Reservoir slope displacement prediction method and system
By combining multimodal data preprocessing and deep learning with visual prediction, the problems of insufficient multimodal data fusion and neglect of time scale features in reservoir slope displacement prediction are solved, achieving high-precision, real-time deformation trend prediction and intelligent early warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-26
- Publication Date
- 2026-04-07
AI Technical Summary
Existing technologies struggle to achieve high-precision, real-time, and reliable reservoir slope displacement prediction. Insufficient multimodal data fusion, lack of modal weight adaptive mechanisms, and failure to consider multi-timescale features and spatial topological relationships result in inconsistent prediction results.
By employing multimodal data preprocessing, spatial and temporal modeling, modal adaptive weights, and graph convolutional networks, combined with visual prediction for deep learning, spatial correlation features among various modal data of reservoir slopes are extracted to enable high-precision, real-time deformation trend prediction.
It achieves high-precision, real-time, and reliable prediction of reservoir slope displacement, improves the comprehensiveness and accuracy of monitoring, reduces false alarms and missed alarms, and has intelligent early warning capabilities.
Smart Images
Figure CN121808200A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of geological disaster monitoring and artificial intelligence prediction, and in particular to a method and system for predicting reservoir slope displacement. Background Technology
[0002] Reservoir banks are subject to the combined effects of multiple factors, including rainfall, seepage, water level fluctuations, geological structure, and climate conditions, making them highly susceptible to deformation, collapse, or landslide disasters. Traditional monitoring systems typically use a single sensor (such as an inclinometer or pore pressure gauge) for monitoring, which can only reflect the local condition and is difficult to achieve multi-factor coupled analysis and accurate prediction.
[0003] In recent years, with the development of sensor and computer vision technologies, multi-source information fusion and visual displacement monitoring systems have emerged. Infrared visual monitoring can achieve millimeter-level displacement detection, pore pressure and tilt sensors can reflect internal stress changes, and rainfall, temperature, and humidity can characterize external environmental conditions. However, existing methods still have the following problems: (1) There are differences in sampling frequency and nonlinear coupling among multimodal data, and traditional fusion methods cannot make full use of feature information; (2) Deep learning models mostly remain at the single-modal input level and lack modal weight adaptive mechanism; (3) The combined impact of multiple time scales (short-term disturbances and long-term changes) on the forecast was not considered; (4) The spatial topology and terrain constraints of the sensor space were ignored, resulting in insufficient spatial consistency of the prediction results; (5) Visual monitoring data is mostly used for detection rather than prediction, and lacks effective coupling with depth prediction models.
[0004] In summary, existing technologies lack deep learning methods for predicting reservoir slope displacement that integrate multimodal monitoring data, possess spatial and temporal modeling capabilities, adaptive modal weights, and combine visual prediction. As a result, it is difficult to achieve high-precision, real-time, and reliable deformation trend prediction and risk warning. Summary of the Invention
[0005] To overcome the shortcomings of existing technologies, the present invention aims to propose a method and system for predicting reservoir slope displacement. The method integrates multimodal monitoring data, has spatial and temporal modeling capabilities, adaptive modal weights, and can combine visual prediction to perform deep learning-based reservoir slope displacement prediction, thereby achieving high-precision, real-time, and reliable deformation trend prediction and risk warning.
[0006] To achieve the objectives of this invention, the following technical solution is adopted: One of the objectives of this invention is: A method for predicting reservoir slope displacement, the method comprising the following steps: Acquire multimodal data of reservoir slopes in the target area and preprocess the multimodal data of reservoir slopes; Based on the preprocessed multimodal data of the reservoir slope, the spatial variation characteristics of the reservoir slope are extracted, and the displacement of the reservoir slope is determined according to the spatial variation characteristics of the reservoir slope. Based on the preprocessed multimodal data of reservoir slope and the displacement of reservoir bank slope, spatial correlation features between the various modal data of reservoir slope in the target area are extracted; Based on the spatial correlation characteristics between various modal data of the reservoir slope, the displacement of the reservoir slope is predicted, and the prediction results are output.
[0007] In the above technical solution, preprocessing the acquired multimodal data of reservoir slopes improves the reliability and stability of the data, thereby accelerating the efficiency of data computation. Extracting the spatial variation characteristics of the reservoir slopes to determine the reservoir slope displacement compensates for the inadequacy (error) of single-frame detection, which only reflects instantaneous displacement, and expands the information dimension of visual displacement detection from a temporal perspective, enhancing the comprehensiveness and accuracy of monitoring. Extracting the spatial correlation features between the various modal data of the reservoir slopes in the target area improves spatial perception, enabling the identification of local slip zone characteristics and regional coordinated changes, thus enhancing spatial consistency. Finally, based on the spatial correlation features between the various modal data of the reservoir slopes, reservoir slope displacement prediction is performed. This combines visual prediction with deep learning-based reservoir slope displacement prediction, achieving high-precision, real-time, and reliable deformation trend prediction.
