Digital twin hydraulic engineering operation and maintenance monitoring system and method based on multi-modal data

Abstract

The invention provides a digital twin hydraulic engineering operation and maintenance monitoring system and method based on multi-modal data, and belongs to the technical field of hydraulic engineering and digital twin. The method comprises the following steps: determining a plurality of sub-unit systems of a target water conservancy project to be monitored, and constructing a corresponding digital twin sub-model for each sub-unit system to obtain a digital twin engineering model of the target water conservancy project; acquiring multi-modal historical monitoring data related to the target water conservancy project; based on the multi-modal data, determining a construction environment of a digital twinborn model; mapping each digital twinning sub-model to a construction environment, and running a digital twinning engineering model; and acquiring multi-modal real-time monitoring data related to the target water conservancy project in real time, inputting the multi-modal real-time monitoring data into the digital twin engineering model, and performing operation and maintenance monitoring on the target water conservancy project. According to the method, the rigid-flexible container adaptive model is adopted, so that the cloud resource calling cost can be reduced while the visual operation and maintenance of the water conservancy project are realized and the monitoring efficiency and accuracy are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of water conservancy projects and digital twin technology, and particularly relates to a digital twin water conservancy project operation and maintenance monitoring system and method based on multi-modal data. BACKGROUND

[0002] With the continuous expansion of water conservancy project construction scale and the increasing refinement of operation and maintenance requirements, digital twin technology, with its core characteristics of "virtual-real mapping, real-time interaction, and dynamic optimization", has gradually become a key technical support in the field of water conservancy project operation and maintenance monitoring. Digital twin technology realizes dynamic simulation, state monitoring and intelligent decision-making of the whole life cycle of the project by constructing a digital mirror of the physical entity, providing a new idea for solving the problems of "information lag, one-sided monitoring, and passive decision-making" in traditional water conservancy project operation and maintenance. At present, digital twin technology has been preliminarily applied in the operation and maintenance monitoring of large-scale reservoirs, inter-basin water transfer projects, and dike systems and other water conservancy infrastructure. By integrating data collected by sensing devices and engineering design parameters, a simplified digital model is constructed to realize real-time tracking of key indicators such as water level changes and structural deformations, which has to some extent improved the operation and maintenance efficiency and risk warning capability. For example, the digital twin water conservancy project operation safety monitoring system and operation method proposed in Chinese patent application CN115688227A, the water conservancy monitoring method based on digital twin proposed in Chinese patent application CN116625324A, and the water conservancy monitoring method and system based on digital twin proposed in Chinese patent application CN113139659A.

[0003] The data sources involved in water conservancy project operation and maintenance monitoring have significant multi-modal characteristics, including not only static sensing data reflecting the inherent properties of the project, but also dynamic sensing data changing dynamically over time. There are significant differences in time scale and data dimension between static data and dynamic data, and related technologies lack a unified fusion framework for multi-modal data. At the same time, the running efficiency and accuracy of the digital twin model are highly dependent on the rationality of the construction environment based on cloud servers, and different types of sub-unit systems have significant differences in the demand for computing resources, response speed, and dynamic adjustment capability. However, related technologies do not consider such differences in different types of sub-unit systems when constructing the model environment, making it difficult for the operation and maintenance monitoring system to accurately depict all factors and the whole cycle of water conservancy projects, and unable to meet the needs of large-scale water conservancy projects for high reliability and high real-time operation and maintenance decision-making. Cloud server resources are not used reasonably and efficiently. SUMMARY

[0004] To solve the above technical problems, the present application provides a digital twin water conservancy project operation and maintenance monitoring system and method based on multi-modal data.

[0005] In a first aspect of the present application, a digital twin water conservancy project operation monitoring method based on multi-modal data is proposed.

[0006] The method comprises the following steps: determining a plurality of sub-unit systems of a target water conservancy project to be monitored, constructing a corresponding digital twin sub-model for each sub-unit system, and obtaining a digital twin project model of the target water conservancy project; acquiring multi-modal historical monitoring data related to the target water conservancy project, the multi-modal historical monitoring data comprising static sensing data and dynamic sensing data; determining a construction environment of the digital twin model based on the multi-modal data, the construction environment comprising M rigid container environments with fixed specifications and N flexible container environments with variable specifications, wherein M and N are positive integers and M < N; mapping each digital twin sub-model to the construction environment and running the digital twin project model; real-time acquisition of multi-modal real-time monitoring data related to the target water conservancy project, input into the digital twin project model, and operation monitoring of the target water conservancy project.

[0007] The mapping of each digital twin sub-model to the construction environment and the running of the digital twin project model specifically comprises: mapping at least a first digital twin sub-model to a first rigid container environment and at least a second digital twin sub-model to a second flexible container environment.

[0008] After mapping at least a first digital twin sub-model to a first rigid container environment, the method further comprises: determining whether a sub-model migration condition is met, if so, redirecting the first digital twin sub-model to a third flexible container environment.

[0009] The real-time acquisition of multi-modal real-time monitoring data related to the target water conservancy project, the input into the digital twin project model, and the operation monitoring of the target water conservancy project further comprise: based on the multi-modal real-time monitoring data, predicting multi-modal prediction fitting data; inputting the multi-modal real-time monitoring data and the multi-modal prediction fitting data into the digital twin project model in sequence, and operation monitoring of the target water conservancy project.

[0010] The method further comprises: based on the results of the operation monitoring of the target water conservancy project, correcting the digital twin project model of the target water conservancy project.

[0011] The process of obtaining multimodal prediction fitting data based on multimodal real-time monitoring data specifically includes: extracting trend terms from multimodal real-time monitoring data using a time series decomposition algorithm, constructing a prediction model using a long short-term memory neural network, and generating multimodal prediction fitting data for the next 24 hours based on a sliding window mechanism.

[0012] In a second aspect of the invention, to implement the method described in the first aspect, a digital twin water conservancy project operation and maintenance monitoring system based on multimodal data is also proposed.

[0013] The system includes: The model building module is used to determine multiple sub-unit systems of the target water conservancy project to be monitored, and to build a corresponding digital twin sub-model for each sub-unit system to obtain the digital twin engineering model of the target water conservancy project. The historical data acquisition module is used to acquire multimodal historical monitoring data related to the target water conservancy project, including static sensor data and dynamic sensor data. The environment configuration module is used to determine the construction environment of the digital twin model based on the multimodal data. The construction environment includes M rigid container environments with fixed specifications and N flexible container environments with variable specifications, where M and N are both positive integers and M... <N; The model execution module is used to map each digital twin sub-model to the construction environment and run the digital twin engineering model; The real-time monitoring module is used to collect multimodal real-time monitoring data related to the target water conservancy project in real time, and to predict multimodal prediction fitting data based on the multimodal real-time monitoring data; the multimodal real-time monitoring data and the multimodal prediction fitting data are sequentially input into the digital twin engineering model to perform operation and maintenance monitoring of the target water conservancy project; The model correction module is used to correct the digital twin engineering model of the target water conservancy project based on the results of operation and maintenance monitoring of the target water conservancy project.

[0014] The model running module includes a mapping unit and a migration unit. The mapping unit is used to map at least a first digital twin sub-model to a first rigid container environment and to map at least a second digital twin sub-model to a second flexible container environment. The migration unit is used to redirect the first digital twin sub-model to a third flexible container environment when the sub-model migration conditions are met.

[0015] The model correction module is used to calculate the output deviation of the digital twin engineering model based on the operation and maintenance monitoring results. When the output deviation value exceeds the preset threshold for three consecutive cycles, the model parameter optimization process is triggered to dynamically correct the physical parameters and algorithm coefficients of the digital twin engineering model.

[0016] The system also includes: The visualization unit is used to visualize the results of operation and maintenance monitoring of the target water conservancy project.

[0017] This invention systematically integrates static and dynamic sensor data to construct a multimodal data fusion framework, effectively solving the "data silo" problem in existing technologies. Static data provides the foundational support for the model, while dynamic data reflects real-time changes in operating conditions. Both are input into the digital twin model through a unified mapping mechanism. This multimodal fusion strategy allows the model to depict both inherent engineering attributes and dynamic changes, avoiding trend drift or lag issues caused by reliance on single data sources, and comprehensively improving the accuracy and timeliness of operation and maintenance monitoring. Addressing the rigid configuration problem of existing container environments, this invention designs a differentiated adaptation scheme of M rigid containers and N flexible containers. Rigid containers provide a stable computing environment, ensuring operational stability; flexible containers provide elastic resource scheduling. Simultaneously, a sub-model migration mechanism enables dynamic resource allocation, avoiding computational waste and ensuring model continuity under sudden changes in operating conditions. Finally, this invention innovatively adopts a multimodal real-time data-driven prediction-correction mechanism to solve the problems of single prediction and lag in correction in existing technologies. Through time series decomposition and long short-term memory neural networks, 24-hour prediction fitting data is generated based on a sliding window, combined with real-time data as dual input to the model. Meanwhile, the model correction module automatically calculates the output deviation, triggering parameter optimization when the deviation exceeds the threshold for three consecutive cycles, dynamically adjusting the physical parameters and algorithm coefficients. This closed-loop mechanism improves prediction accuracy, avoids the subjectivity of manual correction, and ensures long-term consistency between the model and the physical entity.

[0018] Further specific advantages and implementation principles of the present invention will be further detailed in the specific embodiments section in conjunction with the accompanying drawings. Attached Figure Description

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

[0020] Figure 1 This is a schematic diagram of the main execution flow of a digital twin water conservancy project operation and maintenance monitoring method based on multimodal data according to an embodiment of the present invention; Figure 2 yes Figure 1 A further preferred embodiment of the method is illustrated in the diagram; Figure 3This is a schematic diagram of the hardware unit composition of a digital twin water conservancy project operation and maintenance monitoring system based on multimodal data according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a human-computer interaction scenario in the practical application of a digital twin water conservancy project operation and maintenance monitoring system based on multimodal data, according to an embodiment of the present invention. Detailed Implementation

[0021] In the specific embodiments of this application, if the embodiments of the relevant technical solutions involve user-related data, then when the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0022] In this section, the present invention provides several method or product embodiments. Each method or product embodiment constitutes an independent technical solution and makes at least one contribution to the prior art, capable of solving one or more technical problems mentioned in the background art, and having one or more of the aforementioned outstanding effects or advantages.

