Data query analysis method based on large model driving

Abstract

The invention discloses a data query analysis method based on large model driving, and aims to solve the problems of weak natural language understanding, low professional term recognition accuracy, poor multi-dimensional data storage access efficiency and the like in the existing marine data query, and realize intelligent mapping and efficient retrieval from natural languages to professional data. The method comprises the following steps: constructing a five-dimensional ontology marine terminology library containing a terminology layer, a synonymous mapping layer and the like, and extracting spatio-temporal information to establish an index; configuring a term conversion engine, identifying term variants and detecting synonymous conflicts; deploying a Zarr / MinIO storage system, adding marine exclusive metadata, performing hierarchical storage according to hot / warm / cold data, and performing adaptive partitioning; a multi-engine retrieval system is established, and query and data paths are mapped through storage-semantic intelligent routing. The term recognition accuracy of the method reaches 92.7%, 100GB data set retrieval is only 1.3 seconds, the storage compression efficiency is improved by 40%, the method can support the fields of ocean scientific research, environment monitoring and the like, and technical support is provided for intelligent application of ocean big data.

Description

Technical Field

[0001] This invention relates to the field of intelligent marine data processing technology, and in particular to a data query and analysis method based on large model-driven approaches. Background Technology

[0002] With the rapid development of marine observation technology and the continuous improvement of observation networks, marine scientific data is growing at an unprecedented rate. Data types encompass various sources, including satellite remote sensing, buoy observation, shipborne surveys, and deep-sea exploration, and data formats include various professional formats such as NetCDF, HDF5, and GRIB. This marine data exhibits significant multidimensional characteristics, typically containing multiple dimensions such as longitude, latitude, depth, and time, with individual data files often reaching hundreds of GB or even TB levels. Marine researchers and related industry users face multiple challenges when querying and analyzing this data, including its massive volume, complex formats, and numerous technical terms. Traditional data management and retrieval methods are no longer sufficient to meet the application needs of the marine big data era.

[0003] Existing marine data query and analysis technologies primarily rely on traditional relational database management systems and file system retrieval methods, which have significant limitations when processing marine data. First, traditional databases struggle to efficiently store and index multidimensional array data, lack sufficient support for scientific data formats such as NetCDF, and require complex data conversion and preprocessing, severely impacting data access efficiency. Second, existing systems lack the ability to intelligently understand marine terminology; users must use precise technical vocabulary and complex queries to retrieve the required data, creating a high barrier to entry for non-specialist users. Furthermore, traditional methods cannot effectively handle natural language queries, failing to understand the semantic connections and implicit information in user query intent; for example, they cannot automatically associate "red tide" with related concepts such as "harmful algal bloom" or "abnormal chlorophyll concentration."

[0004] Furthermore, existing technologies have shortcomings in handling the spatiotemporal characteristics of marine data, lacking optimized storage strategies and retrieval mechanisms tailored to the specific features of marine data. Marine data typically exhibits significant spatiotemporal correlation and continuity, but traditional storage systems employ generic data organization methods that fail to consider access patterns and query characteristics, leading to inefficiencies in time-series analysis or spatial-range queries. Simultaneously, existing systems lack knowledge graph and semantic association capabilities, making it impossible to establish causal relationships and correlation networks between marine phenomena, hindering support for complex comprehensive analytical queries, such as queries involving multiple factors and processes, like "the impact of typhoons on the marine environment." Therefore, a new method for marine data query and analysis is urgently needed, capable of understanding natural language, recognizing technical terms, optimizing storage access, and supporting intelligent retrieval. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problems existing in current marine data query and analysis technologies, such as insufficient natural language understanding capabilities, low accuracy in identifying technical terms, poor efficiency in accessing multidimensional data storage, and slow response speed for complex queries. This invention provides a data query and analysis method based on a large model-driven approach, enabling intelligent mapping and efficient retrieval from natural language to technical data.

