A high-confidence UAV game decision-making intelligent perception system

Abstract

The present invention discloses a high-confidence drone game decision-making intelligent perception system. The system includes: a user dialogue and input module, a multi-task integrated intelligent agent module, a drone radiation source identification knowledge graph module, a drone radiation source waveform feature extraction module, and an identification result game decision analysis module. The user dialogue and input module receives drone radiation source identification questions and drone radiation source waveform data; the multi-task integrated intelligent agent module performs comprehensive intelligent analysis based on large language model technology, breaks down tasks and coordinates the execution of subtasks, and calls each submodule to complete automated intelligent analysis; the drone radiation source identification knowledge graph module builds a dynamic knowledge base to achieve structured representation and rapid retrieval; the drone radiation source waveform feature extraction module performs feature analysis on waveform signals, generates feature description text, and converts it into an embedded vector; the identification result game decision analysis module evaluates radiation source characteristics and threat level, outputs information and threat situation for game decision analysis, thereby improving game decision-making effectiveness. The advantages of the present invention are improved recognition accuracy and generalization ability, adaptability to dynamic electromagnetic environments, self-learning and optimization capabilities, effective support for electronic countermeasures requirements, enhanced information support capabilities for command decision-making, and improved game decision-making effectiveness.

Description

Technical Field

[0001] The present invention belongs to the field of multi-task intelligent identification game decision-making of unmanned aerial vehicle radiation sources, and in particular relates to a high-confidence unmanned aerial vehicle game decision-making intelligent perception system. Background Art

[0002] In modern electronic environments, mastering electromagnetic space dominance is a key factor in determining environmental game decisions. With the continuous advancement of technology, the electromagnetic environment has become increasingly complex. In particular, the emergence of highly dynamic and variable electromagnetic interference and a large number of heterogeneous radiation sources poses a serious challenge to traditional UAV emitter identification methods. Traditional methods typically rely on pulse descriptors (such as frequency, pulse width, and arrival time) for target identification. However, in complex electromagnetic environments, this single-dimensional data set is no longer sufficient for the rapid and accurate identification of emitters. First, these methods struggle to accurately identify signals in high-density electromagnetic environments. Due to the static nature of their parameters, they cannot effectively adapt to dynamic environmental changes. Second, when faced with multi-source signal interference, traditional methods often perform poorly due to their lack of ability to process complex data associations. Finally, these technologies cannot quickly adapt to and learn from unknown or emerging signal types, limiting their effectiveness in fast-paced electronic warfare scenarios.

[0003] By constructing a rich semantic network of entities, attributes, relationships, and events, knowledge graphs can integrate and understand large amounts of scattered, unstructured electromagnetic data. This technology identifies and links information scattered across multiple data sources, providing a structured and dynamically updated knowledge base. This enables electronic warfare decision support systems to conduct deeper analysis and understanding of complex electromagnetic scenarios. Through knowledge graphs, the system can correlate information based on known radiation source knowledge. The embedding of auxiliary knowledge significantly improves recognition accuracy, especially when dealing with unknown signals and dynamically changing electromagnetic environments.

[0004] Intelligent agent technology based on a large language model introduces an unprecedented level of intelligence to drone emitter identification. Within the drone emitter identification system, the intelligent agent parses operator commands, automatically extracts task-related information, and performs complex analytical tasks, achieving integrated multi-task intelligent solution. The large language model has excellent generalization performance. Combined with the domain knowledge of the drone emitter knowledge graph, it further enhances its reasoning capabilities. By continuously updating the knowledge graph, it can adapt to new electronic environmental conditions, improving overall game-playing decision-making flexibility.

[0005] Through this technological fusion, the present invention aims to overcome the limitations of traditional UAV radiation source identification methods in complex electromagnetic environments, provide a comprehensive, adaptive and highly intelligent solution, effectively support the electronic countermeasures needs in complex electromagnetic environments, further enhance the information support capabilities of command decision-making, and achieve improved game decision-making effects. Summary of the Invention

[0006] The purpose of the present invention is to address the deficiencies of the existing technology and provide a high-confidence drone game decision-making intelligent perception system.

