Water conservancy project supervision system based on big data

Abstract

The invention belongs to the technical field of hydraulic engineering supervision, and particularly relates to a hydraulic engineering supervision system based on big data. Comprising a data acquisition and preprocessing module used for collecting and processing a multi-source heterogeneous data set; the multi-modal feature extraction and alignment module is used for generating a node feature matrix; the space-time degradation modeling and fusion module is used for predicting and generating an evolution trajectory of the key performance indexes of the water conservancy project; the health trajectory prediction and risk decision module is used for calculating the future failure probability of the structure and determining whether to start intervention or not by adopting a reinforcement learning decision agent; the active intervention strategy optimization module is used for generating a group of optimal maintenance strategies; the parameterization maintenance instruction generation module is used for converting the operation instruction into a standardized operation instruction; and the digital twin visualization and execution feedback module is used for feeding back and updating the time-space diagram attention network model. According to the method, the accuracy and the reliability of predicting the long-term evolution trend of the key performance index are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy project supervision technology, specifically a water conservancy project supervision system based on big data. Background Technology

[0002] In the field of water conservancy engineering, the long-term safe and stable operation of large-scale infrastructure is of paramount importance. Its supervision process involves continuous monitoring of the health status of the engineering structure, accurate prediction of performance degradation trends, and scientific decision-making on corresponding maintenance intervention measures. Traditional supervision models mainly rely on periodic manual inspections and alarm mechanisms based on fixed thresholds. This approach has significant shortcomings in processing and utilizing the massive, multi-source, and heterogeneous data generated throughout the entire life cycle of the project.

[0003] While existing technologies employ various sensors and monitoring techniques, insufficient data fusion is a common problem. Data from different dimensions, such as sensors, drone images, hydrological and meteorological data, and historical archives, has not been effectively integrated, creating information silos. Furthermore, existing performance degradation prediction models are mostly purely data-driven models or simplified physical mechanism models. The former may produce predictions that defy common sense due to a lack of physical constraints and have weak generalization ability, while the latter struggles to capture the complex and implicit spatiotemporal relationships in the real world, resulting in insufficient prediction accuracy. In addition, maintenance decisions often rely on the experience and judgment of engineers, lacking quantitative assessment of future risks and systematic optimization of the entire life cycle cost, which can easily lead to insufficient or excessive maintenance, making it difficult to achieve the optimal balance between safety, economy, and service life.

[0004] Therefore, how to build an intelligent regulatory system that can deeply integrate data-driven approaches and physical mechanisms to solve the technical problems of insufficient data integration, inaccurate degradation prediction, and unintelligent maintenance decisions in the traditional model is a problem that urgently needs to be solved by those skilled in the art.

[0005] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention discloses a water conservancy project monitoring system based on big data. Specifically, this invention aims to provide a monitoring system, the technical solution of which includes: A big data-based water conservancy project monitoring system includes: The data acquisition and preprocessing module is used to collect and process multi-source heterogeneous datasets; The multimodal feature extraction and alignment module is used to generate a node feature matrix based on the multi-source heterogeneous dataset; The spatiotemporal degradation modeling and fusion module is used to receive the node feature matrix and use a spatiotemporal graph attention network model guided by physical mechanism to predict the evolution trajectory of key performance indicators of water conservancy projects. The health trajectory prediction and risk decision-making module is used to calculate the probability of future structural failure based on the evolution trajectory, and to use a reinforcement learning decision agent to decide whether to initiate intervention. An active intervention strategy optimization module is used to generate a set of optimal maintenance strategies when the intervention is initiated. The parameterized maintenance instruction generation module is used to convert the strategy selected from the optimal maintenance strategy into standardized work instructions. The digital twin visualization and execution feedback module is used to visualize the operation instructions and receive the actual execution data of the maintenance operation to update the spatiotemporal graph attention network model.

[0007] Preferably, the multimodal feature extraction and alignment module is specifically used for: A pre-trained convolutional neural network is used to process drone images in the multi-source heterogeneous dataset to generate crack feature tensors; The historical inspection text records in the multi-source heterogeneous dataset are processed using a natural language processing model to generate text event vectors. Based on the component IDs and acquisition timestamps contained in the building information model in the multi-source heterogeneous dataset, the crack feature tensor and the text event vector are spatiotemporally aligned. The aligned crack feature tensor and the text event vector are integrated to generate the node feature matrix.

[0008] Preferably, the spatiotemporal degradation modeling and fusion module is specifically used for: Based on the physical connection relationships and spatial proximity of the components contained in the building information model in the multi-source heterogeneous dataset, a topological graph structure of the water conservancy project is constructed. A graph attention network layer is used to aggregate spatial neighborhood information in the topological graph structure; A serial temporal convolutional network or gated recurrent unit is used to capture the temporal evolution of the node feature matrix; By combining the spatial neighborhood information and the temporal evolution law, a spatiotemporal graph attention network model guided by the physical mechanism is constructed.

