Geological disaster intelligent monitoring and early warning method and system

Abstract

The present application relates to the technical field of geological disaster monitoring, and more particularly to a geological disaster intelligent monitoring and early warning method and system. The method obtains multi-source heterogeneous spatio-temporal data and standardizes them, adopts tensor decomposition and spatio-temporal attention collaborative modeling to form a spatio-temporal dependency graph, fuses the spatio-temporal dependency graph with a geological disaster domain knowledge graph, uses comparative learning and reinforcement learning to obtain a regional hidden danger risk assessment result, and updates the knowledge graph based on the assessment result and outputs the spatial distribution of hidden danger points and the risk level. The present application can realize the cross-regional migration of disaster-causing mechanisms and effectively improve the accuracy and regional adaptability of geological disaster early warning.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring technology, and in particular to intelligent monitoring and early warning methods and systems for geological disasters. Background Technology

[0002] In the field of intelligent monitoring and early warning of geological disasters, current conventional practices typically rely on single-type sensor data or limited historical records, preprocessing multi-source data through sampling methods with fixed spatiotemporal resolution. For example, for disasters such as landslides and debris flows, independent data sources such as GPS displacement monitoring, rain gauges, and remote sensing images are often used, each standardized according to a preset grid or time interval, and then fused using simple weighted averaging or threshold rules. The extraction of spatiotemporal features is mostly based on traditional statistical regression models or shallow machine learning methods, such as support vector machines and random forests, attempting to learn precursory patterns of disasters from historical events. Some studies have attempted to introduce convolutional neural networks or recurrent neural networks to model spatiotemporal sequences, but the model structures are relatively fixed and lack dynamic adaptability to the evolution of disasters at different scales. The application of knowledge graphs in the field of geological disasters is still in its early stages, usually based on static ontologies manually defined by expert experience, used for structured descriptions of disaster events and simple relational queries, with reasoning methods limited to rule-based logical deduction or graph matching.

[0003] The aforementioned conventional practices have significant drawbacks. On the one hand, the processing methods for multi-source heterogeneous spatiotemporal data are too rigid. Using a fixed spatiotemporal resolution makes it difficult to match the non-stationary characteristics of geological disaster data distribution—in areas with dense monitoring networks, low resolution will lose key details of changes; while in remote mountainous areas with sparse data, high resolution leads to a large amount of invalid interpolation, introducing false information. Simple weighting or rule-based fusion between data from different sources cannot effectively preserve the unique spatiotemporal distribution characteristics and complementary relationships of each data source. Instead, it amplifies the noise of individual sensors or masks local anomaly signals, making it difficult to guarantee the quality of the basic data on which subsequent modeling is based.

[0004] On the other hand, existing disaster early warning models suffer from severe generalization limitations. These models are typically trained on local historical data from specific regions. When applied to new areas with varying hydrogeological conditions, the limited coverage of disaster evolution patterns in the training samples makes it difficult for the models to extract cross-regional migratory mechanisms from known disasters. The inference process is often a single, static computation, lacking a mechanism to dynamically adjust the analysis path based on the confidence level of intermediate results. Once the statistical distribution of the input data shifts due to environmental changes, or when encountering novel precursor combinations not included in the training set, the model's predictive performance drops sharply. Furthermore, knowledge graphs and data-driven modeling are often separate technical approaches—knowledge graphs, after independent construction, serve only as semantic labels for the results, without participating in guiding the extraction and inference of spatiotemporal features. They also cannot automatically update the internal relational strength by learning new cases, leading to continuous aging of domain knowledge and an inability to dynamically reflect the evolutionary trends of the geological disaster system. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for intelligent monitoring and early warning of geological disasters, which can solve the problems in the prior art.

[0006] A first aspect of the present invention provides a method for intelligent monitoring and early warning of geological disasters, comprising:

[0007] Acquire multi-source heterogeneous spatiotemporal data of the target region, dynamically adjust the spatiotemporal resolution according to the density distribution of the multi-source heterogeneous spatiotemporal data, and map spatiotemporal data from different sources to a unified spatiotemporal reference framework to form a standardized spatiotemporal dataset.

[0008] Tensor decomposition and spatiotemporal attention are performed on the standardized spatiotemporal dataset. The spatiotemporal coupling relationship of the standardized spatiotemporal dataset is extracted by tensor decomposition and multi-scale disaster evolution patterns are captured. The latent factors obtained by tensor decomposition are used as constraints for attention calculation. The attention weights are fed back to guide the rank optimization of the tensor decomposition, forming a spatiotemporal dependency graph.

[0009] A knowledge graph in the field of geological hazards is constructed. The spatiotemporal dependency graph is fused with the knowledge graph in the field of geological hazards to obtain a fusion graph. Reasoning is performed on the fusion graph. Feature mapping between known disaster areas and unlabeled areas is established by using contrastive learning. The reasoning path is dynamically adjusted according to the reasoning confidence through reinforcement learning to realize the cross-regional migration of disaster-causing mechanisms and obtain regional hidden danger risk assessment results.

[0010] Based on the feedback of the regional hidden danger risk assessment results, the relationship strength in the geological disaster knowledge graph is updated, and the spatial distribution and risk level of hidden danger points are output.

[0011] Tensor decomposition and spatiotemporal attention co-modeling are performed on the standardized spatiotemporal dataset. Tensor decomposition is used to extract the spatiotemporal coupling relationships of the standardized spatiotemporal dataset and capture multi-scale disaster evolution patterns, including:

[0012] The standardized spatiotemporal dataset is constructed as a time-space-attribute third-order tensor, and tensor decomposition is performed on the third-order tensor to extract the time factor matrix, space factor matrix, and attribute factor matrix.

[0013] Spatiotemporal attention weights are calculated based on the time factor matrix and the spatial factor matrix, with the time factor matrix serving as the initial constraint for the time attention weights and the spatial factor matrix serving as the initial constraint for the spatial attention weights.

[0014] The spatiotemporal attention weights are fed back to the tensor decomposition process. The rank selection of the time factor matrix, the spatial factor matrix, and the attribute factor matrix is ​​dynamically adjusted according to the spatiotemporal attention weights. Through iterative optimization, the time factor matrix, the spatial factor matrix, the attribute factor matrix, and the spatiotemporal attention weights are made to achieve synergy and consistency.

[0015] Based on the time factor matrix, spatial factor matrix, and attribute factor matrix after coordination and consistency, the spatiotemporal coupling relationship of the standardized spatiotemporal dataset is reconstructed to capture multi-scale disaster evolution patterns.

[0016] Using the latent factors obtained from tensor decomposition as constraints for attention computation, and feeding back the attention weights to guide the rank optimization of the tensor decomposition, a spatiotemporal dependency graph is formed, including:

[0017] Extract latent factor vectors from the temporal and spatial factor matrices obtained from tensor decomposition, calculate the dominant directions of the latent factor vectors in the temporal and spatial dimensions, and use the dominant directions as search constraints for spatiotemporal attention weights.

[0018] Under the search constraints, the spatiotemporal attention weights of the standardized spatiotemporal dataset are calculated, and the sparse distribution pattern of the spatiotemporal attention weights is statistically analyzed. When the sparse distribution pattern is concentrated, the rank parameter of the tensor decomposition is reduced to enhance feature focusing. When the sparse distribution pattern is dispersed, the rank parameter of the tensor decomposition is increased to expand the feature capture range.

[0019] Based on the adjusted rank parameter, the tensor decomposition is re-executed to obtain the updated time factor matrix and the updated space factor matrix. The updated latent factor vector is extracted from the updated time factor matrix and the updated space factor matrix. The change in the dominant direction of the updated latent factor vector is calculated. The iteration stops when the change in the dominant direction is lower than the convergence threshold.

[0020] Based on the converged spatiotemporal attention weights, a node association structure is constructed to form a spatiotemporal dependency graph.

[0021] A knowledge graph for the geological hazard domain is constructed. The spatiotemporal dependency graph is fused with the geological hazard domain knowledge graph using heterogeneous graph fusion to obtain a fused graph. Inference is performed on the fused graph, and feature mapping between known hazard areas and unlabeled areas is established using contrastive learning, including:

[0022] Disaster type nodes, disaster-causing factor nodes, and environmental condition nodes are extracted from historical disaster cases. Based on the disaster evolution patterns among the disaster type nodes, disaster-causing factor nodes, and environmental condition nodes, causal link edges and condition trigger edges are constructed to form a knowledge graph in the field of geological disasters.

[0023] Identify the semantic correspondence between spatiotemporal nodes in the spatiotemporal dependency graph and disaster type nodes in the geological disaster domain knowledge graph, establish cross-graph bridging edges to connect the spatiotemporal nodes and the disaster type nodes, and integrate the spatiotemporal association edges of the spatiotemporal dependency graph with the causal link edges and condition trigger edges of the geological disaster domain knowledge graph into a unified edge set to form a fusion graph;

[0024] Extract the neighborhood subgraph structure of known disaster area nodes and the neighborhood subgraph structure of unlabeled area nodes from the fusion graph. Calculate the topological pattern features of cross-graph bridging edges, causal link edges, and spatiotemporal correlation edges in the neighborhood subgraph structure. Use the topological pattern features as the structural anchor point for contrastive learning.