[0007] The objective of the present invention is achieved through the following technical solutions: a high-confidence UAV game decision-making intelligent perception system, the system comprising: a user dialogue and input module, a multi-task integrated intelligent agent module, a UAV radiation source identification knowledge graph module, a UAV radiation source waveform feature extraction module, and an identification result game decision analysis module;

[0008] The user dialogue and input module is used to receive the problem of drone radiation source identification and drone radiation source waveform data input by the user, convert the drone radiation source identification problem into a question query vector, and package the drone radiation source waveform data into a waveform data file that meets the system requirements;

[0009] The multi-task integrated intelligent agent module performs comprehensive intelligent analysis of the problem of drone radiation source identification based on large language model technology, decomposes and analyzes multiple tasks, manages and coordinates the execution of each subtask, calls the drone radiation source identification knowledge graph module and the drone waveform feature extraction module to execute the subtasks, integrates the results of each subtask, generates prompt words suitable for input into the large language model, and inputs the output of the large language model into the recognition result game decision-making and analysis module to generate a detailed report or real-time feedback;

[0010] The UAV radiation source identification knowledge graph module is used to build and maintain a dynamic knowledge base of UAV radiation sources, including data preprocessing and entity labeling, knowledge extraction and relationship inference, and storage and retrieval of graph knowledge. Through these functions, the module can achieve structured representation and rapid retrieval of UAV radiation source knowledge data.

[0011] The UAV radiation source waveform feature extraction module is used to perform accurate and comprehensive feature analysis on the received UAV radiation source waveform data. It includes a comprehensive UAV waveform feature analysis tool library, which performs diversified feature analysis on the received UAV waveform, facilitating more accurate UAV radiation source identification and judgment.

[0012] The identification result game decision analysis module is used to evaluate the characteristics and threat level of the drone radiation source, comprehensively apply the drone waveform characteristics and radiation source knowledge graph information, and output the drone radiation source information and game decision threat situation.

[0013] Furthermore, the user dialogue and input module is used to receive the user's input of the drone radiation source identification question and the drone radiation source waveform data, convert the drone radiation source identification question into a question query vector, and package the drone radiation source waveform data into a waveform data file that meets the system requirements, specifically:

[0014] The user dialogue and input module is used to receive the user's input of the drone radiation source identification question and the drone radiation source waveform data, and use the pre-trained language embedding model to convert the user's input drone radiation source identification question into a question query vector through a conversion operation; and call the Python interpreter to package the drone radiation source waveform data into a waveform data file in the .pkl format that meets the system requirements.

[0015] Furthermore, the conversion operation of the pre-trained language embedding model specifically includes the following sub-steps:

[0016] a.1) First, the input text Text is tokenized using the tokenization tool Tokenizator, converting the text Text into a basic unit of token: token = Tokenizator(Text), where Tokenizator(·) is the tokenization operation;

[0017] a.2) The basic unit token is then embedded through the embedding layer, mapping it to an embedding vector e of fixed size: e = Embedding(token), where Embedding(·) is the embedding operation;

[0018] a.3) To preserve the sequential order of the input, we need to add the positional encoding PE to the embedding vector e, obtaining the embedding vector E: E = e + PE;

[0019] The calculation formula of position encoding PE is:

[0020] Among them, pos is the position index of the basic unit of the symbol in the input sequence, i is the dimension index of the embedding vector e, and d model The dimension of the pre-trained language embedding model.

[0021] Furthermore, the multi-task integrated intelligent agent module performs a comprehensive intelligent analysis of the problem of drone radiation source identification based on the large language model technology, decomposes and analyzes multiple tasks, manages and coordinates the execution of each subtask, calls the drone radiation source identification knowledge graph module and the drone waveform feature extraction module to execute the subtasks, integrates the results of each subtask, generates prompt words suitable for input into the large language model, and inputs the output of the large language model into the recognition result game decision-making module to generate a detailed report or real-time feedback. This is specifically implemented through the following sub-steps:

[0022] b.1) Task Decomposition and Scheduling: Based on the complexity of the input content and user needs, the intelligent agent decomposes the problem of drone emitter identification into multiple subtasks, such as knowledge graph construction, waveform feature extraction, or threat level assessment. Based on the thought chain, these subtasks are prioritized and assigned to the corresponding modules for processing.

