Abnormal state prediction method and large model training method for urban drainage pipe network

Abstract

The application relates to the technical field of urban drainage pipe network monitoring, in particular to an abnormal state prediction method and a large model training method for a urban drainage pipe network, which comprises the following steps: acquiring real-time climate data and sensor monitoring data of a target collection node, and acquiring a flow dynamic threshold value of the target collection node; obtaining corresponding theoretical errors according to the sensor monitoring data, the real-time climate data and historical correlation data; acquiring sensor monitoring data corresponding to each collection node in an upstream and downstream topological network of the target collection node based on sensor position data of the target collection node and a pre-set spatial topological network, and obtaining a flow comparison result corresponding to the target collection node based on historical correlation data and a flow comparison algorithm; and obtaining an abnormal state evaluation result of the target collection node according to the flow dynamic threshold value of the target collection node, the flow comparison result and the theoretical errors. The application significantly improves the accuracy of urban drainage pipe network defect monitoring.

Description

Technical Field

[0001] This application relates to the field of urban drainage network monitoring technology, and in particular to a method for predicting abnormal states of urban drainage networks and a training method for a large-scale model for predicting abnormal states of urban drainage networks. Background Technology

[0002] As a key component of urban infrastructure, urban drainage networks bear the core functions of rainwater collection, transportation, storage, and sewage discharge. Abnormalities in their operation directly impact urban flood control, water environment safety, and the order of residents' lives and production. With accelerated urbanization, increased hardening of urban underlying surfaces, and frequent extreme rainfall events, coupled with the generally old age, complex topology, and concealed nature of drainage networks, problems such as siltation, damage, blockage, overflow, and sensor malfunctions are becoming increasingly prominent. Traditional operation and maintenance models relying on manual inspections and periodic dredging suffer from low efficiency, delayed response, and difficulty in achieving accurate monitoring around the clock, failing to meet the demands of refined and intelligent management of modern urban drainage systems.

[0003] Currently, the mainstream approach to monitoring the health status of drainage pipe networks is still based on a single threshold method. This involves pre-setting fixed flow or water level thresholds, and identifying anomalies when monitoring data exceeds these thresholds. This method fails to fully consider the dynamic impact of climatic factors such as rainfall intensity, season, and time of day, as well as the differences in hydraulic characteristics between different regions and pipe diameters. It is prone to false alarms during light rain or normal water periods, or to missing anomalies during heavy rain or peak flow periods. Furthermore, this method relies solely on single-point monitoring data and does not incorporate flow balance verification based on the upstream and downstream topology of the pipe network. It cannot effectively distinguish between sensor data distortion and actual network faults (such as local blockages or leaks), resulting in insufficient accuracy in anomaly detection. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] In view of the above-mentioned shortcomings and deficiencies of the existing technology, this application provides an abnormal state prediction method and a large model training method for urban drainage pipe networks. It solves the technical problems of the single threshold judgment method, which does not consider the differences in hydraulic characteristics of pipe networks of different regions and pipe diameters, and is prone to false alarms and missed anomalies. It also solves the technical problems of the method that relies on single-point monitoring data, does not combine the upstream and downstream topological relationship of the pipe network for flow balance verification, and cannot effectively distinguish between sensor data distortion and actual pipe network faults (such as local blockage and leakage), resulting in insufficient accuracy of abnormal state judgment.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the main technical solutions adopted in this application include:

[0008] In a first aspect, embodiments of this application provide a method for predicting abnormal states of urban drainage pipe networks, including:

[0009] The system acquires real-time climate data and sensor monitoring data of target collection nodes in the urban drainage network. Based on the real-time climate data, the sensor monitoring data of target collection nodes, and pre-set historical correlation data and dynamic threshold algorithm, it obtains the dynamic threshold of flow for the target collection nodes. The sensor monitoring data includes sensor location data and flow data.

[0010] Based on sensor monitoring data, real-time climate data, pre-set flow reconstruction algorithms, and historical correlation data, the flow data of the target acquisition node is reconstructed using theoretical values ​​to obtain the theoretical reconstructed data of the target acquisition node, and the theoretical error between the flow data of the target acquisition node and the theoretical reconstructed data is calculated.

[0011] Based on the sensor location data of the target acquisition node and the pre-set spatial topology network, the upstream and downstream topology networks of the target acquisition node are determined, and the sensor monitoring data corresponding to each acquisition node in the upstream and downstream topology networks are obtained.

[0012] Based on the sensor monitoring data of each acquisition node in the upstream and downstream topology network and the sensor monitoring data of the target acquisition node, as well as the pre-set historical correlation data and traffic comparison algorithm, the traffic comparison result corresponding to the target acquisition node is obtained.

[0013] Based on the target acquisition node's dynamic traffic threshold, traffic comparison results, and theoretical errors, the abnormal status judgment result of the target acquisition node is obtained.

[0014] Optionally, the historical correlation data includes historical climate data for multiple time periods, and the historical climate data for each time period includes historical flow data for all collection nodes in the urban drainage network;

[0015] Based on real-time climate data and sensor monitoring data from the target acquisition nodes, as well as pre-set historical correlation data and dynamic threshold algorithms, the dynamic threshold of traffic flow for the target acquisition nodes is obtained, including:

[0016] Based on real-time climate data and historical correlation data, the similarity between real-time climate data and historical climate data for each time period in historical correlation data is calculated, and all historical climate data with similarity reaching a pre-set similarity threshold are selected.

[0017] Based on the selected historical climate data and sensor location data of the target acquisition node, the historical flow data of the target acquisition node corresponding to each historical climate data is determined, and based on all the historical flow data determined by the target acquisition node, the historical statistical benchmark threshold and historical threshold fluctuation coefficient are obtained.

[0018] Based on historical statistical benchmark thresholds and historical threshold fluctuation coefficients, the dynamic threshold of traffic for the target acquisition node is obtained.

[0019] Optionally, based on all historical traffic data determined by the target acquisition node, historical statistical benchmark thresholds and historical threshold fluctuation coefficients are obtained, including:

[0020] Based on all historical traffic data determined by the target collection node, calculate the mean historical traffic and the standard deviation of historical traffic for the city corresponding to the target collection node.

[0021] Based on the historical average traffic flow and the standard deviation of historical traffic flow in the city corresponding to the target data collection node, the historical statistical benchmark threshold and the historical threshold fluctuation coefficient are obtained.

[0022] Optionally, the real-time climate data includes rainfall intensity data labeled with seasonal and rainfall duration tags;

[0023] Based on historical statistical benchmark thresholds and historical threshold fluctuation coefficients, the dynamic thresholds of traffic for the target data collection node are obtained, including:

[0024] The rainfall intensity data, seasonal labels, and rainfall duration labels were standardized to obtain the rainfall intensity correction coefficient, seasonal correction coefficient, and rainfall duration correction coefficient.

[0025] Based on the standardized rainfall intensity correction coefficient, seasonal correction coefficient, and rainfall duration correction coefficient, as well as the pre-set correction weights, the historical threshold fluctuation coefficient is scaled and corrected to obtain the final fluctuation coefficient.

[0026] Based on the final fluctuation coefficient and historical statistical benchmark threshold, the dynamic threshold of traffic at the target acquisition node is obtained.

[0027] Optionally, based on sensor monitoring data, real-time climate data, and pre-set flow reconstruction algorithms and historical correlation data, the flow data of the target acquisition node is reconstructed using theoretical values ​​to obtain the theoretically reconstructed data of the target acquisition node, including:

[0028] Based on real-time climate data and historical correlation data, the similarity between real-time climate data and historical climate data for each time period in historical correlation data is calculated, and all historical climate data with similarity reaching a pre-set similarity threshold are selected.

[0029] Based on the selected historical climate data and sensor location data of the target acquisition node, the historical flow data of the target acquisition node corresponding to each historical climate data is determined, and the average historical flow of the target acquisition node is calculated based on all the historical flow data determined for the target acquisition node.

[0030] Based on the pre-set physical baseline traffic and historical average traffic of the target acquisition node, the corresponding physical correction coefficient is obtained. The physical correction coefficient is the ratio of the difference between the physical baseline traffic and the historical average traffic to the physical baseline traffic.

[0031] The traffic data of the target acquisition node is reconstructed based on the physical correction coefficient to obtain the theoretical reconstructed data of the target acquisition node.

[0032] Optionally, based on the sensor monitoring data corresponding to each acquisition node in the upstream and downstream topology network and the sensor monitoring data of the target acquisition node, as well as pre-set historical correlation data and traffic comparison algorithms, the traffic comparison result corresponding to the target acquisition node is obtained, including:

[0033] Based on the sensor monitoring data corresponding to each acquisition node in the upstream and downstream topology network and the sensor monitoring data of the target acquisition node, a first upstream flow ratio and a second downstream flow ratio are obtained. The first upstream flow ratio is the ratio of the flow data of the target acquisition node to the flow data corresponding to the upstream node of the target acquisition node in the upstream and downstream topology network, and the second downstream flow ratio is the ratio of the flow data corresponding to the downstream node of the target acquisition node to the flow data of the target acquisition node in the upstream and downstream topology network.

