Hydraulic engineering safety monitoring method and system based on data processing

Abstract

The invention relates to the technical field of data processing, in particular to a hydraulic engineering safety monitoring method and system based on data processing. The method comprises the following steps: constructing a multi-source data acquisition scheme for a water conservancy project to obtain a multi-source data acquisition scheme; performing multi-source data acquisition on the hydraulic engineering according to the multi-source data acquisition scheme to obtain an original data stream; performing monitoring node grading according to a multi-source data acquisition scheme to obtain an initial node set; performing initial edge construction and type definition according to the initial node set to obtain an initial edge list; performing different types of initial edge weight assignment on the initial edge list, and performing initial graph structure generation to obtain an initial graph structure; and carrying out correlation analysis on the original data stream, and carrying out dynamic edge weight adjustment processing on the initial graph structure to obtain a dynamic topological graph. According to the invention, intelligentization, automation and refinement of hydraulic engineering safety monitoring are realized through a data processing technology, and the engineering safety guarantee level is significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to a water conservancy project safety monitoring method and system based on data processing. Background Art

[0002] Due to the complex structure of water conservancy projects, the ever-changing operating environment, and the long-term influence of factors such as water flow, sediment, and temperature changes, various potential safety hazards are likely to occur, such as structural deformation, seepage, landslides, etc. Therefore, it is of great significance to conduct safety monitoring on water conservancy projects to timely detect and warn of potential risks and ensure the safety of the project and the safety of people's lives and property.

[0003] Traditional methods usually adopt simple threshold alarms, that is, an alarm is issued when a certain monitoring parameter exceeds a preset threshold. Therefore, it is difficult to handle complex data correlations and the early warning ability is insufficient.

[0004] The disadvantages of being difficult to handle complex data correlations are as follows: A water conservancy project is a complex system, and there are complex interactions and influences among various monitoring parameters. For example, an increase in rainfall leads to an increase in the reservoir water level, which in turn affects the seepage pressure of the dam body. Analyzing a single parameter in isolation is likely to cause misjudgment and missed judgment, ignoring the interaction between parameters. The safety state of a water conservancy project is not only related to the current monitoring data but also related to historical data and spatial location. For example, a sudden change in the displacement of a certain monitoring point indicates a risk, but if analyzed in combination with the displacement data of surrounding monitoring points, it will be found that this is only a local disturbance and does not pose a risk. Traditional methods are difficult to capture this spatio-temporal correlation.

[0005] The disadvantages of insufficient early warning ability are as follows: The setting of the threshold needs to consider various factors, such as project type, operating conditions, environmental conditions, etc. A fixed threshold is difficult to adapt to the complex and changeable actual situation, easily leading to false alarms and missed alarms. Traditional methods can only analyze the current monitoring data and cannot predict future risk trends. This makes the early warning lag and it is difficult to take preventive measures in advance. Summary of the Invention

[0006] Based on this, it is necessary to provide a water conservancy project safety monitoring method and system based on data processing to solve at least one of the above technical problems.

[0007] To achieve the above object, a water conservancy project safety monitoring method based on data processing includes the following steps:

[0008] Step S1: Construct a multi-source data acquisition plan for the water conservancy project to obtain a multi-source data acquisition plan; collect multi-source data for the water conservancy project according to the multi-source data acquisition plan to obtain an original data stream;

[0009] Step S2: Classify the monitoring nodes according to the multi-source data acquisition scheme to obtain the initial node set; construct the initial edges and define their types based on the initial node set to obtain the initial edge list; assign weights to different types of initial edges in the initial edge list and generate the initial graph structure to obtain the initial graph structure; perform correlation analysis on the original data stream to obtain the monitoring point correlation matrix; use the monitoring point correlation matrix to perform dynamic edge weight adjustment on the initial graph structure to obtain the dynamic topology map;

[0010] Step S3: Obtain the prediction requirement data of the water conservancy project; construct the ST-GCN model architecture according to the prediction requirement data of the water conservancy project to obtain the model structure configuration; construct the ST-GCN model based on the model structure configuration and the dynamic topology map, and perform the ST-GCN model;

[0011] Step S4: Use the ST-GCN model to predict the future state to obtain the predicted state sequence; conduct a preliminary risk judgment on the predicted state sequence to obtain the individual risk assessment result; perform a multi-parameter comprehensive risk assessment based on the Bayesian network according to the individual risk assessment result to obtain the comprehensive risk assessment result; generate a risk assessment report according to the comprehensive risk assessment result to obtain the risk assessment report; judge the early warning trigger conditions for the risk assessment report and issue early warning information to implement the water conservancy project safety monitoring task.

[0012] The present invention realizes the comprehensive collection, format unification and quality improvement of multi-source heterogeneous data of water conservancy projects by constructing a multi-source data collection scheme and collecting and preprocessing data according to the scheme, laying a solid data foundation for subsequent analysis. It effectively solves the problems of single data source and uneven data quality in traditional methods, and improves the reliability and availability of monitoring data. By constructing a dynamic topology map, the structural information of water conservancy projects, the spatial relationship between monitoring points and the temporal correlation of monitoring data are organically combined, providing more expressive and informative input data for subsequent deep learning models. It effectively solves the problem that traditional methods are difficult to handle complex data correlations, enabling the model to better learn the operation rules and risk patterns of water conservancy projects. By constructing and optimizing the ST-GCN model, the model can effectively learn the spatio-temporal features in the dynamic topology map and accurately predict the future state of water conservancy projects. The spatio-temporal convolution structure, attention mechanism and parameter optimization strategy of the model jointly ensure the prediction accuracy and generalization ability of the model, providing reliable prediction results for risk assessment. Using the trained ST-GCN model to predict the future state and combining with the Bayesian network for multi-parameter comprehensive risk assessment, the comprehensive and accurate assessment of the risks of water conservancy projects is realized, and automatic early warning is carried out according to the preset early warning strategy, and finally a risk assessment report is formed and early warning information is released. This step effectively solves the problems of strong subjectivity and lagging early warning in traditional methods, and realizes the intelligence and automation of the safety monitoring of water conservancy projects. Therefore, the present invention provides a method for safety monitoring of water conservancy projects based on data processing, effectively solving the drawbacks of traditional safety monitoring methods of water conservancy projects, realizing the intelligence, automation and refinement of safety monitoring, and significantly improving the engineering safety guarantee level.

[0013] Preferably, step S1 includes the following steps:

[0014] Step S11: Arrange sensors on the water conservancy project and construct a data transmission network to obtain a multi-source data collection scheme;

[0015] Step S12: Collect multi-source data of the water conservancy project according to the multi-source data collection scheme to obtain a real-time multi-source data stream;

[0016] Step S13: Perform data format conversion and verification on the real-time multi-source data stream to obtain formatted verification data;

[0017] Step S14: Integrate multi-source data and synchronize time for the formatted verification data to obtain an original data stream.

[0018] This section, by making a detailed sensor layout plan for the water conservancy project and constructing a reliable data transmission network, enables the multi-source data acquisition scheme to comprehensively cover key monitoring areas, ensuring the integrity and representativeness of the collected data and providing a reliable data basis for subsequent analysis and evaluation. The application of LoRaWAN technology reduces the power consumption and cost of data transmission, realizes remote monitoring and real-time data acquisition, and improves the monitoring efficiency. Clear sensor models and installation specifications ensure the accuracy and consistency of data acquisition, avoiding data errors caused by equipment differences or improper installation. Data acquisition according to the pre-established multi-source data acquisition scheme can obtain data from different types of sensors, forming a real-time multi-source data stream that reflects the real-time operation status of the water conservancy project. The preset sampling frequency can capture the dynamic changes in the project status and reasonably allocate sampling resources according to the characteristics of different parameters, avoiding data redundancy or information loss. Real-time data transmission and storage provide guarantee for timely detection of abnormal situations and rapid response. By performing data format conversion and verification on the real-time multi-source data stream, the original data collected by different sensors can be converted into a unified physical quantity unit and standard format, facilitating subsequent data processing and analysis. Strict data verification rules, such as range checking and outlier detection, can effectively identify and eliminate incorrect data, improve data quality, and ensure the accuracy and reliability of the analysis results. The automated process of data verification reduces the workload of manual intervention and improves the data processing efficiency. By performing multi-source data integration and time synchronization on the formatted and verified data, the data from different sensors and different sampling frequencies can be aligned to a unified time axis, forming a complete and time-synchronized original data stream. The application of the linear interpolation method effectively solves the problem of data misalignment caused by different sampling frequencies and ensures the accuracy of data analysis. Data integration and time synchronization provide the necessary data preparation for subsequent construction of dynamic topology maps and spatio-temporal analysis, laying a foundation for the training and application of deep learning models.

[0019] Preferably, step S2 includes the following steps:

[0020] Step S21: Obtain the structural design drawing of the water conservancy project; perform monitoring node mapping on the multi-source data acquisition scheme and the structural design drawing of the water conservancy project to obtain a node mapping set;

[0021] Step S22: Classify the importance of nodes in the node mapping set to obtain an initial node set;

[0022] Step S23: Construct initial edges and define their types according to the structural design drawing of the water conservancy project and the initial node set to obtain an initial edge list; assign weights to different types of initial edges in the initial edge list to obtain a weighted initial edge list;

[0023] Step S24: Generate an initial graph structure based on the weighted initial edge list and the initial edge list to obtain the initial graph structure;

[0024] Step S25: Conduct a correlation analysis of the original data stream based on historical data to obtain a monitoring point correlation matrix; perform preprocessing on the original data of the original data stream to obtain a preprocessed data stream;

[0025] Step S26: Use the monitoring point correlation matrix to adjust the dynamic edge weights of the initial graph structure to obtain a dynamically weighted graph;

[0026] Step S27: Perform data fusion on the dynamically weighted graph and the preprocessed data stream to obtain a dynamic topology graph.

[0027] By mapping the sensor information in the multi-source data acquisition scheme to the hydraulic engineering structure design drawing, the present invention establishes a spatial correspondence relationship between the sensor data and the engineering structure, connects the abstract sensor data with the specific engineering parts, and provides a basis for subsequent graph structure construction and spatial analysis. The accurate three-dimensional coordinate information of the monitoring points can accurately reflect the positions of the monitoring points in the engineering structure and improve the accuracy of subsequent analysis. The establishment of the node mapping set provides the necessary data support for subsequent node importance grading and edge connection relationship determination. Grading the importance of nodes highlights the roles and statuses of different monitoring points in engineering safety monitoring, enables the model to pay more attention to the monitoring data of key parts, and improves the prediction accuracy of the model and the accuracy of risk assessment. The graded initial node set provides a basis for subsequent construction of a more targeted graph structure and differential data analysis. By combining the hydraulic engineering structure design drawing and the initial node set, an initial edge list containing multiple edge types is constructed, which can more comprehensively reflect the spatial relationship and mutual influence between monitoring points. The weight assignment of different types of edges, such as weight assignment based on material strength and distance, makes the edge weights more physically meaningful, can more accurately express the association strength between monitoring points, and provides more effective information for subsequent graph convolution operations and feature learning. Using the weighted initial edge list and the initial node set, the initial graph structure of the hydraulic engineering is constructed, abstracting the engineering structure into a graph model, and providing a data structure basis for subsequent graph convolution operations and deep learning. This graph structure not only contains the spatial position information of the monitoring points, but also contains the connection relationship and association strength between the monitoring points, providing an important information carrier for the model to learn the structural characteristics and state change rules of the hydraulic engineering. Through the correlation analysis of the original data stream based on historical data, a monitoring point correlation matrix is obtained, which quantifies the data correlation between different monitoring points and provides a basis for subsequent dynamic adjustment of edge weights. At the same time, the original data stream is preprocessed, including outlier removal, missing value filling, and data normalization, improving the quality and consistency of the data and providing more reliable input data for subsequent model training and prediction. Using the monitoring point correlation matrix to dynamically adjust the edge weights of the initial graph structure enables the graph structure to reflect the real-time changing data correlation between monitoring points, improving the sensitivity and adaptability of the model to engineering state changes. The dynamically adjusted edge weights enable the graph structure to better capture abnormal changes in the engineering state and provide more accurate information for risk assessment and early warning. By fusing the dynamically weighted graph and the preprocessed data stream, a dynamic topological graph containing spatio-temporal information is generated, providing high-quality input data for the training and prediction of the ST-GCN model. The dynamic topological graph not only contains the topological information of the engineering structure, but also contains the real-time monitoring data of each monitoring point, enabling the model to simultaneously learn the static characteristics and dynamic change rules of the engineering structure, and thus more accurately predict future states and assess risks.

