Basin maintenance intelligent management system based on internet of things and intelligent decision

Abstract

The application discloses a basin maintenance intelligent management system based on Internet of Things and intelligent decision, and relates to the technical field of computers, comprising a multi-source sensing and edge access module for realizing multi-protocol access of industrial data, low-power point data, video and thermal image data, time and session alignment, data quality labeling, breakpoint continuation and edge side abnormal preliminary screening.In the application, the multi-source sensing and edge access module can process various types of data such as industrial data, low-power point data, video and thermal image data, solve technical problems such as multi-protocol access, time and session alignment, data quality labeling and breakpoint continuation, ensure the integrity and reliability of data on the edge side, and the edge side abnormal preliminary screening function significantly improves the timeliness of initial fault discovery, reduces network bandwidth pressure, effectively solves the problems of abnormal collaborative detection and alarm flooding in complex systems, and improves the accuracy and timeliness of fault warning.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to an intelligent management system for watershed maintenance based on the Internet of Things and intelligent decision-making. Background Technology

[0002] Watershed maintenance work plays a vital role in ensuring the safe and stable operation of water conservancy facilities, improving water resource utilization efficiency, and maintaining ecological balance.

[0003] Traditional watershed maintenance management methods rely heavily on manual inspections, experience-based judgments, and regular maintenance, which have many drawbacks: data collection and processing are lagging and incomplete, asset management and knowledge application are inefficient, equipment status assessment and anomaly warnings are inaccurate, maintenance plans and scheduling lack intelligent support, simulation and impact assessment capabilities are insufficient, spare parts management and operational support are inadequate, and safety and compliance risks exist. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent watershed maintenance management system based on the Internet of Things and intelligent decision-making.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent management system for watershed maintenance based on the Internet of Things and intelligent decision-making, comprising: a multi-source sensing and edge access module for realizing multi-protocol access, time and session alignment, data quality labeling, breakpoint resumption, and edge-side anomaly screening of industrial data, low-power point data, and video and thermal imaging data; an asset ledger and knowledge graph module for establishing a joint ontology covering spatiotemporal, working condition, and procedure dimensions, performing cross-document information extraction, structured constraint verification and minimum repair before knowledge storage, and performing source credibility fusion and tracing of knowledge entries; a status assessment and anomaly detection module for calculating causal consistent health assessment results under the condition that exogenous variables are intervened as a baseline, generating residuals based on counterfactual predictions, outputting the confidence interval of remaining lifetime, and performing graph spatiotemporal anomaly identification and alarm priority fusion on the asset coupling graph; and a maintenance plan and intelligent scheduling module for using a multi-objective optimization framework. The system jointly optimizes the project schedule, cost, downtime risk, and carbon emission intensity, introducing opportunity constraints and conditional risk value constraints to address hydrological uncertainties, and reconstructs the published plan with minimal disturbance within the rolling time domain. The digital twin simulation and impact assessment module performs differentiable calibration of equipment and hydrodynamic processes, weighted fusion of multi-fidelity models based on uncertainties, uses neural operator agents for rapid solution evaluation, and propagates external uncertainties to form robustness scores. The spare parts, energy materials, and operation support module enables safety stock and cross-site or cross-domain allocation based on classification and prediction, and provides mobile operation records and multimodal evidence collection. The safety and compliance module implements zero-trust access control, fine-grained permission management, ticket interlocking, data encryption, and tamper-proof traceability. All modules interact via a data and rule bus between the platform layer and the collaboration layer to form a complete process control from data acquisition, status assessment, solution simulation to planning, scheduling, and closed-loop operation.

[0006] As a further description of the above technical solution: The multi-source sensing and edge access module dynamically adjusts the sampling frequency based on the health status of the equipment and runs a lightweight anomaly detection model on the edge side to trigger local linkage control. At the same time, it performs self-iterative updates of thresholds that can be applied in grayscale. The spare parts, energy materials and operation support module maintains the correspondence between spare parts and components, models and compatibility based on a knowledge graph. Under cross-site conditions, it performs optimal allocation based on delivery time, transportation costs and business impact and achieves full lifecycle tracking. The safety and compliance module interlocks electronic work tickets with on-site control. It automatically blocks dangerous operations when the conditions of handover, double confirmation and isolation tagging are not met, and encrypts and stores evidence of key operations for audit traceability.

[0007] As a further description of the above technical solution: The asset ledger and knowledge graph module models the three dimensions of the joint ontology as spatiotemporal positioning, operating status, and procedural constraints, respectively. It stores the relationships between assets, components, sensor points, defects, procedures, spare parts, work tickets, risks, and geographic topology in a graph database. The asset ledger and knowledge graph module comprehensively considers the matching score of the source, the source quality weight, and the time decay of evidence when evaluating the credibility of candidate triples. The fusion weight is a configurable or learnable parameter. Before knowledge is stored in the database, the asset ledger and knowledge graph module eliminates structured constraint conflicts by solving the minimum repair problem. The repair cost is determined based on a weighted average of credibility and business impact.

[0008] As a further description of the above technical solution: The status assessment and anomaly detection module calculates causally consistent health assessment results based on the deviation between the expected value and the actual observed value when the exogenous variable is fixed as a baseline. The difference between the counterfactual prediction and the actual observation is used as the residual for anomaly identification. The remaining lifespan estimate is obtained by random deactivation sampling to obtain the distribution, and quantile consistency calibration is performed based on an independent calibration set to obtain a confidence interval with controllable coverage. The adjacency relationship of the asset coupling is updated by jointly using a time series model and a graph convolution model, and the strength of the deviation between the predicted value and the actual value is used as the anomaly score.

[0009] As a further description of the above technical solution: The optimization objectives of the maintenance plan and intelligent scheduling module simultaneously consider the total project duration, comprehensive cost, expected conditions of downtime risk, and carbon emission intensity. Opportunity constraints are set for hydrological uncertainties to limit the probability of constraint violation to no more than a set threshold. Feasible initial solutions trained from historical best solutions are used as the solver for warm start-up, and reconstruction and release are carried out in the rolling time domain in a way that minimizes the predetermined work order time offset.

[0010] As a further description of the above technical solution: The digital twin simulation and impact assessment module minimizes both observation errors and physical equation residuals during parameter calibration to ensure consistency with field observations and compliance with physical laws. When fusing multi-fidelity models, it adopts a weighting strategy inversely proportional to uncertainty and uses neural operator agents to quickly predict key indicators in milliseconds. It samples external uncertainties, comprehensively considers the expected value and volatility of impact indicators to generate robustness scores, and feeds back sensitive boundaries and constraints to the scheduler solver.

[0011] As a further description of the above technical solution: A watershed maintenance method based on the Internet of Things and intelligent decision-making includes: S1, Access and Governance: performing multi-protocol data collection, time and session alignment, quality labeling, and breakpoint resumption; S2, Knowledge Construction: performing cross-document extraction based on a joint ontology, performing structured constraint verification and minimum repair before data entry, and completing source credibility fusion and traceability data entry; S3, State Assessment: calculating causally consistent health assessment results under the condition that exogenous variables are intervened as a baseline, generating residuals based on counterfactual predictions, forming confidence intervals for remaining lifetime, and applying them to a graph structure. S4, Scheme Evaluation: Rapidly evaluate candidate maintenance windows using differentiable twin simulation, multi-fidelity fusion, and neural operator agents, and perform uncertainty propagation and robustness scoring; S5, Planning and Scheduling: Solve the maintenance plan under multi-objective optimization, opportunity constraints, and conditional risk-value constraints, employing learning-enhanced warm start and rolling time-domain minimum perturbation release; S6, On-site Execution and Closed-Loop Accumulation: Complete the operation closed loop through ticket interlocking, mobile operation traceability, and multimodal evidence collection, and update the knowledge and model in a versioned manner.

[0012] As a further description of the above technical solution: A watershed maintenance method based on the Internet of Things and intelligent decision-making is proposed. In step S2, the credibility of candidate triples is calculated by weighted fusion of source matching score, source quality, and evidence timeliness decay. Minimum repair is performed to eliminate constraint violations based on a cost model based on credibility and business impact. In step S3, the health status is jointly judged using causal consistency health assessment results and residuals based on counterfactual predictions. The interval estimation of remaining lifetime is calibrated using random deactivation sampling and quantile consistency. In step S3, a time series model and a graph convolution model are jointly used on the asset coupling graph to obtain the predicted value for the next time step. An anomaly score is formed based on the strength of prediction bias, and then compared with the business impact. The order priority is formed by weighted fusion of levels. In S4, differentiable twin parameter calibration and multi-fidelity model weighted fusion are used, and neural operator agent is used for rapid evaluation. After uncertainty propagation, robustness score is calculated and sensitive boundary is fed back to the scheduling module. In S5, a multi-objective function including schedule, cost, expected downtime risk conditions and carbon emission intensity is used for modeling, and opportunity constraints are set to limit the probability of violation. At the same time, learning-enhanced warm start and minimum disturbance reconstruction are used for rolling release. In S6, zero trust and fine-grained access control are used to realize dynamic authorization of personnel, equipment and tasks, and ticket interlocking, encrypted storage and audit traceability are integrated into the operation closed loop.

[0013] As a further description of the above technical solution: An electronic device includes a processor and a memory, the memory storing a program that can run on the processor, the program being executed to cause the electronic device to perform the steps of a method for a watershed maintenance intelligent management system based on the Internet of Things and intelligent decision-making.

[0014] The present invention has the following beneficial effects: 1. In this invention, firstly, through a multi-source sensing and edge access module, the invention can process various types of data such as industrial data, low-power point data, video and thermal imaging data, solving technical challenges such as multi-protocol access, time and session alignment, data quality labeling, and breakpoint resumption, ensuring the integrity and reliability of data at the edge. The edge-side anomaly screening function significantly improves the timeliness of initial fault detection and reduces network bandwidth pressure. The asset ledger and knowledge graph module establishes a joint ontology covering spatiotemporal, operating condition, and procedure dimensions, structurally integrating scattered maintenance knowledge. Through cross-document information extraction and structured constraint verification and minimum repair mechanisms, the quality and consistency of knowledge storage are ensured. The consistency, source credibility fusion and traceability functions effectively solve the problem of knowledge source reliability, and improve the practicality and credibility of knowledge graphs. The status assessment and anomaly detection module calculates causally consistent health assessment results by using the intervention of exogenous variables as a baseline, avoiding the problem of confusion between correlation and causality in traditional assessments. The residuals obtained based on counterfactual prediction can more accurately identify potential anomalies. The output of the remaining lifetime confidence interval provides a quantitative time window for maintenance decisions. The graph spatiotemporal anomaly identification and alarm priority fusion on the asset coupling graph effectively solves the problem of anomaly collaborative detection and alarm proliferation in complex systems, and improves the accuracy and timeliness of fault early warning.

[0015] 2. In this invention, the maintenance planning and intelligent scheduling module, under a multi-objective optimization framework, comprehensively considers the project duration, cost, downtime risk, and carbon emission intensity, achieving a shift from single-objective to multi-objective optimization. This results in more comprehensive decision-making. The introduction of opportunity constraints and conditional risk-value constraints effectively addresses external risks such as hydrological uncertainties, making the plan more robust. The minimum disturbance reconstruction mechanism in the rolling time domain ensures the flexibility and stability of the published plan. The digital twin simulation and impact assessment module ensures the physical consistency and accuracy of simulation results through differentiable calibration of equipment and hydrodynamic processes. The application of multi-fidelity model weighted fusion based on uncertainty and neural operator agents enables rapid evaluation of maintenance plans, shortening the decision-making cycle. External uncertainty propagation and robustness are also addressed. The formation of a robust scoring system makes risk assessment more quantitative and comprehensive. The spare parts, energy materials, and operational support module, based on classification and prediction, enables dynamic management of safety stock and optimized allocation across sites or domains, effectively reducing inventory costs and improving spare parts availability. Mobile operation records and multimodal evidence collection functions enhance the standardization, transparency, and traceability of on-site operations, providing real data support for management decisions. The safety and compliance module implements zero-trust access control, fine-grained permission management, ticket interlocking, data encryption, and tamper-proof traceability, constructing a multi-layered and comprehensive security protection system. In particular, the ticket interlocking mechanism effectively prevents dangerous operations, significantly improves the safety of maintenance operations, and ensures the integrity and auditability of data and operations. Attached Figure Description

[0016] Figure 1 This is a system architecture diagram of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Reference Figure 1This invention provides an embodiment of an intelligent watershed maintenance management system based on the Internet of Things and intelligent decision-making, comprising: a multi-source sensing and edge access module for realizing multi-protocol access, time and session alignment, data quality labeling, breakpoint resumption, and edge-side anomaly screening of industrial data, low-power point data, and video and thermal imaging data; an asset ledger and knowledge graph module for establishing a joint ontology covering spatiotemporal, working condition, and procedure dimensions, performing cross-document information extraction, structured constraint verification and minimum repair before knowledge storage, and performing source credibility fusion and tracing of knowledge entries; a status assessment and anomaly detection module for calculating causal consistent health assessment results under the condition that exogenous variables are intervened as a baseline, generating residuals based on counterfactual predictions, outputting the confidence interval of remaining lifetime, and performing graph spatiotemporal anomaly identification and alarm priority fusion on the asset coupling graph; and a maintenance plan and intelligent scheduling module for managing the project schedule, The system jointly optimizes costs, downtime risks, and carbon emission intensity, introducing opportunity constraints and conditional value-at-risk constraints to address hydrological uncertainties, and reconstructs published plans with minimal perturbation within the rolling time domain. The digital twin simulation and impact assessment module performs differential calibration of equipment and hydrodynamic processes, weighted fusion of multi-fidelity models based on uncertainties, employs neural operator agents for rapid solution evaluation, and propagates external uncertainties to form robustness scores. The spare parts, energy materials, and operational support module enables safety stock and cross-site or cross-domain allocation based on classification and prediction, and provides mobile operation records and multimodal evidence collection. The safety and compliance module implements zero-trust access control, fine-grained permission management, ticket interlocking, data encryption, and tamper-proof traceability. All modules interact via a data and rule bus between the platform layer and the collaboration layer to form a complete process control from data acquisition, status assessment, solution simulation to planning, scheduling, and operational closure.