[0025] Based on the structural anchor points, positive and negative sample pairs are constructed. By maximizing the feature similarity of the positive sample pairs and minimizing the feature similarity of the negative sample pairs, a feature mapping between known disaster areas and unlabeled areas is established.

[0026] Identifying the semantic correspondence between spatiotemporal nodes in the spatiotemporal dependency graph and disaster type nodes in the geological disaster domain knowledge graph, and establishing cross-graph bridges to connect the spatiotemporal nodes and the disaster type nodes includes:

[0027] Spatial location features and temporal evolution features of spatiotemporal nodes are extracted from the spatiotemporal dependency graph. Disaster-causing mechanism features and environmental constraint features of disaster type nodes are extracted from the knowledge graph of geological disaster domain. The spatial location features and the temporal evolution features are mapped to the data-driven semantic space, and the disaster-causing mechanism features and the environmental constraint features are mapped to the knowledge-driven semantic space.

[0028] Construct a cross-spatial semantic alignment function, and project the spatiotemporal node representations in the data-driven semantic space and the disaster type node representations in the knowledge-driven semantic space onto a unified semantic representation space through the cross-spatial semantic alignment function, and calculate the semantic matching degree between the spatiotemporal node representations and the disaster type node representations in the unified semantic representation space;

[0029] Spatiotemporal nodes with semantic matching degree exceeding a preset matching threshold and disaster type nodes are identified as node pairs with semantic correspondence. For the node pairs, the set of spatiotemporal adjacent nodes of the spatiotemporal nodes in the spatiotemporal dependency graph and the set of disaster adjacent nodes of the disaster type nodes in the geological disaster domain knowledge graph are extracted. The structural consistency metric between the set of spatiotemporal adjacent nodes and the set of disaster adjacent nodes is calculated.

[0030] Cross-graph bridge edges are jointly established based on the semantic matching degree and the structural consistency metric, and the spatiotemporal nodes are connected to the disaster type nodes.

[0031] By dynamically adjusting the inference path based on the inference confidence level through reinforcement learning, the disaster-causing mechanism can be transferred across regions, resulting in regional hazard risk assessment results, including:

[0032] Construct a set of candidate inference paths from known disaster area nodes to unlabeled area nodes on the fusion graph, and calculate the path-level inference confidence of each candidate inference path in the set of candidate inference paths;

[0033] A reinforcement learning agent is constructed, and the path-level inference confidence is used as a state input to the reinforcement learning agent. The reinforcement learning agent outputs a path selection strategy, and the path selection strategy dynamically allocates the activation weights of each candidate inference path according to the path-level inference confidence.

[0034] Based on the activation weights, candidate inference paths are selected to form a set of migration inference paths. Along the set of migration inference paths, the disaster-causing mechanism patterns of known disaster area nodes are propagated to unlabeled area nodes, forming a cross-regional migration disaster-causing mechanism representation of unlabeled area nodes.

[0035] Based on the disaster-causing mechanism of cross-regional migration, the hazard risk score of the unlabeled regional nodes is calculated, and the evaluation reward signal generated by the hazard risk score is fed back to the reinforcement learning agent to update the path selection strategy, so as to obtain the regional hazard risk assessment result.

[0036] Based on the feedback from the regional hazard risk assessment results, the relationship strength in the geological hazard knowledge graph is updated, and the spatial distribution and risk level of hazard points are output, including:

[0037] Extract the correctly predicted migration inference paths and the incorrectly predicted migration inference paths from the regional hidden danger risk assessment results, statistically analyze the activation frequency distribution of the causal link edges and conditional trigger edges involved in the migration inference paths, and calculate the confidence correction coefficient of each edge based on the activation frequency distribution.

[0038] The relationship strength of causal link edges and conditional trigger edges in the geological disaster domain knowledge graph is updated according to the confidence correction coefficient. Based on the updated relationship strength, inference propagation is re-executed on the geological disaster domain knowledge graph. The corrected risk score of unlabeled area nodes is calculated. According to the numerical distribution of the corrected risk score, the unlabeled area nodes are divided into different risk level intervals, and risk level labels are assigned to each unlabeled area node.

[0039] Extract the geographic coordinate attributes of the unlabeled area nodes of the risk level label, construct a spatial mapping relationship between the geographic coordinate attributes and the risk level label, and output the spatial distribution and risk level of the hazard points containing the spatial mapping relationship.

[0040] A second aspect of the present invention provides an intelligent monitoring and early warning system for geological disasters, comprising:

[0041] The data fusion unit is used to acquire multi-source heterogeneous spatiotemporal data of the target area, dynamically adjust the spatiotemporal resolution according to the density distribution of the multi-source heterogeneous spatiotemporal data, and map spatiotemporal data from different sources to a unified spatiotemporal reference framework to form a standardized spatiotemporal dataset.

[0042] The spatiotemporal modeling unit is used to perform tensor decomposition and spatiotemporal attention collaborative modeling on the standardized spatiotemporal dataset. It extracts the spatiotemporal coupling relationship of the standardized spatiotemporal dataset and captures multi-scale disaster evolution patterns through tensor decomposition. It uses the latent factors obtained by tensor decomposition as constraints for attention calculation and feeds back the attention weights to guide the rank optimization of the tensor decomposition, forming a spatiotemporal dependency graph.

[0043] The risk assessment unit is used to construct a knowledge graph in the field of geological hazards. It integrates the spatiotemporal dependency graph with the knowledge graph in the field of geological hazards to obtain a fusion graph. Reasoning is performed on the fusion graph. It uses contrastive learning to establish feature mapping between known disaster areas and unlabeled areas. Through reinforcement learning, it dynamically adjusts the reasoning path according to the reasoning confidence to realize the cross-regional migration of disaster-causing mechanisms and obtain regional hidden danger risk assessment results.

[0044] The result output unit is used to update the relationship strength in the geological disaster knowledge graph based on the feedback of the regional hidden danger risk assessment results, and output the spatial distribution and risk level of hidden danger points.

[0045] A third aspect of the present invention provides an electronic device, comprising:

[0046] processor;

[0047] Memory used to store processor-executable instructions;

[0048] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0049] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0050] By dynamically adjusting the spatiotemporal resolution of multi-source heterogeneous spatiotemporal data and mapping it to a unified benchmark, information misalignment and redundancy caused by differences in data sources are effectively eliminated, improving the accuracy and consistency of data fusion. The standardized dataset retains key spatiotemporal features, providing high-quality input for subsequent modeling, enabling monitoring and early warning to adapt to sparse or dense data distributions in different regions, and significantly enhancing adaptability to complex geological environments.

[0051] By leveraging a tensor decomposition and spatiotemporal attention synergy mechanism, this study deeply explores the coupling relationships between spatiotemporal data and multi-scale disaster evolution patterns. Attention computation is guided by latent factors, while the tensor decomposition rank is optimized through attention weight feedback, avoiding model bias caused by unreasonable rank presuppositions in traditional methods. The resulting spatiotemporal dependency graph accurately depicts the dynamic relationships between disaster-causing factors, enabling the model to capture subtle spatiotemporal changes and significantly improving the sensitivity to identifying disaster precursor features and the accuracy of predicting long-term evolution trends.

[0052] By heterogeneously fusing a spatiotemporal dependency graph with a knowledge graph in the field of geological hazards, and then establishing feature mappings between known and unlabeled areas through comparative learning, combined with reinforcement learning to dynamically adjust the inference path, the cross-regional transfer of disaster-causing mechanisms is achieved. This process effectively solves the challenge of risk assessment in areas with scarce samples, enabling the model to learn from the experience of similar geological disaster events, significantly improving the generalization ability and reliability of regional hazard risk assessment. Feedback updates the relationship strength in the knowledge graph, allowing the risk inference results to continuously self-correct with new data, continuously optimizing the spatial distribution of hazard points and the positioning accuracy of risk levels, providing dynamic and accurate decision support for geological disaster early warning. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the intelligent monitoring and early warning method for geological disasters according to an embodiment of the present invention;

[0054] Figure 2This is a flowchart illustrating the feedback optimization process for updating the knowledge graph relationship strength based on regional hazard risk assessment results, as described in an embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0057] Figure 1 This is a flowchart illustrating the intelligent monitoring and early warning method for geological disasters according to an embodiment of the present invention.

[0058] Intelligent monitoring and early warning methods for geological disasters include:

[0059] Acquire multi-source heterogeneous spatiotemporal data of the target region, dynamically adjust the spatiotemporal resolution according to the density distribution of the multi-source heterogeneous spatiotemporal data, and map spatiotemporal data from different sources to a unified spatiotemporal reference framework to form a standardized spatiotemporal dataset.

[0060] Tensor decomposition and spatiotemporal attention are performed on the standardized spatiotemporal dataset. The spatiotemporal coupling relationship of the standardized spatiotemporal dataset is extracted by tensor decomposition and multi-scale disaster evolution patterns are captured. The latent factors obtained by tensor decomposition are used as constraints for attention calculation. The attention weights are fed back to guide the rank optimization of the tensor decomposition, forming a spatiotemporal dependency graph.

