An Adaptive Identification Method for Abnormal Deformation of Dam Slope Based on Dynamic Graph Neural Network

Abstract

This application relates to an adaptive identification method for abnormal deformation of dam slopes based on a dynamic graph neural network. This method collects deformation monitoring data from multiple monitoring points on the dam slope, performs time alignment and preprocessing on the multi-source monitoring data, and constructs multi-time-step features reflecting the deformation evolution characteristics of the monitoring points. The monitoring points are modeled as graph nodes, and a dynamically updated monitoring graph is constructed based on the spatial relationships, geological zoning relationships, and deformation correlations between the monitoring points. The dynamic graph neural network is used to jointly learn the spatiotemporal features of the monitoring data to construct a normal deformation feature space, and the deviation of the current monitoring state from the normal state is calculated. An adaptive anomaly identification threshold is then used to identify abnormal deformation. This method can adapt to changes in the monitoring network structure, improve the continuity and reliability of abnormal deformation identification, and is suitable for long-term safety monitoring and early warning of dam slopes.

Description

Technical Field

[0001] This application relates to the field of safety monitoring and intelligent analysis technology for water conservancy projects, and in particular to an adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks. Background Technology

[0002] The slope of a reservoir dam is a crucial component ensuring the safe operation of water conservancy projects, and its stability directly affects the structural safety of the dam and the safety of life and property in the downstream area. During the long-term operation of the dam, slope deformation is influenced by a variety of factors, including water level fluctuations, rainfall infiltration, reservoir bank geological structure, construction disturbances, and environmental changes, often exhibiting strong time-varying characteristics, significant nonlinearity, and complex spatial correlations. Therefore, how to continuously monitor the deformation state of dam slopes and promptly identify abnormal deformation behavior has always been an important research topic in the field of water conservancy project safety monitoring.

[0003] Currently, the main technologies for identifying abnormal deformation of dam slopes include the following categories:

[0004] One type of discrimination method is based on empirical thresholds or rules, which determines abnormal states by setting fixed thresholds for monitoring indicators such as displacement, velocity, or acceleration. This type of method is simple to implement and has low computational cost, but the thresholds often depend on engineering experience and are difficult to adapt to the dynamic changes in deformation characteristics under different working conditions, which can easily lead to false alarms or missed alarms.

[0005] Another type of method is based on statistical analysis or traditional time series models, such as regression analysis and ARIMA models. These methods establish single-point deformation evolution models using historical data for anomaly detection. While these methods are effective in handling linear or weakly nonlinear problems, they typically only model a single monitoring point and struggle to depict the spatial relationships between different monitoring points, thus limiting their ability to identify overall deformation patterns in complex slopes.

[0006] As the scale of monitoring data continues to grow, anomaly detection methods based on machine learning and deep learning are gradually being applied, such as Long Short-Term Memory (LSTM) networks, Convolutional Neural Networks (CNNs), and their combined models. These methods can automatically learn deformation features from large amounts of historical data, improving prediction and recognition accuracy to some extent. However, existing deep learning methods are usually modeled based on fixed input structures or static network structures, assuming that the number of monitoring points and spatial relationships remain unchanged.

[0007] In practical engineering, dam slope monitoring systems often face the following situations during long-term operation:

[0008] First, some monitoring points have become ineffective due to equipment aging, construction interference, or communication failures.

[0009] Second, new monitoring points are constantly being added as engineering needs change;

[0010] Third, changes in local geological conditions may alter the relationships between measuring points.

[0011] In the above situations, existing static model-based methods often require rebuilding the model or retraining the entire system, making it difficult to guarantee the continuity and real-time nature of monitoring and analysis, and limiting their engineering applicability.

[0012] In addition, most existing anomaly identification methods focus on time-dimensional analysis, paying insufficient attention to the spatial topological relationships between monitoring points and their dynamic characteristics over time, making it difficult to effectively reflect the spatial propagation characteristics and associated evolution patterns of slope deformation.

[0013] Therefore, existing technologies still have shortcomings in dynamically adapting to changes in network structure, comprehensively utilizing spatial correlation and temporal evolution information, and achieving continuous and stable identification of abnormal deformations, and further improvements are urgently needed. Summary of the Invention

[0014] This application provides an adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks, in order to solve the problems existing in the related technologies for identifying abnormal deformation of dam slopes, such as difficulty in adapting to dynamic changes of monitoring points, ignoring spatial correlation between monitoring points, and insufficient model stability and continuity.

[0015] Firstly, an adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks is provided, which includes the following steps:

[0016] Multiple monitoring points were set up in the dam slope area to continuously collect multi-source monitoring data to characterize the slope deformation state;

[0017] The collected multi-source monitoring data is time-aligned to map the data from different monitoring points to a unified time axis. The time-aligned monitoring data is then subjected to outlier identification, missing value repair, and normalization to construct the temporal feature vectors of each monitoring point in multiple consecutive time steps.

[0018] Each monitoring point is mapped to a graph node. The connection relationship between the graph nodes is determined based on the spatial location relationship, geological zoning relationship and / or deformation correlation calculated based on historical monitoring data. Edge weights that characterize the correlation strength are set for the connection relationship between nodes, thereby constructing a monitoring dynamic graph to describe the spatial correlation relationship of the monitoring points.

[0019] When the number of monitoring points changes or the relationship between monitoring points changes, the node set and / or edge weights in the monitoring dynamic graph are updated.

[0020] Based on the updated monitoring dynamic graph, the multi-time step temporal feature vectors corresponding to each monitoring point are used as node features and input into the dynamic graph neural network. By performing neighborhood feature aggregation and node feature update operations in the dynamic graph neural network, the deformation features of the monitoring point itself are fused with the deformation features of adjacent monitoring points to obtain a feature representation that characterizes the overall spatiotemporal deformation state of the dam slope.

[0021] Based on the feature representations corresponding to the monitoring data during the historical normal operation phase, a normal deformation feature space is constructed to characterize the distribution of normal deformation behavior of the dam slope.

[0022] The feature representation corresponding to the current monitoring state is compared with the normal deformation feature space, and the deviation of the current monitoring state from the normal deformation feature space is calculated as an anomaly scoring index.

[0023] The anomaly identification threshold is determined based on the statistical distribution characteristics of the anomaly scoring indicators in historical monitoring data, and the anomaly identification threshold is updated according to the statistical characteristics of the anomaly scoring indicators within a preset time window during system operation.

[0024] The anomaly score index corresponding to the current monitoring status is compared with the updated anomaly identification threshold to determine whether the dam slope has undergone abnormal deformation, and the anomaly identification result is output.

[0025] In some embodiments, the monitoring points are set at different spatial locations on the dam slope to obtain multi-source monitoring data that can reflect the slope deformation state. The monitoring points include at least one or more of displacement monitoring points, inclination monitoring points, distributed optical fiber monitoring points, and seepage pressure monitoring points.

[0026] Different types of monitoring points are used to collect monitoring data in different physical quantities, including but not limited to slope displacement, deformation rate, strain information and seepage pressure change information. Each monitoring point collects monitoring data according to a preset sampling period. Since different monitoring devices have differences in sampling frequency, sampling start time and data transmission method, before entering the subsequent modeling, the multi-source monitoring data are uniformly mapped to the same time axis through time alignment and resampling processing.

[0027] By unifying the processing of data collected from multiple types of monitoring points, monitoring data from different sources and with different dimensions can be jointly analyzed within the same dynamic graph framework, thus providing a consistent data foundation for subsequent feature learning and abnormal deformation identification based on dynamic graph neural networks.

[0028] In some embodiments, the construction of the time-series features is based on the monitoring data of the monitoring point in the current time and multiple consecutive time steps before it. By setting the time window length, the monitoring data of multiple consecutive sampling times are combined to form a multi-time-step feature vector.

