AI prediction system for reservoir dam safety assessment

Abstract

The invention relates to the technical field of reservoir dam safety management, and discloses an AI prediction system for reservoir dam safety assessment, comprising: a monitoring data acquisition unit for receiving monitoring data streams representing seepage, displacement and stress states; the geometric space mapping unit is used for determining grid nodes according to the dam space grid model, establishing a space corresponding relation between the monitoring indexes and the grid nodes, and performing physical space-based geometric indexing on the monitoring indexes; the self-adaptive weight distribution module is used for extracting a feature vector according to the change rate of the monitoring index, and locking the weight change gradient of adjacent grid nodes by using an adjacent incidence matrix, so that the mutually connected grid nodes are synchronously corrected in weight calculation; according to the system and the method, time-space association between monitoring indexes and physical entities is established, local risks are amplified through a geometric structure connectivity trend, and the sensitivity of capturing dam body hidden dangers by the system is improved.

Description

Technical Field

[0001] This invention relates to an AI prediction system for reservoir dam safety assessment, belonging to the field of reservoir dam safety management technology. Background Technology

[0002] Currently, in reservoir dam safety assessments, data processing systems are used to integrate multi-source monitoring data and assist management decisions. The commonly used technical approach is based on an evaluation index weighting system. By collecting physical quantities from sensors such as water level, displacement, and seepage, the dam safety assessment level is calculated. This type of technical solution provides a standardized processing logic for dam risk management and has certain applicability in safety assessments under normal service environments. However, as the service life of water conservancy projects increases and extreme environmental loads fluctuate, the evolution of dam safety status exhibits complex nonlinear characteristics. Traditional solutions typically treat monitoring indicators such as seepage, stress, and displacement as independent numerical variables. This approach shows significant limitations under pressure conditions because it separates the monitoring data from the physical structural topology of the dam entity. This results in the assessment results failing to reflect the transmission path of risk within the dam structure, leading to a lack of spatial topological perception in the data processing process.

[0003] To address the aforementioned limitations, methods such as increasing sensor density or applying general statistical models cannot resolve the underlying contradictions. Increasing the number of sensors generates massive amounts of redundant data, increasing the system's computational load, and does not change the fact that monitoring indicators are logically isolated. General mathematical regression models lack physical geometric constraints, making it difficult to simulate the stress and deformation patterns of dams under specific three-dimensional boundary conditions. Evaluation results relying solely on mathematical fitting often deviate from engineering reality. When local deterioration occurs in the dam body, the system struggles to accurately capture its continuous impact on adjacent structures. For example, the Chinese invention with publication number CN121409344A... The patent application discloses a safety monitoring method and system for small reservoir dams. It comprehensively assesses risks from the dimensions of seepage flow, seepage pressure and surface deformation, and introduces an environmental risk impact coefficient correction. The core algorithm logic relies on the premise that the monitoring points are linearly independent or that the risk distribution has an ideal statistical distribution. The existing technology lacks the physical geometric constraints of the dam entity, and the assessment process is disconnected from the physical spatial topology. It cannot reflect the risk transmission path inside the three-dimensional structure. When the dam body undergoes non-ideal working conditions and local deterioration, it lacks the ability to perceive spatial connectivity, making it difficult to capture and predict the continuous impact of hidden dangers on adjacent structures, resulting in assessment lag or false alarms.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a data processing logic that anchors dynamic monitoring indicators to the grid nodes of the dam's three-dimensional geometric model, so that the calculation of evaluation weights is controlled by the spatial topological constraints of physical entities, and achieves accurate prediction of structural degradation trends based on grid connectivity. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: An AI prediction system for reservoir dam safety assessment, comprising:

[0006] The monitoring data acquisition unit is used to receive the real-time monitoring data stream of the dam. The real-time monitoring data stream includes multiple monitoring indicators that characterize the seepage state, displacement state, and stress state.

[0007] The geometric space mapping unit is used to determine multiple grid nodes based on the spatial grid model of the dam and to establish the spatial correspondence between each monitoring index and each grid node, so as to provide a physical space geometric index for the monitoring index.

