A water conservancy project operation safety monitoring system based on digital twinning

Abstract

The application discloses a kind of based on digital twinning water conservancy project operation safety monitoring system, it is related to digital twinning technical field, including: data regularizing module, generates standard state data;Twin registration module, constructs hydraulic twin body;Trajectory through module, generates response trajectory atlas;Phase encoding module, generates lag fingerprint matrix;Causal search module, executes MMHC algorithm, generates phase adjacency matrix;Trajectory reasoning module, constructs improved TrajGRU model, generates operation state prediction result and residual record;Twin evolution module, generates phase twin graph;Risk output module, outputs safety monitoring result.The application realizes the accurate identification of water conservancy project operation risk, improves lag response monitoring capacity and safety early warning reliability.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, and in particular to a water conservancy project operation safety monitoring system based on digital twins. Background Technology

[0002] With the development of digital monitoring, online inspection, and operation scheduling management of water conservancy projects, data on water level, flow rate, seepage pressure, seepage, displacement, settlement, cracks, vibration, gate opening, pump station status, environmental status, image inspection, and engineering configuration are continuously generated. To identify safety risks in the operation of water conservancy projects, existing technologies collect multi-source monitoring data, perform time alignment, missing data completion, and anomaly removal, and generate safety early warning results using threshold judgment, statistical analysis, digital twin display, or neural network models.

[0003] Existing water conservancy project operation safety monitoring systems mostly focus on displaying single-point monitoring values ​​and fixed threshold alarms. They are insufficient in expressing the registration relationship between project blocks, monitoring nodes, equipment nodes, and digital twin nodes. There is a lack of traceable trajectory connection structure between water condition changes, seepage response, structural deformation, and equipment disturbances, making it difficult to form a hydraulic twin and response trajectory map consistent with the actual project parts.

[0004] In actual operation, the seepage pressure response, structural deformation, and apparent damage caused by changes in water load exhibit significant hysteresis propagation characteristics. Existing methods often directly input standardized time-series data into prediction models, lacking support from hysteresis fingerprint matrices, phase adjacency matrices, and phase twin graphs. This results in insufficient ability to identify risk paths, risk levels, sensor anomaly markers, and residual backflow, easily leading to problems such as missed hysteresis risks, misjudgment of scheduling disturbances, and weak interpretability of monitoring results.

[0005] Therefore, how to provide a water conservancy project operation safety monitoring system based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a water conservancy project operation safety monitoring system based on digital twins. This invention fully utilizes hydraulic twins, the MMHC algorithm, and the improved TrajGRU model to construct response trajectory maps, lag fingerprint matrices, and phase twin maps, thereby realizing the identification of water conservancy project operation risks. It has the advantages of traceable risk paths, accurate lag response identification, and strong interpretability of monitoring results.

[0007] A water conservancy project operation safety monitoring system based on digital twin according to an embodiment of the present invention includes:

[0008] The data acquisition module collects hydraulic engineering monitoring data, performs preprocessing on the hydraulic engineering monitoring data, and generates standard state data.

[0009] The twin mapping module performs twin registration of engineering nodes on standard state data, and generates hydraulic twins according to the node registration relationship between engineering blocks, monitoring nodes, equipment nodes and digital twin nodes;

[0010] The trajectory construction module performs trajectory connection processing on the hydraulic twin, including the water flow transmission chain, structural connection chain, equipment action chain, and historical response path, to generate a response trajectory map.

[0011] The fingerprint construction module performs source-target node response pairing on standard state data, performs phase lag encoding according to response direction, response type, lag time and water load phase, and generates a lag fingerprint matrix.

[0012] The causal search module executes the MMHC algorithm based on the response trajectory map and the hysteresis fingerprint matrix, performing path memory search, skeleton reconstruction and hysteresis edge search to generate a phase adjacency matrix;

[0013] The risk identification module constructs an improved TrajGRU model. The phase adjacency matrix and response trajectory map are input into the improved TrajGRU model. After processing by the response trajectory generator, the pressure hysteresis gate cell and the residual backflow unit, the operating status prediction results and residual records are generated.

[0014] The twin write-back module constructs a phase twin graph based on the runtime prediction results, residual records, and phase adjacency matrix;

[0015] The monitoring output module writes the phase twin image into the hydraulic twin to generate safety monitoring results.

[0016] Optionally, the hydraulic monitoring data includes water condition monitoring data, seepage monitoring data, structural deformation data, equipment operation data, environmental monitoring data, image inspection data, and engineering configuration data.

[0017] Optionally, the preprocessing of the hydraulic monitoring data includes time alignment, spatial binding, missing data completion, and anomaly marking.

[0018] Optionally, the generation of the hydraulic twin includes:

[0019] Read standard status data, extract project block number, monitoring node number, equipment node number, digital twin node number, node type, node coordinates, water flow direction field, structural connection field and equipment association field from the project configuration data, extract the collection timestamp from the standard status data, and generate node registration records;

[0020] According to the project block number, the monitoring node number and the equipment node number are assigned to the corresponding project block, and the digital twin node number is bound to the project block number, the monitoring node number and the equipment node number to generate the node registration relationship;

[0021] The node registration relationship is arranged according to the collection timestamp. The node registration relationship, node type, node coordinates, water flow direction field, structural connection field, equipment association field, and node status field extracted from water monitoring data, seepage monitoring data, structural deformation data, equipment operation data, environmental monitoring data and image inspection data are written into the twin object record.

[0022] The twin object records are written into the object index according to the project block number, and the object index is associated with the digital twin node number to generate a hydraulic twin.

[0023] Optionally, the generation of the response trajectory map includes:

[0024] Read the engineering block number, monitoring node number, equipment node number, digital twin node number, node type and node coordinates from the hydraulic twin, and establish a graph node;

[0025] A water flow conduction chain is generated based on the water flow direction field in the hydraulic twin; a structural connection chain is generated based on the structural connection field; and a device action chain is generated based on the device association field.

[0026] Read the node status field of the same node under adjacent collection timestamps from the standard status data, calculate the change between adjacent node status fields, generate node status change records according to the order of occurrence of the changes, and generate historical response paths based on the node status change records.

[0027] The graph nodes, water flow conduction chains, structural connection chains, equipment action chains, and historical response paths are written into the trajectory record. The trajectory record includes the starting node, ending node, chain type, trajectory number, and collection timestamp.

[0028] Connect the trajectory records according to the starting node and the ending node, and bind the trajectory records with the digital twin node numbers to generate a response trajectory map.

[0029] Optionally, the generation of the hysteresis fingerprint matrix includes:

[0030] Read the node number and node type in the hydraulic twin, read the collection timestamp in the standard state data, extract the water condition status from the water condition monitoring data, extract the seepage status from the seepage monitoring data, extract the structural deformation status from the structural deformation data, extract the equipment operation status from the equipment operation data, extract the environmental status from the environmental monitoring data, extract the inspection status from the image inspection data, and generate a node status sequence according to the node number and collection timestamp.

[0031] The source node and target node are determined according to the historical response path in the response trajectory map. The source node state change time, target node state change time, source node state change direction and target node state change direction are extracted from the node state sequence.

[0032] The response direction is determined based on the order of the source node status change time and the target node status change time. The response type is determined based on the source node type, target node type, and inspection status. The lag time is determined based on the time interval between the two changes. The water load phase is determined based on the water level change direction and flow rate change direction in the water situation status, as well as the gate opening change status and pump station start / stop status in the equipment operation status.

[0033] Write the source node, target node, response direction, response type, lag time, water load phase, inspection status, source node status change direction, and target node status change direction into the matrix elements.

[0034] Arrange the matrix elements according to the source node number and the target node number, and bind the matrix elements to the collection timestamp to generate a hysteresis fingerprint matrix.

[0035] Optionally, generating the phase adjacency matrix includes:

[0036] Read the response trajectory map and the hysteresis fingerprint matrix, use the map nodes as the variable set of the MMHC algorithm, and use the historical response paths and matrix elements as candidate relationship records;

[0037] During the MMPC candidate parent-child node search phase, a path memory search is performed. The adjacent node relationships in the candidate relationship records are checked according to the historical response paths. The adjacent nodes that pass the MMPC conditional independence test are written into the candidate parent-child node set.

[0038] In the Hill-Climbing structural optimization stage, skeleton reconstruction is performed. Candidate parent-child node sets are read according to the water load phase. Within the node range defined by the candidate parent-child node sets, edge addition, edge deletion, and edge reverse search are performed. The phase causal skeleton is generated through the Hill-Climbing structural score.

[0039] During the edge search phase, a delayed edge search is performed. Based on the delayed fingerprint matrix, the delayed time, response type, inspection status, source node state change direction, and target node state change direction are written into the causal edges in the phase causal skeleton.

[0040] The causal edges are arranged with the source node number as the matrix row and the target node number as the matrix column, and the water load phase is used as the matrix layer label. The phase adjacency matrix is ​​generated according to the directed acyclic graph output process.

[0041] Optionally, the generation of runtime state prediction results and residual records includes:

[0042] An improved TrajGRU model is constructed, which includes the trajectory generation network, hidden state sampling process, reset gate calculation process, update gate calculation process, candidate hidden state calculation process, and output mapping process of the TrajGRU model, and is configured with a response trajectory generator, osmotic hysteresis gate cell, and residual backflow unit.

[0043] Read the phase adjacency matrix, response trajectory map and standard state data, arrange the node state fields according to the acquisition timestamp, and generate the current input features, initial hidden state and initial residual record;

[0044] The trajectory generation network generates a trajectory offset index based on the current input features and the hidden state of the previous acquisition timestamp. The response trajectory generator performs node matching between the trajectory offset index and the causal edges in the phase adjacency matrix and the trajectory records in the response trajectory graph to generate the engineering response trajectory input.

[0045] The hidden state sampling process reads the hidden state of the previous acquisition timestamp of the corresponding source node according to the engineering response trajectory input, and generates the trajectory sampling hidden state;

[0046] The reset gate calculation process and the update gate calculation process generate the reset gate state and the update gate state based on the current input features, the trajectory sampling hidden state, and the engineering response trajectory input.

