Deep learning-based intelligent monitoring system and method for bank collapse risk of reservoir area

Abstract

The invention provides a deep learning-based intelligent monitoring system and method for reservoir bank collapse risk, and relates to the technical field of bank monitoring, and the method comprises the steps: obtaining and preprocessing all monitoring point data of a reservoir bank monitoring network, and constructing a total monitoring point set; screening key base points and quantifying the inherent risk characterization quantity of the key base points; a dynamic association network of the common monitoring points and the key base points is established, and the final clustering association degree of the common monitoring points relative to the key base points is obtained; calculating a radiation risk value of each key base point to the associated common monitoring points in combination with the inherent risk characterization quantity and the final clustering association degree; fusing the radiation risk values of all the key base points to obtain a comprehensive risk value of the common monitoring point; constructing a directed weighted graph by taking the key base point as a source point, and dividing risk communities by adopting a graph neural network community discovery algorithm; and based on the community comprehensive risk value statistical distribution and the highest risk level of the contained key base points, assigning a unified risk level to each monitoring point and outputting the risk level.

Description

Technical Field

[0001] This invention relates to the field of shoreline monitoring technology, specifically a smart monitoring system and method for reservoir shoreline collapse risk based on deep learning. Background Technology

[0002] As a core component of water conservancy projects, the stability of reservoir banks directly impacts reservoir operation, surrounding safety, and ecological sustainability. Current technologies for monitoring reservoir bank collapse risks have significant shortcomings, failing to meet the intelligent monitoring needs arising from the integration of the digital economy and high-end equipment manufacturing industries. Specifically, the following problems exist: Existing monitoring methods rely on manual intervention in many aspects, resulting in a cumbersome and inefficient process. They cannot achieve real-time dynamic monitoring of bank collapse risks or accurate risk level output, and thus cannot meet the core needs of smart water conservancy management.

[0003] At the same time, each monitoring point is generally regarded as an independent assessment unit, without the construction of a scientific dynamic correlation network of monitoring points, and without the effective characterization of risk transmission characteristics. This makes it impossible to accurately assess the radiation impact of key risk points on the surrounding monitoring areas, thereby affecting the comprehensiveness and accuracy of risk assessment.

[0004] In conclusion, existing monitoring technologies are insufficient for accurately classifying the risk of bank collapse in reservoir areas.

[0005] Therefore, this paper presents a deep learning-based intelligent monitoring system and method for reservoir bank collapse risk, which can accurately classify bank collapse risks and greatly improve the accuracy of bank collapse risk prevention and control. Summary of the Invention

[0006] To address the aforementioned technical problems, the present invention aims to provide a deep learning-based intelligent monitoring system and method for reservoir bank collapse risk, which can accurately classify bank collapse risks and greatly improve the precise prevention and control effect of bank collapse risks.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based intelligent monitoring method for reservoir bank collapse risk, comprising: Acquire shore area monitoring data and related data from all monitoring points in the reservoir shore monitoring network, preprocess the shore area monitoring data, and construct a total set of monitoring points; Based on the monitoring data and related data of the shore area, several key points with high or extremely high risk levels are selected from the total set of monitoring points according to the preset screening rules, and the inherent risk characterization of each key point is quantified. Based on each ordinary monitoring point in the total set of monitoring points except for the key base points, a dynamic association network between ordinary monitoring points and key base points is established. The direct association degree is calculated, the indirect association degree is determined by iteratively applying the transmission attenuation mechanism, and the chain association degree product operation is performed backtracking based on the risk transmission path to obtain the final cluster association degree of each ordinary monitoring point relative to each key base point. For each key baseline, combining the inherent risk characterization of the key baseline with the final clustering correlation of the corresponding ordinary monitoring point, a correction function, background risk attributes, and relevant coastal area data are introduced to calculate the radiation risk value that each key baseline radiates to its associated ordinary monitoring point individually. For any ordinary monitoring point, the radiation risk values ​​corresponding to all key baseline points are integrated to obtain the comprehensive risk value corresponding to the ordinary monitoring point; Using key baselines as the source, a directed weighted graph is constructed based on the final clustering correlation and risk transmission path. A community detection algorithm based on graph neural networks is used to divide risk communities. Finally, based on the statistical distribution of comprehensive risk values ​​within each community and the highest risk level of the key baselines contained therein, a unified risk level is assigned to each monitoring point and output.

[0008] Preferably, the risk levels include extremely high risk, high risk, medium risk, and low risk; Based on shoreline monitoring data, the process of selecting several key monitoring points with high or extremely high risk levels from the total set of monitoring points according to preset screening rules, and quantifying the inherent risk characteristics of each key monitoring point, includes: Based on the coastal area monitoring data and the preset screening rules, several candidate base points are selected from the total set of monitoring points to form a preliminary candidate base point set; based on each candidate base point in the preliminary candidate base point set, the initial inherent risk characterization quantity corresponding to the candidate base point is calculated through a pre-trained risk assessment model. Centered on the candidate base point, within a preset initial association radius, neighboring monitoring points are found to form a local influence domain of the candidate base point. Based on a preset risk radiation model and the initial inherent risk characterization of the candidate base point, the simulated risk values ​​of all neighboring monitoring points within the local influence domain are calculated. Based on the distribution of the simulated risk values ​​within the local influence domain, the risk level of the candidate base point is determined. Candidate base points with risk levels of high risk or extremely high risk are used as key base points, and the inherent risk characterization of each key base point is calculated.

