Urban flood disaster probability early warning method based on unmanned aerial vehicle mobile edge computing

Abstract

The application provides a kind of city flood disaster probability early warning method based on unmanned aerial vehicle mobile edge computing, it is related to the field of city flood disaster prevention and control, including: monitoring device deployment, target monitoring device set generation, unmanned aerial vehicle dynamic path planning, unmanned aerial vehicle edge intelligent computing and on-site evidence, cloud early warning response.The method realizes the accurate early warning of flood disaster through unmanned aerial vehicle mobile collection and edge computing, solves the problems of systematic failure of existing flood disaster early warning methods in extreme disaster scenarios, low reliability of sensor networks, lag in early warning response, and human influence leading to loss of protection, etc., improves the response rate, avoids irreversible personnel casualties and property losses caused by untimely early warning.

Description

Technical Field

[0001] This invention relates to the field of urban flood disaster prevention and control technology, specifically to a method for probabilistic early warning of urban flood disasters based on mobile edge computing of unmanned aerial vehicles. Background Technology

[0002] Urban flooding refers to the phenomenon of water accumulation in cities caused by heavy or continuous rainfall exceeding the city's drainage capacity. It is characterized by its suddenness, rapid changes in time and space, wide range of impact, and obvious chain reactions. It mainly includes urban waterlogging, external floods / mountain floods entering the city, and storm surge floods, which can easily cause personal injury, traffic paralysis, infrastructure damage, property loss, and public health risks.

[0003] With the intensification of global climate change and the rapid advancement of urbanization, the frequency, intensity, and impact range of urban extreme rainstorm events are all showing a significant upward trend. Urban flooding has become a major public safety issue threatening people's lives and property and hindering sustainable urban development. Currently, urban flood disaster early warning mainly relies on real-time monitoring through fixed sensor networks and communication links to upload a large amount of raw sensor data to the cloud for centralized processing. However, during rainstorms, communication base stations are prone to overload, link congestion, or even interruption. Furthermore, enclosed spaces such as underground mines and parking garages suffer from wireless signal attenuation, leading to data upload delays or complete data loss, directly causing early warning failures and poor communication reliability. At the same time, existing monitoring devices generally rely on battery power, and battery life is significantly shortened in harsh environments such as underwater and humid conditions, resulting in high maintenance costs. Moreover, surface sensors are easily covered by fallen leaves and mud, or stolen or maliciously damaged, leading to data distortion or loss. Single-point failures can also directly create monitoring blind spots in the corresponding area. In addition, existing flood risk assessment and early warning models rely on historical data, which has problems such as single and outdated data sources, poor adaptability of assessment models, and delayed emergency response. They cannot operate reliably, provide accurate early warnings, or make timely decisions in extreme environments, leading to irreversible impacts or damage caused by urban floods. Summary of the Invention

[0004] To address the problems existing in the prior art, the present invention aims to provide a method for probabilistic early warning of urban flood disasters based on mobile edge computing using unmanned aerial vehicles (UAVs). This method achieves accurate early warning of flood disasters through mobile data collection by UAVs and edge computing, effectively solving problems such as systemic failure in extreme disaster scenarios, low reliability of sensor networks, delayed early warning response, and lack of protection due to human influence in existing flood disaster early warning methods. It improves response speed and avoids irreversible casualties and property losses caused by untimely early warning.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A method for probabilistic early warning of urban flooding disasters based on mobile edge computing using unmanned aerial vehicles (UAVs) includes: Step S1, Monitoring device deployment: Monitoring devices are deployed at main drainage pipes, low-lying areas along roads, main river channels, and rainfall monitoring locations. Ant colony optimization is used to achieve clustered redundant deployment and generate the optimal deployment points for sensor clusters. Step S2, Target Monitoring Device Set Generation: First, retrieve the raw meteorological radar data and obtain the real-time ground precipitation intensity field; then, use graph neural network dynamic runoff modeling to obtain the runoff volume and water accumulation risk value of each node; finally, output the target monitoring device set. Step S3, UAV dynamic path planning: First, generate the initial route of the UAV based on the improved Clark-Rattle algorithm, and then use the centralized initial route as the benchmark and adopt a decentralized bee colony self-organizing architecture for dynamic adjustment; Step S4, UAV edge intelligent computing and on-site evidence collection: First, the probability of flooding is calculated using a federated learning model, then the probability of regional flooding is calculated using a spatial Bayesian network, and finally on-site evidence collection is carried out (for verification and corroboration). Step S5, Cloud-based early warning response: First, verify the information, then determine the level of early warning.

[0007] Based on further optimization of the above scheme, in step S1, a Doppler flow meter is used to monitor the flow velocity of the main drainage pipe, an immersion level gauge or a radar level gauge is used to monitor the surface water depth at low-lying points along the road, a radar level gauge or an ultrasonic level gauge is used to monitor the water level of the main river, and a tipping bucket rain gauge is used to monitor the rainfall at the rainfall monitoring point.

[0008] Based on the further optimization of the above scheme, in step S1, for the monitoring locations of main drainage pipelines, low-lying points on roads, and main river channels, the optimal deployment points for generating sensor clusters are optimized by using the ant colony algorithm to achieve clustered redundant deployment. Specifically, the urban historical flood risk map, population density map, drainage system network map, and set of deployable locations are taken as input. First, initialize the ant colony algorithm parameters, including the number of ants. N ant pheromone volatility coefficient Pheromones importance Heuristic function importance With maximum number of iterations N ddc ; The heuristic function is: ; In the formula: R qf (j) Indicates position jHistorical flood risk value; P qf (j) Indicates position j Population density; D qf (j) Indicates position j The reciprocal of the drainage capacity; C qf (j) Indicates deployment at location j The cost; Each ant starts from a random location and chooses its next deployment location based on the transition probability: ; In the formula: Indicates time t At time, position i Arrive at the location j Pheromone concentration along the path; Indicates from position i Transfer to location j Inspirational information Indicates from position i Transfer to location s The inspirational information; allowed k Indicates the first k Only ants at all times t A set of next deployment locations that can be selected; Iteratively update pheromones: ; In the formula: Ants k In the path ij The pheromones released from above; The iteration terminates when the set of all deployable locations is empty, the preset maximum number of sensors is reached, or the maximum number of iterations is reached. The optimal deployment scheme is selected, and sensors with a distance of less than 200m are grouped into a cluster, and the optimal sensor deployment set is output. Each sensor cluster is .

[0009] Configure a differential sampling synchronization mechanism for each sensor cluster: that is, the time synchronization error of the sensors within the cluster is less than 10ms. When the reading deviation of more than two sensors within the cluster exceeds 20%, it is automatically judged as a sensor failure and abnormal data is removed.

[0010] Based on further optimization of the above scheme, in step S1, the arrangement of rainfall monitoring sites is specifically as follows: First, acquire existing rainfall monitoring devices in the city and preset two sampling intervals. L cy(Used to select representative points within a city) and coverage radius R fg (Two parameters used to determine whether the monitoring devices are dense enough); then, within the city limits, according to the sampling interval L cy A batch of sampling points is generated. For each sampling point, its distance to the nearest rainfall monitoring device is calculated (initially, the only rainfall monitoring devices available in the city are those already in use). Then, the distances are sorted from largest to smallest, and the sampling point with the largest distance is selected. l max Its maximum distance is d max ;like d max ≤ R fg This indicates that the density of urban rain gauges meets the requirements and the deployment is complete. d max > R fg This indicates areas with insufficient monitoring, requiring the installation of additional rainfall monitoring devices; in this case, priority should be given to sampling points. l max Rainfall monitoring devices are installed at the sampling points. l max If rainfall monitoring devices cannot be installed at the location, then the distance from the sampling point should be... l max The nearest and the distance is less than or equal to R fg Install rainfall monitoring devices at the location; During the deployment of rainfall monitoring sites, the priority of ant colony algorithm site selection is increased, prioritizing deployment in areas with a history of high flood risk.

[0011] Based on further optimization of the above scheme, in step S2, the inversion of the original meteorological radar data and the acquisition of the real-time ground precipitation intensity field specifically involves: calculating the ground precipitation intensity using variational Doppler radar inversion based on the original reflectivity data Z from the local S-band meteorological radar. ; In the formula: This indicates the proportionality coefficient (reflecting the overall concentration and size distribution of precipitation particles). Indicates the relationship index (reflecting the spectral distribution characteristics of precipitation particles); When the drone retrieves ground rainfall data y rain,i At that time, the relationship proportional coefficient and relationship index are updated in real time using the least squares method: ; In the formula: yrain,i Indicates the first i The measured rainfall intensity at each rain gauge station Z i Indicates the first i Linear values ​​of radar reflectivity at the corresponding locations of each rain gauge station N ture This indicates the number of validly calibrated rain gauges; The final output shows precipitation forecast fields for the next 15 minutes, 30 minutes, and 60 minutes. .