[0008] The adaptive weight allocation module is used to extract feature vectors based on the rate of change of each monitoring indicator relative to the preset initial value, and to determine the weight of each monitoring indicator using a weighted algorithm. In the calculation process, an adjacency correlation matrix based on the physical space geometric index is introduced as a weight distribution constraint factor to limit the weight change gradient of adjacent grid nodes based on the connection state between grid nodes, so that the monitoring indicators corresponding to interconnected grid nodes are synchronously corrected during the weight calculation process.

[0009] The safety risk early warning module is used to calculate the safety quantitative score of the dam based on the calculated weights, and output risk early warning instructions according to the mapping result between the safety quantitative score and the preset risk range.

[0010] Preferably, when extracting feature vectors, the adaptive weight allocation module adjusts the priority of monitoring indicators in the weight allocation algorithm based on the rate of change at the current moment, and when the water level change or seepage pressure value exceeds the preset sensitivity threshold, it increases the corresponding component in the feature vector to improve the response speed of the safety risk warning module to risk features.

[0011] Preferably, the multiple monitoring indicators characterizing the seepage state, displacement state, and stress state include seepage pressure, crack opening and closing degree, dam surface displacement, internal stress, and steel reinforcement corrosion depth.

[0012] Preferably, the adaptive weight allocation module extracts the change characteristics of each monitoring indicator based on the following function: Y=(X(t) / X'-1)×100%, where Y is the rate of change, X(t) is the current monitoring value of the monitoring indicator, and X' is the preset initial value of the monitoring indicator.

[0013] Preferably, when determining the grid nodes, the geometric space mapping unit divides the geometric surface and internal structure of the dam into interconnected grid topology units, and assigns a unique spatial coordinate code to each grid node as a physical space geometric index.

[0014] Preferably, the adaptive weight allocation module uses an adjacency correlation matrix to constrain the monitoring indicators that have undergone abrupt changes. When the monitoring indicators of a specific grid node fluctuate abnormally, the module synchronously increases the proportion of adjacent regions in the weight allocation based on spatial correspondence, so as to amplify the trend of local risk characteristics through the connectivity of the geometric structure.

[0015] Preferably, the evaluation model constructed by the safety risk early warning module includes four management dimensions: on-site inspection, monitoring and analysis, flood control capacity, and numerical model analysis. The safety risk early warning module generates a predicted value of the dam's safety status during its subsequent service life by logically arranging the quantitative indicators of the four management dimensions.

[0016] Preferably, the preset risk range includes an extremely high risk range, a high risk range, a medium risk range, and a low risk range. The safety risk warning module triggers the corresponding level of risk control logic based on the preset risk range in which the safety quantitative score is located.

[0017] Preferably, the monitoring data acquisition unit is also used to acquire meteorological monitoring data and water level fluctuation data of the environment where the dam is located, and input the meteorological monitoring data and water level fluctuation data as external constraint variables into the safety risk early warning module; the safety risk early warning module performs weighted offset correction on the safety quantitative score based on the meteorological monitoring data and water level fluctuation data, so as to eliminate the interference of environmental noise on the safety status assessment.

[0018] Preferably, it also includes a self-calibration feedback unit, which is used to compare the predicted safety status value output by the safety risk early warning module with the actual service status value of the dam in the future, and adjust the proportional coefficient of the weight distribution constraint factor in the adaptive weight allocation module in reverse according to the generated comparison residual, so as to realize the dynamic correction of the weight calculation logic.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. In the safety assessment of reservoir dams, the Delaunay triangulation algorithm is used to discretize the computer-aided design three-dimensional geometric model into a continuous finite geometric topological mesh, realizing the deep integration of the spatial attributes of monitoring data and the logical processing system. This mechanism anchors the extracted dynamic index vector to the nearest geometric topological mesh node based on the spatial coordinates of the physical sensors, constructing a three-dimensional topological attribute model with dynamic monitoring data. This transforms the originally isolated discrete monitoring values ​​into a structured data stream that carries the spatial correlation of physical entities. When optimizing the consistency ratio, the topological adjacency matrix representing spatial connectivity is introduced as a constraint term, and the adjustment path of the weights is limited by the boundary conditions of the geometric model. This effectively avoids the logical distortion caused by deviating from the physical topological laws in traditional mathematical iteration, and ensures that the prediction results are highly consistent with the physical deterioration path of the dam entity.