[0009] Preferably, the training of the spatiotemporal graph attention network model guided by the physical mechanism is specifically used for: Calculate data-driven loss , wherein The mean squared error between the model's predicted values ​​and the actual observed values; Calculate physical mechanism loss , wherein The consistency between the model's predicted value and the expected value derived from the previous state based on a preset physical mechanism model; According to the formula In conjunction with the above The above and preset weighting coefficients Generate a hybrid loss function ; By minimizing the hybrid loss function The spatiotemporal graph attention network model guided by the physical mechanism is trained.

[0010] Preferably, the health trajectory prediction and risk decision-making module is specifically used for: Monte Carlo simulation is used to generate external load samples by random sampling based on the probability distribution of future external loads. The evolution trajectories of the external load samples and the key performance indicators are input into the spatiotemporal graph attention network model to generate multiple possible evolution trajectories of the key performance indicators. The frequency of exceeding the preset design threshold in the evolution trajectories of the multiple possible key performance indicators is statistically analyzed. Based on the frequency, the probability of failure of the structure at any future time is calculated. .

[0011] Preferably, the reinforcement learning decision agent is specifically used for: Obtain a value that includes the current structural health metric and the failure probability. status ; Based on the state Select an action from a set that includes "Continue Monitoring" actions and preset maintenance actions. implement; The selection is based on a preset reward function. The reward function is defined as follows: If the action If the reward is "continued monitoring", then the expected risk cost for the next stage is negative, and this expected risk cost is related to the probability of failure at the next moment. and estimated economic losses caused by project failure Proportional; If the action If a maintenance action is preset, its reward is the sum of the direct cost and indirect loss of that maintenance action.

[0012] Preferably, the active intervention strategy optimization module is specifically used for: Establish a multi-objective optimization problem with the objectives of minimizing total maintenance cost, minimizing the average failure probability over the entire life cycle, and maximizing the remaining useful life of the structure. The multi-objective optimization problem is solved using a multi-objective non-dominated sorting genetic algorithm to generate the optimal maintenance strategy; The optimal maintenance strategy is a Pareto optimal solution set.

[0013] Preferably, the digital twin visualization and execution feedback module is specifically used for: Load and highlight the components to be maintained in the work instructions onto the three-dimensional digital twin model; By using a mobile terminal equipped with augmented reality technology, the virtual work guidance information contained in the work instructions can be accurately superimposed onto the physical entity; Receive the actual execution data after the maintenance operation is completed; The actual execution data is used as new samples to incrementally update and train the spatiotemporal graph attention network model.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a prediction model driven by both data and physical mechanisms. It can not only learn complex spatiotemporal correlations from massive historical data, but also incorporate engineering physical laws such as material degradation formulas as constraints into the model training. This effectively avoids the problem of prediction results violating physical common sense and significantly improves the accuracy and reliability of predicting the long-term evolution trend of key performance indicators.

[0015] 2. This invention introduces a reinforcement learning decision agent, which can automatically weigh the cost of immediate repair against the risk of delayed repair, making intervention decisions with the goal of minimizing the total cost over the entire lifecycle. When initiating intervention, a multi-objective optimization algorithm is used to comprehensively consider multiple objectives such as cost, safety, and remaining lifespan, generating a set of optimal maintenance strategies covering different trade-off points for the decision-maker, transforming the decision-making process from human experience-based judgment to a scientific optimization process.

[0016] 3. This invention can integrate and process various types of data, including sensor data, UAV images, hydrological and meteorological data, historical text records, and building information models. By employing advanced artificial intelligence models such as convolutional neural networks and natural language processing, it automatically and accurately extracts deep features such as cracks and leaks from unstructured data such as images and text, and aligns and fuses them under a unified spatiotemporal benchmark, providing unprecedentedly rich and high-quality input for high-precision prediction models.

[0017] 4. This invention utilizes digital twin and augmented reality technologies to accurately visualize and convey optimized maintenance instructions to on-site operations. More importantly, the system can receive actual execution data of maintenance operations and feed it back to the system as new data samples for incremental updates and continuous optimization of the core predictive model. This enables the entire monitoring system to have the ability to continuously learn and adaptively evolve, and the accuracy of its predictions and decisions will continuously improve over time. Attached Figure Description

