Construction environment sudden change risk field prediction method fusing geological and meteorological data

Abstract

The invention discloses a construction environment sudden change risk field prediction method fusing geological and meteorological data, and belongs to the technical field of construction engineering risk prediction, and the method comprises the steps: obtaining geological data, meteorological data and construction progress data of a construction area, carrying out the structural processing of the multi-modal heterogeneous data, and carrying out the construction engineering risk prediction. And generating graph data containing the multi-modal features. According to the method, a space-time adaptive graph neural network prediction model based on an attention mechanism is adopted, weights of different features can be dynamically calculated and allocated according to real-time environment and construction data, and a space-time propagation relationship of risks among geographic nodes is captured by using a message passing mechanism of a graph. According to the method, data of different modals can be deeply fused, accurate prediction of the construction environment sudden change risk is realized, corresponding early warning is triggered according to the risk level, the accuracy and timeliness of risk prediction are improved, and scientific support is provided for intelligent decision making of constructional engineering.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of construction engineering risk prediction, in particular to a construction environment mutation risk field prediction method fusing geological and meteorological data. BACKGROUND

[0002] In the field of construction engineering, safety risk prediction of construction environment is a key link to ensure personnel life and property safety and smooth project progress. With the development of information technology, through the fusion of geological exploration, meteorological monitoring and construction progress and other multi-source data, intelligent risk early warning using data-driven computing models has become an important development direction. These methods aim to analyze historical and real-time data to identify potential sudden risks such as landslides, collapses, etc. in advance.

[0003] In the prior art, some risk prediction methods attempt to use multiple models to process different types of data. For example, one model is used to analyze geological data to assess the inherent stability of the slope, while another independent model is used to analyze meteorological data to predict the impact of heavy rainfall events, and finally the output results of the two models are simply weighted and fused or logically judged to obtain the final risk level. This multi-model serial processing procedure is a common implementation method in current technology.

[0004] However, the above prior art solution has obvious defects. First, the data from different sources is processed by independent models, ignoring the complex nonlinear interaction between these data modalities, resulting in a one-sided understanding of the causes of risk. Second, this type of method usually uses fixed model parameters and analysis logic, which cannot dynamically adjust the prediction focus according to the changes in construction stages or environmental conditions. In addition, most existing prediction models focus on the risk state of a single monitoring point, making it difficult to effectively simulate the spatial correlation and transmission process of risk in the entire construction area. SUMMARY

[0005] To solve the above problems, the present application provides a construction environment mutation risk field prediction method fusing geological and meteorological data, which uses a spatio-temporal adaptive graph neural network model to dynamically fuse multi-source heterogeneous data and simulate the spatio-temporal propagation of risk, thereby improving the accuracy and foresight of the prediction.

[0006] The above object can be achieved by the following solution: A construction environment mutation risk field prediction method fusing geological and meteorological data, comprising: obtaining geological data, meteorological data and construction progress data of a construction area, and structuring these data; Based on the structured data, the construction area is divided into a plurality of geographical nodes, and edges between the geographical nodes are constructed according to geographical spatial relationship and geological correlation to generate graph data containing multi-modal features; fusing and encoding features of different modalities in the graph data to generate a unified feature vector; inputting the unified feature vector into a spatio-temporal adaptive graph neural network prediction model, dynamically calculating and distributing weights of different modalities according to real-time construction progress data and meteorological data through an attention mechanism inside the spatio-temporal adaptive graph neural network prediction model; capturing spatio-temporal propagation relationships of risks between geographical nodes by using a message passing mechanism of a graph, and outputting mutation risk values of the nodes; generating a mutation risk field map of the construction environment according to the mutation risk values, and triggering corresponding risk warnings according to risk levels.

[0007] Optionally, the obtaining of the geological data, meteorological data and construction progress data of the construction area comprises: obtaining the geological data through a geological sensor deployed at the construction site; obtaining the meteorological data through an automatic weather station; and obtaining the construction progress data by accessing a construction plan database.

[0008] Optionally, the constructing of edges between geographical nodes according to geographical spatial relationships and geological correlations comprises: identifying adjacent geographical nodes in geographical positions, and establishing first-type edges between the adjacent geographical nodes; identifying geographical nodes located in the same geological structure unit, and establishing second-type edges between the geographical nodes located in the same geological structure unit; and merging the first-type edges and the second-type edges to form a complete edge set of the graph data.

[0009] Optionally, the fusing and encoding of features of different modalities in the graph data to generate a unified feature vector comprises: processing the geological data by using a convolutional neural network to extract spatial features and generate a geological feature vector; processing the meteorological data and the construction progress data by using a recurrent neural network to extract time sequence features and generate a time sequence feature vector; and performing feature cross fusion on the geological feature vector and the time sequence feature vector to generate the unified feature vector.

[0010] Optionally, the feature cross fusion on the geological feature vector and the time sequence feature vector comprises: splicing the geological feature vector and the time sequence feature vector to form a combined vector; performing linear transformation on the combined vector, and applying a nonlinear activation function to a result of the linear transformation to generate the unified feature vector.

