Water conservancy project dam body crack intelligent early warning system and method
Through structural perception attention variational denoising and timing convolutional network combined with dynamic semantic maps, the future trend of dam fractures in water conservancy engineering was predicted, and the problem of fracture feature extraction and multi-time and spatial scale evolution laws were solved in complex noise environments, and high-precision early warning decisions and automatic responses were achieved.
Patent Information
- Application Number
- CN202510638400.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-19
- Publication Date
- 2025-08-05
AI Technical Summary
The prior art is difficult to effectively extract the detailed characteristics of dam fractures in water conservancy engineering under complex noise environments, and the transparency and credibility of dynamic modeling and early warning decision-making mechanisms of crack evolution laws on multiple time and space scales are insufficient.
The structure-perceptual attention variation denoising method is used to suppress noise, combined with the time-space evolution law of cracks modeled by time-sequential convolutional network, the generation neural network predicts future crack trends through dynamic semantic structural maps and map enhancement, and SHAP interpretability analysis is introduced to construct a risk score function to realize pop-up warning and automatic response.
It significantly improves the accuracy and early warning timeliness of dam crack monitoring of water conservancy projects, and enhances the credibility and engineering applicability of early warning results.
Smart Images

Figure CN120431434A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of water conservancy project structure safety monitoring, and in particular to an intelligent early warning system and method for cracks in a water conservancy project dam body. Background Art
[0002] As the core infrastructure for ensuring flood control safety, hydropower generation and water resource regulation, the structural health monitoring and damage warning of water conservancy project dams are crucial to the safe operation of the projects. As a typical form of structural damage, the early and accurate detection and evolution trend prediction of dam cracks are key technical challenges in preventing dam break accidents.
[0003] The Chinese invention application with publication number CN119540130A discloses a wall crack detection system based on image recognition, including an image enhancement and denoising module, a geometric feature extraction module, a structural health prediction module, a crack classification and rating module, a crack extension monitoring module, a repair suggestion generation module, and a crack detection verification module. This image recognition-based wall crack detection system significantly improves the accuracy and real-time performance of wall crack detection by adopting crack geometric feature extraction, a structural health prediction model based on a graph convolutional network, and a crack extension monitoring module combining time series analysis and optical flow method, realizes trend prediction and adaptive feedback of crack extension, provides an efficient and intelligent solution for building structure health management, and effectively prevents potential structural failures.
[0004] In recent years, with the rapid development of multimodal sensing technology, deep learning, and dynamic graph reasoning, intelligent monitoring methods that integrate multi-source heterogeneous data such as acoustic emission, thermal imaging, and three-dimensional laser scanning have gradually become a research hotspot. However, the effective extraction of crack detail features in complex noise environments, the dynamic modeling of crack evolution laws at multiple spatiotemporal scales, and the improvement of the transparency and credibility of early warning decision-making mechanisms are still core issues that need to be broken through. Summary of the Invention
[0005] The purpose of the present invention is to address the problems existing in the background technology and to propose an intelligent early warning system and method for cracks in dam bodies of water conservancy projects.
[0006] The technical solution of the present invention is an intelligent early warning method for cracks in dam bodies of water conservancy projects, comprising the following specific implementation steps:
[0007] S1, periodically collecting multi-source image data of the dam body to form an image data set;
[0008] S2, based on the structure-aware attention variational denoising method, generates a structural response map to highlight the crack area, uses attention-weighted precoding to enhance the micro-crack features, adopts a variational autoencoder to suppress noise and maintain edge structure consistency, combines learnable weights to fuse channel residuals, and outputs a fused image;
[0009] S3. Construct a spatiotemporal sequence by fusing multi-source fusion images, use a temporal convolutional network to model the spatiotemporal evolution of cracks, generate semantic labels based on crack feature maps, introduce hidden states to enhance spatiotemporal context information, design constraint functions to quantify crack stability and interaction, and predict microcrack states.
[0010] S4. Map the microcrack state into a dynamic node containing spatial position, morphological characteristics and semantic energy level. Construct the structural-semantic fusion edge weight based on geometric distance, phase synchronization factor and semantic energy level difference. Evaluate the node risk through local excitation factor and integrate the attribute set to form a dynamic semantic structure map.
[0011] S5. Build a graph-enhanced generative neural network. By splitting node features into structural and semantic encodings, use a graph attention network to model topology and semantic energy levels respectively. Combined with a variational generation module, it infers potential evolutionary states and predicts future crack trends through time-aware decoding. Focusing on high-risk areas across the entire graph, the network is aggregated based on semantic energy level attention.
[0012] S6. Construct a risk scoring function, combine it with the SHAP method for interpretability analysis, and build an early warning response function to trigger pop-up warnings through dynamic thresholds and link automatic response plans with expert collaborative diagnosis.
[0013] Preferably, any image in the image data set includes three channel components: a grayscale image, an infrared image, and a thermal imaging image.
