LNG storage tank knowledge modeling and intelligent optimization integrated method and system

Abstract

This invention discloses an integrated method and system for knowledge modeling and intelligent optimization of LNG (liquefied natural gas) storage tanks, belonging to the field of data processing. It aims to solve the problems of data silos, lack of automated model generation, and high optimization costs in LNG storage tank design. The method includes: acquiring design task information; constructing a domain ontology containing three-level concepts (assembly, component, and part), composition and contact relationships, and location, geometry, and material properties; converting this ontology ontology into a graph database through schema-level and instance-level mapping; adaptively generating multidisciplinary models based on user-specified types using this graph database; embedding finite element analysis into the main loop of a genetic algorithm; using iteratively accumulated samples to train and dynamically update a deep neural network surrogate model online; outputting optimal parameters and writing them back to the graph database. This invention achieves data homogeneity and automatic derivation of multidisciplinary models, significantly reducing optimization costs and improving design efficiency and quality.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an integrated method and system for knowledge modeling and intelligent optimization of LNG storage tanks. Background Technology

[0002] The structure of large, fully containment steel LNG storage tanks, such as... Figure 1 As shown, it includes the inner tank wall, insulation layers (bottom insulation layer and side insulation layer), inner tank bottom plate, outer tank wall, outer tank bottom plate, pile cap, lifting rods, steel dome, and outermost steel accessories. As a novel tank structure, unlike traditional concrete storage tanks, it adopts a double-layer thin-walled steel structure system, with both the inner and outer tanks constructed of steel plates with a thickness of 8 to 22 millimeters. The design of LNG storage tanks is a highly complex systems engineering project, involving deep collaboration among multiple disciplines such as structural mechanics, fluid mechanics, materials engineering, and construction management. A complete LNG storage tank design project typically consists of multiple parallel professional branches, each requiring the generation of its own dedicated engineering model based on unified design data, and achieving structural lightweighting while meeting safety constraints.

[0003] Under the traditional design paradigm, LNG storage tank engineering practice mainly faces the following two levels of technical challenges:

[0004] The first level presents the challenge of multidisciplinary model generation and data management. Based on the same design data foundation, at least three parallel professional model generation needs exist: structural engineers use finite element analysis software to build mechanical analysis models to verify the structural safety of storage tanks under conditions such as self-weight, internal pressure, wind load, and earthquakes; building information modeling engineers use software such as Revit and CATIA to build building information models for multidisciplinary clash detection, construction simulation, and asset management; and for visualization or virtual reality walkthroughs, geometric models need to be exported as surface mesh models in STL format. The finite element models, industrial foundation models, and STL models used in these three scenarios are traditionally built independently and stored in different file systems or databases. The geometric dimensions, material properties, and design parameters of the same storage tank need to be manually entered and modified repeatedly in multiple unrelated systems such as ANSYS and Revit, which is prone to data conflicts due to version inconsistencies. Data exchange between heterogeneous models relies on intermediate file formats or API conversion, and problems such as geometric interpretation deviations, loss of attribute information, and broken semantic associations persist during the conversion process. There is a lack of a unified automated mapping path from the same design data base to different professional models, and the generation of each model requires independent manual operation or dedicated conversion tools. In recent years, ontology technology and graph databases have been applied in the field of architectural engineering, but existing research mainly focuses on information retrieval and semantic reasoning for single model types, and has not yet established an adaptive generation framework from a unified graph data base to multiple professional models.

[0005] The second level of challenges stems from the high computational cost and lack of intelligent guidance in structural optimization. In the structural design of LNG storage tanks, an optimal balance between safety and economy must be sought. The optimization process typically involves dozens of continuous dimensional variables, integer quantity variables, and discrete cross-section type variables, constituting a complex mixed-integer nonlinear programming problem. Traditional optimization methods directly embed finite element analysis into iterative loops. Each evaluation of a candidate design requires a complete nonlinear finite element calculation. For a tank structure composed of thousands of components, the entire optimization process requires hundreds or thousands of evaluations, resulting in extremely high computational costs. To reduce the computational burden, surrogate model technology has been introduced into the field of structural optimization. However, most existing surrogate model-assisted optimization methods employ offline training strategies: the model is pre-trained using a large number of FEA samples and then fixed for the optimization process. Offline samples are insufficient to cover the entire design space, leading to insufficient prediction accuracy in unknown areas. Furthermore, the surrogate model cannot utilize newly generated FEA data during the optimization process for self-improvement. More critically, when design data changes, existing finite element models, industrial foundation models, and STL models all need to be manually regenerated, further exacerbating the efficiency bottleneck of iterative design. Summary of the Invention

[0006] This invention provides an integrated method and system for knowledge modeling and intelligent optimization of LNG storage tanks. The technical problems it aims to solve are: existing LNG storage tank designs lack a unified data foundation and an automatic generation path for multidisciplinary models; finite element analysis in structural optimization is costly and offline proxy models cannot be dynamically updated; and there is a lack of a closed-loop linkage mechanism between knowledge modeling and intelligent optimization.

[0007] To address the aforementioned technical problems, this invention provides an integrated method for LNG storage tank knowledge modeling and intelligent optimization, comprising the following steps:

[0008] S1. Construct an ontology graph database based on various design task information of liquefied natural gas storage tanks, including:

[0009] Obtain the design task information of the LNG storage tank, and construct the LNG storage tank domain ontology based on the design task information. The LNG storage tank domain ontology includes a concept set, a relationship set, and an attribute set. The concept set includes structural concepts at three levels: assembly level, component level, and part level. The relationship set includes compositional relationships and contact relationships. The attribute set includes location attributes, geometric attributes, and material attributes.

[0010] The LNG storage tank ontology is converted into the architectural constraints of a graph database, and the engineering data of the LNG storage tank is instantiated into the graph database according to the architectural constraints to generate an ontology graph database.

[0011] S2. Based on the ontology graph database, an adaptive multidisciplinary model of the liquefied natural gas storage tank is generated, wherein the multidisciplinary model includes at least a finite element analysis model;

[0012] S3. Based on the ontology graph database and the finite element analysis model, online intelligent optimization of the LNG storage tank structure is performed using a deep neural network surrogate model and a genetic algorithm to obtain the optimal design parameters of the liquefied natural gas storage tank.

[0013] This invention also provides an integrated system for LNG storage tank knowledge modeling and intelligent optimization. Applying the aforementioned integrated method for LNG storage tank knowledge modeling and intelligent optimization, the system includes:

[0014] The graph database construction module is used to build an ontology graph database based on various design task information of liquefied natural gas storage tanks;

[0015] The model generation module is used to adaptively generate a multidisciplinary model of a liquefied natural gas storage tank based on the ontology graph database.

[0016] The proxy-assisted optimization module is used to perform online intelligent optimization of the LNG storage tank structure based on the ontology graph database and the finite element analysis model in the multidisciplinary model, and based on the deep neural network proxy model and genetic algorithm to obtain the optimal design parameters of the liquefied natural gas storage tank.

[0017] The integrated method and system for LNG storage tank knowledge modeling and intelligent optimization provided by this invention has the following advantages:

[0018] 1. Unified graph database eliminates data silos: Construct an ontology graph database with three levels of assembly-component-part, providing a common semantic foundation for finite element, IFC and STL models, fundamentally solving the problems of duplicate data entry and version conflicts across multiple systems.

[0019] 2. Adaptive generation of multidisciplinary models: Based on the target type specified by the user (finite element / IFC / STL), the model automatically completes the lossless conversion from graph data to a special model through node, relationship and attribute mapping rules, without manual intervention or intermediate format conversion.

[0020] 3. Incremental update triggers model linkage synchronization: Supports two incremental operations: structural adjustment (inserting / deleting components along the path) and attribute modification. It automatically identifies affected components and regenerates the corresponding model, avoiding full reconstruction and achieving efficient iteration of design changes and model synchronization.

[0021] 4. Online proxy model significantly reduces optimization costs: Deep neural networks are embedded in the main loop of genetic algorithms, and periodic online retraining is carried out using accumulated FEA samples, and the optimal hyperparameters are automatically searched; elite individuals are injected through virtual screening, reducing the number of high-fidelity analyses from hundreds to 50-80, and the computation time is compressed by 70%-80%, and the optimal steel consumption is reduced by 13.6% compared with traditional GA.

[0022] 5. Forming a closed loop of "modeling-generation-optimization-update": The ontology graph database provides a unified input for model generation and optimization. The optimization results can be written back to the database and automatically trigger the model to be regenerated, realizing the integrated collaborative operation of knowledge management, multi-disciplinary output and intelligent optimization, and providing a general paradigm for complex industrial facilities. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the structure of a liquefied natural gas storage tank provided in an embodiment of the present invention;

[0024] Figure 2 A flowchart illustrating the integrated method for LNG storage tank knowledge modeling and intelligent optimization provided in this embodiment of the invention.

[0025] Figure 3 A schematic diagram illustrating a progressive ontology development strategy for multi-task applications provided in an embodiment of the present invention;

[0026] Figure 4 This is a schematic diagram illustrating the evolution of the core body of a liquefied natural gas storage tank provided in an embodiment of the present invention;

[0027] Figure 5 A schematic diagram of a two-stage mapping mechanism from network ontology language to tag attribute graph provided in an embodiment of the present invention;

[0028] Figure 6 This is a schematic diagram illustrating the dynamic evolution of graph databases and multi-modal data extraction process provided in an embodiment of the present invention;

[0029] Figure 7 A flowchart of a finite element analysis model generation method based on a graph database provided in an embodiment of the present invention;

[0030] Figure 8 A flowchart of a method for generating an industrial basic engineering management model based on a graph database, provided in an embodiment of the present invention;

[0031] Figure 9 A flowchart of a lightweight STL visualization model generation method based on a graph database provided in an embodiment of the present invention;

[0032] Figure 10 This is an example diagram illustrating the application of target component query and 3D model mapping based on graph database provided in an embodiment of the present invention.