[0023] However, it is understood that it is not required that every embodiment solves all the technical problems mentioned in the background art or achieves all the technical effects (e.g., the multiple effects / advantages mentioned later). For each individual embodiment, as long as it makes at least one contribution relative to the prior art and achieves an improved technical effect (e.g., has at least one of the multiple advantages / effects mentioned later), has outstanding substantive features or significant progress, the technical solution constituted by that individual embodiment should possess novelty and inventiveness in the sense of patent law.

[0024] Furthermore, in several embodiments of this section, "functional unit" and "functional module" are equivalent, that is, "XX unit" and "XX module" are merely the names of the main body that performs the corresponding function (step), which can be understood as the corresponding hardware execution component or the software module / process that performs the corresponding function. The technical solutions of this invention will also include descriptions such as "first," "second," and "third." It should be understood that "first," "second," and "third" are used only to distinguish different functional entities and do not represent order or sequence. For example, in the phrase "mapping at least a first digital twin model to a first rigid container environment, and mapping at least a second digital twin model to a second flexible container environment," the "first digital twin model" and "second digital twin model" can both be any one of the aforementioned "multiple sub-unit systems" that meets the corresponding conditions; similarly, the "first rigid container environment" is any one of the aforementioned "M rigid container environments with fixed specifications," and the "third flexible container environment" and "second flexible container environment" are any two different "flexible container environments" that meet the corresponding conditions from the aforementioned "N flexible container environments with variable specifications."

[0025] To better understand the motivation behind the improvement of the technical solution of this invention, we will first briefly introduce the problems existing in the prior art, thereby leading to the technical problem that this invention aims to solve.

[0026] In recent years, with the deepening of smart water conservancy construction, the application of digital twin technology in the field of water conservancy project operation and maintenance monitoring has ushered in a period of rapid development. Digital twins, by constructing a digital model highly synchronized with the physical water conservancy project, realize real-time perception, dynamic simulation, and intelligent decision-making of the project's status, and have become a core technical means to improve the efficiency of water conservancy project operation and maintenance and reduce safety risks. Currently, large-scale water conservancy projects in China, such as the Three Gorges Dam and the South-to-North Water Diversion Project, have gradually introduced digital twin technology. By deploying sensor networks to collect data such as water level, flow rate, and structural deformation, and combining this with BIM models to construct digital mirrors, they have initially achieved remote monitoring and early warning of key parts of the project. At the application level, digital twin water conservancy operation and maintenance monitoring has evolved from early static modeling to dynamic interaction. Some systems can achieve hourly model updates, providing operation and maintenance personnel with an intuitive visualization interface of the project's status. However, at the data processing level, existing technologies generally overlook the multimodal characteristics of water conservancy project monitoring data, resulting in the data's value not being fully explored. Water conservancy project monitoring data exhibits significant multimodal characteristics, including both static data reflecting the inherent attributes of the project, such as long-term stable structured data like dam concrete strength, foundation geological structure, and gate mechanical parameters, and dynamic data that changes over time, such as high-frequency unstructured data like real-time water level, flow velocity, structural vibration frequency, and equipment operating temperature. Static data forms the basis of model building, determining the physical accuracy of digital twins; dynamic data reflects the real-time status of the project and is the core of achieving dynamic monitoring. However, existing technologies often employ a singular data acquisition and processing strategy: some systems focus solely on the real-time transmission of dynamic data, treating static data as fixed parameters input to the model only, leading to a decrease in model accuracy due to static parameter drift over long-term operation; other systems rely on static data for offline modeling, using dynamic data only as an auxiliary reference, failing to achieve real-time model updates. More importantly, existing technologies lack a framework for fusing multimodal data. Static and dynamic data differ significantly in time scale and data dimension, making it difficult for traditional data stitching methods to establish effective correlations, resulting in a situation where "static data is idle and dynamic data is isolated." For example, in dam structure monitoring, existing systems mostly analyze dynamic data collected by vibration sensors in isolation, without combining it with static data such as the age of the dam concrete for comprehensive evaluation. This leads to an inability to accurately distinguish between normal structural vibrations and abnormal vibrations caused by potential cracks, easily resulting in false alarms or missed alarms. At the model runtime environment level, existing technologies have not achieved differentiated adaptation between rigid and flexible containers, making it difficult to balance model runtime efficiency and stability. The operation of digital twin models relies on the support of the underlying computing environment. Different types of hydraulic engineering sub-systems have fundamentally different requirements for computing resources: For structurally stable subsystems such as dam foundations and spillways, whose model parameters change slowly and whose computing resource requirements are relatively fixed, they are suitable for operation in a rigid container environment with stable specifications to ensure computational accuracy and stability. However, for subsystems significantly affected by hydrological conditions, such as hydrodynamics and sediment deposition, whose model parameters need to be dynamically adjusted with real-time data, their computing resource requirements fluctuate greatly, urgently requiring a flexible container environment with variable specifications to provide elastic support. However, existing technologies generally adopt a uniform container environment configuration mode: some systems, in pursuit of stability, use rigid container environments exclusively, causing dynamic sub-models such as hydrodynamics to be unable to update parameters in a timely manner due to resource constraints, resulting in simulation results lagging behind actual hydrological changes; other systems, to meet dynamic adjustment requirements, over-rely on flexible container environments, causing static sub-models to occupy redundant computing resources, significantly increasing operating costs. More seriously, existing technologies lack dynamic migration mechanisms between container environments. When water conservancy projects encounter extreme weather or sudden operating conditions (such as a sudden rise in water level due to heavy rainfall), the structural sub-models originally operating in rigid containers need to switch to flexible containers to adapt to drastic parameter changes. However, existing systems cannot achieve smooth migration, often leading to model interruption or data loss, severely affecting monitoring continuity. For example, in gate operation and maintenance monitoring, existing systems fix the gate mechanical model in a rigid container. When floods cause sudden changes in the gate's stress mode, the model cannot migrate to a flexible container for dynamic adjustment, easily leading to deviations in mechanical wear prediction and creating potential safety hazards. Furthermore, the lack of systematic design in existing model adaptation mechanisms further exacerbates the limitations of operation and maintenance monitoring. The accuracy of digital twin models depends on the precise matching between the model and the operating environment. However, existing technologies have not established adaptation standards between model types and container environments, leading to widespread "model-environment" mismatches. For example, deploying a water flow model that requires high-frequency updates in a rigid container results in a computation cycle that cannot match the rate of water flow change; deploying a structurally stable basic model in a flexible container causes unnecessary accuracy fluctuations due to dynamic resource scheduling. This adaptation deficiency makes it difficult for digital twin models to fully realize their potential, failing to guarantee long-term stable monitoring of static subsystems or meet the real-time response requirements of dynamic subsystems, ultimately significantly reducing the comprehensiveness and reliability of operation and maintenance monitoring.

[0027] To address at least one of the technical problems raised above, several embodiments of the technical solution of the present invention are proposed herein.

[0028] First see Figure 1 , Figure 1This is a schematic diagram of the main execution flow of a digital twin water conservancy project operation and maintenance monitoring method based on multimodal data according to an embodiment of the present invention.

[0029] For the sake of brevity in the description of the following steps, this will be... Figure 1 The steps of the method are numbered as follows (step numbers are omitted in the attached figures): S1: Identify multiple sub-systems of the target water conservancy project to be monitored. S2: Construct a corresponding digital twin sub-model for each sub-unit system to obtain the digital twin engineering model of the target water conservancy project; S3: Acquire multimodal historical monitoring data related to the target water conservancy project, wherein the multimodal historical monitoring data includes static sensor data and dynamic sensor data; S4: Based on the multimodal data, determine the construction environment of the digital twin model. The construction environment includes M rigid container environments with fixed specifications and N flexible container environments with variable specifications, where M and N are both positive integers and M... <N; S5: Map each digital twin sub-model to the building environment and run the digital twin engineering model; S6: Collect multimodal real-time monitoring data related to the target water conservancy project in real time, input it into the digital twin engineering model, and perform operation and maintenance monitoring of the target water conservancy project.

[0030] Figure 2 yes Figure 1 A further preferred embodiment of the method is illustrated in the diagram. Figure 2 In step S5, each digital twin sub-model is mapped to the construction environment, and the digital twin engineering model is run. Specifically, this includes: S500: Map at least a first digital twin model to a first rigid container environment, and map at least a second digital twin model to a second flexible container environment.

[0031] Furthermore, step S6 further includes: S610: Real-time acquisition of multimodal real-time monitoring data related to the target water conservancy project; S620: Based on the multimodal real-time monitoring data, multimodal prediction fitting data is obtained; S630: The multimodal real-time monitoring data and the multimodal prediction fitting data are successively input into the digital twin engineering model to perform operation and maintenance monitoring of the target water conservancy project; Specifically, step S620, which involves predicting multimodal prediction fitting data based on real-time multimodal monitoring data, includes: extracting trend terms from the real-time multimodal monitoring data using a time series decomposition algorithm, constructing a prediction model using a long short-term memory neural network, and generating multimodal prediction fitting data for the next 24 hours based on a sliding window mechanism.

[0032] Furthermore, after step S6, the method further includes: S7: Based on the results of operation and maintenance monitoring of the target water conservancy project, revise the digital twin engineering model of the target water conservancy project and return to step S6.

[0033] As a further preferred embodiment, after mapping at least the first digital twin model to the first rigid container environment in step S500, the method further includes: Determine whether the first digital twin sub-model satisfies the sub-model transfer condition. If so, the first digital twin model is redirected to the third flexible container environment.

[0034] Next, in order to better understand the steps of the above method embodiments, this section will elaborate on them in conjunction with more specific application scenarios.

[0035] It should be noted that the specific application scenarios given here are merely illustrative and not exhaustive. The specific parameter names and sizes in the corresponding scenarios are only for calculation or principle illustration and do not represent that all embodiments have such type or size limitations. In actual implementation of the technical solution of this invention, the parameter names and sizes in the corresponding scenarios are not limited thereto, and the scope of protection of this invention is determined by the content of the claims.

[0036] Step S1 is based on the structural composition, functional attributes and operation and maintenance requirements of the water conservancy project. The target project is systematically decomposed and divided into sub-unit systems with clear boundaries and independent functions.

