Dam monitoring data processing method and system

Abstract

The invention belongs to the technical field of water conservancy project monitoring and data processing, and provides a dam monitoring data processing method and system, which is characterized in that a space-time collaborative sensing architecture of multi-source heterogeneous monitoring data is constructed, and physical field constraints and a data driving model are fused. High-robustness quality control, anomaly traceability analysis and dynamic reconstruction compensation of dam global monitoring data are realized, so that the integrity, accuracy and real-time availability of the monitoring data are guaranteed under complex working conditions, and high-confidence data support is provided for dam safety assessment and early warning decision.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy engineering monitoring and data processing technology, specifically, it relates to a method and system for processing dam monitoring data. Background Technology

[0002] Modern dam monitoring systems rely on a multi-type sensor network deployed across the dam body, foundation, and surrounding environment to continuously collect key physical parameters such as displacement, seepage pressure, stress, and temperature, forming a high-dimensional, high-frequency time-series data stream. This data is not only the foundation for assessing the structural health status but also a crucial input for triggering early warning mechanisms and guiding operational decisions. However, monitoring data is highly susceptible to environmental interference, equipment drift, communication interruptions, and other factors during acquisition, transmission, and storage, resulting in inconsistent data quality. This severely restricts the reliability of subsequent analysis models and the response accuracy of early warning systems.

[0003] Existing technologies generally focus on two levels: gross error identification and early warning model construction, attempting to filter noise and predict potential risks through algorithms. However, current mainstream methods reveal systemic flaws when dealing with complex operating conditions: existing technologies rely excessively on the integrity and stability of historical data, leading to a sharp decline in model robustness in scenarios with missing data or sudden noise interference; their gross error identification process often uses isolated point detection or single-point threshold judgment, failing to effectively model the spatial topological relationships and physical coupling mechanisms between monitoring points, causing local anomalies to be misjudged as global trends or vice versa; at the same time, most solutions only stay at the "identification-removal" level, lacking source analysis and guidance for the repair of anomalies, resulting in broken data chains and hindered model iteration. Summary of the Invention

[0004] To address the problems in existing technologies, this invention provides a method and system for processing dam monitoring data. By constructing a spatiotemporal collaborative sensing architecture for multi-source heterogeneous monitoring data and integrating physical field constraints and data-driven models, it achieves highly robust quality control, anomaly tracing and analysis, and dynamic reconstruction compensation for dam-wide monitoring data. This ensures the integrity, accuracy, and real-time availability of monitoring data under complex operating conditions, providing high-confidence data support for dam safety assessment and early warning decisions.

[0005] According to one aspect of the present invention, a method for processing dam monitoring data is provided, comprising: Multi-type monitoring data of the entire dam structure are collected synchronously through a distributed sensor network. The multi-type monitoring data includes displacement data, seepage pressure data, stress and strain data, temperature data, and environmental load data. Spatiotemporal alignment and sampling frequency normalization are performed on multiple types of monitoring data to construct a multidimensional monitoring data matrix under a unified spatiotemporal coordinate system, wherein the time dimension is divided according to a preset sampling period and the spatial dimension is mapped according to the physical coordinates of the monitoring points. The initial quality screening of the multidimensional monitoring data matrix is ​​carried out based on the physical field constraint model, which includes the elasticity equilibrium equation, the seepage continuity equation, and the heat conduction differential equation. By calculating the absolute value of the residual between the data of each monitoring point and the theoretical value of the physical field, data points that exceed the preset physical threshold are marked as suspicious data. Spatial neighborhood correlation verification is performed on suspicious data. Spatial neighborhood correlation verification is performed by constructing a three-dimensional spatial neighborhood window centered on the suspicious data point, calculating the eigenvalue spectrum of the covariance matrix of all monitoring point data within the window, and determining that the suspicious data point is an isolated anomaly if the proportion of the principal eigenvalue is lower than the preset proportion threshold. Multimodal source tracing analysis is performed on isolated anomalies. The multimodal source tracing analysis includes comparison of sensor status self-test signals, matching of environmental disturbance event logs, and backtracking of historical data of similar working conditions. Anomaly cause classification labels are generated, which include three categories: sensor hardware failure, transient environmental interference, and local structural mutation. Based on the anomaly cause classification label, differentiated data reconstruction is performed on isolated anomalies. If the anomaly is caused by sensor hardware failure, spatial interpolation is used to generate alternative values ​​based on neighboring monitoring point data using the Kriging interpolation algorithm. If the anomaly is caused by transient environmental interference, time series prediction is used to generate alternative values ​​based on the historical data of the monitoring point using an autoregressive moving average model. If the anomaly is caused by local structural mutation, the original data is retained and a mutation event label is added. The reconstructed monitoring data is input into the dynamic coupling early warning model, which consists of a physical driving sub-model and a data driving sub-model in parallel. The physical driving sub-model calculates the theoretical value of the dam structure response in real time based on the finite element simulation framework, while the data driving sub-model learns the temporal evolution law of the monitoring data in real time based on the long short-term memory neural network. The output results of the two are weighted and fused to generate a comprehensive early warning index. Based on the comparison between the comprehensive early warning index and the preset classification threshold, the corresponding early warning response mechanism is triggered. The early warning response mechanism includes issuing data resampling instructions, generating monitoring point inspection tasks, and automatically compiling structural safety assessment reports.

