Multi-satellite platform intelligent query and retrieval method based on natural language interaction

Abstract

The invention discloses a multi-satellite platform intelligent query and retrieval method based on natural language interaction. The method comprises the following steps: receiving a natural language query of a user and preprocessing the query; different satellite platform naming specifications are processed in a unified manner through a multi-format satellite identification intelligent identification technology; a mixed decision mechanism combining deep learning and a rule engine is adopted to deeply analyze query semantics; accurately positioning a target knowledge base based on a self-adaptive routing strategy; and obtaining an optimal retrieval result by using a parallel mixed retrieval technology. According to the method, through a mixed understanding mechanism combining deep learning and a rule engine, the accuracy and speciality of understanding the natural language query instruction by the user are remarkably improved, and then the accuracy of retrieval query is improved. In addition, through a parallel mixed retrieval strategy, two methods of semantic vector retrieval and keyword retrieval are comprehensively utilized to retrieve the query instruction, and the retrieval results of the two methods are fused, so that the retrieval effect is further optimized.

Description

Technical Field

[0001] This invention relates to a multi-satellite platform intelligent query and retrieval method based on natural language interaction, belonging to the field of satellite communication applications. Background Technology

[0002] With the rapid development of satellite technology and the widespread adoption of multi-satellite platform collaborative operations, satellite telemetry, tracking, and command (TT&C) personnel and technical experts need to quickly and accurately obtain relevant knowledge information from various satellite platforms using natural language. However, existing retrieval systems struggle to effectively handle natural language queries containing information from multiple satellite platforms. User queries often include satellite identifiers with different naming conventions, and existing systems lack a unified identification and standardization mechanism, leading to inaccurate association between queries and knowledge content, severely impacting retrieval accuracy. Furthermore, traditional keyword-based retrieval methods cannot deeply understand the semantic connotations of natural language queries; specialized queries in the satellite field typically contain complex technical intents and specific application scenarios, making it difficult for simple keyword matching to accurately capture users' true information needs, resulting in poor relevance of search results. In addition, a single retrieval strategy cannot adapt to the diverse characteristics of knowledge from multiple satellite platforms; technical documents from different satellite platforms have different structural features and professional focuses, making it difficult for fixed retrieval strategies to achieve consistent retrieval results across platforms, affecting overall system performance. Finally, existing systems lack intelligent routing mechanisms and cannot dynamically select the most relevant knowledge base based on query characteristics, resulting in low retrieval efficiency and unstable result quality, failing to meet the high real-time requirements of satellite operation and maintenance scenarios. All of these problems result in existing retrieval methods failing to accurately understand users' natural language query commands and producing poor query results. Summary of the Invention

[0003] The technical problem solved by this invention is to overcome the shortcomings of existing technologies and provide a multi-satellite platform intelligent query and retrieval method based on natural language interaction. This method, through a hybrid understanding mechanism combining deep learning and a rule engine, can accurately understand the intent of the user's natural language query instructions. Furthermore, by employing a parallel hybrid retrieval strategy, it comprehensively utilizes semantic vector retrieval and keyword retrieval to optimize retrieval results, thus solving the problems of inaccurate understanding of user natural language query instructions and poor query performance in existing retrieval methods.

[0004] The technical solution of this invention is:

[0005] A multi-satellite platform intelligent query and retrieval method based on natural language interaction, comprising the following steps:

[0006] (1) Preprocess the user's query request in natural language form to obtain the natural language query text;

[0007] (2) By using contextual semantic analysis and regular expression pattern matching, the naming identifiers of multiple satellite platforms pre-queried by the user are extracted from the natural language query text, and the naming identifiers of multiple satellite platforms are uniformly converted into a standard format to obtain the pre-query satellite standard format; after extracting the naming identifiers of multiple satellite platforms pre-queried by the user from the natural language query text, the part of the natural language query text that has not been extracted is defined as the remaining query information; the pre-query satellite standard format and the remaining query information are combined to form the pre-query instruction information;

[0008] (3) Input the pre-query instruction information into the large language model that has been fine-tuned by satellite domain knowledge for deep semantic parsing, and output the user's deep query intent and query type; input the pre-query satellite standard format into the rule engine based on domain knowledge graph to extract professional term features; combine the user's deep query intent, query type and professional term features to obtain structured query understanding results;

[0009] (4) Establish an intelligent mapping model between query type and knowledge base content; based on the query type in the query understanding results, the intelligent mapping model generates the optimal knowledge base combination containing multiple knowledge bases and the retrieval filtering conditions corresponding to each knowledge base in the knowledge base combination.

[0010] (5) Multi-threaded concurrency technology is used to enable each knowledge base in the optimal knowledge base combination to perform semantic vector retrieval and keyword retrieval simultaneously; semantic vector retrieval takes the query understanding results as input, and performs similarity matching in the optimal knowledge base combination based on the deep intent of the user query in the query understanding results to obtain the initial semantic query results; keyword retrieval takes the query understanding results as input, and performs precise matching in the optimal knowledge base combination based on the professional terminology features in the query understanding results to obtain the initial keyword query results; the initial semantic query results and initial keyword query results are filtered based on the retrieval filtering conditions corresponding to each knowledge base to obtain the corresponding semantic query results and keyword query results;

[0011] (6) The semantic query results and keyword query results are reordered and merged to obtain the final search results.

