Dynamic risk assessment method for underground engineering geological safety based on multi-source data fusion

Abstract

The present invention discloses a dynamic risk assessment method for underground engineering geological safety based on multi-source data fusion, which relates to the field of risk assessment technology. The method includes collecting multi-source heterogeneous data related to underground engineering and extracting implicit information, modeling underground engineering geological safety risk factors as a risk network, calculating the comprehensive importance of nodes based on the Stacking integration algorithm, and identifying key risk factors; using the spatiotemporal characteristics of implicit information to obtain characteristic parameters of key risk factors, introducing a random walk mechanism to obtain a dynamic accident prediction chain, and using a graph attention network to learn and represent the probability distribution of accident evolution paths; assessing the vulnerability of connecting edges in the risk network, establishing a dynamic risk assessment model based on node importance and edge vulnerability, and obtaining the dynamic risk value corresponding to the accident based on the probability of accident occurrence and risk mitigation factor. The present invention realizes intelligent identification, dynamic assessment, and precise early warning of risk factors, improving the accuracy and real-time performance of risk identification and assessment.

Description

Technical Field

[0001] The present invention relates to the technical field of risk assessment, and more specifically, to a dynamic risk assessment method for underground engineering geological safety based on multi-source data fusion. Background Art

[0002] Underground projects face complex geological conditions and dynamically changing safety risks during construction and operation. Traditional geological safety risk assessment methods primarily rely on single data sources or static analysis, such as geological exploration reports, field monitoring data, or empirical models. These methods fail to fully reflect the dynamic changes in risk during construction and operation. With the advancement of information technology, multi-source data fusion technology has provided a new solution for underground engineering geological safety risk assessment. It can integrate multi-dimensional information such as geology, construction, monitoring, and the environment to achieve dynamic risk identification and assessment. Traditional underground engineering geological safety risk assessment primarily relies on the following methods: geological exploration and empirical models, single-source monitoring data analysis, and static assessment models. These methods are limited by isolated data and delayed updates, making them incapable of dynamic risk perception and early warning, increasing the probability of engineering accidents such as landslides, water inrush, and rockbursts.

[0003] Multi-source data fusion technology significantly improves the accuracy and real-time nature of risk assessment by integrating data from diverse sources and scales (such as geological exploration data, InSAR remote sensing monitoring, real-time sensor data, construction logs, and environmental factors), combined with machine learning, deep learning, and Bayesian networks. The geological conditions of underground projects are inherently hidden, mutating, and uncertain, making traditional static assessment methods inadequate for project safety management. Therefore, the challenge of implementing dynamic risk assessment methods for underground engineering geological safety based on multi-source data is urgent, providing more scientific and reliable technical support for the safe construction and operation of underground projects. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes a dynamic risk assessment method for underground engineering geological safety based on multi-source data fusion. By integrating multi-dimensional data such as geological exploration, construction monitoring, and environmental factors, and combining complex network theory with machine learning algorithms, dynamic assessment and accurate early warning of risk factors are achieved.

[0005] A first aspect of the present invention provides a method for dynamic risk assessment of underground engineering geological safety based on multi-source data fusion, comprising the following steps:

[0006] Collecting multi-source heterogeneous data related to underground engineering, preprocessing the multi-source heterogeneous data, extracting implicit information from the preprocessed multi-source heterogeneous data, and analyzing the spatiotemporal characteristics of the implicit information;

[0007] Obtain underground engineering geological safety risk factors, model the underground engineering geological safety risk factors into a risk network, calculate the comprehensive importance of nodes in the risk network based on the Stacking integration algorithm, and identify key risk factors;

[0008] The spatial and temporal characteristics of implicit information are used to obtain the characteristic parameters of key risk factors. A random walk mechanism is introduced to obtain a dynamic accident prediction chain. The graph attention network is used to learn and represent the dynamic accident prediction chain to obtain the probability distribution of the accident evolution path.

[0009] Assess the vulnerability of the connecting edges in the risk network, establish a dynamic risk assessment model based on the comprehensive importance of nodes and edge vulnerability, obtain the dynamic risk value corresponding to the accident according to the probability of accident occurrence and risk mitigation factor, and visualize underground engineering accidents and dynamic risk values.

[0010] In this solution, multi-source heterogeneous data related to underground engineering is collected, pre-processed, and implicit information is extracted from the pre-processed multi-source heterogeneous data. Specifically,

[0011] Collect geological exploration data, construction process data, real-time monitoring data, environmental data, and historical accident data involved in the entire life cycle of underground projects as multi-source heterogeneous data, perform missing value filling, outlier processing, and normalization on the structured data in the multi-source heterogeneous data, and perform text mining on the unstructured data in the multi-source heterogeneous data;

[0012] Performing spatiotemporal registration on the preprocessed multi-source heterogeneous data, clustering the spatiotemporal registered multi-source heterogeneous data according to data characteristics, obtaining a low-dimensional embedding representation through graph Laplacian matrix eigendecomposition, constructing a multimodal similarity matrix, and using the multimodal similarity matrix to determine and generate clusters to obtain the final clustering results;

[0013] Perform graph processing on spatiotemporal correlation data to construct a spatiotemporal graph, where nodes are monitoring points and edges contain spatial adjacency and temporal correlation. A graph convolution layer is used to extract spatial features, coupled with an LSTM layer to capture temporal dynamics and output corresponding data features.

[0014] For physically coupled data, a multi-task learning network is used to obtain corresponding data features through different task branches. For environmental response data, an attention mechanism is used in combination with a Transformer network for feature extraction to obtain corresponding data features.

[0015] The data features of different clusters are integrated to generate heterogeneous features, the heterogeneous features are subjected to feature standardization, and the mutual information between the data features of different clusters is calculated. Based on the mutual information, high-order interaction features are constructed, and implicit information is generated according to the heterogeneous features and the high-order interaction features.