[0019] The multi-source sensing and edge access module is the front end of the system, responsible for the acquisition and preprocessing of heterogeneous data. This module enables multi-protocol access of industrial data (such as SCADA and PLC data), low-power point data (such as wireless sensor network data), and video and thermal imaging data (such as drone inspection images and infrared temperature measurement). To address the issue of inconsistent timestamps from different data sources, this module provides time and session alignment functions to ensure consistency of all relevant data in the time dimension. A data quality annotation function is also provided to assess data quality and provide a basis for subsequent analysis. For situations of network instability or transmission interruption, it supports breakpoint resumption to ensure the reliability of data transmission. At the edge, a lightweight anomaly detection model is run to perform initial screening of edge-side anomalies. For initially identified anomalies, local linkage control (e.g., audible and visual alarms, automatic valve closure, etc.) can be triggered for rapid response. The sampling frequency is dynamically adjusted according to the equipment health status. When the equipment health is poor or there are potential anomalies, the system can increase the sampling frequency of relevant sensors to obtain more refined data for analysis; when the equipment health is good, the sampling frequency is reduced to save resources. The lightweight anomaly detection model operating at the edge has a threshold that can be updated iteratively with grayscale implementation. This means that the anomaly detection threshold is gradually adjusted and optimized based on historical performance and expert feedback, making it more adaptable to changes in actual operating conditions and improving detection accuracy and recall. Sampling frequency adaptive parameter: The lowest sampling frequency of a low-power sensor. The highest sampling frequency for critical industrial equipment (pump sets, valves) under abnormal conditions; sampling adjustments are based on health indices. The interval: → , → , → Edge anomaly detection threshold: Vibration amplitude threshold: Temperature rise rate threshold: Voiceprint energy offset threshold: ,in, , The steady-state statistical standard deviation. Gray-scale threshold iteration coefficients: initial weights. Update rules: , The expert corrected the signal.

[0020] The asset ledger and knowledge graph module is the core of the system's knowledge base, used to manage dispersed watershed maintenance knowledge in a structured and intelligent manner. This module establishes a joint ontology covering spatiotemporal, operating condition, and procedural dimensions, uniformly modeling information such as equipment, geographical location, operating status, and maintenance procedures within the watershed. Entities and relationships are extracted from various unstructured or semi-structured documents (such as maintenance reports, operation logs, and technical manuals) using cross-document information extraction technologies (such as natural language processing and information retrieval). Before knowledge is stored in the database, structured constraint verification and minimal repair are performed to ensure logical consistency and integrity of the knowledge and eliminate conflicts. Source credibility fusion and tracing are performed on knowledge entries to quantify the reliability of knowledge sources and record the knowledge generation path, thereby improving the transparency and credibility of the knowledge. The three dimensions of the joint ontology are modeled as spatiotemporal positioning (such as latitude and longitude, geographical region, and installation location), operating condition status (such as real-time parameters and their changing trends such as temperature, pressure, and vibration), and procedural constraints (such as maintenance standards, operating procedures, and safety requirements). Simultaneously, a graph database is used to store the relationships between assets, components, sensor points, defects, procedures, spare parts, work orders, risks, and geographical topology, enabling efficient relational queries and reasoning. The asset ledger and knowledge graph module comprehensively considers the source matching score (e.g., the accuracy of the information extraction model), source quality weight (e.g., the authority of expert knowledge bases is higher than that of ordinary documents), and the time decay of evidence (i.e., the older the evidence, the lower its credibility) when evaluating the credibility of candidate triples (entity-relationship-entity). The fusion weights are configurable or learnable parameters to adapt to different business scenarios and data characteristics. Before knowledge is stored in the database, the asset ledger and knowledge graph module eliminates structured constraint conflicts by solving a minimum repair problem (e.g., when two information sources describe the same equipment attributes inconsistently, information is selected or fused by minimizing the repair cost). The repair cost is determined based on a weighted average of credibility and business impact, prioritizing the repair of conflicts with low credibility and high business impact. Credibility fusion parameters: Source matching score weight: Source quality weight: Time-related decay weight: Time-related decay function: ,in Minimum repair cost model parameters: Credibility weight coefficient: Business impact weight: Repair threshold: Knowledge update trigger condition: the same entity in If there are more than 3 conflicting entries, a gray-scale update will be triggered.

[0021] The asset ledger and knowledge graph module uses a hierarchical asset ledger as its framework, unifying the modeling of facilities, systems, equipment, components, and sensor points under the same naming system. This connects unique identifiers and hierarchical topologies, forming a computable and traceable knowledge base. The facility layer covers reservoirs, dams, pumping stations, gates, pipelines, and hydrological stations, recording geographical location, design parameters, operational boundary conditions, service functions, and a list of systems under their jurisdiction. The system layer covers electrical systems, hydraulic systems, and automated control systems, characterizing the electrical parameters of busbars, feeders, transformers, and distribution cabinets; the hydraulic characteristics of pump sets, pipe sections, and valves; and the control attributes of sensor networks, PLC units, and remote communication modules. The equipment layer reaches the management granularity of hoists, pump sets, transformer cabinets, and valves, storing model numbers, nameplate parameters, operating conditions, maintenance records, status tags, and system references. The component layer refines motors, bearings, oil pumps, valve bodies, and control units, recording material parameters, service life, wear status, maintenance records, and associated spare parts models. The sensor point layer includes pressure points, flow points, liquid points, pressure points, and temperature points, labeled with sensor type, installation location, sampling frequency, range, accuracy level, and data quality tags. A top-down hierarchical structure ensures that data from any sensor point can be traced along the chain to components, equipment, systems, and facilities, forming a comprehensive view from point to surface. The ledger is built from design drawings, as-built archives, maintenance documents, and existing databases, uniformly generating facility codes, system codes, equipment codes, component codes, and sensor point codes. A data dictionary and a set of required fields are established, and a graph database is used to construct directed edges for membership and association. Any asset change generates a new version while retaining the mapping of the old version, enabling historical playback and difference comparison. Above the hierarchical ledger, three dimensions—spatiotemporal, operating conditions, and procedures—are introduced to construct a spatiotemporal-operating condition-procedural joint ontology, placing facts, constraints, and operational processes in the same computable space. The spatiotemporal dimension characterizes the watershed, watershed divisions, stations, facility locations, geographic coordinates, terrain fragments, river channels or canals, pipeline segments, and topological nodes, accompanied by coordinate systems, altitudes, mileage, adjacency relationships, upstream and downstream relationships, spatial precision, and timestamps. The operational condition dimension characterizes the operating mode, load levels, boundary conditions, shutdown windows, start / stop events, health levels, and data quality levels, answering the question of how to operate under what boundaries. The procedure dimension carries maintenance procedures, ticket templates, work permits, personnel qualifications, interlock constraints, risk classification, and compliance clauses, providing calculation entry points for procedure numbers, clause numbers, work steps, hazard sources, control measures, qualification types, interlock conditions, and compliance status. The joint ontology is implemented in a graph database through a three-level schema of classes, attributes, and relationships, clearly defining the hierarchy between equipment and stations, the subordination between components and equipment, the constraint correspondence between work permits and procedures, and the fulfillment relationship between personnel qualifications and procedure clauses. It also retains direction, strength, timeliness, source, and credibility scores on cross-dimensional directed edges, achieving semantic consistency, traceability, and auditability.Naming follows a unified standard, with namespaces distinguishing asset domains, spatiotemporal domains, operating condition domains, and procedure domains. All entities and relationships have globally unique identifiers, with relationship identifiers consisting of a start point, an end point, and a relationship type, ensuring cross-domain queries are conflict-free and replayable. Cross-document information extraction transforms facts from maintenance reports, inspection records, defect reports, and work logs into candidate relationship families of faults—components—spare parts—procedures—operating conditions. The process first performs layout parsing, time alignment, and asset alignment, anchoring text paragraphs with unique identifiers for facilities, systems, equipment, components, and sensor points, and removing header and footer noise and template redundancy. Subsequently, dictionary matching, syntactic patterns, and few-shot learning are combined to identify entities and relationships: dictionary matching covers component names, fault types, spare part models, procedure numbers, work ticket numbers, and operating condition terms; syntactic patterns map expressions such as replacing spare parts, executing procedure clauses, and operating under specified loads into structured relationships; and few-shot learning incorporates writing differences between different units and work groups. The identification results are merged with evidence from the same time window and the same asset object to generate candidate triples, along with evidence fragments, source identifiers, page number locations, and timestamps, before entering structured and business constraint verification. On the structure side, it ensures that components can be traced back to equipment, procedure numbers belong to the currently effective version, and operating time and shutdown window are consistent. On the business side, it verifies that concurrent shutdowns of the same pump group are prohibited, electrical maintenance requires dual qualifications, and single-person operation is prohibited in high-risk conditions. When conflicts exist, a minimum-cost repair strategy is adopted, proposing demotion, delayed entry into the database, or transfer to the pending review area, and recording the triggering rules, conflict fields, and evidence chains. If the candidate relationship exceeds the existing schema, the system automatically generates an ontology evolution proposal, including a temporary name for the new entity or relationship, a suggested parent class, a set of necessary attributes, adjacent relationships, a list of affected queries and rules, and representative evidence. This proposal enters expert review and shadow release; after confirmation of no destructive impact, it is solidified as a new version, and backward compatibility and rapid rollback are ensured through mapping tables and read-only aliases. The credibility score quantifies the admissibility of relation entries, comprehensively considering source quality, evidence decay over time, and multi-source fusion weights under a unified time window, object granularity, and ontology version. Source quality is dynamically calibrated based on channel reliability, approval completeness, extraction path stability, and historical accuracy. Data interfaces already online, issued work orders, and maintenance reports with approval flows receive higher weights, while OCR parsing and instant messaging transcription receive lower weights. Failure over time differentiates relation types: equipment and operating conditions, and equipment and defects experience faster weight declines in the short term; components and their affiliations, and equipment and their systems maintain higher validity over the long term. During the flood season, evidence related to water levels and inflow shortens the effective window to quickly reflect on-site changes. Multi-source fusion emphasizes diversity, consistency, and interpretability. Heterogeneous evidence receives diversity bonuses, highly consistent evidence receives consistency weights, and conflicts in key fields trigger adjudication. High-quality sources and versions with new timestamps are prioritized for retention, and conflict penalties are applied to the final credibility score. The contribution ratio and reasons for adoption of each piece of evidence are output.The scoring results, along with quality, stability, and availability tags, are written to the main database. High-trust relationships directly participate in inference, gray-zone relationships initiate proactive learning and review, and low-trust relationships are downgraded and archived for incremental correction. When outputting externally, the source composition, time distribution, and adjudication summary are provided to facilitate the setting of thresholds and weighting strategies for planning, scheduling, and security compliance modules. Pre-database constraint verification ensures structural consistency, business compliance, and timeliness. In the object alignment phase, the subject, object, and time window of candidate entries are anchored to a unique asset identifier and a unified timeline, and semantic definition is completed by referencing the classes, attributes, and relationships of the joint ontology. In the rule binding phase, corresponding constraint templates and triggering conditions are automatically loaded based on the entry type, facility and system, operating condition, and applicable procedures. Rules such as prohibiting concurrent shutdowns of the same pump group, requiring dual qualifications for electrical maintenance, and prohibiting single-person operation in high-risk conditions are instantiated into computable predicates and time window relationships, clarifying the required contextual evidence and authorization boundaries. During the evidence collection and adjudication phase, work orders, qualification files, scheduling instructions, shutdown windows, interlock status, and risk levels are retrieved from the knowledge base and operational archives. Evaluation is performed according to template priority and short-circuit strategies. If a prohibitive constraint is hit, a conflict is identified, and the rule number, triggering field, and evidence source are recorded. If a conflict is established, the system prioritizes data removal with a reduced weight as a reference, or suggests delaying data removal with a minimum adjustable time window, or sends the data to the review area and generates a supplementary materials list and reviewer. When multiple remediation paths coexist, a comprehensive evaluation is conducted based on four costs: disruption to live network queries and rule execution, violation of historical consistency, impact on security boundaries, and increased data removal latency. A default recommendation is selected, and manual one-click switching is allowed. The entire process generates a violation interpretation path, connecting rule instantiation, evidence retrieval, hit criteria, adjudication conclusion, and remediation actions. The rule number, data source, version number, hit fields, time window, and participants are recorded. After remediation, the results, review conclusion, and final data removal version are recorded to ensure replayability and auditability. Version governance employs dual control via namespaces and effective time. Namespaces distinguish different ontology schemas and knowledge entries. During major structural upgrades, old namespaces are converted to read-only mode, and cross-version access consistency is maintained through federated views. Effective dates mark the applicable range of entries or rules, allowing the same relation to have different values ​​and interpretations at different times. Version evolution follows evolution proposals, expert reviews, shadow releases, canary releases, and official implementation, recording effective dates, impact lists, and rollback points. When necessary, a rapid rollback restores the old namespaces and mapping tables and freezes writes involving new structures, ensuring continuous availability and compliance auditing. Trust scoring parameters, source weights, timeliness curves, and fusion thresholds are incorporated into the same version governance. Any policy changes are recorded for applicable objects and impact assessments for auditing and playback. The data channel employs batch processing and streaming parallel mechanisms, completing both one-time cleaning, identification, and batch writing of historical data, as well as online extraction, matching, and verification of new work tickets, real-time data collection, and manual entry.All inbound requests are intercepted before entering the constraint engine. Violations and inconsistencies are blocked and sent to the repair or review path. Qualified entries, carrying their version number, effective date, and source link, are written to the main database, supporting time replay and difference comparison. The graph database uses an attribute graph model to represent nodes and edges. Nodes represent facilities, systems, equipment, components, sensor points, operating conditions, boundary conditions, shutdown windows, start / stop events, maintenance procedures, work tickets, personnel qualifications, interlock constraints, and compliance clauses. Edges express affiliation, location, monitoring, connection, dependence, constraint, and role, and record timeliness, reliability, and traceability links on the edges. Indexes are built around sites, equipment types, procedure numbers, and time windows to improve retrieval efficiency. Operational inference and services revolve around scheduling, safety compliance, and anomaly detection. The system can derive feasible shutdown windows from equipment operating conditions and boundary conditions, determine work ticket compliance status based on personnel qualifications and interlock constraints, calculate cross-site linkage impacts by combining geographical topology and adjacency relationships, and feed the results back to upper-level modules as factual basis and rule constraints. The credibility scoring of knowledge relationships and the explanation path for violations jointly ensure the reliability and traceability of reasoning input; unified naming and version governance ensure cross-period comparison and rapid rollback capabilities; batch processing and streaming parallelism ensure the synchronous updating of knowledge and on-site status. Through the complete description of hierarchical ledgers, semantic alignment of joint ontologies, evidence aggregation from cross-document extraction, dynamic quantification of credibility scores, access control for constraint verification, replayable evolution of version governance, and efficient retrieval and reasoning of graph databases, assets, operating conditions, defects, procedures, spare parts, work tickets, risks, and geographical topology are unified into a computable, verifiable, and evolvable knowledge graph. This mechanism reduces information fragmentation and procedure misuse, improves the automatic violation detection rate, and reduces the scheduling recalculation failure rate, continuously providing a stable and reliable knowledge foundation for watershed-oriented planning, safety compliance, and anomaly detection. Multi-source credibility fusion formula: , Knowledge Entries The final credibility. Support knowledge entries The total number of sources of evidence. : No. The fusion weight of each source satisfies . : No. Each source for the entry The matching score. : No. Quality weights for each source. : No. The time-degradation coefficient of each source. : No. Evidence from different sources is often outdated.