[0006] According to another aspect of the present invention, a dam monitoring data processing system is provided, comprising: The multi-source data synchronous acquisition module is used to synchronously acquire various types of monitoring data across the entire dam structure through a distributed sensor network. These data include displacement data, seepage pressure data, stress and strain data, temperature data, and environmental load data. The spatiotemporal data normalization module is used to perform spatiotemporal alignment and sampling frequency normalization on multiple types of monitoring data, and to construct a multidimensional monitoring data matrix under a unified spatiotemporal coordinate system, wherein the time dimension is divided according to the preset sampling period, and the spatial dimension is mapped according to the physical coordinates of the monitoring points. The physical field constraint screening module is used to perform initial quality screening on the multidimensional monitoring data matrix based on the physical field constraint model. The physical field constraint model includes the elasticity equilibrium equation, the seepage continuity equation, and the heat conduction differential equation. By calculating the absolute value of the residual between the data of each monitoring point and the theoretical value of the physical field, data points that exceed the preset physical threshold are marked as suspicious data. The spatial correlation verification module is used to verify the spatial neighborhood correlation of suspicious data. The spatial neighborhood correlation verification is carried out by constructing a three-dimensional spatial neighborhood window centered on the suspicious data point, calculating the eigenvalue spectrum of the covariance matrix of all monitoring point data within the window, and if the proportion of the principal eigenvalue is lower than the preset proportion threshold, the suspicious data point is determined to be an isolated anomaly. The multimodal source analysis module is used to perform multimodal source analysis on isolated anomalies. The multimodal source analysis includes sensor status self-test signal comparison, environmental disturbance event log matching, and historical data backtracking of similar working conditions to generate anomaly cause classification labels. The anomaly cause classification labels include three categories: sensor hardware failure, transient environmental interference, and local structural mutation. The differential data reconstruction module is used to perform differential data reconstruction on isolated anomalies based on the anomaly cause classification label. If the anomaly is caused by sensor hardware failure, spatial interpolation is used to generate alternative values ​​based on neighboring monitoring point data using the Kriging interpolation algorithm. If the anomaly is caused by transient environmental interference, time series prediction is used to generate alternative values ​​based on the historical data of the monitoring point using an autoregressive moving average model. If the anomaly is caused by a local structural mutation, the original data is retained and a mutation event label is added. The dynamic coupling early warning module is used to input the reconstructed monitoring data into the dynamic coupling early warning model. The dynamic coupling early warning model is composed of a physical driving sub-model and a data driving sub-model in parallel. The physical driving sub-model calculates the theoretical value of the dam structure response in real time based on the finite element simulation framework, while the data driving sub-model learns the temporal evolution law of the monitoring data in real time based on the long short-term memory neural network. The output results of the two are weighted and fused to generate a comprehensive early warning index. The graded early warning response module is used to trigger the corresponding level of early warning response mechanism based on the comparison result between the comprehensive early warning index and the preset graded threshold. The early warning response mechanism includes issuing data resampling instructions, generating monitoring point inspection tasks, and automatically compiling structural safety assessment reports.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention effectively overcomes the problems of high false positive rates and poor adaptability caused by relying solely on statistical models in existing technologies by constructing a dual quality control mechanism that integrates physical field constraints and data-driven approaches. The physical field constraint model, based on the fundamental laws of mechanics, seepage, and thermodynamics, provides a solid theoretical boundary for data screening, significantly reducing false alarms caused by environmental noise or sensor drift. The spatial neighborhood correlation verification mechanism fully utilizes the spatial continuity characteristics of the dam structure to accurately identify isolated anomalies, avoiding misjudging local mutations as systemic failures. The multimodal source tracing analysis module enables refined classification of anomaly causes, providing a clear basis for subsequent differentiated data reconstruction and greatly improving the rationality and engineering practicality of data repair.

