A data processing method for urban drainage network risk prediction

Abstract

The application is suitable for the technical field of urban infrastructure risk management, and provides a data processing method for urban drainage pipe network risk prediction, comprising the following steps: step S5, multi-disaster coupling risk assessment: based on the risk probability of each node obtained in step S4, combining a pre-constructed waterlogging inundation model, calculating a multi-disaster coupling risk index, and generating a dynamic distribution map of urban drainage pipe network risk; step S6, intelligent scheduling of emergency knowledge and resources: based on the urban pipe network emergency knowledge graph, using a path optimization algorithm, generating an optimal scheduling path from an emergency resource point to a risk point, and outputting an emergency decision scheme. The data processing method for urban drainage pipe network risk prediction provided by the application, in the aspect of data processing, in step S1, the pre-constructed unified data dictionary is used for semantic alignment and standardized processing of multi-source heterogeneous data, so that the semantic consistency and high quality of the data are ensured.

Description

Technical Field

[0001] This invention belongs to the field of urban infrastructure risk management technology, and in particular relates to a data processing method for predicting risks in urban drainage pipe networks. Background Technology

[0002] As the core of urban lifeline engineering, the safe operation of urban drainage networks is crucial to public safety and urban resilience. Currently, my country's drainage system faces multiple challenges, including frequent extreme weather events, aging facilities, data heterogeneity, and delayed emergency response. Existing technologies have significant shortcomings in the following aspects:

[0003] First, there is the difficulty in integrating multi-source data. Pipeline networks involve various types of data, such as geographic information, IoT sensing, meteorology, and historical disaster data. These data are scattered, have significantly different formats, and lack semantic consistency. Traditional manual input or simple splicing cannot bridge the semantic gap, resulting in data fragmentation and making it difficult to form high-quality datasets to support risk analysis.

[0004] Second, digital twin modeling lags behind. Traditional pipeline network models are mostly static structures, relying on periodic manual surveying, and cannot be dynamically linked with real-time IoT sensing data. Modeling of ultra-large-scale urban pipeline networks is inefficient and has a long update cycle, making it difficult to support real-time risk monitoring and dynamic simulation.

[0005] Third, risk feature extraction is one-sided. Traditional methods are limited to isolated analysis of single-point monitoring data and do not consider the spatial dependencies between nodes in the pipeline network topology. For example, they cannot effectively trace whether the abnormal liquid level was caused by upstream water inflow or downstream drainage obstruction, resulting in a lack of global perspective in risk identification.

[0006] In summary, there is an urgent need for a comprehensive technical solution that can integrate multi-source data, achieve dynamic model updates, fuse topological features, quantify the coupling effects of multiple disasters, and support intelligent emergency dispatch. Summary of the Invention

[0007] The purpose of this invention is to provide a data processing method for predicting risks in urban drainage pipe networks, thereby addressing the aforementioned problems.

[0008] This invention is implemented as follows: a data processing method for risk prediction of urban drainage pipe networks, comprising the following steps:

[0009] Step S1: Multi-source heterogeneous data fusion and standardization processing: Collect basic geographic information data of drainage pipe network, IoT sensing data, meteorological data and historical disaster data; based on the pre-built unified data dictionary, perform semantic alignment and standardization processing on the collected multi-source data to generate a standardized fusion dataset;

[0010] Step S2: Holographic digital twin modeling based on spatiotemporal grid: The urban drainage network is divided and encoded into a three-dimensional spatial grid to construct a holographic digital twin model of the network that supports dynamic updates;

[0011] Step S3, Multimodal Risk Feature Extraction: Extract pipeline network operation status features, environmental features, and historical disaster features from the fused dataset; construct a graph representation of the pipeline network topology, and generate a risk feature vector for each pipeline network node through a graph neural network;

[0012] Step S4: Risk probability prediction based on dynamic Bayesian network: Construct a dynamic Bayesian network model, which is defined by the initial state distribution, state transition probability and observation likelihood function; take the risk feature vector as the observation input, and recursively calculate the posterior probability of each node in each risk state through Bayesian filtering, which is used as the risk probability output of the node.

[0013] Step S5, Multi-hazard Coupling Risk Assessment: Based on the risk probabilities of each node obtained in Step S4, and combined with the pre-constructed urban flooding model, calculate the multi-hazard coupling risk index and generate a dynamic risk distribution map of the urban drainage network.

[0014] Step S6, Emergency Knowledge and Resource Intelligent Scheduling: Based on the urban pipeline network emergency knowledge graph, a path optimization algorithm is used to generate the optimal scheduling path from emergency resource points to risk points and output emergency decision-making solutions.

[0015] In a further technical solution, step S1 further includes:

[0016] By introducing R2RML mapping rules, unstructured historical disaster data is converted into resource description framework triple format, which is then integrated with structured data to construct the foundation layer of the urban drainage network knowledge graph.

[0017] In a further technical solution, step S2 further includes:

[0018] A three-dimensional spatial grid partitioning technique is used to encode the pipeline network in three dimensions, establishing a dynamic mapping relationship between grid units and pipeline components. When IoT sensing data is updated, the corresponding component is located according to the grid coding index, triggering a local update of the digital twin model.

[0019] In a further technical solution, step S3 further includes:

[0020] The drainage pipe network topology is modeled as a graph structure, where nodes represent pipe network nodes and edges represent pipe segment connections. A graph convolutional network is used to aggregate and update node features, and the node features output by the network are used as risk feature vectors.

[0021] In a further technical solution, step S4 further includes:

[0022] The initial state distribution is the probability that a node is in each risk state at the initial moment;

[0023] The state transition probability is the probability that a node will transition from a risk state at the previous moment to a risk state at the current moment, and it is positively correlated with the service life of the pipeline and the soil corrosivity.

[0024] The observation likelihood function is the probability density of a node observing the current risk feature vector when it is in a certain risk state.