[0012] Based on further optimization of the above scheme, in step S2, the specific steps for obtaining the flow rate and water accumulation risk value of each node using graph neural network dynamic flow modeling are as follows: First, the coordinates of all manholes, drainage outlets, and river confluences are extracted from the Geographic Information System (GIS) data to serve as base nodes. Then, the locations of all monitoring devices are matched to the nearest base node to form a node set. V Based on the flow direction of the pipeline and the river, a set of directed edges E is formed: if there exists a node... i To the node j If the water flow path is such that an edge is added... e ij ∈ E ; Calculate initial weights based on pipeline / river channel parameters: ; In the formula: C ij Representing an edge e ij The corresponding design drainage capacity of the pipeline / river channel C max This represents the maximum design drainage capacity of all edges; n cx Represents the roughness coefficient. A mj Indicates the cross-sectional area of ​​the water passage. R water Indicates the hydraulic radius. S pd Indicates the slope of the pipe / river channel; Initial adjacency matrix: ; Then, the gridded precipitation intensity is matched to each node. Obtain each node v At any moment t Feature vectors: ; In the formula: vpipe (t) This represents the pipeline flow velocity monitoring value. h water (t) This represents the monitoring value of surface water depth. h river (t) This indicates the river water level monitoring value; Z-score normalization is performed on each feature dimension to eliminate dimensions: ; In the formula: They represent the first f Historical mean and standard deviation of each feature; Obtain the past T in The feature input tensor for the time period: ; Subsequently, a Spatio-Temporal Graph Convolutional Network (ST-GCN) model was constructed, consisting of three spatio-temporal convolutional blocks and two fully connected layers. Each spatio-temporal convolutional block is composed of a graph convolutional layer, a temporal convolutional layer, batch normalization, and dropout. ; in, Dropout The ratio is 0.3 to prevent overfitting; Convolutional layer GraphConv() Used to capture the spatial correlation of drainage system topology: ; In the formula: Indicates the first l The output feature matrix of the layer; W(t) Represents a dynamic adjacency matrix; Add a self-loop adjacency matrix. I Represents the identity matrix. express The degree matrix; Indicates the first l Layer-learnable weight matrix; Represents the ReLU activation function; Temporal convolutional layer TimeConv() To capture the temporal correlation of water flow evolution, 1D convolution is used: ; In the formula: Conv1D() This represents a 1D convolution operation; Indicates the first l Layer convolution kernel weights; After flattening the output of the last spatiotemporal convolutional block, the sink flow prediction is obtained by passing it through two fully connected layers: ; ; In the formula: t x Indicates the index of the prediction time step; H (3) This represents the output feature matrix of the third spatiotemporal convolutional block; FC1() This indicates the first fully connected layer. FC2() This indicates the second fully connected layer; These represent the weight matrices of the corresponding fully connected layers. b 1. b 2 represents the bias term of the corresponding fully connected layer; Then, for each edge e ij Calculate its real-time drainage capacity coefficient: ; In the formula: v design,i Represents a node i The corresponding design flow rate of the pipeline; if This indicates that the pipeline may be severely blocked and its drainage capacity may be significantly reduced. And calculate the historical correction coefficient based on the historical congestion records of this edge: ; In the formula: N block,ij This indicates the number of times the edge became congested in the past year. N total,ij Indicates the total number of monitoring sessions; Update dynamic edge weights: ;but: ; Next, the sink flow rate output by the inverse normalization model (the sink flow rate output by the model is the normalized value): ; In the formula: These represent the historical mean and standard deviation of the inflow rate, respectively. Calculate the risk value of water accumulation: ; In the formula: C v-max Represents a node v Maximum drainage capacity; when Risk v (t + t x )A value less than 0.5 indicates sufficient drainage capacity (low risk); a value less than or equal to 0.5 indicates insufficient drainage capacity (low risk). Risk v (t + t x ) A value less than 1 indicates that drainage capacity is strained (medium risk). Risk v (t + t x ) A value of ≥1 indicates water accumulation (high risk). According to node priority Priority v The sorting provides a basis for subsequent drone mission scheduling: ; In the formula: Pop v Represents a node v Normalized population density of the covered area; Finally, model training and updates are performed, specifically including: inputting the feature sequence of the past 10 minutes as input samples, labeling them as the actual sink flow of the next 15 minutes (calculated from monitoring data), and dividing them into training, validation, and test sets in a 7:2:1 ratio, with mean squared error (MSE) as the loss function. ; In the formula: Represents a node v exist t + t x The actual flow rate at any given moment; Indicates the size of the node set V; t out This represents the time step of the model's output prediction of future runoff volume; the Adam optimizer is used with an initial learning rate of 1e-3 and a learning rate decay coefficient of 0.9, decaying once every 10 epochs; every morning, the model is fine-tuned using the previous day's new data to update the model parameters, ensuring that the model can adapt to seasonal changes and the long-term evolution of the drainage system.

[0013] Based on further optimization of the above scheme, in step S2, the output target monitoring device set specifically refers to: the precipitation forecast field... Input a digital twin to preview the development of urban flooding in the next 15 minutes, 30 minutes, and 60 minutes, and generate a risk evolution heat map; Nodes with a water accumulation risk value not less than the water accumulation risk threshold are selected. Add the corresponding monitoring devices to the target set; select monitoring devices less than 500m away from high-risk areas and add them to the target set; for each river, calculate the average rainfall intensity of its catchment area, and if it is greater than 20mm / h, add all monitoring devices on that river to the target set. Prioritize each target monitoring device The target set is then sorted from highest to lowest priority to obtain the target monitoring device set: .

[0014] Based on further optimization of the above scheme, in step S3, generating the initial route of the UAV based on the improved Clack-Rattle algorithm specifically involves: First, the target monitoring device set is as follows , among which, the i The coordinates of the monitoring devices are ( x i ,y i ); drone swarms are The drone take-off and landing point is B The coordinates are ( x b ,y b The maximum flight time of the drone is t UAV-max The drone hovering time above each monitoring device is t UAV-a The drone's cruising speed is v UAV The wireless communication radius is R comm The laser communication radius is R laser ; For any two points The distance between them is: ; drones from o a Fly to o b The time is: ; The total time for any route is: ,in, t flight The flight time is represented by the flight time for each segment of the journey. The result is obtained by addition; N q Indicates the number of monitoring devices accessed along the flight path; Then, the set Each monitoring device is treated as an independent single-point route: for any monitoring device Generate the initial route r i The access sequence is r i =[ a i The actual flight distance of this route is: B → a i → B The initial set of routes is then: R init ={ r 1 ,r 2 ,…,r Nj}, Nj Indicates the number of monitoring devices; Then, the two routes r i =[ a i ]and r j =[ a j Merged into one route r ij =[ a i , a j ](voyage B → a i → a j → B The saved flight distance is the amount saved. S(i, j) : ; Traverse all disordered monitoring device pairs ( a i , a j )( i < j (To avoid double counting), calculate the savings for each pair. S(i, j) Then, sort all point pairs in descending order of savings to obtain a list of point pairs to be merged: ; In the formula: M represents the total number of point pairs, and S(i1, j1) ≥ S(i2, j2) ≥ … ≥ S(iM, jM) ; Finally, route merging is performed, specifically including: Step 1, from the list of point pairs to be merged... Point In the middle, take the pair with the maximum current savings, that is... Step 2, if and If the point is already on the same route, then the point pair will be moved from... Point Remove from the list and return to step 1; Step 3, if... and If not all points are endpoints of their respective routes, then pair that point with the other point. Point Remove from the list and return to step 1; Step 4, Assume it includes The route is denoted as route 1, and its access sequence is sequence 1, which includes... The route is denoted as route 2, and its access sequence is sequence 2: like a i It is the end point of sequence 1 and a j If it is the starting point of sequence 2: then sequence 1 is directly appended to the beginning of sequence 2 to form "sequence 1 + sequence 2"; like a i The starting point of sequence 1 and a j If it is the end point of sequence 2: then sequence 2 is directly appended to the beginning of sequence 1 to form "sequence 2 + sequence 1"; like a i , a j Since each sequence is the starting point of its own sequence, sequence 1 is first reversed and then concatenated with sequence 2 to form "reverse sequence 1 + sequence 2". like a i , a j Since each of these is the end point of its own sequence, we first reverse sequence 2 and then concatenate it with sequence 1 to form "sequence 1 + reverse sequence 2". Step 5: Calculate the total time for the merged new route. t hb-total ,like t hb-total ≤ t UAV-max If the condition is met, the pair will be merged and removed from the "Pair List"; if... t hb-total > t UAV-max If the merger is abandoned, the original route will be retained; Step 6: Repeat steps 1 through 5 until the point-to-point list is empty, obtaining the final set of initial routes. ; in, mroad The number of routes after merging is calculated, with each route corresponding to an initial mission for one drone, meaning that at least one drone needs to be deployed in this round of missions. m road A drone.