[0008] Furthermore, the preprocessing of the multimodal data of the reservoir slope includes: The acquired multimodal data of the reservoir slope in the target area includes rainfall monitoring data, soil temperature and humidity data, infrasound and ground acoustic wave data, reservoir slope tilt angle data, reservoir water level data, slope seepage status data, phreatic line data, air temperature and humidity data, and visual displacement monitoring data. The acquired reservoir slope multimodal data were subjected to unified time synchronization, missing value imputation, outlier removal and normalization processing to obtain a reservoir slope multimodal dataset containing time series data.
[0009] Furthermore, the process of extracting the spatial variation characteristics of reservoir bank slopes based on the preprocessed multimodal data of reservoir slopes includes: Visual displacement monitoring data of reservoir slope multimodal data is acquired, wherein the visual displacement monitoring data includes fixed target image monitoring data of the bank slope; A preset depth detection network is used to extract the spatial coordinate data of the target bounding box as the reservoir bank slope shifts from the image monitoring data of the fixed target on the bank slope, and spatial coordinate data below the preset confidence threshold are removed to obtain the target spatial coordinate data that changes with the reservoir bank slope shift.
[0010] Furthermore, the process of determining the reservoir bank slope displacement based on the aforementioned spatial variation characteristics includes: The initial target center coordinates are obtained based on the target spatial coordinate data that changes with the reservoir bank slope displacement, and the current target centroid coordinates are calculated using a preset sub-pixel edge model. The reservoir bank slope displacement is calculated based on the initial target center coordinates and the current target centroid coordinates.
[0011] Furthermore, the process of calculating the current target centroid coordinates using a preset sub-pixel edge model includes: A local quadratic surface is constructed based on the edge pixel data of the target bounding box, expressed as follows:
[0012] in, This represents the pixel edge coordinates of the target bounding box. This represents the parameters of the local quadratic surface that need to be solved; Based on local quadratic surface The coordinates of the sub-pixel edge points of the current target are calculated using a preset sub-pixel edge model, expressed as follows:
[0013] The sub-pixel edge coordinates of each current target are obtained by solving the problem. ); Based on the sub-pixel edge point coordinates of each current target Calculate the current target centroid coordinates The expression is:
[0014] in, This represents the weight of the i-th edge; represents the coordinates of the i-th sub-pixel edge point; r represents the radius of the target's circular structure.
[0015] Furthermore, the expression for calculating the reservoir bank slope displacement based on the initial target center coordinates and the current target centroid coordinates is as follows:
[0016] in, This represents the subpixel coordinates of the target circle's center in the initial reference frame.
[0017] Furthermore, the process of extracting spatial correlation features among different modal data of the reservoir slope in the target area includes: Modal feature encoding is performed on each modal data point in the multimodal data of the reservoir slope and the displacement of the reservoir slope to obtain the feature codes of each modal data point. Then, each modal feature is mapped to a feature space of the same dimension through a linear mapping layer. The expression is as follows:
[0018] in, Indicates the first i Modal primitive features, This represents the encoded feature representation; Indicates the bias term; Represents a linear transformation; Based on the modal features mapped to the same dimensional feature space, a spatial topology is constructed according to the spatial coordinates of the sensors used to acquire these modal features. The expression is as follows:
[0019] in, Indicates sensor i and j Spatial correlation weights between them; p Indicates monitoring point nodes; This represents the Euclidean distance between the i-th sensor and the j-th sensor in three-dimensional space. The bandwidth parameter represents the kernel function; A graph convolutional network is used to perform feature propagation and aggregation on the spatial topology, extracting spatial correlation features between different monitoring points, as expressed in the following expression:
[0020] in, Represents the modal characteristic matrix. Indicates the convolution weights; This represents the spatial correlation weights between sensors.
[0021] In the above technical solution, the spatial layout relationship of the sensor is modeled by graph convolutional network (GCN), which enables the model to have spatial perception capability, recognize local sliding zone features and regional cooperative changes, and improve spatial consistency.
[0022] Furthermore, the process of predicting reservoir slope displacement based on the spatial correlation characteristics between various modal data of the reservoir slope includes: Based on the spatial correlation characteristics between various modal data of reservoir slope, a pre-trained multimodal depth prediction model is used to predict reservoir slope displacement. The multimodal depth prediction model includes a modal adaptive weight layer, a multi-timescale Transformer layer, and a reservoir slope displacement prediction layer. The modal adaptive weighting layer is used to dynamically calculate the contribution of each modality, and the expression is as follows:
[0023] in, Indicates the i-th modal feature The corresponding learnable weight vector transpose; Represents the j-th modal feature The transpose of the corresponding learnable weight vector; Weighted fusion is performed based on the contribution of each modality to obtain the weighted fusion feature, expressed as:
[0024] Based on the weighted fusion features of each modality's contribution, the multi-timescale Transformer layer simultaneously extracts the fast response features from short-term disturbances and the trend features from long-term trend information, and then fuses the short-term disturbance and long-term trend features, as expressed in the following expression:
[0025] in, and All of these represent trainable weight parameters. This indicates a rapid response characteristic in short-term disturbances. Indicates the trend characteristics within a long-term trend; The reservoir slope displacement prediction layer is based on a fusion of short-term disturbance and long-term trend characteristics. To predict reservoir slope displacement and calculate the losses incurred during the prediction process, the expression is as follows:
[0026] Where N represents the number of gravity gradient data samples used for reservoir slope displacement prediction; This indicates the predicted output displacement value; y Indicates the actual displacement; Represents the regularization coefficient; This indicates the prediction confidence level.