[0047] The pressure hysteresis gated cell performs candidate hidden state calculation based on the reset gate state, update gate state, current input features and trajectory sampling hidden state, and generates the current hidden state according to the water load state formed by the water condition state, the pressure response state formed by the seepage state, the structural response state formed by the structural deformation state and the apparent damage state formed by the inspection state.

[0048] The residual backflow unit reads the residual records according to the acquisition timestamp. The first acquisition timestamp reads the initial residual record, and the subsequent acquisition timestamps read the residual record of the previous acquisition timestamp. The read residual records are written into the current hidden state, and the running state prediction result is generated through the output mapping process.

[0049] The difference between the predicted running status and the corresponding real-time status record in the standard status data is calculated, and the difference results are arranged according to node number, collection timestamp and residual type to generate residual records.

[0050] Optionally, constructing the phase twin map includes:

[0051] Read the operation status prediction results, residual records, phase adjacency matrix and hydraulic twin, and match the project blocks, monitoring nodes, equipment nodes and causal edges according to the digital twin node number;

[0052] Write the residual records into the corresponding causal edges in the phase adjacency matrix, and generate risk node records according to the residual source node and residual target node;

[0053] Risk paths are generated based on causal edges, lag time, response type, inspection status, and risk node records. Risk types, risk levels, and sensor anomaly markers are generated according to the number of consecutive causal edges, residual persistence status, and node association status in the risk paths.

[0054] Risk node records, risk paths, risk types, risk levels, and sensor anomaly markers are written into the graph structure to generate a phase twin graph.

[0055] Optionally, the generation of security monitoring results includes:

[0056] Read the phase twin diagram and the hydraulic twin, and write back the risk node records and risk paths in the phase twin diagram to the hydraulic twin according to the digital twin node numbers;

[0057] Extract the project block number, monitoring node number, equipment node number, risk path, risk type, risk level, and sensor anomaly marker from the written-back hydraulic twin to generate a safety monitoring record;

[0058] The safety monitoring records are arranged according to the collection timestamp, and the safety monitoring records are bound to the phase twin map to generate safety monitoring results.

[0059] The beneficial effects of this invention are:

[0060] This invention generates standard state data by collecting water condition monitoring data, seepage monitoring data, structural deformation data, equipment operation data, environmental monitoring data, image inspection data, and engineering configuration data. It then performs engineering node twin registration on the standard state data to form node registration relationships between engineering blocks, monitoring nodes, equipment nodes, and digital twin nodes. This allows the safety monitoring of water conservancy projects to move beyond single-point curve display and fixed threshold alarms, enabling the accurate binding of monitoring data to real engineering parts and improving the entity correspondence and state expression completeness of the hydraulic twin.

[0061] This invention constructs a response trajectory map and a hysteresis fingerprint matrix to uniformly organize the water flow conduction chain, structural connection chain, equipment action chain, historical response path, response direction, response type, hysteresis time, and water load phase. This creates a traceable data chain between water level changes, seepage response, structural deformation, equipment disturbance, and inspection anomalies, enabling the expression of the hysteresis conduction process caused by water load changes. It improves the identification ability of seepage pressure hysteresis, structural deformation delay, and apparent damage expansion, and reduces missed alarms and false judgments caused by single-index threshold alarms.

[0062] This invention generates a phase adjacency matrix using the MMHC algorithm, and generates operational status prediction results and residual records by improving the response trajectory generator, seepage hysteresis gate cell, and residual backflow unit in the TrajGRU model. Then, a phase twin map is constructed and written into the hydraulic twin, enabling risk node records, risk paths, risk types, risk levels, and sensor anomaly markers to be presented in a digital twin space, thereby improving the accuracy of risk identification in water conservancy projects, the interpretability of risk propagation paths, and the traceability of safety monitoring results. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They explain the invention together with the embodiments of the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is an overall flowchart of a water conservancy project operation safety monitoring system based on digital twin proposed in this invention;

[0065] Figure 2 This is a schematic diagram of the improved TrajGRU model structure of a water conservancy project operation safety monitoring system based on digital twin proposed in this invention;

[0066] Figure 3 The flowchart of the MMHC algorithm for a water conservancy project operation safety monitoring system based on digital twin proposed in this invention is shown below. Detailed Implementation

[0067] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0068] refer to Figure 1 , Figure 2 and Figure 3 A water conservancy project operation safety monitoring system based on digital twins, comprising:

[0069] The data acquisition module collects hydraulic engineering monitoring data, performs preprocessing on the hydraulic engineering monitoring data, and generates standard state data.

[0070] The twin mapping module performs twin registration of engineering nodes on standard state data, and generates hydraulic twins according to the node registration relationship between engineering blocks, monitoring nodes, equipment nodes and digital twin nodes;

[0071] The trajectory construction module performs trajectory connection processing on the hydraulic twin, including the water flow transmission chain, structural connection chain, equipment action chain, and historical response path, to generate a response trajectory map.

[0072] The fingerprint construction module performs source-target node response pairing on standard state data, performs phase lag encoding according to response direction, response type, lag time and water load phase, and generates a lag fingerprint matrix.

[0073] The causal search module executes the MMHC algorithm based on the response trajectory map and the hysteresis fingerprint matrix, performing path memory search, skeleton reconstruction and hysteresis edge search to generate a phase adjacency matrix;

[0074] The risk identification module constructs an improved TrajGRU model. The phase adjacency matrix and response trajectory map are input into the improved TrajGRU model. After processing by the response trajectory generator, the pressure hysteresis gate cell and the residual backflow unit, the operating status prediction results and residual records are generated.

[0075] The twin write-back module constructs a phase twin graph based on the runtime prediction results, residual records, and phase adjacency matrix;

[0076] The monitoring output module writes the phase twin image into the hydraulic twin to generate safety monitoring results.

[0077] In this embodiment, the hydraulic monitoring data includes water condition monitoring data, seepage monitoring data, structural deformation data, equipment operation data, environmental monitoring data, image inspection data, and engineering configuration data.

[0078] In this embodiment, the preprocessing of hydraulic monitoring data includes time alignment, spatial binding, missing data completion, and anomaly marking.

[0079] In this embodiment, generating the hydraulic twin includes:

[0080] Read standard status data, extract project block number, monitoring node number, equipment node number, digital twin node number, node type, node coordinates, water flow direction field, structural connection field, and equipment association field from the project configuration data, extract the acquisition timestamp from the standard status data, and generate node registration records. Specifically, generating node registration records involves:

[0081] Read the acquisition timestamp and data source marker from the standard status data; read the engineering block table, monitoring node table, equipment node table, and digital twin node table from the engineering configuration data; extract the engineering block number, block boundary coordinates, and block center coordinates from the engineering block table; extract the monitoring node number, monitoring node type, and monitoring node coordinates from the monitoring node table; extract the equipment node number, equipment node type, and equipment node coordinates from the equipment node table; extract the digital twin node number, digital twin node type, and digital twin node coordinates from the digital twin node table; extract the flow direction field from the hydraulic connection table in the engineering configuration data; extract the structural connection field from the structural connection table in the engineering configuration data; extract the equipment association field from the equipment connection table in the engineering configuration data; arrange the engineering block number, monitoring node number, equipment node number, digital twin node number, node type, node coordinates, flow direction field, structural connection field, equipment association field, and acquisition timestamp in field order to generate a node registration record;

[0082] According to the project block numbering, the monitoring node number and equipment node number are assigned to the corresponding project block. The digital twin node number is then bound to the project block number, monitoring node number, and equipment node number to generate a node registration relationship. Specifically, the generation of the node registration relationship is as follows:

[0083] Read the project block number, block boundary coordinates, monitoring node number, monitoring node coordinates, equipment node number, equipment node coordinates, digital twin node number, and digital twin node coordinates from the node registration record. Perform a point-to-point determination between the monitoring node coordinates and the block boundary coordinates within the polygon. If the determination result is within the block boundary, write the monitoring node number to the corresponding project block number. If the determination result is not within the block boundary, calculate the shortest distance from the monitoring node coordinates to each block boundary line segment. Write the monitoring node number with the shortest distance not exceeding 2 meters to the project block number with the smallest distance. Perform the same grouping process on the equipment node coordinates. Calculate the x-coordinate difference, y-coordinate difference, and elevation coordinate difference between the digital twin node coordinates and the monitoring node coordinates. Sum the squares of the three differences and take the square root to obtain the monitoring node registration distance. Bind the digital twin node number with a registration distance not exceeding 1 meter to the monitoring node number. Calculate the registration distance between the digital twin node coordinates and the equipment node coordinates. Bind the digital twin node number with a registration distance not exceeding 1 meter to the equipment node number to generate the node registration relationship.

[0084] The node registration relationships are arranged according to the acquisition timestamp. The node registration relationships, node types, node coordinates, water flow direction fields, structural connection fields, equipment association fields, and node status fields extracted from hydrological monitoring data, seepage monitoring data, structural deformation data, equipment operation data, environmental monitoring data, and image inspection data are written into the twin object record.

[0085] The node status fields extracted from water condition monitoring data, seepage monitoring data, structural deformation data, equipment operation data, environmental monitoring data, and image inspection data are as follows:

[0086] Read the node number, acquisition timestamp, and data source marker from the standard state data; extract water level, flow rate, and direction of water level change from hydrological monitoring data; extract seepage pressure, seepage flow rate, and direction of seepage pressure change from seepage monitoring data; extract displacement, settlement, crack width, vibration, and direction of structural deformation change from structural deformation data; extract gate opening, pump station start / stop status, equipment current, and direction of equipment movement change from equipment operation data; extract rainfall, temperature, wind speed, and direction of environmental change from environmental monitoring data; and extract data from images. The inspection data is used to extract the inspection defect category, crack pixel length, apparent anomaly confidence level, and inspection status. The extracted fields are arranged in the following order: water level value, flow rate value, water level change direction, seepage pressure value, seepage flow rate, seepage pressure change direction, displacement value, settlement value, crack width, vibration value, structural deformation change direction, gate opening, pump station start / stop status, equipment current value, equipment action change direction, rainfall, temperature value, wind speed value, environmental change direction, inspection defect category, crack pixel length, apparent anomaly confidence level, and inspection status to generate node status fields.