[0011] Optionally, the unified feature vector is input into a spatio-temporal adaptive graph neural network prediction model, and weights of different modal features are dynamically calculated and distributed according to real-time construction progress data and meteorological data through an attention mechanism inside the spatio-temporal adaptive graph neural network prediction model, including: encoding the construction progress data and the meteorological data into a context query vector; calculating attention weights of each modal feature according to the context query vector; weighting the corresponding modal features in the unified feature vector using the attention weights to generate a weighted feature vector, and using the weighted feature vector to capture the spatio-temporal propagation relationship of the risk between geographical nodes.

[0012] Optionally, calculating the attention weights of each modal feature according to the context query vector includes: inputting the context query vector into a multi-layer perception machine to calculate original weight scores of each modal feature; applying a softmax function to the original weight scores for normalization processing to generate the attention weights.

[0013] Optionally, the spatio-temporal propagation relationship of the risk between geographical nodes is captured using a message passing mechanism of a graph, and a mutation risk value of each node is output, including: for each geographical node in the graph, aggregating feature information of neighbor nodes to form an aggregated information vector; combining the feature information of the geographical node itself with the aggregated information vector to update the feature representation of the geographical node; iteratively performing the aggregation and updating steps to simulate multi-step propagation of risk information in the graph structure.

[0014] Optionally, a mutation risk field map of the construction environment is generated according to the mutation risk value, and a corresponding risk warning is triggered according to the risk level, including: comparing the mutation risk value of each node with a multi-level risk threshold system to determine the risk level of each node; generating a structured warning signal according to the risk level; and sending the warning signal to a construction management system to trigger a response action.

[0015] Based on the same inventive concept, the application also provides a construction environment mutation risk field prediction system for fused geology and meteorological data, comprising: a data acquisition module for acquiring geology data, meteorological data and construction progress data of a construction area, and structurally processing the data; a graph construction module for dividing the construction area into a plurality of geographic nodes based on the structurally processed data, and constructing edges between the geographic nodes according to geographic spatial relationship and geology correlation to generate graph data containing multi-modal features; a feature fusion module for fusion encoding of features of different modes in the graph data to generate a unified feature vector; a prediction module for inputting the unified feature vector into a spatio-temporal adaptive graph neural network prediction model, dynamically calculating and distributing weights of different modal features according to real-time construction progress data and meteorological data through an attention mechanism inside the spatio-temporal adaptive graph neural network prediction model, and capturing spatio-temporal propagation relationship of risks between geographic nodes by using a message passing mechanism of the graph to output mutation risk values of each node; and an output module for generating a mutation risk field graph of the construction environment according to the mutation risk values, and triggering corresponding risk warnings according to risk levels.

[0016] Compared with the prior art, the application has the following advantages: 1. The application realizes deep fusion of multi-source heterogeneous data such as geology, meteorology and construction by constructing a unified spatio-temporal adaptive graph neural network model. This method discards the simple feature splicing or multi-model concatenation in traditional technologies, and can explore the internal coupling relationship between different modal information from the data bottom layer, so as to more comprehensively understand the complex causes of sudden risks and improve the accuracy of risk identification.

[0017] 2. The attention mechanism introduced in the application gives the prediction model dynamic adaptive ability. The model can autonomously adjust the attention degree to different data sources according to the real-time changes of the construction stage and environmental conditions, and focus the calculation on the most critical risk driving factors at present. This intelligent analysis mode makes the risk prediction more timely and targeted, especially enhancing the capture ability of sudden and instantaneous risks.

[0018] 3. The application utilizes the natural topological structure advantage of graph neural network to realize effective simulation of the spatial propagation path of risks. Not only can the risk level of a single position be predicted, but also the conduction effect and influence range of risks in the construction site can be revealed, which improves the risk management from discrete single-point monitoring to continuous field situation awareness, provides a scientific basis for global disaster prevention and mitigation decision-making, and enhances the practicality and forward-looking of the prediction results.

[0019] Other features and advantages of the present application will be set forth in the description that follows, and in part will be apparent from the description, or can be learned by practice of the application. The purposes and other advantages of the present application will be realized and attained by the structures described in the description, claims, and accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings described below are some embodiments of the present application, and the person skilled in the art can obtain other drawings according to these drawings without any creative effort.

[0021] Figure 1 is a flowchart of a construction environment mutation risk field prediction method of fusing geological and meteorological data according to an embodiment of the present application.

[0022] Figure 2 is a trend graph of different node risk values changing with time according to an embodiment of the present application.

[0023] Figure 3 is a thermal map of different node mutation risk values according to an embodiment of the present application.

[0024] Figure 4 is a structural schematic diagram of a construction environment mutation risk field prediction system of fusing geological and meteorological data according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by the person skilled in the art without any creative effort are within the protection scope of the present application.