[0021] 2. By using an index classification function to calculate the rate of change of dam safety evaluation parameters in real time, a scientific decoupling of static and dynamic indicators is achieved. This mechanism identifies the degree of deviation of the current performance value from the initial performance value and dynamically extracts feature vectors such as seepage, displacement, and stress during the data preprocessing stage. This processing method can significantly reduce the information redundancy of the system when processing large-scale monitoring data, and automatically shift the focus of data processing resources towards indicators with obvious performance degradation characteristics. This improves the sensitivity of the prediction system to capturing instantaneous risks under extreme conditions such as sudden rise in water level or abnormal seepage.

[0022] 3. Based on the adaptive weight adjustment mechanism of the improved analytic hierarchy process and the synergistic gain of the topological penalty factor, a weight optimization closed loop with physical constraints is constructed. When the data of a certain grid node in the three-dimensional topological attribute model is distorted, the system locks the lower limit of the weight difference between adjacent nodes based on the topological adjacency matrix, and forcibly increases the correlation weight of the physically connected area in the evaluation system. This makes local small displacements or leakage hazards no longer regarded as isolated numerical fluctuations, but amplified in the prediction through the connectivity of the geometric structure. This achieves a leap from experience-based numerical fitting to structured prediction based on geometric topology, and enhances the stability of the system in dealing with complex nonlinear risk evolution. Attached Figure Description

[0023] Figure 1 A diagram showing the components and dynamic weight allocation architecture of an AI prediction system;

[0024] Figure 2 This is a logic diagram for the coordinated early warning of physical space constraints and environmental factors.

[0025] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0027] An AI prediction system for reservoir dam safety assessment includes:

[0028] The monitoring data acquisition unit is used to receive the real-time monitoring data stream of the dam. The real-time monitoring data stream includes multiple monitoring indicators that characterize the seepage state, displacement state, and stress state.

[0029] The geometric space mapping unit is used to determine multiple grid nodes based on the spatial grid model of the dam and to establish the spatial correspondence between each monitoring index and each grid node, so as to provide a physical space geometric index for the monitoring index.

[0030] The adaptive weight allocation module is used to extract feature vectors based on the rate of change of each monitoring indicator relative to the preset initial value, and to determine the weight of each monitoring indicator using a weighted algorithm. In the calculation process, an adjacency correlation matrix based on the physical space geometric index is introduced as a weight distribution constraint factor to limit the weight change gradient of adjacent grid nodes based on the connection state between grid nodes, so that the monitoring indicators corresponding to interconnected grid nodes are synchronously corrected during the weight calculation process.

[0031] The safety risk early warning module is used to calculate the safety quantitative score of the dam based on the calculated weights, and output risk early warning instructions according to the mapping result between the safety quantitative score and the preset risk range.

[0032] Preferably, when extracting feature vectors, the adaptive weight allocation module adjusts the priority of monitoring indicators in the weight allocation algorithm based on the rate of change at the current moment, and when the water level change or seepage pressure value exceeds the preset sensitivity threshold, it increases the corresponding component in the feature vector to improve the response speed of the safety risk warning module to risk features.

[0033] Preferably, the multiple monitoring indicators characterizing the seepage state, displacement state, and stress state include seepage pressure, crack opening and closing degree, dam surface displacement, internal stress, and steel reinforcement corrosion depth.

[0034] Preferably, the adaptive weight allocation module extracts the change characteristics of each monitoring indicator based on the following function: Y=(X(t) / X'-1)×100%, where Y is the rate of change, X(t) is the current monitoring value of the monitoring indicator, and X' is the preset initial value of the monitoring indicator.

[0035] Preferably, when determining the grid nodes, the geometric space mapping unit divides the geometric surface and internal structure of the dam into interconnected grid topology units, and assigns a unique spatial coordinate code to each grid node as a physical space geometric index.

[0036] Preferably, the adaptive weight allocation module uses an adjacency correlation matrix to constrain the monitoring indicators that have undergone abrupt changes. When the monitoring indicators of a specific grid node fluctuate abnormally, the module synchronously increases the proportion of adjacent regions in the weight allocation based on spatial correspondence, so as to amplify the trend of local risk characteristics through the connectivity of the geometric structure.

[0037] Preferably, the evaluation model constructed by the safety risk early warning module includes four management dimensions: on-site inspection, monitoring and analysis, flood control capacity, and numerical model analysis. The safety risk early warning module generates a predicted value of the dam's safety status during its subsequent service life by logically arranging the quantitative indicators of the four management dimensions.

