Deep Learning-Based Method and System for Monitoring Deformation of Hydraulic Engineering

Abstract

This invention discloses a deep learning-based method and system for monitoring deformation in hydraulic engineering projects, belonging to the field of hydraulic engineering deformation monitoring technology. The method specifically includes: collecting a hydraulic engineering deformation monitoring dataset, organizing the monitoring data into a benchmark format to obtain a sequence of total deformation at monitoring points, a sequence of structural attribution of monitoring points, a sequence of operating condition characteristics, and a sequence of operating condition stages; constructing a back-loop closed-loop separation network based on the sequence of total deformation at monitoring points, the sequence of operating condition characteristics, and the sequence of operating condition stages to obtain a sequence of recoverable operating condition response deformation at monitoring points and a sequence of initial and residual deformation at monitoring points; and constructing a partition boundary-preserving graph neural inference network based on the sequence of total deformation at monitoring points, the sequence of recoverable operating condition response deformation at monitoring points, the sequence of initial and residual deformation at monitoring points, and the sequence of structural attribution of monitoring points to obtain the hydraulic engineering deformation monitoring results. This application improves the accuracy and stability of hydraulic engineering deformation monitoring results.

Description

Technical Field

[0001] This invention belongs to the field of deformation monitoring of water conservancy projects, specifically involving a method and system for deformation monitoring of water conservancy projects based on deep learning. Background Technology

[0002] In the deformation monitoring of water conservancy projects, deep learning-based deformation analysis requires acquiring the displacement changes of measuring points on monitoring objects such as dams, dikes, sluice gates, and slopes, and combining this with water level changes and operational processes to determine the deformation state of the engineering structure. However, water conservancy projects are constantly affected by conditions such as water storage, flood discharge, gate opening and closing, and water level drop, making the sources of measuring point displacement changes quite complex. These displacement changes may originate from normal responses caused by water level fluctuations, or from dam settlement, joint misalignment, or cumulative structural deformation, making it difficult to directly interpret the deformation results.

[0003] In existing technologies, deep learning-based methods for monitoring deformation in hydraulic engineering typically use historical displacement, water level, and operational records as unified inputs, then directly output the total deformation or deformation prediction results at the measuring points through a deep learning model. However, the total deformation of hydraulic engineering projects includes both recoverable responses caused by water level fluctuations and operational changes, as well as residual deformations formed by long-term structural changes. Existing methods struggle to effectively distinguish between recoverable responses and residual deformations, leading to the misjudgment of recoverable deformations during normal operation as abnormalities, and the easy masking of early residual deformations by normal operating conditions. This results in poor accuracy and stability of deformation monitoring results for hydraulic engineering projects. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for monitoring deformation of hydraulic engineering based on deep learning.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] On the one hand, this invention provides a deep learning-based method for monitoring deformation in hydraulic engineering projects, comprising:

[0007] Collect deformation monitoring datasets for water conservancy projects, organize the monitoring data into a benchmark, and obtain the total deformation sequence of measuring points, the structural attribution relationship of measuring points, the working condition characteristic sequence, and the working condition stage sequence.

[0008] Based on the total deformation sequence of the measuring points, the working condition characteristic sequence, and the working condition stage sequence, a return closed separation network is constructed to obtain the recoverable working condition response deformation sequence of the measuring points and the initial residual deformation sequence of the measuring points.

[0009] Based on the total deformation sequence of measuring points, the recoverable working condition response deformation sequence of measuring points, the initial residual deformation sequence of measuring points, and the structural attribution relationship of measuring points, a partition boundary-preserving graph neural inference network is constructed to obtain the deformation monitoring results of water conservancy projects.

[0010] Specifically, a deformation monitoring dataset of water conservancy projects is collected, and the monitoring data is standardized and organized to obtain the total deformation sequence of monitoring points, the structural attribution relationship of monitoring points, the working condition characteristic sequence, and the working condition stage sequence, including:

[0011] Collect data sets for deformation monitoring of water conservancy projects;

[0012] Based on the displacement observation data of the measuring points, the displacement observation values ​​of each monitoring point at each monitoring time are determined, and the displacement observation values ​​are processed by benchmark difference to obtain the total deformation sequence of the measuring points.

[0013] Based on the data of the measurement point layout and the structural zoning data, the structural zoning corresponding to each monitoring point is determined, and the structural zoning is mapped to the total shape sequence of the measurement points to obtain the structural affiliation relationship of the measurement points;

[0014] Based on water level data and scheduling operation records, the working condition data corresponding to each monitoring time in the total deformation sequence of the measuring points are extracted, and the water level change and scheduling status are generated to obtain the working condition feature sequence.

[0015] Based on the water level change and scheduling status, the direction of water level change and scheduling operation status corresponding to each monitoring moment are determined, and stage markings are performed to obtain the working condition stage sequence.

[0016] Specifically, based on the total deformation sequence of the measuring points, the operating condition characteristic sequence, and the operating condition stage sequence, a return closed-loop separation network is constructed to obtain the recoverable operating condition response deformation sequence and the initial residual deformation sequence of the measuring points, including:

[0017] Based on the total deformation sequence of the measuring points and the working condition stage sequence, the total deformation of each monitoring point in the water level loading segment, water level unloading segment and water level recovery segment is segmented to obtain a set of deformation segments.

[0018] Based on the set of deformation fragments and the sequence of working conditions, a set of deformation fragment attributes is generated, and a water level return deformation map is constructed.

[0019] Based on the water level return deformation map and the deformation segment attribute set, the return transmission cost is constructed, and the return closure matching and partial optimal transmission solution are performed to obtain the return matching deformation segment and the return unclosed deformation difference.

[0020] Based on the back-travel matching deformation segment and the back-travel unclosed deformation difference, a back-travel closed separation network is constructed to establish a recoverable response state and a residual cumulative state, and then the separation and update are performed to obtain the recoverable working condition response deformation sequence and the initial residual deformation sequence of the measuring point.

[0021] Specifically, based on the set of deformation fragments and the sequence of working conditions, a set of deformation fragment attributes is generated, and a water level return deformation map is constructed, including:

[0022] Based on the set of deformation segments, determine the start and end times of each deformation segment and the working condition stage to which the segment belongs;

[0023] Based on the working condition feature sequence, the average water level of each deformation segment, the direction of water level change, and the direction of deformation change are determined to obtain the attribute set of the deformation segment.

[0024] Based on the attribute set of deformable fragments, each deformable fragment is mapped to a deformable fragment node, and each deformable fragment node is paired to obtain a set of candidate return node pairs;

[0025] Based on the deformation segment attribute set, the candidate return node pair set is judged by water level proximity, working condition stage difference, and stage sequence to obtain the effective return node pair set;

[0026] Based on the set of effective return node pairs, return edges are generated, and combined with deformation fragment nodes, a water level return deformation map is constructed.

[0027] Specifically, based on the water level return deformation map and the attribute set of deformation segments, the return transmission cost is constructed, and return closure matching and partial optimal transmission solutions are performed to obtain the return matching deformation segment and the return unclosed deformation difference, including:

[0028] Based on the water level return deformation map and the deformation segment attribute set, the deformation segment node pairs connected by the return edge are extracted, and the deformation segment node pairs are divided into source end deformation segments and target end deformation segments to obtain the return candidate transmission pair set;

[0029] Based on the backhaul candidate transmission pair set, water level backhaul comparison, deformation recovery direction comparison, and stage sequence comparison are performed to obtain the candidate transmission attribute set and generate the backhaul transmission cost corresponding to each backhaul candidate transmission pair.

[0030] Based on the backhaul transmission cost, backhaul closure matching and partial optimal transmission solution are performed on the backhaul candidate transmission pair set to obtain the partial optimal transmission matching result.

[0031] Based on the partial optimal transmission matching results, the return closed matching segment and the target end unmatched deformation are extracted, and the target end unmatched deformation is subjected to water level recovery segment time sequence expansion to obtain the return matching deformation segment and the return unclosed deformation difference.

[0032] Specifically, based on the backhaul transmission cost, backhaul closure matching and partial optimal transmission solutions are performed on the backhaul candidate transmission pair set to obtain partial optimal transmission matching results, including:

[0033] Based on the backhaul candidate transmission pair set, determine the transmission quality corresponding to each source end deformation segment and each target end deformation segment;

[0034] Based on the backhaul transmission cost, backhaul closure matching is performed on the backhaul candidate transmission pair set to determine the allowed transmission pairs and obtain the allowed transmission pair set.

[0035] Based on transmission quality and allowed transmission pairs, the transmission range of the source-end deformed segment is limited, while the unmatched transmission quality of the target-end deformed segment is preserved, thus constructing partial transmission constraints.

[0036] Based on partial transmission constraints, the minimum transmission cost is solved for the set of allowed transmission pairs to obtain partially optimal transmission matching results.

[0037] Specifically, based on the partially optimal transmission matching results, the return closed matching segment and the target end unmatched deformation are extracted, and the target end unmatched deformation is subjected to time-series expansion of the water level recovery segment to obtain the return matched deformation segment and the return unclosed deformation difference, including:

[0038] Based on the partial optimal transmission matching results, the deformation segment pairs that have completed the return closed matching are extracted, and the unmatched deformation variables retained in the target end deformation segments are extracted to obtain the set of closed matching segment pairs and the set of unmatched retained deformation variables;

[0039] Based on the set of closed matching segments, the source-end deformation segment and the target-end deformation segment in each closed matching segment pair are determined as the return matching deformation segment;

[0040] Based on the set of unmatched retained deformation variables, and combined with the direction of change of the fragment deformation variables corresponding to the deformation fragment at the target end, the unclosed candidate deformation variables corresponding to the deformation fragment at the target end are determined.

[0041] Based on the unclosed candidate deformation, the unclosed deformation corresponding to the water level recovery segment is determined, and it is expanded according to the corresponding monitoring time to obtain the return unclosed deformation difference value.

[0042] Specifically, based on the return matching deformation segment and the return unclosed deformation difference, a return closure separation network is constructed to establish the recoverable response state and the residual cumulative state, and then separated and updated to obtain the recoverable working condition response deformation sequence and the initial residual deformation sequence of the measuring point, including:

[0043] Based on the retrace matching deformation segment and the retrace unclosed deformation difference, the retrace matching response and unclosed residual input corresponding to each monitoring time are generated.

[0044] Based on the operating condition characteristic sequence, the back-run matched response quantity and the unclosed residual input quantity, a back-run closed separation network is constructed, and a recoverable response state and a residual cumulative state are established in the back-run closed separation network.

[0045] Based on the unclosed residual input quantity and the operating condition characteristic sequence, the return closure gating quantity is generated, the return matching response quantity is allocated to the recoverable response state, and the unclosed residual input quantity is allocated to the residual accumulation state, thus obtaining the recoverable response state sequence and the residual accumulation state sequence.

[0046] Based on the recoverable response state sequence, generate the recoverable working condition response deformation sequence of the measuring points;

[0047] Based on the residual cumulative state sequence, the initial residual deformation sequence of the measuring point is generated.

