A self-learning urban rainstorm flood disaster chain conduction path rapid prediction method and system
By processing video stream images and camera data at the edge, and combining a lightweight convolutional neural network and a physical model, the system learns the surface and underground topology networks, thus solving the problem of insufficient accuracy in predicting disaster chain propagation paths in existing technologies and achieving high-precision and real-time disaster hazard forecasting.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SUN YAT SEN UNIV
- Filing Date
- 2026-02-12
- Publication Date
- 2026-04-24
Smart Images
Figure CN121744942B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of urban flooding and hydrological application technology, and in particular to a self-learning method and system for rapid forecasting of urban rainstorm and flood disaster chain transmission paths. Background Technology
[0002] Against the backdrop of global climate change, extreme rainstorm events are occurring more frequently and their intensity is continuously escalating. Coupled with problems such as hardening of the underlying surface and insufficient carrying capacity of the pipeline network caused by the accelerated urbanization process, the suddenness and chain reaction of urban rainstorm and flood disasters are becoming increasingly prominent.
[0003] Rainstorms and floods are no longer simply surface waterlogging disasters. Instead, they spread and escalate through a chain-like path: surface runoff – underground pipe overflow – backflow into low-lying areas – traffic disruption – damage to underground spaces – paralysis of public facilities. This forms a multi-stage, multi-dimensional disaster chain, posing a serious threat to urban operational safety and the lives and property of residents. Current urban flood control and disaster reduction systems lack sufficient holistic and dynamic prediction capabilities regarding the transmission of disaster chains. Traditional forecasting models are ill-suited to the complex scenarios of rapidly evolving disaster chains. There is an urgent need to develop efficient and accurate rapid forecasting methods for disaster chain transmission paths to provide advance warning support for emergency response and risk prevention.
[0004] Existing rapid forecasting methods for urban rainstorm and flood disaster chains rely on acquiring historical rainstorm parameters of the target area at the edge and comparing them with various pre-set flood calculation schemes. However, the edge still runs pre-defined hydrological models or offline calculation schemes. These pre-set models have inherent structural deficiencies due to the lack of core data on underground pipe networks, resulting in fundamental deviations in their output even with high edge computing speeds. More importantly, due to the lack of a closed-loop correction mechanism for "visual ground truth," existing edge computing is mostly limited to processing numerical data such as rain gauge readings or simple image compression. It cannot directly calculate high-precision ground truth of water depth at the edge, thus failing to provide real-time physical residual feedback for cloud-based deep learning models. This makes it difficult for the models to dynamically optimize with the rainstorm process and become more accurate over time, ultimately leading to poor disaster forecast accuracy. Summary of the Invention
[0005] This invention provides a self-learning method and system for rapid forecasting of urban rainstorm and flood disaster chain transmission paths. It solves the technical problem that existing rapid forecasting methods for urban rainstorm and flood disaster chain transmission paths cannot provide real-time physical residual feedback for cloud-based deep learning models, ultimately resulting in poor disaster forecast accuracy.
[0006] The first aspect of this invention provides a rapid forecasting method for the transmission path of urban rainstorm and flood disaster chains that can be self-learned, comprising:
[0007] Acquire multiple consecutive frames of images from a video stream and camera intrinsic and extrinsic parameter data, and preprocess the multiple consecutive frames of images from the video stream based on a lightweight convolutional neural network to output a set of pixel coordinates of the land-water boundary line;
[0008] Using trigonometric relationships and a pinhole camera imaging model, the measured water depth is calculated based on the camera's intrinsic and extrinsic parameters and the set of pixel coordinates of the land-water boundary line, and the final measured water depth is output.
[0009] Construct a discrete surface cellular grid and a fully connected topological network, and convert the network topology and hydrological state data corresponding to the fully connected topological network into a high-dimensional tensor of state features;
[0010] Based on the discretized Manning formula and the water balance law, according to the state feature high-dimensional tensor and the fully connected topology network, the future water depth, real-time dynamic roughness and virtual connection edge conduction weight of multiple surface cells in the discrete surface cell grid are output.
[0011] A physical information composite loss function is constructed, which includes a data fitting term, a structural sparsity term, and a parameter smoothing term. When a predefined sliding time window is reached, the artificial intelligence model is updated based on the mini-batch stochastic gradient descent mechanism using the physical information composite loss function according to the final measured water depth, the future water depth of multiple surface cells, the real-time dynamic roughness, and the conduction weight of virtual connection edges. The updated artificial intelligence model is determined through multiple self-learning processes.
[0012] Based on the updated artificial intelligence model, a visualized disaster chain transmission map is output, and the intrusion forecast of potential urban rainstorm and flood disaster hazards is realized.
[0013] Optionally, the lightweight convolutional neural network includes an encoder and a decoder; the preprocessing of multiple consecutive frames of the video stream based on the lightweight convolutional neural network to output a set of pixel coordinates for the land-water boundary line includes:
[0014] Rain and fog are removed from multiple consecutive frames of the video stream to obtain multiple consecutive frames of the video stream after rain and fog removal.
[0015] Background modeling is performed on multiple consecutive frames of the video stream after rain and fog removal to obtain clear static background frames;
[0016] The clear static background frame is encoded using an encoder to obtain encoded features;
[0017] The encoded features are classified at the pixel level using a decoder based on a channel attention mechanism to obtain a binarized mask image.
[0018] The Canny edge detection operator is used to extract the set of pixel coordinates of the land-water boundary line from the binarized mask image.
[0019] Optionally, the step of using triangulation and pinhole camera imaging models to calculate the measured water depth based on the camera's intrinsic and extrinsic parameters and the set of pixel coordinates of the land-water boundary line, and outputting the final measured water depth, includes:
[0020] Based on the pinhole camera imaging model, a mapping relationship between the image pixel coordinate system and the world physical coordinate system is constructed.
[0021] Based on the mapping relationship between the image pixel coordinate system and the world physical coordinate system and the camera intrinsic and extrinsic parameters, calculate the vertical pixel offset of each water-land boundary pixel coordinate in the set of water-land boundary pixel coordinates;
[0022] Select the coordinates of multiple key points of the water-land boundary pixel in the set of water-land boundary pixel coordinates;
[0023] Using the triangular geometric relationship, based on the camera's intrinsic and extrinsic parameters and the vertical pixel offset of the coordinates of multiple key points of the water-land boundary line pixels, the physical water depth of each key point of the water-land boundary line pixels is calculated.
[0024] The physical water depth is calculated by averaging the coordinates of the key points of each pixel on the land-water boundary line to obtain the final measured water depth.
[0025] Optionally, the construction process of the fully connected topology network is as follows:
[0026] Construct a discrete surface cell grid and a discrete underground node set, wherein the discrete surface cell grid includes multiple surface cells and the discrete underground node set includes multiple underground nodes;
[0027] Virtual connection edges with initial weights of zero or minimal values are established between the surface cells and the underground nodes, as well as between the underground nodes themselves, and a fully connected topology network is constructed.
[0028] Optionally, the step of outputting the future water depth, real-time dynamic roughness, and virtual connection edge conduction weights of multiple surface cells in the discrete surface cell grid based on the discretized Manning formula and the water balance law, according to the high-dimensional tensor of the state characteristics and the fully connected topology network, includes:
[0029] Spatiotemporal encoding of high-dimensional hydraulic features is performed on the high-dimensional tensor of the state features to obtain a high-dimensional hidden state feature map;
[0030] Decoding the high-dimensional hidden state feature map yields the real-time dynamic roughness of multiple surface cells;
[0031] Based on the high-dimensional hidden state feature map, the attention coefficients of potential connection pairs in the fully connected topology network are calculated, and the attention coefficients of the potential connection pairs are converted into virtual connection edge conduction weights.
[0032] The discretized Manning formula is used to calculate the flow rate between adjacent surface cells based on the real-time dynamic roughness of multiple surface cells.
[0033] Based on the virtual connection edge conduction weight, calculate the potential underground transmission traffic;
[0034] The future water depth of each surface cell is calculated using the water balance law based on the overflow between adjacent surface cells and the potential underground transport flow.
[0035] Optionally, it also includes:
[0036] The data fitting term is used to measure the error between the future water depth and the final measured water depth.
[0037] The structural sparsity term is used to apply L1 regularization constraints to the conduction weights of virtual connection edges, forcing the weights of virtual connection edges without hydraulic connections to approach zero, thereby filtering out sparse underground conduction paths.
[0038] The parameter smoothing term is used to apply spatiotemporal continuity constraints to the real-time dynamic roughness to prevent physical parameters from undergoing drastic, non-physical changes.
[0039] Optionally, the step of outputting a visualized disaster chain transmission map based on the updated artificial intelligence model and realizing the forecast of urban rainstorm and flood disaster risks includes:
[0040] Based on the target virtual connection edge conduction weight and target global water accumulation depth distribution output by the updated artificial intelligence model;
[0041] Based on the target virtual connection edge conduction weights and the target global water depth distribution, a disaster chain transmission map is generated;
[0042] The disaster chain transmission map is visualized to obtain a visualized disaster chain transmission map;
[0043] The disaster chain cascading risk index is calculated using the target virtual connection edge conduction weight;
[0044] When the cascading disaster risk index exceeds a preset threshold, an urban rainstorm and flood disaster hazard intrusion forecast is issued.
[0045] The second aspect of this invention provides a self-learning rapid forecasting system for urban rainstorm and flood disaster chain transmission paths, comprising:
[0046] The acquisition module is used to acquire multiple consecutive frames of images from the video stream and camera intrinsic and extrinsic parameter data, and to preprocess the multiple consecutive frames of images from the video stream based on a lightweight convolutional neural network to output a set of pixel coordinates of the land-water boundary line.
[0047] The output module is used to calculate the measured water depth based on the camera's intrinsic and extrinsic parameters and the set of pixel coordinates of the land-water boundary line using triangular geometric relationships and a pinhole camera imaging model, and outputs the final measured water depth.
[0048] The construction module is used to construct discrete surface cell grids and fully connected topological networks, and to convert the network topology and hydrological state data corresponding to the fully connected topological network into high-dimensional tensors of state features;
[0049] The calculation module is used to output the future water depth, real-time dynamic roughness, and virtual connection edge conduction weight of multiple surface cells in the discrete surface cell grid based on the discretized Manning formula and the water balance law, according to the state feature high-dimensional tensor and the fully connected topology network.
[0050] The self-learning module is used to construct a physical information composite loss function that includes data fitting terms, structural sparsity terms, and parameter smoothing terms. When a predefined sliding time window is reached, based on the mini-batch stochastic gradient descent mechanism, the physical information composite loss function is used to update the artificial intelligence model according to the final measured water depth, the future water depth of multiple surface cells, the real-time dynamic roughness, and the virtual connection edge conduction weights. The updated artificial intelligence model is determined through multiple self-learning processes.
[0051] The forecasting module is used to output a visualized disaster chain transmission map based on the updated artificial intelligence model, and to realize the forecast of potential urban rainstorm and flood disaster intrusion.
[0052] A third aspect of the present invention provides an electronic device, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the self-learning rapid forecasting method for urban rainstorm and flood disaster chain transmission paths as described above.
[0053] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, it implements the self-learning rapid forecasting method for urban rainstorm and flood disaster chain transmission paths as described above.
[0054] As can be seen from the above technical solutions, the present invention has the following advantages:
[0055] The present invention provides a self-learning method for rapid forecasting of urban stormwater flood disaster chain transmission paths. It acquires multiple consecutive frames of video stream images and camera intrinsic and extrinsic parameter data, and preprocesses these images using a lightweight convolutional neural network to output a set of pixel coordinates for the land-water boundary. Using triangulation and a pinhole camera imaging model, it calculates the measured water depth based on the camera extrinsic and extrinsic parameter data and the set of pixel coordinates for the land-water boundary, outputting the final measured water depth. It constructs a discrete surface cellular grid and a fully connected topological network, and converts the network topology and hydrological state data corresponding to the fully connected topological network into a high-dimensional tensor of state features. Based on discretization... Based on the Manning formula and the water balance law, and using a high-dimensional tensor of state characteristics and a fully connected topological network, the model outputs the future water depth, real-time dynamic roughness, and virtual connection edge conduction weights for multiple surface cells in a discrete surface cell grid. A physical information composite loss function is constructed, including data fitting, structural sparsity, and parameter smoothing terms. When a predefined sliding time window is reached, based on a mini-batch stochastic gradient descent mechanism, the physical information composite loss function is used to update the artificial intelligence model according to the final measured water depth, the future water depth of multiple surface cells, the real-time dynamic roughness, and the virtual connection edge conduction weights. Through multiple self-learning iterations, the model is further refined to a higher level. The updated AI model outputs a visualized disaster chain transmission map and enables forecasting of urban rainstorm and flood hazard intrusion. Based on the above scheme, this invention preprocesses images using a lightweight convolutional neural network and extracts the land-water boundary. Combined with geometric modeling involving both inside and outside the camera, it directly calculates the true value of measured water depth at the edge, providing a real-time and reliable source of physical residual feedback of "visual truth" for the cloud-based AI model. Simultaneously, by constructing a fully connected topological network and introducing structural sparsity constraints, it achieves adaptive discovery (structural self-learning) of underground hidden transmission paths, compensating for the lack of underground pipe networks in existing pre-built models. To address the structural gaps caused by insufficient data, the model combines discretized Manning's formula and the water balance law with parameter smoothing constraints. This allows the model's output parameters, such as future water depth and real-time dynamic roughness, to have a solid physical process support and to adaptively invert with environmental changes (parameter self-learning). Furthermore, through a small-batch stochastic gradient descent update mechanism under a sliding time window, the model deeply integrates the measured true value of water depth with the multi-dimensional parameters output by the model. This enables the model to dynamically iterate and learn during the rainstorm process, ensuring that the model becomes more accurate with each calculation. This significantly improves the accuracy of disaster chain transmission path forecasts and provides precise and timely technical support for the prevention and control of urban rainstorm and flood disaster hazards. Attached Figure Description
[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0057] Figure 1 This is a flowchart illustrating the steps of a self-learning rapid forecasting method for urban rainstorm and flood disaster chain transmission paths provided in Embodiment 1 of the present invention.