[0012] Furthermore, the user's query request in natural language form in step (1) includes query requests in text form and query requests in voice form;

[0013] The preprocessing process for text-based query requests is as follows: text cleaning and word segmentation are performed on the text-based query requests, and irrelevant characters are removed to finally obtain the natural language query text;

[0014] The preprocessing process for voice query requests is as follows: the voice query requests are filtered to remove noise, and the noise-removed voice is converted into text, ultimately yielding the natural language query text.

[0015] Furthermore, the specific process of extracting the naming identifier of the satellite platform pre-queried by the user from the natural language query text in step (2) and uniformly converting the naming identifier of the satellite platform into a standard format is as follows:

[0016] (2.1) Identify the naming identifiers of satellite platforms contained in natural language query text by using Chinese character, number, and letter combination recognition algorithms and contextual semantic analysis technology;

[0017] (2.2) The identified satellite platform naming identifiers are mapped to a unified naming system through pattern matching and a standardization mapper; the unified naming system is a standard naming system of "Chinese characters + numbers + letters"; the satellite platform naming identifiers converted into the unified naming system are the pre-query satellite standard format.

[0018] Furthermore, the query types in step (3) include four types: telemetry query, remote control query, fault query, and general query; the technical terminology features include the parameter code, parameter name, instruction name, instruction code, fault name, and fault number of the user query; the query understanding results also include the natural language query text, the query confidence level corresponding to each query type, the keywords in the natural language query text, and the pre-query satellite standard format.

[0019] Furthermore, in step (4), based on the query understanding results, the intelligent mapping model generates an optimal combination of knowledge bases containing multiple knowledge bases and the retrieval filtering conditions corresponding to each knowledge base in the knowledge base combination as follows:

[0020] (4.1) The intelligent mapping model selects the knowledge bases that form the optimal knowledge base combination based on the query type in the query understanding results. If the query type is a telemetry query, multiple general knowledge bases and telemetry professional knowledge bases are combined to form the optimal knowledge base combination. If the query type is a remote control query, multiple general knowledge bases and remote control professional knowledge bases are combined to form the optimal knowledge base combination. If the query type is a fault query, multiple general knowledge bases and fault professional knowledge bases are combined to form the optimal knowledge base combination. If the query type is a general query, multiple general knowledge bases are combined to form the optimal knowledge base combination.

[0021] (4.2) When the query type is a general query, no operation is performed on the optimal knowledge base combination; otherwise, the intelligent mapping model dynamically adjusts the optimal knowledge base combination based on the query confidence in the query understanding results; the intelligent mapping model dynamically adjusts the optimal knowledge base combination as follows: if the query confidence is high, the general knowledge base in the optimal knowledge base combination is deleted; if the query confidence is low, no operation is performed on the knowledge base.

[0022] (4.3) The intelligent mapping model generates different search filtering conditions for different knowledge bases in the optimal knowledge base combination;

[0023] If the optimal knowledge base combination does not contain a professional knowledge base, then for each general knowledge base in the optimal knowledge base combination, the intelligent mapping model generates general search filtering conditions.

[0024] If the optimal knowledge base combination contains a professional knowledge base, the intelligent mapping model generates professional search filtering conditions for the professional knowledge base in the optimal knowledge base combination; and generates general search filtering conditions for the general knowledge base in the optimal knowledge base combination.

[0025] The professional knowledge base refers to the telemetry professional knowledge base, the remote control professional knowledge base, and the fault professional knowledge base; the general search filtering conditions include satellite ID and document type; the professional search filtering conditions include satellite ID, document type, parameter code, parameter name, command code, remote control code, command name, and fault name.

[0026] Furthermore, the method of using multi-threaded concurrency technology in step (5) to enable each knowledge base in the optimal knowledge base combination to simultaneously perform semantic vector retrieval and keyword retrieval is as follows: using thread pool technology, an independent vector retrieval sub-thread and BM25 retrieval sub-thread are created for each knowledge base in the optimal knowledge base combination; the keyword retrieval adopts the BM25 retrieval method.

[0027] This invention also proposes a multi-satellite platform intelligent query and retrieval system based on natural language interaction, comprising: a natural language query unit, a multi-satellite identifier recognition unit, a hybrid query understanding engine, an adaptive routing decision-maker, and a parallel hybrid retrieval system;

[0028] The natural language query unit preprocesses the user's query request in natural language form to obtain the natural language query text;

[0029] The multi-satellite identifier recognition unit extracts the naming identifiers of multiple satellite platforms pre-queried by the user from the natural language query text through contextual semantic analysis and regular expression pattern matching methods, and converts the naming identifiers of multiple satellite platforms into a standard format to obtain the pre-queried satellite standard format; after extracting the naming identifiers of multiple satellite platforms pre-queried by the user from the natural language query text, the unextracted part of the natural language query text is defined as the remaining query information; the pre-queried satellite standard format and the remaining query information are combined to form the pre-query instruction information;

[0030] The hybrid query understanding engine inputs pre-query instruction information into a large language model fine-tuned by satellite domain knowledge for deep semantic parsing, outputting the user's deep query intent and query type; at the same time, it inputs the pre-query satellite standard format into a rule engine based on domain knowledge graph to extract professional terminology features; combining the user's deep query intent, query type, and professional terminology features, it obtains structured query understanding results.