[0034] Based on real-time climate data and historical correlation data, the similarity between real-time climate data and historical climate data for each time period in historical correlation data is calculated, and all historical climate data with similarity reaching a pre-set similarity threshold are selected.

[0035] Based on the sensor location data of the target acquisition node of historical climate data and each selected historical climate data, the historical flow data of the target acquisition node corresponding to each historical climate data is determined, and the average historical flow corresponding to the target acquisition node is calculated based on all the historical flow data determined by the target acquisition node.

[0036] Based on the traffic data of the target collection node and the historical average traffic, the third historical traffic ratio is calculated, where the third historical traffic ratio is the ratio of the traffic data of the target collection node to the historical average traffic.

[0037] The traffic comparison results corresponding to the target acquisition node include the first upstream traffic ratio, the second downstream traffic ratio, and the third historical traffic ratio.

[0038] Optionally, historical correlation data, spatial topology networks, real-time climate data, and sensor monitoring data of target acquisition nodes in urban drainage networks are all obtained through a pre-established data platform;

[0039] The data platform is built on a lake-warehouse integrated architecture, which includes a data lake and a data warehouse.

[0040] Optionally, the method further includes:

[0041] Based on the pre-set DOM data, DEM data, urban element vector data and real geographic data of the urban drainage network, an L2 level urban three-dimensional geographic base is constructed.

[0042] Based on the pre-set BIM model and the pre-established urban drainage network model, construct the digital base plate of L3 level drainage facilities;

[0043] A digital twin platform is constructed based on the L2-level urban 3D geographic base map, the L3-level drainage facility digital base map, and the data platform.

[0044] Optionally, based on the target acquisition node's dynamic traffic threshold, traffic comparison results, and theoretical errors, the defect status judgment result of the target acquisition node is obtained, including:

[0045] Based on the dynamic threshold of the target acquisition node, the flow comparison results, and the theoretical error, when the flow data of the target acquisition node is found to be abnormal, the location and abnormality type of the problematic pipe section in the urban drainage pipe network are determined based on the dynamic threshold of the flow, the flow comparison results, the theoretical error, and the pre-set spatial topology network.

[0046] Secondly, embodiments of this application provide a training method for a large-scale model for predicting abnormal states of urban drainage pipe networks, including:

[0047] A general-purpose large model is trained based on a pre-set model training set to obtain a trained defect prediction large model. During the training process, the general-purpose large model is subjected to structured pruning based on the model training set, and the structured-pruned general-purpose large model is fine-tuned based on historical climate datasets and historical sensor monitoring datasets of any urban pipe network in the model training set. The model training set includes historical climate datasets and historical sensor monitoring datasets of multiple different urban pipe networks.

[0048] The general large model is used to implement the above-mentioned method for predicting abnormal states of urban drainage pipe networks.

[0049] (III) Beneficial Effects

[0050] This application presents a method for predicting abnormal states in urban drainage pipe networks. It employs a dynamic threshold algorithm to obtain dynamic thresholds adapted to real-time operating conditions, effectively avoiding alarm and missed detection problems caused by differences in the hydraulic characteristics of the pipe network. Furthermore, it uses a flow reconstruction algorithm to obtain theoretically reconstructed data and calculate theoretical errors, achieving accurate comparison between single-point data and historical operating conditions. Simultaneously, a flow comparison algorithm enables linkage verification between single-point data and upstream / downstream topological flow, effectively distinguishing between sensor faults and actual pipe network faults, significantly improving the accuracy of abnormal state monitoring in urban drainage pipe networks. Attached Figure Description

[0051] Figure 1 A flowchart illustrating an abnormal state prediction method for urban drainage pipe networks provided in this application embodiment;

[0052] Figure 2 This is a schematic diagram of the abnormal state judgment result generation process provided in the embodiments of this application;

[0053] Figure 3 This is a schematic diagram of the urban drainage network abnormal state prediction process provided in an embodiment of this application;

[0054] Figure 4 This is a schematic diagram illustrating the process of establishing a digital twin platform as provided in this application embodiment. Detailed Implementation

[0055] To better explain and facilitate understanding of this application, the following detailed description of the application is provided in conjunction with the accompanying drawings and specific embodiments.

[0056] Urban drainage pipe networks are core urban infrastructure, undertaking key functions such as rainwater collection and transportation, and sewage discharge. Their stable operation is directly related to urban flood control, water environment safety, and residents' daily lives. With the advancement of urbanization, the hardening rate of the underlying surface has increased, and extreme rainfall has become more frequent. In addition, the pipe networks generally have problems such as being old, having complex topologies, and being highly concealed. As a result, hidden dangers such as siltation, blockage, overflow, and sensor failure are becoming increasingly prominent. The traditional manual inspection and regular dredging mode is inefficient and has a slow response time, making it difficult to meet the needs of refined and intelligent operation and maintenance.

[0057] Current methods for monitoring abnormal conditions in drainage pipe networks mostly employ a single fixed threshold method, judging anomalies by setting flow or water level thresholds. This method fails to consider the dynamic effects of climate such as rainfall intensity and season, as well as the differences in hydraulic characteristics of pipe networks of different regions and diameters. It is prone to false alarms during light rain and normal water periods, and misses anomalies during peak rainfall periods. Furthermore, it relies only on single-point data and does not combine upstream and downstream topology for flow balance verification, making it unable to distinguish between sensor distortion and actual pipe network faults, resulting in insufficient accuracy in judgment.

[0058] The proposed method for predicting abnormal states in urban drainage pipe networks uses a dynamic threshold algorithm to generate thresholds that adapt to real-time operating conditions, thus avoiding false leaks caused by differences in the hydraulic characteristics of the pipe network. It combines a flow reconstruction algorithm to obtain theoretical data and calculate errors, and uses a flow comparison algorithm to achieve single-point and upstream / downstream topology flow linkage verification, which greatly improves the accuracy of abnormal state monitoring and provides reliable support for intelligent operation and maintenance.

[0059] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0060] This application provides a method for predicting abnormal states in urban drainage pipe networks. This method is based on flow sensors pre-deployed at various data collection nodes within the urban drainage pipe network. The method is as follows: Figure 1 , Figure 2 and Figure 3 As shown, it includes:

[0061] S1. Acquire real-time climate data and sensor monitoring data of target acquisition nodes in the urban drainage network, and obtain the dynamic threshold of flow of the target acquisition nodes based on the real-time climate data, sensor monitoring data of target acquisition nodes, as well as pre-set historical correlation data and dynamic threshold algorithm; wherein, sensor monitoring data includes sensor location data and flow data;

[0062] S2. Based on sensor monitoring data, real-time climate data, and pre-set flow reconstruction algorithm and historical correlation data, reconstruct the flow data of the target acquisition node using theoretical values ​​to obtain the theoretical reconstruction data of the target acquisition node, and calculate the theoretical error between the flow data of the target acquisition node and the theoretical reconstruction data.

[0063] S3. Based on the sensor location data of the target acquisition node and the pre-set spatial topology network, determine the upstream and downstream topology networks of the target acquisition node, and obtain the sensor monitoring data corresponding to each acquisition node in the upstream and downstream topology networks.

[0064] S4. Based on the sensor monitoring data corresponding to each acquisition node in the upstream and downstream topology network and the sensor monitoring data of the target acquisition node, as well as the pre-set historical correlation data and traffic comparison algorithm, obtain the traffic comparison result corresponding to the target acquisition node.

[0065] S5. Based on the target acquisition node's dynamic traffic threshold, traffic comparison results, and theoretical error, obtain the abnormal status judgment result of the target acquisition node.

[0066] This embodiment provides a method for predicting abnormal states in urban drainage pipe networks. It employs a dynamic threshold algorithm to obtain dynamic thresholds that adapt to real-time operating conditions, effectively avoiding alarm and missed detection problems caused by differences in the hydraulic characteristics of the pipe network. Furthermore, it obtains theoretical reconstructed data and calculates theoretical errors through a flow reconstruction algorithm, and simultaneously achieves linkage verification between single-point data and upstream and downstream topological flow through a flow comparison algorithm, significantly improving the accuracy of abnormal state monitoring in urban drainage pipe networks.

[0067] Optionally, in a specific embodiment, the historical correlation data includes historical climate data for multiple time periods, and the historical climate data for each time period includes historical flow data for all collection nodes in the urban drainage network.

[0068] Based on real-time climate data and sensor monitoring data from the target acquisition nodes, as well as pre-set historical correlation data and dynamic threshold algorithms, the dynamic threshold of traffic flow for the target acquisition nodes is obtained, including:

[0069] Based on real-time climate data and historical correlation data, the similarity between real-time climate data and historical climate data for each time period in historical correlation data is calculated, and all historical climate data with similarity reaching a pre-set similarity threshold are selected.

[0070] Based on the selected historical climate data and sensor location data of the target acquisition node, the historical flow data of the target acquisition node corresponding to each historical climate data is determined, and based on all the historical flow data determined by the target acquisition node, the historical statistical benchmark threshold and historical threshold fluctuation coefficient are obtained.