[0025] Input the risk feature vector at the current moment, and calculate the posterior probability at the current moment recursively using Bayesian filtering based on the current observation likelihood value, the posterior probability at the previous moment, and the state transition probability.

[0026] A further technical solution is that the method for determining the state transition probability further includes:

[0027] Historical accident data were collected, and the state transition probability matrix was fitted using maximum likelihood estimation. For sparse data regions, a physical model based on pipeline aging patterns was pre-defined.

[0028] A further technical solution is that the method for determining the observation likelihood function further includes:

[0029] Historical monitoring data of nodes under each risk state are collected, and the probability distribution of the observed data under each risk state is fitted using the kernel density estimation method to obtain the observation likelihood function.

[0030] A further technical solution is that the calculation of the multi-hazard coupling risk index in step S5 further includes:

[0031] Calculate the coupling risk index of node i at the current time, which is obtained by weighted summation of three terms:

[0032] The first term is the first weighting coefficient multiplied by the probability of waterlogging risk at node i;

[0033] The second term is the second weighting coefficient multiplied by the probability of gas leakage risk in the region where node i is located;

[0034] The third term is the third weighting coefficient multiplied by the product of the probability of waterlogging risk and the probability of gas leakage risk;

[0035] The sum of the first, second, and third weighting coefficients is 1, and the third weighting coefficient is used to characterize the risk amplification effect when waterlogging and leakage occur simultaneously.

[0036] In a further technical solution, step S6 further includes:

[0037] The objective function is to minimize the emergency response time, and the constraint is the road traffic status. The road traffic status is dynamically updated based on the real-time water depth. The emergency resource scheduling path is generated using Dijkstra's algorithm or differential evolution algorithm.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] This invention provides a data processing method for risk prediction of urban drainage pipe networks. In terms of data processing, in step S1, this application performs semantic alignment and standardization on multi-source heterogeneous data through a pre-constructed unified data dictionary, ensuring semantic consistency and high quality of the data. This enables subsequent risk prediction models to be trained and inferred based on more reliable and comprehensive data, avoiding prediction bias caused by data quality issues.

[0040] This invention provides a data processing method for risk prediction of urban drainage pipe networks. In terms of digital twin modeling, in step S2, this application constructs a dynamically updated holographic digital twin model by performing three-dimensional spatial grid subdivision and encoding of the urban drainage pipe network. When the liquid level data of a certain inspection well is updated, the system can quickly locate and update the corresponding component in the digital twin model according to its grid encoding, achieving real-time synchronization between the model and the physical pipe network. This dynamic updating capability significantly improves the timeliness of risk monitoring, enabling timely reflection of the latest operating status of the pipe network.

[0041] The present invention provides a data processing method for risk prediction of urban drainage pipe network. In terms of risk feature extraction, in step S3, this application not only extracts multimodal features such as pipe network operation status, environment and historical disasters from the fused dataset, but more importantly, it constructs a graph representation of the pipe network topology and generates a risk feature vector for each pipe network node through a graph neural network.

[0042] This invention provides a data processing method for risk prediction in urban drainage pipe networks. Regarding risk probability prediction, in step S4, a dynamic Bayesian network model is constructed, and Bayesian filtering is used to recursively calculate the posterior probability of each node in each risk state, achieving dynamic quantification of risk state evolution over time. Further, in step S5, based on node risk probabilities combined with an urban flooding model, a multi-hazard coupled risk index is calculated, and a dynamic risk distribution map of the urban drainage pipe network is generated. By considering the interaction between urban flooding and gas leaks, the system can identify areas where urban flooding may trigger gas leaks and quantify the amplification effect of this coupled risk, providing a more comprehensive and accurate risk situation awareness than single-hazard assessment.

[0043] This invention provides a data processing method for risk prediction in urban drainage pipe networks. In terms of emergency decision support, in step S6, based on an urban pipe network emergency knowledge graph, a path optimization algorithm is used to generate the optimal scheduling path from emergency resource points to risk points, and an emergency decision plan is output. This intelligent scheduling capability significantly improves the efficiency and accuracy of emergency response, effectively reducing disaster losses. Attached Figure Description

[0044] Figure 1 A schematic diagram of the data processing method for risk prediction of the municipal drainage pipe network;

[0045] Figure 2 A schematic diagram illustrating the steps involved in data processing for risk prediction of the city's drainage pipe network. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0047] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0048] like Figures 1-2 As shown, a data processing method for risk prediction of urban drainage pipe networks provided in an embodiment of the present invention includes the following steps:

[0049] In step S1, multi-source heterogeneous data fusion and standardization are performed. To comprehensively understand the operational status and potential risks of the urban drainage network, it is necessary to collect basic geographic information data, IoT sensing data, meteorological data, and historical disaster data. For example, basic geographic information data can include static attributes such as the layout, material, and burial depth of the network, which can be obtained through a Geographic Information System (GIS) database or manual input. IoT sensing data covers real-time dynamics within the network, such as data collected by devices like level sensors, flow meters, and water quality monitors, typically transmitted to a data center via wireless communication networks. Meteorological data, such as rainfall, temperature, and wind speed, can be obtained through weather stations or meteorological service interfaces. Historical disaster data, such as past flooding events, pipe rupture records, and equipment failure reports, can be collected from urban management department archives, reports, or unstructured text. Since data from different sources may use different naming conventions, units of measurement, or data formats, direct integration could lead to information chaos. Therefore, based on a pre-built unified data dictionary, semantic alignment and standardization are performed on the collected multi-source data. For example, all liquid level data can be standardized to "meters," and timestamps can be standardized to ISO 8601 format. During processing, conversion scripts can be manually written to map raw data fields to standard fields defined in the data dictionary, performing unit conversions, missing value imputation, and other operations. This approach ensures that all data maintains semantic and format consistency, generating a standardized fusion dataset that provides a unified, high-quality data foundation for subsequent analysis and modeling.