[0009] Preferably, the step of establishing a dynamic correlation network between ordinary monitoring points and key base points based on each ordinary monitoring point in the total set of monitoring points (excluding key base points), and calculating the direct correlation degree and iteratively applying the propagation attenuation mechanism to determine the indirect correlation degree includes: Based on the set of key baselines and the set of ordinary monitoring points, a dynamic interconnected network topology is constructed with key baselines as the core and ordinary monitoring points as related nodes. Based on the constructed dynamic correlation network, the direct correlation degree between each key baseline and each ordinary monitoring point is calculated to generate an initial correlation matrix; Based on the strong correlation threshold, ordinary monitoring points with a direct correlation degree higher than the strong correlation threshold are assigned as the first layer of correlation points of the corresponding key base points; at the same time, from the first layer of correlation points, the points with the highest direct correlation degree and whose inherent risk characterization exceeds the secondary source threshold are selected and added to the candidate secondary radiation source set; the remaining points are added to the undetermined point set. Based on the preset promotion criteria, ordinary monitoring points that meet the promotion criteria are selected from the set of candidate secondary radiation sources as formal secondary radiation sources. The set of key base points and the formal secondary radiation sources are used together as the radiation source set for the current round. The indirect correlation between each point in the radiation source set and each point in the undetermined point set is calculated based on the conduction attenuation mechanism. In each iteration, the undetermined points whose calculated maximum indirect correlation is higher than the weak correlation threshold are assigned to the radiation source corresponding to the maximum indirect correlation, and their level and risk transmission path are recorded; at the same time, the points in the newly assigned undetermined points that meet the promotion conditions are added to the candidate secondary radiation source set for the next round of evaluation. The iteration terminates when the set of points to be determined no longer changes or when the preset maximum number of iterations is reached.

[0010] Preferably, the step of performing a chain-like correlation product operation based on risk transmission path backtracking to obtain the final clustering correlation degree of each ordinary monitoring point relative to each key base point includes: For each ordinary monitoring point that has been assigned, the risk transmission path is traced back to the key base point, and the direct or indirect correlation of each segment on the path is calculated by chain multiplication according to the transmission attenuation mechanism to obtain the final cluster correlation of the ordinary monitoring point relative to the key base point.

[0011] Preferably, the calculation of the radiation risk value radiated individually from each key baseline to its associated ordinary monitoring point includes: Based on the inherent risk characterization quantity corresponding to the key baseline and the final clustering correlation degree of ordinary monitoring points relative to the key baseline, the final clustering correlation degree is corrected based on a preset correction function to obtain the corrected correlation degree. Based on the corrected correlation degree and the corresponding inherent risk characterization quantity, the preliminary radiation risk value is calculated; Based on the background risk attributes of the ordinary monitoring points and the multiple correction coefficients corresponding to the relevant data in the coastal area, the preliminary radiation risk value is corrected to obtain the final radiation risk value.

[0012] Preferably, the step of fusing the radiation risk values ​​of all key baselines corresponding to any ordinary monitoring point to obtain the comprehensive risk value corresponding to the ordinary monitoring point includes: Obtain information on the risk transmission direction between each key baseline and its corresponding ordinary monitoring point; Based on any ordinary monitoring point, for each key base point that generates radiation risk value to the ordinary monitoring point, the corresponding risk transmission direction vector is determined in combination with the risk transmission direction information; Calculate the sum of the risk transmission direction vectors of all key reference points that generate radiation risk values ​​for the ordinary monitoring points, and obtain the resultant vector; Based on the obtained composite vector, the final comprehensive risk value of the ordinary monitoring point is calculated.

[0013] Preferably, assigning a uniform risk level to each monitoring point and outputting the following includes: Using key baselines as initial source nodes and ordinary monitoring points with determined final clustering correlations as subsequent nodes, a directed weighted graph is constructed based on hierarchical affiliation and risk transmission paths; the direction of the edges in the graph represents the risk transmission direction, and the weight represents the corresponding final clustering correlation. Based on the directed weighted graph, a community detection algorithm is used to identify a subset of nodes with tight internal connections and sparse external connections, and each subset constitutes a risk community. For each identified risk community, calculate the statistical characteristics of the comprehensive risk value of all nodes within the risk community; Based on the statistical characteristics and the highest risk level of the key nodes in the risk community, a dominant risk level is assigned to the risk community, and all nodes in the community inherit the dominant risk level.