[0048] Specifically, based on the total deformation sequence of measuring points, the recoverable working condition response deformation sequence of measuring points, the initial residual deformation sequence of measuring points, and the structural attribution relationship of measuring points, a partition boundary-preserving graph neural inference network is constructed to obtain the deformation monitoring results of hydraulic engineering projects, including:

[0049] Based on the total deformation sequence of the measuring points, the recoverable working condition response deformation sequence of the measuring points, the initial residual deformation sequence of the measuring points, and the structural attribution relationship of the measuring points, a feature set of measuring point nodes with structural partitioning labels is generated.

[0050] Based on the feature set of measurement point nodes, measurement point nodes within the same structural partition are connected as aggregated edges within the same partition, and measurement point nodes on both sides of the boundary of adjacent structural partitions are connected as boundary preservation edges, thus obtaining a partition boundary preservation deformation map.

[0051] Based on the partition boundary preservation deformation map, a partition boundary preservation graph neural inference network is constructed to perform same-region aggregation and boundary difference preservation on the initial residual deformation sequence of the measurement points, so as to obtain the same-region aggregated residual deformation and the boundary difference residual deformation.

[0052] Based on the residual deformation variables of the same region and the residual deformation variables of the boundary difference, the initial residual deformation variable sequence of the measuring points is corrected by structural partitioning to obtain the residual deformation variables of the structural body.

[0053] Based on the total deformation sequence of the measuring points, the recoverable working condition response deformation sequence of the measuring points, and the residual deformation of the structural body, deformation monitoring results of hydraulic engineering are generated.

[0054] On the other hand, the present invention provides a deep learning-based hydraulic engineering deformation monitoring system, comprising:

[0055] The data acquisition module is used to collect deformation monitoring datasets of water conservancy projects, organize the monitoring data into a benchmark, and obtain the total deformation sequence of measuring points, the structural attribution relationship of measuring points, the working condition characteristic sequence, and the working condition stage sequence.

[0056] The separation network module constructs a return closed separation network based on the total deformation sequence of the measuring points, the working condition characteristic sequence, and the working condition stage sequence, to obtain the recoverable working condition response deformation sequence of the measuring points and the initial residual deformation sequence of the measuring points;

[0057] The deformation monitoring module constructs a partition boundary-preserving graph neural inference network based on the total deformation sequence of the measuring points, the recoverable working condition response deformation sequence of the measuring points, the initial residual deformation sequence of the measuring points, and the structural attribution relationship of the measuring points, to obtain the deformation monitoring results of the hydraulic engineering project.

[0058] Compared with existing technologies, the beneficial effects of this invention include: by constructing a water level return deformation map and performing return closure matching and partial optimal transmission solutions, deformation segments during water level rise, water level fall, and water level recovery are mapped, while retaining deformations that are not matched after water level recovery. This prevents deformations that cannot recover with the water level return from being incorporated into the normal operating condition response, reduces the masking of early residual deformations, and improves the reliability of residual deformation extraction. By constructing a return closure separation network, the returned matched deformation segments are used as the basis for the recoverable response state, and the returned unclosed deformation difference is used as the basis for the residual accumulation state. This separates the total deformation at the measuring point into recoverable operating condition response deformations and initial residual deformations, reducing misjudgments caused by judging solely based on the total deformation, improving the accuracy of water conservancy project deformation monitoring results, and enhancing the stability of water conservancy project deformation monitoring results. Attached Figure Description

[0059] Figure 1 The flowchart of the deep learning-based deformation monitoring method for water conservancy projects provided by this invention is shown below.

[0060] Figure 2 This is a schematic diagram of water level return loop closure matching and unclosed deformation extraction provided by the present invention;

[0061] Figure 3 The structural diagram of the deep learning-based hydraulic engineering deformation monitoring system provided by this invention is shown. Detailed Implementation

[0062] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0063] Example 1:

[0064] See Figures 1-2 This embodiment provides a deep learning-based method for monitoring deformation in hydraulic engineering, including the following specific steps:

[0065] Step S1: Collect the deformation monitoring dataset of the water conservancy project, organize the monitoring data into a benchmark, and obtain the total deformation sequence of the measuring points, the structural attribution relationship of the measuring points, the working condition characteristic sequence, and the working condition stage sequence.

[0066] The specific steps of step S1 are as follows:

[0067] Step S101: Collect data sets for deformation monitoring of water conservancy projects.

[0068] In this embodiment, a deformation monitoring dataset of a water conservancy project is collected. The data includes displacement observation data of measuring points, water level data, temperature data, rainfall data, seepage data, scheduling and operation records, measuring point layout data, and structural zoning data.

[0069] Displacement observation data is acquired through automated displacement monitoring equipment deployed on the surfaces of reservoir dams, dikes, sluice gates, slopes, or canal slopes. Each monitoring point is assigned a unique numerical number. When the monitoring equipment directly outputs the deformation displacement, it directly records the displacement observation value of each monitoring point at each original sampling time. When the monitoring equipment outputs three-dimensional spatial coordinates, the principal deformation direction vector of each monitoring point is first determined. The first full hour after the completion and acceptance of the water conservancy project and the stable operation of water storage is selected as the reference time. The coordinate difference between the coordinates of each subsequent original sampling time and the coordinates of the reference time is calculated, and then the coordinate difference is projected onto the principal deformation direction vector to obtain the displacement observation value at the corresponding time.

[0070] Water level data is collected synchronously through radar water level gauges and pressure water level monitoring stations deployed in the reservoir area, along the river, and upstream and downstream of sluice gates.

[0071] Temperature data is divided into ambient temperature and dam internal temperature. Ambient temperature is collected by temperature and humidity sensors deployed in unobstructed areas around the project, while dam internal temperature is collected by thermal temperature sensors buried in layers at different depths and locations within the dam.

[0072] Rainfall data is collected in time intervals using distributed rain gauges within the watershed. Seepage data includes dam seepage pressure, dam foundation phreatic line height, downstream seepage flow, and dam pore water pressure, which are collected in real time by piezometers, phreatic line monitors, electromagnetic flowmeters, and pore water pressure sensors, respectively.

[0073] The scheduling and operation records are directly extracted from the central scheduling and monitoring system of water conservancy projects. They cover four types of operational information: gate opening and closing status, flood discharge tunnel opening and closing status, reservoir water storage and regulation status, and ecological water release status. Discrete operational statuses are uniformly converted into binary numerical features, with equipment on-state operation assigned a value of 1 and equipment off-state operation assigned a value of 0.

[0074] The monitoring point layout data includes the plane spatial coordinates of each monitoring point, the monitoring type of the monitoring point, and the dam section or embankment section number to which the monitoring point belongs. It is directly extracted from the completed monitoring point layout archives of the water conservancy project. The plane spatial coordinates of the monitoring points include the plane horizontal coordinate, the plane vertical coordinate, and the elevation value.

[0075] The structural zoning data includes the boundaries of dam sections, the independent zoning range of gate piers, the boundaries of dike sections, the functional zoning range of slope protection, the spatial location of structural expansion joints and joints, and the topological relationships between adjacent zoning areas. It is obtained by vectorization and analysis based on the construction design drawings of hydraulic engineering structures.

[0076] Extract the original sampling periods of various sensors such as displacement, water level, temperature, and seepage, and find the least common divisor of all sampling periods to obtain a unified monitoring time interval.

[0077] A unified monitoring time sequence consisting of consecutive hourly times is constructed. The start time of the sequence is selected as the start time of the stable water storage operation of the project, and the end time is selected as the end time of the continuous monitoring cycle.

[0078] For example, the original sampling period for the displacement sensor is 1 hour, the original sampling period for the water level sensor is 10 minutes, the original sampling period for the temperature sensor is 30 minutes, and the original sampling period for the seepage sensor is 1 hour. The least common divisor of all sampling periods is 1 hour, and the unified monitoring time interval is 1 hour.

[0079] Step S102: Based on the displacement observation data of the measuring points, determine the displacement observation value corresponding to each monitoring point at each monitoring time, and perform benchmark difference processing on the displacement observation value to obtain the total deformation sequence of the measuring points.

[0080] In this embodiment, the collected original displacement observation data of the measuring points are subjected to time-series interpolation and normalization, and all time points of the unified monitoring time sequence are matched to obtain the original displacement observation value of each monitoring point at each unified monitoring time.

[0081] Historical water level time series data of the project during 7 consecutive days of operation without scheduling were selected. The water level change at adjacent times was calculated to obtain a water level change sample sequence. The arithmetic mean and standard deviation of the water level change sample sequence were calculated. According to the 3σ anomaly discrimination criterion, the arithmetic mean and three times the standard deviation of the sample were summed to obtain the stable water level fluctuation threshold.

[0082] The first hour within a seven-day period after the completion and acceptance of a water conservancy project, during which the water level fluctuation is less than the stable water level fluctuation threshold, is selected as the benchmark time.

[0083] All original displacement observations at each monitoring point are subjected to benchmark difference processing. The displacement observations at each time point are subtracted from the displacement observations at the benchmark time point to obtain the deformation at each time point relative to the benchmark time point.

[0084] All deformation variables at each monitoring point are arranged sequentially according to the time order of a unified monitoring time sequence to obtain the total deformation variable sequence of the monitoring points.

[0085] For example, water level data during a stable period of 7 consecutive days of the project were selected, and the mean of the water level change sample sequence was calculated to be 0.001 meters per hour, the standard deviation was 0.003 meters per hour, and the stable water level fluctuation threshold was 0.01 meters per hour.

[0086] Time 0 was selected as the reference time, at which the original displacement observation value of monitoring point No. 1 was 12.5 mm.

[0087] The original displacement observation value of the measuring point at time 1 was 13.7 mm, and at time 2 it was 15.0 mm. After benchmark difference processing, the deformation values ​​at the corresponding times were 0 mm, 1.2 mm, and 2.5 mm, respectively. Arranged in chronological order, the total deformation sequence of the first three times of measuring point 1 was obtained.

[0088] Step S103: Based on the measurement point layout data and structural zoning data, determine the structural zoning corresponding to each monitoring measurement point, and match the structural zoning with the total deformation sequence of the measurement points to obtain the structural affiliation relationship of the measurement points.

[0089] In this embodiment, the planar spatial coordinates of each monitoring point are extracted from the measurement point layout data, and the closed polygon boundary coordinates of each structural partition are extracted from the structural partition data.

[0090] The ray method is used to determine the classification of measurement points into zones. An infinitely long ray is drawn from the plane coordinates of a certain monitoring point to the positive horizontal direction. The number of intersections between the ray and each side of the closed polygon of the corresponding structural zone is counted. If the number of intersections is odd, the measurement point is determined to be inside the structural zone. If the number of intersections is even, the measurement point is determined to be outside the structural zone.

[0091] Each monitoring point is assigned to a specific zone, and a unique structural zone number is determined for each monitoring point.

[0092] The structural partition number of each monitoring point is associated with the total deformation sequence of the corresponding monitoring point, and all associated information is organized in ascending order of the monitoring point number to obtain the structural affiliation relationship of the monitoring points.

[0093] For example, the plane coordinates of monitoring point No. 1 are (523682.5, 3412769.2), and the coordinates of the closed polygon boundary of the upstream dam section No. 1 are (523650.0, 3412740.0), (523720.0, 3412740.0), (523720.0, 3412790.0), (523650.0, 3412790.0).

