A flood intelligent deduction method and system based on global integration

Abstract

The present invention relates to the field of smart city environmental perception technology, and in particular to a flood intelligent deduction method and system based on global fusion. The method includes the following steps: deploying ground-nailed intelligent node units in a preset area, and collecting regional waterlogging data through the ground-nailed intelligent node units; constructing a geographic adjacency graph for the regional waterlogging data to obtain proximity graph data; performing local cross-domain graph attention calculation on the proximity graph data to obtain proximity perception graph data; performing cross-domain meta-correlation perception on the proximity perception graph data to obtain cross-domain meta-correlation data; performing waterlogging trend prediction on the cross-domain meta-correlation data to obtain waterlogging trend data; performing global flood simulation based on the waterlogging trend data to obtain global flood data. The present invention combines edge computing units to perform global flood simulation, generates spatially continuous and semantically clear risk layers, and provides efficient and explainable decision support for urban-level flood control scheduling.

Description

Technical Field

[0001] The present invention relates to the field of smart city environmental perception technology, and in particular to a flood intelligent deduction method and system based on global fusion. Background Art

[0002] Existing flood warning systems mostly rely on fixed water level gauges or manual observations. These systems have sparse sensing points, low update frequency, and insufficient spatial coverage, making it difficult to fully perceive the dynamics of accumulated water in complex terrain. Cloud platforms are located in remote data centers, and all data must be uploaded to the cloud for processing before decision-making results can be fed back. In sudden disasters such as floods, this delay can lead to missing the optimal time for intervention. Furthermore, traditional methods often rely on rule-based modeling or statistical regression models for trend prediction, which makes it difficult to capture the nonlinear evolution caused by sudden floods or periodic disturbances, resulting in insufficient prediction accuracy and timeliness. Summary of the Invention

[0003] In order to solve the above technical problems, the present invention proposes a flood intelligent deduction method and system based on global fusion to solve at least one of the above technical problems.

[0004] This application provides a flood intelligent deduction method based on global fusion, the method comprising:

[0005] S1. Deploy ground-nailed smart node units in a preset area and collect regional waterlogging data through the ground-nailed smart node units.

[0006] S2. Construct a geographic adjacency graph for the regional waterlogging data to obtain proximity graph data; perform local cross-domain graph attention calculation on the proximity graph data to obtain proximity perception graph data;

[0007] S3. Perform cross-domain meta-association perception on the neighboring perception graph data to obtain cross-domain meta-association data; perform water accumulation trend prediction on the cross-domain meta-association data to obtain water accumulation trend data;

[0008] S4. Conduct a full-area flood simulation based on the water accumulation trend data to obtain full-area flood data.

[0009] The present invention deploys ground-nailed intelligent node units that integrate millimeter-wave radars, pressure sensors, and neuromorphic chips in preset areas, which can achieve low-power, localized dynamic data collection and preliminary processing of waterlogging, reducing dependence on central servers. By constructing a geographic adjacency graph and performing local cross-domain graph attention calculations, the system not only captures the spatial topological relationship between nodes, but also integrates multi-domain features such as trend changes, modal heterogeneity, and historical evolution to generate a structured proximity perception graph, effectively improving data representation capabilities. Combining multi-perspective meta-feature aggregation and semantic partitioning operations, it is possible to model the potential diffusion mechanism of waterlogging between regions, generate a highly reliable cross-domain meta-correlation map, and then decouple and predict different waterlogging evolution patterns through linear, exponential, and periodic multi-branch trend inference networks, thereby improving the ability to respond to sudden flood trends.

[0010] Optionally, S1 includes:

[0011] Deploy ground-nailed intelligent node units in a preset first area, collect water accumulation height and force per unit area, and obtain water accumulation data for the first area;

[0012] Deploy ground-nailed intelligent node units in a preset second area, and collect water accumulation height and force per unit area to obtain water accumulation data for the second area, wherein the first area and the second area are different areas;

[0013] Performing event-driven encoding on the first area waterlogging data and the second area waterlogging data to obtain event-driven data;

[0014] The event-driven data is enhanced with temporal and spatial offset perception to obtain regional waterlogging data.

[0015] In the present invention, ground-nailed intelligent node units are deployed in the first area and the second area respectively, and combined with millimeter-wave radar and pressure sensors, the water accumulation height and unit area force data are collected respectively. This not only realizes high-precision perception of the water accumulation status in the area, but also can capture hidden water accumulation trends through pressure changes, such as undercurrents or accumulation of water in low-lying areas. By sparsely processing the original perception data of different areas through event-driven coding, it is possible to effectively filter out redundant static information and only retain key events such as mutations, water level jumps, and pressure surges, thereby reducing communication and computing loads and improving data response efficiency. On this basis, the event-driven data is further enhanced by performing spatiotemporal offset perception enhancement, which can not only identify the temporal dislocation of cross-regional waterlogging evolution, but also capture the potential correlation between local sudden waterlogging and neighboring area trends, thereby improving the structural consistency and trend expression capabilities of regional waterlogging data.

[0016] Optionally, the geographic adjacency graph construction includes:

[0017] Extract regional nodes from regional waterlogging data to obtain regional node data;

[0018] Perform terrain semantic classification on regional node data to obtain regional classification data;

[0019] The mobility weighted directional adjacent edges are constructed for the regional classification data to obtain adjacent edge data;

[0020] Perform local structural similarity screening on the adjacent edge data to obtain adjacent edge screening data;

[0021] The graph is constructed based on the regional node data and adjacent edge screening data to obtain the proximity graph data.

[0022] In the present invention, regional nodes are extracted from regional waterlogging data, and points with representative change characteristics or terrain-sensitive points are preferentially selected as graph nodes, thereby avoiding redundant calculations and structural noise caused by blind global mapping. Subsequently, through terrain semantic classification, the nodes are divided into catchment areas, transition areas, and highland relief areas according to factors such as terrain height, slope changes, and drainage patency, thereby realizing the structural prior injection of potential waterlogging paths. The introduction of liquidity weights and directional judgment mechanisms when constructing adjacent edges can simulate the natural trend of actual water bodies flowing along the slope direction, making the graph structure physically reasonable. Local structural similarity calculations are introduced in the adjacent edge screening link to effectively eliminate invalid edges that do not have semantic consistency or terrain coupling.

[0023] Optionally, the local cross-domain graph attention calculation includes:

[0024] Extract node multi-domain features from the proximity graph data to obtain node multi-domain feature data;

[0025] Perform cross-domain weighted graph attention allocation on the node multi-domain feature data to obtain cross-domain weighted graph attention data;

[0026] The neighboring graph data is momentum suppressed and fused according to the cross-domain weighted graph attention data to obtain the neighboring perception graph data.

[0027] In the present invention, multi-domain features of each node in the neighborhood graph are extracted, covering multiple dimensions such as spatial position, water accumulation trend, perceptual modality and historical response, to construct a structured and multi-semantic node representation, breaking through the limitations of traditional graph neural networks that are only based on topology or single attribute modeling. Subsequently, a cross-domain weighted graph attention mechanism is introduced, which can dynamically allocate the propagation weights of adjacent edges according to the correlation strength between different feature domains, so that the information aggregation process not only considers the connection relationship, but also reflects the fusion characteristics of trend orientation and perceptual semantics. In order to prevent some high-weight nodes from experiencing information overload or representation offset during the process of graph information propagation, a momentum suppression fusion mechanism is introduced to regulate the information absorption intensity through the feature change rate and local gradient response, thereby improving the stability and discrimination of the overall node embedding.

[0028] Optionally, the cross-domain meta-association perception includes:

[0029] Perform multi-dimensional feature element pairing between nodes on the proximity perception graph data to obtain node feature element pair data;

[0030] Perform multi-view meta-feature aggregation on the node feature meta-pair data to obtain node feature aggregation data;

[0031] Generate a meta-association graph based on node feature aggregation data to obtain meta-association graph data;

[0032] Perform semantic partition projection on the meta-association graph data to obtain cross-domain meta-association data.

[0033] In the present invention, multi-dimensional feature meta-pairs are constructed for any pair of nodes in the neighboring perception graph, and the differences in multi-source features such as spatial relationships, water accumulation trends, perceptual modalities and historical evolution are made explicit, and converted into feature pair expressions with structural, behavioral and semantic tension, thereby improving the resolution ability of high-order relationship modeling between nodes. Through the multi-perspective meta-feature aggregation mechanism, association clues are extracted from multiple perspectives such as spatial dependency, trend collaboration, and modal complementarity, and an attention fusion strategy is introduced to dynamically weight the perspective contribution, thereby enhancing the adaptability and expression accuracy of the feature aggregation results to cross-domain semantic relationships. The constructed meta-association graph not only reflects the explicit connections between nodes, but also depicts potential trend isomorphism areas and implicit coupling structures, making the graph more explanatory and generalizable.