[0087] Write the twin object record, specifically as follows:

[0088] Read the node registration relationship, node type, node coordinates, water flow direction field, structural connection field, equipment association field, and node status field. Sort all node status fields corresponding to the same digital twin node number in ascending order according to the acquisition timestamp. Use the digital twin node number as the primary key field and the project block number, monitoring node number, equipment node number, node type, node coordinates, water flow direction field, structural connection field, equipment association field, acquisition timestamp, and node status field as attribute fields. Write the primary key field and attribute fields into the same row of records to generate a twin object record.

[0089] The twin object records are written into the object index according to the project block number, and the object index is associated with the digital twin node number to generate a hydraulic engineering twin, wherein:

[0090] Write the object index, specifically:

[0091] Read the twin object records, sort the twin object records in ascending order according to the project block number, sort the twin object records in ascending order according to the digital twin node number within the same project block number, sort the twin object records in ascending order according to the collection timestamp within the same digital twin node number, concatenate the project block number, digital twin node number and node type in sequence to generate the object index key, record the start row number and end row number of the twin object record corresponding to the object index key in the storage table, and write the object index key, start row number, end row number, project block number and digital twin node number into the object index;

[0092] The hydraulic twin is generated as follows:

[0093] Read the twin object records, node registration relationships, and object indexes. Use the twin object records as state layer data, the node registration relationships as registration layer data, the object indexes as retrieval layer data, and the flow direction field, structural connection field, and equipment association field as connection layer data. Establish a block directory according to the project block number and a node directory according to the digital twin node number. Write the state layer data, registration layer data, retrieval layer data, and connection layer data into the corresponding block directory and node directory to generate a hydraulic twin.

[0094] In this embodiment, generating the response trajectory map includes:

[0095] Read the project block number, monitoring node number, equipment node number, digital twin node number, node type, and node coordinates from the hydraulic twin, and establish a graph node. Specifically, establishing a graph node involves:

[0096] Read the engineering block number, monitoring node number, equipment node number, digital twin node number, node type, and node coordinates from the hydraulic twin. Use the digital twin node number as the node primary key, and use the engineering block number, monitoring node number, equipment node number, node type, and node coordinates as node attributes. Split the node coordinates into x-coordinate, y-coordinate, and elevation coordinate. Write the node primary key and node attributes corresponding to each digital twin node number into one line of graph node record. Sort the graph node records in ascending order according to the engineering block number. Within the same engineering block number, sort the graph node records in ascending order according to the digital twin node number to generate graph nodes.

[0097] A flow conduction chain is generated based on the flow direction field in the hydraulic twin; a structural connection chain is generated based on the structural connection field; and a device action chain is generated based on the device association field.

[0098] The water flow conduction chain is generated as follows:

[0099] Read the flow direction field and graph nodes from the hydraulic twin. Extract the upstream node number, downstream node number, flow direction mark, and hydraulic connectivity mark from the flow direction field. Use the upstream node number that is hydraulically connected as the starting node and the corresponding downstream node number as the ending node. Calculate the x-coordinate difference, y-coordinate difference, and elevation coordinate difference between the coordinates of the starting node and the ending node. Sum the squares of the three differences and take the square root to obtain the flow node distance. Write the starting node, ending node, flow direction mark, and flow node distance into the flow conduction edge. Connect all flow conduction edges in the direction from the upstream node number to the downstream node number to generate the flow conduction chain.

[0100] The structure connection chain is generated as follows:

[0101] Read the structural connection field and graph nodes from the hydraulic twin. Extract the adjacent project block number, structural component number, connection type, and connection strength mark from the structural connection field. Use the digital twin node number corresponding to the adjacent project block number as the structural connection node. Establish structural connection edges for structural connection nodes with the same connection type and valid connection strength mark. Calculate the x-coordinate difference, y-coordinate difference, and elevation coordinate difference between the coordinates of two structural connection nodes. Sum the squares of the three differences and take the square root to obtain the structural node distance. Write the structural connection node, connection type, structural node distance, and structural component number into the structural connection edge. Connect the structural connection edges in ascending order of project block number to generate a structural connection chain.

[0102] The generation device action chain is as follows:

[0103] Read the equipment association field and graph nodes from the hydraulic twin. Extract the equipment node number, controlled monitoring node number, action type, and equipment action direction from the equipment association field. Take the digital twin node corresponding to the equipment node number as the equipment start node and the digital twin node corresponding to the controlled monitoring node number as the action end node. Calculate the x-coordinate difference, y-coordinate difference, and elevation coordinate difference between the coordinates of the equipment start node and the action end node. Sum the squares of the three differences and take the square root to obtain the equipment action distance. Write the equipment start node, action end node, action type, equipment action direction, and equipment action distance into the equipment action edge. Connect the equipment action edges in ascending order of equipment node number to generate the equipment action chain.

[0104] Read the node status field of the same node at adjacent acquisition timestamps from the standard status data, calculate the change between adjacent node status fields, generate node status change records according to the order in which the changes occur, and generate historical response paths based on the node status change records, where:

[0105] Generate node state change records, specifically as follows:

[0106] Read the node status field of the same node under adjacent acquisition timestamps in the standard status data. Take the node status field of the previous acquisition timestamp as the preceding status and the node status field of the next acquisition timestamp as the current status. Calculate the difference between the current status value and the preceding status value for each of the following: water level, flow rate, seepage pressure, seepage flow, displacement, settlement, crack width, vibration, gate opening, equipment current, rainfall, temperature, wind speed, crack pixel length, and apparent anomaly confidence. Write the field with a difference greater than 0 into an upward change flag, the field with a difference less than 0 into a downward change flag, and the field with a difference equal to 0 into a hold change flag. Write the acquisition timestamp, node number, field name, difference, change flag, and data source into the node status change record.

[0107] The historical response path is generated as follows:

[0108] Read node status change records, water flow conduction chains, structural connection chains, and equipment action chains. Sort node status change records in ascending order of acquisition timestamp. Nodes whose change markers are not maintained within adjacent acquisition timestamps are designated as response nodes. Match response nodes with the start and end nodes in the water flow conduction chain, structural connection chain, and equipment action chain. Establish a response connection when two response nodes have a water flow conduction edge, structural connection edge, or equipment action edge. Designate the earlier response node as the path start point and the later response node as the path end point. Use the acquisition timestamp difference between two response nodes as the response interval. Write the path start point, path end point, response connection type, and response interval into the historical response path.

[0109] The graph nodes, water flow conduction chains, structural connection chains, equipment action chains, and historical response paths are written into the trajectory record. The trajectory record includes the start node, end node, chain type, trajectory number, and acquisition timestamp. Specifically, writing the graph nodes, water flow conduction chains, structural connection chains, equipment action chains, and historical response paths into the trajectory record is as follows:

[0110] Read the graph nodes, water flow conduction chains, structural connection chains, equipment action chains, and historical response paths. Write each water flow conduction edge in the water flow conduction chain into a trajectory record with chain type "water flow conduction". Write each structural connection edge in the structural connection chain into a trajectory record with chain type "structural connection". Write each equipment action edge in the equipment action chain into a trajectory record with chain type "equipment action". Write each response connection in the historical response path into a trajectory record with chain type "historical response". Arrange the start node, end node, chain type, trajectory number, and acquisition timestamp in each trajectory record according to the field order. The trajectory number is numbered consecutively starting from 1 in ascending order according to the acquisition timestamp to generate a trajectory record.

[0111] Connect the trajectory records according to the starting and ending nodes, and bind the trajectory records with the digital twin node numbers to generate a response trajectory map. Specifically, generating the response trajectory map involves:

[0112] Read the trajectory records and graph nodes, connect the trajectory records into a directed graph structure according to the start node and end node, match the start node in the trajectory record with the digital twin node number in the graph node, match the end node in the trajectory record with the digital twin node number in the graph node, write the successfully matched digital twin node number into the node binding field of the trajectory record, write the unmatched trajectory records into the verification queue, arrange the trajectory records with completed node binding according to the project block number, chain type and trajectory sequence number, write the graph node, trajectory record, node binding field and verification queue into the same graph structure container, and generate the response trajectory graph.

[0113] In this embodiment, generating the hysteresis fingerprint matrix includes:

[0114] Read the node number and node type in the hydraulic twin, read the collection timestamp in the standard state data, extract the water condition status from the water condition monitoring data, extract the seepage status from the seepage monitoring data, extract the structural deformation status from the structural deformation data, extract the equipment operation status from the equipment operation data, extract the environmental status from the environmental monitoring data, extract the inspection status from the image inspection data, and generate a node status sequence according to the node number and collection timestamp.

[0115] The source and target nodes are determined according to the historical response paths in the response trajectory map. The source node state change time, target node state change time, source node state change direction, and target node state change direction are extracted from the node state sequence, where:

[0116] The source and target nodes are determined as follows:

[0117] Read the historical response paths in the response trajectory map, extract the path start point, path end point, response connection type and response interval from each historical response path, write the path start point into the source node field and the path end point into the target node field. When the historical response path contains continuous path segments, connect the path segments with the same path end point as the next path start point. After connection, the first path start point is still used as the source node and the last path end point is used as the target node to generate source and target node pairing records.

[0118] Extract the source node state change time, target node state change time, source node state change direction, and target node state change direction, specifically as follows:

[0119] Read the source and target node pairing records and node status sequences. Match the node numbers in the node status sequences according to the source node number. Read the water status, seepage status, structural deformation status, equipment operation status, environmental status, and inspection status of the source node at each acquisition time stamp. Write the earliest acquisition time stamp of the change marker that is not the one that keeps the change marker into the source node status change time. Write the corresponding change marker into the source node status change direction. Perform the same matching process according to the target node number. Write the earliest acquisition time stamp of the change marker that is not the one that keeps the change marker and whose acquisition time stamp is not earlier than the source node status change time into the target node status change time. Write the corresponding change marker into the target node status change direction.