[0023] b.2) Subtask Execution and Coordination: Based on task requirements, the UAV radiation source identification knowledge graph module, the UAV waveform feature extraction module, and the identification result game decision-making and analysis module are called. The UAV radiation source identification knowledge graph module is used to construct the knowledge graph and perform data association and reasoning. The UAV waveform feature extraction module is used to extract waveform features from UAV radiation source waveform data. The identification result game decision-making and analysis module is used to summarize information from UAV radiation source waveform data and assess the threat level. The intelligent agent verifies the execution results of each subtask to ensure that all subtasks are completed as required and that the output content meets actual requirements.

[0024] b.3) Result Integration and Output: The intelligent agent integrates the results of each subtask and generates prompt words suitable for input into the large language model. Based on the output of the large language model, the results are analyzed and input into the recognition result game decision-making module, generating detailed reports or real-time feedback, providing intuitive operation and decision-making support.

[0025] b.4) Continuous Learning and Updating: The multi-task UAV emitter identification system based on the intelligent agent knowledge graph uses user feedback and operation results to continuously optimize the performance of the large language model, update the knowledge graph, and adjust the processing strategy to adapt to environmental changes and new threat types.

[0026] Furthermore, the UAV radiation source identification knowledge graph module is used to build and maintain a dynamic knowledge base of UAV radiation sources, including data preprocessing and entity annotation, knowledge extraction and relationship inference, and storage and retrieval of graph knowledge. Through these functions, the module can achieve structured representation and rapid retrieval of UAV radiation source knowledge data, specifically:

[0027] The knowledge graph module for drone radiation source identification calls the multi-task integrated intelligent agent module to perform named entity recognition on the problem of drone radiation source identification, and obtains the entity names and attribute names related to the drone radiation source in the problem of drone radiation source identification; then, based on the multi-task integrated intelligent agent module, knowledge extraction and relationship inference are performed on the identified entity names and attribute names, and each entity and the corresponding attribute are matched, and stored in the format of triples, with the storage type being [entity A, attribute name of entity A, attribute value of entity A]; the relationships between different entities are also matched and stored in the format of triples, with the storage type being [entity A, relationship between entity A and entity B, entity B]; each triple is then converted into a knowledge embedding vector through a pre-trained language embedding model and stored in the dynamic knowledge base of drone radiation sources.

[0028] Furthermore, the UAV radiation source waveform feature extraction module is used to perform accurate and comprehensive feature analysis on the received UAV radiation source waveform data. It includes a comprehensive UAV waveform feature analysis tool library, which performs diversified feature analysis on the received UAV waveform, which is conducive to more accurate UAV radiation source identification and judgment, specifically:

[0029] The UAV waveform feature extraction module reads a waveform data file packaged in a .pkl format that meets system requirements. The UAV waveform feature analysis tool library in the UAV waveform feature extraction module first calls the pulse description word feature analysis tool to extract the frequency, power, pulse width, modulation mode, arrival angle, arrival time and other radiation source characteristic attributes of each waveform in the waveform data file, and calls the multi-task integrated intelligent agent module to perform entity matching and behavior analysis of the UAV radiation source, and constructs the feature information of each waveform in the form of a triplet [waveform number k, attribute name of waveform number k, attribute value of waveform number k]. The triples under the same waveform number are combined into a feature description text, which is then converted into a feature query vector through a pre-trained language embedding model.

[0030] Furthermore, the identification result game decision analysis module is used to evaluate the characteristics and threat level of the drone radiation source, comprehensively apply the drone waveform characteristics and radiation source knowledge graph information, and output the drone radiation source information and game decision threat situation, specifically:

[0031] The recognition result game decision-making module first performs a similarity search on the question query vector in the dynamic knowledge base of the drone radiation source to find the K knowledge embedding vectors with the largest Gaussian weighted cosine similarity; then performs a similarity search on any feature query vector in the drone waveform feature extraction module in the dynamic knowledge base of the drone radiation source to find the K knowledge embedding vectors with the largest Gaussian weighted cosine similarity; then, the question query vector and its corresponding K knowledge embedding vectors and the feature query vector and its corresponding K knowledge embedding vectors are used to construct a context, and the context constructed by the 2K+2 embedding vectors is used as the prompt word of the large language model and input into the multi-task integrated intelligent agent module to finally obtain the information summary and threat level assessment of the drone radiation source waveform data; wherein, the construction method of the prompt word Prompt of the large language model is as follows:

[0032] Prompt=E question ||E feature ||E knowledge ;

[0033] Among them, E question is the question query vector, E feature is the feature query vector, E knowledge is the knowledge embedding vector.

