Intelligent well lid road surface water multi-mode detection method and system

Abstract

The application discloses a smart well lid road surface water multi-mode detection method and system, relates to the technical field of drainage pipe network data processing, and the method obtains a plurality of road surface water alarm information and a drainage pipe network topology map; determines a multi-mode confidence degree based on a ground surface water accumulation state and an underground water accumulation state corresponding to the road surface water alarm information, and screens alarm information to be processed; determines node initial features based on the alarm information to be processed and the drainage pipe network topology map; determines a water power propagation weight based on physical properties of the drainage pipe network topology map, and enhances the node initial features; locates along a water flow reverse path based on the enhanced node features and the water power propagation weight, and performs cause-effect aggregation based on a source coincidence degree to obtain a root cause node to determine a fault type. The application fuses physical water power rules and multi-mode data filtering, and realizes high-precision cause-effect aggregation and root cause positioning.

Description

Technical Field

[0001] This application relates to the field of drainage network data processing technology, specifically to a method and system for multimodal detection of surface water on intelligent manhole covers. Background Technology

[0002] Currently, in the operation and maintenance management of urban drainage pipe networks, the processing of road flooding alarm information typically employs clustering methods based on spatiotemporal proximity or textual semantic similarity for event aggregation and root cause inference. However, this approach relies heavily on the surface similarity of data, neglecting the complex hydraulic propagation mechanisms within the drainage pipe network. In practical applications, spatially adjacent flooding points may be caused by completely independent upstream blockages, while flooding points that are far apart but located on the same propagation chain often share a common root cause. Existing methods cannot identify this shared origin relationship, leading to misjudgment of root causes or the dispersed dispatch of work orders with the same origin. Summary of the Invention

[0003] This application provides a multimodal detection method and system for road surface water on intelligent manhole covers. It can embed physical hydraulic propagation weights into the feature enhancement process and combine reverse tracing for causal aggregation, thereby realizing physical homogeneous aggregation of alarm information and improving the accuracy and reliability of root cause localization.

[0004] This application provides a multimodal detection method for surface water accumulation on intelligent manhole covers, including:

[0005] Obtain multiple road surface water accumulation alarm information and drainage pipe network topology map;

[0006] Based on the surface water accumulation status and groundwater accumulation status corresponding to the road water accumulation alarm information, a multimodal confidence level is determined, and based on the multimodal confidence level, an alarm information to be processed is determined from multiple road water accumulation alarm information.

[0007] Based on the alarm information to be processed and the drainage network topology, the initial characteristics of the nodes are determined;

[0008] Based on the physical properties of the drainage network topology, hydraulic propagation weights are determined, and the initial features of the nodes are enhanced based on the hydraulic propagation weights to obtain enhanced node features.

[0009] Based on the node enhancement features and the hydraulic propagation weights, upstream positioning is performed along the reverse water flow path, and the source overlap of the positioning path is used to perform causal aggregation on the alarm information to be processed to obtain the root cause node corresponding to each alarm information to be processed.

[0010] The fault type is determined based on the root cause node.

[0011] Optionally, determining the hydraulic propagation weights based on the physical properties of the drainage network topology includes:

[0012] Obtain the pipe diameter, slope, and length of each pipe segment in the drainage network topology diagram;

[0013] Based on the pipe diameter and the slope, a first evaluation value that is positively correlated with the hydraulic propagation weight is determined;

[0014] Based on the pipe length, a second evaluation value that is negatively correlated with the hydraulic propagation weight is determined;

[0015] The hydraulic propagation weight is determined based on the first evaluation value and the second evaluation value.

[0016] Optionally, the step of enhancing the initial node features based on the hydraulic propagation weights to obtain enhanced node features includes:

[0017] A directed adjacency matrix is ​​generated based on the drainage network topology, and the hydraulic propagation weight is used as the edge weight of the directed adjacency matrix to obtain a hydraulically weighted adjacency matrix.

[0018] The hydraulically weighted adjacency matrix is ​​normalized to obtain a normalized hydraulic adjacency matrix.

[0019] The feature enhancement network calculates the enhanced features of the nodes by performing feature propagation calculations on the initial features of the nodes based on the normalized hydraulic adjacency matrix and the preset weight matrix.

[0020] The feature propagation calculation uses the normalized hydraulic adjacency matrix to weight and aggregate the initial features of the nodes, and combines a preset weight matrix and activation function to perform a nonlinear transformation, and then propagates layer by layer to obtain the enhanced features of the nodes.

[0021] Optionally, the upstream localization based on the node enhancement features and the hydraulic propagation weights along the reverse path of the water flow includes:

[0022] Based on the node enhancement features and the attributes of the alarm information to be processed, the severity score of each node is calculated;

[0023] Nodes with severity scores greater than a preset score threshold are identified as candidate problem nodes;

[0024] Starting from the candidate problem node, recursively locate the upstream node along the reverse path of the water flow;

[0025] During the localization process, for multiple upstream branches of the current node, the branch contribution intensity of each upstream branch is calculated based on the severity score of the current node and the proportion of the hydraulic propagation weight of each upstream branch in the total in-degree hydraulic weight.

[0026] The upstream node corresponding to the upstream branch whose contribution intensity is greater than a preset contribution threshold is determined as the next positioning node.

[0027] Optionally, the causal aggregation of the alarm information to be processed based on the source overlap of the location path to obtain the root cause node corresponding to each alarm information to be processed includes:

[0028] Obtain the set of nodes contained in the location path corresponding to each of the alarm messages to be processed;

[0029] For any two alarm messages to be processed, calculate the intersection and union of their node sets;

[0030] The source overlap ratio is determined based on the ratio of the number of elements in the intersection to the number of elements in the union.

[0031] If the source overlap ratio is greater than or equal to a preset overlap threshold, and the upstream node depth shared by the two node sets is greater than or equal to a preset depth threshold, then the two alarm messages to be processed will be aggregated into the same causal aggregation package.

[0032] The upstream node in the causal aggregation package is identified as the root cause node.

[0033] Optionally, determining the multimodal confidence level based on the surface water accumulation status and groundwater accumulation status corresponding to the road surface water accumulation alarm information includes:

[0034] Obtain the road surface water depth and manhole water depth corresponding to the road surface water accumulation alarm information;

[0035] If the depth of water accumulation on the road surface and the depth of water accumulation in the well meet the depth judgment conditions, then the multimodal confidence level is determined to be the first confidence level; the depth judgment conditions include that the depth of water accumulation on the road surface is greater than zero and the depth of water accumulation in the well is greater than a preset waterless zone threshold, then the multimodal confidence level is determined to be the first confidence level.

[0036] If the depth of water accumulation on the road surface and the depth of water accumulation in the well do not meet the depth judgment conditions, then the multimodal confidence level is determined to be the second confidence level; the second confidence level is less than the first confidence level.

