Multi-source sensing fusion-based road drainage system blockage monitoring method and system

Abstract

This invention belongs to the technical field of road drainage system blockage monitoring methods, and particularly relates to a multi-source sensor fusion method and system for road drainage system blockage monitoring. It constructs a dynamic topology sensing network for the pipeline network, deploys multimodal integrated sensor monitoring units at each node of the network, constructs and updates a digital twin model of the pipeline network topology in real time through lidar scanning, and synchronously collects data on water flow status and structural changes in the network. Simultaneously, it employs adaptive sensor correction and diversion flow monitoring schemes for special nodes in the network to achieve accurate data collection for these special nodes. Based on the topology digital twin model, a topology-hydrological coupled blockage propagation model is constructed, integrating real-time network operation data and topology parameters. Numerical simulation methods are used to clarify the migration patterns of water flow and blockages, and quantify the diffusion characteristics of blockages under different topologies.

Description

Technical Field

[0001] This invention belongs to the technical field of road drainage system blockage monitoring methods, and particularly relates to a road drainage system blockage monitoring method and system using multi-source sensor fusion. Background Technology

[0002] In the current field of road drainage system blockage monitoring, existing monitoring methods often suffer from limitations such as single sensor types and insufficient data fusion capabilities. They frequently rely on single water level or flow sensors to collect data, making it difficult to simultaneously capture changes in the pipe network's flow status and structure. Furthermore, there is a lack of suitable sensor calibration and flow monitoring solutions for special nodes such as drop manholes and intercepting wells, resulting in low monitoring data accuracy. Simultaneously, pipe network topology models are mostly statically constructed and cannot be updated in real-time based on actual conditions such as pipe network modifications and structural deformations. Discrepancies exist between digital twin models and the actual pipe network state, making it difficult to provide an accurate topological basis for blockage monitoring and prediction, and easily leading to distorted subsequent analysis results.

[0003] Existing blockage prediction and source tracing technologies still have many shortcomings. On the one hand, prediction models often rely solely on data-driven or physical principles. Purely data-driven models are prone to violating hydraulic dynamics laws, while purely physical models struggle to adapt to complex pipe network topologies and variable operating conditions. Furthermore, they lack sufficient feature mining for extreme rainfall scenarios, resulting in low prediction time-series accuracy. On the other hand, blockage source tracing algorithms do not fully integrate the spatiotemporal correlation of monitoring data with pipe network topology weights, exhibiting poor adaptability to special topologies such as ring-shaped and multi-branched structures, and their positioning accuracy is insufficient to meet engineering requirements. In addition, the lack of a hierarchical processing architecture with edge-cloud collaboration means that models cannot achieve online self-calibration through operational feedback and topology changes. Moreover, cross-departmental data is not effectively linked for cross-validation, leading to delayed blockage warnings, low source tracing efficiency, and difficulty in quickly responding to the risk of pipe network blockage, especially under extreme weather conditions. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for monitoring blockages in a road drainage system using multi-source sensor fusion, the method comprising:

[0005] A dynamic topology sensing network for the pipeline network is constructed. Multimodal integrated sensing and monitoring units are deployed at each node of the pipeline network. A digital twin model of the pipeline network topology is constructed and updated in real time through lidar scanning. Data on the water flow status and structural changes in the pipeline network are collected synchronously. At the same time, appropriate sensing correction and diversion flow monitoring schemes are adopted for special nodes of the pipeline network to achieve accurate collection of monitoring data for special nodes.

[0006] A topology-hydrology coupled blockage propagation model is constructed based on a topology digital twin model. By integrating real-time pipeline network operation data and topology parameters, the migration patterns of water flow and blockages are clarified through numerical simulation methods, and the diffusion characteristics of blockages under different topologies are quantified.

[0007] Establish a multi-dimensional blockage classification index system, integrate key hydraulic characteristic indicators, use the objective weighting method to determine the index weights and classify blockage levels, and dynamically correct the classification thresholds in combination with pipeline network structure parameters.

[0008] A hybrid prediction model combining data-driven and physical constraints is constructed, integrating historical data time-series characteristics with hydraulic physics principles, and combining multi-source environmental and pipeline load data to achieve prediction of blockage escalation under normal scenarios and time-series prediction of blockage development under extreme weather conditions.

[0009] A blockage location reverse tracing algorithm is adopted, which relies on the spatiotemporal correlation of multi-node monitoring data and combines the pipeline topology weight to locate the source of blockage. The algorithm is optimized for special topology structures to improve the positioning accuracy.

[0010] Build an edge-cloud collaborative computing architecture to achieve hierarchical processing of congestion level early warning and large-scale propagation prediction, and simultaneously establish an online model self-calibration mechanism to dynamically update model parameters through operation and maintenance feedback data and topology changes;

[0011] Establish a multi-source data collaborative verification system, connect with relevant data from different departments to conduct cross-verification, and introduce special data to optimize the congestion prediction time series for extreme rainfall scenarios.

[0012] Furthermore, embodiments of the present invention also provide a multi-source sensor fusion road drainage system congestion monitoring system, characterized in that it includes:

[0013] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to perform the above-described multi-source sensor fusion road drainage system congestion monitoring method by executing the machine-executable instructions.

[0014] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the above-described multi-source sensor fusion method for monitoring blockages in a road drainage system.

[0015] Based on the above, through multi-source sensor fusion and dynamic topology digital twin technology for pipeline networks, accurate and synchronous acquisition of pipeline network flow status and structural change data was achieved. Monitoring deviations at special nodes are controllable, and the digital twin model can be incrementally updated in real time according to actual changes in the pipeline network, laying a precise topological and data foundation for blockage monitoring. The topology-hydrology coupling model accurately quantifies the blockage diffusion characteristics under different topologies. The multi-dimensional blockage grading index system, through objective weighting and dynamic correction of structural parameters, achieves accurate determination of blockage levels. The hybrid prediction model, which integrates data-driven and physical constraint fusion, takes into account both the escalation trend prediction of conventional scenarios and the fine-grained temporal prediction of extreme weather. The dual-mode collaboration significantly improves prediction accuracy. The blockage location reverse tracing algorithm, combining spatiotemporal correlation and topological weights, optimizes the adaptability to special topologies, achieving accurate location of the blockage source. The pipe segment-level location error meets engineering requirements.

[0016] The edge-cloud collaborative computing architecture built in this invention enables hierarchical processing of congestion tiered early warning and large-scale propagation prediction. The lightweight edge model ensures rapid response to short-term warnings at the node level, while the high-precision cloud model completes long-term regional predictions and overall risk assessment. The hierarchical data transmission strategy significantly reduces bandwidth consumption and improves data interaction efficiency. A dual-trigger online model self-calibration mechanism dynamically updates model parameters through operational feedback and topology changes, continuously optimizing model performance. A multi-source data collaborative verification system improves data reliability through cross-departmental data verification. The introduction of extreme rainfall-specific data further optimizes the congestion prediction timeline. Overall, this invention achieves intelligent monitoring, prediction, source tracing, and early warning of road drainage system congestion, significantly improving the efficiency and intelligence level of pipeline network operation and maintenance, effectively reducing the risks of water accumulation caused by pipeline congestion under both normal and extreme weather conditions, and possesses significant engineering application value. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the execution flow of the multi-source sensor fusion method for monitoring blockages in road drainage systems provided in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of exemplary hardware and software components of the road drainage system congestion monitoring system provided in an embodiment of the present invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a method for monitoring blockages in a road drainage system using multi-source sensor fusion, as provided in one embodiment of the present invention. The following is a detailed description of this method.

[0020] Step S110: Construct a dynamic topology sensing network for the pipeline network, deploy multimodal integrated sensing and monitoring units at each node of the pipeline network, construct and update the digital twin model of the pipeline network topology in real time through lidar scanning, synchronously collect data on the water flow status and structural changes of the pipeline network, and adopt appropriate sensing correction and diversion flow monitoring schemes for special nodes of the pipeline network to achieve accurate collection of monitoring data for special nodes.

[0021] An ultrasonic water level sensor (range 0-5m, accuracy ±1mm), a radar flow sensor (range 0.1-10m / s, accuracy ±2%), and a vibration sensor (range 0-50g, frequency response 10-1000Hz) are selected and integrated into an IP68-rated multimodal monitoring unit. This unit is equipped with a low-power MCU (standby power consumption ≤10mA) and a LoRa / 5G / NB-IoT dual-mode communication module (LoRa band 433MHz, NB-IoT supports 800 / 900MHz). The SLAM-200 lidar device (point cloud density 1000 points / cm², ranging accuracy ±3mm) is used and is mounted or fixedly installed on the pipe cross-section via a pipeline robot (CR-300 type, traveling speed 0.5m / s). For special nodes such as drop wells and intercepting wells, a dual-sensor comparison and correction method is adopted (water level sensor spacing 0.5m). The diversion flow monitoring adopts a closed-loop calculation scheme of "main pipeline flow - branch flow". The edge computing unit corrects the data deviation in real time to ensure that the relative error of the monitoring data of special nodes is ≤±3%. It realizes the synchronous and accurate acquisition of water flow status (water level, flow rate, flow velocity) and structural change (pipeline deformation, node loosening) data, providing a real-time data source for the digital twin model.

[0022] Step S111: Conduct a general survey of basic pipeline information through drawing retrieval and on-site survey to clarify the pipeline topology type, critical path nodes and the distribution of special nodes such as inspection wells, drop wells and interception wells, collect physical parameters such as pipe diameter, pipe material, slope and design flow rate, and establish a basic pipeline information database.

[0023] By retrieving as-built drawings and operation and maintenance records of the 1:500 drainage pipe network in the target area from the urban construction archives management department, and conducting on-site surveys using a total station (angle measurement accuracy ±2″) and an inclinometer (accuracy ±0.1°), the survey scope covered the entire pipe network and ancillary facilities. The pipe network topology type (branched, ring, or mixed) was identified, and the specific coordinates (GPS positioning accuracy ±1m) of key path nodes (main pipe intersection nodes, flow control nodes) and special nodes such as inspection wells, drop wells, and intercepting wells were marked. Core physical parameters were collected: pipe diameter was measured using a laser rangefinder (error ≤ ±5mm), pipe material was confirmed through visual identification combined with material testing reports, and slope was measured using a level (accuracy ≤ ±0.1%). Design flow was obtained from the pipe network design documents and corrected based on actual operating conditions. A structured pipe network basic information database was established, with fields including node ID, topology affiliation, coordinate location, physical parameters, and special node identifiers. The database supports data addition, deletion, modification, querying, and batch export, with a response time ≤ 500ms, providing basic data support for subsequent model construction and algorithm calculations.

[0024] Step S112: Select an ultrasonic water level sensor, a radar flow sensor, and a vibration sensor, and integrate them into an integrated multimodal monitoring module, along with a low-power data acquisition unit and a LoRa / 5G / NB-IoT wireless communication module;

[0025] The ultrasonic water level sensor is model UWM-01 (range 0-5m, resolution 0.1mm, water pressure resistance ≥0.6MPa), the radar flow sensor is model RFM-02 (non-contact, range 0.1-10m / s, measurement accuracy ±2%FS), and the vibration sensor is model VS-03 (piezoelectric, range 0-50g, frequency response 10-1000Hz). All three are integrated into a single 15cm×10cm×8cm metal housing. A low-power data acquisition unit is provided, using an STM32L476 microcontroller (operating voltage 3.3V, acquisition frequency 1Hz, standby power consumption ≤10mA). The communication module is a LoRa / 5G / NB-IoT dual-mode module (LoRa communication distance ≤3km, 5G / NB-IoT supports public network access, data transmission latency ≤100ms). The module features a dual power supply solution of a built-in lithium battery (capacity 10000mAh) and a solar charging panel (power 5W), ensuring continuous operation for ≥6 months. Each sensor interface adopts a standardized design (RS485 protocol) and supports hot-swapping replacement. The acquisition unit has a built-in data cache module (storage capacity ≥16GB) that can cache nearly 3 months of monitoring data to avoid data loss due to network interruption.

[0026] Step S113: Based on the complexity of the pipeline network topology and the characteristics of the flow distribution, fix the multimodal monitoring module on the well wall or inner wall of the pipeline at the path nodes and special nodes, adopt a waterproof sealing installation process to avoid the area directly impacted by the water flow, and simultaneously calibrate the sensor measurement accuracy to meet the engineering monitoring standards.

[0027] Based on the complexity of the pipeline topology (one monitoring module every 200m in a branched network and one every 150m in a ring network) and the flow distribution characteristics (densified installation in high-flow sections with a spacing ≤100m), the installation locations of the monitoring modules were determined: path nodes were installed 1.5m from the bottom of the manhole on the manhole wall, and special nodes were installed at the horizontal centerline of the inner wall of the pipeline, all avoiding areas directly impacted by water flow (≥0.3m from the centerline of the pipeline). The installation process used expansion bolts for fixing + IP68-grade sealing rings + epoxy resin potting to ensure waterproofing, corrosion resistance, and vibration resistance. After installation, the sensors were calibrated for accuracy: the water level sensor was calibrated at 5 points (0.5m, 1m, 2m, 3m, 4m) using a standard water level tank (accuracy ±0.1mm); the flow sensor was calibrated at 3 sets of flow velocities (0.5m / s, 2m / s, 5m / s) using a standard flow device (accuracy ±1%); and the vibration sensor was calibrated for sensitivity using a standard vibration table (frequency 50Hz, amplitude 0.1mm). After calibration, the sensor measurement error must meet the requirements of the "Technical Specification for Monitoring of Urban Drainage Pipeline Network" (CJJ / T210-2014): water level measurement error ≤ ±2mm, flow rate measurement error ≤ ±3%, and vibration measurement error ≤ ±5%.

[0028] Step S114: Deploy lidar equipment suitable for the internal environment of the pipeline network. The equipment can be mounted on a pipeline robot or fixedly installed on the cross-section. Set the point cloud density, scanning angle, and ranging range parameters to ensure coverage of the inner wall of the pipeline, the internal structure of the nodes, and the connection parts.

[0029] The lidar equipment selected is the SLAM-300 3D lidar (dustproof and waterproof rating IP67, temperature range -10℃~60℃), with a ranging range of 0.5-50m, ranging accuracy ±3mm, point cloud density of 1000 points / cm², scanning angle of 360° (horizontal) × 90° (vertical), and scanning frequency of 10Hz. There are two mounting methods: for conventional pipe sections, a CR-300 pipeline robot is used (traveling speed 0.5m / s, climbing ability ≤30°); for special nodes (such as large-diameter inspection wells and tee joints), a fixed mounting bracket (adjustable height 0.5-2m) is used for fixed installation. The scanning parameters are set as follows: point cloud density 1000 points / cm², horizontal scanning step distance 0.1°, vertical scanning step distance 0.2°, and ranging range 0.5-50m. The device has a built-in inertial navigation module (positioning accuracy ±0.5m / 100m) and a GPS module (positioning accuracy ±1m), which synchronously record the scanning position coordinates and timestamps (error ≤10ms). To adapt to the dim and humid environment inside the pipeline network, the device is equipped with an infrared supplementary lighting module (effective supplementary lighting distance ≤10m) to ensure the integrity of point cloud data acquisition and to cover the inner wall of the pipeline, the internal structure of the nodes, and the connection parts without blind spots.

