A knowledge graph generation method based on a large language model

Abstract

The application provides a knowledge graph generation method based on a large language model, and relates to the technical field of knowledge graph generation.The real-time radar data and text data are acquired, the frequency domain features of the radar data are extracted, the text semantic feature vectors of the text data are extracted, and the time-space alignment is performed, the features are fused through the constructed fusion model, the fusion feature vectors are output, the relationship of the fusion feature vectors is extracted through the established large language model to generate triplets, the distribution of the triplet data is determined according to the similarity of the triplets, the triplet data is completed in the position where the distribution is sparse, the balance of the data distribution of the triplets is ensured, the triplets are formed into a triplet chain according to the continuity of the radar confrontation, the knowledge graph generation model is constructed according to the triplet chain, the knowledge graph is output, and the knowledge graph is updated according to the real-time radar data and text data.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph generation technology, specifically a knowledge graph generation method based on a large language model. Background Technology

[0002] In modern radar warfare, radar technology and jamming technology have become crucial components of the confrontation. The enemy, through advanced radar jamming techniques such as range-gate dragged pin (RGPO), can often effectively confuse or deceive friendly radar systems, thereby gaining an informational advantage. This jamming technique alters the characteristics of radar signals, interfering with the radar system's signal reception, reducing detection accuracy, and affecting the tactical decisions of both sides. Therefore, accurately identifying enemy radar jamming patterns in complex environments and effectively countering them with appropriate strategies has become a pressing technical challenge.

[0003] Currently, radar jamming technology analysis largely relies on traditional signal processing methods, such as frequency domain analysis and time domain analysis. These methods have limitations when dealing with high-frequency, complex jamming patterns and struggle to respond in real-time to ever-changing battlefield environments. Meanwhile, the tactical strategies and enemy-friendly relationships embedded in textual data are often overlooked or underutilized. Therefore, the lack of effective fusion and in-depth analysis of multimodal information results in poor adaptability of existing systems in complex environments, leading to low accuracy in jamming pattern recognition, lag in countermeasure strategy generation, and the inability to dynamically model the spatiotemporal correlations of jamming techniques, hindering the generation of accurate countermeasure strategies in real time. Existing knowledge graph construction technologies primarily focus on static knowledge bases, lacking the ability to adaptively update real-time battlefield data, particularly in cross-modal feature alignment and sparse data completion, significantly limiting the timeliness and accuracy of tactical decision-making.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a knowledge graph generation method based on a large language model to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A knowledge graph generation method based on a large language model, comprising the following steps:

[0008] Step 1: Acquire real-time radar data and text data from both sides during radar detection and counter-jamming processes; extract frequency domain features from radar data to identify jamming patterns; and extract semantic feature vectors from text data to identify relationships between subjects.

[0009] Step 2: Spatiotemporally align the extracted radar frequency domain features and text semantic feature vectors. Based on the aligned radar frequency domain features and text semantic feature vectors, perform feature fusion through the constructed fusion model using a cross-modal attention mechanism, and output the fused feature vector.

[0010] Step 3: By establishing a large language model, the fused feature vector is used to extract relations and generate triples. The distribution of triple data is determined based on the similarity of the triples. Triple data is filled in for sparse distributions to ensure the balance of triple data distribution.

[0011] Step 4: Form a triplet chain based on the continuity of radar countermeasures. Based on the triplet chain, construct a knowledge graph generation model, output a knowledge graph, and update the knowledge graph based on real-time radar data and text data.

[0012] Furthermore, the radar data includes the radar's operating frequency, waveform, and transmission power;

[0013] The specific process for extracting frequency domain features from radar data is as follows:

[0014] Radar data is acquired at the same time intervals. After sampling and preprocessing by maximum normalization, the DC component in the signal is removed. Interval time The radar data is divided into time windows, and frequency domain features are extracted from the radar data within the time window.

[0015] The formula for removing the DC component from a signal is: in, To remove the DC component from the signal, In order to be in Preprocessed radar signals at specific times This represents the total number of signal samples.