[0071] Based on historical statistical benchmark thresholds and historical threshold fluctuation coefficients, the dynamic threshold of traffic for the target acquisition node is obtained.

[0072] Specifically, three rain gauges around the urban drainage network monitor the rainfall intensity from 10 minutes ago to the present moment, and the average value of the monitoring values ​​from the three rain gauges is taken.

[0073] The system acquires the current flow data monitored by the flow sensors at the target acquisition node and extracts pre-stored historical correlation data. This data includes historical climate data for multiple time periods, and the historical climate data for each time period corresponds to the historical flow data of all acquisition nodes in the urban drainage network.

[0074] Based on the hydraulic characteristics of the pipeline network, calculate the minimum normal flow rate and the maximum design flow rate to ensure that the threshold does not exceed the physical limit (to avoid safety risks caused by unreasonable threshold settings).

[0075] Minimum normal flow rate is generally set as the average flow rate of the target data collection node over a period of time when there is no rainfall.

[0076] The maximum design flow rate (the limit flow rate to avoid pipeline overflow) is generally calculated as follows: Maximum safe flow rate = Pipeline flow capacity × Safety factor (the safety factor is generally taken as 0.9, i.e., a 10% safety margin is reserved). Wherein, the pipeline flow capacity (full pipe flow) is based on Manning's formula Q = (A × R) (2 / 3) ×I (1 / 2) ) / n, where A is the pipe cross-sectional area, R is the hydraulic radius, I is the pipe slope, and n is the pipe roughness coefficient. Alternatively, the maximum safe flow rate can be determined based on engineering experience using data from similar pipe networks.

[0077] The minimum normal flow rate and the maximum design flow rate constitute the normal operating range of the pipeline network.

[0078] The data is filtered based on the similarity between real-time climate data and historical climate data for each time period in the historical correlation data. For example, data from the summer of 2021-2023 (June-August), with rainfall intensity of 25-30 mm / h and morning between 8-9 am, are filtered out, and a total of 86 valid data points are extracted. The historical mean flow rate μ and the standard deviation of the city's historical flow rate σ are calculated for the 86 data points.

[0079] The historical statistical baseline threshold is calculated using μ±2σ, which covers 95% of normal data to reduce misjudgment. The historical threshold fluctuation coefficient is calculated using σ / μ×100%.

[0080] By learning the correlation between "features and traffic fluctuations" in historical data through algorithms, differentiated weights are assigned to features in the current scenario (features with greater impact have higher weights), providing a basis for threshold correction.

[0081] Generally, both real-time and historical climate data include rainfall intensity data labeled with seasonal and rainfall duration labels; feature weight adaptation calculation, that is, by using scene weights, "scaling correction" (adjusting the fluctuation range) is applied to historical statistical benchmark thresholds to finally generate dynamic traffic thresholds that are adapted to the current scene (which must be within the range of physical constraint thresholds).

[0082] The scaling correction is used to adjust the fluctuation amplitude. The correction formula is: Final fluctuation coefficient = Historical threshold fluctuation coefficient × (1 + ∑(Feature correction coefficient × Weight corresponding to each feature correction coefficient)). Here, the feature correction coefficient is the feature value after standardizing the real-time climate data. Specifically, the feature correction coefficient includes the correction coefficients corresponding to rainfall intensity, seasonal label, and duration label. The specific values ​​of these feature correction coefficients need to be determined according to the pre-deployed correction coefficient strategy. For example, the rainfall intensity correction coefficient is +0.2 for 28 mm / h (moderate rain), +0.15 for the summer flood season, and +0.05 for the duration label correction coefficient at 9 AM (weekday). The weight corresponding to each feature correction coefficient is a weight set based on engineering experience, and its sum must be 1.

[0083] Using historical statistical benchmark thresholds as the center and combining them with the final fluctuation coefficient, the upper and lower limits of the dynamic flow threshold are calculated. That is, the upper limit of the dynamic flow threshold = historical statistical benchmark threshold × (1 + final fluctuation coefficient), and the lower limit of the dynamic flow threshold = historical statistical benchmark threshold × (1 - final fluctuation coefficient). The above upper and lower limits are combined to form the dynamic flow threshold range of the target acquisition node under the current real-time climate conditions, which serves as the basis for determining whether the flow is abnormal in subsequent abnormal state judgment.

[0084] This application's embodiments significantly reduce the probability of misjudgment by screening historical climate and flow data under the same operating conditions. Furthermore, based on the standardization and differentiated weighting of rainfall intensity, season, and time period characteristics, the fluctuation coefficient is precisely scaled and corrected using formulas, enabling the threshold to dynamically adapt to real-time operating condition differences and improving adaptability to complex climate scenarios. Simultaneously, it incorporates physical limit constraints of the pipeline network, determining the minimum normal flow and maximum safe flow through Manning's formula or engineering experience to ensure that the dynamic threshold does not exceed the pipeline's flow capacity and safety boundaries, avoiding safety risks caused by unreasonable threshold settings. The resulting dynamic threshold range, adapted to the current operating conditions and taking into account both data statistical patterns and physical characteristics, provides an accurate, reliable, and safe basis for subsequent judgment of abnormal states in the drainage network.

[0085] Optionally, in one specific embodiment, the flow data of the target acquisition node is reconstructed using theoretical values ​​based on sensor monitoring data, real-time climate data, a pre-set flow reconstruction algorithm, and historical correlation data to obtain theoretically reconstructed data for the target acquisition node, including:

[0086] Based on real-time climate data and historical correlation data, the similarity between real-time climate data and historical climate data for each time period in historical correlation data is calculated, and all historical climate data with similarity reaching a pre-set similarity threshold are selected.

[0087] Based on the selected historical climate data and sensor location data of the target acquisition node, the historical flow data of the target acquisition node corresponding to each historical climate data is determined, and the average historical flow of the target acquisition node is calculated based on all the historical flow data determined for the target acquisition node.

[0088] Based on the pre-set physical baseline traffic and historical average traffic of the target acquisition node, the corresponding physical correction coefficient is obtained. The physical correction coefficient is the ratio of the difference between the physical baseline traffic and the historical average traffic to the physical baseline traffic.

[0089] The traffic data of the target acquisition node is reconstructed based on the physical correction coefficient to obtain the theoretical reconstructed data of the target acquisition node.

[0090] Specifically, historical correlation data is extracted from the pre-established spatiotemporal data base (lake-warehouse integration + ClickHouse), including historical climate data for multiple time periods (rainfall intensity, season, time period, temperature, etc.); the historical climate data for each time period corresponds to the historical flow data (with timestamp, node number, and flow value) of all collection nodes in the urban drainage network.

[0091] Historical data is preprocessed to remove outliers (such as negative flow values, data exceeding the physical range, data with obvious abrupt changes and no rainfall support); an index is created by "time period - climate characteristics - node flow" to facilitate rapid matching.

[0092] Calculate the similarity between real-time and historical climate data and filter historical data under the same operating conditions. For each historical climate data point, extract common feature dimensions (such as rainfall intensity, season, and time period). Use a preset similarity algorithm (such as Euclidean distance, cosine similarity, or weighted Manhattan distance) to calculate the similarity value between the real-time climate and each historical climate data point. Set a similarity threshold (such as 0.85) and filter out all historical climate data with a similarity ≥ the threshold to form a subset of historical climate data under the same operating conditions. A list of historical climate data under the same operating conditions (N data points, N≥30 to ensure statistical validity) is generated.

[0093] Subsequently, the historical flow data of the target acquisition node under the same operating conditions is matched and extracted. This is done by using a subset of historical climate data under the same operating conditions and the sensor location data (node ​​number / latitude and longitude) of the target acquisition node. Based on the time period identifier of each historical climate data under the same operating conditions, the historical flow table of the full node corresponding to that time period is located in the historical associated data. Using the target acquisition node number / location as the key, the historical flow value of that node in that time period is extracted from the historical flow table of the full node. By traversing all historical climate data under the same operating conditions, the historical flow dataset of the target acquisition node under the same operating conditions is collected.

[0094] Statistical calculations are performed on the historical traffic data set of the target acquisition node under the same working conditions to obtain the historical traffic mean. If there are obvious outliers, the step mean or median can be used instead of the arithmetic mean to improve robustness.

[0095] Based on the fundamental attribute data of the target data collection nodes (such as service area, surface type, runoff time, runoff coefficient, etc.), and using conventional urban hydrological models (such as inference formula method, SCS model) or engineering experience, the physical baseline flow rate of the target data collection nodes under no rainfall / baseline conditions is calculated to reflect the base flow / sewage volume of the pipe network. This is the baseline value of the flow rate, and the specific calculation formula is Q. base =φ·F·q0, where φ is the Jinghe River runoff coefficient (determined according to the pre-set underlying surface type), F is the node service area, and q0 is the baseline rainfall intensity or sewage quota. If it cannot be calculated directly, the historical average flow during periods without rainfall is used as the physical baseline flow.