[0034] Furthermore, the similarity search process is specifically as follows:

[0035] Given any query vector E query For each knowledge embedding vector E in the dynamic knowledge base of the drone radiation source K , calculate the corresponding Gaussian weighted cosine similarity WGC(E query ,E K ,γ):

[0036]

[0037] Where γ is the bandwidth parameter of the Gaussian function, which controls the influence of distance; ||E query -E K || 2 Represents the Euclidean distance between two vectors; E query ·E K is the vector dot product; ||E query || and ||E K || are the moduli of the two vectors.

[0038] The beneficial effects of the present invention are:

[0039] 1) By integrating large language model intelligent agents and knowledge graph technology, this system can accurately process and analyze complex drone emitter data, improving the accuracy and generalization of emitter identification. The intelligent agent uses advanced large language model technology to achieve a deep understanding and precise answers to drone emitter questions, while automatically adjusting the identification strategy to adapt to the dynamic changes in the electromagnetic environment. In addition, the application of knowledge graphs enables the system to integrate and utilize prior knowledge of drone emitters to build a structured and continuously updated drone emitter knowledge base.

[0040] 2) This system also has the ability to self-learn and optimize. As time goes by and data accumulates, the system's performance and recognition accuracy will continue to improve. This learning mechanism is particularly suitable for processing unknown signals and highly dynamic electromagnetic interference, ensuring the application effect in fast-paced electronic countermeasure scenarios. Through this adaptive learning and optimization, this system can effectively support the needs of electronic countermeasures in complex electromagnetic environments, provide strong information support for command and decision-making, and achieve improved game decision-making effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a structural diagram of a high-confidence UAV game decision-making intelligent perception system;

[0042] In the figure, 1-user dialogue and input module; 2-multi-task integrated intelligent agent module; 3-UAV radiation source identification knowledge graph module; 4-UAV radiation source waveform feature extraction module; 5-identification result game decision-making module. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to illustrate the present invention, rather than to represent all embodiments. All other embodiments obtained by persons of ordinary skill in the art based on the embodiments of the present invention without creative work are within the scope of protection of the present invention.

[0044] Example 1

[0045] like Figure 1 As shown, the present invention provides a high-confidence UAV game decision-making intelligent perception system, which includes: a user dialogue and input module 1, a multi-task integrated intelligent agent module 2, a UAV radiation source identification knowledge graph module 3, a UAV radiation source waveform feature extraction module 4 and an identification result game decision analysis module 5.

[0046] The user dialogue and input module 1 is used to receive the problem of drone radiation source identification and drone radiation source waveform data input by the user, convert the drone radiation source identification problem into a question query vector, and package the drone radiation source waveform data into a waveform data file that meets the system requirements, specifically:

[0047] The user dialogue and input module is used to receive the user's input of the drone radiation source identification question and the drone radiation source waveform data, and use the pre-trained language embedding model (BERT, Bidirectional Encoder Representations from Transformers) to convert the user's input drone radiation source identification question into a question query vector through a transformation operation; and call the Python interpreter to package the drone radiation source waveform data into a waveform data file in the .pkl format that meets the system requirements.

[0048] The conversion operation of the pre-trained language embedding model specifically includes the following sub-steps:

[0049] a.1) First, the input text Text is tokenized using the tokenization tool Tokenizator, converting the text Text into a basic unit of token: token = Tokenizator(Text), where Tokenizator(·) is the tokenization operation;

[0050] a.2) The basic unit token is then embedded through the embedding layer, mapping it to an embedding vector e of fixed size: e = Embedding(token), where Embedding(·) is the embedding operation;

[0051] a.3) To preserve the sequential order of the input, we need to add the positional encoding PE to the embedding vector e, obtaining the embedding vector E: E = e + PE;

[0052] The calculation formula of position encoding PE is:

[0053] Among them, pos is the position index of the basic unit of the symbol in the input sequence, i is the dimension index of the embedding vector e, and d model The dimension of the pre-trained language embedding model.

