A large model-based water conservancy project quality supervision method and system

Abstract

The application discloses a water conservancy project quality supervision method and system based on a large model, relates to the field of water conservancy projects, and comprises the following steps: collecting multi-source heterogeneous data of a water conservancy project and performing pretreatment to form a spatiotemporally aligned original data pool; constructing a water conservancy project knowledge cube through a large model, inputting a natural language supervision query of the water conservancy project knowledge cube into the large model; decomposing the natural language supervision query into structured subtasks, performing reasoning on data source evidence retrieved from the water conservancy project knowledge cube by using the structured subtasks, and generating a traceable supervision report; inputting an adversarial prompt instruction into the large model, generating an adversarial risk scenario description, and reversely constructing a minimum virtual modification set of part of entities or attributes in the water conservancy project knowledge cube by the large model. The application realizes a fundamental change from information presentation to intelligent insight in water conservancy project quality supervision.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering, and in particular to a method and system for quality supervision of water conservancy projects based on a large model. Background Technology

[0002] In the field of water conservancy engineering, quality supervision is crucial for ensuring safety and efficiency. With the development of IoT, big data, and AI technologies, the supervision model is transforming from traditional manual methods to digital and intelligent approaches. Currently, building an intelligent monitoring platform based on digital twins and multi-source information fusion is essential. By integrating data from sensor networks, BIM models, GIS, and project management systems, a virtual mapping of the project can be constructed. Furthermore, rule engines or machine learning models can be used to achieve threshold-based early warning and anomaly monitoring, thereby improving the automation and visualization of data processing and reducing the workload of manual labor.

[0003] When faced with complex and open professional queries and in-depth risk insights, existing technologies rely on preset models and fixed rules, mainly to match known patterns and issue alerts. They cannot dynamically understand complex professional queries that are cross-domain and not preset, such as comprehensively evaluating the long-term performance of a structure under multiple factors. They lack the ability to autonomously plan analysis paths, build evidence chains, and perform interpretable reasoning. Existing systems are more like data dashboards and rule alarms than intelligent assistants that can perform deep semantic interaction and provide evidence-based analysis reports. When dealing with cutting-edge and exploratory quality risks, their proactive discovery and forward-looking early warning capabilities are insufficient. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a water conservancy project quality supervision method based on a large model to solve the problem that existing technologies cannot dynamically understand complex professional queries, autonomously perform interpretable reasoning, and proactively explore forward-looking risks.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for quality supervision of water conservancy projects based on a large model, which includes collecting multi-source heterogeneous data of water conservancy projects and preprocessing them to form a spatiotemporally aligned raw data pool. A knowledge cube of water conservancy projects is constructed through a large model, and natural language regulatory queries of the knowledge cube of water conservancy projects are input into the large model. The natural language regulatory query is decomposed into structured subtasks. The data source evidence retrieved from the knowledge cube of water conservancy projects is used to perform reasoning to generate a traceable regulatory report. Input adversarial prompts into the large model to generate adversarial risk scenario descriptions. The large model then constructs a set of minimal virtual modifications to some entities or attributes in the water conservancy engineering knowledge cube, forming a counterfactual modification scheme. The counterfactual modification scheme is then virtually applied to generate an adversarial knowledge cube variant. Based on the adversarial knowledge cube variant, natural language regulatory queries are input into the large model again to generate a stress test regulatory report. The differences between the traceable regulatory report and the stress test regulatory report are compared and analyzed to generate a proactive risk exploration and verification report.

[0007] As a preferred embodiment of the water conservancy project quality supervision method based on a large model as described in this invention, the specific steps for forming the spatiotemporally aligned original data pool are as follows: Collect multi-source heterogeneous data from water conservancy projects, including design BIM models, construction logs, material testing reports, sensor time-series data, drone inspection images, and industry standard texts; A unified timestamp operation is performed on the multi-source heterogeneous data of water conservancy projects. A spatial coordinate alignment operation is then performed on the multi-source heterogeneous data of water conservancy projects after the timestamps are unified. The data is then transformed and registered into the unified global coordinate system of the project to form a spatiotemporally aligned original data pool.

[0008] As a preferred embodiment of the water conservancy project quality supervision method based on a large model as described in this invention, the specific steps of the natural language supervision query are as follows: By leveraging a spatiotemporally aligned raw data pool, a large model is driven to perform semantic parsing and entity attribute extraction on unstructured text. By utilizing a spatiotemporally aligned raw data pool, a large model is driven to perform entity recognition and defect localization on image data; Based on the semantic parsing and entity attribute extraction results of unstructured text by the large model and the entity recognition and defect localization results of image data by the large model, a knowledge cube of water conservancy engineering is constructed. Based on the knowledge cube of water conservancy projects, natural language regulatory queries are input into the large model.

[0009] As a preferred embodiment of the large-model-based water conservancy project quality supervision method described in this invention, the specific steps for generating a traceable supervision report are as follows: Based on natural language regulatory queries, a large model is used to perform deep semantic deconstruction and task planning on natural language regulatory queries, forming structured sub-tasks; By utilizing structured subtasks, precise graph queries are performed from the hydraulic engineering knowledge cube to retrieve information fragments corresponding to each structured subtask, along with references to the original data locations, thus forming evidence of the data source. By utilizing the large model to invoke internalized domain rules, logical judgments and calculations are performed on each structured subtask, generating inference sub-conclusions with complete evidence citation chains. By using a large model to comprehensively summarize the reasoning sub-conclusions and evidence citation chains, a traceable regulatory report is generated.

[0010] As a preferred embodiment of the large-model-based water conservancy project quality supervision method described in this invention, the specific steps for describing the adversarial risk scenario are as follows: Based on the knowledge cube of water conservancy projects and traceable regulatory reports, adversarial prompting instructions are constructed to stimulate the exploration of potential compliance margins and long-term complex failure modes. Input adversarial prompts into the large model to drive it to perform inference and synthesis based on the adversarial prompts, the water conservancy engineering knowledge cube, and the traceable regulatory report; Based on adversarial prompts and instructions, and through the deduction and synthesis of large models, adversarial risk scenario descriptions are generated.

[0011] As a preferred embodiment of the large-model-based water conservancy project quality supervision method described in this invention, the specific steps of the counterfactual modification scheme are as follows: Based on the adversarial risk scenario description, a large model is used to perform causal analysis and extract key elements from the adversarial risk scenario description, and identify some entities or attributes in the water conservancy engineering knowledge cube that are directly related to the adversarial risk scenario description. Based on the adversarial risk scenario description and some identified entities or attributes, counterfactual inference is performed using a large model to determine the hypothetical minimal virtual modification set that triggers the consequences defined in the adversarial risk scenario description. Based on the adversarial risk scenario description, the identified entities or attributes, and the determined minimum set of virtual modifications, a counterfactual modification scheme is formed that explicitly records the virtual modifications to some entities or attributes in the knowledge cube of water conservancy projects.

[0012] As a preferred embodiment of the large-model-based water conservancy project quality supervision method described in this invention, the specific steps of the adversarial knowledge cube variant are as follows: Based on the counterfactual modification scheme, a complete copy of the water conservancy engineering knowledge cube is created in memory to generate a replicated knowledge cube; Based on the counterfactual modification scheme and the replicated knowledge cube, all modification operations defined in the counterfactual modification scheme are virtually executed on the replicated knowledge cube to generate a copy of the knowledge cube with the applied modifications. Based on the application-modified knowledge cube copy, logical consistency and data integrity verification are performed on the application-modified knowledge cube copy to generate adversarial knowledge cube variants.

[0013] As a preferred embodiment of the water conservancy project quality supervision method based on a large model as described in this invention, the specific steps of the pressure test supervision report are as follows: Based on the adversarial knowledge cube variant, and using the adversarial knowledge cube variant as the basis for analysis, inputs are fed into the large model along with the initial analytical intent. Figure 1 Natural language regulatory queries; Based on natural language regulatory queries and adversarial knowledge cube variants input into a large model, the large model is used to perform a traceable decomposition, retrieval, and reasoning analysis process on the natural language regulatory queries. Based on a traceable decomposition, retrieval, and reasoning analysis process executed by a large model, stress test regulatory reports are generated.

[0014] As a preferred embodiment of the large-model-based water conservancy project quality supervision method described in this invention, the specific steps of the proactive risk exploration and verification report are as follows: Based on the traceable regulatory report and the stress test regulatory report, a structured comparison of the traceable regulatory report and the stress test regulatory report is conducted in multiple dimensions, including conclusions, reasoning paths and cited evidence chains, to identify all the differences caused by the counterfactual modification scheme; Based on all identified discrepancies and counterfactual modification schemes, the causal relationship between the discrepancies and the specific modification operations in the counterfactual modification schemes is analyzed to locate the key entities, attributes and relationships that lead to the change in the engineering status evaluation, thus forming the location of key vulnerable links. Based on the identification of key vulnerable links, the description of adversarial risk scenarios, and all identified discrepancies, a comprehensive analysis is conducted to generate a proactive risk exploration and verification report.