[0096] Based on the physical baseline flow rate and the historical average flow rate under the same operating conditions, a physical correction factor is calculated, where the formula for calculating the physical correction factor is k. phy =(Q base -μ) / Q base k phy This is the physical correction factor, reflecting the degree of deviation of the historical statistical mean from the physical baseline flow. If μ > Q base Then k phy <0 indicates that the historical flow rate is generally higher than the base flow rate. If μ < Q base Then k phy >0 indicates that the historical flow is generally lower than the base flow, for k phy Limit the range, such as to [-0.5, +0.5], to avoid extreme values ​​that could cause reconstruction distortion.

[0097] Based on the real-time traffic monitoring data, physical correction coefficient, and physical baseline traffic of the target acquisition node, the theoretical reconstructed data is calculated using the following formula: Theoretical reconstructed data Q recon =Q base ·(1-k phy ).

[0098] The physical rationality of the theoretically reconstructed data is verified based on the physical constraints of the pipeline network.

[0099] The theoretical error is calculated based on real-time traffic monitoring data and traffic reconstruction data. This theoretical error includes the absolute error Err. abs =|Q real -Q recon |, relative error Err abs =|Q real -Q recon | / Q recon.

[0100] This embodiment employs a data spatiotemporal foundation to efficiently extract and preprocess historical correlation data. It filters samples under similar operating conditions using climate similarity and calculates historical flow averages to ensure the statistical rationality of the reconstructed data. By combining physical baseline flow and correction coefficients, it organically integrates data-driven statistical patterns with the hydrological and physical characteristics of the pipe network, effectively avoiding the problem of pure data models deviating from actual operating conditions. Through limiting the correction coefficient amplitude and verifying physical rationality, the robustness and reliability of the theoretically reconstructed data are further improved, avoiding interference from extreme and unreasonable values. The resulting theoretically reconstructed data can serve as a benchmark for the expected flow under normal operating conditions at the target acquisition node. Combined with absolute and relative error calculations, it accurately reflects the deviation between real-time monitoring data and theoretical values, providing crucial support for distinguishing sensor data distortion from actual pipe network faults and improving the accuracy of abnormal state judgment. Simultaneously, it considers both statistical patterns and physical constraints, making the flow reconstruction more closely aligned with the actual operating characteristics of urban drainage pipe networks.

[0101] Optionally, in one specific embodiment, based on the sensor location data of the target acquisition node and a pre-set spatial topology network, the method includes:

[0102] The spatial topology network data of the urban drainage network is retrieved from the pre-established drainage network GIS database or data spatiotemporal base. This data includes the spatial location (latitude and longitude / node number) of all collected nodes and pipe segments, connection relationships (upstream node → pipe segment → downstream node), network hierarchy and zoning information, forming a structured node-pipe segment topology relationship table.

[0103] By matching the sensor location data (node ​​number / latitude and longitude / pipeline code) of the target acquisition node with the node information in the spatial topology network, the unique identifier and spatial coordinates of the target acquisition node in the topology network are accurately located, and its level and area in the pipeline network are determined.

[0104] Based on the located target acquisition node, the node-pipe segment connection relationship of the spatial topology network is traversed. All directly / indirectly connected upstream acquisition nodes are traced upstream, and all directly / indirectly connected downstream acquisition nodes are traced downstream, forming a subset of the upstream and downstream topology network of the target acquisition node. The pipe segment connection, flow direction relationship and distance / weight information of each upstream and downstream node with the target acquisition node are clarified.

[0105] The extracted upstream and downstream topology networks are verified to ensure that there are no broken links, loops, or isolated nodes, and that all upstream and downstream nodes are valid data acquisition nodes of deployed sensors. If there are any missing or abnormalities, the network topology is automatically supplemented to ensure the reliability of the topology foundation for subsequent traffic comparisons.

[0106] Optionally, in a specific embodiment, based on the sensor monitoring data corresponding to each acquisition node in the upstream and downstream topology network and the sensor monitoring data of the target acquisition node, as well as pre-set historical correlation data and traffic comparison algorithms, the traffic comparison result corresponding to the target acquisition node is obtained, including:

[0107] Based on the sensor monitoring data corresponding to each acquisition node in the upstream and downstream topology network and the sensor monitoring data of the target acquisition node, a first upstream flow ratio and a second downstream flow ratio are obtained. The first upstream flow ratio is the ratio of the flow data of the target acquisition node to the flow data corresponding to the upstream node of the target acquisition node in the upstream and downstream topology network, and the second downstream flow ratio is the ratio of the flow data corresponding to the downstream node of the target acquisition node to the flow data of the target acquisition node in the upstream and downstream topology network.

[0108] Based on real-time climate data and historical correlation data, the similarity between real-time climate data and historical climate data for each time period in historical correlation data is calculated, and all historical climate data with similarity reaching a pre-set similarity threshold are selected.

[0109] Based on the selected historical climate data and sensor location data of the target acquisition node, the historical flow data of the target acquisition node corresponding to each historical climate data is determined, and the average historical flow corresponding to the target acquisition node is calculated based on all the historical flow data determined for the target acquisition node.

[0110] Based on the traffic data of the target collection node and the historical average traffic, the third historical traffic ratio is calculated, where the third historical traffic ratio is the ratio of the traffic data of the target collection node to the historical average traffic.

[0111] The traffic comparison results corresponding to the target acquisition node include the first upstream traffic ratio, the second downstream traffic ratio, and the third historical traffic ratio.

[0112] Specifically, it involves acquiring and verifying sensor monitoring data based on the target acquisition node and the established upstream and downstream topology network (upstream node set and downstream node set), as well as sensor monitoring data of each acquisition node in the upstream and downstream topology network.

[0113] In other words, verify the integrity of upstream and downstream node data: ensure that all upstream / downstream nodes have valid traffic data (non-empty, non-outlier, and within the physical range). Handle missing / outlier data: for short-term missing data, use time-series interpolation or the average of nodes in the same region to fill in the gaps; for long-term missing / invalid data, mark the node as unavailable and remove it from the topology comparison set to avoid affecting the ratio calculation. Align the traffic data of the target acquisition node with the traffic data of all upstream and downstream nodes using the same timestamp (e.g., all using the current 5-minute average) to ensure time-series consistency.

[0114] Obtain the effective upstream node traffic set, the effective downstream node traffic set, and the real-time traffic data of the target collection node.

[0115] Calculate the total upstream flow (considering the confluence characteristics of the pipe network, usually the sum of the flow of all effective upstream nodes (generally only the first-level upstream nodes). Effective upstream nodes are all nodes of the input target collection node. If there are flow splitting / convergence weights of upstream nodes (such as differences in pipe diameter, slope, and confluence area), the weighted sum can be calculated according to the weights). Calculate the first upstream flow ratio based on the total upstream flow and the flow data of the target collection node. The calculation formula is: First upstream flow ratio = Flow data of the target collection node / Total upstream flow, to reflect the proportion of the target collection node flow to the total upstream inflow, reflecting the confluence / loss / overflow situation, which is generally between 0.8 and 1.2.

[0116] The calculation of the downstream total flow is similar to that of the upstream total flow. The second upstream flow ratio is calculated by combining the downstream total flow and the flow data of the target collection node. The second upstream flow ratio = downstream total flow / flow data of the target collection node. This ratio reflects the proportion of the flow of the target collection node relative to the total upstream inflow, and reflects the confluence / loss / overflow situation. It is generally between 0.8 and 1.2.

[0117] Screening historical climate data under the same operating conditions (reusing the similarity logic of dynamic threshold / flow reconstruction) is similar to the screening method in the above embodiments. Historical climate data with similarity ≥ similarity threshold are screened to form a subset of historical climate data under the same operating conditions (ensuring sample size N≥30). The historical climate data under the same operating conditions are deduplicated and abnormal periods (such as equipment maintenance, rainstorm extreme events) are removed to improve sample quality.

[0118] Based on the historical climate subset under the same operating conditions, the sensor location data of the target acquisition node, and the historical flow table of all nodes in the historical correlation data, the historical flow table of the corresponding node for each historical climate under the same operating conditions is located according to the time period identifier of each historical climate under the same operating conditions. The historical flow value of the node for that time period is extracted by using the target acquisition node number / location as the key. All historical climate data under the same operating conditions are traversed to collect the historical flow dataset of the target acquisition node under the same operating conditions. The average historical flow is calculated.

[0119] Based on the real-time traffic and historical average traffic of the target acquisition node, the third historical traffic ratio is calculated using the formula: Third Historical Traffic Ratio = Real-time Traffic Data of Target Acquisition Node / Historical Average Traffic. This ratio reflects the deviation of the current real-time traffic from the historical average level under the same operating conditions, and is generally between 0.7 and 1.3. If the historical average traffic is a very small value, such as approximately 0, the absolute difference is used directly to calculate the third historical traffic ratio.

[0120] The three ratios are integrated into a structured comparison result.

[0121] This embodiment ensures the reliability of basic data, calculates the upstream and downstream flow ratio by combining the characteristics of the pipe network confluence, reuses the historical data screening logic under the same operating conditions and optimizes the historical flow ratio calculation, and integrates the three-dimensional ratios to form a structured result, providing a comprehensive and reliable basis for accurately judging the abnormal state of the drainage pipe network and locating the root cause of the abnormality, which is significantly better than the single-dimensional flow judgment method.