[0026] REFERENCE Figure 1One embodiment of the present application proposes a construction environment mutation risk field prediction method that fuses geological and meteorological data. The method uses a spatio-temporal adaptive graph neural network model that can dynamically fuse multi-source heterogeneous data and simulate the spatial propagation of risks. This method discards the simple feature splicing or multi-model concatenation method in traditional technology, and instead builds a unified graph neural network model to explore the internal coupling relationship between different modal information from the data bottom layer, thereby comprehensively understanding the complex causes of sudden risks. The attention mechanism introduced in the present application gives the prediction model dynamic adaptive ability, enabling it to adjust the attention degree to different data sources according to the real-time changes in the construction stage and environmental conditions. This intelligent analysis mode makes the risk prediction more timely and targeted, especially enhancing the ability to capture sudden and transient risks. By taking advantage of the topological structure of graph neural networks, the present application can not only predict the risk level of a single location, but also reveal the conduction effect and influence range of risks within the construction site, improving risk management from discrete single-point monitoring to continuous field situational awareness.

[0027] The method of the embodiment specifically includes: Obtaining geological data, meteorological data and construction progress data of a construction area, and structuring these data; Based on the structured data, dividing the construction area into a plurality of geographic nodes, and constructing edges between the geographic nodes according to the geographic spatial relationship and geological correlation to generate graph data containing multi-modal features; Fusing and encoding the features of different modalities in the graph data to generate a unified feature vector; Inputting the unified feature vector into a spatio-temporal adaptive graph neural network prediction model, and dynamically calculating and assigning the weights of different modal features according to the real-time construction progress data and meteorological data through the attention mechanism inside the spatio-temporal adaptive graph neural network prediction model; Using the message passing mechanism of the graph to capture the spatio-temporal propagation relationship of risks between geographic nodes, and outputting the mutation risk values of each node; Generating a mutation risk field map of the construction environment according to the mutation risk values, and triggering corresponding risk warnings according to the risk levels.

[0028] Optionally, the obtaining of the geological data, meteorological data and construction progress data of the construction area includes: Obtaining geological data through geological sensors deployed at the construction site; Specifically, geological data is acquired from geological sensors. This does not refer to static geological exploration reports, but rather to a series of automated monitoring sensors pre-deployed at key locations in the construction area, such as slopes, foundation pits, and tunnel faces. These geological sensors include, but are not limited to, inclinometers for monitoring slope displacement, pore water pressure gauges for measuring internal pore water pressure of soil, and strain gauges for monitoring structural strain. These sensors periodically and automatically upload measurement data through Internet of Things technology, forming a dynamic geological data stream reflecting the real-time mechanical response of geological bodies under external disturbance.

[0029] Meteorological data is acquired through automatic weather stations; Specifically, real-time meteorological data is acquired from weather stations. To ensure the timeliness and relevance of the data, a set of automatic weather stations is usually deployed at the construction site or adjacent areas. The weather station continuously monitors and records key meteorological parameters, including rainfall, rainfall intensity, wind speed, wind direction, temperature, and humidity. The data collection frequency is very high, which can capture sudden weather events such as short-term heavy rain, and transmit in real time through network interface.

[0030] Construction progress data is acquired by accessing a construction plan database.

[0031] Specifically, construction progress data is acquired from a construction plan database. The data source is usually the project's digital management platform, such as a system integrated with Building Information Modeling (BIM) or a professional project management software. Through an interface, the database is queried regularly to extract information about ongoing construction activities, such as the location and depth of excavation work, the time and amount of blasting work, the distribution and movement trajectory of large machinery, etc. These data represent human dynamic factors that affect the environment. As shown in FIG. 8, the trend of the mutation risk value of different geographical nodes in a specific time period is shown. The line chart clearly depicts the dynamic process of the risk value accumulating over time, especially the rapid upward trend of high-risk nodes, intuitively reflecting the real-time monitoring and prediction capabilities of the method. Figure 2

[0032] Optionally, the edges between the geographical nodes are constructed according to the geographical spatial relationship and geological correlation, and the method further comprises: Adjacent geographical nodes are identified, and a first type of edge is established between adjacent geographical nodes; Specifically, the edges between adjacent nodes are defined based on geographical adjacency relationship. After the construction area is divided into grid-based geographical nodes, all node pairs are traversed. If two geographical nodes are directly adjacent in space, i.e., they share a boundary or vertex, or the Euclidean distance between them is less than a preset threshold, an edge is established between them. This edge represents a direct connection in physical space and is used to simulate the local diffusion effect of risk, such as surface water flow or small-scale soil slip.​

[0033] identify geographical nodes located in the same geological tectonic unit, and establish a second type of edge between geographical nodes located in the same geological tectonic unit; Specifically, the edge between nodes that are not adjacent but geologically continuous is defined based on geological correlation. This step goes beyond simple spatial proximity and aims to capture long-distance risk transmission mediated by underground geological structures. First, the input geological exploration data and geological maps are analyzed to identify macro-geological units that run through the construction area, such as the same fault zone, the same continuous weak interlayer, or the same underground aquifer. Then, it is determined which geographical nodes are located on the same key geological unit. If two or more geographical nodes, even if they are not geographically adjacent, are jointly located in the same geological unit identified above, an edge will be established between them. This edge represents an internal connection in terms of geomechanics or hydrology.