[0028] Preferably, step S23 includes the following steps:

[0029] Step S231: Mark the physical connection edges for the initial node set to obtain physical connection edges; mark the spatial proximity edges for the initial node set to obtain spatial proximity edges;

[0030] Step S232: Mark the water flow direction association edges for the initial node set to obtain water flow association edges; mark the force transmission path association edges for the initial node set to obtain force association edges;

[0031] Step S233: Construct an initial edge list based on the physical connection edges, spatial proximity edges, water flow association edges, and force association edges to obtain an initial edge list;

[0032] Step S234: Assign weights to the physical connection edges based on the material strength parameters to obtain physical connection edge weights; assign weights to the spatial proximity edges based on the reciprocal of the Euclidean distance between nodes to obtain spatial proximity edge weights;

[0033] Step S235: Assign weights to the water flow association edges based on the fixed weight value of the directed edge to obtain water flow direction association edge weights; assign weights to the force association edges based on the stress transfer coefficient to obtain force transmission path association edge weights;

[0034] Step S236: Construct a weighted initial edge list based on the physical connection edge weights, spatial proximity edge weights, water flow direction association edge weights, and force transmission path association edge weights to obtain a weighted initial edge list.

[0035] By marking physical connection edges and spatial proximity edges, two basic connection relationships between monitoring points are clarified. Physical connection edges reflect the inherent connection characteristics of engineering structures, while spatial proximity edges consider the influence of the spatial distance between monitoring points. The combined action of these two types of edges can more comprehensively describe the spatial correlation between monitoring points, laying a foundation for subsequent graph structure construction and spatial analysis. By marking water flow direction correlation edges and force transmission path correlation edges, the influence of water flow and stress transmission paths is introduced, enabling the graph structure to more accurately reflect the physical characteristics and mechanical behavior of hydraulic engineering. Water flow correlation edges reflect the influence of seepage, and force correlation edges reflect the influence of stress transmission. This information is crucial for understanding engineering state changes and risk assessment. Integrating different types of edges into the initial edge list constructs a complete list containing various connection relationships, providing a basis for subsequent weight assignment and graph structure generation. This list clearly records the various connection relationships between monitoring points, providing important topological information for subsequent analysis and modeling. Assigning weights to physical connection edges based on material strength parameters enables the edge weights to reflect the material characteristics of the connection parts. For example, edges corresponding to materials with higher strength have larger weights, more accurately expressing the strength of physical connections. Assigning weights to spatial proximity edges based on the reciprocal of the Euclidean distance considers the influence of spatial distance on the correlation between monitoring points. The closer the distance, the larger the weight, which conforms to the general law of spatial correlation. Assigning a fixed weight value to water flow correlation edges simplifies the modeling process of water flow influence while ensuring the manifestation of water flow direction information in the graph structure. Assigning weights to force correlation edges based on the stress transmission coefficient enables the edge weights to reflect the stress transmission efficiency between monitoring points, thus more accurately expressing the mechanical correlation between monitoring points. Integrating the weight information of different types of edges constructs a weighted initial edge list, providing complete edge information for subsequent generation of the initial graph structure, including connection relationships and weights. This list comprehensively considers various factors such as physical connection, spatial proximity, water flow influence, and stress transmission paths, enabling the initial graph structure to more comprehensively and accurately reflect the actual situation of hydraulic engineering.

[0036] Preferably, step S26 includes the following steps:

[0037] Step S261: Initialize the weights of the initial graph structure based on statistical correlation using the monitoring point correlation matrix to obtain a preliminary dynamic weight graph;

[0038] Step S262: Perform real-time correlation update based on a sliding window on the preprocessed data stream and the preliminary dynamic weight graph to obtain a real-time correlation matrix;

[0039] Step S263: Dynamically fuse and adjust the weights of the preliminary dynamic weight graph according to the real-time correlation matrix to obtain an intermediate dynamic weight graph spectrum;

[0040] Step S264: Perform a correlation correction based on a physical model on the intermediate dynamic weight map to obtain a dynamic weighted graph.

[0041] In the present invention, the edge weights of the initial graph structure are initialized by using the monitoring point correlation matrix, and the statistical correlation of historical data is incorporated into the graph structure, so that the preliminary dynamic weight graph can reflect the long-term data correlation pattern between monitoring points. This provides a basis for subsequent dynamic adjustment based on real-time data and enables the model to better capture the inherent correlation relationship between monitoring points. The real-time correlation update mechanism based on a sliding window can capture the latest data correlation changes between monitoring points and reflect them in the real-time correlation matrix. The use of a sliding window takes into account both the timeliness of data and avoids the influence of data fluctuations at a single time point, making the correlation analysis more stable and reliable. Through dynamic weight fusion and adjustment, the information of the real-time correlation matrix is incorporated into the preliminary dynamic weight graph, so that the graph structure can reflect the latest correlation change trend between monitoring points. The use of the exponential weighted average method balances the influence of historical data and real-time data, retains the contribution of historical information, and highlights the changes in the latest data, making the weight adjustment smoother and more stable. The correlation correction based on a physical model incorporates the prior knowledge and physical laws in the field of water conservancy projects into the graph structure, making up for the possible deficiencies in the correlation analysis that solely relies on data-driven methods. For monitoring points with strong physical correlations, even if the short-term data correlations are weak, the connection strength in the graph structure can be ensured through the correction of the physical model, improving the robustness and reliability of the model. The correlation correction based on a physical model incorporates the prior knowledge and physical laws in the field of water conservancy projects into the graph structure, making up for the possible deficiencies in the correlation analysis that solely relies on data-driven methods. For monitoring points with strong physical correlations, even if the short-term data correlations are weak, the connection strength in the graph structure can be ensured through the correction of the physical model, improving the robustness and reliability of the model

[0042] Preferably, step S3 includes the following steps:

[0043] Step S31: Construct a training data set according to the dynamic topology map to obtain a model training data set;

[0044] Step S32: Obtain water conservancy project prediction demand data; construct an ST-GCN model architecture according to the water conservancy project prediction demand data to obtain a model structure configuration;

[0045] Step S33: Obtain prediction task type data; select a loss function and an optimizer according to the model structure configuration and the prediction task type data to obtain a training parameter configuration;

[0046] Step S34: Train and validate the model according to the model structure configuration, the model training dataset, and the training parameter configuration to obtain an initial ST-GCN model;

[0047] Step S35: Evaluate the performance of the initial ST-GCN model to obtain a model evaluation report;

[0048] Step S36: Optimize the model parameters of the initial ST-GCN model using the model evaluation report to obtain the ST-GCN model.

[0049] In the present invention, by constructing a training dataset based on the dynamic topology map, the structural information and time series data of the water conservancy project are effectively combined, providing a high-quality data basis for the training of the ST-GCN model. The division method of the input-output sequence takes into account the prediction target and time dependence of the model, enabling the model to learn the evolution law of the water conservancy project state. The division of the dataset ensures the independence of model training, validation, and testing, and can more objectively evaluate the performance of the model. Constructing the ST-GCN model architecture according to the water conservancy project prediction demand data ensures that the structure of the model matches the actual prediction task. For example, according to the number and type of target variables to be predicted, the dimension of the output layer of the model is determined; according to the time granularity of the prediction, the parameters of the time convolutional layer of the model are adjusted. The rationality of the model structure configuration directly affects the prediction accuracy and efficiency of the model. Selecting an appropriate loss function and optimizer according to the type of prediction task can effectively guide the training process of the model and improve the prediction performance of the model. For example, for a regression task, it is a common practice to select the mean square error loss function and the Adam optimizer. The rationality of the training parameter configuration is crucial for the convergence speed and final performance of the model. Using the constructed model structure, training dataset, and training parameter configuration for model training and validation enables the model to learn the spatio-temporal evolution law of the water conservancy project state and initially obtain a usable model. The application of the early stopping strategy effectively prevents model overfitting and improves the generalization ability of the model. By evaluating the performance of the initial ST-GCN model and generating an evaluation report, the prediction ability and limitations of the model can be comprehensively understood, providing a reference basis for subsequent model optimization. Using multiple evaluation metrics, such as RMSE, MAE, and R^2, can more objectively evaluate the performance of the model. According to the results of the model evaluation report, optimizing the model parameters, such as adjusting the learning rate, weight decay coefficient, or modifying the network structure, can further improve the prediction accuracy and generalization ability of the model, and finally obtain an ST-GCN model with better performance, providing more reliable prediction results for the safety monitoring of water conservancy projects.

[0050] Preferably, step S32 includes the following steps:

[0051] Step S321: Generate a spatial convolutional layer configuration based on the predicted demand data of the water conservancy project through Chebyshev spectrogram convolution to obtain the spatial convolutional layer configuration;

[0052] Step S322: Construct a temporal convolutional layer according to the spatial convolutional layer configuration to obtain the temporal convolutional layer configuration;

[0053] Step S323: Integrate the temporal attention mechanism according to the spatial convolutional layer configuration and the temporal convolutional layer configuration to obtain the attention mechanism configuration;

[0054] Step S324: Design the model stacking structure based on the spatial convolutional layer configuration, the temporal convolutional layer configuration, and the attention mechanism configuration to obtain the model stacking configuration;

[0055] Step S325: Select an activation function for the model stacking configuration to obtain the activation function configuration;

[0056] Step S326: Design the model output layer according to the model stacking configuration and the activation function configuration to obtain the model structure configuration.

[0057] By adopting Chebyshev spectral graph convolution, the present invention can effectively process graph-structured data, and control the receptive field size of the convolution kernel by setting the order, so as to better capture the spatial dependence relationship between monitoring points. Determining the output feature dimension according to the prediction demand data ensures that the model can learn a sufficiently rich spatial feature representation and provide effective information for subsequent prediction tasks. The construction of the temporal convolution layer enables the model to effectively capture the dynamic change patterns in time series data. By setting appropriate convolution kernel sizes, strides, and padding methods, the receptive field and output sequence length of the temporal convolution can be controlled, so as to better extract temporal features and fuse them with spatial features. Integrating the temporal attention mechanism enables the model to focus on historical moments that have a greater impact on the current state, thereby improving the prediction accuracy of the model. By learning the weights of different time steps, the model can better capture the key information in time series data and ignore irrelevant information, improving the learning efficiency and prediction performance of the model. Combining the spatial convolution layer, the temporal convolution layer, and the attention mechanism into a spatio-temporal convolution block and performing a stacked design can better learn the complex spatio-temporal dependence relationship of the water conservancy project state. The depth and width of the stacked structure, as well as the connection method between blocks, will affect the expressive ability and learning efficiency of the model. Selecting an appropriate activation function, such as the ReLU activation function, can increase the non-linear expressive ability of the model and improve the learning ability and generalization performance of the model. The selection and position of the activation function will affect the training effect and final performance of the model. Designing the output layer according to the model stacking configuration and activation function configuration ensures that the output of the model matches the prediction target. For example, for a regression task, the output layer is usually a fully connected layer, the output dimension of which is the same as the number of prediction target variables, and no activation function is used. The design of the output layer directly affects the prediction result and final performance of the model.

