An Offline Neural Fuzzy Function Control Method for Intelligent Water Resources and Hydrology Assessment

By using an eight-dimensional semantic slot grid modeling method and an offline neurofuzzy function control method based on a neurofuzzy function capsule cluster, the offline adaptation and security issues of intelligent water conservancy and hydrology assessment were solved. This enabled efficient and interpretable intelligent hydrology assessment in the water conservancy intranet environment, adapting to complex and colloquial query scenarios and ensuring the safety, controllability, and traceability of flood control operations.

CN123115389APending Publication Date: 2026-11-17WUHAN XINGHUAN HENGYU INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611415905.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-11
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing intelligent water conservancy and hydrological analysis technologies suffer from problems such as insufficient offline adaptability, lack of risk control system, uncontrollable reasoning logic, high iteration cost, and insufficient data security, which cannot meet the physical isolation and highly rigorous flood control business requirements of the water conservancy intranet.

Method used

An offline neurofuzzy function control method is adopted, which achieves fully offline semantic parsing, risk assessment and tool scheduling through eight-dimensional semantic slot grid modeling, neurofuzzy function capsule clustering, four-fold gating fusion adjudication and evidence ledger full-link evidence storage mechanism, so as to ensure data security and logical controllability.

Benefits of technology

It achieves efficient and interpretable intelligent hydrological analysis within the water conservancy intranet environment without the risk of external network data leakage, adapts to complex and conversational query scenarios, reduces iteration costs, and ensures the safety, controllability, and traceability of flood control operations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115389A_ABST
    Figure CN123115389A_ABST
Patent Text Reader

Abstract

This invention proposes an offline neural fuzzy function control method for intelligent water conservancy and hydrological assessment, comprising: acquiring query questions and performing offline text preprocessing; performing eight-dimensional semantic slot lattice modeling on the text and outputting structured semantic vectors; performing entity compliance verification on the semantic vectors based on a water conservancy knowledge graph; calling a neural fuzzy function capsule cluster deployed in a local offline environment to perform fuzzy semantic quantification assessment and outputting risk assessment results; performing gating fusion adjudication to generate a unified intent slot structure; judging the system operating status based on the intent slot structure and generating a tool invocation plan; performing deterministic hydrological calculations based on measured data, generating business responses and outputting them. This invention abandons reliance on cloud computing power and public network transmission, relies on a general-purpose CPU to achieve localized offline inference, and completes all data processing in a closed loop within the internal network, eliminating the risk of external network data leakage. It meets the strict compliance standards of physical isolation of water conservancy confidential internal networks and is suitable for the routine operation of flood control confidential business.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of water conservancy informatization, edge offline intelligent computing, and water conservancy intelligent risk control technology, specifically to an offline neural fuzzy function control method for intelligent water conservancy and hydrological analysis. This invention is applicable to physically isolated classified water conservancy intranet environments and can be applied to business scenarios such as intelligent analysis of water and rainfall conditions, flood trend analysis, hydrological data verification, flood control auxiliary decision-making, and controlled retrieval of classified water conservancy data. It can meet the hard technical requirements of closed-loop operation of water conservancy business intranet, auditable analysis process, controllable high-risk operations, and lightweight edge deployment. Background Technology

[0002] In the daily operations of water resources and hydrology, including flood control assessment, hydrological consultation, and data verification, there are numerous needs for interactive natural language analysis, covering core business scenarios such as hydrological and rainfall statistics, water level rise and fall rate analysis, rainfall-runoff correlation extrapolation, flood warning determination, dispatch consultation, and historical data verification. These user queries are generally characterized by diverse colloquial expressions, complex nested constraints, and dynamic changes in spatiotemporal dimensions. Furthermore, hydrological monitoring data, flood control dispatch information, and information on water conservancy protection targets are all classified data, and the operation must meet the strict requirements of a closed-loop internal network and no external network data interaction. Therefore, the intelligent water resources and hydrology assessment system must simultaneously possess the technical capabilities of complex semantic parsing, fully offline operation, fine-grained quantitative risk control, and lightweight, low-cost deployment.

[0003] Currently, the main technical approaches to intelligent hydrological analysis in the water conservancy industry fall into the following categories: Manual rule and keyword matching technology: Relying on human experience to enumerate regular rules and business logic, this technology is prone to rule combination explosion problems due to the diverse expression methods of water conservancy statements, numerous site aliases, and dense complex conditions. It has limited scenario coverage, high iteration and maintenance costs, weak generalization ability, and cannot adapt to complex natural language query and judgment scenarios.

[0004] Traditional machine learning classification models can only achieve coarse-grained intent recognition and lack the ability to identify fine-grained risks in the water conservancy field. They cannot detect business risks such as semantic gaps, missing spatiotemporal dimensions, mismatched indicator objects, out-of-bounds permissions, and excessive inference logic. Model training relies on large-scale labeled samples, but grassroots sample resources are scarce, and the output results lack a standardized structure, making it difficult to integrate with risk control verification and tool scheduling processes.

[0005] Cloud-based large-scale models combined with RAG retrieval technology offer strong semantic understanding capabilities, but their operation relies on public network transmission and cloud computing power, posing a risk of leakage of confidential hydrological data and failing to meet compliance requirements for physical isolation within the water conservancy intranet. Furthermore, the large-scale model's reasoning logic is uninterpretable, its operation uncontrollable, and it is prone to generating false hydrological projections. It also lacks standardized evidence tracing links, posing potential flood control safety hazards, and its private deployment costs are high, making large-scale implementation in grassroots water conservancy scenarios difficult.