[0015] Secondly, the present invention provides a water conservancy project quality supervision system based on a large model, including a preprocessing module, which collects multi-source heterogeneous data of water conservancy projects and performs preprocessing to form a spatiotemporally aligned raw data pool. The module constructs a knowledge cube for water conservancy projects using a large model, and inputs natural language regulatory queries of the knowledge cube for water conservancy projects into the large model. The reasoning module decomposes natural language regulatory queries into structured subtasks, uses data source evidence retrieved from the water conservancy engineering knowledge cube from the structured subtasks to perform reasoning, and generates a traceable regulatory report. The construction module inputs adversarial prompts into the large model to generate adversarial risk scenario descriptions. The large model then constructs a minimal virtual modification set of some entities or attributes in the water conservancy engineering knowledge cube to form a counterfactual modification scheme. The counterfactual modification scheme is then virtually applied to generate an adversarial knowledge cube variant. The reporting module, based on an adversarial knowledge cube variant, inputs natural language regulatory queries into the large model again to generate a stress test regulatory report. It then compares and analyzes the differences between the traceable regulatory report and the stress test regulatory report to generate a proactive risk exploration and verification report.

[0016] The beneficial effects of this invention are as follows: It dynamically deconstructs natural language regulatory queries into structured subtasks and performs interpretable reasoning based on data source evidence with precise source identifiers retrieved from the knowledge cube. This enables in-depth understanding and evidence-based analysis of complex and open professional queries, enhancing the credibility and decision support value of regulatory conclusions. Through proactive risk exploration, it uses adversarial prompts to drive the generation of forward-looking risk scenarios in a large model. By performing comparative stress testing analysis on a virtual adversarial knowledge cube variant through counterfactual modifications, it reveals potential vulnerable links and risk transmission paths of projects at the edge of compliance. This upgrades the regulatory model from passive rule response to proactive scenario exploration, endowing the system with the forward-looking ability to anticipate unknown coupled risks, and realizing a fundamental transformation of water conservancy project quality supervision from information presentation to intelligent insight. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without any effort.

[0018] Figure 1 This is a flowchart of a water conservancy project quality supervision method based on a large model.

[0019] Figure 2 This is a schematic diagram of a water conservancy project quality supervision system based on a large model. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. An embodiment appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0023] Reference Figures 1-2 This is one embodiment of the present invention, which provides a method for quality supervision of water conservancy projects based on a large model, including the following steps: S1. Collect multi-source heterogeneous data from water conservancy projects and preprocess them to form a spatiotemporally aligned raw data pool.

[0024] S1.1 Collect multi-source heterogeneous data of water conservancy projects, including design BIM models, construction logs, material testing reports, sensor time-series data, drone inspection images and industry standard texts.

[0025] Furthermore, design BIM models, construction logs, material testing reports, sensor time-series data, UAV inspection images, and industry standard texts are collected from existing information management platforms, IoT platforms, archive databases, and standard document libraries of water conservancy projects. The design BIM model contains geometric, material, and related attribute information of engineering components. Construction logs record daily construction activities, personnel, machinery, and key events in natural language text format. Material testing reports record the physical and mechanical performance indicators of various building materials in structured tables and descriptive text. Sensor time-series data refers to the periodic or continuous reading sequences generated by various monitoring instruments deployed in the engineering structure and environment, such as the outputs of displacement gauges, piezometers, and stress gauges. UAV inspection images are high-resolution images or video streams of the engineering's appearance obtained through aerial photography. Industry standard texts are electronic documents of technical standards and regulations issued by the state and industry. These multi-source heterogeneous data are aggregated into a unified storage area through application programming interfaces, file transfer protocols, or direct database connections.

[0026] S1.2 Perform a timestamp unification operation on the multi-source heterogeneous data of water conservancy projects, perform a spatial coordinate alignment operation on the multi-source heterogeneous data of water conservancy projects after timestamp unification, convert and register it to the unified global coordinate system of the project, and form a spatiotemporally aligned original data pool.

[0027] Furthermore, a timestamp unification operation is performed on the multi-source heterogeneous data of the water conservancy project. Specifically, this involves parsing the time identifiers contained within each data record or in the metadata, and converting all time identifiers to the same time standard, such as Coordinated Universal Time (UTC). After timestamp unification, a spatial coordinate alignment operation is performed on the multi-source heterogeneous data of the water conservancy project. Specifically, this involves identifying the spatial reference information contained in various types of data. For the positioning data of the design BIM model, sensor placement coordinates, and UAV imagery, their spatial coordinates are calculated using coordinate transformation parameters, uniformly converted, and registered to a global coordinate system predefined for the entire water conservancy project. After the timestamp unification and spatial coordinate alignment operations, the multi-source heterogeneous data of the water conservancy project forms a spatiotemporally aligned raw data pool, ensuring that data from different sources and of different types can be associated and compared within a unified temporal and spatial framework. This forms the basis for constructing a spatiotemporally consistent knowledge representation.

[0028] S2. Construct a knowledge cube of water conservancy projects through a large model, and input the natural language regulatory query of the knowledge cube of water conservancy projects into the large model.

[0029] S2.1 Utilize the spatiotemporally aligned raw data pool to drive the large model to perform semantic parsing and entity attribute extraction on unstructured text.

[0030] Furthermore, by utilizing the spatiotemporally aligned raw data pool, unstructured text data such as construction logs, material testing reports, and industry standard texts are extracted. This drives a large-scale model, specifically by using the unstructured text data as input and leveraging the model's language understanding capabilities to identify the entities, events, actions, states, and attributes of the water conservancy project described in the text. Semantic parsing refers to understanding the complete syntactic and contextual meaning of the text. For example, from a construction log, it can be understood that the pouring of C30 concrete in section 3 of the dam describes a pouring event, the initiator being the construction team, the target being section 3 of the dam, and the material used being C30 concrete. Entity attribute extraction, based on semantic parsing, structurally extracts key information. For example, from the aforementioned log, it extracts the event type (pouring), entity (section 3 of the dam), material type (C30 concrete), time attribute (log record time), and possible attributes such as volume and responsible person, and outputs them as entity-relation-attribute triples or similar structured formats. This transforms the professional knowledge recorded in human natural language into structured knowledge units that can be processed and reasoned about by machines.

[0031] Specifically, leveraging the native language understanding capabilities of a general-purpose large-scale model, we perform zero-shot or few-shot deep semantic parsing and information extraction on highly specialized and flexible unstructured texts in the field of water conservancy engineering. Traditional methods rely on pre-defined rule templates or specially trained named entity recognition models tailored to specific entity types and relationships, which have limited generalization capabilities and struggle to cover the varied expressions and complex engineering event descriptions in construction logs. By placing the text in the spatiotemporally aligned raw data pool within its inherent spatiotemporal context, we guide the large-scale model to understand the data using engineering common sense. For example, the large-scale model can not only identify the entity "crack," but also, by combining the spatiotemporal tags of the log, understand that it is a record of a crack discovered at a specific time and in a specific dam section, and can extract its descriptive attributes such as width and orientation. This ensures that the extracted knowledge fragments carry rich contextual semantics and spatiotemporal anchors from their inception, rather than being isolated words. It fully leverages the powerful semantic generalization ability of large models in low-sample scenarios, avoids dependence on massive labeled engineering corpora, and achieves low-cost, high-efficiency, and high-fidelity conversion from free text to structured engineering knowledge, which is key to understanding the knowledge system of engineering narratives.

[0032] S2.2 Utilize the spatiotemporally aligned raw data pool to drive the large model to perform entity recognition and defect localization on image data.

[0033] Furthermore, image data and corresponding spatiotemporal coordinate information are used as input, and the image content is analyzed through the visual understanding capabilities of the large model. Entity recognition refers to locating and classifying different hydraulic engineering components or regions in the image, such as identifying dams, spillways, gates, and slopes in the image. Defect localization, based on entity recognition, further detects and precisely locates apparent defects in the image, such as identifying defect types like cracks, leaks, erosion, and deformation, and marking the specific location and extent of the defects in the image pixel coordinates or a registered global coordinate system. The results of entity recognition and defect localization are usually output in the form of labeled bounding boxes, polygon segmentation masks, or keypoint sets, and are associated with the spatiotemporal metadata of the image, transforming visual information into a structured description with clear semantics and spatial location information, digitizing the appearance of the project, and supplementing intuitive quality status evidence not covered by textual information.