[0058] Preferably, step S4 includes the following steps:

[0059] Step S41: Perform model input preprocessing on the original data stream and the dynamic topology graph to obtain a model input graph;

[0060] Step S42: Input the model input graph into the ST-GCN model for future state prediction to obtain a predicted state sequence;

[0061] Step S43: Extract monitoring point information from the dynamic topology graph to obtain monitoring point information; use the preset safety threshold data and the monitoring point information to perform a preliminary risk judgment on the predicted state sequence to obtain an individual risk assessment result;

[0062] Step S44: Perform multi-parameter comprehensive risk assessment based on the Bayesian network according to the individual risk assessment result to obtain a comprehensive risk assessment result;

[0063] Step S45: Generate a risk assessment report based on the predicted state sequence, monitoring point information, individual risk assessment results, and comprehensive risk assessment results to obtain a risk assessment report.

[0064] Step S46: Determine the risk level of the risk assessment report to obtain the determined risk level; judge the early warning trigger conditions according to the determined risk level and the preset early warning strategy configuration to obtain an early warning trigger signal;

[0065] Step S47: Generate early warning information based on the early warning trigger signal, risk assessment report, and monitoring point information to obtain the early warning information to be sent; publish the early warning information to be sent to obtain the sent early warning record.

[0066] By preprocessing the original data stream and the dynamic topology map, including outlier removal, missing value filling, and data normalization, the present invention ensures the quality and consistency of the input data, and improves the accuracy and stability of model prediction. The preprocessed data is converted into a tensor format to facilitate efficient calculation and processing by the model. The trained ST-GCN model is used to predict the future state of the model input map, obtaining a predicted state sequence, realizing the prediction of the future state of the water conservancy project, and providing a basis for risk assessment and early warning. The prediction results are output in the form of a time series, which can intuitively display the change trend of the water conservancy project state in a future period. The monitoring point information is extracted and combined with the preset safety threshold data to conduct a preliminary risk judgment on the predicted state sequence, obtaining the individual risk assessment results of each monitoring point, providing a basis for subsequent comprehensive risk assessment. By comparing the predicted value with the safety threshold, potential risk points can be quickly identified and their risk levels can be preliminarily evaluated. The multi-parameter comprehensive risk assessment method based on the Bayesian network comprehensively considers the correlation and uncertainty between multiple monitoring parameters, and can more comprehensively and accurately evaluate the overall risk level of the water conservancy project. The inference process of the Bayesian network can effectively process incomplete and uncertain information, improving the reliability of risk assessment. A risk assessment report is generated, integrating key information such as prediction results, risk assessment results, and monitoring point information into a document, facilitating relevant personnel to understand the operation status and risk situation of the water conservancy project. The charts and text descriptions included in the report make the risk information more intuitive and understandable, providing an important reference basis for decision-making. According to the risk assessment report, the risk level is determined, and combined with the early warning strategy configuration, it is judged whether an early warning needs to be triggered, realizing an automated early warning trigger mechanism. The flexibility of the early warning strategy configuration can set different early warning trigger conditions according to different risk levels and early warning requirements, improving the accuracy and efficiency of early warning. According to the early warning trigger signal, risk assessment report, and monitoring point information, early warning information is generated and sent to relevant personnel, realizing timely risk early warning and information transmission. The early warning information can be sent through multiple channels, such as text messages, emails, or a dedicated APP, ensuring the timely delivery of early warning information so that relevant personnel can take measures in time to reduce risks. The sent early warning information is recorded to facilitate subsequent early warning management and effect evaluation.

[0067] Preferably, step S44 includes the following steps:

[0068] Step S441: Obtain the water conservancy project safety evaluation index system and the historical risk event analysis report; define the basic monitoring parameter nodes for the individual risk assessment results to obtain the basic monitoring parameter nodes;

[0069] Step S442: Define intermediate risk index nodes based on the basic monitoring parameter nodes and the water conservancy project safety evaluation index system to obtain intermediate risk index nodes;

[0070] Step S443: Define final risk level nodes based on the intermediate risk index nodes, and construct a Bayesian network node definition table based on the basic monitoring parameter nodes to obtain a Bayesian network node definition table;

[0071] Step S444: Use the preset water conservancy project professional knowledge base to establish directed edges for the nodes in the Bayesian network node definition table to obtain directed edges for the nodes; perform causal relationship modeling on the Bayesian network node definition table to obtain causal relationship data;

[0072] Step S445: Determine the network structure of the Bayesian network node definition table using the historical risk event analysis report to obtain network structure data;

[0073] Step S446: Construct a Bayesian network structure based on the directed edges for the nodes, causal relationship data, and network structure data to obtain a Bayesian network structure diagram;

[0074] Step S447: Determine the conditional probability table for the Bayesian network structure diagram to obtain a Bayesian network conditional probability table;

[0075] Step S448: Input the individual risk assessment result as evidence into the Bayesian network structure diagram, and use the Bayesian network conditional probability table to perform risk inference and assessment to obtain a comprehensive risk assessment result.

[0076] The present invention provides domain knowledge and data support for constructing a Bayesian network by obtaining a safety evaluation index system for water conservancy projects and an analysis report on historical risk events, enabling the Bayesian network to more accurately reflect the actual risk situation of water conservancy projects. Based on the individual risk assessment results, the basic monitoring parameter nodes are defined, connecting the monitoring data with the Bayesian network model and providing input data for subsequent risk reasoning. The intermediate risk index nodes are defined to integrate and abstract the risk information of the basic monitoring parameter nodes, forming higher-level risk indexes such as "structural deformation risk" and "seepage safety risk", making the risk assessment more systematic and hierarchical. The final risk level nodes are defined, and a Bayesian network node definition table is constructed, completing the node definition of the Bayesian network model, clarifying each variable and its value range in the model, and providing a basis for subsequent construction of the network structure and conditional probability table. Using a professional knowledge base to establish a directed edge and causal relationship model between nodes, integrating expert experience and domain knowledge into the Bayesian network, enabling the network structure to more accurately reflect the causal relationship and influence mechanism between monitoring parameters, and improving the scientificity and reliability of risk assessment. Using the historical risk event analysis report to determine the network structure, learning the correlation relationship between risk factors from historical data and reflecting it in the Bayesian network structure, making the model more in line with the actual situation and improving the prediction ability of the model. According to the directed edges between nodes, causal relationship data, and network structure data, a Bayesian network structure diagram is constructed, completing the structure construction of the Bayesian network model, clearly showing the dependence relationship and influence path between each risk factor, and providing a framework for subsequent risk reasoning. Determining the conditional probability table injects probability information into the Bayesian network, quantifying the influence degree between risk factors, and making the risk assessment more refined and quantitative. The determination of the conditional probability table can be based on historical data statistics, expert experience, or machine learning methods to ensure the accuracy and reliability of the probability values. Taking the individual risk assessment results as evidence and inputting them into the Bayesian network, and using the conditional probability table for risk reasoning, a comprehensive risk assessment result is obtained, realizing a comprehensive assessment of the overall risk of water conservancy projects. The reasoning process of the Bayesian network can effectively process uncertain information and improve the accuracy and reliability of risk assessment. Brief Description of the Drawings

[0077] Figure 1 It is a schematic flow chart of the steps of a water conservancy project safety monitoring method based on data processing;

[0078] Figure 2 It is a schematic detailed implementation step flow chart of step S2 in the present invention.

[0079] The implementation, functional features, and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. Detailed Embodiments

[0080] The technical method of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0081] In addition, the accompanying drawings are only schematic diagrams of the present invention and are not necessarily drawn to scale. The same reference numerals in the drawings represent the same or similar parts, and thus their repeated description will be omitted. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor methods and / or microcontroller methods.

[0082] It should be understood that although terms such as "first" and "second" may be used here to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, the first unit can be called the second unit, and similarly the second unit can be called the first unit. The term "and / or" used here includes any and all combinations of one or more of the listed related items.

[0083] To achieve the above object, please refer to Figures 1 to 2 , a water conservancy project safety monitoring method based on data processing, comprising the following steps:

[0084] Step S1: Construct a multi-source data acquisition scheme for the water conservancy project to obtain a multi-source data acquisition scheme; perform multi-source data acquisition on the water conservancy project according to the multi-source data acquisition scheme to obtain an original data stream;

[0085] Step S2: Classify the monitoring nodes according to the multi-source data acquisition scheme to obtain an initial node set; construct and define the types of initial edges according to the initial node set to obtain an initial edge list; assign weights to different types of initial edges in the initial edge list and generate an initial graph structure to obtain an initial graph structure; perform correlation analysis on the original data stream to obtain a monitoring point correlation matrix; use the monitoring point correlation matrix to perform dynamic edge weight adjustment processing on the initial graph structure to obtain a dynamic topology map;

[0086] Step S3: Obtain the predicted demand data of the water conservancy project; construct the ST-GCN model architecture based on the predicted demand data of the water conservancy project to obtain the model structure configuration; construct the ST-GCN model according to the model structure configuration and the dynamic topology map, and perform the ST-GCN model;

[0087] Step S4: Use the ST-GCN model to predict the future state to obtain the predicted state sequence; conduct a preliminary risk judgment on the predicted state sequence to obtain the individual risk assessment result; conduct a multi-parameter comprehensive risk assessment based on the Bayesian network according to the individual risk assessment result to obtain the comprehensive risk assessment result; generate a risk assessment report according to the comprehensive risk assessment result to obtain the risk assessment report; judge the early warning trigger conditions for the risk assessment report and issue early warning information to implement the water conservancy project safety monitoring task.

[0088] In the embodiment of the present invention, reference Figure 1 As shown, it is a schematic diagram of the step flow of the water conservancy project safety monitoring method based on data processing of the present invention. In this example, the water conservancy project safety monitoring method based on data processing includes the following steps:

[0089] Step S1: Construct a multi-source data acquisition plan for the water conservancy project to obtain the multi-source data acquisition plan; collect multi-source data for the water conservancy project according to the multi-source data acquisition plan to obtain the original data stream;

[0090] In the embodiment of the present invention, the core is to collect data from various sensors on the dam and perform preprocessing to make it available for subsequent analysis. It mainly includes: determining the sensor type and location according to the dam structure and monitoring requirements, establishing a data transmission network (such as using LoRaWAN), collecting real-time data streams, converting the original sensor data into physical quantities (such as displacement, pressure), verifying data integrity through range checks and outlier detection (such as the 3σ criterion), and finally synchronizing and integrating the multi-source data into a unified, timestamped data stream.

[0091] Specifically: Deploy a Trimble NetR9 GNSS receiver on the dam crest and a Geokon 4500S-ATM vibrating wire piezometer inside the dam body, and transmit the data to the central server through the LoRaWAN network. Collect data at a preset frequency (such as GNSS 1Hz, piezometer 1 / 60Hz). Convert the original data into physical units using sensor-specific calibration parameters. Data verification includes range checks and 3σ outlier detection. Use linear interpolation for time synchronization and integrate the verified data into a unified CSV file containing timestamps, sensor IDs, and measurement values.