[0022] The Condition Assessment and Anomaly Detection module is responsible for health assessment and anomaly early warning of equipment within the watershed. It calculates causally consistent health assessment results by intervening exogenous variables as a baseline (e.g., assuming constant upstream reservoir flow or fixed ambient temperature), avoiding confusion between causality and correlation. The residual between the expected value obtained from counterfactual prediction (assuming no anomaly occurs) and the actual observed value serves as the basis for anomaly identification. For remaining lifetime estimation, this module outputs the confidence interval of the remaining lifetime, providing a quantitative indicator for maintenance decisions. Graph spatiotemporal anomaly identification is performed on the asset coupling graph, using graph neural networks and other techniques to identify the abnormal state of individual equipment or equipment groups and its temporal and spatial propagation. Finally, alarm priority fusion prioritizes alarms from different sources and with different severity levels, avoiding alarm overload. Causally consistent health assessment results are calculated based on the deviation between the expected value and the actual observed value when exogenous variables are fixed as a baseline, providing a deeper understanding of the root causes of anomalies. The difference between counterfactual prediction and actual observation is used as a residual for anomaly identification; this residual represents the degree to which the equipment state deviates from normal behavior under given external conditions. The remaining lifetime estimate employs a random failure sampling method (SurvivalSampler) to obtain a distribution, and performs quantile consistency calibration based on an independent calibration set to obtain a controllable confidence interval with coverage (e.g., 90% of actual failure times fall within the predicted interval), enhancing the reliability of the prediction. For asset-coupled adjacency relationships, a time series model and a graph convolutional model are jointly used for updates to predict the asset's next-moment state, and the strength of the deviation between the predicted and actual values ​​is used as an anomaly score. This anomaly score can be weighted and fused with the business impact level to generate the final dispatch priority. Causal consistency health index threshold: Normal range: Focus area: Abnormal range: Danger zone: Counterfactual residual threshold: residual mean offset: Continuous overthreshold time: Number of concurrent abnormal channels: →Upgraded to a high-confidence alarm; remaining lifetime confidence interval parameter: number of samples. , inactivation ratio calibrate target coverage If the measured coverage deviates → Trigger recalibration, graph spatiotemporal anomaly detection parameters: adjacency weights ( (Node spatial distance), anomaly propagation threshold: The formula for combining node anomaly scores to determine a valid propagation path is as follows: , Health index deviation score Counterfactual residual score Spatiotemporal residual score.