[0034] merge the first type of edge and the second type of edge to form a complete edge set of the graph data.

[0035] Specifically, all edges defined by the two methods are merged to form a complete edge set of the graph data. The graph structure constructed in this way not only contains geographical neighborhood relationships, but also establishes "shortcuts" through geological correlation edges, enabling the graph neural network to learn and simulate non-local risk propagation patterns. For example, when a settlement occurs at one end of a fault zone, the model can transmit the impact directly to the distant node at the other end of the fault zone through this geological correlation edge. This dual edge definition mechanism enables the model to capture both gradual diffusion and sudden long-distance transmission of risks, thereby improving the prediction ability of systemic and cascading risks under complex geological conditions and making risk field prediction more realistic.

[0036] Optionally, the features of different modalities in the graph data are fused and encoded to generate a unified feature vector, including: a convolutional neural network is used to process the geological data to extract spatial features and generate a geological feature vector; Specifically, for geological data, since it usually exists in the form of geological maps or exploration point distribution maps, it has strong spatial structure characteristics, and therefore the geological data is processed into two-dimensional or three-dimensional grid data. The grid data is processed using a convolutional neural network. The convolutional neural network can automatically learn and extract local spatial patterns and structural information in the geological data, such as the strike of the rock stratum fault zone, the location and range of the weak soil layer, etc., and finally output a geological feature vector that can represent the stability of the static geological environment in the construction area.

[0037] The meteorological data and the construction progress data are processed by using a recurrent neural network to extract time sequence characteristics and generate a time sequence characteristic vector. Specifically, for meteorological data and construction progress data, both are typical time series data, and the value lies in revealing dynamic change trends. Therefore, a recurrent neural network or its variants such as a long short-term memory network is used to encode such time series data. The recurrent neural network can effectively capture the dependence relationship of data in the time dimension through an internal loop structure, for example, the cumulative effect of continuous rainfall, the time sequence influence of a specific construction activity, and the like, thereby generating a time sequence characteristic vector that summarizes the environmental and construction dynamic change law.

[0038] The geological feature vector and the time sequence feature vector are cross-fused to generate a unified feature vector.

[0039] Specifically, after obtaining the geological feature vector representing static spatial information and the time sequence feature vector representing dynamic time sequence information, in order to realize deep fusion instead of simple splicing of the two kinds of information, a feature cross-fusion step is designed. The geological feature vector and the time sequence feature vector are spliced to form a combined vector. In order to enable the model to learn the nonlinear interaction relationship between the two modal features, the combined vector is input into one or more fully connected layers for transformation. The process is represented by the formula: wherein, is the unified feature vector finally generated; represents the geological feature vector; represents the time sequence feature vector; the function represents splicing the two vectors end to end; and represent the weight matrix and the bias vector of the fully connected layer, respectively, which are parameters automatically optimized by the model during the training process; is a nonlinear activation function, such as ReLU. This process maps features of different sources to a unified feature space through a learnable weight matrix , effectively avoiding the problem that data dimensions and physical meanings cannot be directly operated, and enabling the model to autonomously explore and quantify complex coupling relationships such as “what intensity of rainfall” acting on “what type of geological structure” will trigger high risks. The unified feature vector finally output provides an information-comprehensive and content-rich input for the subsequent graph neural network model, improving the accuracy of risk prediction and the sensitivity to sudden events.

[0040] Optionally, the cross-fusion of the geological feature vector and the time sequence feature vector comprises: splicing the geological feature vector and the time sequence feature vector to form a combined vector;​ Specifically, the geological feature vector extracted by the convolutional neural network and the time-series feature vector extracted by the recurrent neural network are concatenated into a higher-dimensional combined vector through a concatenation operation. This concatenation operation physically collocates features representing static geological environment and features representing dynamic environment and construction activities in the same mathematical space, laying the foundation for subsequent interactive learning.

[0041] The combined vector is linearly transformed, and a nonlinear activation function is applied to the result of the linear transformation to generate a unified feature vector.

[0042] Specifically, this concatenated combined vector is input into a pre-set fully connected layer. The fully connected layer is a basic structure in neural networks, and the core is a learnable weight matrix. In this layer, each element in the combined vector will perform matrix multiplication with the weight matrix, which means that each dimension in the geological feature vector will interact with each dimension in the time-series feature vector. During the training process, the model will automatically adjust the parameters of the weight matrix through the backpropagation algorithm, thereby learning the meaningful internal correlation patterns between the two modal features. After linear transformation by the fully connected layer, the output result will be immediately processed by a nonlinear activation function. Nonlinear activation functions, such as the rectified linear unit (ReLU), are key to deep learning models being able to learn complex nonlinear relationships. In the context of construction risk prediction, many risk trigger mechanisms are highly nonlinear, for example, the increase in rainfall is not simply linearly proportional to landslide risk, but there may be a critical threshold. By introducing a nonlinear activation function, the model can effectively fit this complex nonlinear dependence relationship, thereby more accurately depicting the causes of risk.