[0122] Optionally, in a specific embodiment, the abnormal state judgment result of the target acquisition node is obtained based on the dynamic threshold of the target acquisition node, the flow comparison result and the theoretical error, including: when it is determined that the flow data of the target acquisition node is abnormal based on the dynamic threshold of the target acquisition node, the flow comparison result and the theoretical error, the location of the problem pipe section in the urban drainage pipe network is determined based on the dynamic threshold of the flow, the flow comparison result and the theoretical error.

[0123] Specifically, based on the target acquisition node's dynamic traffic threshold, traffic comparison results, and theoretical errors, a multi-dimensional fusion decision is made to ultimately obtain the abnormal state judgment result.

[0124] That is, first, three levels of intervals are preset for each indicator: normal, warning, and abnormal (based on engineering experience and historical data training), which serve as the basis for comprehensive judgment:

[0125] In terms of traffic dynamic thresholds, traffic data within the traffic dynamic threshold is considered normal; traffic data exceeding the traffic dynamic threshold by 5% to 10%, or slightly exceeding the limit for 2 to 3 consecutive periods, is considered a warning; and traffic data exceeding the traffic dynamic threshold by more than 10%, or severely exceeding the limit for multiple consecutive periods, is considered abnormal.

[0126] In the theoretical error dimension, a normal error threshold is set, such as 15%. If the theoretical error is less than or equal to the normal error threshold, it is judged as normal; if the normal error threshold is less than or equal to the pre-set warning error threshold (such as 25%), it is judged as a warning; if the theoretical error is greater than the pre-set warning error threshold, or the absolute error simultaneously exceeds the physical low flow rate threshold (such as >5m), it is judged as a warning. 3 If / h), it is judged as abnormal.

[0127] Similarly, the traffic comparison results are judged using pre-set thresholds.

[0128] Subsequently, each indicator is converted into an abnormal state score (0~100 points, the higher the score, the more abnormal the state) or level (normal / warning / abnormal) according to the degree of deviation. For example, the traffic dynamic threshold score is as follows: when the traffic data is within the traffic dynamic threshold, the traffic dynamic threshold score is 100, which is judged as a warning. When the traffic data exceeds the threshold, the traffic dynamic threshold score is 70~80. When the traffic data is seriously out of bounds, that is, judged as an abnormality, the traffic dynamic threshold score is 0~50, based on the proportion exceeding the threshold. Similarly, the theoretical error score and the traffic ratio score are represented in the same way.

[0129] The final exception state is obtained by using a three-layer logic: veto power, weighted synthesis, and exception root cause localization.

[0130] If any of the following conditions are met, it will be directly judged as abnormal: the dynamic flow threshold is seriously exceeded, the theoretical error is seriously too large, the first upstream / second downstream flow ratio is seriously deviated, or the third historical flow ratio is seriously deviated.

[0131] The score is determined based on the total abnormal state score S (the weighted sum of the scores above). When S ≥ 85, it is an abnormal state; when 70 ≤ S < 85, it is a warning; and when S < 70, it is an abnormal state.

[0132] Subsequently, the root cause of the anomaly was located (distinguishing between sensor faults and actual pipeline faults).

[0133] The core of anomaly root cause localization relies on a combination of three dimensions of data features: dynamic flow threshold, theoretical error, and flow comparison results. Combined with the physical laws of water flow conservation in drainage pipe networks and statistical data logic, it accurately distinguishes between sensor faults and actual pipe network faults, while simultaneously pinpointing the specific fault type, providing clear guidance for operation and maintenance. Specifically:

[0134] The core characteristic of sensor failure is that real-time monitoring data deviates from the normal pattern, but the pipeline topology flow balance and theoretical operating condition matching are normal. That is, the abnormality is at the data level, but there is no corresponding fault manifestation at the physical level. Specifically, the theoretical error is significantly larger (e.g., >25%), but the first upstream flow ratio, the second downstream flow ratio, and the third historical flow ratio in the flow comparison results are all within the normal range. The dynamic flow threshold may be exceeded, but the upstream and downstream flow balance is normal. That is, the theoretical reconstructed data is calculated based on historical data under the same operating conditions and the physical basis flow, reflecting the expected value under normal operating conditions. If the real-time flow deviates greatly from the theoretical value, but the upstream and downstream flow conforms to the law of water conservation, it indicates that the real-time data collected by the sensor is distorted, rather than the actual flow of the pipeline is abnormal. Or, the data of a single node is abnormal, while the data of other nodes in the same area and the same topology link are normal. The data fluctuation is irregular (e.g., sudden jumps, fixed values ​​that do not change) and is not related to climatic characteristics such as rainfall intensity and time period. The repeatability of multiple data collections is poor and exceeds the measurement accuracy range of the equipment.

[0135] The core characteristics of actual pipeline faults (blockage / leakage / overflow / rupture) are real-time data anomalies accompanied by topological flow imbalance, and simultaneous deviations in theoretical errors and historical comparison results. In other words, the data anomalies are consistent with the physical fault manifestations. Specific judgment scenarios and rules are as follows:

[0136] Upstream pipe blockage / siltation: First, the upstream flow ratio is too high (>1.3); second, the downstream flow ratio and theoretical error are normal, but the dynamic flow threshold may be too low (real-time flow data is less than the lower limit of the threshold); third, the historical flow is too low, deviating from the historical average under the same operating conditions, indicating that the upstream pipe blockage has led to a reduction in water inflow and a decrease in the total upstream flow, while the target node flow is relatively stable due to the influence of downstream demand, resulting in a high target node flow / upstream total flow ratio, which is consistent with the characteristics of insufficient confluence caused by blockage; or the upstream node flow is generally low, and the target node flow gradually decreases with the degree of upstream blockage; this is more likely to be observed under no-rainfall conditions, and the flow growth rate during rainfall is significantly lower than the historical level under the same operating conditions.

[0137] Downstream pipe blockage / overflow: The second downstream flow ratio is too low (<0.7), while the first upstream flow ratio is normal; the theoretical error is normal, but the dynamic flow threshold may be too high (real-time flow exceeds the upper limit of the threshold); the third historical flow ratio is too high (>1.3), deviating from the historical average under the same operating conditions, indicating that the blockage in the downstream pipe has obstructed the water flow from the target node, reducing the total downstream flow and making the ratio of total downstream flow to target node flow too low. The flow at the target node has accumulated and increased due to poor drainage, even exceeding the safety threshold; or, the water level at the target node has risen synchronously (if there is water level monitoring), and the flow at the downstream node is generally low; in extreme cases, an overflow occurs, and the flow at the target node suddenly drops and then remains at a low level.

[0138] Pipeline leakage / rupture: Both the first upstream flow ratio and the second downstream flow ratio are abnormal (first upstream flow ratio > 1.2 and second downstream flow ratio < 0.8); the theoretical error is normal, but the dynamic flow threshold may be too low (leakage leads to flow loss); the third historical flow ratio is too small (< 0.7), lower than the historical average under the same operating conditions, indicating that pipeline leakage has caused some flow loss. After the upstream water enters the target node, part of it is lost through the leak point, and the downstream water flow is reduced. Even if the first upstream flow ratio is too large (the target flow is a high percentage of the total upstream flow), the second downstream flow ratio is too small (the downstream flow is a low percentage of the target flow), which is consistent with the characteristics of flow loss; or there is water accumulation or road subsidence around the underground pipeline (if there is inspection data); flow loss is related to the burial depth and material of the pipeline, and it is more likely to occur in old pipelines. The leakage is relatively stable, and there are no sudden jumps in the data.

[0139] Severe pipeline network failure (large-scale blockage / rupture): The theoretical error is significantly larger (>30%), and both the first upstream flow ratio and the second downstream flow ratio deviate significantly (first upstream flow ratio >1.4 or <0.6, second downstream flow ratio >1.4 or <0.6); the dynamic flow threshold is severely exceeded, and the third historical flow ratio deviates significantly (<0.5 or >1.5), indicating that the large-scale failure has caused the hydraulic characteristics of the pipeline network to completely deviate from normal operating conditions, which not only disrupts the upstream and downstream flow balance, but also causes the real-time flow to deviate significantly from the theoretical value and historical average value, which is a complex severe failure; or, multiple nodes in the area have abnormal data, and the flow balance of the topology link is completely broken; the data continues to be abnormal after the failure occurs, with no self-healing trend, and manual intervention is required for repair.

[0140] In some scenarios, data anomalies are not due to equipment or pipeline failures and need to be distinguished separately to avoid misjudgment. That is, if the theoretical error is normal, the first upstream flow ratio and the second downstream flow ratio are normal, and only the third historical flow ratio is abnormal, the dynamic flow threshold may be exceeded. If the anomaly is strongly correlated with extreme weather (such as sudden rainstorms or rare low temperatures) or temporary maintenance (such as valve adjustment or partial maintenance), such anomalies are caused by changes in external operating conditions. The pipeline itself is not faulty, the upstream and downstream flow balance and the theoretical operating condition matching degree are normal, and it only deviates from the historical average value under the same operating conditions. The data can return to normal after the operating conditions are restored.

