A method for tracing the source of pollution in stormwater pipe networks by combining three-dimensional fluorescence fingerprinting and graph neural networks.

Abstract

This invention discloses a method for tracing the source of pollution in stormwater pipe networks by combining three-dimensional fluorescence fingerprinting and graph neural networks, relating to the field of urban stormwater system pollution control technology. The method collects spectral data of the pipe network by deploying a fluorescence spectral sensor array to construct an incrementally updatable fluorescence database; establishes a graph model based on the pipe network topology and hydraulic parameters, and embeds the spatiotemporal features of nodes / pipe segments using a GNN; extracts the spatial features and temporal series correlations of the spectra using a hybrid CNN and LSTM model; calculates the probability distribution of pollution sources through database matching and model inference, and outputs the pollution source type, location, and confidence assessment of the mixed connection path based on the Bayesian inference optimization results. The method introduces anomaly sample identification and multi-source evidence fusion to maintain robustness under complex operating conditions; the results are presented in a tiered alarm and visualization report. This invention constructs a perception-analysis-response closed-loop system to achieve rapid and accurate source tracing of sudden mixed connection pollution.

Description

Technical Field

[0001] This invention relates to the technical field of urban stormwater system pollution control, specifically to a method for tracing the source of stormwater network pollution by combining three-dimensional fluorescent fingerprinting and graph neural networks. Background Technology

[0002] Pollution control in urban stormwater systems is a crucial research area in smart water management and urban sustainable development. Its core lies in identifying and regulating the sources, migration paths, and discharge dynamics of pollutants during stormwater runoff. With rapid urbanization, non-point source pollution loads have increased significantly, and problems such as illegal discharge from sewage pipes into stormwater pipes and initial rainwater runoff are frequent. This results in large amounts of pollutants being discharged into natural water bodies during rainfall, seriously threatening the water quality safety and ecological stability of receiving water bodies.

[0003] In existing technologies, tracing the source of pollution from mixed stormwater pipe networks mainly relies on fixed water quality sampling points and regular manual inspections. This involves detecting changes in key physicochemical indicators such as chemical oxygen demand (COD), ammonia nitrogen concentration, and total phosphorus, and combining these with hydraulic models or finite inversion algorithms to infer the source. Some schemes employ Bayesian networks, the backward tracking matrix method, or the shortest path inversion algorithm for computational reasoning. Static matching methods based on typical pollutant databases also attempt to trace the source using fingerprint identification. However, these technologies generally face problems such as insufficient sensor node coverage, difficulty in describing dynamic changes in pollution propagation paths, fuzzy pollution source characteristics, and high requirements for manual intervention. Furthermore, it is difficult to balance response speed and model accuracy. Changes in hydraulic conditions lead to interactions between pollutants and pipe sediments, resulting in complex signal superposition, distortion of fluorescence peak positions and morphologies, and failure of static database matching methods. These technologies struggle to meet the real-time and accuracy requirements for tracing urban overflow pollution sources under high-frequency rainfall scenarios.

[0004] To address this, we propose a machine learning-based source tracing method for stormwater pipe network pollution that incorporates a dynamic fluorescence spectral library. Summary of the Invention

[0005] The purpose of this invention is to provide a method for tracing the source of rainwater network pollution by combining three-dimensional fluorescent fingerprinting and graph neural networks, which addresses the problems of slow response, low accuracy, and insufficient robustness in existing technologies when facing complex pipeline network conditions, dynamic pollution characteristics, and multi-source mixed connection scenarios.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for tracing the source of pollution in stormwater pipe networks by combining three-dimensional fluorescent fingerprinting and graph neural networks includes the following steps: Step S1: By deploying fluorescence spectral sensor arrays at key nodes of the rainwater pipe network, fluorescence spectral data of water bodies in the pipe network are collected in real time and the data is synchronously transmitted to the cloud processing platform. Step S2: Construct and incrementally update the dynamic fluorescence database to store and index pollutant characteristic fingerprint spectra. The incremental update is based on the distribution offset between real-time data and historical data and is triggered when a new pollution characteristic is detected. Step S3: Use a machine learning model to extract features and recognize patterns from the fluorescence spectral data acquired in real time in step S1. Step S4: Through the synergistic effect of dynamic fluorescence database and machine learning model, combined with pipeline topology constraints, calculate the probability distribution of pollution sources in real time, optimize the source tracing results based on Bayesian inference algorithm, and output the confidence assessment of pollution source type, location and mixed connection path. Step S5: Feed back the source tracing results of step S4 to the regulatory terminal through the visualization interface, trigger the pipeline pollution early warning mechanism, and generate a pollution incident handling suggestion plan simultaneously.