[0006] General intelligent agent tool invocation technology: It has the ability to autonomously schedule tools, but it is not adapted to the specific business constraints of water conservancy and flood control safety standards. The access control is lax and the invocation logic is disordered. It is prone to problems such as unauthorized retrieval of classified data, unauthorized generation of scheduling suggestions, and acceptance of illegal business relationships. The judgment results have low credibility and uncontrollable risks, and cannot meet the high-rigor and traceable control standards of flood control business.

[0007] In summary, existing technologies generally suffer from shortcomings such as insufficient offline adaptability, lack of risk control systems, uncontrollable reasoning logic, high iteration costs, and insufficient data security. Summary of the Invention

[0008] This invention proposes an offline neural fuzzy function control method for intelligent hydrological assessment, which solves the problems of insufficient offline adaptability, lack of risk control system, uncontrollable reasoning logic, high iteration cost, and insufficient data security in existing intelligent hydrological assessment schemes.

[0009] The technical solution of this invention is implemented as follows: This invention provides an offline neural fuzzy function control method for intelligent water conservancy and hydrological assessment, comprising the following steps: The system acquires user-input hydrological natural language query questions and performs fully offline text preprocessing to obtain a standardized text sequence. Eight-dimensional semantic slot lattice modeling is performed on the standardized text sequence to decompose the query into eight semantic dimensions and output a standard structured semantic vector; The structured semantic vector is subjected to entity compliance verification based on the knowledge graph of the water conservancy and hydrology industry to obtain a set of legitimate candidate entities; The system invokes a cluster of neural fuzzy function capsules deployed in a local offline environment to perform fuzzy semantic quantification on the query question and structured semantic vector based on the set of legal candidate entities, and outputs the risk assessment result. The risk assessment results are subjected to gating and fusion adjudication, and a unified intent slot structure is generated by combining deterministic slot information, graph constraints and business rules; The system's operating status is determined based on the unified intent slot structure, and a tool invocation plan is generated and executed based on the system's operating status. Deterministic hydrological calculations are performed based on the measured data returned by the tool call, generating and outputting business responses.

[0010] Specifically, the mathematical model of the eight-dimensional semantic slot grid is as follows: ; in, , , , , , , , These are eight semantic dimensions: object domain, spatial domain, temporal domain, indicator domain, computational domain, relational domain, permission domain, and output domain. This refines the partial order relationship for slots, enabling the mapping from coarse-grained semantics to fine-grained business slots. This is an intersection operation used for slot conflict verification and illegal semantic determination. This is a union operation used to integrate fragmented semantics to generate standardized business slots.

[0011] Specifically, the formula for entity compliance verification is as follows: ; in, For the set of legal candidate entities; A set of entity nodes in a knowledge graph; For entity type; A set of valid query object types; The scope of the physical space to which it belongs; The spatial scope defined for user questions; Query metrics for users; For entities A set of applicable legal hydrological indicators.

[0012] Specifically, the capsule cluster comprises multiple narrow-domain specialized neurofuzzy function capsules, each capsule independently performing fuzzy semantic quantification and judgment. A single narrow-domain specialized neurofuzzy function capsule is defined as follows: ; in, Indicates the first capsule cluster A neurofuzzy function capsule; For the first Each capsule has its own specific business scenario specifications; , The first Standardized input and output paradigms for each capsule; For the first A set of scenario-based classification tags for each capsule; For the first Independent confidence threshold for each capsule; For the first The inherent risk level parameters of each capsule in the scenario; For the first Evidence matching constraints for the analysis of individual capsules; For the first Offline parameter iteration update rules for each capsule; The neurofuzzy function capsule cluster includes at least several of the following: water level rise rate judgment capsule, rainfall level judgment capsule, trough risk detection capsule, index matching verification capsule, permission overstepping identification capsule, causal relationship judgment capsule, forecast boundary verification capsule, and semantic fuzziness evaluation capsule. When performing fuzzy semantic quantization analysis, the neural fuzzy function capsule uses a continuous membership function to map colloquial fuzzy expressions to quantized values ​​in the range of 0 to 1; wherein, the membership function for the water level rise rate quantization scenario is: ; in, The semantic membership degree for rapid water level rise, with a value range of [0,1]. The rate of water level rise in the river section after noise reduction pretreatment; and To adapt to dynamic business thresholds in different river basins and flood seasons.

[0013] Specifically, the gating fusion decision includes setting four independent gating constraints: structural gating, graph gating, evidence gating, and access gating, to determine the adoption of risk assessment results output by a single capsule. ; in, For the first Each capsule outputs an acceptance flag, with a value of 1 indicating acceptance and 0 indicating rejection; For confidence level gating, For the first The confidence level of the output of each capsule. The confidence threshold for the corresponding capsule. This is a conditional indicator function; it takes a value of 1 when the threshold condition is met and 0 otherwise. For structural integrity gating functions, verify the first... Capsule output With deterministic slot information No conflict; For the graph compliance gating function, verify the first... Capsule output The entity relationships involved in the knowledge graph Legal and compliant; For the validity of evidence, the gate function is used to verify the first... Capsule output There is textual basis ; For business permission gating functions, verify the first... Capsule output Not involving business rule sets Unauthorized operations are restricted; the judgment result of the corresponding capsule is only accepted when all four gating mechanisms have been passed. When the capsule reasoning result matches the deterministic slot information or business rules When conflicts exist, weight reduction is applied: when hour, ; in, For the first The original assessment weight of each capsule, The effective weights after weight reduction For weighting coefficients, This is a conflict detection function. When the value is 1, it determines that the reasoning result conflicts with the structured semantics or rigid business rules and automatically triggers a weight reduction constraint.