[0034] Specifically, the visual understanding capabilities of large-scale models are applied to fine-grained, multi-target analysis of water conservancy project inspection images, emphasizing the synergy between entity recognition and defect localization within a unified framework. Leveraging the powerful zero-shot visual grounding and open-vocabulary detection capabilities of large-scale models, the system can simultaneously handle two tasks: identifying the engineering component and identifying the problem within that component. For example, when faced with an image of a dam surface with a complex background, the large-scale model can not only identify the concrete panel as an entity but also locate transverse cracks on the panel in the same process, attributing the defect to that entity. More importantly, because the input images carry precise spatial coordinates from a spatiotemporally aligned raw data pool, the identified entities and defects can be directly mapped into the spatial framework of the water conservancy project knowledge cube, establishing accurate spatial associations with other spatial entities such as BIM models and sensor locations. This achieves end-to-end transformation from pixels to semantics and then to spatial location, efficiently and accurately structuring the apparent state information contained in massive inspection images into spatial semantic knowledge that can be fused with textual knowledge, greatly enriching the knowledge cube's perceptual dimensions of the engineering physical world.

[0035] S2.3. Based on the semantic parsing and entity attribute extraction results of unstructured text by the large model and the entity recognition and defect localization results of image data by the large model, a knowledge cube of water conservancy engineering is constructed.

[0036] Furthermore, based on the semantic parsing and entity attribute extraction results of unstructured text by the large model and the entity recognition and defect localization results of image data by the large model, structured outputs from these two sources are aggregated. The entities, relationships, attributes, events, and spatial geometric information contained in these outputs are stored in a graph database according to a predefined ontology or data pattern. In the graph database, physical components, abstract concepts, inspection indicators, documents, and personnel in water conservancy projects are represented as nodes; the relationships between nodes, such as membership, connection, participation, and description, are represented as edges; node attributes record specific parameters; and events are connected to relevant entity nodes as special nodes with timestamps. Simultaneously, spatial information is integrated into the corresponding entity node attributes or managed by a dedicated spatial database and associated with the graph database. Heterogeneous information from textual semantics and visual perception is fused into a unified, networked knowledge structure containing rich semantic relationships and spatiotemporal attributes. This completes the construction of a water conservancy project knowledge cube, creating a machine-readable, globally searchable, and reasoning-along-path engineering digital knowledge system, integrating scattered data fragments into an organic whole.

[0037] Specifically, a dynamic knowledge graph is constructed with an entity-relationship network as its framework, deeply integrating temporal and spatial dimensions. Traditional digital twins may focus on geometric and physical attributes, while traditional document systems focus on textual associations. The knowledge cube unifies these two at the semantic level. For example, an event extracted from text, namely, grouting construction on entity A at time T, and an event identified from imagery, namely, a leakage defect discovered in adjacent region B of entity A at time T+Δt, are represented in the knowledge cube as event nodes and defect nodes connected to entities A and B, respectively. By querying the spatiotemporal and topological relationship paths between event and defect nodes, potential causal relationships can be automatically inferred. The engineering world is modeled as a spatiotemporal-semantic hypergraph, where each fact (whether textually described or visually observed) is encoded as an element in the graph and connected to other elements through relational edges. This allows subsequent regulatory queries to be answered through traversal, reasoning, and computation on the graph, enabling a multi-dimensional and deep understanding and exploration of the engineering status. It is a core infrastructure supporting traceable reasoning and proactive risk exploration.

[0038] S2.4. Based on the knowledge cube of water conservancy projects, input natural language regulatory queries into the large model.

[0039] Furthermore, based on the hydraulic engineering knowledge cube, users or automated processes can input a query, instruction, or analysis request in natural language regarding engineering quality, safety, progress, or compliance into the large model via a natural language interface. This query constitutes a natural language regulatory query. The content of the natural language regulatory query directly points to the engineering status, entity attributes, or historical events represented by the hydraulic engineering knowledge cube. For example, it can query the health status of a specific structural part, assess the compliance of a certain process, or trace the possible causes of a quality problem. The large model receives natural language regulatory queries as input, providing a high-level interactive interface that conforms to human thinking habits and requires no professional programming skills. This allows complex professional regulatory intentions to be directly transmitted to the intelligent analysis engine, initiating subsequent deep analysis processes.

[0040] Specifically, the large model is placed within a structured and semantically rich engineering knowledge context for question-and-answer interaction, rather than being confronted with raw, chaotic, and unrelated multimodal data. Traditional question-answering systems may rely directly on document retrieval or simple database queries, making it difficult to handle complex questions requiring deep reasoning. In this system, the parsing of natural language-controlled queries and subsequent responses will be closely aligned with the established, machine-understandable engineering world model—the hydraulic engineering knowledge cube. For example, when inputting an assessment of the durability of the No. 3 spillway sidewall, the large model understands that the context of this query is the entire knowledge cube. It recognizes that the No. 3 spillway sidewall is a defined entity node, associated with a series of nodes such as design parameters, material reports, construction records, historical inspection data, and image defect records. This allows the large model to accurately transform ambiguous natural language questions into a series of location, retrieval, and reasoning operations performed on the structured network of the knowledge cube. In essence, it creates a two-way mapping channel between semantic digital engineering and natural language interaction. On the one hand, it enables human professional intentions to be accurately projected into the digital world, and on the other hand, it enables the complex analysis results of the digital world to be presented in an understandable way. It is a key entry point for realizing intelligent assistant-style regulatory interaction.

[0041] S3. Decompose the natural language regulatory query into structured subtasks, and use the data source evidence retrieved from the water conservancy engineering knowledge cube by the structured subtasks to perform reasoning and generate a traceable regulatory report.

[0042] S3.1 Based on natural language regulatory queries, a large model is used to perform deep semantic deconstruction and task planning on natural language regulatory queries, forming structured sub-tasks.

[0043] Furthermore, deep semantic deconstruction refers to the large model's analysis and understanding of the underlying intent, core concerns, and related engineering entities and professional concepts contained in the query. Task planning refers to the large model's decomposition and serialization of this complex, macro-level regulatory intent into a set of specific, explicit, atomic, and independently executable data verification or logical judgment instructions based on the semantic understanding of the query. This set of instructions constitutes the structured subtasks. For example, for a natural language regulatory query assessing the durability of the concrete lining of a spillway tunnel, deep semantic deconstruction identifies the assessment object as the concrete lining and the assessment dimension as durability. Task planning might then generate a series of structured subtasks, such as retrieving the design strength and impermeability grade of the lining concrete, obtaining historical data on concrete carbonation depth and chloride ion content, and querying time-series data on crack opening and closing related to the lining.

[0044] Specifically, based on its internalized engineering common sense and logical reasoning capabilities, the large model can autonomously plan a list of subtasks covering multi-dimensional assessment points. For example, when understanding and assessing the seismic safety of a dam, the large model can not only plan subtasks for retrieving static information such as design seismic parameters and structural dynamic characteristics, but also plan dynamic analysis subtasks for analyzing strong earthquake records and structural responses. Natural language queries can be viewed as high-level instructions that need to be compiled or interpreted, while the large model acts as a compiler, compiling these high-level instructions into a series of low-level, precise opcodes—structured subtasks—that can run within the execution environment of the hydraulic engineering knowledge cube.

[0045] S3.2. Using structured subtasks, perform precise graph queries from the hydraulic engineering knowledge cube to retrieve information fragments corresponding to each structured subtask, along with references to the original data locations, thus forming data source evidence.

[0046] Furthermore, for each structured subtask, a corresponding graph database query statement is constructed or invoked to perform a precise graph query within the hydraulic engineering knowledge cube. The graph query traverses the nodes and edges of the knowledge cube, locating nodes that match the entities, attributes, or conditions described by the structured subtask, and extracts specific parameter values, state descriptions, or document references from these nodes and their associated metadata. Each extracted piece of information is accompanied by a unique identifier within the knowledge cube, which traces back to the original storage location of the information source, such as a specific sensor reading record, a paragraph from a detection report document, or a design drawing entry. The collection of all these information fragments with references to the original data location constitutes the data source evidence supporting the structured subtask.

[0047] Specifically, leveraging the inherent graph structure of the knowledge cube, text-described subtasks are directly transformed into navigational queries based on entities, relationships, and attributes. Traditional keyword full-text retrieval or relational database queries struggle to express complex semantic relationships and cross-modal associations, and the provenance of the returned result set is weak. For example, for the structured subtask of retrieving the water level process line of the piezometer P-05 at a key dam section over the past year, the query engine locates the piezometer node named P-05 in the knowledge cube, finds the key section node along the membership relationship, then finds the water level time-series data node along the monitoring data relationship, and returns all readings stored in that data node along with their timestamps, while also returning a unique identifier pointing to the original sensor database record. This not only retrieves the data but also clearly reveals its semantic position within the engineering context. It enforces the binding of evidence to its source, making every data point used for reasoning auditable.

[0048] S3.3. Utilize the large model to call the internalized domain rules to perform logical judgments and calculations on each structured sub-task, generating inference sub-conclusions with complete evidence citation chains.