[0141] The final output is a structured result containing four core types of information to support operational decisions: overall anomaly status, detailed assessments of each dimension, anomaly root cause location, and operational recommendations. The overall anomaly status includes the judgment level (warning or anomaly), the total anomaly status score, and the judgment timestamp; the detailed assessments of each dimension include dynamic traffic thresholds (whether the traffic data of the target collection node exceeds the limit), theoretical errors, and traffic comparison results; and the anomaly root cause location (if it is a warning / anomaly) includes the anomaly type, anomaly severity, and scope of impact.

[0142] This embodiment uses a three-level interval as a benchmark, combining quantitative scoring, weighted summarization, and a veto rule to improve the objectivity, accuracy, and efficiency of abnormal state judgment. By combining three-dimensional data features and the law of water flow conservation, it accurately distinguishes fault types, locates root causes, and identifies sudden changes in non-fault operating conditions. It outputs structured results including abnormal state levels and operation and maintenance suggestions, providing a reliable basis for the refined and intelligent operation and maintenance of urban drainage pipe networks, and helping to improve operation and maintenance efficiency and the ability to ensure the safe operation of pipe networks.

[0143] Optionally, in a specific embodiment, a lake-warehouse integrated architecture is adopted, combined with the ClickHouse columnar database, to build a high-throughput, high-concurrency data storage and computing platform; through a fully responsive data link, the platform achieves unified access, cleaning, transformation, and fusion of unstructured data (satellite remote sensing, UAV aerial photography), semi-structured data (IoT sensor time-series data), and structured data (manually collected data, business system data); and a standardized model and spatiotemporal indexing mechanism for multi-source heterogeneous data of the urban drainage system are established to enable rapid data retrieval and correlation analysis.

[0144] Specifically, the entire data is inventoried in three categories: structured, semi-structured, and unstructured, with each category having a clearly defined source, format, and update frequency. Structured data includes pipeline ledgers (node / segment attributes, pipe diameter / slope), manual maintenance records, and pump station operation logs (exported from the business system), updated statically or weekly. Semi-structured data includes IoT sensor time-series data (flow rate, water level, rainfall intensity, with timestamps / location tags) and equipment status data, updated second-level or minute-level. Unstructured data includes satellite remote sensing imagery, drone aerial videos, and pipeline inspection PDF reports, updated daily or on demand.

[0145] Establish access standards and specifications: unify basic formats, timestamp formats, coordinate systems, and data units; clarify key fields, all data must include spatiotemporal tags, and structured data should be supplemented with additional business tags.

[0146] Deploy a fully responsive access channel: a real-time access channel that uses the MQTT protocol to connect to IoT sensors and uses Kafka as a message queue to buffer high-concurrency data; a batch access channel that connects to remote sensing images and PDF reports via FTP / SFTP and synchronizes structured data from business systems via JDBC; and a fault tolerance mechanism that configures data retry, breakpoint resumption, and anomaly alarms (such as triggering an alert if sensor data is interrupted for more than 5 minutes).

[0147] Perform data cleaning, data format conversion, and data validation on the data.

[0148] Integrate multi-source data and deploy it as a lake-warehouse integrated storage system. The data lake (HDFS) stores raw preprocessed data and unstructured large files (such as original remote sensing images and complete videos), preserving the original form of the data and supporting subsequent reuse. The data warehouse (ClickHouse columnar database) stores the fused structured / semi-structured data, partitioned by "business theme" (such as "pipeline monitoring", "pump station operation", "rainfall environment").

[0149] Multi-source data fusion modeling: Establishing standardized data models: Modeling by spatiotemporal dimensions + business dimensions. Specifically, a basic node information model (node ​​number, latitude and longitude, pipe diameter, service area); a real-time monitoring data model (node ​​number, timestamp, flow / water level / rainfall, data status); and a historical correlation data model (node ​​number, timestamp, rainfall intensity, scheduling strategy, operational results). Data correlation: Linking different types of data through node number + timestamp.

[0150] Data is stored in tiers: hot data (real-time data for the past 7 days) is stored in the ClickHouse memory cache and supports high-frequency queries; warm data (data for the past 3 months) is stored on the ClickHouse disk and partitioned by time and space; cold data (data older than 3 months) is archived to the data lake and accessed on demand.

[0151] Construct a spatiotemporal two-level index: Spatial index, using the R-tree algorithm, divides data according to latitude and longitude grids to quickly locate node data within a certain spatial range; Time index, using the B-tree algorithm, is a hierarchical index by year, month, and day to quickly filter data within a certain time period; Composite index, combining the spatial and time indexes, supports spatiotemporal combined queries.

[0152] Provide standardized interfaces to enable upper-layer applications to quickly access data. This includes building data association capabilities, developing standardized service interfaces, and implementing permission and security controls.

[0153] This application embodiment constructs a data platform based on a lake-warehouse integrated architecture and integrates the ClickHouse columnar database to achieve full-domain access, unified governance, and efficient service of multi-source heterogeneous data in urban drainage systems.

[0154] Optionally, in one specific embodiment, the construction process of the digital twin platform is as follows: Figure 4 As shown, it includes:

[0155] Data collection includes: DOM (Digital Orthophoto) data: city-level high-resolution imagery (0.1m~0.5m) for surface texture, roads, and building outlines; DEM (Digital Elevation Model) data: city-level elevation data (1m~5m resolution) for terrain undulation, slope, and runoff analysis; urban element vector data: roads, buildings, green spaces, water systems, administrative divisions; drainage network GIS data (nodes, pipe segments, pipe diameter, slope, burial depth, material, elevation); attributes of structures such as pump stations, gates, inspection wells, and sewage treatment plants; BIM model source data (if existing, no need to remodel): BIM models (Revit / IFC) of drainage structures (pump stations, sedimentation wells, valve wells, box culverts); dynamic monitoring data: volume, water level, rainfall, pump station operating status, gate opening, equipment status, etc. (with timestamp + location).

[0156] Data standardization includes unified coordinates (all data are unified to the same coordinate system), unified elevation (unified elevation datum), unified format, and standardized attributes (unified coding of node numbers, pipe segment numbers, structure IDs, service areas, and topological relationships).

[0157] The DEM data is used to construct a terrain triangulation network to generate a three-dimensional urban terrain surface. Slope, aspect, and runoff analysis are performed on the terrain. Texture mapping is performed on the DOM data and applied to the terrain surface to form a realistic surface texture.

[0158] 3D representation of urban elements: Buildings: 3D building volumes are generated by extrude based on building outline vectors and height attributes (or extracted from DOM / LiDAR); Roads: 3D road networks (including lanes, sidewalks, and green belts) are generated based on road centerlines, widths, and elevations; Water systems / green spaces: 3D water system and green space models are generated based on surface vectors; Administrative regions / grids: 3D boundaries and grid divisions are generated (for subsequent regional monitoring).

[0159] It integrates terrain, buildings, roads, water systems, etc. into a unified scene graph, constructs a spatial index (quadtree / octree), supports multi-scale loading (near-fine, far-rough), and outputs L2 level 3D geographic base.

[0160] Based on GIS pipeline network vectors (nodes / segments), the BIM tool generates 3D pipeline segments by pipe diameter, slope, burial depth, and elevation; generates 3D nodes by node type (manhole, starting point, ending point); and automatically establishes the topological connection relationship between pipeline segments and nodes. Parametric modeling of the pipeline network is performed (pipe diameter, slope, and material can be dynamically modified). Collision detection and elevation verification are performed on the pipeline network (to avoid reverse slope, overhangs, and collisions).

[0161] BIM modeling of drainage structures (pump stations / gates / box culverts, etc.): For pump stations, model detailed components such as inlet pools, outlet pools, pump sets, pipes, valves, and control cabinets; for gates, model the gate body, hoist, guide rails, and foundation; for box culverts / culverts, generate 3D models according to cross-sectional dimensions, length, and slope; for ancillary facilities, model maintenance access ports, ladders, ventilation openings, level gauges, etc.

[0162] Enhanced BIM model attributes by adding business attributes to each BIM component: Pipeline network: pipe diameter, slope, burial depth, elevation, material, service area, and runoff range; Pump station: design flow rate, head, number of pumps, operating status, and control logic; Node: type, number, location, upstream and downstream topology, and monitoring point ID; IFC standard attribute set added to the BIM model.

[0163] The L2 geographic base (GIS) and the L3 drainage BIM model are integrated under the same spatial coordinate system to form a unified digital base. That is, the project coordinates of the BIM model are converted into geographic coordinates, ensuring that the elevation of the BIM model is consistent with the elevation of the DEM data, and the BIM model is spatially corrected to make it completely fit the GIS terrain.

[0164] Constructing LOD (Level of Detail): LOD0, city-level overview; LOD1, district-level; LOD2, road-level; LOD3, facility-level.

[0165] Construct a scene hierarchy tree: city → area → road segment → pipeline network → node / pipe segment → structure → component, supporting on-demand loading and progressive rendering.