[0037] Optionally, obtaining the road surface water depth and manhole water depth corresponding to the road surface water accumulation alarm information includes:

[0038] Acquire road surface water level monitoring data, which includes the timestamps and latitude and longitude coordinates of the road surface measuring points;

[0039] Acquire downhole liquid level monitoring data, which includes timestamps of pipeline nodes and water depth;

[0040] Based on spatial distance, the road surface measuring points are mapped to the nearest pipeline node;

[0041] Within a preset time window, extract the road surface water depth and the underground water depth corresponding to the mapped pipeline node;

[0042] The spatial distance is calculated based on the latitude and longitude coordinates of the road surface measuring point and the latitude and longitude coordinates of the pipeline node.

[0043] Optionally, after determining the fault type based on the root cause node, the method further includes:

[0044] Generate a causal aggregation package, which includes a package identifier, a list of associated original work order identifiers, a primary root cause node identifier, a root cause type, and a sequence of nodes along the entire location path;

[0045] The causal aggregation package is output to the work order dispatch system for scheduling and allocation based on the root cause type.

[0046] Optionally, determining the initial characteristics of nodes based on the alarm information to be processed and the drainage network topology includes:

[0047] The alarm information to be processed is mapped to the corresponding node in the drainage pipe network topology diagram;

[0048] For each node, the attributes of the alarm information to be processed mapped to that node are aggregated to obtain the initial node characteristics;

[0049] The initial characteristics of the node include at least one of the following: alarm information count, maximum reported water depth, average water depth, node elevation, in-degree, and out-degree.

[0050] To achieve the above and other related objectives, this application provides an intelligent manhole cover road surface water multimodal detection system, comprising:

[0051] The acquisition module is used to acquire multiple road surface water accumulation alarm information and drainage pipe network topology map;

[0052] The filtering module is used to determine the multimodal confidence level based on the surface water accumulation status and groundwater accumulation status corresponding to the road water accumulation alarm information, and to determine the alarm information to be processed among the multiple road water accumulation alarm information based on the multimodal confidence level;

[0053] The mapping module is used to determine the initial characteristics of nodes based on the alarm information to be processed and the drainage network topology map;

[0054] An enhancement module is used to determine hydraulic propagation weights based on the physical properties of the drainage network topology, and to enhance the initial features of the nodes based on the hydraulic propagation weights to obtain enhanced node features.

[0055] The aggregation module is used to perform upstream positioning along the reverse path of water flow based on the node enhancement features and the hydraulic propagation weight, and to perform causal aggregation on the alarm information to be processed based on the source overlap of the positioning path to obtain the root cause node corresponding to each alarm information to be processed.

[0056] The determination module is used to determine the fault type based on the root cause node.

[0057] As described above, the intelligent manhole cover road surface water multimodal detection method and system provided in this application has the following beneficial effects:

[0058] This application presents a multimodal detection method for road surface water accumulation using intelligent manhole covers. This method determines multimodal confidence levels based on the surface and groundwater conditions of road surface water accumulation alarm information. The method then filters road surface water accumulation alarm information based on these confidence levels, removing noise interference from non-real water accumulation and improving the quality of the input data. Simultaneously, it determines hydraulic propagation weights based on the physical properties of the drainage network topology and embeds these weights into the feature enhancement process of the node's initial features. This makes the feature representation more fluid-mechanically constrained and consistent with natural laws. Based on the node enhancement features and hydraulic propagation weights, upstream location is performed along the reverse water flow path. Causal aggregation of the alarm information to be processed is then performed based on the source overlap of the location path. This accurately identifies faults with the same source along the hydraulic propagation path, effectively distinguishing spatially adjacent but independently caused water accumulation events. This significantly reduces work order redundancy and root cause misjudgment rate, improving the accuracy of root cause localization. Attached Figure Description

[0059] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0060] Figure 1 This is a flowchart illustrating a multimodal detection method for road surface water accumulation using an intelligent manhole cover, as shown in an exemplary embodiment of this application.

[0061] Figure 2This is a structural block diagram of an intelligent manhole cover road surface water multimodal detection system, as illustrated in an exemplary embodiment of this application. Detailed Implementation

[0062] To better understand the above-mentioned objectives, features, and advantages of this application, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments and features described in the embodiments of this application can be combined with each other.

[0063] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of this application is not limited to the specific embodiments disclosed below.

[0064] Please see Figure 1 This is a flowchart illustrating a multimodal detection method for surface water on intelligent manhole covers, as shown in an exemplary embodiment of this application. (See reference...) Figure 1 It can be seen that the intelligent manhole cover road surface water multimodal detection method may include steps S110 to S160.

[0065] Step S110: Obtain multiple road surface water accumulation alarm information and drainage pipe network topology map.

[0066] In one embodiment of this application, road flooding alarm information can originate from various channels, such as citizen complaint work orders in the urban operation management service platform, automatic alarms from IoT water level gauges, and AI-recognized events from video surveillance. This information typically includes original attributes such as the time of occurrence, geographical location, and description of flooding. The drainage network topology map can be a directed graph structure constructed based on data from a municipal drainage network GIS system. Nodes in the drainage network topology map can represent physical entities such as inspection wells, storm drains, pumping stations, or outlets, and directed edges can represent pipe segments and their flow directions.

[0067] For example, a drainage network topology diagram can be represented as follows: , This represents a topology diagram of the drainage pipe network. This represents the set of nodes in the drainage pipe network topology diagram. This represents the set of directed edges. (Node set) Each node can store information such as ID, coordinates, and elevation. (Directed edge set) Each edge stores information such as the start node and the end node. Indicates inspection well, Indicates a pumping station. Indicates the outlet.

[0068] Step S120: Determine the multimodal confidence level based on the surface water status and groundwater status corresponding to the road water accumulation alarm information, and determine the alarm information to be processed from multiple road water accumulation alarm information based on the multimodal confidence level.

[0069] In one embodiment of this application, multimodal confidence can characterize the consistency between surface water accumulation and groundwater accumulation, and can be used to evaluate the authenticity of alarm information. Surface water accumulation can refer to whether there is water accumulation on the road surface or the depth of water on the road surface, while groundwater accumulation can refer to the level of underground liquid. For example, if the surface water accumulation is on the road surface but the corresponding area's underground liquid level is at a normal low level (i.e., the underground liquid level is unchanged), the road surface water accumulation alarm information may originate from ground reflection, sensor failure, or local water accumulation caused by non-drainage factors, and its multimodal confidence is low. If both the surface water accumulation and groundwater accumulation show abnormalities and their changing trends are synchronized, the multimodal confidence is high, for example, the road surface water depth continuously increases and the underground liquid level continuously increases. The alarm information to be processed based on the multimodal confidence can constitute a high-quality input set for subsequent root cause analysis. Executing step S120 can verify the road surface water accumulation alarm information, avoiding the waste of computational resources and root cause inference bias caused by false or irrelevant alarm interference.

[0070] Step S130: Determine the initial characteristics of the nodes based on the alarm information to be processed and the drainage network topology map.

[0071] In one embodiment of this application, the filtered alarm information to be processed can be associated with the spatially nearest or logically related nodes in the drainage network topology map, and aggregated to generate initial node features. These initial node features include not only dynamic indicators reflecting the severity of the current event, such as alarm count, maximum water accumulation depth, and average water accumulation depth, but also static indicators reflecting the inherent attributes of the facility, such as node elevation, historical fault probability, and in-degree and out-degree. Transforming road flooding alarm information into a structured vector ensures that each node's initial features include both the real-time flooding situation and the node's historical risk within the network, providing rich information for subsequent feature enhancement.