[0006] To achieve the above objectives, this invention provides a data query and analysis method based on a large model, comprising the following steps: A marine terminology database was constructed, employing a five-dimensional ontology structure, including a terminology layer, a synonym mapping layer, a spatiotemporal rule layer, a scientific research logic layer, and a graph association layer. The terminology layer captures contextual features of terms through the attention mechanism of a pre-trained large language model and annotates specialized terms using named entity recognition technology. The synonym mapping layer vectorizes terms using embedding technology to generate high-dimensional semantic vector representations. The spatiotemporal rule layer extracts longitude, latitude, depth, and time dimension information by parsing the CF standard metadata of NetCDF files to establish a spatiotemporal index system. Configure a terminology conversion engine, analyze the contextual information of term occurrences, utilize the semantic understanding capabilities of large models to distinguish term usage scenarios, construct a terminology morphology rule base to identify abbreviations, full names, and Chinese-English translations of term variants, and perform synonym conflict detection by calculating semantic similarity and contextual consistency between terms; The storage system is deployed using the Zarr format in conjunction with the MinIO object storage architecture. The standard Zarr format is enhanced with CF metadata standard support and exclusive dimensional attributes such as ocean depth level identifier, tidal cycle marker, and ocean current direction attribute. The data is divided into three levels—hot data, warm data, and cold data—through a hierarchical storage strategy and stored in storage pools with different performance levels. A retrieval system is established that integrates three storage engines: vector database, graph database, and relational database. Through a storage-semantic intelligent routing mechanism, it realizes automatic mapping from natural language queries to data paths. The vector database stores the embedding vectors of terms, the graph database stores the knowledge graph structure, and the relational database stores structured metadata.

[0007] Furthermore, in the five-dimensional ontology structure, the scientific research logic layer constructs a classification system covering physical oceanography, marine biology, marine chemistry, and marine geology subfields by analyzing the discipline system of marine science, and formulates data governance rules and semantic interpretation templates according to the characteristics of different disciplines; the graph association layer integrates the information of each layer to form a marine knowledge graph system, and realizes cross-level and cross-dimensional retrieval through the reasoning ability of large models.

[0008] Furthermore, the terminology conversion engine uses a hybrid approach of rule-based and statistical methods to identify morphological variants, taking into account both linguistic patterns and the statistical features of large-scale corpora. When it detects the use of undefined synonyms or terminological ambiguities in a document, it automatically marks them and provides correction suggestions through a built-in thesaurus and conflict detection algorithm.

[0009] Furthermore, the storage system achieves data format conversion by developing an xarray to Zarr converter, and automatically adds ocean-specific metadata such as data source organization, observation equipment type, and data quality tags during the conversion process; it adopts a depth-adaptive ocean data partitioning strategy, dynamically adjusting the partition size according to data dimensional characteristics and access patterns, prioritizing time dimension partitioning for data that is frequently subjected to time series analysis, and prioritizing spatial dimension partitioning for data that is mainly subjected to spatial analysis.

[0010] Furthermore, in the tiered storage strategy, hot data is stored in an SSD storage pool, warm data is stored in a hybrid storage pool, and cold data is stored in an HDD storage pool; the migration of data between different storage levels is dynamically adjusted according to the data access frequency and importance.

[0011] Furthermore, the retrieval system employs model quantization and inference optimization techniques for local deployment of large-scale models. During query processing, the system first performs semantic parsing on natural language queries using large-scale models to identify query intent and key terms. Then, it selects a combination of engines—vector database, graph database, and relational database—based on the query type. Finally, it integrates and sorts the multi-engine retrieval results using a result fusion algorithm.

[0012] Furthermore, the storage-semantic intelligent routing mechanism automatically generates a data storage path format "ocean-data / {data type} / {time} / {space}.zarr" by parsing the technical terms in the user query. The data type includes temperature, salinity, flow field, and chlorophyll parameters; the time includes year, season, and month identifiers; and the space includes sea area name or latitude and longitude range identifiers.