[0016] The formula for calculating the frequency domain features of radar data within a time window is as follows: in, The frequency domain signal of the radar signal is the first One frequency component, The first in the radar signal One sample, The total number of signal samples. The imaginary unit, Representing the first frequency domain One frequency component, .

[0017] Furthermore, the text data includes the working principle, technical characteristics, tactical applications, countermeasures strategies, and operational logs of radar jamming technology;

[0018] The specific process of extracting text semantic feature vectors from text data is as follows:

[0019] By using a vocabulary to split text data into word units, and adding a [CLS] marker at the beginning of each input text and a [SEP] marker at the end of the text to define text boundaries, the semantic feature vector of the text marked with [CLS] is extracted using RoBERTa.

[0020] Furthermore, the method for constructing the fusion model is as follows:

[0021] Use radar features as queries and text features as keys and values;

[0022] Similarity is calculated using the query and key to obtain attention weights. These attention weights are then applied to the value vector to obtain a weighted output feature vector. in, For query vector, To construct vectors, This is the transpose of the key vector. For value vectors, Let be the dimension of the key vector. The output is the fused feature vector. It is a function;

[0023] The query vector, the construct vector, and the value vector are respectively: in, For the weight matrix of the query, This is the radar frequency domain feature vector. The weight matrix is ​​the key weight matrix. This is a text semantic feature vector.

[0024] Furthermore, the specific method for spatiotemporal alignment is as follows: in, For time-aligned loss functions, This is the alignment mask matrix between radar features and text features. This represents the duration of radar feature acquisition. The number of descriptive words in the text. For radar signals in The feature vector at time step, For text description at location The semantic feature vector at that location.

[0025] Furthermore, the large language model is a joint model of RoBERTa-GCN-EGPLinker, specifically comprising RoBERTa semantic encoding, GCN, and EGPLinker models:

[0026] By using RoBERTa semantic encoding, the fused feature vector Extract the context vector representation, and update the nodes of the context vector representation through GCN. Use the updated context representation and entities in the text as input to the EGPLinker model to connect entity graph relationships.

[0027] Entity graph relationship connections: in, Indicates input sequence data They represent the first The initial and final projection matrices of each entity. For the first The bias of an entity, These represent the extracted entities. , This represents the context of an entity within the text. For entities Relationship score For transpose operation, It is a global pointer model;

[0028] The triplet relationship between entities is determined by a score threshold between the entities, and the triplet is output.

[0029] Furthermore, the method for calculating the distribution of triplet data based on the similarity of triplets is as follows: in, The third in the triplet The probability of each relationship type occurring The third in the triplet Types of relationships The number of relation types in the triple;

[0030] when, When the distribution of triples is sparse, it is determined that the triples are distributed sparsely.

[0031] in, Threshold for judging the sparse distribution of triples;

[0032] The specific method for completing triplet data in sparsely distributed locations is as follows: in, These represent the head entity, tail entity, and relation type in sparsely distributed triples. These are the embedding vectors for the head and tail entities, respectively. The type of candidate relation to be output. The vector is a random noise that follows a standard normal distribution. For relation types in sparsely distributed triples, To generate the learned weight matrix, To determine the learned weight matrix, The complete triplet.

[0033] Furthermore, the method for forming a triplet chain based on the continuity of radar countermeasures is as follows:

[0034] The radar countermeasure process is divided into stages, and a stage label is assigned to each triplet to form a sequence: in, For stage label sequence, For the first stage, The number of stages;

[0035] Statistical analysis of transition frequencies between stages, constructing a probability matrix: in, The transition probability matrix, For the stage Number of times it appears For the stage Transition to Stage The number of times;

[0036] Dynamic programming generates triple chains: in, For the first The triplet is in the stage The cumulative probability, In the stage The maximum cumulative probability, For the first Relationship score for each triplet.