[0072] For example, the initial characteristics of a node can be represented as , This represents the set of nodes in the drainage pipe network topology diagram. This indicates the number of nodes in the node set. Indicates the initial feature dimension. It represents the set of real numbers.

[0073] Step S140: Determine the hydraulic propagation weights based on the physical properties of the drainage network topology, and enhance the initial features of the nodes based on the hydraulic propagation weights to obtain enhanced node features.

[0074] In one embodiment of this application, hydraulic propagation weights can be used as graph structure constraints, and initial node features can be used as node attributes. Both hydraulic propagation weights and initial node features are input into a Graph Convolutional Network (GCN) to perform feature enhancement under physical constraints on the initial node features, resulting in node-enhanced features output by the GCN that include causal propagation information. Hydraulic propagation weights can be calculated based on physical parameters such as pipe diameter, slope, and pipe length of the pipe segment. Hydraulic propagation weights characterize the flow capacity and hydraulic conduction efficiency of the pipe segment. The feature enhancement process embeds the hydraulic propagation weights as constraints into the message passing mechanism of the graph neural network, ensuring that the aggregation and updating of node features strictly follow the laws of water flow propagation. For example, the interaction strength of feature information is greater between nodes connected by pipe segments with strong hydraulic conduction capacity; while feature interaction is suppressed between pipe segments with small diameters or gentle slopes. The node enhancement features obtained in this way integrate the physical perception features of hydraulic causality, which can overcome the shortcomings of graph convolutional networks that ignore the physical characteristics of the pipeline network and easily misassociate spatially adjacent but hydraulically isolated regions, thus ensuring the consistency between the feature representation and the real hydraulic propagation process.

[0075] For example, the hydraulic propagation weight matrix can be represented as , Each element in the matrix can represent the hydraulic transmission capacity between any two nodes, i.e., the hydraulic propagation weight. For example, when nodes u and v are connected by a directed edge, the hydraulic propagation weight can be calculated based on relevant formulas; when nodes u and v are not connected by a directed edge, the hydraulic propagation weight can be set to 0. Node enhancement features can be represented as... , This indicates an enhancement of feature dimensions.

[0076] Step S150: Based on node enhancement features and hydraulic propagation weights, upstream positioning is performed along the reverse water flow path, and causal aggregation of the alarm information to be processed is performed based on the source overlap of the positioning path to obtain the root cause node corresponding to each alarm information to be processed.

[0077] In one embodiment of this application, upstream location refers to the process of recursively tracing potential fault sources against the direction of water flow, starting from downstream nodes exhibiting abnormal water accumulation, based on the risk gradient indicated by hydraulic propagation weights and node enhancement features. This allows for the priority tracking of upstream branches with strong hydraulic conductivity and high contribution to downstream anomalies, thereby quickly locating potential blockage nodes. Causal aggregation can be determined based on the degree of overlap between upstream location paths corresponding to different pending alarm information. If two or more pending alarm information that may be spatially far apart have significantly overlapping reverse tracing paths at an upstream node or pipe segment, they can be determined to have a common hydraulic root cause. These pending alarm information can be aggregated into a causal cluster, and the upstreammost overlapping node can be identified as the root cause node. The causal aggregation method provided by this application embodiment can accurately identify water accumulation events caused by the same upstream blockage, even if these water accumulation points are geographically dispersed, they can still be correctly attributed.

[0078] Step S60: Determine the fault type based on the root cause node.

[0079] In one embodiment of this application, after identifying the root cause node, the specific fault type, such as pipe siltation, structural damage, pump station failure, or illegal discharge, can be inferred based on the attributes of the root cause node and its surrounding pipe sections, historical maintenance records, and the current hydraulic status. The determination of the fault type depends not only on the location information of the root cause node but also on a comprehensive assessment combining the node enhancement features generated in the preceding steps and the abnormal patterns on the location path. For example, if the root cause node is located in an old pipe section and the upstream location path shows a gradual flow attenuation characteristic, it can be determined as chronic siltation; if the root cause node is adjacent to a pump station and accompanied by a sudden change in liquid level, it can be determined as a pump station mechanical failure. The final output fault type can provide a basis for subsequent emergency response and resource scheduling.

[0080] For example, a mapping relationship between nodes and fault types can be pre-set based on historical maintenance records. After determining the root cause node in step S150, the fault type corresponding to the root cause node is determined based on the mapping relationship. Alternatively, the fault type can be determined based on the node type of the root cause node, which may include pump stations, outlets, and inspection wells, etc.

[0081] It should be noted that the intelligent manhole cover road surface water multimodal detection method provided in this application embodiment can be executed by a terminal or a server.

[0082] In one embodiment of this application, the process of determining the hydraulic propagation weight based on the physical properties of the drainage network topology in step S140 may include: obtaining the pipe diameter, slope, and length of each pipe segment in the drainage network topology; determining a first evaluation value that is positively correlated with the hydraulic propagation weight based on the pipe diameter and slope; determining a second evaluation value that is negatively correlated with the hydraulic propagation weight based on the pipe length; and determining the hydraulic propagation weight based on the first evaluation value and the second evaluation value.

[0083] Key physical parameters such as pipe diameter, slope, and length of each pipe segment in the drainage network topology map can be obtained. These parameters can be stored in the database or asset management system of the municipal drainage network system GIS and are inherent attributes characterizing the water conveyance capacity of the pipeline. The flow capacity of the pipeline is directly proportional to the higher power of the pipe diameter and the square root of the slope. The larger the pipe diameter and the steeper the slope, the higher the water flow transmission efficiency, and the greater the corresponding hydraulic correlation strength should be. At the same time, based on the pipe length, a second evaluation value that is negatively correlated with the hydraulic propagation weight is determined. The pipe length represents the path length of water flow transmission. The longer the path, the greater the head loss along the path and the more obvious the signal attenuation. Therefore, its contribution to hydraulic propagation should decrease with increasing length. Finally, based on the first and second evaluation values, the hydraulic propagation weight is determined comprehensively.

[0084] Optionally, for each directed edge The formula for calculating the hydraulic propagation weight can be expressed as:

[0085] ;

[0086] in, This represents the hydraulic propagation weight of the pipe segment connecting node u and node v. This represents the diameter of the pipe segment connecting node u and node v. This indicates the slope of the pipe segment connecting node u and node v. This represents the length of the pipe segment connecting node u and node v. This represents the first constant and can be used to prevent the denominator from being zero. For example, It can be 10 -6 m. The hydraulic propagation weight is associated with each directed edge. Indicates the diameter of the pipe section of Power of 1.

[0087] In this formula, the molecule This can be the primary evaluation value, directly reflecting the theoretical water conveyance capacity of the pipeline; the denominator... This can serve as a second evaluation value, characterizing the drag effect during transmission. The weighting method based on physical mechanisms, compared to correlation analysis based solely on data statistics, effectively avoids the model learning spurious correlations that violate common sense in fluid mechanics, ensuring the physical interpretability of the root cause analysis.

[0088] It should be noted that if there is no directed edge between node u and node v, the hydraulic propagation weight is set to 0.