[0014] A second aspect of the present invention also provides a deep learning-based intelligent monitoring system for reservoir bank collapse risk, comprising: The acquisition module acquires shore area monitoring data and related data from all monitoring points in the reservoir shore monitoring network, and preprocesses the shore area monitoring data to construct a total set of monitoring points. The identification module, based on the shore area monitoring data, selects several key points with a risk level of high risk or extremely high risk from the total set of monitoring points according to preset screening rules, and quantifies the inherent risk characterization of each key point. The analysis module establishes a dynamic correlation network between ordinary monitoring points and key base points based on each ordinary monitoring point in the total set of monitoring points, excluding key base points. It calculates the direct correlation degree, iteratively applies the transmission attenuation mechanism to determine the indirect correlation degree, and performs a chain-like correlation degree product operation based on the risk transmission path backtracking, thereby obtaining the final clustering correlation degree of each ordinary monitoring point relative to each key base point. The risk radiation calculation module, for each key base point, combines the inherent risk characterization of the key base point with the final clustering correlation degree of the corresponding ordinary monitoring point, introduces a correction function, background risk attributes and relevant coastal area data, and calculates the radiation risk value radiated by each key base point to its associated ordinary monitoring point. The comprehensive risk assessment module, for any ordinary monitoring point, integrates the radiation risk values ​​corresponding to all key baselines to obtain the comprehensive risk value corresponding to the ordinary monitoring point; The risk determination module uses key baselines as the source points and constructs a directed weighted graph based on the final cluster correlation degree and risk transmission path. It uses a community detection algorithm based on graph neural networks to divide risk communities. Finally, based on the statistical distribution of comprehensive risk values ​​within each community and the highest risk level of the key baselines contained therein, it assigns a unified risk level to each monitoring point and outputs it.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. Abandoning the blind deployment of existing technology monitoring points, we accurately locate high / extremely high risk key points through screening rules and pre-trained models, quantify their inherent risk characteristics, avoid wasting monitoring resources, and lay a reliable foundation for subsequent analysis.

[0016] 2. Innovatively construct a dynamic correlation network, and through direct / indirect correlation degree calculation and path backtracking, combined with the iterative upgrade of secondary radiation sources, fully capture the risk transmission characteristics and solve the pain point that existing technologies cannot assess the radiation impact of key base points.

[0017] 3. By introducing correction functions and background risk attributes, integrating the radiation effects of multiple key base points, and combining community detection algorithms to divide risk communities, regional collaborative assessment is achieved, overcoming the limitations of existing technologies in terms of single assessment and large bias.

[0018] 4. The entire process is automated without human intervention. It combines deep learning to achieve real-time dynamic monitoring and accurate output, adapting to the needs of smart water management in the context of the digital economy.

[0019] 5. By accurately classifying risks and identifying communities, it clarifies key prevention and control areas and risk spread trends, supports optimal resource allocation and early warning, effectively reduces disaster losses, and has extremely high application value. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, there are other drawings that can be obtained based on these drawings.

[0021] Figure 1 This is a schematic diagram of a deep learning-based intelligent monitoring system for reservoir bank collapse risk.

[0022] Figure 2 This is a schematic diagram of a deep learning-based intelligent monitoring method for reservoir bank collapse risk. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] Example 1 like Figure 1 As shown in the figure, this embodiment discloses a deep learning-based intelligent monitoring method for reservoir bank collapse risk, the method comprising: Acquire shore area monitoring data and related data from all monitoring points in the reservoir shore monitoring network, and preprocess the shore area monitoring data to construct a total set of monitoring points; in this embodiment, the shore area monitoring data includes high-precision latitude and longitude data, maximum deformation, deformation rate, and reservoir water level.

[0026] In this embodiment, the monitoring data of the shore area is preprocessed; specifically, this includes: High-precision latitude and longitude data are converted into an engineering plane rectangular coordinate system (X, Y) using Gauss-Kruger projection, and elevation data are converted into a unified elevation datum to obtain the three-dimensional coordinates (X, Y, Z) of the monitoring point.

[0027] The deformation time series data includes the maximum deformation amount and deformation rate. The deformation rate is calculated by displacement difference within a preset period (e.g., 24 hours) and high-frequency noise is eliminated by moving average filtering.

[0028] The time-series data of reservoir water level and the timestamps of deformation data are strictly synchronized.

[0029] In addition, data is derived from shoreline monitoring data, such as deformation-water level response coefficient, deformation acceleration, and spatiotemporal characteristics. Specifically, the deformation-water level response coefficient is calculated using each significant change in reservoir water level as a window, and the formula is set as follows: . The value represents the change in water level during the corresponding time period; the deformation acceleration is used to capture the acceleration trend of deformation; and the spatiotemporal characteristics are calculated by taking the normal distance from each point to the reservoir shoreline and the distance along the slope.

[0030] To eliminate the influence of different feature dimensions and orders of magnitude, Z-score standardization was then performed on each feature. Finally, a feature vector was constructed for each monitoring point at time t.

[0031] In detail, the relevant data for the shoreline area includes vegetation cover data, meteorological data, geographical data, and human activity data. Specifically, vegetation cover data includes vegetation coverage, vegetation type (trees / shrubs / herbs / bare land), and average vegetation height. Vegetation type is converted into numerical features using One-Hot encoding, and average vegetation height is obtained through inversion from lidar point cloud data, with the unit uniformly in meters. Meteorological data includes: average daily precipitation for the past 3 years, annual extreme rainfall days (daily precipitation ≥ 50 mm), average annual wind speed, and rainstorm intensity (maximum 1-hour rainfall). Geographical data includes: topographic slope, aspect, lithological type (igneous rock / sedimentary rock / metamorphic rock / loose deposits), and distance from fault zones. Human activity data includes: horizontal distance from roads / buildings and other artificial facilities, shoreline engineering disturbance intensity (0-1 quantization, 0 for no disturbance, 1 for severe construction disturbance), and aquaculture / farmland distribution markers (binarized). By adding relevant data from the shoreline area (multi-dimensional data on vegetation cover, meteorology, geography, and human activities), multi-source information can be integrated to effectively compensate for the limitations of single monitoring data and improve the accuracy of identifying the risk level of bank collapse in the reservoir area, the robustness of the model, and the targeting of early warnings.