[0037] Furthermore, the knowledge graph is constructed based on graph neural networks, and the specific process is as follows:

[0038] Based on the entities in each triplet chain being graph nodes, and the relationships in the triplet chain being edges: in, For nodes In graph neural networks, the first Layer representation, For nodes In graph neural networks, the first Layer representation, For nodes The set of neighboring nodes, For nodes , The relationship between them For the first The weight matrix of the layer, It is an aggregate function.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention acquires real-time radar data and text data, extracts frequency domain features from the radar data, extracts text semantic feature vectors from the text data, and performs spatiotemporal alignment. It then performs feature fusion through a constructed fusion model, outputs fused feature vectors, extracts relations from the fused feature vectors to generate triples by establishing a large language model, judges the distribution of triple data based on the similarity of the triples, completes the triple data in sparsely distributed areas to ensure the balance of triple data distribution, forms triple chains based on the continuity of radar countermeasures, and outputs a knowledge graph by constructing a knowledge graph generation model based on the triple chain. The knowledge graph is then updated based on real-time radar data and text data.

[0040] This solution addresses the data silo problem in traditional radar jamming analysis and countermeasure strategy generation by introducing a large language model and multimodal fusion technology. By spatiotemporally aligning radar frequency domain data and text semantic features, and achieving deep fusion of multi-dimensional data based on a cross-modal attention mechanism, it not only improves the accuracy of identifying enemy radar jamming patterns but also enhances the intelligent generation capability of countermeasure strategies. The relation extraction and data completion mechanism based on the large language model solves the knowledge imbalance problem under sparse data distribution. Furthermore, by constructing triple chains and updating the knowledge graph, a continuous event flow can be formed. This means that the knowledge graph construction not only includes static relational information but also incorporates the changes in time and action, making the generated knowledge graph more consistent with actual dynamic characteristics. This solution can dynamically generate countermeasure strategies based on real-time data, providing more accurate support for radar countermeasure decision-making. This method significantly improves the real-time adaptability of radar systems, enabling more efficient identification of enemy jamming and effective countermeasures in complex electronic warfare environments, providing a more comprehensive and accurate basis for countermeasure decision-making. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the overall method flow of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0043] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0044] Example

[0045] Please see Figure 1 The present invention provides a technical solution:

[0046] A knowledge graph generation method based on a large language model, comprising the following steps:

[0047] Step 1: Acquire real-time radar data and text data from both sides during radar detection and counter-jamming processes. Extract frequency domain features from the radar data to identify jamming patterns, and extract semantic feature vectors from the text data to identify the relationships between the subjects.

[0048] Radar data includes information such as the radar's operating frequency, waveform, and transmit power. By extracting frequency domain features from the radar data, interference modes and characteristics of the radar waveform can be identified. Using Fast Fourier Transform (FFT) for frequency domain analysis of radar data transforms the radar signal from the time domain to the frequency domain, facilitating the identification of the intensity, periodicity, and variations of different frequency components. This allows for the effective detection of interference modes and the characteristics of both friendly and enemy radar systems. This process not only helps identify the nature of the interference source but also reflects the radar's operational status, thus providing data support for developing countermeasure strategies.

[0049] The text data includes information on the working principles, technical characteristics, tactical applications, countermeasures, and operational logs of radar jamming technologies. By extracting semantic feature vectors from this text data, various jamming techniques, tactical deployments, and the behavioral relationships between friendly and adversary forces under different circumstances can be identified. The processing of text data not only extracts direct countermeasures and technical details but also reveals the interactions and information flows between different entities, forming a comprehensive knowledge graph.

[0050] In this embodiment, the radar data includes the radar's operating frequency, waveform, and transmission power;

[0051] The specific process for extracting frequency domain features from radar data is as follows:

[0052] Radar data is acquired at the same time intervals. After sampling and preprocessing by maximum normalization, the DC component in the signal is removed. Interval time The radar data is divided into time windows, and frequency domain features are extracted from the radar data within the time window.