[0033] Figure 11 This is a schematic diagram of the overall process of the online structural optimization method for LNG storage tanks based on a deep neural network surrogate model according to the present invention.

[0034] Figure 12 This is a comparison of the convergence curves of the average population weight of GA_DNN and traditional GA during the optimization process in this embodiment of the invention.

[0035] Figure 13 This is a histogram comparison of the final population weight distribution at the termination of optimization between GA_DNN and traditional GA in this embodiment of the invention;

[0036] Figure 14 The curves showing the evolution of the optimal solution weight and key design variables with the number of iterations during the GA_DNN optimization process in this embodiment of the invention are shown.

[0037] Figure 15 The graphs show the mean squared error loss and the curves of the predicted optimal steel consumption in the design during the online training process of the deep neural network surrogate model in this embodiment of the invention, as a function of the number of iterations. Detailed Implementation

[0038] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0039] Example 1

[0040] This invention provides an integrated method for knowledge modeling and intelligent optimization of LNG storage tanks, the principle of which is as follows: Figure 2 As shown, the method as a whole includes the following steps:

[0041] S1: Construct an ontology graph database based on various design task information of liquefied natural gas storage tanks;

[0042] S2: Adaptive generation of finite element analysis models for liquefied natural gas storage tanks based on ontology graph database;

[0043] S3: Based on the ontology graph database and the finite element analysis model, online intelligent optimization of the LNG storage tank structure is performed using a deep neural network surrogate model and a genetic algorithm to obtain the optimal design parameters of the liquefied natural gas storage tank.

[0044] (1) Step S1: Construct the ontology graph database

[0045] Step S1 mainly includes:

[0046] Obtain various design task information for liquefied natural gas (LNG) storage tanks, and construct an ontology for the LNG storage tank domain based on the design task information. The ontology for the LNG storage tank domain includes a concept set, a relation set, and an attribute set. The concept set includes structural concepts at three levels: assembly, component, and part. The relation set includes compositional relations and contact relations. The attribute set includes location attributes, geometric attributes, and material attributes.

[0047] The ontology of liquefied natural gas (LNG) storage tanks is converted into architectural constraints for a graph database. Based on these architectural constraints, the engineering data of the LNG storage tanks is instantiated into the graph database to generate an ontology graph database (or graph model). Nodes in the ontology graph database represent the engineering entities of the LNG storage tanks, edges represent the semantic relationships between entities, and the attributes of nodes include location attributes, geometric attributes, and material attributes.

[0048] The various design task information includes at least structural analysis tasks (corresponding to finite element models, for mechanical analysis, etc.), building information modeling tasks (corresponding to IFC models, for multi-disciplinary clash checks, construction simulation, and asset management, etc.), and visualization tasks (corresponding to STL models, for visualization, design review, or virtual reality walkthroughs, etc.). Based on the acquired design task information, an ontology engineering approach is used to construct a domain ontology, establishing a multi-level conceptual hierarchy of "assembly-component-part" and its semantic relationships. A graph database is used for ontology instantiation and storage, supporting the generation, modification, querying, and retrieval of node labels, edge types, and attribute key-value pairs. This forms a unified digital foundation for the tank structure, supporting an automatic mapping workflow between the structural analysis model, IFC model, and geometric display model, thereby achieving consistent representation and collaborative generation of multi-source models.

[0049] Ontology is a core concept in knowledge engineering, defined as a "conceptualized and clearly defined specification." Simply put, an ontology is like a dictionary that defines what concepts exist within a domain (e.g., "inner tank," "wall panel"), what relationships exist between these concepts (e.g., "composition," "contact"), and what attributes each concept possesses (e.g., thickness, material, location). With ontology, humans and computers can understand and exchange domain knowledge in a consistent manner. In LNG storage tank design, different disciplines may use different terminology for the same thing; ontology eliminates ambiguity through clear definitions, enabling cross-disciplinary collaboration.

[0050] The integrated design of LNG storage tanks involves the collaborative management of multi-source data and various types of entities. For example, statistical data representing the geometry of LNG storage tanks (such as height, diameter, and curvature) can be adjusted during the design process, while structural analysis requires semantic data indicating material types (such as S30408 ​​steel and carbon steel). To uniformly describe this diverse information and the complex relationships between such data and entities, ontology engineering methods are used to formally model the structure of LNG storage tanks.

[0051] Each ontology can be represented as a triple. ,in , and These represent concepts, relationships, and attributes, respectively. Given the diversity of LNG storage tank design tasks, this invention adopts a multi-task-oriented incremental ontology development strategy, the principle of which is as follows: Figure 3 As shown. Reference Figure 3 First, the design scope is defined and an initial ontology is constructed based on multiple LNG tank design tasks. This initial ontology is built upon general engineering foundation classes from ISO 15926 or IFC standards, and then tailored and expanded in conjunction with LNG tank design specifications (such as EN 14620). Subsequently, new concepts are added to the initial ontology through keyword extraction and expansion mechanisms, and / or concepts from different ontologies are combined through a combination mechanism to obtain concept increments. The keyword extraction and expansion mechanism can be implemented using a predefined LNG tank domain dictionary (containing common terms and their synonyms, and English translations) combined with regular expression matching. For new terms not included in the dictionary, a string similarity algorithm (such as edit distance) is used for fuzzy matching with existing concept names, supplemented by expert review. The combination mechanism is used to integrate concepts from different task ontologies, aligning semantically similar concepts through equivalence relations, or mapping heterogeneous names to standard terms. For any given task... The relevant incremental knowledge is organized into , Given tasks The increments of concepts, relationships, and attributes. Then, the increments of concepts. Relationship increment and attribute increment Perform terminology alignment and consistency verification. Verification includes: ① Checking for circular inheritance (e.g., ...). yes subclasses, Again ① Subclasses of the relation; if any exist, report an error and suggest corrections; ② Check for attribute conflicts, i.e., whether the same attribute is assigned an incompatible data type; ③ Verify whether the domain and range of the relation conflict with the existing concept hierarchy; ④ Ensure that the transitivity of the constituent relations is not violated (e.g., ...). composition , composition Then it will be automatically deduced. composition For incremental knowledge that fails validation, the system generates a conflict report and refuses to incorporate it into the core ontology. Finally, validated, reusable, and stable semantics are incorporated into the core ontology library. ,in and Representing the first Wheel and First The composition mechanism combines increments from different tasks when multiple design tasks need to be supported simultaneously.

[0052] The core ontology, developed using Protege (an ontology editing tool), abstracts the LNG storage tank into a multi-level semantic network, containing concepts. ,relation and attributes Top-level structure concept set Strictly mapped to the engineering design granularity: assembly level ( ), component level ( ) and part level ( These levels of granularity support preliminary scheme decisions, detailed parameter specifications, and detailed construction drawing design, respectively. Each specific structural concept is categorized at these levels; for example, the InnerTank is defined as... The subclass InnerShellPlate is defined as Subclasses of. Relationship set. It contains two user-defined object properties: composition relation and contact relationship Compositional relationships Satisfying transitivity axioms to express physical inclusion and contact relationships across levels The symmetry axiom is satisfied to describe the adjacency relationships between components within a level. Together, they establish the topology of vertical decomposition and horizontal association. Finally, the attribute set... Encapsulated location ,geometry and materials The data provides a unified parameter basis for downstream models such as structural analysis and Building Information Modeling (BIM).

[0053] For example, the inner tank (Inner_Tank) as subclasses, in The level is decomposed into inner tank wall plates (Inner_Shell) and inner tank bottom plates (Inner_Bottom), which are further... The hierarchy is refined into inner tank wall plate units (Inner_ShellPlate) and inner tank bottom plate units (Inner_BottomPlate). Subsequently, the task ontology dynamically evolves when performing tank wall design tasks—especially when driven by layered wall thickness design and structural stability verification requirements. To support layered design, it utilizes compositional relationships... The Inner_ShellLayer concept (inner shell layer) is extended under the Inner_Shell node. Simultaneously, to address stability constraints, the InnerStiffener concept (inner stiffening rib) is directly added under the Inner_Tank concept, further... The level is refined to inner tank stiffening rib units. Once verified, these increments are integrated into the core ontology for semantic reuse, and the corresponding evolution results are as follows: Figure 4 As shown.

[0054] LNG storage tank design data exhibits a complex hierarchical structure, highly interconnected entities, and strict topological constraints. Graph databases can intuitively describe entities and their semantic relationships using nodes and edges. Therefore, this embodiment employs a graph database to instantiate and store the ontology model, thereby supporting model generation, updating, information management, and retrieval throughout the design process.

[0055] To facilitate the computationalization of ontology concepts, this embodiment utilizes a Web Ontology Language (OWL) file (hereinafter referred to as OWL file) exported by an ontology editor as a semantic bridge to map abstract domain knowledge to a Graph Database (LPG), resulting in an ontology graph database. The graph database is preferably a database system that supports attribute graph databases, such as Neo4j or JanusGraph. The LPG is formally defined as... In the ontology graph database, engineering entities and directed semantic edges This forms the topological skeleton, while the endpoint functions Tag allocation and property mapping This endows the graph with multidimensional features. The construction of an ontology graph database is achieved through two collaborative mapping phases: schema-level mapping and instance-level mapping, such as... Figure 5 As shown.

[0056] In pattern-level mapping, the concept set Each concept in the graph is mapped to a node label in the graph database, thus connecting the relation sets. Each relation in the graph is mapped to an edge type in the graph database, which maps attribute sets. Each attribute in the graph database is mapped to an attribute key-value pair. Specifically, the ontology concept... and object properties Node label sets were defined respectively. and edge type set The elements. At the same time, data attributes. Converted into an ordered set of specified attribute keys and their data types Furthermore, the domain and range constraints embedded in the OWL file are interpreted as edge endpoint constraint functions. This is to ensure the topological validity of the generated graph.