[0092] Step S2: Classify the monitoring nodes according to the multi-source data acquisition scheme to obtain the initial node set; construct the initial edges and define their types according to the initial node set to obtain the initial edge list; assign weights to the initial edges of different types in the initial edge list and generate the initial graph structure to obtain the initial graph structure; perform correlation analysis on the original data stream to obtain the monitoring point correlation matrix; use the monitoring point correlation matrix to perform dynamic edge weight adjustment processing on the initial graph structure to obtain the dynamic topology graph;

[0093] In the embodiment of the present invention, the goal is to create a dynamic graph representing the dam and its sensor network. It mainly includes: mapping the sensors to the positions on the dam CAD model, assigning importance levels according to the positions and functions of the monitoring points, defining edges based on physical connections and spatial proximity (such as a 5-meter threshold), assigning initial edge weights according to material properties (elastic modulus of physical connections, inverse distance of spatial proximity), calculating dynamic edge weights based on the correlation of real-time and historical data (using the Pearson correlation coefficient with a 7-day sliding window), and finally fusing the dynamic graph structure with the preprocessed sensor data to create a dynamic topology graph.

[0094] Specifically: Use the dam CAD model in AutoCAD to map the sensor IDs to their 3D coordinates and assign importance levels (1, 2, or 3). Create edges between physically connected nodes and nodes within 5 meters, and assign initial weights of 1.0 and inverse distance respectively. Calculate the Pearson correlation coefficient between the time series of sensor data using a 7-day sliding window. Update the edge weights according to these correlations (using a threshold of 0.8 and an update coefficient of 0.7). Combine the dynamic graph structure with the preprocessed sensor data from S1, and store the graph and node features at each time step in a custom binary file format.

[0095] Step S3: Obtain the water conservancy project prediction demand data; construct the ST-GCN model architecture according to the water conservancy project prediction demand data to obtain the model structure configuration; construct the ST-GCN model according to the model structure configuration and the dynamic topology graph, and perform the ST-GCN model;

[0096] In the embodiment of the present invention, the focus is on training a spatio-temporal graph convolutional network (ST-GCN) model. It mainly includes: preparing the training dataset by creating input-output sequences from the dynamic topology graph (e.g., T = 24 hours of input, P = 1 hour of prediction range), defining the ST-GCN architecture (using Chebyshev graph convolution and one-dimensional temporal convolution), selecting the loss function (MSE for regression) and optimizer (Adam), training the model using TensorFlow and using an early stopping strategy to prevent overfitting, evaluating the performance of the model on the test set using RMSE, MAE, and R^2, and finally optimizing the model parameters according to the evaluation results.

[0097] Specifically, a training dataset is prepared by creating sequences of 24-hour inputs and 1-hour outputs from the dynamic topology graph. An ST-GCN model consisting of two blocks is defined, which includes Chebyshev graph convolution (order 2, 64 output features), 1D temporal convolution (kernel size 3, 64 output features), and a multi-head attention mechanism (8 heads, dimension 8). The MSE loss and Adam optimizer (learning rate 0.001, weight decay 0.0001) are used. The model is trained using TensorFlow and early stopping is performed based on the validation loss. The model on the test set is evaluated using RMSE, MAE, and R^2. If the performance is below the expected threshold, the model parameters and architecture are optimized.

[0098] Step S4: Use the ST-GCN model to predict the future state to obtain a predicted state sequence; perform an initial risk judgment on the predicted state sequence to obtain an individual risk assessment result; perform a multi-parameter comprehensive risk assessment based on the Bayesian network according to the individual risk assessment result to obtain a comprehensive risk assessment result; generate a risk assessment report according to the comprehensive risk assessment result; judge the early warning trigger conditions for the risk assessment report and issue early warning information to implement the safety monitoring task of the water conservancy project;

[0099] In the embodiment of the present invention, the trained ST-GCN model is used for prediction and risk assessment, preprocess real-time data and update the dynamic topology graph, use the ST-GCN model to predict the future state (e.g., 1 hour in advance), compare the predicted value with a predefined safety threshold, perform a multi-parameter risk assessment using a Bayesian network (combining expert knowledge and historical data), generate a risk assessment report, determine the risk level according to the output of the Bayesian network, trigger an alarm according to predefined rules and the evaluated risk level (set different alarm levels for different risk severities), and finally send the alarm to the designated recipients via text message, email, or a dedicated application.

[0100] Specifically, preprocess real-time data and update the dynamic topology graph as in step S1. Use the trained ST-GCN model to predict the state 1 hour in the future. Compare the predicted value with the predefined threshold for each monitoring parameter. Input the individual risk assessment into a pre-constructed Bayesian network, use belief propagation to infer the overall risk level, and generate a risk assessment report (PDF). Determine the overall risk level. Trigger an alarm according to the risk level and predefined rules (e.g., an "alarm" level issues an alarm immediately, and a "warning" level issues an alarm after lasting for more than 1 hour). Send the alarm to the designated recipients via text message, email, or application, and record the alarm details.

[0101] Preferably, step S1 includes the following steps:

[0102] Step S11: Install sensors on the water conservancy project and construct a data transmission network to obtain a multi-source data acquisition scheme;

[0103] Step S12: Collect multi-source data from the water conservancy project according to the multi-source data acquisition scheme to obtain a real-time multi-source data stream;

[0104] Step S13: Convert and verify the data format of the real-time multi-source data stream to obtain formatted and verified data;

[0105] Step S14: Integrate multi-source data and synchronize time for the formatted and verified data to obtain an original data stream.

[0106] In the embodiment of the present invention, first, according to the dam design drawings and the predetermined monitoring requirements, determine the key parameters to be monitored, such as dam deformation, seepage pressure, temperature, water level, etc. Then, select a suitable sensor type for each monitoring parameter. For example, a GNSS receiver is used to monitor dam deformation, a vibrating wire piezometer is used to monitor seepage pressure, a PT100 temperature sensor is used to monitor temperature, and an ultrasonic water level gauge is used to monitor water level. Each sensor model must be clearly specified. For example, the GNSS receiver uses Trimble NetR9, and the vibrating wire piezometer uses Geokon 4500S-ATM. Next, according to the dam structure and the accessibility of the monitoring points, determine the specific installation positions and installation methods of each sensor. For example, the GNSS receiver is installed on the control point at the dam top and fixed by bolts; the piezometer is buried at a predetermined depth inside the dam body and connected to the data collector. All installation positions and fixing methods must be detailed in the multi-source data acquisition scheme. Finally, construct a data transmission network. Use LoRa wireless communication technology to deploy LoRaWAN gateways in the dam area and connect each sensor to a LoRaWAN node. The parameters of the LoRaWAN node, such as frequency, spreading factor, coding rate, must be configured according to the actual situation to ensure reliable data transmission. The multi-source data acquisition scheme includes all sensor types, installation positions, installation methods, and specific configuration information of the data transmission network.

[0107] According to the multi-source data acquisition scheme determined in step S11, start the installed various sensors to collect data. Each sensor collects data according to the preset sampling frequency. For example, the GNSS receiver collects data once per second, and the piezometer collects data once per minute. The sampling frequency of the sensor must be clearly defined in the multi-source data acquisition scheme. The collected data is transmitted to the data center server in real time through the LoRaWAN network. The data center server continuously receives the data packets from each LoRaWAN node and parses out the sensor ID, timestamp, and original measurement value contained in each data packet. These data constitute a real-time multi-source data stream, which is stored in the form of a time series. Each data point contains the sensor ID, timestamp, and original measurement value. The format of the data stream must be predefined, such as in JSON format or CSV format.

[0108] Perform format conversion and verification on the real-time multi-source data stream obtained in step S12. First, according to the communication protocol and data format specification of each sensor, convert the original measurement value into the corresponding physical quantity. For example, convert the original data of the GNSS receiver into longitude, latitude coordinates, and elevation values, and convert the frequency data of the piezometer into pressure values. The conversion formula must be clearly listed. For example, pressure value = a * frequency^2 + b * frequency + c, where a, b, and c are calibration coefficients. Then, verify the converted data. The verification rules must be preset in advance. For example, check whether the data is within a reasonable range, whether there are obvious mutations or outliers. The 3σ criterion or other statistical methods can be used for outlier detection. If the data verification fails, record the error information and mark the data point as invalid. The data that passes the verification is added with a verification mark and constitutes the formatted verification data together with the sensor ID and timestamp.

[0109] Perform multi-source data integration and time synchronization on the formatted verification data obtained in step S13 to finally generate the original data stream. First, align the data from different sensors according to the timestamp. Since the sampling frequencies of different sensors may be different, time synchronization processing is required. The linear interpolation method is used to interpolate the data of the sensor with a lower sampling frequency to the same time interval as the sensor with the highest sampling frequency. The interpolation method and parameters must be clearly set. For example, use linear interpolation with a time interval of 1 second. Then, sort the data of all sensors according to the timestamp and integrate them into a unified data structure. The data structure must be predefined, such as in the form of a table containing the timestamp, sensor ID, and corresponding measurement values. Finally, perform a quality check on the integrated data. For example, check whether there is data missing or duplicate. The check method must be clearly defined. For example, check whether there is data from all sensors at each timestamp. The data after integration and time synchronization constitutes the final original data stream for subsequent analysis and processing.

[0110] Preferably, step S2 includes the following steps:

[0111] Step S21: Obtain the structural design drawing of the water conservancy project; perform monitoring node mapping on the multi-source data acquisition scheme and the structural design drawing of the water conservancy project to obtain a node mapping set;

[0112] Step S22: Classify the importance of nodes in the node mapping set to obtain an initial node set;

[0113] Step S23: Perform initial edge construction and type definition according to the structural design drawing of the water conservancy project and the initial node set to obtain an initial edge list; assign weights to different types of initial edges in the initial edge list to obtain a weighted initial edge list;

[0114] Step S24: Generate an initial graph structure according to the weighted initial edge list and the initial edge list to obtain an initial graph structure;

[0115] Step S25: Perform correlation analysis based on historical data on the original data stream to obtain a monitoring point correlation matrix; perform preprocessing on the original data in the original data stream to obtain a preprocessed data stream;

[0116] Step S26: Use the monitoring point correlation matrix to adjust the dynamic edge weights of the initial graph structure to obtain a dynamically weighted graph;

[0117] Step S27: Perform data fusion on the dynamically weighted graph and the preprocessed data stream to obtain a dynamic topology graph.

[0118] As an example of the present invention, refer to Figure 2 As shown, in this example, step S2 includes:

[0119] Step S21: Obtain the structural design drawing of the water conservancy project; perform monitoring node mapping on the multi-source data acquisition scheme and the structural design drawing of the water conservancy project to obtain a node mapping set;

[0120] In an embodiment of the present invention, a CAD file of the structural design drawing of a water conservancy project is obtained. This file contains the geometric shape of the dam, coordinate information of key parts, etc. The CAD file is opened using AutoCAD software, and the three-dimensional coordinate information of all monitoring points on the dam body is extracted. The coordinate accuracy of the monitoring points must reach the millimeter level. Then, according to the multi-source data acquisition scheme generated in step S11, the ID of each sensor in the scheme is mapped one by one to the corresponding monitoring point on the structural design drawing. For example, the GNSS receiver with ID GNSS_001 is mapped to the dam crest control point A, and the piezometer with ID SP_001 is mapped to the internal monitoring point B of the dam body. Each mapping relationship must be clearly recorded to generate a node mapping set containing the sensor ID, monitoring point name, and three-dimensional coordinates. The node mapping set is stored in the CSV file format, including three columns: sensor ID, monitoring point name, X coordinate, Y coordinate, and Z coordinate.

[0121] Step S22: Perform node importance grading on the node mapping set to obtain an initial node set;

[0122] In an embodiment of the present invention, based on the water conservancy project safety monitoring specifications and expert experience, importance grading is performed on each monitoring point in the node mapping set generated in step S21. The grading standard is based on the importance of the monitoring point location and the engineering safety status it reflects. For example, the monitoring points located on the main section of the dam body and the foundation part are classified as first-level nodes with the highest importance; the monitoring points located on the dam shoulder and non-critical parts are classified as second-level nodes; the monitoring points located in the surrounding environment monitoring are classified as third-level nodes with the lowest importance. The importance level of each node is represented by the numbers 1, 2, and 3 and added to the node mapping set to form an initial node set. The initial node set is stored as a CSV file, including four columns: sensor ID, monitoring point name, three-dimensional coordinates, and importance level.