[0058] Figure 2 This is a schematic diagram of a surface-underground dual-layer coupled three-dimensional network structure provided in Embodiment 1 of the present invention;
[0059] Figure 3 This is a block diagram illustrating the principle of parameter inversion and structural deduction for a physical-intelligent hybrid drive provided in Embodiment 1 of the present invention.
[0060] Figure 4 This is a flowchart illustrating a self-learning method for rapid forecasting of urban rainstorm and flood disaster chain transmission paths, provided in Embodiment 1 of the present invention.
[0061] Figure 5 This is a structural block diagram of a self-learning rapid forecasting system for urban rainstorm and flood disaster chain transmission paths, provided in Embodiment 2 of the present invention. Detailed Implementation
[0062] This invention provides a self-learning method and system for rapid forecasting of urban rainstorm and flood disaster chain transmission paths, which solves the technical problem that existing rapid forecasting methods for urban rainstorm and flood disaster chain transmission paths cannot provide real-time physical residual feedback for cloud-based deep learning models, ultimately resulting in poor disaster forecast accuracy.
[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that in the optional embodiments of the present invention, the object information and other related data involved require the permission or consent of the object when the embodiments of the present invention are applied to specific products or technologies, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. That is to say, if the embodiments of the present invention involve data related to the object, it needs to be obtained with the authorization and consent of the object, the authorization and consent of the relevant departments, and in compliance with the relevant laws, regulations, and standards of the country and region. If personal information is involved in the embodiments, the acquisition of all personal information requires the consent of the individual. If sensitive information is involved, the separate consent of the information subject is required, and the embodiments also need to be implemented with the authorization and consent of the object.
[0064] Terminology Explanation:
[0065] 1. Edge Computing
[0066] Edge computing is a technology that offloads computational tasks from a central server to the source of data generation (i.e., devices next to cameras). In this invention, it refers to calculating water depth directly on the roadside and sending the results only to the center, thereby significantly reducing transmission latency.
[0067] 2. Self-learning
[0068] In this invention, the model is characterized by its ability to automatically adjust the neural network weights, physical parameters (such as Manning roughness) and topological structure (such as underground connection weights) online using the backpropagation algorithm based on the physical residuals between the measured true values and the predicted results transmitted back from the edge end during operation, without the need for manual parameter calibration.
[0069] 3. Flood Disaster Chain
[0070] This refers to the spatiotemporal cascading process after a rainstorm, where rainwater falls from the sky to form surface runoff, which then accumulates to form surface water. This water then flows back into underground spaces (such as subways, tunnels, and parking garages) through pipe network inlets and underground gaps, ultimately causing secondary disasters such as infrastructure damage.
[0071] 4. Conduction Path / Transmission Path
[0072] This refers to the specific routes of water flow in the complex urban environment. It includes both visible two-dimensional flow paths on the surface and invisible underground water transport channels such as pipe networks, culverts, and fissures. The core of this invention lies in uncovering these hidden underground transport paths.
[0073] 5. Graph Attention Network (GAT)
[0074] A deep learning model for processing graphical data (such as social networks and molecular structures). Its core mechanism is "attention," which means that when processing a node, it can automatically determine the importance (weight) of its surrounding neighboring nodes, rather than treating them equally.
[0075] 6. Convolutional Long Short-Term Memory Network (Conv-LSTM)
[0076] This is an architecture that combines Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs). CNNs excel at interpreting images (spatial features), while LSTMs excel at recording events (temporal features). Conv-LSTM can simultaneously process spatiotemporal sequence data such as video streams.
[0077] 7. Physics-Informed Loss Function
[0078] This is a set of "reward and punishment rules" used when training an AI model (Artificial Intelligence Model). Ordinary AI only needs the correct result (data-driven), but this function requires the AI not only to produce the correct result, but also that the process conforms to physical laws (such as the conservation of energy). If the AI's calculation violates physical laws, the penalty will be particularly severe.
[0079] 8. End-to-End Gradient Backpropagation
[0080] The "error correction" mechanism in deep learning: After calculating the error between the predicted value and the actual value, the system propagates this error back layer by layer through a mathematical chain (chain rule), telling each parameter in the neural network how to adjust it.
[0081] 9. Cellular Automata (CA)
[0082] A discrete computational model. The space is divided into countless small grids (cells), and each cell updates its own state based on the state of the surrounding cells and preset rules (such as water flowing downhill).
[0083] 10. Manning Roughness Coefficient (n)
[0084] In hydraulics, n is a parameter representing the "roughness of the ground surface". The larger the value, the greater the resistance of the ground to water flow (the slower the flow). For example, the n value is very small for concrete surfaces, while the n value is very large for grass or surfaces covered with garbage.
[0085] 11. Water Balance Principle / Law of Conservation of Volume
[0086] A fundamental law of physics states that the change in water volume within a region equals the amount flowing in minus the amount flowing out, plus the amount of rainwater falling from the sky minus the amount seeping into the ground. Water cannot be created or destroyed out of thin air.
[0087] 12. Monocular Visual Geometric Projection
[0088] A computer vision technology. Using only one camera (monocular), and combining the camera's installation height, angle, and focal length, it uses geometric principles to calculate the actual height or distance of objects (such as the water surface) in the image.
[0089] 13. NVIDIA Jetson Orin
[0090] The NVIDIA Jetson Orin is a high-performance embedded AI computing module. It is small in size but powerful in computing, and is specifically designed to support deep learning inference at the edge. It is the hardware foundation for the "monitoring and solving" functionality of this invention.
[0091] 14. Encoder and Decoder
[0092] Deep learning networks (such as U-Net) have two main components: an encoder, which compresses an image into abstract features (feature extraction), and a decoder, which restores these abstract features into pixel-level classification results (image restoration). This invention uses MobileNetV3 as the encoder (for lightweight and fast processing) and an upsampling layer as the decoder, together completing the conversion from surveillance video footage to a water accumulation mask.
[0093] 15. DEM (Digital Elevation Model)
[0094] A "DEM" is a "digital topographic map" of the ground. It uses a grid of grids with elevation values to digitally represent the undulations of the earth's surface, and is the basic data for computers to identify terrain elevation and determine the direction of water flow.
[0095] 16. BIM (Building Information Modeling)
[0096] "BIM" stands for "digital perspective view" of a building. Unlike maps that only show the surface, BIM contains detailed three-dimensional geometric information such as underground utility tunnels, the internal structure of subway stations, and tunnel routes. This invention uses it to construct a detailed model of underground space.
[0097] 17. Dark Channel Prior (DCP)
[0098] "Dark channel prior" is an image dehazing algorithm based on statistical laws. It uses the rule that "in some areas of a hazy image, at least one color channel pixel value is very low" to infer the fog concentration and remove it, making the hazy, white image of a rainy day clear and transparent.
[0099] 18. Gaussian Mixture Model (GMM)
[0100] Gaussian Mixture Model (GMMM) is a video background modeling algorithm, commonly known as a "spot the difference" algorithm. It can learn from long-term footage to distinguish between stationary road surfaces (background) and moving vehicles and rain lines (foreground), preventing moving cars from being mistaken for standing water.
[0101] 19. GPU Parallel Computing (Graphics Processing Unit)
[0102] "GPU parallel computing" is a technology that utilizes the simultaneous work of tens of thousands of cores in a graphics card. A traditional CPU is like a math professor—it calculates quickly but can only perform calculations one by one; a GPU is like thousands of elementary school students doing addition and subtraction simultaneously. This invention uses it to simultaneously calculate the water flow changes of millions of cells across an entire city, achieving extremely rapid simulation.
[0103] 20. MobileNetV3-UNet
[0104] "MobileNetV3-UNet" is a lightweight deep learning network architecture. MobileNetV3 is responsible for "speed" (low computational cost), and UNet is responsible for "accuracy" (fine feature extraction). This combination is designed to run on edge devices with limited computing power.
[0105] 21.DOM (Digital Orthophoto Map)
[0106] "Digital orthophotos" refers to geometrically corrected aerial or satellite imagery that eliminates perspective distortion and possesses the geometric accuracy of a map. This invention utilizes it in conjunction with a DEM to construct a physical grid of the Earth's surface.
[0107] 22.IPM (Inverse Perspective Mapping)
[0108] "Inverse perspective transformation." A computer vision geometry algorithm used to eliminate perspective distortion (objects appear larger when closer and smaller when farther away) in images, converting a slanted camera view into a top-down planar image to accurately measure actual physical distances and water depths in monocular vision.
[0109] 23. RTSP / ONVIF (Real-Time Streaming Protocol / Open Network Video Interface Forum, a standard protocol for network video transmission and device interfaces)
[0110] This invention utilizes these common protocols to enable edge gateways to seamlessly connect to various existing security or traffic surveillance cameras in the city without the need for hardware replacement.
[0111] 24. CFL Condition (Courant-Friedrichs-Lewy Condition)
[0112] A stability condition in numerical computation. In cellular automata computation, it stipulates that the time step (Δt) cannot be too large, and it must be ensured that the water flow does not flow out of the grid within one step; otherwise, the calculation will diverge (error).
[0113] 25.SWMM (Storm Water Management Model)
[0114] "Storm Flood Management Model". This classic dynamic rainfall-runoff simulation model, developed by the U.S. Environmental Protection Agency, is often used as a representative of traditional physical models in the prior art for comparison with this invention.
[0115] 26. Horton Infiltration Equation
[0116] A classic empirical formula is used to calculate the rate at which soil absorbs water. Water absorption is rapid immediately after rain, but as the soil becomes saturated, the absorption rate decreases exponentially.
[0117] 27. Topological Structure
[0118] In mathematics, it's a discipline that studies the "connections" between objects, without concern for specific shapes and sizes. In a pipeline network, it refers to "which manhole cover is connected to which pipe, and where the pipe leads."
[0119] 28. Moving Horizon Estimation (MWE)
[0120] An online optimization strategy. Instead of calculating all the data at once, it uses a "sliding window" approach, updating the model parameters every few minutes with data from a short period of time (e.g., the last 15 minutes).
[0121] 29. Surface-Subsurface Coupled Network
[0122] A modeling architecture. The top layer is a grid representing the ground, and the bottom layer is nodes representing underground facilities (pipelines, subways). Data is exchanged between the two layers via virtual connections.
[0123] 30. Ground Truth
[0124] In machine learning, this refers to the "true standard answer." In this invention, it specifically refers to the actual physical water depth (hobs) calculated by an edge computing gateway using a monocular vision algorithm. It is the "standard answer" used to train and calibrate cloud-based AI models, and also the benchmark for calculating physical residuals.
[0125] 31. Binary Mask
[0126] A black and white image, where white represents the target area and black represents the background. In this invention, it refers to the output of a semantic segmentation network. All pixels belonging to "water accumulation" are marked as 1 (white), and pixels belonging to "road surface / vehicles" are marked as 0 (black), used to accurately extract the coverage area of the water accumulation.
[0127] 32. SE Channel Attention Mechanism Module (Squeeze-and-Excitation Module)
[0128] An embedded component of a neural network automatically learns the importance of each feature channel, "squeezing out" unimportant signals and "exciting" important signals. In this invention, it is used to enhance the model's ability to identify the edges of turbid water accumulation and suppress noise interference caused by uneven lighting or reflections during rainy days.
[0129] 33. Canny Edge Detector
[0130] A classic image processing algorithm is used to detect edges with drastic brightness changes in an image. In this invention, it is used to extract a closed, continuous water-land boundary line from a binarized mask image, which is a crucial geometric basis for subsequent water depth calculations.
[0131] 34. Waterline
[0132] In this invention, the term refers to the boundary line between the surface of accumulated water and the non-accumulated road surface (or reference object). By calculating the change in the vertical position of this line in the image and combining it with the principles of geometric projection, the physical depth of the accumulated water can be deduced.
[0133] 35. Intrinsic Matrix and Extrinsic Matrix
[0134] Mathematical parameters used to describe a camera in computer vision.
[0135] Internal parameters: describe the camera's own characteristics (such as focal length and optical center), which determine how light forms an image.
[0136] External parameters: describe the camera's attitude in world space (such as installation height and pitch angle).
[0137] In this invention, two matrices are used to establish a mapping relationship between "image pixel coordinates" and "real-world physical coordinates", thereby converting the "pixel distance" on the screen into "meters" in reality.
[0138] 36. Physical Layer
[0139] In this invention, the upper layer refers to the structure of the two-layer coupled network. It is a two-dimensional regular grid (cell) generated by discretizing DEM data, representing the real urban surface. This layer follows explicit physical flow rules (Manning's formula).
[0140] 37. Hidden Layer
[0141] In this invention, the lower layer structure refers to the two-layer coupled network. It consists of a set of virtual nodes with unknown locations, used to simulate underground pipe networks, culverts, or stormwater storage tanks. The structure of this layer is "implicit" and needs to be revealed through AI learning.
[0142] 38. Known Nodes
[0143] In this invention, it refers to underground facility nodes (such as specific storm drains or subway entrances) that are explicitly recorded in GIS / BIM (Geographic Information System / Building Information Model) data. The locations and attributes of these nodes are fixed during model initialization.
[0144] 39. Dark Nodes / Shadow Nodes
[0145] In this invention, "blank placeholder nodes" are automatically pre-positioned by the system in areas lacking pipeline network data. Initially, these nodes have no volume and are not interconnected, but are reserved for activation. Once the AI detects an abnormal disappearance of surface water, it will activate these nodes to act as "underground water storage containers."
[0146] 40. Learnable Interfaces
[0147] In this invention, virtual channels with variable connection states are pre-defined in the model. They are like countless "electronic valves" that are closed by default, and their opening and closing degree is controlled by learnable weights (w).