[0053] The formula for removing the DC component from a signal is: in, To remove the DC component from the signal, In order to be in Preprocessed radar signals at specific times This represents the total number of signal samples.

[0054] The formula for calculating the frequency domain features of radar data within a time window is as follows: in, The frequency domain signal of the radar signal is the first One frequency component, The first in the radar signal One sample, The total number of signal samples. The imaginary unit, Representing the first frequency domain One frequency component, .

[0055] In this embodiment, the text data includes the working principle, technical characteristics, tactical applications, countermeasures strategies, and work logs of radar jamming technology;

[0056] The specific process of extracting text semantic feature vectors from text data is as follows:

[0057] By using a vocabulary to split text data into word units, and adding a [CLS] marker at the beginning of each input text and a [SEP] marker at the end of the text to define text boundaries, the semantic feature vector of the text marked with [CLS] is extracted using RoBERTa.

[0058] Step 2: Spatiotemporally align the extracted radar frequency domain features and text semantic feature vectors. Based on the aligned radar frequency domain features and text semantic feature vectors, perform feature fusion through the constructed fusion model using a cross-modal attention mechanism, and output the fused feature vector.

[0059] Cross-modal attention mechanisms dynamically capture and weight the correlation between two types of information, organically combining them to ensure the final fused features contain more information. This avoids the isolated existence of the two information sources. Cross-modal attention assigns a weight to each feature, highlighting its importance in a specific context. For example, radar data might be more important at one moment, while semantic information from text descriptions might be more crucial at another. This mechanism allows for flexible adjustment of the contribution of the two types of features, enhancing the model's ability to express complex situations. Traditional feature fusion methods often employ simple concatenation or weighted averaging, while cross-modal attention mechanisms learn the interrelationships between modalities and automatically adjust the importance of each modality in different contexts, thus extracting key information more accurately.

[0060] Traditional feature fusion methods (such as concatenation and weighted summation) often overlook the potential dependencies between different modalities. Cross-modal attention mechanisms, by adaptively adjusting the level of attention given to different modalities, effectively avoid this problem, thus better combining the two types of features. In complex tasks, features from a single modality may not be sufficient to fully describe the task's complexity. By fusing radar frequency domain features with textual semantic features, the implicit semantic information in the text can supplement the high-level knowledge that may be lacking in radar data, thereby providing the model with a more multi-dimensional understanding.

[0061] In this embodiment, the method for constructing the fusion model is as follows:

[0062] Use radar features as queries and text features as keys and values;

[0063] Similarity is calculated using the query and key to obtain attention weights. These attention weights are then applied to the value vector to obtain a weighted output feature vector. in, For query vector, To construct vectors, This is the transpose of the key vector. For value vectors, Let be the dimension of the key vector. The output is the fused feature vector. It is a function;

[0064] The query vector, the construct vector, and the value vector are respectively: in, For the weight matrix of the query, This is the radar frequency domain feature vector. The weight matrix is ​​the key weight matrix. This is a text semantic feature vector.

[0065] Radar frequency domain features are based on the spectral analysis of radar signals, reflecting the changes in radar signals in time and frequency. These data have clear spatiotemporal characteristics; for example, the radar's frequency, waveform, and transmit power may change within a specific time window. Text semantic feature vectors, on the other hand, are the result of vectorizing text data through natural language processing, reflecting the semantic information in the text, such as the principles and tactical applications of radar jamming techniques. Because these two types of data originate from different sources (one based on signal processing, the other on text understanding), their spatiotemporal alignment becomes crucial. The purpose of spatiotemporal alignment is to ensure that radar data and corresponding text data are associated within the same time period. For example, changes in radar data at a certain moment may be directly related to technical descriptions or countermeasures within a certain time period. Through spatiotemporal alignment, it is possible to ensure the temporal synchronization of the two types of features, enabling them to be associated within the same time dimension, thereby more accurately reflecting the dynamic changes of the system and the countermeasure process.