[0054] The multi-task integrated intelligent agent module 2 performs a comprehensive intelligent analysis of the problem of drone radiation source identification based on the large language model technology, decomposes and analyzes multiple tasks, manages and coordinates the execution of each subtask, calls the drone radiation source identification knowledge graph module and the drone waveform feature extraction module to execute the subtasks, integrates the results of each subtask, generates prompt words suitable for input into the large language model, and inputs the output of the large language model into the recognition result game decision-making and analysis module to generate a detailed report or real-time feedback. This is specifically implemented through the following sub-steps:

[0055] b.1) Task Decomposition and Scheduling: Based on the complexity of the input content and user needs, the intelligent agent decomposes the problem of drone emitter identification into multiple subtasks, such as knowledge graph construction, waveform feature extraction, or threat level assessment. Based on the thought chain, these subtasks are prioritized and assigned to the corresponding modules for processing.

[0056] b.2) Subtask Execution and Coordination: Based on task requirements, the UAV radiation source identification knowledge graph module, the UAV waveform feature extraction module, and the identification result game decision-making and analysis module are called. The UAV radiation source identification knowledge graph module is used to construct the knowledge graph and perform data association and reasoning. The UAV waveform feature extraction module is used to extract waveform features from UAV radiation source waveform data. The identification result game decision-making and analysis module is used to summarize information from UAV radiation source waveform data and assess the threat level. The intelligent agent verifies the execution results of each subtask to ensure that all subtasks are completed as required and that the output content meets actual requirements.

[0057] b.3) Result Integration and Output: The intelligent agent integrates the results of each subtask and generates prompt words suitable for input into the large language model. Based on the output of the large language model, the results are analyzed and input into the recognition result game decision-making module, generating detailed reports or real-time feedback, providing intuitive operation and decision-making support.

[0058] b.4) Continuous Learning and Updating: The multi-task UAV emitter identification system based on the intelligent agent knowledge graph uses user feedback and operation results to continuously optimize the performance of the large language model, update the knowledge graph, and adjust the processing strategy to adapt to environmental changes and new threat types.

[0059] The UAV radiation source identification knowledge graph module 3 is used to build and maintain a dynamic knowledge base of UAV radiation sources, including data preprocessing and entity labeling, knowledge extraction and relationship inference, and storage and retrieval of graph knowledge. Through these functions, the module can achieve structured representation and rapid retrieval of UAV radiation source knowledge data, specifically:

[0060] The knowledge graph module for drone radiation source identification calls the multi-task integrated intelligent agent module to perform named entity recognition on the problem of drone radiation source identification, and obtains the entity names and attribute names related to the drone radiation source in the problem of drone radiation source identification; then, based on the multi-task integrated intelligent agent module, knowledge extraction and relationship inference are performed on the identified entity names and attribute names, and each entity and the corresponding attribute are matched, and stored in the format of triples, with the storage type being [entity A, attribute name of entity A, attribute value of entity A]; the relationships between different entities are also matched and stored in the format of triples, with the storage type being [entity A, relationship between entity A and entity B, entity B]; each triple is then converted into a knowledge embedding vector through a pre-trained language embedding model and stored in the dynamic knowledge base of drone radiation sources.

[0061] The drone radiation source identification knowledge graph module 3 will also ask questions to the multi-task integrated intelligent agent module 2 to extract knowledge and infer relationships for the identified entity names and attribute names. The questioning method is as follows: "Please construct a knowledge graph triple between the named entities and attributes in the sentence, in the format: [entity A, relationship, entity B], [entity, attribute, attribute value]; [JYL-1 drone, mobile 3D surveillance drone, S band]", the multi-task integrated intelligent agent module 2 will output "[JYL-1 drone, band, S band], [JYL-1 drone, type, mobile 3D surveillance drone]", thereby achieving the matching between entities and attributes, obtaining the relationship between entities, and storing them in the triple format; each triple is then converted into a knowledge embedding vector through the pre-trained language model BERT and stored in the same database, so as to construct an embedding vector knowledge base based on the knowledge graph.

[0062] The UAV radiation source waveform feature extraction module 4 is used to perform accurate and comprehensive feature analysis on the received UAV radiation source waveform data. It includes a comprehensive UAV waveform feature analysis tool library, which performs diversified feature analysis on the received UAV waveform, which is conducive to more accurate UAV radiation source identification and judgment. Specifically:

[0063] The UAV waveform feature extraction module reads a waveform data file packaged in a .pkl format that meets system requirements. The UAV waveform feature analysis tool library in the UAV waveform feature extraction module first calls the pulse description word feature analysis tool to extract the frequency, power, pulse width, modulation mode, arrival angle, arrival time and other radiation source characteristic attributes of each waveform in the waveform data file, and calls the multi-task integrated intelligent agent module to perform entity matching and behavior analysis of the UAV radiation source, and constructs the feature information of each waveform in the form of a triplet [waveform number k, attribute name of waveform number k, attribute value of waveform number k]. The triples under the same waveform number are combined into a feature description text, which is then converted into a feature query vector through a pre-trained language embedding model.