[0023] The condition assessment and anomaly detection module takes multimodal data of watershed-level equipment and operating conditions as input, focusing on health assessment, remaining life interval prediction, time series and image anomaly detection, and multidimensional alarm fusion and priority calculation. Its goal is to stably output actionable handling suggestions and dispatching basis under unstable hydrological and complex scheduling backgrounds. The collected features cover vibration, acoustic signature, oil, temperature rise, load, start / stop, seepage pressure, and thermal imaging channels. Cleaning, normalization, and stability checks are performed at the edges and on both sides of the platform: short gaps are filled forward or interpolated; continuous missing segments are labeled as low-quality for weight reduction; range out-of-range and noise spikes are removed or isolated, and spectral and sliding window statistics are used to suppress false anomalies; channels with zero-point drift and systematic offset are biased based on benchmark points or redundant points; high-jitter time series are smoothed using low-pass or moving average; channels with different dimensions are subjected to linear normalization, zero-mean unit variance standardization, or robust scaling based on median-interquartile range to enter a unified modeling space. The platform maintains a quality tag library, categorizing data into three levels: high-quality (available), medium-quality (downgraded), and low-quality (isolated). The stability test results are written back to the asset ledger and knowledge graph to update the health tags of sensor points. To ensure data can serve as a basis for long-term health assessment, the system simultaneously examines trend stability, variance stability, physical correlation of redundant points, and cross-modal clock consistency over time. Correction or quality degradation is triggered if long-term correlation deviations or timestamp drift exceed thresholds. To prevent external hydrological and meteorological disturbances from masking true equipment degradation, a structural causal model is introduced to calculate the causal consistency health index. After scenario deconstruction, a graph structure is built using stations as basic units. Nodes include weather, inflow, water level, load, temperature rise, vibration, and start / stop. Upstream gate opening and pump speed are used as control nodes to distinguish between controlled and natural changes. Directed edges follow the priori link of weather → inflow → water level → load → temperature rise, vibration, and allow upstream control to simultaneously affect water level and load while prohibiting reverse pointing from temperature rise or vibration to exogenous nodes. Multi-source timestamps are aligned to a unified timeline. Only segments that pass cleaning and stability testing are included in causal identification. The sliding window length and step size are adjusted based on the dynamic response of the equipment to explicitly characterize the hysteresis effect of load on temperature rise and vibration. Structural learning adopts a parallel strategy of prior-driven and identifiability checks. Without damaging the main frame, gate opening and pump speed are used as instrumental variables to block hybrid paths. Robustness of edge direction and strength is ensured through comparative fitting between stable and disturbed periods, structural migration verification at adjacent sites, and error replay verification.Based on this, the system implements "benchmark intervention" for weather, inflow, and water level. It uses a conventional water level band derived from seasonal steady-state statistics and ledger topology as an exogenous benchmark to generate expected trajectories for load, temperature rise, and vibration while maintaining current start-up, shutdown, and scheduling. The expected trajectories are compared with observations moment-by-moment to obtain relative load offsets, relative temperature rise offsets, and relative vibration offsets. Modal contributions are further decomposed and aggregated into a single-value health index according to preset or adaptively learned weights for the equipment group. When significant anomalies occur in key channels (such as a sharp increase in temperature rise rate), the index calculation automatically increases the weight of the corresponding channel to avoid averaging masking. The index also includes source quality markers and interval reliability markers to adjust the influence in subsequent fusion. To reduce misjudgments, the system employs a combination of constraints at the index interpretation level, including external disturbance stripping, control differentiation, propagation path consistency, multimodal verification, historical steady-state comparison, and threshold grayscale bands. An event is only classified as abnormal when all three conditions—duration, number of consecutive occurrences, and amplitude—are simultaneously met. Differential handling is provided for typical boundary scenarios such as peak flood inflows, upstream gate anomalies, and planned start-stop operations: the index is downweighted during external disturbance periods and validated after the disturbance subsides; during consistent propagation from upstream to downstream, it is prioritized as an externally driven event; and start-stop operations and load step windows utilize separate exclusion strategies, retaining only persistent offsets for evaluation. In addition to the causal consistency health index, counterfactual analysis is used to improve sensitivity to endogenous faults. The system freezes exogenous variables at audited baseline levels, keeping control quantities and start / stop states unchanged. A deep time-series prediction model generates theoretical expectations for temperature rise and vibration under the assumption of "no external disturbances but preservation of current control." The model typically employs a long short-term memory network or a temporal convolutional network capable of characterizing time-series dependencies and hysteresis effects. Inputs include the baseline exogenous sequence, load, start / stop status, and operating mode; output is the expected trajectory of the target channel. The difference between actual observations and expectations is the counterfactual residual, representing endogenous biases that cannot be explained by external factors. In scenarios such as bearing wear, insufficient lubrication, and abnormal cooling, the residuals manifest as the coupling of increased vibration characteristic band energy and slow temperature rise, the concurrence of increased temperature rise rate and enhanced micro-impact, and a step-like, persistent difference caused by a slower rate of temperature fall under the same load. The system also labels residuals with data quality and model applicability, and uses a three-threshold joint logic control based on duration, number of consecutive occurrences, and deviation magnitude to escalate alarms. When both temperature rise and vibration show positive residuals within a reasonable lag, the anomaly confidence level is significantly increased. If the residual is only a brief single-channel shift that quickly falls back, it is marked as an observation. Finally, the counterfactual residuals and the causal consistency health index are fused and ranked together to maintain robustness while identifying weak endogeneous risks at an early stage. The prediction of remaining lifespan is based on time-series deep learning, capturing a hybrid pattern of periodic and sudden events through multi-scale inputs.After data cleaning, normalization, and stability testing, it enters the model center for training. The model population mainly uses LSTM or TCN, combining historical operating segments with the latest real-time data, covering key channels such as load, vibration, temperature rise, start-up / shutdown, and operating modes. Training adopts a parallel strategy of site-based and population-based approaches, saving time ranges, feature dictionaries, loss convergence, and validation performance, and finally deploying it online as a containerized inference service. Online inference generates point estimates of lifetime at the current moment using a sliding window, while simultaneously enabling random inactivation sampling to perform multiple forward passes on the same input segment to approximate the lifetime distribution. The number of samplings, inactivation ratio, and random seed are fixed before deployment and incorporated into version control. In extreme operating conditions, robust binning and truncation are used for outlier samples to avoid a few anomalies dominating the distribution. To ensure the statistical significance of the intervals, the system performs quantile calibration using an independent calibration set strictly isolated from training, mapping the original distribution to lifetime intervals under the target coverage. The calibration set covers seasonal, load levels, and typical operating condition differences. Calibration parameters and applicable ranges are versioned and stored in the database. If on-site sampling finds that the coverage deviates, recalibration and gray-scale release are triggered until it returns to the target interval. The output consists of point estimates, calibrated upper and lower bounds of the interval, and confidence labels. Confidence labels record data quality, model version, sampling configuration, and calibration mapping version. Interval boundaries employ time continuity constraints and extreme value suppression to mitigate short-window jumps while preserving audit trails. To avoid exogenous disturbances misleading lifetime assessments, the causal consistency health index and counterfactual residuals serve as both auxiliary inputs to the model and post-event consistency checks. If apparent lifetime contraction occurs due to flood peaks or backflow, the system uses rules to suppress false alarms. On the maintenance side, risk levels are set by superimposing the lower bound of the interval with the process safety boundary, and priority is fine-tuned based on interval trends and counterfactual residual evolution. When the interval span is large, crossing thresholds, the labeling uncertainty is high, requiring weighted review based on the business impact level. When coverage sampling deviates from the target, a more conservative strategy can be selected by relying on the backtracking version of the confidence labels. Anomaly propagation identification under coupling effects relies on spatiotemporal modeling of the graph. The system automatically generates a graph structure based on asset ledgers and geographic topology. Nodes cover pump sets, motors, gate hoists, power distribution cabinets, valves, and various sensor points. Node metadata retains equipment type, operating mode, geographical location, site affiliation, upstream and downstream roles, health tags, and data quality tags. Vibration, temperature rise, oil levels, and acoustic signatures of the same equipment are expressed as subordinate child nodes or attributes for cross-modal alignment. Edge types describe electrical coupling, hydraulic coupling, water coupling, and control links, recording direction, strength, timeliness, interlock markers, and reliability scores. The strength and effectiveness are dynamically updated in time slices during situations such as valve closure, loop maintenance isolation, or peak water inflow. The system generates a graph snapshot sequence with a fixed step size, uses window aggregation and interpolation to align channels with different sampling frequencies to a unified time grid, and explicitly adds hysteresis-derived channels to the node features for quantities with physical lag.The inference mechanism employs an alternating "space-time-space" or "time-space-time" structure, using graph convolution or graph attention for neighborhood aggregation, and gated recurrent or temporal convolution for multi-scale temporal pattern capture. Causal consistency health index and counterfactual residuals are used as independent channels in attention weight calculation to suppress exogenous disturbances and amplify endogenous offsets. Control edges are input separately during propagation to distinguish between controlled changes and anomalous propagation. At each time step, the model generates the expectation of key channels and aligns it with observations to obtain graph spatiotemporal residuals. Subsequently, anomalous scores for nodes and edges are calculated: for nodes, the comprehensive health index offset, counterfactual residuals, graph spatiotemporal residual strength, and business impact level are considered; for edges, abnormal coupling strength, residual propagation that cannot be explained by the control chain, and edge credibility scores are examined. When adjacent nodes deviate sequentially before and after their occurrence, the system backtracks along directed edges, calculates path consistency, and outputs the most probable propagation path by jointly ranking multiple candidates based on consistency, credibility scores, and business impact relevance. Leading clues are provided for combinations of low-amplitude, continuous upstream offsets and gradually increasing weak downstream responses. To suppress transient noise, anomaly scoring employs window aggregation with three thresholds: duration, cumulative intensity, and number of consecutive occurrences. Planned start / stop and maintenance isolation windows are suppressed through control edges and work order constraints. All anomalies and paths are accompanied by evidentiary fingerprints involving nodes, edges, time windows, key feature contributions, and excluded control interpretations. When confidence is insufficient, active learning and verification are automatically initiated to accumulate incremental samples and improve generalization. In the alarm fusion and priority ranking stages, the degree of deviation of the causal consistency health index, the degree to which remaining lifetime approaches the threshold, the magnitude of counterfactual residuals, the intensity of graph spatiotemporal residuals, and the level of business impact are aligned and scaled uniformly at a completely consistent time window and object granularity, preserving both original and standardized dual-track evidence. Robust pruning and soft saturation are applied to outliers, and evidence is downweighted or isolated according to quality labels to avoid amplifying and misleading low-quality evidence. The fusion weighting system incorporates two mechanisms: a configuration path and a learning path. Configuration weights are set and approved by security and operations teams in the early stages or during periods of data sparseness. Learning weights, after sufficient sample size, fit historical closed-loop data to determine "which factors best predict real events and significant impacts." Both are versioned and have time-limited effectiveness. When the two conflict, "security first" is prioritized to increase sensitivity to the level of business impact and the approach to lifespan thresholds. A consistency check is performed before fusion: if the health index and counterfactual residuals are significant, but the graph spatiotemporal residuals are weak, the bias is towards localized endogenous faults; if the graph spatiotemporal residuals are significant and spread along directed edges, but the local offset is limited, the bias is towards propagational risks. The check results do not overwrite the original values ​​but are fed back to the weight scheduling module to dynamically adjust the weight allocation for the current fusion. Control switching, planned start / stop, and maintenance isolation time periods are suppressed by rule triggers to prevent planned disturbances from being amplified.The merged system outputs a single priority score, while imposing continuity constraints and hysteresis control over time. Three thresholds—duration, cumulative intensity, and number of consecutive occurrences—jointly determine whether to escalate to an event-level event and prevent high-frequency repetitive dispatches. The score is then mapped to four handling levels: Emergency, Warning, Attention, and Observation. Within each level, the system compares business impact, proximity to the lower bound of lifetime, path consistency and residual strength, and the combined magnitude of health index and counterfactual residuals. When all factors are close, data quality and evidence diversity are used as the tiebreaker. Generated dispatch candidates include device identifier, handling level, priority score, evidence composition, suggested handling time limit, and qualification prompts, and undergo dispatchability verification under resource and interlock constraints. If dispatch is temporarily unavailable, the item is retained with the reason for the blockage and the earliest executable time, while simultaneously triggering spare parts pre-allocation and resource coordination suggestions. Each score includes an explanation package explaining "why it was ranked this position," and intermediate results and version information are archived to support audit playback. An active learning loop is introduced to continuously reduce false alarms and improve generalization. When high-uncertainty situations arise, such as scores approaching thresholds, multimodal signal conflicts, or prediction interval spans exceeding upper limits, the system automatically triggers manual review. The system pushes the time-period data, scores, and evidence chain to operations and maintenance personnel and experts for confirmation. The review conclusions are fed back into the training set for incremental training in subsequent cycles, and the model's decision boundary gradually converges with the accumulation of samples. The triggering strategy is based on uncertainty metrics and multi-branch difference thresholds, prioritizing samples most sensitive to performance improvements to save on manual costs. With closed-loop operation, the discrimination capability of boundary samples is enhanced, the false alarm rate is reduced, and the model's transfer stability under new operating conditions and on new equipment is improved. In terms of engineering implementation, data preprocessing is completed collaboratively by the edge appliance and the platform to ensure consistency and integrity before data entry. Model training relies on the platform's AI model center to complete large-scale time-series and graph structure training on the graphics processor, and is deployed in a containerized form as an online inference service. Lifetime interval calibration uses a fixed calibration set and a fixed random seed to ensure repeatability. The graph spatiotemporal structure is automatically generated by the asset ledger and geographical topology and is incrementally refreshed with on-site changes. All models, rules, and weights are versioned and managed with effective time, and support canary releases and rapid rollbacks. Evidence fingerprints, explanatory texts, and intermediate results of abnormal events, propagation paths, and fusion sorting are uniformly archived to achieve frame-by-frame playback and full-link auditing. Online operation shows that the system can identify early degradation with an average lead time of more than five days, the lifetime interval remains within a stable range under the target 90% coverage, the F1 score for propagation anomaly identification meets the target requirements, and the false alarm rate is further significantly reduced after the introduction of active learning. Based on the joint output of causal consistency health index, counterfactual residuals, graph spatiotemporal anomalies, and calibrated lifespan intervals, the module translates predictive maintenance into concrete resource scheduling and on-site execution, forming a closed-loop system from data cleaning—causal modeling—counterfactual inference—lifespan interval—graph spatiotemporal detection—fusion ranking—active learning, continuously providing robust and reliable health assessment and dispatch priority basis in complex hydrological and scheduling environments.Formula for a causal consistency health index: , : The causal consistency health index of time-based devices. Total number of monitoring dimensions. : No. The weights of each monitoring dimension satisfy the following: . : No. A robust normalization function for each dimension. : Time of the first Actual observations in each dimension. : The expected observed value when the exogenous variable is fixed as the baseline. : The set of exogenous variables. The baseline value of the exogenous variable. Counterfactual residual formula: , : The counterfactual residual of time. : Key observation indicators of time-lapse equipment. When the exogenous variable is fixed as the baseline, The predicted observations at time points are predicted by either a structural causal model (SCM) or a causal forest model. : Set of exogenous variables and benchmark values.

[0024] The maintenance planning and intelligent scheduling module is the brain of the system, responsible for formulating and optimizing maintenance plans. Within a multi-objective optimization framework, it jointly optimizes the schedule, cost, downtime risk, and carbon emission intensity to find the optimal maintenance solution. To address external disturbances such as hydrological uncertainty, opportunity constraints (limiting the probability of constraint violation to no more than a set threshold) and Conditional Value at Risk (CVaR) constraints are introduced to ensure the robustness of the plan. In the rolling time domain, the published plan is reconstructed in a minimal-disturbance manner; that is, when new situations arise, modifications to the existing plan are minimized to maintain the stability of the plan. The optimization objectives simultaneously consider the total schedule (minimization), comprehensive cost (minimization, including maintenance, downtime, spare parts, etc.), the conditional expectation of downtime risk (minimization, such as the expected value of downtime loss in the worst-case scenario), and carbon emission intensity (minimization, such as carbon emissions from equipment operation and transportation). Opportunity constraints are set for hydrological uncertainty to limit the probability of constraint violation to no more than a set threshold (e.g., at a 95% confidence level, the water level within the maintenance window cannot exceed a certain limit). Furthermore, using feasible initial solutions trained from historical best solutions as a warm start for the solver can accelerate the optimization process and avoid starting the search from scratch. Within the rolling time domain, reconstruction and deployment are performed by minimizing the predetermined work order time offset, ensuring that even in the event of unforeseen circumstances, the impact on scheduled maintenance tasks is minimized. Multi-objective optimization weight template (normal water period): Work order weighting Cost weight Risk weight carbon emission weight Opportunity constraints and risk thresholds: upper limit of probability of violation (i.e., 95% confidence that there will be no default), conditional value at risk level Rolling disturbance control parameters: trust region radius (Relative time window length), plan offset tolerance Publish and re-solve trigger conditions: If the rate of change of the hard constraint If the health threshold crosses the risk range, a rolling reconstruction will be triggered.