[0061] A knowledge graph in the field of geological hazards is constructed. The spatiotemporal dependency graph is fused with the knowledge graph in the field of geological hazards to obtain a fusion graph. Reasoning is performed on the fusion graph. Feature mapping between known disaster areas and unlabeled areas is established by using contrastive learning. The reasoning path is dynamically adjusted according to the reasoning confidence through reinforcement learning to realize the cross-regional migration of disaster-causing mechanisms and obtain regional hidden danger risk assessment results.

[0062] Based on the feedback of the regional hidden danger risk assessment results, the relationship strength in the geological disaster knowledge graph is updated, and the spatial distribution and risk level of hidden danger points are output.

[0063] In one optional implementation, tensor decomposition and spatiotemporal attention co-modeling are performed on the standardized spatiotemporal dataset. The tensor decomposition extracts the spatiotemporal coupling relationships of the standardized spatiotemporal dataset and captures multi-scale disaster evolution patterns, including:

[0064] The standardized spatiotemporal dataset is constructed as a time-space-attribute third-order tensor, and tensor decomposition is performed on the third-order tensor to extract the time factor matrix, space factor matrix, and attribute factor matrix.

[0065] Spatiotemporal attention weights are calculated based on the time factor matrix and the spatial factor matrix, with the time factor matrix serving as the initial constraint for the time attention weights and the spatial factor matrix serving as the initial constraint for the spatial attention weights.

[0066] The spatiotemporal attention weights are fed back to the tensor decomposition process. The rank selection of the time factor matrix, the spatial factor matrix, and the attribute factor matrix is ​​dynamically adjusted according to the spatiotemporal attention weights. Through iterative optimization, the time factor matrix, the spatial factor matrix, the attribute factor matrix, and the spatiotemporal attention weights are made to achieve synergy and consistency.

[0067] Based on the time factor matrix, spatial factor matrix, and attribute factor matrix after coordination and consistency, the spatiotemporal coupling relationship of the standardized spatiotemporal dataset is reconstructed to capture multi-scale disaster evolution patterns.

[0068] Constructing a standardized spatiotemporal dataset into a time-space-attribute third-order tensor is a fundamental step in tensor decomposition and spatiotemporal attention co-modeling. Specifically, the time dimension, space dimension, and attribute dimension of the standardized spatiotemporal dataset are respectively mapped to the three mode axes of the tensor to form a third-order tensor. ,in Indicates the number of time steps. Indicates the number of nodes in spatial location. This indicates the number of attribute feature dimensions. These dimensions cover various geological hazard-related monitoring indicators, including rainfall, slope displacement, groundwater level, and soil moisture content. Through this third-order tensor organization method, spatiotemporal data from different monitoring sources are integrated within a unified tensor structure, avoiding the loss of spatiotemporal coupling information caused by dimension expansion in traditional matrix processing.

[0069] For third-order tensors Perform Tucker decomposition to decompose it into a time factor matrix. Spatial factor matrix and attribute factor matrix and kernel tensor ,in , , These are the initial rank parameters for the corresponding dimensions. The decomposition process uses an iterative solution with alternating least squares to minimize the reconstruction error. Continuously decreasing, among which Denotes the Frobenius norm. Indicates the tensor along the first Multiplication of each pattern. Time factor matrix. The evolutionary patterns between different time steps were captured, and the spatial factor matrix was used. Describes the association structure between spatial location nodes, attribute factor matrix This reflects the covariance relationship between different disaster monitoring indicators.

[0070] Based on the extracted time factor matrix and spatial factor matrix The temporal attention weights and spatial attention weights are calculated separately. The calculation of the temporal attention weights is based on... Each row vector serves as the initial query and key vector, and a time attention matrix is ​​obtained through a scaled dot product attention mechanism. , where matrix elements Indicates the first The time step for the first The initial constraint of the level of attention at each time step means that attention computation no longer relies entirely on the statistical distribution of the original data, but incorporates the low-rank temporal structure revealed by tensor decomposition. The spatial attention weights are similarly determined by... Each row vector in the matrix is ​​used as an initial constraint to calculate the spatial attention matrix. ,element Reflecting spatial nodes With nodes The strength of the correlation between them. and As an initial constraint rather than a direct replacement for attention computation, it effectively avoids the weight degradation problem caused by the lack of prior structural information in sparse data regions, which is particularly important in scenarios where geological disaster monitoring stations are unevenly distributed.

[0071] Spatiotemporal attention weights and This feedback is fed back to the tensor decomposition process to dynamically adjust the rank selection of each dimension. Specifically, it is based on the temporal attention matrix. The attention entropy value at each time step assesses the complexity of the information at that time step, and the time factor dimension is appropriately increased for time periods with higher information entropy. For time periods with high information redundancy, the corresponding rank parameter is compressed, thus achieving adaptive rank selection in the time dimension. The rank adjustment mechanism in the spatial dimension is similar; for spatial node clusters corresponding to areas prone to geological disasters, based on... Appropriately improve the clustering structure To preserve finer spatial resolution information. The rank of the attribute factor matrix. The key monitoring indicators revealed by the spatiotemporal attention weights are then adjusted, and attribute dimensions that contribute less to disaster early warning are merged and downranked to reduce computational redundancy.

[0072] During the iterative optimization process, tensor decomposition and spatiotemporal attention weights are updated alternately, forming a two-way constraint mechanism. In each iteration, the current spatiotemporal attention weights are first fixed, and the three factor matrices and the kernel tensor are updated; then, the updated factor matrices are fixed, and the spatiotemporal attention weights are recalculated. The iteration terminates when the change in the factor matrix between two consecutive iterations is less than a preset threshold. Or it may reach the maximum number of iterations limit. When the iteration converges, the time factor matrix... Spatial factor matrix Attribute factor matrix Spatiotemporal attention weights , Achieving a state of synergy and consistency means that the low-rank structure reflected by the factor matrix and the key dependencies emphasized by the attention weights corroborate each other, without any contradictions or oscillations.

[0073] Based on the three factor matrices after synergistic consistency, the spatiotemporal coupling relationship of the standardized spatiotemporal dataset is reconstructed, and multi-scale disaster evolution patterns are captured. On the time scale, through... Multi-resolution analysis was conducted to identify short-term (hourly to daily) triggering response patterns of sudden rainfall, medium-term (weekly to monthly) slope creep accumulation patterns, and long-term (seasonal to grade) geological tectonic activity cycle patterns. Spatially, this was achieved using... The clustering structure identifies disaster propagation links between different geomorphic units (such as valleys, slopes, and plateaus), revealing how local triggering factors can expand into regional disaster events through spatial neighborhood effects. At the attribute scale, the attribute factor matrix... This study reveals the synergistic response patterns among various indicators, including rainfall, displacement, and groundwater level. For example, it demonstrates the lag coupling relationship between peak rainfall and accelerated slope displacement under specific geological conditions, a pattern quantified using low-rank representations of attribute factors. The aforementioned multi-scale disaster evolution models are ultimately output as spatiotemporal dependency graphs, providing structured data-driven priors for subsequent heterogeneous graph fusion with knowledge graphs in the field of geological hazards.

[0074] In one optional implementation, the latent factors obtained from tensor decomposition are used as constraints for attention computation, and the attention weights are fed back to guide the rank optimization of the tensor decomposition, forming a spatiotemporal dependency graph, including:

[0075] Extract latent factor vectors from the temporal and spatial factor matrices obtained from tensor decomposition, calculate the dominant directions of the latent factor vectors in the temporal and spatial dimensions, and use the dominant directions as search constraints for spatiotemporal attention weights.

[0076] Under the search constraints, the spatiotemporal attention weights of the standardized spatiotemporal dataset are calculated, and the sparse distribution pattern of the spatiotemporal attention weights is statistically analyzed. When the sparse distribution pattern is concentrated, the rank parameter of the tensor decomposition is reduced to enhance feature focusing. When the sparse distribution pattern is dispersed, the rank parameter of the tensor decomposition is increased to expand the feature capture range.

[0077] Based on the adjusted rank parameter, the tensor decomposition is re-executed to obtain the updated time factor matrix and the updated space factor matrix. The updated latent factor vector is extracted from the updated time factor matrix and the updated space factor matrix. The change in the dominant direction of the updated latent factor vector is calculated. The iteration stops when the change in the dominant direction is lower than the convergence threshold.

[0078] Based on the converged spatiotemporal attention weights, a node association structure is constructed to form a spatiotemporal dependency graph.

[0079] Time factor matrix obtained from tensor decomposition and spatial factor matrix In the process, latent factor vectors corresponding to each dimension are extracted. For the time dimension, from... Extract the first A time latent factor vector ,in The range of values ​​is to Regarding spatial dimensions, from Extract the first spatial latent factor vectors ,in The range of values ​​is to After extraction, singular value analysis or principal component analysis is performed on all latent factor vectors in both the time and spatial dimensions to calculate the dominant direction, i.e., the direction with the largest variance contribution in each dimension. Dominant direction in the time dimension. The dominant direction of the spatial dimension is determined by the first right singular vector of the matrix composed of all temporal latent factor vectors. The first right singular vector of the matrix composed of all spatial latent factor vectors is determined. The dominant direction characterizes the distribution axis where information is most concentrated in the factor space, representing the most significant spatiotemporal variation patterns captured by the current tensor decomposition. and As a search constraint for subsequent attention weight calculation, the attention mechanism allocates weights within the principal component neighborhood of the factor space, thereby avoiding random diffusion of attention weights in irrelevant subspaces and improving the physical interpretability of attention calculation.