[0014] Preferably, the encoding process of enhancing microcrack features using attention weighted precoding is as follows:
[0015] S21, for each image I in the image dataset i Use shallow edge direction feature filter to generate structural response map:
[0016]
[0017] Among them, R i (x,y) represents the structural response intensity of the original image at the pixel position (x,y); represents the edge response along the direction θ = {0°, 45°, 90°, 135°}; I i (x,y) represents the pixel value of the original image at the pixel position (x,y); Represents the two-dimensional gradient vector of the image; represents the unit vector in the direction θ;
[0018] S22. Introducing structure-aware attention weight map A i (x,y), defined as:
[0019] Among them, A i (x,y) represents the normalized structural response, A i (x,y)∈[0,1];max(R i ) represents the maximum value in the response graph; ε represents a small constant;
[0020] Apply weights to the structural consistency weighted precoding to form a precoding graph
[0021]
[0022] Preferably, the generation process of the fused image is as follows:
[0023] A1. Use the variational autoencoder network guided by the image semantic structure to train the weighted image Perform low-dimensional semantic encoding and noise reconstruction suppression:
[0024] Coding phase:
[0025] Decoding stage:
[0026] Denoising loss function:
[0027] Among them, zi represents the image potential representation vector; represents the approximate posterior distribution of the encoder network output, parameterized as a neural network q φ ;φ represents the parameters of the encoder network; D θ represents the structure-aware encoder and decoder network, whose parameters are θ; represents the Gaussian distribution of the encoder output in the form of a latent space distribution; μ i and Represent the mean and variance vector of the image in the latent space respectively; represents the reconstructed image; Represents the input image, i.e. the pre-coded image; L denoise represents the denoising loss function; represents pixel reconstruction error; D KL (·) represents the variational KL divergence; represents the edge structure consistency loss; λ1 represents the degree of adjustment of semantic distribution balance; λ2 represents the adjustment of edge preservation strength; Represents the edge gradient of the reconstructed image; Represents image I i Medium crack structure response diagram R i Gradient operation of (x,y);
[0028] A2. Crack channel sensitive fusion, outputting the final fused image
[0029]
[0030] Among them, α, β, and λ represent channel fusion weights; CrackBoost(·) represents high-frequency enhancement of the heat map channel with the clearest crack details; Represent the reconstructed grayscale image, infrared image, and thermal imaging image respectively.
[0031] Preferably, the microcrack state prediction process is as follows:
[0032] S31, based on the acquired multi-source fusion image Construct crack spatiotemporal sequence image set Γ i :
[0033]
[0034] in, represents the fused image at time t; Γ i represents an image sequence;
[0035] The spatiotemporal sequence of cracks is modeled using a time-series convolutional neural network to capture the evolution of cracks in different time and space:
[0036] Among them, h t represents the hidden state at time t; h t-1 represents the hidden state at time t-1; f θ' (·) represents the time series modeling function;
[0037] S32, combined with crack characteristic diagram R i (x,y) and the spatial evolution process of the crack, based on the crack semantic label generation mechanism, output the crack semantic label:
[0038]
[0039] Among them, S i (x,y,t) represents the semantic label of the crack at time t and spatial position (x,y); Represents the crack feature map extracted by convolutional neural network;
[0040] The hidden state h t As the contextual information of the spatiotemporal sequence, it is used to enhance the semantic label:
[0041] S enh (x, y, t) = κ·S i (x,y,t)+(1-κ)·ht ;
[0042] Among them, κ represents the weight of controlling the fusion of semantic labels and spatiotemporal information; S enh (x, y, t) represents the enhanced semantic label;
[0043] S33. When constructing the mutual influence model between cracks, the constraint function of crack evolution is constructed:
[0044]
[0045] Among them, D(S enh (x,y,t)-S enh (x,y,t-1)) represents the measurement of the current crack label S enh (x,y,t) and the crack label S at the previous moment enh (x,y,t-1) changes; C(S enh (x, y, t)) represents the crack interaction metric based on the spatial position (x, y); λ2 represents the weight coefficient of the evolution constraint term; L evolution represents the constraint function;
[0046] S34. Combined with the evolution constraint calculation, the crack evolution prediction model is used to predict the evolution of microcracks. The model output is:
[0047]
[0048] in, represents the predicted microcrack state, i.e. the probability and evolution trend of cracks in time t and space (x, y); f θ” (·) represents the crack evolution prediction model, which is learned based on spatiotemporal features and evolution constraints.
[0049] Preferably, the dynamic semantic structure graph construction process is as follows:
[0050] S41. Node construction, semantic micro-crack node embedding: Each crack region is represented by its evolving semantic state. Based on this, we construct a dynamic node embedding vector
[0051]
[0052] The semantic energy level of each node is defined as a function of its evolution rate and potential field strength, that is, the node semantic energy level:
[0053] Among them, (x i ,y i ) indicates the center position of microcracks; represents the crack gradient; ε i (t) represents the node semantic energy level; φ i (t) represents the phase of the semantic potential field; Indicates the crack state change rate; ω1 and ω2 are adjustment coefficients;
[0054] S42. Edge construction, structural adjacency and semantically driven dynamic edge weights: A structural-semantic fusion edge set is constructed between micro-crack nodes. The edge weight is expressed as:
[0055]
[0056] in, represents the edge weight between node i and node j at time t; d ij represents the geometric distance between node i and node j; cos(φ i (t)-φ j (t)) represents the semantic phase synchronization factor; Indicates adjustable parameters;
[0057] S43. Define the local excitation factor of each node:
[0058] If γ i (t)>Tth, then node i is regarded as a high-energy semantically heterogeneous point;
[0059] Wherein, Tth represents the set threshold; N i represents the set of adjacent nodes of node i; j represents the adjacent nodes of node i; γ i (t) represents the local excitation factor of node i;
[0060] S44. Constructing a dynamic semantic structure graph:
[0061]
[0062] Among them, V t Represents a node set; represents the connection relationship and connection strength between crack regions i and j; F t ={ε i (t),φ i (t),γ i (t)} represents an attribute set.
[0063] Preferably, the future crack trend prediction process is as follows:
[0064] S51, each node v i ∈V t Represents a spatial microcrack point and constructs the initial input vector
[0065]
[0066] Among them, (x i ,y i ) represents node v i The two-dimensional space coordinates of Represents the predicted microcrack state at the node at time t; ε i (t) represents the node semantic energy level; φ i (t) represents the phase of the semantic potential field; γ i (t) represents the local excitation factor of node i; MLP init (·) represents a multilayer perceptron module with standardized input;
[0067] S52, split the representation of each node into two parts: structural encoding and semantic encoding, and pass them through the structural graph attention network GAT s and semantic graph attention network GAT σ To model:
[0068]
[0069] in, Indicates node structure encoding; Represents semantic encoding; N i represents the topological neighbor set of node i; GAT s (·) indicates structured attention network; GAT σ (·) represents the semantic attention network;
[0070] S53. Construct a variational generation module to infer the potential state of the node in the future from the current structure and semantic information:
[0071] First, the structural-semantic embedding is jointly encoded and its variational posterior distribution is derived:
[0072]
[0073] Then perform parameter resampling and output the potential node representation
[0074] Among them, z i represents the potential evolution representation of node i; μ i ,Σ i They represent the mean and covariance of the potential state of node i respectively; MLP(·) represents the parameter learning network, and the mean μ is obtained by multi-layer perceptron MLP learning. i and covariance Σ i , the input is structure embedding and semantic embedding splicing; N(·) represents normal distribution; ⊙ represents element-by-element multiplication; ε' represents the perturbation factor of standard normal distribution sampling;
[0075] S54, represent the potential node Decoding into microcrack state estimation at future time
[0076]
[0077] in, represents the crack state prediction value at the future time t+Δt; Δt represents the prediction time step; MLP evolve (·) represents the time-aware decoding network, which is used to simulate the semantic-structural coupling evolution trend.