[0008] 2. The dynamic coupling early warning model proposed in this invention organically combines physical simulation with machine learning. It retains the strong interpretability of physical models for structural mechanisms while leveraging the strong fitting ability of data models for complex nonlinear patterns. Through a dynamic weight adjustment mechanism, the early warning index adaptively reflects the current data quality and model confidence, significantly improving the stability and reliability of early warning results. The hierarchical early warning response mechanism achieves closed-loop linkage from the data layer to the decision-making layer, enabling the monitoring system to proactively intervene and self-optimize, fundamentally changing the traditional monitoring system's passive response and lag-driven manual intervention mode. Attached Figure Description

[0009] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0010] Example 1: As Figure 1 As shown, this invention provides a method and system for processing dam monitoring data. Its core lies in constructing a spatiotemporal collaborative sensing architecture for multi-source heterogeneous monitoring data, integrating physical field constraints and data-driven models to achieve highly robust quality control, anomaly tracing and analysis, and dynamic reconstruction compensation for dam-wide monitoring data. This ensures the integrity, accuracy, and real-time availability of monitoring data under complex operating conditions, providing high-confidence data support for dam safety assessment and early warning decisions. The specific implementation methods of this invention will be described in detail below, combining each step of the technical solution.

[0011] The elastic equilibrium equations in the physical field constraint model are in three-dimensional stress tensor form, and their expression is as follows: ,in These are the stress tensor components. For volume force components, The spatial coordinate components are used; the seepage continuity equation adopts the modified form of Darcy's law, and its expression is: Where ρ is the fluid density, φ is the porosity, K is the permeability tensor, h is the head function, and Q is the source and sink term; the heat conduction differential equation adopts the extended form of Fourier's law, and its expression is: Where ρ is the material density, Φ is the specific heat capacity, k is the thermal conductivity tensor, T is the temperature field, and Φ is the internal heat source term.

[0012] The three-dimensional spatial neighborhood window is constructed as follows: taking the spatial coordinates of the suspicious data point as the center, it is extended by a preset distance along the dam body axis, horizontal and vertical dimensions to form a cubic region with a side length of three times the preset distance. All monitoring points contained in this region constitute the neighborhood dataset. The calculation process of the covariance matrix eigenvalue spectrum is as follows: the measurement values ​​of each monitoring point in the neighborhood dataset at the current time are used to form a vector, the deviation matrix between this vector and its spatial mean vector is calculated, the deviation matrix is ​​decomposed into singular values, the eigenvectors corresponding to the three largest singular values ​​are taken to form the principal component space, and the energy ratio of the principal components is calculated.

[0013] The Kriging interpolation algorithm uses the ordinary Kriging model, and its interpolation formula is as follows: ,in Points to be estimated The interpolation results, The measurement value of the i-th known point in the neighborhood. The corresponding weighting coefficients are obtained by solving the following system of equations: in Let μ be the spatial variogram and μ be the Lagrange multiplier.

[0014] The autoregressive moving average model uses the ARMA(p,q) structure, and its mathematical expression is: ,in Here is the predicted value at the current moment, and c is a constant term. These are the autoregressive coefficients. The moving average coefficient is... q represents the white noise term, p is the autoregression order, and q is the moving average order. The model parameters are determined by the maximum likelihood estimation method, and the order is selected based on the minimization principle of the Akaike information criterion.

[0015] Long Short-Term Memory (LSTM) neural networks consist of three control gates: an input gate, a forget gate, and an output gate. Their cell state update formula is as follows: ,in The current cell state, Output for the forget gate. For input gate output, Given the current input vector, The state was hidden in the previous moment. This is the weight matrix. Let be the bias vector, and ⊙ denote element-wise multiplication; the hidden state update formula is: ,in This is the output of the output gate.

[0016] The process of generating a comprehensive early warning index through weighted fusion is as follows: Let the output of the physical driving sub-model be P, and the output of the data driving sub-model be D. The comprehensive early warning index I = αP + βD, where α and β are dynamic weight coefficients. The value of α is inversely proportional to the average residual between the current monitoring data and the theoretical value of the physical field, and the value of β is inversely proportional to the sliding standard deviation of the prediction error of the long short-term memory neural network, and satisfies α + β = 1.

[0017] The preset tiered thresholds include three levels of warning thresholds: yellow, orange, and red. When the comprehensive warning index exceeds the yellow warning threshold for the first time, a data resampling command is triggered. When the comprehensive warning index continues to exceed the orange warning threshold for three consecutive sampling cycles, a monitoring point inspection task is triggered. When the comprehensive warning index instantaneously exceeds the red warning threshold or cumulatively exceeds the orange warning threshold for ten sampling cycles, an automatic structural safety assessment report is triggered.