[0016] In this solution, the temporal and spatial characteristics of implicit information are analyzed, specifically:

[0017] Using an improved CEEMDAN to perform time series decomposition on the implicit information, adding a specific proportion of white noise through multiple iterations to construct a preprocessing sequence, performing EMD decomposition on the preprocessing sequence to obtain intrinsic mode components, and screening intrinsic mode components greater than a preset correlation coefficient threshold;

[0018] The eigenmodal components obtained by screening are divided into long-term trend items, periodic fluctuation items, and sudden anomaly items. The Hurst exponent analysis is performed using the sliding window improved rescaled range method. The Hurst exponents of different components are associated with construction log events to generate the temporal characteristics of implicit information.

[0019] Calculate the range and sill value by rock layer based on the geological data of the underground project site, model physical constraints through Kriging difference, generate a continuous spatial distribution corresponding to the implicit information based on the physical constraints and the CNN model, calculate the spatial gradient of the continuous spatial distribution, and generate the spatial characteristics of the implicit information;

[0020] The temporal and spatial features of the implicit information are tensor-unified, and weights are calculated for the tensor-unified temporal and spatial features based on the attention mechanism. The weights are used for feature fusion to obtain the temporal and spatial features of the implicit information.

[0021] In this solution, underground engineering geological safety risk factors are obtained and modeled as a risk network. The comprehensive importance of nodes in the risk network is calculated based on the Stacking integration algorithm to identify key risk factors. Specifically,

[0022] Use big data retrieval methods to obtain multidimensional risk factors for underground engineering geological safety accidents, and construct an interaction matrix between multidimensional risk factors. Convert each risk factor and accident case into a network node, and use directed edges to represent causal directions to construct a network topology to generate a risk network.

[0023] Based on the risk network, the degree centrality, betweenness centrality, clustering coefficient, and closeness centrality indicators of the nodes are calculated. A stacking ensemble strategy is introduced to model node importance. Random forest, gradient boosting tree, and graph attention network are used as base learners. Random forest handles the nonlinear interaction between degree centrality and betweenness centrality, gradient boosting tree captures the monotonic relationship between clustering coefficient and closeness centrality, and graph attention network learns the implicit features of network topology.

[0024] Generate a three-dimensional feature vector for each node based on the features output by the base learner, add a construction stage label and a geological module code to the three-dimensional feature vector, use the three-dimensional feature vector to train a meta-learner, add an attention weighting mechanism and engineering constraints to the meta-learner, and output a comprehensive importance;

[0025] A comprehensive importance threshold is preset, nodes with an importance greater than the preset comprehensive importance threshold are marked, and key risk factors are identified based on the marked nodes.

[0026] In this solution, the temporal and spatial characteristics of implicit information are used to obtain the characteristic parameters of key risk factors, and a random walk mechanism is introduced to obtain a dynamic accident prediction chain. Specifically:

[0027] Based on the spatiotemporal characteristics of implicit information corresponding to multi-source heterogeneous data related to underground engineering, a characteristic parameter system for key risk factors is constructed. Initial values are assigned to key risk factor nodes, and a MetaPath random walk template is preset. The spatiotemporal feature similarity of nodes is obtained based on the initial values of key risk factor nodes, and the basic transition probability matrix is obtained by combining the network edge weights.

[0028] Obtain historical accident cases at different construction stages, analyze the occurrence frequencies of different accident categories based on the historical accident cases, obtain the accident preference coefficient of the current construction node as a correction term based on the occurrence frequencies of different accident categories, and correct the basic transition probability matrix to generate a dynamic transition probability matrix;

[0029] Select key risk factor nodes whose initial values are greater than the preset threshold, initiate walks from the selected key risk factor nodes simultaneously, use the dynamic transition probability matrix to sample the walk paths, dynamically adjust the walk granularity based on the comprehensive importance of the key risk factor nodes, and merge similar node sequences to build a prediction chain;

[0030] The prediction chain is compared with the historical accident chain in a dynamic time warping mode, and the prediction chain is enhanced using the historical accident chain that meets the dynamic time warping distance requirements. Chain weights are constructed based on the comprehensive importance of key risk factor nodes, edge weights, and time attenuation factors. The chain weights are used to perform priority sorting and select a preset number of dynamic accident prediction chains.

[0031] In this solution, a graph attention network is used to learn and represent the dynamic accident prediction chain and obtain the probability distribution of the accident evolution path. Specifically:

[0032] Obtain a preset number of dynamic accident prediction chains, use a graph attention network to learn and represent the dynamic accident prediction chains, set up an engineering semantics-guided multi-head attention mechanism in the graph attention layer, and match attention heads with different engineering semantics;

[0033] The multi-head attention mechanism is used to calculate dynamic attention, and a deformable convolution kernel is used to adapt to the changes in spatial relationships in different construction stages. LSTM modules are inserted between graph attention layers to memorize the state evolution history of nodes.

[0034] Through three layers of graph attention, neighborhood information is gradually aggregated to obtain the context-aware embedding representation of each node and the edge embedding representation. Special position encoding is applied to the first and last nodes of the dynamic accident prediction chain, and the context-aware embedding representation and edge embedding representation of each node are used to calculate the path scoring function.

[0035] The relative probability is calculated through the softmax method adapted to the construction stage, and the probability distribution of different accident evolution paths is generated, and the top 5 accident evolution paths with the highest probability are displayed in real time.

[0036] In this solution, the vulnerability of the connecting edges in the risk network is evaluated. Based on the comprehensive importance of nodes and the vulnerability of edges, a dynamic risk assessment model is established. The dynamic risk value corresponding to the accident is obtained according to the probability of accident occurrence and the risk mitigation factor. Specifically,

[0037] In the risk network, the vulnerability of the connection edge is obtained based on the dynamic edge betweenness, dynamic load rate and failure impact, and a dynamic risk assessment model is established. The risk network status is quantified in the dynamic risk assessment model, and the node status index and edge activity are calculated based on the node comprehensive importance and edge vulnerability.