[0007] Preferably, in step S2, a dynamic fluorescence database for non-steady-state pollution events in stormwater pipe networks is constructed. This database characterizes the time-varying evolution of pollutant spectral features under rainfall-induced hydraulic condition abrupt changes, including but not limited to rainfall-induced hydraulic condition abrupt changes and source tracing under normal sunny conditions. By judging the degree of distribution deviation between real-time collected data and historical stable operating condition data, an incremental update process is triggered only when the pollution characteristics are determined to exceed the range of existing stable patterns. Specifically, this includes: Step S21: Perform preprocessing operations for event detection on the newly added fluorescence spectral data to separate the background changes and pollution fingerprint background characteristics caused by external water inflow and infiltration (groundwater infiltration, river water backflow, residual rainwater, etc.), thereby providing a stable reference for subsequent pollution event identification; Step S22: Dynamically classify the pollution source feature fingerprint spectrum based on the unsupervised clustering analysis algorithm, and perform cluster center expansion operation when the following conditions are met simultaneously within the same pollution event cycle: The fingerprint appears repeatedly within multiple non-continuous time windows; Its distance from existing cluster centers in the feature space remains consistently higher than a set threshold; Its occurrence is accompanied by abnormal changes in the hydraulic state of the pipeline network; Step S23: Combining the pipeline network topology and water flow dynamics model, weights are assigned to the spatial correlation of pollution events, and a database index with spatiotemporal attributes is established.

[0008] Preferably, in the data cleaning stage of the preprocessing operation in step S21, a dual denoising strategy combining wavelet transform and Savitzky-Golay filtering is adopted, which includes an energy threshold function for high-frequency noise and an adaptive window smoothing algorithm for baseline drift. The wavelet transform is performed using the Daubechies-4 wavelet basis for multi-scale decomposition. Let the original fluorescence spectral signal be... The high-frequency noise threshold uses a threshold function based on energy distribution: ; in These are wavelet high-frequency coefficients. The empirical adjustment factor is N, where N is the number of coefficients. In baseline drift filtering, the Savitzky-Golay filter window width is adaptively adjusted according to signal stationarity to ensure curve shape preservation and smoothness. The pollution source feature anomaly detection mechanism combines local density ratio to calculate the isolation degree of the pollution source fingerprint vector, defining the frequency anomaly scoring function as follows: ; in Let i be the frequency of occurrence of the i-th type of contaminated fingerprint within the current time window. For its local density estimation in the feature space, when If the threshold is exceeded, it is automatically marked as a high-risk pollution source and manual review is triggered; the graph embedding process is based on the node vectors trained by the GNN model. By using cosine similarity for fast path matching and dynamically updating path retrieval weights, the accuracy of identifying contaminated paths can be improved. In step S22, the dynamic classification of the pollution source feature fingerprint spectrum adopts the K-means clustering algorithm, and a frequency anomaly detection mechanism is introduced in the process: when a certain type of pollution source feature appears abnormally frequently under low rainfall background conditions, and its frequency growth trend is inconsistent with the changes in the hydraulic parameters of the pipeline network, the system determines that the feature does not belong to the hydraulically driven background disturbance, and marks it as a high-risk new pollution source to trigger the manual review process. After the review is passed, the feature is included in the standard database. In step S23, a graph neural network (GNN) model based on the pipeline network topology is used to map the pollution propagation process into a path state space constrained by the direction of water flow and the connectivity of nodes. This is used to exclude propagation paths that are physically inaccessible under the current hydraulic conditions when multiple candidate source tracing paths exist. Step S2 also includes establishing a database version control mechanism, retaining historical version data after each update, and automatically triggering a rollback command to the previous stable version and recalculating when the confidence level of the tracing result is lower than a preset threshold.

[0009] Preferably, in step S3, the machine learning model adopts a hybrid architecture of convolutional neural network (CNN) and long short-term memory network (LSTM). The CNN module is activated only when a spectral morphological change segment is detected to capture the local fingerprint structure at the moment of triggering the pollution event. The LSTM module is used to model the evolution trajectory of the pollution fingerprint with changes in hydraulic conditions within the time window before and after the event to distinguish between transient noise disturbances and real pollution propagation behavior. The feature extraction process of the machine learning model includes the following steps: Step S31: Standardize the fluorescence spectral data, use Z-score normalization to eliminate dimensional differences between sensors, and remove data redundancy through principal component analysis (PCA) to map the high-dimensional spectral signal to a compact feature space. Step S32: Construct a feature extraction layer containing a multi-branch parallel convolution kernel group, including three sizes: 1×3, 1×5, and 1×7, to simultaneously capture the fine morphological features of local peaks and global trend features in spectral data. Step S33: Introduce an attention mechanism into the LSTM network layer to strengthen the response weights at key time points before and after the pollution event, so as to avoid the dilution effect of historical data under long-term stable flow conditions on the characteristics of sudden pollution events. Step S34: Using a transfer learning strategy, the model parameters pre-trained on pipeline network data in other regions are used as initial values ​​and fine-tuned using local measured data to achieve rapid model adaptation. Step S35: Set up the online incremental learning module for the model and set the traceability deviation monitoring threshold. When the deviation between the traceability result and the manual verification exceeds the threshold, the incremental training process will be started automatically and the model parameters will be updated.