[0123] Step S23: Perform initial edge construction and type definition according to the water conservancy project structural design drawing and the initial node set to obtain an initial edge list; perform weighted assignment of different types of initial edges on the initial edge list to obtain a weighted initial edge list;

[0124] In an embodiment of the present invention, an initial edge list is constructed based on the structural design drawing of the water conservancy project and the initial node set. The connection rules are as follows: for any two monitoring points, if they are directly connected physically (for example, located in the same dam section), a "physical connection" edge is established between them; if their spatial distance is less than 10 meters, a "spatial proximity" edge is established. The type of edge must be clearly defined. The initial edge list is stored as a CSV file, containing three columns: starting node ID, ending node ID, and edge type. Then, the initial edge weights of different types are assigned to the initial edge list. The initial weight of the physically connected edge is set to 1.0, and the initial weight of the spatially adjacent edge is set to the inverse of the distance between nodes. The weight value must be clearly set. The weight value is added to the initial edge list to form a weighted initial edge list, which is stored as a CSV file, containing four columns: starting node ID, ending node ID, edge type, and weight.

[0125] Step S24: generating an initial graph structure according to the weighted initial edge list and the initial edge list to obtain an initial graph structure;

[0126] In the embodiment of the present invention, the NetworkX graph processing library is used to construct the initial graph structure based on the weighted initial edge list generated in step S23. Each monitoring point in the initial node set is used as a node of the graph, and each edge in the weighted initial edge list is used as an edge of the graph. The node attributes include the monitoring point name, three-dimensional coordinates, and importance level. The edge attributes include edge type and weight. The generated initial graph structure is saved in the GraphML file format, including node information, edge information, and their attributes.

[0127] Step S25: performing a correlation analysis on the original data stream based on historical data to obtain a monitoring point correlation matrix; performing original data preprocessing on the original data stream to obtain a preprocessed data stream;

[0128] In an embodiment of the present invention, the original data stream obtained in step S14 is subjected to a correlation analysis based on historical data. The historical data of the past year is selected to calculate the Pearson correlation coefficient of the time series data corresponding to any two monitoring points. The calculation method of the Pearson correlation coefficient must be clearly defined. The calculation results are stored in an N×N matrix, where N is the number of monitoring points, and each element of the matrix represents the correlation coefficient between the corresponding two monitoring points, forming a monitoring point association matrix. The matrix is ​​stored in CSV file format. Then, the original data stream is preprocessed. The 3σ criterion is used to eliminate outliers, and the linear interpolation method is used to fill in missing values. The preprocessing method and parameters must be clearly set. The data is normalized and the value of each monitoring parameter is scaled to the [0,1] interval. The normalization method must be clearly defined, for example, using Min-Max normalization. The preprocessed data is stored as a CSV file to form a preprocessed data stream.

[0129] Step S26: Dynamically adjust the edge weights of the initial graph structure using the monitoring point correlation matrix to obtain a dynamically weighted graph;

[0130] In the embodiment of the present invention, based on the monitoring point correlation matrix generated in step S25, dynamically adjust the edge weights of the initial graph structure generated in step S24. For each edge in the initial graph structure, if the corresponding element value (i.e., the correlation coefficient) of the two connected monitoring points in the monitoring point correlation matrix is greater than 0.8, then update the weight of this edge to this correlation coefficient; otherwise, keep the initial weight of the edge unchanged. The weight adjustment rule must be clearly defined. The adjusted graph structure is called a dynamically weighted graph and is still saved in the GraphML file format, containing the updated edge weight information.

[0131] Step S27: Perform data fusion on the dynamically weighted graph and the preprocessed data stream to obtain a dynamic topology map;

[0132] In the embodiment of the present invention, perform data fusion on the dynamically weighted graph generated in step S26 and the preprocessed data stream generated in step S25 to generate a dynamic topology map. Use the monitoring data at each time step in the preprocessed data stream as the feature vector of the corresponding node. The dynamic topology map contains the graph structure (nodes, edges, weights) at each time step and the feature vector of the nodes. The dynamic topology map is stored in a custom binary file format, and the data of one graph is stored for each time step, including node information, edge information, weight information, and node feature vectors.

[0133] Preferably, step S23 includes the following steps:

[0134] Step S231: Mark the physically connected edges for the initial node set to obtain physically connected edges; mark the spatially adjacent edges for the initial node set to obtain spatially adjacent edges;

[0135] Step S232: Mark the water flow direction associated edges for the initial node set to obtain water flow associated edges; mark the force transmission path associated edges for the initial node set to obtain force associated edges;

[0136] Step S233: Construct an initial edge list according to the physically connected edges, spatially adjacent edges, water flow associated edges, and force associated edges to obtain an initial edge list;

[0137] Step S234: Assign weights to the physically connected edges based on the material strength parameters to obtain physically connected edge weights; assign weights to the spatially adjacent edges based on the reciprocal of the Euclidean distance between nodes to obtain spatially adjacent edge weights;

[0138] Step S235: Assign weights to the water flow associated edges based on a fixed weight value for directed edges to obtain the weights of the water flow direction associated edges; assign weights to the force associated edges based on the stress transfer coefficient to obtain the weights of the force transfer path associated edges.

[0139] Step S236: Construct a weighted initial edge list based on the physical connection edge weights, spatial proximity edge weights, water flow direction associated edge weights, and force transfer path associated edge weights to obtain the weighted initial edge list.

[0140] In the embodiment of the present invention, read the initial node set CSV file to obtain the ID and three-dimensional coordinates of each monitoring point. Then, determine the physical connection relationship according to the hydraulic engineering structure design drawings. For example, for an arch dam, there is a physical connection between adjacent measurement points; for a gravity dam, there is a physical connection between the measurement points within the same dam section. Record the node pairs with physical connections and mark them as "physical connection" edges, and store them as a CSV file, including two columns of the starting node ID and the ending node ID. Next, calculate the Euclidean distance between any two nodes in the initial node set. Calculate the distance using the three-dimensional coordinates, and the formula is `distance = sqrt((x1 - x2)^2+(y1 - y2)^2+(z1 - z2)^2)`. If the distance between two nodes is less than or equal to 5 meters, establish a "spatial proximity" edge between these two nodes. Save the spatial proximity edges as a CSV file, including two columns of the starting node ID and the ending node ID.

[0141] Read the initial node set CSV file to obtain the ID, location information, and monitoring parameter type of each monitoring point. Determine the water flow direction associated edges according to the hydraulic engineering design drawings and the water flow direction. For example, for seepage monitoring points, establish directed edges from the upstream monitoring points to the downstream monitoring points according to the water flow direction from upstream to downstream. These edges are marked as "water flow associated" edges and stored as a CSV file, including two columns of the starting node ID and the ending node ID. Then, determine the force transfer path associated edges according to the finite element analysis results or other stress analysis methods. For example, for stress monitoring points, establish directed edges from the force application point to the force receiving point according to the stress transfer path. These edges are marked as "force associated" edges and also stored as a CSV file, including two columns of the starting node ID and the ending node ID.

[0142] Read the four CSV files generated in steps S231 and S232: physically connected edges, spatially adjacent edges, water flow associated edges, and force associated edges. Merge the edge information in these four files into a new CSV file to form an initial edge list. This CSV file contains three columns: starting node ID, ending node ID, and edge type. The edge type field indicates the type of the edge, such as "physically connected", "spatially adjacent", "water flow associated", or "force associated". Ensure that there are no duplicate edges during the merging process. If duplicate edges occur, keep one and record the duplicate information.

[0143] Read the initial edge list generated in step S233 and the hydraulic engineering material parameter table. For each "physically connected" edge, find the corresponding elastic modulus from the material parameter table according to the material type and strength grade of the connection part. Set the weight of the edge to the elastic modulus value. For example, if the material of the connection part is C30 concrete, the weight of the edge is set to 3.25×10^10 Pa. For each "spatially adjacent" edge, calculate the Euclidean distance `d` between the two nodes, and set the weight of the edge to `1 / d`. Keep four significant figures in the calculation result. Store the calculated weight values in two new CSV files, corresponding to the physically connected edge weights and the spatially adjacent edge weights respectively. Each file contains three columns: starting node ID, ending node ID, and weight.

[0144] Read the initial edge list generated in step S233 and the finite element analysis result file. For each "water flow associated" edge, assign a fixed weight value of 0.5 to represent the influence of the water flow direction. Save the result to a CSV file containing three columns: starting node ID, ending node ID, and weight. For each "force associated" edge, extract the stress transfer coefficient connecting the two nodes from the finite element analysis result file. Set the weight of the edge to the absolute value of the stress transfer coefficient. Save the result to another CSV file containing three columns: starting node ID, ending node ID, and weight.

[0145] Read the four CSV files generated in steps S234 and S235: physically connected edge weights, spatially adjacent edge weights, water flow associated edge weights, and force associated edge weights. Create a new CSV file named weighted initial edge list. Merge all the edge information in the four files into the weighted initial edge list. This file contains four columns: starting node ID, ending node ID, edge type, and weight. Ensure that the weight value of each edge matches its corresponding edge type.

[0146] Preferably, step S26 includes the following steps:

[0147] Step S261: Use the monitoring point correlation matrix to perform weight initialization based on statistical correlation for the initial graph structure to obtain a preliminary dynamic weight graph;

[0148] Step S262: Perform real-time correlation update based on a sliding window on the preprocessed data stream and the preliminary dynamic weight graph to obtain a real-time correlation matrix;

[0149] Step S263: Perform dynamic weight fusion and adjustment on the preliminary dynamic weight graph according to the real-time correlation matrix to obtain an intermediate dynamic weight graph;

[0150] Step S264: Perform correlation correction based on a physical model on the intermediate dynamic weight graph to obtain a dynamically weighted graph.

[0151] In the embodiment of the present invention, the initial graph structure (GraphML format) generated in step S24 and the monitoring point correlation matrix (CSV format) generated in step S25 are read. A new graph structure is created, and all node and edge attributes of the initial graph structure are copied. For each edge in the graph, the IDs of the two connected nodes are obtained (for example, Node_A and Node_B). The Pearson correlation coefficient corresponding to these two nodes is found in the monitoring point correlation matrix (for example, `correlation(Node_A,Node_B)`). The absolute value of this correlation coefficient is used as the initial dynamic weight of the edge. If there is no corresponding correlation coefficient for the two nodes in the monitoring point correlation matrix (for example, due to data missing), the initial dynamic weight of the edge is set to the weight of the edge in the initial graph structure. The updated edge weight information is saved to the new graph structure, and the graph structure is saved in GraphML format and named the preliminary dynamic weight graph.

[0152] The preprocessed data stream (CSV format) generated in step S25 and the preliminary dynamic weight graph (GraphML format) generated in step S261 are read. The sliding window size is set to 7 days, that is, the data of the most recent 168 hours. For the current time *t*, the data of 168 hours before *t* in the preprocessed data stream is extracted. For each edge in the preliminary dynamic weight graph, the IDs of the two connected nodes are obtained. The Pearson correlation coefficient of the data sequences of these two nodes within the sliding window is calculated. The calculated correlation coefficients are stored in an N×N matrix, where N is the number of monitoring points, to form a real-time correlation matrix. If the data of the two nodes within the sliding window is insufficient to calculate the correlation coefficient, the weight of the corresponding edge in the preliminary dynamic weight graph is used as the substitute value of the correlation coefficient. The real-time correlation matrix is saved in CSV format.