[0029] The multi-time-step feature vector is used to characterize the deformation evolution process of the monitoring point over a period of time, so that when the model identifies abnormal deformation, it not only considers the instantaneous state at the current moment, but also comprehensively reflects the deformation development trend and historical evolution information. The time-series features constructed in this way can enhance the model's ability to identify slow-evolutionary anomalies and sudden anomalies, and avoid misjudgment caused by relying only on data from a single time point.

[0030] The length of the time window can be set according to the actual needs of the project and the sampling characteristics of the monitoring data. While maintaining the integrity of the time features, it also takes into account the computational efficiency, thereby providing a stable and continuous time dimension input for the subsequent feature learning of the dynamic graph neural network.

[0031] In some embodiments, the connection relationship between monitoring points in the monitoring dynamic graph is determined based on at least one of the following: spatial distance threshold between monitoring points, consistency of geological units or engineering zones, and deformation correlation.

[0032] In the specific construction process, when the spatial distance between two monitoring points is less than a preset threshold, a connection relationship is established between the corresponding nodes; or when two monitoring points are located in the same geological unit or engineering zone, a connection relationship is established between the corresponding nodes; or when the deformation correlation calculated based on historical monitoring data is higher than a preset threshold, a connection relationship is established between the corresponding nodes.

[0033] The monitoring dynamic map constructed in the above manner can reflect the spatial layout characteristics and deformation synergy of the measuring points in the dam slope monitoring system, providing a structured foundation for the subsequent feature propagation and information fusion of the dynamic graph neural network under graph structure constraints.

[0034] In some embodiments, the edge weights between nodes in the monitoring dynamic graph are determined based on the similarity between the deformation features of the monitoring points, and are used to quantify the strength of the correlation between different monitoring points;

[0035] The edge weights are set based on the differences in feature vectors of the monitored points within the same time period. By mapping the differences in feature vectors, points with similar deformation features have larger edge weights, while points with large differences in deformation features have smaller edge weights. By introducing edge weights, the monitoring dynamic graph can not only express whether there is a correlation between the points, but also further characterize the strength of the correlation.

[0036] The edge weights are dynamically adjusted as the monitoring data is updated, so that the dynamic monitoring map can continuously reflect the spatial propagation characteristics of dam slope deformation and its dynamic changes over time.

[0037] In some embodiments, the dynamic graph neural network adopts a multi-layer network structure, and realizes the propagation and fusion of measurement point deformation information at multiple spatial levels through multiple neighborhood feature aggregations;

[0038] In each layer of the dynamic graph neural network, node features are updated by aggregating the features of their neighboring nodes and combining them with edge weights. This allows nodes to integrate deformation information from neighboring measurement points while updating their own features. As the number of network layers increases, node features gradually integrate monitoring measurement point information over a larger range, thereby achieving layer-by-layer modeling from local deformation features to regional-scale deformation features.

[0039] This multi-layer feature propagation mechanism enables the model to characterize the overall deformation behavior of the dam slope, rather than being limited to a single measuring point or a local area.

[0040] In some embodiments, the anomaly scoring index includes at least the deviation of the current monitoring state from the normal deformation feature space;

[0041] The normal deformation feature space is constructed based on monitoring data from historical normal operation phases. It is used to characterize the distribution characteristics of deformation behavior of dam slope under normal working conditions. By comparing the feature representation corresponding to the current monitoring state with the normal deformation feature space, the degree of deviation is calculated, thereby quantifying the difference between the current deformation state and the historical normal state.

[0042] The greater the deviation, the more the current monitoring state deviates from the normal deformation pattern, and the higher the degree of abnormality. By introducing an anomaly scoring index based on the deviation of the feature space, the abnormal deformation judgment is based on historical normal behavior, thereby improving the rationality and stability of anomaly identification.

[0043] In some embodiments, the anomaly identification threshold is determined based on the statistical distribution of anomaly scoring indicators in historical monitoring data and is dynamically updated through a sliding time window;

[0044] In the initial stage of system operation, the anomaly identification threshold is initialized using the statistical characteristics of anomaly scores during the historical normal operation phase; during the continuous operation of the system, the threshold is updated based on the statistical characteristics of anomaly scores in the recent period, so that the anomaly identification threshold can be adaptively adjusted as the distribution of monitoring data changes.

[0045] This approach avoids the problem of false alarms or missed alarms caused by using fixed thresholds at different operating stages, and improves the adaptability of abnormal deformation identification in long-term operation.

[0046] In some embodiments, during the abnormal deformation discrimination process, the monitoring point is determined to have abnormal deformation only when the same monitoring point meets the abnormal discrimination conditions in multiple consecutive time steps.

[0047] By introducing a continuous time step determination mechanism, the abnormal misjudgment caused by short-term fluctuations or occasional interference within a single time step is avoided. The number of continuous time steps can be set according to the actual engineering needs to achieve a balance between the sensitivity and stability of abnormal identification.

[0048] This judgment method improves the reliability of anomaly identification results and helps reduce invalid alarms in engineering applications.

[0049] In some embodiments, when multiple monitoring points located in the same spatial region or geological unit are determined to be abnormal within a similar time period, the region is determined to be an abnormal deformation region, and the regional-level abnormal deformation identification result is output.

[0050] The regional anomaly identification is based on the joint analysis of anomaly results from multiple measuring points. It can reflect the spatial clustering characteristics of slope anomalies. By identifying regional anomalies, it is beneficial to identify the overall or local structural anomaly deformation trend of the dam slope, providing operation and management personnel with more engineering-significant decision-making basis.

[0051] The beneficial effects of the technical solution provided in this application include:

[0052] 1. This invention abstracts the dam slope monitoring system into a dynamic graph structure that evolves over time, and dynamically updates the graph structure when monitoring points are added or removed or when the correlation changes, so that the abnormal deformation identification process can adapt to changes in the monitoring network structure, thereby maintaining the continuity and stability of abnormal identification under long-term operating conditions.

[0053] 2. This invention utilizes a dynamic graph neural network to perform feature learning on deformation data from multiple measurement points under graph structure constraints, thereby achieving the fusion analysis of spatial correlation information between measurement points and deformation time evolution characteristics. This avoids information loss caused by relying solely on single-point time series and improves the overall integrity and accuracy of abnormal deformation identification.

[0054] 3. This invention integrates feature deviation with prediction residual to construct an anomaly scoring index, so that anomaly identification not only depends on the current data state, but also comprehensively considers the deviation of historical normal behavior and evolution trend, thereby improving the robustness of anomaly identification under complex working conditions.

[0055] 4. Based on the statistical distribution characteristics of anomaly scoring, this invention combines time window updates, operating condition corrections, and smoothing control mechanisms to adaptively adjust the anomaly identification threshold, enabling the anomaly judgment criteria to be dynamically updated with changes in the operating stage and data distribution, thereby reducing the probability of false alarms and missed alarms.

[0056] 5. This invention does not rely on frequent retraining during model operation. Instead, it achieves adaptive model updates through dynamic graph structure updates and threshold adaptation mechanisms, enabling the abnormal deformation recognition method to maintain continuous and stable operation when monitoring conditions change, thereby reducing manual intervention and system maintenance costs.

[0057] 6. This invention outputs the anomaly identification results in a structured manner and combines anomaly persistence determination and regional summary analysis to provide graded early warning for abnormal deformation, so that the identification results can serve the operation and management of dam slopes in an intuitive and traceable way, thereby improving the practicality of anomaly identification results in engineering applications. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a schematic diagram of the overall process of the adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks of the present invention;

[0060] Figure 2 This is a schematic diagram of the multi-source monitoring data acquisition and preprocessing process for dam slopes provided in an embodiment of this application.