[0030] Step S115: Advance the full-domain scanning of the lidar in segments according to the pipeline topology path, simultaneously record the GPS / inertial navigation coordinates of the scanning position, collect three-dimensional point cloud data of pipeline centerline, pipe diameter change, node structure morphology, and pipeline connection relationship, and perform multi-angle repeated scanning of special nodes.

[0031] The pipeline network was divided into several scanning segments, each 50m long, and the LiDAR system performed a full-area scan from the start to the end of the network. During the scan, the pipeline robot moved at a constant speed (0.5m / s) along the pipeline centerline, recording GPS / inertial navigation coordinates in real time (positioning frequency 10Hz), and simultaneously acquiring 3D point cloud data (storage format LAS1.4) of the pipeline centerline, diameter changes, node structure, and pipeline connection relationships. For special nodes such as inspection wells, drop wells, and intercepting wells, three multi-angle scans (60° intervals) were used to ensure the capture of complex internal structures (such as ladders, flow channels, and branch outlets). Before scanning, the equipment was preheated (10 minutes) and parameters were calibrated. During the scan, the point cloud data quality was monitored in real time (effective point rate ≥95%). If data loss or excessive noise occurred, the scan was immediately paused and the equipment position was adjusted for a re-scan. After scanning, timestamps, location tags, and segment IDs are added to each point cloud data segment to ensure data traceability. At the same time, the original point cloud data is transmitted to cloud storage (transmission bandwidth ≥10Mbps) via wireless communication module and backed up locally.

[0032] Step S116: Use Gaussian filtering and statistical filtering algorithms to remove noise points in point cloud data, stitch together data from different sections using point cloud registration technology, extract pipeline axis and cross-sectional contour using RANSAC algorithm, identify node types and connection relationships through cluster analysis, and construct a pipeline network topology skeleton model.

[0033] A combined algorithm of Gaussian filtering (standard deviation σ=0.5) and statistical filtering (10 neighborhood points, threshold multiplier 2) is used to remove noise points (such as isolated points formed by water droplets or dust) from the point cloud data, ensuring that the effective point rate after filtering is ≥98%. Point cloud registration uses an improved ICP algorithm (50 iterations, convergence threshold 0.001) to stitch together point cloud data from different sections, eliminating stitching errors (≤±3mm). The RANSAC algorithm (1000 iterations, interior point threshold 0.01m) is used to extract the pipeline axis (straight line fitting error ≤±2mm) and cross-sectional profile (circular fitting error ≤±5mm). Cluster analysis (K-means algorithm, K value preset to 3-5 classes) is used to identify node types (inspection wells, tees, crosses, etc.) and connection relationships (such as orthogonal, oblique). Based on the extracted axis, contour and node information, a pipeline topology skeleton model is constructed (output format is IGES). The model includes core information such as pipe segment centerline, node position, and connection method, ensuring that the consistency error between the skeleton model and the actual pipeline topology is ≤±5mm, laying the foundation for the subsequent construction of a three-dimensional geometric model.

[0034] Step S117: Based on the topological skeleton model and physical parameters, build a three-dimensional geometric model of the pipeline network, map the physical properties of pipe diameter, pipe material, and slope, integrate the position coordinates and data interface of the multimodal monitoring module, realize the real-time association between the model and sensor data, and develop a visualization platform that supports three-dimensional display of topological structure, node query and data traceability.

[0035] Based on a topological skeleton model, a 3D geometric model of the pipeline network was built using AutoCAD Civil 3D software, recreating the 3D morphology of pipes and nodes at a 1:1 scale with a model accuracy of ≤±5mm. By linking node IDs to the pipeline network's basic information database, physical attributes such as pipe diameter, pipe material, slope, and design flow rate are mapped to corresponding components in the model, achieving a binding between physical attributes and geometric shapes. The location coordinates (GPS positioning accuracy ±1m) and data interface (MQTT protocol) of the multimodal monitoring module are integrated, marking sensor installation locations in the 3D model and supporting real-time retrieval of collected water level, flow rate, and vibration data (data update frequency 1Hz) by clicking sensor icons. A visualization platform (based on Unity3D) was developed, comprising three main modules: 3D display, node query, and data traceability. The 3D display module supports model scaling, rotation, and sectioning, allowing viewing of the internal structure of the pipeline network; the node query module supports searching by node ID, type, and location, displaying detailed node information and real-time monitoring data; the data traceability module allows viewing of sensor data trend curves for the past 3 months (such as water level change trends and flow rate fluctuation curves), and supports data export (in Excel and CSV formats). The platform is deployed in the cloud and supports access from both web and mobile devices, with a response time of ≤1 second.

[0036] Step S118: Set a regular update cycle and connect to the network operation and maintenance management system. When receiving topology modification and maintenance construction information, automatically trigger the lidar directional scan, identify the topology change area by comparing the old and new point cloud data, and incrementally update the geometric structure and physical parameters of the digital twin model.

[0037] The model is scheduled for regular updates: a full-area scan update is conducted monthly, and a scan update of key areas (high-flow pipe sections, easily clogged nodes) is conducted weekly. Simultaneously, the system connects to the pipeline network operation and maintenance management system via API. When information such as topology modifications and maintenance work is received (e.g., pipe replacement, node addition, pipe diameter adjustment), a directional LiDAR scan is automatically triggered (completed within 24 hours). The directional scan focuses on the changed area and related pipe sections (extending up to 50m), collecting new and old point cloud data, which are compared using CloudCompare software (difference threshold 0.05m) to identify areas of topology change (e.g., pipe position shifts, node structure changes). Incremental updates are performed on the changed areas: corresponding parts of the 3D geometric model are modified (e.g., updating pipe length, adjusting node shape), and physical parameters in the database are updated synchronously (e.g., replacing pipe diameter, updating pipe material information), with an update time ≤ 2 hours. After the update is completed, an update report (including the changed area, update content, and parameter change records) is generated and pushed to the operation and maintenance management system to ensure that the geometric structure and physical parameters of the digital twin model are always consistent with the actual pipeline network.

[0038] Step S119: The edge computing unit receives water flow status data from the multimodal monitoring module and structural monitoring data from the lidar in real time, cross-validates the accuracy of the model topology and the validity of the sensor data. When abnormal data or suspected topological changes are detected, a secondary scan and model calibration process is triggered to ensure that the digital twin model is dynamically consistent with the actual pipeline network status.

[0039] The edge computing unit uses NVIDIA Jetson Xavier NX (21 TOPS computing power) to receive real-time water flow status data (sampling frequency 50Hz) from the multimodal monitoring module and structural monitoring data from the lidar (sampling frequency 1Hz) via the MQTT protocol. The cross-validation logic is as follows: Based on the topology and physical parameters of the digital twin model, the theoretical flow rate and water level values ​​for each node are derived and compared with the sensor's measured values. If the deviation exceeds 5%, the data is considered abnormal. Simultaneously, the latest topology scanned by the lidar is compared with the historical structure stored in the model. If the deviation exceeds 10mm, a suspected topology change is identified. When an abnormal data or suspected topology change is detected, a secondary scan (lidar directional scan, scan time ≤ 30 minutes) and model calibration process are immediately triggered: the least squares method is used to correct the model parameters (e.g., adjusting pipe diameter and slope values), and the theoretical values ​​are re-derived until the deviation between the measured and theoretical values ​​is ≤ ±3%. The edge computing unit provides real-time feedback on the dynamic consistency between the model and the actual pipeline network. The consistency criteria are: topology deviation ≤ ±5mm, and deviation between sensor data and model derivation value ≤ ±3%. This ensures that the digital twin model can accurately reflect the real-time status of the actual pipeline network, providing reliable model support for subsequent blockage prediction and source tracing.

[0040] Step S120: Construct a topology-hydrology coupled blockage propagation model based on the topology digital twin model, integrate real-time pipeline network operation data and topology parameters, clarify the migration law of water flow and blockage through numerical simulation methods, and quantify the diffusion characteristics of blockage under different topologies;

[0041] Based on the existing digital twin model of the pipeline network topology, a bidirectional coupled topology-hydrology blockage propagation model was built. A multiphysics coupling architecture was adopted to integrate pipeline topological constraints with hydrodynamic calculations and blockage transport patterns. The model incorporates real-time operational data such as water level, flow rate, and rainfall from the pipeline network, and synchronously calls topological parameters such as pipe diameter, slope, and node connection relationships. A numerical simulation method combining the finite volume method and the mass tracking method was used to dynamically simulate the water flow state, blockage initiation and transport, and sedimentation processes within the pipeline network. By simulating operating conditions under different initial blockage locations and blockage degrees, the model accurately depicts changes in water flow regime and blockage diffusion paths. Independent simulations were conducted for branched, ring-shaped, and mixed topologies of the pipeline network, quantifying core characteristics such as blockage diffusion speed, impact range, and gradation rate. The model clarifies the regulatory role of different topologies on blockage propagation, providing numerical simulation support for blockage source tracing and early warning. The simulation time step was controlled within 1-5 seconds to ensure dynamic response accuracy.

[0042] Step S121: Extract geometric and topological feature data of the entire pipeline network segments and nodes based on the topological digital twin model. The feature data includes pipe diameter, pipe length, laying slope, pipe material roughness, node connection relationship, topology type and special node structure parameters. After standardizing the feature data, construct a pipeline network topology graph structure with nodes as vertices and pipe segments as edges to complete the geometric and topological basic modeling of the coupled model.

[0043] Geometric and topological feature data of pipe segments and nodes across the entire pipeline network are extracted in batches from the topological digital twin model. This includes pipe diameter, pipe length, laying slope, pipe roughness, node connection relationships, topological type, and special node structural parameters such as drop manholes and intercepting manholes. The data extraction accuracy is consistent with the twin model, with geometric parameter errors ≤ ±5mm. Min-max normalization and dimensionless processing methods are used to standardize feature data of different dimensions, eliminating the interference of data magnitude differences on model calculations. A directed weighted topological graph is constructed with pipeline nodes as vertices and pipe segments as edges. The weight of a pipe segment is determined by a combination of pipe diameter, slope, and roughness. The weight of a node is associated with the number of pipe segments and the flow carrying capacity, forming a topological graph structure combining an adjacency matrix and an edge attribute table. Based on this, geometric and topological modeling of the coupled model is completed, ensuring a 1:1 match between the model mesh and the actual pipeline network, laying a structural foundation for subsequent hydraulic calculations and blockage simulations.

[0044] Step S122: Access the real-time water level, flow rate, flow velocity data and rainfall monitoring data collected by the multi-modal monitoring unit, align the multi-source monitoring data with the corresponding nodes and pipe segments in the topological digital twin model in time and space, remove abnormal fluctuation data and complete data filtering and normalization processing to form a real-time input dataset for the topology-hydrological coupling model.

[0045] Real-time water level, flow rate, velocity, and rainfall data collected by a multimodal monitoring unit are accessed via a wireless communication interface. The data sampling frequency is 1Hz, and the transmission delay is ≤100ms. Based on node ID and timestamp, the multi-source monitoring data is spatiotemporally aligned with the corresponding nodes and pipe segments in the topological digital twin model, achieving spatial location matching and temporal series synchronization. The 3σ criterion is used to remove abnormal fluctuations such as sudden changes in flow rate and anomalous jumps in water level. A moving average filtering algorithm is used to remove high-frequency noise, preserving the true trend of data changes. The filtered data undergoes min-max normalization, mapping the values ​​to the [0,1] interval to eliminate the influence of different physical quantity dimensions, forming a unified format for the real-time input dataset of the topological-hydrological coupling model. The dataset is updated every 5 minutes, with 24-hour historical data cached locally to ensure the continuity and effectiveness of the model input, providing a reliable data source for solving the hydraulic equations.

[0046] Step S123: Establish a hydraulic control equation system adapted to complex pipe network topology. The one-dimensional Saint-Venant equation is used as the core of water flow control. Hydraulic constraints are constructed in combination with the flow balance conditions of pipe network nodes. Water flow distribution rules and boundary conditions are set for different topological structures such as branching, ringing and mixed. Hydraulic connection relationships are defined separately for special nodes such as drop wells and intercepting wells, forming a control equation set that integrates topological constraints and hydrodynamic constraints.

[0047] A hydraulic control equation system adapted to complex pipe network topologies is established, using the one-dimensional unsteady flow Saint-Venant equation as the core control equation for water flow motion. This system includes continuity and momentum equations, accurately describing the spatiotemporal variations of water level and flow rate in the pipe network. Hydraulic constraints are constructed based on the flow balance conditions of pipe network nodes, ensuring the conservation of inflow and outflow at nodes and forming closed hydraulic calculation constraints. Flow distribution rules are set for three types of topologies: branched, ringed, and mixed. A unidirectional recursive calculation rule is used for branched pipe networks, a ring network flow balance correction is introduced for ringed pipe networks, and corresponding calculation rules are applied to mixed pipe networks in different zones. Hydraulic connections are defined separately for special nodes such as drop structures and intercepting wells, considering characteristics such as drop head loss and flow distribution between and between streams. This deeply integrates topological and hydrodynamic constraints to form a control equation set suitable for complex pipe networks, ensuring the accuracy of hydraulic calculations under different topologies and node types.

[0048] Step S124: The finite volume method is used to spatially discretize the control equations. The calculation units are divided according to the topological relationship between the pipe network segments and nodes. The numerical flux calculation method of the unit interface is determined. The discretization accuracy is adaptively adjusted according to the pipe material, pipe diameter, and slope, and a reasonable time step is set to realize the numerical iterative solution of the water flow motion state.

[0049] The finite volume method (FLC) is used to spatially discretize the topology-hydrology coupled control equations. Based on the topological relationship between pipe segments and nodes, the pipe network is divided into discrete computational grids with individual pipe segments as units, and nodes as interfaces between these units. Upwind configurations are selected as the numerical flux calculation method for unit interfaces to ensure the stability and convergence of the flow calculations. The discretization accuracy is adaptively adjusted according to pipe material type, pipe diameter, and laying slope. High-precision discretization is used for large-diameter, low-slope pipe segments, while the discretization accuracy is appropriately simplified for small-diameter, high-slope pipe segments, balancing computational efficiency and accuracy. An adaptive time step is set based on CFL conditions, with a time step of 1s for normal operating conditions and reduced to 0.5s for extreme rainfall conditions to avoid numerical divergence. The discretized algebraic equations are solved using an iterative method, with a convergence threshold set to 10. -6 When the difference between two adjacent iterations is less than the threshold, the numerical solution of the water flow motion state is completed, ensuring that the accuracy of hydraulic parameter calculation meets engineering requirements.