[0148] 41. Physical Links
[0149] In this invention, it refers to the channel connecting "surface cells" and "known nodes". It represents a real-world rainwater grate or well opening, and its conductivity weight is 1 (enabled) by default.
[0150] 42. Latent Links
[0151] In this invention, it refers to a virtual channel connecting "surface cells" and "dark nodes," or between "dark nodes." It represents an unknown culvert or fissure. Its initial weight is close to 0 (off), and it only increases (activated) when the AI determines that a hydraulic connection exists.
[0152] 43. Physics-Intelligent Tensor Mapping
[0153] In this invention, discrete physical grid attributes (such as water depth h, rainfall P, and terrain Z) are transformed into multidimensional arrays (tensors X) that can be read by deep learning models. t This process involves building a data bridge between the "physical world" and the "AI neural network."
[0154] 44. Structure Learning
[0155] This invention specifically refers to the process of using a Graph Attention Network (GAT) to analyze the correlation of water level changes, thereby calculating the connection weights w. In other words, it's the process by which AI automatically "draws" an underground pipe network map.
[0156] 45. Parameter Learning
[0157] This invention specifically refers to the process of using Conv-LSTM to analyze the spatiotemporal characteristics of the flow field, thereby retrieving the dynamic Manning roughness n. In other words, it refers to the process by which AI automatically "senses" the degree of road congestion.
[0158] 46. Composite Loss Function
[0159] In this invention, L refers to the overall objective function used to train AI. total It includes not only "accuracy of prediction" (data fitting term), but also "correctness of structure" (sparseness constraint) and "smoothness of parameters" (smoothness constraint), guiding the model to converge in a direction that conforms to physical common sense.
[0160] 47. Online Learning
[0161] One machine learning strategy involves a model that is not static after training but is continuously fine-tuned by receiving new data during operation. In this invention, the system continuously updates the network weights using real-time ground truth values transmitted every minute during a rainstorm, making the model increasingly accurate as the rainfall continues.
[0162] 48. Automatic Differentiation
[0163] The core technology of deep learning frameworks (such as PyTorch / TensorFlow) can automatically calculate the gradient of complex functions. This invention utilizes this technology to treat physical equations (such as Manning's formula and water balance) as differentiable computational graph nodes, allowing error gradients to pass through the physical formulas and be fed back to the neural network, thus achieving "physics-driven AI training".
[0164] 49.SGD (Stochastic Gradient Descent)
[0165] "Stochastic gradient descent" is a deep learning optimization algorithm used to update the model's weight parameters based on the calculated gradient direction during backpropagation, thereby continuously reducing prediction error.
[0166] 50. The cloud refers to a high-performance computing cluster (HPC) or cloud server, which is mainly responsible for running large-scale parallel cellular automata models.
[0167] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a self-learning rapid forecasting method for urban rainstorm and flood disaster chain transmission paths, as provided in Embodiment 1 of the present invention.
[0168] This invention provides a rapid forecasting method for the transmission path of urban rainstorm and flood disaster chains that can learn independently, comprising:
[0169] Step 101: Obtain multiple consecutive frames of images from the video stream and camera intrinsic and extrinsic parameter data, and preprocess the multiple consecutive frames of images from the video stream based on a lightweight convolutional neural network to output a set of pixel coordinates of the land-water boundary line.
[0170] A video stream consists of multiple consecutive frames of images captured in real time by an image acquisition device and arranged in a time sequence. It is a fundamental data source for image preprocessing and water-land boundary extraction.
[0171] Camera intrinsic and extrinsic parameters refer to parameters such as focal length and principal point coordinates that reflect the camera's own imaging characteristics, and parameters such as mounting height, pitch angle, and rotation / translation matrix that reflect the relationship between the camera and the world's physical coordinate system. These are the key basis for mapping pixel coordinates to physical coordinates.
[0172] It should be noted that the video stream first undergoes basic processing such as frame selection and noise reduction on multiple consecutive frames. Then, it is input into a lightweight convolutional neural network to complete feature extraction and pixel-level segmentation of the land and water areas, accurately locating the land-water boundary. Finally, it outputs a closed and continuous set of pixel coordinates of the land-water boundary line, providing core coordinate input for subsequent calculation of the measured water depth.
[0173] Further, step 101 may include the following sub-steps:
[0174] S11. Remove rain and fog from multiple consecutive frames of the video stream to obtain multiple consecutive frames of the video stream after rain and fog removal.
[0175] S12. Perform background modeling on multiple consecutive frames of the video stream after removing rain and fog to obtain clear static background frames;
[0176] S13. Encode the clear static background frame using an encoder to obtain the encoded features;
[0177] S14. The decoder uses a channel attention mechanism to perform pixel-level classification of the encoded features to obtain a binarized mask image.
[0178] S15. The Canny edge detection operator is used to extract the set of pixel coordinates of the land-water boundary line in the binary mask image.
[0179] Lightweight convolutional neural networks consist of an encoder and a decoder.
[0180] It should be noted that this invention constructs an "edge-side visual perception" system, which not only solves the problems of high bandwidth consumption and poor timeliness in traditional modes, but also provides high-precision "observation ground truth," i.e., the final measured water depth, for subsequent deep learning cellular models, forming the foundation for realizing a closed loop of intelligent forecasting. This invention deploys embedded edge computing gateways (using edge computing devices based on the NVIDIA Jetson Orin platform) at the front-end monitoring nodes of urban flood-prone areas, achieving real-time quantization and structured output of water depth through "edge-side inference." The specific implementation process includes four sub-steps: video stream access and enhancement, water body semantic segmentation, monocular visual depth calculation, and structured data transmission.
[0181] First, the edge computing gateway connects to existing roadside security or traffic cameras via the RTSP / ONVIF standard protocol to acquire real-time video streams (i.e., continuous multi-frame images) with a resolution of at least 1920×1080 and a frame rate of at least 25fps. To address the issues of blurred images and low contrast caused by insufficient light and rain / fog during heavy rain, image enhancement algorithms are deployed in the gateway's ISP (Image Signal Processing) pipeline. The Dark Channel Prior (DCP) algorithm is used to remove fog and improve clarity. A Gaussian Mixture Model (GMM) is used to model the background of the continuous multi-frame images, filtering out interference from dynamic rain streaks and moving vehicles, and extracting clear static background frames as the basic input for subsequent water accumulation identification.
[0182] Furthermore, an improved MobileNetV3-UNet lightweight convolutional neural network model is run within the GPU (Graphics Processing Unit) of the edge gateway. This model uses MobileNetV3 as the encoder, leveraging its depthwise separable convolutional layers to extract image features, significantly reducing computational cost. A Squeeze-and-Excitation (SE) channel attention mechanism is introduced at the skip connections of the decoder to automatically weight key feature channels, enhancing the model's accuracy in recognizing the edges (blurred boundaries) of murky water. The preprocessed image is input into the network for edge-side inference. The model outputs a pixel-level binarized mask, accurately classifying each pixel in the image as either a "waterlogged area (foreground)" or a "non-waterlogged area (background)." The Canny edge detection operator is applied to the mask to extract a closed, continuous set of pixel coordinates for the waterline boundary.
[0183] The edge computing gateway mentioned in this invention is an embedded edge computing device with GPU computing power, which in actual engineering usually refers to the NVIDIA Jetson Orin / Xavier series embedded AI computing box.
[0184] In this embodiment, the Dark Channel Prior (DCP) algorithm is used to remove rain and fog from multiple consecutive frames of the video stream, eliminating the interference of rain and fog on image quality and improving image clarity, resulting in a series of consecutive frames of the video stream after rain and fog removal. Based on this, a Gaussian Mixture Model (GMM) is used to model the background of the rain and fog-removed images, removing dynamic interference factors and preserving stable scene background information, resulting in clear static background frames. Subsequently, an encoder extracts and compresses features from the clear static background frames to obtain encoded features containing key scene information. A decoder then uses a channel attention mechanism to specifically weight the encoded features, strengthening the representation ability of features related to water accumulation areas, thereby completing pixel-level classification and outputting a binary mask map that can distinguish between water and land areas. Finally, the Canny edge detection operator is used to accurately locate the water-land boundary region in the binary mask map and extract a set of closed and continuous water-land boundary line pixel coordinates. This invention, through progressive image preprocessing and feature extraction operations, effectively improves the accuracy of water-land boundary recognition, providing high-quality core data support for subsequent calculations of measured water depth.
[0185] Step 102: Using trigonometric geometry and a pinhole camera imaging model, the measured water depth is calculated based on the camera's intrinsic and extrinsic parameters and the set of pixel coordinates of the land-water boundary line, and the final measured water depth is output.
[0186] It should be noted that, firstly, based on the pinhole camera imaging model, the mapping relationship between the image pixel coordinate system and the world physical coordinate system is established by combining the camera's intrinsic and extrinsic parameter data. Then, the pixel coordinate set of the water-land boundary line is associated, and the physical spatial position corresponding to each boundary pixel is calculated through trigonometric relationships. In this way, the actual water depth is calculated, and finally, the accurate final measured water depth is output.
[0187] Furthermore, step 102 may include the following sub-steps:
[0188] S21. Based on the pinhole camera imaging model, construct the mapping relationship between the image pixel coordinate system and the world physical coordinate system;
[0189] S22. Based on the mapping relationship between the image pixel coordinate system and the world physical coordinate system and the camera intrinsic and extrinsic parameters, calculate the vertical pixel offset of each water-land boundary pixel coordinate in the water-land boundary pixel coordinate set.
[0190] S23. Select multiple key point coordinates of water-land boundary line pixels from the set of water-land boundary line pixel coordinates;
[0191] S24. Using trigonometric relationships, based on camera intrinsic and extrinsic parameters and the vertical pixel offset of multiple key point coordinates of land and water boundary pixels, calculate the physical water depth of each key point coordinate of land and water boundary pixels.
[0192] S25. Calculate the average value of the physical water depth for the key point coordinates of each water-land boundary pixel to obtain the final measured water depth.
[0193] The pinhole camera imaging model refers to the geometric model constructed by simulating the imaging principle of a pinhole camera. It is used to establish the corresponding mapping relationship between the image pixel coordinate system and the world physical coordinate system, and is the core model for realizing the transformation from pixel coordinates to physical coordinates.
[0194] Trigonometric relationships refer to the mathematical principles of geometric operations and position calculations based on the correspondence between the sides and angles of a triangle. They are the core calculation basis for converting vertical pixel offsets into physical water depth.
[0195] It should be noted that, since ordinary surveillance cameras lack depth sensors, this invention utilizes the principle of Monocular Vision Inverse Perspective Mapping (IPM) to calculate the physical water depth in conjunction with pre-calibrated reference objects.
[0196] Furthermore, in the initial stage of system deployment, parameter calibration is performed on each camera. For intrinsic parameter calibration: obtain the camera's intrinsic parameter matrix K (including focal length f). x , f y Principal point coordinates cx , c y For external parameter calibration: obtain the camera's mounting height H relative to the road surface. cam The pitch angle θ and the rotation and translation matrix [R|T]. For reference object calibration: Select a fixed reference object with a known physical height (such as a curb, bollard, or ruler) in the image as the reference and record the physical coordinates of its base.
[0197] Furthermore, using a pinhole camera imaging model, an image pixel coordinate system (u,v) and a world physical coordinate system (X) are established. w ,Y w Z w The mapping relationship between ( ) is assumed to be a flat plane (i.e., Z). w =0), for any point p(u,v) on the land-water boundary detected in the image, its corresponding physical depth D and actual water depth h obs Solve the problem using the following steps:
[0198] Pixel-to-physical coordinate transformation equation:
[0199] ;
[0200] in, Represents the homogeneous coordinates in the image pixel coordinate system, where u and v are the horizontal and vertical pixel coordinates of the water accumulation edge points on the image; The depth value of the target point in the camera coordinate system (distance along the optical axis, in meters); This is the camera's intrinsic matrix, which includes the focal length (f). x , f y ) and the center point of the optical axis (c x , c y ), used to describe the internal optical characteristics of a camera, in the form of ; For camera extrinsic parameters (rotation and translation matrices), the rotation matrix R (containing elements) to ) and translation vector T (including t) x ,t y ,t z It consists of ) and is used to describe the camera's mounting attitude (pitch, yaw, altitude) relative to the world coordinate system. Let Z be the homogeneous coordinates in the world physical coordinate system, where Z w It is usually set to 0 (assuming the road surface is the reference plane).
[0201] Furthermore, based on the vertical pixel offset Δv of the land-water boundary line on the reference object, combined with the camera pitch angle θ and installation height H... camThe depth of water accumulation, i.e., the physical depth h of the water-land boundary pixel key point coordinates, is calculated using trigonometric geometric relationships. obs :
[0202] ;
[0203] in, The calculated actual physical depth of water accumulation (unit: meters). This represents the vertical pixel offset (in pixels) of the land-water boundary line relative to the base reference in the image. The depth distance of the reference base in the camera coordinate system (unit: meters). The pitch angle of the camera (unit: radians or degrees), which is the angle between the lens optical axis and the horizontal plane. This is the vertical focal length (in pixels) in the camera intrinsic parameter matrix. The angle of incidence of light (in radians) corresponding to a pixel in the image is calculated from the pixel coordinates.
[0204] To eliminate single-point calculation errors, the system selects multiple key points on the land-water boundary line to calculate the water depth separately, and takes the average value as the final measured water depth of the monitoring point.