[0034] Optionally, the multi-view meta-feature aggregation includes:

[0035] Perform perspective division on the node feature element pair data to obtain perspective division data;

[0036] Perform multi-view sub-network aggregation on the view division data to obtain multi-visual aggregated data;

[0037] Semantic attention fusion is performed on multi-view aggregated data to obtain node feature aggregated data.

[0038] In the present invention, the data is divided into perspectives by node feature pairs, and the original feature pairs are split into multiple sub-perspective channels according to semantic dimensions such as spatial structure differences, trend evolution consistency, and modal response complementarity, thereby avoiding the semantic confusion problem caused by forcibly fusing heterogeneous features in a unified space. A structurally differentiated sub-network is introduced for each perspective channel for parallel aggregation. Each sub-network can adopt different modeling strategies (such as local convolution, attention mechanism, or periodic coding) according to the perspective characteristics, effectively retaining the specificity and significance of various relationships. A semantic attention fusion mechanism is introduced based on the aggregation results of all perspectives to adaptively learn the contribution of each perspective in the current context, highlight the key perspective, and suppress noise information, thereby forming a node feature aggregation expression with clear semantics and a clear structure.

[0039] Optionally, the water accumulation trend prediction includes:

[0040] Constructing a cross-domain meta-graph for the cross-domain meta-association data to obtain cross-domain meta-graph data;

[0041] Perform multi-branch trend inference on cross-domain meta-graph data to obtain multi-branch trend data;

[0042] The multi-branch trend data are subjected to time series inversion verification to obtain the waterlogging trend data.

[0043] In the present invention, a cross-domain meta-graph is constructed based on cross-domain meta-association data. By introducing high-order structural coupling relationships, potential semantic information such as trend collaboration, response time difference and modal complementarity between nodes is converted into graph structure connections, which significantly enhances the structural support and propagation path expression of trend modeling. Subsequently, a multi-branch trend inference mechanism is introduced to model different types of trend evolution patterns (such as linear water accumulation growth, exponential surge, and periodic fluctuation) as structurally heterogeneous prediction sub-networks, and dynamically select the prediction path that best suits the current node trend through trend feature encoding, taking into account local regularity and global diversity, and improving the model's fitting and generalization capabilities for complex trend changes. The prediction results are reversely reconstructed and verified through a time series inversion verification mechanism, and the stability correction and jump error suppression of the prediction results are achieved in combination with historical trend trajectories, effectively improving the robustness and time series consistency of the water accumulation trend output.

[0044] Optionally, the multi-branch trend reasoning includes:

[0045] Extract trend features from cross-domain meta-graph data to obtain trend feature data;

[0046] Perform linear regression calculation on trend characteristic data to obtain linear water accumulation data;

[0047] Perform exponential curve fitting on trend characteristic data to obtain exponential surge data;

[0048] The trend feature data is processed by frequency domain convolution network to obtain periodic fluctuation data;

[0049] The credibility trend of linear water accumulation data, exponential surge data and periodic fluctuation data is fused to obtain multi-branch trend data.

[0050] In the present invention, trend features are extracted from cross-domain meta-graph data, and the temporal response, modal difference and spatial behavior features of the nodes are encoded into a unified trend expression to ensure that the input has cross-dimensional consistency and time sensitivity. A linear regression model, an exponential curve fitting model and a frequency domain convolutional network are constructed respectively to identify and model three types of typical water accumulation evolution patterns: linear water accumulation reflects a conventional steadily rising trend, exponential surge targets the rapid water level changes caused by short-term rainstorms, and frequency domain processing can identify fluctuation patterns caused by drainage regulation or periodic rainfall. The branch models are structurally heterogeneous and functionally complementary, avoiding the problem that a single model is difficult to be compatible with complex trend forms. Through the credibility fusion mechanism, weights are dynamically allocated according to factors such as the fitting residuals of each branch in historical data, its own prediction stability and the consistency of the surrounding structure, thus realizing an adaptive combination of trend results.

[0051] Optionally, the global flood simulation includes:

[0052] Perform regional state trend processing based on the water accumulation trend data to obtain regional state trend data;

[0053] Perform graph diffusion evolution deduction on regional state trend data to obtain graph diffusion evolution data;

[0054] The graph diffusion evolution data is partitioned into risk layers to obtain the global flood data.

[0055] In the present invention, regional state trend processing is performed based on waterlogging trend data, and the waterlogging growth rate, trend duration and spatial impact range of each node are uniformly encoded into regional state trend data, so that each node has behavior-driven attributes, which effectively enhances the dynamic response capability during the simulation process. Through graph diffusion evolution deduction, the trend state is injected into the graph structure as a source item, and the mobility weight, directional factor and terrain constraint conditions of the adjacent edges are combined to simulate the propagation path and intensity changes of waterlogging between different regions, and construct a graph diffusion dynamic system that approximates the physical mechanism. This diffusion process not only captures the spatiotemporal continuity of trend propagation, but also depicts the heterogeneous conduction behavior between regions, such as the multi-stage evolution characteristics of "waterlogging source-transition zone-waterlogging blind area". By partitioning the graph diffusion evolution results into risk layers, a spatially visualized flood situation map is formed, and key areas such as high-risk waterlogging areas, fluctuation areas and controllable areas are clearly identified, realizing a closed-loop linkage from trend prediction to spatial response.

[0056] Optionally, the present application further provides a flood intelligent deduction system based on global fusion, which is used to execute the flood intelligent deduction method based on global fusion as described above. The flood intelligent deduction system based on global fusion includes:

[0057] The node deployment and collection module is used to deploy ground-nailed intelligent node units in a preset area and collect regional water accumulation data through the ground-nailed intelligent node units;

[0058] The graph neural perception fusion module is used to construct a geographic adjacency graph for regional waterlogging data to obtain proximity graph data; it performs local cross-domain graph attention calculation on the proximity graph data to obtain proximity perception graph data;

[0059] The high-order trend evolution reasoning module is used to perform cross-domain meta-correlation perception on the neighboring perception graph data to obtain cross-domain meta-correlation data; and to perform waterlogging trend prediction on the cross-domain meta-correlation data to obtain waterlogging trend data;

[0060] The global flood evolution simulation module is used to simulate global floods based on water accumulation trend data to obtain global flood data.

[0061] The present invention aims to achieve low-power, localized, high-frequency, multimodal waterlogging data collection by deploying ground-based intelligent node units integrated with millimeter-wave radar, pressure sensors, and neuromorphic computing chips within a predefined area. This approach effectively addresses the slow response and sparse distribution of traditional water level monitoring methods. The constructed geographic adjacency graph not only considers the spatial distance between nodes but also factors such as topographic structure, drainage paths, and flow directions. A local cross-domain graph attention mechanism integrates trend, modal, and historical features, improving the local graph's expressiveness and noise immunity. By aggregating multi-perspective meta-features and semantic partitioning to generate a high-order meta-correlation graph, the system captures implicit trend propagation paths and risk linkage mechanisms. A multi-branch trend prediction network is then used to model linear evolution, sudden surge events, and periodic disturbances, outputting more stable and sensitive trend data. By combining topographic data with adjacency topology to simulate global flood diffusion, a city-scale risk map is generated, significantly enhancing the ability to accurately predict and dispatch urban flood control measures during extreme climate conditions or sudden rainstorms. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0063] Figure 1 A flowchart showing the steps of a flood intelligent deduction method based on global fusion according to an embodiment is shown;

[0064] Figure 2A flowchart showing the steps of a node deployment and collection method according to an embodiment is shown;

[0065] Figure 3 A flowchart showing the steps of a method for constructing a geographic adjacency graph according to an embodiment is shown;

[0066] Figure 4 A flowchart showing the steps of a cross-domain meta-association perception method according to an embodiment is shown;

[0067] Figure 5 A flowchart showing the steps of a method for simulating global flood evolution according to an embodiment is shown;

[0068] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0069] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0070] Furthermore, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0071] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0072] See also Figures 1 to 5 , this application provides a flood intelligent deduction method based on global fusion, the method comprising:

[0073] S1. Deploy ground-nailed smart node units in a preset area and collect regional waterlogging data through the ground-nailed smart node units.