[0080] When calculating spatiotemporal attention weights under search constraints, the temporal attention matrix... The Middle The time step for the first Attention weights at each time step By projecting the time factor vector to the dominant direction After similarity calculation, the projection operation confines the attention computation domain to the subspace spanned by the dominant direction. Similarly, the spatial attention matrix... Middle node With nodes The strength of the correlation between By projecting the spatial factor vector to The cosine similarity is then calculated. The projected similarity is then normalized using softmax to ensure that the sum of the weights for each row is 1, forming a probabilistic attention distribution. After completing the attention weight calculation, the... and The weights in the time attention matrix are analyzed for sparse distribution patterns, specifically using the Gini coefficient or entropy as sparsity metrics. Let the row sparsity metric of the time attention matrix be... The row sparsity metric of the spatial attention matrix is Preset centralized judgment threshold and dispersion determination threshold .when This indicates that the time attention weights are highly concentrated on a few time steps, and the time dimension rank parameter... Correspondingly reduced to strengthen the focus on the dominant time pattern; when This indicates that the time attention weights are distributed across a large number of time steps. The corresponding increase is made to expand the range of capture for diverse temporal evolution patterns. Spatial dimension rank parameter in accordance with and , The relationship is adjusted symmetrically. Attribute dimension rank parameter. The settings remain unchanged in this iteration, and adjustments are only made when the attribute attention shows a similar sparse distribution shift, in order to ensure the independence and specificity of the rank parameter adjustments for each dimension.

[0081] Based on the adjusted , , The third-order tensor constructed from the standardized spatiotemporal dataset Re-perform the Tucker decomposition to obtain the updated kernel tensor. Updated time factor matrix Updated spatial factor matrix and the updated attribute factor matrix The Tucker decomposition is solved using alternating least squares, optimizing each factor matrix sequentially while keeping other factor matrices fixed, until the reconstruction error converges. Extract the updated set of time latent factor vectors from Extract the updated set of spatial latent factor vectors and calculate the updated temporal dominant directions for each. and the dominant direction of space The change in dominant direction is obtained by calculating the cosine distance between the two dominant direction vectors. The change in dominant direction in the time dimension is defined as follows: The change in the dominant direction of spatial dimension is defined as .when and If both conditions are met simultaneously, the iteration is considered converged and the iteration is stopped. This is the preset iteration termination threshold. If any condition is not met, then... and The current dominant direction is replaced, and the sparse distribution pattern judgment and rank parameter adjustment are re-executed with the updated attention weights, entering the next iteration. The entire iterative process constitutes a two-way collaborative optimization closed loop between tensor decomposition and the spatiotemporal attention mechanism, continuously guiding tensor decomposition towards a more physically meaningful factor structure through the sparse pattern feedback of the attention weights.

[0082] After iterative convergence, based on the final converged time attention matrix... Spatial attention matrix Construct a node association structure. In the time dimension, Exceeding the preset significance threshold Time step Defined as a pair of nodes with significant temporal dependencies, a directed edge is established on the time axis; in the spatial dimension, Exceed Spatial node pairs Undirected edges are established in the spatial graph for node pairs defined as having significant spatial relationships. Temporal and spatially related edges together constitute a heterogeneous node relationship structure. Nodes include two categories: time-step nodes and spatial location nodes, and edges are categorized as temporally dependent edges and spatially related edges. Based on this heterogeneous node relationship structure, the attribute feature vectors of each spatial location node at each time step are used as node features, and the attention weights of temporally dependent and spatially related edges are used as edge weights, forming a spatiotemporal dependency graph with node features and weighted edges. This spatiotemporal dependency graph fully depicts the coupling dependencies of geological hazard-related elements within the target area in both temporal evolution and spatial propagation dimensions, providing a structured data-driven foundation for subsequent heterogeneous graph fusion with geological hazard knowledge graphs.

[0083] In practical applications, for geological disaster monitoring scenarios, time nodes correspond to different monitoring times, and spatial nodes correspond to different monitoring stations or grid units. Time-dependent edges reveal the hysteresis response relationship between rainfall precursors and slope displacement, while spatially correlated edges reveal the stress transmission and hydrological connectivity relationships between adjacent slopes. Through the aforementioned bidirectional collaborative iterative mechanism, the spatiotemporal dependency graph can adaptively capture the multi-scale spatiotemporal coupling patterns corresponding to different geological disaster types, effectively supporting subsequent disaster-causing mechanism analysis and risk assessment.

[0084] In one optional implementation, a knowledge graph for the geological hazard domain is constructed, and the spatiotemporal dependency graph is heterogeneously fused with the geological hazard domain knowledge graph to obtain a fused graph. Inference is performed on the fused graph, and feature mapping between known hazard areas and unlabeled areas is established using contrastive learning, including:

[0085] Disaster type nodes, disaster-causing factor nodes, and environmental condition nodes are extracted from historical disaster cases. Based on the disaster evolution patterns among the disaster type nodes, disaster-causing factor nodes, and environmental condition nodes, causal link edges and condition trigger edges are constructed to form a knowledge graph in the field of geological disasters.

[0086] Identify the semantic correspondence between spatiotemporal nodes in the spatiotemporal dependency graph and disaster type nodes in the geological disaster domain knowledge graph, establish cross-graph bridging edges to connect the spatiotemporal nodes and the disaster type nodes, and integrate the spatiotemporal association edges of the spatiotemporal dependency graph with the causal link edges and condition trigger edges of the geological disaster domain knowledge graph into a unified edge set to form a fusion graph;

[0087] Extract the neighborhood subgraph structure of known disaster area nodes and the neighborhood subgraph structure of unlabeled area nodes from the fusion graph. Calculate the topological pattern features of cross-graph bridging edges, causal link edges, and spatiotemporal correlation edges in the neighborhood subgraph structure. Use the topological pattern features as the structural anchor point for contrastive learning.

[0088] Based on the structural anchor points, positive and negative sample pairs are constructed. By maximizing the feature similarity of the positive sample pairs and minimizing the feature similarity of the negative sample pairs, a feature mapping between known disaster areas and unlabeled areas is established.

[0089] When extracting structured knowledge from a historical disaster case database, corresponding disaster type nodes are established for different disaster types such as landslides, debris flows, and collapses. Each node carries attributes such as disaster category label, historical occurrence frequency, and typical characteristic descriptions. Disaster-causing factor nodes cover physical triggering factors such as rainfall intensity, seismic intensity, slope and aspect, soil and rock properties, and groundwater level. Each disaster-causing factor node records its quantitative threshold range and mechanism of action. Environmental condition nodes include background conditions such as vegetation cover, land use type, fault structure distribution, and topographic relief. After node construction, based on the disaster evolution patterns recorded in historical cases, causal link edges are established between disaster type nodes and disaster-causing factor nodes. The weight of the edge reflects the contribution intensity of the disaster-causing factor to a specific disaster type. Conditional trigger edges are established between disaster-causing factor nodes and environmental condition nodes. These trigger edges carry logical expressions for trigger conditions, such as "a landslide is triggered when rainfall intensity exceeds a critical threshold and the slope is greater than a specific angle." Through the systematic organization of these nodes and edges, a geological disaster knowledge graph capable of expressing the mechanism of disaster occurrence is formed.

[0090] When fusing spatiotemporal dependency graphs with knowledge graphs in the field of geological hazards, it is necessary to address the heterogeneity of the two types of graphs in terms of node semantic space. Spatiotemporal nodes in the spatiotemporal dependency graph carry numerical features such as sensor observations and remote sensing feature vectors, while hazard type nodes in the knowledge graph carry symbolic semantic attributes. Semantic correspondences are identified by calculating the cosine similarity between the feature vectors of spatiotemporal nodes and the semantic embedding vectors of hazard type nodes. When the similarity exceeds a preset semantic matching threshold, a cross-graph bridging edge is established between the corresponding spatiotemporal node and the hazard type node. The weight of the cross-graph bridging edge is directly assigned by the semantic similarity value, thereby quantifying the association strength between the two types of heterogeneous nodes. In the edge set integration stage, spatiotemporal association edges reflecting spatial proximity and temporal correlation in the spatiotemporal dependency graph, causal link edges in the knowledge graph, and conditional trigger edges are uniformly included in the same edge set, and different edge type labels are assigned to each type of edge to distinguish the semantic meaning of different types of edges in subsequent graph neural network calculations. The resulting fusion graph contains multiple node types and multiple edge types, forming a complete heterogeneous graph structure that can simultaneously carry spatiotemporal observation information and domain prior knowledge.