[0064] For example, if a user uploads a waveform data file "wave.pkl", the multi-task integrated intelligent agent module 2 will call the drone waveform feature analysis tool library to read this waveform data file, obtain the pulse description word features corresponding to the drone radiation source waveform in the waveform data file, and organize these feature information together according to the same waveform serial number to form the feature description text of the input waveform, such as "[waveform 1, frequency, 3100MHz], [waveform 1, pulse width, 10μs], [waveform 1, modulation mode, LFM]", and then use the pre-trained language embedding model to convert the feature description text of the input waveform into an embedding vector to construct a feature query vector.

[0065] The identification result game decision analysis module 5 is used to evaluate the characteristics and threat level of the drone radiation source, comprehensively apply the drone waveform characteristics and radiation source knowledge graph information, and output the drone radiation source information and game decision threat situation, specifically:

[0066] The recognition result game decision-making module first performs a similarity search on the question query vector in the dynamic knowledge base of the drone radiation source to find the K knowledge embedding vectors with the largest Gaussian weighted cosine similarity; then performs a similarity search on any feature query vector in the drone waveform feature extraction module in the dynamic knowledge base of the drone radiation source to find the K knowledge embedding vectors with the largest Gaussian weighted cosine similarity; then, the question query vector and its corresponding K knowledge embedding vectors and the feature query vector and its corresponding K knowledge embedding vectors are used to construct a context, and the context constructed by the 2K+2 embedding vectors is used as the prompt word of the large language model and input into the multi-task integrated intelligent agent module to finally obtain the information summary and threat level assessment of the drone radiation source waveform data; wherein, the construction method of the prompt word Prompt of the large language model is as follows:

[0067] Prompt=E question ||E feature||E knowledge ;

[0068] Among them, E question is the question query vector, E feature is the feature query vector, E knowledge is the knowledge embedding vector.

[0069] The similarity search process is as follows:

[0070] Given any query vector E query For each knowledge embedding vector E in the dynamic knowledge base of the drone radiation source K , calculate the corresponding Gaussian weighted cosine similarity WGC(E query ,E K ,γ):

[0071]

[0072] Where γ is the bandwidth parameter of the Gaussian function, which controls the influence of distance; ||E query -E K || 2 Represents the Euclidean distance between two vectors; E query ·E K is the vector dot product; ||E query || and ||E K || are the moduli of the two vectors.

[0073] Compared with the traditional cosine similarity-based similarity to calculate the similarity of query vectors, Gaussian weighted cosine similarity is calculated by Gaussian term exp(-γ||E query -E K || 2 ), the distance factor is introduced, which can more accurately reflect the similarity between vectors, thereby ensuring that the system is more inclined to retrieve and query vector E query Knowledge embedding vector E that is similar in both direction and distance K , which helps the system retrieve more relevant data, thereby achieving high-confidence intelligent perception and game decision-making.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A high-confidence UAV game decision-making intelligent perception system, characterized by: The system includes: a user dialogue and input module, a multi-task integrated intelligent agent module, a UAV radiation source identification knowledge graph module, a UAV radiation source waveform feature extraction module, and an identification result game decision-making and analysis module; The user dialogue and input module is used to receive the problem of drone radiation source identification and drone radiation source waveform data input by the user, convert the drone radiation source identification problem into a question query vector, and package the drone radiation source waveform data into a waveform data file that meets the system requirements; The multi-task integrated intelligent agent module performs comprehensive intelligent analysis of the problem of drone radiation source identification based on large language model technology, decomposes and analyzes multiple tasks, manages and coordinates the execution of each subtask, calls the drone radiation source identification knowledge graph module and the drone waveform feature extraction module to execute the subtasks, integrates the results of each subtask, generates prompt words suitable for input into the large language model, and inputs the output of the large language model into the recognition result game decision-making and analysis module to generate a detailed report or real-time feedback; The UAV radiation source identification knowledge graph module is used to build and maintain a dynamic knowledge base of UAV radiation sources, including data preprocessing and entity labeling, knowledge extraction and relationship inference, and storage and retrieval of graph knowledge. Through these functions, the module can achieve structured representation and rapid retrieval of UAV radiation source knowledge data. The UAV radiation source waveform feature extraction module is used to perform accurate and comprehensive feature analysis on the received UAV radiation source waveform data. It includes a comprehensive UAV waveform feature analysis tool library, which performs diversified feature analysis on the received UAV waveform, facilitating more accurate UAV radiation source identification and judgment. The identification result game decision analysis module is used to evaluate the characteristics and threat level of the drone radiation source, comprehensively apply the drone waveform characteristics and radiation source knowledge graph information, and output the drone radiation source information and game decision threat situation.