[0074] In one embodiment, the system sets up multiple preset monitoring points in low-lying or flood-prone areas in the city, and divides the area into several preset sensing areas based on the actual terrain and waterlogging risk distribution. 3 to 5 ground-nailed intelligent node units are deployed in each sensing area to improve the spatial resolution and fault tolerance of monitoring. Each ground-nailed intelligent node unit integrates a millimeter-wave radar module and a pressure sensor module, which are respectively used to monitor the current water height and the pressure change per unit area of ​​the ground in real time. The system sets the sampling frequency to 2Hz, that is, two sets of data are collected per second, namely the water height value and the force value per unit area at the current moment, which are recorded as the water height: , represents the depth of the surface water layer at the current moment; force per unit area: , which represents the vertical pressure value measured by the pressure sensor at the moment of sampling. The above two types of data are converted into digital signals by the analog-to-digital conversion module (ADC) and temporarily stored in the local cache module. In order to reduce data redundancy and network load, the system uses an event trigger mechanism to judge and process the sampled data. When any of the following trigger conditions is met, the sample is judged as an "event point": Condition 1: The change in the height of the accumulated water exceeds the set threshold, that is, the difference between the current height and the previous moment Greater than the set threshold , indicating a sudden change in water level; Condition 2: the rate of change of force per unit area per unit time Greater than the preset threshold , indicating a rapid increase in external water pressure. When any of the conditions are met, the system encapsulates the current data into an event data packet. Each event data packet contains the following five fields: Node number ( ), used to identify the source of the event; timestamp ( ), indicating the specific time when the event occurred; the change value of the water height ( ), used to describe water level fluctuations; force change value ( ), used to assess water pressure trends; event type identification ( ), used to distinguish different types of emergencies, such as "sudden water level change" or "abnormal pressure." Once the event data packet is constructed, the system reports it asynchronously via a low-power wide-area communication network. Supported communication methods include the LoRa communication protocol and the NB-IoT protocol.

[0075] S2. Construct a geographic adjacency graph for the regional waterlogging data to obtain proximity graph data; perform local cross-domain graph attention calculation on the proximity graph data to obtain proximity perception graph data;

[0076] In one embodiment, the system reads the geographic coordinates (including longitude and latitude) of all event nodes in the area and the water level information at the corresponding time point. Based on the above node information, the system uses the Delaunay triangulation method to generate an initial spatial adjacency structure graph. This graph is denoted as G=(V,E), where V represents the node set and E represents the edge set. The system generates an initial spatial adjacency structure graph for any two nodes. and Calculate Euclidean distance and height difference slope: If the Euclidean distance between two nodes is less than the set threshold , and the slope formed by the height difference and horizontal distance is greater than the minimum threshold , an initial connecting edge is established between the two. The system selects directional valid edges with physical rationality based on the degree of consistency between the direction of the edge and the direction of regional terrain flow. Specifically, the angle between the direction vector of the connection between nodes and the dominant water flow direction in the region is calculated. If the angle is less than 45 degrees, the edge is retained as a valid connecting edge; otherwise, it is discarded. Through the above operations, the system obtains a proximity graph structure that can reflect the local terrain diversion characteristics. The system constructs a feature vector containing multi-dimensional heterogeneous features for each node, specifically including the following five components: the current collected water height; the time variation amplitude of the water height (such as the difference between two consecutive time points); the current unit area pressure value; the trend direction extracted by the node in multiple time segments (such as rising, stable, and falling); the node's perception modality identification (such as millimeter wave, pressure sensing, visual monitoring, etc.). For each edge in the graph, the system uses a multi-head graph attention mechanism to calculate the weight. Specifically: for each attention head, the system configures linear transformation matrices for the query node and target node respectively to extract their attention representations; after weighting the calculation results of multiple attention heads, the system performs a Softmax normalization operation on the target node set to generate the attention weight of the source node to each neighbor node; based on the attention weight, the system performs a weighted summation of the adjacent node feature vectors of each node to form the perception enhancement vector of the node in the graph structure.

[0077] S3. Perform cross-domain meta-association perception on the neighboring perception graph data to obtain cross-domain meta-association data; perform water accumulation trend prediction on the cross-domain meta-association data to obtain water accumulation trend data;

[0078] In one embodiment, the system constructs a multidimensional feature tuple for all connected node pairs in a proximity perception graph to characterize the differential relationships between nodes. This tuple consists of two parts: a concatenated collection of node features to preserve absolute state information; and differences between node features to capture relative trends. Each set of node feature tuples includes information on the following dimensions: spatial position difference, trend similarity, and differences between perceptual modalities. Based on the perspective dimensions involved in the feature tuple (e.g., spatial, trend, or modality), these features are input into the corresponding multilayer perception network for feature transformation. Each perspective subnetwork independently outputs a perspective correlation score between nodes. The system performs a weighted fusion of the correlation scores from different perspectives. The fusion weights are determined through an attention mechanism, which is normalized based on the importance of each perspective's response. This outputs a unified correlation strength value. The system uses the combined correlation strength values ​​of all these node pairs as edge weights to construct a meta-correlation graph structure. Only edges with strength values ​​above a set threshold are retained in the graph. For each node in the meta-association graph, the system constructs a trend state sequence based on the historical water level sequence. This sequence typically consists of the magnitude of water level changes over consecutive time slices (e.g., three time steps), reflecting the node's current dynamic evolution of water level. The system also incorporates three trend identification submodules to address different types of trends: linear trend detection, which performs linear regression modeling on the state sequence within a sliding window to characterize uniformly rising or falling trends; exponential surge detection, which fits an exponential model and uses a dynamic threshold mechanism to identify abnormal patterns of water level surges; and periodic fluctuation detection, which uses Fourier transforms to extract frequency domain features and perform phase alignment detection to identify periodic water level fluctuations. The system assigns a confidence score to each of these three types of prediction results and uses this score as the basis for weighted fusion. The confidence score is calculated inversely proportional to the average prediction error (e.g., mean squared error) of the trend model on historical data. Smaller errors correspond to higher confidence scores and a greater fusion weight. The fused trend prediction value serves as the output of the estimated water level trend at the node's future time point.

[0079] S4. Conduct a full-area flood simulation based on the water accumulation trend data to obtain full-area flood data.

[0080] In one embodiment, the estimated water accumulation trend for each node, predicted in the previous phase, is injected into the corresponding graph node as the initial water level state. Each node represents an actual monitoring point or virtual computing unit within a geographic area. The system performs an iterative graph diffusion calculation, gradually updating the water level state of each node. The update rule considers the water level difference between adjacent nodes and the feasibility of water flow propagation. In each iteration, the water level state of each node is adjusted based on the current water level state of its neighboring nodes. The propagation intensity between nodes is determined by the following factors: terrain elevation difference (affecting water flow directionality), resistance factors caused by pavement materials or landform structure, and drainage capacity parameters of the area. The system only allows water level flow simulation along edges where the terrain angle is less than a set threshold (e.g., 45°) to conform to gravity flow. The entire graph diffusion process typically iterates for 20 to 50 rounds (i.e., propagation time steps) until the overall system state stabilizes or the simulation evolution reaches a preset termination condition. After the diffusion evolution is complete, the system analyzes the spatial distribution patterns of the simulation output and uses clustering algorithms (such as density-based or graph-cut-based regional clustering) to categorize the flooding status within the region, creating a flood risk zoning map. The resulting zoning generally includes the following three types of areas: High-risk flooding areas: areas with persistently high water accumulation, drastic water level fluctuations, and converging propagation paths; Propagation buffer zones: areas with moderate water level fluctuations but serving as flow transition nodes; and Potential risk source zones: areas with low initial water accumulation but connected to multiple high-risk areas through topographic pathways, indicating potential for evolution. The output includes a unique code for each area, a corresponding risk level label, and zone boundary data for visualization.

[0081] Optionally, S1 includes:

[0082] S11. Deploy ground-nailed intelligent node units in a preset first area, and collect water accumulation height and force per unit area to obtain water accumulation data for the first area;

[0083] In one embodiment, a typical low-lying and flood-prone area is selected in the first area, and no less than three ground-nailed intelligent node units are deployed along the key terrain change areas. Each ground-nailed intelligent node unit integrates the following sensing and acquisition modules, including a millimeter-wave radar module. The module operates at a frequency of 60GHz and the sampling frequency is set to 2 times per second (2Hz). It is used to calculate the current water level by the time delay of the echo reflected by the accumulated water. The measurement accuracy of this module can reach plus or minus 1 mm, and the water level data obtained is expressed as the water level at the current moment, recorded as The pressure sensor module is used to collect the vertical pressure value per unit area, which is expressed as The measured pressure is dynamically compensated to eliminate non-water accumulation interference caused by external disturbances such as human trampling, so that the obtained pressure data mainly reflects the actual water accumulation pressure caused by the static pressure of the water column. In addition, the system calculates the rate of change of water level and pressure data based on a three-second sliding time window. Specifically, the water level change rate is the difference between the current water level and the water level of the previous second, expressed as ,Right now ; The pressure change rate is the difference between the current pressure and the pressure one second ago, expressed as ,Right now The system records the average rate of change within the sliding window once per second as a standard sampling point. The system structures and encapsulates information such as sampling time, node number, water level, pressure per unit area, water level change rate, and pressure change rate, generating a standard data packet. This packet includes the following fields: current timestamp; node unique identifier; current water level; current pressure per unit area; current water level change rate; and current pressure change rate.