[0013] Furthermore, the spatiotemporal rule layer stores the extracted longitude, latitude, depth, and time attributes as attributes of nodes and edges in the graph database, establishing a graph-based retrieval system based on spatiotemporal constraints. When a query containing spatiotemporal constraints is received, the spatiotemporal index of the graph database is used to quickly locate data nodes that meet the conditions, and the corresponding Zarr format data file path is returned.

[0014] The above method constructs a complete intelligent data query and analysis system. First, it establishes a marine terminology database based on a five-dimensional ontology structure. Through a layered design of terminology layer, synonym mapping layer, spatiotemporal rule layer, scientific research logic layer, and graph association layer, it achieves comprehensive coverage and deep understanding of marine terminology. The terminology layer utilizes the attention mechanism and named entity recognition technology of a pre-trained large language model to accurately capture and label marine terminology; the synonym mapping layer establishes a semantic association network between terms through embedding vectorization technology; the spatiotemporal rule layer parses CF standard metadata to construct a spatiotemporal index system; the scientific research logic layer formulates a classification system and governance rules based on the characteristics of marine science; and the graph association layer integrates multi-layered information to form a complete marine knowledge graph, supporting cross-dimensional intelligent retrieval.

[0015] In terms of terminology understanding and conversion, this invention designs an intelligent terminology conversion engine, which possesses three core capabilities: context awareness, morphological variant recognition, and synonym conflict detection. By analyzing the contextual information of terms and utilizing the semantic understanding capabilities of large models, the system can accurately distinguish the meanings of terms in different scenarios; by constructing a morphological rule base and employing a combination of rule and statistical methods, it achieves accurate identification of various variant forms of terms; and through a thesaurus and conflict detection algorithms, it automatically discovers and corrects ambiguities in terminology usage. This multi-layered terminology processing mechanism ensures accurate conversion from user natural language input to professional terminology.

[0016] In terms of data storage optimization, this invention adopts an innovative architecture combining the Zarr format with MinIO object storage, and has been deeply optimized for the characteristics of marine data. The system enhances the standard Zarr format, adding support for the CF metadata standard and marine-specific dimensional attributes, including unique identifiers such as ocean depth levels, tidal cycles, and ocean current directions. By implementing a hierarchical storage strategy, data is allocated to storage pools of different performance levels based on access frequency and importance. Hot data is stored on SSDs for fast access, while warm and cold data are stored in hybrid storage and HDDs respectively to balance performance and cost. Particularly important is the system's implementation of a depth-adaptive block partitioning strategy, dynamically adjusting the partitioning method according to data access patterns, significantly improving the access efficiency of multidimensional data.

[0017] In terms of intelligent retrieval, this invention constructs a multi-engine integrated retrieval system, integrating three storage engines: vector database, graph database, and relational database. It achieves intelligent query parsing and result fusion through localized deployment of a large model. The system establishes a storage-semantic intelligent routing mechanism, which can automatically map natural language queries to specific data storage paths. Through model quantization and inference optimization techniques, the system can achieve millisecond-level query response in a typical server environment. During the retrieval process, the system first parses the user's query intent, identifies key terms, then selects an appropriate engine combination for retrieval, and finally integrates multi-source results through a fusion algorithm to return the most relevant data.