[0032] Based on the monitoring data of the shore area, several key points with a risk level of high risk or extremely high risk are selected from the total set of monitoring points according to the preset screening rules, and the inherent risk characterization of each key point is quantified; the risk level includes extremely high risk, high risk, medium risk and low risk.

[0033] It should be noted that the process of selecting several key monitoring points with high or extremely high risk levels from the total set of monitoring points based on shoreline monitoring data and according to preset screening rules, and quantifying the inherent risk characterization of each key monitoring point, includes: Based on the monitoring data of the shore area and the preset screening rules, a number of candidate base points are selected from the total set of monitoring points to form a preliminary set of candidate base points; Based on each candidate base point in the preliminary candidate base point set, and based on the shore area monitoring data corresponding to the candidate base point, the initial inherent risk characterization quantity corresponding to the candidate base point is calculated through a pre-trained risk assessment model. Centered on the candidate base point, within a preset initial association radius, the neighboring monitoring points are found to form the local influence domain of the candidate base point; Within the local influence domain, based on the preset risk radiation model and the initial inherent risk characterization of the candidate base point, the simulated risk value of all neighboring points within the local influence domain is calculated, and the risk level of the candidate base point is determined based on the distribution of the simulated risk value within the local influence domain. Candidate base points with high or very high risk levels are selected as key base points and included in the final set of key base points. The inherent risk characterization is then recalculated for each key base point in the final set of key base points.

[0034] Specifically, there are various preset screening rules. For example, monitoring points that meet any two of the following criteria are selected into the preliminary candidate base point set: those whose deformation acceleration exceeds the standard for N consecutive cycles (e.g., 3 periods), whose deformation rate exceeds the standard, or whose deformation-water level response coefficient exceeds the standard.

[0035] The feature vectors of the candidate base points are used as input to the corresponding risk assessment model, and the output risk probability is the corresponding initial inherent risk representation. . Let be the i-th monitoring point in the candidate base point set C.

[0036] The local influence domain is based on the candidate base point Centered on, radius , The chief of the reservoir bank.

[0037] The preset risk radiation model adopts an exponential decay model. For monitoring points within the local influence domain, the calculation logic for the simulated risk value is as follows: In the formula, For point The simulated risk value, For point With point The actual distance between them along the slope This is the attenuation length constant. Let j be the j-th monitoring point within the local influence domain. With point The actual distance between them along the slope is through a point With point The three-dimensional coordinates are calculated by combining the actual slope of the reservoir area slope (fitted by the elevation difference and horizontal distance between adjacent monitoring points).

[0038] The risk level of the candidate baseline is determined by calculating the proportion of points within the local influence domain whose simulated risk values ​​exceed a threshold. This proportion determines the corresponding risk level: extremely high risk, high risk, medium risk, or low risk. If a point is not considered high or extremely high risk, it is downgraded to a regular monitoring point. The inherent risk characterization is calculated as follows: ; For the final set of key points, For the i-th monitoring point in the final set of key base points K, This is the inherent risk characterization of the i-th monitoring point in the final set of key reference points K. for Spatial effect index.

[0039] Based on each ordinary monitoring point in the total set of monitoring points except for the key base points, a dynamic association network between ordinary monitoring points and key base points is established. The indirect association degree is determined by calculating the direct association degree and iteratively applying the transmission attenuation mechanism. The chain association degree product operation is then performed backtracking based on the risk transmission path to obtain the final cluster association degree of each ordinary monitoring point relative to each key base point. It should be noted that, based on each ordinary monitoring point in the total set of monitoring points excluding key baselines, a dynamic association network is established between ordinary monitoring points and key baselines. The direct association degree is calculated, the indirect association degree is determined by iteratively applying a transmission attenuation mechanism, and a chain-like association degree product operation is performed based on the risk transmission path backtracking. This yields the final clustering association degree of each ordinary monitoring point relative to each key baseline, including: Based on the set of key baselines and the set of ordinary monitoring points, the direct correlation degree between each key baseline and each ordinary monitoring point is calculated, and an initial correlation matrix is ​​generated. Based on the strong correlation threshold, ordinary monitoring points with a direct correlation degree higher than the strong correlation threshold are assigned as the first layer of correlation points of the corresponding key base points; at the same time, from the first layer of correlation points, the points with the highest direct correlation degree and whose inherent risk characterization exceeds the secondary source threshold are selected and added to the candidate secondary radiation source set; the remaining points are added to the undetermined point set. Based on the preset promotion criteria, ordinary monitoring points that meet the promotion criteria are selected from the set of candidate secondary radiation sources as formal secondary radiation sources. The set of key base points and the formal secondary radiation sources are used together as the radiation source set for the current round. The indirect correlation between each point in the radiation source set and each point in the undetermined point set is calculated based on the conduction attenuation mechanism. In each iteration, the undetermined points whose calculated maximum indirect correlation is higher than the weak correlation threshold are assigned to the radiation source corresponding to the maximum indirect correlation, and their level and risk transmission path are recorded; at the same time, the points in the newly assigned undetermined points that meet the promotion conditions are added to the candidate secondary radiation source set for the next round of evaluation. The iteration terminates when the set of points to be determined no longer changes or the preset maximum number of iterations is reached. For each ordinary monitoring point that has been assigned, the risk transmission path is traced back to the key base point, and the direct or indirect correlation of each segment on the path is calculated by chain multiplication according to the transmission attenuation mechanism to obtain the final cluster correlation of the ordinary monitoring point relative to the key base point.