[0084] S12. Deploy ground-nailed intelligent node units in a preset second area, and collect water accumulation height and force per unit area to obtain water accumulation data for the second area, wherein the first area and the second area are different areas;

[0085] In one embodiment, the second area is another preset monitoring area that is physically isolated from the first area or discontinuous in terrain structure, such as a geographical unit located in a different block, intersection or drainage node. This area is also equipped with no less than three, preferably three to five, ground-nailed intelligent node units for independent data collection. The hardware structure, collection method, sensor type and configuration parameters of the ground-nailed intelligent node units in the second area are consistent with those in the first area, including a millimeter-wave radar module for obtaining the current water level height, recorded as Its measurement principle, frequency, sampling rate and accuracy are the same as those in the first area; the pressure sensor module is used to measure the vertical pressure per unit area, which is recorded as , and dynamically compensate for non-waterlogged pressure disturbances to ensure that the pressure data accurately reflects the waterlogging status; for continuous data points, the system uses a three-second sliding time window to calculate the rate of change of water level and pressure. The water level change rate is expressed as the current water level minus the water level of the previous second, that is, ; The pressure change rate is the current pressure minus the pressure of the previous second, that is The system extracts the average value within the sliding window as the standard sampling point per second, forming a standard structured data record. Each record includes the following fields: current sampling timestamp; node unique number; current water level; current unit area pressure; current water level change rate; current pressure change rate. The timestamps of data collected by all nodes are synchronized with a unified master control time base, preferably using a high-precision Global Navigation Satellite System (GNSS) clock to achieve timing consistency among multiple nodes.

[0086] S13, performing event-driven encoding on the first area waterlogging data and the second area waterlogging data to obtain event-driven data;

[0087] In one embodiment, the system filters and encodes the water accumulation data from the first area and the second area according to the preset water accumulation state triggering rules. The specific rules include the following three types of event triggering conditions: Water level sudden increase event: When the water level change amplitude per unit time exceeds the set threshold This event is triggered when the difference between the current water level and the water level at the previous time point is greater than the threshold, which is expressed as "the water level change is greater than ",in The water level sudden increase threshold is set to 20 mm per minute. Abnormal pressure event: When the pressure change rate per unit time exceeds the set threshold This event is triggered when the pressure change rate per unit time is greater than ",For example It can be set to 50 Pascals per second. Water pressure-water level inconsistency event: When the current pressure value is significantly higher than the static water pressure corresponding to the same water level, the system determines that there is abnormal water accumulation (such as sewer backflow) and triggers this event. The judgment condition is: "The current unit area pressure value is greater than the static pressure value converted from the water column height plus the error margin", that is, the following relationship is met: ,in is the force per unit area of ​​the node, is the water density, is the gravitational acceleration constant, is the current water level of the node, The margin is empirically set to preferably be greater than 300 Pascals. When waterlogging data meets any of the above trigger conditions, the system identifies it as an event record and encodes it into an event vector. Each event vector includes the following fields: the timestamp of the current event; the unique ID of the node that triggered the event; the event type label (such as "water level surge," "abnormal pressure," "water pressure and water level inconsistency," etc.); the current water level change; the current pressure change; and the node location encoding information. This event data is stored uniformly in the edge cache module, forming event-driven data.

[0088] S14. Enhance the spatiotemporal offset perception of event-driven data to obtain regional waterlogging data.

[0089] In one embodiment, the system calculates the time difference between the occurrence of the same type of flooding events between different nodes and constructs a response offset matrix. For the same type of events at the i-th node and the j-th node, the response time difference is recorded as , defined as: response offset This is equal to the time at which the event at node j occurred minus the time at which the event at node i occurred, expressed in seconds. When multiple intelligent nodes experience the same type of flooding event consecutively within a set time window, the system considers it a spatiotemporal coordinated event group. The system implements a spatiotemporal fusion mechanism based on a Gaussian kernel function to enhance regional water level data. ,in For nodes The water level enhancement value is obtained by space and time weighted fusion correction. is the node order item, is the spatial distance attenuation factor, is the time delay attenuation factor, For nodes In time The water level value, is the time parameter, For nodes Relative Node The event time delay of each node j is represented by the water level contribution weight, which consists of two parts: (1) the spatial distance attenuation factor, which is expressed as a Gaussian kernel function with the physical distance between the two nodes as the variable, ,in is the spatial distance attenuation factor, Gaussian kernel function, used to represent nodes and nodes The spatial distance between them decays, For nodes and The spatial distance between is the spatial scale control parameter; (2) the time delay attenuation factor, which is expressed as a Gaussian kernel function with the time offset as a variable, in the form of ,That is the time delay attenuation factor, Gaussian kernel function, used to represent the node Relative Node The event time delay decays, For nodes Relative Node The event time delay, This is a time scale control parameter that adjusts the impact of time offset. The current water level at each node is dynamically fused and corrected based on spatial proximity and response synchronization. The system represents the processed regional flooding status as a three-dimensional data tensor. This tensor contains the following: spatial location coordinates (x, y); current time t; enhanced water level at that location (regional fused water level); and enhanced pressure at that location (regional fused pressure). This data structure is represented as: Regional flooding tensor = {(x, y, t), enhanced regional water level, enhanced regional pressure}.

[0090] Optionally, the geographic adjacency graph construction includes:

[0091] S21, extracting regional nodes from regional waterlogging data to obtain regional node data;

[0092] In one embodiment, the input data of the regional node extraction operation is the regional waterlogging data tensor, which is expressed as: ,in Represents spatial coordinates; Indicates a point in time; Indicates that at the coordinate point ( ) place, time point Height of accumulated water; Indicates the pressure per unit area at the corresponding position. To extract representative regional nodes, the system first constructs a spatial grid structure on the region and assigns each grid unit In each spatial grid cell, it is determined whether there is significant water accumulation fluctuation or pressure change in the most recent time window. The length of the time window is set to Specifically, if any of the following conditions are met at any time point in the past T time range within the grid: the maximum value of the first-order change rate of the water level (i.e. the difference between adjacent time points) exceeds the preset threshold ,Right now: , or the maximum value of the pressure per unit area in the time window exceeds the pressure threshold ,Right now: , then determine the center point of the grid For each regional node that is determined to be valid, the system records its corresponding water level and pressure values, and calculates the trend slope value (i.e., linear trend indicator) by performing a linear fit on the water level data of the node in the last five time steps.

[0093] S22, performing terrain semantic classification on the regional node data to obtain regional classification data;

[0094] In one embodiment, the elevation value of the corresponding position of the node is obtained ; Calculate the following semantic indicators, local elevation difference: ,in For nodes The local elevation difference of the node The elevation of its neighboring nodes The difference between the average elevations of For nodes In position The elevation value at For nodes The average value of the elevation values ​​of the neighboring nodes, For nodes The set of elevation values ​​of all adjacent nodes; local slope (flow potential): , For nodes The local slope of For nodes Water flow gradient amplitude at; drainage accessibility index (TWI): ,in For nodes The drainage accessibility index, For nodes The catchment area of ​​the node The water flow convergence area; according to the above indicators, the nodes are divided into three categories, including water confluence nodes: , transition node: , slope or high ground node: ,in is the low slope threshold, is the high slope threshold, is the high drainage accessibility index threshold, is the low drainage accessibility index threshold, and outputs node labeled data.

[0095] S23, constructing a mobility weighted directional adjacent edge for the regional classification data to obtain adjacent edge data;

[0096] In one embodiment, for each node ,exist Retrieve neighboring nodes within range ,like ,in For nodes With node The Euclidean distance between is the maximum distance threshold, limiting the node and nodes The maximum adjacency distance between For nodes and nodes The angle between them represents the angle between the topographic flow direction and the water flow gradient. For nodes and nodes The angle between the flow direction and the water gradient direction, For nodes and nodes The flow direction vector between For nodes or node , then add the edge set E and assign a flow direction weight to each edge: ,in For nodes and nodes The directional weight of the mobility between Towards Possibility of migration, is the coefficient of mobility weight, which is used to control the influence of water flow gradient on weight. For nodes The water flow gradient value, For nodes The water flow gradient value, is the coefficient of trend weight, which is used to control the impact of trend changes on weight. For nodes The water accumulation trend value, For nodes The water accumulation trend value, The larger it is, the more the water Towards Migrate and output adjacent edge data.