[0091] When extracting the neighborhood subgraph structure from the fusion graph, for each known disaster area node, its neighborhood subgraph is sampled with that node as the center, according to a fixed number of hops (e.g., 2-hop neighborhood). The neighborhood subgraph contains spatiotemporal nodes, disaster type nodes, disaster-causing factor nodes, and environmental condition nodes that are directly or indirectly connected to the central node, as well as cross-graph bridging edges, causal link edges, conditional trigger edges, and spatiotemporal association edges connecting these nodes. For unlabeled area nodes, the same neighborhood sampling strategy is used to extract their neighborhood subgraph structure. After the neighborhood subgraph structure is determined, its topological pattern features are calculated. The calculation of topological pattern features comprehensively considers the distribution of three types of edges in the subgraph: the number and weight distribution of cross-graph bridging edges are statistically analyzed to reflect the semantic association density between the nodes in the area and the knowledge graph; the type combination of causal link edges is statistically analyzed to capture the disaster causal link structure involved in the area; and the temporal span and spatial range of spatiotemporal association edges are statistically analyzed to characterize the spatiotemporal evolution characteristics of the area. The three types of statistics mentioned above are concatenated into a fixed-dimensional feature vector, which serves as the topological pattern feature of the neighborhood subgraph of the node and is defined as the structural anchor in contrastive learning. The design of the structural anchor ensures that the feature mapping process not only relies on the node's own attribute features but also fully utilizes graph structure information, thereby enhancing the discriminative power of the feature representation.

[0092] When constructing contrastive learning sample pairs based on structural anchors, positive sample pairs consist of known disaster area nodes and their closest unlabeled region nodes in the topological pattern feature space. This similarity is measured using Euclidean distance in the structural anchor feature space. Specifically, for each known disaster area node, several candidate nodes with the smallest structural anchor feature vector distance are retrieved from the set of unlabeled region nodes. These candidate nodes are paired with the known disaster area node to form positive sample pairs, indicating that both have similar graph structure backgrounds in disaster evolution patterns. Negative sample pairs consist of known disaster area nodes and their more distant unlabeled region nodes in the topological pattern feature space. The selection of negative samples employs a difficult negative sample mining strategy, prioritizing unlabeled nodes with moderate feature distances rather than randomly selecting the farthest node, thereby increasing the training difficulty of contrastive learning and the model's discriminative ability.

[0093] The objective function of contrastive learning optimizes the feature mapping network by maximizing the feature similarity of positive sample pairs and minimizing the feature similarity of negative sample pairs. Let the feature vector of a known disaster area node encoded by the graph neural network be... The feature vector of the node corresponding to the unlabeled region of the positive sample is , No. The feature vectors of nodes in the unlabeled regions of each negative sample are: The contrast loss function is:

[0094] ;

[0095] in Represents the cosine similarity function. This is a temperature hyperparameter used to control the smoothness of the similarity distribution. The number of negative samples corresponding to each positive sample. By minimizing The feature mapping network is optimized to map known disaster areas and unlabeled areas with similar topological patterns to adjacent locations in the feature space, while mapping areas with large differences in topological patterns to distant locations.

[0096] After training convergence, the feature vectors of nodes in known disaster areas are used as reference anchors to perform nearest neighbor retrieval on the feature vectors of nodes in unlabeled areas. Nodes in unlabeled areas whose features are most similar to those of known disaster areas are identified as potentially high-risk hazard areas, thus achieving cross-regional transfer of disaster-causing mechanisms. This process does not require manually labeled disaster category information in unlabeled areas; it relies solely on structural information in the fused map and the labels of known disaster areas to complete knowledge transfer from labeled to unlabeled areas, significantly expanding the coverage of intelligent geological disaster monitoring.

[0097] In one optional implementation, identifying the semantic correspondence between spatiotemporal nodes in the spatiotemporal dependency graph and disaster type nodes in the geological disaster domain knowledge graph, and establishing cross-graph bridges to connect the spatiotemporal nodes and the disaster type nodes includes:

[0098] Spatial location features and temporal evolution features of spatiotemporal nodes are extracted from the spatiotemporal dependency graph. Disaster-causing mechanism features and environmental constraint features of disaster type nodes are extracted from the knowledge graph of geological disaster domain. The spatial location features and the temporal evolution features are mapped to the data-driven semantic space, and the disaster-causing mechanism features and the environmental constraint features are mapped to the knowledge-driven semantic space.

[0099] Construct a cross-spatial semantic alignment function, and project the spatiotemporal node representations in the data-driven semantic space and the disaster type node representations in the knowledge-driven semantic space onto a unified semantic representation space through the cross-spatial semantic alignment function, and calculate the semantic matching degree between the spatiotemporal node representations and the disaster type node representations in the unified semantic representation space;

[0100] Spatiotemporal nodes with semantic matching degree exceeding a preset matching threshold and disaster type nodes are identified as node pairs with semantic correspondence. For the node pairs, the set of spatiotemporal adjacent nodes of the spatiotemporal nodes in the spatiotemporal dependency graph and the set of disaster adjacent nodes of the disaster type nodes in the geological disaster domain knowledge graph are extracted. The structural consistency metric between the set of spatiotemporal adjacent nodes and the set of disaster adjacent nodes is calculated.

[0101] Cross-graph bridge edges are jointly established based on the semantic matching degree and the structural consistency metric, and the spatiotemporal nodes are connected to the disaster type nodes.

[0102] In constructing heterogeneous graph fusion, it is necessary to establish a cross-graph bridging relationship between the spatiotemporal dependency graph and the knowledge graph in the field of geological hazards. Nodes in the spatiotemporal dependency graph carry spatiotemporal information extracted from multi-source data such as sensor data, remote sensing images, and topographic data, while nodes in the knowledge graph originate from prior knowledge in fields such as geology and hydrology. The two types of nodes belong to different semantic spaces, and directly comparing their representation vectors is meaningless. Therefore, it is necessary to first extract the features of each type of node separately, and then establish a comparable unified representation through a semantic alignment mechanism.

[0103] When extracting features from spatiotemporal nodes in a spatiotemporal dependency graph, spatial location features and temporal evolution features are obtained for each node. Spatial location features include static geographic attributes such as the node's geographic coordinates, slope, aspect, topographic relief, and stratigraphic lithology encoding, as well as the distribution of the node's topological connectivity with surrounding nodes in the spatiotemporal dependency graph. Temporal evolution features are obtained by extracting features from the node's observation sequence within a time window, specifically including feature vectors that characterize the dynamic process of disaster formation, such as displacement rate trends, slope changes in the cumulative rainfall response curve, and periodic fluctuations in soil moisture content. After concatenating these two types of features, a multi-layer fully connected transformation is applied to map them onto a data-driven semantic space, yielding the representation vector of each spatiotemporal node in that space. Its dimensions are .

[0104] When extracting features from disaster type nodes in a knowledge graph of geological hazards, the hazard-causing mechanism features and environmental constraint features are obtained for each disaster type node. The hazard-causing mechanism features are derived from structured knowledge recorded in the knowledge graph, such as disaster occurrence conditions, triggering factors, and physical and mechanical mechanisms. After encoding the knowledge graph using a graph neural network, the neighborhood aggregation representation of the node is extracted. The environmental constraint features describe the sensitivity constraints of the disaster type to environmental conditions such as topography, climate, and geological structure. Based on relation triples in the knowledge graph, the constraint conditions are encoded into dense vectors using relation embedding methods. After concatenating the hazard-causing mechanism features and environmental constraint features, a multi-layer fully connected transformation is performed to map them to a knowledge-driven semantic space, obtaining the representation vector of each disaster type node in this space. Its dimensions are .

[0105] Since the dimensions and distributions of data-driven semantic spaces and knowledge-driven semantic spaces are different, it is impossible to directly calculate similarity between the two spaces. Therefore, a cross-space semantic alignment function is constructed. This function consists of two independent projection networks, which respectively project... and Projected to dimension A unified semantic representation space. Specifically, the unified representation of spatiotemporal nodes is as follows: The unified representation of disaster type nodes is as follows: ,in and All are multilayer perceptrons with normalization layers. After projection, the semantic matching degree between spatiotemporal nodes and disaster type nodes is calculated in a unified semantic representation space. The normalized cosine similarity calculation method is adopted, that is The projection network parameters of the cross-spatial semantic alignment function are trained on labeled samples by supervised nodes. The correspondence between spatiotemporal observation data and disaster types in known disaster event records is used as a supervision signal to reduce the distance between similar node pairs in a unified semantic representation space and increase the distance between dissimilar node pairs.

[0106] When the semantic matching degree between a certain spatiotemporal node and a certain disaster type node Exceeding the preset matching threshold When the node pair is identified as a candidate node pair with a semantic correspondence, it enters the structural consistency verification stage. For spatiotemporal nodes in the candidate node pair, their first-order adjacent node set is extracted from the spatiotemporal dependency graph. This set contains all spatiotemporal nodes directly connected to the given spatiotemporal node; for disaster type nodes in candidate node pairs, their first-order adjacent node set is extracted from the geological disaster domain knowledge graph. This set contains all knowledge nodes that are directly connected to the disaster type node through any relation type.

[0107] Structural consistency measure It is calculated by comparing the semantic similarity distribution between two sets of adjacent nodes. Specifically, for Each spatiotemporal adjacency node in the and For each knowledge-adjacent node in the dataset, its representation vector is also calculated in the unified semantic representation space. Then, the distribution distance between the two sets is calculated using the optimal transmission method. This distance is then transformed using a monotonically decreasing method to obtain the structural consistency metric. When the semantic distributions of two adjacent node sets are more similar, The higher the value, the more consistent the candidate node pairs are not only in their own semantics, but also in their local graph structure, thus reducing the false positive rate of semantic alignment.