[0089] In one embodiment of this application, the process of enhancing the initial features of nodes based on hydraulic propagation weights in step S140 to obtain enhanced node features may include: generating a directed adjacency matrix based on the drainage network topology graph, and using the hydraulic propagation weights as the edge weights of the directed adjacency matrix to obtain a hydraulically weighted adjacency matrix; normalizing the hydraulically weighted adjacency matrix to obtain a normalized hydraulic adjacency matrix; and using a feature enhancement network, performing feature propagation calculations on the initial features of nodes based on the normalized hydraulic adjacency matrix and a preset weight matrix to obtain enhanced node features; wherein, the feature propagation calculation uses the normalized hydraulic adjacency matrix to weight and aggregate the initial features of nodes, and combines the preset weight matrix and activation function to perform nonlinear transformations, propagating layer by layer to obtain enhanced node features.

[0090] A directed adjacency matrix is ​​generated based on the drainage network topology, and the hydraulic propagation weights are used as the edge weights of the directed adjacency matrix to obtain a hydraulically weighted adjacency matrix. For example, if there is a directed pipe segment connection between node j and node i, the element at the corresponding position in the matrix is ​​assigned the hydraulic propagation weight of that pipe segment; otherwise, it is assigned a value of 0. Subsequently, the hydraulically weighted adjacency matrix is ​​normalized to obtain a normalized hydraulic adjacency matrix. The normalization operation can use the symmetric normalized Laplace method. By normalizing the hydraulically weighted adjacency matrix, the influence of node degree differences on the feature aggregation amplitude is eliminated, preventing gradient explosion or numerical instability during deep network propagation, while preserving the relative magnitude of the hydraulic weights. Through a feature enhancement network, based on the normalized hydraulic adjacency matrix and the preset weight matrix, feature propagation calculation is performed on the initial features of the nodes to obtain the node enhanced features. In this process, the feature propagation calculation uses a normalized hydraulic adjacency matrix to weight and aggregate the initial features of nodes, and then performs a nonlinear transformation based on a preset weight matrix and activation function. After propagation layer by layer, the enhanced features of the nodes are obtained. The feature enhancement network can be a graph convolutional network.

[0091] Construct a directed adjacency matrix A. The formula for determining each element in directed adjacency matrix A can be expressed as:

[0092] ;

[0093] in, This indicates that there is an edge from node j to node i;

[0094] Constructing a hydraulically weighted adjacency matrix Hydraulic weighted adjacency matrix The formula for determining each element in the formula can be expressed as:

[0095] ;

[0096] Determine the water force matrix Water force matrix It is a diagonal matrix, and the hydraulic matrix is... The formula for determining each element in the formula can be expressed as:

[0097] ;

[0098] Normalized hydraulic adjacency matrix The formula for determining it can be expressed as:

[0099] ;

[0100] The propagation equation can be expressed as:

[0101] ;

[0102] in, This represents the feature matrix of the nodes in the l-th layer of the graph convolutional network. Includes the initial node characteristics of all nodes; The learnable weight matrix for the l-th layer, i.e., the preset weight matrix, can be determined through supervised learning; This represents a non-linear activation function, such as the ReLU function; Let L represent the feature matrix of the nodes in the (l+1)th layer of the graph convolutional network. The graph convolutional network has L layers, and the final output of the graph convolutional network is... This indicates node augmentation features, including node augmentation features of all nodes. A weighted summation of neighbor node features was implemented, with the weighting coefficients strictly following the hydraulic propagation weights, indicating that significant information exchange only occurs between hydraulically connected nodes with strong transmission capabilities; and and This endows graph convolutional networks with the ability to fit complex nonlinear patterns. Through multi-layer propagation computation, the final feature of each node not only includes its own alarm attributes but also incorporates the accumulated risk state along the upstream hydraulic propagation path. By embedding physical mechanisms into the data-driven model, the generated node enhancement features are both data-driven and constrained by fluid dynamics laws. When processing sparse and noisy water accumulation alarm data, it can capture the true causal propagation pattern more accurately than a purely data-driven model, significantly improving the robustness of subsequent root cause localization.

[0103] For example, a training set can be constructed using historical work orders and root cause annotations. A supervised classification task is then built: for nodes in the test set, their historical annotations determine whether they are upstream blockage nodes or downstream water accumulation nodes, and a graph convolutional network is embedded into a binary classification network. A weighted cross-entropy loss function is used, assigning higher weights to root cause nodes. A pre-defined weight matrix can be obtained through training.

[0104] In one embodiment of this application, step S150, which involves upstream positioning along the reverse flow path based on node enhancement features and hydraulic propagation weights, may include: calculating the severity score of each node based on the node enhancement features and the attributes of the alarm information to be processed; identifying nodes with severity scores greater than a preset score threshold as candidate problem nodes; recursively locating upstream nodes along the reverse flow path starting from the candidate problem nodes; during the positioning process, for multiple upstream branches of the current node, calculating the branch contribution intensity of each upstream branch based on the severity score of the current node and the proportion of the hydraulic propagation weight of each upstream branch in the total in-degree hydraulic weight; and identifying the upstream node corresponding to the upstream branch with a branch contribution intensity greater than a preset contribution threshold as the next positioning node.

[0105] Starting from the downstream node exhibiting abnormal behavior, the physical source of the water accumulation is determined by moving against the direction of water propagation. First, based on the node's enhanced features and the attributes of the alarm information to be processed, a severity score is calculated for each node. The severity score characterizes the current water accumulation risk status of the node.

[0106] The formula for calculating the severity score can be expressed as:

[0107] ;

[0108] in, The severity score characterizes node v. The number of alarm messages to be processed, mapped to node v; The maximum reported water depth in the unprocessed alarm information mapped to node v can represent the instantaneous severity over time. The node augmentation features characterizing node v Representing the mean vector, , The node augmentation features characterizing node u Represents the squared Euclidean distance. As the first weighting coefficient, This is the second weighting coefficient. This is the third weighting coefficient.

[0109] Node augmentation features representing node v Compared with the global normal state mean The square of the Euclidean distance between them is used to detect hidden anomalies where the number of alarms is small but the hydraulic pattern deviates significantly from the normal. , and These are preset weighting coefficients, which can be calibrated based on historical operation and maintenance data.

[0110] Nodes with severity scores exceeding a preset threshold are identified as candidate problem nodes. Based on the severity score, areas requiring analysis can be filtered out in the network diagram, eliminating background noise nodes that, while exhibiting minor alarms, do not reach a fault level. Starting from a candidate problem node, recursively locating upstream nodes along the reverse flow path. The reverse flow path can be a path opposite to the directed edge direction in the drainage network topology diagram. During the location process, for multiple upstream branches from the starting point, the branch contribution intensity of each upstream branch is calculated based on the severity score of the starting point and the proportion of each upstream branch's hydraulic propagation weight in the total in-degree hydraulic weight. Upstream nodes corresponding to upstream branches with branch contribution intensity exceeding a preset contribution threshold are identified as the next location node. This calculation and filtering process is recursively executed until preset termination conditions are met, such as reaching the network source, exceeding a preset upper limit in backtracking depth, or all upstream branches having contribution intensities below a threshold. In this way, each candidate problem node generates one or more reverse location paths pointing to potential fault sources.