[0014] Furthermore, after executing the gating fusion decision but before determining the system's operating status, the process also includes: performing semantic uncertainty budget quantization and calculating the comprehensive uncertainty risk budget value. ; in, For users to query questions The overall uncertainty risk budget value; , , , , , , These are the weighted coefficients for semantic slot missing risk, soft assumption redundancy risk, entity matching redundancy risk, business association ambiguity risk, statement semantic ambiguity risk, permission out-of-bounds risk, and confidential data security risk, respectively. This represents the number of missing slots. This represents the number of soft hypothesis candidates. The number of candidate entities; This represents the number of candidate associations. The semantic ambiguity risk value of the statement; This represents the risk value for exceeding the permission limit. The security risk value for classified data; Budget based on comprehensive uncertainty risk The numerical range automatically switches between four working modes: full execution, degraded execution, user clarification, and operation blocking.

[0015] Furthermore, based on the aforementioned comprehensive uncertainty risk budget value And the capsule output execution risk monotonic irreversible constraint mechanism, including: Calculate the comprehensive risk score for high-risk scenarios: ; in, For users to query questions Comprehensive risk assessment for high-risk scenarios; For the first The effective weight of each capsule after weight reduction; For the first The confidence level of the analysis output of each capsule; For the first The business decision boundary output by each capsule; This is a set of high-risk boundaries; For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. Risk level transitions satisfy the monotonic irreversible constraint: ; in, For the current moment System risk level, The system risk level for the next moment; To add new markers for measured hydrological evidence; Update the identifier for knowledge graph constraints; This is a confirmation mark for manual compliance review; only applicable to... , or If at least one of the conditions is true, the risk level can be reduced; otherwise, the risk level will only increase.

[0016] Specifically, the system operating state is represented as follows: ; in, Indicates the system's operating status; , , , , These represent the following states: fully executed, user-clarified, operation blocked, degraded execution, and manual review. The system combines multiple parameters such as semantic slot integrity, entity compliance, risk level, and tool availability to achieve automatic state transition. When the system is in user clarification, operation blocking, or manual review status, the system status is only allowed to be reassessed after at least one new event occurs, such as user supplementation of missing fields, manual confirmation of graph relationships, or completion of review by authorized personnel.

[0017] Specifically, the tool invocation plan uses a link scoring formula to selectively filter execution links: ; in, For the first A comprehensive score for the execution chain of the tool; , , , These are the normalized weighted coefficients; For the first Bar tools and intent slots Match score; For the first Tools for Knowledge Graphs Coverage integrity score; To execute the first The tool solution quality score; To execute the first Risk deduction for each tool link; The prerequisite for tool execution is that the tool can only be invoked when the system is in a fully executed state or a degraded execution state.

[0018] Preferably, the control method further includes: constructing a multi-objective joint loss function, and using manually reviewed samples to perform closed-loop optimization of the system parameters, wherein the loss function is as follows: ; in, For the multi-objective joint total loss function; , , , These are the weighting coefficients for each loss term; This is to account for the loss due to slot identification error; Losses due to missed detections at the business boundary; Generate error loss for the tool call plan; Losses due to missed detection of high-risk items; Based on the total loss function Targeted adjustment of the confidence decision threshold of the neural fuzzy function capsule The system includes knowledge graph relationships and business rule parameters; at the same time, it retains user query questions, structured semantic vectors, output parameters of each capsule, gating decision basis, tool call logs and final judgment conclusions to form a full-link evidence ledger, and manually reviewed abnormal samples are fed back for iterative optimization.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention abandons the reliance on cloud computing power and public network transmission, and relies on general-purpose CPU to realize local offline reasoning. Data operation, semantic parsing, risk assessment and tool scheduling are all completed in the closed loop of the internal network. There is no risk of external network data leakage. It meets the hard compliance standard of physical isolation of water conservancy confidential internal network and is suitable for the normal operation of flood control confidential business. (2) This invention uses eight-dimensional lattice theory mathematical modeling to replace traditional manual rule enumeration, and constructs a slot lattice mathematical model covering eight semantic dimensions: object, space, time, index, calculation, association, permission, and output. Through partial order relation, intersection operation and union operation, it automatically completes semantic slot detection, conflict verification, intelligent completion and redundancy removal, so that the system has the ability to generalize and adapt to complex colloquial and multi-condition nested hydrological query scenarios, reduces the system iteration and maintenance cost, and solves the problem of manual rule combination explosion. (3) This invention constructs a hydrological-specific fuzzy membership function to achieve continuous mathematical quantification of colloquial fuzzy descriptions such as water level rise rate, rainfall change, and gradual change of working conditions. This makes up for the shortcomings of traditional rules that cannot quantify gradual working conditions and general models that lack business basis for quantification, thereby improving the refinement and interpretability of hydrological working condition assessment. (4) This invention establishes the priority principle of flood control business rules through a triple risk control mechanism of four-dimensional gating fusion adjudication, risk monotonic irreversible constraint, and five-level state machine management, which solves the industry defects of high confidence false inference, unauthorized judgment, and arbitrary risk downgrading of general large models, and ensures the safety and controllability of flood control judgment from the bottom of the algorithm. (5) The present invention adopts an algorithm kernel and industry map decoupling architecture, and the general inference algorithm and risk control logic are fixed and solidified. For different watersheds and different regions, only the local map and site parameters need to be updated. There is no need for model reconstruction and retraining, which reduces the cost of grassroots deployment, transformation and iteration. (6) This invention establishes a full-link evidence ledger storage mechanism to fully retain the logs of semantic parsing, quantitative analysis, risk control decision-making, tool calls, and result output, so as to make the analysis process traceable, verifiable, and auditable; at the same time, it constructs a multi-objective joint loss function and uses manual review samples to perform closed-loop optimization of system parameters, which is suitable for low-resource operation and maintenance scenarios at the grassroots level. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the offline neural fuzzy function control method for intelligent water conservancy and hydrology analysis according to the present invention.