[0153] Read the preliminary dynamic weight graph (in GraphML format) generated in step S261 and the real-time correlation matrix (in CSV format) generated in step S262. Set the weight update coefficient α = 0.7. Create a new graph structure and copy all node and edge attributes of the preliminary dynamic weight graph. For each edge in the graph, obtain the IDs of the two connected nodes (e.g., Node_A and Node_B). Obtain the real-time correlation coefficient of these two nodes from the real-time correlation matrix (e.g., `realtime_correlation(Node_A,Node_B)`). Update the weight of the edge using the exponential weighted average method: `W_new(Node_A,Node_B) = α * |realtime_correlation(Node_A,Node_B)| + (1 - α) * W_old(Node_A,Node_B)`, where `W_old(Node_A,Node_B)` is the weight of this edge in the preliminary dynamic weight graph. Limit the updated weight within the range of [0, 1]. Save the updated edge weight information into the new graph structure and save this graph structure in GraphML format, named the intermediate dynamic weight graph.

[0154] Read the intermediate dynamic weight graph (in GraphML format) generated in step S263 and the pre-established physical correlation knowledge base. The physical correlation knowledge base is a CSV file containing three columns: the starting node ID, the ending node ID, and the correlation strength. The correlation strength is a numerical value representing the strength of the physical association between two nodes. For example, 1 represents a strong association, 0.5 represents a medium association, and 0 represents no association. Create a new graph structure and copy all node and edge attributes of the intermediate dynamic weight graph. For each edge in the graph, obtain the IDs of the two connected nodes. Look up the correlation strength of these two nodes in the physical correlation knowledge base. If the correlation strength is greater than 0.8 and the weight of the corresponding edge in the intermediate dynamic weight graph is less than 0.5, update the weight of the edge to 0.5. This ensures that even if the short-term data correlation is low, the edge weights between physically strongly associated nodes still remain at a reasonable level. Save the updated edge weight information into the new graph structure and save this graph structure in GraphML format, named the dynamic weighted graph.

[0155] Preferably, step S3 includes the following steps:

[0156] Step S31: Construct a training dataset according to the dynamic topology graph to obtain a model training dataset;

[0157] Step S32: Obtain the water conservancy project prediction demand data; construct the ST-GCN model architecture according to the water conservancy project prediction demand data to obtain the model structure configuration;

[0158] Step S33: Obtain the prediction task type data; select the loss function and optimizer according to the model structure configuration and the prediction task type data to obtain the training parameter configuration;

[0159] Step S34: Perform model training and validation according to the model structure configuration, the model training dataset, and the training parameter configuration to obtain the initial ST-GCN model;

[0160] Step S35: Evaluate the performance of the initial ST-GCN model to obtain a model evaluation report;

[0161] Step S36: Optimize the model parameters of the initial ST-GCN model using the model evaluation report to obtain the ST-GCN model.

[0162] In the embodiment of the present invention, the dynamic topology map data generated in step S27 is read. This data includes the graph structure (nodes, edges, weights) and node features (sensor monitoring data) at each time step. Set the input time step length to T = 24 (i.e., data for the past 24 hours), and the prediction time step length to P = 1 (i.e., predict the state for the next 1 hour). Extract continuous time series data from the dynamic topology map to construct training samples. Each training sample contains an input sequence and an output sequence. The input sequence is a sequence composed of the graph data of the past T time steps, and the output sequence is the node features of the next P time steps. Divide the dataset into a training set, a validation set, and a test set according to the ratio of 7:2:1. Save the training set, the validation set, and the test set as three independent binary files to form the model training dataset.

[0163] Obtain the prediction demand data for the water conservancy project. This data clarifies the target variables to be predicted, such as dam displacement, seepage pressure, etc., and the time granularity of the prediction, such as hourly, daily. According to the prediction demand data, construct the ST-GCN model architecture. The model contains two spatio-temporal convolutional blocks, and each block consists of a spatial graph convolutional layer and a temporal convolutional layer. The spatial graph convolutional layer uses Chebyshev graph convolution with an order of 2. The temporal convolutional layer uses one-dimensional convolution with a kernel size of 3. The model also contains an output layer for outputting the prediction result. Record the specific parameters of the model architecture, including the kernel size, feature dimension, number of layers, etc., in the model structure configuration file and store it in JSON format.

[0164] Obtain the data of the prediction task type to determine whether the prediction task is a regression task or a classification task. Since the data predicted in the present invention is continuous numerical data (such as displacement, pressure), the prediction task type is a regression task. According to the model structure configuration and the regression task type, select the mean squared error (MSE) as the loss function and the Adam optimizer as the optimizer for model training. Set the learning rate of the Adam optimizer to 0.001 and the weight decay coefficient to 0.0001. Record the selected loss function, optimizer, and corresponding hyperparameters into the training parameter configuration file and store it in JSON format.

[0165] Use the deep learning framework TensorFlow to construct the ST-GCN model according to the model structure configuration generated in step S32. Load the model training dataset generated in step S31, and set the loss function and optimizer according to the training parameter configuration generated in step S33. Use the training set data to train the model, and evaluate the model performance using the validation set data after each epoch. During the training process, monitor the changes in the training loss and validation loss, and use the early stopping strategy to prevent overfitting. For example, if the validation loss does not decrease for 5 consecutive epochs, stop the training. Save the model with the best performance during the training process as the initial ST-GCN model.

[0166] Load the initial ST-GCN model obtained in step S34 and the test set data generated in step S31. Use the test set data to evaluate the model, and calculate the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R^2) of the model on the test set. Record the evaluation results into the model evaluation report and store it in text file format.

[0167] Analyze the model evaluation report generated in step S35. If the model performance does not meet the preset requirements, for example, the RMSE is greater than the preset threshold, the model parameters need to be optimized. The hyperparameters of the model can be adjusted, such as the learning rate, weight decay coefficient, convolution kernel size, etc., or the model structure can be modified, such as increasing or decreasing the number of network layers. Repeat steps S34 and S35 until the model performance meets the requirements. Save the finally obtained model as the ST-GCN model file.

[0168] Preferably, step S32 includes the following steps:

[0169] Step S321: Generate the configuration of the spatial convolution layer based on the Chebyshev spectral graph convolution according to the prediction demand data of the water conservancy project to obtain the spatial convolution layer configuration;

[0170] Step S322: Construct the temporal convolution layer according to the spatial convolution layer configuration to obtain the temporal convolution layer configuration;

[0171] Step S323: Integrate the temporal attention mechanism according to the spatial convolutional layer configuration and the temporal convolutional layer configuration to obtain the attention mechanism configuration;

[0172] Step S324: Design the model stacking structure based on the spatial convolutional layer configuration, the temporal convolutional layer configuration, and the attention mechanism configuration to obtain the model stacking configuration;

[0173] Step S325: Select an activation function for the model stacking configuration to obtain the activation function configuration;

[0174] Step S326: Design the model output layer according to the model stacking configuration and the activation function configuration to obtain the model structure configuration.

[0175] In the embodiment of the present invention, the prediction demand data of the water conservancy project is read, and the number and type of the prediction target variables are determined. For example, the prediction target variables are the dam displacement and the seepage pressure, with a total of 2 variables. Based on this, the spatial convolutional layer is configured. The Chebyshev spectral graph convolution is used as the convolution operator. The order of the Chebyshev polynomial is set to 2 to capture the local graph structure information. The output feature dimension of the spatial convolutional layer is set to 64. The convolution kernel parameters are initialized using the Glorot uniform distribution initialization method. These configuration parameters, including the convolution operator type, order, output feature dimension, and initialization method, are saved to the spatial convolutional layer configuration file (JSON format).

[0176] Read the spatial convolutional layer configuration generated in step S321. According to this configuration, a temporal convolutional layer is constructed. The one-dimensional convolution is used as the temporal convolution operator. The convolution kernel size is set to 3, the stride is set to 1, and the padding method is "same" to keep the time series length unchanged. The output feature dimension of the temporal convolutional layer is set to be the same as that of the spatial convolutional layer, that is, 64. The convolution kernel parameters are initialized using the Glorot uniform distribution initialization method. These configuration parameters, including the convolution operator type, convolution kernel size, stride, padding method, output feature dimension, and initialization method, are saved to the temporal convolutional layer configuration file (JSON format).

[0177] Read the spatial convolutional layer configuration generated in step S321 and the temporal convolutional layer configuration generated in step S322. Based on these configurations, integrate the temporal attention mechanism. Use the multi-head self-attention mechanism and set the number of attention heads to 8. Set the dimensions of the query, key, and value for each attention head to be the output feature dimension of the spatial convolutional layer and the temporal convolutional layer divided by the number of attention heads, i.e., 64 / 8 = 8. Use the scaled dot-product attention calculation formula: `Attention(Q,K,V) = softmax(QK^T / sqrt(d_k))V`, where `d_k` is the dimension of the key. Save these configuration parameters, including the attention mechanism type, the number of attention heads, the query / key / value dimensions, and the calculation formula, to the attention mechanism configuration file (in JSON format).

[0178] Read the configurations generated in steps S321, S322, and S323. Design the model stacking structure. The model contains two spatio-temporal convolutional blocks. Each spatio-temporal convolutional block consists of a spatial convolutional layer, a temporal convolutional layer, and a temporal attention mechanism, connected in this order. Add a batch normalization layer and a ReLU activation function after each spatio-temporal convolutional block. Stack these two spatio-temporal convolutional blocks together. Save the stacking structure information, including the number of blocks, the layer types and connection order within each block, and the connection method between blocks, to the model stacking configuration file (in JSON format).

[0179] Read the model stacking configuration generated in step S324. Select the activation function for each spatio-temporal convolutional block. After the batch normalization layer of each spatio-temporal convolutional block, use the ReLU activation function to increase the non-linear expression ability of the model. Save the selection and position information of the activation function to the activation function configuration file (in JSON format).

[0180] Read the model stacking configuration generated in step S324 and the activation function configuration generated in step S325. Design the model output layer. Add a fully connected layer at the end of the model, and its output dimension is equal to the number of predicted target variables (e.g., dam displacement and seepage pressure, the output dimension is 2). Since the prediction target is a continuous value, the output layer does not use an activation function. Add the type, output dimension, and activation function information of the output layer to the model stacking configuration and the activation function configuration to form a complete model structure configuration, and save it as a JSON format file.

[0181] Preferably, step S4 includes the following steps:

[0182] Step S41: Perform model input preprocessing on the original data stream and the dynamic topology map to obtain the model input map;

[0183] Step S42: Input the model input map into the ST-GCN model for future state prediction to obtain the predicted state sequence;

[0184] Step S43: Extract monitoring point information from the dynamic topology graph to obtain the monitoring point information; use the preset safety threshold data and the monitoring point information to initially judge the risk of the predicted status sequence to obtain the individual risk assessment result;

[0185] Step S44: Perform multi-parameter comprehensive risk assessment based on the Bayesian network according to the individual risk assessment result to obtain the comprehensive risk assessment result;

[0186] Step S45: Generate a risk assessment report for the predicted status sequence, monitoring point information, individual risk assessment result, and comprehensive risk assessment result to obtain the risk assessment report.

[0187] Step S46: Determine the risk level of the risk assessment report to obtain the determined risk level; judge the early warning trigger condition according to the determined risk level and the preset early warning strategy configuration to obtain the early warning trigger signal;

[0188] Step S47: Generate early warning information according to the early warning trigger signal, risk assessment report, and monitoring point information to obtain the to-be-sent early warning information; publish the to-be-sent early warning information to obtain the sent early warning record.