[0050] Step S125: Construct a numerical model for blockage migration and deposition, set critical conditions for blockage initiation, migration, deposition, and accumulation by combining water flow velocity and water level rise, correlate blockage particle size, deposition morphology, blockage cross-sectional area changes with water flow hydraulic parameters, embed the blockage evolution rules into the water flow numerical solution process, and realize synchronous coupling simulation of water flow movement and blockage migration.

[0051] A numerical model of blockage migration and deposition was constructed. Based on the pipe network's flow velocity and water level rise, critical conditions such as the critical flow velocity for blockage initiation, the critical water depth for migration, the critical flow rate for deposition, and the accumulation density were set. These critical parameters were calibrated using indoor hydraulic tests and field measurements. The blockage was treated as non-cohesive solid particles, and correlation functions were established between blockage particle size, deposition morphology, and changes in blockage cross-sectional area and hydraulic parameters such as flow velocity, water level, and shear force. The blockage deposition rate showed a negative correlation with flow velocity and a positive correlation with flow saturation. The evolutionary rules of blockage initiation, migration, deposition, and accumulation were embedded into the numerical solution process. A sequential coupling solution method was adopted: first, the flow state was calculated, and then the blockage distribution and deposition state were updated based on hydraulic parameters. This achieved synchronous coupling simulation of flow movement and blockage migration, realistically reproducing the entire physical process of blockage occurrence and development.

[0052] Step S126: Conduct blockage propagation simulation calculations for three types of pipe network topologies: branched, ringed, and mixed. Set initial blockage conditions at different locations and to different degrees in the model, iteratively calculate the changes in water flow state and blockage diffusion process within the pipe network, and record the curves of water level, flow rate, and blockage degree of each node and pipe segment over time after the blockage occurs.

[0053] For three typical pipe network topologies—branched, ring-shaped, and mixed—multi-condition blockage propagation simulations were conducted. Initial blockage conditions were defined in the coupled model, covering different locations such as main pipes, branch pipes, and special nodes, and varying degrees of blockage (mild, moderate, and severe). The initial blockage cross-sectional area was set at 10%, 30%, and 50% of the pipe cross-section. Using real-time hydrological data and topology parameters as boundary conditions, numerical iterative calculations were initiated, with the iteration step size consistent with the flow calculations, dynamically tracking changes in flow state and blockage diffusion within the pipe network. During the simulation, parameters such as water level, flow rate, velocity, blockage cross-sectional area percentage, and blockage level were recorded in real-time for each node and pipe segment after blockage occurred, storing time-series data at a frequency of 1 minute / time to form a complete dynamic curve of blockage propagation. At least 10 sets of simulations were conducted for each type of pipe network topology, covering both normal and extreme rainfall scenarios to ensure the comprehensiveness and representativeness of the simulation results.

[0054] Step S127: Establish a quantitative index system for blockage diffusion characteristics. Extract characteristic parameters such as propagation time, number of affected pipe segments, backlog range, flow attenuation magnitude, and blockage level evolution rate of blockage diffusion through numerical simulation results. Compare the simulation data under different topologies to form a quantitative database of blockage diffusion patterns for various topology pipe networks.

[0055] A multi-dimensional quantitative index system for blockage diffusion characteristics was established. Core characteristic parameters such as blockage diffusion propagation time, number of affected pipe segments, backlog range, flow attenuation magnitude, and blockage level evolution rate were extracted from numerical simulation results, and the definitions and calculation methods of each parameter were clarified. Comparative analysis was conducted on simulation data of three types of pipe networks: branched, ring-shaped, and mixed. The mean, extreme values, and variation patterns of blockage diffusion parameters under different topologies were statistically analyzed, and the correlation between topology structure and blockage diffusion characteristics was extracted. The simulated characteristic parameters, topology types, operating conditions, and simulation results were structured and stored to construct a quantitative database of pipe network blockage diffusion patterns. Database fields include topology type, initial operating conditions, characteristic parameters, and simulation curves, supporting data retrieval and comparative analysis. This database provides standardized data support for subsequent blockage prediction, early warning threshold setting, and pipe network topology optimization design, enhancing the engineering application value of the model.

[0056] Step S128: Use the field-measured blockage event data to calibrate and verify the parameters of the coupled model, adjust the pipe roughness, blockage transport coefficient, and node hydraulic loss coefficient, so that the error between the model simulation results and the actual blockage diffusion characteristics meets the engineering monitoring requirements, and complete the model calibration and optimization.

[0057] Historical blockage event data, including measured data on blockage location, diffusion range, propagation time, and water level / flow rate changes, were collected as calibration and validation samples for the topology-hydrology coupled model. A combination of trial and error and least squares methods was used to iteratively adjust key parameters in the model, such as pipe roughness, blockage transport coefficient, and nodal hydraulic loss coefficient. Using the error between simulation results and measured data as the optimization objective, engineering allowable error thresholds were set: propagation time error ≤ 10%, backwater height error ≤ 5%, and flow rate attenuation error ≤ 8%. Parameters were adjusted multiple times and simulations were repeated until the model calculation results met the error requirements, completing model calibration. After calibration, the model was validated using independent measured blockage events, achieving a validation pass rate ≥ 95%, ensuring that the coupled model can accurately reflect the blockage diffusion patterns of the actual pipe network and possesses reliability and accuracy for engineering applications.

[0058] Step S129: Establish a real-time linkage interface between the model and the topology digital twin. When the digital twin model completes the topology update due to pipeline renovation and structural changes, the coupled model automatically reads the updated geometric and topology data, reconstructs the computing units and boundary conditions, and adapts to the new pipeline structure.

[0059] A real-time linkage interface was established between the topology-hydrological coupling model and the topology digital twin model, using the MQTT communication protocol to achieve bidirectional data transmission, with an interface response time of ≤1s. A topology update trigger mechanism was established; when the digital twin model completes an incremental topology update due to pipeline renovation, maintenance, structural damage, etc., the interface automatically pushes an update notification, and the coupling model reads the updated geometric parameters, topological relationships, node attributes, and other data in real time. Based on the new pipeline topology, the numerical calculation unit was automatically reconstructed, the discrete mesh was re-divided, and the flow boundary conditions and topological constraint rules were adjusted to quickly adapt to the changed pipeline structure. The reconstruction process requires no manual intervention, and the computational mesh reconstruction time is ≤5min, ensuring that the model topology remains consistent with the actual pipeline network, avoiding simulation failure due to pipeline topology changes, and guaranteeing the continuous and stable operation of the coupling model.

[0060] Step S1210: The spatiotemporal data and quantitative characteristic results of the blockage diffusion obtained from the simulation are transmitted back to the topological digital twin model in real time, and the blockage propagation path, development trend and impact range are visualized in the three-dimensional twin scene.

[0061] The spatiotemporal data and quantitative characteristics of blockage propagation simulated by the topology-hydrology coupling model are transmitted back to the topology digital twin model via a real-time linkage interface at a frequency of 1 minute per transmission. Dynamic rendering technology is used in the 3D twin scene to visualize the blockage propagation path, development trend, and impact range. Different colors are used to indicate the blockage level: red for severe blockage, yellow for moderate blockage, and blue for mild blockage. Dynamic arrows indicate the direction of blockage propagation, and numerical labels display quantitative indicators such as propagation time and flow attenuation. The system supports replay, fast forward, and pause operations of the blockage propagation process, allowing users to view the blockage status and hydraulic parameters of any pipe segment at any time. The visualization results are simultaneously pushed to the operation and maintenance terminal, providing intuitive 3D decision-making support for on-site blockage handling and emergency dispatch, achieving deep integration of blockage simulation data and the twin visualization scene.

[0062] Step S130: Establish a multi-dimensional blockage classification index system, integrate key hydraulic characteristic indicators, use the objective weighting method to determine the index weights and classify blockage levels, and dynamically correct the classification thresholds in combination with pipeline network structure parameters.

[0063] This step constructs a multi-dimensional blockage classification index system covering hydraulic response and pipeline structure. It integrates four core hydraulic characteristic indicators: flow attenuation rate, water level rise, proportion of blockage cross-sectional area, and decrease in water velocity. Abandoning subjective weighting methods, it adopts the entropy weighting method as an objective weighting means to determine the weight of each indicator. Weights are allocated based on the indicators' ability to distinguish blockage states. Combining pipeline operation and maintenance industry standards and historical blockage cases, blockage levels are classified into three levels: mild, moderate, and severe. For differences in pipeline structures with different pipe diameters, laying slopes, and pipe material roughness, a dynamic correction mechanism for classification thresholds is established. This correlates topological parameters with hydraulic index thresholds, allowing the classification thresholds to adaptively match different pipeline topologies such as branched, ringed, and mixed structures, avoiding classification deviations caused by uniform thresholds. The index system balances hydraulic response sensitivity and structural adaptability, with a classification judgment response time ≤1 second and a matching degree of ≥95% between the classification results and the actual pipeline blockage state, providing a quantitative basis for blockage early warning and treatment.

[0064] Step S131: Screen relevant characteristic indicators of hydraulic response to pipe network blockage, determine the indicator set as flow rate decay rate, water level rise, proportion of blockage cross-sectional area, and water velocity decrease, and clarify the definition, monitoring method and collection frequency of each indicator;

[0065] Core characteristics of the hydraulic response to pipe network blockage were selected, and the indicator set was determined to be flow attenuation rate, water level rise, proportion of blockage cross-sectional area, and flow velocity decrease. The flow attenuation rate is the ratio of the actual flow rate to the design flow rate at a node; the water level rise is the difference between the real-time monitored water level and the normal operating water level; the proportion of blockage cross-sectional area is the ratio of the cross-sectional area of ​​the blockage to the total cross-sectional area of ​​the pipe; and the flow velocity decrease is the ratio of the difference between the actual flow velocity and the design flow velocity to the design flow velocity. Flow rate and velocity were monitored using radar flow sensors, water level using ultrasonic water level sensors, and the blockage cross-sectional area was obtained through a combination of lidar scanning and hydraulic inversion. All indicators were collected at a frequency of 1 minute. The monitoring points, collection methods, and data formats for each indicator were clearly defined to ensure that the indicators are measurable, sensitive, and independent, with a monitoring error ≤ ±3%, meeting the quantitative judgment requirements for blockage classification.

[0066] Step S132: Collect measured data of various blockage scenarios and normal operation under different topologies, pipe diameters and operating conditions. After removing anomalies by the 3σ criterion and supplementing missing data by linear interpolation, eliminate the dimensions by min-max standardization and map the data to the [0,1] interval to form a standardized index dataset.

[0067] Measured data were collected under various pipeline network topologies, pipe diameters, and operating conditions, covering all scenarios from normal operation to minor, moderate, and severe blockage. Data sources included real-time data from multimodal monitoring units, historical blockage maintenance records, and hydraulic simulation data. The 3σ criterion was used to remove outliers, eliminating invalid data such as sudden changes in flow rate and abnormal water level fluctuations. Linear interpolation was used to fill in missing data periods to ensure data sequence continuity. The min-max standardization algorithm was used to normalize the data for each indicator, eliminating dimensional differences between indicators such as flow rate, water level, and cross-sectional area, mapping all data to the [0,1] value range. The preprocessed data was structured by node ID and timestamp to form a standardized indicator dataset. The dataset had an effective data rate ≥98%, free from outliers and missing values, providing a reliable data foundation for subsequent indicator weighting and classification.

[0068] Step S133: Using the entropy weight method as an objective weighting tool, first calculate the proportion of the i-th indicator and the j-th sample, then calculate the entropy value of the indicator based on the proportion, and use the entropy value to reflect the degree of disorder of the indicator information, and then calculate the indicator difference coefficient. Based on the proportion of the difference coefficient, determine the weight of each indicator, and verify the rationality of the weight through consistency test.

[0069] The entropy weight method is used for objective weighting of indicators. First, the numerical proportion of the j-th sample for the i-th indicator is calculated to eliminate the influence of data magnitude. Then, the information entropy value of each indicator is calculated based on the sample proportion. The smaller the entropy value, the higher the dispersion of the indicator data and the stronger the ability to distinguish blockage states. The index difference coefficient is obtained by subtracting the entropy value from 1. The final weights of four indicators—flow attenuation rate, water level rise, proportion of blockage cross-sectional area, and water velocity decrease—are determined based on the proportion of the difference coefficient. After the weight calculation is completed, a consistency test is used to verify its rationality. If the test coefficient CR < 0.1, the weight allocation is considered valid, avoiding weight imbalance that could lead to distorted classification. The entropy weight method involves no subjective human intervention throughout the process, and the weighting results closely reflect the actual operating patterns of the pipeline network. The indicator weights can be dynamically adjusted as the dataset is updated, ensuring the objectivity and applicability of the classification system.

[0070] Step S134: Combining industry operation and maintenance standards, historical blockage handling cases and hydraulic simulation results, the K-means clustering algorithm is used to perform cluster analysis on standardized indicator data, determine the indicator values ​​corresponding to the cluster centers, and classify the blockage into three levels: light, moderate and severe. It is clarified that light blockage is when a single indicator slightly deviates from the normal range and does not affect the overall drainage efficiency; moderate blockage is when two or more indicators exceed the standard and the drainage capacity is significantly reduced; and severe blockage is when all indicators significantly exceed the standard and there is a risk of water accumulation. At the same time, the threshold range of the indicators corresponding to each level is quantified.

[0071] Combining industry standards for drainage network operation and maintenance, historical blockage handling cases, and topology-hydrological coupling simulation results, the K-means clustering algorithm was used to perform unsupervised clustering analysis on the standardized indicator dataset. Three clusters were set, corresponding to three levels of blockage: mild, moderate, and severe. The threshold ranges for each blockage level were determined by the cluster center values. Mild blockage was defined as a slight deviation of a single indicator from the normal range, with no impact on the overall drainage efficiency of the network; moderate blockage was defined as two or more indicators exceeding the standard, resulting in a significant decrease in the network's drainage capacity; and severe blockage was defined as all indicators significantly exceeding the standard, posing a risk of surface water accumulation in the network. The clustering results were compared and optimized with actual blockage scenarios, quantifying the threshold boundaries for each indicator level. The threshold classification accuracy was ≤0.01, and the consistency rate between the classification results and the blockage status measured by the pipeline inspection robot was ≥95%, forming a quantitative classification standard for blockage levels that can be directly applied in engineering.

