A multi-agent cooperative pipe network fault intelligent disposal method

Abstract

This invention relates to a multi-agent collaborative intelligent fault handling method for pipeline networks, comprising the following steps: S1: Acquiring relevant pipeline network data and preprocessing it to obtain a joint dataset D. t Snapshot S of the current network operating conditions t S2: The monitoring and early warning agent performs anomaly detection on each DMA and performs multi-source consistency verification to obtain event object E0; S3: Based on event object E0, the diagnostic and location agent performs anomaly classification under actual operating conditions and provides classification probabilities to obtain the updated event E1; S4: The diagnostic and location agent performs constraint propagation on the topology map to obtain a set of candidate pipe segments; S5: Based on the set of candidate pipe segments, for each candidate location, the theoretical response when an anomaly occurs at that location is calculated to obtain the final set of abnormal pipe segments; S6: Based on the final set of abnormal pipe segments, the impact assessment agent calculates for each abnormal pipe segment separately and outputs the event level based on the calculation results. This invention effectively improves the efficiency of daily refined management and fault handling of pipeline networks.

Description

Technical Field

[0001] This invention relates to the field of pipeline fault handling, and in particular to an intelligent method for handling pipeline faults through multi-agent collaboration. Background Technology

[0002] Urban pipe networks (water supply, gas, heating, drainage, etc.) are typically characterized by large coverage areas, complex topologies, highly variable operating conditions, and heterogeneous equipment. During operation, they are susceptible to factors such as pipe aging, damage from third-party construction, valve malfunctions / maloperation, pump station or pressure regulating equipment failures, sudden load changes, and extreme weather, leading to events such as pipe bursts, leaks, abnormal pressure, insufficient supply, and water quality (or concentration / temperature) exceeding limits. Once a failure occurs, the impact often spreads across regions, has significant chain reactions, and requires timely response, potentially causing large-scale supply outages, secondary safety risks, and fluctuations in public services.

[0003] Current pipeline fault handling relies heavily on SCADA alarms, manual experience-based judgment, and on-site investigation, which presents several problems: First, data silos and spatiotemporal inconsistencies exist. Information from SCADA / IoT, GIS, work orders, and complaints is difficult to integrate quickly within the same semantic and topological framework, resulting in incomplete judgment. Second, false alarms and missed alarms coexist. Sensor drift, communication jitter, and single-point anomalies can easily trigger ineffective handling, while real faults may only manifest as weak anomalies in the early stages and be overlooked. Third, location is difficult and highly uncertain. Relying on a single measuring point or manual investigation usually takes a long time and it is difficult to provide a confident ranking of candidate locations. Fourth, handling solutions lack quantitative evaluation. Operations such as valve isolation, pump station switching, and zoned scheduling lack rapid and verifiable simulation and comparison of the impact range, critical user protection, and secondary risks (such as low-pressure backflow / gas accumulation / heat network imbalance), which can easily lead to over-isolation or under-handling. Fifth, the collaborative chain is long, involving multiple departments such as monitoring, dispatching, customer service, emergency repair, and safety supervision. Information transmission and decision-making approval at multiple stages can easily cause delays, and there is insufficient record-keeping and post-mortem learning throughout the process. . Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide a multi-agent collaborative intelligent fault handling method for pipeline networks, effectively improving the efficiency of daily refined management and fault handling of pipeline networks.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A multi-agent collaborative intelligent fault handling method for pipeline networks includes the following steps: S1: Obtain relevant pipeline network data and preprocess it to obtain the joint dataset D. t Snapshot S of the current network operating conditions t ; S2: Based on the joint dataset D tSnapshot S of the current network operating conditions t The monitoring and early warning agent performs anomaly detection on each DMA and performs multi-source consistency verification. If the anomaly reaches the triggering condition, an event is created and the key data fragments of the trigger window are frozen for traceable analysis, resulting in event object E0. S3: Based on event object E0, the diagnostic and location agent performs real-world anomaly classification and provides the classification probability to obtain the updated event E1; S4: Based on the updated event E1, the diagnostic localization agent performs constraint propagation on the topology map, taking the strongest abnormal measurement point as the center and expanding along the upstream and downstream reachable paths. Combined with the valve boundary, the search space is limited to the set of pipe segments that can affect the measurement point. Then, the inspection geographical clues are integrated to form spatial priors and obtain the candidate pipe segment set. S5: Based on the candidate pipe segment set, for each candidate location, calculate the theoretical response when an anomaly occurs at that location, match it with the actual residual, and update the probability using a Bayesian approach combined with the failure prior; at the same time, perform physical consistency filtering to obtain the final set of abnormal pipe segments; S6: Based on the final set of abnormal pipe segments, the impact assessment agent calculates the impact level for each abnormal pipe segment and outputs the event level based on the calculation results.

[0006] Further preprocessing includes data cleaning, unit and dimension unification, time synchronization, spatial mapping, and topology verification; simultaneously, health and data quality labels are calculated for each measurement point, and the original values, corrected values, reasons for correction, quality annotations, clock uncertainty, and mapping confidence metadata are all retained; resulting in the joint dataset D. t With network operating condition snapshot S t The joint dataset D t This is a collection of multiple tables aligned to a unified timeline, including real-time quantity time series, event and work order timelines, equipment and asset dimension tables, GIS topology maps, DMA boundary and valve state sequences, and quality and health labels; the network condition snapshot S t This includes supply and demand balance indicators for each DMA, pressure distribution at key nodes, main flow direction, pump operation combinations, boundary valve connectivity patterns, anomaly candidate lists and their confidence levels.

[0007] Furthermore, the monitoring and early warning agent performs stratified detection on each DMA, including pressure anomalies, flow anomalies, flow-pressure relationship anomalies, water quality and concentration anomalies, and equipment condition anomalies, specifically as follows: Pressure anomalies include sudden pressure drops or rises, sustained low pressure, pressure oscillations, and deviations from historical baselines, combined with the time-period characteristics and seasonal patterns of nighttime low pressure and daytime peak pressure; Flow anomalies include nighttime underflow rises, daytime water consumption curve anomalies, DMA inflow-outflow imbalances, and short-term shock flows; Flow-pressure relationship anomalies include pressure drops without flow increases, flow increases with excessive pressure drops, deviations from hydraulic model predictions, and abnormal physical inconsistencies in pump pressure characteristic curves; Water quality and concentration anomalies include excessive or insufficient residual chlorine, abnormal turbidity, pH shifts, and pipeline pollution indicators; Equipment condition anomalies include frequent pump start-stops, abnormal valve position jumps, abnormal electrical parameters, and prolonged sensor jamming; The detection algorithm of the monitoring and early warning agent uses statistical thresholds and physical constraints for checks, and weights the detection results for reliability based on the health status of the measuring points and data quality labels.