[0057] (1)

[0058] (2)

[0059] (3)

[0060] (4)

[0061] Formulas (1) and (2) define the set of node labels for the ontology graph database. and edge type set .in All concepts in the ontology Through mapping function The resulting labels constitute, and Then, all relations in the ontology The mapping is obtained. These two formulas together establish the foundation for the mapping from ontology to graph database schema: each concept corresponds to a type of node, and each relation corresponds to a type of edge, ensuring that the graph obtained by subsequent instantiation is semantically strictly consistent with the ontology. Formula (3) maps all attributes in the ontology. Convert to a collection of key-value pairs Each record contains the key name of the attribute. and data types This provides a type-safe definition for storing the attributes of graph nodes. Equation (4) defines an edge endpoint constraint function. It will be for each edge type Mapped to a pair of node labels This means that edges can only connect nodes with specific labels, thus ensuring the legality of the graph topology at the structural level.

[0062] After the schema-level mapping is completed, the system creates corresponding node labels, edge types, and attribute key definitions in the database based on the architectural constraints generated by formulas (1)-(4), and uses the constraint mechanism provided by the database to perform mandatory validation of the edge endpoint types to ensure that any subsequently inserted edge meets the domain and value range restrictions. When using Neo4j, the mapped labels, relation types, and attribute keys can be directly mapped to Label, Relationship Type, and Property Key in the database, thereby ensuring seamless implementation of ontology semantics.

[0063] In instance-level mapping, obtain the engineering data instance record set of the LNG storage tank. , set of engineering data instances Each data instance in the diagram is populated from bottom to top under architectural constraints, following the logic described in formulas (5)-(7). Engineering Data Instance Record Set Originating from structural analysis models (such as neutral files from ANSYS), BIM models (such as IFC files), or design parameter tables (such as wall thickness tables in Excel format), these data are parsed and uniformly converted into an intermediate representation in the form of triples: entity instances, relation assertions, and attribute assertions. Specifically, they originate from instance record sets. Each data instance Instantiated as a graph node Based on the class hierarchy defined by the subClassOf axiom, all parent class tags are inherited (e.g., instances of InnerTank have both InnerTank and Assembly tags), thus supporting queries at any design granularity. Simultaneously, a globally unique identifier is assigned to each project entity to ensure lifecycle traceability. Then, a set of relationship assertions... Driver endpoint function and tag assignment Generate oriented edges with type, and attribute value sets. Through property mapping Attach it to the corresponding node.

[0064] (5)

[0065] (6)

[0066] (7)

[0067] in, Represents a data instance The most specific class (i.e., direct type) to which it belongs in the ontology, and according to Axiom: This instance should inherit from all its parent class types. Therefore, the node... tag collection Not only includes It also includes the tags of each of its parent classes. For example, if an InnerTank instance's direct type is InnerTank and its parent class is Assembly, then the tag set for that node is... For attribute mapping functions This article defines it as the operation of persisting key-value pairs to node properties: Let... This indicates that the key is... Value The attributes are written into the node's attribute set, while This is the concrete expression of the operation, that is, the operation on the node. Assign the key name Value Attributes.

[0068] Formula (5) describes the instantiation process of a graph node: for the instance record set Each data instance in The system creates a graph node. And the label set of this node. It must contain the direct type of the instance. And all its parent types. This means that a node of an "inner tank" instance will have tags such as InnerTank and Assembly, thus supporting data querying at different design granularities. Formula (6) converts relation assertions into directed edges in a graph database. For each relation assertion The system is at the corresponding graph node and Generate a directed edge between them The direction of the side is from point to The label of the edge Take as relation name In this way, the relationships such as "composition" or "contact" defined in the ontology become traversable paths in the graph database, facilitating complex relationship queries. Formula (7) asserts attributes. Stored on graph nodes. The system uses attribute mapping functions. At the node Chinese key-value pairs The attribute value is recorded in a specific format. This allows the geometric parameters (such as thickness and diameter) and material properties (such as steel grade) of each engineering entity to be saved as node attributes, supporting subsequent read and modification operations.

[0069] Ultimately, the fusion of schema-level semantics and instance-level data resulted in a complete ontology graph database (graph database). Through the unified framework of OWL bridging, the object structure, semantic relationships, and parameter information in LNG storage tank design are efficiently organized, providing a strong digital foundation for subsequent complex queries and knowledge reuse.

[0070] LNG storage tank design is essentially a continuous, iterative optimization process. Faced with constantly changing operating conditions and specification constraints, engineers must frequently adjust geometric parameters, component schemes, and material selections. Since these adjustments directly alter the engineering entity and its semantic network, the graph database must possess the ability to dynamically evolve in sync with the design process. This requires specific modification requests. The following embodiment will use a graph database. The update behavior is formally defined as follows: According to the graph evolution model, these updates are divided into two basic operations, as shown in Equation (8):

[0071] (8)

[0072] Formula (8) defines the graph database in modification requests. The dynamic evolution behavior under this framework is divided into two types: structural adjustment and attribute update. Structural adjustment follows a specified path. The path, consisting of a series of alternating nodes and edges, is represented as follows: ,in , , As the source node, For the target node, This refers to the alternating nodes and edges between the source and target nodes. For insertion operations, the system first parses the new component type, target parent node, and expected location carried in the modification request. Then, it creates new nodes and the required relational edges (including compositional and contact relationships) at the corresponding locations on the path. The new nodes inherit the labels and attribute templates of their respective concepts, and the system automatically verifies the validity of edge endpoints based on ontology constraints. For removal operations, the system deletes nodes and their associated edges on the specified path. If a node has child components (connected through compositional relationships), the user must specify whether to cascade deletion or only remove the parent-child relationship. Regardless of whether it is insertion or removal, an integrity check is performed after the operation to ensure that the graph database still satisfies the edge endpoint constraint functions. And the axioms of relational transitivity / symmetry. For example, when a new ring of stiffening ribs (InnerStiffener) needs to be added under the Inner Shell, the user request includes the target path. The system creates an InnerStiffener node at the end of this path and establishes... The relationship is defined, and positional and geometric parameters are set simultaneously. Attribute updates directly modify the nodes. The internal parameters, whose new values ​​are derived from the attribute mapping. Return. This mechanism enables incremental updates to persistent graph instances, avoiding the need to rebuild the entire graph database with every design change.

[0073] To effectively enable diverse downstream application scenarios, graph databases must be able to both dynamically update and retrieve data. This embodiment defines three core graph extraction mechanisms, unified in the extraction function. As shown in formula (9):

[0074] (9)

[0075] The first method is entity extraction, which uses a predicate filtering function. Precisely isolate individual nodes that meet the specified query conditions, based on their labels. and attributes The decision is made. The second method is path extraction, which involves selecting from a specified set of edge types. Under constraints, this operation occurs at the source node. and target node Traverse between and extract from Three pairs The third method is subgraph extraction, which utilizes extraction functions to create a continuous sequence of relationships, thereby revealing the deep topological transmission logic between components. Isolate by specifying nodes The connected subgraph rooted at the specified root will ultimately output the local node set. Edge set And their role in property mapping On the restrictions The path extraction returns a continuous sequence of relationships without duplicate nodes or edges, indicating a simple path. If the query requires traversing a loop structure, subgraph extraction should be used to obtain the complete connected components. All extraction results are serialized into JSON format for use by downstream modules.

[0076] These two fundamental capabilities—dynamic evolution and multi-pattern extraction—work together to establish the complete read / write lifecycle of a graph database. The dynamic evolution and multi-pattern extraction process of a graph database can be referenced. Figure 6 .like Figure 6As shown, incremental updates are performed on the persistent graph instance (left), effectively avoiding the computational overhead of a complete reconstruction. Subsequently, this updated graph database seamlessly serves as the basis for the execution of the three extraction modes (right). Finally, all extracted topology and parameter data are uniformly serialized into a lightweight JSON (JavaScript Object Notation) format. This standardized data interface eliminates transmission redundancy and ensures seamless integration between the evolving graph database and downstream structural analysis computation modules.

[0077] (2) Step S2: Adaptive generation of LNG multidisciplinary model

[0078] This step adaptively generates the corresponding target professional model based on the user's different model generation requests, specifically including:

[0079] S21. Receive user input for model generation request: The user selects the target model type to be generated through the system interface. The target model type is identified as any one of finite element analysis model, industrial basic engineering management model or STL lightweight visualization model.

[0080] For example, when structural engineers need to perform seismic safety checks, they should select the "Finite Element Analysis Model" type; when building information modeling engineers need to perform clash checks or export bills of quantities, they should select the "Industrial Infrastructure Engineering Management Model" type; and when design teams need to conduct web-based design reviews or virtual reality walkthroughs, they should select the "STL Lightweight Visualization Model" type.

[0081] S22. Determine the corresponding set of mapping rules based on the target model type identifier.

[0082] The three target model types each have their own set of mapping rules, and each set of mapping rules contains at least node mapping rules, relation mapping rules, and attribute mapping rules.

[0083] As an example, for a finite element analysis model, the node mapping rule is: component type → element type; the relationship mapping rule is: composition relationship → assembly constraint; contact relationship → contact boundary; and the attribute mapping rule is: geometric attribute → mesh parameter; material attribute → material definition. For an industrial foundation-type engineering management model, the node mapping rule is: concept level → industrial foundation-type entity inheritance level; the relationship mapping rule is: composition relationship → industrial foundation-type aggregation relationship; and the attribute mapping rule is: geometric attribute → industrial foundation-type geometric representation; and material attribute → industrial foundation-type material association. For an STL lightweight visualization model, the node mapping rule is: component type → sweep parameterization scheme; the relationship mapping rule is: composition relationship → model assembly level; and the attribute mapping rule is: geometric attribute → section and path parameters.

[0084] S23. Extract a subset of graph data from the graph database that matches the set of mapping rules.