[0022] Figure 2 This is a flowchart of a four-dimensional gating + risk monotonic irreversible risk control embodiment of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] This invention can be deployed on ordinary intranet servers or edge industrial control machines, requiring no high-end GPU computing power, relying only on basic CPU computing resources and a local intranet database, and is suitable for physically isolated deployment environments within water conservancy-related confidential intranets. Before deployment, a localized water conservancy knowledge graph needs to be pre-built, and localized basic data such as river sections, hydrological stations, rainfall stations, warning water levels, flood season thresholds, station aliases, and business permission constraints need to be entered, and the algorithm kernel and risk control rules need to be solidified and configured.

[0025] Reference Figure 1 This invention provides an offline neural fuzzy function control method for intelligent water conservancy and hydrological assessment, comprising the following steps: Step S1: Obtain the user-input hydrological natural language query and perform fully offline text preprocessing to obtain a standardized text sequence.

[0026] The fully offline text preprocessing is performed in a local offline environment and specifically includes: cleaning the input text by removing invalid characters, filtering colloquial particles, normalizing sentence structure, and removing redundant information, eliminating parsing biases caused by differences in input sentence structure, and outputting a standardized text sequence to provide a unified input basis for subsequent structured modeling and fuzzy quantitative analysis. The entire process involves no external network data interaction, ensuring the security of confidential data.

[0027] Step S2: Perform eight-dimensional semantic slot lattice modeling on the standardized text sequence, decompose the query into eight semantic dimensions, and output a standard structured semantic vector.

[0028] The mathematical model of the eight-dimensional semantic slot grid is as follows: ; in, , , , , , , , These are eight semantic dimensions: object domain, spatial domain, temporal domain, indicator domain, computational domain, relational domain, permission domain, and output domain. This refines the partial order relationship for slots, enabling the mapping from coarse-grained semantics to fine-grained business slots. This is an intersection operation used for slot conflict verification and illegal semantic determination. This is a union operation used to integrate fragmented semantic data to generate standardized business slots. Through the above lattice-theoretic quantization operations, the system can automatically complete semantic slot missing detection, conflict verification, intelligent completion, and redundancy removal without manual rule intervention, effectively solving the industry's technical problems of poor adaptability and combinatorial explosion caused by manual rules.

[0029] The system creates a question record for each question, containing eight core fields: object, scope, time, metric, calculation, relation, authority, and output. Fields can be empty, but each non-empty value must include a source. Sources are categorized into four types: original_text (original question), conversation_default (confirmed conversation default value), graph_match (local knowledge graph matching result), and system_clock (system clock). The output of the neural function is not considered a fact source and is only recorded as neural_candidate. The field contents of the question record are shown in Table 1 below: Table 1 Problem Record Fields

[0030] When a value is obtained from the session default range or graph, the record also retains the session round of the default value, the graph entity number, and the matching alias, so as to distinguish between "the user did not say" and "the user omitted it in this round but the session has confirmed it" during subsequent review.

[0031] Step S3: Perform entity compliance verification on the structured semantic vector based on the knowledge graph of the water conservancy and hydrology industry to obtain a set of legitimate candidate entities.

[0032] This invention adopts a decoupled architecture between the algorithm kernel and the local graph. The general core algorithm and risk control logic are fixed and solidified, and regionally differentiated business adaptation only requires replacing the local graph parameters, reducing the difficulty of system iteration and transformation. The decoupled adaptation calculation formula is as follows: ; in, For the overall capabilities of the system; To solidify the universal algorithm kernel; Parameters for the knowledge graph of local water system monitoring stations; Parameters for the regional station and river section alias database; These are the parameters for mapping local business interfaces.

[0033] The formula for entity compliance verification is as follows: ; in, For the set of legal candidate entities; A set of entity nodes in a knowledge graph; For entity type; A set of valid query object types; The scope of the physical space to which it belongs; The spatial scope defined for user questions; Query metrics for users; For entities A set of applicable legal hydrological indicators.

[0034] This step relies on the water conservancy knowledge graph to complete the compliance verification of entities and indicators, filter illegal business associations, and screen legal and valid judgment entities. It is used to intercept illegal indicator matching scenarios such as rain gauges checking water levels and river sections checking rainfall, ensuring that the query semantics are compliant and valid.

[0035] The knowledge graph stores the relationships between administrative regions, water systems, rivers, river sections, monitoring stations, indicators, protection objects, and data interfaces. Instead of handing the entire database text to the neural function, the system first retrieves the graph using the names and types in the original sentences, and then generates a short JSON context. The short context preservation strategies under different input scenarios are shown in Table 2. Table 2. Short Context Preservation Strategies for Graphs

[0036] The graph relationships include the attribute confirmed. Relationships between rainfall stations and hydrological stations with confirmed=false are only used to narrow down the candidate pool and cannot be used to answer "certain impact" or "inevitable cause".

[0037] Step S4: Invoke the neural fuzzy function capsule cluster deployed in the local offline environment, perform fuzzy semantic quantification on the query question and structured semantic vector based on the set of legitimate candidate entities, and output the risk assessment result.

[0038] Addressing the practical application scenarios of water conservancy and hydrological queries, characterized by colloquial language, ambiguous gradual changes in operating conditions, confidential and isolated deployment environments, and limited computing power of grassroots equipment, this invention adopts an offline neural fuzzy function capsule cluster lightweight inference architecture. It abandons the traditional large-model overall coupled inference method and decomposes complex hydrological comprehensive judgment tasks into multiple sets of lightweight inference units with independent boundaries and specialized functions. All units are deployed locally and can achieve offline parallel computing on the intranet using a general-purpose CPU, without the need for cloud computing power or external network data interaction.