[0066] In this embodiment, the specific method for spatiotemporal alignment is as follows: in, For time-aligned loss functions, This is the alignment mask matrix between radar features and text features. This represents the duration of radar feature acquisition. The number of descriptive words in the text. For radar signals in The feature vector at time step, For text description at location The semantic feature vector at that location.

[0067] Step 3: By establishing a large language model, the fused feature vector is used to extract relations and generate triples. The distribution of triple data is determined based on the similarity of the triples. Triple data is filled in for sparse distributions to ensure the balance of triple data distribution.

[0068] The RoBERTa-GCN-EGPLinker joint model is a composite deep learning model that combines a pre-trained language model, a graph convolutional network (GCN), and a graph reasoning module. It is designed to handle tasks related to knowledge graph generation, relation extraction, reasoning, and data completion. By combining these three different techniques, the model can perform triple extraction, relation reasoning, and knowledge graph optimization more efficiently and accurately on large-scale data, thereby improving the completeness, accuracy, and balance of the knowledge graph.

[0069] RoBERTa is an improved model based on BERT. It is optimized by increasing the amount of training data, training time, and training strategies, significantly improving BERT's performance in natural language processing tasks. RoBERTa is widely used in relation extraction tasks because it can understand and capture fine-grained semantic relationships in text, identify entities and their relationships in fused feature vectors, and effectively handle complex linguistic features and implicit grammatical information, thus providing foundational data for subsequent graph convolutional networks and graph inference modules. GCN is a deep learning model for processing graph data, especially excelling in graph-structured tasks. GCN effectively captures dependencies between nodes in a graph, enhancing node representation through information passing between adjacent nodes. In knowledge graph generation, GCN can encode relationships between entities into graph node features based on graph structural information, helping to better understand the connections between entities. EGPLinker is an efficient global pointer model for named entity recognition and relation extraction tasks. It addresses some limitations of traditional sequence labeling models through a pointer network based on inner product computation. Its core idea is to transform entity recognition and relation extraction tasks into prediction tasks for the position of each pair of words, thereby enabling the direct detection of the start and end positions of entities, as well as the identification of different types of entities and relations.

[0070] In this embodiment, the large language model is a joint model of RoBERTa-GCN-EGPLinker, specifically including RoBERTa semantic encoding, GCN, and EGPLinker models:

[0071] By using RoBERTa semantic encoding, the fused feature vector Extract the context vector representation, and update the nodes of the context vector representation through GCN. Use the updated context representation and entities in the text as input to the EGPLinker model to connect entity graph relationships.

[0072] RoBERTa semantic encoding: in, Represented as a context vector, This is the output fused feature vector;

[0073] Updating the relationships between text nodes based on GCN: in, For the first The node representation matrix of a layer is initially represented by a context vector. For the first The node representation matrix of the layer, As the symmetry normalization factor, It is an adjacency matrix. For degree matrix, Update the weight matrix for the nodes. For activation functions;

[0074] Entity graph relationship connections: in, Indicates input sequence data They represent the first The initial and final projection matrices of each entity. For the first The bias of an entity, These represent the extracted entities. , This represents the context of an entity within the text. For entities Relationship score For transpose operation, It is a global pointer model;

[0075] The triplet relationship between entities is determined by a score threshold between the entities, and the triplet is output.

[0076] Triple completion is a key task in knowledge graph construction, especially when certain domains or relationships are scarce in the graph. By using triple similarity analysis to complete the graph, the blank areas in the knowledge graph can be effectively filled and information loss can be avoided.

[0077] By judging the distribution based on similarity, we can identify which domains have relatively scarce or uneven triple data and then supplement them accordingly. This helps to address the problem of excessive concentration of low-frequency triples or high-frequency triples, preventing the knowledge graph from becoming overly biased towards certain domains or relationships.