[0025] The goal of the maintenance planning and intelligent scheduling module is to achieve a holistic balance between safety and compliance, water supply and flood control, energy consumption and carbon constraints, resource capacity, and equipment health risks. It generates executable, auditable, and rollback-capable maintenance and operation plans, and coordinates with status assessment, digital twins, and knowledge graphs within a rolling window. The module operates on a cycle of requirement generation, solution orchestration, constraint solving, simulation verification, release execution, and closed-loop learning. It covers both medium- and long-term annual maintenance master plans and weekly and daily rolling adjustments, ensuring the robustness and availability of the plan under conditions of sudden external disturbances and resource changes. The input side consists of four core facts. The first category is equipment-side facts, including causal consistency health indices, counterfactual residuals, remaining life intervals and their quantile coverage, graph spatiotemporal anomaly paths and propagation strength, as well as hierarchical topology, interlocking relationships, redundancy capabilities, criticality, and compliance clause bindings provided by the asset ledger. The second category is resource-side facts, covering team calendars, personnel qualifications and annual review validity periods, cross-certification matrices, outsourcing vendor SLAs, spare parts inventory and ordering status, tools and special aircraft, vehicle and hoisting resources, temporary navigation closures or traffic constraints, and safety permit approval sequences. The third category is process and environmental-side facts, including hydrological and meteorological forecasts, inflow boundaries and scheduling target curves, load corridors and peak / valley electricity prices, carbon emission factors, environmentally sensitive periods, upstream and downstream linkage windows, geographical topological accessibility, and on-site construction enclosure sequences. The fourth category is strategies and constraints, including hard safety boundaries, regulatory compliance requirements, enterprise standards, outage concurrency limits, minimum available redundancy for critical equipment, cross-site linkage rules, service levels and breach of contract costs, energy consumption and carbon constraint weights, and seasonal supply priority. All inputs are uniformly named and versioned in the knowledge graph, aligned temporally to the scrolling window, and spatially to the site and equipment levels. Quality tags are used to drive weight reduction or suppression in subsequent solutions. Demand generation transforms health risks and business requirements into plan candidates. Based on the degree of approach to the lower bound of the remaining lifespan and the persistence of gray-scale deviations in the health index, "risk-driven" maintenance demands are generated. Combined with preventative periodic demands from regulations and statutory inspection requirements, "system-driven" work orders are formed. Work orders for handling sudden failures and linkage anomalies are then treated as "event-driven" emergency demands. These three types of demands are merged and deduplicated at a unified object granularity. Urgency is calculated comprehensively based on equipment criticality, redundancy capacity, business impact, and cross-modal evidence strength, generating candidate time windows and pre- and post-delay dependencies. If multiple demands exist for the same equipment, the module attempts to merge and package them within an acceptable downtime, or split them into multiple short-term shutdowns when redundancy is insufficient or risk is high, reducing peak disturbances to water supply and flood control. The scheme orchestration maps candidate demands to schedulable work segments. Each maintenance requirement generates work steps, ticket links, isolation points and interlocking conditions, a list of special tools and equipment, and safety control measures under the constraints of the knowledge graph. It also derives the lower limit of personnel qualification combinations, minimum and recommended team size, estimated working time intervals, quality re-inspection and trial operation time, and post-recovery performance observation period.For operations requiring shutdown or load reduction, the module and scheduling target curve are aligned with twin sensitivity to calculate feasible load corridors, allowable start-stop sequences, and maximum concurrency for each operation. If spare parts are scarce or their transit times are close, the system generates pre-allocation and expediting strategies, providing a delay risk score and alternative parts compatibility check results. If alternatives are not feasible and the risk is high, the operation is marked as a special segment requiring "supply protection," and can only be deployed during periods of high redundancy and low load. Constraint solving employs hierarchical multi-objective optimization. The outer layer uses a combination of mixed-integer programming and constraint programming for modeling, while the inner layer uses heuristics and neighborhood search for acceleration. The objective function is weighted hierarchically based on priorities such as safety, supply, compliance, energy consumption and carbon emissions, travel and waiting costs, cross-site coordination costs, and handover and start-stop losses. Hard constraints include interlocking to prohibit concurrency, minimum available capacity, personnel qualifications and validity periods, pre-approval of tickets, isolation point coverage, geographical accessibility and traffic windows, navigation closure and enclosure periods, operational environment boundaries, upstream water release and downstream water level thresholds, equipment cooling and reset waiting, minimum start-up and shutdown intervals, and trial operation verification and observation period coverage. Soft constraints include penalties for deviations from the scheduling target curve, peak energy consumption penalties, carbon emission intensity penalties, insufficient cross-station peak shifting penalties, batch grouping deviations, cross-shift handover times, and nighttime disturbance penalties. To avoid solver oscillations caused by discrete start-up and shutdown and nonlinear load response, the module employs a trust region rolling strategy and gentle relaxation to limit changes within each window, and introduces a differentiable twin linearized sensitivity as a cost slope hint to improve convergence speed and solution executability. Resource planning occurs concurrently with the solution process, constraining peak personnel, vehicle, and dedicated aircraft capacity using calendar-driven capability curves and calculating the number of effective combinations through a cross-certification matrix. When key qualifications are insufficient in the short term, the module triggers cross-site support and outsourcing vendor recall processes, calculating dispatch mileage and arrival time, and incorporating travel costs and response time penalties into the objective function. On the spare parts side, material availability is verified through inventory, order, and safety stock windows. If long-lead-time parts exist and their lifespan lower bound is approaching, the system automatically generates a comparison between "advance material preparation and conservative operation," selecting the path with lower total cost while maintaining the safety margin. For tool conflicts, the module attempts to adjust start time and concurrency; if unsuccessful, it triggers an alternative equipment list and leasing channel. All resource conflict-resolving processes are recorded as an auditable explanatory chain. Two-way linkage with the digital twin runs throughout the solution process. After candidate solutions are generated, a differentiable twin is invoked to perform forward simulation under the current boundary conditions, producing water supply pressure, flood control capacity, energy consumption and carbon emissions, boundary trigger probability and minimum safe distance, and calculating the local sensitivity of control variables to key indicators. If the simulation shows that certain concurrent shutdown schemes lead to an increase in the probability of exceeding constraints or a narrowing of the safety margin, the system applies soft and hard constraints to the corresponding operations, tightens the allowable time period or reduces the concurrency. If energy consumption and carbon emissions differ significantly under the same risk, the objective function is dynamically weighted according to the sensitivity slope, prioritizing low-energy-consumption and low-carbon schemes.For constraints with strong nonlinearity, the module labels the linearized failure region and forces small-step exploration or conservative evaluation using a surrogate model within this region to avoid gradient misleading. A rolling recalculation mechanism ensures the continuous availability of the plan under real-world disturbances. The module updates observations and forecasts hourly or on an event-triggered basis, inputting new hydrological data, load, equipment health, and resource status, and comparing changes in the feasible region of the plan. When hard constraints or resource unavailability render the current solution infeasible, the system performs a minimum disturbance reconstruction, prioritizing the movement of low-urgency operations, merging packaged operations, exchanging teams with similar qualifications, fine-tuning start / stop step sizes within the trust region, and providing an improved solution in a short time through parallel multi-solution solving and solution pool management. If on-site feedback indicates that a task is completed early or delayed, the module automatically reclaims the released time and resources, repackages them, and inserts them in a suboptimal manner to avoid wasting time. If a sudden alarm enters the "emergency level," the system reorders tasks using a preemptive strategy, forcibly prioritizing them according to safety boundaries and supply guarantee weights, while simultaneously triggering linked verification and rapid twin verification at related sites to ensure that the risk of propagation is addressed proactively. During the deployment and execution phase, the module slices the solution into work assignment instructions, ticket readiness, and isolation schemes, generating travel routes and entry / exit controls according to the geographical location and time sequence of the task and connecting them to the access control system; the ticket link is constrained by the knowledge graph. The system automatically verifies the completeness of approvals and the interlocking unlocking sequence; any unmet prerequisites are marked as incomplete and prohibited from issuance. On-site mobile terminals receive instructions and lists, input start, pause, resumption, and completion times, and upload images and instrument data. The system updates the process parameter library based on actual backfilling time, number of starts and stops, isolation switching duration, and trial operation stability, for retraining the construction cycle model and continuous improvement of time estimation. Compliance and safety governance are integrated throughout the entire process. Before being added to the database, the module performs consistency and compliance checks on all planned items to ensure that the validity period of tickets covers the work period and that the qualification combination meets regulations. The system implements dual-condition interlock release, prohibits single-person operation under high-risk conditions, ensures that cross-station concurrent shutdowns do not trigger the redundancy limit of the same pump group, and provides full coverage of electrical and hydraulic isolation points with the signature of the reviewer. For items that trigger prohibitive constraints, the system provides a minimum-cost remediation path, including downgrading to reference, delaying entry into the database, transferring to the pending review area, or adding supporting materials. Each decision generates an interpretation path, linking the rule number, triggering field, evidence source, and remediation decision, supporting playback and auditing at any time. Multi-objective trade-offs are reflected in the objective function and threshold governance. It provides security, supply assurance, and energy security. The system employs a weighted approach to energy consumption, carbon emissions, cost, and reliability, supporting strategy templates for flood season, dry season, and peak summer demand. During flood season and critical supply periods, the weight of safety and supply assurance is significantly increased, with energy consumption and carbon optimization giving way to redundancy assurance. During periods of normal water levels and low load, marginal targets for energy consumption and carbon emissions are raised, encouraging the migration of non-critical operations to low-carbon units and off-peak electricity pricing periods. Weights and thresholds are version-based, allowing for phased rollouts and rapid rollbacks to complete strategy switching without disrupting continuity. Historical versions and impact lists are permanently stored. Geographic and route constraints are centered on accessibility and traffic windows.Using GIS data from inside and outside the station and temporary lockdown information, the system calculates round-trip time and on-site movement paths, considering factors such as fencing, hoisting radius, danger zone isolation, and conflicts with shared passageways to avoid simultaneous large-scale hoisting and densely populated operations. For cross-station linkage windows, the system provides a full-cost comparison of serial and parallel solutions and assesses the impact on upstream and downstream water levels and water supply stability, prioritizing the solution with the least disturbance to key indicators. Cost and carbon accounting are performed upfront at the planning level. Based on the energy consumption curves of equipment and processes, time-of-use electricity prices, and carbon factors, the system calculates estimated energy consumption and carbon emissions for the planning period, combining hoisting and vehicle mileage, outsourcing SLAs, and overtime policies to calculate direct costs and opportunity costs. When two solutions are equivalent in terms of safety and supply assurance, the solution with lower unit risk reduction cost and better carbon intensity is selected. If the cost is lower but the carbon intensity increases significantly and exceeds the quarterly quota, a hard carbon budget constraint is triggered, or a low-carbon substitution window is used. The knowledge loop is driven by proactive learning and indicators. The actual start-up and shutdown sequence, actual working hours, ticket approval rate, isolation switchover time, trial operation failure rate, energy consumption and carbon deviation, water supply and water level exceeding time limits, and false alarm and missed alarm review conclusions after the plan is executed are uniformly fed back. The module performs periodic recalibration on health thresholds, priority weights, and process time models, and establishes monthly and quarterly dashboards for management purposes, including plan achievement rate, supply guarantee achievement rate, number of events exceeding constraints, energy consumption achievement rate, carbon achievement rate, average recovery time, and the rate of decrease in false alarm rate, driving continuous strategy improvement. Anomaly and emergency response strategies are pre-configured within the plan. For critical equipment approaching the lower bound of RUL and with spare parts shortages, the system generates a dual-path strategy package of operational derating, hot standby switching, and short-term maintenance. It also conducts pre-simulation drills for several flood season and peak flood scenarios using twins, forming executable backoff curves and minimum redundancy limits. For propagation anomalies caused by upstream gate jamming or downstream backflow, the module pre-configures cross-station coordinated maintenance and monitoring to ensure local isolation or flow restriction is completed before the intensity of the abnormal path increases. The relevant work permit links and qualifications are aggregated and approved in one go during plan preparation. In terms of system governance, all plans, solutions, versions, constraints, and interpretation links have globally unique identifiers and effective time ranges, supporting time replay and cross-version comparison. External interfaces and reports are consistently stable under namespaces and compatible views. Major structural changes are implemented using shadow releases and canary rollouts to ensure consistent queries from external systems before and after the switch. The permission model provides fine-grained authorization based on roles and domains, and key actions are reviewed by two people and the operation watermark ensures that changes are traceable; when collaborating across units, the granularity of the plan and the evidence summary are shared through a secure aggregation service, without exposing the privacy details of personnel and assets.In terms of engineering implementation, a containerized microservice architecture is adopted, with the solver kernel and simulation kernel deployed separately. The solution pool and scene cache support parallel multi-solution competition. Event bus-driven rolling recalculation completes minimal disturbance reconstruction within seconds to minutes. The edge appliance is responsible for on-site data cleaning and quality labeling, while the platform-level model center provides version governance for solver parameters, sensitivity, thresholds, and weights. The continuous integration pipeline performs static checks and regression tests on strategy and parameter changes to avoid deployment regression. To improve stability, the solver warms up and restarts from the most recent feasible solution after abnormal exit or timeout, and degenerates to a conservative strategy template when necessary, ensuring that a feasible schedule that meets safety red lines and minimum supply constraints can still be provided even in the worst case. Through the above mechanisms, the maintenance plan and intelligent scheduling module connect health risks, resource capabilities, physical boundaries, and uncertainties with a unified optimization and simulation language, enabling the plan to reduce costs and increase efficiency under normal operating conditions, while maintaining safety and resilience under extreme scenarios. Tight coupling with state assessment, digital twins, and knowledge graphs ensures that every operation, from requirement generation to execution closure, is traceable, measurable, and replayable, ultimately achieving the operational goals of less downtime, shorter downtime, stable supply, low carbon emissions, and auditability. Multi-objective optimization formula. , The overall objective function value represents the total cost; the smaller the value, the better the plan. : Scheduling decision variables. : Work option weight. Total construction period. Cost weighting. Total cost. Risk weight. Conditional Value at Risk (VaR). : Confidence level of CVaR. :plan In uncertainty The risk of loss. Uncertain variables. Carbon emission weights. Total planned carbon emissions.

[0026] The digital twin simulation and impact assessment module provides simulation verification capabilities for maintenance decisions. Differentiable calibration of equipment and hydrodynamic processes allows for parameter optimization of the simulation model via backpropagation, ensuring consistency between the model and actual observation data without violating physical laws. Weighted fusion of multiple fidelity models is performed based on uncertainties (such as sensor errors and model parameter uncertainties), combining the advantages of models with different levels of accuracy to provide accurate and computationally efficient simulation results. A neural operator agent is used for rapid scheme evaluation, accelerating the computation of complex physical models to the millisecond level and supporting real-time decision-making. External uncertainties (such as changes in weather conditions and fluctuations in hydrological conditions) are propagated to form a robustness score, quantifying the performance of maintenance schemes under different uncertainty conditions. Finally, sensitive boundaries and constraints are fed back to the scheduling solver to guide adjustments to the plan. During parameter calibration, both observation errors (the difference between model output and actual observations) and physical equation residuals (the degree of deviation from the physical laws within the model) are minimized to ensure that the model is consistent with field observations while adhering to physical laws. When fusing multi-fidelity models, a weighting strategy inversely proportional to uncertainty is adopted, assigning higher weights to models with lower uncertainty to obtain more reliable comprehensive predictions. By using neural operator agents to quickly predict key indicators (such as downtime and economic losses under maintenance plans) in milliseconds, the efficiency of plan evaluation is greatly improved. External uncertainties are sampled, and a robustness score is generated by comprehensively considering the expected value and volatility of the influencing indicators; a higher score indicates a more stable performance of the plan under uncertainty. Sensitive boundaries (such as which small changes in parameters significantly affect robustness) and constraints are fed back to the scheduling solver so that the scheduling module considers these risk factors during optimization. Differentiable twin calibration parameters: physical constraint weights. Parameter convergence threshold Calibration frequency: once every 24 hours, or when the observation bias is >5%. Multi-fidelity model weighting threshold: Variance ratio threshold for each layer of the model. →Lower-weighted, low-fidelity layer with the lowest fusion weight Stability window Used for variance sliding estimation, robustness score composition ratio: pass rate Performance convergence Boundary margin Scoring threshold: For a robust solution, For high-risk scenarios, neural operator proxy parameters include: learning rate. Batch size Early termination criterion: The verification error decreases by less than 0.1% for five consecutive rounds.