[0015] Based on the further optimization of the above scheme, in step S3, the centralized initial route is used as a benchmark, and a decentralized bee colony self-organizing architecture is adopted for dynamic adjustment, specifically as follows: First, all drones form a self-organizing network through a dual-link communication system of "wireless communication + laser communication" to share mission status and environmental information; An auction-based task allocation mechanism is adopted: When the risk value at a certain monitoring point suddenly rises above the preset risk threshold Risk th2 At that time, the task request is broadcast to the entire network; nearby drones calculate the bid based on remaining battery power, distance, and current task load: ; In the formula: Indicates drone u i Normalized remaining charge; Load(u i ) Indicates drone u i Task load factor; The drone that bids the highest wins the mission, while the other drones automatically adjust their routes. Then, multi-agent deep reinforcement learning (MADRL) is used for dynamic path planning: State space: ,in, pos u This indicates the location of the drone. E u Indicates the remaining battery power. Risk a This indicates the risk value at each monitoring point. Obstacle This represents visual obstacle avoidance data. pos other Indicates the location of other drones; Action space: That is, flight speed and the monitoring point for the next visit next a ; Reward function: ,in, I(a i-end ) This indicates the task completion indicator function (i.e., the drone has arrived). ai (1 indicates data retrieval is complete, 0 indicates otherwise). t d-total This represents the total time taken for a single step (calculated from the drone's flight speed and distance). Collsion This represents the collision indication function. These represent the corresponding weight coefficients; The model uses the PPO algorithm, is pre-trained in a digital twin environment, and deployed on each drone edge node; When the communication range is insufficient, the tethered drone is automatically dispatched to a height of 150m to serve as a communication relay station.

[0016] Based on further optimization of the above scheme, in step S4, the federated learning model calculates the probability of flooding as follows: First, the drone flies above the monitoring device, establishes a connection via short-range communication, and retrieves historical data. t - t history ,t Observation sequence within ] Y , t history Indicates the length of the history window; A lightweight isolated forest algorithm is deployed at the edge to perform anomaly detection and calculate anomaly score for each data point: ; In the formula: E(h(y k )) Representing data points y k Average path length in an isolated tree c qw (E) This represents the expected value of the average path length. Data points with outlier scores greater than 0.7 were removed to obtain the preprocessed observation sequence. Y’ ; Then, the local federated learning model is called. Model local And calculate the statistics: Mean: ;in, This indicates the number of valid samples after preprocessing. Standard deviation: ; Exceeding threshold ratio: For surface water depth monitoring devices, river water level monitoring devices, and rainfall monitoring devices: For drainage pipe flow velocity monitoring devices: ;in, y th Indicates the threshold for flood assessment ，l()Indicates an indicator function; Next, calculate the overall risk score: ; In the formula: Indicates the weighting coefficient. This indicates a smoothing term to prevent the denominator from being zero; Finally, the overall risk score will be considered. S zhfx Mapped to the probability of flooding: ; In the formula: k flood Indicates the slope parameter; After each day's task is completed, each drone uploads its local model gradient updates to the cloud, where the cloud aggregates and updates the global model. ; In the formula: K UAV This indicates the number of drones participating in this round of federal aggregation.

[0017] Based on further optimization of the above scheme, in step S4, calculating the probability of regional flooding using a spatial Bayesian network specifically involves: Construct a spatial Bayesian network, where nodes represent the flood probability of each monitoring point and edge weights represent spatial correlation. g ij Calculate conditional probability: ; In the formula: Nir i Represents a node i The set of neighboring nodes; Indicates the symbol for consecutive products; g ij This indicates the spatial correlation between monitoring points i and j; The probability of obtaining a regional level through fusion: ; In the formula: This represents the probability weight.

[0018] Based on further optimization of the above scheme, in step S4, the on-site evidence collection specifically involves: when P flood (t) When the value is greater than 0.5, the drone automatically adjusts its flight altitude to 20m, and takes three visible light photos, one near-infrared photo, and one line structured light scan data from different angles to complete on-site evidence collection. The final output package is then completed. , ID iIndicates the monitoring device ID. type i Indicates the type of monitoring device. t Indicates the task timestamp. P flood (t) Indicates the probability of flooding at a single point. P region (t) Indicates the probability of regional flooding. Photo data This refers to the data from photos taken at the scene as evidence.

[0019] Based on the further optimization of the above scheme, in step S5, the verification is specifically as follows: for result packages with on-site photos, manual verification is performed in conjunction with the photos; if the photos clearly show no flooding, the monitoring device is determined to be abnormal and corresponding maintenance is arranged. Meanwhile, this result package is ignored in this process. The tiered early warning determination is specifically based on a set of preset thresholds in the cloud. 0<P th1 <P th2 <P th3 <P th4 <1 And classify the warning levels: Level 1 Warning (Blue Warning): P flood ∈[ P th1 , P th2 ); Level II Warning (Yellow Warning): P flood ∈[ P th2 , P th3 ); Level 3 Warning (Orange Warning): P flood ∈[ P th3 , P th4 ); Level IV Warning (Red Alert): P flood ∈[ P th4 , 1 ]; Similarly, the probability of regional flooding P region The same threshold system is also used to classify the warning levels.

[0020] The following are the technical effects of the present invention: This invention, during the deployment of monitoring devices, combines multi-objective optimization using ant colony optimization, clustered redundant deployment, differential sampling synchronization, and iterative deployment of rain stations. This not only solves the monitoring blind spots caused by single-point failures but also avoids problems such as inconsistent sensor data, misjudgments due to failures, and insufficient rainfall monitoring density and poor representativeness. Through the target monitoring device aggregation generation step, it uses meteorological radar to invert precipitation intensity and generate real-time precipitation forecast fields, avoiding the problems of historical data lag and inaccurate precipitation input. It uses ST-GCN dynamic runoff modeling to predict the runoff volume and water accumulation risk value of each node, solving the problem that traditional models cannot adapt to complex pipe network topologies. It uses a digital twin model to generate a risk evolution heat map and screen target monitoring devices, effectively avoiding the problems of delayed early warning response and inability to identify high-risk points in advance. By employing dynamic path planning for drones, an improved Clark-Rattle algorithm is used to generate initial routes and minimize flight time, avoiding issues such as low initial scheduling efficiency and significant wasted endurance. A decentralized swarm self-organizing and auction mechanism is used to allocate unexpected tasks, combining centralized initial scheduling with distributed self-organizing adjustments. This avoids the slow response and inability to handle sudden risk changes inherent in centralized scheduling, as well as communication link interruptions and single-unit failures leading to task failure. Edge intelligent computing and on-site evidence collection are used to calculate flood probabilities using federated learning models, avoiding the response lag and complete failure of early warnings during communication interruptions caused by centralized cloud computing. Spatial Bayesian networks are used to fuse terrain, pipelines, roads, rivers, and historical data to calculate regional flood probabilities, avoiding the inability of single-point monitoring to reflect regional cascading risks and the inability of static models to adapt to temporary changes in operating conditions. Cloud-based early warning response not only avoids abnormal monitoring caused by sensor false alarms, malicious damage, or theft, but also achieves accurate early warning responses at both the single-point and regional levels.

[0021] The early warning method of this invention can significantly improve communication reliability, enhance the robustness of sensor networks, accelerate early warning response speed, and improve prediction accuracy. At the same time, the system has strong adaptive capabilities, can conduct evidence collection at high-risk points, and supports manual verification, thereby supporting emergency command and dispatch. It effectively solves the problems of communication failure, lack of perception, model lag, and response delay in existing urban flood early warning methods under extreme weather conditions. Attached Figure Description

[0022] Figure 1 This is a flowchart of the urban flood disaster probability early warning method in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below. In the following description, specific details such as specific system structures and technologies are presented for illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention.