[0040] Specifically, the direct correlation is the direct correlation between monitoring points i and j, which is the product of the spatial factor, the characteristic factor, and the temporal factor. The spatial factor is equal to the negative power of the natural exponent, where the exponent is the distance between i and j along the slope. The square of the result is divided by (2 multiplied by the square of the spatial scale parameter σ_s). The feature factor is equal to the cosine similarity between the feature vector obtained by encoding the deformation rate and response coefficient of monitoring point i through a shallow neural network and the feature vector obtained by encoding the deformation rate and response coefficient of monitoring point j. The time factor is the larger of 0 and (1 minus the absolute value of the maximum correlation lag time between the deformation sequences of i and j divided by the maximum allowable lag).

[0041] Indirect correlation degree: The indirect correlation degree of point A to point C through secondary radiation source B is equal to the direct correlation degree between A and B multiplied by the interlayer conduction efficiency attenuation coefficient γ, and then multiplied by the direct correlation degree between B and C.

[0042] The final cluster association degree is the final association degree from the key node K to the target node P. For path The final correlation is not a simple product, but a discretization of the path integral concept: the direct correlation between K and S1 is multiplied by the product of the direct correlation between each pair of adjacent secondary radiation sources Sm and Sm+1 from S1 to Sn-1 and γ (multiplication), and then multiplied by the direct correlation between Sn and P. The integral result has a stronger attenuation effect on long paths. It is a secondary radiation source.

[0043] For each key baseline, combining the inherent risk characterization of the key baseline with the final clustering correlation of the corresponding ordinary monitoring point, a correction function, background risk attributes, and relevant coastal area data are introduced to calculate the radiation risk value that each key baseline radiates to its associated ordinary monitoring point individually. It should be noted that, for each key baseline, the calculation of the radiation risk value radiated by each key baseline to its associated ordinary monitoring point, by combining the inherent risk characterization of the key baseline with the final clustering correlation degree of the corresponding ordinary monitoring point, and by introducing a correction function, background risk attributes, and relevant coastal area data, includes: Based on the inherent risk characterization quantity corresponding to the key baseline and the final clustering correlation degree of ordinary monitoring points relative to the key baseline, the final clustering correlation degree is corrected based on a preset correction function to obtain the corrected correlation degree. Based on the corrected correlation degree and the corresponding inherent risk characterization quantity, the preliminary radiation risk value is calculated; Based on the background risk attributes of the ordinary monitoring points and the multiple correction coefficients corresponding to the relevant data in the coastal area, the preliminary radiation risk value is corrected to obtain the final radiation risk value.

[0044] Specifically, the correction function is a Sigmoid-like function with activation and saturation: In the formula To correct the correlation, The activation coefficient, This is the saturation threshold.

[0045] The preliminary radiation risk value is the product of the corrected correlation coefficient and the inherent risk characterization of the key baseline. Background risk attributes include the geological type of the area where the monitoring point is located, such as silty clay area, gravel area, etc., with different geological types assigned different background risk coefficients. Multiple correction coefficients include vegetation protection coefficient, rainstorm erosion coefficient, terrain stability coefficient, and human disturbance coefficient.

[0046] For any ordinary monitoring point, the radiation risk values ​​corresponding to all key baseline points are integrated to obtain the comprehensive risk value corresponding to the ordinary monitoring point; It should be noted that, for any ordinary monitoring point, the comprehensive risk value corresponding to that ordinary monitoring point, obtained by integrating the radiation risk values ​​of all key reference points, includes: Obtain information on the risk transmission direction between each key baseline and its corresponding ordinary monitoring point; Based on any ordinary monitoring point, for each key base point that generates radiation risk value to the ordinary monitoring point, the corresponding risk transmission direction vector is determined in combination with the risk transmission direction information; Calculate the sum of the risk transmission direction vectors of all key reference points that generate radiation risk values ​​for the ordinary monitoring points, and obtain the resultant vector; Based on the obtained composite vector, the final comprehensive risk value of the ordinary monitoring point is calculated.