[0085] Based on the target model type identifier, determine the required data dimensions. Data dimensions must include at least one or more of the following: geometric dimension, material dimension, and topological dimension.

[0086] Different target model types have different requirements for data dimensions:

[0087] (1) The finite element analysis model requires geometric dimension (for mesh generation), material dimension (for assigning material parameters) and topological dimension (for setting boundary conditions and contact constraints).

[0088] (2) Industrial basic engineering management models require geometric dimensions (for geometric representation of industrial basics), material dimensions (for material association) and topological dimensions (for spatial hierarchy construction).

[0089] (3) The STL lightweight visualization model mainly requires geometric dimensions (for sweep mapping to generate mesh).

[0090] Based on the data dimensions, nodes and edges in the graph database are filtered to extract subsets of graph data that meet the specified conditions. For example, when generating a finite element model, all part-level nodes and their geometric and material properties, as well as the topological structures connected by compositional and contact relationship edges, are extracted.

[0091] S24. Based on the set of mapping rules, convert the subset of graph data into a model file in the target format and output it.

[0092] Depending on the target model type identifier, step S24 executes one of the following three conversion processes.

[0093] When the target model type is identified as a finite element analysis model, the conversion process is as follows: Figure 7 As shown, it includes:

[0094] S241A: Determine the element type. Based on the node labels of each node in the graph data subset, determine the component type (shell, solid, beam) of each engineering entity. According to the component type, assign the corresponding finite element type to each engineering entity according to the component-element type mapping table shown in Table 1.

[0095] Table 1 Mapping between component types and element types

[0096]

[0097] S242A: Generate geometric mesh. Extract the geometric attributes (section shape type, section size, path control point coordinates, spatial pose, etc.) of specific entities from the graph data subset, and generate a geometrically discretized mesh for each engineering entity according to the assigned element type. For beam-type components, generate a one-dimensional mesh along the path direction and assign section attributes; for solid-type components, use a three-dimensional solid mesh generation algorithm.

[0098] S243A: Assign material parameters, element types, and material types. Extract material parameters (Young's modulus, Poisson's ratio, yield strength, mass density, etc.), element types (shell elements, beam elements, and solid elements, etc.), and connection types (conformal topology connections shared by nodes, multi-point motion coupling with motion constraints, and nonlinear contact interfaces of contact pairs, etc.) of specific entities from the graph data subset, and assign corresponding material parameters, element types, and connection types to each engineering entity.

[0099] S244A: Set boundary constraints and extract load information. Extract edge relationship types from the graph data subset—for compositional edges, establish corresponding assembly constraints or bonded contacts, locking 6 degrees of freedom; for contact edges, establish corresponding contact pairs, setting friction coefficients and contact behaviors (such as frictional sliding contact or bonded coupling). Simultaneously, establish partial constraints based on boundary nodes in the graph data (such as tank bottom nodes). Determine relevant information for different loads, such as points, lines, and surfaces.

[0100] S245A: Assembly Output. The generated mesh, material parameters, element types, connection types, boundary constraints, and load information are combined into an analysis file, which is transmitted as a JSON file. This file is further parsed to generate a finite element model file in a local coordinate system (Cartesian coordinate system, cylindrical coordinate system, spherical coordinate system, etc.) (such as ANSYS's CDB format or ABAQUS's INP format), which can then be used for structural analysis (structural simulation, result extraction, fitness evaluation, etc.).

[0101] In addition, key simulation results (such as nodal displacements and element stresses) can be fed back to the graph database through an inverse mapping path to enrich the attribute information of the graph data subset.

[0102] When the target model type is identified as an industrial basic engineering management model, the conversion process is as follows: Figure 8 As shown, it includes:

[0103] S241B: Constructing a spatial structure hierarchy. Based on the node hierarchy topology in the graph data subset, construct a spatial structure hierarchy for industrial foundation standards. Graph nodes representing project information, tank sites, and LNG tanks are mapped to project entities, site entities, and tank entities, respectively, forming a top-down spatial hierarchy.

[0104] S242B: Mapping Entity Types. This maps assembly-level nodes in the project entity to assembly entities, component-level nodes to component entities, and part-level nodes to their corresponding part entities. Specifically, taking the dome system as an example: at the part level, the dome skin panel is mapped to a plate entity, and the grid shell beam component is mapped to a beam entity; at the component level, both the dome skin system and the grid shell are represented by building component proxy entities; at the assembly level, these proxy substructures are further aggregated into a dome assembly entity represented by an assembly entity.

[0105] S243B: Generate Geometric Representation. Extract the geometric attributes of each node from the graph data subset to generate an industrial-grade geometric representation. Rotationally swept solids are used for rotationally symmetric components (such as tank walls, including information on cross-sections, rotation axes, and angles); stretched solids are used for components with invariant cross-sections (such as ring beams, including cross-sections, stretching direction, and height); and boundary representation solids are used for domes or irregular curved surfaces (including information on vertices, faces, and closed shells). Cross-sectional parameters, dimensional attributes, and spatial pose are derived from node attributes in the graph database.

[0106] S244B: Associated Material Information. Material attributes of each node are extracted from the graph data subset, and material information is associated with the corresponding industrial base class objects through material association entities. Project-specific metadata is encapsulated in attribute set entity instances linked by attribute definition entities.

[0107] S245B: Generate Bill of Quantities. Extract the geometric parameters of each node from the subset of drawing data, automatically calculate and generate the bill of quantities. The bill of quantities maintains a consistent correspondence between industrial base component entities, multi-level data diagrams, and quantity information, including basic information such as component category, key geometric parameters (e.g., diameter, thickness, height), material grade, and unique identifiers. At the main structural level, the bill of quantities covers four assembly systems: inner tank, outer tank, insulation layer, and dome; within each assembly system, the bill of quantities is further decomposed to the component level.

[0108] S246B: Assembly Output. Assembles the generated spatial structure, solid objects, geometric representations, material relationships, and bill of quantities into an Industrial Foundation Class Model (IFC) file.

[0109] When the target model type is identified as an STL lightweight visualization model, the system performs the following process for each component in the graph data subset, and the principle of parameter sweep mapping to generate the surface mesh is as follows: Figure 9 As shown, it includes:

[0110] S241C: Parametric Decomposition. Geometric information of the target component is retrieved from a subset of graph data and reformulated as a decoupled representation of cross-sections, paths, and coordinate systems.

[0111] S242C: Sweep Mapping. The cross-sectional parameters are propagated along the path and mapped to three-dimensional space under the local frame, generating discretized surface points (sweep point set) of the component based on this mapping.

[0112] S243C: Mesh Triangulation. A quadrilateral mesh is constructed based on the adjacency relationships in the parametric mesh, and then each quadrilateral cell is divided into two triangular faces. The two resulting triangles share a diagonal, thus completely converting the quadrilateral mesh into the triangular mesh required by the STL format.

[0113] S244C: Normal vector calculation. Calculate the outward unit normal vector for each triangular facet:

[0114] (10)

[0115] in, It is the first The outward unit normal vector of a facet , , These are three vertices arranged in a counter-clockwise order.

[0116] S245C: STL encoding. Each triangle facet is represented as a quadruple in ASCII or binary format. Write to an STL file.

[0117] The system executes the above process on each part-level node in the graph data subset to generate independent STL files for each component, or merges them into a single overall STL file based on the composition relationship edges.

[0118] S25. Synchronous update.

[0119] In response to incremental updates to the graph database, the system identifies the subset of graph data affected by the incremental update. The identification logic is based on the modification records in the graph database: if the attribute of a node is modified (e.g., the wall panel thickness is changed from 12 mm to 10 mm), then the component corresponding to that node and all its parent nodes are identified as the affected subset of graph data; if the composition relationship of a node changes (adding or deleting child nodes), then the affected parent nodes and all their child nodes are identified as the affected subset of graph data.

[0120] Based on the set of mapping rules, the corresponding model files are automatically regenerated. Since the update scope is limited to the affected subset of graph data, the computational overhead of full reconstruction is avoided.

[0121] (3) Step S3: Adaptive agent-assisted optimization

[0122] like Figure 10 As shown in the flowchart, step S3 specifically includes:

[0123] S31. Modeling of the optimization problem: Establish an optimization problem model for the LNG storage tank. The optimization problem model includes overall design variables, an objective function aimed at minimizing the total steel consumption, constraints on the structural response not exceeding the allowable value, and the boundary range of each design variable.

[0124] S32. Construction of a deep neural network surrogate model: Based on the optimization problem model, a deep neural network surrogate model is constructed, which includes the structural configuration of the input layer, hidden layer and output layer and the setting of training hyperparameters;

[0125] S33. Baseline optimization driven by genetic algorithm: Based on the optimization problem model, perform baseline optimization driven by genetic algorithm to generate an initial population containing a preset number of individuals and iteratively evolve. For each individual in the population, call the finite element analysis model generated in step S2 to obtain the real structural response, and store the qualified samples into the sample library.

[0126] S34. Training and updating of the surrogate model: During the iteration of the genetic algorithm, the sample library accumulated in step S33 is used as the training set to dynamically train and update the deep neural network surrogate model constructed in step S32 online. At least during the first training, the optimal hyperparameter configuration is automatically searched and used for subsequent training.

[0127] S35. Output of optimal design parameters: After each update of the deep neural network surrogate model in step S34, the current surrogate model is used to perform virtual response prediction on the newly generated candidate individuals after the genetic operation. The individuals with the best prediction are selected and injected into the current population. The injected population is reordered according to the true fitness and a preset number of individuals are retained to enter the next generation, forming a new generation population. The process returns to step S33 to continue iterating until the convergence condition is met, at which point the optimal design parameters are output. The optimization results are written back to the ontology graph database in step S1 through an incremental update mechanism, triggering the regeneration of the corresponding model file in step S2.