[0108] Based on semantic matching degree and structural consistency measurement The final cross-graph bridging strength is calculated using a weighted fusion method. ,Right now ,in This is the weighting coefficient for semantic matching, with a value range of [value range missing]. This can be adjusted based on data quality and the completeness of the knowledge graph. When Exceeding the cross-graph bridge establishment threshold At that time, a cross-graph bridge edge is established between the spatiotemporal node and the disaster type node, and the cross-graph bridge edge is established between the spatiotemporal node and the disaster type node. The initial weight of the bridging edge is stored in the edge attributes of the fusion graph. Each cross-graph bridging edge simultaneously records its corresponding semantic matching degree component and structural consistency component, which facilitates the differentiation and weighting of evidence from different sources during subsequent reasoning.

[0109] In practical applications, establishing cross-graph bridging edges between the same spatiotemporal node and multiple disaster type nodes indicates that the region faces threats from multiple disaster types simultaneously. In this case, all bridging edges that meet the threshold conditions are retained, and multi-label inference is performed on the fused graph, thereby supporting the risk identification of complex geological disasters. The establishment of cross-graph bridging edges allows observational information in the spatiotemporal dependency graph to propagate along the bridging edges to the knowledge graph side, activating relevant disaster-causing mechanism knowledge nodes; simultaneously, constraint information in the knowledge graph can also propagate back to the spatiotemporal nodes, injecting prior knowledge constraints into the data-driven feature representation, ultimately achieving bidirectional enhancement of data and knowledge on the fused graph.

[0110] In one optional implementation, reinforcement learning is used to dynamically adjust the inference path based on the inference confidence level, thereby enabling the cross-regional migration of disaster-causing mechanisms and obtaining regional hazard risk assessment results, including:

[0111] Construct a set of candidate inference paths from known disaster area nodes to unlabeled area nodes on the fusion graph, and calculate the path-level inference confidence of each candidate inference path in the set of candidate inference paths;

[0112] A reinforcement learning agent is constructed, and the path-level inference confidence is used as a state input to the reinforcement learning agent. The reinforcement learning agent outputs a path selection strategy, and the path selection strategy dynamically allocates the activation weights of each candidate inference path according to the path-level inference confidence.

[0113] Based on the activation weights, candidate inference paths are selected to form a set of migration inference paths. Along the set of migration inference paths, the disaster-causing mechanism patterns of known disaster area nodes are propagated to unlabeled area nodes, forming a cross-regional migration disaster-causing mechanism representation of unlabeled area nodes.

[0114] Based on the disaster-causing mechanism of cross-regional migration, the hazard risk score of the unlabeled regional nodes is calculated, and the evaluation reward signal generated by the hazard risk score is fed back to the reinforcement learning agent to update the path selection strategy, so as to obtain the regional hazard risk assessment result.

[0115] After the fusion graph is constructed, for the connectivity between known disaster area nodes and unlabeled area nodes, all directed walk sequences that satisfy the path length constraint are enumerated to form a candidate inference path set. Each candidate inference path consists of a series of alternating nodes and edges, with edge types covering spatiotemporal adjacency, disaster causal relationships in the knowledge graph, and cross-graph bridge edges. The upper limit of path length is usually set to 3 to 5 hops to achieve a balance between expressive power and computational complexity. For each path in the candidate inference path set, its path-level inference confidence is calculated. Specifically, the relational confidence of each edge on the path (determined jointly by the relational strength in the knowledge graph and the edge weights in the spatiotemporal dependency graph) is comprehensively aggregated with the feature similarity of the nodes at the path endpoints. Let a certain candidate path... Depend on Composed of 3 sides, the first The reliability of the relationship between the edges is Then the path-level inference confidence of this path By calculating the weighted geometric average of the confidence levels along each edge of the path, and introducing a path length penalty factor to suppress noise accumulation caused by excessively long paths, short and high-confidence paths are given higher inference priority.

[0116] The state space of the reinforcement learning agent consists of the confidence vectors of all paths in the current candidate inference path set, while the action space is a continuous output of activation weights assigned to each candidate path. The agent is implemented using a policy gradient framework. Its policy network receives the path-level inference confidence vectors as input and outputs the activation weight distribution corresponding to each candidate path. The activation weights are normalized using softmax to ensure a sum of 1, and the weight values ​​are positively correlated with the path confidence—paths with higher confidence receive larger activation weights, thus undertaking a more significant role in knowledge dissemination during subsequent transfer inference. The policy network employs a multi-layer fully connected structure, with residual connections introduced in the hidden layers to stabilize the training process, and dropout regularization to prevent overfitting. In the early stages of training, the agent randomly perturbs the activation weight distribution with a high exploration rate. As the training epochs increase, the exploration rate is gradually reduced, causing the policy to converge to a stable path selection preference.

[0117] When filtering candidate inference paths based on activation weights, an activation weight threshold is set. Only paths with activation weights exceeding a certain threshold are retained to form the transfer inference path set. This filtering mechanism effectively eliminates redundant paths with low confidence and high noise, ensuring that subsequent propagation of the disaster-causing mechanism focuses on the most information-rich structural channels. The number of paths in the transfer inference path set is usually much smaller than the original candidate set, but it covers the most critical causal propagation links in the fusion graph. In extreme cases, if the activation weights of all candidate paths are lower than a certain threshold... If the threshold is adaptively reduced to 50% of the current maximum activation weight, it will ensure that there is at least one valid migration path and avoid interruption in the inference process.

[0118] When propagating the disaster-causing mechanism along the set of transfer reasoning paths, starting from the known disaster area nodes, the disaster-causing mechanism pattern representation (encoded by the feature vector obtained in the contrastive learning phase) is passed hop-by-hop along the path to the unlabeled area nodes. During each hop of propagation, the reliability of the relationship at the current edge is used as the basis for the propagation. As a propagation attenuation coefficient, the disaster-causing mechanism representation during propagation is weighted and modulated to ensure that the information contribution is reduced accordingly after propagation through low-confidence edges. When multiple migration inference paths point to the same unlabeled region node, the propagation results from different paths are weighted and fused according to the activation weights of each path to form a cross-regional migration disaster-causing mechanism representation for that unlabeled region node. Let the unlabeled region node be... Received from The propagation signal of the migration path, the first The propagation vector corresponding to each path is Activation weight is ,but The mechanism of disaster caused by cross-regional migration By weighted summation The calculation shows that the vector comprehensively encodes information about different combinations of disaster-causing factors carried by multiple propagation paths.

[0119] Based on the disaster-causing mechanism of cross-regional migration, a quantitative score is given for the hidden risks of nodes in unlabeled areas. Input a risk scoring network, which outputs a scalar risk score. The risk scoring network is defined by a 0-point scale, where 0 represents no risk and 1 represents extremely high risk. It consists of two fully connected layers and a sigmoid activation function. Supervised pre-training is performed on labeled data from known disaster areas, followed by fine-tuning using reward signals during the reinforcement learning phase. The reward signal design considers two factors: firstly, a positive reward is given if the current risk score matches subsequent observed disaster events (i.e., a disaster actually occurred in the high-risk area); secondly, a negative reward is given if the score leads to a large number of false positives or false negatives. The reward signal is fed back to the policy network using a temporal difference method to update the parameters of the path selection policy, enabling the agent to gradually learn path activation patterns that generate more accurate risk assessments.

[0120] After multiple rounds of iterative training, the path selection strategy stabilized, and the risk scores of unlabeled nodes converged. All node risk scores were mapped to discrete risk levels according to a pre-defined risk level classification rule: nodes with scores below 0.3 were classified as low risk, those between 0.3 and 0.6 as medium risk, those between 0.6 and 0.8 as high risk, and those above 0.8 as extremely high risk. Combining the spatial coordinates of the nodes, a spatial distribution map of hazard points and a risk level zoning map of the target area were generated, which constitutes the regional hazard risk assessment result. This result was subsequently fed back into the geological hazard knowledge graph to update the relationship strength between related nodes, forming an online adaptive update loop for the knowledge graph, allowing subsequent reasoning processes to continuously benefit from newly accumulated assessment experience.

[0121] In one optional implementation, the relationship strength in the geological hazard knowledge graph is updated based on the feedback of the regional hazard risk assessment results, and the spatial distribution and risk level of the hazard points are output, including:

[0122] Extract the correctly predicted migration inference paths and the incorrectly predicted migration inference paths from the regional hidden danger risk assessment results, statistically analyze the activation frequency distribution of the causal link edges and conditional trigger edges involved in the migration inference paths, and calculate the confidence correction coefficient of each edge based on the activation frequency distribution.

[0123] The relationship strength of causal link edges and conditional trigger edges in the geological disaster domain knowledge graph is updated according to the confidence correction coefficient. Based on the updated relationship strength, inference propagation is re-executed on the geological disaster domain knowledge graph. The corrected risk score of unlabeled area nodes is calculated. According to the numerical distribution of the corrected risk score, the unlabeled area nodes are divided into different risk level intervals, and risk level labels are assigned to each unlabeled area node.