[0018] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the system of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0020] Example 1: A big data-based water conservancy project monitoring system includes: The data acquisition and preprocessing module is used to collect and process multi-source heterogeneous datasets; The multimodal feature extraction and alignment module is used to generate node feature matrices based on multi-source heterogeneous datasets; The spatiotemporal degradation modeling and fusion module is used to receive the node feature matrix and use a spatiotemporal graph attention network model guided by physical mechanisms to predict the evolution trajectory of key performance indicators of water conservancy projects. The health trajectory prediction and risk decision-making module is used to calculate the probability of future structural failure based on the evolution trajectory, and to use a reinforcement learning decision agent to decide whether to initiate intervention. The proactive intervention strategy optimization module is used to generate a set of optimal maintenance strategies when intervention is initiated. The parameterized maintenance instruction generation module is used to convert the strategy selected from the optimal maintenance strategy into standardized work instructions. The digital twin visualization and execution feedback module is used to visualize job instructions and receive actual execution data of maintenance jobs to update the spatiotemporal graph attention network model. This embodiment discloses a water conservancy project monitoring system based on big data, aiming to build an intelligent monitoring system that deeply integrates data-driven and physical mechanisms, and solve the technical problems of insufficient data integration, inaccurate degradation prediction, and unintelligent maintenance decision-making in the traditional model; The system comprises a series of functionally decoupled but process-connected modules, which together form a complete technical closed loop from data acquisition to decision execution. The data acquisition and preprocessing module aims to construct a comprehensive and standardized engineering data volume, providing a high-quality data source for all subsequent analyses. In this embodiment, this module continuously collects multi-source heterogeneous datasets covering the entire lifecycle of the project through various preset data interfaces. This dataset includes strain and temperature data acquired in real time using fiber Bragg grating sensors. High-resolution images of the dam surface obtained through regular aerial photography by drones. Reservoir water level and rainfall data obtained by accessing the database of meteorological and hydrological departments. Inspection and maintenance records formed after digitizing historical paper archives ; and the Building Information Model (BIM) as the anchoring basis for all data spaces. After acquisition, the data preprocessing unit automatically performs cleaning, noise reduction, and format standardization operations on the raw data; for example, for strain sensor signals. Wavelet transform is used for filtering to remove high-frequency noise; for image data Perform white balance and distortion correction; The multimodal feature extraction and alignment module's core task is to transform the diverse data from previous modules, sourced and formatted differently, into a standardized node feature matrix that can be directly used by machine learning models. This module employs targeted algorithms for feature engineering on different types of data and then applies the extracted features based on... The model provides component IDs and data acquisition timestamps for precise spatiotemporal alignment, ultimately ensuring accuracy at each discrete time point. Generate a comprehensive node feature matrix. ; The spatiotemporal degradation modeling and fusion module aims to accurately predict the future evolution trajectory of key performance indicators of water conservancy projects. This module is the core technology of this invention; it receives a series of historical node feature matrices generated by the preceding module. As input; in this embodiment, the module adopts a physical mechanism-guided spatiotemporal graph attention network model, hereinafter referred to as the PI-STGAT model; the model learns complex spatiotemporal correlations from historical data, and integrates known engineering physics laws, such as material degradation formulas, as a regularization constraint during the model training process, thereby predicting the evolution trajectory of each key performance indicator in the future. The health trajectory prediction and risk decision-making module performs a forward-looking risk assessment based on the KPI evolution trajectory predicted by the preceding module, and intelligently decides whether maintenance intervention is needed. This module, based on the predicted KPI evolution trajectory and combined with probabilistic modeling of future external load uncertainties, calculates the probability of structural failure at any future time through large-scale simulation. A reinforcement learning (RL)-based decision agent analyzes the current structural health status and future failure probability curves, and automatically decides whether to continue monitoring or initiate maintenance intervention immediately based on a preset strategy aimed at minimizing the total lifecycle cost. When the decision result is to initiate intervention, the proactive intervention strategy optimization module is activated. Its purpose is to find the optimal maintenance strategy from many possible maintenance solutions. In this embodiment, the module models the maintenance strategy optimization problem as a multi-objective optimization problem. Its optimization objectives typically include: minimizing the total maintenance cost, minimizing the average failure probability over the entire life cycle, and maximizing the remaining service life of the structure. The system uses the multi-objective non-dominated sorting genetic algorithm NSGA-II to perform a global search on the complex solution space containing variables such as maintenance time, location, materials, and processes. Finally, it generates a set of optimal maintenance strategies that achieve different trade-offs among the various objectives, which is the Pareto optimal solution set. After decision-makers select a specific solution from the optimal strategy set, the parameterized maintenance instruction generation module is responsible for transforming this abstract strategy into specific, executable, and standardized work instructions. Based on a preset instruction template, this module automatically parses and encapsulates information such as time, spatial location, maintenance measures, and required materials lists contained in the selected strategy into a series of structured parameterized maintenance instructions. ; The digital twin visualization and execution feedback module, serving as the interface between the system and the physical world, is responsible for the visual transmission of instructions and the closed-loop feedback of execution information. This module loads the previously generated work instructions into the 3D digital twin platform, highlighting the specific components and work areas to be maintained on the model. Simultaneously, it is also responsible for receiving the actual execution data after the maintenance work is completed. This on-site feedback data will be transmitted back to the system's data acquisition and preprocessing module as new data samples to incrementally update and continuously optimize the spatiotemporal graph attention network model, thereby achieving adaptive evolution and continuous performance improvement of the entire system. This embodiment constructs a complete intelligent closed-loop management system encompassing data, model, decision-making, execution, and feedback through the collaborative work of the aforementioned modules. Compared to the traditional model that relies on manual inspections and fixed threshold alarms, this system can fully integrate multi-source heterogeneous data. Through a prediction model driven by both data and physical mechanisms, it achieves accurate prediction of the evolution trajectory of structural health status. Based on a quantitative assessment of future risks, it automatically generates globally optimal proactive intervention maintenance instructions. This greatly enhances the predictability, scientific rigor, and economic efficiency of water conservancy project supervision, effectively avoids problems caused by insufficient or excessive maintenance, and ensures the safety of the project throughout its entire life cycle.