[0078] Preferably, the process of constructing the aggregate representation of high-risk areas in the entire map is as follows:
[0079] Introducing attention weights based on semantic energy levels, we perform focused fusion on generated node representations, emphasizing high-energy-level risk areas:
[0080]
[0081]
[0082] Among them, α i represents the semantic attention weight of node i; w T represents the transpose of the learnable parameter vector; exp(·) represents the natural exponential; H risk It represents the aggregated representation of high-risk areas in the entire map, that is, the total comprehensive risk value of high-risk areas identified by the system; N represents the total number of sub-areas.
[0083] Preferably, the warning response function construction process is as follows:
[0084] S61. Map the crack prediction values and high-risk areas output by the graph-enhanced generative neural network (GE-GNN) into quantifiable risk indicators and construct a risk scoring function:
[0085]
[0086] Among them, R i represents the risk score of the i-th node; represents the crack prediction value of node i at the future moment; Represents the graph structure centrality of nodes in high-risk areas; N i represents the set of neighbor nodes of node i; w ij represents the graph connection weight between node i and its neighbor node j; β1, β2, and β3 represent risk contribution weight parameters;
[0087] S62. Explain the contribution of input features to prediction results in the GE-GNN model through SHAP values:
[0088]
[0089] in, represents the SHAP explanation value of the k-th feature on the i-th node; represents the kth input feature of node i;
[0090] S63. Construct an early warning response function:
[0091] in, represents the warning probability of node i; represents the contribution value of the most significant factor in the interpretation; σ(·) represents the Sigmoid function; η1 and η2 represent the weights of risk and explanation factors; τ represents the manually set trigger threshold.
[0092] The technical solution of the present invention is an intelligent early warning system for cracks in a dam body of a water conservancy project, which is used to implement an intelligent early warning method for cracks in a dam body of a water conservancy project as described in the claims, comprising:
[0093] Multimodal sensing and acquisition module, used to deploy sensor networks, periodically collect multi-source image data of key areas of the dam body, and perform standardization and time synchronization preprocessing;
[0094] The multimodal data preprocessing module is used to suppress noise interference and preserve crack details based on the structure-aware attention variational denoising method. It uses structural consistency constraints and crack-sensitive channel enhancement strategies to ultimately output a multi-source fused image with a high signal-to-noise ratio through channel residual fusion.
[0095] A microcrack evolution semantic field construction module is used to model the spatiotemporal sequence of fused images using a temporal convolutional network, combine crack feature maps with hidden state enhancement semantic labels, and introduce a crack evolution constraint function to predict crack state changes;
[0096] A dynamic semantic structure graph construction module is used to map the predicted crack state into dynamic nodes. Based on the geometric distance between cracks, semantic energy level differences, and phase synchronization factors, the edge weights of the structure-semantic coupling are constructed to form a dynamic semantic structure graph for characterizing the crack aggregation state and its evolution trend.
[0097] The graph-enhanced generative neural network module is used to extract node features by splitting the graph attention network of structural encoding and semantic encoding, combining it with the variational generation module to predict future crack states and use semantic energy-level attention to focus on high-risk areas;
[0098] The multi-warning decision-making and human-computer interaction module is used to perform risk scoring and SHAP interpretability analysis based on the output of the graph-enhanced generative neural network, build a dynamic warning response function, and realize pop-up warning, automatic response plan generation and expert collaborative diagnosis functions.
[0099] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects:
[0100] This paper designs an intelligent early warning system and method for cracks in dam bodies of water conservancy projects. By integrating technologies such as multimodal sensor networks, structure-aware variational denoising, dynamic semantic graph reasoning, and generative neural networks, it significantly improves the accuracy of monitoring cracks in dam bodies of water conservancy projects and the timeliness of early warning.
[0101] The multimodal data fusion method effectively integrates the features of grayscale, infrared and thermal imaging channels, suppresses noise interference while fully retaining the detailed features of cracks; the temporal convolutional network is combined with the dynamic semantic graph to accurately model the spatiotemporal evolution law of cracks, and through the construction of node and edge weights of structure-semantic coupling, it reveals the crack aggregation state and its interaction mechanism; the graph-enhanced generative neural network splits the structural code and semantic code, combines the variational generation module to predict the future crack propagation trend, and focuses on high-risk areas based on the semantic energy level attention mechanism; SHAP interpretability analysis is introduced to quantify the decision-making basis of the model, and a closed-loop decision-making system of pop-up warning, automatic response plan generation and expert collaborative diagnosis is realized through the risk scoring function and dynamic threshold trigger mechanism, which improves the robustness of the monitoring system while enhancing the credibility and engineering applicability of the warning results. BRIEF DESCRIPTION OF THE DRAWINGS
[0102] Figure 1 This is a system architecture diagram of an intelligent early warning system for dam cracks in water conservancy projects proposed by the present invention;
[0103] Figure 2 This is a flow chart of a method for intelligent early warning of cracks in dam bodies of water conservancy projects proposed by the present invention. DETAILED DESCRIPTION
[0104] Example 1, as Figure 1 As shown, the present invention proposes an intelligent early warning system for cracks in dam bodies of water conservancy projects, comprising: a multimodal perception and acquisition module, a multimodal data preprocessing module, a microcrack evolution semantic field construction module, a dynamic semantic structure graph construction module, a graph enhancement generation neural network module, and a multi-warning decision-making and human-computer interaction module.