[0128] Step S31 addresses the issues of mixed design variable types and variable dimensions through a unified mixed-integer encoding strategy, transforming the constrained optimization problem into an unconstrained form suitable for iterative solution, thus providing standardized mathematical input for subsequent steps. Step S32 constructs a multi-layer deep neural network, which, due to its powerful nonlinear approximation capability, accurately fits the highly nonlinear mechanical response relationship in LNG storage tanks, laying the foundation for subsequent rapid prediction. Step S33, while performing baseline optimization using a genetic algorithm, continuously accumulates finite element analysis samples, overcoming the deficiency of insufficient sample space coverage in offline methods, and providing dynamic support for the online training of the surrogate model. The real data source for state expansion; step S34 achieves dynamic retraining during the optimization process through a preset algebraic triggering mechanism, enabling the accuracy of the surrogate model to continuously improve with the accumulation of samples, and ensuring performance consistency between different training batches through automatic hyperparameter search, completely overcoming the fundamental limitation that offline surrogate models cannot improve themselves; step S35 uses the current best surrogate model to perform millisecond-level virtual screening on a large number of candidate individuals after genetic operations, selecting predicted elites to inject into the population, guiding the genetic algorithm to concentrate on searching high-potential regions with extremely low computational cost, effectively alleviating the premature convergence problem caused by the lack of directional guidance in standard genetic algorithms. The five steps are interconnected, ultimately enabling the optimization process to be completed within an acceptable computational cost, and significantly improving the quality of the population solution set, achieving the dual optimization goals of computational cost and design quality while ensuring structural security.

[0129] The present invention will now be described in further detail with reference to the accompanying drawings.

[0130] Step S31: Modeling the Optimization Problem

[0131] This invention relates to an above-ground double-walled, fully containment metal LNG storage tank, whose main superstructure includes a steel inner tank, a steel outer tank, and a dome. Design variables are based on... The organization consists of three structural assemblies, and the overall design vector is defined as follows:

[0132] (11)

[0133] Among them, superscript , and These correspond to the inner tank, outer tank, and dome, respectively. These are the inner tank design vector, outer tank design vector, and dome design vector, respectively, which are further... and Level expansion aligns the optimization variables with the predefined three-level design granularity in the ontology.

[0134] For inner tank and outer tank ( Both are steel shell systems and share the same parametric scheme. Their design vectors are written as follows:

[0135] (12)

[0136] in, The inner diameter of the storage tank (unit: m). The height of the cylindrical shell (unit: m). For the first Shell plate thickness (unit: mm) (Numbered from bottom to top) The thickness of the base plate (unit: mm). The number of stiffening components (integer). For the first A unified parameter vector for each stiffening component. The reason for dividing the wall panel into 10 layers is that the dynamic water pressure at the bottom of the tank wall is the greatest under seismic action, gradually decreasing along the height direction. Adopting a variable thickness design can significantly reduce the amount of steel used while meeting strength requirements.

[0137] Each stiffening member is uniformly encoded using a fixed-length vector to accommodate mixed optimization of different cross-sectional types:

[0138] (13)

[0139] in, These are discrete section type indicators, for example, 0 indicates unstiffened, 1 indicates L-shaped steel, 2 indicates I-shaped steel, and 3 indicates channel steel; to For the corresponding geometric or layout parameters, for L-shaped steel, the first two parameters represent the leg length and leg thickness, respectively, and the rest are filled with 0; for I-shaped steel, the first four parameters represent the section height, flange width, web thickness, and flange thickness, respectively, and the rest are filled with 0. Unused entries are uniformly set to zero to ensure that the input dimension is fixed, so that the deep neural network can process discrete and continuous variables simultaneously.

[0140] The dome design vector is defined as follows:

[0141] (14)

[0142] in, The radius of curvature of the dome (unit: m). Skin thickness (unit: mm). The numbers (integers) represent the number of the central ring rod, circumferential members, and radial members, respectively. A unified cross-sectional parameter vector is provided for the corresponding (i.e., the central ring rod, circumferential rods, and radial rods) (formatted as in equation (13)). This expression incorporates the continuous geometric parameters of the dome and the number of discrete rods into the optimization variables, taking into account the collaborative design of the skin and the shell structure.

[0143] The optimization objective is to minimize the total steel consumption of the storage tank system. :

[0144] (15)

[0145] In the formula, This represents the total number of structural components. For the first The material density of each structural component (e.g., S30408 ​​stainless steel) ), For the overall design vector The determined volume of the components. This objective is directly related to the project's economics, as steel costs account for the vast majority of the main tank's construction cost.

[0146] For the seismic conditions considered in this invention, the structural safety constraint adopts a normalized response ratio function. express:

[0147] (16)

[0148] in, For the first The structural response at key locations, including von Mises equivalent stress, circumferential stress, and axial stress; The corresponding allowable value shall be determined in accordance with standard GB50341; To constrain the total number. When A condition is met if the requirement is met, otherwise it indicates a violation of the constraint. All constraints are standardized into a dimensionally normalized form for easy handling during optimization.

[0149] To handle constraints, the external penalty function method is adopted, and the penalized objective function is defined as follows:

[0150] (17)

[0151] in, This is a penalty factor, which is gradually increased during the iteration process (initially set to...). (Multiply by 1.5 for every 10 generations). Let be the positive part function (maximum function) of the constraint violation. This objective function only penalizes violated constraints, and the penalty is proportional to the square of the violation severity to ensure the smoothness of the penalty function. When all constraints are satisfied, the penalty term is zero, and the objective function equals the original steel usage; when a violation exists, the penalty term forces the iterative solution to move towards the feasible region.

[0152] Overall design variables Boundary constraints must also be met:

[0153] (18)

[0154] in, and These are the overall design variables. The lower and upper bound vectors are determined by design specifications and engineering experience. For example, the typical range for diameter is 20–40 m, shell thickness is 6–40 mm, and the number of stiffening members is 0–60. These boundary constraints ensure the engineering rationality of the design results while reducing the search space.

[0155] Since the design vector simultaneously includes continuous size variables, integer quantity variables, and discrete cross-section type variables, the optimization problem of this invention is formulated as a mixed integer nonlinear programming problem, requiring a global optimization algorithm capable of handling mixed variables.

[0156] II. Step S32: Construction of the Deep Neural Network Agent Model

[0157] This invention employs a deep neural network as a surrogate prediction model for structural response, used to approximate the mapping relationship from design variables to 10 key structural response indicators.

[0158] The 10 key structural response indicators include: von Mises equivalent stress of the dome skin plate, von Mises equivalent stress of the inner tank wall plate, von Mises equivalent stress of the outer tank wall plate, circumferential stress of the inner tank wall plate, circumferential stress of the outer tank wall plate, axial stress of the inner tank wall plate, axial stress of the outer tank wall plate, axial stress of the dome grid shell members, axial stress of the inner tank stiffeners, and axial stress of the outer tank stiffeners. These indicators cover the most critical strength verification items of the storage tank under seismic loading and can comprehensively reflect the structural safety.

[0159] Deep neural network proxy models include:

[0160] 1) Input Layer: The input dimension equals the total number of design variables. Since the number of stiffening components is variable, the actual length of the design vector is not fixed. This invention adopts a fixed-length padding strategy, uniformly padding all possible design variables to 80 dimensions (this value is greater than the maximum number of variables in any feasible solution), with unused dimensions filled with 0. This approach enables the neural network to handle tank designs of different scales while maintaining the stability of the network structure.

[0161] 2) Hidden Layers: A multi-layer fully connected feedforward network is employed. The optimal configuration, determined through optimization, consists of 5 hidden layers with 100, 150, 200, 300, and 200 neurons in each layer. The Rectified Linear Unit (ReLU) activation function is used for the hidden layers. The ReLU function is computationally simple and alleviates gradient vanishing, making it suitable for deep networks. Batch normalization is introduced after each layer to accelerate training and stabilize convergence. Dropout regularization (with a ratio of 0.2) is used to randomly drop some neurons to prevent overfitting.

[0162] 3) Output Layer: The output layer has a dimension of 10, corresponding to the 10 structural response indices mentioned above. The output layer uses a linear activation function and does not apply any nonlinear transformation to ensure that the predicted values ​​can cover stress values ​​within any real number range.

[0163] 4) Loss Function and Optimizer: The training loss function uses Mean Squared Error (MSE), which is defined as:

[0164] (19)

[0165] in, This represents the batch sample size. For the first The first sample The true FEA value of each response metric, These are the corresponding neural network predictions, and all the neural network predictions constitute the output vector. The optimizer chosen is Adam (Adaptive Moment Estimation), with a learning rate of 0.001 and an exponential decay rate estimated by the first moment. The exponential decay rate estimated by the second moment The Adam optimizer can adaptively adjust the learning rate of each parameter and exhibits good convergence performance when dealing with high-dimensional non-convex optimization problems.

[0166] 5) Regularization Strategy: In addition to Dropout, this invention also employs L2 weight regularization, adding a penalty term for the sum of squared weights to the loss function, with the penalty coefficient (weight decay) set to... This strategy effectively controls the magnitude of network weights, further suppressing overfitting.

[0167] 6) Number of training rounds and early stopping: The maximum number of training rounds is set to 200. At the same time, an early stopping strategy is adopted: when the validation set loss does not decrease in 20 consecutive training rounds, training is terminated early and the optimal model weights are restored to avoid overfitting and invalid calculations.

[0168] 7) Data partitioning: During each training session, the accumulated FEA samples are randomly divided into training and validation sets at a ratio of 80% and 20% respectively, for model training and performance monitoring.

[0169] III. Step S33: Genetic Algorithm-Driven Benchmark Optimization

[0170] This invention uses a standard genetic algorithm as the global search framework, and its basic process includes population initialization, fitness evaluation, selection, crossover, and mutation.

[0171] 1) Population Initialization: An initial population is generated using a Latin hypercube sampling strategy to ensure uniform coverage of the design space in the initial stage. The population size is set to 50. Each individual represents a complete LNG storage tank design scheme, encoded as a hybrid vector, where continuous variables are directly encoded as real numbers, integer variables are encoded as integers, and discrete section types are encoded as integer indices.