[0038] Calculate the path activation probability for each dynamic accident prediction chain based on the node status index and edge activity, integrate the global probability based on the path activation probability, introduce structural measure factors and management measure factors as risk mitigation factors, and calculate the dynamic risk value based on the global probability and risk mitigation factors;

[0039] Risk levels are divided according to the dynamic risk values, and warning information is generated based on the divided risk levels.

[0040] The second aspect of the present invention provides a dynamic risk assessment system for underground engineering geological safety based on multi-source data fusion, which includes a multi-source data acquisition and preprocessing module, an implicit information mining module, a risk network modeling and analysis module, an accident chain prediction module, a dynamic risk assessment module, and a visual early warning module;

[0041] The multi-source data acquisition and preprocessing module is responsible for collecting multi-source heterogeneous data related to underground engineering, preprocessing the multi-source heterogeneous data, and realizing standardized access and fusion processing of the multi-source heterogeneous data;

[0042] The implicit information mining module is responsible for extracting implicit information from pre-processed multi-source heterogeneous data and analyzing the temporal and spatial characteristics of the implicit information;

[0043] The risk network modeling and analysis module is responsible for obtaining underground engineering geological safety risk factors, modeling the underground engineering geological safety risk factors into a risk network, calculating the comprehensive importance of nodes in the risk network based on the Stacking integration algorithm, identifying key risk factors, and evaluating the vulnerability of connecting edges in the risk network;

[0044] The accident chain prediction module is responsible for generating potential accident evolution paths based on the random walk mechanism, obtaining dynamic accident prediction chains, using graph attention networks to learn and represent dynamic accident prediction chains, and obtaining the probability distribution of accident evolution paths;

[0045] The dynamic risk assessment module is responsible for establishing a dynamic risk assessment model based on node importance and edge vulnerability, and obtaining the dynamic risk value corresponding to the accident according to the accident probability and risk mitigation factor;

[0046] The visual warning module visualizes underground engineering accidents and dynamic risk values, and displays high-risk areas and risk time series changes through a multi-dimensional warning dashboard.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The present invention is based on a dynamic risk assessment method based on multi-source data fusion. By integrating multi-dimensional data such as geological exploration, construction monitoring, and environmental factors, and combining complex network theory with machine learning algorithms, it realizes the closed-loop management of the entire chain of underground engineering risks, including data perception-network modeling-intelligent prediction-dynamic assessment-decision support, significantly improving the accuracy of risk identification and the timeliness of early warning. Based on complex network theory, a risk factor interaction network is constructed, and the importance of nodes is dynamically calculated through the Stacking integration algorithm to reflect the risk evolution path in real time. A random walk mechanism is introduced to generate an accident prediction chain, and the risk propagation model is learned by combining the graph attention network. The real-time monitoring data is assimilated through the Bayesian network to achieve minute-level updates of risk probabilities, thereby enhancing the accuracy and interpretability of risk early warnings. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or exemplary descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to these drawings without paying any creative work.

[0050] Figure 1 A flow chart of a method for dynamic risk assessment of underground engineering geological safety based on multi-source data fusion is shown;

[0051] Figure 2A flow chart showing the calculation of the comprehensive importance of risk network nodes to identify key risk factors;

[0052] Figure 3 A flow chart of introducing the random walk mechanism to obtain a dynamic accident prediction chain is shown;

[0053] Figure 4 The block diagram of the underground engineering geological safety dynamic risk assessment system based on multi-source data fusion is shown. DETAILED DESCRIPTION

[0054] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0055] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0056] Figure 1 A flow chart of a dynamic risk assessment method for underground engineering geological safety based on multi-source data fusion is shown.

[0057] like Figure 1 As shown, this embodiment provides a method for dynamic risk assessment of underground engineering geological safety based on multi-source data fusion, including:

[0058] S102, collecting multi-source heterogeneous data related to underground engineering, preprocessing the multi-source heterogeneous data, extracting implicit information from the preprocessed multi-source heterogeneous data, and analyzing the spatiotemporal characteristics of the implicit information;

[0059] S104, obtaining underground engineering geological safety risk factors, modeling the underground engineering geological safety risk factors into a risk network, calculating the comprehensive importance of nodes in the risk network based on a Stacking integration algorithm, and identifying key risk factors;

[0060] S106, using the spatiotemporal characteristics of implicit information to obtain characteristic parameters of key risk factors, introducing a random walk mechanism to obtain a dynamic accident prediction chain, using a graph attention network to learn and represent the dynamic accident prediction chain, and obtaining the probability distribution of the accident evolution path;

[0061] S108, evaluate the vulnerability of the connecting edges in the risk network, establish a dynamic risk assessment model based on the comprehensive importance of nodes and edge vulnerability, obtain the dynamic risk value corresponding to the accident according to the probability of accident occurrence and risk mitigation factor, and visualize underground engineering accidents and dynamic risk values.

[0062] It should be noted that the geological exploration data, construction process data, real-time monitoring data, environmental data and historical accident data involved in the entire life cycle of underground projects are collected as multi-source heterogeneous data; the address exploration data include drilling data, geological profiles, rock mechanics parameters, hydrogeological data, etc.; the construction process data include excavation progress, support parameters, construction logs, quality inspection records, etc.; the real-time monitoring data include displacement monitoring, stress monitoring, seepage pressure monitoring, microseismic monitoring and other time series data, and the environmental data include rainfall, temperature, vibration monitoring and other external factor data; the historical accident data include accident case libraries and disposal plans for similar projects. The structured data (monitoring data, construction parameters) in the multi-source heterogeneous data are filled with missing values, processed for outliers, and normalized, and the unstructured data (geological reports, construction logs) in the multi-source heterogeneous data are subjected to text mining. The pre-processed multi-source heterogeneous data are subjected to spatiotemporal registration to establish a mapping relationship between the data. For example, the geological exploration data (drilling coordinates), monitoring point layout locations, and construction area models are converted into coordinate systems and unified into the engineering coordinate system, and the time series data (monitoring sensors, construction records) are timestamp-standardized.