[0038] Preferably, the preset risk range includes an extremely high risk range, a high risk range, a medium risk range, and a low risk range. The safety risk warning module triggers the corresponding level of risk control logic based on the preset risk range in which the safety quantitative score is located.

[0039] Preferably, the monitoring data acquisition unit is also used to acquire meteorological monitoring data and water level fluctuation data of the environment where the dam is located, and input the meteorological monitoring data and water level fluctuation data as external constraint variables into the safety risk early warning module; the safety risk early warning module performs weighted offset correction on the safety quantitative score based on the meteorological monitoring data and water level fluctuation data, so as to eliminate the interference of environmental noise on the safety status assessment.

[0040] Preferably, it also includes a self-calibration feedback unit, which is used to compare the predicted safety status value output by the safety risk early warning module with the actual service status value of the dam in the future, and adjust the proportional coefficient of the weight distribution constraint factor in the adaptive weight allocation module in reverse according to the generated comparison residual, so as to realize the dynamic correction of the weight calculation logic.

[0041] Example 1: When the system faces a load condition of continuous rainfall causing the reservoir water level to rise, the stress distribution and seepage field inside the dam's physical entity undergo dynamic evolution. Traditional monitoring systems treat seepage pressure and dam surface displacement as isolated numerical variables. Relying on a fixed scoring matrix, they cannot perform real-time adjustments based on the dynamic fluctuations of environmental loads. This makes it impossible for the data processing system to accurately calculate the spread trend of potential hazards to adjacent load-bearing units based on structural connectivity when faced with local displacement or seepage risks at the dam abutment, resulting in topological blind spots in data assessment and delayed early warning. The monitoring data acquisition unit continuously receives real-time monitoring data streams of multiple monitoring indicators characterizing the dam's current service status, such as seepage pressure and dam surface displacement. Simultaneously, the geometric space mapping unit maps the dam's spatial topology based on the dam's... The inter-grid model determines multiple grid nodes and establishes a spatial correspondence between each monitoring indicator and each grid node, using a unique physical spatial geometric index as the basis for the monitoring indicators. Based on this, the adaptive weight allocation module extracts the change characteristics of each monitoring indicator relative to the preset initial value using the function Y=(X(t) / X'-1)×100%, where Y is the rate of change, X(t) is the current monitoring value of the monitoring indicator, and X' is the preset initial value of the monitoring indicator. When the absolute value of the rate of change of the seepage pressure index at a specific grid node exceeds the preset sensitivity threshold, the adaptive weight allocation module introduces the adjacency correlation matrix established based on the physical spatial geometric index as a weight distribution constraint factor during the calculation process of determining the weight of each monitoring indicator using a weighted algorithm.

[0042] The weight change gradient of adjacent grid nodes is limited by the connection state between grid nodes. Specifically, during the algorithm's iterative optimization, the lower limit of the difference between the weight components of grid nodes with geometric topological adjacency is locked. This ensures that the monitoring indicators of interconnected grid nodes are synchronously corrected during weight calculation. This processing logic integrates the performance degradation rate based on time series with the three-dimensional spatial connectivity constraint, allowing local single-point data anomalies to directly increase the priority and proportion of adjacent regions in weight allocation through the adjacency correlation matrix. This resolves the technical contradiction between discrete sensor values ​​and the physical transmission laws of continuous physical media. As adjacent grid nodes complete their weight allocation in the weight allocation logic... The spatial trend is magnified and updated. The safety risk early warning module extracts meteorological monitoring data and water level fluctuation data of the dam's environment as external constraint variables. Based on the meteorological monitoring data and water level fluctuation data, weighted offset correction is performed on various quantitative indicators to reduce the interference of environmental noise on the safety status assessment. Based on the determined weights, the safety quantitative score of the dam is calculated. According to the mapping relationship between the safety quantitative score and the preset risk interval, the current state is compared with the extremely high risk interval or high risk interval level. Finally, a risk early warning command carrying specific physical spatial geometric index attributes is output. This execution process transforms isolated data monitoring into risk evolution numerical inference under specific three-dimensional boundary conditions by relying on grid topology constraints.

