A multi-modal data quality control driven sewer network diagnosis method and intelligent agent

Abstract

This invention discloses a multimodal data quality control-driven diagnostic method and intelligent agent for drainage pipe networks, relating to the fields of intelligent operation and maintenance and artificial intelligence technology for urban drainage pipe networks. The method includes: acquiring and preprocessing multi-source heterogeneous data from the drainage pipe network; constructing the spatial topology of the pipe network and establishing spatiotemporal correlations of the multimodal data; constructing a multimodal data quality control model and calculating a comprehensive quality score and reliability weight; extracting modal features from the multimodal data to obtain a multimodal fusion state vector; retrieving historical fault cases and generating diagnostic prompts; loading a generative diagnostic model to perform diagnostic inference on the diagnostic prompts; and acquiring and updating the diagnostic feedback data corresponding to the structured diagnostic results. This invention integrates data quality control throughout the entire process of data acquisition, fusion, inference, and optimization, improving the accuracy, reliability, and interpretability of pipe network anomaly identification in complex environments, and providing technical support for risk warning and intelligent operation and maintenance of urban drainage facilities.

Description

Technical Field

[0001] This invention relates to the fields of intelligent operation and maintenance of urban drainage pipe networks, artificial intelligence, and multimodal data fusion technology. Specifically, it relates to a multimodal data quality control-driven diagnostic method and intelligent agent for drainage pipe networks. Background Technology

[0002] Urban drainage networks are crucial infrastructure for ensuring urban drainage safety, flood control, and water environment management. Their operational status directly impacts the normal operation of the city and the daily lives of residents. Because drainage pipes are buried underground for extended periods and are continuously affected by factors such as sewage corrosion, pipe aging, foundation settlement, external loads, foreign object deposition, and construction disturbances, they are prone to blockages, ruptures, leaks, settlement, and structural damage. When network faults are not detected and accurately located in a timely manner, they can lead to reduced drainage capacity, sewage overflows, road flooding, and even risks to local structural safety. Therefore, achieving timely perception of the operational status of drainage networks and intelligent fault diagnosis is of great significance for improving the operational safety and maintenance efficiency of urban drainage facilities.

[0003] With the development of IoT, GIS, machine vision, and AI technologies, a large amount of multi-source data has been accumulated during the operation and maintenance of drainage pipe networks. This includes time-series data from sensor monitoring such as liquid level, flow rate, pressure, and water quality; GIS spatial topology attribute data such as pipe diameter, material, burial depth, slope, and network connection relationships; CCTV images and video data of pipelines; and operation and maintenance text data such as inspection records, maintenance records, historical fault reports, and engineering technical documents. How to comprehensively utilize these different types of data to achieve fault diagnosis in drainage pipe networks has become an important research direction in the field of intelligent operation and maintenance of drainage pipe networks.

[0004] While existing methods for diagnosing drainage pipe networks can utilize sensor monitoring data, GIS spatial topology data, pipeline CCTV visual data, and historical operation and maintenance knowledge for condition analysis, they still have at least the following problems: First, data from different sources vary significantly in terms of completeness, reliability, and timeliness, while existing multimodal fusion methods lack a dynamic data quality evaluation mechanism for current diagnostic tasks. Secondly, different modal data are usually processed separately, which does not fully utilize the consistency between different data of the same pipeline network object and the same time window. When multiple data conflict with each other, it is difficult to determine the credibility of different diagnostic evidence. Third, existing multimodal fusion methods typically employ fixed fusion or direct feature fusion, meaning that changes in data quality cannot effectively affect the contribution of different modal features to the overall state representation. Fourth, there is a lack of effective correlation between the current operating status and historical failure cases, engineering knowledge and expert experience, making it difficult for the diagnostic model to make full use of the accumulated field failure and maintenance experience; Fifth, while some AI diagnostic methods can output the type or probability of anomalies, they lack sufficient criteria for fault diagnosis, risk level, scope of impact, and specific maintenance recommendations, making it difficult to directly meet the needs of engineering operation and maintenance decision-making. Sixth, the results of on-site manual verification, actual maintenance results, and post-maintenance operation status failed to form an effective feedback loop, making it difficult for the diagnostic system to be continuously updated and optimized based on actual engineering operation results.

[0005] In summary, there is a need for an intelligent diagnostic technology for drainage networks that can comprehensively evaluate the integrity, reliability, timeliness, and cross-modal consistency of multi-source heterogeneous data from drainage networks, dynamically adjust the contribution of different modal information in the state fusion process based on data quality, combine historical fault cases, engineering knowledge, and expert rules for diagnostic reasoning, and continuously optimize diagnostic capabilities using on-site maintenance feedback. Summary of the Invention

[0006] The main objective of this invention is to disclose a multimodal data quality control-driven method for diagnosing drainage pipe networks, comprising: Multi-source heterogeneous data of drainage pipe network is acquired and preprocessed to generate multimodal data; the multi-source heterogeneous data includes: sensor monitoring time series data, GIS spatial topology attribute data, pipeline CCTV visual inspection data and operation and maintenance knowledge text data; Specifically, the preprocessing includes: The sensing monitoring time series data is divided according to a preset time window, and the sensing monitoring time series data collected by different sensing devices are subjected to time synchronization, anomaly detection, missing data repair and data standardization preprocessing. The time synchronization refers to the process of converting different sensor monitoring time-series data with different sampling frequencies to a unified time axis by resampling, linear interpolation, and spline interpolation. The anomaly detection refers to the use of sliding window statistics, Z-score, interquartile range, and isolated forest to identify anomaly sensing and monitoring time-series data, which serves as the basis for subsequent data reliability evaluation. The missing data repair refers to: for short-term missing sensor monitoring time-series data, interpolation repair is performed using adjacent valid data; for long-term continuous missing sensor monitoring time-series data, no forced filling is performed, and the corresponding missing state is retained for use in subsequent data quality control steps to reduce the data integrity index corresponding to that mode. The data standardization refers to the elimination of dimensional differences between different sensor monitoring time series data by using Z-score standardization or min-max normalization methods. The pipeline CCTV visual inspection data is subjected to image enhancement, target area detection, and defect feature extraction to generate pipeline defect information; the pipeline defect information includes: cracks, deformation, leakage, deposits, obstacles, interface abnormalities, and corrosion; The operation and maintenance knowledge text data is parsed and semantic information is extracted to generate fault description information, maintenance experience information and engineering knowledge semantic information.