[0205] In this embodiment, based on the geometric imaging principle of the pinhole camera imaging model, a precise mathematical mapping relationship between the image pixel coordinate system and the world physical coordinate system is first constructed, clarifying the conversion rules from pixel coordinates to physical space coordinates. Then, based on this mapping relationship and camera intrinsic and extrinsic parameter data, the vertical pixel offset corresponding to each pixel coordinate in the water-land boundary pixel coordinate set is calculated in the image. This offset reflects the difference in the vertical position of the pixel point and the camera optical axis, providing a key intermediate parameter for subsequent physical depth calculation. Next, multiple water-land boundary pixel key point coordinates with uniform distribution and significant features are selected from the water-land boundary pixel coordinate set to avoid the random error of single-point calculation. Then, using trigonometric geometric relationships, combined with camera intrinsic and extrinsic parameter data and the vertical pixel offset of each key point, the physical water depth corresponding to each key point is obtained through geometric deduction. Finally, the average value of the physical water depth of all key points is calculated to eliminate the random error of single-point measurement and output a stable and reliable final measured water depth. This invention achieves the calculation of the true value of water depth at the edge by constructing a precise coordinate mapping, screening key points and combining triangular geometric relationships for calculation, and then calibrating the mean to obtain the final measured water depth. This provides a real-time and reliable physical residual feedback basis for cloud-based artificial intelligence models.
[0206] It's worth noting that after completing the above calculations, the edge gateway does not send back the original video stream, but instead generates a lightweight JSON message. The message content is defined as: {DeviceID, Timestamp, WaterDepth (hobs),FloodedAreaCoordinates}. This message is sent to the central system in real time via the 4G / 5G / NB-IoT network. This data is not only used for real-time alarms, but also serves as crucial "ground truth" input into the subsequent "deep learning-cell coupling model." The central system uses these measured ground truths to calculate the prediction residuals of the physical model, thereby driving the deep neural network to perform backpropagation, achieving the identification of the disaster chain propagation path and adaptive parameter correction.
[0207] Step 103: Construct a discrete surface cell grid and a fully connected topological network, and convert the network topology and hydrological state data corresponding to the fully connected topological network into a high-dimensional tensor of state features.
[0208] It should be noted that, firstly, the discrete surface cell grid is divided according to the topographic features of the study area, and the spatial range and adjacency relationship of each cell are clarified. Based on the adjacency relationship, a fully connected topological network is constructed to represent the hydraulic conduction relationship between cells. Then, the network topology information of the fully connected topological network and the hydrological state data are integrated, and after feature encoding and dimension normalization, they are converted into a high-dimensional state feature tensor suitable for subsequent model calculations.
[0209] The process of constructing a fully connected topology network may include the following sub-steps:
[0210] S31. Construct a discrete surface cell grid and a discrete underground node set, wherein the discrete surface cell grid includes multiple surface cells and the discrete underground node set includes multiple underground nodes.
[0211] S32. Establish virtual connection edges with initial weights of zero or minimal values between surface cells and underground nodes, and between underground nodes themselves, and construct a fully connected topology network.
[0212] Discrete surface cell grids refer to multiple non-overlapping surface unit grids formed by dividing the surface topography of the study area according to preset rules. They are the basic carrier for characterizing the hydrological response characteristics of different areas of the surface and simplifying the calculation of hydraulic processes.
[0213] A fully connected topology network refers to a network structure that covers both surface and underground areas, connecting nodes (surface cells and underground nodes) through virtual connection edges, and can fully represent the hydraulic conduction paths and relationships between nodes.
[0214] Network topology refers to the layout and connection pattern of nodes (surface cells and underground nodes) and virtual connecting edges in a fully connected topological network. It reflects the structural characteristics of the network and provides a structural basis for the analysis of hydraulic transmission paths.
[0215] Hydrological status data refers to key data that characterizes the surface and groundwater hydrological conditions of the study area, including parameters such as water depth, water level, and flow rate. It is the core input data that reflects the dynamic changes in hydrology.
[0216] Discrete underground node set refers to a set composed of key locations such as underground pipeline manholes and junctions, based on the distribution of underground pipeline networks. Each underground node corresponds to a core connected location of the underground pipeline network and is used to characterize the connectivity status of the underground pipeline network.
[0217] A surface cell refers to a single unit in a discrete surface cell grid. It is the basic unit for calculating hydrological processes in a surface region and can independently characterize the hydrological response characteristics of that region.
[0218] An underground node refers to a single node in a discrete set of underground nodes. It corresponds to a key connection point in the underground pipeline network, is used to connect the conduction path of the underground pipeline network, and characterizes the hydraulic connectivity of the underground area.
[0219] Virtual connection edges refer to the virtual links established between surface cells and underground nodes, and between underground nodes themselves. They are used to characterize the hydraulic transmission relationship between different nodes and are the core link carrier for constructing a fully connected topology network.
[0220] It should be noted that, as Figure 2 As shown, this step forms the physical foundation for "path self-learning." Addressing the limitations of traditional hydrodynamic models, which rely on high-precision underground pipe network data and have fixed topologies, making them unable to simulate unknown underground conduction paths, this invention constructs a two-layer coupled network of "surface physical grid + underground virtual nodes" within the central system. This architecture pre-configures numerous "virtual connection interfaces" that can be activated by deep learning models (such as Conv-LSTM), transforming discrete physical urban spaces into a computational graph structure that can be controlled by AI in real time. The specific implementation process includes four sub-steps: surface physical grid discretization, underground virtual node mapping, pre-configuration of potential conduction interfaces, and physical-intelligent tensor mapping.
[0221] First, based on a high-precision DEM (Digital Elevation Model) and DOM (Digital Orthophoto Map), the continuous urban surface is discretized into a regular two-dimensional cell matrix. Set the grid spatial resolution to × (In this embodiment, the size is 2m × 2m), covering the entire computational domain.
[0222] For any surface cell Extract and bind the following static properties, which remain unchanged during the derivation process: 1) Center elevation 1) Initial potential energy of surface gravity flow, directly sampled from DEM data. 2) Baseline permeability ( The value is determined based on the underlying surface material (e.g., asphalt, grass, permeable bricks) and serves as the baseline value for physical infiltration calculation. Here, i is the row index and j is the column index.
[0223] For the real-time dynamic roughness at time t Initialize to standard Manning roughness (e.g., 0.013 for asphalt pavement). However, in this invention, this parameter is defined as a "time-varying state variable," with a reserved data interface to receive real-time correction coefficients from the deep learning model to simulate the accumulation of road debris or the water-blocking effect of vehicles during heavy rain.
[0224] For underground spaces, this invention abandons the high-cost full 3D voxel modeling and instead constructs a discrete "set of underground nodes S". sub “”, used to carry out undercurrent exchange.
[0225] Furthermore, this invention utilizes GIS / BIM data to map identified underground space entrances (including storm drains, subway entrances / exits, and underground parking garage ramps) as independent underground Node cells. k Record its physical properties: 1) Elevation of the base plate 1) Determines the potential energy of the groundwater level. 2) Storage capacity. : Determines the water storage capacity of the node.
[0226] In areas identified by DEM analysis as low-lying and flood-prone locations but lacking pipeline network data, the system automatically pre-defines several virtual water storage nodes. shadow These nodes are initially disconnected and have a volume of 0. They are pre-embedded "containers" in the system, to be activated when the AI identifies hidden underground storage spaces (such as culverts or cavities) in subsequent steps. "Virtual links" are established between surface cells and underground nodes, as well as between underground nodes, to construct a potentially fully connected topology network.
[0227] It is worth mentioning that this invention defines a connecting edge. For a unidirectional flow channel from cell u to cell v, its conductivity is controlled by weight w:
[0228] 1) Type A: Physical Links
[0229] Object: Vertical connection between a surface cell and its corresponding known storm drain node.
[0230] Initialization: Initial conduction weights are set to... =1.0.
[0231] Physical meaning: Represents a drainage channel that is confirmed to exist on the design drawings, and is enabled by default.
[0232] 2) Type B: Latent Links
[0233] Target: Surface depression cells Subsurface dark nodes (simulating unknown infiltration / backflow fractures).
[0234] underground nodes Underground nodes (simulating unknown horizontal pipe networks).
[0235] Initialization: Initial conduction weights are set to... ( 0 (default is off).
[0236] Learning mechanism: The weights w of all types B are labeled as "trainable parameters". In subsequent simulations, the deep neural network will dynamically increase the value of w on specific edges based on the spatiotemporal anomaly pattern of surface water accumulation, thereby "activating" an unknown underground disaster transmission path in the physical model.
[0237] Furthermore, in order for the physical cells to be directly driven by the embedded deep learning model (such as Conv-LSTM) of the central system, the above spatial structure needs to be transformed into a computer-readable high-dimensional tensor format. Specifically, a structure of shape (M, N, C) is constructed. in The feature map (i.e., the high-dimensional tensor of state features) is used as the input to the AI model (artificial intelligence model) at time t. Channel C in Include:
[0238] Channel 1: Current surface water depth field H t .
[0239] Channel 2: Cumulative Rainfall Distribution P t .
[0240] Channel 3: Topographic Gradient Field (Reflecting the orientation of the terrain).
[0241] Channel 4: Historical water depth variation rate (Reflects the trend of water accumulation).
[0242] For the output of the AI model, construct a shape of (M, N, C) out The output control graph of channel C{out} is used to control the physical cells in reverse.
[0243] Channel 1: Roughness Correction Coefficient Diagram ( Used for real-time updates of the physical layer. .
[0244] Channel 2: Potential Interface Activation Threshold Map This is used to determine whether to activate the underground virtual connection in a certain area (i.e., when the AI determines that a hidden path exists there). (Increases, driving the w value to increase).
[0245] In this embodiment, the study area is first divided into discrete surface cell grids and discrete underground node sets based on the surface topography and underground pipe network distribution. The discrete surface cell grids divide the surface into multiple non-overlapping surface cells according to a preset size, which are used to represent the hydrological response units of different areas of the surface. The discrete underground node sets select key locations such as manholes and junctions of the underground pipe network as multiple underground nodes, which are used to represent the core connected nodes of the underground pipe network. Then, based on the hydraulic conduction correlation logic between the surface and underground, virtual connection edges are established between surface cells and corresponding underground nodes (reflecting the hydraulic exchange relationship between surface water and underground pipe network) and between underground nodes (reflecting the connectivity and conduction relationship of underground pipe network). This constructs a fully connected topological network that covers the surface-underground coupling area and fully represents the hydraulic conduction path. Then, the topological structure information of this network and hydrological state data are integrated and converted into a high-dimensional tensor of state features suitable for subsequent model calculations. This invention, by incorporating underground nodes and virtual connecting edges, compensates for the shortcomings of existing methods due to the lack of underground pipeline data in the model structure. This enables the fully connected topology network to completely depict the basic architecture of disaster chain transmission between the surface and underground, providing comprehensive structural support for subsequent accurate calculation of parameters such as water depth and dynamic roughness.
[0246] Step 104: Based on the discretized Manning formula and the water balance law, and according to the state characteristic high-dimensional tensor and fully connected topology network, output the future water depth, real-time dynamic roughness, and virtual connection edge conduction weights of multiple surface cells in the discrete surface cell grid.
[0247] It should be noted that, firstly, a surface cell hydraulic conduction calculation model is constructed based on the discretized Manning formula. The water balance law is combined to clarify the water balance logic among surface cells. Then, the network topology and hydrological state information contained in the high-dimensional tensor of state features are used as the core input. Relying on the node association and conduction path depicted by the fully connected topological network, the future water depth of multiple surface cells is solved through real-time calculation, and the real-time dynamic roughness related to the surface roughness is dynamically updated. At the same time, the hydraulic conduction capacity of each virtual connection edge is quantified, and finally, the conduction weight of the virtual connection edge is output.
[0248] Furthermore, step 104 may include the following sub-steps:
[0249] S41. Spatiotemporal encoding of high-dimensional hydraulic features is performed on the high-dimensional tensor of state features to obtain a high-dimensional hidden state feature map.
[0250] S42. Decode the high-dimensional hidden state feature map to obtain the real-time dynamic roughness of multiple surface cells;
[0251] S43. Calculate the attention coefficients of potential connection pairs in a fully connected topological network based on the high-dimensional hidden state feature map, and convert the attention coefficients of potential connection pairs into virtual connection edge conduction weights.
[0252] S44. Using the discretized Manning formula, calculate the flow rate between adjacent surface cells based on the real-time dynamic roughness of multiple surface cells.
[0253] S45. Calculate the potential underground transmission traffic based on the virtual connection edge conduction weight;
[0254] S46. Using the water balance law, calculate the future water depth of each surface cell based on the overflow between adjacent surface cells and the potential underground transport flow.
[0255] It should be noted that this step executes a "Physics-AI Hybrid" deduction loop within the high-performance computing unit of the central system. Its core technical logic lies in breaking the constraints of traditional hydrodynamic models' "fixed structure and static parameters," employing a deep learning model to take over the uncertainties in the physical model in real time. Specifically, the AI model does not directly output the water level at the next moment (to avoid potential physical non-conservation phenomena from black-box models), but instead predicts the "dynamic resistance parameters" and "hidden connection weights" in the physical model by analyzing spatiotemporal characteristics. Subsequently, these AI-output control variables are substituted into rigorous fluid dynamics equations, achieving both automatic discovery of unknown underground paths and ensuring that the simulation process strictly adheres to the law of mass conservation. The specific implementation process includes five detailed sub-steps: high-dimensional spatiotemporal feature encoding, online dynamic roughness inversion, graph attention-based path activation, hybrid flux calculation, and global state update.
[0256] First, the generated high-dimensional tensor X of the state features is... t The data (including current water depth, cumulative rainfall, topographic gradient field, historical rate of change, etc.) are input into a Conv-LSTM (Convolutional Long Short-Term Memory Network). This network combines the spatial feature extraction capability of a Convolutional Neural Network (CNN) with the temporal series memory capability of a Long Short-Term Memory Network (LSTM), enabling it to capture the deep-seated patterns of water accumulation evolution.
[0257] Let the input tensor at time t be X.t The hidden state at the previous moment was H. t-1 The Conv-LSTM unit filters and memorizes information flow through a gating mechanism (input gate, forget gate, output gate). Its computation process is described as follows:
[0258] ;
[0259] in, The output at time t is the high-dimensional hidden state feature map (Hidden State) with size (M, N,Chidden). The convolutional kernel weights and bias terms contained in the network are the knowledge carriers that the model learns through training.