2. A high-confidence UAV game decision-making intelligent perception system according to claim 1, characterized in that: The user dialogue and input module is used to receive the problem of drone radiation source identification and drone radiation source waveform data input by the user, convert the drone radiation source identification problem into a question query vector, and package the drone radiation source waveform data into a waveform data file that meets the system requirements, specifically: The user dialogue and input module is used to receive the user input of the drone radiation source identification question and the drone radiation source waveform data, and use the pre-trained language embedding model to convert the user input of the drone radiation source identification question into a question query vector through a conversion operation; And call the Python interpreter to package the drone radiation source waveform data into a .pkl format waveform data file that meets the system requirements.

3. A high-confidence UAV game decision-making intelligent perception system according to claim 2, characterized in that: The conversion operation of the pre-trained language embedding model specifically includes the following sub-steps: a.1) First, the input text Text is tokenized using the tokenization tool Tokenizator, converting the text Text into a basic unit of token: token = Tokenizator(Text), where Tokenizator(·) is the tokenization operation; a.2) The basic unit token is then embedded through the embedding layer, mapping it to an embedding vector e of fixed size: e = Embedding(token), where Embedding(·) is the embedding operation; a.3) To preserve the sequential order of the input, we need to add the positional encoding PE to the embedding vector e, obtaining the embedding vector E: E = e + PE; The calculation formula of position encoding PE is: Among them, pos is the position index of the basic unit of the symbol in the input sequence, i is the dimension index of the embedding vector e, and d model The dimension of the pre-trained language embedding model.

4. A high-confidence UAV game decision-making intelligent perception system according to claim 3, characterized in that: The multi-task integrated intelligent agent module performs a comprehensive intelligent analysis of the problem of drone radiation source identification based on the large language model technology, decomposes and analyzes multiple tasks, manages and coordinates the execution of each subtask, calls the drone radiation source identification knowledge graph module and the drone waveform feature extraction module to execute the subtasks, integrates the results of each subtask, generates prompt words suitable for input into the large language model, and inputs the output of the large language model into the recognition result game decision-making and analysis module to generate a detailed report or real-time feedback. This is specifically implemented through the following sub-steps: b.1) Task Decomposition and Scheduling: Based on the complexity of the input content and user needs, the intelligent agent decomposes the problem of drone emitter identification into multiple subtasks, such as knowledge graph construction, waveform feature extraction, or threat level assessment. Based on the thought chain, these subtasks are prioritized and assigned to the corresponding modules for processing. b.2) Subtask Execution and Coordination: Based on task requirements, the UAV radiation source identification knowledge graph module, the UAV waveform feature extraction module, and the identification result game decision-making and analysis module are invoked. The UAV radiation source identification knowledge graph module is used to construct the knowledge graph and perform data association and reasoning. The UAV waveform feature extraction module is used to extract waveform features from UAV radiation source waveform data. The identification result game decision-making and analysis module is used to summarize information from UAV radiation source waveform data and assess threat levels. intelligent The agent verifies the execution results of each subtask to ensure that all subtasks are completed as required and that the output content meets actual needs; b.3) Result Integration and Output: The intelligent agent integrates the results of each subtask and generates prompt words suitable for input into the large language model. Based on the output of the large language model, the results are analyzed and input into the recognition result game decision-making module, generating detailed reports or real-time feedback, providing intuitive operation and decision-making support. b.4) Continuous Learning and Updating: The multi-task UAV emitter identification system based on the intelligent agent knowledge graph uses user feedback and operation results to continuously optimize the performance of the large language model, update the knowledge graph, and adjust the processing strategy to adapt to environmental changes and new threat types.