[0094] A ray is drawn from the coordinates of the measuring point in the positive horizontal direction, intersecting with the right boundary of the No. 1 upstream dam section at (523720.0, 3412769.2). The number of intersection points is 1, and it is determined that the measuring point belongs to the No. 1 upstream dam section.

[0095] Associate the upstream dam section number with the total shape sequence of the first measuring point, and complete the association of all measuring points in the same way to obtain the structural affiliation of the measuring points.

[0096] Step S104: Based on water level data and scheduling operation records, extract the operating condition data corresponding to each monitoring time in the total deformation sequence of the measuring points, and generate water level change and scheduling status to obtain the operating condition feature sequence.

[0097] In this embodiment, 1 / 6 to 1 / 2 of the uniform monitoring time interval is taken as the single-sided time intercept width. By intercepting symmetrically from left to right, a preset time range is formed, and the water level data, temperature data, rainfall data, seepage data and scheduling operation records are time-aligned.

[0098] Based on the unified monitoring time sequence, the water level values ​​corresponding to each monitoring time are extracted from the time-aligned water level data, and the gate opening and closing status, flood discharge tunnel opening and closing status, water storage status, and water release status corresponding to each monitoring time are extracted from the scheduling operation records.

[0099] The difference between the current water level and the previous water level is calculated at each time step to obtain the water level change.

[0100] The extracted scheduling operation statuses are converted into binary numerical features. The device is assigned a value of 1 when it is running and a value of 0 when it is shut down or stationary, thus obtaining the scheduling status at each time point.

[0101] The ambient temperature and the internal temperature of the dam body at each moment are extracted from the time-aligned temperature data. The difference between the current temperature value and the previous temperature value is calculated moment by moment to obtain the temperature change.

[0102] The cumulative rainfall at each moment is extracted from the time-aligned rainfall data, and the seepage pressure at each moment is extracted from the time-aligned seepage data. The difference between the seepage pressure at the current moment and the seepage pressure at the previous moment is calculated moment by moment to obtain the seepage pressure change.

[0103] The water level, water level change, ambient temperature, temperature change, cumulative rainfall, seepage pressure, seepage pressure change, and scheduling status at each time point are combined into a single-time condition feature vector. The condition feature vectors of all times are then arranged in chronological order according to a unified monitoring time sequence to obtain the condition feature sequence.

[0104] For example, the uniform monitoring time interval is 1 hour, the single-sided time truncation width is 10 minutes, and the preset time range is 10 minutes before and after.

[0105] Cross-correlation time delay analysis revealed that the maximum lag time delay of rainfall affecting dam deformation was 2 hours, and the cumulative rainfall statistical duration was taken as 2 hours.

[0106] The water level at time 1 is 120.52 meters, and the water level at time 0 is 120.49 meters. The calculated water level change is 0.03 meters.

[0107] At this moment, the gate opening status is 1, the spillway status is 0, the water storage status is 1, the water release status is 0, and the scheduling status combination is [1, 0, 1, 0].

[0108] At this moment, the ambient temperature is 25.2 degrees Celsius, the temperature change is 0.5 degrees Celsius, the cumulative rainfall is 0 mm, the seepage pressure is 0.21 MPa, and the seepage pressure change is 0.01 MPa. Combining these values, the characteristic vector of the working condition at this moment is [120.52, 0.03, 25.2, 0.5, 0, 0.21, 0.01, 1, 0, 1, 0]. The characteristic vectors of all moments are generated and arranged in the same way to obtain the working condition characteristic sequence.

[0109] Step S105: Based on the water level change and scheduling status, determine the direction of water level change and scheduling operation status corresponding to each monitoring time, and mark the stages to obtain the working condition stage sequence.

[0110] In this embodiment, the direction of water level change is determined based on the amount of water level change at each moment. When the amount of water level change is greater than 0, the direction of water level change is determined to be rising; when the amount of water level change is less than 0, the direction of water level change is determined to be falling; and when the amount of water level change is equal to 0, the direction of water level change is determined to be stable.

[0111] Based on the scheduling status at each moment, the scheduling operation status is determined. When the status of any scheduling device is 1, the scheduling operation status is determined to be running. When the status of all scheduling devices is 0, the scheduling operation status is determined to be stationary.

[0112] Historical water level time series data were selected from the period of 30 consecutive days of stable and uninterrupted operation of the project without scheduling disturbances. The water level change was calculated moment by moment between adjacent moments to obtain a water level change sample sequence. The arithmetic mean and sample standard deviation of the sample sequence were calculated. According to the 3σ anomaly discrimination criterion, the arithmetic mean and three times the sample standard deviation were summed to obtain the water level change threshold.

[0113] The minimum number of consecutive water level changes exceeding the water level change threshold within this stable period is counted, and this number is set as the number of consecutive monitoring moments.

[0114] When there are no fewer than a certain number of consecutive monitoring times and the water level changes in the direction of rising at each time point, the continuous period is marked as the water level loading segment.

[0115] When there are no fewer than a certain number of consecutive monitoring times and the water level changes in the direction of decrease at each time point, the continuous period is marked as the water level unloading segment.

[0116] After the water level unloading process is completed, when there are no fewer than a number of consecutive monitoring times and the direction of water level change is stable at each time, and the scheduling operation status is static at each time, the continuous period is marked as the water level recovery period.

[0117] All other time periods are uniformly marked as stable periods. The operating condition stages of all monitoring times are arranged sequentially according to the time order of the unified monitoring time sequence to obtain the operating condition stage sequence.

[0118] For example, water level data during a stable period of 30 consecutive days of the project were selected, and the mean of the water level change sample sequence was calculated to be 0.001 meters per hour, the standard deviation was 0.003 meters per hour, and the water level change threshold was 0.01 meters per hour.

[0119] The minimum number of consecutive anomalies was found to be 3, and the number of consecutive monitoring times was 3.

[0120] The water level changes were greater than 0.01 meters per hour for 13 consecutive time periods from time 0 to time 12, and the direction of water level change was upward. This was marked as the water level loading segment.

[0121] For 15 consecutive time periods from time 24 to time 38, the water level change was less than -0.01 meters per hour, and the direction of water level change was downward. This period was marked as the water level unloading segment.

[0122] The absolute value of the water level change for nine consecutive time periods from time 39 to time 47 does not exceed 0.01 meters per hour, the direction of water level change is stable, and the scheduling operation status is static at all times. This is marked as the water level recovery period. The remaining time periods are marked as the stable period. Arranging all the marks in chronological order yields the working condition stage sequence.

[0123] Step S2: Based on the total deformation sequence of the measuring points, the working condition characteristic sequence, and the working condition stage sequence, construct a return closed separation network to obtain the recoverable working condition response deformation sequence of the measuring points and the initial residual deformation sequence of the measuring points.

[0124] The specific steps of step S2 are as follows:

[0125] Step S201: Based on the total deformation sequence of the measuring points and the working condition stage sequence, the total deformation of each monitoring measuring point in the water level loading segment, water level unloading segment and water level recovery segment is segmented to obtain a deformation segment set.

[0126] In this embodiment, the deformation segment length is set to cover the complete deformation response cycle caused by a single water level scheduling of a water conservancy project, so as to avoid the deformation trend being lost due to excessively short segments and the mixing of multiple working conditions due to excessively long segments. The deformation segment length is set to 3 to 12 times the uniform monitoring time interval.

[0127] The total deformation sequence of each monitoring point is continuously extracted using a sliding window method, with a sliding step size of one uniform monitoring time interval.

[0128] For each obtained deformation segment, extract the water level segment and working condition feature segment within the corresponding time period, and count the working condition stage that appears most frequently in the segment as the working condition stage to which the segment belongs.

[0129] All the extracted deformation segments, corresponding water level segments, working condition characteristic segments, and their respective working condition stages are integrated to obtain a set of deformation segments.

[0130] For example, the uniform monitoring time interval is 1 hour, the deformation segment length is 6, and the sliding step size is 1 hour.

[0131] The total deformation sequence of measuring point 1 is truncated by sliding, and the first deformation segment covers time 0 to time 5. The working condition stage that appears most frequently in this segment is the water level loading segment. Therefore, the working condition stage to which this segment belongs is the water level loading segment.

[0132] Step S202: Based on the deformation fragment set and the working condition feature sequence, generate the deformation fragment attribute set and construct the water level return deformation map.

[0133] The specific steps of step S202 are as follows:

[0134] Step S2021: Based on the set of deformation segments, determine the start and end times of each deformation segment and the working condition stage to which the segment belongs.

[0135] In this embodiment, for each deformation segment in the deformation segment set, the start and end monitoring times corresponding to its sliding interception are extracted to obtain the segment start and end times of each deformation segment.

[0136] Extract the working condition stage to which each deformation segment belongs, and complete the extraction of the basic attributes of all deformation segments.

[0137] Step S2022: Based on the working condition feature sequence, determine the average water level of each deformation segment, the direction of water level change, and the direction of deformation change, and obtain the attribute set of the deformation segment.

[0138] In this embodiment, for each deformation segment, the arithmetic mean of the water level values ​​at all times within the corresponding water level segment is calculated to obtain the segment average water level.

[0139] Calculate the difference between the water level at the end of the segment and the water level at the beginning of the segment to obtain the water level change of the segment. If the water level change of the segment is greater than 0, the water level change direction of the segment is determined to be rising; if it is less than 0, it is determined to be falling; and if it is equal to 0, it is determined to be stable.

[0140] Calculate the difference between the deformation at the end of the segment and the deformation at the beginning of the segment to obtain the change in the deformation of the segment. If the change in the deformation of the segment is greater than 0, the direction of change of the deformation of the segment is determined to be increasing; if it is less than 0, it is determined to be decreasing; and if it is equal to 0, it is determined to be stationary.

[0141] The start and end times of each deformation segment, the working condition stage to which it belongs, the average water level of the segment, the direction of water level change of the segment, and the direction of deformation change of the segment are integrated to obtain the deformation segment attribute set.

[0142] Step S2023: Based on the deformable fragment attribute set, map each deformable fragment to a deformable fragment node, and pair each deformable fragment node to obtain a candidate backhaul node pair set.

[0143] In this embodiment, each deformation segment is mapped to an independent node in the water level return deformation diagram, and the node number is consistent with the number of the corresponding deformation segment.

[0144] All deformed segment nodes are paired up to generate all possible node pairs, resulting in a candidate backhaul node pair set.

[0145] Step S2024: Based on the deformation fragment attribute set, perform water level proximity discrimination, working condition stage difference discrimination, and stage sequence discrimination on the candidate return node pair set to obtain the effective return node pair set.

[0146] In this embodiment, a water level approach threshold is set, a dataset of water level changes during a single water level scheduling process in the project's history is selected, the arithmetic mean of all single scheduling water level changes is calculated, and 5% to 20% of this mean is taken as the water level approach threshold.

[0147] For each pair of nodes in the candidate backhaul node pair set, calculate the absolute value of the difference between the average water levels of the corresponding segments of the two nodes. If the absolute value is less than or equal to the water level proximity threshold, then the water level proximity is determined.