[0027] The digital twin simulation and impact assessment module is designed for rapid, reliable, and quantifiable assessment of operational status before and after maintenance. Its core is based on equipment-level twins and one-dimensional / two-dimensional hydrodynamic processes, outputting indicators such as water supply, flood control, energy consumption, risk, and carbon emission reduction. Sensitivity and constraints are then fed back into the scheduling solver, forming a closed loop of data-driven + physical constraints. It characterizes the efficiency curves and fault evolution of key equipment such as gate hoists and pump sets, and can also simulate channel capacity and energy loss in simplified hydrodynamic models. This allows for quantitative pre-simulation of risks and energy conservation before shutdowns or load reductions, making it particularly suitable for pressure testing in extreme hydrological conditions. Differentiable twins are the foundation for bidirectional linkage. On the equipment side, head-flow relationship, efficiency curves, and mechanical and electrical losses are expressed parametrically, using continuously differentiable basis functions or piecewise smooth approximations within the operating domain. On the process side, conservation equations, boundary fluxes, local losses, and topographic effects are assembled into a differentiable operator link. The discrete format uses smooth interpolation that is stable in both time and space, ensuring the differentiability of state updates to control variables and parameters. This results in an end-to-end differentiable mapping from control variables such as gate opening and pump speed to key indicators. The calibration phase employs a joint objective of "observational consistency + physical consistency": minimizing observational biases in flow rate, level, pressure, power, and temperature rise drives parameter convergence, while embedding conservation, boundary conditions, monotonicity, and stability into inviolable constraints. Equipment parameters are calibrated locally in stages, followed by joint fine-tuning of process parameters, with cross-validation performed at independent time intervals to avoid overfitting and scaling instability. Differentiable twins support automatic differentiation to obtain local sensitivity. At the operating point of candidate schemes, the system calculates the gradients and directional derivatives of indicators such as water supply pressure, flood discharge capacity, energy consumption, and equipment load relative to gate opening and pump speed, while simultaneously propagating partial derivatives from both the equipment and process sides to reflect true coupling. Before reinjection, gradients undergo scaling, numerical stability screening, and confidence assessment, retaining only high-confidence directions consistent with active constraints. Each rolling optimization limits the trust region radius to prevent discrete decisions from being misled by excessively strong gradients, ensuring stable progress in small steps. Hard boundaries, such as minimum navigable water level, maximum allowable water level rise, pump unit safe load, and gate hoist limits, are reinjected between calibration thresholds and buffer zones as hard constraints or highest-priority soft constraints. Optimizable continuous control variables provide local linear approximations and cost slopes, assisting the solver to converge faster under multi-objective conditions. Rolling operation follows a cycle of "calibration—simulation—sensitivity—reinjection" in coordination with scheduling. Each window first fine-tunes parameters using the latest observations to eliminate drift; then, it performs batch simulations of alternative shutdown / load reduction schemes and generates multi-dimensional indices, while simultaneously calculating gradients and boundary activity; the scheduler updates weights, tightens or relaxes relevant constraints, adjusts the trust region, and obtains a new solution with a warm start; the control variables of the new solution are fed back to the twin for rapid verification. If over-constraints or decreased robustness are found, the module provides reverse suggestions and step size limits until the solution is acceptable on both the physical and solution sides. Closed-loop recording of each reinjection version, threshold, gradient summary, and convergence status supports auditing and playback.To balance accuracy and real-time performance, the module employs a multi-fidelity federated twin. The device-level high-fidelity model focuses on local efficiency, power, and thermal load, exhibiting the highest sensitivity to detail but incurring significant computational overhead. The watershed-level medium-fidelity model simplifies the one-dimensional / two-dimensional hydrodynamic characterization of upstream inflow, downstream water level, and river network topology, excelling at global constraints and channel linkages. The low-fidelity layer provides trend judgments and conservative estimates using empirical rules and steady-state approximations, minimizing latency. The three layers are evaluated in parallel under a unified scenario, time base, and entity mapping. The outputs are dynamically fused using uncertainty weighting: weights are jointly determined by source credibility (historical residuals and drift rate), scenario matching (whether it falls within the high-confidence working domain of each layer), and real-time stability (short-window variance, anomaly sensitivity, and observation bias). If a threshold conflict occurs between layers, rapid recalibration and boundary verification are triggered first, followed by weight reduction or agent switching to maintain a physically reasonable range. The fused output is compared with the observation in each window. If the deviation exceeds the threshold, the calibration frequency of the relevant layers is increased, and the high-sensitivity boundary is directly reinjected into the scheduling to ensure that uncertainty is transformed into consumable constraint and weight signals. Cross-unit / cross-basin scenarios adopt a federated approach to maintain twin consistency and protect privacy. Each participant locally maintains high / medium / low-fidelity models and parameters, which are evaluated in parallel under a unified scenario template. Only the statistics and uncertainty characterization of key indicators (such as cross-sectional water level quantiles, flood capacity confidence bands, equipment efficiency stability intervals, and sensitivity intervals) are shared. After secure aggregation and reduction in a dense state, the coordinating end only sees the aggregated distribution and confidence intervals, and based on this, the fusion weights, scenario weights, and boundary thresholds are unified and then broadcast back as parameters. Before federation starts, entity and time alignment is completed based on the joint ontology, and the start and end points and granularity of the rolling window are fixed. Periodically drifting probes monitor whether the output of any domain exceeds the global allowable range, triggering enhanced calibration of that domain or switching to a conservative agent to maintain cross-domain consistency and safe boundaries. To achieve millisecond-level evaluation during the initial selection and rolling recalculation of solutions, a neural operator agent is introduced as a rapid calculation layer. The agent takes boundary conditions, terrain parameters, and control variables as input, learning the mapping from boundary / control to state and operational indicators. A single forward pass can output signals for water supply pressure, flood control capacity, energy consumption, carbon emission reduction, stability, and sensitivity. Training data primarily consists of high / medium-fidelity model simulations with field segments for alignment correction, covering low / normal / high water levels, various start-stop strategies, and linkage methods. Training is divided into three phases: first, learning large-scale responses under global conservation; then, introducing fine-grained control and equipment efficiency; and finally, covering cross-site linkage and rolling disturbances. Online deployment employs versioning and a unified test baseline, with applicable domains clearly marked. If the confidence level decreases due to scenario shifts or boundary mutations, the baseline review ratio is automatically increased, the release step size is shortened, or a more conservative agent is used. During batch evaluation, the agent first quickly performs a coarse-grained selection and eliminates significantly infeasible solutions. For the "final round," a differentiable twin / medium-fidelity review is then scheduled, balancing timeliness and rigor. It incorporates uncertainty propagation and robust evaluation, with a closed loop consisting of scenario generation, batch simulation, statistical evaluation, and constraint back-injection.For input uncertainties related to inflow, rainfall, and upstream / downstream water levels, the system adaptively switches between Latin hypercube sampling and multinomial chaotic expansion based on time budget and analysis objectives: the former covers the high-dimensional input space with fewer samples, while the latter characterizes higher-order interactions and supports variance decomposition through orthogonal basis expansion. Batch simulations simultaneously drive neural operators and differentiable twins, outputting metrics with numerical convergence and stability labels, and removing outliers to avoid statistical bias. In the statistical phase, the average, peak-to-valley difference, and out-of-bounds duration of water supply pressure, the expected value and lower quantile of flood control capacity, the time integral and peak value of energy consumption, and the relative amount and fluctuation range of carbon emission reduction are calculated. Based on these, a robustness score is constructed, comprising a single-valued metric that integrates three parts—probability of compliance, performance convergence, and boundary margin—after weighting. The score and summary are written back to the scheduling side as input for multi-objective trade-offs. For highly sensitive boundaries, the module generates two types of backfeeds: a safety redline generates hard constraints with a minimum safety distance and effective time period; a tradeoff boundary generates soft constraints, providing a penalty slope, sensitivity range, and trust region radius, allowing the solver to compromise among multiple objectives and limiting the step size; for strongly nonlinear regions near the critical point, linearized failure markers and alternative approximation ranges are provided, prompting the invocation of a conservative proxy or mandatory twin verification. Each rolling window executes a backfeed-solve-verify-solidify short loop; when hard constraints lead to infeasibility, minimum perturbation reconstruction is triggered; when soft constraints change the tradeoffs, control variables are updated within the trust region; if the verification shows that the over-constraint probability has not decreased or the robustness score has not increased, the trust region is shrunk and rolled back until the improvement is successful. The interface is consistent with the upstream module. The asset ledger and knowledge graph provide topology, boundaries, interlocks, qualifications, and compliance terms, driving differentiable twin assembly and boundary setting. The causal consistency health index and counterfactual residuals from the state assessment module serve as distinguishing variables between health status and controlled changes, inputting into the twin and agent and used for result interpretation. Maintenance planning and intelligent scheduling obtain hard / soft constraints, gradient hints, robustness scores, and boundary activity from this module, while simultaneously sending back control quantity slices of new solutions for review. All interactive objects are versioned, recording parameters, mappings, thresholds, random seeds, and applicable domains to support playback and comparison. The project adopts a containerized, layered approach: data access interfaces with hydrological, meteorological, topographic, asset, and monitoring data, achieving spatiotemporal alignment and scale normalization; a twin kernel assembles differentiable operators at the equipment and hydrodynamic levels into an end-to-end graph, providing interfaces for automatic differentiation, calibration, and sensitivity; a proxy service handles online inference and domain determination for neural operators; an uncertainty engine is responsible for scene generation, parallel simulation, and statistical evaluation; an injection service publishes hard / soft constraints, gradients, and trust domain parameters to the scheduler solver using structured messages; and an audit and governance layer records version, evaluation, residual, drift, and federated aggregation metadata. During cross-domain collaboration, a secure aggregation service performs encrypted reduction, preventing the original data from being stored locally, and logs and reports are managed using access control and watermarking.Through the above design, the digital twin simulation and impact assessment module provides differentiability sensitivity, robust statistics, and federated consistency while maintaining physical consistency. In batch comparison and rolling recalculation, the neural operator agent significantly reduces latency while maintaining rigor in "benchmark sampling." Uncertainty propagation transforms risk into consumable constraints and weights. Linkage with the small-step trust region of the scheduling solver avoids solution oscillations and enhances executability. Ultimately, the module enables maintenance plans to achieve an auditable balance between safety red lines, water supply security, flood control capacity, energy consumption, and carbon targets, and provides managers with replayable, interpretable, and implementable decision-making support under extreme conditions. The joint calibration formula for the differentiable twin is: , The parameter set of the twin model. : The actual on-site observation value at that moment. Model parameters are hour, Predicted value at time, L2 norm squared, Physical constraint weights. Physical constraint function, multifidelity uncertainty weighted fusion formula: , The final predicted value after multi-fidelity fusion. : Fidelity hierarchy index : No. The predicted mean of the layer model, : No. The prediction variance of the layer model Regularization constant, neural operator agent training formula: , The parameter set of neural operators. Neural operator model : No. Boundary conditions for each training sample. : No. Control variables for each training sample : No. High-fidelity solutions (labels) for each sample. : Training sample index.

[0028] The Spare Parts, Energy, and Operations Support module is responsible for spare parts management and on-site operational assistance. Based on classification and forecasting, it achieves dynamic management and optimization of safety stock. It supports cross-site or cross-domain allocation to improve spare parts utilization. It provides mobile operation records and multimodal evidence collection (such as text records, voice, images, and video) to ensure transparency and traceability of the operation process. Based on a knowledge graph, it maintains a spare parts knowledge graph, recording in detail the correspondence between spare parts and components, models, and compatibility, improving the accuracy of spare parts selection. Under cross-site conditions, it performs optimal allocation based on arrival time, transportation costs, and business impact. For example, when a site urgently needs spare parts but its own inventory is insufficient, the system automatically selects the spare parts with the lowest allocation cost from other sites and achieves full lifecycle tracking.

[0029] The security and compliance module is the cornerstone of the entire system's security. It implements zero-trust access control, rigorously verifying all access requests. It achieves fine-grained access control, ensuring users can only access their authorized data and functions. Through a ticket interlocking mechanism, it establishes a linkage between electronic work tickets and on-site physical controls to prevent unauthorized operations. Data encryption ensures data security during transmission and storage. Tamper-proof record keeping and audit traceability ensure that all operations and data changes are traceable to meet compliance requirements. It interlocks electronic work tickets with on-site controls. When an electronic work ticket fails to meet security conditions such as briefing, double confirmation, and isolation tagging, the system automatically blocks dangerous operations; for example, without a safety briefing, on-site equipment operation permissions cannot be unlocked. Simultaneously, critical operations are encrypted, documented, and audited to ensure the legality and non-repudiation of all critical operations. Interlock conditions for verification: If double confirmation is not completed, the operation is locked; if security briefing is not signed, the system refuses to issue commands; if isolation tagging is not completed, the locked state remains. Encryption and audit parameters: Data encryption algorithm: AES-256; minimum retention period for operation logs: 10 years; audit traceability delay threshold: no more than 5 seconds.