[0010] Preferably, in step S33, the attention mechanism employs a time-weighted function based on contamination evolution dynamics to assign differentiated attention to the hidden states output at each time step of the LSTM network, defining the attention weight vector as follows: ; in Let represent the hidden state at time step t. Let ψ be the context vector, and ψ be a scoring function constructed based on a bidirectional GRU to characterize the correlation between temporal variation features and the pollution fingerprint response. This mechanism is used to enhance the model's response sensitivity to pollution outbreak nodes. The online learning module is triggered by an error-driven incremental update strategy. Let the current source tracing prediction result be... The manual verification label is y, and its loss function is... Exceeding the dynamic threshold The fine-tuning process is initiated at that time, in which The standard deviation of recent sample loss is γ, which is an adjustment factor. This mechanism ensures that the model maintains continuous adaptability and source tracing accuracy under pollution type drift and spectral changes.

[0011] Preferably, the method for calculating the pollution source probability distribution map in step S4 further includes: Step S41: Simulate the transmission and diffusion process of pollutants based on the pipeline hydraulic model, and correct the propagation velocity parameters by combining real-time flow data; Step S42: Generate a candidate set of pollution sources using the Monte Carlo method, and optimize the convergence efficiency of candidate solutions through Markov Chain Monte Carlo (MCMC) sampling; Step S43: Construct a multi-objective optimization function to establish consistency constraints among candidate pollution source locations, occurrence times, and spectral fingerprint matching results, so as to avoid bias in the source tracing results caused by minimizing a single error; Step S44: Introduce evidence theory (DS theory) to fuse multi-sensor data and reduce the impact of single-node data anomalies on the source tracing results; Step S45: The output results include the confidence ellipse region for locating the pollution source, the topological sequence of the most likely confluence paths, and the timeline prediction of the pollution event development.

[0012] Preferably, in step S43, the multi-objective optimization function constructs a joint loss function by fusing spatial error, temporal error, and spectral feature error. ; in Indicates the predicted location of the pollution source The Euclidean distance error between the actual position P and the actual position P. Indicates the predicted time of the event. The absolute difference from the actual time T, This represents the cosine similarity error between the predicted spectral feature vector and the best-matching fingerprint vector in the database. These are the error term weight coefficients; this optimization function guides MCMC sampling to cluster in high-confidence regions, while also incorporating the support functions of each node in DS evidence fusion. Conflict coefficient Confidence distribution correction is performed to make the final pollution source confidence area show a stable convergence trend, significantly reducing the interference of outliers on the shape and path prediction of the distribution map.

[0013] Preferably, the visualization interface and early warning mechanism in step S5 include the following steps: Step S51: Develop a 3D pipeline visualization engine based on the WebGL standard to support dynamic rendering of pollution propagation paths and interactive retrospective analysis of the entire process of historical events. Step S52: Establish a multi-dimensional hierarchical early warning and response mechanism. Based on the confidence interval of the source tracing results, the early warning level is divided into four levels, and a differentiated emergency response time limit and standardized handling plan are set for each level. Step S53: The disposal suggestion generation module integrates the expert system rule base, including pipeline blocking priority assessment, diversion scheme optimization algorithm and emergency resource scheduling model; Step S54: Establish an interpretability report generation system for the source tracing results, and automatically label the database entries and model feature contributions of key decision-making criteria; Step S55: Standardize the data interface with the municipal GIS platform to support cross-verification of traceability results with pipeline maintenance records and sewage discharge permit information.

[0014] Preferably, in step S53, the diversion scheme optimization algorithm in the treatment suggestion generation module is constructed based on a multi-objective constrained programming model, assuming the pollutant concentration is... The capacity of the pipe section is The state of the control valve is a binary variable. The objective function to be minimized is the weighted sum of total pollutant exposure and emergency response costs, i.e., minimizing... ; in These are the weighting factors for response costs and environmental risks, respectively. The function is solved under constraints, including the processing capacity limitations of each node and the flow conservation condition. The final output is the optimal diversion path set that satisfies the controllability of the pipeline network, thereby improving the spatial coordination efficiency of emergency response and suppressing the spread of pollutants to highly sensitive areas.