[0166] Establish a one-to-one mapping relationship between BIM components, GIS elements, and data acquisition node IDs: for example, BIM pipe segment ID → GIS pipe segment ID → sensor ID → monitoring data; establish an upstream and downstream topology relationship table; and establish a service scope association table.

[0167] The system integrates dynamic monitoring data, achieving spatiotemporal alignment and fusion. Dynamic data access channels include: real-time data (flow, water level, rainfall, pump station status) via MQTT / Kafka / HTTP interfaces; and historical data (batch access from the data spatiotemporal base (ClickHouse + Lake Warehouse)). The data format is standardized as timestamp + monitoring point ID + numerical value + quality identifier. Spatiotemporal alignment involves: time alignment, unifying all dynamic data to the same time base; and spatial alignment, binding data to the 3D model based on the mapping relationship between collection node ID, BIM component ID, and GIS node ID.

[0168] Perform data fusion and quality verification, and define mapping rules for data and visual representation.

[0169] Synchronize 3D visualization rendering, interactive operation, and bidirectional mapping; perform multi-scale, multi-view, and multi-level rendering; set interactive operation methods; and establish mapping and synchronous updates between real and digital scenes.

[0170] This embodiment constructs an L2-level geographic base and an L3-level digital base for drainage facilities, and achieves deep integration of GIS and BIM. Combined with a data platform, it completes the access and verification of dynamic monitoring data, forming a digital twin platform that supports multi-scale visualization and bidirectional mapping synchronization. This provides three-dimensional visualization support for the prediction of abnormal states and operation and maintenance management of urban drainage networks, significantly improving the level of refined and intelligent management of drainage systems and the efficiency of emergency response.

[0171] Furthermore, embodiments of this application provide a training method for a large-scale model for predicting abnormal states of urban drainage pipe networks, including:

[0172] A general-purpose large model is trained based on a pre-set model training set to obtain a trained large-scale model for predicting abnormal states. During the training process, the general-purpose large model is subjected to structured pruning based on the model training set, and the structured-pruned general-purpose large model is fine-tuned based on historical climate datasets and historical sensor monitoring datasets of any urban pipe network in the model training set. The model training set includes historical climate datasets and historical sensor monitoring datasets of multiple different urban pipe networks.

[0173] The general large model is used to implement the above-mentioned method for predicting abnormal states of urban drainage pipe networks.

[0174] Furthermore, the model training set needs to contain two types of data from multiple different urban pipe networks, and the data dimensions must strictly correspond to the input of the anomaly prediction method. The historical climate dataset includes time-series data such as rainfall intensity, season, and time period for each city (covering different working conditions such as light rain, heavy rain, and normal water period); the historical sensor monitoring dataset includes historical flow data and sensor location data of all collection nodes in the drainage pipe network of each city, and needs to be associated with the topological relationship of the corresponding urban drainage pipe network.

[0175] All data in the model training set underwent data preprocessing, which included data cleaning, format standardization, and labeling. Data cleaning involved removing outliers (such as negative flow rates, data exceeding the pipe's physical range, and sudden flow changes without rainfall support) and filling in missing values ​​(short-term missing values ​​were interpolated using time-series interpolation, while long-term missing values ​​were marked as invalid). Format standardization included unifying timestamps (e.g., converting all data to 5-minute averages), coordinate systems (sensor location data were standardized to latitude and longitude), and data units (rainfall intensity was standardized to mm / h, and flow rate to m³). 3 / h); Tag labeling: Add an abnormal status label (normal / warning / abnormal) and a fault type label (such as blockage, leakage, sensor failure) to each data entry. The label generation should be based on the fusion judgment logic of dynamic threshold, theoretical error, and flow comparison results in the attachment (such as if the flow exceeds the dynamic threshold by more than 10% and the theoretical error is greater than 25%, it is labeled as abnormal - pipeline blockage).

[0176] Choose a general-purpose large model that supports time series data processing and spatial topology modeling (such as a Transformer-based time series model, a graph neural network (GNN) model, etc.), which should meet the requirements of "structured pruning and adaptability to high-dimensional time series and spatial data".

[0177] Set pre-training objectives to ensure that the model's learning goals align with the actual prediction task:

[0178] Core Task 1: Learn the mapping relationship between "climate data → dynamic flow threshold", enabling the model to grasp the dynamic impact of rainfall intensity, season, and time period on flow threshold; Core Task 2: Learn the calculation logic of "flow data + climate data → theoretical reconstruction data + theoretical error", understanding the relationship between the physical basic flow of the pipeline network and historical flow; Core Task 3: Learn the pattern of "topology data + flow data → flow comparison results", mastering the flow balance relationship between upstream and downstream nodes; Core Task 4: Learn the fusion judgment logic of "dynamic threshold + theoretical error + flow comparison results → abnormal state", able to output the abnormal state level (normal / warning / abnormal) and the location of the problem pipeline segment.

[0179] The training set is input into a general large model, and training is carried out using the "batch iterative training" method. The number of iterations is set according to the model's convergence (e.g., 50-100 rounds), and the accuracy of abnormal state prediction on the validation set (e.g., accuracy ≥ 85%) is used as the convergence criterion. The AdamW optimizer is used, and the loss function is selected as cross-entropy loss (adapted to abnormal state classification) + MSE loss (adapted to regression outputs such as dynamic threshold of traffic and theoretical error).

[0180] The structured pruning process is performed on the large model to remove redundant parameters and adapt to the real-time monitoring requirements. The core objective is to reduce parameter redundancy and inference latency of the pre-trained model, and ensure that the model can meet the engineering requirements of real-time abnormal state judgment and low computing power consumption of urban drainage pipe network.

[0181] When the accuracy of the pre-trained model in predicting abnormal states on the validation set stabilizes (e.g., accuracy fluctuation ≤ 1% for 5 consecutive rounds) and the parameter size exceeds the engineering deployment threshold (e.g., number of parameters > 100 million), structured pruning is initiated. Redundant layers / parameters unrelated to the prediction of abnormal states in the drainage network are prioritized for removal, while core functional modules are retained. These include the temporal feature extraction layer (processing climate and flow time-series data), the spatial topology modeling layer (processing upstream and downstream relationships of nodes), and the multi-dimensional fusion layer (fusion threshold, error, and comparison results). Redundant layers in general large models adapted to unrelated tasks such as natural language and images are removed, as well as ineffective neurons with parameter weights close to 0.

[0182] The pruning strategy adopts a structured pruning approach that combines "layer pruning + channel pruning": the importance of parameters in each layer is calculated based on the training set (e.g., by judging through the magnitude of parameter gradients); the pruning ratio is set (e.g., removing 30%-50% of redundant parameters) to ensure that the accuracy loss of the model on the validation set after pruning is ≤3%; after pruning, the model is "weight reinitialized" to avoid degradation of local optima.

[0183] After pruning, perform preliminary validation by testing the pruned model on the validation set. The accuracy of abnormal state prediction should be ≥82% (with a loss of ≤3% compared to before pruning); the inference latency should be reduced by ≥40% compared to before pruning (to meet the low latency requirements of real-time monitoring); and the parameter size should be controlled within the acceptable range for engineering deployment (e.g., ≤50 million parameters).

[0184] After structured pruning, the model is fine-tuned to ensure that the pruned model accurately adapts to the specific characteristics of a single city's pipe network (such as the siltation characteristics of an old pipe network and the topographical confluence characteristics of a city) while retaining the general patterns of multiple cities, thereby improving the prediction accuracy of local scenarios.

[0185] Select a complete dataset of the pipeline network of any target city from the model training set (including the city's historical climate data, sensor monitoring data from all collection nodes, and pipeline topology data). The data must include typical abnormal state scenarios of the city (such as local blockage, pipeline leakage, and sensor failure). Use a small learning rate (e.g., 1e-5 to 1e-4) to avoid covering the general rules learned in the pre-training stage; set the number of iterations to 10-20 rounds, and use the validation set accuracy of the city data (e.g., ≥88%) as the convergence criterion; fine-tuning is limited to updating only the parameters of the top fully connected layer and fusion layer of the model, while retaining the core weights of the bottom temporal feature extraction and spatial topology modeling layers (key parameters retained after pruning).

[0186] The fine-tuning task requires key optimizations: localization of dynamic threshold calculation, enabling the model to learn the threshold fluctuation coefficient correction logic of the target city's pipe network under different climates (such as the threshold amplification ratio during summer rainstorms in the city); specific optimization of topology flow comparison, adapting to the characteristics of the target city's pipe network such as pipe diameter distribution and node spacing, and adjusting the judgment of the normal range of upstream and downstream flow ratios; and improving the accuracy of abnormal state root cause location, optimizing the ability to distinguish between "sensor failure" and "actual pipe network failure" (such as the common error characteristics of sensors in a certain area of ​​the city).

[0187] Finally, the performance of the fine-tuned model was verified, and the model was deployed and integrated.