[0061] Figure 3 A dynamic diagram illustrating the structure of a dam slope monitoring network provided in this embodiment of the application;

[0062] Figure 4 This is a schematic diagram of node and edge weight modeling in the monitoring dynamic graph provided in the embodiments of this application;

[0063] Figure 5 This is a schematic diagram of the dynamic graph neural network feature learning process provided in an embodiment of this application;

[0064] Figure 6 This is a schematic diagram of multi-timestep spatiotemporal feature fusion provided in an embodiment of this application;

[0065] Figure 7 This is a schematic diagram illustrating the abnormal deformation discrimination principle provided in the embodiments of this application;

[0066] Figure 8 This is a schematic diagram of the adaptive adjustment process for the anomaly identification threshold provided in the embodiments of this application;

[0067] Figure 9 This is a schematic diagram illustrating the adaptive updating and continuous operation of the model provided in the embodiments of this application;

[0068] Figure 10 This is a schematic diagram illustrating the abnormal result output and hierarchical early warning provided in the embodiments of this application. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0070] This application provides an adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks, which can solve the problems of existing dam slope abnormal deformation identification technologies, such as difficulty in adapting to dynamic changes of monitoring points, ignoring spatial correlation between monitoring points, and insufficient model stability and continuity.

[0071] Please see Figures 1-10 The present invention aims to achieve joint analysis of deformation data from multiple monitoring points on dam slopes by introducing a dynamically updated monitoring network modeling mechanism. Even when the number of monitoring points, spatial relationships, or working conditions change, it can still continuously and stably identify abnormal deformations, thereby improving the intelligence level and engineering applicability of dam slope safety monitoring.

[0072] To achieve adaptive identification of abnormal deformation of dam slopes, this invention proposes an adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks. This method achieves dynamic adaptation to changes in the monitoring network structure by performing dynamic graph modeling and joint analysis on monitoring data from multiple measuring points on the dam slope, and on this basis, completes the continuous identification of abnormal deformation.

[0073] The technical solution of this invention mainly includes the following steps:

[0074] (I) Monitoring Data Acquisition and Preprocessing Steps

[0075] In this invention, the monitoring data acquisition and preprocessing steps are the foundation for the adaptive identification of abnormal deformation of dam slopes. The purpose is to address the practical engineering problems of diverse monitoring data sources, inconsistent sampling, and the coexistence of noise and missing data. By standardizing the original monitoring data, a stable and reliable data foundation can be provided for the subsequent construction of dynamic graph structures and anomaly identification models.

[0076] 1. Multi-source monitoring data acquisition

[0077] Multiple monitoring points are deployed along the dam slope to conduct long-term, continuous monitoring of the slope's operational status. These monitoring points include, but are not limited to, inclinometers, GNSS displacement monitoring points, distributed fiber optic sensors, piezometers, or other monitoring equipment capable of reflecting the slope's deformation status.

[0078] Each monitoring point collects monitoring data according to a preset sampling period, forming a multi-point, multi-time-scale raw monitoring data sequence. Let's assume a time... At that time, the first The raw monitoring data collected from each monitoring point is represented as follows: ;

[0079] in, : Indicates the monitoring point At any moment The amount of deformation, strain value, or seepage pressure value;

[0080] : Monitoring point number, used to distinguish monitoring points in different spatial locations;

[0081] The monitoring data collection time corresponds to a specific point in time on a unified timeline.

[0082] 2. Time alignment and resampling processing

[0083] Since the sampling frequency and sampling start time of different types of monitoring equipment may differ, directly using the raw monitoring data will lead to inconsistencies in the time dimension, thus affecting the effect of subsequent joint modeling.

[0084] Therefore, this invention first performs time alignment processing on the monitoring data from multiple monitoring points. Through time interpolation or resampling, the data from each monitoring point are uniformly mapped to the same time series. ;

[0085] in, Discrete sampling times obtained after time alignment and resampling; The total number of sampling times contained in the time series.

[0086] This allows for the construction of a multi-point monitoring data set under a unified time axis, ensuring that the data from each monitoring point are comparable at the same time.

[0087] 3. Identification of outlier and missing data

[0088] During the long-term operation of the monitoring system, abnormal mutation values ​​or missing values ​​may exist in the original monitoring data due to factors such as equipment failure, communication interruption or external disturbance.

[0089] To identify abnormal data, this invention uses historical statistical characteristics to determine anomalies in the monitoring data.

[0090] When measuring point At any moment Monitoring data is considered outlier if it meets the following conditions: ;

[0091] in, Monitoring points The average value in historical monitoring data; Monitoring points Standard deviation in historical monitoring data; : Preset discrimination coefficient; The absolute value operator is used to indicate the degree of deviation of data.

[0092] This discrimination method can effectively identify abnormal monitoring data caused by equipment malfunctions or sudden interference.

[0093] 4. Repair of Abnormal and Missing Data

[0094] To prevent the interruption of the monitoring data sequence from affecting subsequent analysis, this invention performs repair processing on the identified outliers or missing values.

[0095] The repair methods can include linear interpolation, weighted estimation of adjacent time points, or smoothing estimation based on historical trends, thereby ensuring the continuity and stability of the monitoring data in the time dimension.

[0096] 5. Data normalization processing

[0097] Since the data collected from different monitoring points may have significant differences in dimensions and numerical ranges, if joint modeling is carried out directly, it is easy for some monitoring point data to dominate the model.

[0098] Therefore, this invention normalizes the repaired monitoring data to eliminate dimensional differences between different measuring points. The normalization process can be carried out in the following manner: ;

[0099] in, Normalized monitoring points At any moment Data values;

[0100] Monitoring points The minimum value in historical monitoring data;

[0101] Monitoring points The maximum value in historical monitoring data.

[0102] 6. Multi-timestep feature construction

[0103] To fully characterize the temporal evolution of dam slope deformation, this invention further constructs multi-time-step input features during the preprocessing stage.

[0104] For monitoring points At any moment The time series feature vector is represented as:

[0105] ;

[0106] in, Monitoring points At any moment Constructed time series feature vectors;

[0107] : Normalized monitoring data values;

[0108] : Time window length, used to describe the evolution of monitoring data over a period of time;

[0109] to : Each consecutive sampling moment within the time window.

[0110] By performing the above processing on each monitoring point, a multi-point time-series feature set is formed for subsequent modeling:

[0111] ;

[0112] in, This indicates the total number of monitoring points currently participating in the analysis.

[0113] 7. Technical Effects and Engineering Significance

[0114] Through the above-mentioned monitoring data acquisition and preprocessing steps, this invention can construct high-quality multi-point time-series input data while ensuring data continuity and consistency, providing reliable data support for subsequent monitoring dynamic map construction and abnormal deformation identification.

[0115] (II) Steps for constructing the monitoring dynamic graph

[0116] After completing the monitoring data acquisition and preprocessing, in order to characterize the spatial relationship between multiple monitoring points on the dam slope and their dynamic characteristics over time, this invention further abstracts the monitoring system into a monitoring dynamic graph structure that evolves over time, which is used to describe the relationship between the monitoring points and the propagation of their deformation information.

[0117] 1. Graph node modeling of monitoring points; In this invention, each monitoring point deployed in the dam slope area is considered as a node in the graph. Let the current time... The set of monitoring points involved in the analysis is as follows: ;in, : Current analysis time; : No. Each monitoring point corresponds to a graph node; : The number of monitoring points currently participating in the analysis.

[0118] Each node The corresponding node feature vector is represented by the multi-timestep feature constructed in the previous section, that is: ;

[0119] in, Monitoring points At any moment The node feature vector is used to reflect the deformation evolution state of the measuring point over a period of time.

[0120] 2. Initial Monitoring Map Structure Construction: To depict the spatial relationships between different monitoring points, an edge set of the monitoring map is constructed in the initial stage based on the actual engineering situation. This edge set represents the relationships between monitoring points, denoted as: ;

[0121] in, :time The set of edges in the monitoring graph below; : Indicates the monitoring point With measuring points There are connections between them; , : Nodes in the monitoring graph.

[0122] In practice, the connection relationship between monitoring points can be determined based on one or more of the following criteria:

[0123] (1) Spatial distance criterion, when monitoring points With measuring points The spatial distance between them is less than a preset threshold At that time, establish connections between the corresponding nodes: ;

[0124] in, Monitoring points With measuring points Spatial distance between them;

[0125] : Logical implication symbol, indicating that a connection is established when a condition is met.