[0111] In one possible implementation, the upstream location process may include: Step 1, assuming node v is a candidate problem node currently set as the starting point, initializing the path. Step 2: Find the set of all upstream adjacent nodes of node v. Step 3, if That is, there are multiple upstream branches, and the branch contribution intensity of each upstream node can be calculated; Step 4, add upstream nodes whose branch contribution intensity is greater than the preset contribution threshold to the path, and recursively execute steps 2 to 4 for each newly added node in the path until one of the preset termination conditions is met: no upstream node, reaching the maximum depth level 5, or no new node is found for three consecutive hops.

[0112] The formula for calculating the contribution intensity of a branch can be expressed as:

[0113] ;

[0114] in, This represents the branch contribution strength of upstream node u to the current downstream node v. This represents the severity score of node v. Let x represent the hydraulic propagation weight from node u to v, and let x represent the upstream node that can flow directly to node v via a directed edge. This represents the hydraulic propagation weight from node x to v. The first term in the calculation formula represents the severity score of node v divided by the sum of the hydraulic weights of all incoming edges, representing the distribution cardinality of the total propagation intensity on the incoming edges. The second term in the calculation formula represents the directed edge... The hydraulic weight represents the proportion of all incoming edge weights, i.e., the proportion of propagation undertaken by that upstream branch. The first term normalizes the downstream anomaly level according to the total incoming hydraulic weight, while the second term further weights it based on the hydraulic weight proportion of a specific upstream branch.

[0115] For example, if node v experiences severe water accumulation and has two branches upstream, one being a large-diameter, steep-slope main pipe (with high hydraulic weight) and the other a small-diameter, gently sloping branch pipe (with low hydraulic weight), the algorithm will prioritize the main pipe as being more likely to transmit the anomaly, thus assigning it a higher contribution intensity.

[0116] In one embodiment of this application, step S150, which involves causal aggregation of the alarm information to be processed based on the source overlap of the location path to obtain the root cause node corresponding to each alarm information to be processed, may include: obtaining the set of nodes contained in the location path corresponding to each alarm information to be processed; calculating the intersection and union of the node sets of any two alarm information to be processed; determining the source overlap ratio based on the ratio of the number of elements in the intersection to the number of elements in the union; if the source overlap ratio is greater than or equal to a preset overlap threshold, and the depth of the upstream node shared by the two node sets is greater than or equal to a preset depth threshold, then aggregating the two alarm information to be processed into the same causal aggregation package; and determining the upstream node in the causal aggregation package as the root cause node.

[0117] In urban drainage networks, water accumulation points that are far apart may converge on the same main pipe or even the same pumping station outlet when traced back. If only the source overlap ratio is considered without considering depth, it is easy to mistakenly aggregate water accumulation caused by independent faults in different branches into a single root cause. Only when two paths not only overlap at nodes but also at a sufficiently deep upstream level are they considered to have a common root cause. Finally, the upstreamst node in the causal aggregation package is determined as the root cause node. Compared to the spatial proximity (i.e., common source) criterion in related technologies, the path overlap (i.e., common source) causal criterion used in this application improves the accuracy of root cause determination.

[0118] For example, dummy nodes A and B are located at the ends of two parallel branches, with their reverse positioning paths being {A,A1,A2,Main1} and {B,B1,B2,Main1}, respectively. Although nodes A and B intersect at node Main1, because the shared node is only Main1 and its depth is shallow (only 3 hops from their respective starting points, while the preset depth threshold is 4 hops), or because the source overlap ratio of their intersection {Main1} relative to the union {A,A1,A2,B,B1,B2,Main1} is extremely low, the system determines that they do not meet the aggregation condition and outputs root causes A2 and B2 independently. If the paths of nodes C and D are {C,C1,C2,C3} and {D,D1,C2,C3}, respectively, and their depths overlap at nodes C2 and C3, with a high source overlap ratio and sufficient depth, then nodes C and D are aggregated, and the upstream node C3 is identified as the common root cause. This aggregation mechanism based on deep overlap of physical paths ensures that each causal aggregation package corresponds to a physical fault point, providing a solid basis for subsequent accurate order dispatch.

[0119] In one embodiment of this application, the process of determining the multimodal confidence level in step S120 based on the surface water accumulation status and groundwater accumulation status corresponding to the road water accumulation alarm information may include: obtaining the road water accumulation depth and underground water accumulation depth corresponding to the road water accumulation alarm information; if the road water accumulation depth and the underground water accumulation depth meet the depth judgment condition, then the multimodal confidence level is determined to be a first confidence level; the depth judgment condition includes that the road water accumulation depth is greater than zero and the underground water accumulation depth is greater than a preset waterless zone threshold, then the multimodal confidence level is determined to be a first confidence level; if the road water accumulation depth and the underground water accumulation depth do not meet the depth judgment condition, then the multimodal confidence level is determined to be a second confidence level; the second confidence level is less than the first confidence level.

[0120] The depth of water accumulation on the road surface can be determined using data reported by road surface water level gauges, video AI recognition, or mobile terminals. The depth of water accumulation in manholes can also be collected using ultrasonic level sensors installed on smart manhole covers or via pipeline level gauges.

[0121] Road surface flooding alarm messages with a first confidence level can be identified as alarm messages to be processed. Road surface flooding alarm messages with a second confidence level can be manually reviewed to filter out any genuine road surface flooding alarms.

[0122] For example, the first confidence level can be set to 1, and the second confidence level can be set to 0.

[0123] In some embodiments, the process of obtaining the road surface water depth and manhole water depth corresponding to the road surface water accumulation alarm information may include: acquiring road surface water level monitoring data, which includes the timestamp and latitude / longitude coordinates of the road surface measuring points; and simultaneously acquiring manhole liquid level monitoring data, which includes the timestamp and water depth of the pipeline network nodes. Since the road surface monitoring equipment and underground pipeline facilities are not spatially one-to-one, it is necessary to map the road surface measuring points to the nearest pipeline network node based on spatial distance. The spatial distance can be calculated based on the latitude / longitude coordinates of the road surface measuring points and the pipeline network node. The Haversine formula can be used to calculate the spherical distance between two points, or an approximate calculation can be performed directly using the Euclidean distance in the projected coordinate system. The target nodes for mapping are typically limited to nodes with hydraulic connectivity, such as inspection wells or storm drains, excluding atypical water accumulation observation points such as pump stations or outlets, to ensure consistency with the node definition in the drainage pipeline network topology map.

[0124] After spatial mapping is completed, the surface water depth and manhole water depth corresponding to the mapped nodes need to be extracted within a preset time window. Since there is a certain hydraulic lag time between road runoff and underground pipe network, and the sampling frequencies of different sensors may differ, setting a reasonable time window can correlate asynchronously observed data to the same physical event, avoiding misjudgments caused by instantaneous sampling misalignment. The preset time window can include the first preset time before the reporting time of the surface water alarm information to the second preset time after the reporting time. For example, if the surface water level reaches its peak at time t, while the manhole water level only responds at t+5 minutes, comparing only the data at time t might incorrectly classify it as a low-confidence situation. However, by introducing a time window for alignment, it can be correctly identified as a high-confidence synchronous water accumulation event.