[0024] Example 1: A method for probabilistic early warning of urban flooding disasters based on mobile edge computing using unmanned aerial vehicles (UAVs) includes: Step S1: Deployment of monitoring devices: Monitoring devices are deployed at main drainage pipes, low-lying areas along roads, main river channels, and rainfall monitoring locations. Doppler flow meters are used to monitor the flow velocity in the main drainage pipes. Submersible level gauges or radar level gauges are used to monitor the surface water depth at low-lying areas along roads. Radar level gauges or ultrasonic level gauges are used to monitor the water level in the main river channels. Tipping bucket rain gauges are used to monitor the rainfall at the rainfall monitoring locations.

[0025] For monitoring locations of main drainage pipelines, low-lying areas along roads, and main rivers, the ant colony algorithm is used to optimize clustered redundant deployment and generate the optimal deployment points for sensor clusters (that is, sensors less than 200m apart are divided into a cluster, and devices within the same cluster are redundant backups of each other). Specifically, the urban historical flood risk map, population density map, drainage system network map, and set of deployable locations are used as inputs. First, initialize the ant colony algorithm parameters, including the number of ants. N ant (Generally 50), pheromone volatility coefficient (Generally 0.1), pheromone importance (Generally 1), Heuristic function importance (Generally 2) and the maximum number of iterations N ddc (Usually 100); The heuristic function is: ; In the formula: R qf (j) Indicates position j Historical flood risk value; P qf (j) Indicates position j Population density; D qf (j) Indicates position j The reciprocal of the drainage capacity; C qf (j) Indicates deployment at location j The cost; Each ant starts from a random location and chooses its next deployment location based on the transition probability: ; In the formula: Indicates time t At time, position i Arrive at the location j Pheromone concentration along the path; Indicates from position i Transfer to location j The heuristic information (in this scheme, the core objective of the ant colony algorithm is to select the sensor deployment sites with the highest comprehensive value from the set of deployable locations; the "transfer" behavior of ants is from the selected deployment sites) i Select the next location to deploy. j During this process, the position j The deployment value is determined solely by its own attributes and the currently selected location. i Irrelevant, each candidate position j The deployment value is determined by a combination of its inherent attributes, including historical flood risk, population density, reciprocal drainage capacity, and deployment cost. These attributes are related to the location. j It is inherent in itself, and from which position i The "transfer" has nothing to do with it; it's about evaluating whether to choose a location. j At that time, only need to consider j It is itself, without needing to consider its relationship with the previous position. i The "distance" or "connection cost" between them. Therefore, from location i Transfer to location j The heuristic information depends only on the candidate location. j The deployment value, namely ); Indicates from position i Transfer to location s Heuristic information (reusing heuristic functions, i.e.) ); allowed k Indicates the first k Only ants at all times t The next set of deployment locations that can be selected (obtained from the set of deployable locations, initially) allowed k =Set of deployable locations; if the ant does not select a location. j , will j from allowed k Remove, when allowed k The search for this ant ends when the number of sensors is empty or the preset total number of sensors is reached. Iteratively update pheromones: ; In the formula: Ants k In the path ij The pheromones released from the ants (only when ants release pheromones) k The deployment scheme includes paths ij The time and time are not zero; it equals the pheromone intensity constant divided by the objective function value of the deployment scheme built by Antminer k, where the objective function value of the deployment scheme built by Antminer k includes a comprehensive evaluation value of total deployment cost, coverage risk, etc. The iteration terminates when the set of all deployable locations is empty, the preset maximum number of sensors is reached, or the maximum number of iterations is reached (any one of these conditions is sufficient). The optimal deployment scheme is selected, and sensors with a distance of less than 200m are grouped into a cluster, and the optimal sensor deployment set is output. Each sensor cluster is .

[0026] A differential sampling synchronization mechanism is configured for each sensor cluster: the time synchronization error of the sensors within the cluster is less than 10ms. When the reading deviation of more than two sensors within the cluster exceeds 20%, it is automatically determined as a sensor failure and abnormal data is removed. (That is, all sensors within the cluster achieve time synchronization through differential GPS or Beidou satellite, with an error of less than 10ms; when the reading deviation of more than two sensors within the same cluster exceeds 20%, it is automatically determined as a sensor failure and abnormal data is removed.)

[0027] The deployment of monitoring devices for main drainage pipelines, low-lying areas of roads, and main waterways is optimized using an ant colony algorithm to achieve clustered redundancy and generate the optimal deployment points for sensor clusters.

[0028] The specific deployment of rainfall monitoring sites is as follows: First, acquire information on existing rainfall monitoring devices in the city and pre-set two sampling intervals. L cy (Used to select representative points within a city) and coverage radius R fg (Two parameters used to determine whether the monitoring devices are dense enough); then, within the city limits, according to the sampling interval L cy Generate a batch of sampling points (specifically, this includes: firstly, extracting and obtaining vector boundary data of the urban built-up area from a Geographic Information System (GIS) as the effective range for sampling; then, based on sampling intervals...). L cyUsing the side length as an example, a square or rectangular grid is generated within the aforementioned effective range, and all vertices of the grid become the initial sampling points. Then, the initial sampling points are filtered, removing those falling into undeployable areas (such as water bodies, mountains, military restricted areas, permanent basic farmland, etc.), thus obtaining the sampling point set. For each sampling point, its distance to the "nearest rainfall monitoring device" is calculated (at the initial execution, the rainfall monitoring device is limited to those already existing in the city). Then, the distances are arranged from largest to smallest, obtaining the sampling point with the largest distance. l max Its maximum distance is d max ;like d max ≤ R fg This indicates that the density of urban rain gauges meets the requirements and the deployment is complete. d max > R fg This indicates areas with insufficient monitoring, requiring the installation of additional rainfall monitoring devices; in this case, priority should be given to sampling points. l max Rainfall monitoring devices are installed at the sampling points. l max If rainfall monitoring devices cannot be installed at the location, then the distance from the sampling point should be... l max The nearest and the distance is less than or equal to R fg Install rainfall monitoring devices at the designated location (if this location cannot be found, it indicates the sampling point). l max The harsh geographical conditions at the location prevented the deployment of rainfall monitoring devices in large surrounding areas, thus failing to meet the required standards. d max ≤ R fg Under these conditions, sampling points should be directly removed. l max (Because there is no large-scale human activity in this area, flooding will not have a serious impact.) During the deployment of rainfall monitoring stations, the ant colony algorithm is prioritized for site selection, with priority given to deployment in historically high-risk flood areas. Specifically, the deployment of rainfall monitoring stations uses the output of the ant colony algorithm as the site selection priority, combined with the aforementioned coverage radius and sampling interval for density verification (coverage radius...). R fg and sampling interval L cyUsed to quantitatively determine whether the density of rain gauges meets the monitoring requirements. When there is insufficient monitoring area, high-value locations selected by the ant colony algorithm are prioritized for deployment. The two are complementary to "priority determination" and "density verification", rather than substitutes for each other.

[0029] Step S2, Target Monitoring Device Assembly Generation: First, the raw meteorological radar data is inverted to obtain the real-time ground precipitation intensity field. Specifically, based on the raw reflectivity data Z of the local S-band meteorological radar (which can be obtained through the National Meteorological Science Data Center of the China Meteorological Administration, the data sharing platform of the provincial meteorological bureau, or the local radar receiving system), the ground precipitation intensity is calculated using variational Doppler radar inversion. ; In the formula: The proportionality coefficient represents the overall concentration and size distribution of precipitation particles, typically ranging from 100 to 400. The relation index (reflecting the spectral distribution characteristics of precipitation particles, generally 1.2 to 2); When the drone retrieves ground rainfall data y rain,i At that time, the relationship proportional coefficient and relationship index are updated in real time using the least squares method: ; In the formula: y rain,i Indicates the first i The measured rainfall intensity at each rain gauge station Z i Indicates the first i Linear values ​​of radar reflectivity at the corresponding locations of each rain gauge station N ture This indicates the number of validly calibrated rain gauges; The final output shows precipitation forecast fields for the next 15 minutes, 30 minutes, and 60 minutes. Specifically, it includes: First, based on the local S-band weather radar, a set of volume scan reflectivity factor data is acquired every 5 minutes. Z(x, y, t) ,in,( x, y ) represents spatial grid coordinates, t Indicates the scan time; generates a time-continuous reflectance factor sequence.