[0039] The capsule cluster comprises multiple narrow-domain specialized neurofuzzy function capsules, each performing independent fuzzy semantic quantification and judgment. A single narrow-domain specialized neurofuzzy function capsule is defined as follows: ; in, Indicates the first capsule cluster A neurofuzzy function capsule; For the first Each capsule has its own specific business scenario specifications; , The first Standardized input and output paradigms for each capsule; For the first A set of scenario-based classification tags for each capsule; For the first Independent confidence threshold for each capsule; For the first The inherent risk level parameters of each capsule in the scenario; For the first Evidence matching constraints for the analysis of individual capsules; For the first The offline parameter iteration update rules for each capsule; through the above standardized structure definition, a single neurofuzzy capsule has independent scene adaptation, input and output definition, threshold determination, risk identification and iteration update capabilities. Each capsule is completely decoupled and can be independently iterated and optimized without changing the overall system architecture.

[0040] The neural fuzzy function capsule cluster includes at least several of the following: water level rise rate judgment capsule, rainfall level judgment capsule, trough risk detection capsule, indicator matching verification capsule, permission overstepping identification capsule, causal relationship judgment capsule, forecast boundary verification capsule, and semantic fuzziness evaluation capsule. Each capsule operates in parallel and independently, and the results are aggregated to the downstream risk control module to complete compliance adjudication, effectively avoiding the problems of missed judgments and misjudgments by a single model.

[0041] When performing fuzzy semantic quantization analysis, the neural fuzzy function capsule uses a continuous membership function to map colloquial fuzzy expressions to quantized values ​​in the range of 0 to 1; wherein, the membership function for the water level rise rate quantization scenario is: ; in, The semantic membership degree for rapid water level rise, with a value range of [0,1]. The rate of water level rise in the river section after noise reduction pretreatment; and To adapt to dynamic business thresholds in different river basins and flood seasons.

[0042] The membership value, used as a weighted reference parameter for risk control, must undergo multi-layered gating verification and risk constraints before it can be used as a valid basis for judgment, ensuring that the results of intelligent reasoning do not override the rigid business rules for flood control. By constructing a hydrological-specific fuzzy membership function, continuous mathematical quantification is achieved for colloquial and fuzzy descriptions such as water level rise rate, rainfall changes, and gradual changes in operating conditions, thus overcoming the shortcomings of traditional rules in quantifying gradual operating conditions and the lack of business basis for quantification in general models.

[0043] Step S5: Perform gating fusion adjudication on the risk assessment results, and generate a unified intent slot structure by combining deterministic slot information, graph constraints and business rules.

[0044] To prevent the high-confidence fuzzy inference results of the model from breaking the rigid safety rules of flood control, this invention sets up four independent gating constraints: structure, graph, evidence, and permissions. The corresponding judgment result can only be accepted when all gating checks pass.

[0045] like Figure 2 As shown, the gating fusion decision includes: setting four independent gating constraints: structural gating, graph gating, evidence gating, and access gating, to determine the adoption of the risk assessment results output by a single capsule. ; in, For the first Each capsule outputs an acceptance flag, with a value of 1 indicating acceptance and 0 indicating rejection; For confidence level gating, For the first The confidence level of the output of each capsule. The confidence threshold for the corresponding capsule. This is a conditional indicator function; it takes a value of 1 when the threshold condition is met and 0 otherwise. For structural integrity gating functions, verify the first... Capsule output With deterministic slot information No conflict; For the graph compliance gating function, verify the first... Capsule output The entity relationships involved in the knowledge graph Legal and compliant; For the validity of evidence, the gate function is used to verify the first... Capsule output There is textual basis ; For business permission gating functions, verify the first... Capsule output Not involving business rule sets Unauthorized operations are restricted; the judgment result of the corresponding capsule is only accepted when all four gating mechanisms have been passed. When the capsule reasoning result matches the deterministic slot information or business rules When conflicts exist, the system automatically applies weight reduction constraints to decrease the weight of incorrect inference results and avoid interference from erroneous judgments. The weight reduction calculation formula is as follows: when hour, ; in, For the first The original assessment weight of each capsule, The effective weights after weight reduction For weighting coefficients, This is a conflict detection function. When the value is 1, it determines that the reasoning result conflicts with the structured semantics or rigid business rules and automatically triggers a weight reduction constraint.

[0046] This step uses four-dimensional gating fusion adjudication to perform four independent verifications on the output of the neurofuzzy function capsule: structure, graph, evidence, and permissions. It rigidly establishes the priority principle of flood control business rules and solves the industry defects of general large models such as high-confidence false inferences, unauthorized judgments, and arbitrary risk downgrading.

[0047] Step S6: Perform semantic uncertainty budget quantification and calculate the comprehensive uncertainty risk budget value.

[0048] To achieve more refined, tiered, and quantifiable risk assessment and control, this invention constructs a multi-dimensional semantic uncertainty budget model, calculates the comprehensive risk level of query statements using weighted averages, identifies semantic defects and security risks, and provides a quantitative basis for differentiated risk control strategies.

[0049] The formula for calculating the budgeted value of uncertainty risk is as follows: ; in, For users to query questions The overall uncertainty risk budget value; , , , , , , These are the weighted coefficients for semantic slot missing risk, soft assumption redundancy risk, entity matching redundancy risk, business association ambiguity risk, statement semantic ambiguity risk, permission out-of-bounds risk, and confidential data security risk, respectively. This represents the number of missing slots. This represents the number of soft hypothesis candidates. The number of candidate entities; This represents the number of candidate associations. The semantic ambiguity risk value of the statement; This represents the risk value for exceeding the permission limit. The security risk value for classified data; Budget based on comprehensive uncertainty risk The numerical range automatically switches between four working modes: full execution, degraded execution, user clarification, and operation blocking.