[0072] Step S135: Select pipe diameter, laying slope, and pipe roughness as correction parameters, collect hydraulic test data of the pipeline network under different combinations of structural parameters, construct a multiple linear regression correction model, with pipe diameter D and slope i as independent variables and the grading threshold T of each index as dependent variables, and obtain the correction formula T'=T×(aD+bi+c) by least squares fitting, where a, b, and c are fitting coefficients, and set roughness correction coefficients for different pipe types such as concrete pipe, plastic pipe, and cast iron pipe;

[0073] Pipe diameter, laying slope, and pipe roughness were selected as the core correction parameters for the classification threshold. Hydraulic test and measured data of the pipeline network under different parameter combinations were collected to construct a multiple linear regression correction model. With pipe diameter D and laying slope i as independent variables, and the classification threshold T for each hydraulic index as the dependent variable, the least squares method was used to fit the threshold correction formula T'=T×(aD+bi+c), where a, b, and c are fitting coefficients, determined through iterative solutions using multiple sets of experimental data. For different pipe types such as concrete pipes, plastic pipes, and cast iron pipes, corresponding roughness correction coefficients were set according to the differences in pipe wall roughness, and the threshold was fine-tuned a second time. The correction model can adapt to pipeline network structures with different pipe diameters, slopes, and pipe materials, eliminating classification errors caused by differences in pipeline physical parameters. The deviation after threshold correction is ≤±2%, enabling the classification index system to have cross-scenario and cross-topology adaptability.

[0074] Step S136: Input the pipeline network structure parameters into the correction model to obtain personalized classification thresholds, select typical pipeline network topologies to carry out field tests, combine the actual blockage status obtained by the pipeline inspection robot, compare and verify the classification results of the index system and calculate the accuracy rate, iteratively adjust and optimize for classification deviation or index sensitivity issues, and form a stable and reliable multi-dimensional blockage classification index system.

[0075] The structural parameters of the target pipeline network, such as pipe diameter, laying slope, and pipe material type, are input into a threshold correction model to calculate a personalized grading threshold suitable for the network. Field verification tests are conducted on three typical pipeline topologies: branched, ring-shaped, and mixed. Pipeline inspection robots are used to detect the blockage status of pipe sections in the field, obtaining the actual blockage level as a verification standard. The automatic grading results of the indicator system are compared with the measured results to calculate the grading accuracy, which is required to be ≥95%. If the grading deviation exceeds 5%, the indicator weights, clustering parameters, or model coefficients are iteratively adjusted; if the sensitivity of a single indicator is insufficient, the indicator monitoring method or data preprocessing algorithm is optimized. Through multiple rounds of verification and iterative optimization, a stable and reliable multi-dimensional blockage grading indicator system is formed, which can accurately determine the pipeline network blockage level in real time, supporting subsequent blockage prediction and emergency response.

[0076] Step S140: Construct a hybrid prediction model that combines data-driven and physical constraints, integrates historical data time series characteristics with hydraulic physics principles, and integrates multi-source environmental and pipeline load data to achieve prediction of blockage escalation under normal scenarios and time series prediction of blockage development under extreme weather conditions;

[0077] A hybrid prediction model combining data-driven and physical constraints is constructed. Long Short-Term Memory (LSTM) networks are used to mine the temporal evolution characteristics of historical congestion data, and one-dimensional Saint-Venant equations and nodal flow balance equations are used to construct hydraulic physical constraints, achieving a deep integration of data patterns and hydraulic principles. The model integrates multi-source data on historical congestion events, real-time pipeline operation, meteorological environment, pipeline load, and extreme weather, taking into account the prediction needs of both normal and extreme rainfall conditions. Under normal scenarios, it can predict the escalation trend of congestion severity within 1-6 hours, and under extreme weather conditions, it can accurately predict the congestion development timeline and the arrival time of severe congestion. The model's prediction response time is ≤100 milliseconds, and the temporal prediction error is ≤10%. The model is adaptable to different pipeline topologies and can be linked with a topological digital twin model to update parameters in real time, providing core algorithmic support for comprehensive pipeline congestion early warning.

[0078] Step S141: Collect multi-source basic data on historical blockage events, pipeline operation, environment, pipeline load and extreme weather, use outlier detection algorithm to remove invalid data, use interpolation to supplement missing data, and construct a standardized input dataset after spatiotemporal alignment and normalization.

[0079] Five types of multi-source basic data were collected: historical blockage events, real-time pipeline operation, meteorological environment, pipeline load, and extreme weather data. The data covers all dimensions of information, including blockage time, location, level, water level, flow rate, rainfall, and radar echo. A 3σ outlier detection algorithm was used to remove invalid data such as sudden flow changes and distorted water levels. Linear interpolation was used to fill in missing values ​​for data periods. Using node ID as the spatial reference and 1-minute time granularity, spatiotemporal alignment of the multi-source data was achieved, eliminating spatial misalignment and temporal asynchrony issues. Min-max normalization was used to eliminate dimensional differences between different physical quantities, mapping the data to the [0,1] interval to form a standardized input dataset. The dataset was divided into training and testing sets in a 7:3 ratio, with a data validity rate ≥98%, providing standardized, high-quality input data for the training and validation of the hybrid prediction model.

[0080] Step S142: Using the Saint-Venant equation as the core, and combining the flow balance equation and energy equation of the pipeline node, a physical constraint system is constructed, and the physical constraints are transformed into mathematical hard constraints, which limit the prediction results to conform to the principles of hydraulic dynamics.

[0081] Using the one-dimensional unsteady flow Saint-Venant equation as the core hydraulic governing equation, combined with the flow balance equation and energy equation of the pipeline network nodes, a physical constraint system adapted to complex pipeline networks is constructed. Hydraulic principles such as flow continuity, momentum conservation, and flow balance are transformed into mathematical hard constraints, ensuring that the predicted flow rate, water level, and velocity parameters conform to the laws of hydraulic dynamics. Constraint boundary conditions are set separately for different pipeline network topologies, including branched, ring-shaped, and mixed networks; and hydraulic connection constraints are defined separately for special nodes such as drop wells and intercepting wells. Physical constraints are embedded throughout the model prediction process, avoiding unreasonable results that violate physical laws from data-driven models. The calculation error of the constraint conditions is ≤±3%, ensuring that the prediction results both conform to historical data patterns and the actual hydraulic operation logic of the pipeline network, thus improving the rationality and reliability of the model prediction.

[0082] Step S143: Extract time-series features and correlation features from standardized data, strengthen the weights of key extreme weather features through attention mechanisms, and complete model feature engineering;

[0083] Feature engineering was performed on standardized input data to extract two core feature categories: temporal features and correlation features. Temporal features included the water level and flow rate trends in the 24 hours prior to blockage, cumulative water level increments, and flow rate decay rates. Correlation features included the coupling coefficient between rainfall and flow rate, and the correlation between pipeline load rate and blockage probability. For extreme rainfall scenarios, an attention mechanism was introduced to strengthen the weights of key features such as rainfall intensity, radar echo intensity, and short-duration rainfall peaks, improving the model's feature recognition ability for extreme weather. Feature data underwent screening and dimensionality reduction, eliminating redundant features and retaining the top 80% of core features contributing to blockage prediction, with a feature dimensionality compression rate of ≤30%. After feature engineering, the model's input features became more targeted, improving prediction accuracy for extreme weather scenarios by ≥10%, effectively enhancing the model's scenario adaptability.

[0084] Step S144: A data-driven model framework is built using a long short-term memory network, an overfitting suppression mechanism is introduced, physical constraints are embedded into the model training process, a hybrid prediction fusion architecture is constructed, and a coupled iterative prediction module for rainfall, flow rate and blockage under extreme weather conditions is built based on the rainstorm intensity formula and the hydraulic carrying capacity of the pipe network.

[0085] A data-driven model framework is built using a Long Short-Term Memory (LSTM) network with 3-5 hidden layers. Dropout layers are inserted between the hidden layers, and L2 regularization terms are added to construct an overfitting suppression mechanism and improve the model's generalization ability. Hydraulic physical constraints are embedded into the model's loss function, constructing a hybrid prediction architecture that integrates prediction loss and physical constraint loss. Both types of losses are simultaneously optimized through backpropagation. A coupled iterative prediction module based on regional rainfall intensity formulas and the hydraulic ultimate bearing capacity of the pipe network is built, calculating the blockage development timeline iteratively with a 1-minute step size under extreme weather conditions. The model is trained using the Adam optimizer, with an early stopping mechanism to avoid overtraining. It automatically switches between normal and extreme modes, with a model training convergence time ≤ 4 hours. The prediction accuracy meets engineering monitoring requirements, enabling real-time and accurate prediction of blockage development trends.

[0086] Step S1441: Configure the input layer, 3-5 hidden layers, and output layer of the LSTM network according to the input feature dimension, set the corresponding activation function, and use the Xavier method to initialize the network weights;

[0087] The LSTM network structure is configured based on the total dimensions of the input features (temporal features + correlation features + physical constraint parameters): the number of neurons in the input layer corresponds one-to-one with the feature dimensions, with 3-5 hidden layers (128-256 neurons per layer), and 2 neurons in the output layer (corresponding to the progression probabilities of mild → moderate and moderate → severe). Activation functions are selected: the hidden layers use the ReLU activation function (to alleviate gradient vanishing), and the output layer uses the Sigmoid activation function (mapping probabilities to the [0,1] interval). The network weights are initialized using the Xavier method, adaptively allocating initial weights by calculating the input and output dimensions to ensure consistent output variance across layers and avoid gradient explosion or vanishing in the early stages of training. After initialization, the mean weight is controlled within the [-0.05, 0.05] interval to ensure stable model training and lay the foundation for subsequent feature extraction and constraint fusion.

[0088] Step S1442: Insert dropout layers between LSTM hidden layers and add an L2 regularization term to the loss function to build an overfitting suppression mechanism.

[0089] Dropout layers are inserted between adjacent hidden layers in the LSTM network, with a dropout probability of 0.2-0.3. During training, some neurons are randomly deactivated to break excessive dependencies between features. Simultaneously, an L2 regularization term is added to the model's loss function with a regularization coefficient λ=0.001. By penalizing the network weight parameters, the absolute value of the weights is limited, reducing model complexity. The overfitting suppression mechanism must meet the following requirements: the loss deviation between the training and test sets ≤10%, and the decrease in prediction accuracy on new datasets ≤5%. This mechanism effectively improves the model's generalization ability and avoids prediction failures in extreme weather or new topological scenarios due to overfitting of training data.

[0090] Step S1443: Using the discrete form of the Saint-Venant equation and the nodal flow balance equation as hydraulic hard constraints, construct a total constraint loss function that includes prediction loss and physical constraint error;

[0091] Using the one-dimensional Saint-Venant equations (continuity equation + momentum equation) and nodal flow balance equations discretized by the finite volume method as hydraulic hard constraints, they are transformed into the mathematical expression: Q pred =f(H pred ,D,i,n) (Q is the flow rate, H is the water level, D is the pipe diameter, i is the slope, n is the roughness), ΣQ in =ΣQ out Construct the total constraint loss function: Loss total =Loss pred +α×Loss phy Loss pred The mean squared error (MSE) is used to calculate the deviation between the predicted value and the true label. Loss phy The deviation between the predicted flow / water level and the derived values ​​of the hydraulic equations is calculated, with α being the constraint weight (initially 0.3-0.5). This function confines the prediction results within the physically feasible region, ensuring that the output does not violate the laws of hydraulic dynamics, with a physical constraint error ≤8%.

[0092] Step S1444: Using a parallel input mode, the temporal features, correlation features, and physical constraint parameters are simultaneously input into the LSTM model to build a hybrid prediction fusion architecture. The prediction loss and physical constraint loss are simultaneously optimized through backpropagation.

[0093] A hybrid prediction and fusion model is built using a parallel input architecture. First, the dimensions of temporal features, correlation features, and physical constraint parameters are adapted, expanding static features to a sequence dimension. Then, three independent input branches are constructed to extract three types of features and adapt them to the output dimension. Feature concatenation and the main LSTM layer capture fusion patterns, outputting the congestion escalation probability. A dual loss function is defined, and the Adam optimizer is used to optimize the total loss through backpropagation. Phased training is implemented (50 epochs of pre-training + 100-200 epochs of fusion training), dynamically adjusting the physical constraint weights. During training, gradient clipping (clipnorm=1.0), a batch size of 64, and an early stopping mechanism (verifying that the loss does not decrease for 10 consecutive epochs) ensure stability, ultimately achieving coordinated convergence of the prediction loss and the physical constraint loss.

[0094] Step S14441: Perform feature dimension adaptation preprocessing on temporal features, correlation features and physical constraint parameters, and extend the correlation features and physical constraint parameters to the sequence dimension that matches the temporal features to ensure that the time step dimensions of the three are consistent to meet the parallel input requirements;

[0095] Input features undergo dimensionality adaptation preprocessing: The time-series features are sequence-type data with [number of samples, time step, number of time-series features], while the correlation features (static) and physical constraint parameters (static) initially have the shape [number of samples, number of features]. A dimensionality expansion algorithm is used to repeatedly expand the static features along the time step dimension, unifying their shape to [number of samples, time step, number of features], consistent with the time step dimension of the time-series features (time step size set to 24-48, corresponding to 1-2 hours of monitoring data). During the expansion process, timestamp alignment ensures data temporal consistency and adapts to parallel input requirements.

[0096] Step S14442: Construct a parallel LSTM network submodule with three independent input branches. The temporal feature branch extracts temporal dependency features through 1-2 layers of LSTM sublayers. The correlation feature branch and the physical constraint parameter branch extract corresponding features through fully connected layers, and the output feature dimensions of each branch are adapted.

[0097] Three independent parallel input branches are constructed: the temporal feature branch receives sequential data, concatenates 1-2 LSTM sub-layers (128 neurons per layer, ReLU activation), extracts temporal dependent features, and outputs a dimension of [time step, 128]; the correlation feature branch receives expanded static features, extracts nonlinear correlation features through a fully connected layer (64 neurons, ReLU activation), and outputs a dimension of [time step, 64]; the physical constraint parameter branch maps parameters such as pipe diameter and slope to the feature space through a fully connected layer (64 neurons, ReLU activation), and outputs a dimension of [time step, 64].

[0098] Step S14443: The output features of the three branches are concatenated and fused and then input into the main LSTM layer. After feature temporal pattern capture and global average pooling, the probability of escalation of congestion level under normal scenarios is output through the output layer.

[0099] The features output from the three branches are concatenated along the channel dimension, resulting in a feature dimension of [time step, 256] (128+64+64). The fused features are then input into two main LSTM layers (256 neurons per layer, ReLU activation) to capture the temporal evolution of the fused features and their correlation with physical constraints, maintaining the [time step, 256] output dimension. A global average pooling layer compresses the sequence output into

[256] -dimensional static features, eliminating redundancy in the time step dimension. This is then connected to the output layer (2 neurons, Sigmoid activation) to output the probability of congestion escalation from mild to moderate and from moderate to severe within 1-6 hours under normal scenarios. The feature retention rate during pooling is ≥90%.

[0100] Step S14444: Define a dual loss function that includes prediction loss and physical constraint loss. The prediction loss is calculated using mean square error to determine the deviation between the model output and the true label. The physical constraint loss is calculated using the deviation between the model prediction value and the hydraulic equation derivation value. Construct a total loss function to balance the contributions of the two types of losses.