[0030] Then, the sparse iterative optical flow method (a conventional method in radar echo extrapolation and nowcasting of precipitation, including Shi-Tomasi corner detection + Lucas-Kanade local tracking + pyramid hierarchical iteration + semi-Lagrange extrapolation) is used to extract discrete points with obvious rainfall intensity gradients from radar images as "feature points," i.e., key points most likely to reflect the echo movement trend; the Lucas-Kanade optical flow algorithm is used to obtain the displacement vector of each feature point between consecutive frames; the original image is progressively scaled down by constructing an image pyramid, first estimating the approximate motion displacement on the small-scale image, and then upsampling and refining the calculation results layer by layer, finally obtaining sub-pixel level motion vector accuracy; based on the tracked feature point motion trajectory, linear extrapolation is performed to the positions of future prediction times) to calculate the motion field of the reflectivity factor images at adjacent time points, obtaining the motion vector of each grid point. It indicates that the echo is in x and y Speed ​​of movement in a certain direction; Then, based on the current moment t Reflection factor field of 0 Z(x, y, t 0 ) and motion vector field ( u, v The reflectivity factor field at future times Δt (i.e., 15 min, 30 min, and 60 min) is calculated using the semi-Lagrange extrapolation method. ; Finally, the extrapolated reflectivity factor field is combined with the relational parameters calibrated in real time using the least squares method. , Output precipitation forecast field: .

[0031] Then, graph neural network dynamic flow modeling is used to obtain the flow rate and water accumulation risk value of each node, specifically: First, the coordinates of all manholes, drainage outlets, and river confluences are extracted from the Geographic Information System (GIS) data to serve as base nodes. Then, the locations of all monitoring devices are matched to the nearest base node to form a node set. V Based on the flow direction of the pipeline and the river, a set of directed edges E is formed: if there exists a node... i To the node j If the water flow path is such that an edge is added... e ij ∈ E ; Calculate initial weights based on pipeline / river channel parameters: ; In the formula:C ij Representing an edge e ij The corresponding design drainage capacity of the pipeline / river channel C max This represents the maximum design drainage capacity of all edges; n cx This represents the roughness coefficient (0.013 for concrete pipes and 0.03 for earthen canals). A mj Indicates the cross-sectional area of ​​the water passage. R water Indicates the hydraulic radius. S pd Indicates the slope of the pipe / river channel; Initial adjacency matrix: ; Then, the gridded precipitation intensity is matched to each node. Obtain each node v At any moment t Feature vectors: ; In the formula: v pipe (t) This represents the pipeline flow velocity monitoring value. h water (t) This represents the monitoring value of surface water depth. h river (t) This indicates the river water level monitoring value; Z-score normalization is performed on each feature dimension to eliminate dimensions: ; In the formula: They represent the first f Historical mean and standard deviation of each feature; Obtain the past T in The feature input tensor for the time period: ; Subsequently, a Spatio-Temporal Graph Convolutional Network (ST-GCN) model was constructed, consisting of three spatio-temporal convolutional blocks and two fully connected layers. Each spatio-temporal convolutional block is composed of a graph convolutional layer, a temporal convolutional layer, batch normalization, and dropout. ; in, Dropout The ratio is 0.3 to prevent overfitting; Convolutional layer GraphConv() Used to capture the spatial correlation of drainage system topology: ; In the formula: Indicates the first l The output feature matrix of the layer; W(t) Represents a dynamic adjacency matrix; Add a self-loop adjacency matrix. I Represents the identity matrix. express The degree matrix; Indicates the first l Layer-learnable weight matrix; Represents the ReLU activation function; Temporal convolutional layer TimeConv() To capture the temporal correlation of water flow evolution, 1D convolution is used: ; In the formula: Conv1D() This represents a 1D convolution operation (kernel size is 3, stride is 1). Indicates the first l Layer convolution kernel weights; After flattening the output of the last spatiotemporal convolutional block, the sink flow prediction is obtained by passing it through two fully connected layers: ; ; In the formula: t x This represents the prediction time step index (i.e., the prediction time from the current time t; generally). t x =1,2,…,15); H (3) This represents the output feature matrix of the third spatiotemporal convolutional block; FC1() This indicates the first fully connected layer. FC2() This indicates the second fully connected layer; These represent the weight matrices of the corresponding fully connected layers. b 1. b 2 represents the bias term of the corresponding fully connected layer; Then, for each edge e ij Calculate its real-time drainage capacity coefficient: ; In the formula: v design,i Represents a node i The corresponding design flow rate of the pipeline; if This indicates that the pipe may be severely blocked and its drainage capacity may be significantly reduced (if the side...). e ij Without real-time monitoring data, (Only used for historical correction coefficients) And calculate the historical correction coefficient based on the historical congestion records of this edge: ; In the formula: N block,ij This indicates the number of times the edge became congested in the past year. N total,ij Indicates the total number of monitoring sessions; Update dynamic edge weights: ;but: ; Next, the sink flow rate output by the inverse normalization model (the sink flow rate output by the model is the normalized value): ; In the formula: These represent the historical mean and standard deviation of the inflow rate, respectively. Calculate the risk value of water accumulation: ; In the formula: C v-max Represents a node v Maximum drainage capacity; when Risk v (t + t x ) A value less than 0.5 indicates sufficient drainage capacity (low risk); a value less than or equal to 0.5 indicates insufficient drainage capacity (low risk). Risk v (t + t x ) A value less than 1 indicates that drainage capacity is strained (medium risk). Risk v (t + t x ) A value of ≥1 indicates water accumulation (high risk). According to node priority Priority v The sorting provides a basis for subsequent drone mission scheduling: ; In the formula: Pop v Represents a node v Normalized population density of the covered area (by normalizing to the [0,1] interval); Finally, model training and updates are performed, specifically including: inputting the feature sequence of the past 10 minutes as input samples, labeling them as the actual sink flow of the next 15 minutes (calculated from monitoring data), and dividing them into training, validation, and test sets in a 7:2:1 ratio, with mean squared error (MSE) as the loss function. ; In the formula: Represents a node v exist t + t x The actual flow rate at any given time (calculated from historical monitoring data and the status of the drainage system); Indicates the size of the node set V; t out This represents the number of time steps (typically 15) for the model's future runoff prediction; the Adam optimizer is used with an initial learning rate of 1e-3 and a learning rate decay factor of 0.9, decaying once every 10 epochs; every morning, the model is fine-tuned using the previous day's new data to update the model parameters, ensuring that the model can adapt to seasonal changes and the long-term evolution of the drainage system.

[0032] Finally, the output target monitoring device set is as follows: the precipitation forecast field Input a digital twin to preview the development of urban flooding in the next 15 minutes, 30 minutes, and 60 minutes, and generate a risk evolution heat map; Nodes with a water accumulation risk value not less than the water accumulation risk threshold are selected. ( Risk th1 (Generally 0.3), add the corresponding monitoring device to the target set; screen monitoring devices less than 500m away from high-risk areas and add them to the target set; for each river, calculate the average precipitation intensity of its catchment area, and if it is greater than 20mm / h, add all monitoring devices on that river to the target set; Prioritize each target monitoring device The target set is then sorted from highest to lowest priority to obtain the target monitoring device set: .