[0037] Taking a large reservoir water conservancy project as an example, the specific implementation process of step S1 is as follows: First, a comprehensive structural survey of the target project was conducted to identify its core components. Based on the core functional chain of the reservoir project—"water retention-discharge-transfer-monitoring"—the main structural units were preliminarily identified, including primary structural modules such as the dam structure, spillway system, water conveyance tunnel, gate opening and closing equipment, reservoir hydrological monitoring area, and dam foundation seepage prevention system. Secondly, a two-tiered breakdown is conducted based on key operational and maintenance monitoring points. For the dam structure, it is divided into sub-units according to material properties and stress areas, such as the concrete gravity dam section, the dam abutment seepage prevention curtain area, and the dam drainage system. Among these, the concrete gravity dam section requires key monitoring of structural deformation and stress distribution, while the dam abutment seepage prevention curtain area requires attention to changes in seepage flow. The spillway system is broken down into three sub-units: the overflow weir, the stilling basin, and the gate control mechanism. The overflow weir focuses on monitoring the water level-flow relationship, while the gate control mechanism emphasizes monitoring the mechanical operating status. Furthermore, dynamic interaction relationships are considered to supplement the sub-units. In view of the coupling characteristics of "water flow-structure-equipment" in reservoir projects, a new hydrodynamic sub-unit (covering the reservoir flow field and spillway flow morphology) and an equipment linkage control sub-unit (related to the coordinated operation of gates, pumps and other equipment) are added to ensure that the sub-unit system covers the key aspects of the dynamic operation of the project. Finally, the sub-unit division is optimized through boundary verification. The functional independence (e.g., the gate hoist stator model can independently reflect the equipment's operating status) and data availability (each sub-unit must correspond to a specific sensor monitoring point) of each sub-unit are verified. Redundant units are eliminated and functionally overlapping parts are merged, ultimately forming a system of 12-15 core sub-units, providing clear entity mapping objects for subsequent sub-model construction. For example, after step S1, the determined sub-unit system of a large reservoir includes: concrete dam segment sub-units, dam foundation seepage prevention sub-units, spillway overflow sub-units, arc gate sub-units, water conveyance tunnel structure sub-units, reservoir water level monitoring sub-units, dam temperature monitoring sub-units, gate hoist electronic sub-units, hydrodynamic sub-units, sedimentation monitoring sub-units, dam vibration monitoring sub-units, and emergency drainage sub-units.

[0038] Building upon this foundation, proceed to step S2. Step S2, based on the characteristics of the sub-unit systems defined in step S1, combines multi-dimensional data with professional modeling methods to construct a high-fidelity digital image. Its core implementation logic is "data-driven + physical modeling + feature adaptation," meaning that differentiated modeling strategies are adopted for the structural attributes, operational patterns, and monitoring needs of different sub-units to ensure that the sub-models can accurately reflect the physical entity status and support subsequent operation and maintenance monitoring functions. Taking the large reservoir sub-unit system divided in step S1 as an example, the specific implementation process of step S2 is as follows: First, a special data collection and preprocessing process was conducted for sub-unit modeling. For the concrete dam segment sub-units, static data was collected, including structural parameters such as concrete grade (e.g., C30 / C40), pouring layer thickness, and reinforcement density, as well as geological data such as the compressive strength of the dam foundation rock mass and fault distribution. Dynamic data covered monitoring data of the dam body's vertical displacement (accuracy ±0.1mm), horizontal displacement (accuracy ±0.2mm), and joint opening and closing degree over the past 5 years, with environmental interference noise removed using filtering algorithms. For the hydrodynamic sub-units, underwater topographic data of the reservoir area (1m×1m resolution), historical hydrological yearbooks (including annual maximum flow and water level fluctuations), and real-time velocity profile data (acquired by ADCP sensors) were collected and spatiotemporally aligned to form a standardized data matrix. Secondly, a differentiated modeling method was adopted to construct sub-models. For structural sub-units, a "physical mechanism + finite element" modeling method was used: the concrete dam segment sub-units were modeled using ANSYS to establish a three-dimensional finite element model, dividing the dam body into more than 80,000 hexahedral elements, importing material constitutive parameters (elastic modulus / Poisson's ratio) and boundary conditions (dam foundation constraints, water pressure load), simulating the structural stress distribution through the static analysis module, and simulating the seismic response characteristics through the dynamic module; the water conveyance tunnel structural sub-units were modeled using FLAC³D software to construct a discrete element model, focusing on characterizing the interaction between the tunnel lining and the surrounding rock, setting key parameters such as surrounding rock mechanical parameters (cohesion, internal friction angle) and lining thickness (0.8-1.2m). For the dynamic interactive sub-units, a "data-driven + numerical simulation" integrated modeling approach is adopted: the hydrodynamic sub-unit is based on a three-dimensional hydrodynamic model built on MIKE3, with the computational domain set as the reservoir area and downstream river channel (total area 50km²), and the grid size 5m×5m. Parameters such as roughness coefficient (0.025 for the reservoir area and 0.035 for the river channel) and gravitational acceleration are imported, and the flow field distribution is simulated by coupling the SSTk-ω turbulence model; the gate hoist electronic sub-unit adopts a multi-domain modeling method, with the mechanical part based on a three-dimensional gear transmission model built on SolidWorks, and the electrical part using a motor control loop model built on MATLAB / Simulink. The two achieve collaborative simulation through interface variables (such as motor speed → gear torque). Furthermore, the characteristic parameters of the sub-models were calibrated and dynamically adapted. For the arc-shaped gate sub-unit, the relationship curve between the gate opening angle (0°-90°) and the flow coefficient was obtained through field tests, and the hydraulic loss coefficient in the model was corrected. For the sedimentation monitoring sub-unit, based on the measured sedimentation thickness data (deep measurements after the annual flood season), empirical formulas (such as Zhang Ruijin's sediment transport formula) were used to calibrate the settling velocity coefficient (1.2-1.5 cm / s) and sediment-carrying force parameters in the model. In particular, for the dam body temperature monitoring sub-unit, the temperature field distribution data collected by fiber optic grating sensors (sampling interval of 10 min) was combined with the inversion algorithm to optimize the hydration heat parameters in the model, so that the deviation between the simulated temperature and the measured value was controlled within ±1℃. Finally, sub-model association interfaces and engineering-level integration were constructed. Data interaction channels were established between the sub-models: the water level output of the hydrodynamic sub-model served as the input load for the arc gate sub-model, and the gate opening and closing state parameters were fed back to the flow model to adjust the flow area; the acceleration data of the dam vibration monitoring sub-model was associated with the fatigue damage calculation module of the concrete dam sub-model. The 12 sub-unit models were visualized and integrated using a digital twin integration platform (such as Unity3D), with a unified timestamp synchronization mechanism (10ms accuracy). This resulted in a complete digital twin engineering model containing geometric models, physical attributes, behavioral rules, and interaction relationships, laying the foundation for subsequent environmental mapping and operation and maintenance monitoring.

[0039] Step S3, "Acquiring multimodal historical monitoring data related to the target water conservancy project," provides fundamental data support for the construction of the digital twin model and environmental configuration. Taking a large reservoir water conservancy project as an example, the specific implementation process of step S3 is as follows: First, the scope and period for collecting multimodal historical monitoring data were defined. Based on the 12 sub-unit systems divided in step S1, the spatial boundary for data collection was defined as the reservoir area (including the dam, spillway, and water conveyance tunnel) and the surrounding 5km hydrological and meteorological monitoring area. The time span covered the complete operational cycle after the project was completed (for newly built projects, the trial operation period plus the data of the past 3 years was collected; for projects already in operation, the data of the past 10 years was collected). For the concrete dam segment sub-units, the collection cycle was implemented according to the principle of "one-time collection of basic data + periodic accumulation of dynamic data." Structural design parameters were collected during the project completion and acceptance phase, and dynamic monitoring data were accumulated at a daily collection frequency. For the hydrodynamic sub-units, complete hydrological cycle data for the high-water season, low-water season, and normal-water season were collected simultaneously to ensure coverage of historical samples under different hydrological conditions. Secondly, data is collected in two main categories: static sensing data and dynamic sensing data. Static sensing data focuses on the inherent properties and long-term stable parameters of the project, specifically including: Structural parameters, such as the pouring elevation of the concrete dam (accuracy ±5cm), the dam crest length (320m), the cross-sectional dimensions of the dam (upstream slope ratio 1:0.7, downstream slope ratio 1:1.5), the inner diameter of the water conveyance tunnel (3.5m), and the lining material grade (C25), are obtained through a combination of digital extraction from engineering design drawings and on-site laser scanning (accuracy ±2mm). Material properties: Collect concrete cube compressive strength (35 MPa at 28 days) and elastic modulus (3.0 × 10⁻⁶). 4 Material test data such as MPa), steel bar yield strength (335MPa), and geological survey data such as uniaxial compressive strength (80MPa) and integrity coefficient (0.85) of the dam foundation rock mass must all be associated with the corresponding test report number and test time; Basic environmental data, including the reservoir's geographical location (latitude and longitude coordinates accurate to the second), drainage area (1200 km²), and multi-year average rainfall (1200 mm), are obtained by connecting to the national hydrological database and local meteorological station historical records. Dynamic sensing data focuses on real-time changes in parameters during engineering operation, and the specific data collected includes: For structural dynamics, total stations and inclinometers were deployed at different elevations on the dam body to collect vertical displacement data (cumulative maximum value 35 mm) and horizontal displacement data (cumulative maximum value 22 mm) over the past 10 years. Piezometers were used to collect uplift pressure data of the dam foundation (range 0-500 kPa, accuracy ±0.5% FS). The data sampling frequency was once per hour. For hydrological dynamics, water level changes (highest water level 165m, lowest water level 130m) are recorded daily using water level gauges (accuracy ±1cm) deployed in the reservoir area. Flow data (range 0-500m³ / s, accuracy ±2%) is collected by ultrasonic flow meters installed at the spillway inlet. Cross-sectional velocity distribution data is obtained by combining ADCP flow meters (sampling interval 10 minutes). For equipment operation, parameters such as motor operating current (0-50A), voltage (380V±5%), operating temperature (30-60℃), and gate opening degree (0-100%) are collected for the gate hoist subunit. These parameters are exported from the historical database of the equipment control system, with a data recording interval of once per minute. Finally, standardized preprocessing and integrated storage of multimodal historical data are performed. For static sensor data, outliers are removed using a combination of manual review and automatic verification. Scanned copies of paper design drawings are converted to CAD vector format, and geological exploration data is stored in a three-dimensional matrix structure of "borehole number-depth-parameter". For dynamic sensor data, a moving average filtering algorithm is used to remove instantaneous interference noise (such as removing abrupt changes caused by sensor malfunctions), and linear interpolation is used to fill in missing data segments (each missing segment not exceeding 3 hours). The data is then uniformly converted to a time series format. By constructing a multimodal data platform, a spatiotemporal database (such as PostgreSQL + PostGIS) is used to store spatially correlated data, and a time series database (such as InfluxDB) is used to store high-frequency dynamic data. This results in a standardized historical dataset containing 8 major categories and 32 sub-parameters, with a total data volume of no less than 500GB, providing comprehensive data support for the parameter configuration of the environment in step S4.