[0021] Example 2: The multimodal feature extraction and alignment module is specifically used for: A pre-trained convolutional neural network is used to process drone images from a multi-source heterogeneous dataset to generate crack feature tensors. A natural language processing model is used to process historical inspection text records in multi-source heterogeneous datasets to generate text event vectors. Based on the component IDs and acquisition timestamps contained in the building information model of the multi-source heterogeneous dataset, spatiotemporal benchmark alignment is performed on the crack feature tensor and text event vector. Integrate the aligned crack feature tensor with the text event vector to generate a node feature matrix; Based on the system in Example 1, this example further defines and optimizes the specific implementation of the multimodal feature extraction and alignment module; In this embodiment, in order to extract effective information from unstructured data more deeply, the multimodal feature extraction and alignment module adopts a variety of advanced artificial intelligence models; Specifically, for from Drone images acquired from the dataset This module employs a pre-trained convolutional neural network for processing; in this embodiment, the U-Net network architecture is selected. This network is used to perform pixel-level semantic segmentation on high-resolution images of the dam surface, automatically identifying crack regions in the image and accurately quantifying key geometric information such as their length, maximum width, and area, ultimately forming a structured crack feature tensor. ; For historical inspection records This module employs a natural language processing model; in this embodiment, the BERT model based on the Transformer architecture is selected. This model, through pre-training on a large amount of engineering corpus, is able to deeply understand the semantic information of the text. It is used to perform named entity recognition and relation extraction on inspection records, automatically identifying and extracting structural defects such as leakage, spalling, and steel corrosion, as well as descriptive information such as the location and severity of these events, ultimately forming standardized text event vectors. ; After completing feature extraction from various types of data, the core task of the module is to perform spatiotemporal alignment. Based on the component IDs and acquisition timestamps contained in the building information model of the multi-source heterogeneous dataset, the module aligns the crack feature tensor and text event vectors to a spatiotemporal reference. For example, a crack identified through image recognition... It will be associated with The unique wall component ID and the time of the drone aerial photography; similarly, a leakage event extracted from the text, its It will also be associated with the corresponding information described in the report. Component ID and recording time; This module integrates the aligned crack feature tensor with the text event vector, as well as other features such as sensor time-series feature vectors. Finally, at any discrete time point On the dam All features of each monitoring point or key structural unit are integrated into one 3D node feature matrix ;in, It is a preset node feature dimension, and its value needs to be set empirically based on the richness of feature engineering to ensure that it contains enough information for subsequent degradation modeling. By employing advanced AI models such as U-Net and BERT, this embodiment significantly improves the system's ability to automatically and accurately extract deep features from unstructured data such as images and text, reducing the subjectivity and workload of manual interpretation; more importantly, by combining with... Strict spatiotemporal alignment effectively integrates previously isolated and heterogeneous multimodal and polymorphic information into a unified feature matrix. This provides unprecedentedly rich and high-quality input for the subsequent construction of high-precision spatiotemporal prediction models, which is a key foundation for improving the prediction accuracy of the entire system.

[0022] Example 3: The spatiotemporal degradation modeling and fusion module is specifically used for: Based on the physical connection relationships and spatial proximity of the components contained in the building information model in a multi-source heterogeneous dataset, a topological graph structure of the water conservancy project is constructed. A graph attention network layer is used to aggregate spatial neighborhood information in the topological graph structure; Serialized temporal convolutional networks or gated recurrent units are used to capture the temporal evolution of node feature matrices; By combining spatial neighborhood information with temporal evolution patterns, a spatiotemporal graph attention network model guided by physical mechanisms is constructed. Based on the system in Example 1, this example elaborates on the internal construction method of the spatiotemporal graph attention network model guided by physical mechanism in the spatiotemporal degradation modeling and fusion module; In this embodiment, the model construction process fully considers the physical characteristics and time evolution of the hydraulic engineering structure; Firstly, to enable the model to understand the mechanical transmission and influence relationships within the engineering structure, this module constructs a topological graph of the hydraulic engineering project based on the physical connections and spatial proximity of components contained in the building information model from a multi-source heterogeneous dataset. In this graph structure, each node This represents a key structural unit or sensor measurement point, with the number of nodes being [number missing]. ;side This indicates that there is a direct physical connection or high spatial proximity between two units; this topological graph structure is represented in the form of an adjacency matrix in subsequent calculations, providing a basic framework for the model to capture spatial dependencies; Secondly, in order to effectively aggregate neighborhood information and identify key influencing nodes at each time step, this module adopts a graph attention network layer to aggregate spatial neighborhood information in the topological graph structure. Unlike traditional graph convolutional networks that assign the same weight to all neighbors, GAT can dynamically assign different attention weights to each neighbor node. This means that the model can autonomously learn the key path of stress or damage transmission in the structure, and GAT can capture the importance of this difference. Third, in order to capture the degradation of structural performance over time, this module connects temporal convolutional networks or gated recurrent units to capture the temporal evolution of node feature matrices; these time series models are embedded between GAT layers and are responsible for handling the dependencies of each node's own features over time. Based on the above construction, this module combines spatial neighborhood information and temporal evolution patterns to construct a spatiotemporal graph attention network model guided by physical mechanisms; specifically, at each time step, the input... First, spatial information is aggregated through the GAT layer to obtain node representations that take into account the influence of neighborhood; then, this output is fed into the TCN or GRU layer for time-dimensional information updates; multiple layers are stacked alternately in this way to achieve deep modeling of complex spatiotemporal coupled dynamic processes. This embodiment constructs an explicit topological graph structure and adopts a model architecture that combines GAT and TCN / GRU, so that the prediction model is no longer a black box that simply processes time series. The model can simultaneously and dynamically capture the spatial transmission effect and temporal cumulative effect of damage. This refined modeling of spatiotemporal dependence greatly improves the accuracy and interpretability of the model prediction.