[0031] The adaptive routing decision-maker has an internal intelligent mapping model; based on the query type in the query understanding results, the intelligent mapping model generates the optimal combination of multiple knowledge bases and the retrieval filtering conditions corresponding to each knowledge base in the combination.

[0032] The parallel hybrid retrieval engine employs multi-threaded concurrency technology to enable each knowledge base in the optimal knowledge base combination to simultaneously perform semantic vector retrieval and keyword retrieval. Semantic vector retrieval takes query understanding results as input, and based on the user's deep query intent in the query understanding results, performs similarity matching in the optimal knowledge base combination to obtain initial semantic query results. Keyword retrieval takes query understanding results as input, and based on the professional terminology features in the query understanding results, performs precise matching in the optimal knowledge base combination to obtain initial keyword query results. Subsequently, based on the retrieval filtering conditions corresponding to each knowledge base, the initial semantic query results and initial keyword query results are filtered to obtain the corresponding semantic query results and keyword query results. Finally, the parallel hybrid retrieval engine reorders and merges the semantic query results and keyword query results to obtain the final retrieval results.

[0033] Furthermore, the query instruction information is input into the large oracle model fine-tuned by satellite domain knowledge, and the output of the large oracle model fine-tuned by satellite domain knowledge also includes technical terminology features; the query instruction information is input into the rule engine based on the knowledge domain graph, and the output of the rule engine based on the knowledge domain graph also includes the user's deep query intent and query type.

[0034] Furthermore, the hybrid query understanding engine adopts a dual-processing architecture, featuring a dual-processing collaborative working mode and a single-processing working mode.

[0035] Furthermore, the dual-processing architecture includes a large language model branch and a rule engine branch; the large language model branch is equipped with a large oracle model fine-tuned with satellite domain knowledge, and the rule engine branch is equipped with a rule engine based on a knowledge domain graph; the single-processing working mode includes a single-processing working mode for the large language model branch and a single-processing working mode for the rule engine branch.

[0036] When both the big oracle model branch and the rule engine branch are working normally, the dual-processing architecture adopts a dual-path collaborative working mode. The big language model branch inputs the pre-query instruction information into the big language model, which has been fine-tuned by satellite domain knowledge, for deep semantic parsing and outputs the user's deep query intent and query type. The rule engine branch inputs the pre-query instruction information into the rule engine based on the domain knowledge graph and extracts professional terminology features. Combining the user's deep query intent and query type output by the big language model, which has been fine-tuned by satellite domain knowledge, and the professional terminology features output by the rule engine based on the domain knowledge graph, a structured query understanding result is obtained.

[0037] When an anomaly occurs in the Big Prophecy model branch, the dual-processing architecture adopts a single-processing mode for the rule engine branch. The rule engine branch inputs the pre-query instruction information into the rule engine based on the domain knowledge graph. The rule engine based on the domain knowledge graph outputs the technical terminology features, the user's deep query intent, and the query type. Combining the technical terminology features, the user's deep query intent, and the query type output by the rule engine based on the domain knowledge graph, a structured query understanding result is obtained.

[0038] When the rule engine branch encounters an anomaly, the dual-processing architecture adopts a single-processing mode for the large language model branch. The large language model branch inputs the pre-query instruction information into the large language model that has been fine-tuned by satellite domain knowledge. The large language model that has been fine-tuned by satellite domain knowledge outputs the user's deep query intent, query type, and technical terminology features. Combining the technical terminology features, user's deep query intent, and query type output by the large language model that has been fine-tuned by satellite domain knowledge, a structured query understanding result is obtained.

[0039] The beneficial effects of this invention compared to the prior art are:

[0040] (1) This invention utilizes a hybrid understanding mechanism that combines deep learning and rule engines, which fully leverages the semantic understanding capabilities of large language models and incorporates the professional knowledge of domain experts, significantly improving the accuracy and professionalism of users' natural language query understanding, thereby enhancing the accuracy of retrieval queries.

[0041] (2) This invention uses a parallel hybrid retrieval strategy to combine semantic vector retrieval and keyword retrieval methods to retrieve query commands, and integrates the results of the two methods to further optimize the retrieval effect.

[0042] (3) The present invention dynamically selects the optimal knowledge base combination and the corresponding retrieval filtering conditions based on the characteristics of the user's natural language query instructions, and then selects the query results from the optimal knowledge base combination, thereby further ensuring the accuracy of the retrieval.