[0124] Extract the geographic coordinate attributes of the unlabeled area nodes of the risk level label, construct a spatial mapping relationship between the geographic coordinate attributes and the risk level label, and output the spatial distribution and risk level of the hazard points containing the spatial mapping relationship.

[0125] like Figure 2 As shown, the method includes:

[0126] After obtaining the regional hazard risk assessment results, the correct and incorrect reasoning information contained in the assessment results needs to be fed back to the geological hazard knowledge graph to continuously correct the strength of various relationship edges in the graph, thereby achieving adaptive updating of the knowledge graph and accurate output of hazard points.

[0127] Based on the regional hazard risk assessment results and the actual disaster occurrence of the labeled samples, the migration inference paths are divided into a set of correctly predicted migration inference paths and a set of incorrectly predicted migration inference paths. Correctly predicted paths indicate that the causal link edges and conditional trigger edges traversed by the path have high explanatory power in the current geological environment, and their activation behavior should be positively reinforced. Incorrectly predicted paths indicate that there are edges in the path with overestimated relationship strength or logical chains that are not suitable for the current regional characteristics, and they need to be negatively corrected. The two sets of paths are traversed separately, and the number of times each causal link edge and conditional trigger edge is activated in the correct path is counted. and the number of times it was activated in the wrong path. This forms an activation frequency distribution. The activation frequency distribution reflects the degree of participation of each edge in different inference quality paths and is the direct basis for calculating the confidence correction coefficient.

[0128] Calculate each edge based on the activation frequency distribution. Confidence correction coefficient The confidence correction coefficient takes into account the relative proportions of positive and negative activation frequencies. A higher positive activation frequency and a lower negative activation frequency result in a larger correction coefficient, and vice versa. Specifically, Depend on and The normalized difference between them determines the smoothing term. To prevent the denominator from being zero, express it as ,in This is a regularization smoothing constant, and its value is usually set to a small positive number to ensure numerical stability. The range of values ​​is A positive value indicates that the relationship strength of that edge should be strengthened, while a negative value indicates that it should be weakened.

[0129] Based on confidence level correction coefficient The strength of the original relationships between causal links and conditional triggering edges in the knowledge graph of the geological disaster domain. Update the relationship to obtain the updated relationship strength. The update rule is as follows: ,in The update step size for relation strength is used to control the magnitude of each round of feedback updates. and These are the lower and upper bounds of the relation strength, respectively. The updated relation strength is constrained within a reasonable range using a clip operation to prevent extreme values ​​from disrupting the overall structural consistency of the knowledge graph. This update is performed on all causal link edges and conditional trigger edges involved in the current round of reasoning in the knowledge graph; edges not activated by any reasoning path retain their original relation strength.

[0130] After updating the relation strength, inference propagation is re-executed on the fusion graph based on the updated knowledge graph. The inference propagation process transmits disaster-causing mechanism information from known disaster area nodes to unlabeled area nodes along each migration path. The propagation weight of each edge on the path is determined by its updated relation strength. It is determined that edges with higher relationship strength contribute more to information transmission. After re-inference propagation, a revised risk score is calculated for each unlabeled region node. The revised risk score integrates the weighted aggregation results of disaster-causing mechanism vectors propagated to the node from multiple migration paths, reflecting the degree of geological hazard risk at the node after the knowledge graph relationship strength is updated.

[0131] Based on the revised risk score The numerical distribution of the risk scores was used to divide all unlabeled nodes into different risk level intervals. The risk level division employed a quantile method, setting several quantile cut-off points based on the distribution of the modified risk scores across all nodes, thus dividing the nodes into four risk levels: low, medium, high, and extremely high. The selection of quantile cut-off points balanced the balance of the number of nodes within each risk level interval with the non-linear distribution characteristics of actual geological hazard risks. For modified risk scores exceeding the extremely high risk threshold... Regardless of their quantile position, all nodes are forcibly assigned an extremely high-risk label to ensure that high-risk potential points are not underestimated. After the risk level classification is completed, each unlabeled area node obtains a corresponding risk level label, including the level category identifier and the specific value of the corrected risk score, which provides a quantitative basis for subsequent spatial distribution output.

[0132] Extract the geographic coordinate attributes of all unlabeled area nodes with assigned risk level labels. The geographic coordinate attributes are derived from the spatial location information of each node in the standardized spatiotemporal dataset, including longitude. with latitude The system consists of two components, with precision consistent with the spatiotemporal reference framework of the target area. Geographic coordinate attributes are associated with the risk level labels of corresponding nodes to construct a spatial mapping relationship. This spatial mapping relationship uses nodes as the basic unit; each node's spatial mapping record includes its geographic coordinates, risk level category, revised risk score, and identifiers of the main migration paths involved in generating that score. This ensures that the output results possess complete spatial positioning information and traceable reasoning basis.

[0133] After the spatial mapping relationship is constructed, the spatial distribution and risk level of potential hazards are output in a structured form. The output includes the geographical location, risk level label, and corrected risk score of all identified potential hazards within the target area, supporting multiple search methods such as filtering by risk level, querying by spatial range, and sorting by score. For high-risk and extremely high-risk nodes in adjacent geographical locations, further spatial clustering is performed, grouping nodes with a spatial distance less than a preset aggregation radius. Nodes of the same level are merged into a hidden danger area. The coordinates of its outer rectangular boundary and the highest corrected risk score within the area are output as a unit, which makes it easier for disaster prevention and control departments to quickly locate key prevention and control areas and formulate targeted early warning and response measures.

[0134] A second aspect of the present invention provides an intelligent monitoring and early warning system for geological disasters, comprising:

[0135] The data fusion unit is used to acquire multi-source heterogeneous spatiotemporal data of the target area, dynamically adjust the spatiotemporal resolution according to the density distribution of the multi-source heterogeneous spatiotemporal data, and map spatiotemporal data from different sources to a unified spatiotemporal reference framework to form a standardized spatiotemporal dataset.

[0136] The spatiotemporal modeling unit is used to perform tensor decomposition and spatiotemporal attention collaborative modeling on the standardized spatiotemporal dataset. It extracts the spatiotemporal coupling relationship of the standardized spatiotemporal dataset and captures multi-scale disaster evolution patterns through tensor decomposition. It uses the latent factors obtained by tensor decomposition as constraints for attention calculation and feeds back the attention weights to guide the rank optimization of the tensor decomposition, forming a spatiotemporal dependency graph.

[0137] The risk assessment unit is used to construct a knowledge graph in the field of geological hazards. It integrates the spatiotemporal dependency graph with the knowledge graph in the field of geological hazards to obtain a fusion graph. Reasoning is performed on the fusion graph. It uses contrastive learning to establish feature mapping between known disaster areas and unlabeled areas. Through reinforcement learning, it dynamically adjusts the reasoning path according to the reasoning confidence to realize the cross-regional migration of disaster-causing mechanisms and obtain regional hidden danger risk assessment results.

[0138] The result output unit is used to update the relationship strength in the geological disaster knowledge graph based on the feedback of the regional hidden danger risk assessment results, and output the spatial distribution and risk level of hidden danger points.

[0139] A third aspect of the present invention provides an electronic device, comprising:

[0140] processor;

[0141] Memory used to store processor-executable instructions;

[0142] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0143] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0144] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent monitoring and early warning of geological disasters, characterized in that, include: Acquire multi-source heterogeneous spatiotemporal data of the target region, dynamically adjust the spatiotemporal resolution according to the density distribution of the multi-source heterogeneous spatiotemporal data, and map spatiotemporal data from different sources to a unified spatiotemporal reference framework to form a standardized spatiotemporal dataset. Tensor decomposition and spatiotemporal attention are performed on the standardized spatiotemporal dataset. The spatiotemporal coupling relationship of the standardized spatiotemporal dataset is extracted by tensor decomposition and multi-scale disaster evolution patterns are captured. The latent factors obtained by tensor decomposition are used as constraints for attention calculation. The attention weights are fed back to guide the rank optimization of the tensor decomposition, forming a spatiotemporal dependency graph. A knowledge graph in the field of geological hazards is constructed. The spatiotemporal dependency graph is fused with the knowledge graph in the field of geological hazards to obtain a fusion graph. Reasoning is performed on the fusion graph. Feature mapping between known disaster areas and unlabeled areas is established by using contrastive learning. The reasoning path is dynamically adjusted according to the reasoning confidence through reinforcement learning to realize the cross-regional migration of disaster-causing mechanisms and obtain regional hidden danger risk assessment results. Based on the feedback of the regional hidden danger risk assessment results, the relationship strength in the geological disaster knowledge graph is updated, and the spatial distribution and risk level of hidden danger points are output.