[0050] In step S2, holographic digital twin modeling based on spatiotemporal grids is performed. To achieve refined management and dynamic monitoring of urban drainage networks, spatial grid technology can be used to model the network. For example, the entire urban area can be divided into a series of regular two-dimensional or three-dimensional grid units, each assigned a unique code. Components in the drainage network, such as pipes, manholes, and pumping stations, can be mapped to their respective grid units. This gridded representation allows for efficient organization and indexing of the network's geometric and attribute information. When the physical state of the network changes, such as when the liquid level data of a pipe is updated, the corresponding component in the digital twin model can be quickly located using its corresponding grid code, and its attributes can be updated. This update mechanism can be implemented using periodic full refresh or event-triggered local update, thereby constructing a dynamically updated holographic digital twin model of the network and maintaining real-time synchronization between the digital twin model and the physical network.

[0051] In step S3, multimodal risk feature extraction is performed. To comprehensively assess pipeline network risk, multiple features need to be extracted from the aforementioned fused dataset. Pipeline network operational status features can include real-time liquid level, flow rate, flow velocity, water quality parameters, etc., which directly reflect the current operational status of the pipeline network. Environmental features cover external factors, such as real-time rainfall, soil moisture, and groundwater level, which may affect the pipeline network's carrying capacity and structural stability. Historical disaster features can include the frequency of past flooding, the type and location of pipeline damage, maintenance records, etc., which help identify weak links and potential risk patterns in the pipeline network. These features can be extracted from the raw data using methods such as statistical analysis and time series analysis. Drainage pipeline networks are complex interconnected systems, and the risk of a single node is often closely related to the status of adjacent nodes. To capture this spatial dependence, a graph representation of the pipeline network topology can be constructed, where pipeline nodes (e.g., manholes, connection points) are nodes of the graph, and pipe segments (e.g., pipes) are edges of the graph. Each node can be associated with its extracted operational status, environmental, and historical disaster features. Subsequently, these node features are processed using a graph neural network. For example, a simple multilayer perceptron (MLP) can be designed as a layer of a graph neural network. By iteratively aggregating the neighbor information of each node and updating it in combination with its own features, a risk feature vector containing risk information of itself and its neighborhood can be generated for each pipeline node.

[0052] In step S4, risk probability prediction based on a dynamic Bayesian network is performed. To achieve dynamic prediction of pipeline network risks, a dynamic Bayesian network model can be constructed. This model can describe the evolution of the risk state of pipeline network nodes over time. The initial state distribution represents the probability that each pipeline network node is in a different risk state (e.g., normal, warning, high risk) at the start of prediction. The state transition probability defines the likelihood that a node will transition from a certain risk state in the previous time step to another risk state in the current time step. The observation likelihood function describes the probability of observing a specific risk feature vector given that a node is in a certain risk state. These parameters can be initially set through expert experience, historical data statistics, or machine learning methods. At each time step, the aforementioned risk feature vector is used as the observation input. The Bayesian filtering algorithm uses the current observation value, the posterior probability of the previous time step, and the state transition probability to recursively update the posterior probability of each node in each risk state through iterative calculations in the prediction and update steps. For example, the prediction step predicts the prior probability of the current time step based on the posterior probability and state transition probability of the previous time step, while the update step corrects the prior probability by incorporating the current observation likelihood function to obtain the posterior probability of the current time step. These posterior probabilities serve as the risk probability output of the node, reflecting the likelihood of it being in different risk states.

[0053] In step S5, a multi-hazard coupling risk assessment is performed. The risks of urban drainage networks are not limited to a single hazard; they often involve the interaction of multiple hazards such as flooding and gas leaks. After obtaining the risk probabilities of each node, a pre-constructed flooding model can be used as a basis, based on the risk probabilities obtained in step S4. The flooding model can simulate the depth and extent of surface water accumulation in the city based on real-time rainfall and the network's operational status. For example, the flooding risk probability can be linearly superimposed with the gas leak risk probability to obtain a preliminary coupling risk index. By calculating the coupling risk index of all nodes and visualizing it on a geographic information system, a multi-hazard coupling risk index is calculated, generating a dynamic risk distribution map of the urban drainage network, intuitively displaying the risk levels and types of different areas.

[0054] In step S6, intelligent scheduling of emergency knowledge and resources is performed. To improve the efficiency of emergency response, an emergency knowledge graph of the urban pipeline network can be used. This graph contains information on emergency resource points such as emergency teams, material warehouses, and repair equipment, as well as information on the geographical location and risk level of different risk points, and the relationships between them are established. When a high-risk area is identified, relevant emergency resources can be obtained using this knowledge graph. Subsequently, path optimization algorithms, such as breadth-first search (BFS) or depth-first search (DFS), are used to calculate the shortest path from the nearest emergency resource point to the risk point, taking into account road conditions (e.g., simply treating all roads as passable). Finally, the optimal scheduling path from the emergency resource point to the risk point is generated, and an emergency decision-making plan is output to guide emergency personnel in a rapid response.

[0055] In this embodiment, a closed-loop, intelligent urban drainage network risk prediction and emergency response solution is formed through a series of innovative technologies, including multi-source heterogeneous data fusion, dynamic digital twin modeling, multi-modal risk feature extraction based on fused topology, dynamic prediction of multi-hazard coupled risks based on dynamic Bayesian networks, and intelligent scheduling of emergency knowledge and resources. This solution overcomes the shortcomings of existing technologies in data processing, model timeliness, comprehensiveness of risk perception, dynamic prediction, and intelligent emergency response, providing strong technical support for the safe operation of urban drainage networks.

[0056] In a preferred embodiment of the present invention, step S1 further includes:

[0057] By introducing R2RML mapping rules, unstructured historical disaster data is converted into resource description framework triple format, which is then integrated with structured data to construct the foundation layer of the urban drainage network knowledge graph.