[0120] The response direction is determined based on the chronological order of source node status changes and target node status changes; the response type is determined based on the source node type, target node type, and inspection status; the lag time is determined based on the time interval between the two changes; and the water load phase is determined based on the direction of water level and flow rate changes in the hydrological status, as well as the gate opening changes and pump station start / stop status in the equipment operation status.

[0121] Determine the direction of the response, specifically:

[0122] Read the state change time of the source node and the state change time of the target node, calculate the time difference by subtracting the state change time of the source node from the state change time of the target node. When the time difference is greater than 0, write the response direction as the source node pointing to the target node. When the time difference is equal to 0, read the historical response path direction in the response trajectory map and write the historical response path direction as the response direction. When the time difference is less than 0, write the response direction as the target node pointing to the source node.

[0123] Determine the response type, specifically:

[0124] Read the source node type, target node type, inspection status, and source-target node pairing record. When the source node type is a water condition node and the target node type is a seepage node, write the water condition to seepage response type. When the source node type is a seepage node and the target node type is a structural node, write the seepage to structural response type. When the source node type is a structural node and the target node type is an inspection node, write the structural to apparent response type. When the source node type is an equipment node and the target node type is a water condition node, write the equipment to water condition response type. When the apparent anomaly valid flag in the inspection status is valid, write the apparent anomaly additional flag in the response type and generate the response type.

[0125] The lag time is determined as follows:

[0126] Read the source node's state change time and the target node's state change time, convert both times into second-level timestamps, calculate the target node's second-level timestamp minus the source node's second-level timestamp to obtain the time interval in seconds, take the absolute value of the time interval in seconds to obtain the lag time, write a synchronization response flag when the lag time is 0, and write a lag response flag when the lag time is greater than 0, and write the lag time and response flag into the source and target node pairing record;

[0127] The water load phase is determined as follows:

[0128] Read the direction of water level change and flow rate change in the water situation status, read the gate opening change status and pump station start / stop status in the equipment operation status, write the rising water phase when the water level change direction is marked as rising water level and the flow rate change direction is marked as increasing flow rate, write the receding water phase when the water level change direction is marked as falling water level and the flow rate change direction is marked as decreasing flow rate, write the stabilizing water phase when the water level change direction is marked as maintaining water level and the flow rate change direction is marked as maintaining flow rate, write the scheduling disturbance phase when the gate opening change status is not marked as maintaining gate opening or the pump station start / stop status is not marked as maintaining, and write the water load phase into the source target node pairing record;

[0129] Write the source node, target node, response direction, response type, lag time, water load phase, inspection status, source node status change direction, and target node status change direction into the matrix elements.

[0130] Arrange the matrix elements according to the source node number and the target node number, and bind the matrix elements to the collection timestamp to generate a hysteresis fingerprint matrix. Specifically, generating the hysteresis fingerprint matrix involves:

[0131] Read all matrix elements, arrange matrix rows in ascending order of source node number, arrange matrix columns in ascending order of target node number, and arrange matrix layers in ascending order of water load phase. The water load phases are arranged in the order of rising water phase, receding water phase, stable water phase, and scheduling disturbance phase. Arrange multiple matrix elements under the same source node number, the same target node number, and the same water load phase in ascending order of acquisition timestamp. Bind matrix elements to acquisition timestamps. Write matrix row index, matrix column index, matrix layer index, matrix elements, and acquisition timestamps into the same matrix container to generate a hysteresis fingerprint matrix.

[0132] In this embodiment, generating the phase adjacency matrix includes:

[0133] Read the response trajectory map and the hysteresis fingerprint matrix, use the map nodes as the variable set of the MMHC algorithm, and use the historical response paths and matrix elements as candidate relationship records;

[0134] During the MMPC candidate parent-child node search phase, a path memory search is performed. Adjacent node relationships in the candidate relationship records are checked according to historical response paths. Adjacent nodes that pass the MMPC conditional independence test are written into the candidate parent-child node set, where:

[0135] During the MMPC candidate parent-child node search phase, a path memory search is performed, specifically as follows:

[0136] Read candidate relationship records and historical response paths. Count the number of consecutive occurrences of the same node pair in the historical response path according to the source node number and target node number. Write a path memory hit mark when the number of consecutive occurrences is not less than 3, and write a path memory miss mark when the number of consecutive occurrences is less than 3. Write the node pair with the path memory hit mark into the MMPC queue to be checked first, and write the node pair with the path memory miss mark into the tail of the MMPC queue to be checked in ascending order of lag time to generate path memory search results.

[0137] Neighboring nodes that pass the MMPC conditional independence test are added to the candidate parent-child node set, specifically:

[0138] Read the path memory search results, standard state data, and hysteresis fingerprint matrix. Align the source node state sequence and target node state sequence of the node pair to be tested with the same collection timestamp. Calculate the correlation statistics between the source node state sequence and the target node state sequence. Add the node state sequences in the candidate adjustment node set to the condition set one by one. Recalculate the correlation statistics under the condition set. Set the conditional independence test threshold to 0.05. The threshold is based on the 5% significance level allowed in water conservancy project safety monitoring to eliminate weakly correlated incidental connections. When the test result is not greater than 0.05, retain the adjacent node relationship. When the test result is greater than 0.05, delete the adjacent node relationship. Write the retained adjacent nodes into the candidate parent-child node set.

[0139] In the Hill-Climbing structural optimization phase, skeleton reconstruction is performed. Candidate parent-child node sets are read according to the water load phase. Within the node range defined by the candidate parent-child node sets, edge addition, edge deletion, and edge reverse search are performed. A phase causal skeleton is generated using the Hill-Climbing structural score, where:

[0140] During the Hill-Climbing structural optimization phase, skeleton reconstruction is performed, specifically as follows:

[0141] Read the water load phase in the candidate parent-child node set and the lag fingerprint matrix. Divide the candidate parent-child node set into four phase subsets according to the rising water phase, receding water phase, stable water phase and scheduling disturbance phase. Read the candidate node pairs and corresponding lag times in each phase subset. Generate undirected candidate edges according to the candidate node pairs. Delete duplicate candidate edges and retain the same-direction candidate edge with the smallest lag time. Write the undirected candidate edges in each phase subset into the corresponding phase skeleton layer to generate the phase skeleton to be scored.

[0142] A phase causal skeleton is generated through Hill-Climbing structure scoring, specifically as follows:

[0143] Read the phase skeleton to be scored, standard state data, and candidate parent-child node set. Use the phase skeleton to be scored as the initial structure, and sequentially perform edge addition, edge deletion, and edge reverse search, changing only one edge at a time. After each change, check whether a directed cycle is formed. If a directed cycle is formed, discard the change. If no directed cycle is formed, calculate the structure score. The structure score is scored using the Bayesian information criterion. The score consists of the node conditional likelihood term minus the complexity penalty term. The complexity penalty term is calculated based on the number of parent nodes and the number of samples. If the score increases, accept the change. If the score does not increase, revert to the structure before the change. Stop the search when the score does not increase after 10 consecutive searches, and generate the phase causal skeleton.

[0144] During the edge search phase, a delayed edge search is performed. Based on the delayed fingerprint matrix, the delayed time, response type, inspection status, source node state change direction, and target node state change direction are written into the causal edges in the phase causal skeleton. Specifically, the delayed edge search during the edge search phase is performed as follows:

[0145] Read the phase causal skeleton and hysteresis fingerprint matrix. According to the source node number, target node number and water load phase retrieval matrix elements of the causal edge, extract all the hysteresis times corresponding to the same causal edge. Write the hysteresis time with the most occurrences into the causal edge hysteresis field. For hysteresis times with the same occurrences, take the smaller value. Write the response type, inspection status, source node status change direction and target node status change direction into the causal edge attribute table in ascending order of collection timestamp to generate phase causal edges with hysteresis attributes.

[0146] The causal edges are arranged with the source node number as the matrix row and the target node number as the matrix column. The water load phase is used as the matrix layer label. The phase adjacency matrix is ​​generated according to the directed acyclic graph output process. Specifically, the generation of the phase adjacency matrix according to the directed acyclic graph output process is as follows:

[0147] Read the causal edges of the phases with hysteresis attributes, use the source node number as the matrix row index, the target node number as the matrix column index, and the water load phase as the matrix layer index. The water load phase layers are arranged according to the rising water phase, receding water phase, stable water phase, and scheduling disturbance phase. Write the edge existence mark, hysteresis time, response type, inspection status, source node state change direction, and target node state change direction at the corresponding row, column, and layer positions. Perform a directed acyclic graph check on each phase layer. If a directed cycle is found, delete the edge with the lowest structural score and recheck until there are no directed cycles in the phase layer. Then merge all phase layers to generate a phase adjacency matrix.

[0148] In this embodiment, the generation of the running status prediction results and residual records includes:

[0149] An improved TrajGRU model is constructed, which includes the trajectory generation network, hidden state sampling process, reset gate calculation process, update gate calculation process, candidate hidden state calculation process, and output mapping process of the TrajGRU model. A response trajectory generator, osmotic hysteresis gate cell, and residual backflow unit are configured. Specifically, the construction of the improved TrajGRU model involves:

[0150] The system reads node number, acquisition timestamp, and node status fields from the standard state data; it reads the graph node, trajectory record, and node binding fields from the response trajectory graph; and it reads the matrix row index, matrix column index, matrix layer index, edge existence marker, lag time, response type, inspection status, source node state change direction, and target node state change direction from the phase adjacency matrix. The improved TrajGRU model is a rewritten version of the original TrajGRU model, retaining the trajectory generation network, hidden state sampling process, reset gate calculation process, update gate calculation process, candidate hidden state calculation process, and output mapping process. The trajectory output representing the two-dimensional spatial offset position in the TrajGRU model is rewritten as a 4D trajectory offset index, where the 4D trajectory offset index sequentially represents the candidate source node number, candidate target node number, trajectory sequence number, and lag time index. This is then applied to the trajectory generation network. Then, a response trajectory generator is inserted. The response trajectory generator matches the 4D trajectory offset index with the matrix row index, matrix column index, matrix layer index in the phase adjacency matrix and the trajectory record in the response trajectory map. The hidden state sampling in the TrajGRU model oriented towards two-dimensional spatial location is rewritten as hidden state sampling oriented towards digital twin nodes and engineering response paths. During the candidate hidden state calculation process, a seepage hysteresis gate cell is inserted. The input of the candidate hidden state calculation process is expanded from the current input features and trajectory sampling hidden state to the current input features, trajectory sampling hidden state, water load state, seepage response state, structural response state and apparent damage state. Before the output mapping process, a residual backflow unit is inserted to write the residual record of the previous acquisition timestamp into the current hidden state. This results in an improved TrajGRU model that includes engineering response trajectory matching, seepage hysteresis gate update and residual closed-loop correction.