[0049] Furthermore, from its internalized domain knowledge base, i.e., the domain rule set, the model matches specific rules applicable to the judgment intent of the current subtask. The matching is based on the semantic description of the structured subtask and the type of data source evidence. The large model extracts the specific data content carried by the evidence value field from the data source evidence set, using this specific data content as input parameters, and substituting it into the logical conditions or calculation formulas defined by the matched specific rules for calculation. The calculation process simulates the thought process of a domain expert interpreting data for compliance or assessing its status based on regulatory provisions. The calculation produces a clear intermediate conclusion, such as qualified, exceeding standards, high risk, or a calculated value. The large model generates a structured output, i.e., a reasoning sub-conclusion. This reasoning sub-conclusion not only includes the aforementioned intermediate conclusion but must also record, in the form of a reference list, the original data location reference of each data source evidence upon which this conclusion is based, as well as the identifier of the specific rule applied.

[0050] Specifically, the large model is used as a rule interpreter and logic calculator, and the input, rules, and output are forcibly and structurally linked. It represents the inference sub-conclusion, the structured sub-task, the evidence set, the domain rule set, the specific rules applied, and the specific values ​​extracted from the evidence. Its capabilities are decomposed and constrained within a transparent framework of retrieving evidence, applying rules, producing conclusions, and referencing them. Traditional machine learning models, when performing anomaly detection or classification, have implicit and difficult-to-trace internal decision-making logic. For example, for the sub-task of determining whether the concrete pouring temperature exceeds the standard, the large model matches the rule that the pouring temperature must not exceed a certain threshold, extracts the temperature value from the evidence, and if the condition is true, the conclusion is that it exceeds the standard. The inference sub-conclusion explicitly cites the temperature sensor reading as evidence and the standard clause number as a rule.

[0051] The expression for the inference sub-conclusion is: ; in, For the first A reasoning sub-conclusion, For the first A structured subtask, For the first A set of evidence, For a set of rules for a domain, The first step to make a judgment on structured subtasks Specific rules, This refers to the specific data content carried by the evidence value field in the data source evidence object. Index for subtasks For evidence index, For rule indexing.

[0052] S3.4. Utilize the large model to comprehensively summarize the reasoning sub-conclusions and evidence citation chains to generate a traceable regulatory report.

[0053] Furthermore, the large model receives all inference sub-conclusions generated for the same natural language regulatory query. Each inference sub-conclusion includes its own judgment result, a list of data source evidence citations, and the domain rule identifiers applied. Analyzing the logical relationships between these inference sub-conclusions—such as whether they support a single main argument, or whether causal, progressive, or contradictory relationships exist—the large model integrates information and resolves contradictions based on a review and understanding of the original intent of the natural language regulatory query and the engineering state reflected in each inference sub-conclusion. The large model generates a comprehensive, user-friendly, and readable document: the source tracing regulatory report. This report not only presents a final comprehensive answer and recommendation to the original query but also clearly presents, in structured or natural language, the complete logical chain and source tracing information from the initial query decomposition, to the formulation of each structured subtask, the retrieval of data source evidence, the application of domain rules, and finally to each inference sub-conclusion.

[0054] Specifically, a traceable regulatory report is defined as the final output, and it is stipulated that it must be generated by comprehensively summarizing all atomized inference sub-conclusions and their complete evidence citation chains. This ensures that the final report is not a simple listing of intermediate results, nor a regeneration of the large model detached from evidence, but rather a logical sublimation and summary based on all verifiable intermediate inferences. When generating the report, the large model's thinking is strictly constrained by the evidence network constituted by all inference sub-conclusions. For example, when answering an overall structural durability query by comprehensively considering multiple inference sub-conclusions such as concrete strength, carbonation depth, and crack width, the large model must perform weighted evaluation and comprehensive judgment based on the mutual corroboration or contradiction relationships between these sub-conclusions. The report should indicate that the overall conclusion of insufficient durability is mainly supported by the two sub-conclusions of excessive carbonation depth and active cracks, and guide users to review the detailed evidence and rules for these two sub-conclusions. A conclusion pyramid is constructed, with the base being massive amounts of raw data, the body being the rule-processed, cited inference sub-conclusions, and the apex being the comprehensively summarized traceable regulatory report. In this way, any questioning of the top-level conclusions can be traced down the chain of citations provided in the report, level by level, all the way to the original data at the base of the pyramid. This achieves global interpretability and debatability of the analytical conclusions, firmly anchoring the intelligent output of the large model to verifiable facts and rules, and completely solving the black box and trust problems when artificial intelligence is applied to high-risk fields.

[0055] S4. Input adversarial prompts into the large model to generate adversarial risk scenario descriptions.

[0056] S4.1 Based on the knowledge cube of water conservancy projects and traceable regulatory reports, construct adversarial prompting instructions designed to stimulate the exploration of potential compliance margins and long-term complex failure modes.

[0057] Furthermore, the analysis examines the current state of engineering knowledge reflected in the hydraulic engineering knowledge cube, as well as the existing analytical conclusions and known risk points revealed in the traceable regulatory report. The adversarial prompts aim to guide thinking beyond conventional compliance checks, focusing instead on exploring complex failure possibilities that may only manifest themselves within the boundaries of existing regulations through the slow, long-term interaction of multiple factors. The prompts typically include specific role settings, objective constraints, and thinking frameworks. For example, a large model might be set up as a senior safety analyst dedicated to uncovering hidden vulnerabilities in engineering projects, requiring them to envision a systemic failure scenario where, based on their understanding of the current engineering state, the project formally meets all known regulatory provisions, but experiences a specific time window or encounter with a particular environmental sequence, triggered by the coupling of complex mechanisms such as material performance evolution, load accumulation effects, and multi-component interactions.

[0058] Specifically, adversarial prompts aim to pose a more challenging meta-question: Given complete compliance with known specifications, in what unforeseen ways might a project fail? This is akin to introducing red team thinking or the devil's advocate role into artificial intelligence analysis. For example, based on a traceable regulatory report confirming that all monitoring indicators of a dam are within design limits, a constructible adversarial prompt might be: Assuming you are the dam's designer, from the worst-case scenario, imagine a detailed situation where, assuming no major flaws in the existing design, construction, and monitoring data, the dam's structural behavior might abruptly change at a certain point in the future due to the long-term synergistic effects of alkali-aggregate reaction in concrete, steel corrosion, and freeze-thaw cycles. By using the hydraulic engineering knowledge cube and the traceable regulatory report as the known factual basis and constraints for thought experiments within a large model, the generated deductions are not far-fetched fantasies but rather logically rigorous hypothetical scenarios rooted in engineering reality.

[0059] S4.2 Input adversarial prompts into the large model to drive the large model to perform inference and synthesis based on the adversarial prompts, the water conservancy engineering knowledge cube, and the traceable supervision report.

[0060] Furthermore, the driving force of the large-scale model, based on the input adversarial prompts and instructions, combined with the factual evidence and analytical baseline provided by the hydraulic engineering knowledge cube and the traceable regulatory report, conducts an in-depth, logically consistent deduction and synthesis. Deduction refers to the large-scale model simulating adversarial roles, utilizing its internalized interdisciplinary knowledge of engineering mechanics, materials science, and hydrology, as well as complex systems thinking, to construct in its mind a causal chain that evolves from the current engineering state under the constraints set by the adversarial prompts, gradually leading to a non-obvious failure. Synthesis refers to the large-scale model organizing and outputting this logical chain in a coherent and detailed natural language narrative, forming a complete and descriptive scenario story, i.e., an adversarial risk scenario description. This adversarial risk scenario description must include a clear explanation of the initial triggers, intermediate evolutionary processes, key coupling mechanisms, and the final failure mode.

[0061] Specifically, driven by adversarial prompts, the large model needs to extract the attributes and relationships of relevant entities from the knowledge cube, understand the current safety margin from traceable regulatory reports, and then, like writing a rigorous technical novel, conceive a logically possible bad story. For example, when conceiving a scenario about the sudden instability caused by long-term fretting wear of the hinges of a floodgate, the large model needs to call upon information from the knowledge cube about the gate structure, material properties, lubrication records, and water flow pulsation loads, and consider the conclusion from the traceable regulatory report that the vibration monitoring data showed no abnormalities but was near the upper limit of the specification, thus constructing a progressive failure narrative of fretting wear accumulation leading to changes in fit tolerances, triggering dynamic instability under a specific hydraulic impact. The large model can be viewed as a super simulator capable of counterfactual causal reasoning and long-chain logical narratives. By traversing and combining its vast knowledge parameters, it searches and generates a potential path from the current state to the accident under given constraints, conforming to engineering logic.

[0062] S4.3. Based on the adversarial prompts and instructions, and the deduction and synthesis of large models, generate adversarial risk scenario descriptions.