[0047] Specifically, the risk transmission direction vector is obtained by... point to Unit geometric vector, based on key pivot point The principal strain direction of the local deformation field (estimated through displacement difference of neighboring points), the projection direction of gravitational acceleration on the slope (downhill direction), and three other angles are calculated and synthesized accordingly. Based on the obtained resultant vector, the final comprehensive risk value of the ordinary monitoring point is calculated by taking the length of the resultant vector. The resultant vector, which represents the direction and magnitude of the resultant force of the combined effect of multiple risk sources, is transformed into a scalar value that only represents the magnitude of the total risk intensity. This value is the direct basis for subsequent risk level classification and visualization.

[0048] Using key baselines as the source, a directed weighted graph is constructed based on the final clustering correlation and risk transmission path. A community detection algorithm based on graph neural networks is used to divide risk communities. Finally, based on the statistical distribution of comprehensive risk values ​​within each community and the highest risk level of the key baselines contained therein, a unified risk level is assigned to each monitoring point and output.

[0049] It should be noted that the process involves using key reference points as the source, constructing a directed weighted graph based on the final clustering correlation and risk transmission path, and employing a community detection algorithm based on graph neural networks to divide risk communities. Finally, based on the statistical distribution of the comprehensive risk value within each community and the highest risk level of the key reference points it contains, a unified risk level is assigned to each monitoring point, and the following is output: Using key baselines as initial source nodes and ordinary monitoring points with determined final clustering correlations as subsequent nodes, a directed weighted graph is constructed based on hierarchical affiliation and risk transmission paths; the direction of the edges in the graph represents the risk transmission direction, and the weight represents the corresponding final clustering correlation. Based on the directed weighted graph, a community detection algorithm based on graph neural networks is used to identify a subset of nodes with tight internal connections and sparse external connections, and each subset constitutes a risk community. For each identified risk community, calculate the statistical characteristics of the comprehensive risk value of all nodes within the risk community; Based on the statistical characteristics and the highest risk level of the key nodes in the risk community, a dominant risk level is assigned to the risk community, and all nodes in the community inherit the dominant risk level.

[0050] Specifically, the node set includes all final key baselines and ordinary monitoring points whose affiliations have been determined. Node attributes: Each node carries its spatial coordinates, final comprehensive risk value, inherent risk representation (specific to key baselines), and a deep feature vector generated by the preprocessing steps.

[0051] The core model of the community detection algorithm based on graph neural networks adopts a combination of a graph autoencoder framework and a graph attention network encoder. Specifically, the corresponding model architecture includes an input layer, a GAT encoder (first layer) for capturing local neighbor information and aggregating neighbor features based on attention weights, a GAT encoder (second layer) for fusing multiple attention features to enhance global node correlation, an average pooling layer, and a decoder. The specific logic of the GAT encoder is as follows: for a neighbor node j of node i, the original attention score is first calculated, then the original attention score is normalized to obtain attention weights, and neighbor node features are aggregated based on these attention weights to obtain the first-layer embedding vector of node i. A dual loss function of "reconstruction loss + community constraint loss" is used to train the entire model. After training, the output vector of the GAT encoder (second layer) is extracted as the input for subsequent clustering. Based on the output vector of the GAT encoder (second layer), community division is achieved through spectral clustering. The validity of the results is ensured through validity verification and anomaly handling. Specifically, spectral clustering involves: calculating the similarity between nodes using cosine similarity based on the output vector to form a corresponding similarity matrix; outputting a normalized Laplacian matrix based on the similarity matrix; reducing the dimensionality of the normalized Laplacian matrix based on the normalized Laplacian matrix and a preset number of communities to output a low-dimensional feature matrix; then, based on the low-dimensional feature matrix and the preset number of communities, assigning community labels to each node; and assigning the node to the community containing the nearest cluster center by calculating the Euclidean distance between the node and the cluster center. Closely connected clusters are quantized as those where the average Euclidean distance between nodes within the corresponding community is less than a preset first distance, while sparsely connected clusters are quantized as those where the average Euclidean distance between corresponding communities is greater than a preset second distance, and the preset first distance is less than the preset second distance. The statistical features include the mean, 90th percentile, and standard deviation of the hazard value.

[0052] like Figure 2 As shown in the figure, this embodiment also discloses a deep learning-based intelligent monitoring system for reservoir bank collapse risk, including: The acquisition module acquires shore area monitoring data and related data from all monitoring points in the reservoir shore monitoring network, and preprocesses the shore area monitoring data to construct a total set of monitoring points. The identification module, based on the shore area monitoring data, selects several key points with a risk level of high risk or extremely high risk from the total set of monitoring points according to preset screening rules, and quantifies the inherent risk characterization of each key point. The analysis module establishes a dynamic correlation network between ordinary monitoring points and key base points based on each ordinary monitoring point in the total set of monitoring points, excluding key base points. It calculates the direct correlation degree, iteratively applies the transmission attenuation mechanism to determine the indirect correlation degree, and performs a chain-like correlation degree product operation based on the risk transmission path backtracking, thereby obtaining the final clustering correlation degree of each ordinary monitoring point relative to each key base point. The risk radiation calculation module, for each key base point, combines the inherent risk characterization of the key base point with the final clustering correlation degree of the corresponding ordinary monitoring point, introduces a correction function, background risk attributes and relevant coastal area data, and calculates the radiation risk value radiated by each key base point to its associated ordinary monitoring point. The comprehensive risk assessment module, for any ordinary monitoring point, integrates the radiation risk values ​​corresponding to all key baselines to obtain the comprehensive risk value corresponding to the ordinary monitoring point; The risk determination module uses key baselines as the source points and constructs a directed weighted graph based on the final cluster correlation degree and risk transmission path. It uses a community detection algorithm based on graph neural networks to divide risk communities. Finally, based on the statistical distribution of comprehensive risk values ​​within each community and the highest risk level of the key baselines contained therein, it assigns a unified risk level to each monitoring point and outputs it.