[0126] (2) Geological zoning criteria: When two monitoring points are located in the same geological unit, structural plane or engineering zone, a connection relationship shall be established between the corresponding nodes.

[0127] (3) Deformation correlation criterion: Based on historical monitoring data, the measurement points are calculated. With measuring points The correlation between the deformations is determined, and a connection is established when the correlation exceeds a preset threshold.

[0128] The initial monitoring map structure is constructed in the above manner, so that the map structure can reflect the spatial layout and engineering characteristics of the dam slope monitoring system.

[0129] 3. Graph edge weight modeling

[0130] Based on the constructed monitoring map, to further characterize the differences in the correlation strength between monitoring points, this invention introduces weight parameters to the map edges. Let the monitoring points... With measuring points The edge weights between them are represented as follows: ;

[0131] in, , Monitoring point number; : Current analysis time.

[0132] In practical implementation, edge weights can be calculated based on the similarity of deformation data between measurement points, for example, based on the similarity definition of normalized monitoring data: ;

[0133] in, :node At any moment eigenvectors;

[0134] :node At any moment eigenvectors;

[0135] Euclidean norm, used to measure the difference between two eigenvectors;

[0136] : Weight decay adjustment parameter, used to control the degree of influence of similarity on edge weights;

[0137] Exponential function: used to map feature differences to weight values.

[0138] By introducing edge weights, the monitoring graph can reflect not only whether the monitoring points are related, but also the strength of the relationship.

[0139] 4. Monitor the dynamic update mechanism of the dynamic graph.

[0140] During long-term monitoring of dam slopes, the monitoring network structure may change over time. Therefore, this invention defines the monitoring map as a dynamic graph that evolves over time. ;

[0141] in, :time The monitoring dynamic graph model; : A set of nodes; : Edge set; : The set of edge weights, containing all .

[0142] The dynamic graph update mechanism includes the following scenarios:

[0143] (1) Add new monitoring points; when a new monitoring point is put into operation, it is added to the node set as a new node. And establish connection relationships with existing nodes based on spatial location and deformation correlation.

[0144] (2) Monitoring points fail or data is missing for a long time; when a monitoring point has a long-term missing data, reduce its node weight or remove its corresponding connection relationship to avoid abnormal data from interfering with the overall analysis.

[0145] (3) Dynamic adjustment of the correlation between measuring points; as the operating conditions change, the deformation correlation between measuring points may change. This invention dynamically adjusts the edge weights based on real-time monitoring data, so that the monitoring graph structure can reflect the current deformation propagation characteristics.

[0146] 5. Project Significance and Subsequent Connections

[0147] By constructing and dynamically updating the monitoring dynamic graph structure, this invention can continuously and stably describe the overall state of the dam slope monitoring system when the number of monitoring points, spatial layout, or correlation changes.

[0148] The dynamic graph structure serves as the input for subsequent dynamic graph neural network feature learning steps, providing a structured foundation for the joint modeling and recognition of abnormal deformations.

[0149] (III) Feature Learning Steps of Dynamic Graph Neural Networks

[0150] After completing the construction of the monitoring dynamic map, in order to fully explore the spatial correlation characteristics and dynamic laws of the evolution of the deformation data of multiple measuring points on the dam slope over time, this invention further introduces a dynamic graph neural network to perform feature learning and information fusion processing on the monitoring dynamic map.

[0151] The dynamic graph neural network is used to jointly model the temporal deformation characteristics of each monitoring point under the constraints of the dynamic graph structure, thereby obtaining a high-level feature representation that can reflect the overall deformation state of the slope.

[0152] 1. Dynamic graph input modeling

[0153] At any moment The monitoring dynamic diagram constructed in the previous section can be represented as: ;

[0154] Among them, the node set This represents the monitoring points currently participating in the analysis, and the edge set. The weight set represents the relationship between measurement points. This indicates the correlation strength between measurement points.

[0155] For each monitoring point node Its initial node feature vector is defined as: ;

[0156] in, This is the multi-time-step temporal deformation feature vector constructed in Section 1, used to describe the deformation evolution process of the measurement point over a certain period of time.

[0157] 2. Neighborhood Feature Aggregation Mechanism

[0158] In dynamic graph neural networks, the propagation and fusion of deformation information between measurement points is achieved through a neighborhood feature aggregation mechanism. For nodes... At any moment The set of neighboring nodes is represented as: ;

[0159] in, , : Monitor the graph nodes corresponding to the monitoring points; : Indicates the measuring point With measuring points At any moment There is a connection.

[0160] Based on the characteristics of neighboring nodes and edge weights, the nodes The neighborhood features are weighted and aggregated to obtain the neighborhood information representation:

[0161] ;

[0162] in, Monitoring points At any moment The neighborhood aggregated feature vector;

[0163] :time measuring point With measuring points The correlation weight between them;

[0164] This indicates a weighted summation of the features of all adjacent measurement points.

[0165] This aggregation process enables each monitoring point to integrate deformation information from its surrounding points, thereby characterizing the collaborative deformation features of the local area.

[0166] 3. Node Feature Update

[0167] After completing the neighborhood feature aggregation, the node's own features and the aggregated neighborhood features are jointly updated to obtain the node's feature representation in the current layer: ;

[0168] in, Monitoring points Features of nodes after the first layer of the graph neural network; , Learnable parameter matrix; Bias vector; : Non-linear activation function.

[0169] This update method allows node features to simultaneously include their own historical deformation information and the collaborative deformation features of neighboring measuring points, which is beneficial for reflecting the overall deformation state of the slope.

[0170] 4. Multi-layer feature propagation and dynamic adaptation: In order to obtain spatial correlation information over a wider range, this invention can adopt a multi-layer dynamic graph neural network structure, and realize multi-hop neighborhood information propagation by repeating neighborhood aggregation and node update steps.

[0171] In the In a layered network, the node feature update process can be represented as:

[0172] ;

[0173] in, Measurement point In the Node features after layered graph neural networks;

[0174] Measurement point The node features in the previous layer;

[0175] , : No. The learnable weight matrix of a layered network;

[0176] : Graph neural network layer index.

[0177] As the number of network layers increases, node features can integrate monitoring point information over a wider range, thereby characterizing the overall deformation pattern at the regional scale of the slope.

[0178] When the structure of the monitored dynamic graph changes (such as adding nodes or adjusting edge weights), the dynamic graph neural network can directly propagate features to the updated graph structure while keeping the parameter structure unchanged, thereby achieving dynamic adaptation to changes in the monitored network.

[0179] 5. Temporal Feature Fusion and State Representation

[0180] Since the deformation of dam slopes has obvious time evolution characteristics, this invention integrates the node features at different time steps in the dynamic graph neural network feature learning process to form a comprehensive state representation.

[0181] Set nodes The characteristics within a continuous time period are represented as follows Then a node can be constructed at time. The combined state vector: ;

[0182] in, : A mapping function used to fuse features from nodes across multiple time steps; : Length of time window.

[0183] The integrated state vector is used to reflect the deformation state of the monitoring points under the dual constraints of spatial correlation and temporal evolution.

[0184] 6. Feature Output and Subsequent Connections

[0185] Through the above dynamic graph neural network feature learning steps, the monitoring points at time [time] are finally obtained. High-dimensional feature representation set:

[0186] ;

[0187] in, This represents the number of monitoring points currently participating in the analysis.

[0188] The feature representation set serves as the input for subsequent abnormal deformation discrimination steps, and is used to perform abnormal analysis and identification of the deformation behavior of the monitored measurement points.

[0189] (iv) Abnormal Deformation Judgment Steps

[0190] After completing the dynamic graph neural network feature learning, in order to identify abnormal deformation behavior of dam slopes, this invention further uses the obtained node feature representation to make anomaly judgments on the deformation state of monitoring points.

[0191] The core of the abnormal deformation discrimination step is to determine whether the current monitoring state deviates from the normal operation state by considering the historical deformation characteristics of the measuring point itself and the coordinated deformation relationship between it and the surrounding measuring points.