[0063] Furthermore, relevant facts extracted from the hydraulic engineering knowledge cube, baseline judgments obtained from traceable regulatory reports, and logical causal deductions performed by the large model itself are integrated and woven into a narrative description with a beginning, development, and conclusion. The description of adversarial risk scenarios needs to clearly articulate a complete storyline starting from the current or an initial state, through a series of interactions and temporal evolutions of physical, chemical, or mechanical mechanisms, ultimately leading to some unexpected performance degradation, functional loss, or structural failure in the project. The description should include key risk triggers, intermediate evolutionary stages, coupling relationships between different factors, and the final consequence representation, and the entire narrative should be self-consistent in engineering logic.

[0064] Specifically, a scenario describing a resilient risk, such as that of an arch dam using novel fiber-reinforced concrete, might be described as follows: Under long-term wet-dry cycles and temperature stress, progressive damage occurs at the fiber-matrix interface. Accumulated damage leads to a slight degradation of the effective elastic modulus in localized areas. Although this degradation is not apparent in conventional strength and deformation monitoring, it alters the overall dynamic characteristics of the dam. Under the influence of a far-field seismic wave, local dynamic mismatch causes excessive stress concentration in a specific dam section, ultimately inducing the initiation and propagation of macroscopic cracks in areas with existing micro-damage. This clearly outlines the cross-scale, cross-temporal chain of action from microscopic material damage to macroscopic structural response, treating complex engineering risks as an event sequence or state evolution path rather than isolated event points. By requiring large models to generate such narratives, forcing them to construct a deep and coherent logical framework, the potential key control variables and sensitive links in the risk evolution process are exposed.

[0065] S5. The large model is reverse-engineered to construct a minimal virtual modification set of some entities or attributes in the knowledge cube of water conservancy projects, forming a counterfactual modification scheme.

[0066] S5.1 Based on the adversarial risk scenario description, a large model is used to perform causal analysis and extract key elements from the adversarial risk scenario description, and identify some entities or attributes in the water conservancy engineering knowledge cube that are directly related to the adversarial risk scenario description.

[0067] Furthermore, based on the adversarial risk scenario description, the text describing the adversarial risk scenario is used as input, and a large model is employed to perform causal analysis and key element extraction on the description. Causal analysis refers to the large model analyzing the cause-and-effect relationships in the event chain described in the adversarial risk scenario, identifying the key driving factors, necessary conditions, and evolutionary stages leading to the final outcome. Key element extraction, based on causal analysis, locates and extracts the engineering elements that play a decisive role in the occurrence of the scenario from the natural language narrative of the adversarial risk scenario description. These elements correspond to specific objects or features existing in the hydraulic engineering knowledge cube. For example, from a scenario describing excessive concrete carbonation depth leading to cover failure, which in turn causes steel reinforcement corrosion, expansion, and cracking, causal analysis identifies excessive carbonation depth as the initial cause, cover failure as the intermediate state, and steel reinforcement corrosion and expansion as the direct cause. Key element extraction identifies the concrete carbonation depth attribute, the concrete cover thickness attribute, and the steel reinforcement entity. The large model then queries and matches the corresponding nodes in the hydraulic engineering knowledge cube based on the extracted element names and types, identifying specific entities or attributes located in the knowledge cube that are directly related to the description of adversarial risk scenarios.

[0068] Specifically, the large-scale model needs to deconstruct risk scenarios generated by itself or others, much like deconstructing a technical case, to identify the actual anchor points within the hydraulic engineering knowledge cube upon which the scenario depends. For example, for a scenario describing how, under the combined effects of periodic freeze-thaw cycles and an alkaline environment, the filter layer of a dam's drainage pipe gradually becomes clogged, leading to increased uplift pressure and inducing seepage deformation of the dam foundation, the large-scale model needs to extract key attributes such as the permeability coefficient of the drainage pipe filter layer, the number of freeze-thaw cycles, the pH of the ambient water, the uplift pressure, and the permeability coefficient of the dam foundation. It then needs to locate the entity nodes representing the drainage pipe and the dam foundation, as well as the nodes representing these attributes, within the hydraulic engineering knowledge cube. This ensures that any hypothetical modifications target real, semantically clear elements within the engineering knowledge system, rather than vague concepts.

[0069] S5.2 Based on the adversarial risk scenario description and some identified entities or attributes, use a large model to perform counterfactual inference to determine the hypothetical minimal virtual modification set that triggers the consequences defined in the adversarial risk scenario description.

[0070] Furthermore, to ensure the anticipated consequences in the adversarial risk scenario description occur, assuming the engineering state reflected in the hydraulic engineering knowledge cube remains unchanged, what minimum, hypothetical changes are needed to some of the identified entities or attributes? The large model understands the sufficient condition chain for the consequences to occur from the adversarial risk scenario description, and selects elements from the identified entities or attributes that are located at key nodes in this condition chain and whose state changes are sufficient to significantly drive the scenario evolution. Then, the large model assigns a hypothetical new state value or relationship to each of these selected elements, deviating from the state recorded in the current knowledge cube, forming a set of virtual modifications. The principle for constructing this set is minimization, that is, using the fewest element modifications and the smallest modification magnitude to satisfy the logical probability of triggering the consequences. For example, if the identified relevant attributes include concrete strength, crack width, and ambient humidity, counterfactual inference may determine that only a virtual modification of the ambient humidity attribute from the currently recorded moderate to continuous saturation is needed to logically trigger the accelerated carbonization consequence of the scenario description, thus forming a minimal set of virtual modifications containing only a single-point modification of the ambient humidity attribute.

[0071] Specifically, the concept of counterfactual inference, a philosophical and causal inference field, is transformed into an automated, minimization-based attribution operation performed by a large-scale model for engineering risk analysis. It not only asks, "If A occurs, what will B do?", but more subtly asks, "What minimal changes need to be made to A to make B occur?" This seeks to identify the simplest cause or the most sensitive trigger point for complex risk scenarios. Traditional sensitivity analysis typically perturbs multiple parameters sequentially and observes output changes, but it struggles to automatically determine the minimum perturbation combination. The large-scale model, based on its understanding of the inherent logic of adversarial risk scenario descriptions, acts as an attribution detective. For example, considering a scenario about cavitation damage in a spillway, the description involves multiple factors such as water flow velocity, volume curve, concrete surface smoothness, and aeration facilities. Through counterfactual inference, the large-scale model might determine that, based on the current situation, only a virtual modification to the effectiveness of the aeration facilities—changing it from normal operation to partial failure—is sufficient to logically increase the cavitation risk, without simultaneously modifying the volume curve or flow velocity. This judgment is based on the internalized understanding of the cavitation mechanism by the large model, and it conducts target-oriented vulnerability detection. It does not test all parameters aimlessly, but rather seeks the most economical modification path that can lead to a specific, undesirable consequence.

[0072] S5.3 Based on the adversarial risk scenario description, the identified partial entities or attributes, and the determined minimum set of virtual modifications, a counterfactual modification scheme is formed that explicitly records the virtual modifications to partial entities or attributes in the knowledge cube of water conservancy projects.

[0073] Furthermore, citing the adversarial risk scenario description on which the modification is based, it is stated that the goal of the modification is to verify or explore the specific risk. The target entities or attributes located in the hydraulic engineering knowledge cube are listed, the objects of the modification operation are clarified, and the specific content of each modification in the set of minimizing virtual modifications is specified in detail, including which field of which entity or attribute is being targeted, and what hypothetical state its value is being virtually modified from its current state (recorded in the hydraulic engineering knowledge cube) to.

[0074] Specifically, the generated counterfactual modification scheme is a self-contained experimental design document with clear intent and detailed operational instructions. For example, the scheme might state: To verify the adversarial risk scenario description regarding hydraulic fracturing of the core wall, a specified virtual modification is made to the tensile strength attribute of the core wall soil entity, the rise rate attribute of the reservoir water level entity, and the treatment quality attribute of the construction joint entity within the hydraulic engineering knowledge cube. By referencing the adversarial risk scenario description and the identified entity attributes, it is clearly revealed that this virtual modification aims to test the coupling effect of insufficient tensile strength, sudden water level rise, and weak joints. This establishes a foundation for the experimental repeatability and interpretability of conclusions in AI-driven engineering analysis. The counterfactual modification scheme, like a pre-registration of a scientific experiment, clearly defines the intervention variables, control group, and experimental group.

[0075] S6, Virtual application counterfactual modification scheme, generating adversarial knowledge cube variants.

[0076] S6.1 Based on the counterfactual modification scheme, a complete copy of the water conservancy engineering knowledge cube is created in memory to generate the copied knowledge cube.

[0077] Furthermore, based on the counterfactual modification scheme, a data copy operation is triggered. The source object for the copy operation is the hydraulic engineering knowledge cube. In memory, a logically independent and content-completely identical copy of the hydraulic engineering knowledge cube is created. This copying process must ensure that the copy not only contains all core data elements such as entity nodes, relation edges, and attribute values ​​from the original knowledge cube, but also related metadata, index structure, and spatial-temporal association information. The resulting new data object, the copied knowledge cube, is initially completely identical to the original hydraulic engineering knowledge cube in terms of information content, but is isolated from it in memory address space. This allows subsequent virtual modification operations to be performed on the copied knowledge cube without affecting or polluting the original hydraulic engineering knowledge cube.