[0148] Determine whether the two nodes belong to different operating conditions. If they do, then use the difference in operating conditions to make the distinction.

[0149] Compare the start times of the segments corresponding to the two nodes. If the start time of one node is earlier than the start time of the other node, then the stage sequence is determined.

[0150] Node pairs that pass all three criteria are considered valid backhaul node pairs. All valid backhaul node pairs are then integrated to obtain a set of valid backhaul node pairs.

[0151] For example, a dataset of water level changes from 100 single water level scheduling processes in the project's history was selected, and the average water level change from a single scheduling process was calculated to be 1.0m. 10% of the average value was taken as the water level approach threshold, i.e., 0.10m.

[0152] The average water level of the segment corresponding to the first node of measuring point 1 is 120.50m, belonging to the water level loading segment, with a start time of time 0. The average water level of the segment corresponding to the tenth node is 120.45m, belonging to the water level unloading segment, with a start time of time 9. The absolute value of the difference in average water levels between the two nodes is 0.05m, which is less than or equal to 0.10m, and is determined by the proximity of the water levels. The operating segments are the loading segment and the unloading segment, which are different, and are determined by the difference in operating segments. The start time of the first node is earlier than that of the tenth node, and is determined by the sequence of stages. Therefore, this node pair is a valid return node pair.

[0153] Step S2025: Based on the set of valid return node pairs, generate return edges and combine them with deformation fragment nodes to construct a water level return deformation map.

[0154] In this embodiment, for each pair of nodes in the effective backhaul node pair set, an undirected backhaul edge is generated between the two nodes. The attributes of the backhaul edge include the numbers of the two nodes and the corresponding average water level difference.

[0155] By integrating all deformation segment nodes and all return edges, a water level return deformation map corresponding to each monitoring point is constructed. The water level return deformation map uses nodes to represent deformation segments and edges to represent the water level return correlation between two deformation segments.

[0156] For example, there are 20 deformation segment nodes at measuring point 1. Among them, 8 pairs of nodes passed the three criteria and are considered valid return node pairs. Therefore, 8 undirected return edges are generated between these 8 pairs of nodes. By integrating the 20 nodes and 8 edges, the water level return deformation diagram corresponding to measuring point 1 is obtained.

[0157] like Figure 2 As shown, step S203: Based on the water level return deformation map and the deformation segment attribute set, construct the return transmission cost, and perform return closure matching and partial optimal transmission solution to obtain the return matching deformation segment and the return unclosed deformation difference value.

[0158] The specific steps of step S203 are as follows:

[0159] Step S2031: Based on the water level return deformation map and the deformation segment attribute set, extract the deformation segment node pairs connected by the return edge, and divide the deformation segment node pairs into source end deformation segments and target end deformation segments to obtain the return candidate transmission pair set.

[0160] In this embodiment, all pairs of deformed segment nodes connected by the return edge are extracted from the water level return deformation map.

[0161] For each pair of nodes, the nodes belonging to the water level loading stage are divided into source-end deformation segments, and the nodes belonging to the water level unloading stage or water level recovery stage are divided into target-end deformation segments.

[0162] All the source and destination node pairs that have been divided are integrated to obtain a set of candidate backhaul transmission pairs.

[0163] For example, eight pairs of nodes connected by the return edge are extracted from the water level return deformation diagram of measuring point 1.

[0164] The first pair of nodes includes the first node and the tenth node. The working condition stage to which the first node belongs is the water level loading segment, and the working condition stage to which the tenth node belongs is the water level unloading segment. Therefore, the first node is divided into the source end deformation segment, and the tenth node is divided into the target end deformation segment.

[0165] The same method was used to divide all 8 pairs of nodes, resulting in a set of candidate backhaul transmission pairs containing 8 source-end and destination-end node pairs.

[0166] Step S2032: Based on the backhaul candidate transmission pair set, perform water level backhaul comparison, deformation recovery direction comparison, and stage sequence comparison to obtain the candidate transmission attribute set and generate the backhaul transmission cost corresponding to each backhaul candidate transmission pair.

[0167] In this embodiment, for each backhaul candidate transmission pair, the normalized water level backhaul distance is calculated. The water level backhaul distance is calculated as follows: first, the absolute value of the average water level difference between the source end and the target end segment is calculated, and then the absolute value is divided by the difference between the maximum water level and the minimum water level in the current monitoring period. To avoid the denominator being zero, a minimum value is added to the denominator.

[0168] To calculate the shape distance of the standardized deformation segments, first, the deformation segments at the source and target ends are standardized with zero mean, and then the average of the absolute values ​​of the time differences between the two standardized segments at corresponding times is calculated to obtain the shape distance.

[0169] Calculate the cost of deformation recovery direction. If the water level changes in the source end and the target end are in opposite directions and the deformation changes in opposite directions, the cost is set to the first preset value. If the water level changes in opposite directions but the deformation changes in zero direction, the cost is set to the second preset value. In other cases, the cost is set to the third preset value.

[0170] The cost is calculated based on the order of the working conditions. If the source end is the loading segment and the target end is the unloading segment or the recovery segment, and the timing of the source end is earlier than that of the target end, then the cost is taken as the first preset value. If the source end is the unloading segment and the target end is the recovery segment, and the timing of the source end is earlier than that of the target end, then the cost is taken as the second preset value. In other cases, the cost is taken as the third preset value.

[0171] The Analytic Hierarchy Process (AHP) is used to determine the weights of water level return distance, shape distance, deformation recovery direction cost, and operating condition stage sequence cost. The sum of the four weights is 1. The four costs are multiplied by their respective weights and then summed to obtain the return transmission cost corresponding to the candidate return transmission pair.

[0172] The transmission costs of all candidate transmission pairs are integrated to obtain a candidate transmission attribute set.

[0173] The first, second, and third preset values ​​are set according to the physical laws of deformation recovery in water conservancy projects. The first preset value is 0, which corresponds to the ideal return recovery state where the water level and deformation change directions are completely opposite, and the transmission cost is the lowest. The second preset value is 0.5, which corresponds to the partial recovery state where the water level change direction is opposite but the deformation does not change significantly, and the transmission cost is moderate. The third preset value is 1, which corresponds to the non-recovery state where the water level or deformation change direction is the same, and the transmission cost is the highest.

[0174] For example, the first preset value is 0, the second preset value is 0.5, and the third preset value is 1.

[0175] The weights calculated using the analytic hierarchy process (AHP) are: water level return distance weight 0.35, shape distance weight 0.25, deformation recovery direction cost weight 0.25, and working condition stage priority cost weight 0.15.

[0176] The maximum water level during the current monitoring period is 125.00m, and the minimum water level is 115.00m.

[0177] The average water level of the segment at the first source node of measuring point 1 is 120.50m, and the average water level of the segment at the tenth target node is 120.45m. The absolute value of the difference between the two is 0.05m. The calculated water level return distance is 0.05 divided by 10.00 plus 10 to the power of negative 6, which gives a result of 0.005.

[0178] The average absolute value of the time difference between the standardized deformation segments of the two nodes is 0.12, that is, the shape distance is 0.12.

[0179] The water level at the source end changes in the direction of rising, and the deformation changes in the direction of increasing; the water level at the target end changes in the direction of falling, and the deformation changes in the direction of decreasing, therefore the deformation recovery direction cost is 0.

[0180] The source starts at time 0 and the target starts at time 9. The source timing is earlier than the target timing, so the cost of the different phases is 0.

[0181] The final calculated return transmission cost for this transmission pair is 0.35 multiplied by 0.005 plus 0.25 multiplied by 0.12, resulting in 0.03175.

[0182] Step S2033: Based on the backhaul transmission cost, perform backhaul closure matching and partial optimal transmission solution on the backhaul candidate transmission pair set to obtain the partial optimal transmission matching result, which includes the matched transmission quality and the unmatched transmission quality.

[0183] The specific steps of step S2033 are as follows:

[0184] Step S20331: Based on the backhaul candidate transmission pair set, determine the transmission quality corresponding to each source end deformation segment and each target end deformation segment.

[0185] In this embodiment, for each source-end deformation segment, the absolute value of the difference between the deformation at the end time and the deformation at the beginning time of the segment is calculated as the transmission quality of the source-end node.

[0186] For each target end deformation segment, calculate the absolute value of the difference between the deformation at the end time and the deformation at the beginning time of the segment, and use it as the transmission quality of the target end node.

[0187] Set a minimum transmission quality, which is one percent of the transmission quality of all nodes. If the transmission quality of a node is less than the minimum transmission quality, then set its transmission quality to the minimum transmission quality to avoid unstable matching due to excessively low transmission quality.

[0188] For example, one percent of the transmission quality of all nodes is 0.01mm.

[0189] The deformation at the start time of the first source node at measurement point 1 is 0 mm, and the deformation at the end time is 3.0 mm. The absolute value of the difference is 3.0 mm. Therefore, the transmission quality of this source node is 3.0 mm.

[0190] The deformation at the start time of the 10th target node is 3.0 mm, and the deformation at the end time is 0.3 mm. The absolute value of the difference is 2.7 mm. Therefore, the transmission quality of this target node is 2.7 mm.

[0191] Step S20332: Based on the backhaul transmission cost, perform backhaul closure matching on the backhaul candidate transmission pair set to determine the allowed transmission pairs and obtain the allowed transmission pair set.

[0192] In this embodiment, the median and standard deviation of the backhaul cost for all backhaul candidate transmission pairs are calculated, and the sum of the median and a standard deviation is used as the closure decision cost threshold.

[0193] For each candidate backhaul transmission pair, if its backhaul transmission cost is less than or equal to the closure decision cost threshold, it is determined to be an allowed transmission pair; if it is greater than the closure decision cost threshold, it is determined to be a prohibited transmission pair and transmission matching is not allowed.

[0194] All allowed transport pairs are combined to obtain the allowed transport pair set.

[0195] For example, the transmission costs of the eight backhaul candidate transmission pairs at measurement point 1 are 0.03175, 0.035, 0.04, 0.042, 0.045, 0.055, 0.06, and 0.07, respectively.

[0196] The calculated median is 0.0435 and the standard deviation is 0.0135. Therefore, the closure decision cost threshold is 0.0435 plus 0.0135, which is 0.057.

[0197] The transmission cost of the first transmission pair is 0.03175, which is less than or equal to 0.057, and is therefore an allowed transmission pair; the transmission cost of the sixth transmission pair is 0.055, which is less than or equal to 0.057, and is therefore an allowed transmission pair; the transmission cost of the seventh transmission pair is 0.06, which is greater than 0.057, and is therefore a prohibited transmission pair.

[0198] The final result is a set of allowed transmission pairs containing 6 allowed transmission pairs.

[0199] Step S20333: Based on the transmission quality and allowed transmission pairs, limit the transmission range of the source-end deformed segment, and retain the unmatched transmission quality of the target-end deformed segment to construct partial transmission constraints.

[0200] In this embodiment, the complete optimal transmission constraints of the prior art are: the sum of the quality of each source node's transmission to all target nodes must be equal to the transmission quality of the source node itself; the sum of the quality of all source node transmissions received by each target node must be equal to the transmission quality of the target node itself; and all transmission qualities must be greater than or equal to zero.