[0027] In the above technical solution, the dynamic allocation of the weights of each sensor data is realized through the modal adaptive weighting layer, which can automatically adjust the modal contribution according to the environmental conditions at different stages (such as the rainy season and the period of sudden water level change), thereby improving the prediction sensitivity and robustness. The multi-timescale Transformer layer with short-term and long-term dual branches is adopted to realize the joint modeling of fast disturbances and slow trends, which significantly improves the prediction stability and response speed.
[0028] Furthermore, the method further includes the following steps: An adaptive early warning threshold is set based on the reservoir slope displacement prediction results, expressed as follows:
[0029] in, Indicates the empirical benchmark threshold. This represents the predicted value of reservoir slope displacement. express The adjustment coefficient, This indicates the confidence level corresponding to the predicted displacement of the reservoir slope. express The adjustment coefficient; Based on the aforementioned adaptive early warning threshold Set early warning levels, including Level 1, Level 2, and Level 3, and send early warning information to the monitoring center in real time according to the warning level; The Level 1 warning indicates that the reservoir slope displacement shows a slight deformation trend, the Level 2 warning indicates that the reservoir slope displacement shows a continuous acceleration of deformation, and the Level 3 warning indicates that the reservoir slope displacement has a high risk of landslide.
[0030] In the above technical solution, by setting an adaptive early warning threshold based on the reservoir slope displacement prediction results, it is possible to achieve intelligent early warning with credibility perception, thereby reducing false alarms and missed alarms.
[0031] The second objective of this invention is: A reservoir slope displacement prediction system, the system comprising: The data acquisition module is used to acquire multimodal data of reservoir slopes in the target area; The data processing module is used to preprocess the multimodal data of the reservoir slope; The data processing module is also used to extract the spatial variation characteristics of the reservoir bank slope based on the preprocessed reservoir bank slope multimodal data, so as to determine the reservoir bank slope displacement based on the spatial variation characteristics of the reservoir bank slope. The data processing module is also used to extract spatial correlation features between different modal data of the reservoir slope in the target area based on the preprocessed multimodal data of the reservoir slope and the displacement of the reservoir bank slope. The reservoir slope displacement prediction module is used to predict the reservoir slope displacement based on the spatial correlation characteristics between various modal data of the reservoir slope, and output the prediction results.
[0032] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a method and system for predicting reservoir slope displacement. By preprocessing the acquired multimodal data of the reservoir slope, the reliability and stability of the data can be improved, thereby accelerating the efficiency of data calculation. By extracting the spatial variation characteristics of the reservoir slope to determine the reservoir slope displacement, the method can compensate for the shortcomings (errors) of single-frame detection, which can only reflect instantaneous displacement, and expand the information dimension of visual displacement detection from the time dimension, improving the comprehensiveness and accuracy of monitoring. By extracting the spatial correlation characteristics between the various modal data of the reservoir slope in the target area, the spatial perception ability is improved, so as to identify the local slip zone characteristics and regional coordinated changes, and improve spatial consistency. Finally, reservoir slope displacement is predicted based on the spatial correlation characteristics between the various modal data of the reservoir slope. It can combine visual prediction with deep learning to predict reservoir slope displacement, achieving high-precision, real-time, and reliable deformation trend prediction. Attached Figure Description
[0033] Figure 1 A flowchart illustrating the steps of a reservoir slope displacement prediction method provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the depth detection network provided in the embodiments of this application; Figure 3 This is a flowchart illustrating the data processing of the multimodal depth prediction model provided in the embodiments of this application. Figure 4 This is a schematic diagram of a reservoir slope displacement prediction system provided in an embodiment of this application. Detailed Implementation
[0034] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.
[0035] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.
[0036] Example: This embodiment provides a method for predicting reservoir slope displacement. (See also...) Figure 1 The method includes the following steps: Step S1: Obtain multimodal data of the reservoir slope in the target area and preprocess the multimodal data of the reservoir slope; Step S2: Extract the spatial variation characteristics of the reservoir bank slope based on the preprocessed multimodal data of the reservoir bank slope, and determine the displacement of the reservoir bank slope based on the spatial variation characteristics of the reservoir bank slope; Step S3: Based on the preprocessed multimodal data of the reservoir slope and the displacement of the reservoir bank slope, extract the spatial correlation features between the various modal data of the reservoir slope in the target area; Step S4: Based on the spatial correlation characteristics between the modal data of the reservoir slope, predict the displacement of the reservoir slope and output the prediction results.