[0030] The intelligent watershed maintenance management system based on the Internet of Things and intelligent decision-making operates through a five-layer logical closed loop: perception, cognition, decision-making, execution, and feedback. The interaction relationships between the modules are as follows: The multi-source perception and edge access module is responsible for collecting and preprocessing heterogeneous data from the field (industrial signals, video, thermal imaging, low-power points), and uploading it to the platform layer after initial screening and synchronization at the edge. The asset ledger and knowledge graph module aligns the accessed data with existing asset structures, maintenance procedures, and geographical topology knowledge, providing knowledge support and constraints for condition assessment and scheduling. The condition assessment and anomaly detection module utilizes the topological structure and physical boundaries in the knowledge graph to calculate the causal consistency health index and counterfactual residuals, and generates remaining lifetime confidence intervals and alarm priorities. After receiving the condition assessment results, the digital twin simulation and impact assessment module performs multi-fidelity differentiable twin simulations on candidate maintenance schemes, outputting risk, energy consumption, carbon intensity, and robustness scores. The maintenance planning and intelligent scheduling module integrates the above information, generates the optimal maintenance scheme under multi-objective constraints, and achieves dynamic adjustment through a trust domain rolling optimization mechanism. The spare parts, energy materials, and operational support module executes material allocation and operational support based on scheduling results, and provides feedback on on-site execution data. The safety and compliance module implements zero-trust verification, ticket interlocking, and audit traceability throughout the entire operational chain, forming a safety closed loop. The interactive data flow direction is: perception data flow → knowledge and asset model flow → status and anomaly flow → simulation feedback flow → optimized scheduling flow → execution and compliance flow → closed-loop feedback flow.

[0031] The interaction logic between the multi-source sensing and edge access module and the asset ledger and knowledge graph module is as follows: The edge acquisition end aligns the device code, sensor point number, and sampling time, and then uploads them through a standardized interface (MQTT / OPC-UA). The asset ledger and knowledge graph module searches for matching relationships in the joint ontology based on the device ID and verifies the consistency of sensor point type, range, and operating conditions. When a new device is detected or parameter drift exceeds the threshold, an "asset change entry" is automatically generated in the knowledge graph, and a shadow update is initiated. Modules are linked through credibility tags and data quality tags; low-credibility data will be downgraded or isolated. Typical interaction parameters: sensor point code (SensorID), timestamp accuracy Δt≤10ms, data integrity rate>99.5%, update trigger threshold: parameter offset>5% or status code change>3 times / hour. The interaction logic between the status assessment and anomaly detection module and the asset ledger and knowledge graph module is as follows: The status assessment module calls the structured constraint rules of the knowledge graph (such as device interlocking, shutdown sequence, and compliance qualifications). The knowledge graph module returns the corresponding equipment's operating status window, procedure number, and interlock status, forming the context constraints for causal modeling. Anomaly detection results (health index, counterfactual residual, lifetime confidence interval) are written back to the knowledge graph to form asset health records. If the anomaly type matches an existing defect pattern, the knowledge graph inference module automatically generates a "risk management suggestion" entity node. Interaction event trigger conditions: (Health threshold), continuous counterfactual residual offset The anomaly duration is >30 minutes. The interaction logic between the state assessment and anomaly detection module and the digital twin simulation and impact assessment module is as follows: When an abnormal trend or lifespan interval approaches a threshold, the state assessment module transmits a health data package (health index sequence, key observation channels, control variables) to the digital twin simulation module. Based on this input, the simulation module generates initial boundary conditions for a differentiable twin model, simulating the risk scenario of "continuing operation under the current anomaly." If the robustness score of the simulation output... If this occurs, the scheduling module will automatically reschedule the maintenance window. The interaction logic between the digital twin simulation and impact assessment module and the maintenance plan and intelligent scheduling module is as follows: the simulation module provides a performance index matrix for each candidate solution. The scheduling module receives the metrics, constructs a multi-objective optimization function, and performs a rolling solution. When the simulation detects active boundary constraints or increased uncertainty, it generates soft constraint backfeeds (trust region radius). (Adjustments). After the scheduling module completes the solution, it sends the optimal control variables (maintenance window, task sequence) back to the simulation module for quick verification. The interaction logic between the maintenance plan and intelligent scheduling module and the spare parts, energy, and work support module: The scheduling module publishes the approved maintenance plan and work order information to the spare parts module, including equipment model, time window, and estimated material usage. The spare parts module calls the knowledge graph to verify material compatibility and performs allocation path optimization. After allocation, it feeds back the ETA (Estimated Time of Arrival) and risk level to the scheduling module for dynamically adjusting the work sequence. After the work is completed, the mobile terminal uploads construction records and video evidence, which are then synchronized to the asset ledger and knowledge graph. The interaction logic between the spare parts, energy, and work support module and the safety and compliance module: Before each maintenance task is started, the safety and compliance module must verify the work ticket, double confirmation, and isolation tagging status. If the ticket does not meet the conditions, the safety module directly refuses to issue the work instruction. During work execution, the field terminal transmits multimodal evidence in real time with encryption, and the safety module records the operation traces using AES-256. After the task is completed, the audit traceability report is synchronously sent back to the spare parts module and the knowledge graph for traceability updates. A watershed maintenance method based on the Internet of Things and intelligent decision-making includes: S1, Access and Governance: This involves multi-protocol data acquisition, time and session alignment, quality labeling, and breakpoint resumption. A multi-source sensing and edge access module is responsible for collecting industrial data, low-power point data, and video and thermal imaging data from various sensors, industrial control systems, and video surveillance equipment. This data undergoes time and session alignment to ensure the correlation and consistency of data from different sources in the time dimension. The collected data is quality-labeled to identify abnormal or missing data, and a breakpoint resumption mechanism ensures the reliability of data transmission. Simultaneously, a lightweight anomaly detection model is run at the edge for preliminary anomaly screening, and the sampling frequency is dynamically adjusted based on the equipment health status. S2, Knowledge Construction: This involves cross-document extraction based on a joint ontology, performing structured constraint verification and minimal repair before data entry, and completing source credibility fusion and traceability for data entry. An asset ledger and knowledge graph module constructs a joint ontology covering spatiotemporal, operational, and procedural dimensions as the basis for knowledge representation. Entities, attributes, and relationships are automatically extracted from unstructured and semi-structured documents using technologies such as natural language processing to form candidate triples. Before storing this knowledge in the database, rigorous structured constraint verification is performed to identify potential conflicts. If conflicts exist, a minimum repair problem is solved to correct them at minimal cost, ensuring knowledge consistency. Simultaneously, source credibility calculations, fusion, and tracing are performed on knowledge entries to quantify their reliability. Finally, this structured and reliable knowledge is stored in a graph database. The credibility of candidate triples is calculated using a weighted fusion of source matching score, source quality, and evidence timeliness decay. For example, information from authoritative specification documents has higher source quality and matching score, while more recent evidence has higher timeliness. Minimum repair is performed based on a cost model based on credibility and business impact to eliminate constraint violations. For example, when two sources describe the temperature threshold for the same device inconsistently, the system will prioritize the threshold with higher credibility and greater business impact (such as causing greater security risks). S3, State Assessment: Calculates causally consistent health assessment results under the baseline condition of exogenous variable intervention, generates residuals based on counterfactual predictions, forms confidence intervals for remaining lifetime, and performs spatiotemporal anomaly detection and priority fusion on a graph structure. The state assessment and anomaly detection module first calculates causally consistent health assessment results for the equipment by setting baseline conditions (e.g., simulating ideal operating conditions without external interference), eliminating the confounding effect of external interference on the assessment. It generates residuals by comparing counterfactual predictions (e.g., predicting how the equipment's state would change if no anomalies occurred) with actual observations; these residuals more sensitively reflect abnormal equipment behavior. Based on historical data and methods such as random deactivation sampling, the remaining lifetime of the equipment is estimated, and confidence intervals with controllable coverage are provided.On an asset coupling graph constructed from a knowledge graph, time series models and graph convolutional models are jointly used to identify spatiotemporal anomaly patterns in individual devices or groups of related devices, and alarm priorities are fused based on the severity of the anomalies and the level of business impact. For health status, causally consistent health assessment results are jointly judged with residuals based on counterfactual predictions to provide a more comprehensive and accurate basis for anomaly identification. For interval estimation of remaining lifetime, random deactivation sampling and quantile consistency calibration are used to improve the statistical reliability of the prediction results. Time series models and graph convolutional models are jointly used on the asset coupling graph to obtain predicted values ​​for the next time step, and an anomaly score is formed based on the strength of the prediction bias; the higher the score, the more severe the anomaly. The anomaly score is then weighted and fused with the level of business impact to form a dispatch priority, ensuring that urgent and important anomalies are handled first. S4, Scheme Evaluation: Candidate maintenance windows are rapidly evaluated using differentiable twin simulation, multi-fidelity fusion, and neural operator agents, followed by uncertainty propagation and robustness scoring. The digital twin simulation and impact assessment module simulates different maintenance schemes (such as different maintenance times and contents) by constructing digital twin models of the equipment. These twin models undergo differentiable calibration to ensure consistency with the real physical process. To balance simulation accuracy and computational efficiency, multi-fidelity model fusion technology is employed, and neural operator agents are used to achieve millisecond-level rapid prediction of key indicators. During the simulation, uncertainties in the external environment (such as future hydrological conditions and weather changes) are propagated, quantifying the performance of maintenance schemes under various uncertain conditions and generating robustness scores. Higher scores indicate stronger resilience. Differentiable twin parameter calibration and weighted fusion of multi-fidelity models (i.e., assigning higher weights to more accurate or more confident models) are used. Rapid evaluation via neural operator agents significantly shortens the scheme evaluation time. After uncertainty propagation, robustness scores are calculated based on the expected value and volatility of maintenance results under different risk scenarios, and sensitive boundaries are fed back to the scheduling module to indicate which uncertainties have the greatest impact on decision-making. S5, Planning and Scheduling: The maintenance plan is solved under multi-objective optimization, opportunity constraints, and conditional risk value constraints, using learning-enhanced warm start and rolling time-domain minimum disturbance release; S6, On-site Execution and Closed-Loop Accumulation: The operation closure is completed through ticket interlocking, mobile operation tracking, and multimodal evidence collection, and the knowledge and model are updated in a versioned manner.

[0032] System Deployment and Operation Mechanism: Unified Interface Standards Between Modules: Interfaces are defined based on JSON-RPC and GraphQL protocols, supporting semantic queries and version control. Data Bus Mechanism: All modules interact through the platform-level "Data and Rules Bus," employing Kafka streaming channels to ensure low latency. Version Control and Canary Release: Each parameter, model, and weight update is tagged with a version number, allowing for replay and comparison with older versions. Traceability Design: Each decision-making process outputs an explanatory chain (trigger source → calculation path → rule ID → execution result). Security and Audit Mechanisms: The entire chain employs AES-256 encryption, zero-trust authentication, and operation watermark auditing, supporting 10-year traceability and rapid auditing. A watershed maintenance method based on the Internet of Things and intelligent decision-making is proposed. In step S2, the credibility of candidate triples is calculated by weighted fusion of source matching score, source quality, and evidence timeliness decay. Minimum repair is performed to eliminate constraint violations based on a cost model based on credibility and business impact. In step S3, the health status is jointly judged by causal consistent health assessment results and residuals based on counterfactual prediction. The interval estimation of remaining lifetime is calibrated by random inactivation sampling and quantile consistency. In step S3, time series models and graph convolution models are jointly used on the asset coupling graph to obtain the predicted value of the next time step. An anomaly score is formed by the intensity of prediction bias, and then weighted fusion with the business impact level to form the dispatch order. Prioritization in S4 employs differentiable twin parameter calibration and weighted fusion of multi-fidelity models, with rapid evaluation via a neural operator agent. After uncertainty propagation, robustness scores are calculated, and sensitive boundaries are fed back to the scheduling module. S5 models a multi-objective function encompassing project duration, cost, expected downtime risk conditions, and carbon emission intensity, setting opportunity constraints to limit violation probabilities. Learning-enhanced warm-start and minimum-disturbance reconstruction are used for rolling deployment. The maintenance planning and intelligent scheduling modules model maintenance plan formulation as a multi-objective optimization problem, considering multiple objectives such as total project duration, comprehensive cost, expected downtime risk conditions, and carbon emission intensity, striving for an optimal balance among them. Opportunity constraints are introduced to address hydrological uncertainties, limiting the probability of violating constraints such as water level and flow rate to within preset thresholds. Conditional risk value constraints are also employed to control extreme losses from downtime risks. During the optimization process, a learning-enhanced warm-start strategy trained from historical best solutions is used to quickly find a high-quality initial solution. When actual conditions change, the system, within the rolling time domain, performs minimal disturbance reconstruction and release of the plan based on minimizing the time offset of published work orders. A multi-objective function, incorporating expected factors such as duration, cost, downtime risk conditions, and carbon emission intensity, is used to achieve comprehensive optimization of the maintenance plan. Opportunity constraints are set to limit the probability of constraint violations to a set threshold, enhancing the plan's resilience to uncertainty. Simultaneously, learning-enhanced warm start (using a machine learning model to predict a good initial solution) and minimal disturbance reconstruction are employed for rolling release, enabling the plan to flexibly respond to unforeseen circumstances while maintaining overall stability. S6 implements dynamic authorization of personnel, equipment, and tasks through zero-trust and fine-grained access control, and integrates ticket interlocking, encrypted storage, and audit traceability throughout the closed-loop operation. During the on-site execution phase, the safety compliance module ensures that all hazardous operations comply with safety procedures through a ticket interlocking mechanism; for example, on-site operations cannot proceed without the approval of electronic work tickets and confirmation of safety measures. The spare parts, energy materials and operation support module provides mobile terminal support, allowing workers to record operations and collect evidence in a multimodal manner (photos, videos, and voice recordings) on-site, ensuring transparency and traceability of the operation process.Upon completion of the task, all task data, encountered problems, solutions, and newly generated knowledge (such as new defect types and improved maintenance procedures) will be transmitted back to the asset ledger and knowledge graph module for updating and maintaining the knowledge graph. Simultaneously, data accumulated during system operation can be used to continuously optimize the status assessment model and scheduling optimization model, enabling version updates and iterative optimization of knowledge and models, forming a complete maintenance management closed loop. Zero-trust and fine-grained access control enable dynamic authorization of personnel, equipment, and tasks, ensuring that only authorized personnel can execute authorized tasks on authorized equipment. Ticket interlocking, encrypted evidence storage, and audit traceability are integrated throughout the entire task closed loop process. From pre-task approval and execution to post-task review, all stages are subject to strict security supervision and recording, ensuring compliance, security, and auditability of operations.