[0018] The beneficial effects of this invention are: According to actual tests, the terminology recognition accuracy reached 92.7%, a 43% improvement compared to directly using general large models, significantly lowering the barrier for users to access professional data. In terms of data access performance, the system achieved rapid retrieval of a 100GB dataset within 1.3 seconds, a 5.8-fold speedup compared to the traditional NetCDF sequential reading method, and improved storage compression efficiency by 40%, effectively addressing the storage pressure of petabyte-scale marine data. More importantly, the system supports complex natural language queries and multidimensional association analysis. Users can describe their query needs using everyday language, and the system can understand the query intent and return accurate results, which is of great significance for promoting the widespread application of marine data. The method of this invention has demonstrated enormous application potential in multiple fields such as marine scientific research, environmental monitoring, distant-water fisheries, and offshore energy exploration, providing strong technical support for the intelligent application of marine big data. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a system architecture diagram of the present invention, which shows the complete system architecture from natural language query input to final data storage; Figure 2 The diagram shows the hierarchical structure of the five-dimensional ontology terminology library of this invention, illustrating the five-layer architecture of the terminology library; Figure 3 The diagram shows the architecture of the Zarr / MinIO hierarchical storage system of this invention, illustrating the data conversion process from the original format to the Zarr format. Figure 4 This is a flowchart of the terminology conversion engine of the present invention, which illustrates the three core steps of terminology conversion in flowchart form; Figure 5 The diagram shows the multi-engine integrated architecture of the retrieval system of this invention, illustrating the localized deployment of large models, intelligent routing decisions, and the parallel retrieval and result fusion mechanism of three database engines. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0022] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0023] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0024] See Figures 1 to 5 As shown This invention provides a data query and analysis method based on a large model. This method achieves a complete technical closed loop from natural language query to professional data retrieval by constructing a marine terminology database, deploying an intelligent storage system, and implementing a multi-engine retrieval mechanism.

[0025] Implementation Methods for Constructing a Marine Terminology Database The terminology database is constructed using a five-dimensional ontology structure, comprising five core layers: a terminology layer, a synonym mapping layer, a spatiotemporal rule layer, a research logic layer, and a knowledge graph association layer. In the terminology layer, the system uses a large-scale model to understand and extract contextual semantics, accurately identifying specialized phrases. Specifically, marine science literature is preprocessed by inputting the text into a pre-trained large-scale language model. The model's attention mechanism captures the contextual features of terms, and named entity recognition technology is used to annotate potential specialized terms. Preliminary screening is then performed based on the frequency of terminology occurrence and contextual relevance. The synonym mapping layer utilizes a large-scale terminology model for interpretation and enhancement techniques to generate concise and accurate draft definitions or explanatory descriptions for candidate or existing terms. During this process, the system vectorizes terms using embedding technology to generate high-dimensional semantic vector representations, storing these vectorized terms in a knowledge graph database to form a semantic association network between terms.

[0026] The implementation focus of the spatiotemporal rules layer lies in parsing and constructing the underlying metadata knowledge graph. The system extracts dimensional information such as longitude, latitude, depth, and time by parsing the CF standard metadata of NetCDF files, establishing a spatiotemporal index system, and storing these spatiotemporal attributes as node and edge attributes in a graph database to achieve graph-based retrieval based on spatiotemporal constraints. The scientific research logic layer classifies, governs, and interprets data based on professional marine scientific research logic. The system analyzes the disciplinary system and research paradigms of marine science, constructing a classification system covering subfields such as physical oceanography, marine biology, marine chemistry, and marine geology, and formulating corresponding data governance rules and semantic interpretation templates according to the characteristics of different disciplines. The graph association layer, as the highest level, is responsible for constructing a multidimensional knowledge graph and achieving fusion retrieval based on a large model. This layer integrates the information from the aforementioned layers to form a complete marine knowledge graph system, and achieves intelligent retrieval across levels and dimensions through the reasoning capabilities of the large model.