[0188] This embodiment uses a process of "pre-training with multi-city data → structured pruning for efficiency improvement → fine-tuning and adaptation with single-city data" to generate a large-scale abnormal state prediction model that can cover general scenarios in multiple cities, adapt to the specificities of a single city, and meet the needs of real-time monitoring and precise positioning.

[0189] In addition, this application provides an abnormal state prediction system for urban drainage pipe networks, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the above-mentioned abnormal state prediction method for urban drainage pipe networks.

[0190] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0191] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0192] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0193] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0194] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for predicting abnormal states in urban drainage pipe networks, characterized in that, include: The system acquires real-time climate data and sensor monitoring data of target collection nodes in the urban drainage network. Based on the real-time climate data, the sensor monitoring data of target collection nodes, and pre-set historical correlation data and dynamic threshold algorithm, it obtains the dynamic threshold of flow for the target collection nodes. The sensor monitoring data includes sensor location data and flow data. Based on sensor monitoring data, real-time climate data, pre-set flow reconstruction algorithms, and historical correlation data, the flow data of the target acquisition node is reconstructed using theoretical values ​​to obtain the theoretical reconstructed data of the target acquisition node, and the theoretical error between the flow data of the target acquisition node and the theoretical reconstructed data is calculated. Based on the sensor location data of the target acquisition node and the pre-set spatial topology network, the upstream and downstream topology networks of the target acquisition node are determined, and the sensor monitoring data corresponding to each acquisition node in the upstream and downstream topology networks are obtained. Based on the sensor monitoring data of each acquisition node in the upstream and downstream topology network and the sensor monitoring data of the target acquisition node, as well as the pre-set historical correlation data and traffic comparison algorithm, the traffic comparison result corresponding to the target acquisition node is obtained. Based on the target acquisition node's dynamic traffic threshold, traffic comparison results, and theoretical errors, the abnormal status judgment result of the target acquisition node is obtained.

2. The method for predicting abnormal states of urban drainage pipe networks according to claim 1, characterized in that, The historical correlation data includes historical climate data for multiple time periods, and the historical climate data for each time period includes historical flow data for all collection nodes in the urban drainage network. Based on real-time climate data and sensor monitoring data from the target acquisition nodes, as well as pre-set historical correlation data and dynamic threshold algorithms, the dynamic threshold of traffic flow for the target acquisition nodes is obtained, including: Based on real-time climate data and historical correlation data, the similarity between real-time climate data and historical climate data for each time period in historical correlation data is calculated, and all historical climate data with similarity reaching a pre-set similarity threshold are selected. Based on the selected historical climate data and sensor location data of the target acquisition node, the historical flow data of the target acquisition node corresponding to each historical climate data is determined, and based on all the historical flow data determined by the target acquisition node, the historical statistical benchmark threshold and historical threshold fluctuation coefficient are obtained. Based on historical statistical benchmark thresholds and historical threshold fluctuation coefficients, the dynamic threshold of traffic for the target acquisition node is obtained.

3. The method for predicting abnormal states of urban drainage pipe networks according to claim 2, characterized in that, Based on all historical traffic data determined by the target acquisition nodes, historical statistical baseline thresholds and historical threshold fluctuation coefficients are obtained, including: Based on all historical traffic data determined by the target collection node, calculate the mean historical traffic and the standard deviation of historical traffic for the city corresponding to the target collection node. Based on the historical average traffic flow and the standard deviation of historical traffic flow in the city corresponding to the target data collection node, the historical statistical benchmark threshold and the historical threshold fluctuation coefficient are obtained.

4. The method for predicting abnormal states of urban drainage pipe networks according to claim 2, characterized in that, The real-time climate data includes rainfall intensity data labeled with seasonal and rainfall duration tags; Based on historical statistical benchmark thresholds and historical threshold fluctuation coefficients, the dynamic thresholds of traffic for the target data collection node are obtained, including: The rainfall intensity data, seasonal labels, and rainfall duration labels were standardized to obtain the rainfall intensity correction coefficient, seasonal correction coefficient, and rainfall duration correction coefficient. Based on the standardized rainfall intensity correction coefficient, seasonal correction coefficient, and rainfall duration correction coefficient, as well as the pre-set correction weights, the historical threshold fluctuation coefficient is scaled and corrected to obtain the final fluctuation coefficient. Based on the final fluctuation coefficient and historical statistical benchmark threshold, the dynamic threshold of traffic at the target acquisition node is obtained.

5. The method for predicting abnormal states of urban drainage pipe networks according to claim 1, characterized in that, Based on sensor monitoring data, real-time climate data, and pre-set flow reconstruction algorithms and historical correlation data, the flow data of the target acquisition node is reconstructed using theoretical values ​​to obtain the theoretically reconstructed data of the target acquisition node, including: Based on real-time climate data and historical correlation data, the similarity between real-time climate data and historical climate data for each time period in historical correlation data is calculated, and all historical climate data with similarity reaching a pre-set similarity threshold are selected. Based on the selected historical climate data and sensor location data of the target acquisition node, the historical flow data of the target acquisition node corresponding to each historical climate data is determined, and the average historical flow corresponding to the target acquisition node is calculated based on all the historical flow data determined for the target acquisition node. Based on the pre-set physical baseline traffic and historical average traffic of the target acquisition node, the corresponding physical correction coefficient is obtained. The physical correction coefficient is the ratio of the difference between the physical baseline traffic and the historical average traffic to the physical baseline traffic. The traffic data of the target acquisition node is reconstructed based on the physical correction coefficient to obtain the theoretical reconstructed data of the target acquisition node.

6. The method for predicting abnormal states of urban drainage pipe networks according to claim 1, characterized in that, Based on the sensor monitoring data corresponding to each acquisition node in the upstream and downstream topology network and the sensor monitoring data of the target acquisition node, as well as the pre-set historical correlation data and traffic comparison algorithm, the traffic comparison result corresponding to the target acquisition node is obtained, including: Based on the sensor monitoring data corresponding to each acquisition node in the upstream and downstream topology network and the sensor monitoring data of the target acquisition node, a first upstream flow ratio and a second downstream flow ratio are obtained. The first upstream flow ratio is the ratio of the flow data of the target acquisition node to the flow data corresponding to the upstream node of the target acquisition node in the upstream and downstream topology network, and the second downstream flow ratio is the ratio of the flow data corresponding to the downstream node of the target acquisition node to the flow data of the target acquisition node in the upstream and downstream topology network. Based on real-time climate data and historical correlation data, the similarity between real-time climate data and historical climate data for each time period in historical correlation data is calculated, and all historical climate data with similarity reaching a pre-set similarity threshold are selected. Based on the sensor location data of the target acquisition node of historical climate data and each selected historical climate data, the historical flow data of the target acquisition node corresponding to each historical climate data is determined, and the average historical flow corresponding to the target acquisition node is calculated based on all the historical flow data determined by the target acquisition node. Based on the traffic data of the target collection node and the historical average traffic, the third historical traffic ratio is calculated, where the third historical traffic ratio is the ratio of the traffic data of the target collection node to the historical average traffic. The traffic comparison results corresponding to the target acquisition node include the first upstream traffic ratio, the second downstream traffic ratio, and the third historical traffic ratio.

7. The method for predicting abnormal states of urban drainage pipe networks according to claim 4, characterized in that, Historical correlation data, spatial topology networks, real-time climate data, and sensor monitoring data of target acquisition nodes in urban drainage networks are all obtained through a pre-established data platform; The data platform is built on a lake-warehouse integrated architecture, which includes a data lake and a data warehouse.

8. The method for predicting abnormal states of urban drainage pipe networks according to claim 7, characterized in that, The method further includes: Based on the pre-set DOM data, DEM data, urban element vector data and real geographic data of the urban drainage network, an L2 level urban three-dimensional geographic base is constructed. Based on the pre-set BIM model and the pre-established urban drainage network model, construct the digital base plate of L3 level drainage facilities; A digital twin platform is constructed based on the L2-level urban 3D geographic base map, the L3-level drainage facility digital base map, and the data platform.

9. The method for predicting abnormal states of urban drainage pipe networks according to claim 1, characterized in that, Based on the target acquisition node's dynamic traffic threshold, traffic comparison results, and theoretical errors, the defect status judgment result of the target acquisition node is obtained, including: Based on the dynamic threshold of the target acquisition node, the flow comparison results, and the theoretical error, when the flow data of the target acquisition node is found to be abnormal, the location and abnormality type of the problematic pipe section in the urban drainage pipe network are determined based on the dynamic threshold of the flow, the flow comparison results, the theoretical error, and the pre-set spatial topology network.

10. A training method for a large-scale model for predicting abnormal states of urban drainage pipe networks, characterized in that, include: A general-purpose large model is trained based on a pre-set model training set to obtain a trained defect prediction large model. During the training process, the general-purpose large model is subjected to structured pruning based on the model training set, and the structured-pruned general-purpose large model is fine-tuned based on historical climate datasets and historical sensor monitoring datasets of any urban pipe network in the model training set. The model training set includes historical climate datasets and historical sensor monitoring datasets of multiple different urban pipe networks. The general large model is used to implement the abnormal state prediction method for urban drainage pipe networks as described in any one of claims 1 to 9.

Images