[0007] Based on GIS spatial topology attribute data, a pipeline spatial topology structure is constructed, the spatiotemporal correlation of multimodal data is established, and a multimodal correlation data set is generated. Specifically, the construction of the pipeline network spatial topology includes: Based on the GIS spatial topology attribute data, drainage network nodes, pipe segments, and their connection relationships are obtained, and a spatial topology structure of the network is constructed, including: treating network nodes as graph structure nodes, pipe segment connection relationships as graph structure edges, and associating network spatial attributes with corresponding graph structure nodes or graph structure edges, as shown below: ; in, Represents the set of nodes in the drainage pipe network. This indicates the pipe segments and their connection relationships between different nodes; the nodes include: inspection wells, rainwater inlets, pumping stations, and other drainage network connection facilities; the pipe segments represent the drainage pipes connecting the corresponding nodes; Establish corresponding pipeline object identifiers based on pipeline nodes and pipeline segments, which are used to determine the specific pipeline objects corresponding to data from different sources during subsequent data processing; Furthermore, establishing the spatiotemporal correlation of multimodal data includes: Establish the association between GIS spatial topology attribute data and pipeline network objects, and then... The GIS spatial topology attribute data corresponding to each pipeline network object is represented as follows: ;in, Represents GIS spatial topological attribute data; a unified index based on object associations. Map the pipe diameter, material, length, burial depth, slope, spatial location, and upstream and downstream connection relationships in the GIS spatial topology attribute data to the corresponding pipe network objects; Based on the multimodal data, time-series status data corresponding to pipeline nodes or pipe segments are obtained. According to the pipeline object identifier and spatial location, the time-series status data is mapped to the corresponding pipeline node or pipe segment. The time-series status data includes: liquid level, flow rate, flow velocity, pressure, water quality, and pump station operating status. Associate pipeline defect information with the corresponding pipe section and the corresponding preset time window; Link fault description information and maintenance experience information to the corresponding pipeline nodes or pipe sections and the corresponding preset time windows; Based on the multimodal data, cross-modal association relationships are established with pipeline network objects and preset time windows to form a multimodal association data set for data quality evaluation and multimodal feature fusion.

[0008] Constructing a multimodal data quality control model includes: evaluating the quality of the multimodal associated data set; determining data integrity indicators, data reliability indicators, data timeliness indicators, and cross-modal consistency indicators based on a preset time window; and calculating a comprehensive quality score and a credibility weight; wherein the preset time window is a continuous time interval on the time axis. Specifically, the comprehensive quality score is based on the current diagnosis. The calculation is performed on modal data, and the calculation method is expressed as follows: ; in, Indicates the first The overall quality score of modal data represents the first... Data integrity metrics for modal data Indicates the first Data reliability metrics for modal data Indicates the first Data timeliness indicators for modal data Indicates the first Cross-modal consistency metrics for modal data; , , , The corresponding evaluation index weights, satisfying the constraints, are expressed as follows: , ; No. Confidential weights corresponding to modal data The overall quality score of various modal data involved in the current diagnosis is used. Confirmed, indicated as: ; in, This represents the total number of valid data modalities participating in the current diagnosis. Let be the confidence weight adjustment parameter, and: , ; Furthermore, the data integrity indicator refers to: based on the data integrity index within the preset time window... Determining data integrity metrics based on missing data in modal data , used to characterize the The degree of effective data completeness of modal data; The data reliability index refers to: based on the preset time window... Anomaly detection results for modal data determine data reliability indicators , used to characterize the The degree of credibility of valid information in modal data, and the corresponding data reliability index as the degree of data anomaly increases. The lower; The data timeliness indicator refers to: based on the preset time window... The time difference between the effective data acquisition time of modal data and the current diagnostic time determines the data timeliness index. , used to characterize the The temporal validity of modal data relative to the current diagnostic time; The cross-modal consistency index refers to: based on the first [time window] belonging to the same pipeline network object and the corresponding preset time window The degree of consistency between modal data and other related modal data in describing the operating status of the same pipeline network is used to determine the cross-modal consistency index. ; Specifically, the dynamic weighted fusion includes: The modal features of the multimodal data are uniformly represented and corresponding feature mapping functions are set for each, transforming the modal features to a unified representation. 3D feature space; for the current diagnosis of the th Modal data, represented as: ; in, Indicates the first The feature mapping function corresponding to the modality consists of a fully connected layer, layer normalization, and a non-linear activation function. The first to participate in the current diagnosis The modal feature vector corresponding to the modal data is represented as follows: Determine the multimodal fusion state vector corresponding to the current operating state of the drainage pipe network. , represented as: ; in, For the first The confidence weights corresponding to modal data The total number of valid data modalities participating in the current diagnosis; To achieve data quality-driven trusted fusion, a trusted weighting approach is adopted. Dynamically weighted and fused multimodal features currently involved in diagnosis are represented as follows: ; in, Indicates the first The pipeline object in the first The multimodal fusion state vector corresponding to each diagnostic time window; Indicates the first The feature vector corresponding to the class mode; Indicates the first The credibility weights corresponding to the modalities.

[0009] Based on the multimodal fusion state vector, historical fault cases are retrieved, engineering knowledge and expert rules are obtained, and diagnostic prompts are generated. Specifically, the retrieved historical fault cases include: The first The state features corresponding to each historical failure case are represented as a historical case state vector. Calculate the current multimodal fusion state vector With the historical case state vector Cosine similarity between , represented as: ; in, Indicates the current operating status of the drainage network and the first Cosine similarity between historical failure cases; Based on the cosine similarity The historical failure cases are sorted to obtain historical failure cases whose similarity meets the preset conditions; Based on the obtained historical fault cases, engineering knowledge and expert rules related to the current operation status of the drainage network are obtained from the engineering knowledge base and expert rule base. The current operation status of the drainage network, the retrieved historical fault cases, engineering knowledge and expert rules are combined to form a search enhancement context.

[0010] A generative diagnostic model is loaded, which performs diagnostic reasoning on diagnostic prompts and outputs structured diagnostic results. The structured diagnostic results include: fault type, risk level, scope of impact, and maintenance and handling plan. Specifically, the diagnostic reasoning includes: inputting diagnostic prompts into the generative diagnostic model, and having the generative diagnostic model combine the current operating status of the drainage network with retrieved historical fault cases, engineering knowledge, and expert rules to perform diagnostic reasoning and generate the structured diagnostic results; The fault type refers to one or more of the following fault types used to characterize the current pipeline network object: pipeline rupture, blockage, leakage, settlement, and structural damage. The risk level refers to the level determined based on the current degree of anomaly, the scope of the fault's impact, and the development trend of the operational status. The scope of influence refers to the range determined based on the faulty pipeline node or segment and its connection relationship in the pipeline network spatial topology. The maintenance and handling plan refers to a plan determined based on the fault type, risk level, scope of impact, retrieved historical fault cases, engineering knowledge, and expert rules, and includes at least maintenance priority and recommended handling measures.

[0011] The diagnostic feedback data corresponding to the structured diagnostic results is obtained and correlated to form a diagnostic feedback sample. The historical fault case library, engineering knowledge base, multimodal data quality control model and generative diagnostic model are then corrected and updated. The diagnostic feedback data includes: manual verification information, on-site maintenance feedback data and the operational status data of the drainage network after diagnosis.

[0012] On the other hand, the present invention provides a multimodal data quality control-driven intelligent agent for drainage network diagnosis, used to execute a multimodal data quality control-driven drainage network diagnosis method, including: The multimodal data perception module is used to acquire and preprocess multimodal data of the drainage network, establish the spatiotemporal correlation between the multimodal data and the corresponding network objects, and form a multimodal correlation data set. The data quality control module is used to evaluate the quality of multimodal associated data sets, determine data integrity indicators, data reliability indicators, data timeliness indicators, and cross-modal consistency indicators, and construct a comprehensive quality score and credibility weight. The multimodal trusted fusion module is used to extract modal features from multimodal data, and dynamically weight and fuse the modal features according to the trusted weights to obtain a multimodal fusion state vector; The intelligent diagnostic reasoning module retrieves historical fault cases, acquires engineering knowledge and expert rules based on the multimodal fusion state vector, generates diagnostic prompts, and outputs structured diagnostic results using a generative diagnostic model. The feedback learning module is used to acquire manual verification information, on-site maintenance feedback data, and drainage network operation status data, and to update the historical fault case library, engineering knowledge base, multimodal data quality control model, and generative diagnostic model.