[0027] Implementation method of terminology conversion engine The core functions of the terminology conversion engine include three aspects: context awareness of terms within a large model, morphological variant recognition, and synonym conflict detection. In the implementation of context awareness, the system analyzes the contextual information in which terms appear and utilizes the semantic understanding capabilities of the large model to distinguish the correct usage scenarios of terms. For example, when a user inputs "red tide," the system can determine from the context that it refers to the marine biological phenomenon of "harmful algal blooms," rather than any other meaning. The implementation of morphological variant recognition involves constructing a morphological rule base for terms to identify different forms of terminology, including abbreviations, full names, and Chinese-English translations. The system employs a hybrid approach based on rules and statistics, considering both linguistic rules and utilizing the statistical features of a large-scale corpus to ensure accurate identification of various forms of professional terminology. Synonym conflict detection is implemented by establishing a thesaurus and a conflict detection algorithm. When an undefined synonym is found in a document or when terminology ambiguity exists, the system automatically marks it and provides correction suggestions. This process determines the existence of synonym conflicts by calculating the semantic similarity and contextual consistency between terms.

[0028] Implementation Methods of Zarr / MinIO Storage System The storage system employs an architecture combining the Zarr format with MinIO object storage to achieve efficient storage and management of marine data. In the specific implementation of data encapsulation, the system enhances the standard Zarr format by adding support for the CF metadata standard and adds unique dimensional attributes tailored to the characteristics of marine data, including ocean depth level identifiers, tidal cycle markers, and ocean current direction attributes—metadata fields specific to the ocean. The storage cluster deployment adopts a tiered storage strategy using MinIO, classifying data into three levels—hot data, warm data, and cold data—based on access frequency and importance. Hot data is stored in a high-performance SSD storage pool to ensure fast access, warm data is stored in a hybrid storage pool to balance performance and cost, and cold data is stored in a large-capacity HDD storage pool for long-term archiving.

[0029] The data transformation tool was implemented by developing an xarray to Zarr converter. This converter not only converted the data format but also automatically added ocean-specific metadata during the conversion process, including information such as the data source organization, observation equipment type, and data quality tags. During the conversion, the system employs a depth-adaptive ocean data partitioning strategy, dynamically adjusting the partition size based on the data's dimensional characteristics and access patterns. For example, data frequently used for time-series analysis is partitioned with a time-dimension-first approach, while data primarily used for spatial analysis is partitioned with a spatial-dimension-first approach. This adaptive partitioning mechanism significantly improves data access efficiency.

[0030] Implementation methods of retrieval system The retrieval system adopts a multi-engine integrated architecture, including three storage engines: a vector database, a graph database, and a general relational database. It achieves intelligent query parsing and result fusion through localized deployment of large models. The vector database stores embedding vectors for terms, supporting semantic similarity-based retrieval; the graph database stores the knowledge graph structure, supporting complex relation-based queries; and the relational database stores structured metadata, supporting exact match queries. The system establishes a storage-semantic intelligent routing mechanism to automatically map natural language queries to specific data paths. For example, when a user queries data related to "mesoscale eddies," the system automatically routes the query to the corresponding data file under the path "ocean-data / eddies / {id}.zarr".

[0031] The local deployment of the large model employs model quantization and inference optimization techniques to reduce computational resource requirements while ensuring inference accuracy, enabling the system to achieve millisecond-level query response times in a typical server environment. The query processing flow first uses the large model to semantically analyze the user's natural language query, identifying the query intent and key terms. Then, it selects an appropriate combination of search engines based on the query type. Finally, a result fusion algorithm integrates and sorts the search results from multiple engines, returning the most relevant data results.

[0032] Example Example 1: Ocean Temperature Anomaly Inquiry When researchers input the query "warming areas of the western Pacific Ocean in the summer of 2023," the system's processing flow is as follows: First, the terminology conversion engine identifies "warming of seawater" as the corresponding technical term "sea surface temperature anomaly (SSTAnomaly)," and maps "western Pacific Ocean" to a specific latitude and longitude range (120°E-180°E, 0°-40°N). Then, the spatiotemporal rule layer resolves the time constraint as June to August 2023 and the spatial constraint as the western Pacific Ocean region. The retrieval system uses an intelligent routing mechanism to locate the storage path "ocean-data / sst / 2023 / summer / wpacific.zarr" and reads the corresponding temperature data blocks from the Zarr file. The system completes the retrieval of the 100GB dataset within 1.3 seconds, returning a temperature anomaly distribution map and statistical analysis results, including key information such as the area of ​​the anomaly region, the maximum temperature deviation value, and the duration.