[0192] 1. Determination of the abnormality detection object; In this invention, the abnormal deformation detection object includes at least one of the following: the deformation state of a single monitoring point at the current moment; the deformation state of a local area composed of multiple adjacent monitoring points; the changing trend of the deformation state of the monitoring point within a continuous time period.

[0193] For ease of description, the following explanation uses a single monitoring point as an example, but this does not constitute a limitation on the scope of protection of this invention.

[0194] 2. Construction of the feature space for normal deformation

[0195] Based on dam slope monitoring data from historical operation phases, a feature space reflecting slope deformation behavior under normal operating conditions is constructed using node feature representations output by a dynamic graph neural network.

[0196] Set up monitoring points The set of comprehensive state vectors corresponding to the normal historical operation phase is as follows:

[0197] ;

[0198] in, : Indicates the monitoring point The normal deformation feature space constructed under normal historical operating conditions;

[0199] : Indicates the monitoring point At a historical moment The comprehensive state feature vector obtained after feature learning by a dynamic graph neural network;

[0200] : Represents the first feature space used to construct the normal deformation feature space A historic moment;

[0201] : Indicates the number of historical normal operation samples selected.

[0202] The normal deformation feature space is used to characterize the distribution characteristics of deformation behavior of monitoring points under normal working conditions.

[0203] 3. Calculation of deviation from current deformation state

[0204] When the measuring point is obtained At the present moment Feature representation Then, the abnormal deformation is quantitatively described by calculating the degree of deviation between it and the normal deformation feature space.

[0205] In one implementation, the difference between the current state and the normal state can be measured by a feature distance, and the deviation can be expressed as:

[0206] ;

[0207] in, : Indicates the monitoring point At the present moment The degree of deviation of the deformation state from the normal deformation feature space;

[0208] : Indicates the monitoring point At the present moment The comprehensive state feature vector;

[0209] : Represents the normal deformation feature space Any historical feature vector in;

[0210] : Represents the Euclidean norm, used to measure the distance between two eigenvectors;

[0211] : This indicates that the historical feature closest to the current state is selected from the normal deformation feature space to quantify the minimum deviation.

[0212] Deviation The larger the value, the more significant the difference between the current deformed state and the historical normal state.

[0213] 4. Anomaly-assisted detection based on prediction residuals

[0214] In addition to feature deviation, this invention can also combine predicted residuals to assist in the identification of abnormal deformations.

[0215] Based on the prediction results of the deformation state of the measuring point using a dynamic graph neural network, let the predicted features be represented as follows: Then the prediction residual can be expressed as: ;

[0216] When the prediction residual continues to increase, it indicates that the current deformation behavior of the measuring point has significantly deviated from the normal evolution law learned by the model.

[0217] 5. Construction of Anomaly Scoring Metrics

[0218] To comprehensively reflect the degree of anomaly at the measurement points, this invention integrates feature deviation and prediction residual to construct an anomaly scoring index:

[0219] ;

[0220] in, : Indicates the monitoring point At the present moment The abnormal deformation scoring index is used to quantitatively characterize the degree of abnormal deformation of the monitoring point at the current moment;

[0221] : Represents the coefficient used to adjust the weight of deviation in outlier ratings;

[0222] : This represents the coefficient used to adjust the weight of the predicted residual in the anomaly score.

[0223] 6. Rules for Judging Abnormal Deformation

[0224] The abnormality scoring index is compared with a preset or adaptively adjusted abnormality discrimination threshold to complete the abnormality deformation determination.

[0225] Determine the measuring point when the following conditions are met. At any moment Abnormal deformation behavior exists: ;

[0226] in, : Indicates the monitoring point At any moment The corresponding threshold for detecting abnormal deformation.

[0227] The threshold can be dynamically adjusted based on the distribution of historical monitoring data, changes in operating conditions, or statistical characteristics, thereby improving the stability of anomaly detection.

[0228] 7. Regional-level anomaly deformation identification

[0229] In one implementation, the abnormal scores of multiple adjacent monitoring points can be jointly analyzed. When multiple monitoring points in a certain local area show an abnormal increase in scores at the same time, it is determined that there is an abnormal deformation trend in that area.

[0230] Regional anomaly identification is helpful in recognizing overall or local structural abnormal deformation behavior of slopes.

[0231] (v) Adaptive adjustment steps for anomaly detection threshold

[0232] In the process of abnormal deformation detection, the setting of the abnormality identification threshold directly affects the accuracy and stability of abnormality identification. If a fixed threshold is used, false alarms or missed alarms are likely to occur at different operating stages or under different working conditions.

[0233] Therefore, based on anomaly detection, this invention further introduces an adaptive adjustment step for the anomaly recognition threshold to improve the adaptability of anomaly deformation recognition during long-term operation.

[0234] 1. Basic definition of anomaly detection threshold

[0235] In this invention, the monitoring points At any moment The determination of abnormal deformation is based on the abnormality scoring index. .

[0236] The corresponding anomaly detection threshold is expressed as follows: ;

[0237] When abnormal scoring Exceeding the threshold When the time is right, it is determined that there is abnormal deformation behavior at the corresponding measuring point.

[0238] The threshold is not a fixed constant, but is dynamically adjusted according to the distribution of monitoring data and the operating status.

[0239] 2. Threshold initialization based on historical anomaly score distribution

[0240] In the initial stage of system operation or model initialization, monitoring data from historical normal operation phases are used to statistically analyze the distribution characteristics of abnormal scores at measurement points, which serve as the basis for threshold setting.

[0241] Set up measuring points The set of anomaly scores corresponding to the normal historical operation phase is as follows: ;

[0242] in, Measurement point At a historical moment Abnormal score values;

[0243] Historical monitoring moments;

[0244] : The number of historical samples used to construct statistical features of normal operating conditions.

[0245] In one implementation, an initial threshold can be determined based on the statistical characteristics of the anomaly scores: ;

[0246] in, Measurement point The initial anomaly detection threshold; : Abnormal scoring set The mean; : Abnormal scoring set Standard deviation; Threshold sensitivity adjustment coefficient, used to control the strictness of anomaly detection.

[0247] 3. Time-adaptive update mechanism for thresholds

[0248] During the long-term operation of the dam slope, the statistical characteristics of the monitoring data may change over time due to factors such as seasonal water level changes, changes in rainfall conditions, and adjustments to operating conditions.

[0249] Therefore, this invention uses a sliding time window to dynamically update the distribution of abnormal scores.

[0250] Let's assume at the current time Based on the length of the time window Select the set of abnormal scores from the most recent period:

[0251] ;

[0252] Based on the statistical characteristics of the anomaly scores within this time window, the anomaly identification threshold is updated as follows:

[0253] ;

[0254] in, : Set of abnormal scores within the time window The mean;

[0255] : Set of abnormal scores within the time window The standard deviation.

[0256] The anomaly detection threshold can be dynamically adjusted over time through a sliding window update mechanism.

[0257] 4. Adaptive threshold correction under changing operating conditions

[0258] During certain operational phases, such as periods of rapid water level changes, concentrated rainfall, or construction impacts, the overall fluctuation range of monitoring data may increase.

[0259] If a uniform threshold is still used, it is easy to misjudge normal operating condition fluctuations as abnormalities.

[0260] Therefore, this invention introduces a working condition-related correction factor to correct the anomaly identification threshold. Let the working condition correction factor be expressed as... Then the threshold is updated as follows: ;

[0261] in, It is determined based on changes in operating conditions or environmental parameters and is used to reflect the impact of current operating conditions on the fluctuation range of monitoring data.

[0262] 5. Threshold smoothing and stability control

[0263] To avoid drastic fluctuations in the anomaly detection threshold within a short period, this invention performs a smoothing process on the updated threshold. Let the smoothed anomaly detection threshold be expressed as: ;

[0264] in, The threshold for smooth anomaly detection in the previous time step;

[0265] The current anomaly detection threshold after working condition correction;

[0266] : Smoothing coefficient, used to control the smoothness of threshold updates.