[0058] In this embodiment, R2RML mapping rules are a W3C recommendation standard. Their core function is to define how to map data from a relational database to a Resource Description Framework (RDF) dataset, i.e., how to convert database tables, columns, rows, and other elements into RDF triples (subject-verb-object). In this embodiment, natural language processing techniques are first used to extract entities and relations from unstructured historical disaster data, transforming it into structured data. Then, R2RML mapping rules are used to convert it into RDF triple format. These rules can be written in XML or Turtle formats, explicitly specifying the correspondence between database tables and RDF classes, columns and RDF attributes, and how to generate Uniform Resource Identifiers (URIs).

[0059] Unstructured historical disaster data refers to data that does not have a predefined data model or fixed format, and usually exists in the form of text, images, audio, or video. In the field of urban drainage pipe networks, this may include text files such as historical accident reports, maintenance logs, and user complaint records, which contain natural language descriptions of the time, location, type, scope of impact, and handling process of the disaster; or on-site photos and video materials, recording visual information of the disaster scene such as flooding, pipe rupture, and facility damage.

[0060] The resource description framework triple format is a standard data model for describing Web resources, with a basic structure of "subject-verb-object". This format can represent the relationships between entities in a machine-understandable way. For example, triples can be represented using standard serialization formats such as Turtle, N-Triples, RDF / XML, or JSON-LD, or these triples can be stored and managed through RDF storage systems (such as triple stores or graph databases) for efficient querying and reasoning.

[0061] Structured data refers to data with a well-defined data model, fixed format, and pre-defined fields, typically stored in relational databases, spreadsheets, or XML files. In the field of urban drainage networks, this may include basic geographic information data of the drainage network, such as pipe ID, material, diameter, burial depth, and coordinates, usually stored in a geographic information system database.

[0062] The foundational layer of the urban drainage network knowledge graph refers to the core knowledge system that comprises urban drainage network entities and their interrelationships, constructed through the fusion, standardization, and semantic processing of multi-source heterogeneous data. It provides unified, high-quality knowledge services for upper-layer applications. For example, ontology modeling tools (such as Protégé) can be used to define concepts, attributes, and relationships in the drainage network domain, forming the schema layer of the knowledge graph, upon which instance data can be populated.

[0063] In a preferred embodiment of the present invention, step S2 further includes:

[0064] The pipeline network is 3D spatially encoded using a three-dimensional spatial grid partitioning technique (the pipeline network space is recursively partitioned using an octree structure, with each leaf node corresponding to a spatial unit no larger than 1m×1m×1m, and Morton code is used for encoding), establishing a dynamic mapping relationship between grid units and pipeline network components; when IoT sensing data is updated, the corresponding component is located according to the grid encoding index, triggering a local update of the digital twin model.

[0065] In this embodiment, the three-dimensional spatial mesh partitioning technology aims to divide the complex three-dimensional space of the urban drainage pipe network into regular or irregular mesh units and assign a unique code to each mesh unit. This technology can discretize and digitize the physical spatial structure of the pipe network, providing a precise spatial positioning and indexing basis for pipe network components. For example, an octree structure can be used for spatial partitioning, recursively dividing the three-dimensional space into eight subspaces until a preset minimum mesh granularity is reached or the number of pipe network components contained meets the requirements. The path of each subspace node constitutes its three-dimensional spatial code.

[0066] Building upon this foundation, establishing a dynamic mapping relationship between grid cells and pipeline components means establishing a real-time, updatable association between these 3D spatially encoded grid cells and actual physical pipeline components (such as pipes, manholes, and pumping stations) in the digital twin model. This mapping relationship allows the system to quickly locate the corresponding physical component through the grid code, and vice versa. This mapping relationship can be maintained through a database or key-value store system. For example, a mapping table can be created in the database, containing fields such as grid code, component ID, component type, and component spatial location. When the pipeline component information is updated, this mapping table is updated synchronously.

[0067] When IoT sensing data is updated, such as data changes from level sensors, flow meters, and water quality sensors, and is received by the system, the system can locate the corresponding component based on the grid code index. IoT sensing data typically contains sensor IDs or location information. The system can pre-bind sensor IDs to the grid codes of the pipeline components they are located on. When a data update is received from a sensor, the system looks up the corresponding grid code through the sensor ID, thereby locating the pipeline component.

[0068] Once a component is located, the system triggers a partial update to the digital twin model. This means that only the parts of the digital twin model relevant to the updated IoT sensing data are modified and adjusted, rather than rebuilding or globally refreshing the entire model. For example, after locating a corresponding pipeline component, the system can call the component's corresponding representation in the digital twin model (e.g., 3D geometric model, attribute data, operating status parameters, etc.) and update its attributes or status based on the new sensing data. Partial updates can also be implemented through an event-driven mechanism. When the sensing data of a pipeline component changes, an event is triggered, carrying the updated data and the component ID. The digital twin platform subscribes to and processes this event, thereby updating only the affected component's model data and visualization status.

[0069] In a preferred embodiment of the present invention, step S3 further includes:

[0070] The drainage pipe network topology is modeled as a graph structure, where nodes represent pipe network nodes and edges represent pipe segment connections. A graph convolutional network is used to aggregate and update node features, and the node features output by the network are used as risk feature vectors.

[0071] In this embodiment, modeling the drainage network topology as a graph structure means abstracting the complex drainage network system into a mathematical model composed of points and lines. This graph structure can intuitively represent the various components of the network and their interconnections. This can be achieved by parsing the network data stored in a geographic information system to extract the network's geometric information and topological connections.

[0072] In this context, a node represents a key location or facility within the drainage network. These nodes can include, but are not limited to, manholes, storm drains, pumping stations, intercepting wells, discharge outlets, and connection points of pipe sections. They can be implemented by extracting point features with specific attributes from GIS data and using them as nodes in the graph.

[0073] Meanwhile, the edge representation of pipe segment connections refers to the line segment connecting two nodes in a graph structure, representing the physical connection or hydraulic connectivity of pipe segments in a drainage network. These pipe segment connections can be actual pipe connections or paths for water flow from one node to another. This can be achieved by analyzing the start and end points of line features in GIS data to determine which nodes have direct pipe segment connections and constructing the graph edges accordingly, or by establishing logical connections between nodes based on the hydraulic model and flow direction rules of the network.