[0125] For example, if the alarm time for road surface water accumulation is t0=14:05, then the preset time window can be [t0-3, t0+3]=[14:02, 14:08]. The maximum value of the data in the time window [14:02, 14:08] can be found in the downhole liquid level monitoring data as the depth of downhole water accumulation.

[0126] It should be noted that if there is no downhole liquid level monitoring data within the preset time window, the road surface water accumulation alarm information can be marked and then manually filtered.

[0127] In one embodiment of this application, after determining the fault type based on the root cause node, the method further includes generating a causal aggregation package and outputting the causal aggregation package to the work order dispatch system. The causal aggregation package is a data container that encapsulates a complete chain of evidence and handling recommendations. It includes a package identifier, a list of associated original work order identifiers, a primary root cause node identifier, a root cause type, and a sequence of nodes along the entire location path.

[0128] The package identifier is a unique index of the causal aggregation package within the global system, used for cross-system tracking and status synchronization. The associated list of original work order identifiers records the original IDs of all pending alarm messages aggregated into the package, allowing dispatchers to view and handle multiple alarm events with different appearances but the same root cause at once, significantly reducing work order redundancy. The primary root cause node identifier indicates the location of the upstream fault point determined through reverse tracing, serving as the target coordinates for the maintenance team's on-site operations. The root cause type is a classification label for the nature of the fault, such as pipe siltation, structural damage, pump station failure, or illegal sewage discharge. This type is typically determined based on the attributes of the root cause node, historical maintenance records, and the current hydraulic status.

[0129] The entire location path node sequence records the complete topological path tracing back from the downstream abnormal node to the primary root cause node. When on-site verification reveals no obvious anomalies at the primary root cause node, maintenance personnel can check intermediate nodes step by step along this path, avoiding blind searches. It may also include root cause confidence, a quantitative assessment of the reliability of this root cause inference. Its value can be derived from the weighted pass-through of the aforementioned multimodal confidence scores, or mapped based on the magnitude of the source overlap ratio. In the work order dispatch system, this score directly determines the priority ranking of tasks; high-confidence work orders will be automatically prioritized or trigger an emergency response process. It may also include spatial data of the impact area, describing the geographical area affected by the root cause. This data can be a catchment area polygon generated based on the drainage network topology, or a dynamic buffer zone generated based on the root cause node, used to assist in delineating traffic control areas or public warning areas. Recommended treatment measures are standardized operational suggestions automatically generated based on the root cause type and severity, such as recommending a high-pressure flushing truck for severe siltation, or recommending an electromechanical maintenance team for pump station failures.

[0130] The causal aggregation package is output to the work order dispatch system for scheduling and allocation based on root cause type. Through this structured data interaction method, this application achieves a seamless connection between technical analysis and business execution. Compared to related technologies that only output an isolated fault label, the rich contextual information carried by the causal aggregation package is more conducive to automated scheduling and can also reduce the cognitive load of manual judgment. The causal aggregation package can also contain more dimensions of information, such as related meteorological and rainfall data, links to nearby video surveillance, or references to similar historical cases.

[0131] In one embodiment of this application, step S130, which determines the initial characteristics of a node based on the alarm information to be processed and the drainage network topology map, may include: mapping the alarm information to be processed to the corresponding node in the drainage network topology map; for each node, aggregating the attributes of the alarm information to be processed mapped to that node to obtain the initial characteristics of that node; wherein, the initial characteristics of the node include at least one of the following: alarm information count, maximum reported water depth, average water depth, node elevation, in-degree, and out-degree.

[0132] The alarm information to be processed is mapped to the corresponding node in the drainage network topology diagram. This mapping process is usually based on the principle of spatial proximity. For example, the Haversine formula mentioned in the previous embodiment is used to calculate the spherical distance between the alarm point and the network node, and each alarm information is associated with the nearest inspection well or storm drain node. Mapping can also be based on hydraulic connectivity, that is, alarm information is preferentially associated with the downstream inlet node within its catchment area, rather than simply the geometrically nearest node, to more accurately reflect the hydraulic attribution relationship.

[0133] For each node, the attributes of the alarm information to be processed mapped to that node are aggregated to obtain the node's initial characteristics. The construction of the node's initial characteristics adopts a hybrid representation strategy that combines dynamic monitoring data with static topological attributes to comprehensively characterize the node's water accumulation risk status. Among them, the node's initial characteristics include at least one of the following: alarm information count, maximum reported water accumulation depth, average water accumulation depth, node elevation, in-degree, and out-degree.

[0134] Alarm message count refers to the number of pending alarm messages mapped to this node within a preset statistical period, directly reflecting the frequency and intensity of water accumulation events at that location. The maximum reported water accumulation depth and average water accumulation depth capture the peak severity and overall flooding level of the event, respectively. Combining these two metrics can effectively distinguish between instantaneous deep water accumulation caused by short-duration heavy rainfall and continuous shallow water accumulation caused by long-duration rainfall. It should be noted that the maximum reported water accumulation depth in the road water accumulation alarm messages can be determined as the road water accumulation depth, and the average water accumulation depth can also be determined as the road water accumulation depth.

[0135] In terms of static properties, node elevation serves as the potential energy benchmark for the gravity flow system of drainage pipe networks. Nodes with low elevations naturally have a higher risk of water catchment and backflow. In-degree and out-degree describe the connectivity characteristics of nodes in the topology. Nodes with large in-degree and small out-degree are often water catchment bottlenecks or points prone to blockage, while nodes with large out-degree are usually the main transmission channels.

[0136] The features mentioned above together constitute a multidimensional feature vector. Where d represents the total number of feature dimensions. In practical applications, to eliminate the impact of differences in feature dimensions on subsequent graph convolution calculations, it is usually necessary to standardize or normalize the initial features of nodes, such as using Z-Score standardization or Min-Max normalization. In other implementations, more features can be introduced according to specific business needs, such as the area of ​​the catchment area to which the node belongs, the impermeability rate, the real-time operating status of the upstream pumping station, and / or the real-time rainfall of the local meteorological station. Through a feature construction method that combines dynamic and static elements and integrates spatiotemporal data, the subsequent graph neural network can not only perceive the real-time water accumulation situation but also deeply understand the inherent physical vulnerability and topological conduction characteristics of the pipeline network, thereby significantly improving the accuracy and robustness of root cause analysis.

[0137] In one embodiment of this application, the drainage network topology map can be obtained by acquiring node data and pipe segment data of the drainage network. Node data includes node identifiers, coordinates, and manhole bottom elevations, while pipe segment data includes upstream and downstream node identifiers, pipe diameter, pipe length, and slope. Specifically, this data typically originates from the municipal drainage facility management system (GIS), the water bureau's drainage department asset management system, or field survey records. The manhole bottom elevation is a core physical parameter characterizing the potential energy state of a gravity flow drainage system, directly determining the natural flow direction of water. Pipe diameter, pipe length, and slope are the basic inputs for subsequent calculations of hydraulic propagation weights. Node data may also include attributes such as node type (e.g., inspection well, storm drain, pumping station, outlet), and catchment area number, while pipe segment data may include information such as pipe material and roughness coefficient.