[0013] According to this invention, by constructing a multimodal data quality control model, a four-dimensional quality evaluation of the completeness, reliability, timeliness, and cross-modal consistency of different modal data participating in the current diagnosis is performed. This enables the diagnostic process to determine the credibility of different data sources based on the actual state of the data, avoiding interference from low-quality data in the diagnostic judgment. By establishing multimodal data association relationships based on pipeline network objects and time windows, sensor monitoring data, GIS spatial data, pipeline CCTV visual data, and operation and maintenance text data can be jointly analyzed around the same drainage pipeline network object. Furthermore, the contribution of different modal features to the comprehensive state representation is dynamically adjusted according to the credibility weight, effectively suppressing the impact of missing data, anomalies, insufficient timeliness, or cross-modal conflicts on the diagnostic results, and significantly improving the multimodal state representation. The reliability of the diagnostic model is enhanced by providing external information relevant to the current diagnostic task through historical fault case databases, engineering knowledge bases, and expert rule bases. This information is combined with the spatial topology of the pipeline network to determine the scope of fault impact, outputting fault type, risk level, and maintenance disposal plan including maintenance priority and recommended measures. This strengthens the engineering basis, interpretability, and operational decision support capabilities of the diagnostic results. Furthermore, by feeding back manual verification, on-site maintenance results, and post-maintenance operational status to the historical fault case database, engineering knowledge base, and related models, a continuous optimization loop is formed, linking diagnosis, verification, maintenance, and learning. This continuously improves the accuracy of pipeline anomaly identification, diagnostic reliability, and interpretability in complex environments, providing technical support for risk warning and intelligent operation and maintenance of urban drainage facilities. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the method flow provided according to an embodiment of the present invention; Figure 2 This is a system architecture diagram provided according to an embodiment of the present invention. Detailed Implementation

[0015] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.

[0016] like Figure 1 As shown, this invention discloses a multimodal data quality control-driven method for diagnosing drainage pipe networks, comprising the following steps: Step S100: Obtain multi-source heterogeneous data of drainage pipe network and perform preprocessing to generate multimodal data; Multimodal data includes: Time-series data from sensor monitoring: Time-series data from sensor monitoring of liquid level, flow rate, flow velocity, pressure, water quality, and pump station operating status during the operation of the drainage pipe network; GIS spatial topology attribute data: related attribute information such as pipes, pipe segments, drainage network nodes, pipe diameter, pipe material, length, burial depth, slope, construction time, renovation time, and spatial coordinates; Pipeline CCTV visual inspection data: visual data; Operation and maintenance knowledge text data: inspection records, maintenance records, and historical fault texts.

[0017] Specifically, the preprocessing of multi-source heterogeneous data from drainage pipe networks includes: 1) The sensor monitoring time-series data is divided according to a preset time window (a preset continuous time interval on the time axis), and time synchronization, anomaly detection, missing data repair, and data standardization preprocessing are performed on the sensor monitoring time-series data collected by different sensing devices. Specifically, this includes: Time synchronization: For sensor monitoring time series data with different sampling frequencies, resampling, linear interpolation, and spline interpolation are used to transform the different data to a unified time axis; Anomaly detection: Sliding window statistics, Z-score, interquartile range, and isolated forest are used to identify anomaly detection in time-series sensor monitoring data. The anomaly detection results are used as the basis for data preprocessing and subsequent data reliability evaluation. Missing data repair refers to: for short-term missing sensor monitoring time-series data, interpolation repair is performed using adjacent valid data; for long-term continuous missing sensor monitoring time-series data, no forced filling is performed, and the corresponding missing state is retained for use in subsequent data quality control steps to reduce the data integrity index corresponding to that mode. Data standardization refers to the elimination of dimensional differences between different sensor monitoring time series data by using Z-score standardization or min-max normalization methods.

[0018] 2) Perform image enhancement, target area detection, and defect feature extraction on the CCTV visual inspection data of the pipeline to generate pipeline defect information; The pipeline defect information includes: cracks, deformation, leakage, deposits, obstructions, interface abnormalities, and corrosion.

[0019] 3) Perform text parsing and semantic information extraction on the operation and maintenance knowledge text data to obtain fault description information, maintenance experience information and engineering knowledge semantic information.

[0020] Step S110: Construct the spatial topology structure of the pipeline network based on GIS spatial topology attribute data, establish the spatiotemporal correlation of multimodal data, and generate a multimodal correlation data set; Specifically, the steps for constructing the pipeline network spatial topology are as follows: Step 1: Establish the spatial topology of the drainage network based on the drainage network nodes, pipe segments, and their connections in the GIS spatial topology attribute data, represented as follows: ; in, Represents the set of nodes in the drainage pipe network. This indicates the pipe segments and their connection relationships between different nodes; the nodes include: inspection wells, rainwater inlets, pumping stations, and other drainage network connection facilities; the pipe segments are drainage pipes connecting the corresponding nodes.

[0021] Step 2: Treat the drainage network nodes as graph structure nodes, the pipe segment connection relationships as graph structure edges, and associate the network attributes such as pipe diameter, material, length, burial depth, slope and spatial location with the corresponding nodes or pipe segments; Step 3: Associate pipe diameter, pipe material, length, burial depth, slope, construction time, renovation time, and spatial coordinates with the corresponding nodes or pipe segments; establish a unique network object identifier for each node and pipe segment in the drainage network, represented as: ;in, Indicates the index of the pipeline object. Used to uniquely identify the corresponding drainage network node or pipe segment, and as a unified index to establish object associations between different data sources; For example, for a certain pipe segment object, it is possible to associate its upstream node, downstream node, pipe diameter, pipe material, length, burial depth, slope, construction time, and spatial location.

[0022] Establishing the spatiotemporal correlation of the multimodal data includes: 1) Establish the association between GIS spatial topology attribute data and pipeline network objects, and link the first... The GIS spatial topology attribute data corresponding to each pipeline network object is denoted as: ; in, Represents GIS spatial topology attribute data; According to the unified index of object association Map the pipe diameter, material, length, burial depth, slope, spatial location, and upstream and downstream connection relationships in the GIS spatial topology attribute data to the corresponding pipeline network objects.

[0023] 2) Based on the multimodal data, obtain the time-series status data corresponding to the network nodes or pipe segments, and map the S100 time-series status data to the corresponding network nodes or pipe segments according to the object identifier and spatial location of the drainage network.

[0024] 3) Associate the S100 pipeline defect information with the corresponding pipe section and the preset time window; 4) Associate the S100 fault description information and maintenance experience information with the corresponding pipeline node or pipe section and the corresponding preset time window.