[0101] Predicted Loss pred The mean squared error is used for calculation, and the formula is Loss. pred =(1 / N)×Σ(y pred -y true )² (N is the number of samples, y pred To predict the probability, y true (For real labels); Physical constraint loss phy The deviation between the predicted flow / water level and the derived value from the hydraulic equation is calculated using the formula Loss. phy =(1 / N)×[Σ(Q pred -Q phy )²+Σ(H pred -H phy )²](Q phy H phy (These are the derived values ​​of the equation). The total loss function is Loss. total =Loss pred +α×Loss phy The initial value of α is 0.3-0.5, used to balance the contributions of the two types of losses. This function enables the model training to simultaneously optimize prediction accuracy and physical consistency, avoiding distortion caused by a single loss.

[0102] Step S14445: The Adam optimizer is used to backpropagate the total loss gradient to each input branch and the main LSTM layer, decompose the gradient components and update the corresponding branch weights, and apply L2 regularization to the physical constraint parameter branch.

[0103] The Adam optimizer (initial learning rate 0.001, adaptively decaying with each training epoch) is used to backpropagate the total loss gradient to the three input branches and the main LSTM layer. The gradient propagation path is: output layer → main LSTM layer → feature concatenation layer → three input branches, automatically decomposed into gradient components corresponding to the three types of features. The LSTM sub-layer of the temporal feature branch and the fully connected layer of the correlation and physical constraint branches update the weights according to the gradient components. Simultaneously, L2 regularization (coefficient 0.001) is applied to the weights of the physical constraint parameter branch to limit excessive weight growth. During backpropagation, the gradient update step size is ≤0.01 to ensure training stability, and the physical constraint error of the model after weight updates is ≤8%.

[0104] Step S14446: Implement phased training. In the pre-training phase, fix the weights of the physical constraint branch and train the other branches. In the fusion training phase, release the weights of the branch and dynamically adjust the physical constraint weight coefficients to ensure that the total loss and physical constraint error are controllable.

[0105] In the pre-training phase (first 50 rounds), the weights of the physical constraint branches are fixed, and only the temporal and correlation feature branches are trained, aiming to converge the prediction loss to MSE < 0.05. In the fusion training phase (rounds 51-200), the weights of the physical constraint branches are released, and the α value is dynamically adjusted every 10 rounds. If the loss is low... phy >0.5×Loss pred α is increased by 0.1 (maximum 0.8); if Loss phy <0.1×Loss pred α is lowered by 0.1 (minimum 0.2). Phased training avoids excessive interference from initial physical constraints on data pattern learning. Later, weight adjustments achieve synergistic optimization of prediction accuracy and physical consistency, ensuring a continuous decrease in total loss, ultimately reducing the loss. total <0.03.

[0106] Step S14447: Prune the total loss gradient, set the batch size and early stopping mechanism to control training stability, and verify the improvement effect of prediction error, physical constraint error and model prediction accuracy through the test set to ensure that the performance of the fusion architecture meets the standards.

[0107] The total loss gradient is clipped (clipnorm=1.0) to prevent gradient explosion and training divergence. The batch size is set to 64 to balance training efficiency and memory usage. An early stopping mechanism is enabled: training stops when the total loss on the validation set fails to decrease for 10 consecutive rounds to avoid overfitting. The performance of the fusion architecture is validated using a test set: the deviation between the prediction loss test error and training error should be <10%, the physical constraint loss <0.04, the prediction accuracy of the fusion model should be ≥5% higher than the pure LSTM model, and the deviation between the predicted flow / water level and the derived physical equation value should be ≤8%. If these targets are not met, the branch layer structure or hyperparameters are adjusted until the performance requirements are met.

[0108] Step S1445: Based on the rainstorm intensity formula for the target area, calibrate the rainfall calculation coefficient, and combine the pipeline network structure parameters to determine the hydraulic ultimate bearing capacity and the critical velocity threshold of the blockage;

[0109] Based on historical heavy rainfall monitoring data for the target area, the least squares method is used to calibrate the heavy rainfall intensity formula i=A / (t+B). n The regional parameters A, B, and n (calibration error ≤ 5%) are used. Combined with the pipe diameter, laying slope, and pipe material roughness, the hydraulic ultimate bearing capacity of the pipe network is simulated and determined using hydraulic calculation software (such as SWMM), including the maximum allowable flow rate (calculated under full-flow conditions) and the highest allowable water level (≤ 0.1m from the top of the pipe). Through indoor hydraulic tests and field measurements, the critical flow velocity threshold for blockage initiation is calibrated (0.6-0.8 m / s for sandy blockages, 0.8-1.0 m / s for fibrous blockages), providing parameter support for simulating blockage evolution under extreme weather conditions. After parameter calibration, the model's adaptability to operating conditions is ≥ 95%.

[0110] Step S1446: Input the real-time rainfall intensity and the predicted peak rainfall, construct the rainfall-flow-blockage coupled iterative logic, and iteratively calculate the node flow and blockage accumulation status until the iteration converges;

[0111] Input real-time rainfall intensity (1-minute granularity) and radar echo prediction of peak rainfall for the next 15-30 minutes. Calculate the design flow increment for the corresponding time period based on the calibrated storm intensity formula. Combined with the real-time load rate of the pipeline network (actual flow / design flow), calculate the flow change at each node with an iteration step of 1 minute. Establish a correlation function for blockage accumulation rate: accumulation rate = k × (1 / v) × S (k is the accumulation coefficient, v is the flow velocity, and S is the flow saturation), and update the proportion of blockage cross-sectional area synchronously. When the blockage proportion at a node reaches a preset threshold (30% for moderate, 50% for severe), trigger downstream flow redistribution calculation. Convergence is determined when the change in blockage degree between two adjacent iterations is <1%, and the blockage development time series curve is output.

[0112] Step S1447: Establish a collaborative mechanism between the conventional scenario LSTM hybrid model and the extreme weather coupled iterative module, switch the prediction mode according to the meteorological warning and realize data interaction, and output the congestion development time series prediction results;

[0113] A collaborative mechanism is established between a hybrid LSTM model for routine scenarios and an iterative module for extreme weather conditions. This mechanism involves triggering mode switching via weather warnings, implementing data interaction between the two modules through a standardized interface, ensuring data synchronization through memory-level caching, outputting prediction results with differentiated logic, and employing a fault-tolerant rollback mechanism to prevent failures. The routine model focuses on predicting escalation trends over 1-6 hours, while the extreme model focuses on fine-grained time series forecasts over 1-3 hours. The switching latency between the two models is ≤100ms, and data consistency is ≥98%. Historical heavy rainfall data is used to verify the collaborative effect, requiring a trigger threshold matching degree of ≥85%, a time series prediction error of ≤10 minutes, and a model rollback deviation of ≤5%, ensuring continuous and reliable predictions under different weather scenarios.

[0114] Step S14471: Construct a meteorological early warning triggering judgment rule system, connect to the meteorological data interface to obtain rainfall, radar echo intensity and rainstorm warning level data, set extreme weather trigger thresholds and configure a buffer mechanism for continuous periodic judgment, and switch between normal and extreme weather prediction modes accordingly.

[0115] A meteorological warning triggering rule system is established, connecting to the meteorological department's data platform via API to obtain real-time data on hourly rainfall, radar echo intensity, and rainstorm warning level. Extreme weather trigger thresholds are set: 1-hour rainfall ≥ 20mm, radar echo intensity ≥ 45dBZ, or receipt of a blue or higher rainstorm warning; meeting any one of these conditions initiates a candidate trigger state. A threshold buffer mechanism is configured: maintaining the candidate state for three consecutive 5-minute monitoring cycles is required before officially switching to extreme weather prediction mode, avoiding frequent switching due to short-term fluctuations.

[0116] Step S14472: Design a standardized bidirectional data interaction interface across modules, define a transmission data structure including node ID, timestamp, data type, value and confidence level, and realize bidirectional data calls between the conventional LSTM hybrid model and the extreme weather coupled iterative module;

[0117] A standardized bidirectional data interaction interface is designed across modules, using JSON format to define the data structure. Core fields include node ID (unique identifier), timestamp (error ≤ 10ms), data type (status data / prediction result), numerical value (accuracy ≤ 0.01), and confidence level (0-1 range). The interface supports bidirectional access to the current congestion level, escalation probability, and real-time hydraulic parameters output by the conventional LSTM model, and the congestion time-series curve, severe congestion arrival time, and flow redistribution results output by the extreme weather module. The interface uses the MQTT protocol for transmission, with a data transmission latency ≤ 50ms, and supports breakpoint resumption to ensure the real-time performance and integrity of data interaction between the two modules.

[0118] Step S14473: Define the dual-mode operation logic. In the normal mode, the LSTM hybrid model outputs the congestion escalation trend and time independently at a fixed period. After the extreme mode is triggered, the current congestion state output by the LSTM model is used as the initial condition, the cycle of the coupled iteration module is switched and the short-cycle congestion level change and warning time window are output.

[0119] In normal mode, the LSTM hybrid model runs independently with a 5-minute cycle, outputting the congestion escalation trend (mild → moderate, moderate → severe) and escalation time for each node over the next 1-6 hours. When the extreme mode is triggered, the independent LSTM prediction is paused, and its output of the current congestion status (congestion percentage, water level / flow rate) is used as the initial boundary conditions for the coupled iterative module. In extreme mode, the coupled iterative module switches to a 1-minute iteration cycle, outputting the congestion level change every 5 minutes under the peak rainfall over the next 15-30 minutes, and the severe congestion warning time window (accuracy ≤ 5 minutes).

[0120] Step S14474: Use a memory-level data sharing cache as the interaction carrier to complete the read and write synchronization of the two modules on a periodic basis, and combine it with a timestamp verification mechanism to ensure that the data status of the two modules is consistent.

[0121] A Redis in-memory data sharing cache is used as the interaction medium between the two modules. The conventional LSTM model writes the latest node monitoring data and intermediate prediction results to the cache every 5 minutes, with a data retention time of ≥24 hours. The extreme weather module reads cached data every 1 minute, updates the initial conditions for iteration, and writes back the traffic changes and congestion accumulation rates obtained during iteration to the cache. A timestamp verification mechanism is set up to synchronize only the valid data corresponding to the latest timestamp when reading data, discarding expired and redundant data to ensure the consistency of data status between the two modules.

[0122] Step S14475: Construct a logic for fusing and outputting differentiated prediction results. In the normal mode, the model prediction results are directly output. In the extreme mode, the time series curve of the coupled iterative module is used as the core, and the probability of the LSTM model is superimposed as the confidence weight to distinguish the output emphasis of the two modes.

[0123] The standard mode directly outputs the congestion escalation probability and escalation time from the LSTM model, focusing on trend prediction over a 1-6 hour period, providing a reference for daily operations and maintenance. The extreme mode uses the congestion development time-series curve of the coupled iterative module as the core, superimposed with the LSTM model's escalation probability as a confidence weight (e.g., 90% confidence if severe congestion is reached at time T), focusing on fine-grained time-series analysis and risk warnings over a short period of 1-3 hours, providing a basis for emergency response. The output results support visualization and API push. In the extreme mode, warning information is prioritized for push to emergency terminals, with a response latency of ≤30 seconds.

[0124] Step S14476: Add a mode switching fault tolerance and rollback mechanism to achieve a smooth rollback from extreme mode to normal mode based on changes in meteorological data. At the same time, switch emergency predictions and trigger alarms for scenarios where iterative non-convergence occurs.

[0125] In extreme mode, if meteorological data is below the trigger threshold for three consecutive 1-minute periods, a rollback process is automatically initiated. The congestion status of the last iteration of the extreme mode is used as the initial input for restarting the LSTM model to avoid prediction gaps. If the extreme mode iteration fails to converge (the congestion degree changes by more than 1% in 5 consecutive iterations), the system temporarily switches to the LSTM model for emergency prediction, and an alarm is triggered (pushed to the operation and maintenance platform + SMS notification), prompting staff to check the iteration parameters (such as the siltation coefficient and boundary conditions).

[0126] Step S14477: Select historical rainstorm monitoring data to verify the dual-mode collaborative effect, and adjust the trigger threshold, iteration step size and data synchronization frequency according to the preset performance indicators.

[0127] Pipeline monitoring data from historical rainstorm events over the past three years (including more than five severe rainstorm events) were selected to simulate the dual-mode collaborative operation and switching process. Core validation indicators were set as follows: the matching degree between the trigger threshold and the actual rainstorm impact period ≥ 85%, data synchronization latency ≤ 100ms, extreme mode congestion timing prediction error ≤ 10 minutes, and prediction result deviation before and after mode rollback ≤ 5%. Iterative adjustments were made based on the validation results: if the matching degree was insufficient, the trigger threshold was optimized; if the latency exceeded the standard, the cache read / write efficiency was improved; if the error was too large, the iteration step size or data synchronization frequency was adjusted until all indicators met the requirements, ensuring the dual-mode collaborative effect met the standards.

[0128] Step S1448: Train the model using the Adam optimizer, and combine early stopping mechanism with dynamic physical constraint weight adjustment to complete the model training optimization.

[0129] The Adam optimizer was used to train the hybrid model, with an initial learning rate of 0.001 that decayed exponentially with each training epoch (decay coefficient 0.95). An early stopping mechanism was implemented, halting training when the validation set loss did not decrease for 10 consecutive epochs to prevent overtraining and a decline in generalization ability. The physical constraint weights α were dynamically adjusted during training: the loss was calculated every 10 epochs. phy With Loss pred The ratio of Loss phy >0.5×Loss pred α is increased by 0.1 (maximum 0.8); if Loss phy <0.1×Loss pred α was lowered by 0.1 (minimum 0.2) to ensure that the physical constraint error was controllable. After the model training converged, the total loss Loss_total < 0.03%.

[0130] Step S145: Divide the standardized dataset into training and testing sets, train the hybrid model using an optimizer and a preset loss function, and improve prediction accuracy through hyperparameter optimization;

[0131] The standardized dataset was randomly divided into training and test sets in a 7:3 ratio. The training set was used for model parameter learning, and the test set was used for performance validation. The Adam optimizer (learning rate 0.001) was used to train the hybrid model, with mean squared error (MSE) and mean absolute error (MAE) as loss functions to minimize the deviation between predicted and true values. The model hyperparameters were optimized using a grid search method: number of neurons in the LSTM hidden layer (128 / 256 / 512), dropout probability (0.2 / 0.3), and number of iterations (100 / 150 / 200). Each hyperparameter combination was trained three times, and the average performance was taken. After hyperparameter optimization, the model achieved an MSE ≤ 0.02 and an MAE ≤ 0.05 on the test set.