[0043] Optionally, the unified feature vector is input into a spatio-temporal adaptive graph neural network prediction model, and the weights of different modal features are dynamically calculated and assigned according to real-time construction progress data and meteorological data through the attention mechanism inside the spatio-temporal adaptive graph neural network prediction model, including: Encoding the construction progress data and meteorological data into a context query vector; Specifically, the attention weights of different modal features need to be calculated. The model encodes the construction phase data as the current spatio-temporal context and the real-time environmental data into a context query vector. The construction phase data, such as foundation excavation or main body construction, can be encoded as a category vector; the real-time environmental data, such as instantaneous wind speed or cumulative rainfall, can be encoded as a numerical vector. The context query vector is then used to evaluate the importance of each modal component in the unified feature vector.

[0044] According to the context query vector, the attention weights of each modal feature are calculated; Specifically, to achieve this evaluation, the model computes a relevance score for each modality feature to the current context via a small feedforward neural network. The relevance score reflects the potential contribution of a certain data modality, such as geology or weather, to the risk under a specific construction and environmental background. The scores of all modalities are normalized by a softmax function to generate a set of attention weights that sum up to one.

[0045] The corresponding modality features in the unified feature vector are weighted using the attention weights to generate a weighted feature vector, and the weighted feature vector is used to capture the spatiotemporal propagation relationship of the risk between geographical nodes.

[0046] Specifically, after obtaining the attention weights, the model applies these weights to the unified feature vector. Each modality feature component corresponding to each geographical node in the unified feature vector will be multiplied by its corresponding attention weight. This weighting operation dynamically adjusts the contribution of each modality feature in subsequent calculations. For example, under continuous heavy rain weather, the model will adaptively assign higher attention weights to the features of weather data and geological data, thereby amplifying their impact on the final risk prediction. The feature vector adjusted by the attention weight will be used as the input of the message passing mechanism of the graph neural network. In the message passing process, each geographical node will aggregate the weighted feature information passed by its neighbor nodes and update it in combination with its own information. Since the transmitted information has been weighted by attention, it means that the risk propagation simulation will focus more on the most critical risk factors. For example, an upstream node with a high-weighted unstable geological feature due to heavy rain will be more effectively transmitted to downstream nodes, thereby accurately simulating the propagation path of the mudslide risk.

[0047] Optionally, calculating the attention weight of each modality feature according to the context query vector comprises: inputting the context query vector into a multi-layer perceptron to obtain the original weight score of each modality feature; Specifically, the digital description of the current situation involves encoding the current construction phase information and real-time environmental data separately. Construction phase encoding transforms textual descriptions such as "foundation excavation" and "main structure construction" into numerical vectors that the model can understand. Environmental data encoding integrates real-time monitored parameters such as rainfall and wind speed into numerical vectors. These two encoded vectors are then concatenated to form a unified context query vector, which comprehensively describes the "here and now" working conditions and environment. This context query vector is input into a pre-defined multilayer perceptron. A multilayer perceptron is a feedforward neural network that, through internal multilayer nonlinear transformations, can learn and fit extremely complex relationships between inputs and outputs. In this invention, it is trained to learn expert knowledge about "which data modality is more important in which context." The output of the multilayer perceptron is a set of raw, unnormalized weight scores, each corresponding to a data modality such as geology or meteorology, with the numerical value reflecting the initial importance of that modality in the current context.

[0048] The original weight scores are normalized using the softmax function to generate attention weights.

[0049] Specifically, to transform these raw weight scores into a set of normalized values ​​that can be directly used as weights, the system applies the softmax function. The softmax function maps an arbitrary set of real numbers to a probability distribution whose sum is 1. The calculation process is as follows: ; in, It is the first The final attention weights obtained from each modality feature; It is the correlation score calculated between the modality feature and the current context; This is the exponential sum of all modal correlation scores, used for normalization. Here... It is calculated by a small neural network that takes the current context query vector and the corresponding modality features as input, and its parameters are learned during model training. Through this function, the original weight scores are converted into relative importance, with higher scores receiving higher attention weights, and the sum of all weights is exactly 1.

[0050] Optionally, the step of using the graph messaging mechanism to capture the spatiotemporal propagation relationship of risk among geographical nodes and outputting the mutation risk value of each node includes: For each geographical node in the graph, the feature information of the neighboring nodes is aggregated to form an aggregated information vector; Specifically, the message passing mechanism in the present application is the core engine of the graph neural network simulating the spread of risks in space, and the execution process is divided into two closely connected steps of aggregation and update. These two steps are iteratively executed for multiple rounds, and each round of iteration means that the risk information is transmitted one step outwards on the graph structure. In the aggregation step, each geographical node in the graph actively collects feature information from all direct neighbor nodes. Here, "neighbors" are jointly defined by the aforementioned geographical adjacency relationship and geological relevance. Each node receives not the original features, but the feature vectors that have been weighted by the attention mechanism, which means that the transmitted information itself already contains the importance evaluation in the current spatio-temporal context. The aggregation operation is usually completed by a function that is invariant under permutation, such as summing, averaging, or taking the maximum value of the feature vectors of all neighbor nodes, thereby integrating discrete information from multiple neighbors into a single aggregated information vector that summarizes the state of the neighborhood. This process is represented as: ; where, is the aggregated information vector generated for the target node ; represents the aggregation function; is the neighbor node set of node ; is the attention-weighted feature vector of neighbor node .