[0078] Specifically, it profoundly distinguishes between the two distinct analytical objectives of understanding the current state of the project and verifying risk hypotheses, and provides a physically isolated data environment for both. Traditional simulations or analyses may directly modify parameters on the original model, only to revert them after the analysis. This method carries the risk of accidental contamination of the original data and is not conducive to multiple, parallel comparisons of different hypothetical scenarios. By creating replicated knowledge cubes, it essentially creates an independent, initially controllable laboratory petri dish for each risk exploration experiment. For example, for the same hydraulic engineering knowledge cube, multiple replicated knowledge cubes can be created in parallel based on different counterfactual modification schemes to explore the risk impacts of different single or multi-factor couplings, such as material degradation, load overload, and construction defects, while the original knowledge cube always serves as an unchanging control benchmark. The principles of controlled and repeatable experiments are reflected in digital engineering analysis. It ensures that the generation of adversarial knowledge cube variants is a purely incremental operation without side effects, making all analytical conclusions based on the variants clearly attributable to the applied counterfactual modifications, rather than other unknown data perturbations.

[0079] S6.2 Based on the counterfactual modification scheme and the replicated knowledge cube, all modification operations defined in the counterfactual modification scheme are virtually executed on the replicated knowledge cube to generate a copy of the knowledge cube with the applied modifications.

[0080] Furthermore, on the replicated knowledge cube, these modification instructions are located and executed one by one. Location refers to accurately finding the corresponding data node or field in the graph structure of the replicated knowledge cube based on the entity identifier and attribute path described in the instructions. Execution refers to replacing the current value of the located data node or field with the hypothetical new value specified in the counterfactual modification scheme. This modification process is virtual, meaning it only operates on the in-memory data structure of the replicated knowledge cube and does not reflect or change any real original project data or state. Once all the modification operations defined in the counterfactual modification scheme have been executed on the replicated knowledge cube, the state of the replicated knowledge cube has changed accordingly, generating a new data object reflecting the possible state of the project under specific risk assumptions. This new data object is the modified copy of the knowledge cube.

[0081] Specifically, the counterfactual modification scheme is viewed as a domain-specific script for a specific data structure, and a mechanism for interpreting and executing this script is designed. For example, one instruction in the counterfactual modification scheme is to modify the permeability coefficient of the entity dam foundation curtain to the value K'. The execution engine finds the corresponding node in the replicated knowledge cube, updates its attribute value, and may trigger a chain of updates to related derived or associated attributes (such as updating the estimated seepage flow in the relevant area based on Darcy's law). The targeted execution of the virtual modification set is minimized. It does not reconstruct the entire model, but only operates on the most critical and sensitive parts of the knowledge cube. This results in minimal difference between the generated modified knowledge cube copy and the original copy, but this difference is designed to logically trigger the risk consequences of concern.

[0082] S6.3. Based on the application-modified knowledge cube copy, perform logical consistency and data integrity verification on the application-modified knowledge cube copy, and generate adversarial knowledge cube variants.

[0083] Furthermore, verification primarily focuses on two aspects: logical consistency and data integrity. Logical consistency verification checks whether the logical constraints within the knowledge cube are still satisfied after applying counterfactual modifications. For example, whether modified attribute values ​​remain within a reasonable physical dimension, whether there are contradictions in modified entity relationships, and whether temporal or spatial sequences of entities or events with spatiotemporal dependencies exhibit temporal inversions or spatial conflicts. Data integrity verification checks whether the virtual modification operation unintentionally disrupted the structural integrity of the knowledge cube. For example, whether there are nodes that have become isolated due to modification, whether there are edges with invalid references, and whether the core entity-relationship-attribute triple structure remains intact. The verification process can be completed using predefined constraint rules, ontology axioms, or simple range checks. If inconsistencies or incompleteness are found during verification, it may be necessary to correct the modified copy of the knowledge cube or document known limitations.

[0084] Specifically, verification is equivalent to checking and calibrating the experimental setup before a digital experiment. For example, a counterfactual modification might alter the concrete compressive strength to a negative value, and logical consistency verification will catch this error; or modifying the reservoir water level time series might cause it to lose temporal correlation with rainfall records, and data integrity verification will issue a warning. This ensures that while the adversarial knowledge cube variant describes a hypothetical, risky state, this state itself is possible and autonomous within the engineering knowledge system. This allows subsequent stress test regulatory reports generated based on this variant to be interpreted as engineering responses to an internally consistent adverse scenario.

[0085] S7. Based on the adversarial knowledge cube variant, input natural language regulatory queries into the large model again to generate a stress test regulatory report.

[0086] S7.1. Based on the adversarial knowledge cube variant, the adversarial knowledge cube variant is used as the basis for analysis, and the input to the large model is consistent with the initial analysis intent. Figure 1 Natural language regulatory queries.

[0087] Furthermore, based on the adversarial knowledge cube variant, this variant serves as the foundation for a new engineering knowledge state upon which all subsequent analysis steps depend. The adversarial knowledge cube variant is explicitly designated as the current analysis context, and a natural language regulatory query is input into the large model. This input regulatory query must be consistent with, or highly similar in semantics to, the natural language regulatory query input in the initial analysis phase (i.e., when generating the first traceable regulatory report based on the original hydraulic engineering knowledge cube) in terms of analytical intent, objectives, and the core engineering problem of concern, allowing for parallel comparison. For example, if the initial query assesses the overall stability of the current dam body to ensure that both analyses address the same issue, thus making the subsequently generated reports comparable, a complete, standardized intelligent regulatory analysis process is restarted under an engineering state representing a specific risk assumption after applying counterfactual modifications. This allows for observation and evaluation of how the answer to the same regulatory question changes under the risk assumption.

[0088] Specifically, through reuse and initial analysis... Figure 1 The natural language-based regulatory query is equivalent to setting the same standardized test paper for two different engineering knowledge cubes: the baseline state and the risk assumption state. For example, whether for the original knowledge cube or the adversarial variant, the query asks to assess the operational reliability and main risks of the floodgate: ensuring that the difference between the two analysis outputs is solely attributed to the difference in the input knowledge states (hydraulic engineering knowledge cube vs. adversarial knowledge cube variant), excluding interference from differences in question wording. This mimics the rigorous method in scientific experiments where other variables are controlled and only the independent variable is changed to observe the change in the dependent variable. Any discrepancies between the traceable regulatory report and the stress test regulatory report can be clearly and convincingly attributed to the applied counterfactual modifications, thus accurately revealing how these modifications ultimately altered the conclusions and judgments of intelligent regulation by affecting the engineering knowledge state. This is a prerequisite for achieving high-fidelity, interpretable stress testing and vulnerability localization.

[0089] S7.2 Based on natural language regulatory queries and adversarial knowledge cube variants input to a large model, the large model is used to perform a traceable decomposition, retrieval, and reasoning analysis process on the natural language regulatory queries.

[0090] Furthermore, the large model first performs deep semantic deconstruction and task planning on the input natural language regulatory query, generating structured subtasks for the current query. This process is consistent with the analysis flow based on the original knowledge cube, but the context of the query has been switched to an adversarial knowledge cube variant. Next, using the generated structured subtasks, a precise graph query is executed from the adversarial knowledge cube variant (rather than the original hydraulic engineering knowledge cube), retrieving information fragments supporting each subtask and including references to the original data locations, forming data source evidence based on the variant state. The large model invokes internalized domain rules to perform logical judgments on this data source evidence based on the variant state, generating inference sub-conclusions with complete evidence reference chains. The entire decomposition, retrieval, and inference process is methodologically consistent with the process of generating traceable regulatory reports. The only difference is that the source of all information retrieval and the factual basis on which the inference is based have been replaced with the virtually modified engineering knowledge state represented by the adversarial knowledge cube variant.

[0091] Specifically, the same problem is broken down into the same sub-task logic tree, and the same rules are used to judge the evidence. The only change is that the facts retrieved from the adversarial knowledge cube variant change when the sub-task is executed. For example, when assessing structural stress, the elastic modulus property of the material is virtually reduced in the variant, causing the retrieved data source evidence value to change. Consequently, after applying the same strength criterion for reasoning, the stress level sub-conclusion obtained changes from safe to near critical, and the difference in output caused by the change in input is observed. By reusing a traceable process, the generation process of the stress test regulatory report itself becomes interpretable, and the difference in each sub-conclusion can be traced back to a change in a specific data in the adversarial knowledge cube variant.

[0092] S7.3. Based on a large model, a traceable decomposition, retrieval, and reasoning analysis process is implemented to generate a stress test regulatory report.