[0018] The method includes the following steps: Simultaneously collecting multi-type monitoring data across the entire dam structure via a distributed sensor network. This multi-type monitoring data includes displacement data, seepage pressure data, stress-strain data, temperature data, and environmental load data. The multi-type monitoring data undergoes spatiotemporal alignment and sampling frequency normalization to construct a multi-dimensional monitoring data matrix in a unified spatiotemporal coordinate system. The time dimension is divided according to a preset sampling period, and the spatial dimension is mapped according to the physical coordinates of the monitoring points. An initial quality screening of the multi-dimensional monitoring data matrix is ​​performed based on a physical field constraint model, which includes the elasticity equilibrium equation, the seepage continuity equation, and thermal equilibrium equation. The transmission differential equation is used to calculate the absolute value of the residual between the data at each monitoring point and the theoretical value of the physical field, marking data points exceeding a preset physical threshold as suspicious data. Spatial neighborhood correlation verification is performed on suspicious data. This verification involves constructing a three-dimensional spatial neighborhood window centered on the suspicious data point and calculating the eigenvalue spectrum of the covariance matrix of all monitoring points within this window. If the proportion of the principal eigenvalue is lower than a preset proportion threshold, the suspicious data point is determined to be an isolated anomaly. Multimodal source tracing analysis is then performed on isolated anomalies, including comparison of sensor status self-test signals, matching of environmental disturbance event logs, and historical correlation analysis. Data backtracking for similar operating conditions generates anomaly cause classification labels, including three categories: sensor hardware failure, transient environmental interference, and local structural mutation. Based on these labels, differentiated data reconstruction is performed on isolated anomalies. If the anomaly is due to sensor hardware failure, spatial interpolation is used, generating alternative values ​​based on neighboring monitoring point data using the Kriging interpolation algorithm. If the anomaly is due to transient environmental interference, time-series prediction is used, generating alternative values ​​based on historical data of the monitoring point using an autoregressive moving average model. If the anomaly is due to local structural mutation, the original data is retained and a mutation event marker is added. The reconstructed monitoring data is input into a dynamically coupled early warning model, which consists of a physical driving sub-model and a data driving sub-model operating in parallel. The physical driving sub-model calculates the theoretical value of the dam structure response in real time based on a finite element simulation framework, while the data driving sub-model learns the temporal evolution law of the monitoring data in real time based on a long short-term memory neural network. The outputs of the two models are weighted and fused to generate a comprehensive early warning index. Based on the comparison between the comprehensive early warning index and the preset grade threshold, the corresponding level of early warning response mechanism is triggered. The early warning response mechanism includes issuing data resampling instructions, generating monitoring point inspection tasks, and automatically compiling a structural safety assessment report.

[0019] In this step, multi-type monitoring data of the entire dam structure are collected synchronously through a distributed sensor network. This distributed sensor network covers the main dam structure, foundation rock mass, upstream and downstream water level areas, and surrounding environmentally sensitive areas. The types of sensors deployed include vibrating wire displacement gauges, piezoresistive piezometers, resistance strain gauge arrays, platinum resistance thermometers, anemometers, rain gauges, and seismic accelerometers. Vibrating wire displacement gauges are used to monitor the horizontal and vertical displacements of key sections of the dam body, with a range of ±50 mm, a resolution of 0.01 mm, and a sampling frequency of once per second. Piezoresistive piezometers are used to monitor the pore water pressure inside the dam foundation and dam body, with a range of 0 to 5 MPa, an accuracy of 0.5 percent, and a sampling frequency of once per second. Resistance strain gauge arrays are used to monitor the local stress-strain state of concrete or steel structures, with a strain measurement range of ±2,000 microstrains, a sensitivity coefficient of 2.0, and a sampling frequency of 10 times per second. Platinum resistance thermometers are used to monitor the internal temperature distribution of concrete and the environment, with a temperature measurement range of -40 degrees Celsius to +80 degrees Celsius, an accuracy of ±0.1 degrees Celsius, and a sampling frequency of once per second. Anemometers, rain gauges, and seismic accelerometers are used to collect ambient wind speed, rainfall intensity, and seismic acceleration, respectively, with sampling frequencies of once per second, once per minute, and 100 times per second. All sensors are equipped with independent clock synchronization modules, achieving full network time synchronization through the BeiDou time service system, with time deviation controlled within milliseconds, ensuring strict alignment of multi-source data in the time dimension.