[0043] (4) This invention uses multi-threaded concurrency technology to enable each knowledge base in the optimal knowledge base combination to perform semantic vector retrieval and keyword retrieval simultaneously, thus ensuring the efficiency of the retrieval process.

[0044] (5) This invention can process user-input text and voice query requests, enabling users to express their information needs in the most intuitive and diverse way, which greatly reduces the threshold for using the system and improves the user experience.

[0045] (6) By unifying the naming identifiers of satellite platforms into a standard format, this invention can intelligently identify and standardize the mapping of multiple satellite identifiers, thereby establishing a unified multi-satellite platform identifier processing system and effectively solving the identifier compatibility problem in a multi-satellite environment. Attached Figure Description

[0046] Figure 1 This is a flowchart of a multi-satellite platform intelligent query and retrieval method based on natural language interaction according to the present invention;

[0047] Figure 2 This is a system block diagram of a multi-satellite platform intelligent query and retrieval method based on natural language interaction according to the present invention. Detailed Implementation

[0048] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.

[0049] like Figure 1 As shown, this invention provides a multi-satellite platform intelligent query and retrieval method based on natural language interaction.

[0050] The steps are as follows:

[0051] (1) Preprocess the user's query request in natural language form to obtain the natural language query text;

[0052] The user's query request in natural language form in step (1) includes query requests in text form and query requests in voice form;

[0053] The preprocessing process for text-based query requests is as follows: text cleaning and word segmentation are performed on the text-based query requests, and irrelevant characters are removed to finally obtain the natural language query text;

[0054] The preprocessing process for query requests in voice form is as follows: the voice query requests are filtered to remove noise, and the noise-removed voice is converted into text to finally obtain the natural language query text;

[0055] Based on step (1), the present invention can process user-inputted text and voice query requests, enabling users to express their information needs in the most intuitive and diverse way, greatly reducing the threshold for system use and improving user experience.

[0056] (2) By using contextual semantic analysis and regular expression pattern matching, the naming identifiers of multiple satellite platforms pre-queried by the user are extracted from the natural language query text, and the naming identifiers of multiple satellite platforms are uniformly converted into a standard format to obtain the pre-query satellite standard format; after extracting the naming identifiers of multiple satellite platforms pre-queried by the user from the natural language query text, the part of the natural language query text that has not been extracted is defined as the remaining query information; the pre-query satellite standard format and the remaining query information are combined to form the pre-query instruction information;

[0057] The specific process of extracting the naming identifier of the satellite platform pre-queried by the user from the natural language query text and converting the naming identifier of the satellite platform into a standard format in step (2) is as follows:

[0058] (2.1) Identify the naming identifiers of satellite platforms contained in natural language query text by using Chinese character, number, and letter combination recognition algorithms and contextual semantic analysis technology;

[0059] (2.2) The identified satellite platform naming identifiers are mapped to a unified naming system through pattern matching and a standardization mapper; the unified naming system is a standard naming system of "Chinese characters + numbers + letters"; the satellite platform naming identifiers converted into the unified naming system are the standard format of the pre-queried satellites;

[0060] Based on step (2), this invention converts the naming identifiers of satellite platforms into a standard format, enabling intelligent recognition and standardized mapping of multiple satellite identifiers, thereby establishing a unified multi-satellite platform identifier processing system and effectively solving the identifier compatibility problem in a multi-satellite environment.

[0061] (3) Input the pre-query instruction information into the large language model fine-tuned by satellite domain knowledge for deep semantic parsing, and output the user's deep query intent and query type; input the pre-query instruction information into the rule engine based on domain knowledge graph to extract professional term features; combine the user's deep query intent, query type and professional term features to obtain structured query understanding results;

[0062] The query types in step (3) include four types: telemetry query, remote control query, fault query, and general query; the technical terminology features include the parameter code, parameter name, instruction name, instruction code, fault name, and fault number of the user query; the query understanding results also include the natural language query text, the query confidence level corresponding to each query type, the keywords in the natural language query text, and the pre-query satellite standard format;

[0063] Based on step (3), this invention utilizes a hybrid understanding mechanism that combines deep learning and rule engines. This fully leverages the semantic understanding capabilities of large language models and incorporates the professional knowledge of domain experts, significantly improving the accuracy and professionalism of users' natural language query understanding, thereby enhancing the accuracy of retrieval queries.

[0064] (4) Establish an intelligent mapping model between query types and knowledge base content; based on the query understanding results, the intelligent mapping model generates an optimal combination of multiple knowledge bases and search filtering conditions corresponding to each knowledge base in the combination. The specific process is as follows:

[0065] (4.1) The intelligent mapping model selects the knowledge bases that form the optimal knowledge base combination based on the query type in the query understanding results. If the query type is a telemetry query, multiple general knowledge bases and telemetry professional knowledge bases are combined to form the optimal knowledge base combination. If the query type is a remote control query, multiple general knowledge bases and remote control professional knowledge bases are combined to form the optimal knowledge base combination. If the query type is a fault query, multiple general knowledge bases and fault professional knowledge bases are combined to form the optimal knowledge base combination. If the query type is a general query, multiple general knowledge bases are combined to form the optimal knowledge base combination.