[0074] Building upon this, the use of graph convolutional networks to aggregate and update node features refers to utilizing a deep learning model specifically designed for graph-structured data to integrate and optimize the feature information of each network node. Graph convolutional networks iteratively aggregate the feature information of neighboring nodes and combine it with the node's own features to generate new, more representative node feature representations. This can be achieved by defining propagation rules for the graph convolutional layers; for example, spectral domain graph convolutional networks or spatial domain graph convolutional networks can be used to aggregate the features of neighboring nodes.

[0075] Ultimately, using the node features output by the network as risk feature vectors means that after multiple layers of feature aggregation and updating, the graph convolutional network generates a high-dimensional numerical vector for each network node. This vector integrates the node's own attribute information, its position in the network topology, and the influence of its neighboring nodes. These output feature vectors can be directly used as the risk feature vector for that node, for subsequent risk probability prediction. One implementation method is to directly embed the nodes from the last layer of the GCN as risk feature vectors.

[0076] In a preferred embodiment of the present invention, step S4 further includes:

[0077] The initial state distribution is the probability that a node is in each risk state at the initial moment;

[0078] The state transition probability is the probability that a node will transition from a risk state at the previous moment to a risk state at the current moment, and it is positively correlated with the service life of the pipeline and the soil corrosivity.

[0079] The observation likelihood function is the probability density of a node observing the current risk feature vector when it is in a certain risk state.

[0080] Input the risk feature vector at the current moment, and calculate the posterior probability at the current moment recursively using Bayesian filtering based on the current observation likelihood value, the posterior probability at the previous moment, and the state transition probability.

[0081] In this embodiment, the initial state distribution refers to the probability that each node in the urban drainage network is in a different risk state (e.g., normal, slightly abnormal, moderately abnormal, severely abnormal) at the start of the prediction process. This distribution provides the starting point for time series analysis of the entire dynamic risk prediction model. Its implementation can include: statistically analyzing risk events at the initial stage of network operation or at specific points in time based on historical statistical data to obtain the initial probability of each risk state; or, combining expert experience and prior knowledge to set reasonable initial risk probabilities for network nodes of different types and regions.

[0082] State transition probability describes the likelihood that a pipeline node will transition from a certain risk state in the previous moment to another risk state in the current moment after a certain time step. Determining this probability considers key factors such as pipeline service life and soil corrosivity, which directly affect the aging and deterioration process of the pipeline and thus show a positive correlation with the evolution of risk states. Specifically, state transition probability can be determined in the following ways: by collecting a large amount of historical accident data and pipeline maintenance records, and using statistical methods (such as Markov chain models) to fit the state transition patterns under different service lifespans and soil corrosivity conditions; or, based on physical models such as pipeline materials science and corrosion kinetics, the deterioration process of the pipeline under different environmental conditions can be simulated to derive the state transition probability.

[0083] The observation likelihood function is the probability density of the observed risk feature vector (e.g., a vector composed of sensor data such as liquid level, flow rate, water quality, and pressure) when a pipeline node is in a specific risk state. This function bridges the gap between potential risk states and actual observation data, enabling the model to infer the current risk state based on real-time monitoring data. The observation likelihood function can be constructed using various methods: for example, by collecting pipeline monitoring data under different risk states (such as simulated faults or actual fault occurrences), nonparametric estimation methods (such as kernel density estimation) or parametric estimation methods (such as Gaussian mixture models) can be used to fit the probability distribution of the observed data; alternatively, it can incorporate domain expert knowledge to construct a likelihood function based on rules or fuzzy logic.

[0084] Bayesian filtering recursive computation is an algorithm for estimating hidden states in dynamic systems. It accomplishes this by iteratively combining prediction and update steps. In risk prediction, its role is to dynamically update the posterior probability of a pipeline node being in each risk state based on the latest observation data. Specific implementations can include algorithms such as Kalman filtering (suitable for linear Gaussian systems), Extended Kalman Filter (EKF, suitable for nonlinear systems), Unscented Kalman Filter (UKF, suitable for nonlinear systems with higher accuracy than EKF), or Particle Filter (suitable for arbitrary nonlinear non-Gaussian systems). At each time step, these algorithms first predict the prior probability of the current time step using the posterior probability and state transition probability from the previous time step. Then, combining the current time step's risk feature vector (observation input) and observation likelihood function, they calculate the posterior probability of the current time step using the Bayesian update formula.

[0085] In a preferred embodiment of the present invention, the method for determining the state transition probability further includes:

[0086] Historical accident data were collected, and the state transition probability matrix was fitted using maximum likelihood estimation. For sparse data regions, a physical model based on pipeline aging patterns was pre-defined.

[0087] In this embodiment, "collecting historical accident data" refers to systematically acquiring records of various abnormal events or malfunctions related to the operation of the drainage pipe network. This data may include, but is not limited to, the time, location, duration, and scope of impact of events such as pipe blockage, rupture, overflow, and flooding, as well as related environmental conditions (such as rainfall and water levels). Collection methods may include manually entering historical event reports, extracting data from the urban management department's database, or identifying and recording abnormal patterns through IoT sensor data analysis.

[0088] The phrase "fitting the state transition probability matrix using maximum likelihood estimation" refers to using statistical methods to calculate the probability of a drainage network node transitioning from one risk state to another based on collected historical accident data. Maximum likelihood estimation is a commonly used parameter estimation method. Its core idea is to find a set of parameters that maximizes the probability of their occurrence given the observed data. In pipeline network risk prediction, network nodes can be defined as several discrete risk states (e.g., normal, low risk, moderate risk, high risk). By analyzing the state change sequences of nodes at different time points in historical data, and using the maximum likelihood estimation method, a probability matrix describing the transition patterns between these states can be derived.