[0023] Example 4: Training a spatiotemporal graph attention network model guided by physical mechanisms, specifically used for: Calculate data-driven loss ,in, The mean squared error between the model's predicted values ​​and the actual observed values; Calculate physical mechanism loss ,in, The consistency between the model's predicted value and the expected value derived from the previous time step based on the preset physical mechanism model; According to the formula , combined , and preset weighting coefficients Generate a hybrid loss function ; By minimizing the mixed loss function Training a spatiotemporal graph attention network model guided by physical mechanisms; Based on the system in Example 3, this example provides a detailed explanation of the training process of the spatiotemporal graph attention network model guided by physical mechanisms, especially the design of its core loss function. This training method aims to overcome the problems of weak generalization ability and prediction results that may exist in pure data-driven models by deeply integrating data and physical mechanisms. The training objective of this model is to minimize a mixture loss function. To optimize model parameters; To ensure that the model can accurately fit historical observation data, the training process needs to calculate the data-driven loss. ; The mean square error between the model's predicted values ​​and the actual observed values ​​is calculated as follows: ; In this formula, This represents the total number of nodes in the topological graph structure. The total time series step size of the training data; For the model to nodes In time The predicted values ​​of the KPIs are calculated by forward propagation of the model; For nodes In time The actual observed values ​​of the KPIs, their source is Data set; the purpose of this loss term is to constrain the model's output, making it approximate the actual measured data points; To incorporate prior knowledge from the engineering field into the model, the training process also needs to calculate the physical mechanism loss. ; To ensure consistency between the model's predicted values ​​and the expected values ​​derived from the previous time-instance state based on a pre-defined physical mechanism model; taking the carbonation depth of dam concrete as an example, its degradation process can be simplified to a linear model: ,in yes Depth of carbonization at any given moment The carbonization coefficient was obtained through experimental data calibration; based on this physical model, from... Observation status at any moment It can be deduced Expected value at time : ; Here For nodes The corresponding carbonization coefficient is set to a time step of 1. For different key performance indicators, the specific form of the physical mechanism model should be determined based on the main physical degradation process corresponding to that indicator. For example, for KPIs related to material fatigue, a cumulative damage model based on the SN curve can be introduced; for KPIs related to water erosion, an empirical erosion formula based on hydrodynamic parameters can be used. When a precise physical model is difficult to obtain, more general physical constraints can be introduced, such as ensuring that the predicted degradation trend satisfies basic physical laws such as monotonicity or energy conservation. In this way, this invention can flexibly integrate domain prior knowledge into different types of KPI predictions, enhancing the model's generalization ability and physical realism; physical mechanism loss. Then it is defined as the model in Direct prediction of time With this physical derivation, the expected value Mean square error between: ; Through this construction and The units of measurement are all squared in KPI units, achieving dimensional consistency; the role of this loss term is to impose a soft constraint on the interpolation and extrapolation behavior of the model between two real observation points, which conforms to physical dynamics. According to the formula , combined , and preset weighting coefficients Generate a hybrid loss function ;in, It is a dimensionless hyperparameter whose function is to balance the model’s fidelity to the observed data and its adherence to physical laws. Its value can be optimized through methods such as cross-validation. By minimizing the mixed loss function The spatiotemporal graph attention network model guided by physical mechanisms is trained; during the training process, the optimization algorithm simultaneously reduces... and This forces the model to learn complex correlations in the data, and its dynamic prediction process must also follow basic physical laws. This hybrid loss function design is key to achieving a deep integration of data-driven and physical mechanisms. It not only utilizes the complex and implicit correlations contained in big data, but also forces the model to comply with explicit physical laws. Therefore, the trained model has stronger generalization ability and higher reliability compared to a purely data-driven model. Especially when making long-term extrapolation predictions that lack historical data coverage, it can effectively avoid prediction results that violate physical common sense and significantly improve the confidence of the prediction.