[0008] Furthermore, multi-source consistency verification includes priority for multi-point linkage within the same partition, priority for suspecting sensors due to single-point anomalies, and cross-validation of physical consistency, as follows: Prioritizing multi-point linkage within the same partition means that when multiple measurement points within a DMA simultaneously exhibit anomalies in the same direction, and the magnitudes of these anomalies have spatial correlation, they are preferentially identified as real-world operating events; Prioritizing single-point anomalies means that when only a single measurement point is abnormal while neighboring measurement points are stable, and the data quality label of that measurement point shows a decline in health, a risk of drift, or communication anomalies, it is preferentially suspected to be a sensor failure; Cross-validation of physical consistency uses hydraulic models or empirical constraints to verify the physical rationality of the anomaly; After the anomaly passes consistency verification and the confidence level reaches a preset trigger threshold, the monitoring and early warning agent formally creates an event object E0 and freezes key data fragments within the trigger window to ensure the traceability of subsequent analysis. Simultaneously, the agent freezes and packages the original data, corrected data, quality labels, topology snapshots, and device status of the trigger window, writes them to the event storage, and updates E0 to the shared state pool.

[0009] Furthermore, based on event object E0, the diagnostic and location agent performs real-world anomaly classification and provides classification probabilities, resulting in the updated event E1, as follows: The diagnostic localization agent reads the trigger time t0, trigger DMA, trigger measurement point, and freeze window W from E0, and from D... t With S tThe system retrieves pressure, flow, valve position, pump operating conditions, and water quality sequences from multiple points within the same partition of the window. It also incorporates the data quality labels generated by S1 and the health status of the measurement points as weights to construct categorical features. Events are mapped to several real-world operating condition categories, and probabilities are output. The diagnostic and localization agent first uses a multi-class softmax model to provide statistical probabilities for the feature vectors. ; Where c represents the candidate category; p ml (c|x) represents the statistical classification probability of the machine learning model ml classifying the event as category c when the feature is x; x is the feature vector; θ c Here is the model parameter vector for category c; T denotes transpose; exp(·) represents the exponential function used to convert the scores to positive values; ∑ c′ This represents the summation over all candidate categories c′; c′ is the category dummy variable used for the summation. Then, a physical consistency score s is introduced. c Posterior calibration is performed for ∈[0,1], and finally fusion is performed using the calibration formula: ; Where λ represents the strength of the influence of physical consistency on the result; p(c|E0) is the final posterior probability of class c given event object E0; E0 is the initial event object; s c λ represents the physical consistency score for category c; λ represents the physical consistency calibration strength. Finally, the fractal result is written back to the formation update event E1.

[0010] Furthermore, based on the updated event E1, the diagnostic localization agent performs constraint propagation on the topology map, expanding along upstream and downstream reachable paths centered on the strongest anomaly measurement point. Combined with valve boundaries, the search space is limited to the set of pipe segments that can affect that measurement point. Then, by integrating inspection geographical clues to form spatial priors, a candidate pipe segment set is obtained, as follows: The diagnostic and localization agent reads update event E1 and selects the node corresponding to the strongest anomaly detection point on the pipeline topology graph G=(V,E) as the propagation source point v. Subsequently, reachable paths are expanded along both upstream and downstream directions. During expansion, real-time valve opening / closing status, DMA boundary valves, and partition connectivity are introduced as hard constraints: propagation stops upon encountering a closed valve, an isolated boundary, or a disconnected partition; if the valve position is uncertain, connected branches are retained with topological reliability as the weight, but their priority is reduced, which will affect v. The network portion is cut out from the entire network to form a controlled search subgraph; Within the controlled search subgraph, the diagnostic localization agent uses the abnormal patterns given by E1 as the propagation rules. During the propagation process, an interpretability score is accumulated for each edge. The topological distance of the source point, the drivability of valves on the path, the consistency with the time series response of multiple measurement points, and the support of the water balance residual for the branch are comprehensively considered through weighted calculation. Finally, the set of pipe segments with the highest score is output as the candidate pipe segment set, and the evidence chain of each candidate is retained.

[0011] Furthermore, the spatial priors are formed by integrating geographical clues from inspections, as detailed below: Based on the candidate pipeline segment set, spatial priors are formed by integrating inspection and geographical clues, converging into a high-value candidate set, and the diagnostic location of the Agen is determined from D. t Extract inspection records, work orders, complaint hotspots, historical leaks, road construction information, and acoustic alarm location geographic clues adjacent to the event time window, map them onto the GIS space, and project them onto nodes or pipe segments of the topology map; assign preset high spatial prior weights to locations close to the clues; then fuse and sort the topology propagation scores and spatial priors to obtain the final candidate pipe segment set.

[0012] Furthermore, based on the candidate pipe segment set, for each candidate location, the theoretical response when an anomaly occurs at that location is calculated, matched with the actual residual, and the probability is updated using a Bayesian approach combined with failure priors; simultaneously, physical consistency filtering is performed to obtain the final set of anomalous pipe segments, as follows: Based on the final candidate pipe segment set E c ={e j} and the frozen evidence window W for event E1, for each candidate pipe segment location e j Assume a specific anomaly mechanism and calculate the theoretical response under the current topology and boundary valve states. That is, the expected direction, time delay and magnitude of the impact of the anomaly on each observation point i; For each measurement point i, first calculate the actual residual r. i (t): ; Where yi(t) is the actual observed value of measuring point i at time t; Let be the predicted value of measurement point i at time t; Then compare it with candidate position e j Theoretical response Align and calculate matching errors, such as weighted squared error or correlation:

[0013] Where j is the candidate anomaly location index; W is the set of event freeze time windows; w iLet i be the mass weight of measurement point i; To start from candidate position e j The theoretical propagation delay to measurement point i; Let be the variance scale of the residual noise at measurement point i; It is a positive number; err j For candidate e j The overall fitting error; Construct the likelihood from the error: ; in, For candidate position e j Under the premise that the likelihood of observing event evidence E1 is given; ∝ indicates proportionality; exp(·) is an exponential function; Simultaneously, failure priors are introduced. The results are obtained from the asset ledger and historical work orders; the posterior of the candidate pipeline segment is updated using Bayes' theorem. ; Wherein, P(e j |E1) is an anomaly located in candidate e after given event evidence E1. j The posterior probability; If E1 gives the classification probability P(c|E1) of the abnormal mechanism category c, then a mixture likelihood can be performed: This links positioning and fractal analysis and explicitly represents the sources of uncertainty. Where P(c|E1) is the posterior probability of the fractal of event E1; L(E1|e j c) represents the location e in the abnormal position. j And given mechanism c, the conditional likelihood of evidence E1; ∑ c To perform a weighted summation over all mechanism categories; Based on the posterior ranking, physical consistency constraint checks are performed on high-probability candidates to eliminate positions that are statistically matched but physically impossible; after consistency filtering, the final set of abnormal pipe segments is output.