[0063] The multi-source heterogeneous data after spatiotemporal registration are clustered according to the data characteristics, and clustering features are designed according to different data types. For structured data, temporal features (mean, variance, trend slope) and spatial features (correlation of neighboring points) are used as clustering features; for unstructured data, Doc2Vec document vectors and geological entity recognition results are used as clustering features; for image data, directional gradient histograms combined with texture features are extracted as clustering features. A low-dimensional embedding representation is obtained through eigendecomposition of the graph Laplacian matrix, and a multimodal similarity matrix is constructed. Euclidean distance is used for structured data, cosine similarity is used for text, and SSIM structural similarity is used for images. The multimodal similarity matrix is used to determine and generate clusters to obtain the final clustering results. Graph processing is performed on spatiotemporal correlation data to construct a spatiotemporal graph. In the spatiotemporal graph, nodes are monitoring points, and edges contain spatial adjacency relationships and temporal correlations. Spatial features such as deformation propagation patterns are extracted through graph convolution layers, and temporal dynamics such as displacement acceleration trends are captured through coupled LSTM layers to output corresponding data features. For physically coupled data, a multi-task learning network is used to obtain corresponding data features through different task branches. For example, a 1D CNN is used to extract local fluctuation features of stress data. Wavelet packet decomposition and energy spectrum analysis of vibration signals are performed in the first task branch. In the second task branch, regression modeling of construction parameters and stress responses is performed to obtain corresponding dynamic load influence coefficients and construction disturbance sensitivity. An attention mechanism combined with a Transformer network is used to extract features from environmentally responsive data, acquiring corresponding data features. For example, a discrete wavelet transform is applied to seepage pressure data to decompose multi-scale features. Attention weights are used to calculate the differential impact of rainfall events on each monitoring point. Hydrogeological parameters are then fused through cross-modal attention to output hydraulic coupling strength, among other parameters. Data features from different clusters are integrated to generate heterogeneous features, which are then normalized. Mutual information between features from different clusters is calculated, such as the correlation between vibration sensitivity and rock mass deformation synergy. Based on this mutual information, higher-order interactive features are constructed, such as the rainfall-seepage pressure-support stress trimodal coupling index. Implicit information is generated from heterogeneous and high-order interactive features. This efficient conversion of multi-source heterogeneous data from raw information to implicit risk features provides an accurate data foundation for dynamic risk assessment.

[0064] It should be noted that the implicit information is decomposed into time series using an improved CEEMDAN. A preprocessing sequence is constructed by adding a specific proportion of white noise through multiple iterations to effectively separate modal aliasing components. Based on the characteristics of underground engineering data, the noise amplitude is set at 5-10% of the standard deviation of the monitoring data. The preprocessing sequence is subjected to EMD decomposition to obtain intrinsic modal components, and eigenmodal components with a value greater than a preset correlation coefficient threshold are screened. The screened intrinsic modal components are divided into long-term trend terms, cyclical fluctuation terms, and sudden anomaly terms. The long-term trend term is a low-frequency component with a cumulative variance contribution greater than 60%, reflecting the slow evolution of the engineering structure. The cyclical fluctuation term is a mid-frequency component with a significant period, corresponding to the impact of the construction cycle. The sudden anomaly term is a high-frequency residual component that represents abnormal events such as equipment failure or geological mutations. The Hurst exponent analysis is performed using the sliding window improved rescaled range method to determine the persistence of the characteristics. The Hurst exponents of different components are associated with construction log events to generate the temporal characteristics of the implicit information. The range and sill value are calculated for each rock layer based on the geological data of the underground project location. Physical constraints are modeled using Kriging differences to ensure that the prediction conforms to geological laws. The continuous spatial distribution corresponding to the implicit information is generated based on the physical constraints and combined with the CNN model. The spatial gradient of the continuous spatial distribution is calculated. The high-risk transition zone is identified through the spatial gradient field to generate the spatial characteristics of the implicit information. The temporal and spatial characteristics of the implicit information are tensor-unified. The weights of the tensor-unified temporal and spatial characteristics are calculated based on the attention mechanism. The weights are used for feature fusion to obtain the spatiotemporal characteristics of the implicit information.

[0065] Figure 2 A flowchart showing the calculation of the comprehensive importance of risk network nodes and the identification of key risk factors is shown.

[0066] According to an embodiment of the present invention, underground engineering geological safety risk factors are obtained, the underground engineering geological safety risk factors are modeled as a risk network, the comprehensive importance of the nodes in the risk network is calculated based on the Stacking integration algorithm, and the key risk factors are identified, specifically:

[0067] S202, using a big data retrieval method to obtain multi-dimensional risk factors for underground engineering geological safety accidents, and constructing an interaction matrix between the multi-dimensional risk factors, converting each risk factor and accident case into a network node, and using directed edges to represent causal directions, constructing a network topology to generate a risk network;

[0068] S204, based on the risk network, the degree centrality, betweenness centrality, clustering coefficient, and closeness centrality indicators of the nodes are calculated, and a stacking ensemble strategy is introduced to model the node importance. Random forest, gradient boosting tree, and graph attention network are used as base learners, wherein the random forest handles the nonlinear interaction between degree centrality and betweenness centrality, the gradient boosting tree captures the monotonic relationship between clustering coefficient and closeness centrality, and the graph attention network learns the implicit features of the network topology structure;

[0069] S206, generating a three-dimensional feature vector for each node based on the features output by the base learner, adding a construction stage label and a geological module code to the three-dimensional feature vector, using the three-dimensional feature vector to train a meta-learner, adding an attention weighting mechanism and engineering constraints to the meta-learner, and outputting a comprehensive importance;

[0070] S208 , preset a comprehensive importance threshold, mark nodes with a value greater than the preset comprehensive importance threshold, and identify key risk factors based on the marked nodes.