[0138] After acquiring the basic data, the direction of water flow is determined based on the difference in bottom elevation between adjacent nodes, and directed edges are constructed. This process follows the principle of gravity flow in fluid mechanics, that is, water always flows from a higher energy level to a lower energy level. For any pipe segment connecting node u and node v, if the bottom elevation of node u is greater than that of node v, the direction of water flow is determined to be u to v, and a directed edge u→v is constructed; otherwise, a directed edge v→u is constructed. When the bottom elevations of adjacent nodes are equal or the difference is less than a preset measurement error threshold, the slope or diameter of the pipe segment can be further combined for auxiliary determination. For example, the side with a larger slope or smaller diameter can be determined as upstream to solve the problem of ambiguous flow direction in flat areas. The set of directed edges constructed in this way forms a directed graph skeleton that reflects the actual hydraulic transmission path. Topology checks and corrections are performed on the directed edges to obtain the drainage network topology map. Because raw GIS data often suffers from issues such as acquisition errors, data entry mistakes, or lack of synchronization of historical changes, directed graphs directly constructed based on elevation differences may contain topological flaws such as flow loops, isolated subgraphs, or dangling nodes. These flaws can severely interfere with the correctness of graph convolution feature propagation and reverse tracing. Topology checking and correction can include at least one of the following operations: flow reversal detection and elimination, i.e., identifying loops in the graph through depth-first search or strongly connected component algorithms, and reversing or deleting the directed edges that are least physically consistent with the loops based on the cumulative elevation gradient values ​​of each pipe segment in the loop or historical flow monitoring data to break non-physical loops; isolated component connection, i.e. detecting whether there are isolated nodes or subgraphs in the graph that are disconnected from the main connected components. If so, generating virtual connection edges based on the principle of spatial proximity or historical connectivity records, or marking them as independent analysis units at the algorithm level to prevent analysis blind spots caused by missing data; dangling node processing, i.e. identifying abnormal endpoints with zero in-degree and not source nodes or zero out-degree and not outlet nodes, and completing or pruning them in conjunction with the surrounding pipeline structure.

[0139] As another optional implementation, dynamic operational data can be introduced for verification in determining and correcting the water flow direction to adapt to drainage systems with complex operating conditions such as pressure flow, tidal flow, or forced drainage by pumping stations. For example, start-up and shutdown logs of pumping stations or measured time-series data from pipeline flow meters can be obtained. When it is found that the actual flow direction of a certain pipe section is inconsistent with the static flow direction determined based on elevation during a specific period, the pipe section is marked as a bidirectional edge or a time-varying edge, and the flow direction configuration is switched according to the dynamic operating conditions at the time of the event during subsequent feature enhancement and source tracing processes.

[0140] In one possible implementation, in an emergency response scenario of urban flooding caused by a short-term heavy rainfall in a city, the city operation management platform receives hundreds of road flooding alarm messages within a short period of time from IoT water level gauges, video surveillance AI recognition, and citizen hotline reports. The system first performs multimodal confidence filtering, aligning and cross-validating road monitoring data with manhole liquid level data in time and space. For example, for an alarm claiming severe flooding but with a normal liquid level in the corresponding manhole, the system determines it to have a second confidence level and removes it; while alarms with abnormal synchronization between both sides are marked as having a first confidence level and added to the processing queue. The system maps the filtered alarm messages to nodes in the drainage network topology graph, constructing initial node features containing multi-dimensional attributes such as alarm count, water depth, and elevation, and uses graph convolution enhancement based on hydraulic propagation weights calculated based on pipe diameter, slope, and pipe length. Based on this, the system traces the source along the reverse water flow path and aggregates alarm messages scattered across different road sections but sharing the same upstream blockage point into causal aggregation packages based on the source overlap ratio.

[0141] In addition to the method described in the previous embodiments for determining multimodal confidence levels by dividing the data into two levels based on a fixed threshold, fuzzy logic reasoning or Bayesian networks can also be used to achieve continuous probability assessment. The depth of water accumulation on the road surface, the depth of water accumulation in manholes, and their difference can be used as fuzzy input variables, and a continuous confidence score can be derived through a preset membership function and fuzzy rule base. Alternatively, a Bayesian network model can be constructed, using sensor reliability, historical false alarm rate, rainfall intensity, etc., as prior nodes to dynamically calculate the posterior probability that the current alarm indicates actual water accumulation.

[0142] Please see Figure 2 This is a structural block diagram illustrating a multimodal detection system for road surface water accumulation on an intelligent manhole cover, as shown in an exemplary embodiment of this application. (Reference) Figure 2 It can be seen that the intelligent manhole cover road surface water multimodal detection system 200 may include:

[0143] The acquisition module 210 is used to acquire multiple road surface water accumulation alarm information and drainage pipe network topology map;

[0144] The filtering module 220 is used to determine the multimodal confidence level based on the surface water status and groundwater status corresponding to the road water accumulation alarm information, and to determine the alarm information to be processed from multiple road water accumulation alarm information based on the multimodal confidence level.

[0145] Mapping module 230 is used to determine the initial characteristics of nodes based on the alarm information to be processed and the drainage pipe network topology map;

[0146] The enhancement module 240 is used to determine the hydraulic propagation weights based on the physical properties of the drainage network topology, and to enhance the initial features of the nodes based on the hydraulic propagation weights to obtain enhanced node features.

[0147] The aggregation module 250 is used to locate upstream along the reverse path of water flow based on node enhancement features and hydraulic propagation weights, and to perform causal aggregation of the alarm information to be processed based on the source overlap of the location path to obtain the root cause node corresponding to each alarm information to be processed.

[0148] Module 260 is used to determine the fault type based on the root cause node.

[0149] It should be noted that the intelligent manhole cover road surface water multimodal detection system and the intelligent manhole cover road surface water multimodal detection method provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the intelligent manhole cover road surface water multimodal detection system provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0150] Embodiments of this application also provide an electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by one or more processors, enable the electronic device to implement the intelligent manhole cover road surface water multimodal detection method provided in the above embodiments.

[0151] Another aspect of this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a computer's processor, causes the computer to perform the multimodal detection method for surface water on intelligent manhole covers provided in the various embodiments above. This computer-readable storage medium may be included in the electronic devices described in the above embodiments, or it may exist independently and not assembled into the electronic devices.

[0152] Another aspect of this application provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the intelligent manhole cover road surface water multimodal detection method provided in the various embodiments above.

[0153] In the embodiments of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. The terms "comprising" and "including" as used throughout the specification and claims are open-ended terms and should therefore be interpreted as "comprising but not limited to".