[0260] It is worth noting that while Conv-LSTM is a preferred solution for extracting the spatiotemporal features of surface water flow, it is not the only one. In this implementation, 3D-CNN (3D Convolutional Neural Network) and ConvGRU (Convolutional Gated Recurrent Unit) can be used as alternatives. 3D-CNN treats the time dimension as a third spatial dimension, directly performing 3D convolution operations on the (T,H,W) volume of water accumulation data to extract water flow evolution features and use them to invert the dynamic roughness n. ConvGRU, as a simplified variant of LSTM, not only possesses the same spatiotemporal modeling capabilities but also has fewer computational parameters, enabling faster inference speeds. Both solutions can provide high-dimensional feature inputs for subsequent physical parameter inversion; their differences lie only slightly in computational efficiency and long-sequence memory capabilities.
[0261] Furthermore, traditional cellular automata typically use a fixed Manning roughness factor (n), which cannot reflect the real-time changes in resistance caused by floating debris accumulation, vehicle obstruction, etc., during heavy rain. This invention utilizes AI to perceive these changes in real time. This invention uses a 1x1 convolutional layer (equivalent to a pixel-wise fully connected layer) to process high-dimensional feature maps. Decode the abstract features and map them back to the Manning roughness correction coefficient in the physical world:
[0262] ;
[0263] in, Let be the real-time dynamic roughness of cell (i,j) at time t (unit: s / m1 / 3). This is the execution parameter that is ultimately substituted into the physical formula. The initial reference roughness (e.g., 0.013 for asphalt pavement). , These represent the convolutional kernel weights and biases of the decoding layer. The hyperbolic tangent activation function restricts the AI output to the range [-1, 1], ensuring numerical stability. The correction amplitude limit coefficient is set to 0.5 in this embodiment. This ensures that the roughness after AI correction is always within a physically reasonable range (i.e., 0.5nbase~1.5nbase), preventing the model from outputting negative values or extreme outliers, which reflects the design concept of "physically constrained AI".
[0264] It is worth noting that the subsurface latent path activation (Structure Learning) based on graph attention is the core component of this invention for achieving "disaster chain transmission path identification". The system utilizes the graph attention mechanism (GAT) to perform correlation analysis on nodes across the entire domain. Unlike simple neighborhood calculation, GAT can capture long-distance, nonlinear spatiotemporal causal relationships between surface cells and subsurface nodes.
[0265] For any potential connectivity pair (e.g., surface cell u pointing to underground dark node v), calculate the correlation score of its eigenvector. :
[0266] ;
[0267] in, , The feature vectors corresponding to nodes u and v (from H in the above steps) t ). It is a linear transformation matrix used to map features to a subspace suitable for determining connectivity. This is a vector concatenation operation that merges the features of two nodes. This is the original attention score. The larger the value, the stronger the causal relationship the model believes there is between "abnormal water accumulation at point u" and "water level change at point v" (e.g., the water accumulation at point u disappears rapidly, followed by a surge in water level at point v). This represents the transpose of the learnable weight vector of a single-layer feedforward neural network. This vector is responsible for mapping the concatenated high-dimensional features of the nodes to a single scalar real number, i.e., the original attention score. To represent the activation function of a linear rectifier with leakage, (usually the slope of the negative half-axis) (Take 0.2). This function introduces nonlinear characteristics, enabling the model to handle complex nonlinear hydraulic connections between nodes and preventing the gradient vanishing problem during training.
[0268] Furthermore, the Sigmoid activation function is used to transform the correlation scores into conduction weights of the virtual pipelines in the physical model, i.e., the conduction weights of the virtual connection edges. :
[0269] ;
[0270] in, Let be the real-time conduction rate (range [0,1]) of the connection edge u→v. →0: Indicates that the AI determines there is no underground connection here, and the virtual path is closed. If →1: This indicates that the AI has determined that there is a hidden network of pipes here, and the virtual path is started at full speed. This is the activation threshold. Only when the correlation... The path will only be opened when the background noise level significantly exceeds this value, effectively preventing misjudgment. This is the sensitivity coefficient. It controls the steepness of the activation function, determining whether the path transitions gradually from "disconnected" to "connected" or occurs abruptly.
[0271] It is worth mentioning that in the preferred embodiment of this invention, GAT is used to calculate the attention weights between nodes. In addition, the self-attention mechanism in Vision Transformer (ViT) or Graph Transformer also has the ability to calculate the global node correlation and can be used as an alternative implementation. The specific implementation process is as follows: first, the surface cell grid is flattened into a series of token sequences; then, the correlation matrix between any two tokens (corresponding surface cells) is calculated using the Transformer's Attention(Q,K,V) formula; finally, this correlation matrix is normalized using Softmax and directly mapped to the conduction weight w of the underground virtual connection. This implementation can also achieve the discovery of potential underground paths. Its difference from the preferred scheme mainly lies in the computational overhead and the ability to capture local features, while the core technical concept remains consistent.
[0272] Furthermore, GAT calculates weights dynamically, while Improved Graph Convolutional Networks (GCNs) can achieve a similar weight representation function by learning a "learnable adjacency matrix." Specifically, a fully connected adjacency matrix Alearnable is first initialized and used as a trainable parameter of the neural network; then, graph convolution operations are performed... During the process, the values of Alearnable are directly updated through backpropagation. After training, the elements with larger values in the adjacency matrix Alearnable correspond to the actual underground connections. This method belongs to a more direct "structure learning" approach. Although its sensitivity to dynamic changes is not as high as GAT, it can still complete the task of inverting static underground pipe network topology.
[0273] Furthermore, cellular automata (CAs) take over the computational tasks. They utilize the dynamic parameters output by the AI. and structural weights The specific water exchange is calculated using rigorous physical formulas.
[0274] For adjacent physical cells on the Earth's surface (e.g., flow u to v), the surface runoff is calculated using the discretized Manning formula, which is the runoff Q between adjacent surface cells. surf :
[0275] ;
[0276] in, The surface runoff from cell u to cell v is expressed as m³ / s. This is a symbolic function. It is used to determine the direction of water flow, ensuring that water always flows from high head to low head. Zu - Zv is the difference in elevation between the base plates of cells u and v (in meters). The effective water passage width (in meters) is the interface between two cells. In regular grids, the grid side length Δx or Δy is typically used. The effective water depth at the interface (unit: m). To ensure calculation stability, the upstream water level or the maximum of the two is usually taken, i.e., max(h). u ,h v ). This is a key parameter (AI correction term). The dynamic Manning roughness of cell u at time t (unit: s / m^(1 / 3)). This value is not a fixed constant, but is output in real-time by a deep learning model (Conv-LSTM) based on the current water accumulation characteristics. It reflects the real-time resistance of surface debris, road texture, and vehicle obstruction to water flow. , Let be the water depth (in meters) of cells u and v at the current moment. The physical distance between the center points of two cells (unit: m).
[0277] Furthermore, for virtual connections activated by AI (such as from surface u to underground j), the potential flow rate, i.e., the potential underground transmission flow rate Q, is calculated using an AI-weighted simplified pipe flow formula. latent :
[0278] ;
[0279] in, The potential transmission traffic (in m^3 / s) flowing from node u to node j via a virtual path. This is the pipe flow coefficient (dimensionless). It is used to characterize the energy loss of fluid passing through an orifice or pipe, and is usually taken as an empirical value (e.g., 0.6~0.8). Reference cross-sectional area (in m²) for a unit of virtual pipe. A system-preset constant used to convert weights into specific physical flow areas. The acceleration due to gravity is taken as 9.81 m / s^2. This represents the total head difference between the two nodes (unit: m). The calculation formula is: That is, (potential energy at the starting point + pressure energy) minus (potential energy at the ending point + pressure energy). This is a unidirectional flow constraint function. It ensures that flow only occurs when the total head at the starting point is higher than that at the ending point, preventing the AI from erroneously generating backflow when the physical potential energy is insufficient.
[0280] Finally, based on the volume conservation law in fluid mechanics, all source and sink terms for each cell are summarized, and its water depth in the next time step is updated. That is, the future water depth of each surface cell is calculated using the volume conservation law based on the overflow between adjacent surface cells and the potential underground transport flow.
[0281] ;
[0282] in, Let be the water depth (in meters) of cell u at the next time step (t+1), i.e., at a future time. Let be the water depth of cell u at the current time step (t) (in meters). The time step for model calculation (in seconds). The CFL stability condition must be met to prevent numerical divergence. Let be the base area of cell u (in m²). In a regular grid, it is Δx × Δy. The increment of rainfall depth calculated based on the current rainfall intensity (unit: m / s × The obtained depth, i.e., the source term. The current soil infiltration depth loss (unit: m / s × The resulting depth, i.e., the sink term, is usually calculated using the Horton or Green-Ampt infiltration formula. This is the neighborhood set of cell u. It includes physical neighbors (top, bottom, left, and right on the ground) and virtual neighbors (underground nodes connected via AI activation). This represents the total flow rate into the cell (unit: m^3 / s). This represents the total flow rate out of the cell (unit: m^3 / s). Note: The flow rate Q here includes surface runoff Q. surf and underground potential transmission Q latent Two parts.
[0283] like Figure 3As shown, through the above steps, the system completes a closed loop: AI is responsible for "perceiving the environment" and "building the structure," while the physical equations are responsible for "computing the flow" and "constraining conservation." This mechanism ensures that as monitoring data accumulates, AI can increasingly accurately delineate the invisible underground transmission network of the city, thereby achieving true intelligent forecasting.
[0284] It is worth mentioning that in this embodiment, a "graph PDE solver" can be used instead of a cellular automaton (CA). While the original invention used cellular automata (CA) as the core of physical computation, the physical computation part can actually be replaced by a graph-based finite volume method (Graph FVM). In specific implementation, a regular mesh is not emphasized; instead, all nodes (including surface and underground nodes) are treated as an unstructured graph. Simplified Saint-Venant equations or diffusive wave equations are solved directly on the edges of the graph, and the weights w output by the AI are substituted as the "conductivity coefficients" in the equations for solving. This scheme essentially still follows the logic of "physical formula computational fluid dynamics + AI-defined structure," only the specific discrete form of the physical solver is replaced by a graph PDE instead of a cellular automaton (CA), perfectly matching the core concept of "physical-intelligent hybrid drive" of this invention.
[0285] Step 105: Construct a physical information composite loss function that includes a data fitting term, a structural sparsity term, and a parameter smoothing term. When the predefined sliding time window is reached, based on the mini-batch stochastic gradient descent mechanism, the physical information composite loss function is used to update the artificial intelligence model according to the final measured water depth, the future water depth of multiple surface cells, the real-time dynamic roughness, and the conduction weight of virtual connection edges. The updated artificial intelligence model is determined through multiple self-learning processes.
[0286] The data fitting term is used to measure the error between the future water depth and the final measured water depth.
[0287] The structural sparsity term is used to apply L1 regularization constraints to the conduction weights of virtual connection edges, forcing the weights of virtual connection edges without hydraulic connections to approach zero, thereby filtering out sparse underground conduction paths.
[0288] The parameter smoothing term is used to impose spatiotemporal continuity constraints on real-time dynamic roughness to prevent physical parameters from undergoing drastic, non-physical changes.
[0289] It should be noted that this step is a crucial step in achieving the system's "intelligent evolution." Traditional model calibration is typically offline and lagging, while this invention employs an online learning mechanism. The central system utilizes the real-time backhauled water accumulation truth values from the edge terminals to construct a composite objective function that includes data fitting terms and physical structure constraints. Through the automatic differentiation technique of a deep learning framework, the gradient of the prediction error with respect to the physical parameters (n) and structural weights (w) is calculated, and backpropagation is used to update the network weights of the AI model, thereby achieving precise locking of the disaster chain propagation path and adaptive correction of the flow field parameters. The specific implementation process includes three sub-steps: constructing the composite loss function, end-to-end gradient backpropagation, and rolling update of model parameters.
[0290] First, in order to guide the AI model to find the optimal solution while adhering to the laws of physics, this invention defines a composite loss function L consisting of three parts. total This function is calculated once at every time step t.
[0291] ;
[0292] in, , Hyperparameter weights are used to force the model to adjust its parameters so that the prediction results are closer to reality, thereby balancing data accuracy and physical plausibility.
[0293] Furthermore, data fitting loss Used to measure the difference between "predicted water level" and "measured water level":
[0294] ;
[0295] in, The predicted water depth calculated by the DL-CA model (including surface and subsurface interaction results). These are the true observation values returned by the edge computing gateway. A set of observation points on which monitoring equipment is deployed.
[0296] Furthermore, structural sparsity loss This is a key constraint for path recognition. According to Occam's razor, underground disaster transmission paths (pipelines, fissures) should be sparse relative to the overall space. This invention does not aim to have AI connect all nodes in order to fit the data.
[0297] ;
[0298] in, The structural sparsity loss assigns weights to the potential paths output by the GAT network. This forces the GAT network to suppress the weights of most irrelevant paths to zero, retaining only those connections crucial for interpreting water level anomalies. This results in clear and well-defined subsurface conduction pathways, rather than a jumbled mess.
[0299] Furthermore, parameter smoothing loss It can constrain the spatiotemporal continuity of physical parameters (Manning roughness) and prevent physically impossible drastic jumps, thereby ensuring that the inverted roughness field is smoothly transitioned in time and space, conforming to the gradual laws of the physical world.
[0300] ;
[0301] in, Let be the dynamic roughness at time t, which is derived by the AI in the above steps.
[0302] Furthermore, traditional model calibration often treats the physical model as a "black box," relying solely on guesswork to determine parameters. The core advantage of this invention lies in the "differentiability of the physical process." This is because the Manning formula, pipe flow formula, and water balance formula in step three are essentially addition, subtraction, multiplication, division, and exponentiation operations, which are mathematically completely differentiable.
[0303] This means that when the predicted result h^(t+1) does not match the true value, we can use the chain rule, like current conduction, to transmit the "error signal" back to the interior of the deep neural network without loss, and accurately guide the modification of the weights.