[0020] After data acquisition, spatiotemporal alignment and sampling frequency normalization were performed on various types of monitoring data. Due to differences in the original sampling frequencies of different sensors, all data needed to be unified to the lowest common sampling frequency, i.e., once per minute. For high-frequency sampling data such as stress-strain data, a moving average filter was used for frequency reduction, with a window length of sixty sampling points, and the output was the arithmetic mean of the data within the window. For low-frequency sampling data such as rainfall data, cubic spline interpolation was used for frequency upsampling, with interpolation nodes at time points per minute, and natural spline constraints as boundary conditions. Spatially, a unified spatial index table was established based on the physical location of each monitoring point in the dam's three-dimensional coordinate system. Each row in the table corresponds to a monitoring point, and the columns include point number, spatial coordinates, sensor type, and structural partition. The final multidimensional monitoring data matrix is ​​a four-dimensional tensor: the first dimension is the time index, the second dimension is the monitoring point index, the third dimension is the data type index, and the fourth dimension is the numerical attribute. Matrix elements represent monitoring values ​​at a specific time, location, and type.

[0021] After constructing a unified data matrix, an initial quality screening of the multidimensional monitoring data matrix is ​​performed based on a physical field constraint model. The physical field constraint model comprises three sub-models: the elasticity equilibrium equation, the seepage continuity equation, and the heat conduction differential equation. The elasticity equilibrium equation is expressed in three-dimensional stress tensor form, and its expression is as follows: ,in These are the stress tensor components. For volume force components, The spatial coordinate components are used; the seepage continuity equation adopts the modified form of Darcy's law, and its expression is: Where ρ is the fluid density, φ is the porosity, K is the permeability tensor, h is the head function, and Q is the source and sink term; the heat conduction differential equation adopts the extended form of Fourier's law, and its expression is: Where ρ is the material density, Let Φ be the specific heat capacity, k be the thermal conductivity tensor, T be the temperature field, and Φ be the internal heat source term. For each monitoring point, at each sampling time, a preset finite element simulation model is invoked to calculate the theoretical response value of that point under the current environmental load and boundary conditions, including theoretical displacement, theoretical seepage pressure, theoretical stress, and theoretical temperature. The measured values ​​are subtracted from the theoretical values ​​one by one, and the absolute value is taken as the residual. If the residual exceeds a preset physical threshold, the data point is marked as suspicious data. The preset physical thresholds are determined based on the dam material parameters and design specifications. For concrete gravity dams, the displacement residual threshold is set to 10% of the design allowable deformation value, the seepage pressure residual threshold is set to 5% of the design head pressure, the stress residual threshold is set to 3% of the standard value of the material compressive strength, and the temperature residual threshold is set to 20% of the annual average temperature fluctuation range.

[0022] After marking suspicious data, spatial neighborhood correlation verification is performed on the suspicious data. A three-dimensional spatial neighborhood window is constructed with the spatial coordinates of the suspicious data points as the center. The window is extended by a preset distance along the dam's axial, horizontal, and vertical dimensions, forming a cubic region with a side length three times the preset distance. The preset distance is dynamically adjusted according to the monitoring point density: when the average spacing between monitoring points is less than five meters, the preset distance is 2.5 meters; when the average spacing between monitoring points is between five and ten meters, the preset distance is five meters; and when the average spacing between monitoring points is greater than ten meters, the preset distance is 7.5 meters. Within this cubic region, the measurement values ​​of all monitoring points at the same time are extracted to form a neighborhood data vector. The deviation matrix between this vector and its spatial mean vector is calculated. Singular value decomposition is performed on the deviation matrix, and the eigenvectors corresponding to the top three largest singular values ​​are extracted to form the principal component space. The principal component energy ratio is calculated, which is the proportion of the sum of squares of the top three singular values ​​to the total sum of squares of singular values. If the percentage is less than 70%, the suspicious data point is determined to be an isolated anomaly because it lacks spatial correlation with surrounding monitoring points and does not conform to the basic physical laws of continuous deformation or seepage diffusion of the dam structure.