[0093] Furthermore, based on the traceable decomposition, retrieval, and inference analysis process executed by the large model, all intermediate products generated in this process are collected, including structured sub-tasks generated for variants of the adversarial knowledge cube, data source evidence retrieved based on variant retrieval, inference sub-conclusions generated after applying rules, and their complete evidence citation chains. The large model synthesizes, summarizes, and formats these intermediate products to generate a final, user-oriented comprehensive analysis document, namely, the stress test regulatory report. The format, structure, and level of detail of the stress test regulatory report should be highly consistent with the traceable regulatory report to ensure comparability. The report content must clearly describe the analytical conclusions, main findings, identified risk points, and corresponding judgment basis for the input natural language regulatory queries under the specific risk assumptions represented by the variants of the adversarial knowledge cube.

[0094] Specifically, the generated stress test regulatory report is a complete and independent regulatory opinion, but it discusses a hypothetical engineering condition. For example, a traceable regulatory report might conclude that the dam's anti-sliding stability meets the specifications and has sufficient safety margin. However, a stress test regulatory report based on an adversarial knowledge cube variant that simulates the weakening of the strength parameters of the weak interlayers in the dam foundation might conclude that, considering a specific geological condition weakening scenario, the dam's anti-sliding stability safety margin is significantly reduced, with some areas approaching the critical point, and recommending a special investigation and verification. Both reports structurally include a current status assessment, compliance analysis, risk identification, and recommendations, but they are based on different factual foundations.

[0095] S8. Compare and analyze the differences between traceable regulatory reports and stress test regulatory reports to generate a proactive risk exploration and verification report.

[0096] Furthermore, S8.1 Based on the traceable regulatory report and the stress test regulatory report, a structured comparison is made between the traceable regulatory report and the stress test regulatory report in multiple dimensions, including conclusions, reasoning paths, and cited evidence chains, to identify all differences caused by the counterfactual modification scheme.

[0097] Furthermore, in terms of the reasoning path dimension, since both reports have a traceable structure, the structured sub-task sequences generated during their creation can be compared in parallel to identify which identical sub-tasks led to discrepancies in the inference sub-conclusions. In terms of the cited evidence chain dimension, for each inference sub-task where the conclusions diverge, a thorough comparison is made of the data source evidence sets cited by the traceable regulatory report and the stress test regulatory report when answering that sub-task, precisely identifying the differences in content or value of the specific evidence relied upon by the two reports. Through this drill-down comparison from macro-level conclusions to micro-level evidence, a series of chain reaction points triggered by the virtual modification of the water conservancy engineering knowledge cube by the counterfactual modification scheme can be identified throughout the entire regulatory analysis chain. These reaction points constitute all the differences caused by the counterfactual modification scheme, capturing all informational deviations between the baseline analysis (original state) and the stress test analysis (risk assumption state), providing a complete input dataset for in-depth attribution analysis.

[0098] Specifically, the two reports are viewed as two reasoning trees with the same trunk but potentially bearing different fruits on different branches. A parallel depth-first traversal and node comparison are performed on both trees. Traditional report comparisons may rely on manual reading or simple text similarity calculations. Causal differences allow for automated and procedural comparisons. For example, in the reasoning tree of a traceable regulatory report, the sub-task of assessing the compressive strength of concrete concludes as qualified, citing the strength value X from test report A as evidence. However, in the corresponding node of the pressure test regulatory report, the reasoning conclusion changes to unqualified, citing the strength value Y from test report A as evidence. The comparison process can automatically identify this difference and record the specific location of the difference as the compressive strength assessment sub-task, the difference manifested as the conclusion changing from qualified to unqualified, and the direct cause being the change of the cited evidence value from X to Y.

[0099] S8.2 Based on all identified discrepancies and counterfactual modification schemes, analyze the causal relationship between the specific modification operations in the discrepancies and counterfactual modification schemes, locate the key entities, attributes and relationships that lead to the change in the engineering status evaluation, and form the location of key vulnerable links.

[0100] Furthermore, each identified discrepancy should be examined individually, especially those that change at the level of the inference sub-conclusion dimension. The counterfactual modification scheme should be reviewed to clarify the specific virtual modifications performed on the hydraulic engineering knowledge cube, including which entity or attribute was modified and by what value. The potential causal relationship between each discrepancy and the specific modification operations in the counterfactual modification scheme should be analyzed. The knowledge logic of the engineering domain needs to be utilized to determine whether a specific virtual modification, through a series of physical, mechanical, or logical transmissions, will necessarily or likely lead to a change in the evidence value in a specific inference sub-task, thereby causing a change in the inference sub-conclusion. For example, if the counterfactual modification scheme includes modifying the value of attribute A of entity E from V1 to V2, and the discrepancy shows that when evaluating the performance P related to attribute A, the cited evidence value changes from a calculation based on V1 to a calculation based on V2, causing the conclusion to change from C1 to C2, then a causal chain can be established from modifying A to V2 to the change in evidence value and then to the conclusion changing to C2. This analysis can identify entities and attributes whose modifications directly or indirectly lead to a reversal of key regulatory conclusions, as well as the key engineering relationships connecting these modifications to the changes in the final conclusions, such as the relationship between material properties and structural stress, and the relationship between permeability coefficient and uplift pressure.

[0101] Specifically, by changing the parameters of a component and observing voltage changes at various points, the sensitive nodes and key signal paths of the circuit can be located. For example, a counterfactual modification scheme might involve altering two attributes: the alkali content in cement and the ambient humidity. The differences would show a deterioration in the sub-conclusions of the concrete carbonation depth assessment and the steel reinforcement corrosion risk assessment. Through causal analysis, it might be found that the modification of ambient humidity, by affecting the carbonation rate, is the primary driving factor leading to the deterioration of a series of subsequent assessments, while the modification of cement alkali content has a secondary impact in this scenario. The identification of key vulnerable links will highlight the attribute of ambient humidity and its influence on the carbonation rate.

[0102] S8.3 Based on the identification of key vulnerable links, the description of adversarial risk scenarios, and all identified discrepancies, a comprehensive analysis is conducted to generate a proactive risk exploration and verification report.

[0103] Furthermore, the engineering sensitivity insights provided by the identification of critical vulnerabilities, the risk evolution narrative provided by the description of adversarial risk scenarios, and the detailed changing facts provided by all identified discrepancies are integrated, refined, and reorganized. The large-scale model or report generation process requires the writing of a new, higher-level analytical document: the Proactive Risk Exploration and Validation Report. The Proactive Risk Exploration and Validation Report should include: a summary restatement of the adversarial risk scenarios explored; a detailed presentation of all key discrepancies and transmission paths identified through structured comparison and causal analysis; a focus on elucidating the findings of the critical vulnerability identification, that is, clearly identifying the entities, attributes, and their relationships within the engineering system most sensitive to specific risk assumptions, explaining the mechanisms of their sensitivity; based on all the above analyses, providing confirmatory conclusions about the potential risk, an assessment of its probability and impact, and priority recommendations for monitoring, inspection, or reinforcement of the identified critical vulnerabilities.

[0104] Specifically, this study explored a hypothetical scenario of accelerated durability degradation of marine structures under the coupled effects of long-term wet-dry cycles and chloride ion erosion. Through stress testing and comparative analysis, it was verified that when the actual thickness of the concrete cover is lower than the design value and the surface chloride ion concentration increases, the depassivation time of the structural reinforcement will be significantly accelerated, leading to a downgrade of the service life assessment from meeting requirements to not meeting requirements. Furthermore, it was clearly pointed out that the identification of critical vulnerability points revealed that construction deviations in the concrete cover thickness and the rate of surface chloride ion accumulation are the two most sensitive parameters affecting whether the risk in this scenario materializes.

[0105] This embodiment also provides a water conservancy project quality supervision system based on a large model, including: a preprocessing module, which collects multi-source heterogeneous data of water conservancy projects and performs preprocessing to form a spatiotemporally aligned raw data pool; The module constructs a knowledge cube for water conservancy projects using a large model, and inputs natural language regulatory queries of the knowledge cube for water conservancy projects into the large model. The reasoning module decomposes natural language regulatory queries into structured subtasks, uses data source evidence retrieved from the water conservancy engineering knowledge cube from the structured subtasks to perform reasoning, and generates a traceable regulatory report. The construction module inputs adversarial prompts into the large model to generate adversarial risk scenario descriptions. The large model then constructs a minimal virtual modification set of some entities or attributes in the water conservancy engineering knowledge cube to form a counterfactual modification scheme. The counterfactual modification scheme is then virtually applied to generate an adversarial knowledge cube variant. The reporting module, based on an adversarial knowledge cube variant, inputs natural language regulatory queries into the large model again to generate a stress test regulatory report. It then compares and analyzes the differences between the traceable regulatory report and the stress test regulatory report to generate a proactive risk exploration and verification report.