[0304] The target that needs updating is the set of trainable parameters inside the neural network. (Specifically includes the convolutional kernel weights of Conv-LSTM) Graph attention weights of GAT ).
[0305] The error feedback path follows a logical chain of "total error → water level → flow rate → control parameters → neural network weights". The mathematical expression is as follows:
[0306] ;
[0307] in, This indicates the magnitude of the prediction deviation. For example, if the predicted water accumulation at a certain location is 0.5m, but the actual measured water accumulation is 0.8m, the gradient value will be very large and needs to be adjusted. This indicates "how much flow rate needs to be changed to eliminate this water level deviation." It originates from the law of conservation of volume in step three. Because the change in water level is determined by the net flow rate, this term tells the system how many cubic meters of water need to be brought in to compensate for the 0.3m water level difference. This represents the sensitivity of flow rate to roughness (derived from Manning's formula). If the water flow is too slow here, the gradient will indicate: "The resistance n is too large." To tell the Conv-LSTM network to adjust the neuron connections Θlstm, the roughness nt output here is made smaller. The sensitivity of flow rate to valve opening (derived from the pipe flow formula). The system detects: "A large inflow is needed here, but the valve w is currently closed (w→0)." To adjust the graph attention weights Increasing the weight w of this connection activates the underground passage. Through this chain reaction, the error in the measured data is directly transformed into the driving force for activating the underground hidden path.
[0308] Furthermore, this invention does not perform "one-time training," but rather adopts a "learn-as-you-go" strategy, known as "Moving Horizon Estimation (MHE)." This is analogous to an experienced dispatcher whose understanding of the city's pipe network deepens as the torrential rain continues.
[0309] This invention sets a sliding time window T window (For example, data from the past 15 minutes). In each computation step, the system collects all "observation-prediction" sequences within the window and performs a parameter update once:
[0310] ;
[0311] in, , The updated and unupdated neural network weights of the artificial intelligence model. This is the learning rate. It's a hyperparameter that controls the "step size" and determines how quickly the model learns from mistakes. The total gradient calculated in the above steps.
[0312] Based on the above, a physical information composite loss function integrating multi-dimensional physical and data constraints is specifically constructed. This function integrates data fitting terms, structural sparsity terms, and parameter smoothing terms to achieve comprehensive supervision of the model. When the model operation reaches a predefined sliding time window, a mini-batch stochastic gradient descent mechanism is used as the update driver. The final measured water depth is taken as the true supervision value. The future water depth, real-time dynamic roughness, and virtual connection edge conduction weights of multiple surface cells output by the model are substituted into the physical information composite loss function to complete the accurate calculation of the loss value. Then, based on the loss value, the network weights and core parameters of the artificial intelligence model are iteratively optimized and adjusted through gradient backpropagation. Through multiple rounds of online self-learning and parameter correction, the model error gradually converges, and finally the updated artificial intelligence model is determined. This process achieves precise supervision of model updates through a multi-constraint composite loss function, and completes the phased online self-learning iteration of the model by combining a sliding time window and a mini-batch stochastic gradient descent mechanism. It effectively solves the technical problems of static and rigid parameters and lack of real-time measured data feedback correction in traditional models, and significantly improves the accuracy and dynamic adaptability of the model in predicting hydrological parameters related to urban rainstorms and floods. It lays a reliable model foundation for the subsequent output of visualized disaster chain transmission maps and accurate forecasting of urban rainstorm and flood disaster risks.
[0313] Step 106: Output a visualized disaster chain transmission map based on the updated artificial intelligence model, and realize the forecast of urban rainstorm and flood disaster risks.
[0314] It should be noted that, based on the updated artificial intelligence model, the model takes into account core data such as real-time hydrological monitoring data and surface and subsurface topological information. After calculation, the model outputs key parameters and path information of disaster chain transmission, transforming this information into an intuitive and easy-to-understand visual disaster chain transmission map. This map clearly presents the correlation logic of water accumulation evolution and hazard spread. At the same time, relying on the model's optimized predictive capabilities and combining real-time operating conditions and historical data, the model can predict the intrusion range and timing characteristics of disaster hazards in advance, thus achieving the prediction of urban rainstorm and flood hazard intrusion.
[0315] For example, in the initial stage of applying an artificial intelligence model (t=0, the rainstorm has just begun): 1) State: The AI model has not yet been calibrated with measured data, and at this time it "does not know" where the underground pipe network is. 2) Behavior: The system mainly relies on the surface physical model (baseline roughness n0, virtual connection closed w≈0) for inference. At this time, the prediction error may be relatively large.
[0316] In the mid-stage (t↑, water accumulation): 1) State: The edge gateway begins to transmit large amounts of ground truth about the accumulated water. 2) Behavior: The backpropagation mechanism detects "abnormal disappearances" or "abnormal outflows" of surface water. Huge gradients flow into the GAT network, forcing w to increase rapidly at specific connections.
[0317] In the later stages (t→∞, continuing to run): 1) State: The AI model has constructed an underground conduction topology that highly fits reality (learned an accurate W). g ) and dynamic resistance field (learned accurately) 2) Behavior: Predicted Results The results closely match the true values. At this point, the system is not just making predictions, but also creating an "invisible map of the city's groundwater in real time," achieving truly adaptive intelligent forecasting.
[0318] Furthermore, step 106 may include the following sub-steps:
[0319] S61. Based on the target virtual connection edge conduction weight and target global water accumulation depth distribution output by the updated artificial intelligence model;
[0320] S62. Based on the target virtual connection edge conduction weight and the target global water depth distribution, a disaster chain transmission map is generated.
[0321] S63. Visualize the disaster chain transmission map to obtain a visualized disaster chain transmission map;
[0322] S64. Calculate the disaster chain cascading risk index using the target virtual connection edge conduction weight;
[0323] S65. When the cascading risk index of the disaster chain exceeds the preset index threshold, issue a forecast of potential urban rainstorm and flood disaster intrusion.
[0324] The target virtual connection edge conduction weight refers to the weight parameter output by the updated artificial intelligence model, which can accurately characterize the actual hydraulic conduction capacity of each virtual connection edge in a fully connected topology network. It is a core quantitative indicator for characterizing the transmission path of underground disasters.
[0325] The target area water depth distribution refers to the spatial distribution data of water depth in all surface cells within the study area, output by the updated artificial intelligence model. It is the core basis for presenting the situation of surface waterlogging disasters.
[0326] The disaster chain transmission map refers to a map generated based on the virtual connection edge conduction weight of the target and the distribution of water depth in the entire target area, combined with the hydraulic transmission law. It can depict the chain transmission path, intensity and spatial correlation characteristics of urban rainstorm and flood disasters, and is the core carrier reflecting the evolution law of disaster chains.
[0327] It should be noted that this step transforms the overall water depth field derived from the previous steps and the identified hidden transmission paths into visualized decision support information. The system not only outputs "where is the water accumulation," but more importantly, it outputs "from where is the water secretly flowing in," thus providing targeted disaster prevention and dispatching information for urban emergency management departments. The specific implementation process includes three sub-steps: disaster chain map generation, targeted cascading early warning, and emergency dispatching assistance.
[0328] First, based on AI-activated virtual connection weights ( To create a dynamic "urban groundwater topology map".
[0329] Explicit Risk Layer: Displays the distribution of water depth (h) in surface cells in the form of a heat map. u ).
[0330] Implicit Risk Layer (Core): This layer identifies high-weight virtual connections through AI. > The streamline is rendered as a streamer with directional arrows. The thickness of the streamline is related to the calculated potential transport flow (Q). latent It is directly proportional to.
[0331] Visualization: On the dispatch screen, managers can see that the water flow not only spreads across the surface, but also quickly converges from high-water areas to low-lying nodes such as subway stations and underground shopping malls through several bright "invisible channels".
[0332] Furthermore, unlike traditional alarms based solely on water depth, this invention constructs a "disaster chain cascading risk index (I)". chain ")" to provide early warnings to downstream nodes.
[0333] ;
[0334] in, This is the cascading intrusion risk index for downstream node v (such as a subway station entrance). The potential conduction weights between upstream nodes u and v (discovered in real time by AI). This is driven by physical potential energy. Even if the downstream node v is not currently filled with water ( ≈0, but if the upstream u has severe water accumulation and the AI determines that there is a strong connection between the two ( →1), the system will calculate an extremely high I chain This allows for the issuance of a "hazard intrusion warning" 15-30 minutes before the actual arrival of the flood, prompting management personnel to lay down flood control sandbags in advance.
[0335] This invention matches the real-time calculated "flooding-receding" curve with the scheduling capacity of drainage pumping stations. If the model predicts that the floodwater in a certain area recedes extremely slowly (the n value generated by the AI remains consistently high, indicating physical blockage at the drainage outlet), the system automatically generates a work order to guide the emergency response team to specific coordinates for dredging operations, rather than blindly patrolling.
[0336] In this embodiment, relying on the updated artificial intelligence model, core data such as real-time hydrological monitoring data and surface and subsurface topological information are input. Combined with physical rules such as the discretized Manning formula and the water balance law, and intelligent algorithms, the accurate derivation of surface and subsurface hydraulic parameters is completed. This results in the output of the target virtual connection edge conduction weight (quantifying the hydraulic conduction efficiency of the virtual connection edge) characterizing the subsurface hydraulic conduction capacity, and the target global subsurface water depth distribution reflecting the water accumulation distribution pattern in the study area (presenting the spatial pattern of subsurface cell water depth in each region). Using these two core parameters as data support, and combining the conduction path characteristics of the fully connected topological network, the spatial correlation and intensity change information of disaster evolution are integrated. This invention generates a disaster chain transmission map depicting the cascading transmission patterns of rainstorm and flood disasters. The map is then graphically and graphically processed to achieve visualization, resulting in a visually representable disaster transmission path and distribution characteristics. Simultaneously, based on the conduction weights of target virtual connection edges, the disaster transmission efficiency of various underground transmission paths is quantified. Combined with the cascading correlation characteristics between paths, a disaster chain cascading risk index, characterizing the degree of disaster chain reaction, is calculated. Finally, the disaster chain cascading risk index is monitored in real time and a threshold is determined. When the disaster chain cascading risk index exceeds a preset threshold, an immediate urban rainstorm and flood disaster hazard intrusion forecast is issued. This invention generates and visualizes the disaster chain transmission map by outputting precise core parameters from the model. Combined with the quantified cascading risk index, it achieves threshold-based early warning of disaster hazards. This solves the technical problems of existing methods, such as the lack of intuitive depiction of disaster transmission patterns and the lack of quantitative judgment basis for hazard warning. It provides an intuitive reference for transmission patterns and a quantitative risk warning basis for the early prevention and precise emergency response to urban rainstorm and flood disasters, effectively improving the pertinence and timeliness of disaster emergency response.
[0337] For comparison of technical effectiveness, existing technologies can be used as a reference. Existing technologies have formed multiple technical paths for monitoring, simulating, and forecasting urban rainstorm and flood disasters, but all have significant limitations. Edge computing architecture technologies (such as CN118095095A and CN116702993A) solve the problems of high data transmission latency and high bandwidth pressure in traditional cloud models by offloading computing tasks to the edge, thus improving early warning response speed. However, the edge still runs pre-set hydrological models or offline computing schemes, lacking the ability to calculate high-precision ground truth water depth, and cannot provide real-time physical residual feedback to cloud models, making it difficult to achieve model optimization and iteration. Physical mechanism coupled models (such as the CA-UFM model and CN112270115A) improve computational efficiency by simplifying physical processes or improving cellular automata rules, but their operation heavily relies on complete underground pipe network topology data and lacks an "automatic discovery" mechanism for complex underground pipe network structures, making them unable to cope with scenarios where pipe network data is missing. Data-driven intelligent models (such as CN120031390A) utilize deep learning to optimize cellular automata rules, focusing on macro-level ecological risk or disaster trend prediction. However, they do not focus on characterizing hydraulic transmission paths and lack physical mechanism constraints, resulting in a "black box" nature that is uninterpretable. While automatic calibration techniques for physical model parameters (such as CN119647323B and the rectangular generalization calibration method) replace manual parameter tuning and improve parameter determination efficiency, they rely on a known model structure. This approach cannot address the core issue of unknown underground pipe network structures and cannot automatically discover hidden waterways.
[0338] In summary, existing technologies have four major drawbacks: First, edge computing solutions only achieve the downward distribution of computing power, lacking the ability to perceive "visual truth" and evolve models online. If the pre-built model has structural defects, even with fast computing speed, the output results will still be biased, failing to achieve "increasing accuracy with use." Second, purely data-driven models lack hard constraints from physical mechanisms, and the prediction results may violate the common sense of fluid mechanics. The "black box" characteristic reduces the credibility of emergency decision-making. Third, they heavily rely on high-precision underground pipe network topology data, lack the ability to "automatically discover" underground structures, and cannot construct a full-space causal topology of "surface-underground," making it difficult to achieve three-dimensional tracking of disaster chain transmission. Fourth, physical parameters and model structures are static and rigid. Static parameters cannot adapt to the dynamic changes in the environment during rainstorms, and the fixed and uniform model structure cannot balance computational efficiency and simulation accuracy, falling into a zero-sum game of "accuracy and speed."
[0339] Therefore, the technical problem to be solved by this invention can be divided into four parts:
[0340] First, it addresses the problem of inaccurate flood forecasting when underground pipe network maps are unavailable or inaccurate. Current forecasting systems are like navigating a maze, relying solely on surface maps. Figure 1Similarly, the detailed location and dimensions of every underground drainage pipe in the city must be pre-entered. However, in many old urban areas, the underground pipe network maps are often lost, or the actual situation does not match the maps at all (for example, there are unauthorized underground ditches). The problem that this invention aims to solve is: how to "guess" the underground water flow path based on the surface water conditions when there is no underground pipe network "map" or the "map" is incorrect, thereby achieving accurate prediction.