[0078] In traditional knowledge graph generation methods, the imbalance of data sources can lead to an overabundance of triples for certain relation types and a scarcity of triples for others. By analyzing and completing the graph based on the similarity of triples, the distribution of different relations can be balanced, ensuring that all domains and relation types in the graph are adequately represented.

[0079] This balance not only improves the information coverage of the knowledge graph, but also enhances its diversity and practicality. For example, some less common entity types or relationships, after being completed, will receive similar attention as mainstream relationships, thus increasing the breadth and depth of the graph.

[0080] In this embodiment, the calculation method for determining the distribution of triplet data based on the similarity of triplets is as follows: in, The third in the triplet The probability of each relationship type occurring The third in the triplet Types of relationships The number of relation types in the triple;

[0081] when, When the distribution of triples is sparse, it is determined that the triples are distributed sparsely.

[0082] in, Threshold for judging the sparse distribution of triples;

[0083] The specific method for completing triplet data in sparsely distributed locations is through adversarial neural networks, and the specific formula is as follows: in, These represent the head entity, tail entity, and relation type in sparsely distributed triples. These are the embedding vectors for the head and tail entities, respectively. The type of candidate relation to be output. The vector is a random noise that follows a standard normal distribution. For relation types in sparsely distributed triples, To generate the learned weight matrix, To determine the learned weight matrix, The complete triplet.

[0084] By adversarially training the generator and discriminator of an adversarial neural network (ANN), synthetic data similar to the distribution of real data can be generated. In knowledge graph construction, ANNs are used to complete sparse regions with triplet data. Through adversarial training, the generated data not only fills in sparse regions but also ensures that the generated data is consistent with the data distribution in the existing knowledge graph. The generator produces completions similar to real triplet data, while the discriminator judges the authenticity of the generated data. Through this process, the generated triples are more consistent with the existing data patterns in the graph, reducing low-quality completions or inconsistencies with existing data.

[0085] Step 4: Form a triplet chain based on the continuity of radar countermeasures. Based on the triplet chain, construct a knowledge graph generation model, output a knowledge graph, and update the knowledge graph based on real-time radar data and text data.

[0086] Traditional knowledge graphs typically consist of discrete triples, the relationships between which may be isolated and lack clear coherence. By organizing triples in a chain according to the "continuity of radar countermeasures," more logical triple chains can be formed. This allows information in the graph to not only be interconnected but also to be reasoned and analyzed through continuous chains of relationships. For example, relationships between certain entities or events can be traced through continuous chains of triples, thereby gaining a higher level of understanding and reasoning ability.

[0087] The continuity of radar countermeasures refers to associating continuous observations from radar data with triples in a knowledge graph, forming a time-series-based triple chain. With the real-time input of new radar and text data, the system can update the triple chain in a timely manner, seamlessly integrating new information into the existing knowledge graph. This approach, based on continuity and adversarial relationships, can quickly reflect changes in dynamic environments, making it more efficient and intelligent than traditional static update methods. In some application scenarios, knowledge graphs not only need to express semantic relationships between entities but also handle temporal and spatial dimensions. By organizing triples into chains with temporal and spatial continuity, the system can better understand the relationships between entities that change over time. For example, continuous observations based on radar data can help track the changing trajectories or dynamic states of target objects, which is particularly important in fields such as transportation and meteorology. Traditional knowledge graph generation methods often rely on static triples, lacking dynamic updates and continuity processing. This makes knowledge graphs slow to update when faced with real-time data and changing data sources (such as radar and text data), and may not be able to reflect the temporal relationships between data in a timely manner. For example, when new radar data comes in, traditional methods may simply add new triples without forming an effective chain of relationships, resulting in insufficient integration of information.

[0088] In this embodiment, the method for forming a triplet chain based on the continuity of radar countermeasures is as follows:

[0089] The radar countermeasure process is divided into stages, and a stage label is assigned to each triplet to form a sequence: in, For stage label sequence, For the first stage, The number of stages;

[0090] Statistical analysis of transition frequencies between stages, constructing a probability matrix: in, The transition probability matrix, For the stage Number of times it appears For the stage Transition to Stage The number of times;

[0091] Dynamic programming generates triple chains: in, For the first The triplet is in the stage The cumulative probability, In the stage The maximum cumulative probability, For the first Relationship score for each triplet.