[0097] S24, performing local structural similarity screening on the adjacent edge data to obtain adjacent edge screening data;

[0098] In one embodiment, for each edge , extract the local trend distribution in the k-neighborhood (such as k=3) of the nodes at both ends, and count the distribution vector of the water height trend value (i.e. the aforementioned linear slope) of each node in the node neighborhood; the pressure change rate distribution, calculate the time change rate of the force value per unit area of ​​each node in the node neighborhood (i.e. ), and form the pressure change rate vector of the area; terrain elevation difference, calculate the elevation difference between the node and its neighboring nodes (i.e. , For nodes The elevation value of For nodes The elevation value of a node in the neighborhood) forms a local differential feature vector that reflects the terrain undulation. Calculate the cosine similarity of the local structure vector: ,in For nodes and nodes The cosine similarity of the local structure vector between them is used to measure the similarity of the structure between two nodes. For nodes The feature vector set of all nodes in the neighborhood of , including the local trend, pressure change rate and terrain difference of each adjacent node, For nodes The feature vector set of all nodes in the neighborhood of , including the local trend, pressure change rate and terrain difference of each adjacent node, only retains the features that meet The edges of .

[0099] S25. Graph construction is performed based on regional node data and adjacent edge screening data to obtain adjacent graph data.

[0100] In one embodiment, the construction diagram structure ,in is the node set after semantic classification; It is a set of valid flow direction edges after structural similarity screening; each node is accompanied by a real-time water accumulation feature vector; terrain classification label; dynamic trend status; each edge is accompanied by a liquidity weight; direction vector; similarity score; output proximity graph structure.

[0101] Optionally, the local cross-domain graph attention calculation includes:

[0102] Extract node multi-domain features from the proximity graph data to obtain node multi-domain feature data;

[0103] In one embodiment, the input graph structure is: , where each node It has the following original characteristics and water level characteristics: , pressure characteristics: , slope and elevation: , position encoding: , construct three types of perceptual domain features, including physical domain feature vectors: , terrain domain feature vector: , time series domain feature vector (trend in the past T=3 steps): ,in For nodes The water level change in the past step, For nodes The change in water level in the past 2 steps, For nodes The water level changes in the past three steps are concatenated to obtain the feature vector: .

[0104] Perform cross-domain weighted graph attention allocation on the node multi-domain feature data to obtain cross-domain weighted graph attention data;

[0105] In one embodiment, three types of attention mapping functions are defined, including physical domain attention coefficient calculation: ,in is the physical domain attention coefficient, is the normalized activation function, is the physical domain attention mapping weight matrix (preset parameters), For nodes The physical domain eigenvector of For nodes Physical domain feature vector; terrain domain attention coefficient calculation: ,in is the terrain domain attention coefficient, is the normalized activation function, is the terrain domain attention mapping weight matrix (preset parameters), For nodes The terrain domain feature vector of For nodes The terrain domain feature vector; the time series domain attention coefficient calculation: , is the temporal domain attention coefficient, is the normalized activation function, is the temporal domain attention mapping weight matrix, For nodes The time domain feature vector of For nodes The time domain feature vector of ,satisfy , calculate cross-domain attention: ,in is the cross-domain weighted attention coefficient after fusion, is the physical domain weighting coefficient, is the terrain domain weighting coefficient, is the time domain weighting coefficient, outputting cross-domain weighted graph attention data .

[0106] The neighboring graph data is momentum suppressed and fused according to the cross-domain weighted graph attention data to obtain the neighboring perception graph data.

[0107] In one embodiment, assume that the current graph structure is G=(V,E), where V is the node set and E is the edge set. , and its state vector at the tth iteration is recorded as The system firstly follows the node The adjacent node set N(i) and the cross-domain attention weight , the feature vector of the adjacent node Perform weighted summation to form the neighborhood aggregation representation of the current round , calculated as node The neighborhood aggregation feature is equal to: all adjacent nodes The eigenvector of , multiplied by the attention coefficient of the node pair , then sum all adjacent nodes to get the aggregate feature vector The system performs momentum decay processing, performs weighted fusion of the above current aggregation results and the historical state vector, and updates the node state. The update formula is the node state at the next moment Equal to the current aggregate feature Multiply by (1 minus the momentum factor γ), plus the historical state vector Multiply by the momentum coefficient γ. The momentum coefficient γ is a real number in the interval [0,1] and is set between 0.2 and 0.3 to prevent unstable behavior caused by drastic state updates. The above aggregation and fusion process can be iterated for multiple rounds, preferably 2 to 3 rounds (i.e., T = 2 to 3). After T rounds of aggregation, the state vector obtained by each node is , which is regarded as the fusion feature of the node based on local multi-domain semantic enhancement, recorded as The system outputs the fused graph structure, which is defined as the proximity perception graph , where the features of each node already include a multi-layer fusion representation of its surrounding semantic features, historical dynamic features and spatial topological features.

[0108] Optionally, the cross-domain meta-association perception includes:

[0109] S31, performing multi-dimensional feature element pairing between nodes on the neighboring perception graph data to obtain node feature element pair data;

[0110] In one embodiment, the input is a constructed proximity perception graph structure, denoted as ,in represents the set of nodes in the graph, represents the edge set, Represents the fused multi-domain perception feature vector corresponding to each node. For each pair of nodes in the graph that have a connecting edge ( , ), the system constructs the corresponding multi-dimensional feature element pair to characterize the perceptual feature relationship between the two nodes. Specifically, the system constructs the feature combination vector based on the following four items: Node Multi-domain fusion feature vector of node The multi-domain fusion feature vector of ; the feature difference vector between the two (i.e. , used to express local trend differences); the element-wise product vector between the two (i.e. , used to express the interaction strength or similarity between nodes). The above four features are combined to form a complete set of feature element pairs, which are used to describe the fine-grained interaction relationship between two adjacent nodes in terms of spatial structure, perception mode and trend state. Each set of feature element pairs is defined as a candidate association pair and is uniformly stored in the association pair set, recorded as , that is, the set of candidate association pairs It consists of multi-dimensional feature element pairs corresponding to all node pairs connected by edges in the graph.

[0111] S32, performing multi-view meta-feature aggregation on the node feature meta-pair data to obtain node feature aggregation data;

[0112] In one embodiment, three parallel aggregation channels are designed, including distance and perspective aggregation: ,in For nodes and nodes Feature representation from a distance perspective, is a trainable mapping function from a distance perspective (such as a two-layer MLP+activation function), For nodes The eigenvector of For nodes Characteristic vector of ; trend similarity aggregation: ,in For nodes and nodes Feature representation from the perspective of trend similarity, is a trainable mapping function under the trend similarity perspective, For nodes The eigenvector of For nodes Eigenvectors of ; physical coupling aggregation: ,in For nodes and nodes Feature representation from the perspective of physical coupling, is a trainable mapping function from the perspective of physical coupling, For nodes The pressure per unit area, For nodes The height of the water accumulation, For nodes The pressure per unit area, For nodes The water level, each Represents a class of trainable mapping functions (such as two-layer MLP+activation function). Geometric multi-view channels, weighted aggregation through the attention mechanism: ,in For node pairs and The fusion feature representation of is the view order item, with values ​​of 1, 2, 3. For the The attention weight of the view, For the Nodes under perspective and nodes The feature representation of is the index term, is the channel weight score, calculated by the channel self-attention network, is the traversal index of the sum item. Forming the aggregation feature: .

[0113] S33, generating a meta-association graph based on the node feature aggregation data to obtain meta-association graph data;

[0114] In one embodiment, for each pair of candidate nodes in the graph, the system calculates the potential meta-association strength based on the similarity between their aggregated feature vectors. This strength is calculated as follows: first, the aggregated feature vector of one node is multiplied on the left and the aggregated feature vector of the other node is multiplied on the right. A pre-trained weight matrix is ​​introduced to adjust the matching direction within the feature space. The product is then normalized by a sigmoid function to obtain a correlation value between 0 and 1. The system sets a meta-association threshold to determine whether there is a significant high-order semantic or structural association between the nodes. If the potential correlation value of a pair of nodes exceeds the threshold, a valid meta-association relationship is considered between them, and a corresponding meta-edge is established in the graph. All node pairs that meet these conditions are included in the meta-graph edge set, and the system constructs a meta-association graph structure. This meta-graph consists of the original node set and the meta-association edge set generated based on it, retaining the aggregated feature vector of each node as a node attribute of the graph.