[0037] In a preferred embodiment, step S1 involves preprocessing the multimodal data of the reservoir slope, including: The acquired multimodal data of the reservoir slope in the target area includes rainfall monitoring data, soil temperature and humidity data, infrasound and ground acoustic wave data, reservoir slope tilt angle data, reservoir water level data, slope seepage status data, phreatic line data, air temperature and humidity data, and visual displacement monitoring data. The acquired reservoir slope multimodal data were subjected to unified time synchronization, missing value imputation, outlier removal and normalization processing to obtain a reservoir slope multimodal dataset containing time series data.
[0038] Specifically, multiple types of sensors are deployed on the reservoir bank slope to collect data including: rainfall monitoring sensors (recording rainfall intensity and duration); soil temperature and humidity sensors (reflecting surface water content and temperature); infrasound and ground acoustic wave sensors (detecting micro-fracture activity and ground stress fluctuations); tilt sensors (monitoring changes in slope inclination angle); water level sensors (measuring reservoir water level changes and water pressure boundary conditions); pore water pressure sensors (reflecting slope seepage status); seepage line sensors (determining seepage depth and seepage line location); air humidity and temperature sensors (collecting meteorological conditions); and visual displacement monitoring sensors (capturing images of fixed reflective targets using near-infrared cameras). The data from all sensors are then synchronized in time, filled with missing values, removed outliers, and normalized to form a time-series multimodal dataset of the reservoir bank slope.
[0039] In a preferred embodiment, step S2, the process of extracting the spatial variation characteristics of the reservoir slope based on the preprocessed reservoir slope multimodal data, includes: Visual displacement monitoring data of reservoir slope multimodal data is acquired, wherein the visual displacement monitoring data includes fixed target image monitoring data of the bank slope; A preset depth detection network is used to extract the spatial coordinate data of the target bounding box as the reservoir bank slope shifts from the image monitoring data of the fixed target on the bank slope, and spatial coordinate data below the preset confidence threshold are removed to obtain the target spatial coordinate data that changes with the reservoir bank slope shift.
[0040] As a preferred embodiment, the process of determining the reservoir bank slope displacement based on the spatial variation characteristics of the reservoir bank slope includes: The initial target center coordinates are obtained based on the target spatial coordinate data that changes with the reservoir bank slope displacement, and the current target centroid coordinates are calculated using a preset sub-pixel edge model. The reservoir bank slope displacement is calculated based on the initial target center coordinates and the current target centroid coordinates.
[0041] Furthermore, the process of calculating the current target centroid coordinates using a preset sub-pixel edge model includes: A local quadratic surface is constructed based on the edge pixel data of the target bounding box, expressed as follows:
[0042] in, This represents the pixel edge coordinates of the target bounding box. This represents the parameters of the local quadratic surface that need to be solved; Based on local quadratic surface The coordinates of the sub-pixel edge points of the current target are calculated using a preset sub-pixel edge model, expressed as follows:
[0043] The sub-pixel edge coordinates of each current target are obtained by solving the problem. ); Based on the sub-pixel edge point coordinates of each current target Calculate the current target centroid coordinates The expression is:
[0044] in, This represents the weight of the i-th edge; represents the coordinates of the i-th sub-pixel edge point; r represents the radius of the target's circular structure.
[0045] Furthermore, the expression for calculating the reservoir bank slope displacement based on the initial target center coordinates and the current target centroid coordinates is as follows:
[0046] in, This represents the subpixel coordinates of the target circle's center in the initial reference frame.
[0047] Specifically, images of fixed reflective targets on the bank slope are obtained from the visual displacement monitoring data of the multimodal data of the reservoir slope. The preferred image resolution is 2492×1944. The acquired images are first processed by grayscale conversion and Gaussian blur denoising to suppress noise interference and improve the stability of subsequent edge features.
[0048] To achieve automated detection of fixed reflective targets, an improved YOLOv12 depth detection network is used, such as... Figure 2 As shown, the network introduces a multi-scale attention module in the backbone feature extraction layer to enhance the target edge response features; the CIOU loss function is used in the detection head layer to improve the localization accuracy. The network outputs the target bounding box coordinates and confidence scores, and automatically removes targets with confidence scores below a set threshold (0.6).
[0049] Within the detected target area, the initial target center position is obtained using the gray-scale centroid method, and then the current target centroid coordinates are accurately calculated based on a sub-pixel edge model fitted with a quadratic surface. The coordinate difference between the current frame and the initial frame is converted into the actual physical displacement of the reservoir bank slope using camera parameters. The unit is millimeters.
[0050] To extract the short-term dynamic trend of visual displacement, a temporal prediction model based on Convolutional Long Short-Term Memory (ConvLSTM) network and attention mechanism is adopted. The model input is a sequence of target displacements for several consecutive frames (preferably 10 frames). The ConvLSTM layer is responsible for capturing temporal dependencies, and the attention layer is used to enhance key frame information. The output is the displacement change trend curve for several future time moments. The displacement amount within several time moments is calculated based on the trend curve, thereby making up for the deficiency (error) of single-frame detection, which can only reflect instantaneous displacement. This expands the information dimension of visual displacement detection from the time dimension and improves the comprehensiveness and accuracy of monitoring.