[0201] However, this constraint requires all deformation variables to be fully matched, leaving no room for unmatched mass. In hydraulic engineering scenarios, this can lead to the algorithm forcibly matching irreversible residual deformations into the deformations of the loading segment, making the residual deformations unrecognizable. Therefore, to address this issue, three improvements were made to the optimal transmission constraint: the equal-to-source mass constraint was changed to a less-than-equal-to-source constraint, allowing the source deformation mass to not be fully transmitted; an unmatched mass variable was added, explicitly allowing the target end to retain unmatched deformation mass; and a constraint prohibiting forced matching was added, disallowing any transmission when the transmission cost exceeds the closure decision cost threshold.

[0202] The improved transmission constraints are as follows: the sum of the quality transmitted by each source node to all target nodes is greater than or equal to 0 and less than or equal to the transmission quality of the source node itself; the sum of the quality of all source node transmissions received by each target node, plus the unmatched quality of the target node, is equal to the transmission quality of the target node itself; the transmission quality corresponding to the prohibited transmission pair is 0; all unmatched qualities are greater than or equal to 0. These constraints allow the target node to retain unmatched transmission quality, ensuring that only deformations that conform to the return loop closure relationship are matched, and all non-conforming residual deformations are retained.

[0203] Step S20334: Based on partial transmission constraints, solve for the minimum transmission cost of the allowed transmission pair set to obtain the partial optimal transmission matching result.

[0204] In this embodiment, the complete optimal transmission optimization objective in the prior art is to minimize the sum of the products of the transmission cost and the corresponding transmission quality of all transmission pairs. However, this optimization objective only considers minimizing the transmission cost and does not consider the retention of unmatched quality, which will cause the algorithm to prioritize matching all deformation variables, resulting in a high matching cost.

[0205] To address the aforementioned issues, a non-matching penalty term was added to the optimization objective to balance matching accuracy and residual retention. The improved optimization objective minimizes the sum of two parts: the first part is the sum of the products of the transmission cost and corresponding transmission quality of all allowed transmission pairs; the second part is the product of the non-matching penalty coefficient and the sum of the non-matching qualities of all target end nodes. The non-matching penalty coefficient is 0.8 times the closure decision cost threshold. When the matching cost is too high, the non-matching quality can be retained as residual deformation instead of forcibly performing high-cost matching.

[0206] Based on the aforementioned transmission constraints, the linear programming method is used to solve the optimization objective function, obtaining the matched transmission quality for each allowed transmission pair and the unmatched transmission quality for each target end node, and integrating them to obtain the partially optimal transmission matching results.

[0207] For example, the closure decision cost threshold is 0.057, so the non-match penalty coefficient is 0.8 multiplied by 0.057, resulting in 0.0456.

[0208] The transmission cost between the first source node and the tenth target node at measurement point 1 is 0.03175, which is an allowed transmission pair.

[0209] The solution yields a matched transmission quality of 2.4 mm for the transmission pair and an unmatched transmission quality of 0.3 mm for the 10th target node (2.7 mm minus 2.4 mm).

[0210] Step S2034: Based on the partial optimal transmission matching results, extract the back-haul closed matching segment and the target end unmatched deformation, and perform time-series expansion of the target end unmatched deformation segment to obtain the back-haul matched deformation segment and the back-haul unclosed deformation difference.

[0211] The specific steps of step S2034 are as follows:

[0212] Step S20341: Based on the partial optimal transmission matching results, extract the deformation segment pairs that have completed the return closed matching, and extract the unmatched deformation variables retained in the target end deformation segments to obtain the set of closed matching segment pairs and the set of unmatched retained deformation variables.

[0213] In this embodiment, all source-target node pairs with matching transmission quality greater than 0 are extracted from the partial optimal transmission matching results to obtain a closed matching fragment pair set.

[0214] Extract the unmatched transmission quality corresponding to all target end nodes to obtain the unmatched reserved deformation set.

[0215] For example, in the partial optimal transmission matching results of measurement point 1, there are 5 transmission pairs with a matching transmission quality greater than 0, including the transmission pair between the 1st source node and the 10th target node, with a matching transmission quality of 2.4 mm.

[0216] This yields a set of closed matching fragment pairs containing 5 node pairs.

[0217] The unmatched transmission quality of the 10th target node is 0.3mm, so it is added to the unmatched reserved deformation set.

[0218] Step S20342: Based on the set of closed matching segments, determine the source end deformation segment and the target end deformation segment in each closed matching segment pair as the return matching deformation segment.

[0219] In this embodiment, the closed matching segment is integrated with all source-end deformation segments and target-end deformation segments in the set to obtain the return matching deformation segment.

[0220] For example, the closed matching segment pair set of measurement point 1 contains 5 source end nodes and 5 target end nodes. Integrating these 10 deformation segments yields a return matching deformation segment set containing 10 deformation segments.

[0221] Step S20343: Based on the set of unmatched retained deformation variables, and combined with the direction of change of the fragment deformation variables corresponding to the deformation fragment at the target end, determine the unclosed candidate deformation variables corresponding to the deformation fragment at the target end.

[0222] In this embodiment, for each target end deformation segment, the corresponding unmatched retained deformation variable is multiplied by the sign of the deformation variable change direction of the segment to obtain the unclosed candidate deformation variable corresponding to the segment.

[0223] For example, the unmatched retained deformation of the 10th target end node of the first measuring point is 0.3mm, and the deformation of this segment changes in the direction of decreasing and the sign is negative.

[0224] Therefore, the unclosed candidate deformation corresponding to this segment is 0.3mm multiplied by -1, resulting in -0.3mm.

[0225] Step S20344: Based on the unclosed candidate deformation, determine the unclosed deformation corresponding to the water level recovery segment, and expand it according to the corresponding monitoring time to obtain the return unclosed deformation difference value.

[0226] In this embodiment, the prior art usually directly outputs the unmatched total deformation as the residual deformation without performing time-series expansion processing to obtain the scalar total residual value. However, in the deformation monitoring scenario of water conservancy projects, the original total deformation is a time-series sequence, while the residual deformation is a scalar. The two cannot be aligned in the time dimension. The subsequent deep learning model cannot separate the recoverable deformation and residual deformation on a time-by-time basis. It can only obtain the total residual value for the entire monitoring period and cannot achieve real-time, time-by-time deformation monitoring and early warning.

[0227] An improvement was made to address the above issues. Based on the temporal unfolding method of average fragment distribution, the unmatched deformation variables of the target end are evenly distributed to each time step within the fragment according to the fragment length. If a certain time step is covered by multiple fragments, the average value of all covered fragments is taken. Specifically, for each deformation fragment of the target end, its unclosed candidate deformation variable is divided by the deformation fragment length to obtain the unclosed deformation variable corresponding to each time step within the fragment.

[0228] If a certain monitoring moment is covered by multiple target end deformation segments, then the arithmetic mean of all unclosed deformations covering that moment is taken as the final unclosed deformation at that moment.

[0229] Arrange the unclosed deformation values ​​at all monitoring times in chronological order to obtain the sequence of return unclosed deformation difference values.

[0230] For example, the length of the deformed segment is 6.

[0231] The unclosed candidate deformation of the 10th target segment at measuring point 1 is -0.3 mm. The unclosed deformation at each moment within this segment is calculated to be -0.3 mm divided by 6, resulting in -0.05 mm.

[0232] This segment covers time series 9 to 14, therefore the unclosed deformation at these 6 time points is all negative 0.05 mm.

[0233] If time 12 is simultaneously covered by the 10th and 11th target segments, and the unclosed deformation of the 11th segment at that time is -0.06mm, then the final unclosed deformation at time 12 is the sum of -0.05mm and -0.06mm, divided by 2, resulting in -0.055mm.

[0234] like Figure 2 As shown, the rectangle on the left represents the deformation segment formed by the same monitoring point under different water level conditions; the three curves with nodes inside the box represent the trajectory of the total deformation of the monitoring point under different working conditions; the water droplet graphics with different filling states next to the curves are used to distinguish the water level loading segment, water level unloading segment and water level recovery segment; the vertical dashed line indicates the position for segmenting the continuous deformation sequence.

[0235] The short bar-shaped stage axis at the bottom left indicates the sequence of local water level stages corresponding to a single measuring point.

[0236] The long, strip-shaped stage axis running through the bottom of the figure represents the sequential relationship between the water level loading stage, the water level unloading stage, and the water level recovery stage throughout the entire monitoring cycle. Different gray areas in the strip represent different working condition stages, and the upward arrows indicate that the deformation segments corresponding to each working condition stage enter the middle return matching area.

[0237] The arrows from the left to the middle indicate that the set of deformation segments and their corresponding working condition stage information enter the return matching processing area. That is, the deformation segments that have been extracted from the left side according to the water level loading segment, water level unloading segment and water level recovery segment are sent to the middle area to construct the water level return deformation map.

[0238] The large rectangle in the middle represents the solution area for the return closure matching and partial optimal transport in step S203. The multiple columns of circular nodes in the dashed box on the left side of the middle represent the water level return deformation diagram composed of deformation segment nodes. Different columns of nodes correspond to deformation segments under different working conditions. The lines between nodes represent candidate return correspondences. Solid lines represent correspondences that better meet the return closure conditions, while dashed lines represent candidate comparison relationships or incompletely closed correspondences.

[0239] The small box in the lower left corner of the middle section represents the cyclic correspondence feature of the water level return relationship. It is used to indicate that the deformation segment needs to be matched according to the return sequence of water level loading, unloading and recovery, rather than just directly matched according to the time sequence.

[0240] The grid matrix within the dashed box in the middle represents the backhaul cost matrix corresponding to the candidate backhaul transport pairs. The varying shades of the grid represent the cost magnitudes of different candidate transport pairs. The diagonal dashed lines indicate the filtering of candidate transport relationships based on backhaul closure conditions.

[0241] The small grid below the center represents the cost consolidation result formed by the candidate transport attribute set.

[0242] The two columns of circular nodes and dashed arrows within the dashed box on the right side of the middle section represent the partial optimal transmission matching results. The nodes in the left column represent the deformed segments or segments to be matched at the source end, and the nodes in the right column represent the deformed segments at the target end. The arrows indicate the correspondence of segments that have completed transmission matching. The nodes without connected arrows represent the unmatched deformed segments at the target end. The two different dashed arrows in the small box in the lower right corner of the middle section represent the separate output of matched transmission quality and unmatched transmission quality.

[0243] The arrow from the middle to the right indicates that the partial optimal transmission matching result is output to the final extraction result. That is, after the middle part completes the backhaul transmission cost construction, backhaul closure matching and partial optimal transmission solution, the deformation segment that has completed the closure matching is output as the backhaul matching deformation segment. At the same time, the unmatched deformation at the target end is output and expanded into the backhaul unclosed deformation difference value.

[0244] The two rounded boxes on the right represent the two types of results that will be finally output in step S203.