[0066] (4.2) When the query type is a general query, no operation is performed on the optimal knowledge base combination; otherwise, the intelligent mapping model dynamically adjusts the optimal knowledge base combination based on the query confidence in the query understanding results; the intelligent mapping model dynamically adjusts the optimal knowledge base combination as follows: if the query confidence is high, the general knowledge base in the optimal knowledge base combination is deleted; if the query confidence is low, no operation is performed on the knowledge base.

[0067] (4.3) The intelligent mapping model generates different search filtering conditions for different knowledge bases in the optimal knowledge base combination;

[0068] If the optimal knowledge base combination does not contain a professional knowledge base, then for each general knowledge base in the optimal knowledge base combination, the intelligent mapping model generates general search filtering conditions.

[0069] If the optimal knowledge base combination contains a professional knowledge base, the intelligent mapping model generates professional search filtering conditions for the professional knowledge base in the optimal knowledge base combination; and generates general search filtering conditions for the general knowledge base in the optimal knowledge base combination.

[0070] The professional knowledge base refers to the telemetry professional knowledge base, the remote control professional knowledge base, and the fault professional knowledge base; the general search filtering conditions include satellite ID and document type; the professional search filtering conditions include satellite ID, document type, parameter code, parameter name, command code, remote control code, command name, and fault name;

[0071] Based on step (4), the present invention dynamically selects the optimal knowledge base combination and the corresponding retrieval filtering conditions based on the characteristics of the user's natural language query instructions, and then selects the query results from the optimal knowledge base combination, thereby further ensuring the accuracy of the retrieval.

[0072] (5) Multi-threaded concurrency technology is used to enable each knowledge base in the optimal knowledge base combination to perform semantic vector retrieval and keyword retrieval simultaneously; semantic vector retrieval takes the query understanding results as input, and performs similarity matching in the optimal knowledge base combination based on the deep intent of the user query in the query understanding results to obtain the initial semantic query results; keyword retrieval takes the query understanding results as input, and performs precise matching in the optimal knowledge base combination based on the professional terminology features in the query understanding results to obtain the initial keyword query results; the initial semantic query results and initial keyword query results are filtered based on the retrieval filtering conditions corresponding to each knowledge base to obtain the corresponding semantic query results and keyword query results;

[0073] The method for using multi-threaded concurrency technology in step (5) to enable each knowledge base in the optimal knowledge base combination to simultaneously perform semantic vector retrieval and keyword retrieval is as follows: using thread pool technology, an independent vector retrieval sub-thread and BM25 retrieval sub-thread are created for each knowledge base in the optimal knowledge base combination; the keyword retrieval uses the BM25 retrieval method; the present invention enables each knowledge base in the optimal knowledge base combination to simultaneously perform semantic vector retrieval and keyword retrieval through multi-threaded concurrency technology, thus ensuring the efficiency of the retrieval process.

[0074] Based on step (5), this invention uses a parallel hybrid retrieval strategy to comprehensively utilize both semantic vector retrieval and keyword retrieval methods to retrieve query commands, and integrates the results of the two methods to further optimize the retrieval effect.

[0075] (6) The semantic query results and keyword query results are reordered and merged to obtain the final search results.

[0076] like Figure 2 As shown, this invention also proposes a multi-satellite platform intelligent query and retrieval system based on natural language interaction, including a natural language query unit, a multi-satellite identifier recognition unit, a hybrid query understanding engine, an adaptive routing decision-maker, and a parallel hybrid retrieval system;

[0077] The natural language query unit preprocesses the user's query request in natural language form to obtain the natural language query text;

[0078] The multi-satellite identifier recognition unit extracts the naming identifiers of multiple satellite platforms pre-queried by the user from the natural language query text through contextual semantic analysis and regular expression pattern matching methods, and converts the naming identifiers of multiple satellite platforms into a standard format to obtain the pre-queried satellite standard format; after extracting the naming identifiers of multiple satellite platforms pre-queried by the user from the natural language query text, the unextracted part of the natural language query text is defined as the remaining query information; the pre-queried satellite standard format and the remaining query information are combined to form the pre-query instruction information;

[0079] The hybrid query understanding engine inputs pre-query instruction information into a large language model fine-tuned by satellite domain knowledge for deep semantic parsing, outputting the user's deep query intent and query type; at the same time, it inputs the pre-query satellite standard format into a rule engine based on domain knowledge graph to extract professional terminology features; combining the user's deep query intent, query type, and professional terminology features, it obtains structured query understanding results.

[0080] The adaptive routing decision-maker has an internal intelligent mapping model; based on the query type in the query understanding results, the intelligent mapping model generates the optimal combination of multiple knowledge bases and the retrieval filtering conditions corresponding to each knowledge base in the combination.