[0043] Example 2: When the system is deployed on a multi-field coupled physical simulation test platform, traditional isolated index evaluation methods are prone to outputting deviation warning signals when faced with sensor baseline drift and dynamic environmental interference. The experimental data comes from a computational fluid dynamics and finite element analysis joint solution platform. A three-dimensional fluid-structure interaction core model of the dam is constructed based on Darcy's law and Navier-Stokes equations. To verify the anti-interference capability of the scheme in an industrial electromagnetic environment, Gaussian white noise with a signal-to-noise ratio of 20dB is actively superimposed on the original monitoring data stream of seepage pressure and dam surface displacement, and a low-frequency environmental drift disturbance with an amplitude of 5% of the full scale is introduced. The preset sensitivity threshold needs to achieve an optimal balance between anomaly capture real-time performance and global computational load. The system sets the bottom sampling frequency of the physical sensors to 10Hz and constructs a FIFO circular buffer with a depth of 600 sampling points in memory as a sliding time window, i.e., covering continuous The monitoring time domain is extended to 60 seconds; the window movement step size is set to 100 sampling points, that is, global feature vector extraction is triggered every 10 seconds; before extracting the rate of change feature, the system performs median filtering on the 600 raw data points in the window to remove pulse interference noise with amplitude deviation exceeding 20% ​​of the average of the previous cycle, ensuring that the feature vector can accurately characterize the low-frequency structural strain trend of the dam. When the spectral bandwidth of the monitored signal is wide and the variance of the environmental white noise is at a high level, a low threshold will cause the system to misjudge high-frequency noise as local structural damage, causing invalid high-frequency iteration of the grid node weight matrix; a high threshold will cause the small abrupt changes in the initial stage of the seepage channel to be masked by the background. The system extracts the variance of the environmental white noise in the historical rainfall cycle and sets the preset sensitivity threshold to three times the variance value, so as to filter out 99.7% of random physical disturbances and simultaneously capture the real structural strain signal under this quantization rule.

[0044] The experiment constructed a problem intensity gradient control system, setting the superimposed low-frequency environmental drift disturbance amplitude to three increasing levels of 2%, 5%, and 8% of the full scale. The experiment included a control group using a traditional fixed scoring matrix without introducing an adjacency correlation matrix, and an experimental group fully applying the adaptive weight allocation logic of this invention. When simulated heavy rainfall was applied for 12 hours, the raw data stream showed that the seepage pressure monitoring value of a specific grid node located in the dam abutment fissure development zone fluctuated randomly from an initial 0.15 MPa to the range of 0.18 MPa to 0.22 MPa. In the control group, affected by 20 dB Gaussian white noise, the single-point abnormal change rate Y of this specific grid node reached 25.3% under a 2% disturbance amplitude, where Y is the change rate, X(t) is the current monitoring value of the monitoring index, and X' is the preset initial value of the monitoring index. The system of the control group... The weight allocation coefficient of the grid node was independently increased from the initial 0.05 to 0.12, while the weight of the surrounding grid nodes physically adjacent to it remained at 0.05. This caused the local safety quantification score to show a sharp drop within a 10-minute time window. In the experimental group, the adaptive weight allocation module detected that the absolute value of the rate of change of the seepage pressure index of the specific grid node exceeded the preset sensitivity threshold. Subsequently, the adjacency correlation matrix based on the physical space geometric index was introduced as the weight distribution constraint factor. The output data showed that the experimental group locked the lower limit of the difference of the weight components of the grid nodes with geometric topological adjacency to 0.02 during the algorithm calculation process. This forced the four adjacent grid nodes around the specific grid node to complete the weight correction synchronously, so that the weight of the central grid node smoothly transitioned to 0.09, and the weights of the adjacent grid nodes climbed synchronously to 0.07.