[0051] The feature representation of the geographical node is updated by combining its own feature information with the aggregated information vector; Specifically, in the update step, each node will use the aggregated information vector just generated to update its own feature representation. This process is to fuse the original feature information of the node itself with the aggregated information vector from the neighborhood. This fusion is usually implemented through a small neural network layer, which concatenates the original feature vector of the node itself and the aggregated information vector and then performs a nonlinear transformation. Its formula is represented as: ; where, is the updated new feature vector of node ; is the feature vector of node before updating; is a learnable update function, such as a fully connected layer plus a nonlinear activation function. As shown in Figure 3 , the mutation risk values of different geographical nodes at different times are shown. The color depth represents the size of the risk value, allowing managers to identify areas of risk concentration at a glance and grasp the distribution situation of risks in space.

[0052] The aggregation and updating steps are iteratively performed to simulate multi-step propagation of risk information in the graph structure.

[0053] Specifically, through one aggregation and updating, the new feature vector of each node fuses the information of its first-order neighborhood. After multiple iterations, the feature vector of a node will contain the information of multi-order neighborhood, that is, the information realizes long-distance propagation on the graph. The technical effect of the message passing mechanism changes the static and discrete node risk assessment into dynamic and continuous risk field evolution simulation. The model can explicitly capture the conduction effect of risk, for example, how the rainfall saturation state of the upstream node gradually affects the stability of the downstream node through the edges of the graph, so as to realize the global and forward-looking prediction of the risk situation of the entire construction area.

[0054] Optionally, a mutation risk field map of the construction environment is generated according to the mutation risk value, and a corresponding risk warning is triggered according to the risk level, including: The mutation risk value of each node is compared with a multi-level risk threshold system to determine the risk level of each node. Specifically, the step of triggering the risk warning in the present application is to convert the quantitative risk prediction output by the model into actual operational safety management instructions. After the spatiotemporal adaptive graph neural network prediction model completes the calculation, it will output a real-time and continuous mutation risk value for each geographical node in the construction area. This value is a quantitative assessment of the possibility of sudden risk occurring in the future period of time for the node. The core is to compare the mutation risk value with the preset risk threshold. The preset risk threshold is a grading threshold system developed according to safety management regulations, engineering geological conditions and historical disaster data, for example, corresponding to different safety levels such as attention, warning and alarm. The mutation risk value of each node is compared with this threshold system in real time.

[0055] A structured warning signal is generated according to the risk level. Specifically, when the mutation risk value of a certain node exceeds the preset risk threshold of a certain level, the corresponding warning logic will be triggered immediately to automatically generate a structured warning signal. This warning signal is a data packet containing key information such as the geographical node number triggering the warning, the current mutation risk value, the warning level triggered, the timestamp and the main risk contributing factor identified by the model attention mechanism, for example, continuous rainfall or specific construction activities.

[0056] The warning signal is sent to the construction management system to trigger a response action.

[0057] Specifically, the generated early warning signal is sent to the construction management system in real time through a standard application program interface or message queue. After receiving the signal, the construction management system executes the preset response program, such as highlighting the risk area on the electronic sand table or GIS interface, updating the mutation risk field map, and pushing specific and explicit early warning information to project managers, safety officers, and other relevant persons in charge through mobile applications, SMS, or on-site sound and light alarms.

[0058] Based on the same inventive concept, as Figure 4 shown, the present application also provides a construction environment mutation risk field prediction system that fuses geological and meteorological data. The system includes: a data acquisition module for obtaining geological data, meteorological data, and construction progress data in the construction area and structurally processing these data; a graph construction module for dividing the construction area into a plurality of geographic nodes based on the structurally processed data and constructing edges between the geographic nodes according to geographic spatial relationships and geological correlations to generate graph data containing multi-modal features; a feature fusion module for fusion encoding of features of different modalities in the graph data to generate a unified feature vector; a prediction module for inputting the unified feature vector into a spatio-temporal adaptive graph neural network prediction model, dynamically calculating and assigning weights of different modal features according to real-time construction progress data and meteorological data through an attention mechanism inside the spatio-temporal adaptive graph neural network prediction model, and capturing spatio-temporal propagation relationships of risks between geographic nodes using a message passing mechanism of a graph to output mutation risk values of each node; an output module for generating a mutation risk field map of the construction environment according to the mutation risk values and triggering corresponding risk early warnings according to risk levels.