[0089] The phrase "for sparse data areas" refers to pipeline network areas that lack sufficient historical accident data to support reliable statistical analysis. This may include newly constructed pipeline networks, pipeline networks that have recently undergone large-scale renovations, or areas with insufficient data due to inadequate monitoring equipment coverage or incomplete historical records. In these areas, directly relying on historical data for maximum likelihood estimation may lead to large biases and low confidence levels in the estimation results.

[0090] The phrase "pre-setting a physical model based on pipeline aging patterns" refers to, in data-sparse regions, not relying entirely on historical data, but rather constructing or selecting appropriate physical models based on the physical characteristics, material properties, and service environment of drainage network components (such as pipes and manholes) to predict their aging process and its impact on the transition to a risk state. For example, the corrosion rate, fatigue life, and structural strength of pipeline materials over time can be considered, combined with environmental factors such as soil type, water quality, and load, to establish a physical model that can quantify the degree of pipeline deterioration. This model can pre-set the probability of transitioning from a low-risk state to a high-risk state based on information such as the pipeline's service life, design life, and maintenance records, thereby compensating for the lack of historical data.

[0091] This application's solution effectively addresses the challenge of estimating state transition probabilities in data-sparse regions by combining data-driven and physical model approaches, thereby improving the accuracy and robustness of risk prediction. Specifically, in data-rich areas, by collecting historical accident data and fitting the state transition probability matrix using maximum likelihood estimation, the statistical characteristics of actual historical data can be fully utilized to ensure that the probability matrix accurately captures the dynamic transition patterns of risk states. In data-sparse areas, a physical model based on pipeline aging patterns is used to pre-set the state transition probabilities. The model is constructed based on physical mechanisms such as pipeline service life, material properties, and soil corrosivity, avoiding estimation biases caused by insufficient data. This hybrid approach enables the dynamic Bayesian network model to obtain reliable state transition probabilities in pipeline network areas with varying levels of data abundance, thus supporting its stable operation and providing accurate risk probability predictions.

[0092] In a preferred embodiment of the present invention, the method for determining the observation likelihood function further includes:

[0093] Historical monitoring data of nodes under each risk state are collected, and the probability distribution of the observed data under each risk state is fitted using the kernel density estimation method to obtain the observation likelihood function.

[0094] In this embodiment, the scheme aims to provide a sufficient data foundation for constructing the observation likelihood function. Historical monitoring data can come from various sensors deployed in the urban drainage network, such as level sensors, flow sensors, water quality sensors, and pump station operation status sensors. These sensors can collect key parameters of the network operation in real time or near real time. In addition, unstructured or semi-structured data such as manual inspection records, maintenance logs, and historical accident reports can be integrated and transformed into structured data that can be used for model training through data cleaning, labeling, and feature engineering. By collecting data covering different risk states (such as normal operation, minor anomalies, moderate risk, and severe risk), it can be ensured that the probability distribution of the subsequent fitting can comprehensively reflect the real behavior patterns of the network under various operating conditions.

[0095] Kernel density estimation is a non-parametric method for estimating probability density functions. Unlike parametric estimation methods (such as assuming the data follows a Gaussian distribution), kernel density estimation does not presuppose the distribution of the data. Instead, it estimates the overall probability density function of the data by placing a "kernel function" (such as a Gaussian kernel or an Epanechnikov kernel) at each data point and then superimposing these kernel functions. This method can flexibly adapt to various complex data distributions, including multimodal and skewed distributions, avoiding model bias caused by incorrect assumptions about the data distribution. In practical applications, the selection of the kernel function and the optimization of the bandwidth parameter are crucial for kernel density estimation. These can be adjusted using techniques such as cross-validation to achieve the best fit.

[0096] This step utilizes kernel density estimation to fit the probability distribution of historical monitoring data collected under each predefined risk state. Specifically, for each risk state (e.g., "normal state," "mild flooding state," "moderate flooding state," etc.), all historical risk feature vectors belonging to that risk state (e.g., vectors composed of liquid level, flow rate, pump speed, etc.) are used as input, and kernel density estimation is applied to calculate the probability density of observing a specific risk feature vector under that risk state. In this way, a series of conditional probability density functions are obtained, each corresponding to a specific risk state and describing the probability of observed data occurring under that risk state.

[0097] Through the above fitting process, the final conditional probability density function obtained is the observation likelihood function required by the dynamic Bayesian network model. The observation likelihood function P(observation feature vector | risk state) quantifies the probability of observing the current risk feature vector given a specific risk state. For example, if the currently observed liquid level is X, the observation likelihood function can calculate the probability of observing liquid level X under "normal conditions" and the probability of observing liquid level X under "mild flooding conditions," etc. This function is the key input for the dynamic Bayesian network to perform Bayesian filtering recursive calculations. It associates real-time observation data with preset risk states, providing a quantitative basis for subsequent risk probability updates.

[0098] In a preferred embodiment of the present invention, the calculation of the multi-hazard coupling risk index in step S5 further includes:

[0099] Calculate the coupling risk index of node i at the current time, which is obtained by weighted summation of three terms:

[0100] The first term is the first weighting coefficient multiplied by the probability of waterlogging risk at node i;

[0101] The second term is the second weighting coefficient multiplied by the probability of gas leakage risk in the region where node i is located;

[0102] The third term is the third weighting coefficient multiplied by the product of the probability of waterlogging risk and the probability of gas leakage risk;

[0103] The sum of the first, second, and third weighting coefficients is 1, and the third weighting coefficient is used to characterize the risk amplification effect when flooding and leakage occur simultaneously. The gas leakage risk probability is derived from an external gas pipeline network risk monitoring system or obtained through parallel prediction using this method.

[0104] In this embodiment, the coupling risk index is obtained by a weighted sum of three terms. This weighted summation method can comprehensively consider the impact of different risk sources.