[0033] Step S3, UAV dynamic path planning: First, generate the initial route for the UAV based on the improved Clarke-Wright algorithm (the improvement lies in introducing UAV endurance constraints and hovering time constraints to avoid generating routes that exceed the endurance capacity), specifically: First, the target monitoring device set is as follows , among which, the i The coordinates of the monitoring devices are ( xi ,y i ); drone swarms are The drone take-off and landing point is B The coordinates are ( x b ,y b The maximum flight time of the drone is t UAV-max (Generally 30 minutes), the drone's hovering time above each monitoring device is... t UAV-a (Generally 1 minute, hovering time includes data retrieval and analysis time), the drone's cruising speed is... v UAV (Generally 10m / s), the wireless communication radius is R comm (Generally 1km), the laser communication radius is R laser (Usually 2km); For any two points The distance between them is: ; drones from o a Fly to o b The time is: ; The total time for any route is: ,in, t flight The flight time is represented by the flight time for each segment of the journey. The result is obtained by addition; N q Indicates the number of monitoring devices accessed along the flight path; Then, the set Each monitoring device is treated as an independent single-point route: for any monitoring device Generate the initial route r i The access sequence is r i =[ a i The actual flight distance of this route is: B → a i → B The initial set of routes is then: R init ={ r 1 ,r 2 ,…,rNj}, Nj Indicates the number of monitoring devices; Then, the two routes r i =[ a i ]and r j =[ a j Merged into one route r ij =[ a i , a j ](voyage B → a i → a j → B The saved flight distance is the amount saved. S(i, j) (The greater the savings, the more likely it is to be effective) a i and a j The more "convenient" the route is within the same line, the greater the reduction in total flight distance after merging, and the higher the priority. ; Traverse all disordered monitoring device pairs ( a i , a j )( i < j (To avoid double counting), calculate the savings for each pair. S(i, j) Then, sort all point pairs in descending order of savings to obtain a list of point pairs to be merged: ; In the formula: M represents the total number of point pairs, and S(i1, j1) ≥ S(i2, j2) ≥ … ≥ S(iM, jM) ; Finally, route merging is performed, specifically including: Step 1, from the list of point pairs to be merged... Point In the middle, take the pair with the maximum current savings, that is... Step 2, if and If the point is already on the same route, then the point pair will be moved from... Point Remove from the list and return to step 1; Step 3, if... and If not all of them are endpoints of their respective routes (an endpoint is either the first or last monitoring device in a route access sequence), then the point will be paired with the endpoint from... Point Remove from the list and return to step 1; Step 4, Assume it includes The route is denoted as route 1, and its access sequence is sequence 1, which includes... The route is denoted as route 2, and its access sequence is sequence 2: like a i It is the end point of sequence 1 and a j If it is the starting point of sequence 2: then sequence 1 is directly appended to the beginning of sequence 2 to form "sequence 1 + sequence 2"; like a i The starting point of sequence 1 and a j If it is the end point of sequence 2: then sequence 2 is directly appended to the beginning of sequence 1 to form "sequence 2 + sequence 1"; like a i , a j Since each sequence is the starting point of its own sequence, sequence 1 is first reversed and then concatenated with sequence 2 to form "reverse sequence 1 + sequence 2". like a i , a j Since each of these is the end point of its own sequence, we first reverse sequence 2 and then concatenate it with sequence 1 to form "sequence 1 + reverse sequence 2". Step 5: Calculate the total time for the merged new route. t hb-total ,like t hb-total ≤ t UAV-max If the condition is met, the pair will be merged and removed from the "Pair List"; if... t hb-total > t UAV-max If the merger is abandoned, the original route will be retained; Step 6: Repeat steps 1 through 5 until the point-to-point list is empty, obtaining the final set of initial routes. ; in, m road The number of routes after merging is calculated, with each route corresponding to an initial mission for one drone, meaning that at least one drone needs to be deployed in this round of missions. m road A drone.

[0034] Then, using the centralized initial route as a baseline, a decentralized swarm self-organizing architecture (i.e., an organizational structure where drones negotiate in a distributed manner rather than relying on a ground control center) is adopted for dynamic adjustments, specifically: First, all drones form a self-organizing network through a dual-link communication system of "wireless communication + laser communication" to share mission status and environmental information; An auction-based task allocation mechanism is adopted: When the risk value at a certain monitoring point suddenly rises above the preset risk threshold Risk th2 When the value is typically 0.7, a task request is broadcast to the entire network; nearby drones calculate their bids based on remaining battery power, distance, and current task load. ; In the formula: Indicates drone u i Normalized remaining charge; Load(u i ) Indicates drone u i The task load factor (i.e., the busyness of the tasks currently assigned to the drone, reflecting its ability to take on new tasks). The drone that bids the highest wins the mission, while the other drones automatically adjust their routes. Then, multi-agent deep reinforcement learning (MADRL) is used for dynamic path planning: State space: ,in, pos u This indicates the location of the drone. E u Indicates the remaining battery power. Risk a This indicates the risk value at each monitoring point. Obstacle This represents visual obstacle avoidance data. pos other Indicates the location of other drones; Action space: That is, flight speed and the monitoring point for the next visit next a ; Reward function: ,in, I(a i-end ) This indicates the task completion indicator function (i.e., the drone has arrived). a i (1 indicates data retrieval is complete, 0 indicates otherwise). t d-total This represents the total time taken for a single step (calculated from the drone's flight speed and distance). CollsionThis represents the collision indication function (1 if and only if the distance detected by the lidar is less than 0.5m, otherwise 0). These represent the corresponding weighting coefficients (generally) ); The model uses the PPO algorithm, is pre-trained in a digital twin environment, and deployed on each drone edge node; When the communication range is insufficient (such as when there is a blind spot or interruption in the communication link of the drone swarm), the tethered drone is automatically dispatched to take off at an altitude of 150m to serve as a communication relay station.

[0035] Step S4, UAV Edge Intelligent Computing and On-site Evidence Collection: First, a federated learning model is used to calculate the probability of flooding, specifically: First, the drone flies above the monitoring device, establishes a connection via short-range communication (such as WiFi, Bluetooth, LoRa, acoustic backscatter, etc.), and retrieves historical data. t - t history ,t Observation sequence within ] Y , t history Indicates the length of the history window (usually 30 minutes); A lightweight isolated forest algorithm is deployed at the edge to perform anomaly detection and calculate anomaly score for each data point: ; In the formula: E(h(y k )) Representing data points y k Average path length in an isolated tree c qw (E) This represents the expected value of the average path length. Data points with outlier scores greater than 0.7 were removed to obtain the preprocessed observation sequence. Y’ ; Then, the local federated learning model is called. Model local (This model is deployed on the edge node of the drone and is synchronously updated by a global federated learning model in the cloud), and statistics are calculated: Mean: ;in, This indicates the number of valid samples after preprocessing (the number of valid observation data points remaining in the historical window after removing outliers). Standard deviation: ; Exceeding threshold ratio: For surface water depth monitoring devices, river water level monitoring devices, and rainfall monitoring devices: For drainage pipe flow velocity monitoring devices: ;in, y th Indicates the threshold for flood assessment ，l() This indicates an indicator function (1 if the condition is true, 0 if the condition is false). Next, calculate the overall risk score: ; In the formula: This represents the weighting coefficient (generally between [0,1]). This represents a smoothing term to prevent the denominator from being zero (typically 1e-6). Finally, the overall risk score will be considered. S zhfx Mapped to the probability of flooding: ; In the formula: k flood This represents the slope parameter (generally greater than 0, 5 in this scheme); After each day's task is completed, each drone uploads its local model gradient updates to the cloud, where the cloud aggregates and updates the global model. ; In the formula: K UAV This indicates the number of drones participating in this round of federal aggregation.

[0036] Then, spatial Bayesian networks are used to calculate the probability of regional flooding, specifically: Construct a spatial Bayesian network, where nodes represent the flood probability of each monitoring point and edge weights represent spatial correlation. g ij (Determined from the spatial correlation matrix, based on topography, drainage network, road network, river system, and historical water accumulation data); Calculate conditional probability: ; In the formula: Nir i Represents a node i The set of neighboring nodes; Indicates the symbol for consecutive products; g ij This represents the spatial correlation between monitoring points i and j (calculated based on topography, drainage network, road network, river system, historical water accumulation data, etc.). The probability of obtaining a regional level through fusion: ; In the formula: This represents the probability weight (typically 0.6).

[0037] Finally, on-site evidence collection is conducted (for verification and corroboration), specifically: when P flood (t) When the value is greater than 0.5 (preset threshold), the drone automatically adjusts its flight altitude to 20m and takes three visible light photos (4K resolution), one near-infrared photo (wavelength 750-900nm) from different angles, and acquires one line structured light scan data (1mm resolution) to complete on-site evidence collection. The final output package... For use in cloud-based manual verification or subsequent model training; ID i Indicates the monitoring device ID. type i Indicates the type of monitoring device. t Indicates the task timestamp. P flood (t) Indicates the probability of flooding at a single point. P region (t) Indicates the probability of regional flooding. Photo data This refers to the data from photos taken at the scene as evidence.

[0038] Step S5, Cloud-based early warning response: First, verify the data, then determine the level of early warning; the verification process is as follows: for result packages with on-site photos, manually verify the data in conjunction with the photos; if the photos clearly show no flooding, then determine that the monitoring device is abnormal and arrange for corresponding maintenance, while ignoring this result package in this process. The tiered early warning determination is specifically based on a set of preset thresholds in the cloud. 0<P th1 <P th2 <P th3 <P th4 <1 (generally P th1 0.3 P th2 0.5 P th3 0.7 P th4 The value is 0.9), and warning levels are divided as follows: Level 1 Warning (Blue Warning): P flood ∈[ P th1 , P th2 We suggest you continue to monitor this situation. Level II Warning (Yellow Warning): Pflood ∈[ P th2 , P th3 It is recommended to conduct an inspection. Level 3 Warning (Orange Warning): P flood ∈[ P th3 , P th4 It is recommended to prepare for emergency drainage. Level IV Warning (Red Alert): P flood ∈[ P th4 , 1 It is recommended to activate the emergency response immediately. Similarly, the probability of regional flooding P region The same threshold system is also used to classify the warning levels.