[0151] In the improved TrajGRU model, the node state fields are arranged in the following order: water level, flow rate, seepage pressure, seepage flow, displacement, settlement, crack width, vibration, gate opening, equipment current, rainfall, temperature, wind speed, crack pixel length, and apparent anomaly confidence, forming a 15-dimensional current input vector. The hidden state dimension is set to 32 dimensions. The trajectory generation network reads the 15-dimensional current input vector and the 32-dimensional hidden state from the previous acquisition timestamp and concatenates them into a 47-dimensional trajectory input vector. The 47-dimensional trajectory input vector is then processed by a 64-dimensional intermediate... The layer mapping is a 4D trajectory offset index. The response trajectory generator reads the 4D trajectory offset index, phase adjacency matrix, and response trajectory map and outputs the engineering response trajectory input. The hidden state sampling process reads the 32D hidden state of the corresponding source node based on the engineering response trajectory input and outputs the 32D trajectory sampling hidden state. The reset gate calculation process and update gate calculation process read the 54D gating input vector and output the 32D reset gate state and 32D update gate state, respectively. The 54D gating input vector consists of a 15D current input vector, a 3D phase adjacency matrix, a phase adjacency matrix, and a phase adjacency matrix, and outputs the 3D trajectory sampling hidden state. The hidden state of the 2D trajectory sampling is obtained by concatenating the 4D response type encoding, the 2D inspection state encoding, and the 1D hysteresis time normalization value. The pressure hysteresis gate cell reads the 32D reset gate state, the 32D update gate state, and the 63D candidate input vector and outputs the 32D current hidden state. The 63D candidate input vector is obtained by concatenating the 15D current input vector, the 32D reset trajectory sampling hidden state, the 4D water load state vector, the 4D pressure response state vector, the 5D structural response state vector, and the 3D apparent damage state vector. The residual return unit... Read the 15-dimensional residual vector of the previous acquisition timestamp and map it to a 32-dimensional residual embedding vector. Add the 32-dimensional residual embedding vector to the 32-dimensional current hidden state element by element to generate a 32-dimensional backflow hidden state. The output mapping process maps the 32-dimensional backflow hidden state to a 15-dimensional running state prediction vector. Subtract the 15-dimensional running state prediction vector from the 15-dimensional real-time state vector of the next acquisition timestamp item by item to generate a 15-dimensional residual vector. Generate residual records by arranging them according to node number, acquisition timestamp, field name, predicted value, real-time value, and difference.

[0152] The training process of the model is as follows: Historical standard state data, historical response trajectory maps, and historical phase adjacency matrices are read. The historical standard state data is sorted in ascending order according to the digital twin node number and the acquisition timestamp. The water level, flow rate, seepage pressure, seepage flow, displacement, settlement, crack width, vibration, gate opening, equipment current, rainfall, temperature, wind speed, crack pixel length, and apparent anomaly confidence of each node at each acquisition timestamp are arranged into a 15-dimensional input vector. The 15-dimensional input vector of the current acquisition timestamp is used as the training input, and the 15-dimensional node state vector of the next acquisition timestamp is used as the training input. Training labels are used to divide samples into training, validation, and test samples in ascending order of collection timestamps, with training samples comprising 70%, validation samples 15%, and test samples 15%. Model weights are initialized using a uniform distribution, with the boundary values ​​set to the square root of the inverse of the input dimension. Initial weights are uniformly drawn between positive and negative boundaries. The initial bias value is set to 0 to avoid gating saturation due to excessively large initial hidden states. The maximum number of training epochs is set to 200, the batch size is set to 64, and the learning rate is set to 0.001 to ensure a stable decrease in the prediction error of the node state sequence. The loss function is 1. The weighted sum of the squared mean of the prediction differences for the five fields is used. The field loss weight is calculated as the reciprocal of the standard deviation of the corresponding field in the training samples. The 15 reciprocals are summed to obtain the normalized denominator. The reciprocal of each field is divided by the normalized denominator to obtain the corresponding field weight. The setting is based on the fact that water level, seepage pressure, crack width, and apparent anomaly confidence have different dimensions, and it is necessary to reduce the dominance of large-value fields on the training objective. During training, the training samples are input into the improved TrajGRU model, which sequentially generates the engineering response trajectory input, trajectory sampling hidden state, reset gate state, update gate state, current hidden state, backflow hidden state, and 15-dimensional operating state prediction. The predicted vector is measured, and the weighted average of the squared differences between the predicted vector and the training label is calculated. The trainable parameters in the trajectory generation network, reset gate calculation process, update gate calculation process, candidate hidden state calculation process, residual backflow unit and output mapping process are updated using the Adam optimizer. After each round of training, the validation sample is read to calculate the validation loss. Training is stopped when the validation loss does not decrease for 20 consecutive rounds. The model parameters corresponding to the round with the lowest validation loss are read as the final model parameters. Then, the prediction difference, residual ratio and residual direction consistency rate of 15 fields are calculated using test samples to generate the improved TrajGRU model that has been trained.

[0153] Read the phase adjacency matrix, response trajectory spectrum, and standard state data; arrange the node state fields according to the acquisition timestamp; and generate the current input features, initial hidden state, and initial residual record. Specifically, generating the current input features, initial hidden state, and initial residual record involves:

[0154] Read the phase adjacency matrix, response trajectory spectrum and standard state data. Sort the node state fields corresponding to the same node number in ascending order according to the acquisition timestamp. Write the 15-dimensional input vector under each acquisition timestamp into the current input feature. Stack the current input features of all nodes under the same acquisition timestamp in ascending order according to the node number to generate a node input matrix. Set the initial hidden state to a zero matrix of the number of nodes multiplied by 32. Set the initial residual record to a zero matrix of the number of nodes multiplied by 15. Write 0 to each element in the zero matrix. The initial hidden state and the initial residual record are both bound to the first acquisition timestamp.

[0155] The trajectory generation network generates a trajectory offset index based on the current input features and the hidden state of the previous acquisition timestamp. The response trajectory generator performs node matching between the trajectory offset index and the causal edges in the phase adjacency matrix and the trajectory records in the response trajectory graph to generate the engineering response trajectory input, where:

[0156] The response trajectory generator includes:

[0157] Trajectory Index Parser: Reads the 4D trajectory offset index output by the trajectory generation network, parses the candidate source node number, candidate target node number, trajectory sequence number and lag time index, and generates trajectory index parsing results;

[0158] Causal edge matcher: Reads the matrix row index, matrix column index, matrix layer index, edge existence marker and lag time from the phase adjacency matrix, matches the candidate source node number with the matrix row index, matches the candidate target node number with the matrix column index, matches the lag time index with the lag time, and generates causal edge matching results;

[0159] Track record matcher: Reads the track record, start node, end node, chain type, track number and collection timestamp from the response track map, matches the track number in the track offset index with the track record, and generates the track record matching result;

[0160] Hydraulic path validator: Reads the causal edge matching results and trajectory record matching results, performs consistency checks on candidate source nodes, candidate target nodes, start nodes and end nodes, and generates hydraulic path verification results;

[0161] Lag Time Aligner: Reads the lag time and current sampling timestamp from the causal edge, subtracts the lag time from the current sampling timestamp to obtain the sampling timestamp, and generates a hidden state sampling time index;

[0162] Response attribute splicer: Reads response type, inspection status, source node status change direction, target node status change direction, chain type and water load phase, splices the above fields in order, and generates an engineering response attribute vector;

[0163] Sampling address outputter: Reads candidate source node number, candidate target node number, hidden state sampling time index and engineering response attribute vector, and generates engineering response trajectory input. The engineering response trajectory input is used to read the hidden state of the corresponding source node at the previous collection timestamp during the hidden state sampling process.

[0164] The trajectory generation network generates a trajectory offset index based on the current input features and the hidden state of the previous acquisition timestamp, specifically as follows:

[0165] Read the current input features and the hidden state of the previous acquisition timestamp. Concatenate the 15-dimensional input vector from the current input features with the 32-dimensional hidden state of the previous acquisition timestamp of the corresponding node in the field order to generate a 47-dimensional trajectory input vector. Input the 47-dimensional trajectory input vector into the trajectory generation network. After linear mapping and hyperbolic tangent activation, a 4-dimensional trajectory output vector is obtained. Map the first-dimensional trajectory output vector to the candidate source node number according to the number of nodes. Map the second-dimensional trajectory output vector to the candidate target node number according to the number of nodes. Map the third-dimensional trajectory output vector to the trajectory sequence number according to the number of trajectory records in the response trajectory map. Map the fourth-dimensional trajectory output vector to the lag time index according to the lag time value sequence in the phase adjacency matrix to generate the trajectory offset index.

[0166] The input for generating the engineering response trajectory is as follows:

[0167] Read the trajectory offset index, phase adjacency matrix, and response trajectory map. Match the candidate source node number with the matrix row index of the phase adjacency matrix, match the candidate target node number with the matrix column index of the phase adjacency matrix, match the lag time index with the lag time in the phase adjacency matrix, and match the trajectory sequence number with the trajectory record in the response trajectory map. When the matrix row index, matrix column index, lag time, and trajectory sequence number are all successfully matched, read the corresponding causal edge and trajectory record. When there are unmatched items, search the trajectory record in the response trajectory map that is closest to the current trajectory sequence number according to the same source node number and target node number. Merge the response type, inspection status, source node status change direction, and target node status change direction in the causal edge with the starting node, ending node, chain type, and acquisition timestamp in the trajectory record according to the field order to generate the engineering response trajectory input.