[0024] Example 5: The health trajectory prediction and risk decision-making module is specifically used for: Monte Carlo simulation is used to generate external load samples by random sampling based on the probability distribution of future external loads. The evolution trajectories of external load samples and key performance indicators are input into the spatiotemporal graph attention network model to generate multiple possible evolution trajectories of key performance indicators. Statistically analyze the frequency of exceeding the preset design threshold in the evolution trajectories of multiple possible key performance indicators; Based on the frequency, the probability of structural failure at any future time is calculated. ; Based on the system in Example 1, this example calculates the future failure probability of the structure in the health trajectory prediction and risk decision-making module. The specific methods are explained in detail; the purpose of these methods is to quantify the impact of future uncertainties on structural safety. In this embodiment, the module uses the Monte Carlo simulation method to evaluate the failure probability; the specific implementation steps are as follows: This module fits a probability distribution of future external loads based on historical hydrological and meteorological data. For example, by analyzing flood records from the past 50 years, a probability distribution function for encountering a specific level of flood within the next year can be fitted. Based on this probability distribution, large-scale random sampling is performed to generate a large number of external load samples. In this embodiment, the number of sampling times... It is a parameter that can be set by the user, for example, to 10,000 times; For each randomly sampled external load sample, the module combines its evolution trajectory with that of key performance indicators (KPIs) as input and feeds it back into the pre-trained spatiotemporal graph attention network model. The PI-STGAT model then performs a forward inference based on this specific external load input, generating a possible KPI evolution trajectory under that load condition. This process is repeated. The system will then receive Each bar represents the evolution trajectory of KPIs with different future possibilities; The module is for this The possible evolution trajectories are analyzed, and the frequency of exceeding the preset design threshold is statistically analyzed. The preset design threshold is a structural performance limit value set according to relevant design specifications or safety standards. Its setting logic is to ensure that only statistically significant performance degradation is considered a potential failure. For each future time point... The module will check how many of these 10,000 trajectories have a KPI value that exceeds the threshold at that moment; Based on the frequency, the probability of structural failure at any future time is calculated. The calculation formula is: the number of trajectories exceeding the threshold at time t: ; Through this process, the module ultimately outputs a failure probability evolution curve over a future time span, which intuitively shows the trend of structural risk over time. By introducing Monte Carlo simulation, this embodiment upgrades the prediction of the structure's future from a single, deterministic trajectory to a probabilistic assessment of a series of possible trajectories. This method scientifically quantifies the potential impact of uncertain events such as extreme weather on structural safety, making risk assessment no longer reliant on qualitative judgments; the generated failure probability curve... This provides a solid, quantitative basis for subsequent intelligent decision-making, enabling maintenance decisions to be truly based on the prediction of future risks, thereby achieving a higher level of predictive maintenance.

[0025] Example 6: Reinforcement learning decision agents are specifically used for: Obtain a value that includes the current structural health indicators and failure probability. status ; According to the state Select an action from a set that includes both ongoing monitoring actions and preset maintenance actions. implement; The selection is based on a preset reward function. The reward function is defined as: If action To continue monitoring, the expected risk cost for the next stage, where the reward is negative, is considered in relation to the probability of failure at the next moment. and estimated economic losses caused by project failure Proportional; If action If a maintenance action is preset, its reward is the sum of the direct cost and indirect loss of that maintenance action. Based on the calculation of future failure probability in Example 5, this example describes the specific working mechanism of how the reinforcement learning decision agent decides whether to initiate intervention; the goal of the agent is to learn an optimal policy to minimize the total lifecycle cost. In this embodiment, the reinforcement learning decision agent is defined as a Markov decision process, the core elements of which are as follows: At each decision-making moment, the agent needs to obtain a data set that includes current structural health metrics and failure probability. status This state It is a multidimensional vector, which consists of the current observed values ​​of each KPI, the future KPI evolution trend predicted by the PI-STGAT model, and the failure probability curve for a future period of time calculated by the implementation method of Example 5. ; According to the state The agent selects an action from a set that includes both ongoing monitoring actions and preset maintenance actions. Execution; this set of actions is discrete, for example, continued monitoring, Level 1 maintenance, Level 2 maintenance: ; The agent selects a specific action based on its internal policy function. After training, this function can adjust the current state. Output the optimal action ; The core basis for selecting actions is the preset reward function. Its origin lies in the detailed modeling of risk and cost in engineering economics; the reward function is defined as: If action for (Continued monitoring) indicates no maintenance costs will be incurred during this period, but the risk of structural failure in the next phase must be borne; therefore, its immediate reward... The expected risk cost of the next stage is defined as negative: ; In the formula, This represents the probability of failure at the next moment. The estimated total economic loss caused by project failure is set in advance based on the importance of the project; This is the risk aversion coefficient, set by managers based on their risk tolerance. If action Preset maintenance actions This means that maintenance costs need to be incurred immediately, but the risk of future failure can be reduced; therefore, its immediate reward... The sum of the direct costs and indirect losses of this maintenance action is defined as negative: In this formula, For maintenance measures direct costs, For the purpose of performing maintenance The resulting indirect costs; The goal of an agent is to learn an optimal policy through continuous interaction in a simulated environment. To maximize long-term cumulative discount returns Its learning process follows the Bellman optimal equation; the discount factor in the Bellman optimal equation This determines the weight of future costs in current decisions; By introducing a reinforcement learning decision agent, this embodiment transforms maintenance decisions from static, passive human judgment into a dynamic, adaptive intelligent optimization process. This agent can automatically weigh the cost of immediate repair against the risk of delayed repair, and focus on optimizing the total cost over the entire lifecycle. This decision-making approach is more scientific and rational, and can avoid suboptimal maintenance caused by overly conservative or aggressive decision-makers, thereby maximizing economic benefits while ensuring structural safety.