[0014] Furthermore, based on the final set of abnormal pipe segments, the impact assessment agent calculates the impact for each abnormal pipe segment separately, as follows: Impact assessment agent targets each e j Select a disturbance mechanism consistent with the E1 classification and calculate the hydraulic response caused by the anomaly: ; Where Hj(t) is the water head at that location, and C d,j A j Equivalent leak parameters; For position e j Assume the instantaneous leakage flow rate at the time of leakage; C d,j A is the outflow coefficient; j The area of ​​the equivalent leakage orifice is given by ; g is the acceleration due to gravity. The pressure at each node was then obtained through hydraulic simulation or sensitivity approximation. With supply capacity Based on this, the scope of impact and gap are calculated: set of nodes affected by low pressure. ; in, For time t, when the anomaly occurs at time e j Assume a low-voltage node set; v is the network node index; p min This is the lower limit threshold for service pressure. This means taking the union of all times within the window to obtain the set of nodes that were low-pressure during the window period; Water supply shortfall: ; in, Let d be the unmet demand of node v at time t; max(0,·) ensures the gap is non-negative; v (t) represents the demand of node v at time t; The cumulative gap volume within the window is summed over all nodes and integrated over time. Number of users affected: ; in, For the exception located at e j Under the assumption that the total number of users affected by the low voltage is n v The user weight corresponding to node v; The scale of key users affected; A set of key user nodes; Weighting for key users; This represents the set of nodes that simultaneously meet the criteria of being affected by low voltage and belonging to critical users. Calculate secondary risk proxy: ; Wherein, ρ(e j ) is the position e j The scene weight coefficient.

[0015] Furthermore, the event level is output based on the calculation results, as follows: First, normalize the key impact indicators of each pipe section to [0,1]: ; ; ; ; in, is the normalized dimensionless index; min(1,·) is the cutoff function, where the part exceeding 1 is counted as 1; V0 represents the cumulative water supply gap volume caused by this pipeline section; V0 is the calibration scale for the water supply gap volume. N0 represents the number of users affected by this pipe section; N0 is the calibrated scale for the number of users affected. C0 represents the scale of key users affected by this pipeline segment; C0 is the calibrated measure of the impact on key users. This represents the secondary risk corresponding to this pipe section; R0 is the calibration scale for the secondary risk index. Formation of comprehensive impact at the pipe section level: ; Where α, β, γ, δ are the business weight coefficients for each dimension; Then, by combining the posterior probability of the location, the segment-level risk score R is obtained. j : R j =P(e j |E1)·I j ; Overall event risk is aggregated using expected value: ; Among them, R j Risk classification for the j-th abnormal pipe segment; R max For worst-case scenario risk; R exp For cumulative risk; Select R∈{R max ,R exp Mapped to event level L: ; Where τ1, τ2, τ3 are the graded thresholds; Finally, output the event report, including each segment e j The impact details, ranking, overall level L, list of key users, and recommended handling priorities.

[0016] The present invention has the following beneficial effects: 1. This invention organizes data around events, starting from the joint dataset D. t With working condition snapshot S t Starting from this point, we reduce false alarms through multi-source consistency verification and freeze key data fragments to form event objects E0 when triggered, so as to avoid repeated conclusions due to operating condition drift. 2. This invention integrates mechanistic constraints and data-driven approaches. Topological constraint propagation and valve boundary pruning significantly compress the search space in S4, reducing the instability and computational cost caused by blind search across the entire network. On this basis, S5 constructs the likelihood using theoretical response-actual residual matching, and uses a Bayesian approach to integrate asset failure priors and inspection geographical clues to obtain an interpretable posterior probability ranking. Then, physical consistency filtering is superimposed to eliminate impossible solutions. 3. This invention directly transforms the positioning results into quantitative indicators required for handling, and outputs the event level and priority accordingly, supporting valve isolation, scheduling optimization and notification range determination. The division of labor among multiple agents is clear and the interfaces are well-defined, facilitating modular replacement and incremental iteration. At the same time, priors and weights can be continuously calibrated through historical event accumulation, enabling rapid migration and large-scale deployment across regions and pipeline networks. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: refer to Figure 1 In this embodiment, a multi-agent collaborative intelligent fault handling method for pipeline networks is provided, including the following steps: S1: Acquire pipeline network-related data, including real-time SCADA data (pressure, flow rate, valve position, pump status, etc.), IoT and video / acoustic data, inspection and work order history, GIS topology and asset ledger, zonal DMA and boundary valve status; and preprocess the data to obtain the joint dataset D. t Snapshot S of the current network operating conditions t ; S2: Based on the joint dataset D t Snapshot S of the current network operating conditions t The monitoring and early warning agent performs anomaly detection on each DMA (such as sudden pressure drop, nighttime undercurrent rise, flow-pressure inconsistency, concentration / water quality exceeding limits, etc.) and performs multi-source consistency verification (prioritizing multi-point linkage in the same zone, and prioritizing single-point anomalies to suspect sensors); if the anomaly reaches the trigger condition, an event is created and the key data segments of the trigger window are frozen for traceable analysis, resulting in event object E0 (including trigger time, trigger measurement point, anomaly characteristics, initial confidence level, and data quality overview). S3: Based on event object E0, the diagnostic and location agent performs real-world abnormality classification and provides classification probabilities to obtain the updated event E1, which includes the probability of each fault type and sensor labels. S4: Based on the updated event E1, the diagnostic localization agent performs constraint propagation on the topology map, taking the strongest abnormal measurement point as the center and expanding along the upstream and downstream reachable paths. Combined with the valve boundary, the search space is limited to the set of pipe segments that can affect the measurement point. Then, the inspection geographical clues are integrated to form spatial priors and obtain the candidate pipe segment set. S5: Based on the candidate pipe segment set, for each candidate location, calculate the theoretical response when an anomaly occurs at that location, match it with the actual residual, and update the probability using a Bayesian approach combined with the failure prior; at the same time, perform physical consistency filtering to obtain the final set of abnormal pipe segments; S6: Based on the final set of abnormal pipe segments, the impact assessment agent calculates the impact level for each abnormal pipe segment and outputs the event level based on the calculation results.