[0059] To verify the feasibility of the present application in implementation, the present application is applied to a high-slope excavation engineering project of a mountainous highway. The construction area of the project has complex geological conditions, including multiple fault zones and weak interlayers, and is located in a rainy area, where sudden geological disasters such as landslides and collapses are easily triggered by the coupling effect of heavy rainfall and excavation activities during construction.

[0060] Traditional safety management methods rely on regular manual inspection and static geological survey reports, making it difficult to predict sudden risks in real time and dynamically. The project hopes to use the method of the present application to fuse multi-source heterogeneous data on site and achieve dynamic prediction and real-time early warning of the mutation risk field of the construction environment.

[0061] In this embodiment, the project party first laid out geological sensors such as inclinometers and pore water pressure gauges at key positions in the construction area according to the data acquisition steps of the present application to obtain dynamic geological data; an automatic weather station was set up on site to collect real-time meteorological data such as rainfall and wind speed; and the BIM management platform of the project was connected to obtain construction progress data such as excavation and blasting. Subsequently, the construction area was divided into multiple 10m x 10m geographic nodes, and a graph structure containing two types of edges was constructed based on the geographic adjacency relationship and the strike of the proven geological fault zone.

[0062] To verify the beneficial effects of the present application, a typical working condition during the rainy season was selected for testing and analysis. During this period, a prediction system using the method of the present application was run as the experimental group, and a monitoring system based on the traditional threshold alarm was run as the control group.

[0063] In the feature fusion coding stage, the convolutional neural network is used to process the rasterized geological exploration map to automatically extract spatial features such as fault zone location and rock layer inclination, and generate a geological feature vector; at the same time, the long short-term memory network is used to process 48 hours of continuous rainfall data and construction excavation progress data to capture cumulative effects and time sequence influences, and generate a time sequence feature vector. Subsequently, the feature cross fusion is performed by formula The two types of feature vectors are input into the fully connected layer and processed by the ReLU nonlinear activation function to generate a unified feature vector that can represent the complex coupling relationship between geology, meteorology and construction.

[0064] On a certain day, a new round of blasting and excavation operation was carried out on site, and the weather station monitored a short-term heavy rain. After receiving the above information, the attention mechanism in the spatio-temporal adaptive graph neural network prediction model of the present application was activated. The model encodes the construction phase information of "blasting and excavation" and the real-time rainfall data into a context query vector. According to the vector, the model calculates that the attention weights of meteorological data and geological data should be significantly increased under the current situation. For example, the weight of meteorological features is increased from 0.2 on weekdays to 0.5, while the weight of geological features related to slope stability is increased from 0.4 to 0.45. Subsequently, the feature vectors weighted by attention are subjected to message passing in the graph neural network. The high-risk features of the nodes located upstream of the blasting point and with a sharp rise in pore water pressure are effectively transmitted to the downstream adjacent nodes through the aggregation and update steps of the graph. At the same time, since a node numbered A-15 near the blasting point is located on a known geological fault zone, the model transmits the risk information generated by the blasting disturbance of this node to another node numbered C-08 which is also on the same fault zone but geographically far away through the pre-set geological correlation edges.

[0065] Finally, the model output shows that the mutation risk values of the downstream regions numbered B-15 to B-20 and the C-08 node at the far end of A-15 increase sharply, exceeding the preset "alarm" level threshold. A structured early warning signal is immediately generated, the above-mentioned regions are marked red on the on-site GIS electronic sand table through the construction management system, and the project safety director's mobile phone is pushed with the early warning information: "There is a high landslide risk in the B-15 to B-20 and C-08 regions, and the main risk contributing factors are heavy rain and blasting disturbance. Please evacuate personnel and equipment immediately." The control group only issued a general warning of "orange rain warning in the whole field area", and failed to identify the specific risk concentration area. About 3 hours after the event, a small-scale collapse did occur in the B-18 region. Thanks to the timely warning, the on-site personnel and equipment had been evacuated in advance, avoiding losses.

[0066] Table 1 Comparison of key node data monitoring and risk prediction on a certain day

[0067] The data in Table 1 shows that under the same environmental and construction background, the present application can distinguish the risk levels of different nodes, accurately identify the high-risk areas caused by the coupling of multiple factors, and predict the risk values to be higher than those of other areas. The control group can only provide a non-discriminatory regional warning.

[0068] The comparison results in Table 2 further prove the technical effects of the present application. The present application not only realizes "field perception" of risks, reveals the transmission path and influence range of risks, but also provides highly targeted and operable early warning information, provides forward-looking decision support for on-site managers, realizes the transition from passive response to active avoidance, and improves the intrinsic safety level of the construction site.

[0069] Table 2 Comparison of early warning system performance

[0070] It should be noted that the electrical connections between the various units described above do not necessarily represent direct connections, and indirect connection methods can also be used as long as the purpose of the present application is achieved. The above descriptions are only exemplary embodiments of the present application, and cannot limit the scope of the present application.