[0045] As the perturbation amplitude increased from 2% to 5% and then to 8%, the weight distribution gradient of adjacent grid nodes in the comparison sample group showed a disordered discrete distribution, with the maximum weight range reaching 0.18. In the experimental group, under a 5% perturbation, the weight of the central grid node was 0.11 and that of adjacent nodes was 0.09. Under an 8% perturbation, the weight of the central node climbed to 0.14 and that of adjacent nodes reached 0.12 simultaneously. The monitoring data showed obvious nonlinear physical constraint characteristics. When the perturbation amplitude exceeded the 6.5% critical point of full scale, the weight difference between adjacent grid nodes in the experimental group no longer increased linearly with the perturbation amplitude, but was controlled by the preset lower limit of the difference and converged to the limit boundary. This indicates that the grid topology constraint mechanism of the system entered a rigid saturation defense state of weight distribution under extreme environmental noise, blocking the destruction of the global evaluation matrix by single-point pseudo-extreme values. The safety risk early warning module extracts synchronously acquired meteorological monitoring data and water level fluctuation data, and applies weighted offset corrections to various quantitative indicators to reduce environmental noise interference. Under the 8% extreme disturbance condition, the dam safety quantitative score calculated by the comparative sample group dropped sharply from 92 points to 71 points, outputting an incorrect high-risk interval level early warning instruction. The experimental group, based on the weights corrected by spatial geometric topology, calculated a stable dam safety quantitative score between 85 and 88 points during heavy rainfall, accurately positioning it in the medium-risk interval. The experimental results prove that this invention, by introducing physical spatial geometric indexes and adjacency correlation matrices, transforms the isolated sensor time series evolution into three-dimensional spatial connectivity collaborative inference, eliminating the prediction bias between dynamic environmental noise and the actual structural deterioration trend.

[0046] Example 3: When the system operates in a complex hydrological environment, the monitoring data acquisition unit receives real-time monitoring data streams from the dam. These streams include multiple monitoring indicators characterizing seepage, displacement, and stress states. The geometric space mapping unit determines multiple grid nodes based on the dam's spatial grid model. When establishing the spatial correspondence between each monitoring indicator and each grid node, this unit uses the three-dimensional coordinates of each grid node as the criterion, extracting the monitoring indicator output by the physical sensor with the smallest spatial geometric distance as the state characteristic value of that specific grid node, thus using the physical spatial geometric index as the basis for the monitoring indicators. The adaptive weight allocation module introduces an adjacency correlation matrix established based on the physical spatial geometric index as a weight distribution constraint factor. For any two grid nodes, it identifies whether there is a shared geometric boundary in the spatial grid model. If a corresponding boundary exists, the weight allocation is adjusted accordingly. The element at the corresponding position in the matrix is ​​assigned a value of 1, and a value of 0 is assigned when there is no direct entity connection. When determining the weight of each monitoring indicator using a weighted algorithm, the adaptive weight allocation module calculates the single-point anomaly change rate for a specific grid node and generates an initial update weight. This module extracts the adjacent grid nodes with a value of 1 in the adjacency association matrix, calculates the absolute value of the difference between the initial update weight of the specific grid node and the current weight of each adjacent grid node. If the absolute value of the difference is greater than the preset lower limit of the difference, the adaptive weight allocation module distributes the excess weight difference equally according to the total number of matrix elements with a value of 1, deducts the initial update weight of the specific grid node and adds it equally to the corresponding adjacent grid nodes. This limits the weight change gradient of adjacent grid nodes based on the connection state between grid nodes, so that the monitoring indicators corresponding to interconnected grid nodes are synchronously corrected during the weight calculation process.

[0047] The safety risk early warning module uses a fixed-duration sliding window to extract the current cumulative rainfall and reservoir water level rise rate of the dam's environment as meteorological monitoring data and water level fluctuation data. It also retrieves historical hydrological and meteorological archives of a specific watershed and extracts the maximum single-day rainfall, setting it as the historical extreme rainfall. Get the cumulative rainfall for the current period. With the rate of rise of reservoir water level As an environmental constraint variable, calculate and The ratio determines the dimensionless environmental impact coefficient η, where η characterizes the contribution of external load fluctuations to the seepage and displacement indices collected by the sensors, and the current weights of the grid nodes are used to determine this coefficient. The product of the coefficient η and the coefficient η serves as a weighted offsetting correction term. ,Right now Subtract the correction term from the evaluation input value. The system generates a score and uses linear compensation logic to filter out disturbances in shallow surface physical quantities caused by rainfall runoff. This allows the safety quantification score calculation logic to focus on the nonlinear characteristic response caused by the deterioration of the dam's internal structure. The module divides the current cumulative rainfall by the pre-stored historical extreme rainfall to calculate a dimensionless environmental impact coefficient. The product of the current weight and this environmental impact coefficient is used as a weighted offset correction term. The safety risk warning module subtracts the weighted offset correction term from the current weight of a specific grid node to generate the final weight, reducing the disturbance of fluctuations in shallow surface non-structural indicators caused by surface rainfall runoff. Based on the calculated weights, the safety quantification score of the dam is calculated. The safety risk warning module outputs a risk warning command based on the mapping result between the safety quantification score and the preset risk interval.