[0040] Next, we proceed to step S4. This step involves constructing a hybrid computing architecture combining a rigid container environment and a flexible container environment, based on the characteristic attributes of the multimodal historical monitoring data and the modeling requirements of the sub-unit system. Differentiated resource configuration achieves a balance between model operating efficiency and accuracy. This step must adhere to the principles of "data-driven configuration, requirement-matching adaptation, and elastic resource scheduling" to ensure that the constructed environment meets both the stability requirements of static sub-models and the real-time update requirements of dynamic sub-models.

[0041] Taking a large reservoir water conservancy project as an example, the specific implementation process of step S4 is as follows: First, multimodal data feature analysis and container environment requirement mapping were conducted. Based on the 8 categories and 32 sub-items of multimodal historical data obtained in step S3, the computational requirements of each sub-unit system were analyzed using data feature extraction algorithms: static data accounted for 75% of the concrete dam segment sub-unit, the model parameters were updated less than once per month, and the requirements for the stability of computing resources were high (fluctuation tolerance ≤ ±2%), but the requirements for dynamic adjustment were low; dynamic data accounted for over 90% of the hydrodynamic sub-unit, the model needed to update parameters in real time with changes in water level and flow velocity (update frequency ≥ 5 minutes / time), and the requirements for the elastic expansion of computing resources were significant (peak resource requirements were 3 times the average). By establishing a mapping matrix of "data update frequency - computing resource fluctuation - container type", it was clarified that the rigid container environment is suitable for sub-models dominated by static data and with stable parameters, while the flexible container environment is suitable for sub-models dominated by dynamic data and with high-frequency parameter changes. Secondly, determine the quantity configuration and functional positioning of the rigid container environment and the flexible container environment. According to the classification results of the subunit system, set M = 3 rigid container environments with fixed specifications and N = 8 flexible container environments with variable specifications (meeting the configuration requirement of M < N). The rigid container environment is deployed with physically isolated computing nodes. Each container is allocated a fixed number of CPU cores (8 cores), memory (32GB), and storage resources (1TB SSD). Configure the Linux real-time kernel (RT_PREEMPT patch) to ensure that the computing latency is stable within 50ms. Among them, the first rigid container environment is specially deployed with sub-models of the dam body structure class (concrete dam body segment sub-units, dam foundation anti-seepage sub-units), the second rigid container environment is deployed with sub-models of the gate machinery class (arc gate sub-units, gate hoist stator models), and the third rigid container environment is deployed with sub-models of the basic environment class (reservoir area terrain sub-units). The flexible container environment is built based on the Kubernetes container orchestration platform and uses virtualization technology to achieve dynamic resource allocation. The number of CPU cores for each container can be flexibly adjusted within the range of 2 - 16 cores, the memory supports dynamic expansion from 4 - 64GB, and the storage uses a distributed file system (Ceph) to achieve on-demand expansion. The first flexible container environment is adapted to the hydrodynamic sub-unit (configured with a GPU acceleration card to support parallel computing of the flow field), the second flexible container environment is deployed with the sediment deposition prediction sub-model, and the third to eighth flexible container environments respectively correspond to dynamic sub-models such as hydrological monitoring and equipment status assessment. Data interaction between the flexible containers is achieved through a high-speed Ethernet (10Gbps) to ensure the real-time performance of model co-simulation. Furthermore, perform refined configuration of the container environment specification parameters. The rigid container environment locks the computing resources by presetting resource thresholds, sets hard limits on the upper limit of CPU usage (70%) and the peak memory occupancy (25GB), and disables the dynamic resource scheduling function to avoid model operation interruption caused by resource preemption. Taking the first rigid container environment as an example, for the finite element calculation requirements of the concrete dam body sub-model, configure an Intel Xeon Gold 6248 processor (main frequency 2.5GHz), enable the hardware acceleration instruction set (AVX2) to improve the structural mechanics calculation efficiency, and use a RAID10 array for the storage system to ensure data reliability. The flexible container environment achieves dynamic adaptation through resource elastic scaling strategies. A resource demand prediction model is built based on historical data from step S3, setting automatic scaling trigger conditions (CPU utilization ≥80% for 5 consecutive minutes) and scaling-down thresholds (CPU utilization ≤30% for 10 consecutive minutes). For the first flexible container environment housing the hydrodynamics subunit, an NVIDIA Tesla V100 GPU (32GB VRAM) is configured to support CUDA parallel computing. The number of container instances is automatically adjusted between 1 and 3 using Kubernetes HorizontalPodAutoscaler, while reserving 20% ​​redundant resources to handle sudden computing demands. Furthermore, the flexible container environment is configured with a distributed cache (Redis cluster) to accelerate dynamic data access, with cache expiration time linked to data update frequency (e.g., hydrological data cache automatically refreshes every 5 minutes). Finally, a monitoring and collaboration mechanism for the container environment is constructed. A Prometheus+Grafana monitoring system is deployed to collect real-time metrics such as container CPU utilization, memory usage, and network throughput. Rigid container environment resource fluctuation alarm thresholds are set (alarms are triggered when deviations from the baseline value are ±5%), and flexible container environment resource expansion alarm thresholds are set (warnings are issued when expansion actions are triggered three consecutive times). A container environment collaboration interface is developed to achieve standardized data interaction between rigid and flexible containers (using Protocol Buffers serialization format), with data transmission latency controlled within 100ms. For cross-container dependent sub-models (such as the coupled calculation of the arc gate sub-model and the hydrodynamic sub-model), a time synchronization mechanism between containers is established (NTP service accuracy ±1ms) to ensure the time consistency of model interaction data, laying the environmental foundation for subsequent sub-model mapping and operation.

[0042] Step S5 requires establishing precise mapping rules based on the matching relationship between the characteristics of the sub-model and the attributes of the container environment. Through differentiated deployment of rigid and flexible containers, efficient operation and collaborative simulation of the sub-model can be achieved. This step must adhere to the principles of "characteristic matching, dynamic scheduling, and collaborative operation" to ensure that the sub-model performs optimally in the suitable container environment, while also guaranteeing the overall consistency of the digital twin engineering model.