[0019] In this embodiment, preprocessing includes data cleaning (missing data identification and hierarchical imputation, glitch / jump point suppression, jamming and saturation detection, drift risk assessment), unit and dimension unification (pressure / flow / valve position, etc., are unified to standard units and signs and directions are verified), time synchronization (NTP / offset estimation, arrival delay compensation, unified time raster resampling and aggregation rules), spatial mapping (binding measurement point / equipment IDs to GIS assets, outputting candidate mappings and confidence levels when necessary), and topology verification (valve opening / closing status and connectivity consistency checks, DMA boundary constraint checks, and physical consistency checks with flow and pressure responses). Simultaneously, health and data quality labels (sample-level / window-level / measurement point-level) are calculated for each measurement point, and metadata such as original values, corrected values, correction reasons, quality labels, clock uncertainty, and mapping confidence levels are retained; resulting in the joint dataset D. t With network operating condition snapshot S t The joint dataset D t This is a collection of multiple tables aligned to a unified timeline, including real-time quantity time series, event and work order timelines, equipment and asset dimension tables, GIS topology maps, DMA boundary and valve state sequences, and quality and health labels; the network condition snapshot S t This includes supply and demand balance indicators for each DMA, pressure distribution at key nodes, main flow direction, pump operation combinations, boundary valve connectivity patterns, anomaly candidate lists and their confidence levels.

[0020] In this embodiment, the monitoring and early warning agent performs tiered detection on each DMA, including pressure anomalies, flow anomalies, flow-pressure relationship anomalies, water quality and concentration anomalies, and equipment condition anomalies, specifically as follows: Pressure anomalies include sudden drops or rises in pressure, sustained low pressure, pressure oscillations, and deviations from historical baselines, combined with the time-period characteristics and seasonal patterns of nighttime low pressure and daytime peak pressure; Flow anomalies include nighttime underflow increases (which may indicate increased leakage), abnormal daytime water usage curves, and DMA anomalies. Imbalance between influent and effluent, short-term shock flow; abnormal flow-pressure relationship including pressure drop but no increase in flow, increase in flow but excessive pressure drop, deviation from hydraulic model prediction, abnormal physical inconsistency of pump pressure characteristic curve; abnormal water quality and concentration including excessive or insufficient residual chlorine, abnormal turbidity, pH deviation, pipeline pollution indication; abnormal equipment operating conditions including frequent pump start-stop, abnormal valve position jump, abnormal electrical parameters, and long-term sensor jamming; the detection algorithm of the monitoring and early warning agent adopts statistical threshold (based on historical quantiles) and physical constraint checks (hydraulic balance, connectivity inference), and weights the detection results with credibility based on the health status of the measuring point and data quality labels.

[0021] In this embodiment, multi-source consistency verification includes priority for multi-point linkage within the same partition, priority for suspecting sensors due to single-point anomalies, and cross-validation of physical consistency, as follows: Prioritizing multi-point linkage within the same partition means that when multiple measuring points within a DMA simultaneously exhibit anomalies in the same direction (such as synchronous pressure decreases at multiple points or changes in inflow and outflow in the same direction), and the magnitude of the changes has spatial correlation, it is preferentially judged as a real operating condition event; Prioritizing single-point anomalies means that when only a single measuring point is abnormal while neighboring measuring points are stable, and the data quality label of that measuring point shows a decline in health, a risk of drift, or communication anomalies, it is preferentially suspected to be a sensor failure; Cross-validation of physical consistency uses hydraulic models or empirical constraints (such as pressure-flow relationship, upstream and downstream cascading response, pump pressure characteristics) to verify the physical rationality of the anomaly, eliminating "false anomalies" that obviously violate physical laws; Once the anomaly passes the consistency check and the confidence level reaches the preset trigger threshold (usually a threshold that distinguishes different event types and impact levels), the monitoring and early warning agent formally creates the event object E0 and freezes the key data fragments within the trigger window to ensure the traceability of subsequent analysis. E0 includes the following elements: (1) Trigger time (time of first anomaly detection, time of reaching the trigger threshold, and data window range); (2) Trigger measurement points (main anomaly measurement points, related measurement points participating in the consistency check, and the health and confidence of each measurement point); (3) Anomaly characteristics (anomaly type, deviation magnitude, trend of change, duration, and spatial distribution pattern); (4) Initial confidence level (a comprehensive score based on detection intensity, consistency check results, and historical false alarm rate); (5) Data quality overview (data integrity, correction status, topology confidence, and clock synchronization status within the trigger window). At the same time, the agent freezes and packages the original data, corrected data, quality labels, topology snapshots, and device status of the trigger window, writes them to the event storage, and updates E0 to the shared state pool.