[0172] 2) Fitness Assessment: The true fitness of each individual is calculated with high fidelity using the finite element analysis model obtained in step S2. The assessment includes stress calculation (outputting the von Mises equivalent stress, circumferential stress, and axial stress of each component). To accelerate the assessment, multi-core parallel computing is used to run 4-6 FEA tasks simultaneously.

[0173] 3) Online data cleaning: The mechanical response data is cleaned online immediately after evaluation, including: removing outlier samples that do not converge due to mesh distortion; detecting and removing duplicate or highly similar samples (Euclidean distance less than 1). A consistency check is performed to ensure that the response difference of the same design on different computing cores is within 1%. The cleaned, qualified samples are stored in a sample library for subsequent training of the surrogate model.

[0174] 4) Genetic Operations: A tournament selection strategy (tournament size 2) is adopted, with a crossover probability of 0.8 and a mutation probability of 0.02. For continuous variables, simulated binary crossover and polynomial mutation are used; for integer and discrete variables, uniform crossover and random reset mutation are used. An elite retention strategy ensures that the best individual in each generation directly enters the next generation, preventing the loss of superior genes. Early termination is triggered when the change in optimal fitness is less than 0.1% over 10 consecutive generations, ending the optimization process.

[0175] IV. Step S34: Training and updating the surrogate model

[0176] The core feature that distinguishes this invention from traditional offline agent models lies in the online learning mechanism of the agent model, which specifically includes three stages: training triggering, incremental update, and automatic hyperparameter search.

[0177] 1) Training Trigger Mechanism: The surrogate model is not trained from the beginning, but rather training is initiated for the first time when the genetic algorithm iterates to the 10th generation. At this point, FEA samples from the previous 10 generations have been accumulated, up to a maximum of 50 × 10 = 500 samples, which is sufficient to support the initial model training. Thereafter, training begins every 3 generations, i.e., the 10th generation... generation( The surrogate model is retrained using all currently accumulated samples. This delayed start strategy is based on the following considerations: In the early stages, the GA population is relatively dispersed, with rich sample diversity but insufficient quantity, and early training can easily lead to overfitting; as iterations progress, the sample size gradually increases, and the model accuracy continues to improve.

[0178] 2) Incremental learning strategy: Each time retraining is performed, the entire historical data (rather than just new data) is used for training to avoid catastrophic forgetting problems. At the same time, as the sample size increases (up to about 800-1000 samples), the training time increases slightly, but due to the moderate network size and the use of multi-core parallel acceleration, it is much less than the time required for the same number of FEA evaluations (about several hundred hours), so it is highly economical from the perspective of overall computational cost.

[0179] 3) Automatic Hyperparameter Search: To avoid relying on manual parameter tuning based on human experience, this invention introduces a secondary genetic algorithm to automatically optimize the hyperparameters of the surrogate model. The hyperparameter space explored by the secondary GA includes: number of hidden layers (2-7 layers), number of neurons in each layer (10-500, selected from the candidate set), activation function (sigmoid, ReLU, tanh, LeakyReLU, ELU), optimizer (SGD, Adam, Adagrad, Adadelta, RMSprop), learning rate (0.1, 0.01, 0.001), training set partition ratio (0.6, 0.7, 0.8), and batch size (0.05, 0.1, 0.2). The population size of the secondary GA is 20, the maximum number of iterations is 20, the crossover probability is 0.8, and the mutation probability is 0.05. The fitness function is the average coefficient of determination of 5-fold cross-validation. This metric measures how well the model interprets the validation set data. After the search is complete, the highest-performing model is selected. The combination of hyperparameters is used as the optimal configuration for the current model and is fixed in subsequent retraining (without changing with iterations to maintain consistency).

[0180] 4) Performance evaluation index: This invention uses the coefficient of determination. The mean squared error (MSE) and mean squared error (MSE) are used as evaluation metrics for surrogate models, and their definitions are as follows:

[0181] (20)

[0182] ,(twenty one)

[0183] in, The true value calculated for FEA, For the surrogate model's predicted values, The mean of the true values. To verify the sample size. The closer to 1, the higher the proportion of variance explained by the model and the better the prediction performance; MSE reflects the absolute magnitude of the prediction error.

[0184] V. Step S35: Output of optimal design parameters

[0185] This step is crucial for accelerating online optimization. By integrating the predictive capabilities of the surrogate model into the evolutionary process of the genetic algorithm, it effectively guides the search direction.

[0186] 1) Virtual Screening: After each retraining of the surrogate model (i.e., every 3 generations), the latest DNN surrogate model is used to perform rapid response prediction on all newly generated candidate individuals (i.e., offspring individuals that have undergone crossover and mutation but have not yet undergone FEA evaluation). For each candidate individual... The proxy model outputs predicted values ​​for 10 response metrics. Then calculate the predicted constraint violation amount. and the penalized objective function value This process is entirely based on matrix operations of neural networks, and can complete the evaluation of all 50 offspring individuals in milliseconds, saving several hours of computation time compared to FEA evaluation.

[0187] 2) Elite Injection: Selecting the top-predicted candidates from the virtual screening results. Individuals (in this method) These 5 individuals are then merged with the 50 real individuals evaluated by FEA in the current generation, forming a merged population of 55. The merged population is then reordered according to true fitness (i.e., the penalty objective function value verified by FEA), with the top 50 individuals retained for the next generation. This "prediction-injection-selection" mechanism ensures that the surrogate model's predictions actually influence the direction of evolution, rather than being ignored as mere reference information.

[0188] 3) High-fidelity verification: Although the injected elite individuals are selected through a surrogate model, their true fitness still needs to be evaluated through a complete FEA in the next generation. The surrogate model is only used in the "pre-screening" stage to eliminate obviously poor candidate solutions, thereby concentrating valuable FEA computational resources on the most promising regions. The final optimal solution is still determined based on the FEA verification results, ensuring the reliability of the optimization results.

[0189] 4) Convergence Acceleration Mechanism: Traditional genetic algorithms are prone to getting stuck in local optima during the mid-cycle due to loss of population diversity, causing the objective function to stagnate for a long time. This invention, through the periodic intervention of a surrogate model, continuously injects predicted high-quality individuals into the population, which is equivalent to introducing external knowledge and new search directions into the population. These injected individuals carry design parameter combinations that differ from the current population's average characteristics, helping the population escape the local optimum trap, resulting in a significant stepwise decrease in the objective function after each surrogate intervention. At the same time, the surrogate model itself is continuously refined with the accumulation of samples, and its prediction reliability gradually improves, thus continuously providing valuable guidance information to the population even in the later stages of iteration.

[0190] The experiment selected a double-walled steel full-containment LNG storage tank as an engineering case study, with an effective storage volume of approximately The seismic fortification intensity is 8 degrees, and the basic design seismic acceleration is... The design seismic group is Group 2, and the site category is Class II. The main operating conditions and design basis are summarized in Table 2.

[0191] Table 2 Key Design Information for LNG Storage Tanks

[0192]

[0193] Based on the established ontology model and mapping mechanism, the liquefied natural gas storage tank case is instantiated into a graph database through entity extraction and relation recognition. For example... Figure 11 As shown, the instantiated graph database contains 2330 entity nodes and 16791 relations. In terms of node hierarchy, it includes 1 top-level root node, 5 assembly-level nodes, 126 component-level nodes, and 2198 part-level nodes. At the relational level, two types of relations are defined: composition relations, consisting of 2329 links, represent containment relationships between different levels; contact relations, consisting of 14462 links, describe the physical adjacency relationships between entities. Through the hierarchical organization of nodes and the semantic definition of relations, the graph database provides an explicit representation of the structural composition and local topological connectivity of liquefied natural gas storage tanks.

[0194] Figure 11The application verification process based on a graph database is further demonstrated. First, a target node (e.g., an inner tank wall panel) is selected from the graph database, and its local relational subgraph is retrieved to obtain the corresponding hierarchical relationships, adjacency relationships, and associated attribute information. Then, the retrieved node data is parsed into a structured representation containing node categories, relational semantics, and engineering attributes (corresponding to node labels, variant types, and attribute key-value pairs in the graph database, respectively). For example, a wall panel node can be associated with geometric attributes of a thickness of 9 mm and a height of 2500 mm, as well as the material properties of steel grade S30408. Finally, the query results are mapped to a 3D model to achieve step-by-step location and visualization of the target component and its corresponding structural hierarchy. The node types, relational semantics, and attribute patterns in the graph database maintain strict structural consistency with the ontology definition, which verifies the effective coverage and instantiation operability of the proposed ontology model.

[0195] To systematically evaluate the predictive performance of four machine learning algorithms on the structural response of LNG storage tanks, the coefficient of determination (COP) was used. The mean squared error (MSE) and the mean squared error were used as evaluation metrics. The model that showed the strongest generalization ability was then selected as the surrogate model for the optimization process. The overall performance of each algorithm on a dataset containing 500 samples, as well as the hyperparameters determined by the optimization algorithm, are summarized in Table 3.

[0196] Table 3 Performance comparison and optimal hyperparameter settings for four machine learning algorithms (dataset size: 500)

[0197]

[0198] The results in Table 3 show that all four algorithms achieved high fitting accuracy on the training set. Specifically, the training of KNN... The scores reached 1.00, SVR reached 0.98, and RF and DNN both reached 0.95, demonstrating their strong fitting capabilities. However, their predictive performance on the test set varied significantly. Among the four models, DNN achieved the highest score. The score is 0.90, which is 7.1%, 9.8%, and 20.0% higher than SVR (0.84), RF (0.82), and KNN (0.75), respectively, demonstrating the superiority of our method using DNN as a surrogate model.