[0105] The multimodal perception and acquisition module deploys a sensor network (including but not limited to acoustic emission thermal imagers, visual cameras, and 3D laser scanners) to periodically collect multi-source image data (including grayscale, infrared, and thermal imaging channels) of key areas of the dam body and perform standardization and time synchronization preprocessing;
[0106] The multimodal data preprocessing module is based on the structure-aware attention variational denoising method (SAVD). It uses structural consistency constraints and crack-sensitive channel enhancement strategies to suppress noise interference and preserve crack details. Finally, it outputs a multi-source fused image with a high signal-to-noise ratio through channel residual fusion.
[0107] The microcrack evolution semantic field construction module uses a temporal convolutional network (TCN) to model the spatiotemporal sequence of fused images, combines crack feature maps with hidden state enhancement semantic labels, and introduces crack evolution constraint functions (spatiotemporal stability and spatial interaction) to predict crack state changes.
[0108] The dynamic semantic structure graph construction module maps the predicted crack state into dynamic nodes (including spatial position, morphological characteristics, semantic energy level and phase information). Based on the geometric distance between cracks, semantic energy level difference and phase synchronization factor, it constructs the edge weight of structure-semantic coupling to form a dynamic semantic structure graph, which is used to characterize the crack aggregation state and its evolution trend.
[0109] The graph-enhanced generative neural network module constructs a graph-enhanced generative neural network (GE-GNN). It extracts node features by splitting the graph attention network into structural encoding and semantic encoding, combines it with the variational generation module to predict future crack states, and uses semantic energy-level attention to focus on high-risk areas.
[0110] The multi-warning decision-making and human-computer interaction module is based on the risk score and SHAP interpretability analysis (causal chain diagram and heat map) output by GE-GNN, constructs a dynamic warning response function (comprehensive risk score, explanatory factor contribution and threshold trigger mechanism), realizes pop-up warning, automatic response plan generation and expert collaborative diagnosis functions, and forms a closed-loop intelligent warning and decision-making system.
[0111] Example 2, as Figure 2 As shown, the present invention proposes an intelligent early warning method for cracks in a dam body of a water conservancy project, which is applied to the intelligent early warning system for cracks in a dam body of a water conservancy project proposed in Example 1. The specific implementation steps are as follows:
[0112] S1, the multimodal perception acquisition module periodically collects multi-source image data of the key areas of the dam through the deployed sensor network (including but not limited to acoustic emission thermal imagers, crack width visual cameras, and three-dimensional laser scanning) to form an image dataset Image t ={I1,I2,…,I i ,…,I t};
[0113] Where i = 1, 2, ..., t; t represents the total number of images in the image set; I i Represents the image dataset Imaget Any image I i ;
[0114] It should be noted that each image I i Includes three channel components: grayscale image Infrared image Thermal imaging images To unify the size and format, standardization is first performed: the image is scaled to a uniform resolution W×H; the pixel intensity is normalized to [0,1]; the image is aligned and time-synchronized to form a standard image tensor set.
[0115] S2, the multimodal data preprocessing module, is based on the image structure semantics-guided attention variational denoising method (SAVD), which integrates structural consistency constraints and crack-sensitive channel enhancement strategies to ensure that structural crack details are preserved during image noise removal. The specific implementation process is as follows:
[0116] S21, image crack structure response map generation: for each image I i ∈I t , a shallow edge direction feature filter (structure tensor filtering + guided edge direction detection) is used to generate a structural response map:
[0117]
[0118] Among them, R i (x,y) represents the structural response intensity of the original image at the pixel position (x,y). The larger the value, the more significant the point is in the significant structural area. represents the edge response along multiple directions (set to 0°, 45°, 90°, and 135° in this embodiment), i.e., the direction-guided gradient operator; i (x,y) represents the pixel value of the original image at the pixel position (x,y); Represents the two-dimensional gradient vector of the image; represents the unit vector in the direction θ;
[0119] S22, structural consistency attention weighted precoding:
[0120] Introducing structure-aware attention weight map A i (x,y), defined as:
[0121] Among them, A i (x,y) represents the normalized structural response, highlighting the microcrack area, A i (x,y)∈[0,1];max(R i ) represents the maximum value (normalization factor) in the response graph; ε represents a small constant (such as 10 -6 ), to prevent division by zero;
[0122] Apply weights to the structural consistency weighted precoding to form a precoding graph
[0123]
[0124] S23, semantically guided variational denoising unit (SAVD): using a variational autoencoder network (Structure-Aware VAE) guided by image semantic structure to denoise weighted images Perform low-dimensional semantic encoding and noise reconstruction suppression:
[0125] Coding phase:
[0126] Decoding stage:
[0127] Denoising loss function:
[0128] Among them, z i Represents the image potential representation vector; represents the approximate posterior distribution of the encoder network output, parameterized as a neural network q φ ;φ represents the parameters of the encoder network; D θ represents the structure-aware encoder and decoder network, whose parameters are θ; represents the Gaussian distribution of the encoder output in the form of a latent space distribution; μ i and Represent the mean and variance vectors of the image in the latent space, respectively, predicted by the encoder network and used to construct the Gaussian distribution; represents the reconstructed image; Represents the input image, i.e. the pre-coded image; L denoise represents the denoising loss function; represents pixel reconstruction error; D KL (·) represents the variational KL divergence; represents the edge structure consistency loss; λ1 represents the degree of adjustment of semantic distribution balance; λ2 represents the adjustment of edge preservation strength; Represents the edge gradient of the reconstructed image; Represents image I i Medium crack structure response diagram R i Gradient operation of (x,y);
[0129] S24, crack channel sensitive fusion: Since different channels (grayscale, infrared, thermal map) in the dam crack image have different representations of cracks, the final fused image is output based on the crack channel residual enhancement fusion method.