[0051] The temporal prediction model based on Convolutional Long Short-Term Memory (ConvLSTM) network and attention mechanism finally outputs two-dimensional displacement components (X and Y directions). All results are normalized and then input into the next stage of the multimodal fusion model.
[0052] In this embodiment, the calculation of the current target centroid coordinates first involves representing the pixel data near the target boundary as a local surface:
[0053] in, For spatial pixel position, The parameters of the local quadratic surface that need to be solved are denoted as .
[0054] 1) Local window construction and least squares fitting: For each detected edge point, collect pixel gray values within its neighborhood (e.g., 5*5). The parameters are solved using the least squares method:
[0055] in, Represents the coordinates of the k-th pixel within the neighborhood of the edge point. and grayscale value This allows us to obtain a local quadratic surface model.
[0056] 2) Sub-pixel edge localization: Based on the gradient extrema of the quadratic surface grayscale distribution as the sub-pixel edge location, it can be solved by: ( It is a quadratic surface The gradient (with gradient extrema corresponding to sub-pixel edge positions) is obtained as follows:
[0057] Solving for subpixel coordinates ), that is, the precise edge point location.
[0058] 3) Calculation of sub-pixel centroid coordinates of the current target centroid coordinates: Use all sub-pixel edge point coordinates for geometric reconstruction of the target's circular structure. For circular targets, use weighted least squares circle fitting.
[0059] Using the residuals from edge points to the center of the circle as weights, the coordinates of the center of the circle are iteratively solved:
[0060] in, The weight of the i-th edge is determined by the residual from the edge point to the center of the circle and is used for weighted least squares fitting. Let r be the coordinates of the i-th sub-pixel edge point; r is the radius of the target circular structure.
[0061] Finally, the sub-pixel centroid coordinates of the target were obtained. .
[0062] 4) Calculation of reservoir bank slope displacement: By comparing the coordinates of the center of the current frame with the coordinates of the initial reference frame, the visual displacement (i.e., the displacement of the reservoir bank slope) can be calculated.
[0063] in, The sub-pixel coordinates of the target circle's center in the initial reference frame are used as the benchmark for displacement calculation.
[0064] In a preferred embodiment, step S3, the process of extracting the spatial correlation features between the modal data of the reservoir slope in the target area, includes: First, the sampling frequency of each modal data point in the reservoir slope multimodal data is unified with the reservoir slope displacement, and time alignment is performed. Then, modal feature encoding is performed to obtain the feature codes for each modal data point. Finally, each modal feature is mapped to a feature space of the same dimension through a linear mapping layer, expressed as:
[0065] in, Indicates the first i Modal primitive features, This represents the encoded feature representation; Indicates the bias term; This represents a linear transformation; the modal data includes, but is not limited to, monitoring data output from various sensors (rainfall, temperature and humidity, tilt angle, water level, pore pressure, seepage line, ground sound, air temperature and humidity, and visual displacement); Based on the modal features mapped to a feature space of the same dimension, and according to the spatial coordinates of the sensors used to acquire each modal feature... To construct the spatial topology, the expression is:
[0066] in, Indicates sensor i and j Spatial correlation weights between them; p This indicates monitoring point nodes, which include measurement points from sensors such as rainfall, soil temperature and humidity, tilt angle, water level, pore water pressure, phreatic line, and air temperature and humidity; This represents the Euclidean distance between the i-th sensor and the j-th sensor in three-dimensional space. The bandwidth parameter of the kernel function controls the elements of the spatial distance matrix. The extent of the impact The larger the value, the weaker the attenuation effect of spatial distance, and the more "smooth" the spatial correlation between sensors is reflected in the adjacency matrix; The smaller the value, the stronger the attenuation effect of spatial distance; only sensors that are extremely close together will exhibit strong correlation in the adjacency matrix. This can be achieved by adjusting... It can adapt to the topology modeling needs of sensor networks at different spatial scales; A graph convolutional network (GCN) is used to propagate and aggregate features of the spatial topology, extracting spatial correlation features between different monitoring points. The expression is as follows:
[0067] in, Represents the modal characteristic matrix. Indicates the convolution weights; The spatial correlation weights between sensors are used to characterize the spatial influence relationships between monitoring points. Used as input for multimodal depth prediction models.
[0068] Understandably, by modeling the spatial layout of sensors using graph convolutional networks (GCNs), the model gains spatial awareness, enabling it to identify local sliding zone features and regional cooperative changes, thereby improving spatial consistency.