[0022] In this embodiment, based on event object E0, the diagnostic positioning agent performs real-world anomaly classification and provides classification probabilities to obtain the updated event E1, as follows: The diagnostic localization agent reads the trigger time t0, trigger DMA, trigger measurement point, and freeze window W from E0, and from D... t With S t Pull out pressure, flow, valve position, pump condition, and water quality sequences from multiple points within the same partition of the window. Simultaneously, incorporate the data quality labels generated by S1 and the health status of the measurement points as weights to construct categorization features. These features simultaneously cover: pressure side (e.g., the difference between the median before and after the window). and DMA weighted summation Consistency of pressure linkage ), flow side (e.g., Δq) k Nighttime bottom flow rise ), water balance (such as ), and discrete operating side (such as the intensity of valve position change of boundary valves), and Pump start / stop / frequency sudden change S pump ); The event is mapped to several real-world working condition categories and its probability is output. The diagnostic and localization agent first uses a multi-class softmax model to give statistical probabilities for the feature vectors. ; Where c represents the candidate category (e.g., leakage / pipe burst, valve misoperation / boundary change, pump station anomaly, sudden demand change, water quality incident, etc.); p ml(c|x) represents the statistical classification probability of the machine learning model ml classifying the event as category c when the feature is x (before the introduction of physical consistency calibration); x is the feature vector; θ c Here is the model parameter vector for category c; T denotes transpose; exp(·) represents the exponential function used to convert the scores to positive values; ∑ c′ This represents the summation over all candidate categories c′; c′ is the category dummy variable used for the summation. Then, a physical consistency score s is introduced. c Post-hoc calibration is performed for values ​​in [0,1]. For example, leakage / burst should satisfy the requirement of "pressure drop (-ΔP)". k >0) / Flow rate increase (Δq) k >0) / High multi-point consistency (Consk(p)) and non-valve-dominated; valve events are more consistent with evidence of "significant valve position step and synchronous pressure / flow direction change"; pump anomalies are more consistent with evidence of "sudden pump operating conditions and zone pressure linkage"; finally, the calibration formula was used for fusion: ; Where λ represents the influence of physical consistency on the result; p(c|E0) is the final posterior probability of category c given event object E0 (the result after fusing statistical classification and physical consistency); E0 is the initial event object (trigger time, set of trigger points, freeze window, initial evidence, etc.); s c λ represents the physical consistency score for category c; λ represents the physical consistency calibration strength. Finally, the classification results are written back to the formation update event E1, and traceable basis is provided for subsequent positioning and handling. Based on retaining all fields of E0 and the reference of frozen evidence package, E1 adds (1) the classification results of real working conditions: {(c, p(c∣E0))} are generated in descending order of probability, and the difference between Top-1 and Top-2 is given to indicate whether there are multiple solutions; (2) the summary of key evidence: the core feature values ​​that lead to the classification (including , respectively representing the weighted pressure change of DMA k, the flow change of DMA kk, the minimum nighttime flow, the water balance residual intensity of DMA k, the valve operation intensity / boundary change intensity characteristics of DMA k, and the pump operating condition change intensity characteristics) multi-point consistency score, model residual and its confidence interval; (3) credibility and priority update: combined with the maximum classification probability and the mean weight of effective measurement points within the window Give And indicate whether it is necessary to immediately enter S4 fine positioning.

[0023] In this embodiment, based on the updated event E1, the diagnostic localization agent performs constraint propagation on the topology map, expanding along the upstream and downstream reachable paths with the strongest anomaly measurement point as the center. Combined with valve boundaries, the search space is limited to the set of pipe segments that can affect that measurement point. Then, by fusing inspection geographical clues to form spatial priors, a candidate pipe segment set is obtained, as detailed below: The diagnostic and localization agent reads update event E1 and selects the node corresponding to the strongest anomaly measurement point (the one with the largest pressure drop or the largest residual) on the pipeline topology graph G=(V,E) as the propagation source point v. Subsequently, reachable paths are expanded along both upstream and downstream directions. During expansion, real-time valve opening / closing status, DMA boundary valves, and partition connectivity are introduced as hard constraints: propagation stops upon encountering a closed valve, an isolated boundary, or a disconnected partition; if the valve position is uncertain, connected branches are retained with topological reliability as the weight, but their priority is reduced, which will affect v. The network portion is cut out from the entire network to form a controlled search subgraph; Within the controlled search subgraph, the diagnostic localization agent uses the anomaly patterns given by E1 as propagation rules: for example, for the "leakage / pipe burst" category, it prioritizes propagation along the path that causes changes in pressure gradient and focuses on connected regions that are consistent with pressure measurement points; for the "valve malfunction" category, propagation will revolve around boundary valves, key connected points, and partition interfaces, checking whether abrupt changes in valve positions near the anomaly can explain the response direction of the measurement point; for the "pump anomaly" category, it spreads from the pump station inlet and outlet and its water supply main to the affected partition, and filters paths based on the temporal relationship of pump operating condition changes; during the propagation process, an interpretability score is accumulated for each edge, and a weighted calculation is used to comprehensively consider the topological distance of the source point, the drivability of valves on the path, the consistency with the temporal response of multiple measurement points, and the support of the water balance residual for the branch; finally, the set of pipe segments with the highest score is output as the candidate pipe segment set, and the evidence chain of each candidate is retained (source measurement point, path, which valves constrain it, and which anomaly features it explains).

[0024] In this embodiment, the spatial prior is formed by fusing inspection and geographic clues, specifically as follows: Based on the candidate pipe segment set, the spatial prior is formed by fusing inspection and geographic clues, converging into a high-value candidate set, and the diagnostic location Agen is derived from D. tExtract inspection records, work orders, complaint hotspots, historical leaks, road construction information, and acoustic alarm location geographic clues adjacent to the event time window, map them onto the GIS space, and project them onto nodes or pipe segments of the topology map; assign a preset high spatial prior weight to locations near clues (such as a road segment with multiple recent complaints, manholes pointed to by acoustic alarms, historically high-incidence pipe material / age sections, and near recently repaired interfaces); then fuse and sort the topology propagation score and spatial prior to obtain the final candidate pipe segment set (including the comprehensive score of each segment, corresponding clues, confidence level, and suggested on-site verification points).

[0025] In this embodiment, based on the candidate pipe segment set, for each candidate location, the theoretical response when an anomaly occurs at that location is calculated and matched with the actual residual. A Bayesian approach is then used to update the probability using prior failure information. Simultaneously, physical consistency filtering is performed to obtain the final set of anomalous pipe segments, as detailed below: Based on the final candidate pipe segment set E c ={e j} and the frozen evidence window W for event E1, for each candidate pipe segment location e j Assume a specific anomaly mechanism (determined by E1 classification Top-K, such as leakage / pipe burst, valve malfunction, pump malfunction, etc.), and calculate the theoretical response under the current topology and boundary valve state. That is, the expected direction, time delay, and magnitude of the impact of the anomaly on each observation point i (pressure / flow / water quality); For each measurement point i, first calculate the actual residual r. i (t): ; Where yi(t) is the actual observed value of measuring point i at time t; This is the predicted value of measuring point i at time t (obtained from the hydraulic model). Then compare it with candidate position e j Theoretical response Align and calculate matching errors, such as weighted squared error or correlation:

[0026] Where j is the candidate anomaly location index; W is the set of event freeze time windows; w i Let i be the mass weight of measurement point i; To start from candidate position e j The theoretical propagation delay to measurement point i; Let be the variance scale of the residual noise at measurement point i; It is a positive number; err j For candidate e jThe overall fitting error; Construct the likelihood from the error: ; in, For candidate position e j Under the premise that the likelihood of observing event evidence E1 is given; ∝ indicates proportionality; exp(·) is an exponential function; Simultaneously, failure priors are introduced. The data is obtained from the asset ledger and historical work orders (such as pipe age, material, diameter, soil corrosivity, historical pipe burst density, road load, recent construction, valve operation frequency, etc.); the candidate pipe segment posterior is updated using Bayesian formula: ; Wherein, P(e j |E1) is an anomaly located in candidate e after given event evidence E1. j The posterior probability; If E1 gives the classification probability P(c|E1) of the abnormal mechanism category c, then a mixture likelihood can be performed: This links positioning and fractal analysis and explicitly represents the sources of uncertainty. Where P(c|E1) is the posterior probability of the fractal of event E1 (the output of S3, used to express "mechanistic uncertainty"); L(E1|e j c) represents the location e in the abnormal position. j And given mechanism c, the conditional likelihood of evidence E1; ∑ c To perform a weighted summation over all mechanism categories; Based on the posterior ranking, a physical consistency constraint check is performed on high-probability candidates to eliminate statistically matched but physically impossible locations: including candidates located in subgraphs not connected to the trigger measurement point, or candidates with closed valves on the path causing unreachable impact; candidates that are leaks / pipe bursts but cannot simultaneously explain the directionality of "pressure drop + water inflow increase / increased water balance residual"; candidates that are valve events but lack corresponding valve position steps or operation records to support them; and candidates that are water quality events but lack temporal propagation consistency along the flow direction. After consistency filtering, the final set of abnormal pipe segments is output, and for each pipe segment, the following are provided: posterior probability P(e|E1), evidence (matching error / correlation), prior elements used (pipe age / material / historical work orders, etc.), and suggested on-site verification points (nearest well chamber / valve, acoustic retesting points).

[0027] In this embodiment, based on the final set of abnormal pipe segments, the impact assessment agent calculates the impact for each abnormal pipe segment separately, as follows: Impact assessment agent targets each e jSelect a disturbance mechanism consistent with the E1 classification and calculate the hydraulic response caused by the anomaly: ; Where Hj(t) is the water head at that location, and C d,j A j The equivalent leak parameters are given by S5 empirical prior; blockage / valve anomalies can be represented by changes in the local resistance coefficient; pump anomalies can be represented by deviations in head / flow characteristics; For position e j Assume the instantaneous leakage flow rate at the time of leakage; C d,j A is the outflow coefficient; j The area of ​​the equivalent leakage orifice is given by ; g is the acceleration due to gravity. The pressure at each node was then obtained through hydraulic simulation or sensitivity approximation. With supply capacity Based on this, the scope of impact and gap are calculated: set of nodes affected by low pressure. ; in, For time t, when the anomaly occurs at time e j Assume a low-voltage node set; v is the network node index; p min This is the lower limit threshold for service pressure. This means taking the union of all times within the window to obtain the set of nodes that were low-pressure during the window period; Water supply shortfall (node ​​demand d) v (t): ; in, Let d be the unmet demand of node v at time t; max(0,·) ensures the gap is non-negative; v (t) represents the demand of node v at time t (baseline water demand, flow rate); The cumulative gap volume within the window is summed over all nodes and integrated over time. Number of users affected (node ​​weight n) v ): ; in, For the exception located at e j Under the assumption that the total number of users affected by the low voltage is n v The user weight corresponding to node v; The scale of critical users affected (such as hospitals, schools, key industries, fire protection points, etc.); A set of key user nodes; Weighting for key users; This represents the set of nodes that simultaneously meet the criteria of being affected by low voltage and belonging to critical users. Calculate secondary risk proxy (as proxy volume for road risk, flooding risk, public opinion risk, etc.): ; Wherein, ρ(e j ) is the position e j The scene weighting coefficient (such as road grade, low-lying terrain, proximity to subway / pipe corridor / dense building area, proximity to important facilities, etc.; the larger the coefficient, the higher the external risk brought by the same amount of leakage.

[0028] In this embodiment, the event level is output based on the calculation results, as follows: First, normalize the key impact indicators of each pipeline segment to [0,1] (using operational calibration constants V0, N0, C0, R0 to control the scale): ; ; ; ; in, is the normalized dimensionless index; min(1,·) is the cutoff function, where the part exceeding 1 is counted as 1; V0 represents the cumulative water supply gap volume caused by this pipeline section; V0 is the calibration scale for the water supply gap volume. N0 represents the number of users affected by this pipe section; N0 is the calibrated scale for the number of users affected. C0 represents the scale of key users affected by this pipeline segment; C0 is the calibrated measure of the impact on key users. This represents the secondary risk corresponding to this pipe section; R0 is the calibration scale for the secondary risk index. Formation of comprehensive impact at the pipe section level: ; Where α, β, γ, δ are the business weight coefficients for each dimension; Then, by combining the posterior probability of the location, the segment-level risk score R is obtained. j : R j =P(e j |E1)·I j ; Overall event risk is aggregated using expected value: ; Among them, R j Risk classification for the j-th abnormal pipe segment; R max For worst-case scenario risk; Rexp For cumulative risk; Select R∈{R max ,R exp Mapped to event level L: ; Where τ1, τ2, τ3 are the graded thresholds; Finally, output the event report, including each segment e j The impact details, ranking, overall level L, list of key users, and recommended handling priorities.

[0029] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0030] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0031] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0032] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A multi-agent collaborative intelligent fault handling method for pipeline networks, characterized in that, Includes the following steps: S1: Obtain relevant pipeline network data and preprocess it to obtain the joint dataset D. t Snapshot S of the current network operating conditions t ; S2: Based on the joint dataset D t Snapshot S of the current network operating conditions t The monitoring and early warning agent performs anomaly detection on each DMA and performs multi-source consistency verification. If the anomaly reaches the triggering condition, an event is created and the key data fragments of the trigger window are frozen for traceable analysis, resulting in event object E0. S3: Based on event object E0, the diagnostic and location agent performs real-world anomaly classification and provides the classification probability to obtain the updated event E1; S4: Based on the updated event E1, the diagnostic localization agent performs constraint propagation on the topology map, taking the strongest abnormal measurement point as the center and expanding along the upstream and downstream reachable paths. Combined with the valve boundary, the search space is limited to the set of pipe segments that can affect the measurement point. Then, the inspection geographical clues are integrated to form spatial priors and obtain the candidate pipe segment set. S5: Based on the candidate pipe segment set, for each candidate location, calculate the theoretical response when an anomaly occurs at that location, match it with the actual residual, and update the probability using a Bayesian approach combined with the failure prior; at the same time, perform physical consistency filtering to obtain the final set of abnormal pipe segments; S6: Based on the final set of abnormal pipe segments, the impact assessment agent calculates the impact level for each abnormal pipe segment and outputs the event level based on the calculation results.