[0092] In this embodiment, the knowledge graph is constructed based on a graph neural network, and the specific process is as follows:

[0093] Based on the entities in each triplet chain being graph nodes, and the relationships in the triplet chain being edges: in, For nodes In graph neural networks, the first Layer representation, For nodes In graph neural networks, the first Layer representation, For nodes The set of neighboring nodes, For nodes , The relationship between them For the first The weight matrix of the layer, It is an aggregate function.

[0094] Graph neural networks (GNNs) can learn and represent complex graph structures through the relationships between nodes and edges. In knowledge graphs, entities are typically represented as nodes, and the relationships between entities are represented as edges. Traditional knowledge graph methods often rely solely on simple adjacency matrices or manually constructed rules to represent these relationships. However, GNNs, by utilizing information from adjacent nodes (such as node features and the influence of neighboring nodes), can better capture the complex and multi-layered relationships between entities, thereby enhancing the expressive and reasoning capabilities of knowledge graphs.

[0095] Graph Neural Networks (GNNs) can automatically learn the features of nodes and edges and propagate information through methods such as graph convolution. Through this information propagation, GNNs can capture potential relationships between entities, enhancing the contextual understanding and reasoning capabilities of knowledge graphs. Especially when dealing with data that is complex and dynamically changing, GNNs can adaptively adjust node representations, making knowledge graphs more adaptable to dynamic scenarios. Graph Neural Networks are particularly well-suited for handling dynamically changing graph data. When new triple chains are input in real time, GNNs can adaptively adjust the graph structure by iterating and updating the features of nodes and edges. Compared to traditional static knowledge graphs, GNNs can effectively perform incremental learning, updating entity relationships in real time, especially when facing continuously changing data, ensuring the accuracy and timeliness of the knowledge graph.

[0096] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0097] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0098] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0099] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A knowledge graph generation method based on a large language model, characterized in that, The specific steps include: Step 1: Acquire real-time radar data and text data from both sides during radar detection and counter-jamming processes; extract frequency domain features from radar data to identify jamming patterns; and extract semantic feature vectors from text data to identify relationships between subjects. Step 2: Spatiotemporally align the extracted radar frequency domain features and text semantic feature vectors. Based on the aligned radar frequency domain features and text semantic feature vectors, perform feature fusion through the constructed fusion model using a cross-modal attention mechanism, and output the fused feature vector. Step 3: By establishing a large language model, the fused feature vector is used to extract relations and generate triples. The distribution of triple data is judged based on the similarity of the triples. Triple data is filled in for sparse distributions to ensure the balance of triple data distribution. Step 4: Form a triplet chain based on the continuity of radar countermeasures. Based on the triplet chain, construct a knowledge graph generation model, output a knowledge graph, and update the knowledge graph based on real-time radar data and text data. The method for calculating the distribution of triplet data based on the similarity of triplets is as follows: in, The first in the triplet The probability of each relationship type occurring The first in the triplet Types of relationships The number of relation types in the triple; when, When the distribution of triples is sparse, it is determined that the triples are distributed sparsely. in, Threshold for judging the sparse distribution of triples; The specific method for completing triplet data in sparsely distributed locations is as follows: in, , , These represent the head entity, tail entity, and relation type in sparsely distributed triples. These are the embedding vectors for the head and tail entities, respectively. The type of candidate relation to be output. The vector is a random noise that follows a standard normal distribution. For relation types in sparsely distributed triples, To generate the learned weight matrix, To determine the learned weight matrix, The complete triplet; The method for forming a triplet chain based on the continuity of radar countermeasures is as follows: The radar countermeasure process is divided into stages, and a stage label is assigned to each triplet to form a sequence: in, For stage label sequence, For the first stage, The number of stages; Statistical analysis of transition frequencies between stages, constructing a probability matrix: in, The transition probability matrix, For the stage Number of times it appears For the stage Transition to stage The number of times; Dynamic programming generates triple chains: in, For the first The triplet is in the stage The cumulative probability, In the stage The maximum cumulative probability, For the first Relationship score for each triplet.