[0069] In a preferred embodiment, in step S4, see... Figure 3 The process of predicting reservoir slope displacement based on the spatial correlation characteristics between various modal data of the reservoir slope includes: Based on the spatial correlation characteristics among different modal data of reservoir slope A pre-trained multimodal depth prediction model is used to predict reservoir slope displacement. The multimodal depth prediction model includes a modal adaptive weighting layer (MAWN), a multi-timescale Transformer layer, and a reservoir slope displacement prediction layer. To address the issue of varying importance of different modes under different operating conditions, a modal adaptive weighting mechanism is introduced using the aforementioned modal adaptive weighting layer. This mechanism dynamically calculates the contribution of each mode using a learnable weight vector, expressed as:
[0070] in, Indicates the i-th modal feature The corresponding learnable weight vector transpose is used to encode features for the i-th modality. A weighted calculation is performed to measure the importance contribution of this mode under the current operating conditions; Represents the j-th modal feature The transpose of the corresponding learnable weight vector is used to measure the importance contribution of the j-th mode under the current operating condition; Weighted fusion is performed based on the contribution of each modality to obtain the weighted fusion feature, expressed as:
[0071] This mechanism enables the model to automatically adjust modal weights under different conditions such as heavy rain, low water levels, or drought, thereby achieving dynamic adaptation. To simultaneously capture short-term disturbances and long-term trend information, based on the weighted fusion features of each modality's contribution, the multi-timescale Transformer layer is used to simultaneously extract fast response features from short-term disturbances and trend features from long-term trend information, and then the short-term disturbance and long-term trend features are fused. The expression is as follows:
[0072] in, and All of these represent trainable weight parameters. This indicates a rapid response characteristic in short-term disturbances. Indicates the trend characteristics within a long-term trend; Specifically, the multi-timescale Transformer layer adopts a dual-branch Swing Transformer structure; the short-term branch (inputting a recent time period (preferably 2 hours) monitoring sequence, extracting fast response features) (and long-term branches (inputting a monitoring sequence with a longer time window, preferably 24 hours, to extract trend features)) ); The reservoir slope displacement prediction layer integrates short-term disturbances with long-term trend characteristics. The reservoir slope displacement is predicted by applying a fully connected layer and activation function, and the loss generated during the prediction process is calculated. The expression is as follows:
[0073] Where N represents the number of gravity gradient data samples used for reservoir slope displacement prediction; This indicates the predicted output displacement value; y Indicates the actual displacement; Represents the regularization coefficient; This indicates the prediction confidence level.
[0074] This design can simultaneously optimize prediction accuracy and uncertainty assessment, providing a confidence basis for subsequent adaptive early warning.
[0075] Training and optimization of the multimodal deep prediction model: Supervised training was performed using historical monitoring samples, the preferred optimization algorithm was Adam, and the initial adaptive learning rate was set to 1×10⁻⁶. -4 Once the model training reaches the validation set error convergence, the parameters are fixed for online prediction.
[0076] Understandably, by using a modal adaptive weighting layer to dynamically allocate the weights of each sensor's data, the modal contribution can be automatically adjusted according to the environmental conditions at different stages (such as rainstorms or sudden changes in water levels), thereby improving prediction sensitivity and robustness. By employing a multi-timescale Transformer layer with short-term and long-term branches, joint modeling of rapid disturbances and slow trends can be achieved, significantly improving prediction stability and response speed.
[0077] In a preferred embodiment, the method further includes the following steps: Step S5: Set an adaptive early warning threshold based on the reservoir slope displacement prediction results. The expression is:
[0078] in, Indicates the empirical benchmark threshold. This represents the predicted value of reservoir slope displacement. express The adjustment coefficient, This indicates the confidence level corresponding to the predicted displacement of the reservoir slope. express The adjustment coefficient; this mechanism can automatically adjust the early warning sensitivity based on the model confidence level; Based on the aforementioned adaptive early warning threshold The system sets early warning levels. When the predicted displacement rate or cumulative displacement exceeds the threshold, the system classifies the warnings according to the following rules: Level 1 warning; Level 2 warning; Level 3 warning. The system then sends the warning information to the monitoring center in real time according to the warning level. The Level 1 warning indicates that the reservoir slope displacement shows a slight deformation trend, the Level 2 warning indicates that the reservoir slope displacement shows a continuous acceleration of deformation, and the Level 3 warning indicates that the reservoir slope displacement has a high risk of landslide.
[0079] Specifically, regarding the risk level determination, in order to achieve continuous safety assessment of the monitored targets, this invention divides the early warning level into three levels based on the stress characteristics and historical displacement statistics of the monitored objects, and sets a corresponding threshold for each level; the displacement range corresponding to each early warning level usually has obvious physical meaning, therefore the system can set the threshold according to the following principles: Level 1 Warning (Slight Deformation Trend): When the target displacement exceeds the normal fluctuation range but is still within a controllable range, the system determines it to be a Level 1 warning; the threshold at this time can be based on the average displacement value in historical monitoring. This setting is used to alert users to potential minor anomalies in trends.
[0080] Level 2 Warning (Continuous Accelerated Deformation): When the displacement significantly exceeds the normal operating range and the rate of change increases, the system determines it to be a Level 2 warning; the threshold for this level can be based on the average displacement. Alternatively, a setting of 30% to 50% of the material / structure's allowable deformation limit can be used to indicate the need for timely inspection and intervention.