2. The method according to claim 1, characterized in that, Tensor decomposition and spatiotemporal attention co-modeling are performed on the standardized spatiotemporal dataset. Tensor decomposition is used to extract the spatiotemporal coupling relationships of the standardized spatiotemporal dataset and capture multi-scale disaster evolution patterns, including: The standardized spatiotemporal dataset is constructed as a time-space-attribute third-order tensor, and tensor decomposition is performed on the third-order tensor to extract the time factor matrix, space factor matrix, and attribute factor matrix. Spatiotemporal attention weights are calculated based on the time factor matrix and the spatial factor matrix, with the time factor matrix serving as the initial constraint for the time attention weights and the spatial factor matrix serving as the initial constraint for the spatial attention weights. The spatiotemporal attention weights are fed back to the tensor decomposition process. The rank selection of the time factor matrix, the spatial factor matrix, and the attribute factor matrix is ​​dynamically adjusted according to the spatiotemporal attention weights. Through iterative optimization, the time factor matrix, the spatial factor matrix, the attribute factor matrix, and the spatiotemporal attention weights are made to achieve synergy and consistency. Based on the time factor matrix, spatial factor matrix, and attribute factor matrix after coordination and consistency, the spatiotemporal coupling relationship of the standardized spatiotemporal dataset is reconstructed to capture multi-scale disaster evolution patterns.

3. The method according to claim 1, characterized in that, Using the latent factors obtained from tensor decomposition as constraints for attention computation, and feeding back the attention weights to guide the rank optimization of the tensor decomposition, a spatiotemporal dependency graph is formed, including: Extract latent factor vectors from the temporal and spatial factor matrices obtained from tensor decomposition, calculate the dominant directions of the latent factor vectors in the temporal and spatial dimensions, and use the dominant directions as search constraints for spatiotemporal attention weights. Under the search constraints, the spatiotemporal attention weights of the standardized spatiotemporal dataset are calculated, and the sparse distribution pattern of the spatiotemporal attention weights is statistically analyzed. When the sparse distribution pattern is concentrated, the rank parameter of the tensor decomposition is reduced to enhance feature focusing. When the sparse distribution pattern is dispersed, the rank parameter of the tensor decomposition is increased to expand the feature capture range. Based on the adjusted rank parameter, the tensor decomposition is re-executed to obtain the updated time factor matrix and the updated space factor matrix. The updated latent factor vector is extracted from the updated time factor matrix and the updated space factor matrix. The change in the dominant direction of the updated latent factor vector is calculated. The iteration stops when the change in the dominant direction is lower than the convergence threshold. Based on the converged spatiotemporal attention weights, a node association structure is constructed to form a spatiotemporal dependency graph.

4. The method according to claim 1, characterized in that, A knowledge graph for the geological hazard domain is constructed. The spatiotemporal dependency graph is fused with the geological hazard domain knowledge graph using heterogeneous graph fusion to obtain a fused graph. Inference is performed on the fused graph, and feature mapping between known hazard areas and unlabeled areas is established using contrastive learning, including: Disaster type nodes, disaster-causing factor nodes, and environmental condition nodes are extracted from historical disaster cases. Based on the disaster evolution patterns among the disaster type nodes, disaster-causing factor nodes, and environmental condition nodes, causal link edges and condition trigger edges are constructed to form a knowledge graph in the field of geological disasters. Identify the semantic correspondence between spatiotemporal nodes in the spatiotemporal dependency graph and disaster type nodes in the geological disaster domain knowledge graph, establish cross-graph bridging edges to connect the spatiotemporal nodes and the disaster type nodes, and integrate the spatiotemporal association edges of the spatiotemporal dependency graph with the causal link edges and condition trigger edges of the geological disaster domain knowledge graph into a unified edge set to form a fusion graph; Extract the neighborhood subgraph structure of known disaster area nodes and the neighborhood subgraph structure of unlabeled area nodes from the fusion graph. Calculate the topological pattern features of cross-graph bridging edges, causal link edges, and spatiotemporal correlation edges in the neighborhood subgraph structure. Use the topological pattern features as the structural anchor point for contrastive learning. Based on the structural anchor points, positive and negative sample pairs are constructed. By maximizing the feature similarity of the positive sample pairs and minimizing the feature similarity of the negative sample pairs, a feature mapping between known disaster areas and unlabeled areas is established.

5. The method according to claim 4, characterized in that, Identifying the semantic correspondence between spatiotemporal nodes in the spatiotemporal dependency graph and disaster type nodes in the geological disaster domain knowledge graph, and establishing cross-graph bridges to connect the spatiotemporal nodes and the disaster type nodes includes: Spatial location features and temporal evolution features of spatiotemporal nodes are extracted from the spatiotemporal dependency graph. Disaster-causing mechanism features and environmental constraint features of disaster type nodes are extracted from the knowledge graph of geological disaster domain. The spatial location features and the temporal evolution features are mapped to the data-driven semantic space, and the disaster-causing mechanism features and the environmental constraint features are mapped to the knowledge-driven semantic space. Construct a cross-spatial semantic alignment function, and project the spatiotemporal node representations in the data-driven semantic space and the disaster type node representations in the knowledge-driven semantic space onto a unified semantic representation space through the cross-spatial semantic alignment function, and calculate the semantic matching degree between the spatiotemporal node representations and the disaster type node representations in the unified semantic representation space; Spatiotemporal nodes with semantic matching degree exceeding a preset matching threshold and disaster type nodes are identified as node pairs with semantic correspondence. For the node pairs, the set of spatiotemporal adjacent nodes of the spatiotemporal nodes in the spatiotemporal dependency graph and the set of disaster adjacent nodes of the disaster type nodes in the geological disaster domain knowledge graph are extracted. The structural consistency metric between the set of spatiotemporal adjacent nodes and the set of disaster adjacent nodes is calculated. Cross-graph bridge edges are jointly established based on the semantic matching degree and the structural consistency metric, and the spatiotemporal nodes are connected to the disaster type nodes.

6. The method according to claim 1, characterized in that, By dynamically adjusting the inference path based on the inference confidence level through reinforcement learning, the disaster-causing mechanism can be transferred across regions, resulting in regional hazard risk assessment results, including: Construct a set of candidate inference paths from known disaster area nodes to unlabeled area nodes on the fusion graph, and calculate the path-level inference confidence of each candidate inference path in the set of candidate inference paths; A reinforcement learning agent is constructed, and the path-level inference confidence is used as a state input to the reinforcement learning agent. The reinforcement learning agent outputs a path selection strategy, and the path selection strategy dynamically allocates the activation weights of each candidate inference path according to the path-level inference confidence. Based on the activation weights, candidate inference paths are selected to form a set of migration inference paths. Along the set of migration inference paths, the disaster-causing mechanism patterns of known disaster area nodes are propagated to unlabeled area nodes, forming a cross-regional migration disaster-causing mechanism representation of unlabeled area nodes. Based on the disaster-causing mechanism of cross-regional migration, the hazard risk score of the unlabeled regional nodes is calculated, and the evaluation reward signal generated by the hazard risk score is fed back to the reinforcement learning agent to update the path selection strategy, so as to obtain the regional hazard risk assessment result.

7. The method according to claim 1, characterized in that, Based on the feedback from the regional hazard risk assessment results, the relationship strength in the geological hazard knowledge graph is updated, and the spatial distribution and risk level of hazard points are output, including: Extract the correctly predicted migration inference paths and the incorrectly predicted migration inference paths from the regional hidden danger risk assessment results, statistically analyze the activation frequency distribution of the causal link edges and conditional trigger edges involved in the migration inference paths, and calculate the confidence correction coefficient of each edge based on the activation frequency distribution. The relationship strength of causal link edges and conditional trigger edges in the geological disaster domain knowledge graph is updated according to the confidence correction coefficient. Based on the updated relationship strength, inference propagation is re-executed on the geological disaster domain knowledge graph. The corrected risk score of unlabeled area nodes is calculated. According to the numerical distribution of the corrected risk score, the unlabeled area nodes are divided into different risk level intervals, and risk level labels are assigned to each unlabeled area node. Extract the geographic coordinate attributes of the unlabeled area nodes of the risk level label, construct a spatial mapping relationship between the geographic coordinate attributes and the risk level label, and output the spatial distribution and risk level of the hazard points containing the spatial mapping relationship.

8. A geological disaster intelligent monitoring and early warning system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The data fusion unit is used to acquire multi-source heterogeneous spatiotemporal data of the target area, dynamically adjust the spatiotemporal resolution according to the density distribution of the multi-source heterogeneous spatiotemporal data, and map spatiotemporal data from different sources to a unified spatiotemporal reference framework to form a standardized spatiotemporal dataset. The spatiotemporal modeling unit is used to perform tensor decomposition and spatiotemporal attention collaborative modeling on the standardized spatiotemporal dataset. It extracts the spatiotemporal coupling relationship of the standardized spatiotemporal dataset and captures multi-scale disaster evolution patterns through tensor decomposition. It uses the latent factors obtained by tensor decomposition as constraints for attention calculation and feeds back the attention weights to guide the rank optimization of the tensor decomposition, forming a spatiotemporal dependency graph. The risk assessment unit is used to construct a knowledge graph in the field of geological hazards. It integrates the spatiotemporal dependency graph with the knowledge graph in the field of geological hazards to obtain a fusion graph. Reasoning is performed on the fusion graph. It uses contrastive learning to establish feature mapping between known disaster areas and unlabeled areas. Through reinforcement learning, it dynamically adjusts the reasoning path according to the reasoning confidence to realize the cross-regional migration of disaster-causing mechanisms and obtain regional hidden danger risk assessment results. The result output unit is used to update the relationship strength in the geological disaster knowledge graph based on the feedback of the regional hidden danger risk assessment results, and output the spatial distribution and risk level of hidden danger points.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images