[0115] S34. Perform semantic partitioning projection on the meta-association graph data to obtain cross-domain meta-association data.

[0116] In one embodiment, the system employs a graph neural network to perform node embedding, simultaneously considering the node's aggregated feature information and the topological relationship of the metagraph structure. Specifically, each node is embedded using a graph convolutional network (e.g., GCN) or a graph-structure-aware embedding algorithm (e.g., DGI, GraphSAGE-EM), resulting in a low-dimensional vector representation representing its semantic properties. This embedding process not only preserves the spatial, trend, and physical characteristics of the original node features, but also incorporates contextual information about the cross-domain coupling structure within the metagraph. The system then performs semantic partitioning on the generated node embedding vectors using unsupervised clustering algorithms, including density clustering algorithms (e.g., DBSCAN) to mine local dense structures, or spectral clustering methods to capture natural partitioning trends within the graph based on modularity. The clustering results assign each node to a semantic region, thereby completing the assignment of semantic labels. The system outputs a cross-domain meta-association data structure consisting of three components: a node identifier, a semantic region label, and the original aggregated feature vector. The structure can be formally represented as a set of triples, where each triple corresponds to a node and its semantic attribution and aggregation features.

[0117] Optionally, the multi-view meta-feature aggregation includes:

[0118] Perform perspective division on the node feature element pair data to obtain perspective division data;

[0119] In one embodiment, for each node feature element pair , divide the perspective subsets, including dividing the spatial difference perspective ,extract , represents the difference between two nodes; divide the coupling strength perspective ,extract , used to represent nonlinear interactions; divide the original fusion perspective ,extract , retaining the basic structural features. Each pair of meta-features is divided into three types of view tensors , corresponding to three perception dimensions respectively.

[0120] Perform multi-view sub-network aggregation on the view division data to obtain multi-visual aggregated data;

[0121] In one embodiment, the system perceives the perspective for each type of structure Design the corresponding lightweight sub-network structure and adopt the parameter sharing mechanism to enhance the generalization ability and computational efficiency. Each sub-network is denoted as , dedicated to processing the view tensor Each sub-network The multi-perceptual convolution module (MPConv) consists of two layers in series, which are used to extract local structural trend features and global response features respectively. The calculation process is as follows: Apply the first layer of MPConv convolution and pass the ReLU activation function to extract the preliminary local trend response; the intermediate features are input to the second layer of MPConv convolution layer to further extract high-order response information and form a structural representation vector. The calculation process of the entire sub-network can be expressed as: ,in For the The sub-network function corresponding to the class structure perception perspective is specifically used to process the perspective tensor , is the ReLU activation function, For the The weight matrix of the second convolutional layer from the class perspective is used to extract global response features. For the The weight matrix of the first convolutional layer from the class perspective is used to extract local structural trend features. For the Node pair from the class perspective and The input tensor represents the feature information between nodes and performs dimension alignment on all view results: , summarized as a multi-view feature set: .

[0122] Semantic attention fusion is performed on multi-view aggregated data to obtain node feature aggregated data.

[0123] In one embodiment, for each pair of adjacent nodes and , the system is based on its multi-view feature set , for each perspective Extract the corresponding attention score The attention mechanism is modeled based on trainable parameters and the calculation formula is as follows: ,in For node pairs and In the The attention score under the class perspective indicates the importance of the perspective feature in the fusion. is the index term, is the attention projection vector, is the hyperbolic tangent activation function, is the attention mapping matrix, For the Feature representation from the perspective of class structure, is the index of the viewing angle, ranging from m∈{1,2,3}, For the Node pairs from the class perspective and The system uses the above attention score to perform weighted summation on each perspective feature to obtain the node pair ( , ) is represented by the fusion structure, which is recorded as: , For node pairs and The fusion structure representation represents the feature vector obtained after fusion. is the perspective order item. The system has All adjacent edges (i.e. all adjacent nodes The node feature aggregation vector is defined as follows: , For nodes The polymeric structural characteristics of For nodes The set of adjacent nodes of For nodes The adjacent node index of is the number of adjacent nodes.

[0124] Optionally, the water accumulation trend prediction includes:

[0125] Constructing a cross-domain meta-graph for the cross-domain meta-association data to obtain cross-domain meta-graph data;

[0126] In one embodiment, the system selects characteristic points with significant evolution trends in the time series from cross-domain meta-correlation data as nodes in the graph. The basis for node selection includes criteria such as a significant increase in the rate of change, a continuous change in the trend direction, and frequent occurrence of mutation points. Each graph node contains the following field information, including a geographic location code, which is used to identify the location coordinates or number of the node in two-dimensional space; the average water depth, which represents the average water depth of the node within a certain time window, which is used to measure the severity of the waterlogging; the pressure change rate per unit time, which reflects the dynamic changes in local pressure caused by waterlogging; recent trend labels, such as rising, falling or oscillating, which are used to reflect dynamic characteristics in the short term. The edge connection relationship between nodes is constructed based on the triple similarity criteria, including the spatial proximity criterion. If the geographical distance between any two nodes is less than the preset threshold (for example, 200 meters), they are considered to have spatial similarity; the trend similarity criterion. The consistency of the direction of water accumulation change of the two nodes is calculated within the sliding time window. If the direction consistency index is greater than 0.8 (such as cosine similarity or trend matching after dynamic time warping), the two are considered to have trend similarity; the structural dependency criterion. If the two nodes are located in the same drainage unit, the same low-lying area, or share the same drainage path, they are considered to have structural coupling dependence. The system establishes edge connections between node pairs only when they meet at least two of the above three similarity criteria. For each edge (connected by the node With node The system assigns a set of three-dimensional attribute vectors to it, which are used to represent the linkage characteristics between nodes. The attribute vector includes the spatial distance ( ):node and The Euclidean distance between ): The angle between the line connecting two points and the main slope direction of the terrain; Trend correlation coefficient ( ): The correlation coefficient (e.g., Pearson correlation coefficient) between the water level or pressure change trends of two nodes within a selected time window. The cross-domain metagraph structure constructed by the system is represented as: Cross-domain metagraph structure = (node ​​set V, edge set E, edge attribute matrix A), where the node set V represents all meta-nodes with evolutionary characteristics; the edge set E represents the node pair connection relationship that meets the triple criteria; and the edge attribute matrix A is used to store the three-dimensional attribute vector corresponding to each edge (spatial distance, slope angle, trend correlation).

[0127] Perform multi-branch trend inference on cross-domain meta-graph data to obtain multi-branch trend data;

[0128] In one embodiment, three functional reasoning branches are included, corresponding to the following three types of waterlogging trends, including branch A: linear trend reasoning, which is used to model the steady growth or slow decline of the waterlogging state, and is suitable for the steady-state evolution process caused by continuous rainfall or poor drainage. Branch B: exponential surge reasoning, which is used to simulate the nonlinear dynamics of rapid rise of waterlogging under the background of short-term heavy rainfall, which is typical of sudden waterlogging scenarios. Branch C: periodic fluctuation reasoning, which is used to model the periodic changes of waterlogging state caused by drainage cycle or tidal interference. The three branches are all based on the graph attention neural network structure, with the node features in the cross-domain metagraph as input, and respectively execute the following reasoning processes, linear trend output, ,in For nodes The linear trend inference result of node The trend prediction value obtained by linear transformation after aggregation in the Graph Attention Network (GAT), is the weight matrix of the linear trend branch, which is used to perform linear transformation on the aggregation results of the Graph Attention Network (GAT). is a graph attention network, used to aggregate node neighbor information. It is a cross-domain meta-graph containing a graph structure of nodes and edges, used to input the graph attention network. For nodes The characteristic vector of is used as the input of the graph attention network. The linear trend output of each node is obtained by linear transformation of the graph attention result, which is expressed as node The linear trend output is equal to the weight matrix Acting on its GAT aggregation result. Exponential surge output, ,in For nodes The exponential increase inference result indicates that the node Water accumulation prediction in sudden flood disaster scenarios, is a nonlinear activation function (such as Sigmoid) used to enhance the sensitivity of exponential surges. is the weight matrix of the exponential burst branch, which is used to adjust the combined result of the graph attention network (GAT) and the mutation response factor. For the operation of the Graph Attention Network (GAT), the information of neighbor nodes is aggregated. is an exponential function, For nodes The mutation response factor is used to control the rate of change in the exponential function, node The exponential trend output is composed of the graph attention result and the node mutation response factor The exponential function composed of the two is generated by the weight matrix Adjustment, using nonlinear activation functions (such as Sigmoid) to enhance mutation sensitivity. Periodic trend output, ,in For nodes The periodic trend inference result of the node The predicted value of the cyclical fluctuation, is the weight matrix of the periodic trend branch, which is used to adjust the amplitude of the periodic trend. This branch directly introduces explicit time information and models periodic fluctuations based on the sine function form. The expression includes the periodic frequency term , phase term , and the weight matrix Amplitude adjustment of the control.