[0105] Specifically, the first term is the first weighting coefficient multiplied by the probability of waterlogging risk at node i. The first weighting coefficient is an adjustable parameter reflecting the relative importance of waterlogging risk within the overall coupled risk; its determination can be based on historical data statistical analysis, expert experience assessment, or learning through optimization algorithms. The probability of waterlogging risk at node i represents the likelihood of a waterlogging event occurring at that node, and its value can be derived from the risk probability predicted by the dynamic Bayesian network model in step S4 above.

[0106] The second term is the second weighting coefficient multiplied by the probability of gas leakage risk in the area where node i is located. The second weighting coefficient is also an adjustable parameter, used to reflect the relative importance of gas leakage risk in the overall coupled risk, and its determination method is similar to that of the first weighting coefficient. The probability of gas leakage risk in the area where node i is located can be derived from an independent gas pipeline network risk assessment model, or obtained through statistical analysis and prediction of historical gas leakage event data.

[0107] The third term is the third weighting coefficient multiplied by the product of the probability of waterlogging risk and the probability of gas leakage risk. The third weighting coefficient is a key parameter specifically used to quantify the potential risk amplification effect when waterlogging and gas leakage occur simultaneously. For example, waterlogging may lead to soil liquefaction and pipeline displacement, thereby increasing the risk of gas pipeline rupture; while gas leakage in a flooded environment may obstruct gas diffusion, forming localized high-concentration areas, increasing the risk of explosion. This coefficient can be determined based on statistical analysis of historical coupled disaster events, physical model simulations, or expert knowledge. The product of the probability of waterlogging risk and the probability of gas leakage risk directly reflects the combined probability of the two disasters occurring simultaneously.

[0108] Furthermore, the sum of the first, second, and third weighting coefficients is 1, which ensures the normalization of the coupled risk index, making its result between 0 and 1, which facilitates the classification and comparison of risk levels.

[0109] The proposed solution introduces an innovative weighted summation calculation method to achieve accurate quantification of the multi-hazard coupling risk of urban drainage pipe networks, with particular attention to the risk amplification effect when waterlogging and gas leaks occur simultaneously.

[0110] In step S5, when it is necessary to calculate the coupling risk index of node i at the current time, the system will first obtain the waterlogging risk probability of the node (e.g., through the dynamic Bayesian network prediction result of step S4 above) and the gas leakage risk probability of its area (e.g., through the external gas pipeline network risk model or historical data analysis).

[0111] Subsequently, the system multiplies these two independent risk probabilities by preset first and second weighting coefficients to obtain independent contribution terms for flooding risk and gas leak risk, respectively. Based on this, this scheme further multiplies the flooding risk probability by the gas leak risk probability and then by a dedicated third weighting coefficient to capture and quantify the potential risk amplification effect when both disasters occur simultaneously. For example, flooding may exacerbate corrosion or displacement of gas pipelines, thereby increasing the risk of leaks; conversely, gas leaks in a flooded environment may lead to an expanded explosion radius or make rescue efforts more difficult.

[0112] By weighting and summing these three factors, and ensuring that the sum of all weight coefficients is 1, this scheme generates a comprehensive and normalized coupled risk index. This calculation method not only considers the risk of a single hazard, but more importantly, by introducing a product term and corresponding weight coefficients, it effectively solves the problem of traditional methods failing to quantify the interaction of multiple hazards, especially the risk amplification effect. This makes the risk assessment results of urban drainage networks more accurate and comprehensive, providing a solid foundation for subsequent generation of dynamic risk distribution maps and emergency decision-making.

[0113] In a preferred embodiment of the present invention, step S6 further includes:

[0114] The objective function is to minimize the emergency response time, and the constraint is the road traffic status. The road traffic status is dynamically updated based on the real-time water depth. The emergency resource scheduling path is generated using the Dijkstra algorithm or the differential evolution algorithm, and the Dijkstra algorithm is rerun after each dynamic update.

[0115] In this embodiment, using the shortest emergency response time as the objective function means that when planning emergency resource dispatch routes, the system will optimize by minimizing the time required to travel from the emergency resource point to the risk point. This ensures that the dispatch plan can minimize the arrival time of rescue and improve emergency efficiency. One implementation method is to use the travel time of each segment on the route as a weight, calculate the total response time by accumulating the route weights, and then seek to minimize this sum.

[0116] Using road accessibility as a constraint means that the system considers the actual accessibility of roads during the path planning process. This implies that some roads may be impassable or have extremely low accessibility due to specific conditions (such as flooding, landslides, traffic control, etc.), and these roads will be excluded from the available paths or assigned extremely high access costs. One implementation is to define road accessibility as a binary variable (accessible / impassable), with impassable roads being directly excluded during path search.

[0117] The dynamic updating of road traffic status based on real-time water depth means that the system can acquire real-time information on the water depth on the road and dynamically adjust the road traffic status accordingly. For example, when the real-time water depth of a certain road segment exceeds the safe passage threshold for vehicles, the traffic status of that road segment will change from "passable" to "impassable" or the passage cost will increase significantly. One implementation method is to monitor the water depth in real time using water level sensors or video recognition technology deployed on the road and transmit the data to the dispatch system for updates.

[0118] Generating emergency resource scheduling paths using Dijkstra's algorithm or differential evolution algorithm refers to calculating the optimal path from emergency resource points to risk points using these optimization algorithms, under the premise of satisfying the above objective function and constraints. These algorithms can efficiently find optimal solutions in complex networks. Dijkstra's algorithm is a classic single-source shortest path algorithm, suitable for graphs with non-negative edge weights, and can quickly find the shortest path. In practical applications, the road network can be abstracted as a graph, with the travel time of road segments as the edge weights, and Dijkstra's algorithm can be used to solve the problem. Differential evolution algorithm is a swarm intelligence-based optimization algorithm, suitable for solving optimization problems in continuous spaces, especially performing well in handling nonlinear, multimodal, and constrained complex optimization problems. In path planning, paths can be encoded as individuals, and the shortest path that satisfies the constraints can be found through iterative optimization.