[0267] Threshold smoothing can effectively avoid misjudgments caused by short-term data anomalies and improve the stability of anomaly identification results.

[0268] 6. Coordinated adjustment of regional anomaly thresholds

[0269] In one implementation, multiple monitoring points located within the same spatial region or geological unit can be jointly analyzed.

[0270] When the overall anomaly scores of multiple measuring points in a region increase, the anomaly identification thresholds for measuring points in the region are adjusted collaboratively.

[0271] The regional threshold coordination mechanism is helpful in identifying the overall or local structural abnormal deformation trend of dam slopes.

[0272] 7. Project Significance and Subsequent Connections

[0273] Through the above-described adaptive adjustment steps for the anomaly identification threshold, the present invention can dynamically adjust the anomaly discrimination criteria under different operating stages and conditions, thereby reducing the probability of false alarms and missed alarms.

[0274] The adaptive threshold, as an important component of abnormal deformation discrimination, works together with the aforementioned abnormality scoring index to provide a reliable basis for subsequent adaptive updates and continuous operation of the model.

[0275] (vi) Model Adaptive Update and Continuous Operation Steps

[0276] During the long-term operation and monitoring of dam slopes, the monitoring system faces various dynamic situations, such as the addition or removal of monitoring points, changes in equipment status, and adjustments to operating conditions. To ensure that the abnormal deformation identification method can operate stably and continuously over a long period of time, this invention further introduces a model adaptive update and continuous operation step based on the aforementioned steps.

[0277] The core of this step is to enable the abnormal deformation recognition model to adapt to changes in the monitoring network structure and data distribution without frequently rebuilding or retraining the overall model, thus maintaining the stability and reliability of the recognition results.

[0278] 1. Adaptive response to monitor changes in network structure

[0279] During model operation, when changes occur in the monitored network structure, this invention prioritizes responding through a dynamic graph update monitoring mechanism.

[0280] When a new monitoring point is added to the monitoring system, it is introduced as a new node into the current dynamic monitoring graph. Based on its spatial location, geological zoning, and deformation correlation, a connection is established with existing monitoring points. Since the dynamic graph neural network uses a graph structure as input, this new node can participate in feature propagation and anomaly detection without changing the network parameters.

[0281] When some monitoring points experience long-term data loss due to equipment failure or communication anomalies, the participation weight of the corresponding node is reduced or its connection with other nodes is temporarily removed to weaken its impact on the overall anomaly identification results, thereby ensuring the stability of the model operation.

[0282] 2. Model Adaptation Mechanism under Changing Data Distribution

[0283] During long-term operation, the statistical distribution of dam slope monitoring data may change due to factors such as water level regulation, seasonal changes, and rainfall infiltration.

[0284] This invention uses the aforementioned adaptive adjustment mechanism for anomaly identification threshold to dynamically modify the anomaly discrimination criteria, enabling the model to maintain reasonable anomaly identification capabilities even when data distribution changes.

[0285] During this process, the model parameters remain unchanged, and the model can adapt to the new operating state simply by updating the threshold and adjusting the dynamic graph structure, thereby avoiding frequent model retraining.

[0286] 3. Management of phased transitions between abnormal and normal states

[0287] During certain operational phases, the dam slope may experience short-term abnormal disturbances before returning to a stable state. To avoid long-term impacts of abnormal states on subsequent model operation, this invention introduces a phased switching management mechanism between abnormal and normal states.

[0288] When the anomaly score of a certain monitoring point or area recovers to the threshold range within multiple consecutive time steps, this stage is reinstated into the statistical range of normal operation status to update the anomaly identification threshold and statistical characteristics.

[0289] This method enables dynamic switching management between abnormal and normal states, enhancing the model's adaptability to long-term operational status changes.

[0290] 4. Model stability control during continuous operation

[0291] To ensure the stability of the model during long-term continuous operation, this invention implements stability control during the model's operation.

[0292] During the anomaly identification result output stage, a time consistency check is performed on the anomaly identification results to avoid frequent alarms caused by abnormal fluctuations within a single time step. Simultaneously, the reliability of anomaly alarms can be improved by setting continuous anomaly trigger conditions.

[0293] In addition, when the monitoring system experiences short-term data anomalies or communication interruptions, the model can run temporarily based on the existing monitoring map structure and historical characteristics, and then be updated normally after the data is recovered.

[0294] 5. Continuously operating engineering deployment method

[0295] At the engineering deployment level, the model adaptive update and continuous operation steps of this invention can be executed periodically in the background of the monitoring system, forming a closed-loop operation process by combining monitoring data collection, dynamic graph updates, anomaly detection and threshold adjustment.

[0296] The closed-loop process includes:

[0297] Monitoring data access → graph structure update → feature learning → anomaly detection → threshold adjustment → result output → operation status update.

[0298] This closed-loop process enables the abnormal deformation identification model to continue to play a role in the long-term operation monitoring of dam slopes.

[0299] 6. Engineering significance and technical effects

[0300] Through the above-described adaptive update and continuous operation steps, this invention can maintain the continuity and stability of the abnormal deformation identification process under complex conditions such as changes in network structure, data distribution, and operational conditions.

[0301] This step effectively reduces the cost of manual intervention and repetitive modeling, and improves the engineering applicability and operational reliability of the dam slope safety monitoring system.

[0302] (vii) Abnormal result output and early warning steps

[0303] After completing the abnormal deformation discrimination and adaptive adjustment of the abnormal identification threshold, in order to enable the abnormal identification results to effectively serve the safety management of dam slope operation, this invention further outputs the abnormal identification results in a structured manner, and realizes abnormal deformation early warning on this basis.

[0304] The purpose of the abnormal result output and early warning steps is to transform the abnormal information identified by the model into understandable, traceable, and responsive engineering information, providing decision support for operation and management personnel.

[0305] 1. Structured output of anomaly identification results: In this invention, for each monitoring point... At any moment The anomaly identification results are recorded in a structured manner, and the output information includes at least the following: monitoring point number or spatial location identifier; time information of the anomaly occurrence; and the corresponding anomaly score. The current anomaly detection threshold ; Anomaly determination result (whether it is abnormal).

[0306] The above methods are used to create anomaly identification results records for a single monitoring point, which can be used for subsequent analysis and tracing.

[0307] 2. Determination of Abnormal Persistence and Credibility

[0308] To avoid frequent false alarms caused by short-term data fluctuations or occasional interference, this invention introduces an anomaly persistence determination mechanism in the abnormal result output stage.

[0309] In one implementation, a monitoring point is marked as a valid anomaly only when it meets the anomaly determination criteria for multiple consecutive time steps. Let the number of consecutive anomaly time steps be... An anomaly is determined when the following conditions are met:

[0310] ;

[0311] in, The monitoring point number is used to uniquely identify a specific monitoring point in the dam slope monitoring system.

[0312] The current time of anomaly detection;

[0313] : Represents the time index variable used for anomaly persistence statistics, used to iterate through the current time. Multiple time steps that are traced back;

[0314] : Indicates the monitoring point At time step The corresponding anomaly score value is calculated by the aforementioned anomaly deformation discrimination step and is used to quantify the degree of anomaly of the measuring point at that moment.

[0315] : Indicates the monitoring point In time The corresponding anomaly identification threshold is dynamically obtained through the anomaly identification threshold adaptive adjustment step.

[0316] : Indicates an indicator function, which takes the value 1 when the condition in parentheses is true and takes the value 0 when the condition is false. It is used to count the number of times the anomaly judgment condition is true within a given time window;

[0317] : indicates the time step Time to step Continuous The statistics are accumulated at each time step.

[0318] Continuous evaluation can effectively improve the reliability of anomaly identification results.

[0319] 3. Summarize and output regional-level anomaly results

[0320] In dam slope engineering, abnormal deformation often exhibits spatial clustering characteristics. Therefore, this invention performs regional-level summary analysis of abnormal results from adjacent monitoring points.