[0050] The specific switching logic for each working mode is as follows: When When the threshold is below the low threshold, switch to full execution mode; when the threshold is between the low and medium thresholds, switch to degraded execution mode; when the threshold is between the medium and high thresholds, switch to user clarification mode; when the threshold is above the high threshold, switch to operation blocking mode.

[0051] The full execution mode indicates that all semantic slots are complete, risk-free, and without unauthorized access; the system generates tool contracts and calls internal network data normally. The reduced-authority execution mode indicates that there is a risk of non-critical data missing or ambiguity, but not high-risk; the system executes the available portion and explicitly indicates the missing items. The user clarification mode indicates that critical semantic slots (such as time, range, or object) are missing; the system only queries for the missing items and does not generate data requests. The operation blocking mode indicates that there is a high-risk unauthorized access, no model prediction, or a prohibited access request; the system returns the boundary and the required capabilities or authorizations, and does not execute any data queries. Through semantic uncertainty budget quantification, the risk level of questions can be calculated in a refined and quantitative manner, providing a quantitative basis for differentiated risk control strategies and solving the subjectivity and non-reproducibility problems of traditional experience-based risk control.

[0052] Step S7, based on the comprehensive uncertainty risk budget value The capsule output execution risk monotonic irreversible constraint mechanism includes the following steps; Calculate the comprehensive risk score for high-risk scenarios: ; in, For users to query questions Comprehensive risk assessment for high-risk scenarios; For the first The effective weight of each capsule after weight reduction; For the first The confidence level of the analysis output of each capsule; For the first The business decision boundary output by each capsule; This is a set of high-risk boundaries (including unauthorized scheduling, unauthorized early warning issuance, model-less forecasting, and access restrictions, etc.). For indicator functions, when The value is 1 if it is true, and 0 otherwise.

[0053] Risk level transitions satisfy the monotonic irreversible constraint: ; in, For the current moment System risk level, The system risk level for the next moment; To add new markers for measured hydrological evidence; Update the identifier for knowledge graph constraints; This is a confirmation mark for manual compliance review; only applicable to... , or The risk level is allowed to be reduced when at least one of the criteria is true, and the risk level will only increase in other scenarios. That is, risk downgrading is only allowed in three compliance scenarios: actual data update, graph rule iteration, and manual compliance review. This mechanism prevents the model from voluntarily relaxing the flood control safety judgment standards.

[0054] Step S8: Determine the system operating status based on the unified intent slot structure, and generate and execute a tool call plan based on the system operating status.

[0055] The system operating status is represented as follows: ; in, Indicates the system's operating status; , , , , These represent the following states: fully executed, user-clarified, operation blocked, degraded execution, and manual review. The system integrates multiple parameters, including semantic slot integrity, entity compliance, risk level, and tool availability, to achieve automatic state transitions. It enters a fully executed state when all necessary fields are complete, entities are parsed, and interfaces and permissions are available. It enters a user clarification state when time, scope, object, or metric is missing. It enters an operation blocking state when requests for warnings, unauthorized scheduling, lack of model predictions, or access restrictions occur. It enters a degraded execution state when non-critical data is missing but remaining data can still answer some questions. It enters a manual review state when data jumps, relationship conflicts, or high-risk issues cannot be automatically confirmed. When returned data lacks necessary evidence fields, the original execution state transitions to degraded execution or manual review.

[0056] When the system is in a user clarification, operation blocking, or manual review state, the system status can only be reassessed after at least one new event occurs: the user supplements missing fields, the graph relationship is manually confirmed, or an authorized person completes the review. A new, recordable event must occur for the system to transition from a clarification, blocking, or manual review state back to the execution state.

[0057] The tool invocation plan uses a link scoring formula to selectively filter execution links: ; in, For the first A comprehensive score for the execution chain of the tool; , , , These are the normalized weighted coefficients; For the first Bar tools and intent slots Match score; For the first Tools for Knowledge Graphs Coverage integrity score; To execute the first The tool solution quality score; To execute the first Risk deductions will be applied to each tool link.

[0058] Using the aforementioned scoring function, the system comprehensively considers four dimensions—slot matching degree, map coverage, data quality, and risk level—to intelligently select the execution path with the lowest risk and highest matching. The precondition for tool execution is that tool invocation is only permitted when the system is in a fully executed or de-weighted execution state. The tool invocation plan is a DAG contractual tool scheduling mode with pre-constraints and post-validation. The tool contract must at least include a tool identifier, object scope identifier, start and end times, query indicators, calculation requirements, required evidence fields, and prohibited actions. The system checks the site identifier, timestamp, unit, data quality flag, and data source of the data returned by the tool. Deterministic hydrological calculations are performed using the checked measured data, and the problem record, map context, candidate results, candidate merge results, state transitions, tool request and return summary, and calculation results are written into the evidence record.

[0059] Step S9: Perform deterministic hydrological calculations based on the measured data returned by the tool call, generate a business response, and output it.

[0060] "Rapid rise in water level," "heavy rainfall," and "approaching warning level" are semantically nuanced expressions. The offline neural function uses these to determine which fields are needed, but does not replace hydrological values ​​with semantic confidence scores. When the user completes the time window... to Afterwards, the station s The water level fluctuation and average rate of change are calculated using the following formula: , ; in, For the station s Water level fluctuations; and They are the stations s At the start time and termination time water level; For the station s The average rate of change. and Observation records must come from the same station code, have the same unit, and meet the quality criteria. If a measurement is missing at any time, that station will not participate in the rate ranking, and the missing endpoint will be recorded in the results without using values ​​from neighboring stations.

[0061] When local business has already registered the lower threshold of "rapid growth" a and upper threshold b At that time, the calculated v Perform segmented fuzzing: ; a and bInstead of relying on neural function guessing, the thresholds must be read from the local threshold table, which includes fields for watershed, station type, flood season / non-flood season, effective date, and reviewer. If no thresholds are available, the system only outputs the measured amplitude, rate, and sorting, without specifying a "fast" or "slow" business level.