[0245] The paired node connections in the rounded box at the top right represent the return matching deformation segments, indicating that the recoverable deformation parts that can form a closed corresponding structure during the water level return process are represented by the dashed arrows in the rounded box at the bottom right. The unclosed return deformation difference formed after the unmatched deformation at the target end is expanded in time sequence after the water level recovery segment represents the deformation parts that have not recovered with the return relationship after the water level is recovered.

[0246] Step S204: Based on the retrace matching deformation segment and the retrace unclosed deformation difference, construct the retrace closure separation network, establish the recoverable response state and the residual cumulative state, and perform separation updates to obtain the recoverable working condition response deformation sequence and the initial residual deformation sequence of the measuring point.

[0247] The specific steps of step S204 are as follows:

[0248] Step S2041: Based on the retrace matching deformation segment and the retrace unclosed deformation difference, generate the retrace matching response and unclosed residual input corresponding to each monitoring time.

[0249] In this embodiment, for each monitoring time, all backhaul matching deformation segments containing that time are extracted, the sum of the products of the matching transmission quality of all segments and the average deformation of the corresponding segments is calculated, and then divided by the sum of all matching transmission qualities. To avoid the denominator being zero, a minimum value is added to the denominator to obtain the backhaul matching response amount at that time.

[0250] If there is no corresponding return matching deformation segment at that moment, the return matching response value is 0.

[0251] The value at the corresponding time in the unclosed deformation difference sequence of the return path is directly extracted as the unclosed residual input at that time.

[0252] For example, time 10 is covered by the 10th backhaul matching deformation segment, which has a matching transmission quality of 2.4mm and an average segment deformation of 1.65mm.

[0253] Therefore, the return matching response at that moment is 2.4mm multiplied by 1.65mm and then divided by 2.4mm, resulting in 1.65mm.

[0254] The unclosed residual input at that moment was -0.05 mm.

[0255] Step S2042: Based on the operating condition feature sequence, the back-run matching response quantity and the unclosed residual input quantity, construct the back-run closure separation network, and establish the recoverable response state and the residual cumulative state in the back-run closure separation network.

[0256] In this embodiment, a back-loop closed-loop separation network based on an improved gated loop unit is constructed. The network as a whole adopts an end-to-end timing processing architecture, which does not include pooling layers and downsampling layers, thus fully preserving timing information.

[0257] The network input is an 11-dimensional feature vector at each time point, specifically including 1-dimensional total deformation of the measuring point, 8-dimensional operating condition features: water level value, water level change, ambient temperature, temperature change, cumulative rainfall, seepage pressure value, seepage pressure change and scheduling operation status, 1-dimensional return matching response and 1-dimensional unclosed residual input.

[0258] Two completely independent one-dimensional potential states are set within the network: the recoverable response state and the residual accumulation state. There is no direct information interaction between the two states. They are independently updated and controlled by the return-loop gating variable. The recoverable response state is used to describe the recoverable elastic deformation caused by changes in working conditions such as water level, temperature, and seepage. The residual accumulation state is used to describe the structural plastic residual deformation that is still retained after the water level returns.

[0259] The recoverable response state and residual cumulative state of the network are initialized to 0 at the initial moment. The weights of all fully connected layers are initialized using the Xavier uniform initialization method, and all bias terms are initialized to 0.

[0260] Step S2043: Based on the unclosed residual input quantity and the operating condition characteristic sequence, generate the return closure gating quantity, allocate the return matching response quantity to the recoverable response state, and allocate the unclosed residual input quantity to the residual accumulation state to obtain the recoverable response state sequence and the residual accumulation state sequence.

[0261] In this embodiment, existing deep learning-based deformation monitoring models typically employ a recurrent neural network with a single hidden state. The core update logic is as follows: based on the hidden state of the previous time step and the input features of the current time step, the hidden state of the current time step is calculated through a fully connected layer and a sigmoid activation function; and based on the hidden state of the current time step, the output deformation variable of the current time step is calculated through an output fully connected layer. This model encodes all deformation information into the same hidden state, making it unable to distinguish deformation components from different sources. This presents a problem in hydraulic engineering deformation monitoring scenarios: during training, the model learns the correlation that rising water levels lead to increased deformation, and falling water levels lead to decreased deformation. When residual deformation occurs in the structure, the model treats this residual deformation as part of the normal operating condition response, adjusting the model parameters to fit the total deformation variable including the residual deformation. This makes early, small residual deformations difficult to detect until they accumulate to a sufficiently large size that cannot be explained by the normal response.

[0262] To address the aforementioned issues, existing technologies are improved by designing a dual-state separation back-loop closure separation network. This network splits a single hidden state into two independent latent states: the recoverable response state and the residual cumulative state. These states encode deformation information from different sources. A back-loop closure gating variable is introduced to dynamically control the update ratio of the two states. Specifically, the 4-dimensional gating input features at each time step are extracted, including the unclosed residual input, water level change, seepage pressure change, and scheduling operation status. These features are then input into the gating network, which employs a single-layer fully connected layer structure with an input dimension of 4, an output dimension of 1, and an activation function of sigmoid. The output range is 0 to 1, representing the back-loop closure gating variable.

[0263] Based on the 10-dimensional recoverable input features at the current time, including the recoverable response state at the previous time, 8-dimensional working condition features, and backhaul matching response quantity, the recovery response candidate states are obtained by calculation through the first fully connected layer with an input dimension of 10 and an output dimension of 1, and the activation function is the hyperbolic tangent function.

[0264] Based on the 3D residual input features at the current moment, including the residual cumulative state at the previous moment, the amount of unclosed residual input, and the change in osmotic pressure, the residual cumulative candidate state is obtained by calculation through a second fully connected layer with an input dimension of 3, an output dimension of 1, and a hyperbolic tangent activation function.

[0265] Based on the quiescent closure gating, the two candidate states are updated. The recoverable response state is equal to 1 minus the difference between the quiescent closure gating and the candidate recoverable response state, plus the quiescent closure gating multiplied by the recoverable response state at the previous moment.

[0266] The residual cumulative state equals the return closure gating quantity multiplied by the residual cumulative candidate state, plus 1 minus the difference of the return closure gating quantity multiplied by the residual cumulative state at the previous moment.

[0267] All states are updated sequentially to obtain a recoverable response state sequence and a residual cumulative state sequence.

[0268] The network training adopts a supervised learning approach, dividing the complete monitoring time series data into a training set and a validation set in chronological order. The training set accounts for 80% and the validation set accounts for 20%, and no random shuffling is performed to preserve the continuity of the time series.

[0269] The training batch size was set to 32, the optimizer used the Adam optimization algorithm, the initial learning rate was set to 0.001, the learning rate decay factor was 0.9, and it decayed once every 10 training rounds.

[0270] The total number of training rounds is set to 100. An early stopping mechanism is adopted. When the total loss of the validation set no longer decreases for 10 consecutive rounds, the training is terminated early and the optimal model parameters are saved.

[0271] The total loss function is composed of a weighted average of reconstruction loss, back-path matching consistency loss, unclosed residual correspondence loss, and residual smoothing loss. The weights of the four loss terms are determined using the analytic hierarchy process (AHP). For example, the weight of reconstruction loss is 1.0, the weight of back-path matching consistency loss is 0.5, the weight of unclosed residual correspondence loss is 0.8, and the weight of residual smoothing loss is 0.05. The reconstruction loss is calculated by first summing the recoverable operating condition response deformation and the initial residual deformation at each time step, then calculating the difference between this sum and the measured total deformation at the corresponding time step, and finally averaging all the differences.

[0272] The backhaul matching consistency loss is calculated only for the moments when a backhaul matching relationship exists. For each backhaul matching segment, the average value of the recoverable operating condition response deformation of the network output at all moments within the segment is calculated. Then, the absolute value of the difference between the average value and the matching transmission quality of the segment is calculated. The average of the above absolute values ​​of all matching segments is obtained.

[0273] The loss corresponding to the unclosed residual is first calculated by calculating the difference between the initial residual deformation of the network output at each time step and the initial residual deformation of the previous time step, to obtain the residual deformation update amount at that time step. Then, the difference between this update amount and the unclosed residual input amount at the corresponding time step multiplied by the scaling factor is calculated. The average value is obtained by squaring the above differences at all times. For example, the scaling factor is taken as 0.95.

[0274] The residual smoothing loss is calculated by taking the square of all the differences and then averaging them.

[0275] The total loss is obtained by multiplying each of the four loss items by its corresponding weight and then summing them up.

[0276] With the above improvements, when the return unclosed deformation difference is small, the gating amount is small, and the model mainly updates the recoverable response state, interpreting the current deformation as the normal operating condition response; when the return unclosed deformation difference is large, the gating amount is large, and the model mainly updates the residual cumulative state, interpreting the current deformation as the residual cumulative deformation.

[0277] For example, at time 10, the unclosed residual input is -0.05 mm, the water level change is -0.02 m, the seepage pressure change is -0.005 MPa, and the scheduling operation status is 0.

[0278] After inputting into the fully connected layer of the gated network, the output closure gate value after passing through the Sigmoid activation function is 0.2.

[0279] The recoverable response candidate state is 1.3mm, and the previous recoverable response state was 1.1mm.

[0280] The calculated recoverable response state at that moment is the difference between 1 and 0.2, multiplied by 1.3 mm, plus 0.2 multiplied by 1.1 mm, resulting in 1.26 mm.

[0281] The residual cumulative candidate state is 0.5mm, and the residual cumulative state at the previous time step was 0.4mm.

[0282] The calculated residual cumulative state at that moment is 0.2 multiplied by 0.5 mm. Adding the difference between 1 and 0.2 multiplied by 0.4 mm, the result is 0.42 mm.

[0283] Step S2044: Based on the recoverable response state sequence, generate the recoverable operating condition response deformation sequence of the measuring points.

[0284] In this embodiment, the recoverable response state sequence is input into the first output fully connected layer to obtain the recoverable operating condition response deformation of the measurement point at each time step.

[0285] Arrange the recoverable operating condition response variables at all times in chronological order to obtain the sequence of recoverable operating condition response variables at the measuring points.

[0286] For example, the recoverable response state at time 10 is 1.26mm. After mapping through the first output fully connected layer, the recoverable operating condition response deformation at that time is 1.26mm.

[0287] The recoverable operating condition response deformation variables at all times were calculated using the same method, resulting in the recoverable operating condition response deformation variable sequence for measuring point No. 1.

[0288] Step S2045: Generate the initial residual deformation sequence of the measurement points based on the residual cumulative state sequence.

[0289] In this embodiment, the residual cumulative state sequence is input into the second output fully connected layer to obtain the initial residual deformation of the measurement point at each time step.

[0290] Arrange the initial residual deformations at all times in chronological order to obtain the sequence of initial residual deformations at the measuring points.

[0291] For example, the residual cumulative state at time 10 is 0.42mm. After mapping through the second output fully connected layer, the initial residual deformation at that time is 0.42mm.

[0292] The initial residual deformation at all times was calculated using the same method, resulting in the initial residual deformation sequence for measuring point 1.

[0293] Step S3: Based on the total deformation sequence of the measuring points, the recoverable working condition response deformation sequence of the measuring points, the initial residual deformation sequence of the measuring points, and the structural attribution relationship of the measuring points, construct a partition boundary-preserving graph neural inference network to obtain the deformation monitoring results of the water conservancy project.