[0026] Example 7: The proactive intervention strategy optimization module is specifically used for: Establish a multi-objective optimization problem with the objectives of minimizing total maintenance cost, minimizing the average failure probability over the entire life cycle, and maximizing the remaining useful life of the structure. A multi-objective non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem and generate the optimal maintenance strategy. The optimal maintenance strategy is a Pareto optimal solution set; After the RL agent in Example 6 decides that intervention needs to be initiated, this example provides a detailed explanation of how the proactive intervention strategy optimization module generates a set of optimal maintenance strategies. In this embodiment, the module establishes a multi-objective optimization problem with the objectives of minimizing total maintenance cost, minimizing the average failure probability over the entire life cycle, and maximizing the remaining useful life of the structure. Objective 1: Minimize total cost ; It is a vector of decision variables, representing a specific maintenance plan; Objective 2: Minimize the average failure probability over the entire lifecycle. ; Is to execute the maintenance plan Afterwards, the structure in the future The probability of failure at any given time; This refers to the entire lifecycle duration of the engineering design; Objective 3: Maximize Remaining Useful Life (RUL) ; It is an execution plan After that, the expected service life of the structure; For any candidate maintenance plan The latter two objective function values ​​are quantitatively evaluated. This embodiment uses a simulation-in-the-loop mode; specifically, the evaluation module calls the pre-trained PI-STGAT model; according to the plan... The maintenance measures defined in the code are used to correct the model's input or internal state. Using this virtually repaired state as initial conditions, the PI-STGAT model is run again for long-term forward prediction, generating a new KPI evolution trajectory. Based on this new trajectory, the Monte Carlo simulation is re-executed to calculate the plan. Failure probability curve and remaining service life Thus completing the task and Assessment; The module employs a multi-objective non-dominated sorting genetic algorithm to solve multi-objective optimization problems; the algorithm performs efficient searching in a vast solution space by simulating an evolutionary process. The algorithm's output is not a single optimal solution, but rather generates an optimal maintenance strategy, where the optimal maintenance strategy is a Pareto optimal solution set. Each solution in this set represents a different maintenance strategy that strikes a different balance between the three objectives of cost, security, and lifetime. This embodiment models the maintenance scheme selection as a multi-objective optimization problem and uses the NSGA-II algorithm to solve it, greatly improving the depth and breadth of decision-making. It provides decision-makers with a series of strategy options that achieve Pareto optimality in different performance dimensions. Decision-makers can choose the maintenance strategy that best meets their actual needs based on current budget constraints, security requirements, and long-term planning. This method makes complex engineering trade-offs transparent and scientific, achieving truly refined and optimized decision support.

[0027] Example 8: The digital twin visualization and execution feedback module is specifically used for: Load and highlight the components to be maintained in the job instructions on the 3D digital twin model; By using mobile terminals equipped with augmented reality technology, the virtual work instruction information contained in the work instructions can be accurately overlaid onto the physical entity; Receive the actual execution data after the maintenance work is completed; The actual execution data is used as new samples to incrementally update and train the spatiotemporal graph attention network model; This embodiment elaborates on the specific functions and implementation methods of the digital twin visualization and execution feedback module; this module is a key bridge connecting digital world decision-making and physical world execution, and constitutes a closed loop for the entire system's self-optimization. In this embodiment, when an optimal maintenance strategy is selected and converted into standardized work instructions... Then, this module begins execution; In terms of visualization and interaction, this module loads and highlights the components to be maintained in the work instructions on the 3D digital twin model; for example, instructions The component IDs included will be highlighted in a specific color in the twin model, along with annotation tags for the task details; To provide precise guidance for on-site operations, the module uses a mobile terminal equipped with augmented reality technology to accurately overlay the virtual operation guidance information contained in the operation instructions onto the physical entity. On-site maintenance personnel can use AR devices to see the virtual information projected by the system on the real structural surface, such as the grouting range of cracks, the precise location and depth of boreholes, etc., achieving centimeter-level construction navigation and improving the accuracy and efficiency of the operation. Regarding closed-loop feedback, the core function of the module is to collect real data after execution; after the maintenance job is completed, it receives the actual execution data after the maintenance job is completed. These data include completed photos uploaded by construction workers, lists of materials actually consumed, quality inspection reports, and high-definition aerial images taken by drones for verification. Crucially, the module uses actual execution data as new samples to incrementally update and train the spatiotemporal graph attention network model; After processing, the data is added back to the multi-source heterogeneous dataset at the system's underlying layer. The system can periodically or after receiving a large amount of new data triggers the retraining process of the PI-STGAT model. This embodiment utilizes AR technology to achieve seamless and high-precision integration from digital commands to physical execution, significantly improving the quality and efficiency of maintenance operations. More importantly, by establishing a closed-loop feedback mechanism from the execution end to the data end, each maintenance activity is transformed into a valuable learning opportunity. This enables the entire predictive maintenance system to continuously learn and adaptively optimize. As time goes by and the number of maintenance operations increases, the PI-STGAT model becomes increasingly accurate by continuously acquiring new real-world data, thus forming an ever-evolving intelligent maintenance system.