[0071] That is, any equivalent changes and modifications made in accordance with the teachings of the present application are still within the scope of the present application. Other embodiments of the present application will be readily apparent to those skilled in the art upon considering the specification and practice of the disclosure. The present application is intended to cover any variations, uses, or adaptations of the present application that follow the general principles of the present application and include known or customary technical means in the art not described herein.

Claims

1. A method of predicting a construction environment catastrophe risk field by fusing geophysical and meteorological data, characterized by, The method comprises: obtaining geological data, meteorological data and construction progress data of a construction area, and structuring the data; based on the structured data, dividing the construction area into a plurality of geographic nodes, and constructing edges between the geographic nodes according to geographical spatial relationship and geological correlation to generate graph data containing multi-modal features; fusing and encoding the features of different modalities in the graph data to generate a unified feature vector; inputting the unified feature vector into a spatio-temporal adaptive graph neural network prediction model, dynamically calculating and assigning the weights of different modal features according to real-time construction progress data and meteorological data through the attention mechanism inside the spatio-temporal adaptive graph neural network prediction model; using the message passing mechanism of the graph to capture the spatio-temporal propagation relationship of risks between geographic nodes, and outputting the mutation risk value of each node; generating a mutation risk field map of the construction environment according to the mutation risk value, and triggering corresponding risk warnings according to the risk level.

2. The method of claim 1, wherein the method further comprises: The method comprises: obtaining geological data, meteorological data and construction progress data of a construction area, and structuring the data; obtaining geological data through geological sensors deployed on the construction site; obtaining meteorological data through automatic weather stations; 3. The method of claim 1, wherein the method further comprises: obtaining construction progress data by accessing a construction plan database. The method comprises: identifying adjacent geographic nodes in geographical position, and establishing a first type of edge between adjacent geographic nodes; identifying geographic nodes located in the same geological structure unit, and establishing a second type of edge between geographic nodes located in the same geological structure unit; 4. The method of claim 1, wherein the method further comprises: merging the first type of edge and the second type of edge to form a complete edge set of the graph data. The method comprises: using a convolutional neural network to process the geological data, extract spatial features, and generate a geological feature vector; using a recurrent neural network to process the meteorological data and the construction progress data, extract time series features, and generate a time series feature vector; 5. The method of claim 4, wherein the method further comprises: cross-fusing the geological feature vector and the time series feature vector to generate a unified feature vector. The method comprises: splicing the geological feature vector and the time series feature vector to form a combined vector; 6. The method of claim 1, wherein, performing linear transformation on the combined vector, and applying a nonlinear activation function to the result of linear transformation to generate a unified feature vector. The method comprises: encoding the construction progress data and the meteorological data into a context query vector; calculating the attention weights of each modal feature according to the context query vector; 7. The method of claim 6, wherein the method further comprises: using the attention weights to weight the corresponding modal features in the unified feature vector to generate a weighted feature vector, and using the weighted feature vector to capture the spatio-temporal propagation relationship of risks between geographic nodes. The method comprises: calculating the attention weights of each modal feature according to the context query vector. Input the context query vector into a multi-layer perception machine to calculate original weight scores of each modality feature; Apply a softmax function to the original weight scores for normalization to generate attention weights.

8. The method of claim 1, wherein the method further comprises: The message passing mechanism of the graph captures the spatiotemporal propagation relationship of risks between geographical nodes, and outputs mutation risk values of each node, including: For each geographical node in the graph, aggregate the feature information of neighbor nodes to form an aggregated information vector; Update the feature representation of the geographical node by combining its own feature information and the aggregated information vector; Iteratively perform the aggregation and update steps to simulate the multi-step propagation of risk information in the graph structure.

9. The method of claim 1, wherein the method further comprises: Generate a mutation risk field map of the construction environment according to the mutation risk values, and trigger corresponding risk warnings according to the risk level, including: Compare the mutation risk values of each node with a multi-level risk threshold system to determine the risk level of each node; Generate a structured warning signal according to the risk level; Send the warning signal to the construction management system to trigger a response action.

10. A construction environment catastrophe risk field forecasting system that fuses geophysical and meteorological data, characterized by, The system is used for the construction environment mutation risk field prediction method of fusing geological and meteorological data according to any one of claims 1-9, and the system comprises: A data acquisition module for obtaining geological data, meteorological data and construction progress data in the construction area, and structuring these data; A graph construction module for dividing the construction area into a plurality of geographical nodes based on the structured data, and constructing edges between geographical nodes according to geographical spatial relationships and geological correlations to generate graph data containing multi-modal features; A feature fusion module for fusing and encoding features of different modalities in the graph data to generate a unified feature vector; A prediction module for inputting the unified feature vector into a spatiotemporal adaptive graph neural network prediction model, dynamically calculating and distributing the weights of different modal features according to real-time construction progress data and meteorological data through the attention mechanism inside the spatiotemporal adaptive graph neural network prediction model, and capturing the spatiotemporal propagation relationship of risks between geographical nodes using the message passing mechanism of the graph to output mutation risk values of each node; An output module for generating a mutation risk field map of the construction environment according to the mutation risk values, and triggering corresponding risk warnings according to the risk level.

Images