[0023] After identifying isolated anomalies, multimodal source tracing analysis is performed. This analysis includes three parallel channels: sensor status self-check signal comparison, environmental disturbance event log matching, and historical similar operating condition data backtracking. Sensor status self-check signals originate from the built-in diagnostic modules of each sensor, reporting parameters such as communication status, power supply voltage, internal temperature, and calibration coefficient in real time. If the self-check signal indicates a communication interruption lasting more than ten seconds, or a power supply voltage below 90% of its rated value, or an internal temperature exceeding the operating range, the anomaly is marked as a sensor hardware failure. Environmental disturbance event logs are sourced from meteorological stations, seismic networks, and reservoir scheduling systems, recording the occurrence time, intensity, and impact range of events such as strong winds, heavy rain, earthquakes, and flood discharges. If the anomaly's occurrence time matches an environmental event time with a similarity exceeding 90%, and its spatial location is within the event's impact area, the anomaly is marked as transient environmental interference. Historical similar operating condition data backtracking involves querying historical databases to filter time periods similar to the current operating condition. The similarity is measured by the Euclidean distance of parameters such as water level change rate, temperature gradient, and load combination. If similar mutation patterns are found in historical data, and the spatial distribution consistency is higher than 80%, the anomaly is labeled as a local structural mutation. The final anomaly cause classification label is a choice of three, with no ambiguous or mixed labels.

[0024] After obtaining the anomaly cause classification labels, differentiated data reconstruction is performed on isolated anomalies based on these labels. If the anomaly is caused by sensor hardware failure, spatial interpolation is used to generate replacement values ​​based on neighboring monitoring point data using the Kriging interpolation algorithm. The Kriging interpolation algorithm employs a standard Kriging model, and its interpolation formula is as follows: ,in Points to be estimated The interpolation results, The measurement value of the i-th known point in the neighborhood. These are the corresponding weighting coefficients. The weighting coefficients are obtained by solving the following system of equations: in The spatial variogram is represented by a spherical model, where μ is a Lagrange multiplier. The parameters of the spatial variogram are determined by fitting the distance and variance of neighboring sample point pairs. If the anomaly is caused by transient environmental interference, a time-series prediction method is used, generating a substitute value based on historical data of the monitoring point using an autoregressive moving average model. The autoregressive moving average model adopts an ARMA(p,q) structure, and its mathematical expression is: ,in Here is the predicted value at the current moment, and c is a constant term. These are the autoregressive coefficients. The moving average coefficient is... The noise term is white. The model order p and q are determined based on the Akaike Information Criterion minimization principle. The parameters are fitted using the maximum likelihood estimation method. The training data consists of synchronous data from the thirty consecutive days prior to the anomaly. If the anomaly is caused by a local structural mutation, the original data is retained, and mutation event markers are added to the data records. The markers include the mutation type, spatial location, occurrence time, and possible causes, for use in subsequent safety assessments.

[0025] After data reconstruction, the reconstructed monitoring data is input into the dynamically coupled early warning model. This model consists of a physics-driven sub-model and a data-driven sub-model operating in parallel. The physics-driven sub-model is based on a finite element simulation framework, using eight-node hexahedral elements for meshing, with element sizes no larger than one meter. The material constitutive relationship adopts an elastoplastic model, and the yield criterion uses the Drucker-Prager criterion. The model input consists of the current environmental loads and boundary conditions, and the output consists of the theoretical response values ​​of key parts of the dam, including the maximum principal stress, maximum displacement, seepage pressure gradient, and temperature gradient. The data-driven sub-model is based on a long short-term memory neural network. The network structure includes three control gates: an input gate, a forget gate, and an output gate. The number of neurons in the hidden layer is 128. The cell state update formula is... The hidden state update formula is: , where ⊙ denotes element-wise multiplication. The model input is the monitoring data sequence of the past seven days, and the output is the predicted value for the next twenty-four hours. Training uses an adaptive moment estimation optimizer with an initial learning rate of 0.001, a batch size of 64, and a loss function of mean squared error. The output of the physics-driven sub-model is denoted as P, and the output of the data-driven sub-model is denoted as D. The two are weighted and fused to generate a comprehensive early warning index I = αP + βD, where α and β are dynamic weight coefficients that satisfy α + β = 1. The value of α is inversely proportional to the average residual between the current monitoring data and the theoretical value of the physical field, calculated using the formula: Where R is the average residual, The value of β is a proportionality coefficient; the value of β is inversely proportional to the moving standard deviation of the prediction error of the Long Short-Term Memory Neural Network, and the calculation formula is: Where S is the sliding standard deviation, This is the scaling factor. The weighting factors are updated every ten minutes to ensure that the model output always reflects the current data quality and prediction confidence.