[0132] Step S146: Select measured data from different weather scenarios to verify the model's prediction effect, and adjust the physical constraint weights, model structure, and parameters based on operation and maintenance feedback data to complete model calibration;

[0133] Real-world data (no fewer than 50 sets for each of three typical scenarios—regular weather, light rain, and heavy rain—were selected to validate the model's prediction performance. For regular weather scenarios, the accuracy of congestion escalation probability was validated; for light rain scenarios, the time-series prediction error was validated; and for heavy rain scenarios, the accuracy of the congestion development time window for extreme modules was validated. Pipeline maintenance feedback data (manual handling records and robot measurement results) was collected, and the model's predictions were compared with the actual values. The deviation was calculated: if MAE > 5% or the grading accuracy < 85%, the physical constraint weight α, the number of hidden layers in the LSTM network, or the dropout probability were adjusted; if the sensitivity of a single feature was insufficient, the model was retrained using supplementary scenario data.

[0134] Step S147: Establish an edge-cloud collaborative real-time prediction process. Deploy a lightweight model at the edge to achieve rapid prediction. In the cloud, integrate cross-regional data to optimize the results through a high-precision model and output early warning of congestion escalation and time window for congestion development under extreme weather conditions.

[0135] A lightweight hybrid model, quantized and pruned, is deployed at the edge and mounted on an ARM-based edge computing box (waterproof and low-power). This enables rapid prediction within 100ms, outputting node-level congestion escalation warnings (within 1 hour) and triggering on-site alarm devices. In the cloud, a full-scale high-precision model is deployed based on a GPU distributed cluster. This model receives aggregated data from multiple edge nodes and optimizes prediction results by combining cross-regional meteorological and pipeline data, outputting regional congestion development trends (1-6 hours) and extreme weather congestion time windows. The edge-to-cloud transmission uses an incremental method, synchronizing only warning results and abnormal data, reducing bandwidth consumption and achieving data transmission latency ≤100ms.

[0136] Step S150: The blockage location reverse tracing algorithm is adopted. Based on the spatiotemporal correlation of multi-node monitoring data, the source of blockage is located by combining the pipeline network topology weight. The algorithm is optimized for special topology structures to improve the positioning accuracy.

[0137] A blockage location tracing algorithm is employed to pinpoint the blockage source based on the spatiotemporal correlation of multi-node monitoring data. First, real-time water level and flow data from each node are collected and preprocessed to construct a weighted topology map. Pipe segment weights are integrated with pipe diameter, slope, and historical blockage probability. A dynamically weighted topology map is generated by fusing the spatiotemporal correlation matrix and topology weights. Candidate paths are traversed upstream, and scores are accumulated (path score = pipe segment weight × correlation coefficient × historical frequency coefficient), selecting the top 3 candidate paths. For ring-shaped pipe networks, a new ring network flow balance check is added (prioritizing pipe segments with the smallest flow deviation). The flow distribution deviation rate is calculated for multiple branch nodes (the branch with the largest deviation is the tracing direction). Candidate paths are segmented and refined, and the slope of monitoring data changes is extracted. Confidence is calculated by combining topology weights and correlation strength (≥80% is considered valid) to determine the blockage source. Parameters are iteratively optimized through actual treatment records, achieving a pipe segment-level location error of ≤50 meters.

[0138] Step S151: Collect real-time water level, flow rate data and timestamps of each node, remove outliers by 3σ criterion, supplement missing data by linear interpolation, unify the time granularity, construct a topology graph with network nodes as vertices and pipe segments as edges, assign weights to pipe segments and nodes according to preset formulas to form a weighted topology structure.

[0139] Real-time water level, flow rate, and timestamps are collected from each node. Abrupt outliers (such as flow rate jumps exceeding twice the mean) are removed using the 3σ criterion. Missing data (missing duration ≤ 30 minutes) is supplemented through linear interpolation, with a unified time granularity of 1 minute. A topology graph is constructed with network nodes as vertices and pipe segments as edges. Pipe segments are assigned weights according to a preset formula: Weight = Standardized pipe diameter × (1 / Standardized slope) × Historical blockage probability (pipe diameter and slope are standardized to the [0,1] interval; historical probability is based on statistics from the past 3 years of operation and maintenance data). Node weights are taken as the average weights of connected pipe segments, forming a weighted topology. The topology graph node coordinate error is ≤ ±1m, pipe segment length error is ≤ ±5m, and weight calculation accuracy is ≤ 0.01, providing accurate topology support for subsequent path scoring and source tracing.

[0140] Step S152: Mark abnormal nodes based on multi-dimensional congestion classification indicators, and combine topological connectivity and data spatiotemporal correlation to delineate traceability candidate areas containing abnormal nodes and upstream and downstream specified-level pipe segments;

[0141] Based on multi-dimensional blockage classification indicators (flow attenuation rate, water level rise, etc.), threshold values ​​are set (e.g., flow attenuation rate ≥ 20%, water level rise ≥ 0.3m), and abnormal nodes exceeding these thresholds are marked. The upstream and downstream relationships of abnormal nodes are analyzed using topological connectivity analysis, combined with data spatiotemporal correlation judgment rules: the time difference of anomaly occurrence ≤ the theoretical propagation time of water flow in the pipe segment (propagation time = pipe segment length / average flow velocity), to delineate candidate tracing areas. The area includes the abnormal node and three levels of upstream and downstream pipe segments (each level extends according to the topological path), narrowing the algorithm's calculation scope (the number of pipe segments in the candidate area ≤ 30% of the entire region). During the delineation process, node IDs are matched with the topological path to ensure no key related pipe segments are missed, achieving a regional delineation accuracy of ≥ 90% and improving tracing efficiency.

[0142] Step S153: Calculate the correlation coefficient of flow / water level changes between nodes in the candidate area, and construct a spatiotemporal correlation matrix in combination with the water flow propagation time in the pipe section. The matrix elements reflect the degree of abnormal correlation between nodes.

[0143] Calculate the Pearson correlation coefficient (range [-1, 1]) of flow / water level changes between nodes within the candidate region; a coefficient ≥ 0.7 is considered a strong correlation. Derive the flow propagation time t by combining pipe segment length and average flow velocity. ij = Pipe segment length / Design flow rate (error ≤ 5%), construct the spatiotemporal correlation matrix: Matrix element M(i,j) = Correlation coefficient × (1 / |t i -t j -t ij |)(t i t j(The abnormal times for nodes i and j). Higher element values ​​indicate a stronger correlation between nodes. The matrix dimension is consistent with the number of candidate region nodes (maximum dimension ≤ 50×50). After matrix construction, normalization is performed (element values ​​are mapped to [0,1]) to facilitate subsequent fusion calculation with topological weights. The matrix calculation response time is ≤ 1 second, meeting the real-time traceability requirements.

[0144] Step S154: Integrate the spatiotemporal correlation matrix and topological weights to generate a dynamic weighted topology graph, traverse all possible paths within the candidate region along the upstream direction and accumulate scores, and filter out the candidate paths with the highest scores.

[0145] The spatiotemporal correlation matrix and the pipeline topology weights are fused element-wise. Using nodes as indices and pipe segments as connections, a dynamic weighted topology map is generated. The comprehensive weight of a pipe segment after weighting is the product of the basic topology weight and the corresponding element value of the spatiotemporal correlation matrix. The comprehensive weight is updated in real time every 1 minute to adapt to dynamic changes in monitoring data. Starting from an abnormal node, a depth-first traversal is performed on all connected paths within the candidate area along the upstream direction of the water flow. During the traversal, a cumulative scoring method is used to calculate the total score of a single path. The total score of a path is equal to the sum of the products of the comprehensive weights of each pipe segment within the path and the historical blockage frequency coefficient. The historical blockage frequency coefficient is calibrated based on the actual number of blockages in the pipe segment over the past three years. After the traversal is completed, the paths are sorted from high to low according to their total scores, and the top three candidate paths are selected as the core tracing range for the source of blockage. The calculation response time of the selection process does not exceed 1 second to ensure real-time tracing.

[0146] Step S155: Add ring network flow balance verification for ring network, calculate branch flow distribution deviation rate for multi-branch nodes, and optimize candidate path score and tracing direction;

[0147] For ring network topologies, a new ring network flow balance verification and optimization rule is introduced. Based on the node flow balance equation, the theoretical balanced flow rate of each pipe segment within the ring network is calculated. The flow deviation value is obtained by subtracting the real-time monitored flow rate of each segment from the theoretical balanced flow rate. The candidate path belonging to the segment with the smallest flow deviation is selected, and a correction coefficient of 0.1 is added to the original path score to increase its priority. For multi-branch nodes such as tees and crosses, the deviation rate between the real-time flow distribution ratio of each branch and the standard distribution ratio under normal operating conditions is calculated. The larger the deviation rate, the higher the degree of abnormality in the water flow state of that branch. The branch with the largest deviation rate is determined as the priority tracing direction, and the corresponding path score is simultaneously increased by a correction coefficient of 0.1. Through these special topology optimization rules, the candidate path scores and tracing directions are corrected, avoiding tracing misjudgments caused by ring network circulation interference and multi-branch node flow splitting, thus improving the algorithm's adaptability to complex topologies.

[0148] Step S156: Refine the candidate path into segments, extract the slope of the node monitoring data change, calculate the confidence of the suspected source by combining the degree of topological weight concentration and the spatiotemporal correlation strength, and determine the blockage source whose confidence meets the preset threshold.

[0149] The selected candidate paths were segmented into 5-meter segments, each segment labeled with a pipe segment number and corresponding monitoring node. The temporal slope of the water level and flow rate monitoring data for each segment was extracted. The slope represents the rate of increase in water level or decrease in flow rate per unit time. Segments with a slope abrupt change exceeding 0.05 m / min were identified as suspected blockage sources. Combining the topological weight concentration of the pipe segment and the spatiotemporal correlation strength of the corresponding nodes, a formula for calculating the confidence level of the blockage source was constructed. The confidence level equals the product of the candidate path score percentage, the slope abrupt change coefficient, and the data consistency coefficient, where the data consistency coefficient represents the matching degree between the monitoring data and the hydraulic simulation data. A confidence threshold of 80% was set, and segments with the highest confidence level (greater than or equal to 80%) were selected as the final blockage source. The confidence level calculation accuracy was retained to two decimal places, allowing for source location accuracy down to the single pipe segment level.

[0150] Step S157: Collect actual blockage handling records, compare the algorithm positioning results with the actual values ​​to calculate the positioning error, adjust the topology weight calculation coefficients, spatiotemporal correlation matrix parameters and special topology optimization rules, and update the pipeline topology weights synchronously.

[0151] Collect actual blockage handling records from the pipeline network, including measured data such as the true location of the blockage source, pipe segment number, handling time, and blockage severity. Algorithm-based automatic location results are compared one-to-one with the actual blockage source location. The location error is calculated using pipe segment spacing as the standard; a pipe segment-level location error of no more than 50 meters is considered acceptable. If the location error exceeds the acceptable range, key parameters such as the topology weight calculation coefficient, the spatiotemporal correlation matrix correlation threshold, and special topology optimization correction coefficients are adjusted, and the tracing algorithm is re-executed to verify the location effect. Simultaneously, the actual blockage source location and frequency data are integrated into the pipeline network's historical database, updating the historical blockage probability of each pipe segment and synchronously correcting the pipeline network's basic topology weights. The updated weights are then synchronized in real-time to the topology digital twin model. Algorithm parameters are optimized through closed-loop iterative optimization using measured data.

[0152] Step S160: Build an edge-cloud collaborative computing architecture to realize hierarchical processing of congestion level early warning and large-scale propagation prediction, and simultaneously establish an online model self-calibration mechanism to dynamically update model parameters through operation and maintenance feedback data and topology changes;

[0153] An edge-cloud collaborative computing architecture is established to achieve tiered processing of congestion level early warning and large-scale propagation prediction. A lightweight model is deployed at the edge, focusing on short-term, node-level early warning; a full-scale model is deployed in the cloud for long-term, region-level prediction. An online self-calibration mechanism is established, triggered by both operational feedback data and topology changes. The cloud calibration process is triggered when the amount of operational feedback data reaches 10% of the training sample size or when topology changes meet preset conditions (pipeline length ≥ 10% or number of nodes ≥ 5%). After calibration, optimized parameters are pushed to the edge via an encrypted channel, with hot updates at the edge to ensure dynamic matching of model parameters with the network status. The architecture response latency is ≤100ms, and the tiered accuracy is improved by ≥5% after self-calibration, with a propagation time prediction error of ≤8%, ensuring the real-time performance and accuracy of early warning and prediction.

[0154] Step S161: Select a waterproof, low-power, and interference-resistant ARM architecture edge computing box, deploy a compressed lightweight LSTM hybrid model and a congestion classification index system, ensure that the model inference latency is ≤100ms, build a cloud-based distributed computing cluster based on a GPU server, deploy a complete topology-hydrology coupled congestion propagation model, an extreme weather coupled iteration module, and a full-domain pipeline network digital twin engine, and configure an elastic scaling resource pool.

[0155] A waterproof (IP68), low-power (standby power consumption ≤10mA), and interference-resistant ARM architecture edge computing box (equipped with a Cortex-A53 processor and ≥2GB of memory) was selected to quantize (8-bit quantization) and prune (30% pruning rate) the LSTM hybrid model. A lightweight model and a congestion classification index system were deployed to ensure that the model inference latency was ≤100ms. A cloud-based distributed computing cluster (≥5 nodes) was built based on an NVIDIA A100 GPU server, deploying a complete topology-hydrology coupled congestion propagation model, an extreme weather coupled iterative module, and a full-domain pipeline network digital twin engine. An elastic scaling resource pool was configured to automatically expand according to the computing load (expansion response time ≤5 minutes) to meet the parallel computing needs of thousands of nodes in the pipeline network and the storage requirements of massive data (≥100TB). The peak performance of cloud computing was ≥100TOPS, ensuring the efficiency of large-scale simulation and long-term prediction.

[0156] Step S162: Design an edge-cloud hierarchical data transmission strategy. The edge terminal completes the preprocessing of monitoring data and hierarchical early warning of congestion locally, and only uploads the early warning results, abnormal data and key node status data. The cloud pushes model update parameters and topology change information incrementally to the edge terminal.

[0157] The design employs a hierarchical data transmission strategy between the edge and cloud. At the edge, monitoring data preprocessing (filtering, normalization) and congestion classification early warning calculations are performed locally. Only early warning results (mild / moderate / severe severity indicators), abnormal data (exceeding threshold water levels / flow rates), and status data of key nodes (main pipeline junctions, etc.) are uploaded, with an upload cycle of 5 minutes and a single data entry size ≤1KB. The cloud incrementally pushes model update parameters (weight files, threshold coefficients) and topology change information (pipeline segment additions / modifications) to the edge via the MQTT protocol, transmitting only the changed data (incremental transmission ratio ≤10%) to reduce bandwidth consumption (bandwidth requirement ≤1Mbps). Data transmission utilizes VPN encryption and data anonymization (hiding sensitive location information), with a transmission latency ≤100ms, ensuring the security, efficiency, and continuity of data interaction.