[0071] It should be noted that the multi-dimensional risk factors for underground engineering geological safety accidents are obtained using big data retrieval methods, including geological ontological factors, engineering response factors, and environmental induction factors. The geological ontological factors include rock mass quality indicators, geological structural characteristics, and hydrogeological parameters. The engineering response factors include surrounding rock deformation, support status, construction disturbances, and environmental induction factors. The meteorological and hydrological factors and surrounding loads are included. Based on the principles of geotechnical mechanics, deterministic associations are established as the physical interaction relationships between factors. Potential associations are mined through historical data as statistical correlations between factors. Domain knowledge is integrated to supplement implicit associations as empirical correlations between factors. Each risk factor is converted into a network node. The node attributes include basic attributes such as factor type and monitoring location, as well as dynamic attributes such as current value and change trend. Directed edges between nodes are constructed through different causal relationships, and different edge weights are determined.

[0072] In calculating degree centrality, the weights of all directed edges emanating from the current node are summed to quantify the node's external influence. For example, a high out-degree of an excavation speed node reflects its significant impact on multiple support monitoring points. The weights of all directed edges emanating from the node are summed to characterize the node's susceptibility to influence from other factors. For example, nodes with weak interlayers typically have high in-degree, and the degree values are mapped to the [0, 1] interval. In calculating betweenness centrality, only effective propagation paths of length ≤ 3 are considered. Based on the maximum flow minimum cut theory, the bottleneck role of nodes in risk propagation is analyzed. For example, groundwater pressure nodes often exhibit high betweenness in seepage paths. In calculating the clustering coefficient, only the triangular closure of adjacent nodes within the same geological unit is calculated. For example, clusters of monitoring points within a fault fracture zone typically exhibit high clustering, distinguishing between risk input clustering and risk output clustering. In calculating closeness centrality, 1 / edge weight is used as the distance metric. Local closeness is assessed by construction zone, tracking the closeness evolution of key nodes during construction.

[0073] During the Stacking ensemble modeling phase, random forests, gradient boosting trees, and graph attention networks were used as base learners to generate three-dimensional feature vectors for each node. Engineering features such as construction stage labels and geological unit codes were added for spatiotemporal context enhancement, and the contribution weights of each base model were automatically adjusted based on the current network density. The base learners were independently trained using 5-fold time series cross-validation. Meta-learners were trained using the meta-features output by the base learners. An attention weighting mechanism was designed into the meta-learner to ensure that key construction factors maintained high weights, and the meta-learner was used to output a comprehensive importance score. Through Stacking ensemble analysis, risk factor importance assessment has transitioned from a single perspective to multi-dimensional integration. The established comprehensive importance metric combines statistical significance with engineering interpretability, providing a quantitative decision-making basis for precise risk prevention and control of underground engineering projects.

[0074] Figure 3 A flowchart of introducing the random walk mechanism to obtain a dynamic accident prediction chain is shown.

[0075] According to an embodiment of the present invention, the temporal and spatial characteristics of implicit information are used to obtain characteristic parameters of key risk factors, and a random walk mechanism is introduced to obtain a dynamic accident prediction chain, specifically:

[0076] S302: Build a characteristic parameter system for key risk factors based on the spatiotemporal characteristics of implicit information corresponding to multi-source heterogeneous data related to underground engineering, assign initial values to key risk factor nodes, preset a MetaPath random walk template, obtain the spatiotemporal feature similarity of nodes based on the initial values of key risk factor nodes, and obtain a basic transition probability matrix based on the network edge weights.

[0077] S304: Obtain historical accident cases at different construction stages, analyze the occurrence frequencies of different accident categories based on the historical accident cases, obtain the accident preference coefficient of the current construction node based on the occurrence frequencies of different accident categories as a correction term, and modify the basic transition probability matrix to generate a dynamic transition probability matrix;

[0078] S306: Select key risk factor nodes whose initial values are greater than a preset threshold, initiate walks simultaneously from the selected key risk factor nodes, use a dynamic transition probability matrix to sample walk paths, dynamically adjust the walk granularity based on the comprehensive importance of the key risk factor nodes, and merge similar node sequences to build a prediction chain;

[0079] S308, compare the prediction chain with the historical accident chain in a dynamic time warping mode, use the historical accident chain that meets the dynamic time warping distance requirements to enhance the prediction chain, construct chain weights based on the comprehensive importance of key risk factor nodes, edge weights and time attenuation factors, use the chain weights to prioritize and select a preset number of dynamic accident prediction chains.

[0080] It should be noted that the template of MetaPath random walk is preset through engineering semantics, such as excavation parameters → vibration → loosening zone → settlement, rock layer → seepage pressure → displacement → support stress, rainfall → groundwater level → weak interlayer → slip. The network edge weight is combined with the similarity of the spatiotemporal characteristics of the nodes to generate the basic probability, and the occurrence frequency of different accident categories is analyzed according to the historical accident cases. The accident preference coefficient of the current construction node is obtained as a correction item through the occurrence frequency of different accident categories, such as strengthening the geological structure related path during the support period. Walks are initiated simultaneously from the selected key risk factor nodes, and the dynamic transfer probability matrix is used to sample the walk path. Along the main risk transmission path, the selection probability of nodes with high comprehensive importance is increased, and the search is refined within the local high-risk cluster. The walk granularity is dynamically adjusted according to the comprehensive importance of the key risk factor nodes, and similar node sequences are merged to construct a prediction chain. The Granger causality test is used to verify the temporal relationship for causal verification, and finally the prediction chain is obtained. The chain weight is constructed according to the comprehensive importance of the key risk factor nodes, the edge weight and the time attenuation factor. , and use engineering semantics to annotate chain types, such as seepage erosion type, vibration accumulation type, etc.