[0039] Example 2: As another preferred embodiment of the present invention, based on the above-described embodiment 1, step S3, specifically the PPO algorithm core network structure and model training process, is as follows: A lightweight Actor-Critic network structure is adopted, and the Actor network (policy network) outputs a state vector. S zt Output continuous motion The mean and variance of discrete actions next a Probability distribution; Critic network input state vector S zt Output the value of the current state. Value(S) The specific training process is as follows: Step a: Randomly generate a precipitation intensity field, initialize the positions and battery levels of all UAVs, and randomly generate risk values ​​for monitoring points; Step b: Each UAV interacts with the environment and collects trajectory data based on the current Actor network output actions. ; Step c: Calculate the generalized advantage estimate (GAE) for each time step: ; In the formula: This represents the discount factor (usually 0.99). This represents the GAE parameter (typically 0.95). Step d, PPO trimming loss calculation: Actor loss: ; Critic loss: ; Total loss: , ; In the formula: Indicates the expected time step; Indicating a new strategy In state s t Take action a t The probability of the old strategy The ratio, i.e. ; clip() This represents the clipping function. Indicates the cutting factor; This represents the state value estimation (i.e., the Critic network's assessment of the state). s t (The cumulative reward forecast value). These represent the value loss weight and the entropy regularization term weight, respectively (generally). ); This represents the policy entropy regularization term (which measures the randomness of a policy; the higher the entropy, the more random the policy). Step e: Update network parameters using the Adam optimizer, with a learning rate of 3e-4, a size of 256, and 10 update attempts. Step f: Repeat steps a through e until the model converges, completing the model training.

[0040] Example 3: As another preferred embodiment of the present invention, based on the above-described embodiment 1, in step S4, spatial correlation... g ij The specific method for obtaining it is as follows: First, calculate the terrain correlation matrix separately. g terr (i, j) Correlation matrix of drainage pipe network g pipe (i, j) Road network correlation matrix g road (i, j) River system correlation matrix g river (i, j) : ; In the formula: h i , h j They represent monitoring points respectively. i , j Ground elevation; Dij express i and j The Euclidean distance between them; Indicates the maximum elevation difference within the city limits; ; In the formula: L long-ij express i arrive j The total length of the pipeline, L long-total This indicates the total length of the drainage main network; ; In the formula: L road-total This indicates the total length of the waterway. ; In the formula: L river-total Indicates the total length of the river channel; By integrating the topographic correlation matrix, drainage network correlation matrix, road network correlation matrix, and river system correlation matrix, an initial correlation matrix is ​​obtained. g o (i, j) : ; In the formula: These represent the corresponding weighting coefficients (generally) ); Then, statistics were compiled on all historical flood events, including the number of monitoring points. i Monitoring points during water accumulation j The probability of water accumulation: ; In the formula: N h-ij In history i When water accumulation occurs, j Also, the number of times water accumulation occurred; N h-total In history i Total number of times water accumulation occurred; Weighted calibration correlation matrix obtained: ; In the formula: This represents the corresponding weighting coefficient (generally) ); Finally, perform dynamic updates on relevance: Pipe blockage attenuation: If the spatiotemporal graph convolutional network model detects that the drainage capacity of pipe e has decreased to its original level... ke (t) , k e (t) For all upstream and downstream node pairs passing through this pipeline, the correlation updates are as follows: ∈[0,1] ; Enhanced river overflow: If the water level at a certain river monitoring point exceeds the warning level (i.e., overflow occurs), the correlation between that point and its surrounding low-lying areas is multiplied by 1.5 (with an upper limit of 1, because the overflowing river water will spread to a wider area through the surface). Temporary construction blockade: If a water barrier is set up in a certain area due to construction, then the correlation between this area and its surroundings will be 0 until the construction is completed; After each dynamic update, re-apply to each j All g(i, j, t) Perform normalization to ensure .

Claims

1. A method for predicting urban flood disaster probabilistics based on UAV mobile edge computing, characterized in that: include: Step S1, Monitoring device deployment: Monitoring devices are deployed at main drainage pipes, low-lying areas along roads, main river channels, and rainfall monitoring locations. The optimal deployment points for sensor clusters are generated through clustered redundant deployment optimized by ant colony algorithm. Step S2, Target Monitoring Device Set Generation: First, retrieve the raw meteorological radar data and obtain the real-time ground precipitation intensity field; then, use graph neural network dynamic runoff modeling to obtain the runoff volume and water accumulation risk value of each node; finally, output the target monitoring device set. Step S3, UAV dynamic path planning: First, generate the initial route of the UAV based on the improved Clark-Rattle algorithm, and then use the centralized initial route as the benchmark and adopt a decentralized bee colony self-organizing architecture for dynamic adjustment; Step S4, UAV edge intelligent computing and on-site evidence collection: First, the probability of flooding is calculated using a federated learning model, then the probability of regional flooding is calculated using a spatial Bayesian network, and finally on-site evidence collection is carried out. Step S5, Cloud-based early warning response: First, verify the information, then determine the level of early warning.

2. The urban flood disaster probability early warning method based on UAV mobile edge computing according to claim 1, characterized in that: In step S1, a Doppler flow meter is used to monitor the flow velocity of the main drainage pipe, an immersion level gauge or a radar level gauge is used to monitor the surface water depth at low-lying points along the road, a radar level gauge or an ultrasonic level gauge is used to monitor the water level in the main river, and a tipping bucket rain gauge is used to monitor the rainfall at the rainfall monitoring point.

3. The urban flood disaster probability early warning method based on UAV mobile edge computing according to claim 2, characterized in that: In step S1, for the monitoring locations of main drainage pipelines, low-lying road points, and main river channels, the optimal deployment points of sensor clusters are generated by optimizing the clustered redundant deployment using the ant colony algorithm. Specifically, the urban historical flood risk map, population density map, drainage system network map, and set of deployable locations are used as inputs. First, initialize the ant colony algorithm parameters, including the number of ants. N ant pheromone volatility coefficient Pheromones importance Heuristic function importance With maximum number of iterations N ddc ; The heuristic function is: ； In the formula: R qf (j) Indicates position j Historical flood risk value; P qf (j) Indicates position j Population density; D qf (j) Indicates position j The reciprocal of the drainage capacity; C qf (j) Indicates deployment at location j The cost; Each ant starts from a random location and chooses its next deployment location based on the transition probability: ； In the formula: Indicates time t At time, position i Arrive at the location j Pheromone concentration along the path; Indicates from position i Transfer to location j Inspirational information Indicates from position i Transfer to location s The inspirational information; allowed k Indicates the first k Only ants at all times t A set of next deployment locations that can be selected; Iteratively update pheromones: ； In the formula: Ants k In the path ij The pheromones released from the upper body; The iteration terminates when the set of all deployable locations is empty, the preset maximum number of sensors is reached, or the maximum number of iterations is reached. The optimal deployment scheme is selected, and sensors with a distance of less than 200m are grouped into a cluster, and the optimal sensor deployment set is output. Each sensor cluster is .

4. The urban flood disaster probability early warning method based on UAV mobile edge computing according to claim 3, characterized in that: In step S1, the arrangement of rainfall monitoring sites is specifically as follows: First, acquire existing rainfall monitoring devices in the city and preset two sampling intervals. L cy With coverage radius R fg Two parameters; then, within the city limits, according to the sampling interval. L cy A batch of sampling points is generated. For each sampling point, its distance to the "nearest rainfall monitoring device" is calculated. Then, the distances are arranged from largest to smallest to obtain the sampling point with the largest distance. l max Its maximum distance is d max ;like d max ≤ R fg This indicates that the density of urban rain gauges meets the requirements and the deployment is complete. d max > R fg This indicates that there are areas with insufficient monitoring, and more rainfall monitoring devices need to be installed. At this time, priority should be given to sampling points. l max Rainfall monitoring devices are installed at the sampling points. l max If rainfall monitoring devices cannot be installed at the location, then the distance from the sampling point should be... l max The closest and the distance is less than or equal to R fg Install rainfall monitoring devices at the location; During the deployment of rainfall monitoring sites, the priority of ant colony algorithm site selection is increased, prioritizing deployment in areas with a history of high flood risk.