[0081] Level 3 warning (high risk of landslide): When the displacement approaches or exceeds the structural safety limit, the system triggers a Level 3 warning. Such thresholds are usually determined based on the maximum allowable displacement of the structural design, engineering specifications, or critical instability conditions, such as reaching 70% to 90% of the maximum allowable deformation.
[0082] Understandably, setting adaptive early warning thresholds based on reservoir slope displacement prediction results can achieve intelligent early warning with credibility awareness, reducing false alarms and missed alarms.
[0083] In this embodiment, preprocessing the acquired multimodal data of the reservoir slope improves the reliability and stability of the data, thereby accelerating the efficiency of data computation. Extracting the spatial variation characteristics of the reservoir slope to determine its displacement compensates for the limitations (errors) of single-frame detection, which only reflects instantaneous displacement, and expands the information dimension of visual displacement detection from a temporal perspective, enhancing the comprehensiveness and accuracy of monitoring. Extracting the spatial correlation features between the various modal data of the reservoir slope in the target area improves spatial perception, enabling the identification of local slip zone features and regional coordinated changes, thus enhancing spatial consistency. Finally, based on the spatial correlation features between the various modal data of the reservoir slope, reservoir slope displacement prediction is performed. This combines visual prediction with deep learning-based reservoir slope displacement prediction, achieving high-precision, real-time, and reliable deformation trend prediction.
[0084] In this embodiment, a reservoir slope displacement prediction system is also provided, see [link to relevant documentation]. Figure 4 The system includes: The data acquisition module is used to acquire multimodal data of reservoir slopes in the target area; The data processing module is used to preprocess the multimodal data of the reservoir slope; The data processing module is also used to extract the spatial variation characteristics of the reservoir bank slope based on the preprocessed reservoir bank slope multimodal data, so as to determine the reservoir bank slope displacement based on the spatial variation characteristics of the reservoir bank slope. The data processing module is also used to extract spatial correlation features between different modal data of the reservoir slope in the target area based on the preprocessed multimodal data of the reservoir slope and the displacement of the reservoir bank slope. The reservoir slope displacement prediction module is used to predict the reservoir slope displacement based on the spatial correlation characteristics between various modal data of the reservoir slope, and output the prediction results.
[0085] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A method for predicting reservoir slope displacement, characterized in that, The method includes the following steps: Acquire multimodal data of reservoir slopes in the target area and preprocess the multimodal data of reservoir slopes; Based on the preprocessed multimodal data of the reservoir slope, the spatial variation characteristics of the reservoir slope are extracted, and the displacement of the reservoir slope is determined according to the spatial variation characteristics of the reservoir slope. Based on the preprocessed multimodal data of reservoir slope and the displacement of reservoir bank slope, spatial correlation features between the various modal data of reservoir slope in the target area are extracted; Based on the spatial correlation characteristics between various modal data of the reservoir slope, the displacement of the reservoir slope is predicted, and the prediction results are output.
2. The method for predicting reservoir slope displacement according to claim 1, characterized in that, The preprocessing process for the multimodal data of the reservoir slope includes: The acquired multimodal data of the reservoir slope in the target area includes rainfall monitoring data, soil temperature and humidity data, infrasound and ground acoustic wave data, reservoir slope tilt angle data, reservoir water level data, slope seepage status data, phreatic line data, air temperature and humidity data, and visual displacement monitoring data. The acquired reservoir slope multimodal data were subjected to unified time synchronization, missing value imputation, outlier removal and normalization processing to obtain a reservoir slope multimodal dataset containing time series data.
3. The method for predicting reservoir slope displacement according to claim 2, characterized in that, The process of extracting spatial variation characteristics of reservoir bank slopes based on preprocessed multimodal data includes: Visual displacement monitoring data of reservoir slope multimodal data is acquired, wherein the visual displacement monitoring data includes fixed target image monitoring data of the bank slope; A preset depth detection network is used to extract the spatial coordinate data of the target bounding box as the reservoir bank slope shifts from the image monitoring data of the fixed target on the bank slope, and spatial coordinate data below the preset confidence threshold are removed to obtain the target spatial coordinate data that changes with the reservoir bank slope shift.
4. The reservoir slope displacement prediction method according to claim 3, characterized in that, The process of determining the reservoir bank slope displacement based on the spatial variation characteristics of the reservoir bank slope includes: The initial target center coordinates are obtained based on the target spatial coordinate data that changes with the reservoir bank slope displacement, and the current target centroid coordinates are calculated using a preset sub-pixel edge model. The reservoir bank slope displacement is calculated based on the initial target center coordinates and the current target centroid coordinates.
5. The method for predicting reservoir slope displacement according to claim 4, characterized in that, The process of calculating the current target centroid coordinates using a preset sub-pixel edge model includes: A local quadratic surface is constructed based on the edge pixel data of the target bounding box, expressed as follows: in, This represents the pixel edge coordinates of the target bounding box. This represents the local quadratic surface parameters that need to be solved; Based on local quadratic surface The coordinates of the sub-pixel edge points of the current target are calculated using a preset sub-pixel edge model, expressed as follows: The sub-pixel edge coordinates of each current target are obtained by solving the problem. ); Based on the sub-pixel edge point coordinates of each current target Calculate the current target centroid coordinates The expression is: in, This represents the weight of the i-th edge; represents the coordinates of the i-th sub-pixel edge point; r represents the radius of the target's circular structure.