[0033] Example 2: Analysis of Marine Environmental Changes Under the Influence of Typhoons When researchers queried about "ocean warming caused by typhoons," the system demonstrated its ability to understand complex semantics. The terminology conversion engine first identified the association between "typhoon" and "tropical cyclone," and then, through a scientific logic layer, deduced that the typhoon's impact includes multiple related phenomena such as "decreasing sea surface temperature," "enhanced upper ocean mixing," and "subsurface warming." The knowledge graph association layer retrieved multi-source datasets, including typhoon track data, sea surface temperature data, and ocean vertical profile data, through knowledge graph retrieval. The system automatically constructed a query plan, first obtaining the typhoon track from "ocean-data / typhoon / tracks.zarr," then extracting temperature profile data from "ocean-data / temperature / profiles.zarr" based on the track buffer, and finally generating a comparative analysis of temperature changes before and after the typhoon. The entire query process demonstrated the system's ability to handle complex causal relationship queries.

[0034] Example 3: Data retrieval from monitoring harmful algal blooms When marine environmental monitoring departments query "chlorophyll concentration in areas with recent red tide outbreaks," the system uses a synonym mapping layer to convert "red tide" to "harmful algal bloom (HAB)" and associates it with the indicator parameter "chlorophyll a concentration." Based on data timeliness, the storage system prioritizes retrieving satellite remote sensing data from the hot data storage pool for the past 7 days, via the path "ocean-data / satellite / chlorophyll / realtime.zarr". The system optimizes spatial data reading through a depth-adaptive block partitioning strategy, enabling rapid loading of data over a large area of ​​the ocean. The search results not only include a chlorophyll concentration distribution map but also automatically supplement data on relevant environmental factors such as ocean current direction, water temperature, and nutrients through map correlation, providing comprehensive data support for algal bloom prediction and early warning.

[0035] Example 4: Comprehensive Inquiry of Deep-Sea Environmental Elements When querying "temperature, salinity, and density data below 3000 meters in the Mariana Trench" for a deep-sea research project, the system demonstrated its ability to handle complex multi-dimensional queries. The spatiotemporal rule layer resolved depth constraints (>3000m) and geographic location constraints (Mariana Trench region). The terminology database automatically identified "temperature, salinity, and density" as corresponding to the three physical parameters: "temperature," "salinity," and "density." The storage system quickly located the "ocean-data / deep-sea / mariana / ctd.zarr" dataset through metadata indexing. This dataset employs a depth-first, block-based storage strategy to ensure efficient reading of vertical profile data. When processing comprehensive queries containing multiple parameters, the system optimized query response time by reading data blocks of different parameters in parallel, achieving a 5.8-fold speedup compared to the sequential reading method of the traditional NetCDF format.

[0036] Through the above implementation methods and examples, this invention achieves a term recognition accuracy of 92.7%, a 43% improvement compared to directly using a general large model; the data retrieval response time is reduced to 1.3 seconds (for a 100GB dataset), a 5.8-fold speedup compared to traditional methods; and storage compression efficiency is improved by 40%, effectively saving petabyte-level storage space. This invention successfully overcomes the technical bottleneck of intelligent marine data retrieval, providing efficient data analysis support for marine research institutions, environmental monitoring departments, distant-water fisheries, and offshore energy exploration.