[0130] Among them, α, β, and λ represent channel fusion weights (which can be learned); CrackBoost(·) represents high-frequency enhancement of the heat map channel with the clearest crack details; Represent the reconstructed grayscale image, infrared image, and thermal imaging image respectively.
[0131] S3, the microcrack evolution semantic field construction module introduces the crack evolution semantic field (MESF), combines the semantic information of the image with the spatiotemporal evolution characteristics, and predicts and models the evolution trend of microcracks through a deep learning model. The specific implementation process is as follows:
[0132] S31, based on the acquired multi-source fusion image Construct crack spatiotemporal sequence image set Γ i :
[0133]
[0134] in, represents the fused image at time t; Γ i Represents the image sequence, i.e., the evolution information of the crack during the time change process;
[0135] The temporal and spatial sequence of cracks is modeled based on a temporal convolutional neural network (TCN) to capture the evolution of cracks in different time and space:
[0136] Among them, h t represents the hidden state at the tth moment, indicating the evolution characteristics of the crack; h t-1 represents the hidden state at time t-1; f θ' (·) represents the timing modeling function, designed based on TCN;
[0137] S32, based on the time-space sequence, combined with the crack feature map R i (x,y) and the spatial evolution process of the crack, based on the crack semantic label generation mechanism, output the crack semantic label:
[0138]
[0139] Among them, S i (x, y, t) represents the semantic label of the crack at time t and spatial position (x, y), that is, whether the point is a crack and has an evolution trend; Represents the crack feature map extracted by convolutional neural network;
[0140] The hidden state h t As an additional input for crack label generation, the hidden state ht As the contextual information of the spatiotemporal sequence, it is used to enhance the semantic label:
[0141] S enh (x, y, t) = κ·S i (x,y,t)+(1-κ)·h t ;
[0142] Among them, κ represents the weight of controlling the fusion of semantic labels and spatiotemporal information; S enh (x, y, t) represents the enhanced semantic label;
[0143] S33. Based on the spatiotemporal characteristics and semantic information of cracks, a model of the mutual influence between cracks is constructed, and a constraint function for crack evolution is designed:
[0144]
[0145] Among them, D(S enh (x,y,t)-S enh (x,y,t-1)) represents the measurement of the current crack label S enh (x,y,t) and the crack label S at the previous moment enh The change of (x,y,t-1) reflects the stability of crack evolution; C(S enh (x, y, t)) represents the crack interaction metric based on the spatial position (x, y), that is, the interaction and propagation between cracks; λ2 represents the weight coefficient of the evolution constraint term; L evolution represents the constraint function;
[0146] It should be noted that (x, y) in the constraint function represents each pixel point or spatial position in the image; therefore, the summation range covers the entire image space, that is, all pixels in the image; in this embodiment, the image size is H×W, then x∈[1,H], y∈[1,W]; for each pixel point (x, y), the relevant information of crack evolution is calculated and accumulated;
[0147] S34. Combined with evolution constraint calculation, use crack evolution prediction model (including but not limited to Transformer and graph neural network) to predict microcrack evolution. The model output is:
[0148]
[0149] in, represents the predicted microcrack state, i.e. the probability and evolution trend of cracks in time t and space (x, y); f θ” (·) represents the crack evolution prediction model, which is learned based on spatiotemporal features and evolution constraints.
[0150] S4. The dynamic semantic structure graph construction module maps the predicted microcrack state into dynamic semantic nodes. By combining the structural topology (including but not limited to the connectivity between cracks and the tension conduction path) with the semantic energy level (including but not limited to the crack development stage and semantic weight), a semantic-structural coupling dynamic graph is constructed. The construction process is as follows:
[0151] S41. Node construction, semantic micro-crack node embedding: Each crack region is represented by its evolving semantic state. Based on this, we construct a dynamic node embedding vector Including but not limited to spatial location, morphological features, and semantic level:
[0152] The semantic energy level of each node is defined as a function of its evolution rate and potential field strength, that is, the node semantic energy level:
[0153] Among them, (x i ,y i ) indicates the center position of microcracks; The crack gradient indicates its expansion trend; ε i (t) represents the node semantic energy level; φ i (t) represents the phase of the semantic potential field, which measures its evolution direction; Indicates the crack state change rate, that is, the growth trend; ω1 and ω2 are adjustment coefficients, balancing the time change and spatial potential intensity;
[0154] S42. Edge construction, structural adjacency and semantically driven dynamic edge weights: A structural-semantic fusion edge set is constructed between micro-crack nodes. The edge weight is expressed as:
[0155]
[0156] in, represents the edge weight between node i and node j at time t; d ij represents the geometric distance between node i and node j; cos(φ i (t)-φ j (t)) represents the semantic phase synchronization factor; represents an adjustable parameter that controls distance attenuation and energy level sensitivity;
[0157] S43. Define the local excitation factor of each node:
[0158] Among them, N i represents the set of adjacent nodes of node i; j represents the adjacent nodes of node i; γ i(t) represents the local excitation factor of node i;
[0159] If γ i (t)>Tth (the set threshold), then node i is regarded as a high-energy semantic heterogeneous point;
[0160] S44. Constructing a dynamic semantic structure graph:
[0161]
[0162] Among them, V t Represents a node set; represents the connection relationship and connection strength between crack regions i and j; F t ={ε i (t),φ i (t),γ i (t)} represents an attribute set.