[0015] Preferably, in the interpretability report generation system of step S54, the model feature contribution adopts a quantitative evaluation method based on SHAP values, and a feature input vector is provided. The model outputs the prediction result as follows: Then each feature The contribution value is defined by the SHAP function as: ; Where S is a subset of features and n is the total number of features, this mechanism enables the quantitative interpretation of key pollution features for source tracing decisions, and generates an evidence chain for tracing paths by combining database entry indexes, thereby improving the transparency and verifiability of early warning results.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention overcomes the limitations of traditional static spectral libraries in dealing with complex pipeline network conditions. By using incremental updates of a dynamic fluorescence fingerprint database in conjunction with a multi-scale machine learning model, the system is endowed with continuous learning capabilities. This not only enables rapid response and accurate location of sudden mixed-connection pollution events, but also significantly reduces the time delay of manual verification and significantly improves the robustness of the model and the efficiency of emergency response during long-term operation.

[0017] 2. This invention uses graph neural networks to deeply model the relationship between the topology of the pipeline network and the dynamics of water flow, and vectorizes the pollution propagation path. Combined with Markov chain Monte Carlo optimization and hydraulic model correction, it effectively solves the path ambiguity problem caused by nonlinear diffusion in complex pipeline networks, and significantly improves the logical rationality and positioning accuracy of the source tracing results in terms of spatial resolution.

[0018] 3. This invention innovatively constructs an interpretable model framework, integrates quantitative analysis of SHAP values ​​with a database entry traceability mechanism, and forms a complete closed-loop chain of pollution characteristics-model reasoning-decision basis by matching actual emission records with the database, supporting the traceability of traceability results as evidence and expert review in the municipal management system.

[0019] 4. This invention, through the linkage of a WebGL 3D visualization engine, a hierarchical early warning mechanism, and a multi-objective diversion optimization algorithm, establishes a data link from bottom-level spectral sensing to top-level emergency response. When facing extreme weather conditions or multi-source complex pollution, the system can automatically generate an optimal disposal plan that balances environmental risks and treatment costs, forming a closed-loop system from sensing and analysis to response, significantly improving the intelligence level and resilience of urban drainage systems in responding to emergencies. Attached Figure Description

[0020] Figure 1 A system architecture diagram of a stormwater pipe network pollution tracing method combining three-dimensional fluorescent fingerprinting and graph neural networks is provided for embodiments of the present invention. Figure 2 A flowchart of the dynamic fluorescence database incremental update process for a stormwater pipe network pollution tracing method combining three-dimensional fluorescence fingerprinting and graph neural networks is provided for embodiments of the present invention. Figure 3 A flowchart for calculating the probability distribution map of pollution sources in a stormwater pipe network pollution tracing method combining three-dimensional fluorescent fingerprinting and graph neural networks is provided as an embodiment of the present invention. Figure 4 This invention provides a visualization interface and early warning mechanism architecture diagram for a rainwater pipe network pollution tracing method that combines three-dimensional fluorescent fingerprinting and graph neural networks, as an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0022] Example 1 like Figures 1 to 4 As shown, this embodiment provides a machine learning-based source tracing method for stormwater pipe network pollution combined with a dynamic fluorescence spectral library, including the following steps: Step S1: By deploying fluorescence spectral sensor arrays at key nodes of the rainwater pipe network, fluorescence spectral data of the water in the pipe network is collected in real time and the data is synchronously transmitted to the cloud processing platform.

[0023] Step S2: Construct a dynamic fluorescence database for non-steady-state pollution events in stormwater pipe networks. This database can be used to characterize the time-varying evolution of pollutant spectral characteristics under rainfall-induced hydraulic condition abrupt changes, including but not limited to rainfall-induced hydraulic condition abrupt changes and source tracing under normal sunny conditions. It determines the degree of distribution deviation between real-time collected data and historical stable operating condition data, triggering an incremental update process only when the pollution characteristics are determined to exceed the range of existing stable patterns. Specifically, this includes: Step S21: Perform preprocessing operations for event detection on the newly added fluorescence spectral data to separate the background changes and pollution fingerprint background characteristics caused by external water inflow and infiltration (groundwater infiltration, river backflow, residual rainwater, etc.), thereby providing a stable reference for subsequent pollution event identification; Step S22: Dynamically classify the pollution source feature fingerprint spectrum based on the K-means clustering algorithm, and perform cluster center expansion operation when the following conditions are met simultaneously within the same pollution event cycle: The fingerprint appears repeatedly within multiple non-continuous time windows; Its distance from existing cluster centers in the feature space remains consistently higher than a set threshold; Its occurrence is accompanied by abnormal changes in the hydraulic state of the pipeline network; Step S23: Combining the pipeline network topology and water flow dynamics model, weights are assigned to the spatial correlation of pollution events to optimize database retrieval efficiency.