[0062] Specifically, after checking the site identifier, timestamp, unit, data quality flag, and data source of the data returned by the inspection tool, deterministic calculations are performed using the inspected measured data, and then the results are organized according to the output contract.

[0063] Step S10: Construct a multi-objective joint loss function and use manually verified samples to perform closed-loop optimization of system parameters.

[0064] The loss function is as follows: ; in, For the multi-objective joint total loss function; , , , These are the weighting coefficients for each loss term; This is to account for the loss due to slot identification error; Losses due to missed detections at the business boundary; Generate error loss for the tool call plan; This represents losses due to missed detection of high-risk events.

[0065] Based on the total loss function Targeted adjustment of the confidence decision threshold of the neural fuzzy function capsule The system includes knowledge graph relationships and business rule parameters; at the same time, it retains user query questions, structured semantic vectors, output parameters of each capsule, gating decision basis, tool call logs and final judgment conclusions to form a full-link evidence ledger, and manually reviewed abnormal samples are fed back for iterative optimization.

[0066] The evidence ledger fully retains the original questions, field values ​​and sources, graph versions and hit relationships, function asset hashes, function original outputs, accepted and ignored candidates, state transitions, tool request and return summaries, calculation parameters and final outputs, so that the objects, times, data and calculation methods in the final answer can be traced back to the corresponding sources, and the historical records are not overwritten by subsequent versions.

[0067] When reflowing failed samples, first determine the location of the failure: if there is an error in time or unit normalization, modify the structuration; if there is an error in entity, alias, or relation, modify the graph; if candidate labels are repeatedly incorrect, regenerate the neural function; if there is evidence of candidate coverage, modify the merge gating; if the gold label itself is inconsistent, proceed to manual review and do not directly modify the function.

[0068] When performing regional adaptation, the administrative regions, water systems, river sections, monitoring stations, aliases, local thresholds, and data interface mappings in the local knowledge graph are replaced, while maintaining the input-output pattern of the offline neural function. The model is then re-accepted on the test question set of the adapted region. Employing a decoupled architecture between the algorithm kernel and the industry graph, adaptation to different watersheds and regions only requires updating the local graph and station parameters, eliminating the need for model reconstruction and retraining, significantly reducing the costs of grassroots deployment, modification, and iteration.

[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An offline neural fuzzy function control method for intelligent water conservancy and hydrological analysis, characterized in that, Includes the following steps: The system acquires user-input hydrological natural language query questions and performs fully offline text preprocessing to obtain a standardized text sequence. Eight-dimensional semantic slot lattice modeling is performed on the standardized text sequence to decompose the query into eight semantic dimensions and output a standard structured semantic vector; The structured semantic vector is subjected to entity compliance verification based on the knowledge graph of the water conservancy and hydrology industry to obtain a set of legitimate candidate entities; The system invokes a cluster of neural fuzzy function capsules deployed in a local offline environment to perform fuzzy semantic quantification on the query question and structured semantic vector based on the set of legal candidate entities, and outputs the risk assessment result. The risk assessment results are subjected to gating and fusion adjudication, and a unified intent slot structure is generated by combining deterministic slot information, graph constraints and business rules; The system's operating status is determined based on the unified intent slot structure, and a tool invocation plan is generated and executed based on the system's operating status. Deterministic hydrological calculations are performed based on the measured data returned by the tool call, generating and outputting business responses.

2. The offline neural fuzzy function control method for intelligent water conservancy and hydrological analysis as described in claim 1, characterized in that, The mathematical model of the eight-dimensional semantic slot grid is as follows: ; in, , , , , , , , These are eight semantic dimensions: object domain, spatial domain, temporal domain, indicator domain, computational domain, relational domain, permission domain, and output domain. This refines the partial order relationship for slots, enabling the mapping from coarse-grained semantics to fine-grained business slots. This is an intersection operation used for slot conflict verification and illegal semantic determination. This is a union operation used to integrate fragmented semantics to generate standardized business slots.

3. The offline neural fuzzy function control method for intelligent water conservancy and hydrological analysis as described in claim 1, characterized in that, The formula for entity compliance verification is as follows: ; in, For the set of legitimate candidate entities; A set of entity nodes in a knowledge graph; For entity type; A set of valid query object types; The scope of the physical space to which it belongs; The spatial scope defined for user questions; For users to query metrics; For entities A set of applicable legal hydrological indicators.

4. The offline neural fuzzy function control method for intelligent water conservancy and hydrological analysis as described in claim 1, characterized in that, The capsule cluster comprises multiple narrow-domain specialized neurofuzzy function capsules, each performing independent fuzzy semantic quantification and judgment. A single narrow-domain specialized neurofuzzy function capsule is defined as follows: ; in, Indicates the first in the capsule cluster A neurofuzzy function capsule; For the first Each capsule has its own specific business scenario specifications; , The first Standardized input and output paradigms for each capsule; For the first A set of scenario-based classification tags for each capsule; For the first Independent confidence threshold for each capsule; For the first The inherent risk level parameters of each capsule in the scenario; For the first Evidence matching constraints for the analysis of individual capsules; For the first Offline parameter iteration update rules for each capsule; The neurofuzzy function capsule cluster includes at least several of the following: water level rise rate judgment capsule, rainfall level judgment capsule, trough risk detection capsule, index matching verification capsule, permission overstepping identification capsule, causal relationship judgment capsule, forecast boundary verification capsule, and semantic fuzziness evaluation capsule. When performing fuzzy semantic quantization analysis, the neural fuzzy function capsule uses a continuous membership function to map colloquial fuzzy expressions to quantized values ​​in the range of 0 to 1; wherein, the membership function for the water level rise rate quantization scenario is: ; in, The semantic membership degree for rapid water level rise, with a value range of [0,1]. The rate of water level rise in the river section after noise reduction pretreatment; and To adapt to dynamic business thresholds in different river basins and different flood seasons.