5. A high-confidence UAV game decision-making intelligent perception system according to claim 4, characterized in that: The UAV radiation source identification knowledge graph module is used to build and maintain a dynamic knowledge base of UAV radiation sources, including data preprocessing and entity labeling, knowledge extraction and relationship inference, and storage and retrieval of graph knowledge. Through these functions, the module can achieve structured representation and rapid retrieval of UAV radiation source knowledge data, specifically: The UAV radiation source identification knowledge graph module calls the multi-task integrated intelligent agent module to perform named entity recognition on the UAV radiation source identification problem, and obtains the entity names and attribute names related to the UAV radiation source in the UAV radiation source identification problem; Then, based on the multi-task integrated intelligent agent module, knowledge extraction and relationship inference are performed on the identified entity names and attribute names, each entity is matched with the corresponding attribute, and stored in the triple format with the storage type of [entity A, attribute name of entity A, attribute value of entity A]. The relationships between different entities are also matched and stored in triple format, with the storage type being [entity A, the relationship between entity A and entity B, entity B]. Each triple is then converted into a knowledge embedding vector through a pre-trained language embedding model and stored in the dynamic knowledge base of the drone radiation source.

6. A high-confidence UAV game decision-making intelligent perception system according to claim 5, characterized in that: The UAV radiation source waveform feature extraction module is used to perform accurate and comprehensive feature analysis on the received UAV radiation source waveform data. It includes a comprehensive UAV waveform feature analysis tool library, which performs diversified feature analysis on the received UAV waveform, which is conducive to more accurate UAV radiation source identification and judgment. Specifically: The UAV waveform feature extraction module reads a waveform data file packaged in a .pkl format that meets system requirements. The UAV waveform feature analysis tool library in the UAV waveform feature extraction module first calls the pulse description word feature analysis tool to extract the frequency, power, pulse width, modulation mode, arrival angle, arrival time and other radiation source characteristic attributes of each waveform in the waveform data file, and calls the multi-task integrated intelligent agent module to perform entity matching and behavior analysis of the UAV radiation source, and constructs the feature information of each waveform in the form of a triplet [waveform number k, attribute name of waveform number k, attribute value of waveform number k]. The triples under the same waveform number are combined into a feature description text, which is then converted into a feature query vector through a pre-trained language embedding model.

7. A high-confidence UAV game decision-making intelligent perception system according to claim 6, characterized in that: The identification result game decision analysis module is used to evaluate the characteristics and threat level of the drone radiation source. It comprehensively applies the drone waveform characteristics and radiation source knowledge graph information to output the drone radiation source information and game decision threat situation. Specifically: The recognition result game decision-making module first performs a similarity search on the question query vector in the dynamic knowledge base of drone radiation sources to find the K knowledge embedding vectors with the largest Gaussian weighted cosine similarity; then performs a similarity search on any feature query vector in the drone waveform feature extraction module in the dynamic knowledge base of drone radiation sources to find the K knowledge embedding vectors with the largest Gaussian weighted cosine similarity; Then, the question query vector and its corresponding K knowledge embedding vectors are combined with the feature query vector and its corresponding K knowledge embedding vectors to construct a context. This context, constructed from 2K+2 embedding vectors, is used as the prompt word for the large language model and input into the multi-task integrated intelligent agent module. Ultimately, information summary and threat level assessment of the drone radiation source waveform data are obtained. The prompt word Prompt for the large language model is constructed using the following formula: Prompt=E question ||And feature ||And knowledge ; Among them, E question is the question query vector, E feature is the feature query vector, E knowledge is the knowledge embedding vector.

8. The high-confidence UAV game decision-making intelligent perception system according to claim 7 is characterized in that: The similarity search process is as follows: Given any query vector E query For each knowledge embedding vector E in the dynamic knowledge base of the drone radiation source K , calculate the corresponding Gaussian weighted cosine similarity WGC(E query ,E K ,γ): Where γ is the bandwidth parameter of the Gaussian function, which controls the influence of distance; ||E query -E K || 2 Represents the Euclidean distance between two vectors; E query ·E K is the vector dot product; ||E query || and ||E K || are the moduli of the two vectors.

Images