[0168] The hidden state sampling process reads the hidden state of the previous acquisition timestamp of the corresponding source node according to the engineering response trajectory input, and generates the trajectory sampling hidden state. Specifically, the generation of the trajectory sampling hidden state is as follows:

[0169] Read the starting node, ending node, lag time, and acquisition timestamp from the engineering response trajectory input. Subtract the lag time from the current acquisition timestamp to obtain the sampling timestamp. Search the hidden state corresponding to the starting node in the hidden state sequence table of the previous acquisition timestamp. If a corresponding hidden state exists for the sampling timestamp, read the corresponding hidden state. If no corresponding hidden state exists for the sampling timestamp, read the hidden state with the smallest time difference that is less than the sampling timestamp. Write the read 32-dimensional hidden state to the sampling position of the ending node. If multiple starting node hidden states exist for the same ending node, sort them in descending order according to the structure score in the phase adjacency matrix and select the first-ranked hidden state to generate the trajectory sampling hidden state.

[0170] The reset gate calculation process and the update gate calculation process generate the reset gate state and the update gate state based on the current input features, the trajectory sampling hidden state, and the engineering response trajectory input. Specifically, the reset gate calculation process and the update gate calculation process generate the reset gate state and the update gate state based on the current input features, the trajectory sampling hidden state, and the engineering response trajectory input, as follows:

[0171] Read the current input features, trajectory sampling hidden state, and engineering response trajectory input. Concatenate the 15-dimensional input vector from the current input features, the 32-dimensional hidden vector from the trajectory sampling hidden state, the response type code, the inspection status code, and the lag time normalized value from the engineering response trajectory input according to the field order to generate a gated input vector. Generate a 4-dimensional one-hot vector from the response type code in the order of water condition to seepage, seepage to structure, structure to appearance, and equipment to water condition. Generate a 2-dimensional one-hot vector from the inspection status code based on the appearance anomaly valid marker and appearance anomaly invalid marker. The lag time normalized value is equal to the lag time divided by the maximum lag time in the current training sample. Input the gated input vector into the reset gate linear mapping and obtain the reset gate state through S-shaped activation. Input the gated input vector into the update gate linear mapping and obtain the update gate state through S-shaped activation. Both the reset gate state and the update gate state are 32-dimensional vectors.

[0172] The pressure hysteresis gated cell, based on resetting the gate state, updating the gate state, current input features, and trajectory sampling, performs candidate hidden state calculation according to the water load state formed by the hydrological state, the pressure response state formed by the seepage state, the structural response state formed by the structural deformation state, and the apparent damage state formed by the inspection state, to generate the current hidden state, where:

[0173] The osmolar hysteresis gated cells include:

[0174] Water load state encoder: Reads the direction of water level change and flow rate change in the water situation, converts the water level rise marker, water level fall marker, water level hold marker, flow rate increase marker, flow rate decrease marker and flow rate hold marker into a water load state vector, and generates water load state coding results;

[0175] The pressure response encoder reads the direction of pressure change and the direction of seepage change in the seepage state, converts the pressure increase marker, pressure decrease marker, pressure hold marker, seepage increase marker, seepage decrease marker, and seepage hold marker into a pressure response state vector, and generates the pressure response encoding result.

[0176] Structural response encoder: Reads the displacement change direction, settlement change direction, crack change direction and vibration change direction in the structural deformation state, converts the increase change marker, decrease change marker and maintain change marker into structural response state vector, and generates structural response encoding result;

[0177] Apparent damage encoder: Reads the crack change direction and apparent anomaly valid markers in the inspection status, converts crack expansion markers, crack contraction markers, crack retention markers, apparent anomaly valid markers and apparent anomaly invalid markers into apparent damage state vectors, and generates apparent damage coding results;

[0178] Hysteresis Sequence Arranger: Reads the water load state vector, seepage pressure response state vector, structural response state vector, and apparent damage state vector, arranges them in the order of water load state, seepage pressure response state, structural response state, and apparent damage state, and generates a hysteresis conduction state sequence.

[0179] Reset Actor: Reads the reset gate state and trajectory sampling hidden state, performs element-by-element reset processing on the 32-dimensional trajectory sampling hidden state, and generates the reset trajectory sampling hidden state;

[0180] Candidate State Fusion: Reads the current input features, the reset trajectory sampling hidden state and the hysteresis conduction state sequence, and concatenates the 15-dimensional current input features, the 32-dimensional reset trajectory sampling hidden state, the 4-dimensional water load state vector, the 4-dimensional seepage pressure response state vector, the 5-dimensional structural response state vector and the 3-dimensional apparent damage state vector into a 63-dimensional candidate input vector, and generates candidate hidden states;

[0181] Update gate fusion: Read the update gate state, trajectory sampling hidden state and candidate hidden state, perform element-wise fusion on the 32-dimensional trajectory sampling hidden state and the 32-dimensional candidate hidden state according to the update gate state, and generate the current hidden state;

[0182] Lag gated output: Reads the current hidden state, response type, lag time, and water load phase, binds the current hidden state with the response type, lag time, and water load phase, generates a current hidden state with lag attribute, and outputs it to the residual return unit;

[0183] The osmotic pressure hysteresis gated cell performs candidate hidden state calculations, specifically as follows:

[0184] The system reads the reset gate state, update gate state, current input features, trajectory sampling hidden state, water condition state, seepage state, structural deformation state, and inspection state. It extracts the water level change direction and flow rate change direction from the water condition state and encodes them as a water load state vector. It extracts the seepage pressure change direction and seepage flow change direction from the seepage state and encodes them as a seepage pressure response state vector. It extracts the displacement change direction, settlement change direction, crack change direction, and vibration change direction from the structural deformation state and encodes them as a structural response state vector. It extracts the crack change direction and effective markers of apparent anomalies from the inspection state and encodes them as an apparent damage state vector. It concatenates the water load state vector, seepage pressure response state vector, structural response state vector, apparent damage state vector, current input features, and trajectory sampling hidden state after the reset gate state are applied, according to the field order. It then performs a linear mapping of the input candidate hidden state and activates it with hyperbolic tangent to obtain the candidate hidden state. Finally, it performs element-wise fusion of the trajectory sampling hidden state and the candidate hidden state according to the update gate state to generate the current hidden state.

[0185] The residual backflow unit reads residual records according to the acquisition timestamp. For the first acquisition timestamp, it reads the initial residual record; for subsequent acquisition timestamps, it reads the residual record from the previous acquisition timestamp. The read residual records are written to the current hidden state, and the running state prediction result is generated through the output mapping process. Where:

[0186] The residual reflux unit includes:

[0187] Residual Reader: Reads the initial residual record or the residual record of the previous acquisition timestamp, and generates the residual record to be reflowed;

[0188] Residual Aligner: Aligns the residual records to be reflowed with the current hidden state according to the node number and field name, and generates a 15-dimensional residual vector;

[0189] Residual Embedder: Maps a 15-dimensional residual vector to a 32-dimensional residual embedding vector and generates the residual embedding result;

[0190] Hidden State Fusion: Writes the 32-dimensional residual embedding vector into the 32-dimensional current hidden state and generates a 32-dimensional backflow hidden state;

[0191] Predictor output: Reads the 32-dimensional backflow hidden state and generates a 15-dimensional running state prediction vector through the output mapping process;

[0192] Residual generator: Subtracts the 15-dimensional running state prediction vector from the corresponding real-time state vector item by item to generate residual records;

[0193] The residual reflux unit generates the operational status prediction results, specifically:

[0194] The process reads the current hidden state, initial residual record, previous acquisition timestamp residual record, and acquisition timestamp. When the acquisition timestamp is the first acquisition timestamp, the initial residual record is read; when the acquisition timestamp is not the first acquisition timestamp, the previous acquisition timestamp residual record is read. The read residual records are aligned with the current hidden state according to the node number. The 15-dimensional residual vector is converted into a 32-dimensional residual embedding vector through the residual mapping layer. The 32-dimensional residual embedding vector is added element-wise to the 32-dimensional current hidden state to generate the backflow hidden state. The backflow hidden state is input to the output mapping process. The output mapping process converts the 32-dimensional backflow hidden state into a 15-dimensional prediction vector. The 15 components in the prediction vector correspond to water level, flow rate, seepage pressure, seepage flow, displacement, settlement, crack width, vibration, gate opening, equipment current, rainfall, temperature, wind speed, crack pixel length, and apparent anomaly confidence, respectively, to generate the operation status prediction result.

[0195] The difference between the predicted operational status and the corresponding real-time status record in the standard status data is calculated. The difference results are then arranged according to node number, acquisition timestamp, and residual type to generate residual records. Specifically, generating residual records involves:

[0196] Read the corresponding real-time status records from the operational status prediction results and standard status data. Align the prediction vector and real-time status vector according to the node number and acquisition timestamp. Calculate the difference between the predicted value and the real-time value for each of the following: water level, flow rate, seepage pressure, seepage flow, displacement, settlement, crack width, vibration, gate opening, equipment current, rainfall, temperature, wind speed, crack pixel length, and apparent anomaly confidence level. Write positive residual markers for fields with differences greater than 0, negative residual markers for fields with differences less than 0, and zero residual markers for fields with differences equal to 0. Arrange the node number, acquisition timestamp, field name, predicted value, real-time value, difference, and residual markers in field order to generate residual records.

[0197] In this embodiment, constructing the phase twin map includes:

[0198] Read the operation status prediction results, residual records, phase adjacency matrix and hydraulic twin, and match the project blocks, monitoring nodes, equipment nodes and causal edges according to the digital twin node number;

[0199] The residual records are written into the corresponding causal edges in the phase adjacency matrix, and risk node records are generated according to the residual source node and residual target node, where:

[0200] The residual record is written into the corresponding causal edge in the phase adjacency matrix, specifically as follows:

[0201] Read the residual records and the causal edge mapping table. Retrieve the source node number and target node number in the causal edge mapping table according to the node number in the residual record. When the node number is equal to the source node number, write the residual record into the source node residual field of the causal edge. When the node number is equal to the target node number, write the residual record into the target node residual field of the causal edge. Sort the residual records corresponding to the same causal edge in ascending order according to the collection timestamp. Write the field name, difference, residual tag and collection timestamp into the causal edge residual sequence.