[0081] It should be noted that a preset number of dynamic accident prediction chains are obtained, and the dynamic accident prediction chains are learned and represented using a graph attention network. A multi-head attention mechanism guided by engineering semantics is set in the graph attention layer to match the attention heads with different engineering semantics, where the geological head focuses on features such as rock continuity and structural orientation, the mechanical head focuses on stress-strain transmission relationships, and the construction head tracks the impact of process technology. The multi-head attention mechanism is used to calculate dynamic attention, and a deformable convolution kernel is used to adapt to changes in spatial relationships in different construction stages. The convolution radius is dynamically adjusted with the excavation progress. LSTM modules are inserted between graph attention layers to memorize the state evolution history of the nodes. Model training is performed using extreme working conditions such as a single accident type, a mixture of multiple types, and noise. Neighborhood information is gradually aggregated through three layers of graph attention to eventually obtain the context-aware embedding representation and edge embedding representation of each node, and special position encoding is applied to the head and tail nodes of the dynamic accident prediction chain. The path scoring function is calculated using the context-aware embedding representation and edge embedding representation of each node. , is the learnable parameter matrix, is the type bias term. The relative probability is calculated by the softmax adapted in the construction phase. ,in Using a temperature coefficient, 20% of attention heads are randomly discarded during forward propagation, and 50 samples are taken to obtain the probability distribution variance. Monte Carlo dropout is used to randomly discard 20% of attention heads during forward propagation, and 50 samples are taken to obtain the probability distribution variance. This generates the probability distribution of different accident evolution paths, and displays the top five accident evolution paths in real time. A closed data loop of prediction, action, and verification is established, with the weight of correctly predicted path samples increased by 30%. False positive paths trigger model fine-tuning.

[0082] It should be noted that the vulnerability of connected edges in the risk network is determined based on dynamic edge betweenness, dynamic load ratio, and failure impact. The edge betweenness metric uses an improved flow-sensitive calculation method, counting the proportion of all valid risk transmission paths passing through the edge. The dynamic load ratio calculates the ratio of the risk currently transmitted by the edge (based on node state differences) to its theoretical load-bearing capacity. For example, the load ratio of the edge between the support structure and the surrounding rock = measured axial force / designed allowable axial force × time-varying reduction factor. The failure impact is determined by randomly disconnecting edges through Monte Carlo simulations to measure the decrease in global network efficiency, specifically considering tectonic cascading effects, such as the wider impact of edge failure near fault zones. Peak edge vulnerability periods during construction are identified, and vulnerability benchmark zones are defined based on geological survey results. Similarity correction coefficients are applied to edges within the same rock layer unit, and edges that do not conform to mechanical principles (e.g., long-distance absence of a dielectric) are forcibly set to zero vulnerability.

[0083] A dynamic risk assessment model is established, in which the risk network status is quantified, and the node status index and edge activity are calculated based on the comprehensive importance of nodes and edge vulnerability; the node status index Quantify the node risk level, expressed as , is the material nonlinear coefficient, edge activity Reflects the risk transmission intensity of the connection edge, expressed as , To control the construction quality, the path activation probability is calculated for each dynamic accident prediction chain according to the node status index and edge activity. , characterizes the possibility of the occurrence of the dynamic accident prediction chain, expressed as , is the edge activity of the i-th edge, represents the node status index of each j-th node, is the path length penalty factor, and the global probability is integrated according to the path activation probability, the global risk probability Considering the integrated risk of all critical paths, it is expressed as , is the kth path probability. Structural measures factor and management measures factor are introduced as risk mitigation factors. The structural measures factor is expressed as: , the management measure factor is expressed as: , preferably, the emergency plan score is obtained by collecting historical expert scores. The dynamic risk value is calculated based on the global probability and risk mitigation factor , , The project importance coefficient is used. Risk levels are divided according to the dynamic risk value. A dynamic risk value-loss curve is established based on historical accident data. Thresholds are set according to project risk preferences. The overall threshold is lowered during key construction stages (such as crossing faults), and early warning information is generated based on the divided risk levels.

[0084] Figure 4 The block diagram of the underground engineering geological safety dynamic risk assessment system based on multi-source data fusion is shown.

[0085] The second embodiment of the present invention provides a dynamic risk assessment system 4 for underground engineering geological safety based on multi-source data fusion, comprising: a multi-source data acquisition and preprocessing module 401, an implicit information mining module 402, a risk network modeling and analysis module 403, an accident chain prediction module 404, a dynamic risk assessment module 405, and a visual early warning module 406;

[0086] The multi-source data acquisition and pre-processing module 401 is responsible for acquiring multi-source heterogeneous data related to underground engineering, pre-processing the multi-source heterogeneous data, and realizing standardized access and fusion processing of the multi-source heterogeneous data;

[0087] The implicit information mining module 402 is responsible for extracting implicit information from the pre-processed multi-source heterogeneous data and analyzing the temporal and spatial characteristics of the implicit information;

[0088] The risk network modeling and analysis module 403 is responsible for obtaining underground engineering geological safety risk factors, modeling the underground engineering geological safety risk factors into a risk network, calculating the comprehensive importance of nodes in the risk network based on the Stacking integration algorithm, identifying key risk factors, and evaluating the vulnerability of connecting edges in the risk network;

[0089] The accident chain prediction module 404 is responsible for generating potential accident evolution paths based on the random walk mechanism, obtaining dynamic accident prediction chains, using the graph attention network to learn and represent the dynamic accident prediction chains, and obtaining the probability distribution of the accident evolution paths;

[0090] The dynamic risk assessment module 405 is responsible for establishing a dynamic risk assessment model based on node importance and edge vulnerability, and obtaining a dynamic risk value corresponding to an accident according to the accident probability and risk mitigation factor;

[0091] The visual warning module 406 visualizes underground engineering accidents and dynamic risk values, and displays high-risk areas and risk time series changes through a multi-dimensional warning dashboard.