[0081] The parallel hybrid retrieval engine employs multi-threaded concurrency technology to enable each knowledge base in the optimal knowledge base combination to simultaneously perform semantic vector retrieval and keyword retrieval. Semantic vector retrieval takes query understanding results as input, and based on the user's deep query intent in the query understanding results, performs similarity matching in the optimal knowledge base combination to obtain initial semantic query results. Keyword retrieval takes query understanding results as input, and based on the professional terminology features in the query understanding results, performs precise matching in the optimal knowledge base combination to obtain initial keyword query results. Subsequently, based on the retrieval filtering conditions corresponding to each knowledge base, the initial semantic query results and initial keyword query results are filtered to obtain the corresponding semantic query results and keyword query results. Finally, the parallel hybrid retrieval engine reorders and merges the semantic query results and keyword query results to obtain the final retrieval results.

[0082] Furthermore, the query instruction information is input into the large oracle model fine-tuned by satellite domain knowledge, and the output of the large oracle model fine-tuned by satellite domain knowledge also includes technical terminology features; the query instruction information is input into the rule engine based on the knowledge domain graph, and the output of the rule engine based on the knowledge domain graph also includes the user's deep query intent and query type.

[0083] Furthermore, the hybrid query understanding engine adopts a dual-processing architecture, featuring a dual-processing collaborative working mode and a single-processing working mode.

[0084] Furthermore, the dual-processing architecture includes a large language model branch and a rule engine branch; the large language model branch is equipped with a large oracle model fine-tuned with satellite domain knowledge, and the rule engine branch is equipped with a rule engine based on a knowledge domain graph; the single-processing working mode includes a single-processing working mode for the large language model branch and a single-processing working mode for the rule engine branch.

[0085] When both the big oracle model branch and the rule engine branch are working normally, the dual-processing architecture adopts a dual-path collaborative working mode. The big language model branch inputs the pre-query instruction information into the big language model, which has been fine-tuned by satellite domain knowledge, for deep semantic parsing and outputs the user's deep query intent and query type. The rule engine branch inputs the pre-query instruction information into the rule engine based on the domain knowledge graph and extracts professional terminology features. Combining the user's deep query intent and query type output by the big language model, which has been fine-tuned by satellite domain knowledge, and the professional terminology features output by the rule engine based on the domain knowledge graph, a structured query understanding result is obtained.

[0086] When an anomaly occurs in the Big Prophecy model branch, the dual-processing architecture adopts a single-processing mode for the rule engine branch. The rule engine branch inputs the pre-query instruction information into the rule engine based on the domain knowledge graph. The rule engine based on the domain knowledge graph outputs the technical terminology features, the user's deep query intent, and the query type. Combining the technical terminology features, the user's deep query intent, and the query type output by the rule engine based on the domain knowledge graph, a structured query understanding result is obtained.

[0087] When the rules engine branch malfunctions, the dual-processing architecture adopts a single-processing mode for the large language model branch. The large language model branch inputs the pre-query instruction information into the large language model, which has been fine-tuned with satellite domain knowledge. The large language model, after being fine-tuned with satellite domain knowledge, outputs the user's deep query intent, query type, and technical terminology features. Combining the technical terminology features, user's deep query intent, and query type output by the large language model with satellite domain knowledge, a structured query understanding result is obtained. Through the above process, it is ensured that when one of the hybrid query understanding engines fails, the other can still work normally and complete the semantic understanding function.

[0088] Furthermore, the hybrid query understanding engine adopts a modular design, supports flexible access to multiple large language models and dynamic updates of the rule base, and automatically and seamlessly switches to the rule engine mode when the large language model service is unavailable, ensuring the continuity and reliability of the system service.

[0089] Furthermore, the adaptive routing decision-maker dynamically adjusts the routing strategy based on the query type and query confidence level, taking into account the technical field, complexity, and timeliness requirements of the query, to ensure the practicality and efficiency of the routing decision.

[0090] The parts of this invention not described in detail are common knowledge to those skilled in the art.

Claims

1. A multi-satellite platform intelligent query and retrieval method based on natural language interaction, characterized in that... The steps are as follows: (1) Preprocess the user's query request in natural language form to obtain the natural language query text; (2) By using contextual semantic analysis and regular expression pattern matching, the naming identifiers of various satellite platforms pre-queried by the user are extracted from the natural language query text, and the naming identifiers of various satellite platforms are uniformly converted into a standard format to obtain the standard format of the pre-queried satellites. After extracting the naming identifiers of various satellite platforms that the user pre-queried from the natural language query text, the unextracted parts of the natural language query text are defined as the remaining query information. The pre-query satellite standard format and the remaining query information are combined to form the pre-query instruction information; (3) Input the pre-query instruction information into the large language model that has been fine-tuned by satellite domain knowledge for deep semantic parsing, and output the user's deep query intent and query type; input the pre-query satellite standard format into the rule engine based on domain knowledge graph to extract professional terminology features; By combining the user's deep query intent, query type, and technical terminology characteristics, a structured query understanding result is obtained; (4) Establish an intelligent mapping model between query types and knowledge base content; Based on the query type in the query understanding results, the intelligent mapping model generates the optimal combination of knowledge bases containing multiple knowledge bases and the search filtering conditions corresponding to each knowledge base in the combination. (5) Multi-threaded concurrency technology is used to enable each knowledge base in the optimal knowledge base combination to perform semantic vector retrieval and keyword retrieval simultaneously; semantic vector retrieval takes the query understanding results as input, and performs similarity matching in the optimal knowledge base combination based on the deep intent of the user query in the query understanding results to obtain the initial semantic query results; keyword retrieval takes the query understanding results as input, and performs precise matching in the optimal knowledge base combination based on the professional terminology features in the query understanding results to obtain the initial keyword query results; Based on the retrieval filtering conditions corresponding to each knowledge base, the initial semantic query results and initial keyword query results are filtered to obtain the corresponding semantic query results and keyword query results; (6) The semantic query results and keyword query results are reordered and merged to obtain the final search results.