[0043] Taking a large-scale reservoir water conservancy project as an example, the specific implementation process of step S5 is as follows: First, mapping rules and priority strategies between sub-models and container environments are established. Based on the characteristics of the three rigid container environments and eight flexible container environments determined in step S4, a three-dimensional mapping matrix of "sub-model type - data characteristics - container function" is established. For sub-models with stable structural parameters and a high proportion of static data, they are preferentially mapped to rigid container environments: the concrete dam segment sub-model (75% static data) and the dam foundation seepage prevention sub-model (updated monthly) are assigned to the first rigid container environment; the arc gate sub-model (fixed mechanical structural parameters) and the gate hoist stator model (stable inherent equipment parameters) are assigned to the second rigid container environment; and the reservoir topography sub-model (topography data updated annually) is assigned to the third rigid container environment. For sub-models dominated by dynamic data and requiring frequent parameter updates, they are mapped to flexible container environments: the hydrodynamic sub-model (90% dynamic data, updated every 5 minutes) is assigned to the first flexible container environment; the sedimentation prediction sub-model (dynamically adjusted with changes in water flow) is assigned to the second flexible container environment; and sub-models such as hydrological monitoring and equipment status assessment are assigned to the third to eighth flexible container environments in sequence, ensuring that the load rate of each flexible container is controlled within the optimal range of 60%-70%. Secondly, the sub-model mapping deployment and initialization configuration are executed. A model encapsulation tool is used to package each digital twin sub-model into a standardized container image (OCI standard format), and version management and distribution are achieved through an image repository (Harbor). For the mapping deployment in the rigid container environment, a direct physical machine deployment method is adopted. The container images of the concrete dam segment sub-models are pushed to the computing node of the first rigid container environment via the SSH protocol, configured with a fixed IP address (192.168.1.101) and port (8080), and a local storage volume (1TB SSD) is mounted to save model parameters and historical calculation results. Upon startup, the basic finite element mesh data (over 80,000 elements) and material property parameters (elastic modulus 3.0 × 10⁻⁶) are automatically loaded. 4 After initialization (MPa), the system returns a "ready" status code (200 OK) via the health check interface ( / health). For the mapping and deployment of the flexible container environment, a deployment manifest (DeploymentYAML) is submitted via the Kubernetes API, specifying the resource requests (CPU=4 cores, memory=16GB) and limits (CPU=16 cores, memory=64GB) for the first flexible container environment. Node affinity rules are configured to bind to compute nodes with GPU tags, and the hydrodynamic sub-model image is deployed as 3 replicas (Pods) to achieve load balancing. During deployment, a distributed storage volume (Ceph) is automatically mounted to obtain historical hydrological data, and flow field calculation parameters (roughness coefficient 0.025, gravitational acceleration 9.81m / s²) are injected via ConfigMap. During initialization, a flow field model preheating calculation is performed (using historical data from the past 3 days for iterative convergence) to ensure stable operation within 10 minutes of startup. Furthermore, a data interaction and collaborative operation mechanism between sub-models is implemented. A cross-container communication architecture based on a message queue (Kafka) is constructed. Rigid and flexible containers achieve data subscription and publication through topics: the arc gate sub-model in the second rigid container environment publishes the gate opening angle data (0°-90°) to the "gate_angle" topic, and the hydrodynamic sub-model in the first flexible container environment subscribes to this topic, adjusting the boundary conditions of the flow field calculation in real time; the first flexible container environment publishes the calculated water pressure data to the "flow_pressure" topic, and the concrete dam sub-model in the first rigid container environment subscribes and updates the structural stress analysis parameters. At the same time, a model runtime clock synchronization mechanism is established, controlling the system time deviation of all containers within ±1ms through NTP service, and using timestamp alignment to ensure the temporal consistency of data interaction. For example, the flow field data generated by the hydrodynamic sub-model every 5 minutes carries a timestamp accurate to milliseconds, and the dam structure sub-model updates its calculation cycle accordingly. Finally, the overall operation and status monitoring of the digital twin engineering model are performed. The full quantum model is launched using the model orchestration engine (Orchestrator) of the digital twin platform, following the process of "rigid container initialization first → flexible container startup sequentially → cross-container data link verification → overall simulation startup." During the initialization phase, interface compatibility testing between sub-models is completed (verifying data format and unit consistency). A Prometheus monitoring plugin is deployed to collect operational metrics for each container: rigid containers are monitored primarily for CPU utilization (target ≤70%), memory usage (target ≤25GB), and computation latency (target ≤50ms); flexible containers are monitored primarily for resource scaling times, data transmission throughput (target ≥100Mbps), and model convergence accuracy (flow field calculation error ≤5%). The overall operational status is displayed in real-time through the Grafana dashboard. When abnormal container load occurs (e.g., rigid container CPU utilization remains ≥85% for 5 minutes) or data link interruption occurs, alarms are automatically triggered and pushed to the operations and maintenance terminal to ensure the stable operation of the digital twin engineering model. Step S500, "mapping at least the first digital twin sub-model to a first rigid container environment and mapping at least the second digital twin sub-model to a second flexible container environment," is a crucial operation for the precise matching of sub-models and their construction environments. Its core is to leverage the differences in the characteristics of the sub-models through a targeted mapping strategy to ensure that the rigid container environment supports statically stable sub-models and the flexible container environment adapts to dynamically changing sub-models, laying the foundation for the efficient operation of the digital twin engineering model. This step must adhere to the principles of "characteristic adaptation, precise resource allocation, and verifiable mapping" to ensure that different types of sub-models achieve optimal performance in their respective container environments. Taking a typical sub-model of a large reservoir water conservancy project as an example, the specific implementation process of step S500 is as follows: The first digital twin sub-model selects a segmented sub-model of the concrete dam body and a sub-model of the dam foundation seepage prevention. Both are sub-model types with stable structural parameters and dominated by static data, which meets the adaptation requirements of the rigid container environment of "fixed specifications and stable resources". The specific mapping implementation includes three stages: model encapsulation, environment configuration, and deployment verification. During the model packaging phase, Docker containerization tools were used to package the two sub-models into independent container images. The concrete dam segment sub-model image includes a finite element calculation engine (ANSYS Mechanical Runtime), a preprocessing module (mesh generation algorithm), and a basic parameter library (material properties, geometric dimensions), with the image size controlled within 8GB. The dam foundation seepage prevention sub-model image integrates the seepage calculation core (SEEP / WRuntime) and geological parameter datasets (fault distribution, rock mass permeability coefficient). During the image building process, redundant dependencies were eliminated through multi-stage build, retaining only the binary files and configuration scripts necessary for runtime, and setting image tags (such as v1.0-rigid) to identify the rigid container adaptation version. During the environment configuration phase, dedicated resource allocation was performed for the first rigid container environment: an 8-core Intel Xeon Gold 6248 processor (2.5GHz), 32GB DDR4 memory (ECC-checked), and 1TB SSD storage (RAID 10 array) were configured. The operating system used was Ubuntu 20.04 LTS with the RT_PREEMPT real-time kernel installed, keeping CPU scheduling latency below 20ms. Linux cgroups technology was used to allocate a fixed resource quota of 4 CPU cores and 16GB memory to the concrete dam segment sub-model, and the dam foundation seepage prevention sub-model was also allocated a fixed quota of 4 CPU cores and 16GB memory. Hard resource limits were set (CPU utilization capped at 70%, memory usage capped at 24GB) to prevent resource contention. Simultaneously, a model interaction interface service (based on the gRPC protocol) was deployed in the first rigid container environment, with port 8081 open for receiving real-time data input and port 8082 open for outputting calculation results. The deployment and verification phase ensures mapping validity through a three-step verification process: First, image integrity verification is performed by calculating the image's SHA256 hash value and comparing it with the baseline value (deviation must be ≤0); second, the sub-model container is started, and basic test cases are run (such as stress calculation under the dam's self-weight) to verify the deviation rate between the calculated results and the offline simulation values ​​(must be ≤2%); finally, resource stability testing is conducted, continuously monitoring the container's CPU utilization fluctuations (must be ≤±3%) and memory usage changes (must be ≤±5%) over 12 hours to ensure that the rigid container environment's resource stability meets the sub-model's requirements. After successful verification, the first rigid container environment uses a heartbeat detection mechanism (sending a status message every 30 seconds) to report the sub-model's running status to the digital twin platform, marked as "mapping completed - normal operation". The second digital twin sub-model selects a hydrodynamic sub-model and a sedimentation prediction sub-model. Both are sub-model types dominated by dynamic data and with high-frequency parameter updates, requiring dynamic adaptation based on the "variable specifications and elastic resources" characteristics of the second flexible container environment. The specific mapping implementation includes three stages: elastic resource definition, container orchestration and deployment, and dynamic adaptation verification. The elastic resource definition phase describes the resource requirements of the sub-models based on Kubernetes resource objects. The hydrodynamics sub-model, involving parallel computation of the three-dimensional flow field, defines resource requests as "CPU=4 cores, memory=16GB, GPU=1 card (NVIDIA Tesla T4)," with resource limits of "CPU=12 cores, memory=48GB, GPU=1 card." The sedimentation prediction sub-model defines resource requests as "CPU=2 cores, memory=8GB," with resource limits of "CPU=8 cores, memory=32GB." Simultaneously, HPA (HorizontalPodAutoscaler) rules are configured for the second flexible container environment. When the CPU utilization of the hydrodynamics sub-model remains ≥80% for 5 minutes, horizontal scaling of the container instance is triggered (up to a maximum of 3 replicas); when memory usage remains ≤30% for 10 minutes, vertical scaling down of resources is triggered (memory reduced to 16GB).

[0044] The container orchestration and deployment phase utilizes a Kubernetes Deployment manifest for automated deployment. The manifest specifies the sub-model image (labeled v1.0-flex), the execution command (starting the flow field calculation service: ` / usr / bin / flow_simulator --config / etc / flow / config.yaml`), and environment variables (such as the calculation accuracy level "high" and the data update frequency "5min"). Storage employs a dynamic PVC (PersistentVolumeClaim), mounting a 100GB distributed storage volume (CephRBD) to the hydrodynamics sub-model to store intermediate flow field results. Storage performance parameters (IOPS ≥ 1000) are set via StorageClass. Network configuration exposes internal ports using Service resources (flow model port 9090, sediment model port 9091), and load balancing of external access traffic is achieved through an Ingress controller, ensuring data interaction latency between sub-models is ≤ 50ms. The dynamic adaptation verification phase focuses on testing the resilient response capability of the container environment. A stress testing tool simulates load changes in the hydrodynamic sub-model: under normal water conditions, the model's CPU utilization remains stable at 40%-50%, and the container operates according to its initial resource configuration. When the simulated inflow rate suddenly increases (from 100 m³ / s to 500 m³ / s), the model's computational load rises, and the CPU utilization reaches 85% within 3 minutes, triggering HPA to automatically expand to two container replicas. Resource allocation is dynamically adjusted to an 8-core CPU and 32GB of memory, reducing computation time from 20 seconds to 8 seconds. During verification, resource scaling response time (≤60 seconds) and multi-replica data synchronization accuracy (deviation ≤1%) are recorded to ensure the second flexible container environment can dynamically adapt to changes in the sub-model's load. Finally, the container monitoring panel of the digital twin platform confirms that the mapping status between the second digital twin sub-model and the second flexible container environment is "Resilient Ready - Dynamically Adapted". To ensure clear and traceable mapping relationships, a mapping relationship visualization dashboard is constructed. This dashboard visually displays the correspondence between "first digital twin sub-model → first rigid container environment" and "second digital twin sub-model → second flexible container environment" through a topology diagram. It also labels the resource allocation parameters, operational status indicators (such as CPU utilization and computation latency), and data interaction links for each sub-model. Simultaneously, a mapping relationship configuration library is established. Container image versions, resource quotas, network parameters, and other mapping configuration information are stored in a Git repository for version management. Each mapping adjustment is auditable through change log commits, providing a configuration basis for subsequent sub-model migrations or environment optimizations. Next, step S6 establishes a closed-loop process of "real-time perception - intelligent prediction - dual-mode input - dynamic monitoring" to deeply integrate the real-time status of the physical engineering project with the digital twin model, achieving precise control and risk warning of the project's operational status. Taking a large reservoir water conservancy project as an example, the specific implementation process of step S6 is as follows: S610: Acquisition and Preprocessing of Multimodal Real-time Monitoring Data First, a comprehensive sensing and monitoring network is constructed to achieve high-frequency acquisition of key parameters of water conservancy projects. Differentiated sensing equipment is deployed for different sub-unit systems: vibrating wire strain gauges (range -2000~2000με, accuracy ±0.1%FS) and total stations (sampling frequency 1 time / 5 minutes) are installed at the top, waist, and bottom of the concrete dam sub-units to collect real-time strain and displacement data of the dam body; ultrasonic level gauges (accuracy ±1cm, sampling frequency 1 time / minute) and ADCP current meters (range 0~5m / s, accuracy ±2%) are deployed in the reservoir area and spillway to acquire real-time water level and cross-sectional flow velocity data; current sensors (range 0~100A) and temperature sensors (-40~125℃) are installed at the stator model of the gate hoist to collect the equipment operating current and winding temperature (sampling frequency 1 time / 10 seconds). At the same time, environmental data such as rainfall (accuracy 0.1mm) and wind speed (accuracy 0.1m / s) are collected synchronously by meteorological stations, forming a multimodal real-time data acquisition system covering "structure-hydrology-equipment-environment". The collected data is transmitted to edge computing nodes via industrial Ethernet (10Gbps) and private wireless network (5G) for preprocessing: a moving average filter is used to remove high-frequency noise from the strain gauge data (window size: 5 sampling points); abnormal fluctuations in water level gauge values ​​are identified and eliminated using the 3σ criterion (e.g., a single change exceeding 0.5m is considered abnormal); and the flow velocity data is spatiotemporally aligned (converted to the dam axis coordinate system). The preprocessed data is encapsulated into JSON messages in the format of "sub-unit-parameter type-timestamp" and pushed to a time-series database (InfluxDB) via the MQTT protocol. Each data entry includes metadata such as device ID, collected value, accuracy identifier, and timestamp (accurate to milliseconds) to ensure data traceability. S620: Generation and Optimization of Multimodal Prediction Fitting Data Based on preprocessed real-time data, a three-step method of "time series decomposition - intelligent prediction - error correction" is used to generate multimodal prediction fitting data. The focus of prediction is on the hydrodynamic sub-unit and the dam structure sub-unit. The first step is to use the STL (Seasonal-Trend decomposition using Loess) algorithm to decompose the multimodal real-time data, and decompose the time series data such as water level and flow velocity into trend terms (long-term change patterns), periodic terms (daily / weekly hydrological cycles) and residual terms (random fluctuations). The trend term of water level data is extracted by a sliding window (window size of 7 days), and the periodic term is set to a 24-hour cycle to match the diurnal water level changes. The second step involves constructing a Long Short-Term Memory (LSTM) neural network prediction model. The input layer includes decomposed trend terms, periodic terms, and real-time monitoring features (such as rainfall and gate opening). The hidden layer consists of three layers (64 neurons per layer), and the output layer corresponds to the predicted values ​​for the next 24 hours (with a time granularity of 1 hour). For water flow velocity prediction, the model inputs are the flow velocity sequence, water level sequence, and gate opening data from the past 12 hours, and the output is the flow velocity distribution for the next 24 hours. For dam displacement prediction, the inputs are the strain values ​​and temperature data from the past 24 hours, and the output is the displacement increment for the next 24 hours.