[0092] The third aspect of the present invention provides a computer-readable storage medium, which includes a method program for dynamic risk assessment of underground engineering geological safety based on multi-source data fusion. When the method program for dynamic risk assessment of underground engineering geological safety based on multi-source data fusion is executed by a processor, the steps of the method for dynamic risk assessment of underground engineering geological safety based on multi-source data fusion are implemented.

[0093] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of the modules is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple modules or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms. In addition, the functional modules in the various embodiments of the present invention can all be integrated into one processing module, or each module can be a separate module, or two or more modules can be integrated into one module; the above-mentioned integrated modules can be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0094] Those skilled in the art will understand that all or part of the steps of the above-mentioned method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disks or optical disks, and other media that can store program codes.

[0095] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A dynamic risk assessment method for underground engineering geological safety based on multi-source data fusion, characterized in that: The following steps are involved: Collecting multi-source heterogeneous data related to underground engineering, preprocessing the multi-source heterogeneous data, extracting implicit information from the preprocessed multi-source heterogeneous data, and analyzing the spatiotemporal characteristics of the implicit information; Obtain underground engineering geological safety risk factors, model the underground engineering geological safety risk factors into a risk network, calculate the comprehensive importance of nodes in the risk network based on the Stacking integration algorithm, and identify key risk factors; The spatial and temporal characteristics of implicit information are used to obtain the characteristic parameters of key risk factors. A random walk mechanism is introduced to obtain a dynamic accident prediction chain. The graph attention network is used to learn and represent the dynamic accident prediction chain to obtain the probability distribution of the accident evolution path. Assess the vulnerability of the connecting edges in the risk network, establish a dynamic risk assessment model based on the comprehensive importance of nodes and edge vulnerability, obtain the dynamic risk value corresponding to the accident according to the probability of accident occurrence and risk mitigation factor, and visualize underground engineering accidents and dynamic risk values.

2. The method for dynamic risk assessment of underground engineering geological safety based on multi-source data fusion according to claim 1 is characterized in that: Collect multi-source heterogeneous data related to underground engineering, pre-process the multi-source heterogeneous data, and extract implicit information from the pre-processed multi-source heterogeneous data, specifically: Collect geological exploration data, construction process data, real-time monitoring data, environmental data, and historical accident data involved in the entire life cycle of underground projects as multi-source heterogeneous data, perform missing value filling, outlier processing, and normalization on the structured data in the multi-source heterogeneous data, and perform text mining on the unstructured data in the multi-source heterogeneous data; Performing spatiotemporal registration on the preprocessed multi-source heterogeneous data, clustering the spatiotemporal registered multi-source heterogeneous data according to data characteristics, obtaining a low-dimensional embedding representation through graph Laplacian matrix eigendecomposition, constructing a multimodal similarity matrix, and using the multimodal similarity matrix to determine and generate clusters to obtain the final clustering results; Perform graph processing on spatiotemporal correlation data to construct a spatiotemporal graph, where nodes are monitoring points and edges contain spatial adjacency and temporal correlation. A graph convolution layer is used to extract spatial features, coupled with an LSTM layer to capture temporal dynamics and output corresponding data features. For physically coupled data, a multi-task learning network is used to obtain corresponding data features through different task branches. For environmental response data, an attention mechanism is used in combination with a Transformer network for feature extraction to obtain corresponding data features. The data features of different clusters are integrated to generate heterogeneous features, the heterogeneous features are subjected to feature standardization, and the mutual information between the data features of different clusters is calculated. Based on the mutual information, high-order interaction features are constructed, and implicit information is generated according to the heterogeneous features and the high-order interaction features.

3. The method for dynamic risk assessment of underground engineering geological safety based on multi-source data fusion according to claim 2 is characterized in that: Analyze the spatiotemporal characteristics of implicit information, specifically: Using an improved CEEMDAN to perform time series decomposition on the implicit information, adding a specific proportion of white noise through multiple iterations to construct a preprocessing sequence, performing EMD decomposition on the preprocessing sequence to obtain intrinsic mode components, and screening intrinsic mode components greater than a preset correlation coefficient threshold; The eigenmodal components obtained by screening are divided into long-term trend items, periodic fluctuation items, and sudden anomaly items. The Hurst exponent analysis is performed using the sliding window improved rescaled range method. The Hurst exponents of different components are associated with construction log events to generate the temporal characteristics of implicit information. Calculate the range and sill value by rock layer based on the geological data of the underground project site, model physical constraints through Kriging difference, generate a continuous spatial distribution corresponding to the implicit information based on the physical constraints and the CNN model, calculate the spatial gradient of the continuous spatial distribution, and generate the spatial characteristics of the implicit information; The temporal and spatial features of the implicit information are tensor-unified, and weights are calculated for the tensor-unified temporal and spatial features based on the attention mechanism. The weights are used for feature fusion to obtain the temporal and spatial features of the implicit information.

4. The method for dynamic risk assessment of underground engineering geological safety based on multi-source data fusion according to claim 1 is characterized in that: Obtain underground engineering geological safety risk factors, model the underground engineering geological safety risk factors into a risk network, calculate the comprehensive importance of nodes in the risk network based on the Stacking integration algorithm, and identify key risk factors, specifically: Use big data retrieval methods to obtain multidimensional risk factors for underground engineering geological safety accidents, and construct an interaction matrix between multidimensional risk factors. Convert each risk factor and accident case into a network node, and use directed edges to represent causal directions to construct a network topology to generate a risk network. Based on the risk network, the degree centrality, betweenness centrality, clustering coefficient, and closeness centrality indicators of the nodes are calculated. A stacking ensemble strategy is introduced to model node importance. Random forest, gradient boosting tree, and graph attention network are used as base learners. Random forest handles the nonlinear interaction between degree centrality and betweenness centrality, gradient boosting tree captures the monotonic relationship between clustering coefficient and closeness centrality, and graph attention network learns the implicit features of network topology. Generate a three-dimensional feature vector for each node based on the features output by the base learner, add a construction stage label and a geological module code to the three-dimensional feature vector, use the three-dimensional feature vector to train a meta-learner, add an attention weighting mechanism and engineering constraints to the meta-learner, and output a comprehensive importance; A comprehensive importance threshold is preset, nodes with an importance greater than the preset comprehensive importance threshold are marked, and key risk factors are identified based on the marked nodes.