[0028] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A big data based water conservancy project supervision system, characterized in that, The method comprises the following steps: a data acquisition and preprocessing module for collecting and processing multi-source heterogeneous data sets; a multi-modal feature extraction and alignment module for generating a node feature matrix based on the multi-source heterogeneous data sets; a space-time degradation modeling and fusion module for receiving the node feature matrix and using a physical mechanism guided space-time graph attention network model to predict the evolution trajectory of the key performance indicators of the water conservancy project; a health trajectory prediction and risk decision module for calculating the future failure probability of the structure based on the evolution trajectory and using a reinforcement learning decision agent to decide whether to start intervention; an active intervention strategy optimization module for generating a set of optimal maintenance strategies when the intervention is started; a parameterized maintenance instruction generation module for converting the selected strategy from the optimal maintenance strategy into a standardized operation instruction; a digital twin visualization and execution feedback module for visualizing the operation instruction and receiving actual execution data of the maintenance operation to update the space-time graph attention network model.

2. The water conservancy project supervision system based on big data according to claim 1, wherein The multi-modal feature extraction and alignment module is specifically configured to: use a pre-trained convolutional neural network to process the unmanned aerial vehicle images in the multi-source heterogeneous data sets to generate a crack feature tensor; use a natural language processing model to process the historical inspection text records in the multi-source heterogeneous data sets to generate a text event vector; align the crack feature tensor and the text event vector in space-time reference based on the component ID and the collection timestamp in the building information model in the multi-source heterogeneous data sets; integrate the aligned crack feature tensor and the text event vector to generate the node feature matrix.

3. The water conservancy project supervision system based on big data according to claim 1, characterized in that, The space-time degradation modeling and fusion module is specifically configured to: construct a topological graph structure of the water conservancy project based on the physical connection relationship and spatial proximity of the components in the building information model in the multi-source heterogeneous data sets; use a graph attention network layer to aggregate the spatial neighborhood information in the topological graph structure; use a time convolution network or a gated recurrent unit to capture the time evolution law of the node feature matrix; combine the spatial neighborhood information and the time evolution law to construct the physical mechanism guided space-time graph attention network model.

4. The water conservancy project supervision system based on big data according to claim 3, characterized in that, The training of the physical mechanism guided space-time graph attention network model is specifically configured to: Computing data driven loss wherein the is the mean squared error between the model predicted value and the true observed value; Computing physical mechanism loss wherein the is a consistency between the model predicted value and an expected value derived from a previous time state based on a preset physical mechanism model; According to a formula , in combination with the , the and the preset weight coefficient , a hybrid loss function is generated. by minimizing the hybrid loss function training the physical mechanism guided spatio-temporal graph attention network model.

5. The water conservancy project supervision system based on big data according to claim 1, characterized in that, The health trajectory prediction and risk decision module is specifically configured to: use Monte Carlo simulation to randomly sample external load samples based on the probability distribution of future external loads; input the external load samples and the evolution trajectory of the key performance indicators into the space-time graph attention network model to generate multiple possible key performance indicator evolution trajectories; statistically analyze the frequency of exceeding the preset design threshold in the multiple possible key performance indicator evolution trajectories; depending on said frequency, calculating the failure probability of said structure at any future time .

6. The water conservancy project supervision system based on big data according to claim 5, characterized in that, The reinforcement learning decision agent is specifically configured to: obtaining a state of a current structural health indicator and the failure probability of the structure ; in accordance with the state selecting an action from a set of actions comprising a "continue monitoring" action and a preset maintenance action performing; The selection is based on a preset reward function The reward function is defined as If the action "Continue monitoring" then the reward is the negative of the expected risk cost of the next stage, which is proportional to the failure probability at the next time step and the estimated economic loss due to engineering failure If the action is a preset maintenance action, then its reward is the sum of the direct cost and the indirect loss of the maintenance action.

7. The water conservancy project supervision system based on big data according to claim 1, characterized in that, The active intervention strategy optimization module is specifically configured to: establish a multi-objective optimization problem with the optimization objectives of minimizing the total maintenance cost, minimizing the average failure probability in the whole life cycle, and maximizing the remaining service life of the structure; use a multi-objective non-dominated sorting genetic algorithm to solve the multi-objective optimization problem to generate the optimal maintenance strategy; The optimal maintenance strategy is a Pareto optimal solution set.

8. The water conservancy project supervision system based on big data according to claim 1, characterized in that, The digital twin visualization and execution feedback module is specifically used for: loading and highlighting the components to be maintained in the operation instruction on the three-dimensional digital twin model; precisely superimposing virtual operation guidance information contained in the operation instruction on the physical entity through a mobile terminal equipped with augmented reality technology; receiving the actual execution data after completion of the maintenance operation; using the actual execution data as new samples to incrementally update and train the spatio-temporal graph attention network model.

Images