[0199] To verify the effectiveness of the proposed surrogate-assisted method, GA_DNN (genetic algorithm combined with deep neural network surrogate) was compared with the traditional genetic algorithm (GA), and the comparison results are shown in Table 4. All optimized designs satisfied the stress constraints, the component stress remained below the material yield strength, and the stress ratio of more than 90% of the components was below 0.8.

[0200] Table 4 Comparison of Optimization Results of GA and GA_DNN

[0201]

[0202] As shown in Table 4, in terms of structural weight, GA_DNN achieves an optimal steel consumption of 911.92 tons, a 13.6% reduction compared to the optimal value obtained by traditional GA (1055.35 tons). The average steel consumption of GA_DNN across all operations is 933.92 tons, a 19.31% improvement over the GA baseline (1157 tons). These results confirm the superior optimization capabilities of the surrogate-assisted framework in both best-case and average-case performance.

[0203] Figure 12 The population averages for GA and GA_DNN during training are presented, showing a rapid decrease in the objective function during the first 10 generations. However, their trajectories diverge significantly after this initial phase. Traditional GA enters a near-stagnant state with little further improvement—a typical manifestation of premature convergence due to local optima. In contrast, GA_DNN activates a high-fidelity DNN agent every three generations starting from the 10th generation, performing large-scale virtual screening of the current population and injecting predicted elite solutions into the evolutionary process. Each agent intervention results in a significant stepwise decrease in the objective function. From the 10th to the 16th generation, the average cumulative reduction relative to the plateau reaches approximately 10%, allowing the optimization process to continue making significant progress even after traditional GA has effectively ceased improving.

[0204] Figure 13 The population distribution before and after a representative run of the two methods was further compared. The final population generated by GA was widely distributed, with individual solutions ranging from 1000 to 1300 tons, and a population mean of approximately 1157.41 tons. This indicates that GA has poor convergence and a significant proportion of individuals are located in high-weight regions, suggesting a lack of guidance to escape local optima. In contrast, the final population generated by GA_DNN was highly concentrated, with most individuals clustered between 900 and 950 tons, and a population mean of approximately 923 tons. The concentration of individuals near the superior target region confirms that the DNN agent provides a more informative fitness landscape, guiding the population to move towards promising regions at an earlier stage and achieving deeper convergence towards the global optimum.

[0205] Figure 14The evolution of weight and key design variables associated with the optimal solution throughout the GA_DNN optimization process is described. In the early stages, optimization primarily adjusts macroscopic structural parameters, resulting in a sharp decrease in weight. For example, the height-to-diameter ratio increases from 0.45 in the first generation to 0.69 in the 20th generation, subsequently stabilizing. This indicates that overall geometric proportions are optimized first to achieve a more reasonable global stress distribution. In the later stages, with weight remaining almost constant, the optimization focus shifts to fine-tuning the dimensions of local components, such as reducing the cross-sectional thickness of the first ring plate of the inner tank from 26 mm in the 20th generation to 21 mm in the 30th generation. This evolution reveals that the optimization follows a clear hierarchical path—from global layout adjustments to local structural refinements—reflecting the gradual improvement in solution quality during the iterative search process.

[0206] Figure 15 This demonstrates the performance evolution of the surrogate model during optimization. After initial training, the DNN model had a training MSE of 0.02 and a test MSE of 0.03. With the continuous injection of new high-fidelity samples and dynamic model updates, the prediction accuracy gradually improved. By the [missing information - likely a specific point or stage], the prediction accuracy had reached [missing information - likely a specific level or stage]. During the next iteration ( (For retraining cycles), the training and testing MSE decreased to 0.003 and 0.002, respectively, while The value reached 0.98. These results indicate that the surrogate model becomes increasingly accurate as optimization progresses. Meanwhile, each The predicted optimal design obtained in each iteration is gradually refined: the predicted optimal steel consumption is improved from the first iteration. The amount decreased from 1114.7 tons in the second iteration to the third. The weight of 875.3 tons in the next iteration corresponds to a reduction of approximately 10.6%. This indicates that the adaptive improvement in agent accuracy directly contributes to a more effective global optimization search.

[0207] In summary, this invention addresses the technical challenges in LNG storage tank design, including the difficulty in integrating multi-source heterogeneous data, the lack of automatic generation paths for multidisciplinary models, high computational costs for structural optimization, and the absence of a closed-loop linkage between knowledge modeling and optimization. It proposes an integrated method for LNG storage tank knowledge modeling and intelligent optimization. Its core contributions and beneficial effects are reflected in the following five aspects:

[0208] First, a multi-level domain ontology covering the entire lifecycle of LNG storage tanks was established. Following a three-level granularity of "assembly-component-part", it defined two core semantic associations: compositional relationships and contact relationships, and encapsulated three key attributes: location, geometry, and material. This provides a unified and unambiguous knowledge dictionary for different disciplines such as structural analysis, building information modeling, and visualization, fundamentally eliminating cross-disciplinary terminology conflicts and data silos.

[0209] Secondly, it implements a lossless mapping and dynamic evolution mechanism from ontology to graph database. Through a two-stage mapping process at the schema and instance levels, the OWL ontology is accurately converted into a label-attribute graph, ensuring semantic consistency and computational operability. It also supports two incremental update methods: structural adjustments (adding or deleting components along a path) and attribute modifications, avoiding full reconstruction, and includes supporting entity, path, and sub-database mapping mechanisms. Figure 3 This provides a variety of data extraction modes, offering flexible and efficient data interfaces for downstream applications.

[0210] Third, it pioneered a multidisciplinary model adaptive generation framework based on a unified graph database. According to the user-specified target type (finite element analysis model, IFC engineering management model, or STL visualization model), the system automatically matches mapping rules, extracts a subset of graph data, and converts it into a dedicated model file, all without manual intervention or intermediate format conversion. When design data changes, only an incremental update to the graph database is needed to trigger the regeneration of all related models, completely solving the problems of repetitive data entry and version conflicts in traditional multi-system systems.

[0211] Fourth, an online intelligent optimization strategy integrating a deep neural network surrogate model and a genetic algorithm is proposed. High-fidelity finite element analysis is embedded into the main loop of the genetic algorithm. The real samples accumulated during the iteration process are used to periodically retrain the DNN surrogate model online and automatically search for the optimal hyperparameters. The surrogate model performs millisecond-level virtual screening of candidate individuals, guiding the search to concentrate on high-potential areas. The number of expensive high-fidelity evaluations is reduced from hundreds to 50-80, reducing the computation time by 70%-80%, and the optimal steel consumption is reduced by 13.6% compared to the traditional genetic algorithm.

[0212] Fifth, a closed-loop collaborative system of "knowledge modeling—model generation—intelligent optimization—incremental update" has been constructed. The ontology graph database serves as a unified digital foundation, providing source data for multidisciplinary model generation and structural optimization. The optimal design parameters obtained from optimization can be automatically written back to the graph database, triggering the linked regeneration of model files such as finite element, IFC, and STL. This achieves seamless connection and automated operation throughout the entire design iteration process, providing a reusable and universal paradigm for the design collaboration and intelligent upgrading of large and complex industrial facilities.

[0213] Example 2

[0214] Corresponding to the integrated method for LNG storage tank knowledge modeling and intelligent optimization provided in Embodiment 1 above, this embodiment of the invention also provides an integrated system for LNG storage tank knowledge modeling and intelligent optimization. This system includes:

[0215] The graph database construction module is used to construct an ontology graph database based on various design task information of liquefied natural gas storage tanks; the model generation module is used to adaptively generate multidisciplinary models of liquefied natural gas storage tanks based on the ontology graph database; the surrogate-assisted optimization module is used to perform online intelligent optimization of the LNG storage tank structure based on the ontology graph database and the finite element analysis model in the multidisciplinary model, and based on the deep neural network surrogate model and genetic algorithm, to obtain the optimal design parameters of the liquefied natural gas storage tank.

[0216] The detailed workflow of each module corresponds to steps S1, S2, and S3 in the above method embodiments, and will not be described in detail here.

[0217] The system can be deployed on general computing devices (servers or workstations), or it can be packaged as standalone software or integrated into mainstream engineering design platforms such as ANSYS, ABAQUS, and Revit. It can seamlessly connect with downstream applications through standardized data interfaces (JSON, CDB, IFC, STL).

[0218] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An integrated method of LNG storage tank knowledge modeling and intelligent optimization, characterized in that, Including the following steps: S1. Construct an ontology graph database based on various design task information of liquefied natural gas storage tanks, including: Obtain the design task information of the LNG storage tank, and construct the LNG storage tank domain ontology based on the design task information. The LNG storage tank domain ontology includes a concept set, a relationship set, and an attribute set. The concept set includes structural concepts at three levels: assembly level, component level, and part level. The relationship set includes compositional relationships and contact relationships. The attribute set includes location attributes, geometric attributes, and material attributes. The LNG storage tank ontology is converted into the architectural constraints of a graph database, and the engineering data of the LNG storage tank is instantiated into the graph database according to the architectural constraints to generate an ontology graph database. S2. Based on the ontology graph database, an adaptive multidisciplinary model of the liquefied natural gas storage tank is generated, wherein the multidisciplinary model includes at least a finite element analysis model; S3. Based on the ontology graph database and the finite element analysis model, online intelligent optimization of the LNG storage tank structure is performed using a deep neural network surrogate model and a genetic algorithm to obtain the optimal design parameters of the liquefied natural gas storage tank.