[0321] When multiple monitoring points located in the same spatial area or geological unit are identified as abnormal within a similar time period, the area is marked as an abnormal deformation area, and regional-level anomaly information is output, including the range of the abnormal area, the number of abnormal monitoring points, and the duration of the anomaly.

[0322] Regional-level anomaly outputs help managers quickly identify key risk areas.

[0323] 4. Abnormal Deformation Grading and Early Warning Rules

[0324] Based on the output of abnormal results, this invention provides graded early warning for abnormal deformation according to the magnitude of the abnormal score and the duration of the abnormality.

[0325] In one implementation, abnormal deformation warnings can be divided into multiple levels, such as general abnormality, relatively severe abnormality, and severe abnormality. Different levels of warnings correspond to different abnormality scoring ranges and duration conditions.

[0326] By using a tiered early warning system, managers can take differentiated measures based on the severity of the anomaly.

[0327] 5. Output format of early warning information

[0328] The abnormal warning information of the present invention can be output in various forms, including but not limited to: visual annotation in the monitoring system interface; generating abnormal warning reports or log records; and sending warning information to relevant management systems.

[0329] The format of the early warning information output can be configured according to the specific engineering application requirements, and does not constitute a limitation on the scope of protection of this invention.

[0330] 6. Correlation between early warning results and model operation feedback

[0331] During the process of outputting and issuing warnings for abnormal results, this invention records the warning results in association with the model's running status.

[0332] Once an anomaly warning is confirmed by the engineering team, it can be used as an anomaly sample for subsequent model operation status analysis. When an anomaly warning is determined to be a false alarm, the probability of false alarms in similar situations can be gradually reduced through a threshold adaptive mechanism without changing the model parameters.

[0333] This feedback correlation mechanism creates a closed loop between the output of abnormal results and the continuous operation of the model.

[0334] 7. Engineering significance and technical effects

[0335] Through the above-described abnormal result output and early warning steps, this invention can transform complex anomaly identification calculation results into intuitive and operable engineering early warning information, thereby improving the usability of anomaly identification results in actual engineering.

[0336] This step, together with the aforementioned anomaly detection, threshold adaptation, and continuous model operation steps, forms a complete engineering application closed loop for the dam slope anomaly deformation identification method.

[0337] The working principle of this application is as follows:

[0338] This invention addresses the engineering characteristics of dam slopes during long-term operation, including strong time-varying deformation behavior, complex spatial relationships, and frequent dynamic changes in the monitoring network. It constructs an adaptive anomaly identification method based on a dynamic graph neural network. Its working principle can be summarized as follows: based on multi-source monitoring data, dynamic graph modeling is used to characterize the correlation between measuring points; a dynamic graph neural network is used to fuse spatiotemporal deformation features; and an adaptive anomaly detection and threshold adjustment mechanism is combined to achieve continuous identification and early warning of anomalies.

[0339] (1) The principle of unified time series modeling driven by multi-source monitoring data

[0340] This invention first obtains continuous monitoring data from multiple monitoring points set up on the dam slope, and then maps monitoring data of different types and sampling characteristics to the same time scale through time alignment, anomaly repair and normalization.

[0341] By constructing multi-time-step feature vectors, the input of each monitoring point not only reflects the current state, but also contains deformation evolution information over a period of time, thus providing a stable and continuous temporal basis for subsequent modeling.

[0342] (2) Monitoring Network Dynamic Graph Structure Representation Principle

[0343] To address the issues of potential additions or subtractions, failures, and changes in correlations of monitoring points that may occur during the long-term operation of dam slope monitoring systems, this invention abstracts the monitoring system into a dynamic graph structure that evolves over time.

[0344] In this dynamic graph, monitoring points serve as nodes, while spatial proximity, geological zoning, and deformation correlation among these points are represented by edges. Edge weights characterize the strength of the connections between monitoring points. As monitoring data and operational status change, the graph structure and edge weights are dynamically updated, ensuring that the graph model always reflects the true state of the monitoring network.

[0345] (3) The principle of spatiotemporal feature fusion in dynamic graph neural networks

[0346] Under the constraints of a dynamic graph structure, this invention introduces a dynamic graph neural network to learn features from the monitoring data.

[0347] Through the neighborhood feature aggregation mechanism, each monitoring point updates its own features while integrating deformation information from surrounding monitoring points, thereby enabling the spatial propagation of deformation information. Through multi-layer feature propagation and multi-time step fusion, the node features simultaneously include spatial correlation features and temporal evolution features.

[0348] This process enables the model to characterize the coordinated deformation behavior of the dam slope, both locally and globally, without relying on single-point time series analysis.

[0349] (4) Anomaly detection principle based on deviation and prediction residual

[0350] After obtaining the high-dimensional feature representation of the output of the dynamic graph neural network, this invention constructs a normal deformation feature space and compares the current monitoring state with the historical normal operation state.

[0351] On the one hand, the degree of abnormality of deformation behavior is measured by calculating the deviation between the current features and the normal feature space; on the other hand, the residual between the model prediction results and the actual state is compared to help determine whether the deformation deviates from the existing evolutionary law.

[0352] The two are combined to form an anomaly scoring index, enabling quantitative identification of abnormal deformations.

[0353] (5) Adaptive adjustment principle of anomaly detection threshold

[0354] To avoid the problem of insufficient adaptability of fixed thresholds in different operating stages, this invention constructs anomaly identification thresholds based on the statistical distribution characteristics of anomaly scores, and adaptively adjusts the thresholds through methods such as sliding time windows, operating condition corrections, and smooth control.

[0355] This mechanism enables the anomaly detection criteria to be dynamically updated as time, operating conditions, and data distribution change, thereby maintaining stable identification performance during long-term operation.

[0356] (6) Model adaptive update and continuous operation principle

[0357] During model operation, when changes occur in the monitored network structure or data distribution, this invention does not rely on frequent model retraining, but instead achieves adaptive model updates through dynamic graph structure updates and threshold adaptation mechanisms.

[0358] New monitoring points can be directly added to the dynamic graph as new nodes for analysis, while the impact of failed monitoring points can be reduced through weight adjustment or structural updates, thereby ensuring the continuity and stability of the model in long-term operation.

[0359] (7) Abnormal result output and engineering early warning principle

[0360] After completing the anomaly identification, the present invention outputs the anomaly results in a structured manner and introduces anomaly persistence determination and regional summary analysis to avoid false alarms caused by occasional fluctuations.

[0361] Anomalies are graded and warned based on anomaly scores and durations, and output in the form of visualization, logs, or information push, providing intuitive and actionable safety warning information for dam slope operation and management.

[0362] In summary, this invention achieves intelligent identification and engineering application of abnormal deformation of dam slopes through an overall working mechanism of "data-driven, dynamic graph modeling, spatiotemporal feature fusion, anomaly adaptive discrimination, and continuous operation early warning".

[0363] This method can adapt to dynamic changes in the monitoring network, taking into account both identification accuracy and long-term operational stability, and is suitable for safety monitoring and early warning scenarios of reservoir dam slopes and similar projects.