[0053] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0054] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0055] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more electronic devices to execute all or part of the steps of the methods described in the various embodiments of this application.

[0056] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of the embodiments.

[0057] In the several embodiments provided in this application, it should be understood that the disclosed application can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0058] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0059] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0060] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A deep learning-based intelligent monitoring method for reservoir bank collapse risk, characterized in that, The method includes: Acquire shore area monitoring data and related data from all monitoring points in the reservoir shore monitoring network, preprocess the shore area monitoring data, and construct a total set of monitoring points; Based on the monitoring data of the coastal area, several key points with high or extremely high risk levels were selected from the total set of monitoring points according to the preset screening rules, and the inherent risk characterization of each key point was quantified. Based on each ordinary monitoring point in the total set of monitoring points except for the key base points, a dynamic association network between ordinary monitoring points and key base points is established. The direct association degree is calculated, the indirect association degree is determined by iteratively applying the transmission attenuation mechanism, and the chain association degree product operation is performed backtracking based on the risk transmission path to obtain the final cluster association degree of each ordinary monitoring point relative to each key base point. For each key baseline, combining the inherent risk characterization of the key baseline with the final clustering correlation of the corresponding ordinary monitoring point, a correction function, background risk attributes, and relevant coastal area data are introduced to calculate the radiation risk value that each key baseline radiates to its associated ordinary monitoring point individually. For any ordinary monitoring point, the radiation risk values ​​corresponding to all key baseline points are integrated to obtain the comprehensive risk value corresponding to the ordinary monitoring point; Using key baselines as the source, a directed weighted graph is constructed based on the final cluster correlation degree and risk transmission path, and a community detection algorithm based on graph neural networks is used to divide risk communities; Based on the statistical distribution of comprehensive risk values ​​within each community and the highest risk level of the key baselines contained therein, a unified risk level is assigned to each monitoring point and output.

2. The intelligent monitoring method for reservoir bank collapse risk based on deep learning according to claim 1, characterized in that, The risk levels include extremely high risk, high risk, medium risk, and low risk; Based on shoreline monitoring data, the process of selecting several key monitoring points with high or extremely high risk levels from the total set of monitoring points according to preset screening rules, and quantifying the inherent risk characteristics of each key monitoring point, includes: Based on the monitoring data of the shore area and the preset screening rules, a number of candidate base points are selected from the total set of monitoring points to form a preliminary set of candidate base points; Based on each candidate base point in the preliminary candidate base point set, the initial inherent risk representation quantity corresponding to the candidate base point is calculated through a pre-trained risk assessment model; Centered on the candidate base point, within a preset initial association radius, neighboring monitoring points are found to form a local influence domain of the candidate base point. Based on a preset risk radiation model and the initial inherent risk characterization of the candidate base point, the simulated risk values ​​of all neighboring monitoring points within the local influence domain are calculated. Based on the distribution of the simulated risk values ​​within the local influence domain, the risk level of the candidate base point is determined. Candidate base points with risk levels of high risk or extremely high risk are used as key base points, and the inherent risk characterization of each key base point is calculated.

3. The intelligent monitoring method for reservoir bank collapse risk based on deep learning according to claim 2, characterized in that, The process of establishing a dynamic correlation network between ordinary monitoring points and key base points based on each ordinary monitoring point in the total set of monitoring points (excluding key base points), and calculating the direct correlation degree and iteratively applying the propagation attenuation mechanism to determine the indirect correlation degree includes: Based on the set of key baselines and the set of ordinary monitoring points, a dynamic interconnected network topology is constructed with key baselines as the core and ordinary monitoring points as related nodes. Based on the constructed dynamic correlation network, the direct correlation degree between each key baseline and each ordinary monitoring point is calculated to generate an initial correlation matrix; Based on the strong correlation threshold, ordinary monitoring points with a direct correlation degree higher than the strong correlation threshold are assigned as the first layer of correlation points of the corresponding key base points; at the same time, from the first layer of correlation points, the points with the highest direct correlation degree and whose inherent risk characterization exceeds the secondary source threshold are selected and added to the candidate secondary radiation source set; the remaining points are added to the undetermined point set. Based on the preset promotion criteria, ordinary monitoring points that meet the promotion criteria are selected from the set of candidate secondary radiation sources as formal secondary radiation sources. The set of key base points and the formal secondary radiation sources are used together as the radiation source set for the current round. The indirect correlation between each point in the radiation source set and each point in the undetermined point set is calculated based on the conduction attenuation mechanism. In each iteration, the undetermined points whose calculated maximum indirect correlation is higher than the weak correlation threshold are assigned to the radiation source corresponding to the maximum indirect correlation, and their level and risk transmission path are recorded; at the same time, the points in the newly assigned undetermined points that meet the promotion conditions are added to the candidate secondary radiation source set for the next round of evaluation. The iteration terminates when the set of points to be determined no longer changes or when the preset maximum number of iterations is reached.