[0025] The formation of the multimodal association data set includes: 1) Based on the pipeline network object and time window, establish cross-modal association relationships among the sensor monitoring time series data, GIS spatial topology attribute data, pipeline CCTV visual inspection data, and operation and maintenance knowledge text data.

[0026] 2) Enable joint analysis of different modal data around the same drainage network object and the same time window to form a multimodal association data set for data quality evaluation and multimodal feature fusion.

[0027] Step S120: Construct a multimodal data quality control model, including: evaluating the quality of the multimodal associated data set, determining data integrity indicators, data reliability indicators, data timeliness indicators, and cross-modal consistency indicators, and calculating a comprehensive quality score and a credibility weight; Specifically, for each type of modal data involved in the current diagnosis, quality evaluation is conducted from four dimensions: data integrity, data reliability, data timeliness, and cross-modal consistency. Each quality evaluation indicator is then normalized to ensure that each indicator is within the same evaluation range. The integrity index refers to: based on the first [time window] within the preset time window Determining data integrity metrics based on missing data in modal data , used to characterize the The degree of effective data completeness of modal data; The data reliability index refers to: based on the preset time window... Anomaly detection results for modal data determine data reliability indicators , used to characterize the The degree of credibility of valid information in modal data, and the corresponding data reliability index as the degree of data anomaly increases. The lower; The data timeliness indicator refers to: based on the preset time window... The time difference between the effective data acquisition time of modal data and the current diagnostic time determines the data timeliness index. , used to characterize the The temporal validity of modal data relative to the current diagnostic time; The cross-modal consistency index refers to: based on the first [time window] belonging to the same pipeline network object and the corresponding preset time window The degree of consistency between modal data and other related modal data in describing the operating status of the same pipeline network is used to determine the cross-modal consistency index. .

[0028] The construction of the comprehensive quality score and credibility weights includes: 1) The comprehensive quality score is based on the current diagnosis. The calculation method for the comprehensive quality score based on modal data is expressed as follows: ; in, Indicates the first The overall quality score of modal data represents the first... Data integrity metrics for modal data Indicates the first Data reliability metrics for modal data Indicates the first Data timeliness indicators for modal data Indicates the first Cross-modal consistency metrics for modal data; , , , The weights for data integrity, data reliability, data timeliness, and cross-modal consistency indicators, while satisfying constraints, are expressed as follows: , .

[0029] 2) No. Confidential weights corresponding to modal data The overall quality score of various modal data involved in the current diagnosis is used. Confirmed, indicated as: , ; in, This represents the total number of valid data modalities participating in the current diagnosis. Let be the confidence weight adjustment parameter, and: .

[0030] 3) Data modalities with higher overall quality scores receive higher confidence weights, while data modalities with lower overall quality scores receive lower confidence weights. Specifically, when the quality of a certain modality data changes, its corresponding credibility weight changes synchronously with the overall quality score. The credibility weight is used to adjust the contribution of different modal features to the overall state representation of the drainage network during the subsequent multimodal credibility fusion process.

[0031] Step S130: Extract the modal features of the multimodal data, and dynamically weight and fuse the features according to the trusted weights to obtain a multimodal fusion state vector representing the current operating state of the drainage network; Specifically, modal feature extraction is performed on the S100 multimodal data used in the current diagnosis, including the following steps: Step 1: Extract time-series features from the sensing and monitoring time-series data of the multimodal data to obtain a time-series state feature vector; Step 2: Extract topological features from the GIS spatial topological attribute data of the multimodal data to obtain the spatial topological feature vector of the pipeline network; Step 3: Extract visual features from the multimodal pipeline CCTV visual inspection data to obtain the pipeline defect visual feature vector; Step 4: Extract semantic features of the operation and maintenance knowledge text data from the multimodal data, and perform semantic encoding using a BERT-like text encoder and a pre-trained language model, combining it with the current... The corresponding inspection records, maintenance records, and historical fault text input text semantic encoding model perform semantic encoding on the operation and maintenance knowledge text data in the multimodal associated data set to form text semantic features. ;in, Represents text modality, For the first The pipeline object in the first The corresponding text semantic feature vector within each diagnostic time window.

[0032] The dynamic weighted fusion includes: 1) Perform a unified representation of the modal features of multimodal data and set corresponding feature mapping functions for each, transforming the modal features to a unified representation. 3D feature space; for the current diagnosis of the th Modal data, represented as: ; in, Indicates the first The pipeline object in the first Within the first diagnostic time window The feature vector corresponding to the modal data. This represents the index of the data modality category participating in the current diagnosis. Indicates the first The feature mapping function corresponding to a modality can be composed of a fully connected layer, layer normalization, and a nonlinear activation function.

[0033] 2) The first The modal feature vector corresponding to the modal data is represented as follows: Determine the multimodal fusion state vector corresponding to the current operating state of the drainage pipe network. , represented as: ; in, For the first The confidence weights corresponding to modal data The total number of valid data modalities participating in the current diagnosis; when the first Overall quality score of modal data When it decreases, its corresponding credibility weight Reduce, thereby reducing the modality eigenvector For the multimodal fused state vector . contributions.

[0034] 3) Perform data quality-driven trusted fusion, using S120 trusted weights. The multimodal features currently involved in the diagnosis are dynamically weighted and fused, as follows: ; in, Indicates the first The multimodal fusion state vector corresponding to each pipeline object within the k-th diagnostic time window; Indicates the first The feature vector corresponding to the class mode; Indicates the first The confidence weights corresponding to a modality; when the data quality of a certain modality changes, its corresponding feature pairs The contribution changes synchronously with the credibility weight.

[0035] Step S140: Based on the multimodal fusion state vector, retrieve historical fault cases, obtain engineering knowledge and expert rules, and generate diagnostic prompt information; Specifically, a system of historical failure cases, engineering knowledge, and expert rules will be constructed, including... 1) Build a historical fault case library based on historical fault events that have been manually verified or confirmed by on-site maintenance; Each historical fault case includes: case number, corresponding pipeline object identifier, fault occurrence time, operating status information within the preset time window before the fault occurred, abnormal phenomenon description information, pipeline attribute information, fault impact results, and maintenance and handling measures information.

[0036] 2) Construct an engineering knowledge base, organize engineering knowledge according to semantically complete knowledge fragments, and use a text embedding model to generate corresponding knowledge vectors to support semantic retrieval; used to store engineering knowledge related to the operating status, fault mechanism, fault handling and maintenance experience of drainage pipe network; Among them, engineering knowledge can include the formation mechanism of failures such as pipe blockage, rupture, leakage, settlement and structural damage, typical abnormal manifestations, inspection methods, handling methods, applicable maintenance conditions and maintenance experience; The engineering knowledge base stores general engineering knowledge, while the operation and maintenance knowledge text data mainly represents specific engineering knowledge. The relevant historical inspection and maintenance records serve as general knowledge and historical status information of the object, respectively, during the diagnostic process.

[0037] 3) Construct a basic database for the drainage network to store the identification of drainage network objects, the spatial topology of the network, and basic attribute data such as pipe diameter, material, length, burial depth, slope, construction time, and spatial location. This provides basic support for the spatiotemporal correlation of multimodal data, cross-modal consistency evaluation, and intelligent diagnostic reasoning.