[0045] The third step involves dynamically optimizing the prediction results using a sliding window mechanism: a 12-hour sliding window is set, and the input data is updated hourly (adding the latest 1-hour real-time data and removing the oldest 1-hour historical data). The learning rate (0.001~0.01) and iteration count (50~200) of the LSTM are adjusted using a Bayesian optimization algorithm to ensure that the root mean square error (RMSE) of the prediction is controlled within 3%. For example, if the measured water level for a certain period is 152.3m, and the predicted fitted data is 152.1~152.5m, the error rate is 0.13%, which meets the accuracy requirements. S630: Implementation of Dual-Mode Data Input and Operation & Maintenance Monitoring A dual-mode input strategy of "real-time data-driven current status updates + predictive data-assisted trend early warning" is adopted to sequentially input data into the digital twin engineering model for operation and maintenance monitoring. The specific process is as follows: Real-time data input: Every 5 minutes, pre-processed multimodal real-time data (water level, flow velocity, strain, equipment current, etc.) are input into the corresponding sub-model of the container environment according to sub-unit classification. For example, real-time data of water level 152.3m and flow velocity 2.8m / s are input into the hydrodynamic sub-model of the first flexible container environment to trigger flow field simulation update (calculation time ≤30 seconds); real-time data of dam strain value 125με and displacement 0.02mm are input into the concrete dam sub-model of the first rigid container environment to update the structural stress distribution calculation. Forecast data input: Every hour, the multimodal forecast data for the next 24 hours is input into the model for trend monitoring and risk warning. For example, the fitted data predicting a water level rise of 0.5m in the next 6 hours is input into the flow model to simulate the stress state of the dam under high water levels; the fitted data predicting a dam displacement increment of 0.05mm is input into the structural model to assess whether there is a risk of exceeding the limit (preset displacement threshold 0.1mm / 24 hours). The monitoring process incorporates multiple levels of monitoring indicators and early warning mechanisms: basic indicators include water level deviation (early warning if real-time value deviates from design value by ≥0.3m), structural deformation rate (early warning if ≥0.01mm / h), and equipment temperature (early warning if ≥65℃); advanced indicators include flow field turbulence (early warning if flow velocity gradient ≥0.5m / s・m) and dam stress concentration factor (early warning if ≥1.2). The indicators are visualized through the real-time monitoring module of the digital twin platform (e.g., dam strain cloud map, flow field vector map). When an indicator exceeds its limit, a tiered early warning system is triggered: Level 1 (minor anomaly) pushes the warning to the operation and maintenance terminal; Level 2 (moderate anomaly) triggers an audible and visual alarm; and Level 3 (severe anomaly) automatically triggers an emergency response process (e.g., gate pre-adjustment command). Simultaneously, a data input auditing mechanism is established to ensure the integrity of real-time and predicted data through hash verification. The input time, processing time, and model response results for each data point are recorded, forming a monitoring log stored in a distributed database (MongoDB), providing data support for model correction in subsequent step S7. The entire monitoring process is achieved through collaboration between edge computing nodes and the cloud platform. The edge side is responsible for real-time data processing and rapid early warning, while the cloud side is responsible for global trend analysis and historical data tracing, ensuring the real-time and comprehensive nature of operational monitoring.

[0046] Finally, we proceed to step S7. This step establishes a comprehensive correction mechanism encompassing "monitoring deviation identification - dynamic parameter optimization - model iterative update - effect verification and feedback" to ensure that the digital twin model can adapt to changes in the physical state of the water conservancy project in real time, thereby improving the accuracy and reliability of long-term operation and maintenance monitoring. Taking a typical sub-model of a large reservoir water conservancy project as an example, the specific implementation process of step S7 is as follows: The correction process is triggered based on the operation and maintenance monitoring results of step S6, identifying model mismatch issues by constructing a multi-level deviation evaluation index system. Two types of triggering conditions are set: the basic triggering condition is "within three consecutive monitoring cycles (1 hour per cycle), the deviation rate between the calculated and measured values ​​of the sub-model exceeds a preset threshold"; the advanced triggering condition is "in a single monitoring session, the deviation rate of key indicators (such as dam strain and velocity gradient) exceeds an emergency threshold, or the warning level rises to level two or above." Specific thresholds are set differently according to the sub-model type: the displacement deviation threshold for the concrete dam segment sub-model is 5% (emergency threshold 8%), and the strain deviation threshold is 6% (emergency threshold 10%); the velocity deviation threshold for the hydrodynamic sub-model is 8% (emergency threshold 12%), and the water level deviation threshold is 3% (emergency threshold 5%).

[0047] Taking the segmented sub-model of the concrete dam as an example, during the monitoring in step S6, it was found that the calculated vertical displacement value of the dam crest (12.5 mm) and the measured value by the total station (13.5 mm) had a deviation rate of 7.4%, exceeding the basic threshold of 5% for three consecutive cycles, triggering the model correction process. Through deviation source analysis tools, the deviation was located to be mainly due to the elastic modulus parameter of the dam material (originally set to 3.0 × 10⁻⁶). 4 The pressure (MPa) does not match the actual material properties after aging, and the basic constraint conditions do not take into account the long-term creep effect of the dam foundation rock mass. A combined correction strategy of "physical parameter correction + algorithm coefficient optimization" is adopted for different types of sub-models to ensure the accuracy and relevance of the correction. For structural sub-models (such as segmented concrete dam sub-models and dam foundation seepage prevention sub-models), a physical parameter inversion correction method is used. Taking the concrete dam sub-model as an example: First, an objective function is constructed based on monitoring data, with the optimization objective being "minimizing the sum of the squares of the measured displacement values ​​and the calculated displacement values." Input parameters include the measured displacement sequence, strain sequence, and ambient temperature data for the past 30 days. Second, the particle swarm optimization (PSO) algorithm is used to invert key physical parameters, with the elastic modulus as the optimization variable (search range 2.5 × 10⁻⁶). 4 -3.2×10 4 The constraint condition is the material strength limit (≥28MPa); through iterative calculation (50 iterations), the optimal elastic modulus is obtained as 2.8×10 MPa. 4 MPa, a 6.7% reduction from the original parameter; at the same time, the dam foundation constraint conditions were modified, a rock mass creep coefficient (set to 0.002 / year) was introduced, and the boundary conditions of the finite element model were updated. For dynamic sub-models (such as hydrodynamic sub-models and sediment deposition prediction sub-models), a data-driven algorithm coefficient optimization method is adopted. Taking the hydrodynamic sub-model as an example: its velocity prediction deviation rate reached 9.2% (exceeding the 8% threshold). Deviation tracing showed that the roughness coefficient (originally set to 0.025) was not adjusted in real time with sediment deposition in the reservoir area. By establishing a correlation model between the roughness coefficient and sediment deposition thickness, and inputting the measured sediment deposition thickness data of the past 6 months (monthly post-flood measurements, maximum sediment deposition thickness 0.8m) and velocity deviation data, the roughness coefficient was optimized using the random forest regression algorithm, resulting in a new roughness coefficient value of 0.028 (an increase of 12% from the original value). At the same time, the hidden layer weights of the LSTM prediction model were fine-tuned, and incremental training was performed using the latest 7 days of real-time monitoring data (learning rate 0.001, training 20 rounds) to optimize the dynamic response capability of the prediction algorithm. The correction process is automated through a model parameter management platform: physical parameter corrections for structural sub-models are directly written to the parameter configuration file (such as ANSYS's input.dat file) in the rigid container environment via API, triggering a hot restart of the model after the update (no interruption required, restart time ≤30 seconds); algorithm coefficient optimizations for dynamic sub-models are pushed to the flexible container environment via the Kubernetes ConfigMap hot update mechanism, with the new roughness coefficients and model weights being loaded into the sub-model in real time (effectiveness delay ≤10 seconds). A complete correction record is generated during the correction process, including metadata such as correction time, trigger deviation value, original parameter value, new parameter value, and correction executor, stored in a blockchain system to ensure immutability. After the correction is completed, it must undergo three levels of verification to ensure the effect meets the standards and avoid ineffective or over-correction. The first level of verification is deviation rate recheck, which calculates the deviation rate between the output value of the corrected sub-model and the measured value, and it must be reduced to within the threshold range (e.g., ≤5% for concrete dam sub-models); the second level of verification is trend consistency test, which compares the degree of fit between the model before and after correction and the changing trend of the monitoring indicators (e.g., the consistency of the water level rise trend must be ≥90%); the third level of verification is stability test, which continuously monitors the fluctuation of the calculated results of the container for 24 hours (must be ≤±2%). Taking the hydrodynamic sub-model as an example, the model was validated after correction: Level 1 validation showed that the deviation rate between the calculated and measured flow velocity values ​​decreased from 9.2% to 4.3% (≤8% threshold); Level 2 validation, through trend line comparison, found that the predicted trend of the peak flow velocity of the corrected model matched the measured trend by 92% (originally 76%); Level 3 validation, through continuous monitoring for 24 hours, showed that the fluctuation range of the flow velocity calculation results was ±1.8% (≤±2% standard). After successful validation, the digital twin platform automatically generated a "Model Correction Effect Report," recording the comparison of key indicators before and after correction, parameter adjustment details, and validation data. Once the validated and corrected model is officially implemented, the process returns to step S6 via the closed-loop mechanism of step S7, and operation and maintenance monitoring is re-performed based on the corrected model. If validation fails (e.g., the deviation rate still exceeds the standard), a deep correction process is triggered, adding parameter adjustment dimensions (e.g., simultaneously correcting the elastic modulus and Poisson's ratio) or introducing new monitoring data (e.g., supplementing material test data from dam borehole sampling), and repeating the correction process until the standard is met. Simultaneously, a correction knowledge base is established, accumulating the triggering conditions, parameter adjustment patterns, and effect data for each correction, forming a "problem-solution-effect" correlation graph, providing a reference for the correction of similar sub-models. To ensure transparency and traceability in the correction process, a model correction visualization platform was built. This platform displays key milestones such as correction trigger time, parameter adjustment process, and verification results via a timeline, and provides comparative charts to visually represent deviation changes before and after correction (e.g., trend charts of concrete dam displacement deviation and bar charts of water flow velocity deviation). The platform integrates a correction approval process; major parameter adjustments (e.g., changes in elastic modulus exceeding 10%) require approval from operations and maintenance experts before execution. Simultaneously, correction records are synchronized to the version management system of the digital twin engineering model. Each correction version is uniquely labeled with a version number (e.g., V1.2.1_corrected), supporting version rollback functionality. When a new correction introduces an anomaly, the system can quickly revert to the previous stable version, ensuring the continuity and reliability of model operation.