[0037] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0038] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A large model driving-based data query analysis method, characterized in that, Comprising the following steps: A marine professional term library is constructed, which adopts a five-dimensional ontology structure including a term layer, a synonym mapping layer, a space-time rule layer, a scientific research logic layer, and a graph correlation layer; the term layer captures the context features of the terms through the attention mechanism of a pre-trained large language model, and professional terms are labeled using named entity recognition technology; The synonym mapping layer vectorizes the terms through Embedding technology and generates high-dimensional semantic vector representations; the space-time rule layer extracts longitude, latitude, depth, and time dimension information to establish a space-time index system by analyzing the CF standard metadata of NetCDF files; A term conversion engine is configured to analyze the context information of the terms, use the semantic understanding ability of a large model to distinguish the term usage scenarios, construct a term morphology rule library to identify abbreviations, full names, and English-Chinese term variants, and perform synonym conflict detection by calculating the semantic similarity and context consistency between terms; A storage system is deployed, which uses the Zarr format and the MinIO object storage architecture, adds CF metadata standard support and exclusive dimension attributes such as marine depth level identification, tidal cycle marking, and current direction attributes to the standard Zarr format, and stores data in three levels of hot data, warm data, and cold data in storage pools with different performance through hierarchical storage strategies; A retrieval system is established, which integrates vector databases, graph databases, and relational databases as three storage engines, and realizes automatic mapping from natural language queries to data paths through a storage-semantic intelligent routing mechanism, wherein the vector database stores the Embedding vectors of the terms, the graph database stores the knowledge graph structure, and the relational database stores the structured metadata.

2. The method of claim 1, wherein: In the five-dimensional ontology structure, the scientific research logic layer analyzes the discipline system of marine science to construct a classification system covering physical oceanography, marine biology, marine chemistry, and marine geology sub-fields, and formulates data governance rules and semantic interpretation templates according to the characteristics of different disciplines; the graph correlation layer integrates information from each layer to form a marine knowledge graph system, and realizes cross-level and cross-dimensional retrieval through the reasoning ability of a large model.

3. The method of claim 1, wherein: The term conversion engine identifies morphological variants through a hybrid method based on rules and statistics, considering both linguistic rules and statistical characteristics of large-scale corpora; When undefined synonyms or term ambiguities are found in the document, the established synonym dictionary and conflict detection algorithm automatically mark and provide correction suggestions.

4. The method of claim 1, wherein: The storage system realizes data format conversion by developing an xarray to Zarr converter, automatically adding marine-specific metadata such as data source agencies, observation device types, and data quality labels during the conversion process; and adopts a deep adaptive marine data blocking strategy to dynamically adjust the blocking size according to data dimension characteristics and access patterns, and uses time-dimension-first blocking for data frequently subjected to time series analysis, and space-dimension-first blocking for data mainly subjected to spatial analysis.

5. The method of claim 1, wherein: In the hierarchical storage strategy, hot data is stored in the SSD storage pool, warm data is stored in the hybrid storage pool, and cold data is stored in the HDD storage pool; and the migration of data among different storage levels is dynamically adjusted according to data access frequency and importance.

6. The method of claim 1, wherein: The retrieval system adopts model quantization and inference optimization technology for large model localization deployment; during query processing, the natural language query is first semantically analyzed by the large model to identify the query intent and key terms, then the engine combination of vector databases, graph databases, and relational databases is selected according to the query type, and finally the multi-engine retrieval results are integrated and sorted by a result fusion algorithm.

7. The method of claim 1 or 6, wherein: The storage-semantic intelligent routing mechanism automatically generates a data storage path format "ocean-data / {data type} / {time} / {space}.zarr" by analyzing professional terms in the user query, wherein the data type includes temperature, salinity, flow field, and chlorophyll parameters, the time includes year, season, and month identifiers, and the space includes sea area name or latitude and longitude range identifiers.

8. The method of claim 1, wherein: The spatio-temporal rule layer stores the extracted longitude, latitude, depth, and time attributes as node and edge attributes in the graph database, and establishes a graph-based retrieval based on spatio-temporal constraints; when receiving a query containing spatio-temporal constraints, the graph database's spatio-temporal index is used to quickly locate the data nodes that meet the conditions, and the corresponding Zarr format data file path is returned.

Images