[0154] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0155] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A multi-modal detection method for surface water accumulation on intelligent manhole covers, characterized in that, include: Obtain multiple road surface water accumulation alarm information and drainage pipe network topology map; Based on the surface water accumulation status and groundwater accumulation status corresponding to the road water accumulation alarm information, a multimodal confidence level is determined, and based on the multimodal confidence level, an alarm information to be processed is determined from multiple road water accumulation alarm information. Based on the alarm information to be processed and the drainage network topology, the initial characteristics of the nodes are determined; Based on the physical properties of the drainage network topology, hydraulic propagation weights are determined, and the initial features of the nodes are enhanced based on the hydraulic propagation weights to obtain enhanced node features. Based on the node enhancement features and the hydraulic propagation weights, upstream positioning is performed along the reverse water flow path, and the source overlap of the positioning path is used to perform causal aggregation on the alarm information to be processed to obtain the root cause node corresponding to each alarm information to be processed. The fault type is determined based on the root cause node.

2. The method for multimodal detection of surface water on intelligent manhole covers according to claim 1, characterized in that, The determination of hydraulic propagation weights based on the physical properties of the drainage network topology includes: Obtain the pipe diameter, slope, and length of each pipe segment in the drainage network topology diagram; Based on the pipe diameter and the slope, a first evaluation value that is positively correlated with the hydraulic propagation weight is determined; Based on the pipe length, a second evaluation value that is negatively correlated with the hydraulic propagation weight is determined; The hydraulic propagation weight is determined based on the first evaluation value and the second evaluation value.

3. The method for multimodal detection of surface water on intelligent manhole covers according to claim 2, characterized in that, The feature enhancement is performed on the initial features of the node based on the hydraulic propagation weights. The node enhancement features obtained include: A directed adjacency matrix is ​​generated based on the drainage network topology, and the hydraulic propagation weight is used as the edge weight of the directed adjacency matrix to obtain a hydraulically weighted adjacency matrix. The hydraulically weighted adjacency matrix is ​​normalized to obtain a normalized hydraulic adjacency matrix. The feature enhancement network calculates the enhanced features of the nodes by performing feature propagation calculations on the initial features of the nodes based on the normalized hydraulic adjacency matrix and the preset weight matrix. The feature propagation calculation uses the normalized hydraulic adjacency matrix to weight and aggregate the initial features of the nodes, and combines a preset weight matrix and activation function to perform a nonlinear transformation, and then propagates layer by layer to obtain the enhanced features of the nodes.

4. The method for multimodal detection of surface water on intelligent manhole covers according to claim 1, characterized in that, The upstream localization based on the node enhancement features and the hydraulic propagation weights along the reverse path of the water flow includes: Based on the node enhancement features and the attributes of the alarm information to be processed, the severity score of each node is calculated; Nodes with severity scores greater than a preset score threshold are identified as candidate problem nodes; Starting from the candidate problem node, recursively locate the upstream node along the reverse path of the water flow; During the localization process, for multiple upstream branches of the current node, the branch contribution intensity of each upstream branch is calculated based on the severity score of the current node and the proportion of the hydraulic propagation weight of each upstream branch in the total in-degree hydraulic weight. The upstream node corresponding to the upstream branch whose contribution intensity is greater than a preset contribution threshold is determined as the next positioning node.

5. The method for multimodal detection of surface water on intelligent manhole covers according to claim 4, characterized in that, The source overlap based on the location path is used to perform causal aggregation on the alarm information to be processed, and the root cause node corresponding to each alarm information to be processed is obtained, including: Obtain the set of nodes contained in the location path corresponding to each of the alarm messages to be processed; For any two alarm messages to be processed, calculate the intersection and union of their node sets; The source overlap ratio is determined based on the ratio of the number of elements in the intersection to the number of elements in the union. If the source overlap ratio is greater than or equal to a preset overlap threshold, and the upstream node depth shared by the two node sets is greater than or equal to a preset depth threshold, then the two alarm messages to be processed will be aggregated into the same causal aggregation package. The upstream node in the causal aggregation package is identified as the root cause node.

6. The method for multimodal detection of surface water on intelligent manhole covers according to claim 1, characterized in that, The determination of multimodal confidence based on the surface water accumulation status and groundwater accumulation status corresponding to the road surface water accumulation alarm information includes: Obtain the road surface water depth and manhole water depth corresponding to the road surface water accumulation alarm information; If the depth of water accumulation on the road surface and the depth of water accumulation in the well meet the depth judgment conditions, then the multimodal confidence level is determined to be the first confidence level; the depth judgment conditions include that the depth of water accumulation on the road surface is greater than zero and the depth of water accumulation in the well is greater than a preset waterless zone threshold, then the multimodal confidence level is determined to be the first confidence level. If the depth of water accumulation on the road surface and the depth of water accumulation in the well do not meet the depth judgment conditions, then the multimodal confidence level is determined to be the second confidence level; the second confidence level is less than the first confidence level.

7. The method for multimodal detection of surface water on intelligent manhole covers according to claim 6, characterized in that, The process of obtaining the road surface water depth and manhole water depth corresponding to the road surface water alarm information includes: Acquire road surface water level monitoring data, which includes the timestamps and latitude and longitude coordinates of the road surface measuring points; Acquire downhole liquid level monitoring data, which includes timestamps of pipeline nodes and water depth; Based on spatial distance, the road surface measuring points are mapped to the nearest pipeline node; Within a preset time window, extract the road surface water depth and the underground water depth corresponding to the mapped pipeline node; The spatial distance is calculated based on the latitude and longitude coordinates of the road surface measuring point and the latitude and longitude coordinates of the pipeline node.

8. The method for multimodal detection of surface water on intelligent manhole covers according to claim 1, characterized in that, After determining the fault type based on the root cause node, the method further includes: Generate a causal aggregation package, which includes a package identifier, a list of associated original work order identifiers, a primary root cause node identifier, a root cause type, and a sequence of nodes along the entire location path; The causal aggregation package is output to the work order dispatch system for scheduling and allocation based on the root cause type.

9. The method for multimodal detection of surface water on intelligent manhole covers according to claim 1, characterized in that, The determination of initial node characteristics based on the alarm information to be processed and the drainage network topology includes: The alarm information to be processed is mapped to the corresponding node in the drainage pipe network topology diagram; For each node, the attributes of the alarm information to be processed mapped to that node are aggregated to obtain the initial node characteristics; The initial characteristics of the node include at least one of the following: alarm information count, maximum reported water depth, average water depth, node elevation, in-degree, and out-degree.

10. A multi-modal detection system for surface water accumulation on intelligent manhole covers, characterized in that, include: The acquisition module is used to acquire multiple road surface water accumulation alarm information and drainage pipe network topology map; The filtering module is used to determine the multimodal confidence level based on the surface water accumulation status and groundwater accumulation status corresponding to the road water accumulation alarm information, and to determine the alarm information to be processed among the multiple road water accumulation alarm information based on the multimodal confidence level; The mapping module is used to determine the initial characteristics of nodes based on the alarm information to be processed and the drainage network topology map; An enhancement module is used to determine hydraulic propagation weights based on the physical properties of the drainage network topology, and to enhance the initial features of the nodes based on the hydraulic propagation weights to obtain enhanced node features. The aggregation module is used to perform upstream positioning along the reverse path of water flow based on the node enhancement features and the hydraulic propagation weight, and to perform causal aggregation on the alarm information to be processed based on the source overlap of the positioning path to obtain the root cause node corresponding to each alarm information to be processed. The determination module is used to determine the fault type based on the root cause node.

Images