[0158] Step S163: Divide edge-cloud processing tasks. The edge outputs node-level short-term congestion warnings and links with on-site alarm devices. The cloud conducts large-scale congestion propagation simulations, long-term predictions, and extreme weather risk assessments, and outputs regional-level warning decision-making schemes. Construct an online self-calibration mechanism for the model that is triggered by both operation and maintenance feedback and topology changes. When the amount of operation and maintenance feedback data reaches 10% of the training sample size or the topology changes meet preset conditions, the calibration process is triggered.

[0159] Edge-Cloud Processing Tasks: The edge focuses on real-time congestion classification and early warning for single nodes / local pipe segments, outputting node-level congestion levels, escalation risk alerts, and short-term warnings within 1 hour, directly linking with on-site alarm devices (audible and visual alarms, SMS push notifications, response time ≤30 seconds); the cloud receives aggregated data from multiple edge nodes, combines global topology data with cross-regional meteorological data to conduct basin-wide large-scale congestion propagation simulations, 1-6 hour long-term forecasts, and global extreme weather risk assessments, outputting regional early warning decision-making solutions (such as key prevention and control areas and emergency dispatch suggestions). A dual-trigger self-calibration mechanism is constructed: It connects to the operation and maintenance management system, collects feedback data such as handling records, and triggers calibration when the accumulated training sample size reaches 10%; it monitors digital twin model topology updates in real time, automatically triggering parameter updates when pipe segment length ≥10% or node number ≥5% changes, with a calibration response time ≤1 hour.

[0160] Step S164: Extract the real congestion labels from the cloud-based operation and maintenance feedback data, compare them with the model's historical prediction results to calculate the deviation, combine them with the updated topology parameters, use transfer learning to fine-tune the model weights and recalculate the hierarchical threshold correction coefficients to generate an optimized parameter package;

[0161] After receiving the self-calibration trigger signal, the cloud extracts the actual congestion status labels (such as robot-measured congestion level and handling results) from the operation and maintenance feedback data, compares them with the model's historical prediction results, and calculates the prediction bias (grading accuracy and time-series prediction MAE). Combined with the updated topology parameters (pipe diameter, slope, node connection relationships), a transfer learning method is used to fine-tune the model weights: the feature extraction layer is frozen, and only the output layer and fully connected layer parameters are updated (learning rate 0.001, 50 iterations). Based on the multiple linear regression model, the congestion grading threshold correction coefficient is recalculated, generating an optimized parameter package (file size ≤ 50MB) containing model weight files, threshold coefficients, and topology parameters. The parameter package is validated during generation (validation error ≤ 2%) to ensure parameter validity and provide a reliable basis for edge-end model updates.

[0162] Step S165: Push the parameter package to the edge terminal through an encrypted channel. The edge terminal hot-uploads the new model and records the update log, and simultaneously builds a fault tolerance mechanism for local data caching on the edge terminal and master-slave backup in the cloud.

[0163] The optimized parameter package is pushed to the corresponding edge nodes via a VPN encrypted channel (AES-256 encryption). Upon receiving the package, the edge node automatically verifies data integrity (MD5 checksum). If verification is successful, the old parameters are replaced. A hot update method is used to load the new model without interrupting real-time alert services (update time ≤ 30 seconds). Parameter update logs are recorded synchronously, including fields such as update time, trigger reason, parameter change amount, and update personnel. Data traceability and version rollback are supported (retaining the last 10 update records). A fault tolerance mechanism is established: the edge node is configured with a local data caching module of ≥16GB. In the event of a network interruption, 24-hour monitoring data and alert results are cached, and automatically retransmitted after network recovery. The cloud adopts a 1-master 2-slave backup architecture, synchronizing edge node status and model parameters every minute. In the event of a core node failure, the slave nodes seamlessly take over (switchover time ≤ 5 seconds), ensuring service continuity.

[0164] Step S166: Conduct self-calibration effect verification, requiring that the classification accuracy improves by ≥5% and the propagation time prediction error is ≤8% after calibration. The calibration effect is comprehensively evaluated every quarter and stable parameters are solidified and incorporated into the model version management system.

[0165] Conduct self-calibration effectiveness verification by selecting independent measured blockage event data (no fewer than 20 sets) and calculating the classification accuracy and propagation time prediction error of the model after calibration. The classification accuracy should improve by ≥5% (based on the pre-calibration baseline), and the propagation time prediction error should be ≤8%. If these targets are not met, adjust the number of transfer learning fine-tuning layers (add 1-2 layers), the learning rate, or the topology weight calculation coefficients, and supplement with maintenance feedback data for recalibration. Conduct a comprehensive evaluation of the self-calibration effectiveness quarterly, statistically analyzing indicators such as calibration frequency, prediction accuracy improvement, and maintenance cost reduction (e.g., a ≥20% reduction in fault handling time), and generate an evaluation report. Solidify stable and effective calibration parameters (those meeting the standards in three consecutive evaluations) and incorporate them into the model version management system (version numbers are named according to "year.quarter.serial number") to provide standardized parameter references for adaptation to new pipeline network scenarios.

[0166] Step S170: Establish a multi-source data collaborative verification system, connect with relevant cross-departmental data to conduct cross-verification, and introduce special data to optimize the congestion prediction time series for extreme rainfall scenarios.

[0167] A multi-source data collaborative verification system was established, integrating cross-departmental data from meteorology, water resources, urban management, and emergency response: the meteorological department provides real-time rainfall and radar echo intensity data (updated every 1 minute); the water resources department provides pipeline operation data and topology update information; the urban management department provides inspection records and disposal work orders; and the emergency response department provides data on waterlogging points. Data integration is achieved using RESTful API / MQTT protocols, with a unified data format of JSON. Security is ensured through VPN encryption and data anonymization. Cross-validation is conducted: at the data level, the correlation coefficient between data from different departments is calculated (≥0.7 is considered acceptable); at the model level, the deviation between training results from single and multi-source data is compared (≤8%). For extreme rainfall scenarios, specialized data such as short-duration heavy rainfall radar echoes (≥45dBZ) and historical extreme event pipeline response data are introduced, assigned 2-3 times the feature weight, and the iteration step size is optimized to 1 minute to improve the accuracy of congestion prediction time series (error ≤10 minutes). Retrospective verification is conducted quarterly, adjusting data integration priorities and verification dimensions to ensure the system's effectiveness.

[0168] Based on the same inventive concept, please refer to Figure 2 This paper shows a schematic block diagram of a road drainage system blockage monitoring system 100 provided in this application embodiment for performing the above-described multi-source sensor fusion road drainage system blockage monitoring method. The multi-source sensor fusion road drainage system blockage monitoring system 100 may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0169] In this embodiment, both the machine-readable storage medium 120 and the processor 130 are located within the multi-source sensor fusion road drainage system congestion monitoring system 100 and are separately configured. However, it should be understood that the machine-readable storage medium 120 may also be independent of the multi-source sensor fusion road drainage system congestion monitoring system 100 and may be accessed by the processor 130 via a bus interface. Alternatively, the machine-readable storage medium 120 may be integrated into the processor 130 and may communicate with external systems via the communication unit 110.

[0170] The processor 130 is the control center of the multi-source sensor fusion road drainage system congestion monitoring system 100. It connects to various parts of the system via various interfaces and lines. By running or executing software programs and / or modules stored in the machine-readable storage medium 120, and by calling data stored in the machine-readable storage medium 120, it performs various functions and processes data of the multi-source sensor fusion road drainage system congestion monitoring system 100, thereby providing overall monitoring of the system. Optionally, the processor 130 may include one or more processing cores; for example, the processor 130 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor. The machine-readable storage medium 120 is used to store machine-executable instructions for executing the scheme of this application, and the processor 130 is used to execute the machine-executable instructions stored in the machine-readable storage medium 120 to implement the multi-source sensor fusion road drainage system blockage monitoring method provided in the aforementioned method embodiment.

[0171] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for monitoring blockages in road drainage systems using multi-source sensor fusion, characterized in that: Includes the following steps: A dynamic topology sensing network for the pipeline network is constructed. Multimodal integrated sensing and monitoring units are deployed at each node of the pipeline network. A digital twin model of the pipeline network topology is constructed and updated in real time through lidar scanning. Data on the water flow status and structural changes in the pipeline network are collected synchronously. At the same time, appropriate sensing correction and diversion flow monitoring schemes are adopted for special nodes of the pipeline network to achieve accurate collection of monitoring data for special nodes. A topology-hydrology coupled blockage propagation model is constructed based on a topology digital twin model. By integrating real-time pipeline network operation data and topology parameters, the migration patterns of water flow and blockages are clarified through numerical simulation methods, and the diffusion characteristics of blockages under different topologies are quantified. Establish a multi-dimensional blockage classification index system, integrate key hydraulic characteristic indicators, use the objective weighting method to determine the index weights and classify blockage levels, and dynamically correct the classification thresholds in combination with pipeline network structure parameters. A hybrid prediction model combining data-driven and physical constraints is constructed, integrating historical data time-series characteristics with hydraulic physics principles, and combining multi-source environmental and pipeline load data to achieve prediction of blockage escalation under normal scenarios and time-series prediction of blockage development under extreme weather conditions. A blockage location reverse tracing algorithm is adopted, which relies on the spatiotemporal correlation of multi-node monitoring data and combines the pipeline topology weight to locate the source of blockage. The algorithm is optimized for special topology structures to improve the positioning accuracy. Build an edge-cloud collaborative computing architecture to achieve hierarchical processing of congestion level early warning and large-scale propagation prediction, and simultaneously establish an online model self-calibration mechanism to dynamically update model parameters through operation and maintenance feedback data and topology changes; Establish a multi-source data collaborative verification system, connect with relevant data from different departments to conduct cross-verification, and introduce special data to optimize the congestion prediction time series for extreme rainfall scenarios.

2. The method for monitoring blockages in a road drainage system using multi-source sensor fusion as described in claim 1, characterized in that: The construction of a dynamic topology sensing network for the pipeline network involves deploying multimodal integrated sensing and monitoring units at each node of the pipeline network, and constructing and updating a digital twin model of the pipeline network topology in real time through lidar scanning, including: A comprehensive survey of basic pipeline information was conducted through drawing retrieval and on-site surveys to clarify the pipeline topology, critical path nodes, and the distribution of special nodes such as inspection wells, drop wells, and interception wells. Physical parameters such as pipe diameter, pipe material, slope, and design flow rate were collected to establish a basic pipeline information database. An ultrasonic water level sensor, a radar flow sensor, and a vibration sensor are selected and integrated into an integrated multimodal monitoring module, along with a low-power data acquisition unit and a LoRa / 5G / NB-IoT wireless communication module. Based on the complexity of the pipeline network topology and the characteristics of the flow distribution, the multimodal monitoring module is fixed on the well wall or inner wall of the pipeline at path nodes and special nodes. A waterproof and sealed installation process is adopted to avoid the area directly impacted by the water flow. The sensor measurement accuracy is calibrated simultaneously to meet the engineering monitoring standards. Deploy lidar equipment suitable for the internal environment of the pipeline network. The equipment can be mobilely mounted on the pipeline robot or fixedly installed on the cross-section. Set the point cloud density, scanning angle, and ranging range parameters to ensure coverage of the inner wall of the pipeline, the internal structure of the nodes, and the connection parts. The lidar full-domain scanning is carried out in segments according to the pipeline topology path, and the GPS / inertial navigation coordinates of the scanning position are recorded simultaneously. Three-dimensional point cloud data of pipeline centerline, pipe diameter change, node structure morphology and pipeline connection relationship are collected, and special nodes are scanned repeatedly from multiple angles. Gaussian filtering and statistical filtering algorithms are used to remove noise points in point cloud data. Data from different sections are stitched together using point cloud registration technology. The RANSAC algorithm is used to extract the pipeline axis and cross-sectional profile. Cluster analysis is used to identify node types and connection relationships to construct a pipeline network topology skeleton model. Based on the aforementioned topological skeleton model and physical parameters, a three-dimensional geometric model of the pipeline network is built, mapping the physical properties of pipe diameter, pipe material, and slope. The location coordinates and data interface of the multimodal monitoring module are integrated to realize the real-time association between the model and sensor data. A visualization platform supporting three-dimensional display of topological structure, node query, and data traceability is developed. Set a regular update cycle and connect to the network operation and maintenance management system. When receiving information on topology modification and maintenance construction, automatically trigger LiDAR directional scanning, identify the topology change area by comparing the old and new point cloud data, and incrementally update the geometric structure and physical parameters of the digital twin model. The edge computing unit receives water flow status data from the multimodal monitoring module and structural monitoring data from the lidar in real time, cross-validating the model topology accuracy and the validity of the sensor data. When abnormal data or suspected topology changes are detected, a secondary scan and model calibration process is triggered to ensure that the digital twin model is dynamically consistent with the actual pipeline network status.

3. The method for monitoring blockages in a road drainage system using multi-source sensor fusion as described in claim 1, characterized in that: The aforementioned topology-hydrological coupled blockage propagation model, constructed based on a topological digital twin model, integrates real-time pipeline network operation data and topological parameters. Numerical simulation methods are used to clarify the migration patterns of water flow and blockages, quantifying the diffusion characteristics of blockages under different topologies, including: Based on the topological digital twin model, geometric and topological feature data of the entire pipeline network segments and nodes are extracted. The feature data includes pipe diameter, pipe length, laying slope, pipe material roughness, node connection relationship, topology type and special node structural parameters. After standardizing the feature data, a pipeline network topology graph structure with nodes as vertices and pipe segments as edges is constructed to complete the geometric and topological basic modeling of the coupled model. The system integrates real-time water level, flow rate, flow velocity data and rainfall monitoring data collected by the multimodal monitoring unit. It then performs spatiotemporal alignment of the multi-source monitoring data with the corresponding nodes and pipe segments in the topological digital twin model, removes abnormal fluctuation data, and completes data filtering and normalization to form a real-time input dataset for the topology-hydrological coupling model. A hydraulic control equation system adapted to complex pipe network topology is established, with the one-dimensional Saint-Venant equation as the core of water flow control. Hydraulic constraints are constructed in combination with the flow balance conditions of pipe network nodes. Water flow distribution rules and boundary conditions are set for different topological structures such as dendritic, ring, and mixed structures. Hydraulic connection relationships are defined separately for special nodes such as drop wells and intercepting wells, forming a control equation set that integrates topological constraints and hydrodynamic constraints. The finite volume method is used to spatially discretize the control equations. The calculation units are divided according to the topological relationship between the pipe network segments and nodes. The numerical flux calculation method of the unit interface is determined. The discretization accuracy is adaptively adjusted according to the pipe material, pipe diameter and slope and a reasonable time step is set to realize the numerical iterative solution of the water flow motion state. A numerical model for blockage migration and deposition was constructed. Critical conditions for blockage initiation, migration, deposition, and accumulation were set in combination with water flow velocity and water level rise. The changes in blockage particle size, deposition morphology, and blockage cross-sectional area were correlated with water flow hydraulic parameters. The blockage evolution rules were embedded into the water flow numerical solution process to achieve synchronous coupled simulation of water flow motion and blockage migration. We conducted blockage propagation simulation calculations for three types of pipe network topologies: dendritic, ring-shaped, and mixed. We set initial blockage conditions at different locations and to varying degrees in the model, iteratively calculated the changes in water flow state and blockage diffusion process within the pipe network, and recorded the changes in water level, flow rate, and blockage degree of each node and pipe segment over time after the blockage occurred. Establish a quantitative index system for blockage diffusion characteristics, extract characteristic parameters such as propagation time, number of affected pipe segments, backlog range, flow attenuation magnitude, and blockage level evolution rate through numerical simulation results, and compare simulation data under different topologies to form a quantitative database of blockage diffusion patterns for various topology pipe networks; The coupled model was calibrated and verified using field-measured blockage event data. The pipe roughness, blockage transport coefficient, and node hydraulic loss coefficient were adjusted to ensure that the error between the model simulation results and the actual blockage diffusion characteristics met the engineering monitoring requirements, thus completing the model calibration and optimization. Establish a real-time linkage interface between the model and the topology digital twin. When the digital twin model completes the topology update due to pipeline renovation and structural changes, the coupled model automatically reads the updated geometric and topological data, reconstructs the computing units and boundary conditions, and adapts to the new pipeline structure. The simulated spatiotemporal data and quantitative characteristics of the blockage propagation are transmitted back to the topological digital twin model in real time, and the propagation path, development trend and impact range of the blockage are visualized in the three-dimensional twin scene.