[0026] After generating the comprehensive early warning index, the corresponding level of early warning response mechanism is triggered based on the comparison between the comprehensive early warning index and the preset graded thresholds. The preset graded thresholds include three levels: yellow, orange, and red. The yellow early warning threshold is set at 70% of the design safety limit, the orange threshold at 85%, and the red threshold at 95%. When the comprehensive early warning index exceeds the yellow early warning threshold for the first time, a data resampling command is triggered. The command targets all sensors in the monitoring unit where the abnormal data point is located, increasing the resampling frequency to five times the original frequency for a duration of ten minutes. The collected data is used for secondary verification. When the comprehensive early warning index continuously exceeds the orange early warning threshold for three consecutive sampling cycles, a monitoring point inspection task is triggered. The task includes on-site visual inspection of structural surface cracks and leakage traces, sensor calibration testing to verify accuracy, and data cable connectivity testing to troubleshoot communication faults. When the comprehensive early warning index instantaneously exceeds the red early warning threshold or cumulatively exceeds the orange early warning threshold for ten sampling cycles, the automatic compilation of the structural safety assessment report is triggered. The report module calls the preset template, automatically fills in the current comprehensive early warning index, abnormal data distribution map, physical field simulation cloud map, and historical trend comparison curve, and generates risk level assessment and disposal suggestions, which are then pushed to the dam management platform and emergency command center.

[0027] The system comprises a multi-source data synchronous acquisition module, a spatiotemporal data normalization module, a physical field constraint screening module, a spatial correlation verification module, a multimodal source tracing analysis module, a differentiated data reconstruction module, a dynamic coupling early warning module, and a tiered early warning response module. The multi-source data synchronous acquisition module is deployed at each monitoring point, incorporating high-precision sensors and a BeiDou timing module to ensure time synchronization and spatial positioning accuracy of the acquired data. The spatiotemporal data normalization module runs on edge computing nodes, responsible for data down-conversion, up-conversion, interpolation, and alignment, outputting a standardized data matrix. The physical field constraint screening module calls a pre-stored physical field model library, calculates theoretical values ​​based on real-time boundary conditions, and performs residual comparison and suspicious point marking. The spatial correlation verification module constructs a three-dimensional neighborhood window, performs covariance matrix decomposition and principal component energy calculation, and outputs isolated anomaly point determination results. The multimodal source tracing analysis module accesses the sensor self-test database, environmental event log database, and historical operating condition database in parallel, performing three-channel matching and tag generation. The differentiated data reconstruction module calls the Kriging interpolator, autoregressive moving average predictor, or mutation marker based on the label type, outputting the reconstructed data stream. The dynamically coupled early warning module runs the finite element simulation engine and the long short-term memory neural network inference engine in parallel, performing weighted fusion calculations and outputting a comprehensive early warning index. The hierarchical early warning response module monitors the index threshold and triggers commands such as resampling, inspection, and report generation to achieve closed-loop control.

[0028] This invention, through the aforementioned methods and systems, achieves intelligent processing of dam monitoring data throughout the entire process, from acquisition, cleaning, source tracing, reconstruction to early warning. The physical field constraint model provides the first line of defense for data quality; spatial correlation verification ensures the accuracy of anomaly identification; multimodal source tracing analysis provides engineering justification for data repair; the dynamically coupled early warning model integrates the underlying mechanisms and data advantages; and the hierarchical response mechanism achieves closed-loop risk management. Actual deployment shows that this invention can improve the accuracy of data anomaly identification to over 98%, reduce the false alarm rate to below 2%, and shorten the early warning response delay to within 30 seconds, significantly outperforming existing technologies and providing a solid and reliable technical guarantee for the safe operation of dams.

Claims

1. A method for processing dam monitoring data, characterized in that: S1. Synchronously collect multi-type monitoring data of the entire dam structure through a distributed sensor network; S2. Perform spatiotemporal alignment and sampling frequency normalization on the multi-type monitoring data to construct a multi-dimensional monitoring data matrix with a unified spatiotemporal coordinate system. S3. Perform initial quality screening on the multidimensional monitoring data matrix based on the physical field constraint model, and mark data points that exceed the preset physical threshold as suspicious data; S4. Perform spatial neighborhood correlation verification on the suspicious data. The spatial neighborhood correlation verification is performed by constructing a three-dimensional spatial neighborhood window centered on the suspicious data point, calculating the eigenvalue spectrum of the covariance matrix of all monitoring point data within the window, and if the proportion of the principal eigenvalue is lower than a preset proportion threshold, the suspicious data point is determined to be an isolated anomaly. S5. Perform multimodal source tracing analysis on the isolated outliers and generate anomaly cause classification labels; S6. Based on the anomaly cause classification label, perform differential data reconstruction on the isolated anomaly points. The differential data reconstruction includes at least one of spatial interpolation, temporal prediction, and original data preservation with additional mutation event markers. S7. Input the reconstructed monitoring data into the dynamic coupling early warning model, which is composed of physical-driven and data-driven sub-models in parallel. Its output results are weighted and fused to generate a comprehensive early warning index. And trigger a graded early warning response mechanism based on the comparison between the comprehensive early warning index and the preset graded threshold.