[0038] 4) Establish an expert rule base based on actual operation and maintenance experience of drainage pipe networks. Store fault judgment rules and maintenance and handling rules formed based on drainage pipe network operation parameters, structural attributes, abnormal phenomena and actual operation and maintenance experience. Each expert rule includes: rule number, applicable object, triggering conditions, corresponding fault type, risk judgment conditions and recommended handling measures.

[0039] The specific steps for retrieving historical fault cases include: Step 1: Place the first The state features corresponding to each historical failure case are represented as a historical case state vector. Calculate the current multimodal fusion state vector With the historical case state vector Cosine similarity between , represented as: ; in, Indicates the current operating status of the drainage network and the first Cosine similarity between historical failure cases; Step 2: Based on the cosine similarity Historical failure cases are sorted to obtain those with similarity that meet preset conditions; the search results are then filtered by combining structured information such as pipe diameter, pipe material, pipe network type, and candidate failure type. Step 3: Obtain the basic attribute information of the pipeline network corresponding to the current diagnostic object from the pipeline network basic database, and form engineering knowledge query conditions based on the attributes of the current diagnostic object, the current abnormal status and the historical case search results; Step 4: To ensure that the scope of influence in the diagnostic results is based on the actual spatial structure of the drainage network, according to the current candidate fault objects... Determining the set of candidate influencing objects by identifying connectivity relationships within the GIS spatial topology. For blockage-related faults, the associated nodes and pipe segments within the preset topology range upstream of the current faulty pipe segment are searched first according to the drainage flow direction; for rupture, leakage, settlement or structural damage, the corresponding candidate affected objects can be determined according to the upstream and downstream connection relationship and spatial adjacency relationship of the corresponding fault object. Step 5: The set of candidate affected objects As a spatial constraint condition when analyzing the scope of fault impact in generative diagnostic models, it prevents generative models from generating non-existent impact objects that deviate from the actual GIS pipeline topology. Step 6: Based on the obtained historical fault cases, retrieve engineering knowledge and expert rules related to the current operation status of the drainage network from the engineering knowledge base and expert rule base, and combine the current operation status of the drainage network, the retrieved historical fault cases, engineering knowledge and expert rules into a search enhancement context.

[0040] The diagnostic prompt information structure includes: Diagnostic task: Determine the fault type, risk level, scope of impact, and maintenance and disposal plan based on the current multimodal information of the drainage network.

[0041] Current pipeline network object: Input Information such as pipe diameter, pipe material, burial depth, slope, and upstream and downstream connections.

[0042] Current operating status: The main abnormal states reflected in the input multimodal trusted fusion results.

[0043] Data quality information: Input the overall quality score corresponding to different modalities. and credible weight It is used to characterize the credibility of different diagnostic evidence.

[0044] Similar historical fault cases: Input the abnormal phenomena, actual fault results, maintenance measures, and maintenance results of historical fault cases obtained through retrieval.

[0045] Relevant engineering knowledge: Input relevant knowledge retrieved from the engineering knowledge base.

[0046] Expert rule: Enter the expert rule that currently meets the triggering condition.

[0047] Candidate Impact Objects: Input a set of candidate impact objects determined by the GIS spatial topology. .

[0048] Output requirements: The generative diagnostic model should output the fault type, risk level, scope of impact, and maintenance and handling plan, and perform diagnosis based on the provided operating status, cases, knowledge, and rules.

[0049] Step S150: Load the generative diagnostic model, perform diagnostic reasoning on the diagnostic prompts, and output structured diagnostic results; the structured diagnostic results include: fault type, risk level, scope of impact, and maintenance and handling plan; Specifically, the S140 diagnostic prompt information is input into the generative diagnostic model, which adopts a large language model adapted from the corpus of drainage pipe network related engineering. During the diagnostic process, the generative diagnostic model first analyzes candidate fault types by combining the current operating status and the reliability of each mode. Then, it makes a comprehensive judgment by combining retrieved historical fault cases, engineering knowledge, and expert rules. Based on the current degree of anomaly, the scope of impact, and the development trend of the operating status, it determines the risk level and finally combines the set of candidate affected objects. Repair and handling plans are formulated based on historical repair experience and expert rules; The fault type is used to characterize one or more fault types among pipeline rupture, blockage, leakage, settlement, and structural damage corresponding to the current pipeline network object; The risk level is determined based on the current degree of anomaly, the scope of the fault's impact, and the development trend of the operational status. It can be divided into three levels: low risk, medium risk, and high risk. The scope of influence is determined based on the faulty pipeline node or segment and its connection relationship in the pipeline spatial topology, and the corresponding results are obtained from the set of candidate affected objects; The repair and handling plan is determined based on the fault type, risk level, scope of impact, retrieved historical fault cases, engineering knowledge, and expert rules, and includes at least repair priority and recommended handling measures; The structured diagnostic results can be output in JSON, database fields, or other machine-parseable data formats for writing into the drainage network operation and maintenance management platform.

[0050] Step S160: Obtain the diagnostic feedback data corresponding to the structured diagnostic results and associate them to form a diagnostic feedback sample. Correct and update the historical fault case library, engineering knowledge base, multimodal data quality control model and generative diagnostic model. The diagnostic feedback data includes: manual verification information, on-site maintenance feedback data and drainage network operation status data after diagnosis. Specifically, the manual verification information, on-site maintenance feedback data, and post-diagnosis drainage network operation status data corresponding to the structured diagnostic results are obtained. The manual verification results, actual maintenance results, and post-maintenance drainage network operation status are then correlated with the corresponding structured diagnostic results to form a diagnostic feedback sample.

[0051] The revised and updated historical fault case library, engineering knowledge base, multimodal data quality control model, and generative diagnostic model include: 1) Based on actual fault information verified by manual checks or on-site maintenance, correct the corresponding historical fault cases in the historical fault case database, including: When the actual fault matches the diagnosis result, the current diagnosis event is written as a verified case into the historical fault case library or a corresponding verification record is added. When the actual fault differs from the diagnostic results, the corresponding fault label is corrected based on the actual on-site results, while the original diagnostic results are retained for subsequent analysis of diagnostic deviations.

[0052] 2) Diagnostic feedback samples that meet the preset conditions are added as new historical fault cases and written into the historical fault case library; the engineering knowledge base is updated according to the actual maintenance and handling results. When a certain maintenance measure is verified on site to effectively improve the corresponding fault state, the corresponding maintenance method, applicable fault type, applicable conditions and maintenance effect are added to the engineering knowledge base as new engineering experience.

[0053] 3) Update the evaluation parameters of the multimodal data quality control model and the generative diagnostic model based on the diagnostic feedback samples, including: as the diagnostic feedback samples verified in the field accumulate, analyze the relationship between different quality indicators and the actual diagnostic effect, adjust the evaluation indicator weights and confidence weight adjustment parameters, and use parameter search or constraint optimization methods based on historical verification samples for parameter updates.

[0054] 4) Compile the current status, search context, actual fault results and actual repair effects verified manually or on-site into diagnostic feedback samples. After accumulating a preset number, use the above samples to periodically optimize the generative diagnostic model, or update the diagnostic prompt template, historical case search conditions and expert rules, so that subsequent diagnostic processes can be optimized based on newly added diagnostic feedback samples.