[0048] Example 4: When the system has completed its initial deployment at a specific engineering site and has not yet been connected to the real-time business data stream, the monitoring data acquisition unit extracts the continuous static no-load data stream of the specific dam under standard normal load as the calibration input. The geometric space mapping unit processes the sensor reading time series at each grid node through the mean filtering algorithm and outputs the statistical steady-state baseline matrix. The values ​​in this matrix are used as the preset initial values ​​X' when the adaptive weight allocation module extracts the rate of change features. At the same time, the safety risk early warning module retrieves the historical hydrological and meteorological archives of the specific watershed and extracts the maximum daily rainfall, which is written into the local read-only memory to form the historical limit rainfall parameters required to calculate the environmental impact coefficient. This offline calibration process establishes the underlying parameter benchmark corresponding to the environmental attributes of the specific physical entity.

[0049] In the parallel debugging process of establishing the underlying parameter benchmark, the preset risk interval mapping logic in the safety risk early warning module uses an offline finite element simulation model to complete the boundary threshold filling. The model applies a virtual failure load exceeding the conventional bearing capacity in the spatial grid model and simultaneously extracts the physical characteristic parameters when the grid reaches the local yield critical point. The system sets the simulated safety quantification score corresponding to the yield critical point as the dividing boundary between the medium risk interval and the high risk interval. This data filling procedure anchors the generated risk assessment scale to the physical attenuation evolution law of the physical structure and constructs an objective judgment basis for the assessment interval in the calculation path of converting multi-source monitoring indicators into risk early warning instructions.

[0050] Example 5: When the system performs long-term continuous online evaluation in a complex and ever-changing geological environment, if the weighting algorithm inside the adaptive weight allocation module lacks a quantitative consistency self-check and dynamic correction procedure for the judgment matrix, single-point data jumps caused by sudden environmental changes can easily lead to the global weight allocation matrix deviating from engineering common sense. To establish the logical closed loop of the weighting algorithm, the adaptive weight allocation module constructs a judgment matrix for pairwise comparison of importance when it receives updated monitoring indicators. The system calculates the maximum eigenvalue of this judgment matrix and uses the formula... Calculate the consistency index, where CI is the consistency index and To determine the largest eigenvalue of the judgment matrix and n as the order of the judgment matrix, the system retrieves the pre-stored average random consistency index and marks it as a constant RI. The consistency ratio is calculated using the formula CR=CI / RI, where CR is the consistency ratio. The system then compares the calculated consistency ratio with a preset consistency threshold set to 0.1. If the consistency ratio is less than the preset consistency threshold, the system determines that the logical structure of the current judgment matrix conforms to physical laws and confirms the currently calculated weights.

[0051] When the consistency ratio is greater than or equal to a preset consistency threshold, the adaptive weight allocation module triggers an online fault tolerance procedure. The system extracts the abnormal monitoring indicators corresponding to the specific grid nodes that cause logical conflicts in the judgment matrix, and introduces a preset damping coefficient to perform a forced decay operation on the initial update weight of the specific grid node. This preset damping coefficient is obtained by performing 10 sets of cyclic load simulation tests in increments of 0.05 from 1.00 to 0.50 during the static calibration phase of system power-on initialization, and is finally calibrated to 0.75 based on the convergence speed. Simultaneously, it is used to trigger the fault tolerance procedure. The consistency threshold of 0.10 is determined based on the normal service fluctuation data of a specific dam over 365 days. It is calculated by determining the maximum residual deviation under the 99.7% confidence interval. This is used as the physical boundary for judging whether a logical conflict has occurred in the sensor group. The judgment matrix is ​​reconstructed and a new round of iterative calculation is started until the consistency ratio falls back below the preset consistency threshold. This self-checking and fault-tolerant procedure establishes the operation boundary of the weighted algorithm under extreme conditions by embedding feedback control of the consistency ratio, ensuring the rigor of the dam safety quantitative scoring calculation path and the stability of the output results.