2. The intelligent fault handling method for multi-agent collaborative pipeline network according to claim 1, characterized in that, The preprocessing includes data cleaning, unit and dimension unification, time synchronization, spatial mapping, and topology verification; simultaneously, health status and data quality labels are calculated for each measurement point, and the original values, corrected values, reasons for correction, quality annotations, clock uncertainty, and mapping confidence metadata are all retained; resulting in the joint dataset D. t With network operating condition snapshot S t The joint dataset D t This is a collection of multiple tables aligned to a unified timeline, including real-time quantity time series, event and work order timelines, equipment and asset dimension tables, GIS topology maps, DMA boundary and valve state sequences, and quality and health labels; the network condition snapshot S t This includes supply and demand balance indicators for each DMA, pressure distribution at key nodes, main flow direction, pump operation combinations, boundary valve connectivity patterns, anomaly candidate lists and their confidence levels.

3. The intelligent fault handling method for multi-agent collaborative pipeline network according to claim 1, characterized in that, The monitoring and early warning agent performs stratified detection on each DMA, including pressure anomalies, flow anomalies, flow-pressure relationship anomalies, water quality and concentration anomalies, and equipment condition anomalies, as follows: Pressure anomalies include sudden pressure drops or rises, sustained low pressure, pressure oscillations, and deviations from historical baselines, combined with the time-period characteristics and seasonal patterns of nighttime low pressure and daytime peak pressure; Flow anomalies include nighttime underflow rises, daytime water consumption curve anomalies, DMA inflow-outflow imbalances, and short-term shock flows; Flow-pressure relationship anomalies include pressure drops without flow increases, flow increases with excessive pressure drops, deviations from hydraulic model predictions, and abnormal physical inconsistencies in pump pressure characteristic curves; Water quality and concentration anomalies include excessive or insufficient residual chlorine, abnormal turbidity, pH shifts, and pipeline pollution indicators; Equipment condition anomalies include frequent pump start-stops, abnormal valve position jumps, abnormal electrical parameters, and prolonged sensor jamming; The detection algorithm of the monitoring and early warning agent uses statistical thresholds and physical constraints for checks, and weights the detection results for reliability based on the health status of the measuring points and data quality labels.

4. The intelligent fault handling method for multi-agent collaborative pipeline network according to claim 3, characterized in that, The multi-source consistency verification includes priority for multi-point linkage within the same partition, priority for suspecting sensors due to single-point anomalies, and cross-validation of physical consistency, as follows: Prioritizing multi-point linkage within the same partition means that when multiple measurement points within a DMA simultaneously exhibit anomalies in the same direction, and the magnitudes of these anomalies are spatially correlated, they are preferentially identified as real-world operating events; Prioritizing single-point anomalies means that when only a single measurement point is abnormal while neighboring measurement points are stable, and the data quality label of that measurement point shows a decline in health, a risk of drift, or communication anomalies, it is preferentially suspected to be a sensor failure; Cross-validation of physical consistency uses hydraulic models or empirical constraints to verify the physical rationality of the anomaly; After the anomaly passes the consistency verification, if the confidence level reaches a preset trigger threshold, the monitoring and early warning agent formally creates an event object E0 and freezes key data fragments within the trigger window to ensure the traceability of subsequent analysis. Simultaneously, the agent freezes and packages the original data, corrected data, quality labels, topology snapshots, and device status of the trigger window, writes them to the event storage, and updates E0 to the shared state pool.

5. The intelligent fault handling method for multi-agent collaborative pipeline network according to claim 1, characterized in that, The process involves using the event object E0 to diagnose and locate anomalies in the actual operating conditions, classifying the anomalies, and providing the classification probabilities to obtain the updated event E1, as detailed below: The diagnostic localization agent reads the trigger time t0, trigger DMA, trigger measurement point, and freeze window W from E0, and from D... t With S t The system retrieves pressure, flow, valve position, pump operating conditions, and water quality sequences from multiple points within the same partition of the window. It also incorporates the data quality labels generated by S1 and the health status of the measurement points as weights to construct categorical features. Events are mapped to several real-world operating condition categories, and probabilities are output. The diagnostic and localization agent first uses a multi-class softmax model to provide statistical probabilities for the feature vectors. ; Where c represents the candidate category; p ml (c|x) represents the statistical classification probability of the machine learning model ml classifying the event as category c when the feature is x; x is the feature vector; θ c Here is the model parameter vector for category c; T denotes transpose; exp(·) represents the exponential function used to convert the scores to positive values; ∑ c′ This represents the summation over all candidate categories c′; c′ is the category dummy variable used for the summation. Then, a physical consistency score s is introduced. c Posterior calibration is performed for ∈[0,1], and finally fusion is performed using the calibration formula: ； Where λ represents the strength of the influence of physical consistency on the result; p(c|E0) is the final posterior probability of class c given event object E0; E0 is the initial event object; s c λ represents the physical consistency score for category c; λ represents the physical consistency calibration strength. Finally, the fractal result is written back to the formation update event E1.

6. The intelligent fault handling method for multi-agent collaborative pipeline network according to claim 5, characterized in that, Based on the updated event E1, the diagnostic localization agent performs constraint propagation on the topology map, expanding along upstream and downstream reachable paths centered on the strongest anomaly measurement point. Combined with valve boundaries, the search space is limited to the set of pipe segments that can affect that measurement point. Then, by integrating inspection geographical clues to form spatial priors, a candidate pipe segment set is obtained, as detailed below: The diagnostic and localization agent reads update event E1 and selects the node corresponding to the strongest anomaly detection point on the pipeline topology graph G=(V,E) as the propagation source point v. Subsequently, reachable paths are expanded along both upstream and downstream directions. During expansion, real-time valve opening / closing status, DMA boundary valves, and partition connectivity are introduced as hard constraints: propagation stops upon encountering a closed valve, an isolated boundary, or a disconnected partition; if the valve position is uncertain, connected branches are retained with topological reliability as the weight, but their priority is reduced, which will affect v. The network portion is cut out from the entire network to form a controlled search subgraph; Within the controlled search subgraph, the diagnostic localization agent uses the abnormal patterns given by E1 as the propagation rules. During the propagation process, an interpretability score is accumulated for each edge. The topological distance of the source point, the drivability of valves on the path, the consistency with the time series response of multiple measurement points, and the support of the water balance residual for the branch are comprehensively considered through weighted calculation. Finally, the set of pipe segments with the highest score is output as the candidate pipe segment set, and the evidence chain of each candidate is retained.