[0163] S5. Graph-Enhanced Generative Neural Network Module: A graph-enhanced generative neural network (GE-GNN) is constructed. Based on the dynamic graph, it introduces a semantic energy level attention mechanism, a structural variation generation module, and a temporal evolution director to jointly realize the recognition of crack aggregation and the prediction of future evolution state. The specific implementation process is as follows:
[0164] S51, each node v i ∈V t Represents a spatial microcrack point, extracts its spatial position, crack state and semantic energy level features to construct the initial input vector
[0165]
[0166] Among them, (x i ,y i ) represents node v i The two-dimensional space coordinates of Represents the predicted microcrack state at the node at time t; ε i (t) represents the node semantic energy level; φ i (t) represents the phase of the semantic potential field; γ i (t) represents the local excitation factor of node i; MLP init (·) represents a multilayer perceptron module with standardized input;
[0167] S52, split the representation of each node into two parts: structural encoding and semantic encoding, and pass them through the structural graph attention network GAT s and semantic graph attention network GAT σTo model:
[0168]
[0169] in, Represents node structure encoding, and only uses topological structure for propagation and aggregation; Represents semantic coding, focusing on the semantic energy level evolution trend of the node; N i represents the topological neighbor set of node i; GAT s (·) indicates a structural attention network that only considers edge weights; GAT σ (·) represents the semantic attention network, which only considers the mutual influence between semantic energy levels;
[0170] S53. Construct a variational generation module to infer the potential state of the node in the future from the current structure and semantic information:
[0171] First, the structural-semantic embedding is jointly encoded and its variational posterior distribution is derived:
[0172]
[0173] Among them, z i represents the potential evolution representation of node i; μ i ,Σ i denote the mean and covariance of the potential state of node i respectively; MLP(·) denotes the parameter learning network, which uses the multi-layer perceptron (MLP) to learn the mean μ i and covariance Σ i , the input is structure embedding and semantic embedding splicing; N(·) represents normal distribution;
[0174] Then perform parameter resampling and output the potential node representation
[0175] Where ⊙ represents element-by-element multiplication; ε' represents the perturbation factor of the standard normal distribution sampling;
[0176] S54, represent the potential node Decoding into microcrack state estimation at future time
[0177]
[0178] in, represents the crack state prediction value at the future time t+Δt; Δt represents the prediction time step; MLP evolve (·) represents the time-aware decoding network, which is used to simulate the semantic-structural coupling evolution trend;
[0179] S55. Introduce attention weights based on semantic energy levels to perform focused fusion of generated node representations, emphasizing high-energy-level risk areas:
[0180]
[0181] Among them, α i represents the semantic attention weight of node i; w T represents the transpose of the learnable parameter vector; exp(·) represents the natural exponential; H risk It represents the aggregated representation of high-risk areas in the entire map, that is, the total comprehensive risk value of high-risk areas identified by the system; N represents the total number of sub-areas, that is, the N sub-areas into which the entire monitoring coverage area is divided.
[0182] S6, the multi-warning decision-making and human-computer interaction module, is based on the high-risk area identification results of GE-GNN crack prediction, and combines interpretable risk visualization with a human-computer collaborative warning mechanism. It integrates the graph reasoning results with a dynamic interpretation model, multimodal risk factor fusion, and collaborative interaction mechanism. The specific implementation process is as follows:
[0183] S61. Map the crack prediction values and high-risk areas output by the GE-GNN model into quantifiable risk indicators and construct a risk scoring function:
[0184]
[0185] Among them, R i represents the risk score of the i-th node (region); represents the crack prediction value of node i at the future moment; Represents the graph structure centrality of nodes in high-risk areas; N i represents the set of neighbor nodes of node i; w ij represents the graph connection weight between node i and its neighbor node j; β1, β2, and β3 represent risk contribution weight parameters, which are determined based on historical data training;
[0186] S62. Explain the contribution of input features to the prediction results in the GE-GNN model through SHAP values (Shapley Additive Explanations):
[0187]
[0188] in, represents the SHAP explanation value of the k-th feature on the i-th node; represents the kth input feature of node i;
[0189] Based on this, the interpretation values are superimposed on the equipment topology map through a heat map, and a factor contribution bar chart and a predicted causal chain diagram are generated to assist engineers in understanding the causes of crack development.
[0190] It should be noted that SHAP (Shapley Additive Explanations) is a model interpretation method based on game theory. It quantifies the impact of each input feature on the model output by calculating the marginal contribution of each feature in all possible feature combinations, thereby providing globally consistent and locally accurate explanations for complex models (such as deep neural networks and graph neural networks). Its core concept is derived from Shapley value theory and has the advantages of unique feature allocation, fairness, and model independence.
[0191] S63. Construct an early warning response function:
[0192] in, represents the warning probability of node i; represents the contribution value of the most significant factor in the interpretation; σ(·) represents the Sigmoid function; η1 and η2 represent the weights of risk and interpretation factors; τ represents the manually set trigger threshold;
[0193] like A pop-up warning prompt will be displayed, guiding the engineer to choose the following actions:
[0194] Automatically execute response plans (including but not limited to load adjustment and dispatching inspection tasks);
[0195] View the cause-and-effect diagram for further analysis;
[0196] Upload expert diagnosis results and the system will learn and update parameters.
[0197] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.
Claims
1. An intelligent early warning method for cracks in dam bodies of water conservancy projects, characterized in that: The specific implementation steps include the following: S1, periodically collecting multi-source image data of the dam body to form an image data set; S2, based on the structure-aware attention variational denoising method, generates a structural response map to highlight the crack area, uses attention-weighted precoding to enhance the micro-crack features, adopts a variational autoencoder to suppress noise and maintain edge structure consistency, combines learnable weights to fuse channel residuals, and outputs a fused image; S3. Construct a spatiotemporal sequence by fusing multi-source fusion images, use a temporal convolutional network to model the spatiotemporal evolution of cracks, generate semantic labels based on crack feature maps, introduce hidden states to enhance spatiotemporal context information, design constraint functions to quantify crack stability and interaction, and predict microcrack states. S4. Map the microcrack state into a dynamic node containing spatial position, morphological characteristics and semantic energy level. Construct the structural-semantic fusion edge weight based on geometric distance, phase synchronization factor and semantic energy level difference. Evaluate the node risk through local excitation factor and integrate the attribute set to form a dynamic semantic structure map. S5. Build a graph-enhanced generative neural network. By splitting node features into structural and semantic encodings, use a graph attention network to model topology and semantic energy levels respectively. Combined with a variational generation module, it infers potential evolutionary states and predicts future crack trends through time-aware decoding. Focusing on high-risk areas across the entire graph, the network is aggregated based on semantic energy level attention. S6. Construct a risk scoring function, combine it with the SHAP method for interpretability analysis, and build an early warning response function to trigger pop-up warnings through dynamic thresholds and link automatic response plans with expert collaborative diagnosis.