5. The offline neural fuzzy function control method for intelligent water conservancy and hydrological analysis as described in claim 1, characterized in that, The gated fusion adjudication includes setting four independent gate constraints: structural gate, graph gate, evidence gate, and access gate, to determine the adoption of risk assessment results output by a single capsule. ; in, For the first Each capsule outputs an acceptance flag, with a value of 1 indicating acceptance and 0 indicating rejection; For confidence level gating, For the first The confidence level of the output of each capsule. The confidence threshold for the corresponding capsule. This is a conditional indicator function; it takes a value of 1 when the threshold condition is met and 0 otherwise. For structural integrity gating functions, verify the first... Capsule output With deterministic slot information No conflict; For the graph compliance gating function, verify the first... Capsule output The entity relationships involved in the knowledge graph Legal and compliant; For the evidence validity gating function, verify the first... Capsule output There is textual basis ; For business permission gating functions, verify the first... Capsule output Not involving business rule sets Unauthorized operations are restricted; the judgment result of the corresponding capsule is only accepted when all four gating mechanisms have been passed. When the capsule reasoning result matches the deterministic slot information or business rules When conflicts exist, weight reduction is applied: when hour, ; in, For the first The original assessment weight of each capsule, The effective weights after weight reduction For weighting coefficients, This is a conflict detection function. When the value is 1, it determines that the reasoning result conflicts with the structured semantics or rigid business rules and automatically triggers a weight reduction constraint.

6. The offline neural fuzzy function control method for intelligent water conservancy and hydrological analysis as described in claim 5, characterized in that, After executing the gating fusion decision but before determining the system's operating status, the process also includes: performing semantic uncertainty budget quantization and calculating the comprehensive uncertainty risk budget value. ; in, For users to query questions The overall uncertainty risk budget value; , , , , , , These are the weighted coefficients for semantic slot missing risk, soft assumption redundancy risk, entity matching redundancy risk, business association ambiguity risk, statement semantic ambiguity risk, permission out-of-bounds risk, and confidential data security risk, respectively. This represents the number of missing slots. This represents the number of soft hypothesis candidates. The number of candidate entities; This represents the number of candidate associations. The semantic ambiguity risk value of the statement; This represents the risk value for exceeding the permission limit. This represents the security risk value for classified data. Budget based on comprehensive uncertainty risk The numerical range automatically switches between four working modes: full execution, degraded execution, user clarification, and operation blocking.

7. The offline neural fuzzy function control method for intelligent water conservancy and hydrological analysis as described in claim 6, characterized in that, Based on the aforementioned comprehensive uncertainty risk budget value And the capsule output execution risk monotonic irreversible constraint mechanism, including: Calculate the comprehensive risk score for high-risk scenarios: ; in, For users to query questions Comprehensive risk assessment for high-risk scenarios; For the first The effective weight of each capsule after weight reduction; For the first The confidence level of the analysis output of each capsule; For the first The business decision boundary output by each capsule; This is a set of high-risk boundaries; For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. Risk level transitions satisfy the monotonic irreversible constraint: ; in, For the current moment System risk level, The system risk level for the next moment; To add new markers for measured hydrological evidence; Update the identifier for knowledge graph constraints; This is a confirmation mark for manual compliance review; only applicable to... , or If at least one of the conditions is true, the risk level can be reduced; otherwise, the risk level will only increase.

8. The offline neural fuzzy function control method for intelligent water conservancy and hydrological analysis as described in claim 1, characterized in that, The system operating status is represented as follows: ; in, Indicates the system's operating status; , , , , These represent the following states: fully executed, user-clarified, operation blocked, degraded execution, and manual review, respectively. The system combines multiple parameters such as semantic slot integrity, entity compliance, risk level, and tool availability to achieve automatic state transition. When the system is in user clarification, operation blocking, or manual review status, the system status is only allowed to be reassessed after at least one new event occurs, such as user supplementation of missing fields, manual confirmation of graph relationships, or completion of review by authorized personnel.

9. The offline neural fuzzy function control method for intelligent water conservancy and hydrological analysis as described in claim 1, characterized in that, The tool invocation plan uses a link scoring formula to selectively filter execution links: ; in, For the first A comprehensive score for the execution chain of the tool; , , , These are the normalized weighted coefficients; For the first Bar tools and intent slots Match score; For the first Tools for Knowledge Graphs Coverage integrity score; To execute the first The tool solution quality score; To execute the first Risk deduction for each tool link; The prerequisite for tool execution is that the tool can only be invoked when the system is in a fully executed state or a degraded execution state.

10. The offline neural fuzzy function control method for intelligent water conservancy and hydrological analysis as described in claim 1, characterized in that, The control method further includes: constructing a multi-objective joint loss function, and using manually verified samples to perform closed-loop optimization of system parameters. The loss function is as follows: ; in, For the multi-objective joint total loss function; , , , These are the weighting coefficients for each loss term; This is to account for the loss due to slot identification error; Losses due to missed detections at the business boundary; Generate error loss for the tool call plan; Losses due to missed detection of high-risk items; Based on the total loss function Targeted adjustment of the confidence decision threshold of the neural fuzzy function capsule The system includes knowledge graph relationships and business rule parameters; at the same time, it retains user query questions, structured semantic vectors, output parameters of each capsule, gating decision basis, tool call logs and final judgment conclusions to form a full-link evidence ledger, and manually reviewed abnormal samples are fed back for iterative optimization.