[0048] As a further preferred embodiment, after mapping at least the first digital twin model to the first rigid container environment, the method further includes: Determine whether the sub-model transfer condition is met. If so, the first digital twin model is redirected to the third flexible container environment.

[0049] In this preferred example, the determination of migration conditions needs to be based on a quantitative evaluation system constructed using multi-dimensional monitoring indicators to avoid erroneous or missed migrations caused by subjective judgment. The specific determination criteria include three categories: basic indicators, dynamic indicators, and related indicators, all of which must simultaneously meet threshold requirements for three consecutive monitoring periods (30 minutes per period). The basic indicators focus on model output deviation. The threshold for the deviation rate between the calculated and measured values ​​of the first digital twin sub-model (such as the concrete dam segment sub-model) is set at 8%. When the structural displacement deviation rate is ≥8% for three consecutive cycles (e.g., calculated displacement is 12mm, measured displacement is 13.5mm, deviation rate is 12.5%), a basic condition warning is triggered. The dynamic indicators focus on the rate of change of parameters. Thresholds are set for the volatility of the model input parameters: the rate of change of dam strain ≥ 0.5 με / h, the daily fluctuation of ambient temperature ≥ 10℃, and the rate of sudden rise in water level ≥ 0.3 m / h. When any parameter exceeds the threshold for three consecutive cycles, it is determined that the dynamic conditions meet the standards. The associated indicators take into account the impact of the external environment, including extreme weather warnings (such as orange rainstorm warnings), engineering operation and maintenance (such as emergency opening and closing of gates), and geological disaster risks (such as reservoir landslide warnings). When such associated signals are received, the migration condition determination process is automatically activated. The decision-making process is automated through a migration decision engine: every 30 minutes, the engine retrieves monitoring data from the first rigid container environment from the digital twin platform, calls the deviation calculation module (based on the MAE algorithm) and the trend prediction module (based on the ARIMA model), and generates a migration condition assessment report. When the report shows that all three indicators meet the threshold requirements, the decision engine outputs a "migration condition met" command (confidence ≥ 95%) and pushes it to the model execution module via a message queue; if only some indicators meet the requirements, it outputs an "observation warning" command, extending the monitoring cycle to 15 minutes / time to continuously track the trend of indicator changes. Once the migration conditions are met, the initiator will redirect the sub-model from the first rigid container environment to the third flexible container environment. The entire process must be completed within 10 minutes to ensure continuous monitoring. The process consists of three phases: preparation, migration, and activation. The preparation phase requires reserving resources for the target environment and adapting the model image. The third flexible container environment, based on the Kubernetes API, pre-reserves dedicated resources: 8 CPU cores, 32GB memory, and 1 GPU (for accelerating structural mechanics calculations). A 1TB distributed storage (Ceph) volume is mounted for migrating model data. Simultaneously, a flexible adaptation version image (labeled v1.0-flex-migrate) of the first digital twin model is pulled from the image repository. This image adds a dynamic parameter adjustment module and an elastic computing engine to the rigid version, ensuring resource scheduling characteristics suitable for the flexible environment. The migration phase achieves model state migration through incremental data synchronization. A dual-write mechanism is employed to establish a data channel between the first rigid container environment and the third flexible container environment: First, the basic model parameters (initial values ​​of geometric dimensions and material properties) are synchronized, taking ≤2 minutes; second, the real-time calculation state of the first digital twin sub-model (such as intermediate results of finite element analysis iteratives) is captured using Checkpoint technology, generating a state snapshot (≤500MB in size) and transmitting it to the third flexible container environment. The transmission process uses an incremental synchronization algorithm (transmitting only changed data) to reduce network bandwidth consumption (peak bandwidth ≤100Mbps). During the migration, the first rigid container environment remains operational to ensure uninterrupted monitoring data. The activation phase completes the startup and verification of the new environment model. After loading the model image and state snapshot into the third flexible container environment, the initialization script is automatically executed: configuring a dynamic resource scheduling strategy (automatic scaling up when CPU utilization is ≥80%), starting a real-time parameter adjustment service (response latency ≤50ms), and establishing a new communication link with the data acquisition module (based on the WebSocket protocol). After startup, the model consistency is verified through comparative testing: calculating the model output values ​​of the old and new environments under the same input conditions (such as dam stress distribution), the deviation rate must be ≤3%; at the same time, the resource elasticity response capability of the new environment is tested, simulating a scenario of sudden increase in CPU load (from 50% to 90%), and verifying the automatic scaling time of the container (must be ≤60 seconds). After the sub-model is redirected to the third flexible container environment, targeted adaptation and optimization are required to fully leverage the dynamic adjustment advantages of the flexible environment. Regarding parameter adaptation, dynamic parameter adjustment rules are configured for the migrated sub-model: automatically adjusting the water pressure load coefficient based on real-time water level data (adjustment step size 0.01), correcting the material elastic modulus based on environmental temperature changes (modulus correction ±0.5×10³MPa for every 1℃ temperature change), and optimizing dam foundation constraints based on seepage flow data. The rule engine enables minute-level dynamic parameter updates, improving the response speed by 80% compared to static parameter configuration in a rigid environment.

[0050] For resource optimization, a predictive scheduling strategy is adopted: an LSTM prediction model is trained based on the model resource usage data of the past 7 days to predict peak CPU and memory demand one hour in advance. The third flexible container environment pre-allocates resources according to the prediction results (e.g., if the demand is predicted to be during a flood peak, it is expanded to 12 CPU cores in advance) to avoid computational delays caused by resource shortages. At the same time, a resource reclamation threshold is set. When the resource utilization rate is ≤40% for 30 consecutive minutes, a scaling-down operation is automatically triggered, reducing the number of CPU cores from 8 cores to 4 cores and the memory from 32GB to 16GB, so as to achieve efficient resource utilization. For performance monitoring, a post-migration evaluation system is established: real-time monitoring of model calculation accuracy (deviation rate ≤5%), response time (single frame calculation time ≤2 seconds), and resource utilization (CPU utilization maintained at 60%-70%) is conducted and visualized through a Grafana dashboard. If the post-migration performance is found to be below expectations (e.g., deviation rate still ≥7%), a secondary parameter optimization process is triggered, invoking a Bayesian optimization algorithm to jointly tune the resource configuration of the flexible container and model parameters until the operation and maintenance monitoring requirements are met. After the migration is complete, the digital twin platform automatically updates the model mapping table, marking the operating environment of the first digital twin sub-model as the "third flexible container environment," and records the entire migration process log (including migration time, triggering conditions, resource configuration, and verification results) through the blockchain, forming a complete traceable migration chain. Simultaneously, a model backup of the first rigid container environment is retained, and a rollback mechanism is set up. When the third flexible container environment experiences an anomaly (such as three consecutive calculation failures), the sub-model can be redirected back to the first rigid container environment within 5 minutes, ensuring system reliability in extreme scenarios.

[0051] exist Figures 1-2 Based on the method implementation examples, Figures 3-4 A schematic diagram of the corresponding system implementation is provided.

[0052] It is understood that the system implementation and the method implementation correspond to each other. Therefore, their advantages, specific implementation principles, etc. can be found in the aforementioned method implementation, and will not be described again here.

[0053] Figure 3 This diagram illustrates the hardware components of a digital twin water conservancy project operation and maintenance monitoring system based on multimodal data, according to an embodiment of the present invention.

[0054] exist Figure 3 The image shows a digital twin water conservancy project operation and maintenance monitoring system based on multimodal data.

[0055] The system includes: The model building module is used to determine multiple sub-unit systems of the target water conservancy project to be monitored, and to build a corresponding digital twin sub-model for each sub-unit system to obtain the digital twin engineering model of the target water conservancy project. The historical data acquisition module is used to acquire multimodal historical monitoring data related to the target water conservancy project, including static sensor data and dynamic sensor data. The environment configuration module is used to determine the construction environment of the digital twin model based on the multimodal data. The construction environment includes M rigid container environments with fixed specifications and N flexible container environments with variable specifications, where M and N are both positive integers and M... <N; The model execution module is used to map each digital twin sub-model to the construction environment and run the digital twin engineering model; The real-time monitoring module is used to collect multimodal real-time monitoring data related to the target water conservancy project in real time, and to predict multimodal prediction fitting data based on the multimodal real-time monitoring data; the multimodal real-time monitoring data and the multimodal prediction fitting data are sequentially input into the digital twin engineering model to perform operation and maintenance monitoring of the target water conservancy project; The model correction module is used to correct the digital twin engineering model of the target water conservancy project based on the results of operation and maintenance monitoring of the target water conservancy project.

[0056] Specifically, the model correction module is used to calculate the output deviation of the digital twin engineering model based on the operation and maintenance monitoring results. When the output deviation value exceeds the preset threshold for three consecutive cycles, the model parameter optimization process is triggered to dynamically correct the physical parameters and algorithm coefficients of the digital twin engineering model. Figure 4 This is a schematic diagram of a human-computer interaction scenario in the practical application of a digital twin water conservancy project operation and maintenance monitoring system based on multimodal data, according to an embodiment of the present invention. Figure 3 On this basis, Figure 4 The model operation module is shown to include a mapping unit and a migration unit. The mapping unit is used to map at least a first digital twin sub-model to a first rigid container environment and to map at least a second digital twin sub-model to a second flexible container environment. The migration unit is used to redirect the first digital twin sub-model to a third flexible container environment when the sub-model migration conditions are met.