6. The method for predicting reservoir slope displacement according to claim 5, characterized in that, The expression for calculating the reservoir bank slope displacement based on the initial target center coordinates and the current target centroid coordinates is as follows: in, This represents the subpixel coordinates of the target circle's center in the initial reference frame.
7. The method for predicting reservoir slope displacement according to claim 6, characterized in that, The process of extracting spatial correlation features among different modal data of reservoir slopes in the target area includes: Modal feature encoding is performed on each modal data point in the multimodal data of the reservoir slope and the displacement of the reservoir slope to obtain the feature codes of each modal data point. Then, each modal feature is mapped to a feature space of the same dimension through a linear mapping layer. The expression is as follows: in, Indicates the first i Modal primitive features, This represents the encoded feature representation; Indicates the bias term; Represents a linear transformation; Based on the modal features mapped to the same dimensional feature space, a spatial topology is constructed according to the spatial coordinates of the sensors used to acquire these modal features. The expression is as follows: in, Indicates sensor i and j Spatial correlation weights between them; p Indicates monitoring point nodes; This represents the Euclidean distance between the i-th sensor and the j-th sensor in three-dimensional space. The bandwidth parameter represents the kernel function; A graph convolutional network is used to perform feature propagation and aggregation on the spatial topology, extracting spatial correlation features between different monitoring points, as expressed in the following expression: in, Represents the modal characteristic matrix. Indicates the convolution weights; This represents the spatial correlation weights between sensors.
8. The method for predicting reservoir slope displacement according to claim 7, characterized in that, The process of predicting reservoir slope displacement based on the spatial correlation characteristics among various modal data of reservoir slope includes: Based on the spatial correlation characteristics between various modal data of reservoir slope, a pre-trained multimodal depth prediction model is used to predict reservoir slope displacement. The multimodal depth prediction model includes a modal adaptive weight layer, a multi-timescale Transformer layer, and a reservoir slope displacement prediction layer. The modality adaptive weighting layer is used to dynamically calculate the contribution of each modality, and the expression is as follows: in, Indicates the i-th modal feature The corresponding learnable weight vector transpose; Represents the j-th modal feature The transpose of the corresponding learnable weight vector; Weighted fusion is performed based on the contribution of each modality to obtain the weighted fusion feature, expressed as: Based on the weighted fusion features of each modality's contribution, the multi-timescale Transformer layer simultaneously extracts the fast response features from short-term disturbances and the trend features from long-term trend information, and then fuses the short-term disturbance and long-term trend features, as expressed in the following expression: in, and All of these represent trainable weight parameters. This indicates a rapid response characteristic in short-term disturbances. Indicates the trend characteristics within a long-term trend; The reservoir slope displacement prediction layer is based on a fusion of short-term disturbance and long-term trend characteristics. To predict reservoir slope displacement and calculate the losses incurred during the prediction process, the expression is as follows: Where N represents the number of gravity gradient data samples used for reservoir slope displacement prediction; This indicates the output displacement prediction value; y Indicates the actual displacement; Represents the regularization coefficient; This indicates the prediction confidence level.
9. The method for predicting reservoir slope displacement according to claim 1, characterized in that, The method further includes the following steps: An adaptive early warning threshold is set based on the reservoir slope displacement prediction results, expressed as follows: in, Indicates the empirical benchmark threshold. This represents the predicted value of reservoir slope displacement. express The adjustment coefficient, This indicates the confidence level corresponding to the predicted displacement of the reservoir slope. express The adjustment coefficient; Based on the aforementioned adaptive early warning threshold Set early warning levels, including Level 1, Level 2, and Level 3, and send early warning information to the monitoring center in real time according to the warning level; The Level 1 warning indicates that the reservoir slope displacement shows a slight deformation trend, the Level 2 warning indicates that the reservoir slope displacement shows a continuous acceleration of deformation, and the Level 3 warning indicates that the reservoir slope displacement has a high risk of landslide.
10. A reservoir slope displacement prediction system, characterized in that, The system includes: The data acquisition module is used to acquire multimodal data of reservoir slopes in the target area; The data processing module is used to preprocess the multimodal data of the reservoir slope; The data processing module is also used to extract the spatial variation characteristics of the reservoir bank slope based on the preprocessed reservoir bank slope multimodal data, so as to determine the reservoir bank slope displacement based on the spatial variation characteristics of the reservoir bank slope. The data processing module is also used to extract spatial correlation features between different modal data of the reservoir slope in the target area based on the preprocessed multimodal data of the reservoir slope and the displacement of the reservoir bank slope. The reservoir slope displacement prediction module is used to predict the reservoir slope displacement based on the spatial correlation characteristics between various modal data of the reservoir slope, and output the prediction results.