2. The knowledge graph generation method based on a large language model according to claim 1, characterized in that: The radar data includes the radar's operating frequency, waveform, and transmission power; The specific process for extracting frequency domain features from radar data is as follows: Radar data is acquired at the same time intervals. After sampling and preprocessing by maximum normalization, the DC component in the signal is removed. Interval time The radar data is divided into time windows, and frequency domain features are extracted from the radar data within the time window. The formula for removing the DC component from a signal is: in, To remove the DC component from the signal, In order to be in Preprocessed radar signals at specific times This represents the total number of signal samples. The formula for calculating the frequency domain features of radar data within a time window is as follows: in, The frequency domain signal of the radar signal is the first One frequency component, The first in the radar signal One sample, The total number of signal samples. The imaginary unit, Representing the first frequency domain One frequency component, .

3. The knowledge graph generation method based on a large language model according to claim 1, characterized in that: The text data includes the working principle, technical characteristics, tactical applications, countermeasures strategies, and work logs of radar jamming technology; The specific process of extracting text semantic feature vectors from text data is as follows: By using a vocabulary to split text data into word units, and adding a [CLS] marker at the beginning of each input text and a [SEP] marker at the end of the text to define text boundaries, the semantic feature vector of the text marked with [CLS] is extracted using RoBERTa.

4. The knowledge graph generation method based on a large language model according to claim 1, characterized in that: The method for constructing the fusion model is as follows: Use radar features as queries and text features as keys and values; Similarity is calculated using the query and key to obtain attention weights. These attention weights are then applied to the value vector to obtain a weighted output feature vector. in, For query vector, To construct vectors, This is the transpose of the key vector. For value vectors, Let be the dimension of the key vector. The output is the fused feature vector. It is a function; The query vector, the construct vector, and the value vector are respectively: in, For the weight matrix of the query, This is the radar frequency domain feature vector. Let be the weight matrix of the keys. This is a text semantic feature vector.

5. The knowledge graph generation method based on a large language model according to claim 1, characterized in that: The specific method for spatiotemporal alignment is as follows: in, For time-aligned loss functions, This is the alignment mask matrix between radar features and text features. This represents the duration of radar feature acquisition. The number of descriptive words in the text. For radar signals in The feature vector at time step, For text description at location The semantic feature vector at that location.

6. The knowledge graph generation method based on a large language model according to claim 1, characterized in that: The large language model is a joint model of RoBERTa-GCN-EGPLinker, specifically consisting of RoBERTa semantic encoding, GCN, and EGPLinker models: By using RoBERTa semantic encoding, the fused feature vector Extract the context vector representation, and update the nodes of the context vector representation through GCN. Use the updated context representation and entities in the text as input to the EGPLinker model to connect entity graph relationships. Entity graph relationship connections: in, Indicates input sequence data , , They represent the first The initial and final projection matrices of each entity. For the first The bias of an entity, These represent the extracted entities. , This represents the context of an entity within the text. For entities Relationship score For transpose operation, It is a global pointer model; The triplet relationship between entities is determined by a score threshold between the entities, and the triplet is output.

7. The knowledge graph generation method based on a large language model according to claim 1, characterized in that: Knowledge graphs are constructed based on graph neural networks. The specific process is as follows: Based on the entities in each triplet chain being graph nodes, and the relationships in the triplet chain being edges: in, For nodes In graph neural networks, the first Layer representation, For nodes In graph neural networks, the first Layer representation, For nodes The set of neighboring nodes, For nodes , The relationship between them For the first The weight matrix of the layer, It is an aggregate function.

Images