[0294] The specific steps of step S3 are as follows:

[0295] Step S301: Based on the total deformation sequence of the measuring points, the recoverable working condition response deformation sequence of the measuring points, the initial residual deformation sequence of the measuring points, and the structural affiliation of the measuring points, generate a set of measuring point node features with structural partitioning labels.

[0296] In this embodiment, each monitoring point is mapped to an independent node, and the node number is consistent with the number of the corresponding monitoring point.

[0297] For each node, a 7-dimensional initial feature vector is constructed, which includes the total deformation of the measuring point at the current time, the recoverable working condition response deformation, the initial residual deformation, the structural partition number, the horizontal coordinate, the vertical coordinate, and the elevation value.

[0298] The initial feature vectors of all monitoring points are integrated in order of node number to obtain a feature set of monitoring point nodes with structural partitioning labels.

[0299] For example, the total deformation of monitoring point 1 at time 10 is 1.7 mm, the recoverable working condition response deformation is 1.26 mm, the initial residual deformation is 0.42 mm, the structural partition number is 1, the plane coordinates are (523682.5, 3412769.2), and the elevation value is 125.0 m.

[0300] Therefore, the initial feature vector of this node at time 10 is [1.7, 1.26, 0.42, 1, 523682.5, 3412769.2, 125.0].

[0301] Step S302: Based on the feature set of measurement point nodes, connect the measurement point nodes within the same structural partition as the same-region aggregate edge, and connect the measurement point nodes on both sides of the boundary of adjacent structural partitions as the boundary preservation edge to obtain the partition boundary preservation deformation map.

[0302] In this embodiment, edges are divided into two categories: intra-region aggregation edges and boundary preservation edges, which respectively realize the continuous utilization of intra-region space and the preservation of boundary differences. For any two measuring point nodes, if the structural partition numbers to which the two nodes belong are the same and the Euclidean distance between the two nodes is less than or equal to the neighborhood distance threshold, then an intra-region aggregation edge is established between the two nodes; if the structural partition numbers to which the two nodes belong are different and there are joints, expansion joints or dam section boundaries between the two structural partitions, then a boundary preservation edge is established between the two nodes.

[0303] By integrating all nodes and all edges, we obtain a partition boundary-preserving deformation graph.

[0304] Considering that the length of a single dam section or embankment section in a water conservancy project is usually within 50m, the preferred neighborhood distance threshold is 20m to 50m.

[0305] For example, the neighborhood distance threshold is set to 30m.

[0306] The structural partition number of measuring point 1 is 1, and the structural partition number of measuring point 2 is also 1. The Euclidean distance between the two measuring points is 15m, which is less than 30m. Therefore, a co-regional aggregation edge is established between measuring point 1 and measuring point 2.

[0307] The structural zone number to which measuring point 3 belongs is 2. There is an expansion joint boundary between zone 1 and zone 2. Therefore, a boundary retention edge is established between measuring point 1 and measuring point 3.

[0308] Step S303: Based on the partition boundary preservation deformation map, construct a partition boundary preservation map neural inference network, perform same-region aggregation and boundary difference preservation on the initial residual deformation sequence of the measurement points, and obtain the same-region aggregated residual deformation and boundary difference residual deformation.

[0309] In this embodiment, a partition boundary-preserving graph neural inference network based on a boundary-aware graph attention network is constructed. This network can automatically learn the spatial correlation of test points within the same structural partition, while preserving the deformation difference features at the boundaries of adjacent partitions, thus avoiding feature mixing across partitions.

[0310] The network adopts a three-layer structure, consisting of an input layer, two boundary-aware map attention layers, and a dual-branch output layer.

[0311] The input layer receives a feature set of measurement point nodes with structural partitioning labels, performs standardization processing on the features of each dimension, and maps the values ​​uniformly to the interval [-1, 1].

[0312] The first boundary-aware graph attention layer sets up 8 parallel attention computation heads, each of which outputs 16-dimensional features. The output features of all attention heads are concatenated to obtain 128-dimensional intermediate features. This layer introduces a boundary attention masking mechanism, which only allows feature interaction between nodes connected by the same region aggregation edge, and prohibits feature transfer between cross-region nodes connected by the boundary maintaining edge.

[0313] The second-layer boundary-aware map attention layer is equipped with one attention computation head, which outputs 32-dimensional high-order spatial features to further extract global deformation features within the same partition.

[0314] The output layer contains two independent fully connected branches. The first branch is the same-region aggregation branch, which takes 32-dimensional high-order features as input and outputs the same-region aggregation residual deformation of each measurement point. The second branch is the boundary difference branch, which takes 32-dimensional high-order features as input and outputs the boundary difference residual deformation of each boundary measurement point.

[0315] Manually verified residual deformation data from over three years of engineering history were collected as supervisory labels, including manually measured residual deformation at each measuring point and boundary difference measurements of adjacent zones. Hourly node features over 30 consecutive days were used as a training sample, and the corresponding monthly average residual deformation and boundary difference were used as sample labels.

[0316] All samples were divided into training set and validation set in chronological order, with the training set accounting for 80% and the validation set accounting for 20%.

[0317] The network training uses a weighted mean squared error loss function to calculate the error between the aggregated residual deformation and the manual label in the same region, and the error between the boundary difference residual deformation and the manual label. The two errors are multiplied by their respective weights and then added together to obtain the total loss.

[0318] For example, the weight of the same-region aggregation error is set to 1.0, and the weight of the boundary difference error is set to 1.5. During training, the batch size is set to 16, the AdamW optimization algorithm is used, the initial learning rate is set to 0.0005, the weight decay coefficient is set to 0.0001, and the total number of training rounds is set to 50. When the total loss of the validation set no longer decreases for 5 consecutive rounds, the training is terminated early and the optimal model parameters are saved.

[0319] The feature set of measurement point nodes with structural partition labels and the partition boundary preservation deformation map to be processed are input into the trained partition boundary preservation graph neural inference network. The network automatically performs adaptive weighted aggregation of measurement point features within the same structural partition. Different aggregation weights are learned based on the spatial location and historical deformation features of each measurement point. The network outputs the same-region aggregation residual deformation of each measurement point. At the same time, the network calculates the difference of measurement point features on both sides of the boundary of adjacent partitions and outputs the boundary difference residual deformation of each boundary measurement point.

[0320] For example, after inputting the deformation map of the partition boundary where measuring point 1 is located into the network, the network automatically calculates the aggregation weights of the three neighboring measuring points as 0.47, 0.37 and 0.16 respectively, based on the spatial distance and historical deformation correlation between measuring point 1 and measuring points 2, 3 and 4 in the same area. The network outputs the aggregated residual deformation of measuring point 1 in the same area as 0.4157 mm. At the same time, the network calculates the feature difference between measuring point 1 and measuring point 5 on the other side of the boundary, and outputs the boundary difference residual deformation of measuring point 1 as 0.43 mm.

[0321] Step S304: Based on the residual deformation variables of the same region and the residual deformation variables of the boundary difference, perform structural partitioning correction on the initial residual deformation variable sequence of the measurement points to obtain the structural body residual deformation variables.

[0322] In this embodiment, the distance from each measuring point to the nearest boundary of its structural partition is calculated. If the distance is less than or equal to the boundary influence distance, the boundary preservation coefficient is set to 1; if the distance is greater than the boundary influence distance, the boundary preservation coefficient is set to 0.

[0323] The structural body residual deformation of the measuring point is obtained by adding the boundary preservation coefficient to the residual deformation of the same region and multiplying it by the residual deformation of the boundary difference.

[0324] Considering that the influence range at the structural boundary is usually within 15m, the boundary influence distance is preferably 5m to 15m.

[0325] For example, the boundary influence distance is taken as 10m.

[0326] The distance from measuring point 1 to the nearest structural boundary is 8m, which is less than 10m, so the boundary preservation coefficient is taken as 1.

[0327] The calculated residual deformation of the structure at measuring point 1 is 0.4157 mm. Adding 1 to 0.43 mm gives a result of 0.8457 mm.

[0328] The distance from measuring point 6 to the nearest structural boundary is 20m, which is greater than 10m. Therefore, the boundary preservation coefficient is taken as 0, and the residual deformation of the structural body is equal to the aggregate residual deformation of the same region.

[0329] Step S305: Generate deformation monitoring results for hydraulic engineering based on the total deformation sequence of measuring points, the recoverable working condition response deformation sequence of measuring points, and the residual deformation of the structural body.

[0330] In this embodiment, for each monitoring point, the total deformation, recoverable working condition response deformation, and structural residual deformation at each time step are integrated to obtain the time-by-time deformation monitoring result of that monitoring point.

[0331] For each structural partition, the arithmetic mean of the residual deformation of the structural body at all measurement points within that partition is calculated to obtain the time-by-time residual deformation of that structural partition.

[0332] For each pair of adjacent structural partitions, calculate the maximum value of the residual deformation of the boundary difference at all measuring points at the boundary of the two partitions to obtain the time-by-time boundary difference of the adjacent partitions.

[0333] The deformation monitoring results of all measuring points, the residual deformation of all structural zones, and the boundary differences of all adjacent zones are integrated to obtain the final deformation monitoring results of the hydraulic engineering project.

[0334] For example, the total deformation of measuring point 1 at time 10 is 1.7 mm, the recoverable working condition response deformation is 1.26 mm, and the residual deformation of the structure body is 0.8457 mm. Therefore, the monitoring result of this measuring point is [1.7, 1.26, 0.8457].

[0335] There are 10 measuring points in the upstream section of the No. 1 dam. The average value of the residual deformation of the structure at all measuring points is 0.72 mm. Therefore, the residual deformation of this dam section is 0.72 mm.

[0336] The maximum residual deformation of the boundary difference between partition 1 and partition 2 is 0.52 mm, therefore the boundary difference between these adjacent partitions is 0.52 mm.

[0337] Example 2:

[0338] See Figure 3 This embodiment provides a deep learning-based hydraulic engineering deformation monitoring system, including: a data acquisition module, a separation network module, and a deformation monitoring module.

[0339] The data acquisition module is used to collect data sets of deformation monitoring of water conservancy projects, organize the monitoring data into a benchmark, and obtain the total deformation sequence of measuring points, the structural attribution relationship of measuring points, the working condition characteristic sequence, and the working condition stage sequence.

[0340] The separation network module constructs a return closed separation network based on the total deformation sequence of the measurement points, the operating condition characteristic sequence, and the operating condition stage sequence, thereby obtaining the recoverable operating condition response deformation sequence of the measurement points and the initial residual deformation sequence of the measurement points.

[0341] The deformation monitoring module constructs a partition boundary-preserving graph neural inference network based on the total deformation sequence of the measuring points, the recoverable working condition response deformation sequence of the measuring points, the initial residual deformation sequence of the measuring points, and the structural attribution relationship of the measuring points, to obtain the deformation monitoring results of the hydraulic engineering project.