[0055] The present invention also provides a multimodal data quality control driven intelligent agent for drainage network diagnosis, used to execute the aforementioned multimodal data quality control driven drainage network diagnosis method; like Figure 2 As shown, the multimodal data quality control-driven intelligent agent for drainage network diagnosis is not composed of a single fault identification model, but rather is an intelligent diagnostic agent built around the drainage network diagnosis task, possessing capabilities such as environmental object cognition, multimodal data perception, data quality evaluation, reliable state representation, case and knowledge retrieval, intelligent diagnostic reasoning, and feedback learning; including: The data acquisition layer is used to collect time-series data from sensor monitoring, GIS spatial topology attribute data, pipeline CCTV visual inspection data, and operation and maintenance knowledge text data.

[0056] The multimodal data perception module is used to acquire and preprocess multimodal data of the drainage network in the data acquisition layer, establish the spatiotemporal correlation between the multimodal data and the corresponding network objects, and form a multimodal correlation data set. Specifically, drainage network objects are constructed based on GIS spatial topology attribute data. The spatial topology structure of the drainage network is established according to the nodes, segments, and their connections, and is represented as follows: ; in, Represents the set of nodes in the drainage pipe network. This indicates the pipe segments and their connection relationships between different nodes; the nodes include: inspection wells, rainwater inlets, pumping stations, and other drainage network connection facilities; the pipe segments represent the drainage pipes connecting the corresponding nodes.

[0057] Associate pipe diameter, pipe material, length, burial depth, slope, construction time, renovation time, spatial coordinates, and other attributes related to drainage network diagnosis with corresponding nodes or pipe segments; establish a unique network object identifier for each node and pipe segment in the drainage network, represented as: ; in, Indicates the pipeline object index, the It is used to uniquely identify the corresponding drainage network node or pipe segment, and serves as a unified index for establishing object associations between different data sources.

[0058] For a given pipe segment, you can associate it with attributes such as upstream nodes, downstream nodes, pipe diameter, pipe material, length, burial depth, slope, construction time, and spatial location.

[0059] Preprocessing the sensor monitoring time-series data and establishing spatiotemporal correlations includes: 1) Set the sensor monitoring time series data parameters: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] Each diagnostic time window is denoted as: ;No. The pipeline object in the first The sensor monitoring time series data corresponding to each diagnostic time window are recorded as follows: ;in This represents sensor detection data; The time window is a preset continuous time interval on the time axis; for example, the diagnosis time is... The starting point of the window is The current diagnostic time window is ; The time window length, i.e. Preset parameters; diagnostic time It is a point on the timeline, at which the diagnostic mechanism is triggered; the time window is an interval that includes that point; the window is sliding and continuously updates as the diagnostic process progresses.

[0060] 2) Divide the data of different monitoring devices according to the preset time window, and perform time synchronization, anomaly detection, missing data repair and data standardization.

[0061] 3) Based on the object identifier and spatial location of the drainage network, the time-series status data is mapped to the corresponding network node or pipe segment.

[0062] Preprocessing the GIS spatial topology attribute data and establishing spatiotemporal correlations includes: 1) Set GIS spatial topology attribute data parameters: The GIS spatial topology attribute data corresponding to each pipeline network object is denoted as: ; in, This represents GIS spatial topology attribute data.

[0063] 2) Map the pipe diameter, material, length, burial depth, slope, spatial location and upstream and downstream connection relationship in the GIS spatial topology attribute data to the corresponding pipe network object.

[0064] Preprocessing the CCTV visual inspection data of the pipeline and establishing spatiotemporal correlations includes: 1) Set the parameters for CCTV visual inspection of the pipeline: The CCTV visual inspection data of the pipeline corresponding to each pipeline network object is denoted as follows: ; in, Representing visual data, This indicates the temporal relationship established between the visual information and the current diagnostic time window.

[0065] 2) For the original pipeline CCTV image or video, first perform image enhancement, effective area extraction and defect detection.

[0066] 3) Visual inspection can employ visual processing methods such as object detection, image classification, semantic segmentation, convolutional neural networks, and visual Transformers to obtain information about pipeline defects.

[0067] 4) The extracted pipeline defect information includes: cracks, deformation, leakage, sediment, obstructions, abnormal interfaces, corrosion and other structural or functional defects.

[0068] 5) Each pipeline CCTV inspection result saves the corresponding pipeline network object identifier, inspection time, inspection location, and pipeline defect information; for the same pipeline network object with multiple pipeline CCTV inspection records, the time information corresponding to different inspection records is retained, so that the data quality control module can further determine the time validity of historical visual inspection results relative to the current diagnostic task.

[0069] Preprocessing the maintenance text data and establishing spatiotemporal correlations includes: 1) Perform text parsing and semantic information extraction on the aforementioned operation and maintenance knowledge text data; 2) When there are multiple maintenance records for the same pipeline object, first filter them according to their relevance to the current diagnostic task and time distance, and then aggregate the effective text features.

[0070] 3) Associate the obtained fault description information and maintenance experience information with the corresponding pipeline nodes or pipe sections and the corresponding time windows.

[0071] The data quality control module is used to evaluate the quality of multimodal associated data sets, determine data integrity indicators, data reliability indicators, data timeliness indicators, and cross-modal consistency indicators, and construct a comprehensive quality score and credibility weight.

[0072] The multimodal trusted fusion module is used to extract modal features from multimodal data, and dynamically weight and fuse the modal features according to the trusted weights to obtain a multimodal fusion state vector; Specifically, time-series state feature vectors, pipeline spatial topology feature vectors, pipeline defect visual feature vectors, and textual semantic feature vectors are extracted from sensor monitoring time-series data, GIS spatial topology attribute data, pipeline CCTV visual inspection data, and operation and maintenance knowledge text data, respectively, based on the aforementioned confidence weights. Dynamic weighted fusion of different modal features reduces the contribution of modal data with lower overall quality scores to the fusion result; The intelligent diagnostic reasoning module retrieves historical fault cases, acquires engineering knowledge and expert rules based on the multimodal fusion state vector, generates diagnostic prompts, and outputs structured diagnostic results using a generative diagnostic model. Specifically, the multimodal fusion state vector As a query feature, historical fault cases that match the current operation status of the drainage network are retrieved from the historical fault case database. Engineering knowledge and expert rules related to the current operation status of the drainage network are obtained by combining the engineering knowledge base and expert rule base. Diagnostic prompt information is constructed based on the current operation status of the drainage network, the retrieved historical fault cases, engineering knowledge and expert rules. A structured diagnostic result including fault type, risk level, scope of impact and maintenance and disposal plan is generated through a generative diagnostic model. The feedback learning module is used to acquire manual verification information, on-site maintenance feedback data, and drainage network operation status data, and to update the historical fault case library, engineering knowledge base, multimodal data quality control model, and generative diagnostic model. Specifically, it is used to acquire manual verification information, on-site maintenance feedback data, and post-diagnosis drainage network operation status data, and to update the historical fault case library, engineering knowledge base, data quality evaluation model, and generative diagnostic model based on the acquired data.