2. The intelligent query and retrieval method for multi-satellite platforms based on natural language interaction according to claim 1, characterized in that: The user's query request in natural language form in step (1) includes query requests in text form and query requests in voice form; The preprocessing process for text-based query requests is as follows: text cleaning and word segmentation are performed on the text-based query requests, and irrelevant characters are removed to finally obtain the natural language query text; The preprocessing process for voice query requests is as follows: the voice query requests are filtered to remove noise, and the noise-removed voice is converted into text, ultimately yielding the natural language query text.

3. The intelligent query and retrieval method for multi-satellite platforms based on natural language interaction according to claim 1, characterized in that: The specific process of extracting the naming identifier of the satellite platform pre-queried by the user from the natural language query text and converting the naming identifier of the satellite platform into a standard format in step (2) is as follows: (2.1) Identify the naming identifiers of satellite platforms contained in natural language query text by using Chinese character, number, and letter combination recognition algorithms and contextual semantic analysis technology; (2.2) The identified satellite platform naming identifiers are mapped to a unified naming system through pattern matching and a standardization mapper; the unified naming system is a standard naming system of "Chinese characters + numbers + letters"; the satellite platform naming identifiers converted into the unified naming system are the pre-query satellite standard format.

4. The intelligent query and retrieval method for multi-satellite platforms based on natural language interaction according to claim 1, characterized in that: The query types in step (3) include four types: telemetry query, remote control query, fault query, and general query; the technical terminology features include the parameter code, parameter name, instruction name, instruction code, fault name, and fault number of the user query; the query understanding results also include the natural language query text, the query confidence level corresponding to each query type, the keywords in the natural language query text, and the pre-query satellite standard format.

5. The intelligent query and retrieval method for multi-satellite platforms based on natural language interaction according to claim 4, characterized in that: In step (4), based on the query understanding results, the intelligent mapping model generates the optimal combination of multiple knowledge bases and the retrieval filtering conditions corresponding to each knowledge base in the combination as follows: (4.1) The intelligent mapping model selects the knowledge bases that form the optimal knowledge base combination based on the query type in the query understanding results; If the query type is telemetry query, multiple general knowledge bases and telemetry professional knowledge bases are combined into the optimal knowledge base combination; if the query type is remote control query, multiple general knowledge bases and remote control professional knowledge bases are combined into the optimal knowledge base combination; if the query type is fault query, multiple general knowledge bases and fault professional knowledge bases are combined into the optimal knowledge base combination; if the query type is general query, multiple general knowledge bases are combined into the optimal knowledge base combination. (4.2) When the query type is a general query, no operation is performed on the optimal knowledge base combination; otherwise, the intelligent mapping model dynamically adjusts the optimal knowledge base combination based on the query confidence in the query understanding results; the intelligent mapping model dynamically adjusts the optimal knowledge base combination as follows: if the query confidence is high, the general knowledge base in the optimal knowledge base combination is deleted; if the query confidence is low, no operation is performed on the knowledge base. (4.3) The intelligent mapping model generates different search filtering conditions for different knowledge bases in the optimal knowledge base combination; If the optimal knowledge base combination does not contain a professional knowledge base, then for each general knowledge base in the optimal knowledge base combination, the intelligent mapping model generates general search filtering conditions. If the optimal knowledge base combination contains a professional knowledge base, the intelligent mapping model generates professional search filtering conditions for the professional knowledge base in the optimal knowledge base combination; and generates general search filtering conditions for the general knowledge base in the optimal knowledge base combination. The professional knowledge base refers to the telemetry professional knowledge base, the remote control professional knowledge base, and the fault professional knowledge base; the general search filtering conditions include satellite ID and document type; the professional search filtering conditions include satellite ID, document type, parameter code, parameter name, command code, remote control code, command name, and fault name.

6. The intelligent query and retrieval method for multi-satellite platforms based on natural language interaction according to claim 1, characterized in that: The method for using multi-threaded concurrency technology in step (5) to enable each knowledge base in the optimal knowledge base combination to simultaneously perform semantic vector retrieval and keyword retrieval is as follows: using thread pool technology, an independent vector retrieval sub-thread and BM25 retrieval sub-thread are created for each knowledge base in the optimal knowledge base combination; the keyword retrieval adopts the BM25 retrieval method.