[0202] The risk node record is generated as follows:

[0203] Read the causal edge residual sequence and prediction node mapping table. Write the field where the absolute value of the difference between three consecutive collection timestamps under the same node number is greater than 0 into the residual persistence field. Divide the absolute value of the difference by the absolute value of the real-time value and add 1 to get the residual ratio. Write the field where the residual ratio is not less than 0.1 into the effective residual field. Write the node number that contains both the residual persistence field and the effective residual field into the risk node number. Write the project block number, monitoring node number, equipment node number, field name, collection timestamp, difference, residual ratio and residual tag corresponding to the risk node number into the risk node record.

[0204] Risk paths are generated based on causal edges, lag time, response type, inspection status, and risk node records. Risk types, risk levels, and sensor anomaly markers are then generated according to the number of consecutive causal edges, residual persistence status, and node association status within the risk path.

[0205] Risk paths are generated based on causal edges, lag times, response types, inspection status, and risk node records, specifically as follows:

[0206] Read the causal edges, lag times, response types, and inspection status from the phase adjacency matrix. Read the risk node records and match the risk node numbers with the source node numbers and target node numbers in the causal edges. Write the causal edges containing the risk node numbers into the candidate risk edges. Group the candidate risk edges according to the water load phase. Within the same water load phase, connect the candidate risk edges whose target node numbers are the same as the source node numbers of the previous candidate risk edge and the next candidate risk edge to form a continuous path. Arrange the node numbers, causal edge numbers, lag times, response types, inspection status, and collection timestamps in the continuous path according to the connection order to generate the risk path.

[0207] Risk type, risk level, and sensor anomaly markers are generated based on the number of consecutive causal edges, residual persistence status, and node association status in the risk path. Specifically:

[0208] Read the risk path, risk node record, causal edge residual sequence, and inspection status. When the response type in the risk path includes water condition to seepage response type and the target node residual field includes seepage pressure value, write the seepage hysteresis risk type. When the response type includes seepage to structural response type and the target node residual field includes displacement value, settlement value, crack width, or vibration value, write the structural response risk type. When the response type includes structural to apparent response type and the inspection status includes apparent anomaly valid marker, write the apparent damage risk type. When the number of consecutive causal edges is not less than 3 and there are valid residual fields in 3 consecutive acquisition timestamps, write the level 3 risk level. When the number of consecutive causal edges is 2 and there are valid residual fields in 2 consecutive acquisition timestamps, write the level 2 risk level. When the number of consecutive causal edges is 1 and there is a valid residual field, write the level 1 risk level. When the residual ratio of the same node is not less than 0.3 and there are no corresponding source node residual fields and target node residual fields in adjacent causal edges, write the sensor anomaly marker.

[0209] Risk node records, risk paths, risk types, risk levels, and sensor anomaly markers are written into the graph structure to generate a phase twin graph. Specifically, generating the phase twin graph involves:

[0210] Read risk node records, risk paths, risk types, risk levels, and sensor anomaly markers. Write the risk node records into the node table of the graph structure. Write the causal edge numbers, source node numbers, target node numbers, lag times, and response types in the risk paths into the edge table of the graph structure. Write the risk type, risk level, sensor anomaly markers, water load phase, and acquisition timestamp into the attribute table of the graph structure. Establish graph structure partitions according to the project block numbers and establish graph structure layers according to the water load phases. Write the node table, edge table, and attribute table into the corresponding graph structure partitions and graph structure layers to generate a phase twin graph.

[0211] In this embodiment, the generation of security monitoring results includes:

[0212] Read the phase twin diagram and the hydraulic twin, and write back the risk node records and risk paths in the phase twin diagram to the hydraulic twin according to the digital twin node numbers;

[0213] Extract the project block number, monitoring node number, equipment node number, risk path, risk type, risk level, and sensor anomaly marker from the written-back hydraulic twin to generate a safety monitoring record;

[0214] The safety monitoring records are arranged according to the collection timestamp, and the safety monitoring records are bound to the phase twin map to generate safety monitoring results.

[0215] Example 1: In a continuous operation monitoring scenario at a water conservancy project site, the monitoring object is a medium-sized hydraulic structure complex, comprising 8 engineering blocks, 36 monitoring nodes, 12 equipment nodes, and 56 digital twin nodes. The system receives 72 hours of continuous hydraulic monitoring data, with a data acquisition interval of 60 seconds, resulting in 4320 acquisition moments. 60 data acquisition channels generate 259,200 raw data records. Water level values ​​range from 42.18 to 44.76 meters, flow rate values ​​range from 186.4 to 351.7 cubic meters per second, seepage pressure values ​​range from 71.2 to 96.8 kPa, displacement values ​​range from 0.8 to 7.6 millimeters, and crack widths range from 0.12 to 1.48 millimeters. The raw data suffers from asynchronous sampling, seepage pressure delay, gate scheduling disturbance, and lag in inspection images. Traditional fixed threshold methods only trigger an alarm when the seepage pressure exceeds 95 kPa, failing to identify the continuous risk path consisting of increased seepage pressure delay, structural displacement expansion, and apparent crack propagation after water level rise.

[0216] The data normalization module converts all collected timestamps into yyyyMMddHHmmss format, generates a standard timeline with an alignment interval of 60 seconds, and writes each numerical record to the nearest aligned time point. Data with a time difference exceeding 30 seconds is entered into the completion queue. Before processing, there were 4968 time offset records with an average offset of 18.6 seconds and a maximum offset of 74 seconds. After processing, 257984 valid aligned records and 1216 completion records are generated, and the average time offset is reduced to 0 seconds. The twin registration module reads the engineering block boundary coordinates, monitoring node coordinates, equipment node coordinates, and digital twin node coordinates from the engineering configuration data. It binds 36 monitoring nodes and 12 equipment nodes to 8 engineering blocks and establishes node registration relationships for 56 digital twin nodes.

[0217] The trajectory connectivity module reads the water flow direction field, structural connection field, and equipment association field, generating 42 water flow conduction edges, 68 structural connection edges, and 18 equipment action edges. The system calculates the changes in water level, seepage pressure, displacement, and crack width at adjacent acquisition timestamps from standard state data, forming 137 historical response paths. In a simulated operation segment, the upstream water level at acquisition point 1260 rises from 42.96 meters to 43.41 meters, the flow rate increases from 211.8 cubic meters per second to 264.5 cubic meters per second, the seepage pressure node rises from 78.4 kPa to 84.9 kPa after 18 minutes, the corresponding dam section horizontal displacement increases from 2.1 mm to 3.4 mm after 42 minutes, and the crack pixel length recorded in the inspection image at acquisition point 1336 increases from 86 pixels to 119 pixels. The response trajectory map connects the above processes into continuous trajectories from water level nodes to seepage nodes, from seepage nodes to structural nodes, and from structural nodes to inspection nodes.

[0218] The phase encoding module calculates the difference between the source node's state change time and the target node's state change time, generating a hysteresis fingerprint matrix. The hysteresis time from water condition to seepage response is 600–1680 seconds, with a mean of 1020 seconds; the hysteresis time from seepage to structural response is 900–3600 seconds, with a mean of 2140 seconds; and the hysteresis time from structure to apparent response is 1800–5400 seconds, with a mean of 3360 seconds. The causal search module uses 56 digital twin nodes as the MMHC variable set and 137 historical response paths and hysteresis fingerprint matrix elements as candidate relation records. After path memory search, the number of node pairs is reduced from 284 to 96, the MMPC conditional independence test threshold is 0.05, and 61 node pairs are retained. Hill-Climbing structural optimization performs a phase search according to four types of water loads, generating a phase adjacency matrix containing 61 causal edges.

[0219] The trajectory reasoning module uses historical standard state data, historical response trajectory maps, and historical phase adjacency matrices to train the improved TrajGRU model. The training data comes from 90 running windows, each containing 10 acquisition times, and each acquisition time containing 56 digital twin nodes, forming a total of 50,400 node time-series samples: 35,280 training samples, 7,560 validation samples, and 7,560 test samples. The single training input is a 15-dimensional node state vector, with fields including water level (43.12 m), flow rate (238.6 m / s), seepage pressure (81.7 kPa), seepage flow rate (0.46 L / s), displacement (2.8 mm), settlement (0.9 mm), crack width (0.36 mm), vibration acceleration (0.018 g), gate opening (42%), equipment current (18.7 amps), rainfall (6.2 mm), temperature (22.4 degrees Celsius), wind speed (3.1 m / s), crack pixel length (96 pixels), and anomaly confidence (0.64). The model's hidden state dimension is 32, the batch size is 64, the learning rate is 0.001, training stops at the 86th epoch, and the mean squared difference of the prediction discrepancies on the test set is 0.017.

[0220] Within the same 72-hour simulation window, there were 20 real risk events. The traditional fixed threshold method triggered 15 alarms, with 12 real risk events, 3 false alarms, and 8 missed alarms, resulting in a false alarm rate of 20% and a missed alarm rate of 40%. The ordinary GRU combined with time-correlation post-processing method triggered 20 alarms, with 17 real risk events, 3 false alarms, and 3 missed alarms, resulting in a false alarm rate of 15% and a missed alarm rate of 15%. This invention triggered 20 alarms, with 19 real risk events, 1 false alarm, and 1 missed alarm, resulting in a false alarm rate of 5% and a missed alarm rate of 5%. The average identification delay for pressure hysteresis risk using the ordinary GRU combined with time-correlation post-processing method was 46 minutes, while the average identification delay using this invention was 14 minutes. The fixed threshold method only outputs single-point alarms, with a risk path integrity rate of 0%, while the ordinary GRU combined with time-correlation post-processing method achieved 41.5%, and this invention achieved 87.2%. Five sensor anomaly segments were set in the simulation data; the traditional fixed threshold method identified 2, the ordinary GRU combined with time-correlation post-processing method identified 3, and this invention identified 5. The phase twin graph outputs 23 risk node records, 7 risk paths, 4 level 3 risk nodes, 7 level 2 risk nodes, 12 level 1 risk nodes, and 5 sensor anomaly markers. Simulation results demonstrate that this invention can bind multi-source hydraulic engineering monitoring data to a hydraulic engineering twin, identify hysteresis risk paths through response trajectory maps, hysteresis fingerprint matrices, phase adjacency matrices, and an improved TrajGRU model, thereby improving the accuracy and traceability of safety monitoring results.