[0073] The diagnostic intelligent agent described in this invention also adopts a deployment architecture that coordinates the edge side and the cloud; edge computing devices are set up in drainage pumping stations, monitoring sites, or regional control nodes; The edge side is mainly used to acquire real-time sensor monitoring data and complete time synchronization, basic data cleaning, anomaly detection, missing data statistics, data standardization, and operational status monitoring. When the edge side detects an anomaly, reaches a preset diagnostic cycle, or receives an active diagnostic request, it identifies the current pipeline network object. Real-time sensing status and related data are transmitted to the cloud; cloud-based The system acquires corresponding GIS spatial topology attribute data, pipeline CCTV visual inspection data, and operation and maintenance knowledge text data, and establishes a multimodal associated data set corresponding to the current diagnostic task. Subsequently, the cloud sequentially completes multimodal data quality evaluation, trusted weight generation, multimodal feature extraction, trusted state fusion, historical fault case retrieval, engineering knowledge retrieval, expert rule matching, fault impact range analysis, generative diagnostic reasoning, and structured diagnostic result generation. After the diagnosis is completed in the cloud, the fault type, risk level, scope of impact, and maintenance plan are sent to the edge side. When the diagnosis results meet the preset alarm conditions, corresponding alarm information and operation and maintenance tasks are generated. After the on-site inspection and maintenance are completed, the manual verification information, on-site maintenance data, and post-maintenance operation status are resent to the feedback learning module to update the historical fault case library, engineering knowledge base, and related models.

[0074] Different functional modules in a diagnostic agent can be deployed on the same electronic device, or they can be deployed on edge devices and cloud servers respectively, depending on data real-time requirements, computing resources, and network conditions.

[0075] According to this invention, a multimodal data quality control model is constructed to evaluate the completeness, reliability, timeliness, and cross-modal consistency of different modal data involved in the current diagnosis in four dimensions. This yields a comprehensive quality score and credibility weight for each modal data, improving the credibility of different data sources and avoiding interference from low-quality data in diagnostic judgments. Furthermore, a multimodal data association relationship is established based on pipeline network objects and time windows, enabling joint analysis of multimodal data around the same drainage pipeline network object. The contribution of different modal characteristics to the comprehensive state representation is dynamically adjusted according to the credibility weight, effectively suppressing the impact of missing data, anomalies, insufficient timeliness, or cross-modal conflicts on diagnostic results, and significantly improving the reliability of multimodal state representation. This is further enhanced by utilizing historical fault cases. The database, engineering knowledge base, and expert rule base provide external information relevant to the current diagnostic task for the generative diagnostic model. Combined with the spatial topology of the pipeline network, it determines the scope of fault impact and outputs fault type, risk level, and maintenance disposal plan including maintenance priority and recommended disposal measures. This enhances the engineering basis, interpretability, and operation and maintenance decision support capabilities of the diagnostic results. By feeding back manual verification, on-site maintenance results, and post-maintenance operation status to the historical fault case database, engineering knowledge base, and related models, a continuous optimization closed loop is formed that connects diagnosis, verification, maintenance, and learning. This enables continuous improvement in the accuracy of pipeline network anomaly identification, diagnostic reliability, and interpretability of results in complex environments, providing technical support for risk early warning and intelligent operation and maintenance of urban drainage facilities.

[0076] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A multimodal data quality control-driven diagnostic method for drainage pipe networks, characterized in that, include: Acquire multi-source heterogeneous data of drainage pipe network and preprocess it to generate multimodal data; The multi-source heterogeneous data includes: sensor monitoring time series data, GIS spatial topology attribute data, pipeline CCTV visual inspection data, and operation and maintenance knowledge text data; Based on GIS spatial topology attribute data, a pipeline spatial topology structure is constructed, the spatiotemporal correlation of multimodal data is established, and a multimodal correlation data set is generated. Constructing a multimodal data quality control model includes: evaluating the quality of the multimodal associated data set; determining data integrity indicators, data reliability indicators, data timeliness indicators, and cross-modal consistency indicators based on a preset time window; and calculating a comprehensive quality score and a credibility weight; wherein the preset time window is a continuous time interval on the time axis. Modal features of the multimodal data are extracted, and the features are dynamically weighted and fused according to the trusted weights to obtain a multimodal fused state vector representing the current operating state of the drainage network. Based on the multimodal fusion state vector, historical fault cases are retrieved, engineering knowledge and expert rules are obtained, and diagnostic prompts are generated. A generative diagnostic model is loaded, which performs diagnostic reasoning on diagnostic prompts and outputs structured diagnostic results. The structured diagnostic results include: fault type, risk level, scope of impact, and maintenance and handling plan. The diagnostic feedback data corresponding to the structured diagnostic results is obtained and correlated to form a diagnostic feedback sample. The historical fault case library, engineering knowledge base, multimodal data quality control model and generative diagnostic model are then corrected and updated. The diagnostic feedback data includes: manual verification information, on-site maintenance feedback data and the operational status data of the drainage network after diagnosis.

2. The multimodal data quality control-driven drainage network diagnostic method according to claim 1, characterized in that, The preprocessing includes: The sensing monitoring time series data is divided according to a preset time window, and the sensing monitoring time series data collected by different sensing devices are subjected to time synchronization, anomaly detection, missing data repair and data standardization preprocessing. The time synchronization refers to the process of converting different sensor monitoring time-series data with different sampling frequencies to a unified time axis by resampling, linear interpolation, and spline interpolation. The anomaly detection refers to the use of sliding window statistics, Z-score, interquartile range, and isolated forest to identify anomaly sensing and monitoring time-series data, which serves as the basis for subsequent data reliability evaluation. The missing data repair refers to: for short-term missing sensor monitoring time-series data, interpolation repair is performed using adjacent valid data; for long-term continuous missing sensor monitoring time-series data, no forced filling is performed, and the corresponding missing state is retained for use in subsequent data quality control steps to reduce the data integrity index corresponding to that mode. The data standardization refers to the elimination of dimensional differences between different sensor monitoring time series data by using Z-score standardization or min-max normalization methods. The pipeline CCTV visual inspection data is subjected to image enhancement, target area detection, and defect feature extraction to generate pipeline defect information; the pipeline defect information includes: cracks, deformation, leakage, deposits, obstacles, interface abnormalities, and corrosion; The operation and maintenance knowledge text data is parsed and semantic information is extracted to generate fault description information, maintenance experience information and engineering knowledge semantic information.

3. The multimodal data quality control-driven drainage network diagnostic method according to claim 1, characterized in that, The construction of the pipeline network spatial topology includes: Based on the GIS spatial topology attribute data, drainage network nodes, pipe segments, and their connection relationships are obtained, and a spatial topology structure of the network is constructed, including: treating network nodes as graph structure nodes, pipe segment connection relationships as graph structure edges, and associating network spatial attributes with corresponding graph structure nodes or graph structure edges, as shown below: ; in, Represents the set of nodes in the drainage pipe network. This indicates the pipe segments and their connection relationships between different nodes; the nodes include: inspection wells, rainwater inlets, pumping stations, and other drainage network connection facilities; the pipe segments represent the drainage pipes connecting the corresponding nodes; Establish corresponding pipeline object identifiers based on pipeline nodes and pipeline segments, which are used to determine the specific pipeline objects corresponding to data from different sources during subsequent data processing.