[0341] Secondly, this invention addresses the problem of existing pure AI forecasts being prone to errors that "defy common sense," leading to unreliable results. Currently, it's common to use AI to directly predict flooding, but these AI models are like "black boxes"—they don't understand basic physics (such as water not disappearing into thin air). This can lead to absurd predictions like "the floodwater suddenly disappears even when there are no drainage outlets" during unprecedented torrential rains. The problem this invention aims to solve is: how to "constrain" AI, forcing it to adhere to the conservation laws of the physical world (water must have both flow and release), thereby ensuring that forecast results are scientific and credible.
[0342] Third, this invention addresses the problem of slow response times in forecast systems, which are unable to handle sudden changes such as road blockages during heavy rain. Existing systems typically assume that a city's drainage capacity is constant. However, in reality, during heavy rain, leaves may suddenly block sewer openings, or manhole covers may be pushed open, instantly altering the water situation. If the system still relies on outdated calculations, the forecasts will be inaccurate. The problem this invention aims to solve is: how to enable the system to act like it has eyes, seeing environmental changes in real time (such as where there are blockages), and automatically adjusting calculation parameters to ensure that the forecast keeps pace with the actual situation on the ground.
[0343] Fourth, this invention addresses the problem of only being able to predict where water is on the surface, but not how it flows into underground facilities (such as subways and parking garages). Current systems mostly only generate a surface "water distribution map," indicating which roads are flooded. However, for high-risk areas like subway stations and underground parking garages, managers need to know which unseen underground paths the water "steals" through. The problem this invention aims to solve is: how to not only clearly see the water on the surface, but also track how the water flows from the surface into the underground and then to critical facilities, thus providing rescue personnel with precise "leak-stopping" guidance.
[0344] Therefore, this invention provides a self-learning method for rapid forecasting of urban rainstorm and flood disaster chain transmission paths, such as... Figure 4As shown, Step 1 (Rapid Calculation of Edge Data): Utilizing an edge computing gateway, the true value of water depth is calculated in real time using a monocular vision algorithm, providing the system with high-precision "training labels"; Step 2 (Initial Structure Interpretation): A two-layer coupled network of "surface physical grid + underground virtual nodes" is constructed, with a large number of pre-set potential connection interfaces that can be learned, breaking through the dependence of traditional models on pipeline network data; Step 3 (Transmission Path and Model Parameter Learning): Graph Attention Network (GAT) and Conv-LSTM are innovatively embedded, allowing AI to be responsible for "finding hidden paths" and "correcting dynamic parameters," while cellular automata are responsible for "computing physical flow," achieving a unity of structural self-learning and physical conservation; Step 4 (Dynamic Adjustment): A composite loss function containing physical constraints is constructed, and end-to-end online backpropagation is performed using the chain rule, enabling the model to continuously "self-learn and correct" in practice, becoming more accurate with use; Step 5 (Rapid Forecasting): Finally, a disaster chain map containing hidden transmission paths and cascaded risk warnings are output, realizing a leap from "passive monitoring" to "active prediction."
[0345] Regarding the first drawback (edge perception and online evolution), this invention proposes a closed-loop architecture of "edge visual computation - cloud model evolution." At the edge, the true value of water depth is directly calculated in real time using a monocular vision algorithm, uploading only high-frequency, lightweight water depth data. At the cloud (or central location), this true value is used to calculate the physical residual, and the network weights of the model are updated online through gradient backpropagation. This not only solves the transmission latency problem but also endows the system with the ability to "learn online," enabling it to continuously correct its own defects as the rainstorm progresses, thus achieving continuous improvement in forecast accuracy.
[0346] To address the second drawback (physical conservation): This invention embeds a deep learning model into the computational loop of a cellular automaton (CA), using the water balance equation as a hard constraint on the AI output. The AI is only responsible for predicting parameters and weights; the specific water level evolution is calculated using rigorous physical formulas. This ensures that the forecast results possess both data-driven flexibility and strictly adhere to the law of conservation of mass, eliminating the instability of "black box" models.
[0347] To address shortcomings three and four (dynamic parameter and structural adaptation): This invention is based on a deep learning augmented cellular automata (DL-CA) architecture. Through the collaborative work of Conv-LSTM and graph attention network (GAT), it realizes online inversion of physical parameters and dynamic sparsification adjustment of model structure, thereby achieving a dynamic optimal match between simulation accuracy and computational efficiency.
[0348] To address parameter rigidity—dynamic parameter inversion based on flow field perception: This invention introduces Conv-LSTM (Convolutional Long Short-Term Memory) network to perceive the spatiotemporal evolution characteristics of the flow field in real time. The model can automatically identify environmental changes such as road blockages and pipe siltation based on real-time rainfall and water accumulation monitoring data, and online invert and output the time-varying Manning roughness field (n) across the entire domain. t This allows the physical model to adaptively simulate the complex nonlinear process of "sudden water level changes due to blockage".
[0349] Addressing structural rigidity—Dynamic structural reconstruction based on attention mechanisms: This invention utilizes the learnable adjacency matrix output by GAT (Graph Attention Network) to achieve dynamic adjustment of the model structure. GAT can automatically calculate the connection weights (w) between nodes based on the spatiotemporal response differences of surface water levels. uv For areas with detected hydraulic interactions (such as surface inundation zones and potential underground water-conducting channels), GAT automatically assigns high weights (w→1), which is equivalent to logically "encrypting" the computational connections, ensuring accurate simulation of critical water conveyance paths (especially unknown underground paths). For non-inundation areas or areas without hydraulic connections, GAT automatically suppresses their connection weights (w→0), putting them in a "dormant" state in physical calculations. This is equivalent to dynamically eliminating invalid computational units, significantly reducing computational redundancy.
[0350] Through the aforementioned mechanism, this invention achieves the goal of "using resources wisely." The system can maintain low computational load and operate rapidly during the initial stage of a rainstorm (when water is scarce); during the peak of the rainstorm (when water is abundant and conditions are complex), it automatically allocates computing power to focus on high-risk areas and newly discovered underground transmission paths. This dual self-learning mechanism of "parameters + structure" solves the problem that existing technologies cannot simultaneously achieve "millisecond-level rapid calculation" and "meter-level accurate simulation."
[0351] As can be seen from the above, the core technology of this invention is built around four key points, forming a complete self-learning rapid forecasting technology system for the transmission path of urban rainstorm and flood disaster chains, as detailed below:
[0352] 1) Addressing the issue that existing edge computing only performs "data transfer" and cannot obtain "physical truth," this invention proposes an edge-end water depth truth calculation based on monocular vision and an "edge-cloud" co-evolution architecture. The protected technical solution is a "edge calculation-cloud evolution" system architecture: instead of running complex hydrodynamic models at the edge, a lightweight monocular vision algorithm is deployed to calculate the true value of surface water depth (h) in real time. obsThis technology only uploads high-frequency, low-bandwidth water accumulation data; the cloud uses the ground truth of the water accumulation uploaded from the edge to calculate the physical residual, and updates the weights of the cloud master model (DL-CA) online through gradient backpropagation. The core of this technology lies in the closed-loop feedback mechanism that uses the visual ground truth at the edge as a supervision signal to drive the cloud physical-intelligent hybrid model to become more accurate with use, effectively making up for the shortcomings of existing edge computing, such as the lack of physical ground truth support and the inability to iteratively optimize the model.
[0353] 2) Addressing the core pain point of being unable to model "pipeline blind spots," this invention proposes a "structural self-learning" method for underground hidden hydraulic paths based on the Graph Attention (GAT) mechanism. Existing technologies require a defined pipeline drawing as input, while this invention innovatively proposes a reverse approach of "first creating a virtual fully connected network, then pruning using attention weights." The proposed technical solution involves: constructing a two-layer coupled network of "surface physical grid - underground virtual nodes," initialized as a fully connected heterogeneous graph; using the Graph Attention (GAT) network to analyze the spatiotemporal anomalies of surface water accumulation (such as the causal relationship between rapid water receding in one area and rising water levels in a distant area); and automatically calculating the conduction weights (w) of connected edges using attention coefficients. uv This allows for the reverse identification of unknown underground transmission paths. The key lies in a modeling method that infers underground connectivity structures solely from surface water response, without relying on prior pipe network topology data, successfully solving the modeling challenge in scenarios where pipe network data is lacking.
[0354] 3) To address the problem of "dual rigidity of model structure and parameters," this invention proposes a physical parameter inversion and dynamic sparsification technique based on flow field perception to achieve dynamic adaptation of accuracy and speed. The proposed technical solution includes two core aspects: first, parameter dynamization, utilizing Conv-LSTM to perceive the spatiotemporal characteristics of the flow field in real time and online inverting the time-varying Manning roughness (n) across the entire domain. t The technology employs two main techniques: first, it accurately simulates sudden environmental changes such as blockages and manhole cover popping; second, it dynamically sparsifies the structure by using the sparse adjacency matrix output by GAT to dynamically adjust the density of computational units, activating connections (w→1) in flooded areas and automatically severing connections (w→0) in non-flooded areas, thereby dynamically eliminating invalid calculations and concentrating computing power on critical regions. The key to this technology lies in its self-learning mechanism that automatically adjusts the density (structure) of computational units and the drag coefficient (parameters) based on the flow field state, breaking the zero-sum game between "accuracy and speed" in existing technologies.
[0355] 4) Addressing the contradiction between "black-box AI's lack of conservation" and "the difficulty in training physical models," this invention proposes an end-to-end online reverse correction mechanism based on physical residual fields. The proposed solution involves embedding a deep learning model into the computational loop of a cellular automaton (CA), with the AI output serving as coefficients for the physical equations. A physical information composite loss function is constructed, comprising a data fitting term (approaching the true value at the edge), a structural sparsity term (L1 regularization, forcing clarity of underground paths), and a parameter smoothing term. The chain rule is used to incorporate the physical computation process (water balance equation) into the gradient descent path, achieving end-to-end updates from prediction error to AI network weights. The key is utilizing the water balance equation as a "hard constraint" to ensure that the prediction results strictly adhere to the law of mass conservation during training, thus resolving the instability of pure AI models and reducing the training difficulty of physical models.
[0356] Compared with existing technologies, the primary advantage of this invention is that it breaks through the dependence on prior pipeline network data and achieves "structural self-learning" of groundwater conduction paths. Existing physical coupling models such as the CA-SWMM model and parameter calibration techniques all rely on the premise of "known pipeline network topology," and cannot address the "blind spots" of missing pipeline network data or discrepancies between archives and the current situation in old urban areas. This invention introduces a Graph Attention Network (GAT) to construct a learnable virtual topology layer. By analyzing the spatiotemporal anomaly response characteristics of surface water accumulation (such as the causal relationship between rapid water receding in a certain area and lagging water level rise in a distant area), it reverse-engineers and locks in unknown groundwater connectivity weights (w). uv This allows the method to automatically construct a high-fidelity surface-subsurface coupling model even without pipe network drawings or with incomplete data, achieving a leap from "two-dimensional surface flow" to "full-space three-dimensional tracking".
[0357] The second advantage of this invention is that it achieves a "dynamic optimal fit" between simulation accuracy and computational efficiency, solving the problem of model rigidity. Existing edge computing, physical coupling models, and parameter calibration techniques typically employ static physical parameters (such as fixed Manning roughness) and rigid global uniform mesh structures, failing to balance "millisecond-level" response speed with "meter-level" simulation accuracy. This invention utilizes Conv-LSTM to sense the flow field state in real time and inverts the dynamic Manning roughness (n) online. t This simulates sudden environmental changes such as road blockages; simultaneously, it utilizes GAT's sparse attention mechanism to dynamically adjust the density of computing units—activating connections (densification) in flooded areas and critical paths, and automatically severing connections (pruning) in non-flooded areas. This dual self-learning mechanism of "parameters + structure" enables computing power to automatically focus on key areas of disaster evolution, achieving optimal simulation performance with limited computing resources.
[0358] The third advantage of this invention is the establishment of a closed-loop evolutionary system of "edge ground truth calculation - cloud-based physical correction," which endows the model with online learning capabilities. Existing traditional edge computing only deploys a pre-set hydrological model to the edge, lacking real-time feedback on ground truth water depth, thus preventing the model from self-correcting. This invention deploys a monocular vision algorithm at the edge to directly calculate the ground truth water depth and uses the physical residual between this value and the predicted value to update the weights of the cloud-based master model (DL-CA) online through gradient backpropagation. This "edge-cloud collaborative closed loop" not only solves the data transmission latency problem but also ensures that the model can iterate and learn using each rainstorm process, achieving continuous improvement in forecast accuracy.
[0359] The fourth advantage of this invention is that it combines the flexibility of data-driven approaches with the conservation of physical models, eliminating the risk of a "black box." Existing pure AI models lack physical constraints and often output predictions that violate the law of conservation of mass (such as water disappearing into thin air). This invention embeds a deep learning model into a cellular automaton (CA) computational loop, using the water balance equation as a hard constraint on the AI output. The AI is only responsible for predicting parameters and weights; the specific water level evolution is calculated using rigorous physical formulas. This ensures that the forecast results strictly adhere to the law of conservation of mass under any extreme conditions, providing scientific and reliable data support for emergency decision-making.
[0360] Furthermore, traditional hydraulic models (such as SWMM) suffer from low computational efficiency and insufficient simulation accuracy, with computation times often on the order of hours. They also simplify complex underground spaces into linear structures, making it difficult to accurately depict the three-dimensional intrusion details such as floods rushing into subway stations. Pure AI prediction models (such as LSTM) are "black box" models, lacking the support of physical principles. When faced with rainstorm scenarios not present in the training data, their predictions lack reliability and are prone to "guessing" biases. This invention effectively overcomes the shortcomings of the above two types of models. It introduces a cellular automata (CA) model as the physical mechanism foundation to ensure the accuracy of the general direction of disaster forecasting, and uses ensemble Kalman filtering (EnKF) for AI correction to ensure the accuracy of real-time predictions. Simultaneously, it achieves three-dimensional simulation and forecasting of urban rainstorm and flood disasters, balancing reliability, real-time performance, and detailed characterization capabilities.