5. The urban flood disaster probability early warning method based on UAV mobile edge computing according to claim 4, characterized in that: In step S2, the inversion of raw weather radar data and the acquisition of real-time ground precipitation intensity field specifically involve: based on the raw reflectivity data Z from the local S-band weather radar, the ground precipitation intensity is calculated using variational Doppler radar inversion. ； In the formula: Indicates the proportionality coefficient of the relationship. Indicates the relationship index; When the drone retrieves ground rainfall data y rain,i At that time, the relationship proportional coefficient and relationship index are updated in real time using the least squares method: ； In the formula: y rain,i Indicates the first i The measured rainfall intensity at each rain gauge station Z i Indicates the first i Linear values ​​of radar reflectivity at the corresponding locations of each rain gauge station N ture This indicates the number of validly calibrated rain gauges; The final output shows precipitation forecast fields for the next 15 minutes, 30 minutes, and 60 minutes. .

6. The urban flood disaster probability early warning method based on UAV mobile edge computing according to claim 5, characterized in that: In step S2, the specific steps for obtaining the flow rate and water accumulation risk value of each node using graph neural network dynamic flow modeling are as follows: First, the coordinates of all manholes, drainage outlets, and river confluences are extracted from the geographic information system data to serve as base nodes. Then, the locations of all monitoring devices are matched to the nearest base node to form a node set. V Based on the flow direction of the pipeline and the river, a set of directed edges E is formed: if there exists a node... i To the node j If the water flow path is such that an edge is added... e ij ∈ E ; Calculate initial weights based on pipeline / river channel parameters. The initial adjacency matrix is: ； Then, the gridded precipitation intensity is matched to each node. Obtain each node v At any moment t Feature vectors: ; In the formula: v pipe (t) This represents the pipeline flow velocity monitoring value. h water (t) This represents the monitoring value of surface water depth. h river (t) This indicates the river water level monitoring value; Z-score normalization is performed on each feature dimension to eliminate dimensions; and past data is obtained. T in Feature input tensor for the time period: ; Subsequently, a spatiotemporal graph convolutional network model was constructed, consisting of three spatiotemporal convolutional blocks and two fully connected layers. Each spatiotemporal convolutional block is composed of "graph convolutional layer + temporal convolutional layer + batch normalization + Dropout". Then, for each edge e ij Calculate its real-time drainage capacity coefficient: ； In the formula: v design,i Represents a node i Corresponding to the design flow velocity of the pipeline; And calculate the historical correction coefficient based on the historical congestion records of this edge: ； In the formula: N block,ij This indicates the number of times the edge became congested in the past year. N total,ij Indicates the total number of monitoring sessions; Update dynamic edge weights: ;but: ; Next, the sink flow output by the inverse normalization model: ； In the formula: These represent the historical mean and standard deviation of the inflow rate, respectively. Calculate the risk value of water accumulation: ； In the formula: C v-max Represents a node v Maximum drainage capacity; when Risk v (t+t x ) A value less than 0.5 indicates sufficient drainage capacity; a value less than or equal to 0.5 indicates sufficient drainage capacity. Risk v (t+t x ) A value less than 1 indicates that drainage capacity is strained. Risk v (t+t x ) A value ≥1 indicates that water has accumulated. According to node priority Priority v Sort: ； In the formula: Pop v Represents a node v Normalized population density of the covered area; Finally, model training and updates are performed, specifically including: inputting the feature sequence of the past 10 minutes as input samples, labeling them as the actual sink flow of the next 15 minutes, and dividing them into training set, validation set, and test set in a ratio of 7:2:1, with mean squared error as the loss function.

7. The urban flood disaster probability early warning method based on UAV mobile edge computing according to claim 6, characterized in that: In step S2, the output target monitoring device set specifically refers to: the precipitation forecast field Input a digital twin to preview the development of urban flooding in the next 15 minutes, 30 minutes, and 60 minutes, and generate a risk evolution heat map; Nodes with a water accumulation risk value not less than the water accumulation risk threshold are selected. Add the corresponding monitoring devices to the target set; select monitoring devices less than 500m away from high-risk areas and add them to the target set; for each river, calculate the average rainfall intensity of its catchment area, and if it is greater than 20mm / h, add all monitoring devices on that river to the target set. Assign priority to each target monitoring device The target set is then sorted from highest to lowest priority to obtain the target monitoring device set: 。 8. The urban flood disaster probability early warning method based on UAV mobile edge computing according to claim 7, characterized in that: In step S3, generating the initial route of the UAV based on the improved Clark-Rattle algorithm specifically involves: First, the target monitoring device set is as follows , among which, the i The coordinates of the monitoring devices are ( x i ,y i ); drone swarms are The drone take-off and landing point is B The coordinates are ( x b ,y b The maximum flight time of the drone is t UAV-max The drone hovering time above each monitoring device is t UAV-a The drone's cruising speed is v UAV The wireless communication radius is R comm The laser communication radius is R laser ; For any two points The distance between them is: ; drones from o a Fly to o b The time is: ; The total time for any route is: ,in, t flight The flight time is represented by the flight time for each segment of the journey. The result is obtained by addition; N q Indicates the number of monitoring devices accessed along the flight path; Then, the set Each monitoring device is treated as an independent single-point route: for any monitoring device Generate the initial route r i The access sequence is r i =[ a i The actual flight distance of this route is: B → a i → B The initial set of routes is then: R init ={ r 1 ,r 2 ,…,r Nj }, Nj Indicates the number of monitoring devices; Then, the two routes r i =[ a i ]and r j =[ a j Merged into one route r ij =[ a i , a j The saved flight distance is the amount saved. S(i,j) ; Traverse all disordered monitoring device pairs ( a i , a j ), calculate the savings for each pair. S(i,j) Then, sort all point pairs in descending order of savings to obtain a list of point pairs to be merged: ； In the formula: M represents the total number of point pairs, and S(i1,j1)≥S(i2,j2)≥…≥S(iM,jM) ; Finally, the routes are merged.

9. A method for urban flood disaster probability early warning based on UAV mobile edge computing according to claim 8, characterized in that: In step S3, the centralized initial route is used as a baseline, and dynamic adjustments are made using a decentralized bee colony self-organizing architecture. First, all drones form a self-organizing network through a dual-link communication system of "wireless communication + laser communication" to share mission status and environmental information; An auction-based task allocation mechanism is adopted: when the risk value of a monitoring point suddenly rises above a preset risk threshold... Risk th2 At that time, the task request is broadcast to the entire network; nearby drones calculate the bid based on remaining battery power, distance, and current task load: ； In the formula: Indicates drone u i Normalized remaining charge; Load(u i ) Indicates drone u i Task load factor; The drone that bids the highest wins the mission, while the other drones automatically adjust their routes. Then, multi-agent deep reinforcement learning is used for dynamic path planning: State space: ,in, pos u This indicates the location of the drone. E u Indicates the remaining battery power. Risk a This indicates the risk value at each monitoring point. Obstacle This represents visual obstacle avoidance data. pos other Indicates the location of other drones; Action space: That is, flight speed and the monitoring point for the next visit next a ; Reward function: ,in, I (a i-end ) This indicates a function that indicates task completion. t d-total This indicates the total time taken for a single step. Collation This represents the collision indication function. These represent the corresponding weight coefficients; The model uses the PPO algorithm, is pre-trained in a digital twin environment, and deployed on each drone edge node. When the communication range is insufficient, the tethered drone is automatically scheduled to take off to a height of 150m to serve as a communication relay station.

10. A method for urban flood disaster probability early warning based on UAV mobile edge computing according to claim 9, characterized in that: In step S4, the federated learning model calculates the probability of flooding as follows: First, the drone flies above the monitoring device, establishes a connection via short-range communication, and retrieves historical data. t- t history ,t Observation sequence within ] Y , t history Indicates the length of the history window; A lightweight isolated forest algorithm is deployed at the edge to perform anomaly detection and calculate anomaly score for each data point: ； In the formula: E(h(y k )) Representing data points y k Average path length in an isolated tree c qw (E) This represents the expected value of the average path length. Data points with outlier scores greater than 0.7 were removed to obtain the preprocessed observation sequence. Y’ ; Then, the local federated learning model is called. Model local And calculate statistics, including the mean. Standard deviation and the proportion of exceeding the threshold ; Next, calculate the overall risk score: ； In the formula: Indicates the weighting coefficient. This indicates a smoothing term to prevent the denominator from being zero; Finally, the overall risk score will be considered. S zhfx Mapped to the probability of flooding: ； In the formula: k flood Indicates the slope parameter; After each day's task is completed, each drone uploads its local model gradient updates to the cloud, where the cloud aggregates and updates the global model. ； In the formula: K UAV This indicates the number of drones participating in this round of federal aggregation.

Images