[0221] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A water conservancy project operation safety monitoring system based on digital twin, characterized in that, include: The data acquisition module collects hydraulic engineering monitoring data, performs preprocessing on the hydraulic engineering monitoring data, and generates standard state data. The twin mapping module performs twin registration of engineering nodes on standard state data, and generates hydraulic twins according to the node registration relationship between engineering blocks, monitoring nodes, equipment nodes and digital twin nodes; The trajectory construction module performs trajectory connection processing on the hydraulic twin, including the water flow transmission chain, structural connection chain, equipment action chain, and historical response path, to generate a response trajectory map. The fingerprint construction module performs source-target node response pairing on standard state data, performs phase lag encoding according to response direction, response type, lag time and water load phase, and generates a lag fingerprint matrix. The causal search module executes the MMHC algorithm based on the response trajectory map and the hysteresis fingerprint matrix, performing path memory search, skeleton reconstruction and hysteresis edge search to generate a phase adjacency matrix; The risk identification module constructs an improved TrajGRU model. The phase adjacency matrix and response trajectory map are input into the improved TrajGRU model. After processing by the response trajectory generator, the pressure hysteresis gate cell and the residual backflow unit, the operating status prediction results and residual records are generated. The twin write-back module constructs a phase twin graph based on the runtime prediction results, residual records, and phase adjacency matrix; The monitoring output module writes the phase twin image into the hydraulic twin to generate safety monitoring results.

2. The water conservancy project operation safety monitoring system based on digital twin as described in claim 1, characterized in that, The hydraulic monitoring data includes water condition monitoring data, seepage monitoring data, structural deformation data, equipment operation data, environmental monitoring data, image inspection data, and engineering configuration data.

3. The water conservancy project operation safety monitoring system based on digital twin according to claim 1, characterized in that, The preprocessing of hydraulic monitoring data includes time alignment, spatial binding, missing data completion, and anomaly marking.

4. A water conservancy project operation safety monitoring system based on digital twin as described in claim 1, characterized in that, The generated hydraulic twin includes: Read standard status data, extract project block number, monitoring node number, equipment node number, digital twin node number, node type, node coordinates, water flow direction field, structural connection field and equipment association field from the project configuration data, extract the collection timestamp from the standard status data, and generate node registration records; According to the project block number, the monitoring node number and the equipment node number are assigned to the corresponding project block, and the digital twin node number is bound to the project block number, the monitoring node number and the equipment node number to generate the node registration relationship; The node registration relationship is arranged according to the collection timestamp. The node registration relationship, node type, node coordinates, water flow direction field, structural connection field, equipment association field, and node status field extracted from water monitoring data, seepage monitoring data, structural deformation data, equipment operation data, environmental monitoring data and image inspection data are written into the twin object record. The twin object records are written into the object index according to the project block number, and the object index is associated with the digital twin node number to generate a hydraulic twin.

5. A water conservancy project operation safety monitoring system based on digital twin as described in claim 1, characterized in that, The generated response trajectory map includes: Read the engineering block number, monitoring node number, equipment node number, digital twin node number, node type and node coordinates from the hydraulic twin, and establish a graph node; A water flow conduction chain is generated based on the water flow direction field in the hydraulic twin; a structural connection chain is generated based on the structural connection field; and a device action chain is generated based on the device association field. Read the node status field of the same node under adjacent collection timestamps from the standard status data, calculate the change between adjacent node status fields, generate node status change records according to the order of occurrence of the changes, and generate historical response paths based on the node status change records. The graph nodes, water flow conduction chains, structural connection chains, equipment action chains, and historical response paths are written into the trajectory record. The trajectory record includes the starting node, ending node, chain type, trajectory number, and collection timestamp. Connect the trajectory records according to the starting node and the ending node, and bind the trajectory records with the digital twin node numbers to generate a response trajectory map.

6. A water conservancy project operation safety monitoring system based on digital twin as described in claim 1, characterized in that, The generation of the hysteresis fingerprint matrix includes: Read the node number and node type in the hydraulic twin, read the collection timestamp in the standard state data, extract the water condition status from the water condition monitoring data, extract the seepage status from the seepage monitoring data, extract the structural deformation status from the structural deformation data, extract the equipment operation status from the equipment operation data, extract the environmental status from the environmental monitoring data, extract the inspection status from the image inspection data, and generate a node status sequence according to the node number and collection timestamp. The source node and target node are determined according to the historical response path in the response trajectory map. The source node state change time, target node state change time, source node state change direction and target node state change direction are extracted from the node state sequence. The response direction is determined based on the order of the source node status change time and the target node status change time. The response type is determined based on the source node type, target node type, and inspection status. The lag time is determined based on the time interval between the two changes. The water load phase is determined based on the water level change direction and flow rate change direction in the water situation status, as well as the gate opening change status and pump station start / stop status in the equipment operation status. Write the source node, target node, response direction, response type, lag time, water load phase, inspection status, source node status change direction, and target node status change direction into the matrix elements. Arrange the matrix elements according to the source node number and the target node number, and bind the matrix elements to the collection timestamp to generate a hysteresis fingerprint matrix.

7. A water conservancy project operation safety monitoring system based on digital twin as described in claim 1, characterized in that, The generation of the phase adjacency matrix includes: Read the response trajectory map and the hysteresis fingerprint matrix, use the map nodes as the variable set of the MMHC algorithm, and use the historical response paths and matrix elements as candidate relationship records; During the MMPC candidate parent-child node search phase, a path memory search is performed. The adjacent node relationships in the candidate relationship records are checked according to the historical response paths. The adjacent nodes that pass the MMPC conditional independence test are written into the candidate parent-child node set. In the Hill-Climbing structural optimization stage, skeleton reconstruction is performed. Candidate parent-child node sets are read according to the water load phase. Within the node range defined by the candidate parent-child node sets, edge addition, edge deletion, and edge reverse search are performed. The phase causal skeleton is generated through the Hill-Climbing structural score. During the edge search phase, a delayed edge search is performed. Based on the delayed fingerprint matrix, the delayed time, response type, inspection status, source node state change direction, and target node state change direction are written into the causal edges in the phase causal skeleton. The causal edges are arranged with the source node number as the matrix row and the target node number as the matrix column, and the water load phase is used as the matrix layer label. The phase adjacency matrix is ​​generated according to the directed acyclic graph output process.

8. A water conservancy project operation safety monitoring system based on digital twin according to claim 1, characterized in that, The generated operational status prediction results and residual records include: An improved TrajGRU model is constructed, which includes the trajectory generation network, hidden state sampling process, reset gate calculation process, update gate calculation process, candidate hidden state calculation process, and output mapping process of the TrajGRU model, and is configured with a response trajectory generator, osmotic hysteresis gate cell, and residual backflow unit. Read the phase adjacency matrix, response trajectory map and standard state data, arrange the node state fields according to the acquisition timestamp, and generate the current input features, initial hidden state and initial residual record; The trajectory generation network generates a trajectory offset index based on the current input features and the hidden state of the previous acquisition timestamp. The response trajectory generator performs node matching between the trajectory offset index and the causal edges in the phase adjacency matrix and the trajectory records in the response trajectory graph to generate the engineering response trajectory input. The hidden state sampling process reads the hidden state of the previous acquisition timestamp of the corresponding source node according to the engineering response trajectory input, and generates the trajectory sampling hidden state; The reset gate calculation process and the update gate calculation process generate the reset gate state and the update gate state based on the current input features, the trajectory sampling hidden state, and the engineering response trajectory input. The pressure hysteresis gated cell performs candidate hidden state calculation based on the reset gate state, update gate state, current input features and trajectory sampling hidden state, and generates the current hidden state according to the water load state formed by the water condition state, the pressure response state formed by the seepage state, the structural response state formed by the structural deformation state and the apparent damage state formed by the inspection state. The residual backflow unit reads the residual records according to the acquisition timestamp. The first acquisition timestamp reads the initial residual record, and the subsequent acquisition timestamps read the residual record of the previous acquisition timestamp. The read residual records are written into the current hidden state, and the running state prediction result is generated through the output mapping process. The difference between the predicted running status and the corresponding real-time status record in the standard status data is calculated, and the difference results are arranged according to node number, collection timestamp and residual type to generate residual records.

9. A water conservancy project operation safety monitoring system based on digital twin as described in claim 1, characterized in that, The construction of the phase twin map includes: Read the operation status prediction results, residual records, phase adjacency matrix and hydraulic twin, and match the project blocks, monitoring nodes, equipment nodes and causal edges according to the digital twin node number; Write the residual records into the corresponding causal edges in the phase adjacency matrix, and generate risk node records according to the residual source node and residual target node; Risk paths are generated based on causal edges, lag time, response type, inspection status, and risk node records. Risk types, risk levels, and sensor anomaly markers are generated according to the number of consecutive causal edges, residual persistence status, and node association status in the risk paths. Risk node records, risk paths, risk types, risk levels, and sensor anomaly markers are written into the graph structure to generate a phase twin graph.

10. A water conservancy project operation safety monitoring system based on digital twin according to claim 1, characterized in that, The generated security monitoring results include: Read the phase twin diagram and the hydraulic twin, and write back the risk node records and risk paths in the phase twin diagram to the hydraulic twin according to the digital twin node numbers; Extract the project block number, monitoring node number, equipment node number, risk path, risk type, risk level, and sensor anomaly marker from the written-back hydraulic twin to generate a safety monitoring record; The safety monitoring records are arranged according to the collection timestamp, and the safety monitoring records are bound to the phase twin map to generate safety monitoring results.

Images