2. The intelligent early warning method for dam cracks in a water conservancy project according to claim 1 is characterized in that: Any image in the image dataset includes three channel components: grayscale image, infrared image, and thermal imaging image.
3. The intelligent early warning method for dam cracks in a water conservancy project according to claim 2 is characterized in that: The encoding process of enhancing microcrack features using attention-weighted precoding is as follows: S21, for each image I in the image dataset i Use shallow edge direction feature filter to generate structural response map: Among them, R i (x,y) represents the structural response intensity of the original image at the pixel position (x,y); represents the edge response along the direction θ = {0°, 45°, 90°, 135°}; I i (x,y) represents the pixel value of the original image at the pixel position (x,y); Represents the two-dimensional gradient vector of the image; represents the unit vector in the direction θ; S22. Introducing structure-aware attention weight map A i (x,y), defined as: Among them, A i (x,y) represents the normalized structural response, A i (x,y)∈[0,1];max(R i ) represents the maximum value in the response graph; ε represents a small constant; Apply weights to the structural consistency weighted precoding to form a precoding graph 4. The intelligent early warning method for dam cracks in a water conservancy project according to claim 3 is characterized in that: The generation process of the fused image is as follows: A1. Use the variational autoencoder network guided by the image semantic structure to train the weighted image Perform low-dimensional semantic encoding and noise reconstruction suppression: Coding phase: Decoding stage: Denoising loss function: Among them, z i Represents the image potential representation vector; represents the approximate posterior distribution of the encoder network output, parameterized as a neural network q φ ;φ represents the parameters of the encoder network; D θ represents the structure-aware encoder and decoder network, whose parameters are θ; represents the Gaussian distribution of the encoder output in the form of a latent space distribution; μ i and Represent the mean and variance vector of the image in the latent space respectively; represents the reconstructed image; Represents the input image, i.e. the pre-coded image; L denoise represents the denoising loss function; represents pixel reconstruction error; D KL (·) represents the variational KL divergence; represents the edge structure consistency loss; λ1 represents the degree of adjustment of semantic distribution balance; λ2 represents the adjustment of edge preservation strength; Represents the edge gradient of the reconstructed image; Represents image I i Medium crack structure response diagram R i Gradient operation of (x,y); A2. Crack channel sensitive fusion, outputting the final fused image Among them, α, β, and λ represent channel fusion weights; CrackBoost(·) represents high-frequency enhancement of the heat map channel with the clearest crack details; Represent the reconstructed grayscale image, infrared image, and thermal imaging image respectively.
5. The intelligent early warning method for dam cracks in a water conservancy project according to claim 4 is characterized in that: The microcrack state prediction process is as follows: S31, based on the acquired multi-source fusion image Construct crack spatiotemporal sequence image set Γ i : in, represents the fused image at time t; Γ i represents an image sequence; The spatiotemporal sequence of cracks is modeled using a time-series convolutional neural network to capture the evolution of cracks in different time and space: Among them, h t represents the hidden state at time t; h t-1 represents the hidden state at time t-1; f θ' (·) represents the time series modeling function; S32, combined with crack characteristic diagram R i (x,y) and the spatial evolution process of the crack, based on the crack semantic label generation mechanism, output the crack semantic label: Among them, S i (x,y,t) represents the semantic label of the crack at time t and spatial position (x,y); Represents the crack feature map extracted by convolutional neural network; The hidden state h t As the contextual information of the spatiotemporal sequence, it is used to enhance the semantic label: S enh (x,y,t)=κ·S i (x,y,t)+(1-κ)·h t ; Among them, κ represents the weight of controlling the fusion of semantic labels and spatiotemporal information; S enh (x, y, t) represents the enhanced semantic label; S33. When constructing the mutual influence model between cracks, the constraint function of crack evolution is constructed: Among them, D(S enh (x,y,t)-S enh (x,y,t-1)) represents the measurement of the current crack label S enh (x,y,t) and the crack label S at the previous moment enh (x,y,t-1) changes; C(S enh (x, y, t)) represents the crack interaction metric based on the spatial position (x, y); λ2 represents the weight coefficient of the evolution constraint term; L evolution represents the constraint function; S34. Combined with the evolution constraint calculation, the crack evolution prediction model is used to predict the evolution of microcracks. The model output is: in, represents the predicted microcrack state, i.e. the probability and evolution trend of cracks in time t and space (x, y); f θ″ (·) represents the crack evolution prediction model, which is learned based on spatiotemporal features and evolution constraints.
6. The intelligent early warning method for dam cracks in a water conservancy project according to claim 5, characterized in that: The process of constructing a dynamic semantic structure graph is as follows: S41. Node construction, semantic micro-crack node embedding: Each crack region is represented by its evolving semantic state. Based on this, we construct a dynamic node embedding vector The semantic energy level of each node is defined as a function of its evolution rate and potential field strength, that is, the node semantic energy level: Among them, (x i ,y i ) indicates the center position of microcracks; represents the crack gradient; ε i (t) represents the semantic energy level of the node; φ i (t) represents the phase of the semantic potential field; Indicates the crack state change rate; ω1 and ω2 are adjustment coefficients; S42. Edge construction, structural adjacency and semantically driven dynamic edge weights: A structural-semantic fusion edge set is constructed between micro-crack nodes. The edge weight is expressed as: in, represents the edge weight between node i and node j at time t; d ij represents the geometric distance between node i and node j; cos(φ i (t)-φ j (t)) represents the semantic phase synchronization factor; Indicates adjustable parameters; S43. Define the local excitation factor of each node: If γ i (t)>Tth, then node i is regarded as a high-energy semantically heterogeneous point; Wherein, Tth represents the set threshold; N i represents the set of adjacent nodes of node i; j represents the adjacent nodes of node i; γ i (t) represents the local excitation factor of node i; S44. Constructing a dynamic semantic structure graph: Among them, V t Represents a node set; represents the connection relationship and connection strength between crack regions i and j; F t ={ε i (t),φ i (t),γ i (t)} represents an attribute set.