[0364] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. An adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks, characterized in that, It includes the following steps: Multiple monitoring points were set up in the dam slope area to continuously collect multi-source monitoring data to characterize the slope deformation state; The collected multi-source monitoring data is time-aligned to map the data from different monitoring points to a unified time axis. The time-aligned monitoring data is then subjected to outlier identification, missing value repair, and normalization to construct the temporal feature vectors of each monitoring point in multiple consecutive time steps. Each monitoring point is mapped to a graph node. The connection relationship between the graph nodes is determined based on the spatial location relationship, geological zoning relationship and / or deformation correlation calculated based on historical monitoring data. Edge weights that characterize the correlation strength are set for the connection relationship between nodes, thereby constructing a monitoring dynamic graph to describe the spatial correlation relationship of the monitoring points. When the number of monitoring points changes or the relationship between monitoring points changes, the node set and / or edge weights in the monitoring dynamic graph are updated. Based on the updated monitoring dynamic graph, the multi-time step temporal feature vectors corresponding to each monitoring point are used as node features and input into the dynamic graph neural network. By performing neighborhood feature aggregation and node feature update operations in the dynamic graph neural network, the deformation features of the monitoring point itself are fused with the deformation features of adjacent monitoring points to obtain a feature representation that characterizes the overall spatiotemporal deformation state of the dam slope. Based on the feature representations corresponding to the monitoring data during the historical normal operation phase, a normal deformation feature space is constructed to characterize the distribution of normal deformation behavior of the dam slope. The feature representation corresponding to the current monitoring state is compared with the normal deformation feature space, and the deviation of the current monitoring state from the normal deformation feature space is calculated as an anomaly scoring index. The anomaly identification threshold is determined based on the statistical distribution characteristics of the anomaly scoring indicators in historical monitoring data, and the anomaly identification threshold is updated according to the statistical characteristics of the anomaly scoring indicators within a preset time window during system operation. The anomaly score index corresponding to the current monitoring status is compared with the updated anomaly identification threshold to determine whether the dam slope has undergone abnormal deformation, and the anomaly identification result is output.

2. The adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks as described in claim 1, characterized in that: The monitoring points are set at different spatial locations on the dam slope to obtain multi-source monitoring data that can reflect the slope deformation state. The monitoring points include at least one or more of the following: displacement monitoring points, inclination monitoring points, distributed optical fiber monitoring points, and seepage pressure monitoring points. Different types of monitoring points are used to collect monitoring data in different physical quantities, including but not limited to slope displacement, deformation rate, strain information and seepage pressure change information. Each monitoring point collects monitoring data according to a preset sampling period. Since different monitoring devices have differences in sampling frequency, sampling start time and data transmission method, before entering the subsequent modeling, the multi-source monitoring data are uniformly mapped to the same time axis through time alignment and resampling processing. By unifying the processing of data collected from multiple types of monitoring points, monitoring data from different sources and with different dimensions can be jointly analyzed within the same dynamic graph framework, thus providing a consistent data foundation for subsequent feature learning and abnormal deformation identification based on dynamic graph neural networks.

3. The adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks as described in claim 1, characterized in that: The construction of the time series features is based on the monitoring data of the monitoring point in the current time and multiple consecutive time steps before it. By setting the time window length, the monitoring data of multiple consecutive sampling times are combined to form a multi-time step feature vector. The multi-time-step feature vector is used to characterize the deformation evolution process of the monitoring point over a period of time, so that when the model identifies abnormal deformation, it not only considers the instantaneous state at the current moment, but also comprehensively reflects the deformation development trend and historical evolution information. The time-series features constructed in this way can enhance the model's ability to identify slow-evolutionary anomalies and sudden anomalies, and avoid misjudgment caused by relying only on data from a single time point. The length of the time window can be set according to the actual needs of the project and the sampling characteristics of the monitoring data. While maintaining the integrity of the time features, it also takes into account the computational efficiency, thereby providing a stable and continuous time dimension input for the subsequent feature learning of the dynamic graph neural network.

4. The adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks as described in claim 1, characterized in that: The connection relationship between monitoring points in the monitoring dynamic diagram is determined based on at least one of the following: spatial distance threshold between monitoring points, consistency of geological units or engineering zones, and deformation correlation. In the specific construction process, when the spatial distance between two monitoring points is less than a preset threshold, a connection relationship is established between the corresponding nodes; or when two monitoring points are located in the same geological unit or engineering zone, a connection relationship is established between the corresponding nodes; or when the deformation correlation calculated based on historical monitoring data is higher than a preset threshold, a connection relationship is established between the corresponding nodes. The monitoring dynamic map constructed in the above manner can reflect the spatial layout characteristics and deformation synergy of the measuring points in the dam slope monitoring system, providing a structured foundation for the subsequent feature propagation and information fusion of the dynamic graph neural network under graph structure constraints.

5. The adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks as described in claim 1, characterized in that: The edge weights between nodes in the monitoring dynamic graph are determined based on the similarity between the deformation features of the monitoring points, and are used to quantify the strength of the correlation between different monitoring points. The edge weights are set based on the differences in feature vectors of the monitored points within the same time period. By mapping the differences in feature vectors, points with similar deformation features have larger edge weights, while points with large differences in deformation features have smaller edge weights. By introducing edge weights, the monitoring dynamic graph can not only express whether there is a correlation between the points, but also further characterize the strength of the correlation. The edge weights are dynamically adjusted as the monitoring data is updated, so that the dynamic monitoring map can continuously reflect the spatial propagation characteristics of dam slope deformation and its dynamic changes over time.

6. The adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks as described in claim 1, characterized in that: The dynamic graph neural network adopts a multi-layer network structure and realizes the propagation and fusion of measurement point deformation information at multiple spatial levels through multiple neighborhood feature aggregations. In each layer of the dynamic graph neural network, node features are updated by aggregating the features of their neighboring nodes and combining them with edge weights. This allows nodes to integrate deformation information from neighboring measurement points while updating their own features. As the number of network layers increases, node features gradually integrate monitoring measurement point information over a larger range, thereby achieving layer-by-layer modeling from local deformation features to regional-scale deformation features. This multi-layer feature propagation mechanism enables the model to characterize the overall deformation behavior of the dam slope, rather than being limited to a single measuring point or a local area.

7. The adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks as described in claim 1, characterized in that: The anomaly scoring index includes at least the deviation of the current monitoring state from the normal deformation feature space; The normal deformation feature space is constructed based on monitoring data from historical normal operation phases. It is used to characterize the distribution characteristics of deformation behavior of dam slope under normal working conditions. By comparing the feature representation corresponding to the current monitoring state with the normal deformation feature space, the degree of deviation is calculated, thereby quantifying the difference between the current deformation state and the historical normal state. The greater the deviation, the more the current monitoring state deviates from the normal deformation pattern, and the higher the degree of abnormality. By introducing an anomaly scoring index based on the deviation of the feature space, the abnormal deformation judgment is based on historical normal behavior, thereby improving the rationality and stability of anomaly identification.

8. The adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks as described in claim 1, characterized in that: The anomaly identification threshold is determined based on the statistical distribution of anomaly scoring indicators in historical monitoring data and is dynamically updated through a sliding time window. In the initial stage of system operation, the anomaly identification threshold is initialized using the anomaly scoring statistics from the historical normal operation phase; During continuous system operation, the threshold is updated based on the statistical characteristics of anomaly scores over a recent period, so that the anomaly identification threshold can be adaptively adjusted as the distribution of monitoring data changes. This approach avoids the problem of false alarms or missed alarms caused by using fixed thresholds at different operating stages, and improves the adaptability of abnormal deformation identification in long-term operation.

9. The adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks as described in claim 1, characterized in that: In the process of abnormal deformation identification, the monitoring point is determined to have abnormal deformation only when the same monitoring point meets the abnormal identification conditions in multiple consecutive time steps. By introducing a continuous time step determination mechanism, the abnormal misjudgment caused by short-term fluctuations or occasional interference within a single time step is avoided. The number of continuous time steps can be set according to the actual engineering needs to achieve a balance between the sensitivity and stability of abnormal identification. This judgment method improves the reliability of anomaly identification results and helps reduce invalid alarms in engineering applications.

10. The adaptive identification method for abnormal deformation of dam slopes based on dynamic graph neural networks as described in claim 1, characterized in that: When multiple monitoring points located in the same spatial region or geological unit are identified as abnormal within a similar time period, the region is identified as an abnormal deformation region, and the regional-level abnormal deformation identification result is output. The regional anomaly identification is based on the joint analysis of anomaly results from multiple measuring points. It can reflect the spatial clustering characteristics of slope anomalies. By identifying regional anomalies, it is beneficial to identify the overall or local structural anomaly deformation trend of the dam slope, providing operation and management personnel with more engineering-significant decision-making basis.

Images