[0189] In the embodiment of the present invention, the latest original data stream (CSV format) and dynamic topology graph data are read. The original data stream is preprocessed: outliers are removed using the 3σ criterion, and missing values are filled using the linear interpolation method. Then, the Min-Max normalization method is used to scale the data to the [0,1] interval. The preprocessed data is updated to the dynamic topology graph as node features. According to the input time step length T = 24 defined in step S3, the data of 24 time steps before the current moment t is extracted to construct the model input graph sequence. This sequence is converted into a tensor format and used as the input of the ST-GCN model. The tensor dimension is (1, T, N, F), where 1 represents a single sample, T is the time step length, N is the number of nodes, and F is the node feature dimension.

[0190] Load the ST-GCN model trained in step S36. Input the model input graph sequence prepared in step S41 into the ST-GCN model for inference and prediction. The model outputs the prediction results for the next P = 1 time step, that is, the predicted values for the next 1 hour. The prediction result is a tensor with a dimension of (1, P, N, F), where P is the prediction time step length, N is the number of nodes, and F is the node feature dimension. Convert the tensor into a time series format, including the predicted values of each monitoring point within the next 1 hour, and store it as a CSV file named the predicted status sequence.

[0191] Extract the monitoring point information from the dynamic topology map, including the monitoring point ID, location, monitoring parameter type, etc., and save it as a CSV file named Monitoring Point Information. Read the preset safety threshold data (CSV format), which contains the warning threshold and alarm threshold for each monitoring parameter. Traverse each monitoring point in the prediction status sequence, and compare the predicted value with the corresponding warning threshold and alarm threshold. If the predicted value exceeds the warning threshold, mark the risk level of this monitoring point as "Warning"; if the predicted value exceeds the alarm threshold, mark it as "Alarm"; otherwise, mark it as "Normal". Store the risk level and the magnitude of the value exceeding the threshold for each monitoring point in a CSV file named Individual Risk Assessment Results.

[0192] Load the pre-constructed Bayesian network model. This model defines the dependencies between different monitoring parameters and their impact on the overall risk level. Use the individual risk assessment results obtained in step S43 as evidence and input it into the Bayesian network. Perform probabilistic inference using the belief propagation algorithm to calculate the probability distribution of the overall risk level. Save the probability distribution of the risk level and the most likely risk level to a JSON file named Comprehensive Risk Assessment Results.

[0193] Read the prediction status sequence generated in step S42, the monitoring point information and individual risk assessment results generated in step S43, and the comprehensive risk assessment results generated in step S44. Use the pre-designed report template to generate a risk assessment report (PDF format). The report content includes: a graphical display of the prediction status sequence, the risk level and the magnitude of the value exceeding the threshold for each monitoring point, the probability distribution of the overall risk level and the most likely risk level, as well as specific risk descriptions and recommended measures for high-risk monitoring points.

[0194] Read the risk assessment report generated in step S45, extract the most likely risk level in the comprehensive risk assessment results as the determined risk level. Read the preset warning strategy configuration (JSON format). This configuration defines the warning trigger conditions corresponding to different risk levels. For example, when the determined risk level is "Alarm", trigger the warning immediately; when the determined risk level is "Warning" and the duration exceeds 1 hour, trigger the warning. According to the determined risk level and the warning strategy configuration, determine whether the warning trigger conditions are met. If so, generate a warning trigger signal, including the warning level (such as level 1 warning, level 2 warning) and the trigger reason.

[0195] If a warning trigger signal is generated in step S46, warning information is generated based on the warning trigger signal, the risk assessment report, and the monitoring point information. The warning information includes the warning time, warning level, trigger reason, locations and monitoring parameters of the involved monitoring points, risk description, and recommended handling measures. The warning information is sent to a predefined list of recipients, for example, pushed via text message, email, or a dedicated APP. After the sending is completed, the warning sending time, recipients, and sending status (success or failure) are recorded and saved as a sent warning record (in CSV format).

[0196] Preferably, step S44 includes the following steps:

[0197] Step S441: Obtain the safety evaluation index system of the water conservancy project and the historical risk event analysis report; define the basic monitoring parameter nodes for the individual risk assessment results to obtain the basic monitoring parameter nodes;

[0198] Step S442: Define the intermediate risk index nodes according to the basic monitoring parameter nodes and the safety evaluation index system of the water conservancy project to obtain the intermediate risk index nodes;

[0199] Step S443: Define the final risk level nodes according to the intermediate risk index nodes, and construct a Bayesian network node definition table according to the basic monitoring parameter nodes to obtain the Bayesian network node definition table;

[0200] Step S444: Use the preset professional knowledge base of the water conservancy project to establish the directed edges of the nodes for the Bayesian network node definition table to obtain the directed edges of the nodes; perform causal relationship modeling on the Bayesian network node definition table to obtain the causal relationship data;

[0201] Step S445: Use the historical risk event analysis report to determine the network structure of the Bayesian network node definition table to obtain the network structure data;

[0202] Step S446: Construct a Bayesian network structure according to the directed edges of the nodes, the causal relationship data, and the network structure data to obtain a Bayesian network structure diagram;

[0203] Step S447: Determine the conditional probability table for the Bayesian network structure diagram to obtain the Bayesian network conditional probability table;

[0204] Step S448: Input the individual risk assessment results as evidence into the Bayesian network structure diagram, and use the Bayesian network conditional probability table to perform risk inference and assessment to obtain the comprehensive risk assessment result.

[0205] In an embodiment of the present invention, a pre-defined water conservancy project safety evaluation index system document and a historical risk event analysis report are obtained. The safety evaluation index system document lists all relevant monitoring parameters, such as dam displacement, seepage pressure, reservoir water level, etc. Read the individual risk assessment results (in CSV format) generated in step S43. Based on the individual risk assessment results and the safety evaluation index system, a basic monitoring parameter node is defined for each monitoring parameter. For example, if the individual risk assessment results include the risk levels of dam crest displacement, dam body seepage pressure, and reservoir water level, then three basic monitoring parameter nodes, namely "dam crest displacement risk", "dam body seepage pressure risk", and "reservoir water level risk", are defined respectively. Record the names of all basic monitoring parameter nodes and their corresponding monitoring parameter types into a CSV file named basic monitoring parameter nodes.

[0206] Read the basic monitoring parameter nodes (in CSV format) generated in step S441 and the water conservancy project safety evaluation index system document. Based on the safety evaluation index system, combine the relevant basic monitoring parameter nodes into intermediate risk index nodes. For example, combine the two basic monitoring parameter nodes of "dam crest displacement risk" and "dam foundation settlement risk" into an intermediate risk index node of "structural deformation risk"; combine the "dam body seepage pressure risk" and "drainage volume risk" into an intermediate risk index node of "seepage safety risk". Record the names of all intermediate risk index nodes, their corresponding basic monitoring parameter nodes, and their logical relationships (such as "and", "or") into a CSV file named intermediate risk index nodes.

[0207] Read the intermediate risk index nodes (in CSV format) generated in step S442. Define the final risk level node, such as "overall risk level". The value of this node can be pre-defined risk levels, such as "low", "medium", "high" or "normal", "attention", "warning", "alarm". Record the name and possible values of the final risk level node into a CSV file. Integrate the information of the basic monitoring parameter nodes generated in step S441, the intermediate risk index nodes generated in step S442, and the final risk level node into a CSV file to form a Bayesian network node definition table. This table contains the names, types (basic monitoring parameters, intermediate risk indexes, or final risk levels), and possible values of all nodes.

[0208] Read the Bayesian network node definition table generated in step S443 and the preset professional knowledge base for water conservancy projects. The professional knowledge base is a database containing expert knowledge. For example, excessive dam displacement will lead to an increase in structural risk, and excessive seepage pressure will lead to an increase in seepage risk, etc. According to the professional knowledge base, establish directed edges between each node in the Bayesian network node definition table. For example, establish a directed edge from the "dam crest displacement risk" node to the "structural deformation risk" node, indicating that the dam crest displacement risk will affect the structural deformation risk. Save all the established directed edge information, including the start node and the end node, to a CSV file named node directed edges. At the same time, based on the causal relationships in the field of water conservancy projects, perform causal relationship modeling on the Bayesian network node definition table. For example, an increase in rainfall will lead to an increase in the reservoir water level, and an increase in the reservoir water level will lead to an increase in seepage pressure. Express these causal relationships using mathematical formulas or logical expressions and save them to a text file named causal relationship data.

[0209] Read the Bayesian network node definition table generated in step S443 and the historical risk event analysis report. The historical risk event analysis report records the previous risk events that have occurred and the changes in each monitoring parameter before and after the risk event. According to the historical risk event analysis report, analyze the correlation between each monitoring parameter and the degree of influence on the risk event. For example, if the historical data shows that the dam crest displacement and the dam body seepage pressure both show abnormalities during multiple risk events, it is considered that there is a strong correlation between these two parameters. Use this correlation information to determine the structure of the Bayesian network. For example, add an edge between the "dam crest displacement risk" node and the "dam body seepage pressure risk" node to represent their correlation. Save the finally determined network structure information, including the connection relationship between nodes, to a text file named network structure data.

[0210] Read the node directed edges (CSV format) generated in step S444, the causal relationship data (text format), and the network structure data (text format) generated in step S445. Use a Bayesian network modeling tool (such as GeNIe or BayesFusion) to construct a Bayesian network structure based on these data. Import the node directed edges into the modeling tool to create the connection relationship between nodes. According to the causal relationship data, adjust the network structure to ensure the correct expression of the causal relationship. According to the network structure data, further improve the network structure, such as adding or deleting the connection between nodes. Save the finally constructed Bayesian network structure as a graphic file (such as.xdsl format) named Bayesian network structure diagram.

[0211] Read the Bayesian network structure diagram generated in step S446. For each node in the network, determine its conditional probability table (CPT). The CPT defines the probability distribution of the node when its parent nodes take different values. For the basic monitoring parameter nodes, the frequency of each risk level can be statistically calculated using historical monitoring data as its prior probability distribution. For the intermediate risk index nodes and the final risk level nodes, their CPTs can be determined by combining expert knowledge and historical risk event data. Save the CPTs of all nodes to a CSV file named Bayesian network conditional probability table. The probability values in the CPT should be represented numerically, such as 0.1, 0.5, 0.9, etc.

[0212] Read the Bayesian network structure diagram generated in step S446, the Bayesian network conditional probability table (in CSV format) generated in step S447, and the individual risk assessment results (in CSV format) generated in step S43. Input the individual risk assessment results as evidence into the Bayesian network. For example, if the individual risk assessment result of the "crest displacement risk" node is "early warning", then set the status of this node to "early warning". Use the Bayesian network inference algorithm (such as the joint tree algorithm or variable elimination algorithm) to calculate the posterior probability distribution of the final risk level node. Save the posterior probability distribution and the most likely risk level to a JSON file named comprehensive risk assessment results.

[0213] Preferably, the present invention also provides a water conservancy project safety monitoring system based on data processing for implementing the water conservancy project safety monitoring method based on data processing as described above. The water conservancy project safety monitoring system based on data processing includes:

[0214] A heterogeneous perception module for constructing a multi-source data acquisition scheme for the water conservancy project to obtain a multi-source data acquisition scheme; performing multi-source data acquisition on the water conservancy project according to the multi-source data acquisition scheme to obtain an original data stream;

[0215] A topology encoding module for grading monitoring nodes according to the multi-source data acquisition scheme to obtain an initial node set; constructing initial edges and defining their types according to the initial node set to obtain an initial edge list; assigning weights to different types of initial edges in the initial edge list and generating an initial graph structure to obtain an initial graph structure; performing correlation analysis on the original data stream to obtain a monitoring point correlation matrix; using the monitoring point correlation matrix to perform dynamic edge weight adjustment processing on the initial graph structure to obtain a dynamic topology map;

[0216] A dynamic map learning module for obtaining water conservancy project prediction demand data; constructing an ST-GCN model architecture according to the water conservancy project prediction demand data to obtain a model structure configuration; constructing an ST-GCN model according to the model structure configuration and the dynamic topology map, and performing the ST-GCN model;

[0217] An intelligent early warning module, which is used to predict the future state by using the ST-GCN model to obtain a predicted state sequence; conduct a preliminary risk judgment on the predicted state sequence to obtain an individual risk assessment result; conduct a multi-parameter comprehensive risk assessment based on a Bayesian network according to the individual risk assessment result to obtain a comprehensive risk assessment result; generate a risk assessment report according to the comprehensive risk assessment result to obtain a risk assessment report; judge the early warning trigger conditions for the risk assessment report and issue early warning information to implement the safety monitoring task of the water conservancy project.