[0106] This embodiment also provides a computer device applicable to the water conservancy project quality supervision method based on a large model, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the water conservancy project quality supervision method based on a large model as proposed in the above embodiment.

[0107] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0108] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the water conservancy project quality supervision method based on a large model as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0109] In summary, this invention dynamically deconstructs natural language regulatory queries into structured subtasks and performs interpretable reasoning based on data source evidence with precise source identifiers retrieved from the knowledge cube. This enables in-depth understanding and evidence-based analysis of complex and open-ended professional queries, enhancing the credibility and decision support value of regulatory conclusions. Through proactive risk exploration, it uses adversarial prompts to drive the generation of forward-looking risk scenarios in a large model. By performing comparative stress testing analysis on a virtual adversarial knowledge cube variant through counterfactual modifications, it reveals potential vulnerabilities and risk transmission paths in engineering projects at the edge of compliance. This upgrades the regulatory model from passive rule response to proactive scenario exploration, endowing the system with the forward-looking ability to anticipate unknown coupled risks, and realizing a fundamental transformation in water conservancy project quality supervision from information presentation to intelligent insight.

[0110] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for quality supervision of water conservancy projects based on a large model, characterized in that: This includes collecting multi-source heterogeneous data from water conservancy projects and preprocessing it to form a spatiotemporally aligned raw data pool; A knowledge cube of water conservancy projects is constructed through a large model, and natural language regulatory queries of the knowledge cube of water conservancy projects are input into the large model. The natural language regulatory query is decomposed into structured subtasks. The data source evidence retrieved from the knowledge cube of water conservancy projects is used to perform reasoning to generate a traceable regulatory report. Input adversarial prompts into the large model to generate adversarial risk scenario descriptions. The large model then constructs a set of minimal virtual modifications to some entities or attributes in the water conservancy engineering knowledge cube, forming a counterfactual modification scheme. The counterfactual modification scheme is then virtually applied to generate an adversarial knowledge cube variant. Based on the adversarial knowledge cube variant, natural language regulatory queries are input into the large model again to generate a stress test regulatory report. The differences between the traceable regulatory report and the stress test regulatory report are compared and analyzed to generate a proactive risk exploration and verification report.

2. The water conservancy project quality supervision method based on a large model as described in claim 1, characterized in that: The specific steps for forming the spatiotemporally aligned original data pool are as follows: Collect multi-source heterogeneous data from water conservancy projects, including design BIM models, construction logs, material testing reports, sensor time-series data, drone inspection images, and industry standard texts; A unified timestamp operation is performed on the multi-source heterogeneous data of water conservancy projects. A spatial coordinate alignment operation is then performed on the multi-source heterogeneous data of water conservancy projects after the timestamps are unified. The data is then transformed and registered into the unified global coordinate system of the project to form a spatiotemporally aligned original data pool.

3. The water conservancy project quality supervision method based on a large model as described in claim 2, characterized in that: The specific steps of the natural language monitoring query are as follows: By leveraging a spatiotemporally aligned raw data pool, a large model is driven to perform semantic parsing and entity attribute extraction on unstructured text. By utilizing a spatiotemporally aligned raw data pool, a large model is driven to perform entity recognition and defect localization on image data; Based on the semantic parsing and entity attribute extraction results of unstructured text by the large model and the entity recognition and defect localization results of image data by the large model, a knowledge cube of water conservancy engineering is constructed. Based on the knowledge cube of water conservancy projects, natural language regulatory queries are input into the large model.

4. The water conservancy project quality supervision method based on a large model as described in claim 3, characterized in that: The specific steps for generating a traceable regulatory report are as follows: Based on natural language regulatory queries, a large model is used to perform deep semantic deconstruction and task planning on natural language regulatory queries, forming structured sub-tasks; By utilizing structured subtasks, precise graph queries are performed from the hydraulic engineering knowledge cube to retrieve information fragments corresponding to each structured subtask, along with references to the original data locations, thus forming evidence of the data source. By utilizing the large model to invoke internalized domain rules, logical judgments and calculations are performed on each structured subtask, generating inference sub-conclusions with complete evidence citation chains. By using a large model to comprehensively summarize the reasoning sub-conclusions and evidence citation chains, a traceable regulatory report is generated.

5. The water conservancy project quality supervision method based on a large model as described in claim 4, characterized in that: The specific steps for describing adversarial risk scenarios are as follows: Based on the knowledge cube of water conservancy projects and traceable regulatory reports, adversarial prompting instructions are constructed to stimulate the exploration of potential compliance margins and long-term complex failure modes. Input adversarial prompts into the large model to drive it to perform inference and synthesis based on the adversarial prompts, the water conservancy engineering knowledge cube, and the traceable regulatory report; Based on adversarial prompts and instructions, and through the deduction and synthesis of large models, adversarial risk scenario descriptions are generated.

6. The water conservancy project quality supervision method based on a large model as described in claim 5, characterized in that: The specific steps of the counterfactual modification scheme are as follows: Based on the adversarial risk scenario description, a large model is used to perform causal analysis and extract key elements from the adversarial risk scenario description, and identify some entities or attributes in the water conservancy engineering knowledge cube that are directly related to the adversarial risk scenario description. Based on the adversarial risk scenario description and some identified entities or attributes, counterfactual inference is performed using a large model to determine the hypothetical minimal virtual modification set that triggers the consequences defined in the adversarial risk scenario description. Based on the adversarial risk scenario description, the identified entities or attributes, and the determined minimum set of virtual modifications, a counterfactual modification scheme is formed that explicitly records the virtual modifications to some entities or attributes in the knowledge cube of water conservancy projects.

7. The water conservancy project quality supervision method based on a large model as described in claim 6, characterized in that: The specific steps of the adversarial knowledge cube variant are as follows: Based on the counterfactual modification scheme, a complete copy of the water conservancy engineering knowledge cube is created in memory to generate a replicated knowledge cube; Based on the counterfactual modification scheme and the replicated knowledge cube, all modification operations defined in the counterfactual modification scheme are virtually executed on the replicated knowledge cube to generate a copy of the knowledge cube with the applied modifications. Based on the application-modified knowledge cube copy, logical consistency and data integrity verification are performed on the application-modified knowledge cube copy to generate adversarial knowledge cube variants.

8. The water conservancy project quality supervision method based on a large model as described in claim 7, characterized in that: The specific steps for submitting a stress test regulatory report are as follows: Based on the adversarial knowledge cube variant, the adversarial knowledge cube variant is used as the basis for analysis, and natural language regulatory queries consistent with the initial analysis intent are input into the large model; Based on natural language regulatory queries and adversarial knowledge cube variants input into a large model, the large model is used to perform a traceable decomposition, retrieval, and reasoning analysis process on the natural language regulatory queries. Based on a traceable decomposition, retrieval, and reasoning analysis process executed by a large model, stress test regulatory reports are generated.

9. The water conservancy project quality supervision method based on a large model as described in claim 8, characterized in that: The specific steps for proactive risk exploration and verification reporting are as follows: Based on the traceable regulatory report and the stress test regulatory report, a structured comparison of the traceable regulatory report and the stress test regulatory report is conducted in multiple dimensions, including conclusions, reasoning paths and cited evidence chains, to identify all the differences caused by the counterfactual modification scheme; Based on all identified discrepancies and counterfactual modification schemes, the causal relationship between the discrepancies and the specific modification operations in the counterfactual modification schemes is analyzed to locate the key entities, attributes and relationships that lead to the change in the engineering status evaluation, thus forming the location of key vulnerable links. Based on the identification of key vulnerable links, the description of adversarial risk scenarios, and all identified discrepancies, a comprehensive analysis is conducted to generate a proactive risk exploration and verification report.

10. A water conservancy project quality supervision system based on a large model, based on the water conservancy project quality supervision method based on any one of claims 1 to 9, characterized in that: This includes a preprocessing module that collects multi-source heterogeneous data from water conservancy projects and preprocesses it to form a spatiotemporally aligned raw data pool. The module constructs a knowledge cube for water conservancy projects using a large model, and inputs natural language regulatory queries of the knowledge cube for water conservancy projects into the large model. The reasoning module decomposes natural language regulatory queries into structured subtasks, uses data source evidence retrieved from the water conservancy engineering knowledge cube from the structured subtasks to perform reasoning, and generates a traceable regulatory report. The construction module inputs adversarial prompts into the large model to generate adversarial risk scenario descriptions. The large model then constructs a minimal virtual modification set of some entities or attributes in the water conservancy engineering knowledge cube to form a counterfactual modification scheme. The counterfactual modification scheme is then virtually applied to generate an adversarial knowledge cube variant. The reporting module, based on an adversarial knowledge cube variant, inputs natural language regulatory queries into the large model again to generate a stress test regulatory report. It then compares and analyzes the differences between the traceable regulatory report and the stress test regulatory report to generate a proactive risk exploration and verification report.

Images