5. The method for dynamic risk assessment of underground engineering geological safety based on multi-source data fusion according to claim 1 is characterized in that: The temporal and spatial characteristics of implicit information are used to obtain the characteristic parameters of key risk factors, and a random walk mechanism is introduced to obtain a dynamic accident prediction chain, specifically: Based on the spatiotemporal characteristics of implicit information corresponding to multi-source heterogeneous data related to underground engineering, a characteristic parameter system for key risk factors is constructed. Initial values are assigned to key risk factor nodes, and a MetaPath random walk template is preset. The spatiotemporal feature similarity of nodes is obtained based on the initial values of key risk factor nodes, and the basic transition probability matrix is obtained by combining the network edge weights. Obtain historical accident cases at different construction stages, analyze the occurrence frequencies of different accident categories based on the historical accident cases, obtain the accident preference coefficient of the current construction node as a correction term based on the occurrence frequencies of different accident categories, and correct the basic transition probability matrix to generate a dynamic transition probability matrix; Select key risk factor nodes whose initial values are greater than the preset threshold, initiate walks from the selected key risk factor nodes simultaneously, use the dynamic transition probability matrix to sample the walk paths, dynamically adjust the walk granularity based on the comprehensive importance of the key risk factor nodes, and merge similar node sequences to build a prediction chain; The prediction chain is compared with the historical accident chain in a dynamic time warping mode, and the prediction chain is enhanced using the historical accident chain that meets the dynamic time warping distance requirements. Chain weights are constructed based on the comprehensive importance of key risk factor nodes, edge weights, and time attenuation factors. The chain weights are used to perform priority sorting and select a preset number of dynamic accident prediction chains.

6. The method for dynamic risk assessment of underground engineering geological safety based on multi-source data fusion according to claim 1 is characterized in that: The graph attention network is used to learn and represent the dynamic accident prediction chain and obtain the probability distribution of the accident evolution path. Specifically: Obtain a preset number of dynamic accident prediction chains, use a graph attention network to learn and represent the dynamic accident prediction chains, set up an engineering semantics-guided multi-head attention mechanism in the graph attention layer, and match attention heads with different engineering semantics; The multi-head attention mechanism is used to calculate dynamic attention, and a deformable convolution kernel is used to adapt to the changes in spatial relationships in different construction stages. LSTM modules are inserted between graph attention layers to memorize the state evolution history of nodes. Through three layers of graph attention, neighborhood information is gradually aggregated to obtain the context-aware embedding representation of each node and the edge embedding representation. Special position encoding is applied to the first and last nodes of the dynamic accident prediction chain, and the context-aware embedding representation and edge embedding representation of each node are used to calculate the path scoring function. The relative probability is calculated through the softmax method adapted to the construction stage, and the probability distribution of different accident evolution paths is generated, and the top 5 accident evolution paths with the highest probability are displayed in real time.

7. The method for dynamic risk assessment of underground engineering geological safety based on multi-source data fusion according to claim 1 is characterized in that: Evaluate the vulnerability of the connecting edges in the risk network. Based on the comprehensive importance of nodes and the vulnerability of edges, establish a dynamic risk assessment model. According to the probability of accident occurrence and the risk mitigation factor, obtain the dynamic risk value corresponding to the accident. Specifically: In the risk network, the vulnerability of the connection edge is obtained based on the dynamic edge betweenness, dynamic load rate and failure impact, and a dynamic risk assessment model is established. The risk network status is quantified in the dynamic risk assessment model, and the node status index and edge activity are calculated based on the node comprehensive importance and edge vulnerability. Calculate the path activation probability for each dynamic accident prediction chain based on the node status index and edge activity, integrate the global probability based on the path activation probability, introduce structural measure factors and management measure factors as risk mitigation factors, and calculate the dynamic risk value based on the global probability and risk mitigation factors; Risk levels are divided according to the dynamic risk values, and warning information is generated based on the divided risk levels.

8. A dynamic risk assessment system for underground engineering geological safety based on multi-source data fusion, characterized by: Implementing the underground engineering geological safety dynamic risk assessment method based on multi-source data fusion as described in any one of claims 1 to 7, the system includes a multi-source data acquisition and preprocessing module, an implicit information mining module, a risk network modeling and analysis module, an accident chain prediction module, a dynamic risk assessment module and a visual early warning module; The multi-source data acquisition and preprocessing module is responsible for collecting multi-source heterogeneous data related to underground engineering, preprocessing the multi-source heterogeneous data, and realizing standardized access and fusion processing of the multi-source heterogeneous data; The implicit information mining module is responsible for extracting implicit information from pre-processed multi-source heterogeneous data and analyzing the temporal and spatial characteristics of the implicit information; The risk network modeling and analysis module is responsible for obtaining underground engineering geological safety risk factors, modeling the underground engineering geological safety risk factors into a risk network, calculating the comprehensive importance of nodes in the risk network based on the Stacking integration algorithm, identifying key risk factors, and evaluating the vulnerability of connecting edges in the risk network; The accident chain prediction module is responsible for generating potential accident evolution paths based on the random walk mechanism, obtaining dynamic accident prediction chains, using graph attention networks to learn and represent dynamic accident prediction chains, and obtaining the probability distribution of accident evolution paths; The dynamic risk assessment module is responsible for establishing a dynamic risk assessment model based on node importance and edge vulnerability, and obtaining the dynamic risk value corresponding to the accident according to the accident probability and risk mitigation factor; The visual warning module visualizes underground engineering accidents and dynamic risk values, and displays high-risk areas and risk time series changes through a multi-dimensional warning dashboard.

Images