[0342] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0343] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0344] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring deformation in hydraulic engineering based on deep learning, characterized in that, include: Collect a deformation monitoring dataset of a water conservancy project, organize the monitoring data into a benchmark, and obtain the total deformation sequence of the measuring points, the structural attribution relationship of the measuring points, the working condition characteristic sequence, and the working condition stage sequence. The deformation monitoring dataset of the water conservancy project includes measuring point displacement observation data, water level data, scheduling and operation records, measuring point layout data, and structural zoning data. The working condition stage sequence includes the water level loading stage, the water level unloading stage, and the water level recovery stage. Based on the total deformation sequence, operating condition characteristic sequence, and operating condition stage sequence of the measuring points, a return closure separation network is constructed to obtain the recoverable operating condition response deformation sequence and the initial residual deformation sequence of the measuring points. The recoverable operating condition response deformation sequence and the initial residual deformation sequence of the measuring points are obtained by performing return closure matching and partial optimal transmission solution based on the water level return deformation diagram, and by performing state separation and updating based on the return unclosed deformation difference. Based on the total deformation sequence of measuring points, the recoverable working condition response deformation sequence of measuring points, the initial residual deformation sequence of measuring points, and the structural attribution relationship of measuring points, a partition boundary-preserving graph neural inference network is constructed to obtain the deformation monitoring results of water conservancy projects. Based on the total deformation sequence of the measuring points, the operating condition characteristic sequence, and the operating condition stage sequence, a return closed-loop separation network is constructed to obtain the recoverable operating condition response deformation sequence and the initial residual deformation sequence of the measuring points, including: Based on the total deformation sequence of the measuring points and the working condition stage sequence, the total deformation of each monitoring point in the water level loading segment, water level unloading segment and water level recovery segment is segmented to obtain a set of deformation segments. Based on the set of deformation fragments and the sequence of working conditions, a set of deformation fragment attributes is generated, and a water level return deformation map is constructed. Based on the water level return deformation map and the deformation segment attribute set, the return transmission cost is constructed, and the return closure matching and partial optimal transmission solution are performed to obtain the return matching deformation segment and the return unclosed deformation difference. Based on the back-travel matching deformation segment and the back-travel unclosed deformation difference, a back-travel closed separation network is constructed to establish a recoverable response state and a residual cumulative state, and then the separation and update are performed to obtain the recoverable working condition response deformation sequence and the initial residual deformation sequence of the measuring point. Based on the water level return deformation map and deformation segment attribute set, the return transmission cost is constructed, and return closure matching and partial optimal transmission solutions are performed to obtain the return matching deformation segment and the return unclosed deformation difference value, including: Based on the water level return deformation map and the deformation segment attribute set, the deformation segment node pairs connected by the return edge are extracted, and the deformation segment node pairs are divided into source end deformation segments and target end deformation segments to obtain the return candidate transmission pair set; Based on the backhaul candidate transmission pair set, water level backhaul comparison, deformation recovery direction comparison, and stage sequence comparison are performed to obtain the candidate transmission attribute set and generate the backhaul transmission cost corresponding to each backhaul candidate transmission pair. Based on the backhaul transmission cost, backhaul closure matching and partial optimal transmission solution are performed on the backhaul candidate transmission pair set to obtain the partial optimal transmission matching result, which includes the matched transmission quality and the unmatched transmission quality. Based on the partial optimal transmission matching results, the return closed matching segment and the target end unmatched deformation are extracted, and the target end unmatched deformation is subjected to water level recovery segment time sequence expansion to obtain the return matching deformation segment and the return unclosed deformation difference.

2. The deep learning-based deformation monitoring method for hydraulic engineering projects according to claim 1, characterized in that, The collected data set of deformation monitoring data for water conservancy projects is organized into a standardized format to obtain the total deformation sequence of monitoring points, the structural attribution relationship of monitoring points, the working condition characteristic sequence, and the working condition stage sequence, including: Collect data sets for deformation monitoring of water conservancy projects; Based on the displacement observation data of the measuring points, the displacement observation values ​​of each monitoring point at each monitoring time are determined, and the displacement observation values ​​are processed by benchmark difference to obtain the total deformation sequence of the measuring points. Based on the data of the measurement point layout and the structural zoning data, the structural zoning corresponding to each monitoring point is determined, and the structural zoning is mapped to the total shape sequence of the measurement points to obtain the structural affiliation relationship of the measurement points; Based on water level data and scheduling operation records, the working condition data corresponding to each monitoring time in the total deformation sequence of the measuring points are extracted, and the water level change and scheduling status are generated to obtain the working condition feature sequence. Based on the water level change and scheduling status, the direction of water level change and scheduling operation status corresponding to each monitoring moment are determined, and stage markings are performed to obtain the working condition stage sequence.

3. The deep learning-based deformation monitoring method for hydraulic engineering projects according to claim 1, characterized in that, The process of generating a set of deformation fragment attributes and constructing a water level return deformation map based on a set of deformation fragments and a sequence of working conditions includes: Based on the set of deformation segments, determine the start and end times of each deformation segment and the working condition stage to which the segment belongs; Based on the working condition feature sequence, the average water level of each deformation segment, the direction of water level change, and the direction of deformation change are determined to obtain the attribute set of the deformation segment. Based on the attribute set of deformable fragments, each deformable fragment is mapped to a deformable fragment node, and each deformable fragment node is paired to obtain a set of candidate return node pairs; Based on the deformation segment attribute set, the candidate return node pair set is judged by water level proximity, working condition stage difference and stage sequence to obtain the effective return node pair set; Based on the set of effective return node pairs, return edges are generated, and combined with deformation fragment nodes, a water level return deformation map is constructed.

4. The method for monitoring deformation of hydraulic engineering based on deep learning according to claim 1, characterized in that, The method involves performing backhaul closure matching and partial optimal transmission solution on the backhaul candidate transmission pair set based on backhaul transmission cost, to obtain partial optimal transmission matching results, including: Based on the backhaul candidate transmission pair set, determine the transmission quality corresponding to each source end deformation segment and each target end deformation segment; Based on the backhaul transmission cost, backhaul closure matching is performed on the backhaul candidate transmission pair set to determine the allowed transmission pairs and obtain the allowed transmission pair set. Based on transmission quality and allowed transmission pairs, the transmission range of the source-end deformed segment is limited, while the unmatched transmission quality of the target-end deformed segment is preserved, thus constructing partial transmission constraints. Based on partial transmission constraints, the minimum transmission cost is solved for the set of allowed transmission pairs to obtain partially optimal transmission matching results.

5. The method for monitoring deformation of hydraulic engineering based on deep learning according to claim 1, characterized in that, Based on the partially optimal transmission matching results, the return closed matching segment and the target unmatched deformation are extracted, and the target unmatched deformation is subjected to time-series expansion of the water level recovery segment to obtain the difference between the return matched deformation segment and the return unclosed deformation, including: Based on the partial optimal transmission matching results, the deformation segment pairs that have completed the return closed matching are extracted, and the unmatched deformation variables retained in the target end deformation segments are extracted to obtain the set of closed matching segment pairs and the set of unmatched retained deformation variables; Based on the set of closed matching segments, the source-end deformation segment and the target-end deformation segment in each closed matching segment pair are determined as the return matching deformation segment; Based on the set of unmatched retained deformation variables, and combined with the direction of change of the fragment deformation variables corresponding to the deformation fragment at the target end, the unclosed candidate deformation variables corresponding to the deformation fragment at the target end are determined. Based on the unclosed candidate deformation, the unclosed deformation corresponding to the water level recovery segment is determined, and it is expanded according to the corresponding monitoring time to obtain the return unclosed deformation difference value.

6. The deep learning-based deformation monitoring method for hydraulic engineering projects according to claim 1, characterized in that, The method involves constructing a return-closed separation network based on the return-matching deformation segment and the return-unclosed deformation difference, establishing a recoverable response state and a residual cumulative state, and performing separation updates to obtain the recoverable operating condition response deformation sequence and the initial residual deformation sequence of the measuring point, including: Based on the back-matching deformation segment and the back-unclosed deformation difference, the back-matching response and the unclosed residual input are generated at each monitoring time. Based on the operating condition characteristic sequence, the back-run matched response quantity and the unclosed residual input quantity, a back-run closed separation network is constructed, and a recoverable response state and a residual cumulative state are established in the back-run closed separation network. Based on the unclosed residual input quantity and the operating condition characteristic sequence, the return closure gating quantity is generated, the return matching response quantity is allocated to the recoverable response state, and the unclosed residual input quantity is allocated to the residual accumulation state, thus obtaining the recoverable response state sequence and the residual accumulation state sequence. Based on the recoverable response state sequence, generate the recoverable working condition response deformation sequence of the measuring points; Based on the residual cumulative state sequence, the initial residual deformation sequence of the measuring point is generated.

7. The method for monitoring deformation of hydraulic engineering based on deep learning according to claim 1, characterized in that, Based on the total deformation sequence of measuring points, the recoverable working condition response deformation sequence of measuring points, the initial residual deformation sequence of measuring points, and the structural attribution relationship of measuring points, a partition boundary-preserving graph neural inference network is constructed to obtain the deformation monitoring results of hydraulic engineering projects, including: Based on the total deformation sequence of the measuring points, the recoverable working condition response deformation sequence of the measuring points, the initial residual deformation sequence of the measuring points, and the structural attribution relationship of the measuring points, a feature set of measuring point nodes with structural partitioning labels is generated. Based on the feature set of measurement point nodes, measurement point nodes within the same structural partition are connected as aggregated edges within the same partition, and measurement point nodes on both sides of the boundary of adjacent structural partitions are connected as boundary preservation edges, thus obtaining a partition boundary preservation deformation map. Based on the partition boundary preservation deformation map, a partition boundary preservation graph neural inference network is constructed to perform same-region aggregation and boundary difference preservation on the initial residual deformation sequence of the measurement points, so as to obtain the same-region aggregated residual deformation and the boundary difference residual deformation. Based on the residual deformation variables of the same region and the residual deformation variables of the boundary difference, the initial residual deformation variable sequence of the measuring points is corrected by structural partitioning to obtain the residual deformation variables of the structural body. Based on the total deformation sequence of the measuring points, the recoverable working condition response deformation sequence of the measuring points, and the residual deformation of the structural body, deformation monitoring results of hydraulic engineering are generated.

8. A deep learning-based hydraulic engineering deformation monitoring system, used to implement the deep learning-based hydraulic engineering deformation monitoring method according to any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to collect deformation monitoring datasets of water conservancy projects, organize the monitoring data into a benchmark, and obtain the total deformation sequence of measuring points, the structural attribution relationship of measuring points, the working condition characteristic sequence, and the working condition stage sequence. The separation network module constructs a return closed separation network based on the total deformation sequence of the measuring points, the working condition characteristic sequence, and the working condition stage sequence, to obtain the recoverable working condition response deformation sequence of the measuring points and the initial residual deformation sequence of the measuring points; The deformation monitoring module constructs a partition boundary-preserving graph neural inference network based on the total deformation sequence of the measuring points, the recoverable working condition response deformation sequence of the measuring points, the initial residual deformation sequence of the measuring points, and the structural attribution relationship of the measuring points, to obtain the deformation monitoring results of the hydraulic engineering project.

Images