within the scope of the claims of the present invention.
Claims
1. A knowledge graph-based method for identifying radio stations with resistance to missing data, characterized in that: The method specifically includes the following steps: S1. The collected raw electromagnetic signals are cleaned, missing values are imputed, and normalized to obtain a standardized signal tensor; S2. Construct a radio signal knowledge graph based on the standardized signal tensor. The knowledge graph includes signal entity nodes, radio entity nodes, and time entity nodes, and constructs attribute relationships and temporal relationships between entities. S3. Using the signal to be identified as the central node, extract its K-hop neighborhood association subgraph in the knowledge graph, and convert the association subgraph into a text sequence of natural language description; S4. Input the text sequence into a pre-trained large language model for semantic feature encoding, and output the radio station category of the signal to be identified, thereby realizing individual radio station identification.
2. The method for identifying individual radio stations with data missingness based on knowledge graph enhancement according to claim 1, characterized in that: Step S1 specifically includes: S11. Collect raw radio data and clean the collected raw electromagnetic signals; S12. Set a physical threshold to remove abnormal samples, and use the K-nearest neighbor algorithm to interpolate and fill in the missing features; S13. Construct a globally consistent time axis and normalize each feature to build a standardized signal tensor. Assuming the original signal feature vector The normalization formula is as follows: in The characteristic mean, The standard deviation is given; the final output is a standard structured data sequence containing frequency, bandwidth, amplitude, occurrence time, duration, and station ID.
3. The method for identifying individual radio stations with data missingness based on knowledge graph enhancement according to claim 2, characterized in that: Step S2 specifically includes: S21. Convert discrete signals into knowledge graphs ,in For a collection of entities, For a set of relations, the system first instantiates three types of entity nodes: signal nodes, radio nodes, and time nodes; S22. System constructs multidimensional semantic relationships: In addition to basic attribute relationships, dynamic temporal relationships are constructed; the system groups signals by radio station ID and sorts them by time, and for adjacent signals... and Establish bidirectional time-series edges: ； S23. Calculate the time difference between adjacent signals Explicitly encode it as an edge attribute of the emitter node. The calculated microsecond intervals are written into the graph as relation attributes to generate triples. Finally, all data is serialized and stored in a graph database, forming an electromagnetic knowledge network with complete temporal logic.
4. The method for identifying individual radio stations with data missingness based on knowledge graph enhancement according to claim 3, characterized in that: In step S3, the signal to be identified is used as the central node to extract its... Related subgraphs within skip neighborhoods ; Subsequently, a template mechanism is used to transform the subgraph. Convert to natural language description text This mechanism maps structured triples to natural language sentences, and integrates numerical features with topological structures into sequence data that can be understood by a large language model.
5. The method for identifying individual radio stations with data missingness based on knowledge graph enhancement according to claim 4, characterized in that: In step S4, the bidirectional encoder representation network BERT is constructed and fine-tuned to achieve accurate recognition, generating text. After WordPiece word segmentation, the data is converted into vector representations through an embedding layer. E: The model sets the maximum input length to 512 and utilizes a multi-layer self-attention mechanism to capture long-distance dependencies for the input sequence. The formula for calculating attention is: in These are the query, key, and value matrices, respectively.
6. The method for identifying individual radio stations with data missingness based on knowledge graph enhancement according to claim 5, characterized in that: In step S4, during the training phase, extract Tag hidden state vector The data is mapped to the radio station category space through a fully connected layer; during training, the cross-entropy loss function is used for end-to-end fine-tuning, and the objective function is defined as: in The total number of samples, This is a real radio station tag. The model predicts probabilities for the signal; by learning the "contextual semantics" of the signal, the model can deeply understand the contextual semantics and behavioral patterns of radio signals, thus maintaining extremely high recognition robustness even when some data is missing.
7. The method for identifying individual radio stations with data missingness based on knowledge graph enhancement according to claim 4, characterized in that: In step S3, the value of K in the K-hop neighborhood association subgraph is 2.
8. A knowledge graph-based augmented radio station individual identification system resistant to missing data, characterized in that: The system employs the method described in any one of claims 1 to 7.