[0033] An electronic device includes a processor and a memory. The memory stores a program that can run on the processor. When the processor executes the program, the electronic device performs all the steps of the watershed maintenance method based on the Internet of Things and intelligent decision-making described in the above embodiments. The electronic device can be, but is not limited to, a server, cloud computing device, edge computing device, personal computer, tablet computer, smartphone, or any hardware device with computing and storage capabilities. By executing the program in the memory through the processor, the electronic device can achieve access and governance of multi-source data, knowledge construction, equipment status assessment and anomaly detection, simulation evaluation of maintenance plans, intelligent optimization and scheduling of maintenance plans, and support for on-site operations and safety and compliance management. The processor can be one or more central processing units (CPUs), graphics processing units (GPUs), application-specific integrated circuits (ASICs), or field-programmable gate arrays (FPGAs), etc. The memory can be one or more random access memories (RAMs), read-only memories (ROMs), flash memory, or any other suitable non-transient storage medium. The watershed maintenance method based on the Internet of Things and intelligent decision-making implemented by this electronic device has the same principles and technical effects as the above-described method embodiments, and will not be repeated here.

[0034] Example 1: Early Anomaly Detection and Predictive Maintenance of Upstream Pump Station Bearings 1. Data Access: The multi-source sensing and edge access module detected the vibration amplitude of the pump unit. 2. Knowledge Association: The asset ledger and knowledge graph module locates the pump unit model and component level, and associates it with the historical defect "bearing wear" pattern; 3. Status Assessment: The status assessment and anomaly detection module calculates the causal consistency health index. 4. Simulation Verification: The digital twin simulation and impact assessment module deduced that if the system continued to operate for another 48 hours, the flood discharge efficiency would decrease by 2.8% and energy consumption would increase by 3.2%; 5. Scheduling Decision: The maintenance plan and intelligent scheduling module adjusted the maintenance window to the low-load period at night, optimizing the objective function. 6. On-site execution: Spare parts and materials are allocated to the operation support module to ensure compatibility with bearings. The safety and compliance module verifies the ticket and issues the task. 7. Closed-loop feedback: After the operation is completed, the vibration amplitude returns to normal, and the knowledge graph adds a "early anomaly - bearing wear" relationship node.

[0035] Example 2: Abnormal Propagation and Cross-Station Joint Maintenance of Gate Hydraulic System 1. Detection phase: Upstream gate temperature rise rate 1. Threshold exceeded; the spatiotemporal anomaly model identified resonance at the downstream station with a 10-minute delay; 2. Inference phase: the knowledge graph infers that the two stations share a hydraulic pipeline edge, and the system generates a propagation path; 3. Simulation phase: the digital twin simulation module reproduces the pressure wave propagation under this linkage state and determines that if isolation is not implemented, there are safety risks both upstream and downstream; 4. Scheduling phase: the maintenance plan and intelligent scheduling module trigger linked maintenance, prioritizing downstream flood safety, and setting the trust region radius. 5. Execution phase: The spare parts and energy materials and operation support module performs cross-site allocation of hydraulic oil pump components, and the safety and compliance module enforces ticket interlocking to ensure synchronous shutdown; 6. Retrospective phase: Anomaly propagation paths, time windows and repair actions are archived into the knowledge graph for subsequent risk review.

[0036] Example 3: Rolling Dispatch and Robustness Optimization under Flood Season Conditions 1. External disturbance input: The digital twin simulation module detected a 20% increase in the predicted water inflow, leading to changes in watershed boundary constraints; 2. State assessment update: The health index decreased to 3. Re-solving scheduling: The maintenance plan and intelligent scheduling module are subject to opportunity constraints. 4. Robustness analysis: The simulation module outputs a robustness score. , compliance plan; if This will trigger an alternative solution switch; 5. Execution feedback: On-site execution data is transmitted back to the knowledge graph in real time, forming a "flood season rolling scheduling strategy" node, which is used for subsequent model retraining.

[0037] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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. A watershed maintenance intelligent management system based on the Internet of Things and intelligent decision-making, characterized in that: include: The multi-source sensing and edge access module is used to realize multi-protocol access, time and session alignment, data quality labeling, breakpoint resume, and edge-side anomaly screening of industrial data, low-power point data, and video and thermal imaging data. The asset ledger and knowledge graph module is used to establish a joint ontology covering the dimensions of time, space, working conditions and procedures, perform cross-document information extraction, structured constraint verification and minimum repair before knowledge is stored, and perform source credibility fusion and traceability of knowledge entries. The status assessment and anomaly detection module is used to calculate causally consistent health assessment results under the condition that exogenous variables are intervened as a baseline, generate residuals based on counterfactual predictions, output the confidence interval of remaining life, and perform graph spatiotemporal anomaly identification and alarm priority fusion on the asset coupling graph. The maintenance planning and intelligent scheduling module is used to jointly optimize the schedule, cost, downtime risk and carbon emission intensity under a multi-objective optimization framework. It introduces opportunity constraints and conditional risk value constraints to cope with hydrological uncertainties and reconstructs the published plan in a minimum perturbation manner in the rolling time domain. The digital twin simulation and impact assessment module is used to perform differentiable calibration of equipment and hydrodynamic processes, weighted fusion of multi-fidelity models based on uncertainty, fast scheme evaluation using neural operator agents, and propagation of external uncertainties to form robustness scores. The spare parts, energy materials and operations support module is used to realize safety stock and cross-site or cross-domain transfers based on classification and prediction, and provides mobile operation records and multimodal evidence collection. The security and compliance module is used to implement zero-trust access control, fine-grained permission management, ticket interlocking, data encryption, and tamper-proof traceability. Each module interacts with the data and rules bus between the platform layer and the collaboration layer to form a full-process control from data acquisition, status assessment, scheme simulation to planning, scheduling and operation closure.

2. The intelligent management system for watershed maintenance based on the Internet of Things and intelligent decision-making as described in claim 1, characterized in that: The multi-source sensing and edge access module dynamically adjusts the sampling frequency based on the health status of the equipment and runs a lightweight anomaly detection model on the edge side to trigger local linkage control. At the same time, it performs self-iterative updates of thresholds that can be applied in grayscale. The spare parts, energy materials and operation support module maintains the correspondence between spare parts and components, models and compatibility based on a knowledge graph. Under cross-site conditions, it performs optimal allocation based on delivery time, transportation costs and business impact and achieves full lifecycle tracking. The safety and compliance module interlocks electronic work tickets with on-site control. It automatically blocks dangerous operations when the conditions of handover, double confirmation and isolation tagging are not met, and encrypts and stores evidence of key operations for audit traceability.

3. The intelligent management system for watershed maintenance based on the Internet of Things and intelligent decision-making as described in claim 1, characterized in that: The asset ledger and knowledge graph module models the three dimensions of the joint ontology as spatiotemporal positioning, operating status, and procedural constraints, respectively. It stores the relationships between assets, components, sensor points, defects, procedures, spare parts, work tickets, risks, and geographic topology in a graph database. The asset ledger and knowledge graph module comprehensively considers the matching score of the source, the source quality weight, and the time decay of evidence when evaluating the credibility of candidate triples. The fusion weight is a configurable or learnable parameter. Before knowledge is stored in the database, the asset ledger and knowledge graph module eliminates structured constraint conflicts by solving the minimum repair problem. The repair cost is determined based on a weighted average of credibility and business impact.

4. The intelligent management system for watershed maintenance based on the Internet of Things and intelligent decision-making as described in claim 1, characterized in that: The status assessment and anomaly detection module calculates causally consistent health assessment results based on the deviation between the expected value and the actual observed value when the exogenous variable is fixed as a baseline. The difference between the counterfactual prediction and the actual observation is used as the residual for anomaly identification. The remaining lifespan estimate is obtained by random deactivation sampling to obtain the distribution, and quantile consistency calibration is performed based on an independent calibration set to obtain a confidence interval with controllable coverage. The adjacency relationship of the asset coupling is updated by jointly using a time series model and a graph convolution model, and the strength of the deviation between the predicted value and the actual value is used as the anomaly score.

5. The intelligent management system for watershed maintenance based on the Internet of Things and intelligent decision-making as described in claim 1, characterized in that: The optimization objectives of the maintenance plan and intelligent scheduling module simultaneously consider the total project duration, comprehensive cost, expected conditions of downtime risk, and carbon emission intensity. Opportunity constraints are set for hydrological uncertainties to limit the probability of constraint violation to no more than a set threshold. Feasible initial solutions trained from historical best solutions are used as the solver for warm start-up, and reconstruction and release are carried out in the rolling time domain in a way that minimizes the predetermined work order time offset.

6. The intelligent management system for watershed maintenance based on the Internet of Things and intelligent decision-making as described in claim 1, characterized in that: The digital twin simulation and impact assessment module minimizes both observation errors and physical equation residuals during parameter calibration to ensure consistency with field observations and compliance with physical laws. When fusing multi-fidelity models, it adopts a weighting strategy inversely proportional to uncertainty and uses neural operator agents to quickly predict key indicators in milliseconds. It samples external uncertainties, comprehensively considers the expected value and volatility of impact indicators to generate robustness scores, and feeds back sensitive boundaries and constraints to the scheduler solver.

7. A method for applying to the intelligent watershed maintenance management system based on the Internet of Things and intelligent decision-making as described in any one of claims 1-6, characterized in that: include: S1, Access and Governance: Perform multi-protocol data collection, time and session alignment, quality labeling, and breakpoint resumption; S2, Knowledge Construction: Cross-document extraction is performed based on the joint ontology. Before being stored in the database, structured constraint verification and minimal repair are performed, and source credibility fusion and traceability are completed for storage. S3, State Assessment: Calculate causally consistent health assessment results under the baseline condition that exogenous variables have been intervened, generate residuals based on counterfactual predictions, form confidence intervals for remaining life expectancy, and perform spatiotemporal anomaly detection and priority fusion on the graph structure; S4, Scheme Evaluation: Rapid evaluation of candidate maintenance windows using differentiable twin simulation, multi-fidelity fusion, and neural operator proxy, and uncertainty propagation and robustness scoring. S5, Planning and Scheduling: Solving maintenance plans under multi-objective optimization, opportunity constraints, and conditional risk-value constraints, using learning-enhanced warm start and rolling time-domain minimum perturbation release; S6, On-site Execution and Closed-Loop Data Accumulation: Complete the operation closed loop through ticket interlocking, mobile operation traceability and multimodal evidence collection, and update the knowledge and model in a versioned manner.

8. A watershed maintenance method based on the Internet of Things and intelligent decision-making according to claim 7, characterized in that: In S2, the credibility of candidate triples is calculated using a weighted fusion of source matching score, source quality, and evidence decay. Minimum repair is then performed based on a cost model grounded in credibility and business impact to eliminate constraint violations. In S3, health status is jointly determined using causally consistent health assessment results and residuals based on counterfactual predictions. Interval estimation of remaining lifetime employs random inactivation sampling and quantile consistency calibration. In S3, time series and graph convolution models are jointly used on the asset coupling graph to obtain predicted values ​​for the next time step. Anomaly scores are generated based on the intensity of prediction bias and then weighted fused with the business impact level to form a... In the S4 phase, differentiable twin parameter calibration and multi-fidelity model weighted fusion are used for single-priority dispatch. A neural operator agent is used for rapid evaluation. After uncertainty propagation, a robustness score is calculated and the sensitive boundary is fed back to the scheduling module. In the S5 phase, a multi-objective function is used to model the project, which includes the expected conditions of the project duration, cost, downtime risk, and carbon emission intensity. Opportunity constraints are set to limit the probability of violation. Learning-enhanced warm start and minimum disturbance reconstruction are used for rolling deployment. In the S6 phase, zero trust and fine-grained access control are used to realize dynamic authorization of personnel, equipment and tasks. Ticket interlocking, encrypted storage and audit traceability are integrated into the operation closed loop.

9. An electronic device, characterized in that: It includes a processor and a memory, the memory storing a program that can run on the processor, which, when executed, causes the electronic device to perform the steps of the method of claim 7.

Images