[0129] The multi-branch trend data are subjected to time series inversion verification to obtain the waterlogging trend data.

[0130] In one embodiment, when the historical sequence of waterlogging is known, the trend prediction result is verified by reverse reasoning; the real waterlogging data sequence in the historical time window [tT, t] is used. , push the prediction results back to [t-1,tT] to generate a simulation sequence . Calculate the structural consistency score between the simulated sequence and the historical observations: ,in is the structural consistency score of the reverse reasoning verification, which indicates the degree of match between the predicted results and the real data. is the length of the time window, which indicates the number of time steps used in the historical data. is the time point index, indicating each specific time point in the time window. is the simulation prediction result at time point ti, is the real historical observation data at time point ti, is a smoothing constant used to avoid the denominator being zero and to prevent division by zero errors during calculation. ( is the threshold of the reverse reasoning consistency score, which is used to judge the credibility of a branch prediction), then its corresponding fusion weight decreases in the next round of training; and the branch result is removed from the trend prediction at the current moment. The sequence of credible trend prediction values ​​under error control is obtained .

[0131] Optionally, the multi-branch trend reasoning includes:

[0132] Extract trend features from cross-domain meta-graph data to obtain trend feature data;

[0133] In one embodiment, the input is cross-domain meta-graph data , where each node Contains time series features (such as water level, pressure, rate, etc.); performs fixed-length processing on the time series of each node, with the fixed length size set to w=5 and the step size to 1; extracts the following feature vectors in each window : mean, slope, maximum gradient, the first three orders of Fourier amplitude, volatility; all node features are standardized to form a trend feature matrix , where d is the feature dimension.

[0134] Perform linear regression calculation on trend characteristic data to obtain linear water accumulation data;

[0135] In one embodiment, the least squares method is used to perform linear fitting of the time series: , where the coefficient Depend on: ,in For nodes In time The linear fitting prediction value of the node The predicted value of water accumulation, For nodes The linear regression coefficient represents the slope of the water accumulation change. is the time parameter, For nodes The linear regression intercept of represents the starting value or constant term of the water accumulation change. For each time point in the time series, is the mean of the time points in the time series, For the time point The observed value at is the observed value The mean of .

[0136] Perform exponential curve fitting on trend characteristic data to obtain exponential surge data;

[0137] In one embodiment, an exponential growth model is fitted to the abnormally high slope region: ,in For nodes In time The exponential fitting prediction value of the exponential growth model indicates that the exponential growth model The predicted value of water accumulation, For nodes The exponential model amplitude represents the initial or maximum value of the model. is the exponential function part, which represents the node Over time exponential growth trend, is the exponential growth rate, For nodes The constant offset represents the constant term in the exponential model. The parameter estimation is iteratively optimized using nonlinear least squares fitting. For areas where convergence fails or the slope is extremely low, no results are output and the branches are set to be invalid.

[0138] The trend feature data is processed by frequency domain convolution network to obtain periodic fluctuation data;

[0139] In one embodiment, the main frequency is extracted using discrete Fourier transform (DFT) to construct a spectrum sequence: ,in For nodes The spectrum sequence represents the main frequency information extracted by discrete Fourier transform. is the time step, which represents the sampling moment of the signal, is the total length of the time series, For nodes In time The signal value of is the base of natural logarithms, is the imaginary unit, indicating that the square root of a complex number is -1. is the circumference constant, is the frequency, which represents the frequency component in the discrete Fourier transform. The spectrum amplitude is normalized and then input into the 1D convolutional neural network (FCN). The network structure is Conv1D(16) (one-dimensional convolution layer, the number of channels is set to 16) → ReLU → Conv1D(8) (one-dimensional convolution layer, the number of channels is 8) → MaxPool (maximum pooling layer) → Dense (fully connected layer) → output sine parameters (amplitude, frequency, phase); construct the prediction function: ,in is the periodic fluctuation data, For nodes The spectrum amplitude represents the strength of the periodic signal. For nodes The frequency of the periodic signal is represented by For nodes The phase of , represents the initial phase of the periodic signal, For nodes The offset of , which represents the constant offset of the periodic signal.

[0140] The credibility trend of linear water accumulation data, exponential surge data and periodic fluctuation data is fused to obtain multi-branch trend data.

[0141] In one embodiment, a credibility score is calculated for each branch result. :Linear: ,in For nodes Confidence score in linear trend forecast, is the normalized mean square error of the linear trend, which is used to calculate the confidence score of the linear trend; index: ,in For nodes Credibility score in exponential trend forecast, is the Akaike information criterion of the exponential trend, which is used to calculate the credibility score of the exponential trend. The maximum AIC value among all prediction models is used to standardize the confidence score of the exponential trend; period: , For nodes Credibility score in periodic trend prediction; fusion weight of each branch is obtained through Softmax: ,in For nodes The weight coefficient in trend fusion is used to weight each branch model (linear, exponential, periodic). For nodes Credibility score in trend forecast, is the branch result sequence item, For nodes Credibility scores in other branching models (e.g. linear, exponential, cyclical), is an exponential term, integrating trend values: , For nodes For the trend forecast branch The fusion weight coefficient, For nodes In time Next, based on the branch (e.g. linear, exponential or cyclical models), For nodes In time The fusion trend prediction value under .

[0142] Optionally, the global flood simulation includes:

[0143] S41, performing regional state trend processing based on the water accumulation trend data to obtain regional state trend data;

[0144] In one embodiment, for each regional node in the system, the predicted value of the waterlogging trend in its historical time series is obtained. Specifically, the predicted value of the waterlogging trend of a regional node at multiple consecutive time points is assumed to be a set of time series, and each time point contains the following data elements: the predicted waterlogging height at the current moment; the difference in waterlogging height between the current moment and the previous moment, represented as the waterlogging change rate; the upstream runoff impact factor of the area where the node is located, reflecting the sensitivity of the terrain to waterlogging. Based on the above data, a three-dimensional state vector is constructed for each regional node, including the current predicted waterlogging value (expressed as waterlogging depth); the waterlogging growth rate (i.e., the first-order time difference of the predicted value); and the relative terrain or upstream runoff impact factor (which can be calculated based on the terrain height difference and drainage capacity). Based on the constructed state vector, the system sets a threshold interval to divide the waterlogging state into different levels. The specific rules are as follows: when the state value is higher than the set upper threshold (for example, the severe flooding threshold), the node is marked as "severe waterlogging"; when the state value is between the upper and lower thresholds, it is marked as "moderate waterlogging"; when the state value is lower than the lower threshold, it is marked as "light waterlogging" or "no waterlogging". The above thresholds can be set based on the empirical value of the impact of historical measured water levels on waterlogging conditions or expert evaluation. The system records the state classification results corresponding to all regional nodes at each time point as a regional state trend graph. The graph structure uses nodes as units and time as the dimension, and encodes the waterlogging state of each node over time into a time-region mapping graph. The output result is a data set containing a triple of node identification, time point and state label, which is used for flood propagation modeling and response strategy formulation.

[0145] S42, performing graph diffusion evolution deduction on the regional state trend data to obtain graph diffusion evolution data;

[0146] In one embodiment, the evolution is based on a regional state trend graph, which contains several regional nodes and their state labels at multiple time points. The system combines a pre-built geographic adjacency graph to determine the set of adjacent nodes for each node, forming a dynamic graph structure. Each edge represents the connection between two regions in terms of water or terrain. The system defines a propagation update function for node status. ,in For nodes In time The status value of For nodes In time The status value of is the state propagation rate coefficient, For nodes Neighbor node index, For nodes The set of neighbor nodes of For nodes and nodes The weight factor between For nodes In time The status value of For nodes In time The state value of a node at the next moment is determined by the weighted sum of the difference between the current state and the state of all its adjacent nodes. The weight factor takes into account the following influencing factors: the difference in terrain height between adjacent areas; the smoothness or drainage capacity of potential water flow paths; and the credibility of the physical connectivity of adjacent edges. The state change rate is controlled by an adjustment coefficient to simulate the actual propagation rate and ensure physical rationality. To avoid unlimited amplification of the state value or unreasonable sudden increase, the system performs state suppression. ,in Node after state suppression In time The status value of For nodes In time The status value of is the natural exponential term, It is a Sigmoid function, which is used to limit the state value to change within a reasonable range. is the slope control factor of the Sigmoid function, is the equilibrium position parameter of the Sigmoid function, which applies a nonlinear compression function (such as an S-shaped function) to the evolved state value. This function constrains the state value to a reasonable range by setting a slope control factor and the equilibrium position parameter, more closely matching the evolution of waterlogging in physical space. The system iteratively updates the entire graph structure over multiple rounds, each calculation based on the node state and adjacency propagation relationships at the previous moment. The number of evolution rounds can be set based on a preset simulation time range or state convergence (e.g., 20 to 50 rounds). After each update round, the new state value overwrites the original state graph, forming an evolutionary sequence. The system outputs the evolved graph state matrix, which represents the dynamic evolution of the state trend values ​​of each region at multiple time points as neighborhood interactions occur.