[0361] In this embodiment of the invention, a self-learning method for rapid forecasting of urban stormwater flood disaster chain transmission paths is provided. This method acquires multiple consecutive frames of video stream images and camera intrinsic and extrinsic parameter data, and preprocesses these images using a lightweight convolutional neural network to output a set of pixel coordinates for the land-water boundary. Using triangulation and a pinhole camera imaging model, the method calculates the measured water depth based on the camera extrinsic and extrinsic parameter data and the set of pixel coordinates for the land-water boundary, outputting the final measured water depth. A discrete surface cellular grid and a fully connected topological network are constructed, and the network topology and hydrological state data corresponding to the fully connected topological network are converted into a high-dimensional tensor of state features. Based on the discretized Manning formula and the water balance law, and using a high-dimensional tensor of state characteristics and a fully connected topological network, the system outputs the future water depth, real-time dynamic roughness, and virtual connection edge conduction weights for multiple surface cells in a discrete surface cell grid. A physical information composite loss function is constructed, including data fitting, structural sparsity, and parameter smoothing terms. When a predefined sliding time window is reached, based on a mini-batch stochastic gradient descent mechanism, the physical information composite loss function updates the AI model according to the final measured water depth, the future water depth of multiple surface cells, the real-time dynamic roughness, and the virtual connection edge conduction weights. This process is further refined through multiple self-learning iterations. The updated artificial intelligence model is determined; based on the updated artificial intelligence model, a visualized disaster chain transmission map is output, and the intrusion forecast of potential urban rainstorm and flood disaster hazards is realized; based on the above scheme, this invention preprocesses images and extracts land and water boundaries through a lightweight convolutional neural network, and combines camera-inside and outside geometric modeling to directly solve the high-precision measured ground truth of water depth at the edge, which can provide a real-time and reliable source of physical residual feedback of "visual ground truth" for cloud-based artificial intelligence models. At the same time, a fully connected topological network is constructed and structural sparsity constraints are introduced to realize adaptive discovery of underground hidden transmission paths (structural self-learning), which makes up for the lack of underground pipe in existing pre-built models. To address the structural gaps caused by online data, this paper combines the discretized Manning formula and the water balance law with parameter smoothing constraints. This allows the model's output parameters, such as future water depth and real-time dynamic roughness, to have solid physical process support and adaptive inversion with environmental changes (parameter self-learning). Furthermore, through a small-batch stochastic gradient descent update mechanism under a sliding time window, the measured true value of water depth is deeply integrated with the multi-dimensional parameters output by the model. This enables the model to dynamically iterate and learn during the rainstorm process, ensuring that the model becomes more accurate with each calculation. This significantly improves the accuracy of disaster chain transmission path forecasts and provides precise and timely technical support for the prevention and control of urban rainstorm and flood disaster hazards.
[0362] Please see Figure 5 , Figure 5 This is a structural block diagram of a self-learning rapid forecasting system for urban rainstorm and flood disaster chain transmission paths, provided in Embodiment 2 of the present invention.
[0363] This invention provides a self-learning rapid forecasting system for urban rainstorm and flood disaster chain transmission paths, comprising:
[0364] The acquisition module 501 is used to acquire multiple consecutive frames of images from the video stream and camera intrinsic and extrinsic parameter data, and to preprocess the multiple consecutive frames of images from the video stream based on a lightweight convolutional neural network, outputting a set of pixel coordinates of the land-water boundary line.
[0365] Output module 502 is used to calculate the measured water depth based on camera intrinsic and extrinsic parameters and the set of pixel coordinates of the land-water boundary line using triangular geometric relationships and pinhole camera imaging model, and output the final measured water depth.
[0366] Module 503 is used to construct discrete surface cell grids and fully connected topological networks, and to convert the network topology and hydrological state data corresponding to the fully connected topological network into high-dimensional tensors of state features.
[0367] Calculation module 504 is used to output the future water depth, real-time dynamic roughness, and virtual connection edge conduction weight of multiple surface cells in the discrete surface cell grid based on the discretized Manning formula and the water balance law, according to the state characteristic high-dimensional tensor and fully connected topology network.
[0368] The self-learning module 505 is used to construct a physical information composite loss function that includes a data fitting term, a structural sparsity term, and a parameter smoothing term. When a predefined sliding time window is reached, based on the mini-batch stochastic gradient descent mechanism, the physical information composite loss function is used to update the artificial intelligence model according to the final measured water depth, the future water depth of multiple surface cells, the real-time dynamic roughness, and the virtual connection edge conduction weights. The updated artificial intelligence model is determined through multiple self-learning processes.
[0369] The forecast module 506 is used to output a visualized disaster chain transmission map based on the updated artificial intelligence model and to realize the forecast of potential urban rainstorm and flood disaster intrusion.
[0370] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and modules described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0371] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the self-learning rapid forecasting method for urban rainstorm and flood disaster chain transmission paths as described in the above embodiments.
[0372] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the self-learning rapid forecasting method for urban rainstorm and flood disaster chain transmission paths as described in the above embodiments.
[0373] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.
[0374] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.
[0375] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0376] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0377] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A self-learning rapid forecasting method for urban rainstorm and flood disaster chain transmission paths, characterized in that, include: Acquire multiple consecutive frames of images from a video stream and camera intrinsic and extrinsic parameter data, and preprocess the multiple consecutive frames of images from the video stream based on a lightweight convolutional neural network to output a set of pixel coordinates of the land-water boundary line; Using trigonometric relationships and a pinhole camera imaging model, the measured water depth is calculated based on the camera's intrinsic and extrinsic parameters and the set of pixel coordinates of the land-water boundary line, and the final measured water depth is output. Construct a discrete surface cellular grid and a fully connected topological network, and convert the network topology and hydrological state data corresponding to the fully connected topological network into a high-dimensional tensor of state features; Based on the discretized Manning formula and the water balance law, according to the state feature high-dimensional tensor and the fully connected topology network, the future water depth, real-time dynamic roughness and virtual connection edge conduction weight of multiple surface cells in the discrete surface cell grid are output. A physical information composite loss function is constructed, which includes a data fitting term, a structural sparsity term, and a parameter smoothing term. When a predefined sliding time window is reached, the artificial intelligence model is updated based on the mini-batch stochastic gradient descent mechanism using the physical information composite loss function according to the final measured water depth, the future water depth of multiple surface cells, the real-time dynamic roughness, and the conduction weight of virtual connection edges. The updated artificial intelligence model is determined through multiple self-learning processes. Based on the updated artificial intelligence model, a visualized disaster chain transmission map is output, and the intrusion forecast of potential urban rainstorm and flood disaster hazards is realized.
2. The rapid forecasting method for the transmission path of urban rainstorm and flood disaster chains with self-learning capability as described in claim 1, characterized in that, The lightweight convolutional neural network includes an encoder and a decoder; the preprocessing of multiple consecutive frames of the video stream based on the lightweight convolutional neural network to output a set of pixel coordinates for the land-water boundary line includes: Rain and fog are removed from multiple consecutive frames of the video stream to obtain multiple consecutive frames of the video stream after rain and fog removal. Background modeling is performed on multiple consecutive frames of the video stream after rain and fog removal to obtain clear static background frames; The clear static background frame is encoded using an encoder to obtain encoded features; The encoded features are classified at the pixel level using a decoder based on a channel attention mechanism to obtain a binarized mask image. The Canny edge detection operator is used to extract the set of pixel coordinates of the land-water boundary line from the binarized mask image.
3. The rapid forecasting method for the transmission path of urban rainstorm and flood disaster chains capable of self-learning, as described in claim 1, is characterized in that... The method employs triangulation and a pinhole camera imaging model to calculate the measured water depth based on the camera's intrinsic and extrinsic parameters and the set of pixel coordinates of the land-water boundary line, outputting the final measured water depth, including: Based on the pinhole camera imaging model, a mapping relationship between the image pixel coordinate system and the world physical coordinate system is constructed. Based on the mapping relationship between the image pixel coordinate system and the world physical coordinate system and the camera intrinsic and extrinsic parameters, calculate the vertical pixel offset of each water-land boundary pixel coordinate in the set of water-land boundary pixel coordinates; Select the coordinates of multiple key points of the water-land boundary pixel in the set of water-land boundary pixel coordinates; Using the triangular geometric relationship, based on the camera intrinsic and extrinsic parameters and the vertical pixel offset of the coordinates of multiple key points of the water-land boundary, the physical water depth of each key point of the water-land boundary is calculated. The physical water depth is calculated by averaging the coordinates of the key points of each pixel on the land-water boundary line to obtain the final measured water depth.
4. The rapid forecasting method for the transmission path of urban rainstorm and flood disaster chains capable of self-learning, as described in claim 1, is characterized in that... The construction process of the fully connected topology network is as follows: Construct a discrete surface cell grid and a discrete underground node set, wherein the discrete surface cell grid includes multiple surface cells and the discrete underground node set includes multiple underground nodes; Virtual connection edges with initial weights of zero or minimal values are established between the surface cells and the underground nodes, as well as between the underground nodes themselves, and a fully connected topology network is constructed.
5. The rapid forecasting method for the transmission path of urban rainstorm and flood disaster chains capable of self-learning, as described in claim 1, is characterized in that... Based on the discretized Manning formula and the water balance law, and according to the high-dimensional tensor of the state characteristics and the fully connected topology network, the system outputs the future water depth, real-time dynamic roughness, and virtual connection edge conduction weights of multiple surface cells in the discrete surface cell grid, including: Spatiotemporal encoding of high-dimensional hydraulic features is performed on the high-dimensional tensor of the state features to obtain a high-dimensional hidden state feature map; Decoding the high-dimensional hidden state feature map yields the real-time dynamic roughness of multiple surface cells; Based on the high-dimensional hidden state feature map, the attention coefficients of potential connection pairs in the fully connected topology network are calculated, and the attention coefficients of the potential connection pairs are converted into virtual connection edge conduction weights. The discretized Manning formula is used to calculate the flow rate between adjacent surface cells based on the real-time dynamic roughness of multiple surface cells. Based on the virtual connection edge conduction weight, calculate the potential underground transmission traffic; The future water depth of each surface cell is calculated using the water balance law based on the overflow between adjacent surface cells and the potential underground transport flow.
6. The rapid forecasting method for the transmission path of urban rainstorm and flood disaster chains capable of self-learning, as described in claim 1, is characterized in that... Also includes: The data fitting term is used to measure the error between the future water depth and the final measured water depth. The structural sparsity term is used to apply L1 regularization constraints to the conduction weights of virtual connection edges, forcing the weights of virtual connection edges without hydraulic connections to approach zero, thereby filtering out sparse underground conduction paths. The parameter smoothing term is used to apply spatiotemporal continuity constraints to the real-time dynamic roughness to prevent physical parameters from undergoing drastic, non-physical changes.
7. The rapid forecasting method for the transmission path of urban rainstorm and flood disaster chains capable of self-learning, as described in claim 1, is characterized in that... The process of outputting a visualized disaster chain transmission map based on the updated artificial intelligence model and enabling urban rainstorm and flood disaster hazard intrusion forecasting includes: Based on the target virtual connection edge conduction weight and target global water accumulation depth distribution output by the updated artificial intelligence model; Based on the target virtual connection edge conduction weights and the target global water depth distribution, a disaster chain transmission map is generated; The disaster chain transmission map is visualized to obtain a visualized disaster chain transmission map; The disaster chain cascading risk index is calculated using the target virtual connection edge conduction weight; When the cascading disaster risk index exceeds a preset threshold, an urban rainstorm and flood disaster hazard intrusion forecast is issued.
8. A self-learning rapid forecasting system for urban rainstorm and flood disaster chain transmission paths, characterized in that, include: The acquisition module is used to acquire multiple consecutive frames of images from the video stream and camera intrinsic and extrinsic parameter data, and to preprocess the multiple consecutive frames of images from the video stream based on a lightweight convolutional neural network to output a set of pixel coordinates of the land-water boundary line. The output module is used to calculate the measured water depth based on the camera's intrinsic and extrinsic parameters and the set of pixel coordinates of the land-water boundary line using triangular geometric relationships and a pinhole camera imaging model, and outputs the final measured water depth. The construction module is used to construct discrete surface cell grids and fully connected topological networks, and to convert the network topology and hydrological state data corresponding to the fully connected topological network into high-dimensional tensors of state features; The calculation module is used to output the future water depth, real-time dynamic roughness, and virtual connection edge conduction weight of multiple surface cells in the discrete surface cell grid based on the discretized Manning formula and the water balance law, according to the state feature high-dimensional tensor and the fully connected topology network. The self-learning module is used to construct a physical information composite loss function that includes data fitting terms, structural sparsity terms, and parameter smoothing terms. When a predefined sliding time window is reached, based on the mini-batch stochastic gradient descent mechanism, the physical information composite loss function is used to update the artificial intelligence model according to the final measured water depth, the future water depth of multiple surface cells, the real-time dynamic roughness, and the virtual connection edge conduction weights. The updated artificial intelligence model is determined through multiple self-learning processes. The forecasting module is used to output a visualized disaster chain transmission map based on the updated artificial intelligence model, and to realize the forecast of potential urban rainstorm and flood disaster intrusion.
9. An electronic device, characterized in that, The system includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the self-learning rapid forecasting method for urban rainstorm and flood disaster chain transmission paths as described in any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the self-learning rapid forecasting method for urban rainstorm and flood disaster chain transmission paths as described in any one of claims 1-7.
Citation Information
Patent Citations
Complex terrain flood inundation process simulation method based on cellular automaton
CN112270115A
Flood prediction method and device based on edge calculation and medium
CN116702993A
Mountain torrent forecasting method based on improved hydrological model and edge calculation
CN118095095A
Method and device for rapid determination of parameters of urban waterlogging model based on deep learning
CN119647323B
Cellular automaton model-based urban ecological risk prediction system
CN120031390A