[0119] This application's solution constructs an emergency resource dispatch mechanism that can adapt to real-time changes in the urban environment by taking the shortest emergency response time as the core optimization objective and introducing dynamically changing road traffic conditions as a key constraint. Specifically, upon receiving a risk warning or disaster report, the system first identifies risk points requiring emergency response and available emergency resource points. Subsequently, the system acquires or predicts the water depth of each road segment in real time and dynamically evaluates the traffic status of each road based on preset thresholds or models. For example, it marks road segments with excessively deep water as impassable or assigns extremely high passage costs. This dynamically updated road traffic status information is integrated into the road network model in real time as a constraint condition for the path planning algorithm. Based on this, the path optimization algorithm searches for the optimal path from the emergency resource point to the risk point in the road network that considers the real-time road traffic conditions, with the goal of minimizing the emergency response time. This dynamic path planning method ensures that the generated emergency dispatch path is not only theoretically the shortest but also highly feasible in practice, avoiding path failure or inefficiency caused by sudden situations such as road flooding. By closely integrating with the urban drainage network risk prediction system (i.e., the emergency knowledge and resource intelligent scheduling mentioned in step S6), this solution can quickly initiate emergency response based on accurate risk prediction results and provide highly adaptive scheduling schemes, thereby significantly improving the resilience and efficiency of the city in responding to drainage network risk events.

[0120] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data processing method for risk prediction of urban drainage pipe networks, characterized in that, Includes the following steps: Step S1: Multi-source heterogeneous data fusion and standardization processing: Collect basic geographic information data of drainage pipe network, IoT sensing data, meteorological data and historical disaster data; based on the pre-built unified data dictionary, perform semantic alignment and standardization processing on the collected multi-source data to generate a standardized fusion dataset; Step S2: Holographic digital twin modeling based on spatiotemporal grid: The urban drainage network is divided and encoded into a three-dimensional spatial grid to construct a holographic digital twin model of the network that supports dynamic updates; Step S3, Multimodal Risk Feature Extraction: Extract pipeline network operation status features, environmental features, and historical disaster features from the fused dataset; A graph representation of the pipeline network topology is constructed, and a risk feature vector for each pipeline node is generated using a graph neural network. Step S4: Risk probability prediction based on dynamic Bayesian network: Construct a dynamic Bayesian network model, which is defined by the initial state distribution, state transition probability and observation likelihood function; take the risk feature vector as the observation input, and recursively calculate the posterior probability of each node in each risk state through Bayesian filtering, which is used as the risk probability output of the node. Step S5, Multi-hazard Coupling Risk Assessment: Based on the risk probabilities of each node obtained in Step S4, and combined with the pre-constructed urban flooding model, calculate the multi-hazard coupling risk index and generate a dynamic risk distribution map of the urban drainage network. The calculation of the multi-hazard coupling risk index further includes: Calculate the coupling risk index of node i at the current time, which is obtained by weighted summation of three terms: The first term is the first weighting coefficient multiplied by the probability of waterlogging risk at node i; The second term is the second weighting coefficient multiplied by the probability of gas leakage risk in the region where node i is located; The third term is the third weighting coefficient multiplied by the product of the probability of waterlogging risk and the probability of gas leakage risk; The sum of the first, second, and third weighting coefficients is 1, and the third weighting coefficient is used to characterize the risk amplification effect when waterlogging and leakage occur simultaneously. Step S6, Emergency Knowledge and Resource Intelligent Scheduling: Based on the urban pipeline network emergency knowledge graph, a path optimization algorithm is used to generate the optimal scheduling path from emergency resource points to risk points and output emergency decision-making solutions.

2. The method according to claim 1, characterized in that, Step S1 further includes: By introducing R2RML mapping rules, unstructured historical disaster data is converted into resource description framework triple format, which is then integrated with structured data to construct the foundation layer of the urban drainage network knowledge graph.

3. The method according to claim 1, characterized in that, Step S2 further includes: A three-dimensional spatial grid partitioning technique is used to encode the pipeline network in three dimensions, establishing a dynamic mapping relationship between grid units and pipeline components. When IoT sensing data is updated, the corresponding component is located according to the grid coding index, triggering a local update of the digital twin model.

4. The method according to claim 1, characterized in that, Step S3 further includes: The drainage pipe network topology is modeled as a graph structure, where nodes represent pipe network nodes and edges represent pipe segment connections. A graph convolutional network is used to aggregate and update node features, and the node features output by the network are used as risk feature vectors.

5. The method according to claim 1, characterized in that, Step S4 further includes: The initial state distribution is the probability that a node is in each risk state at the initial moment; The state transition probability is the probability that a node will transition from a risk state at the previous moment to a risk state at the current moment, and it is positively correlated with the service life of the pipeline and the soil corrosivity. The observation likelihood function is the probability density of a node observing the current risk feature vector when it is in a certain risk state. Input the risk feature vector at the current moment, and calculate the posterior probability at the current moment recursively using Bayesian filtering based on the current observation likelihood value, the posterior probability at the previous moment, and the state transition probability.

6. The method according to claim 5, characterized in that, The method for determining the state transition probability further includes: Historical accident data were collected, and the state transition probability matrix was fitted using maximum likelihood estimation. For sparse data regions, a physical model based on pipeline aging patterns was pre-defined.

7. The method according to claim 5, characterized in that, The method for determining the observation likelihood function further includes: Historical monitoring data of nodes under each risk state are collected, and the probability distribution of the observed data under each risk state is fitted using the kernel density estimation method to obtain the observation likelihood function.

8. The method according to claim 1, characterized in that, Step S6 further includes: The objective function is to minimize the emergency response time, and the constraint is the road traffic status. The road traffic status is dynamically updated based on the real-time water depth. The emergency resource scheduling path is generated using Dijkstra's algorithm or differential evolution algorithm.

Images