[0052] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An AI prediction system for reservoir dam safety assessment, characterized in that, include: The monitoring data acquisition unit is used to receive the real-time monitoring data stream of the dam. The real-time monitoring data stream includes multiple monitoring indicators that characterize the seepage state, displacement state, and stress state. The geometric space mapping unit is used to determine multiple grid nodes based on the spatial grid model of the dam and to establish the spatial correspondence between each monitoring index and each grid node, so as to provide a physical space geometric index for the monitoring index. The adaptive weight allocation module is used to extract feature vectors based on the rate of change of each monitoring indicator relative to the preset initial value, and to determine the weight of each monitoring indicator using a weighted algorithm. In the calculation process, an adjacency correlation matrix based on the physical space geometric index is introduced as a weight distribution constraint factor to limit the weight change gradient of adjacent grid nodes based on the connection state between grid nodes, so that the monitoring indicators corresponding to interconnected grid nodes are synchronously corrected during the weight calculation process. The safety risk early warning module is used to calculate the safety quantitative score of the dam based on the calculated weights, and output risk early warning instructions according to the mapping result between the safety quantitative score and the preset risk range.

2. The AI ​​prediction system for reservoir dam safety assessment according to claim 1, characterized in that, When extracting feature vectors, the adaptive weight allocation module adjusts the priority of monitoring indicators in the weight allocation algorithm based on the rate of change at the current moment. When the water level change or seepage pressure value exceeds the preset sensitivity threshold, the module increases the corresponding component in the feature vector to improve the response speed of the safety risk warning module to risk features.

3. The AI ​​prediction system for reservoir dam safety assessment according to claim 1, characterized in that, Multiple monitoring indicators characterizing seepage state, displacement state, and stress state include seepage pressure, crack opening and closing degree, dam surface displacement, internal stress, and steel corrosion depth.

4. The AI ​​prediction system for reservoir dam safety assessment according to claim 1, characterized in that, The adaptive weight allocation module extracts the change characteristics of each monitoring indicator based on the following function: Y=(X(t) / X'-1)×100%, where Y is the rate of change, X(t) is the current monitoring value of the monitoring indicator, and X' is the preset initial value of the monitoring indicator.

5. The AI ​​prediction system for reservoir dam safety assessment according to claim 1, characterized in that, When determining the grid nodes, the geometric space mapping unit divides the geometric surface and internal structure of the dam into interconnected grid topology units and assigns a unique spatial coordinate code to each grid node as a physical space geometric index.

6. The AI ​​prediction system for reservoir dam safety assessment according to claim 1, characterized in that, The adaptive weight allocation module uses an adjacency correlation matrix to constrain the monitoring indicators that have undergone abrupt changes. When the monitoring indicators of a specific grid node fluctuate abnormally, it synchronously increases the proportion of adjacent regions in the weight allocation based on spatial correspondence, so as to amplify the trend of local risk characteristics through the connectivity of geometric structure.

7. The AI ​​prediction system for reservoir dam safety assessment according to claim 1, characterized in that, The evaluation model constructed by the safety risk early warning module includes four management dimensions: on-site inspection, monitoring and analysis, flood control capacity, and numerical model analysis. The safety risk early warning module generates the predicted value of the dam's safety status during its subsequent service life by logically arranging the quantitative indicators of the four management dimensions.

8. The AI ​​prediction system for reservoir dam safety assessment according to claim 1, characterized in that, The preset risk ranges include extremely high risk range, high risk range, medium risk range and low risk range. The security risk warning module triggers the corresponding level of risk control logic based on the preset risk range in which the security quantitative score is located.

9. The AI ​​prediction system for reservoir dam safety assessment according to claim 1, characterized in that, The monitoring data acquisition unit is also used to acquire meteorological monitoring data and water level fluctuation data of the environment where the dam is located, and input the meteorological monitoring data and water level fluctuation data as external constraint variables into the safety risk early warning module; The safety risk early warning module performs a weighted offset correction on the safety quantitative score based on meteorological monitoring data and water level fluctuation data to eliminate the interference of environmental noise on the safety status assessment.

10. The AI ​​prediction system for reservoir dam safety assessment according to claim 1, characterized in that, It also includes a self-calibration feedback unit, which compares the predicted safety status value output by the safety risk early warning module with the actual service status value of the dam, and adjusts the proportional coefficient of the weight distribution constraint factor in the adaptive weight allocation module in reverse according to the generated comparison residual.

Images