7. The intelligent method for handling pipeline network faults through multi-agent collaboration according to claim 5, characterized in that, The spatial prior formed by the fusion of inspection geographic clues is as follows: Based on the candidate pipeline segment set, spatial priors are formed by integrating inspection and geographical clues, converging into a high-value candidate set, and the diagnostic location of the Agen is determined from D. t Extract inspection records, work orders, complaint hotspots, historical leaks, road construction information, and acoustic alarm location geographic clues adjacent to the event time window, map them onto the GIS space, and project them onto nodes or pipe segments of the topology map; assign preset high spatial prior weights to locations close to the clues. The topology propagation score and spatial prior are then fused and sorted to obtain the final candidate pipe segment set.

8. The intelligent fault handling method for multi-agent collaborative pipeline network according to claim 1, characterized in that, Based on the candidate pipe segment set, for each candidate location, the theoretical response when an anomaly occurs at that location is calculated, matched with the actual residual, and the probability is updated using a Bayesian approach combined with failure priors. Simultaneously, physical consistency filtering is performed to obtain the final set of anomalous pipe segments, as detailed below: Based on the final candidate pipe segment set E c ={e j } and the frozen evidence window W for event E1, for each candidate pipe segment location e j Assume a specific anomaly mechanism and calculate the theoretical response under the current topology and boundary valve states. That is, the expected direction, time delay and magnitude of the impact of the anomaly on each observation point i; For each measurement point i, first calculate the actual residual r. i (t): ； Where yi(t) is the actual observed value of measuring point i at time t; Let be the predicted value of measurement point i at time t; Then compare it with candidate position e j Theoretical response Align and calculate matching errors, such as weighted squared error or correlation: ； Where j is the candidate anomaly location index; W is the set of event freeze time windows; w i Let i be the mass weight of measurement point i; To start from candidate position e j The theoretical propagation delay to measurement point i; Let be the variance scale of the residual noise at measurement point i; It is a positive number; err j For candidate e j The overall fitting error; Construct the likelihood from the error: ； in, For candidate position e j Under the premise that the likelihood of observing event evidence E1 is given; ∝ indicates proportionality; exp(·) is an exponential function; Simultaneously, failure priors are introduced. The results are obtained from the asset ledger and historical work orders; the posterior of the candidate pipeline segment is updated using Bayes' theorem. ； Wherein, P(e j |E1) is an anomaly located in candidate e after given event evidence E1. j The posterior probability; If E1 gives the classification probability P(c|E1) of the abnormal mechanism category c, then a mixture likelihood can be performed: This links positioning and fractal analysis and explicitly represents the sources of uncertainty. Where P(c|E1) is the posterior probability of the fractal of event E1; L(E1|e j c) represents the location e in the abnormal position. j And given mechanism c, the conditional likelihood of evidence E1; ∑ c To perform a weighted summation over all mechanism categories; Based on the posterior ranking, physical consistency constraint checks are performed on high-probability candidates to eliminate positions that are statistically matched but physically impossible; after consistency filtering, the final set of abnormal pipe segments is output.

9. The intelligent method for handling pipeline network faults through multi-agent collaboration according to claim 1, characterized in that, Based on the final set of abnormal pipe segments, the impact assessment agent calculates the impact for each abnormal pipe segment separately, as follows: Impact assessment agent targets each e j Select a disturbance mechanism consistent with the E1 classification and calculate the hydraulic response caused by the anomaly: ； Where Hj(t) is the water head at that location, and C d,j A j Equivalent leak parameters; For position e j Assume the instantaneous leakage flow rate at the time of leakage; C d,j A is the outflow coefficient; j The area of ​​the equivalent leakage orifice is given by ; g is the acceleration due to gravity. The pressure at each node was then obtained through hydraulic simulation or sensitivity approximation. With supply capacity Based on this, the scope of impact and gap are calculated: set of nodes affected by low pressure. ； in, For time t, when the anomaly occurs at time e j Assume a low-voltage node set; v is the network node index; p min This is the lower limit threshold for service pressure. This means taking the union of all times within the window to obtain the set of nodes that were low-pressure during the window period; Water supply shortfall: ； in, Let d be the unmet demand of node v at time t; max(0,·) ensures the gap is non-negative; v (t) represents the demand of node v at time t; The cumulative gap volume within the window is summed over all nodes and integrated over time. Number of users affected: ； in, For the exception located at e j Under the assumption that the total number of users affected by the low voltage is n v The user weight corresponding to node v; The scale of key users affected; A set of key user nodes; Weighting for key users; This represents the set of nodes that simultaneously meet the criteria of being affected by low voltage and belonging to critical users. Calculate secondary risk proxy: ； Wherein, ρ(e j ) is the position e j The scene weight coefficient.

10. The intelligent fault handling method for multi-agent collaborative pipeline network according to claim 1, characterized in that, The event level is output based on the calculation results, as follows: First, normalize the key impact indicators of each pipe section to [0,1]: ； ； ； ； in, is the normalized dimensionless index; min(1,·) is the cutoff function, where the part exceeding 1 is counted as 1; V0 represents the cumulative water supply gap volume caused by this pipeline section; V0 is the calibration scale for the water supply gap volume. N0 represents the number of users affected by this pipe section; N0 is the calibrated scale for the number of users affected. C0 represents the scale of key users affected by this pipeline segment; C0 is the calibrated measure of the impact on key users. This represents the secondary risk corresponding to this pipe section; R0 is the calibration scale for the secondary risk index. Formation of comprehensive impact at the pipe section level: ； Where α, β, γ, δ are the business weight coefficients for each dimension; Then, by combining the posterior probability of the location, the segment-level risk score R is obtained. j : R j =P(e j ∣E1)·I j ; Overall event risk is aggregated using expected value: ； Among them, R j Risk classification for the j-th abnormal pipe segment; R max For worst-case scenario risk; R exp For cumulative risk; Select R∈{R max ,R exp Mapped to event level L: ； Where τ1, τ2, τ3 are the graded thresholds; Finally, output the event report, including each segment e j The impact details, ranking, overall level L, list of key users, and recommended handling priorities.

Images