7. The intelligent early warning method for dam cracks in a water conservancy project according to claim 6, characterized in that: The process of predicting future crack trends is as follows: S51, each node v i ∈V t Represents a spatial microcrack point and constructs the initial input vector Among them, (x i ,y i ) represents node v i The two-dimensional space coordinates of Represents the predicted microcrack state at the node at time t; ε i (t) represents the semantic energy level of the node; φ i (t) represents the phase of the semantic potential field; γ i (t) represents the local excitation factor of node i; MLP init (·) represents a multilayer perceptron module with standardized input; S52, split the representation of each node into two parts: structural encoding and semantic encoding, and pass them through the structural graph attention network GAT s and semantic graph attention network GAT σ To model: in, Indicates node structure encoding; Represents semantic encoding; N i represents the topological neighbor set of node i; GAT s (·) indicates structured attention network; GAT σ (·) represents the semantic attention network; S53. Construct a variational generation module to infer the potential state of the node in the future from the current structure and semantic information: First, the structural-semantic embedding is jointly encoded and its variational posterior distribution is derived: Then perform parameter resampling and output the potential node representation Among them, z i represents the potential evolution representation of node i; μ i ,Σ i They represent the mean and covariance of the potential state of node i respectively; MLP(·) represents the parameter learning network, and the mean μ is obtained by multi-layer perceptron MLP learning. i and covariance Σ i , the input is structure embedding and semantic embedding splicing; N(·) represents normal distribution; ⊙ represents element-by-element multiplication; ε' represents the perturbation factor of standard normal distribution sampling; S54, represent the potential node Decoding into microcrack state estimation at future time in, represents the crack state prediction value at the future time t+Δt; Δt represents the prediction time step; MLP evolve (·) represents the time-aware decoding network, which is used to simulate the semantic-structural coupling evolution trend.
8. The intelligent early warning method for dam cracks in a water conservancy project according to claim 7, characterized in that: The process of constructing the aggregated representation of high-risk areas across the entire map is as follows: Introducing attention weights based on semantic energy levels, we perform focused fusion on generated node representations, emphasizing high-energy-level risk areas: Among them, α i represents the semantic attention weight of node i; w T represents the transpose of the learnable parameter vector; exp(·) represents the natural exponential; H risk It represents the aggregated representation of high-risk areas in the entire map, that is, the total comprehensive risk value of high-risk areas identified by the system; N represents the total number of sub-areas.
9. The intelligent early warning method for dam cracks in a water conservancy project according to claim 8, characterized in that: The process of constructing the early warning response function is as follows: S61. Map the crack prediction values and high-risk areas output by the graph-enhanced generative neural network (GE-GNN) into quantifiable risk indicators and construct a risk scoring function: Among them, R i represents the risk score of the i-th node; represents the crack prediction value of node i at the future moment; Represents the graph structure centrality of nodes in high-risk areas; N i represents the set of neighbor nodes of node i; w ij represents the graph connection weight between node i and its neighbor node j; β1, β2, and β3 represent risk contribution weight parameters; S62. Explain the contribution of input features to prediction results in the GE-GNN model through SHAP values: in, represents the SHAP explanation value of the k-th feature on the i-th node; represents the kth input feature of node i; S63. Construct an early warning response function: in, represents the warning probability of node i; represents the contribution value of the most significant factor in the interpretation; σ(·) represents the Sigmoid function; η1 and η2 represent the weights of risk and explanation factors; τ represents the manually set trigger threshold.
10. An intelligent early warning system for cracks in a dam body of a water conservancy project, which is used to implement the intelligent early warning method for cracks in a dam body of a water conservancy project according to any one of claims 1 to 9, characterized in that: include: Multimodal sensing and acquisition module, used to deploy sensor networks, periodically collect multi-source image data of key areas of the dam body, and perform standardization and time synchronization preprocessing; The multimodal data preprocessing module is used to suppress noise interference and preserve crack details based on the structure-aware attention variational denoising method. It uses structural consistency constraints and crack-sensitive channel enhancement strategies to ultimately output a multi-source fused image with a high signal-to-noise ratio through channel residual fusion. A microcrack evolution semantic field construction module is used to model the spatiotemporal sequence of fused images using a temporal convolutional network, combine crack feature maps with hidden state enhancement semantic labels, and introduce a crack evolution constraint function to predict crack state changes; A dynamic semantic structure graph construction module is used to map the predicted crack state into dynamic nodes. Based on the geometric distance between cracks, semantic energy level differences, and phase synchronization factors, the edge weights of the structure-semantic coupling are constructed to form a dynamic semantic structure graph for characterizing the crack aggregation state and its evolution trend. The graph-enhanced generative neural network module is used to extract node features by splitting the graph attention network of structural encoding and semantic encoding, combining it with the variational generation module to predict future crack states and use semantic energy-level attention to focus on high-risk areas; The multi-warning decision-making and human-computer interaction module is used to perform risk scoring and SHAP interpretability analysis based on the output of the graph-enhanced generative neural network, build a dynamic warning response function, and realize pop-up warning, automatic response plan generation and expert collaborative diagnosis functions.
Citation Information
Patent Citations
Wall crack detection system based on image recognition
CN119540130A
Cited By
Dam body monitoring and early warning method and system
CN120948479A
Silicon carbide part stress distribution monitoring and crack risk prediction method
CN121051557A
A method for monitoring stress distribution and predicting crack risk of a silicon carbide component
CN121051557B
Lemon health analysis method
CN121053535A