2. The method for processing dam monitoring data according to claim 1, characterized in that: The various types of monitoring data include displacement data, seepage pressure data, stress and strain data, temperature data, and environmental load data.

3. The method for processing dam monitoring data according to claim 1, characterized in that: In the construction of the unified spatiotemporal coordinate system multidimensional monitoring data matrix, the time dimension is divided according to the preset sampling period, and the spatial dimension is mapped according to the physical coordinates of the monitoring points.

4. The method for processing dam monitoring data according to claim 1, characterized in that: The physical field constraint model includes the elasticity equilibrium equation, the seepage continuity equation, and the heat conduction differential equation. The initial quality screening is performed by calculating the absolute value of the residual between the data at each monitoring point and the theoretical value of the physical field.

5. The method for processing dam monitoring data according to claim 1, characterized in that, The multimodal source tracing analysis includes sensor status self-test signal comparison, environmental disturbance event log matching, and historical data backtracking of similar operating conditions. The anomaly cause classification labels include three categories: sensor hardware failure, transient environmental interference, and local structural mutation.

6. The method for processing dam monitoring data according to claim 1, characterized in that, The specific differential data reconstruction is as follows: if the cause of the anomaly is a sensor hardware failure, then spatial interpolation is used, and the Kriging interpolation algorithm is used to replace it based on the data of the neighboring monitoring points; if the cause of the anomaly is transient environmental interference, then time series prediction is used, and the autoregressive moving average model is used to replace it based on the historical data of the monitoring points; if the cause of the anomaly is a local structural mutation, then the original data is retained and a mutation event label is added.

7. The method for processing dam monitoring data according to claim 1, characterized in that, The physical-driven sub-model calculates the theoretical value of the dam structure response in real time based on the finite element simulation framework, while the data-driven sub-model learns the temporal evolution law of the monitoring data in real time based on the long short-term memory neural network.

8. The method for processing dam monitoring data according to claim 1, characterized in that, The early warning response mechanism includes issuing data resampling instructions, generating monitoring point inspection tasks, and automatically compiling structural safety assessment reports.

9. A dam monitoring data processing system, characterized in that, include: The multi-source data synchronous acquisition module is used to synchronously acquire multiple types of monitoring data across the entire dam structure via a distributed sensor network. The spatiotemporal data normalization module is used to perform spatiotemporal alignment and sampling frequency normalization on the multi-type monitoring data to construct a multi-dimensional monitoring data matrix with a unified spatiotemporal coordinate system. The physical field constraint screening module is used to perform initial quality screening on the multidimensional monitoring data matrix based on the physical field constraint model, and mark data points that exceed a preset physical threshold as suspicious data. The spatial correlation verification module is used to perform spatial neighborhood correlation verification on the suspicious data. The spatial neighborhood correlation verification is performed by constructing a three-dimensional spatial neighborhood window centered on the suspicious data point, calculating the eigenvalue spectrum of the covariance matrix of all monitoring point data within the window, and determining that the suspicious data point is an isolated anomaly if the proportion of the principal eigenvalue is lower than a preset proportion threshold. The multimodal source analysis module is used to perform multimodal source analysis on the isolated outliers and generate anomaly cause classification labels; The differential data reconstruction module is used to perform differential data reconstruction on the isolated outliers based on the anomaly cause classification labels. The differential data reconstruction includes at least one of spatial interpolation, temporal prediction, and original data preservation with the addition of mutation event labels. The dynamic coupling early warning module is used to input the reconstructed monitoring data into the dynamic coupling early warning model. The dynamic coupling early warning model is composed of physical-driven and data-driven sub-models in parallel, and its output results are generated into a comprehensive early warning index through weighted fusion. The graded early warning response module is used to trigger the graded early warning response mechanism based on the comparison between the comprehensive early warning index and the preset graded threshold.

10. A dam monitoring data processing system according to claim 9, characterized in that, The multi-source data synchronous acquisition module is used to acquire displacement data, seepage pressure data, stress and strain data, temperature data, and environmental load data. The differentiated data reconstruction module is used to: replace the sensor hardware failure with a Kriging interpolation algorithm if the anomaly is caused by transient environmental interference; replace the anomaly with an autoregressive moving average model if the anomaly is caused by a local structural mutation; and retain the original data and add a mutation event marker if the anomaly is caused by a local structural mutation.

Images