[0147] S43. Partition the graph diffusion evolution data into risk layers to obtain global flood data.

[0148] In one embodiment, the regional state matrix obtained during the aforementioned graph diffusion evolution process is used as input. Each node in the matrix contains the state evolution value at the current time point. For each regional node, its diffusion rate relative to the adjacent regions is calculated, and the average state difference between the node and all its adjacent nodes is calculated to measure the intensity of the node's external diffusion or external influence. ,in For nodes The diffusion rate reflects the average value of the difference between the state of a node and its adjacent nodes. For nodes The size of the adjacent node set of node The number of neighbors, For nodes The adjacent node index of For nodes The state evolution value at the current time point, For nodes The state evolution value at the current time point. This indicator reflects the spatial gradient of waterlogging spread between regions. The system constructs a two-dimensional risk score vector based on the state value of each node and its diffusion rate. , For nodes The risk score vector of , which combines the node's state value and diffusion rate, is the weighting coefficient, which is used to control the influence weight of the state evolution value in the risk score. For nodes In time The state evolution value of is the weighting coefficient, which is used to control the influence weight of diffusion rate in risk scoring. For nodes The score is composed of two weighted components: the node's own state evolution value, reflecting the severity of the current flooding; and the diffusion rate, reflecting the potential risk of spread. The weighting coefficients can be set empirically or determined through training based on the specific application scenario. Based on the risk score vector, a hierarchical clustering algorithm (such as the density-based clustering algorithm DBSCAN or a pre-set rule partitioning method) is used to cluster and analyze regional nodes. This clustering process can be jointly constrained based on spatial adjacency and score similarity. Based on the clustering results, the system divides all areas into the following three categories: Level 1 high-risk flood zones: corresponding to areas prone to sudden flooding, typically located in low-lying areas, with high scores and diffusion gradients; Level 2 diffusion buffer zones: located in the middle or periphery of the flood propagation path, with moderate scores, serving as a buffer and transition zone; Level 3 safe zones: located on high ground or slopes, with low state values ​​and diffusion gradients, and considered low-risk or no-risk areas. The system outputs a global flood zoning map, which assigns a risk level label to each regional node. The risk level of each node is marked as level 1 (high risk), level 2 (medium risk), and level 3 (low risk).

[0149] Optionally, the present application further provides a flood intelligent deduction system based on global fusion, which is used to execute the flood intelligent deduction method based on global fusion as described above. The flood intelligent deduction system based on global fusion includes:

[0150] The node deployment and collection module is used to deploy ground-nailed intelligent node units in a preset area and collect regional water accumulation data through the ground-nailed intelligent node units;

[0151] The graph neural perception fusion module is used to construct a geographic adjacency graph for regional waterlogging data to obtain proximity graph data; it performs local cross-domain graph attention calculation on the proximity graph data to obtain proximity perception graph data;

[0152] The high-order trend evolution reasoning module is used to perform cross-domain meta-correlation perception on the neighboring perception graph data to obtain cross-domain meta-correlation data; and to perform waterlogging trend prediction on the cross-domain meta-correlation data to obtain waterlogging trend data;

[0153] The global flood evolution simulation module is used to simulate global floods based on water accumulation trend data to obtain global flood data.

[0154] Therefore, no matter from which point of view, the embodiments should be regarded as illustrative and non-restrictive, the scope of the present invention is limited by the attached application documents rather than the above description, and it is intended that all changes that fall within the meaning and scope of equivalent elements of the application documents are included in the present invention.

[0155] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A flood intelligent deduction method based on global fusion, characterized by: The method comprises: S1. Deploy ground-nailed smart node units in a preset area and collect regional waterlogging data through the ground-nailed smart node units. S2. Construct a geographic adjacency graph for the regional waterlogging data to obtain proximity graph data; extract node multi-domain features from the proximity graph data to obtain node multi-domain feature data; perform cross-domain weighted graph attention allocation on the node multi-domain feature data to obtain cross-domain weighted graph attention data; perform momentum suppression fusion on the proximity graph data based on the cross-domain weighted graph attention data to obtain proximity perception graph data; S3. Perform multi-dimensional feature element pairing between nodes on the proximity perception graph data to obtain node feature element pair data; perform multi-perspective meta-feature aggregation on the node feature element pair data to obtain node feature aggregation data; generate a meta-association graph based on the node feature aggregation data to obtain meta-association graph data; perform semantic partitioning projection on the meta-association graph data to obtain cross-domain meta-association data; perform water accumulation trend prediction on the cross-domain meta-association data to obtain water accumulation trend data; S4. Conducting a full-area flood simulation based on the water accumulation trend data to obtain full-area flood data; The multi-view meta-feature aggregation includes: Perform perspective division on the node feature element pair data to obtain perspective division data; Perform multi-view sub-network aggregation on the view division data to obtain multi-visual aggregated data; Semantic attention fusion is performed on multi-view aggregated data to obtain node feature aggregated data.

2. The method according to claim 1, characterized in that S1 includes: Deploy ground-nailed intelligent node units in a preset first area, collect water accumulation height and force per unit area, and obtain water accumulation data for the first area; Deploy ground-nailed intelligent node units in a preset second area, and collect water accumulation height and force per unit area to obtain water accumulation data for the second area, wherein the first area and the second area are different areas; Performing event-driven encoding on the first area waterlogging data and the second area waterlogging data to obtain event-driven data; The event-driven data is enhanced with temporal and spatial offset perception to obtain regional waterlogging data.

3. The method according to claim 1, characterized in that The geographic adjacency graph construction includes: Extract regional nodes from regional waterlogging data to obtain regional node data; Perform terrain semantic classification on regional node data to obtain regional classification data; The mobility weighted directional adjacent edges are constructed for the regional classification data to obtain adjacent edge data; Perform local structural similarity screening on the adjacent edge data to obtain adjacent edge screening data; The graph is constructed based on the regional node data and adjacent edge screening data to obtain the proximity graph data.

4. The method according to claim 1, wherein The water accumulation trend prediction includes: Constructing a cross-domain meta-graph for the cross-domain meta-association data to obtain cross-domain meta-graph data; Perform multi-branch trend inference on cross-domain meta-graph data to obtain multi-branch trend data; The multi-branch trend data are subjected to time series inversion verification to obtain the waterlogging trend data.

5. The method according to claim 4, characterized in that The multi-branch trend reasoning includes: Extract trend features from cross-domain meta-graph data to obtain trend feature data; Perform linear regression calculation on trend characteristic data to obtain linear water accumulation data; Perform exponential curve fitting on trend characteristic data to obtain exponential surge data; The trend feature data is processed by frequency domain convolution network to obtain periodic fluctuation data; The credibility trend fusion of linear water accumulation data, exponential surge data and periodic fluctuation data is performed to obtain multi-branch trend data.

6. The method according to claim 1, characterized in that The global flood simulation includes: Perform regional state trend processing based on the water accumulation trend data to obtain regional state trend data; Perform graph diffusion evolution deduction on regional state trend data to obtain graph diffusion evolution data; The graph diffusion evolution data is partitioned into risk layers to obtain the global flood data.

7. A flood intelligent deduction system based on global integration, characterized by: For executing the flood intelligent deduction method based on global fusion according to claim 1, the flood intelligent deduction system based on global fusion comprises: The node deployment and collection module is used to deploy ground-nailed intelligent node units in a preset area and collect regional water accumulation data through the ground-nailed intelligent node units; The graph neural perception fusion module is used to construct a geographic adjacency graph for regional waterlogging data to obtain proximity graph data; it performs local cross-domain graph attention calculation on the proximity graph data to obtain proximity perception graph data; The high-order trend evolution reasoning module is used to perform cross-domain meta-correlation perception on the neighboring perception graph data to obtain cross-domain meta-correlation data; and to perform waterlogging trend prediction on the cross-domain meta-correlation data to obtain waterlogging trend data; The global flood evolution simulation module is used to simulate global floods based on water accumulation trend data to obtain global flood data.

Images