4. The method for monitoring blockages in a road drainage system using multi-source sensor fusion as described in claim 1, characterized in that: The establishment of a multi-dimensional blockage classification index system integrates key hydraulic characteristic indicators, uses an objective weighting method to determine index weights and classify blockage levels, and dynamically adjusts the classification thresholds based on pipeline network structure parameters, including: Screen relevant characteristic indicators of hydraulic response to pipe network blockage, and determine the indicator set as flow rate decay rate, water level rise, proportion of blockage cross-sectional area, and water velocity decrease rate. Clarify the definition, monitoring method and collection frequency of each indicator. Measured data were collected under different topologies, pipe diameters, and operating conditions, including various blockage scenarios and normal operation. After outliers were removed by the 3σ criterion and missing data were supplemented by linear interpolation, the data were eliminated by min-max standardization and mapped to the [0,1] interval to form a standardized index dataset. The entropy weight method is used as an objective weighting tool. First, the proportion of the i-th indicator and the j-th sample is calculated. Then, the entropy value of the indicator is calculated based on the proportion. The entropy value reflects the degree of disorder of the indicator information. Then, the indicator difference coefficient is calculated. The weight of each indicator is determined according to the proportion of the difference coefficient. The rationality of the weight is verified by the consistency test. Combining industry operation and maintenance standards, historical blockage handling cases, and hydraulic simulation results, the K-means clustering algorithm was used to perform cluster analysis on standardized indicator data to determine the indicator values ​​corresponding to the cluster centers. Blockage levels were divided into three categories: light, moderate, and severe. Light blockage was defined as a single indicator slightly deviating from the normal range without affecting overall drainage efficiency; moderate blockage was defined as two or more indicators exceeding the standard and a significant decrease in drainage capacity; and severe blockage was defined as all indicators significantly exceeding the standard and a risk of water accumulation. At the same time, the threshold ranges of the indicators corresponding to each level were quantified. Pipe diameter, laying slope, and pipe roughness were selected as correction parameters. Hydraulic test data of the pipeline network under different combinations of structural parameters were collected to construct a multiple linear regression correction model. Pipe diameter D and slope i were used as independent variables, and the grading threshold T of each index was used as dependent variables. The correction formula T'=T×(aD+bi+c) was obtained by fitting with the least squares method, where a, b, and c are fitting coefficients. Roughness correction coefficients were set for different pipe types, such as concrete pipe, plastic pipe, and cast iron pipe. By inputting pipeline network structure parameters into the correction model to obtain personalized grading thresholds, selecting typical pipeline network topologies to conduct field tests, and combining the actual blockage status obtained by the pipeline inspection robot, the grading results of the verification index system are compared and the accuracy is calculated. The grading deviation or index sensitivity issues are iteratively adjusted and optimized to form a stable and reliable multi-dimensional blockage grading index system.

5. The method for monitoring blockages in a road drainage system using multi-source sensor fusion as described in claim 1, characterized in that: The aforementioned construction of a data-driven and physically constrained hybrid prediction model integrates historical data time-series characteristics with hydraulic physics principles, and incorporates multi-source environmental and pipeline load data to achieve prediction of blockage escalation under normal scenarios and time-series prediction of blockage development under extreme weather conditions, including: Collect multi-source basic data on historical blockage events, pipeline operation, environment, pipeline load, and extreme weather. Use outlier detection algorithms to remove invalid data, supplement missing data through interpolation, and construct a standardized input dataset after spatiotemporal alignment and normalization. Based on the Saint-Venant equation, a physical constraint system is constructed by combining the flow balance equation and energy equation of the pipeline node. The physical constraints are transformed into mathematical hard constraints, which limit the prediction results to conform to the principles of hydraulic dynamics. Extract time-series and correlation features from standardized data, strengthen the weights of key extreme weather features through an attention mechanism, and complete the model feature engineering. A data-driven model framework is built using a long short-term memory network, an overfitting suppression mechanism is introduced, physical constraints are embedded into the model training process, a hybrid prediction fusion architecture is constructed, and a coupled iterative prediction module for rainfall, flow rate and blockage under extreme weather conditions is built based on the rainstorm intensity formula and the hydraulic carrying capacity of the pipeline network. The standardized dataset is divided into training and testing sets. An optimizer and a preset loss function are used to train a hybrid model. Hyperparameter optimization is used to improve prediction accuracy. Measured data from different weather scenarios were selected to verify the model's prediction performance. Physical constraint weights, model structure, and parameters were adjusted in conjunction with operation and maintenance feedback data to complete model calibration. Establish an edge-cloud collaborative real-time prediction process. Deploy lightweight models at the edge to achieve rapid prediction, and integrate cross-regional data optimization results through high-precision models in the cloud to output early warning of congestion escalation and time windows for congestion development under extreme weather conditions.

6. The method for monitoring blockages in a road drainage system using multi-source sensor fusion as described in claim 5, characterized in that: The aforementioned data-driven model framework employs a long short-term memory network, introduces an overfitting suppression mechanism, embeds physical constraints into the model training process, constructs a hybrid prediction fusion architecture, and simultaneously builds a coupled iterative prediction module for rainfall, flow rate, and blockage under extreme weather conditions based on the rainstorm intensity formula and the hydraulic carrying capacity of the pipe network. This includes: Configure the input layer, 3-5 hidden layers, and output layer of the LSTM network according to the input feature dimension, set the corresponding activation function, and use the Xavier method to initialize the network weights; Insert dropout layers between LSTM hidden layers and add an L2 regularization term to the loss function to build a model overfitting suppression mechanism; Using the discrete form of the Saint-Venant equation and the nodal flow balance equation as hydraulic hard constraints, a total constraint loss function is constructed that includes prediction loss and physical constraint error. A parallel input mode is adopted to simultaneously input temporal features, correlation features and physical constraint parameters into the LSTM model, and a hybrid prediction fusion architecture is built. The prediction loss and physical constraint loss are simultaneously optimized through backpropagation. Based on the formula for the intensity of rainstorms in the target area, the rainfall calculation coefficient is calibrated, and the hydraulic ultimate bearing capacity and the critical velocity threshold of the blockage are determined by combining the pipe network structural parameters. Input the real-time rainfall intensity and the predicted peak rainfall, construct a coupled iterative logic of rainfall-flow-blockage, and iteratively calculate the node flow and blockage accumulation status until the iteration converges; Establish a collaborative mechanism between the LSTM hybrid model for regular scenarios and the coupled iterative module for extreme weather, switch the prediction mode based on meteorological warnings and realize data interaction, and output the time series prediction results of traffic congestion development; The Adam optimizer was used to train the model, and the model training optimization was completed by combining early stopping mechanism and dynamic physical constraint weight adjustment.

7. The method for monitoring blockages in a road drainage system using multi-source sensor fusion as described in claim 6, characterized in that: The method employs a parallel input mode to simultaneously input temporal features, correlation features, and physical constraint parameters into the LSTM model, building a hybrid prediction fusion architecture. It then simultaneously optimizes the prediction loss and physical constraint loss through backpropagation, including: Feature dimension adaptation preprocessing is performed on temporal features, correlation features, and physical constraint parameters. The correlation features and physical constraint parameters are extended to the sequence dimension that matches the temporal features to ensure that the time step dimensions of the three are consistent to meet the requirements of parallel input. Construct a parallel LSTM network submodule with three independent input branches. The temporal feature branch extracts temporal dependency features through 1-2 layers of LSTM sublayers, and the correlation feature branch and physical constraint parameter branch extract corresponding features through fully connected layers, with each branch output feature dimension adapted. The output features of the three branches are concatenated and fused, and then input into the main LSTM layer. After feature temporal pattern capture and global average pooling, the output layer outputs the probability of escalation of congestion level in normal scenarios. Define a dual loss function that includes prediction loss and physical constraint loss. The prediction loss is calculated by using mean squared error to calculate the deviation between the model output and the true label. The physical constraint loss is calculated by calculating the deviation between the model prediction value and the hydraulic equation derivation value. Construct a total loss function to balance the contributions of the two types of losses. The Adam optimizer is used to backpropagate the total loss gradient to each input branch and the main LSTM layer, decompose the gradient components and update the corresponding branch weights, and apply L2 regularization to the physical constraint parameter branch. Implement phased training. In the pre-training phase, fix the weights of the physical constraint branch and train the other branches. In the fusion training phase, release the weights of the branch and dynamically adjust the physical constraint weight coefficients to ensure that the total loss and physical constraint error are controllable. The total loss gradient is pruned, and batch size and early stopping mechanism are set to control training stability. The improvement effect on prediction error, physical constraint error and model prediction accuracy is verified through test set to ensure that the performance of the fusion architecture meets the standards.

8. The method for monitoring blockages in a road drainage system using multi-source sensor fusion as described in claim 6, characterized in that: The mechanism for establishing a collaborative mechanism between a conventional scenario LSTM hybrid model and an extreme weather coupled iterative module, switching prediction modes based on meteorological warnings and achieving data interaction, outputs the time-series prediction results of traffic congestion development, including: Construct a meteorological early warning triggering and judgment rule system, connect to meteorological data interfaces to obtain rainfall, radar echo intensity and rainstorm warning level data, set extreme weather trigger thresholds and configure a buffer mechanism for continuous periodic judgment, and switch between normal and extreme weather prediction modes accordingly; Design a standardized bidirectional data interaction interface across modules, define a transmission data structure that includes node ID, timestamp, data type, value and confidence level, and realize bidirectional data calls between the conventional LSTM hybrid model and the extreme weather coupled iterative module; Define dual-mode operation logic. In normal mode, the LSTM hybrid model outputs the congestion escalation trend and time independently at a fixed period. After the extreme mode is triggered, the current congestion state output by the LSTM model is used as the initial condition, the cycle of the coupled iteration module is switched and short-cycle congestion level changes and warning time windows are output. A memory-level data sharing cache is used as the interaction carrier to complete the read and write synchronization of data between the two modules on a periodic basis, and a timestamp verification mechanism is combined to ensure that the data status of the two modules is consistent. We construct a logic for fusing and outputting differentiated prediction results. In the normal mode, we directly output the model prediction results, while in the extreme mode, we use the time series curve of the coupled iterative module as the core and superimpose the probability of the LSTM model as the confidence weight to distinguish the output emphasis of the two modes. Add a mode switching fault tolerance and rollback mechanism to achieve a smooth rollback from extreme mode to normal mode based on changes in meteorological data. At the same time, it provides emergency prediction and triggers alarms for scenarios where iterative non-convergence occurs. Historical rainstorm monitoring data were selected to verify the synergistic effect of the two modes. The trigger threshold, iteration step size and data synchronization frequency were adjusted according to the preset performance indicators.

9. The method for monitoring blockages in a road drainage system using multi-source sensor fusion as described in claim 1, characterized in that: The aforementioned edge-cloud collaborative computing architecture enables tiered processing of congestion-level early warning and large-scale propagation prediction. Simultaneously, an online model self-calibration mechanism is established to dynamically update model parameters based on operational feedback data and topology changes, including: We selected a waterproof, low-power, and interference-resistant ARM architecture edge computing box, deployed a compressed lightweight LSTM hybrid model and a congestion classification index system to ensure that the model inference latency is ≤100ms, built a cloud-based distributed computing cluster based on GPU servers, deployed a complete topology-hydrology coupled congestion propagation model, an extreme weather coupled iteration module and a full-domain pipeline network digital twin engine, and configured an elastic scaling resource pool. The design of the edge-cloud hierarchical data transmission strategy is as follows: the edge end completes the monitoring data preprocessing and congestion hierarchical early warning locally, and only uploads the early warning results, abnormal data and key node status data. The cloud incrementally pushes model update parameters and topology change information to the edge end. The edge-cloud processing tasks are divided into edge and cloud processing tasks. The edge outputs node-level short-term congestion warnings and links with on-site alarm devices. The cloud conducts large-scale congestion propagation simulations, long-term predictions, and extreme weather risk assessments, and outputs regional warning decision-making schemes. A model online self-calibration mechanism is constructed that is triggered by both operation and maintenance feedback and topology changes. When the amount of operation and maintenance feedback data reaches 10% of the training sample size or the topology changes meet preset conditions, the calibration process is triggered. The actual congestion labels extracted from the cloud-based operation and maintenance feedback data are compared with the historical prediction results of the model to calculate the deviation. Combined with the updated topology parameters, transfer learning is used to fine-tune the model weights and recalculate the hierarchical threshold correction coefficient to generate an optimized parameter package. The parameter package is pushed to the edge through an encrypted channel. The edge hot-uploads the new model and records the update log, and a fault tolerance mechanism is built simultaneously for local data caching at the edge and master-slave backup in the cloud. Conduct self-calibration effect verification, requiring that the accuracy of classification improve by ≥5% and the propagation time prediction error be ≤8% after calibration. The calibration effect is comprehensively evaluated every quarter and stable parameters are solidified and incorporated into the model version management system.

10. A multi-source sensor fusion road drainage system blockage monitoring system, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the multi-source sensor fusion road drainage system blockage monitoring method according to any one of claims 1 to 9 by executing the machine-executable instructions.

Images