[0218] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, it is intended to cover all changes falling within the meaning and scope of the equivalent elements of the application document within the present invention.

[0219] The above are only specific embodiments of the present invention, which enable those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A water conservancy project safety monitoring method based on data processing, characterized in that: The following steps are involved: Step S1: construct a multi-source data collection scheme for a water conservancy project to obtain a multi-source data collection scheme; According to the multi-source data collection scheme, multi-source data collection is carried out on the water conservancy project to obtain the original data stream; Step S2: According to the multi-source data collection scheme, the monitoring nodes are classified to obtain an initial node set; according to the initial node set, initial edges are constructed and type defined to obtain an initial edge list; Assign different types of initial edge weights to the initial edge list, and generate an initial graph structure to obtain an initial graph structure; Perform correlation analysis on the original data stream to obtain the monitoring point correlation matrix; use the monitoring point correlation matrix to dynamically adjust the edge weights of the initial graph structure to obtain a dynamic topology map; Step S3: Acquire the predicted demand data of the water conservancy project; construct the ST-GCN model architecture according to the predicted demand data of the water conservancy project to obtain the model structure configuration; construct the ST-GCN model according to the model structure configuration and the dynamic topology map, and perform the ST-GCN model; Step S4: Use the ST-GCN model to predict the future state and obtain a predicted state sequence; perform a preliminary risk assessment on the predicted state sequence and obtain an individual risk assessment result; According to the individual risk assessment results, a multi-parameter comprehensive risk assessment based on the Bayesian network is performed to obtain a comprehensive risk assessment result; Generate a risk assessment report based on the comprehensive risk assessment results to obtain a risk assessment report; The early warning trigger conditions are judged on the risk assessment report, and the early warning information is released to achieve the safety monitoring task of water conservancy projects.

2. The water conservancy project safety monitoring method based on data processing according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: deploy sensors on the water conservancy project and construct a data transmission network to obtain a multi-source data collection solution; Step S12: performing multi-source data collection on the water conservancy project according to the multi-source data collection scheme to obtain a real-time multi-source data stream; Step S13: performing data format conversion and verification on the real-time multi-source data stream to obtain formatted verification data; Step S14: Perform multi-source data integration and time synchronization on the formatted verification data to obtain the original data stream.

3. The water conservancy project safety monitoring method based on data processing according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21: obtaining a water conservancy project structural design drawing; performing monitoring node mapping on the multi-source data acquisition scheme and the water conservancy project structural design drawing to obtain a node mapping set; Step S22: performing node importance grading on the node mapping set to obtain an initial node set; Step S23: constructing initial edges and defining their types according to the water conservancy project structure design diagram and the initial node set to obtain an initial edge list; assigning weights of different types of initial edges to the initial edge list to obtain a weighted initial edge list; Step S24: generating an initial graph structure according to the weighted initial edge list and the initial edge list to obtain an initial graph structure; Step S25: performing a correlation analysis on the original data stream based on historical data to obtain a monitoring point correlation matrix; performing original data preprocessing on the original data stream to obtain a preprocessed data stream; Step S26: dynamically adjust the edge weights of the initial graph structure using the monitoring point association matrix to obtain a dynamic weighted graph; Step S27: Perform data fusion on the dynamic weighted graph and the preprocessed data stream to obtain a dynamic topology map.

4. The water conservancy project safety monitoring method based on data processing according to claim 3 is characterized in that: Step S23 includes the following steps: Step S231: marking the physical connection edges of the initial node set to obtain physical connection edges; marking the spatial adjacent edges of the initial node set to obtain spatial adjacent edges; Step S232: marking the water flow direction associated edges of the initial node set to obtain water flow associated edges; marking the force transmission path associated edges of the initial node set to obtain force associated edges; Step S233: constructing an initial edge list according to the physical connection edges, spatial adjacent edges, water flow associated edges and force associated edges to obtain an initial edge list; Step S234: assigning weights to the physical connection edges based on the material strength parameters to obtain physical connection edge weights; assigning weights to the spatial neighboring edges based on the inverse of the Euclidean distance between nodes to obtain spatial neighboring edge weights; Step S235: assigning weights to the water flow associated edges based on the fixed weight values ​​of the directed edges to obtain the weights of the water flow direction associated edges; assigning weights to the force associated edges based on the stress transfer coefficient to obtain the weights of the force transfer path associated edges; Step S236: construct a weighted initial edge list according to the physical connection edge weights, the spatial adjacent edge weights, the water flow direction associated edge weights, and the force transfer path associated edge weights to obtain a weighted initial edge list.

5. The water conservancy project safety monitoring method based on data processing according to claim 3 is characterized in that: Step S26 includes the following steps: Step S261: using the monitoring point association matrix to initialize the weight of the initial graph structure based on statistical association, to obtain a preliminary dynamic weight graph; Step S262: performing a real-time relevance update based on a sliding window on the preprocessed data stream and the preliminary dynamic weight map to obtain a real-time relevance matrix; Step S263: performing dynamic weight fusion and adjustment on the preliminary dynamic weight map according to the real-time correlation matrix to obtain an intermediate dynamic weight map; Step S264: Perform correlation correction on the intermediate dynamic weight graph based on the physical model to obtain a dynamic weighted graph.

6. The water conservancy project safety monitoring method based on data processing according to claim 1 is characterized in that: Step S3 includes the following steps: Step S31: construct a training data set according to the dynamic topology map to obtain a model training data set; Step S32: Acquire water conservancy project forecast demand data; construct the ST-GCN model architecture according to the water conservancy project forecast demand data to obtain the model structure configuration; Step S33: Obtain prediction task type data; select a loss function and an optimizer according to the model structure configuration and the prediction task type data to obtain a training parameter configuration; Step S34: Perform model training and verification according to the model structure configuration, model training data set and training parameter configuration to obtain an initial ST-GCN model; Step S35: Evaluate the model performance of the initial ST-GCN model to obtain a model evaluation report; Step S36: Optimize the model parameters of the initial ST-GCN model using the model evaluation report to obtain the ST-GCN model.

7. The water conservancy project safety monitoring method based on data processing according to claim 6 is characterized in that: Step S32 includes the following steps: Step S321: generating a spatial convolution layer configuration based on Chebyshev spectral convolution according to the water conservancy project forecast demand data to obtain a spatial convolution layer configuration; Step S322: constructing a temporal convolution layer according to the spatial convolution layer configuration to obtain a temporal convolution layer configuration; Step S323: Perform temporal attention mechanism integration according to the spatial convolution layer configuration and the temporal convolution layer configuration to obtain an attention mechanism configuration; Step S324: Designing a model stacking structure based on the spatial convolution layer configuration, the temporal convolution layer configuration, and the attention mechanism configuration to obtain a model stacking configuration; Step S325: selecting an activation function for the model stacking configuration to obtain an activation function configuration; Step S326: Design the model output layer according to the model stacking configuration and the activation function configuration to obtain the model structure configuration.

8. The water conservancy project safety monitoring method based on data processing according to claim 1 is characterized in that: Step S4 includes the following steps: Step S41: performing model input preprocessing on the original data stream and the dynamic topology map to obtain a model input map; Step S42: input the model input graph into the ST-GCN model to predict the future state and obtain a predicted state sequence; Step S43: extract monitoring point information from the dynamic topology map to obtain monitoring point information; use the preset safety threshold data and monitoring point information to make a preliminary risk assessment on the predicted state sequence to obtain an individual risk assessment result; Step S44: performing a multi-parameter comprehensive risk assessment based on a Bayesian network according to the individual risk assessment results to obtain a comprehensive risk assessment result; Step S45: Generate a risk assessment report based on the predicted state sequence, monitoring point information, individual risk assessment results, and comprehensive risk assessment results to obtain a risk assessment report. Step S46: determining the risk level of the risk assessment report to obtain the determined risk level; determining the warning trigger condition according to the determined risk level and the preset warning strategy configuration to obtain a warning trigger signal; Step S47: Generate warning information according to the warning trigger signal, risk assessment report and monitoring point information to obtain the warning information to be sent; publish the warning information to be sent to obtain the sent warning record.

9. The water conservancy project safety monitoring method based on data processing according to claim 8 is characterized in that: Step S44 includes the following steps: Step S441: Obtain the water conservancy project safety evaluation index system and the historical risk event analysis report; define the basic monitoring parameter nodes for the individual risk assessment results to obtain the basic monitoring parameter nodes; Step S442: defining intermediate risk indicator nodes according to the basic monitoring parameter nodes and the water conservancy project safety evaluation index system to obtain intermediate risk indicator nodes; Step S443: define the final risk level node according to the intermediate risk indicator node, and construct the Bayesian network node definition table according to the basic monitoring parameter node to obtain the Bayesian network node definition table; Step S444: using a preset water conservancy engineering professional knowledge base to establish node directed edges for the Bayesian network node definition table to obtain node directed edges; performing causal relationship modeling on the Bayesian network node definition table to obtain causal relationship data; Step S445: using the historical risk event analysis report to determine the network structure of the Bayesian network node definition table to obtain network structure data; Step S446: constructing a Bayesian network structure according to the node directed edges, causal relationship data and network structure data to obtain a Bayesian network structure diagram; Step S447: determining the conditional probability table of the Bayesian network structure diagram to obtain the Bayesian network conditional probability table; Step S448: Input the individual risk assessment results as evidence into the Bayesian network structure diagram, and use the Bayesian network conditional probability table to perform risk reasoning and assessment to obtain a comprehensive risk assessment result.

10. A water conservancy project safety monitoring system based on data processing, characterized in that: Used to execute the water conservancy project safety monitoring method based on data processing as claimed in claim 1, the water conservancy project safety monitoring system based on data processing comprises: The heterogeneous perception module is used to construct a multi-source data collection scheme for water conservancy projects and obtain a multi-source data collection scheme; multi-source data collection is performed on water conservancy projects according to the multi-source data collection scheme to obtain the original data stream; The topological coding module is used to classify monitoring nodes according to the multi-source data collection scheme to obtain an initial node set; construct and define initial edges according to the initial node set to obtain an initial edge list; assign different types of initial edge weights to the initial edge list and generate an initial graph structure to obtain an initial graph structure; perform correlation analysis on the original data stream to obtain a monitoring point association matrix; use the monitoring point association matrix to dynamically adjust the edge weights of the initial graph structure to obtain a dynamic topological map; Dynamic graph learning module, used to obtain water conservancy project forecast demand data; construct ST-GCN model architecture according to water conservancy project forecast demand data, and obtain model structure configuration; construct ST-GCN model according to model structure configuration and dynamic topology graph, and conduct ST-GCN model; The intelligent early warning module is used to use the ST-GCN model to predict the future state and obtain the predicted state sequence; to make an initial risk assessment on the predicted state sequence and obtain the individual risk assessment result; to make a multi-parameter comprehensive risk assessment based on the Bayesian network according to the individual risk assessment result and obtain the comprehensive risk assessment result; to generate a risk assessment report according to the comprehensive risk assessment result and obtain the risk assessment report; to judge the early warning trigger conditions of the risk assessment report and to release the early warning information to realize the safety monitoring task of water conservancy projects.

Images