[0057] The system also includes: The visualization unit is used to visualize the results of operation and maintenance monitoring of the target water conservancy project.

[0058] Human-computer interaction unit, used to provide human-computer interaction input components; The system utilizes cloud resources provided by a hybrid cloud server to implement the digital twin operation and maintenance monitoring process of the water conservancy project based on multimodal data.

[0059] Although not shown in the accompanying drawings, a preferred and more common product embodiment may also be an electronic device comprising: a memory and one or more processors. The memory stores one or more application programs adapted to be executed by the one or more processors, as described above, a digital twin-based water conservancy project operation and maintenance monitoring method based on multimodal data.

[0060] Although not shown in the accompanying drawings, further embodiments also include a computer-readable storage medium storing a computer program that, when executed, implements the steps of the aforementioned digital twin water conservancy project operation and maintenance monitoring method based on multimodal data.

[0061] It is understood that the system, product, equipment, and media implementation examples and method implementations correspond to each other and can be referenced by each other, and their principles are similar or the same, so they will not be elaborated again.

[0062] Based on practical application data verification, the advantages of this invention compared to existing technologies include at least the following: (1) This invention effectively solves the "data silo" problem caused by neglecting the multimodal characteristics of data in existing technologies by constructing a multimodal data fusion framework. Existing technologies often rely solely on dynamic monitoring data or static basic data, resulting in either distorted long-term trend analysis due to a lack of static parameter support or an inability to respond to changes in operating conditions in real time due to insufficient dynamic data. This invention clearly divides multimodal historical monitoring data into static sensing data (such as structural design parameters and material mechanical properties) and dynamic sensing data (such as water level changes and water flow pressure), and achieves spatiotemporal alignment and correlation fusion of the two through a standardized preprocessing process. For example, in the monitoring of concrete dams, it integrates static data such as dam concrete grade and geological structure as the model basis, and continuously accesses dynamic data such as real-time strain and displacement, achieving deep coupling between static parameters and dynamic changes through a data platform. This multimodal data fusion mechanism makes the model input more comprehensive. Compared with the single data source of existing technologies, the structural deformation prediction deviation rate is reduced to below 5%, effectively improving the accuracy and reliability of operation and maintenance monitoring. In particular, the model stability under complex hydrological conditions is significantly better than traditional solutions. (2) This invention innovatively designs a hybrid construction environment that combines rigid and flexible containers, overcoming the resource waste and insufficient accuracy problems caused by the "one-size-fits-all" container configuration in existing technologies. Existing technologies either use rigid environments with fixed specifications, which cannot meet the parameter update requirements of dynamic sub-models; or they rely solely on flexible environments, resulting in redundant computing resources. This invention configures the environment differently according to the characteristics of the sub-models: sub-models with fixed structural parameters, such as dam foundations and gate stators, are deployed in M ​​rigid container environments, and their operational stability is ensured by fixed computing resources (such as 8-core CPUs and 32GB of memory); dynamic sub-models such as hydrodynamics and sediment deposition are deployed in N flexible container environments, supporting elastic scheduling of CPU, memory, and other resources within the range of 2-16 cores / 4-64GB. At the same time, a sub-model migration mechanism is established, which automatically redirects the sub-models in the rigid containers to the flexible container environments when the deviation rate exceeds 8% or the parameters change drastically. This adaptation mechanism increases the resource utilization rate of rigid containers to over 70% and improves the dynamic response speed of flexible containers by 80%. Compared with the single environment configuration of existing technologies, it ensures the stable operation of static sub-models while meeting the high-frequency update requirements of dynamic sub-models, and significantly optimizes the efficiency of computing resource configuration. (3) This invention constructs a closed-loop mechanism of "real-time monitoring-prediction fitting-model correction", which makes up for the shortcomings of existing technologies in terms of weak prediction generalization ability and lack of correction mechanism. Existing prediction methods are mostly based on single historical data, without dynamic adjustment combined with real-time data, and model correction relies on human experience. This invention first decomposes multimodal real-time data through the STL algorithm to extract trend and periodic terms; then it combines the LSTM neural network to construct a prediction model, generates fitting data for the next 24 hours based on the sliding window mechanism, and controls the root mean square error of prediction within 3%. At the same time, an automated correction process is established. When the monitoring deviation exceeds the threshold for three consecutive cycles, physical parameter inversion (such as adjusting the elastic modulus of the dam body) and algorithm coefficient optimization (such as updating the roughness coefficient) are triggered, and the model is dynamically updated through hot restart. For example, in the prediction of water flow velocity, the prediction results that integrate real-time water level and historical hydrological data improve the accuracy of single data prediction by 25% compared with the existing technology; the dam structure model is dynamically corrected, and the long-term operation deviation rate is stable below 5%. This closed-loop mechanism enables the model to continuously adapt to the actual state of the project. Compared with the static model of the existing technology, the operation and maintenance early warning response time is shortened to less than 10 minutes, which significantly improves the proactive operation and maintenance capability.

[0063] Other technologies, principles, algorithms, or models not elaborated in detail in this application can be found in the prior art.

[0064] The foregoing has shown and described the method embodiments and systems of the present invention, but it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A digital twin hydraulic engineering operation and maintenance monitoring method based on multi-modal data, characterized in that, The method comprises the following steps: determine a plurality of sub-unit systems of a target water conservancy project to be monitored, construct a corresponding digital twin sub-model for each sub-unit system, and obtain a digital twin engineering model of the target water conservancy project; acquire multi-modal historical monitoring data related to the target water conservancy project, wherein the multi-modal historical monitoring data comprises static sensing data and dynamic sensing data; determine a construction environment of the digital twin model based on the multi-modal data, wherein the construction environment comprises M rigid container environments with fixed specifications and N flexible container environments with variable specifications, M and N are positive integers, and M < N; map each digital twin sub-model to the construction environment and run the digital twin engineering model; collect multi-modal real-time monitoring data related to the target water conservancy project in real time, input the data into the digital twin engineering model, and perform operation and maintenance monitoring on the target water conservancy project.

2. The digital twin hydraulic engineering operation and maintenance monitoring method based on multi-modal data according to claim 1, characterized in that, The method of mapping each digital twin sub-model to the construction environment and running the digital twin engineering model specifically comprises: mapping at least a first digital twin sub-model to a first rigid container environment and mapping at least a second digital twin sub-model to a second flexible container environment.

3. The digital twin hydraulic engineering operation and maintenance monitoring method based on multi-modal data according to claim 2, characterized in that, After mapping at least the first digital twin sub-model to the first rigid container environment, the method further comprises: determining whether a sub-model migration condition is met, if yes, redirecting the first digital twin sub-model to a third flexible container environment.

4. The digital twin hydraulic engineering operation and maintenance monitoring method based on multi-modal data according to claim 1, characterized in that, Wherein, collecting multi-modal real-time monitoring data related to the target water conservancy project in real time, inputting the data into the digital twin engineering model, and performing operation and maintenance monitoring on the target water conservancy project further comprises: predicting multi-modal prediction fitting data based on the multi-modal real-time monitoring data; inputting the multi-modal real-time monitoring data and the multi-modal prediction fitting data into the digital twin engineering model in sequence, and performing operation and maintenance monitoring on the target water conservancy project.

5. The digital twin hydraulic engineering operation and maintenance monitoring method based on multi-modal data according to claim 1, characterized in that The method further comprises: correcting the digital twin engineering model of the target water conservancy project based on the result of the operation and maintenance monitoring on the target water conservancy project.

6. The digital twin hydraulic engineering operation and maintenance monitoring method based on multi-modal data according to claim 4, characterized in that, The method of predicting multi-modal prediction fitting data based on multi-modal real-time monitoring data specifically comprises: extracting a trend item from multi-modal real-time monitoring data by using a time series decomposition algorithm, constructing a prediction model by combining a long short-term memory neural network, and generating multi-modal prediction fitting data for the next 24 hours based on a sliding window mechanism.

7. A digital twin hydraulic engineering operation and maintenance monitoring system based on multi-modal data, characterized in that, The system comprises: a model construction module configured to determine a plurality of sub-unit systems of a target water conservancy project to be monitored, construct a corresponding digital twin sub-model for each sub-unit system, and obtain a digital twin engineering model of the target water conservancy project; a historical data acquisition module configured to acquire multi-modal historical monitoring data related to the target water conservancy project, wherein the multi-modal historical monitoring data comprises static sensing data and dynamic sensing data; an environment configuration module configured to determine a construction environment of the digital twin model based on the multi-modal data, wherein the construction environment comprises M rigid container environments with fixed specifications and N flexible container environments with variable specifications, M and N are positive integers, and M < N; a model running module, configured to map each digital twin sub-model to the construction environment and run the digital twin engineering model; a real-time monitoring module, configured to collect multi-modal real-time monitoring data related to the target water conservancy project in real time, and predict multi-modal prediction fitting data based on the multi-modal real-time monitoring data; and input the multi-modal real-time monitoring data and the multi-modal prediction fitting data into the digital twin engineering model in sequence to perform operation and maintenance monitoring on the target water conservancy project; a model correction module, configured to correct the digital twin engineering model of the target water conservancy project based on the result of operation and maintenance monitoring on the target water conservancy project.

8. The digital twin water conservancy project operation and maintenance monitoring system based on multi-modal data according to claim 7, wherein the model running module comprises a mapping unit and a migration unit; the mapping unit is configured to map at least a first digital twin sub-model to a first rigid container environment and at least a second digital twin sub-model to a second flexible container environment; the migration unit is configured to redirect the first digital twin sub-model to a third flexible container environment when a sub-model migration condition is met.

9. The digital twin water conservancy project operation and maintenance monitoring system based on multi-modal data according to claim 7, wherein the model correction module is configured to calculate an output deviation of the digital twin engineering model based on the operation and maintenance monitoring result, trigger a model parameter optimization process when the output deviation value exceeds a preset threshold for three consecutive periods, and dynamically correct physical parameters and algorithm coefficients of the digital twin engineering model.

10. The digital twin hydraulic engineering operation and maintenance monitoring system based on multi-modal data of claim 7, wherein , and the system further comprises: a visualization unit, configured to visually present the result of operation and maintenance monitoring on the target water conservancy project.

Images