4. The multimodal data quality control-driven drainage network diagnostic method according to claim 1, characterized in that, The establishment of spatiotemporal correlations for multimodal data includes: Establish the association between GIS spatial topology attribute data and pipeline network objects, and then... The GIS spatial topology attribute data corresponding to each pipeline network object is represented as follows: ;in, Represents GIS spatial topological attribute data; a unified index based on object associations. Map the pipe diameter, material, length, burial depth, slope, spatial location, and upstream and downstream connection relationships in the GIS spatial topology attribute data to the corresponding pipe network objects; Based on the multimodal data, time-series status data corresponding to pipeline nodes or pipe segments are obtained. According to the pipeline object identifier and spatial location, the time-series status data is mapped to the corresponding pipeline node or pipe segment. The time-series status data includes: liquid level, flow rate, flow velocity, pressure, water quality, and pump station operating status. Associate pipeline defect information with the corresponding pipe section and the corresponding preset time window; Link fault description information and maintenance experience information to the corresponding pipeline nodes or pipe sections and the corresponding preset time windows; Based on the multimodal data, cross-modal association relationships are established with pipeline network objects and preset time windows to form a multimodal association data set for data quality evaluation and multimodal feature fusion.

5. The multimodal data quality control-driven drainage network diagnostic method according to claim 1, characterized in that, The comprehensive quality score is based on the current diagnosis. The calculation is performed on modal data, and the calculation method is expressed as follows: ； in, Indicates the first The overall quality score of modal data represents the first... Data integrity metrics for modal data Indicates the first Data reliability metrics for modal data Indicates the first Data timeliness indicators for modal data Indicates the first Cross-modal consistency metrics for modal data; , , , The corresponding evaluation index weights, satisfying the constraints, are expressed as follows: ， ； No. Confidential weights corresponding to modal data The overall quality score of various modal data involved in the current diagnosis is used. Confirmed, indicated as: ； in, This represents the total number of valid data modalities participating in the current diagnosis. Let be the confidence weight adjustment parameter, and: ， 。 6. The multimodal data quality control-driven drainage network diagnostic method according to claim 1, characterized in that, The data integrity indicator refers to: based on the data integrity index within the preset time window... Determining data integrity metrics based on missing data in modal data , used to characterize the The degree of effective data completeness of modal data; The data reliability index refers to: based on the preset time window... Anomaly detection results for modal data determine data reliability indicators , used to characterize the The degree of credibility of valid information in modal data, and the corresponding data reliability index as the degree of data anomaly increases. The lower; The data timeliness indicator refers to: based on the preset time window... The time difference between the effective data acquisition time of modal data and the current diagnostic time determines the data timeliness index. , used to characterize the The temporal validity of modal data relative to the current diagnostic time; The cross-modal consistency index refers to: based on the first [time window] belonging to the same pipeline network object and the corresponding preset time window The degree of consistency between modal data and other related modal data in describing the operating status of the same pipeline network is used to determine the cross-modal consistency index. .

7. The multimodal data quality control-driven drainage network diagnostic method according to claim 1, characterized in that, The dynamic weighted fusion includes: The modal features of the multimodal data are uniformly represented and corresponding feature mapping functions are set for each, transforming the modal features to a unified representation. 3D feature space; for the current diagnosis of the th Modal data, represented as: ； in, Indicates the first The feature mapping function corresponding to the modality consists of a fully connected layer, layer normalization, and a nonlinear activation function. The first to participate in the current diagnosis The modal feature vector corresponding to the modal data is represented as follows: Determine the multimodal fusion state vector corresponding to the current operating state of the drainage pipe network. , is represented as: ； in, For the first The confidence weights corresponding to modal data The total number of valid data modalities participating in the current diagnosis; To achieve data quality-driven trusted fusion, a trusted weighting approach is adopted. Dynamically weighted and fused multimodal features currently involved in diagnosis are represented as follows: ； in, Indicates the first The pipeline object in the first The multimodal fusion state vector corresponding to each diagnostic time window; Indicates the first The feature vector corresponding to the class mode; Indicates the first The credibility weights corresponding to the modalities.

8. The multimodal data quality control-driven drainage network diagnostic method according to claim 1, characterized in that, The historical failure cases mentioned include: The first The state features corresponding to each historical failure case are represented as a historical case state vector. Calculate the current multimodal fusion state vector With the historical case state vector Cosine similarity between , is represented as: ； in, Indicates the current operating status of the drainage network and the first Cosine similarity between historical failure cases; Based on the cosine similarity The historical failure cases are sorted to obtain historical failure cases whose similarity meets the preset conditions; Based on the obtained historical fault cases, engineering knowledge and expert rules related to the current operation status of the drainage network are obtained from the engineering knowledge base and expert rule base. The current operation status of the drainage network, the retrieved historical fault cases, engineering knowledge and expert rules are combined to form a search enhancement context.

9. The multimodal data quality control-driven drainage network diagnostic method according to claim 1, characterized in that, The diagnostic reasoning includes: inputting diagnostic prompts into the generative diagnostic model, and having the generative diagnostic model combine the current operating status of the drainage network with retrieved historical fault cases, engineering knowledge and expert rules to perform diagnostic reasoning and generate the structured diagnostic results; The fault type refers to one or more of the following fault types used to characterize the current pipeline network object: pipeline rupture, blockage, leakage, settlement, and structural damage. The risk level refers to the level determined based on the current degree of anomaly, the scope of the fault's impact, and the development trend of the operational status. The scope of influence refers to the range determined based on the faulty pipeline node or segment and its connection relationship in the pipeline network spatial topology. The maintenance and handling plan refers to a plan determined based on the fault type, risk level, scope of impact, retrieved historical fault cases, engineering knowledge, and expert rules, and includes at least maintenance priority and recommended handling measures.

10. A multimodal data quality control driven intelligent agent for drainage network diagnosis, characterized in that, A drainage network diagnostic method driven by multimodal data quality control as described in any one of claims 1-9, comprising: The multimodal data perception module is used to acquire and preprocess multimodal data of the drainage network, establish the spatiotemporal correlation between the multimodal data and the corresponding network objects, and form a multimodal correlation data set. The data quality control module is used to evaluate the quality of multimodal associated data sets, determine data integrity indicators, data reliability indicators, data timeliness indicators, and cross-modal consistency indicators, and construct a comprehensive quality score and credibility weight. The multimodal trusted fusion module is used to extract modal features from multimodal data, and dynamically weight and fuse the modal features according to the trusted weights to obtain a multimodal fusion state vector; The intelligent diagnostic reasoning module retrieves historical fault cases, acquires engineering knowledge and expert rules based on the multimodal fusion state vector, generates diagnostic prompts, and outputs structured diagnostic results using a generative diagnostic model. The feedback learning module is used to acquire manual verification information, on-site maintenance feedback data, and drainage network operation status data, and to update the historical fault case library, engineering knowledge base, multimodal data quality control model, and generative diagnostic model.

Images