7. A multi-satellite platform intelligent query and retrieval system based on natural language interaction, characterized in that... include: Natural language query unit, multi-satellite identifier recognition unit, hybrid query understanding engine, adaptive routing decision-maker, and parallel hybrid retrieval unit; The natural language query unit preprocesses the user's query request in natural language form to obtain the natural language query text; The multi-satellite identifier recognition unit extracts the naming identifiers of multiple satellite platforms that the user has pre-queried from the natural language query text through contextual semantic analysis and regular expression pattern matching methods, and converts the naming identifiers of multiple satellite platforms into a standard format to obtain the standard format of the pre-queried satellites. After extracting the naming identifiers of various satellite platforms that the user pre-queried from the natural language query text, the unextracted parts of the natural language query text are defined as the remaining query information. The pre-query satellite standard format and the remaining query information are combined to form the pre-query instruction information; The hybrid query understanding engine inputs pre-query instruction information into a large language model fine-tuned by satellite domain knowledge for deep semantic parsing, outputting the user's deep query intent and query type; at the same time, it inputs the pre-query satellite standard format into a rule engine based on domain knowledge graph to extract professional terminology features. By combining the user's deep query intent, query type, and technical terminology characteristics, a structured query understanding result is obtained; The adaptive routing decision-maker has an internal intelligent mapping model; Based on the query type in the query understanding results, the intelligent mapping model generates the optimal combination of knowledge bases containing multiple knowledge bases and the search filtering conditions corresponding to each knowledge base in the combination. The parallel hybrid retrieval system employs multi-threaded concurrency technology to enable each knowledge base in the optimal knowledge base combination to simultaneously perform semantic vector retrieval and keyword retrieval. Semantic vector retrieval takes the query understanding results as input and performs similarity matching in the optimal knowledge base combination based on the user's deep query intent in the query understanding results to obtain initial semantic query results. Keyword retrieval takes the query understanding results as input and performs precise matching in the optimal knowledge base combination based on the professional terminology features in the query understanding results to obtain initial keyword query results. Then, based on the retrieval filtering conditions corresponding to each knowledge base, the initial semantic query results and the initial keyword query results are filtered to obtain the corresponding semantic query results and keyword query results; Finally, the parallel hybrid search engine reorders and merges the semantic query results and keyword query results to obtain the final search results.

8. The intelligent query and retrieval system based on natural language interaction across multiple satellite platforms according to claim 7, characterized in that: The query command information is input into the big oracle model fine-tuned by satellite domain knowledge, and the output of the big oracle model fine-tuned by satellite domain knowledge also includes technical terminology features; the query command information is input into the rule engine based on the knowledge domain graph, and the output of the rule engine based on the knowledge domain graph also includes the user's deep query intent and query type.

9. The intelligent query and retrieval system based on natural language interaction across multiple satellite platforms according to claim 8, characterized in that: The hybrid query understanding engine adopts a dual-processing architecture, with both dual-processing collaborative processing mode and single-processing mode.

10. The intelligent query and retrieval system based on natural language interaction across multiple satellite platforms according to claim 9, characterized in that: The dual-processing architecture includes a large language model branch and a rule engine branch; the large language model branch is equipped with a large oracle model fine-tuned with satellite domain knowledge, and the rule engine branch is equipped with a rule engine based on a knowledge domain graph. The single-path processing mode includes the large language model branch single-path processing mode and the rule engine branch single-path processing mode. When both the big oracle model branch and the rule engine branch are working normally, the dual-processing architecture adopts a dual-path collaborative working mode; the big language model branch inputs the pre-query instruction information into the big language model that has been fine-tuned by satellite domain knowledge for deep semantic parsing and then outputs the user's deep query intent and query type; the rule engine branch inputs the pre-query instruction information into the rule engine based on the domain knowledge graph and then extracts professional terminology features. By combining the deep intent and query type of user queries output by the large language model fine-tuned by satellite domain knowledge with the professional terminology features output by the rule engine based on domain knowledge graph, a structured query understanding result is obtained. When an anomaly occurs in the Big Prophecy model branch, the dual-processing architecture adopts the single-processing mode of the rule engine branch. The rule engine branch inputs the pre-query instruction information into the rule engine based on the domain knowledge graph, and the rule engine based on the domain knowledge graph outputs the technical term features, the user's deep query intent, and the query type. By combining the professional terminology features, deep user query intent, and query type output by the rule engine based on domain knowledge graph, a structured query understanding result is obtained; When an anomaly occurs in the rule engine branch, the dual-processing architecture adopts the single-processing mode of the large language model branch. The large language model branch inputs the pre-query instruction information into the large language model that has been fine-tuned by satellite domain knowledge. The large language model that has been fine-tuned by satellite domain knowledge outputs the user's deep query intent, query type, and technical terminology characteristics. By combining the characteristics of specialized terms, the user's deep query intent, and the query type output by the large language model fine-tuned with satellite domain knowledge, a structured query understanding result is obtained.

Images