2. The LNG storage tank knowledge modeling and intelligent optimization integrated method according to claim 1, characterized in that, The construction of the LNG storage tank ontology based on the design task information includes: The design scope is defined jointly by multiple LNG storage tank design tasks, and an initial ontology is constructed based on the design scope. For any design task, the incremental knowledge corresponding to the design task is organized into concept increments, relation increments, and attribute increments; The concept increment, relation increment, and attribute increment are performed with terminology alignment and consistency verification, and the verified incremental knowledge is incorporated into the core ontology library. The process of organizing the incremental knowledge corresponding to the design task into concept increments, relation increments, and attribute increments includes: New concepts are added to the initial ontology through keyword extraction and expansion mechanisms, and / or concepts from different ontologs are combined through a combination mechanism to obtain the concept increment; The composition relation is used to express the physical inclusion relationship between concepts at different levels, and the composition relation satisfies the transitivity axiom; the contact relation is used to express the adjacency relationship between concepts at the same level, and the contact relation satisfies the symmetry axiom.

3. The LNG storage tank knowledge modeling and intelligent optimization integrated method according to claim 2, characterized in that, The process of converting the LNG storage tank ontology into a graph database based on architectural constraints, and instantiating the engineering data of the LNG storage tank into the graph database according to the architectural constraints to generate an ontology graph database, includes: Pattern-level mapping: Parse the concept set, relation set, and attribute set from the network ontology language file of the LNG storage tank domain ontology, map each concept in the concept set to a node label in the graph database, map each relation in the relation set to an edge type in the graph database, map each attribute in the attribute set to an attribute key-value pair in the graph database, and generate edge endpoint constraint functions according to the domain and scope constraints in the network ontology language file; Instance-level mapping: Under the aforementioned architectural constraints, each data instance in the engineering data of the LNG storage tank is mapped to a node in a graph database, wherein the label of the node contains the direct type of the data instance and all its parent types; each relation assertion in the engineering data instance is mapped to a directed edge between corresponding nodes in the graph database, wherein the direction of the directed edge is from the source node of the relation assertion to the target node, and the label of the directed edge is the relation name of the relation assertion; each attribute assertion in the engineering data instance is stored as a key-value pair in the attribute of the corresponding node.

4. The integrated method for knowledge modeling and intelligent optimization of LNG storage tanks according to claim 3, characterized in that, In the schema-level mapping, the node label set of the ontology graph database From the formula Definition, where For the set of concepts, for Any concept in For mapping functions, For the concept The mapped node labels; the set of edge types in the ontology graph database. From the formula Definition, where For the set of relations, for Any relation in, For relationship The edge type obtained from the mapping; the set of attribute key-value pairs in the ontology graph database. From the formula Definition, where For the attribute set, for Any attribute in, For attributes key name, For attributes Data type; edge endpoint constraint function From the formula definition, Map each edge type to a pair of node labels to specify the source node label and the target node label that the edge connects to; The instance-level mapping includes: the engineering data instance record set Any data instance in Create graph nodes , tag collection Includes the data instance direct type and all its parent types; and, the set of relation assertions. Any two data instances Relationship assertion In these two data instances Corresponding graph nodes and Directed edges are generated between them. The directed edge endpoint functions The directed edge The tag is ; and, for data instances Property assertions Property assertions are made through property mapping functions. key-value pairs Stored to the graph node In the attributes.

5. The integrated method for knowledge modeling and intelligent optimization of LNG storage tanks according to claim 1, characterized in that, Step S1 further includes: in response to a modification request, performing an incremental update operation on the ontology graph database; when the modification request is a structural adjustment, generating a topology sequence from the source node to the target node along a specified path, the topology sequence containing nodes and edges, used to insert new components or remove discarded components; when the modification request is an attribute update, directly modifying the internal parameter values ​​of the target node. Step S1 further includes: in response to a data extraction request, extracting target data from the ontology graph database and outputting it; when the data extraction request is entity extraction, filtering based on node labels and node attributes to obtain a single node that meets the query conditions, and outputting the label and attributes of the node; When the data extraction request is path extraction, under the specified edge type constraint, a simple path consisting of a continuous sequence of relations and without duplicate nodes is obtained by traversing from the source node to the target node, and the attributes of each node, each edge and each node in the path are output. When the data extraction request is a subgraph extraction, a connected subgraph is extracted with the specified node as the root, and the node set, edge set, and attributes of each node in the node set of the connected subgraph are output.

6. The integrated method for knowledge modeling and intelligent optimization of LNG storage tanks according to any one of claims 1 to 5, characterized in that, Step S2 specifically includes: S21. Receive user input for model generation request: The user selects the target model type to be generated through the system interface. The target model type is identified as any one of finite element analysis model, industrial basic engineering management model or STL lightweight visualization model. S22. Determine the corresponding set of mapping rules based on the target model type identifier; S23. Extract a subset of graph data from the graph database that matches the mapping rule set; step S23 specifically includes: Based on the target model type identifier, determine the required data dimensions; the data dimensions should include at least one or more of the following: geometric dimension, material dimension, and topological dimension. Based on the data dimensions, the nodes and edges in the graph database are filtered to extract a subset of graph data that meets the conditions. S24. Based on the mapping rule set, convert the graph data subset into a model file in the target format and output it; S25. In response to incremental updates to the graph database, automatically trigger the regeneration of the corresponding model file.

7. The integrated method for knowledge modeling and intelligent optimization of LNG storage tanks according to any one of claims 1 to 6, characterized in that, Step S3 specifically includes: S31. Establish an optimization problem model for LNG storage tanks. The optimization problem model includes overall design variables, an objective function aimed at minimizing the total steel consumption, constraints on the structural response not exceeding the allowable value, and the boundary range of each design variable. S32. Based on the optimization problem model, a deep neural network proxy model is constructed, which includes the structural configuration of the input layer, hidden layer and output layer and the setting of training hyperparameters; S33. Based on the optimization problem model, perform baseline optimization driven by genetic algorithm to generate an initial population containing a preset number of individuals and iteratively evolve iteratively. For each individual in the population, call the finite element analysis model generated in step S2 to obtain the real structural response, and store the qualified samples into the sample library. S34. During the iteration of the genetic algorithm, the sample library accumulated in step S33 is used as the training set to dynamically train and update the deep neural network surrogate model constructed in step S32. At least during the first training, the optimal hyperparameter configuration is automatically searched and used for subsequent training. S35. After each update of the deep neural network surrogate model in step S34, the current surrogate model is used to perform virtual response prediction on the newly generated candidate individuals after the genetic operation. The individuals with the best prediction are selected and injected into the current population. The injected population is reordered according to the true fitness and a preset number of individuals are retained to enter the next generation, forming a new generation population. The process returns to step S33 to continue iterating until the convergence condition is met and the optimal design parameters are output. The optimization results are written back to the ontology graph database in step S1 through an incremental update mechanism, triggering the regeneration of the corresponding model file in step S2.

8. The integrated method for knowledge modeling and intelligent optimization of LNG storage tanks according to claim 7, characterized in that, In step S31, the overall design variables include the inner tank design vector, the outer tank design vector, and the dome design vector; both the inner tank design vector and the outer tank design vector include the inner diameter of the tank, the height of the cylindrical shell, the thickness of each shell layer, the thickness of the bottom plate, the number of stiffening members, and the unified parameter vector of each stiffening member; the dome design vector includes the dome radius of curvature, the skin thickness, the number of central ring rods, the unified cross-sectional parameter vector of the central ring rods, the number of circumferential members, the unified cross-sectional parameter vector of the circumferential members, the number of radial members, and the unified cross-sectional parameter vector of the radial members; both the unified parameter vector of the stiffening members and the unified cross-sectional parameter vector include discrete cross-sectional type indicators and corresponding geometric parameters or layout parameters; In step S31, the objective function is expressed as: ， in, For the overall design vector, This represents the total number of structural components. For the first The material density of each structural component For the reason The determined first The volume of a structural component, min represents minimization; In step S31, the constraint condition is expressed as follows: ， in, This is the normalized response ratio function. For the overall design vector, For the first Structural response at key locations For the corresponding allowable value, To constrain the total number; In step S31, the external penalty function method is used to process the constraints, and the penalized objective function is expressed as: ， in, As a penalty factor, The function constrains the positive part of the violation, and max denotes the function that takes the maximum value.

9. The integrated method for knowledge modeling and intelligent optimization of LNG storage tanks according to claim 8, characterized in that, In step S32, the input layer dimension of the deep neural network surrogate model is equal to the maximum possible dimension of the design variables; the hidden layer adopts a multi-layer fully connected structure, and batch normalization and Dropout regularization are introduced after each hidden layer, and the activation function of the hidden layer is a linear rectified unit; the output layer dimension is equal to the number of preset key structural response indicators, and the output layer activation function is a linear activation function; the preset key structural response indicators include von Mises equivalent stress, circumferential stress, and axial stress; In step S33, the genetic algorithm includes population initialization, fitness evaluation, selection, crossover, and mutation operations; the population initialization adopts the Latin hypercube sampling strategy; the fitness evaluation calculates the stress of each individual by calling a finite element analysis model; the qualified samples are stored in the sample library after online data cleaning, and the online data cleaning includes removing abnormal samples, removing duplicate samples, and consistency checks. In step S34, the dynamic online training and update refers to: starting the first training when the genetic algorithm iterates to the preset trigger generation, and then retraining once every preset generation using all samples; the automatic search for the optimal hyperparameter configuration refers to calling the secondary genetic algorithm to search in the preset hyperparameter space, with the goal of maximizing the cross-validation accuracy index, and applying the searched optimal hyperparameter configuration to the current and subsequent model training.

10. An integrated system for LNG storage tank knowledge modeling and intelligent optimization, employing the integrated method for LNG storage tank knowledge modeling and intelligent optimization as described in any one of claims 1 to 9, characterized in that, The system includes: The graph database construction module is used to build an ontology graph database based on various design task information of liquefied natural gas storage tanks; The model generation module is used to adaptively generate a multidisciplinary model of a liquefied natural gas storage tank based on the ontology graph database. The proxy-assisted optimization module is used to perform online intelligent optimization of the LNG storage tank structure based on the ontology graph database and the finite element analysis model in the multidisciplinary model, and based on the deep neural network proxy model and genetic algorithm to obtain the optimal design parameters of the liquefied natural gas storage tank.

Images