4. The intelligent monitoring method for reservoir bank collapse risk based on deep learning according to claim 3, characterized in that, The chain-like correlation product operation based on risk transmission path backtracking is used to obtain the final clustering correlation degree of each ordinary monitoring point relative to each key base point, including: For each ordinary monitoring point that has been assigned, the risk transmission path is traced back to the key base point, and the direct or indirect correlation of each segment on the path is calculated by chain multiplication according to the transmission attenuation mechanism to obtain the final cluster correlation of the ordinary monitoring point relative to the key base point.

5. The intelligent monitoring method for reservoir bank collapse risk based on deep learning according to claim 4, characterized in that, The calculation of the radiation risk value radiated individually from each key baseline point to its associated ordinary monitoring point includes: Based on the inherent risk characterization quantity corresponding to the key baseline and the final clustering correlation degree of ordinary monitoring points relative to the key baseline, the final clustering correlation degree is corrected based on a preset correction function to obtain the corrected correlation degree. Based on the corrected correlation degree and the corresponding inherent risk characterization quantity, the preliminary radiation risk value is calculated; Based on the background risk attributes of the ordinary monitoring points and the multiple correction coefficients corresponding to the relevant data in the coastal area, the preliminary radiation risk value is corrected to obtain the final radiation risk value.

6. The intelligent monitoring method for reservoir bank collapse risk based on deep learning according to claim 5, characterized in that, For any ordinary monitoring point, the comprehensive risk value corresponding to that ordinary monitoring point is obtained by integrating the radiation risk values ​​of all key reference points, including: Obtain information on the risk transmission direction between each key baseline and its corresponding ordinary monitoring point; Based on any ordinary monitoring point, for each key base point that generates radiation risk value to the ordinary monitoring point, the corresponding risk transmission direction vector is determined in combination with the risk transmission direction information; Calculate the sum of the risk transmission direction vectors of all key reference points that generate radiation risk values ​​for the ordinary monitoring points, and obtain the resultant vector; Based on the obtained composite vector, the final comprehensive risk value of the ordinary monitoring point is calculated.

7. The intelligent monitoring method for reservoir bank collapse risk based on deep learning according to claim 6, characterized in that, The process of assigning a uniform risk level to each monitoring point and outputting the following includes: Using key baselines as initial source nodes and ordinary monitoring points with determined final clustering correlations as subsequent nodes, a directed weighted graph is constructed based on hierarchical affiliation and risk transmission paths; the direction of the edges in the graph represents the risk transmission direction, and the weight represents the corresponding final clustering correlation. Based on the directed weighted graph, a community detection algorithm based on graph neural networks is used to identify a subset of nodes with tight internal connections and sparse external connections, and each subset constitutes a risk community. For each identified risk community, calculate the statistical characteristics of the comprehensive risk value of all nodes within the risk community; Based on the statistical characteristics and the highest risk level of the key nodes in the risk community, a dominant risk level is assigned to the risk community, and all nodes in the community inherit the dominant risk level.

8. A deep learning-based intelligent monitoring system for reservoir bank collapse risk, implementing the deep learning-based intelligent monitoring method for reservoir bank collapse risk as described in any one of claims 1 to 7, characterized in that, include: The acquisition module acquires shore area monitoring data and related data from all monitoring points in the reservoir shore monitoring network, and preprocesses the shore area monitoring data to construct a total set of monitoring points. The identification module, based on the shore area monitoring data, selects several key points with a risk level of high risk or extremely high risk from the total set of monitoring points according to preset screening rules, and quantifies the inherent risk characterization of each key point. The analysis module establishes a dynamic correlation network between ordinary monitoring points and key base points based on each ordinary monitoring point in the total set of monitoring points, excluding key base points. It calculates the direct correlation degree, iteratively applies the transmission attenuation mechanism to determine the indirect correlation degree, and performs a chain-like correlation degree product operation based on the risk transmission path backtracking, thereby obtaining the final clustering correlation degree of each ordinary monitoring point relative to each key base point. The risk radiation calculation module, for each key base point, combines the inherent risk characterization of the key base point with the final clustering correlation degree of the corresponding ordinary monitoring point, introduces a correction function, background risk attributes and relevant coastal area data, and calculates the radiation risk value radiated by each key base point to its associated ordinary monitoring point. The comprehensive risk assessment module, for any ordinary monitoring point, integrates the radiation risk values ​​corresponding to all key baselines to obtain the comprehensive risk value corresponding to the ordinary monitoring point; The risk determination module uses key baselines as the source points and constructs a directed weighted graph based on the final cluster correlation degree and risk transmission path. It uses a community detection algorithm based on graph neural networks to divide risk communities. Finally, based on the statistical distribution of comprehensive risk values ​​within each community and the highest risk level of the key baselines contained therein, it assigns a unified risk level to each monitoring point and outputs it.

Images