[0024] Step S3: Use a machine learning model to extract features and recognize patterns from the real-time acquired fluorescence spectral data. The model adopts a hybrid architecture of convolutional neural network (CNN) and long short-term memory network (LSTM). The CNN module is activated only when it detects abrupt changes in spectral morphology, and is used to capture the local fingerprint structure at the moment of triggering a pollution event. The LSTM module is used to model the evolution trajectory of the pollution fingerprint with changes in hydraulic conditions within the time window before and after the event, so as to distinguish between transient noise disturbances and real pollution propagation behavior.

[0025] Step S4: Through the synergistic effect of dynamic fluorescence database and machine learning model, calculate the probability distribution map of pollution sources in real time, optimize the source tracing results based on Bayesian inference algorithm, and output the confidence assessment of pollution source type, location and mixing path. Step S5: Feed back the source tracing results to the regulatory terminal through a visual interface, trigger the pipeline pollution early warning mechanism, and simultaneously generate a pollution incident handling suggestion plan.

[0026] Example 2 Based on Example 1, the incremental update algorithm for the dynamic fluorescence database in step S2 further includes: In the preprocessing operation of step S21, the data cleaning stage adopts a dual denoising strategy combining wavelet transform and Savitzky-Golay filtering, and designs adaptive thresholds for high-frequency noise and baseline drift respectively. Furthermore, in this embodiment, the wavelet transform in the data cleaning stage uses the Daubechies-4 wavelet basis for multi-scale decomposition. Let the original fluorescence spectral signal be... The high-frequency noise threshold uses a threshold function based on energy distribution: ; in These are wavelet high-frequency coefficients. The empirical adjustment factor is N, where N is the number of coefficients. In baseline drift filtering, the Savitzky-Golay filter window width is adaptively adjusted according to signal stationarity to ensure curve shape preservation and smoothness. The pollution source feature anomaly detection mechanism combines local density ratio to calculate the isolation degree of the pollution source fingerprint vector, defining the frequency anomaly scoring function as follows: ; in Let i be the frequency of occurrence of the i-th type of contaminated fingerprint within the current time window. For its local density estimation in the feature space, when If the threshold is exceeded, it is automatically marked as a high-risk pollution source and manual review is triggered; the graph embedding process is based on the node vectors trained by the GNN model. By using cosine similarity for fast path matching and dynamically updating path retrieval weights, the accuracy of identifying contaminated paths can be improved.

[0027] In step S22, during the dynamic classification of pollution source feature fingerprint spectrum, an anomaly detection mechanism is introduced: when a certain type of pollution source feature appears abnormally frequently under low rainfall background conditions, and its frequency growth trend is inconsistent with the changes in pipeline hydraulic parameters, the system determines that the feature does not belong to hydraulically driven background disturbance, and marks it as a high-risk new pollution source to trigger the manual review process. After the review is passed, the feature is included in the standard database. In step S23, a graph neural network (GNN) model based on the pipeline network topology is used to map the pollution propagation process into a path state space constrained by the direction of water flow and the connectivity of nodes. This is used to exclude propagation paths that are physically unreachable under the current hydraulic conditions when multiple candidate source tracing paths exist. Furthermore, in this embodiment, step S2 also includes establishing a database version control mechanism to retain historical version data after each update. When the confidence level of the tracing result is lower than a preset threshold, it automatically rolls back to the previous stable version and recalculates.

[0028] Example 3 Based on Example 2, the feature extraction process of the machine learning model in step S3 specifically includes: Step S31: Standardize the fluorescence spectral data, use Z-score normalization to eliminate dimensional differences between sensors, and reduce the dimension to the feature space through principal component analysis (PCA). Step S32: Construct a multi-scale convolution kernel group, including three sizes: 1×3, 1×5, and 1×7, to capture local detail features and global trend features of spectral data, respectively. Step S33: Introduce an attention mechanism into the LSTM layer to strengthen the response weights of key time nodes before and after the pollution event, so as to avoid the dilution effect of historical data under long-term stable flow on the characteristics of sudden pollution events. Step S34: Using a transfer learning strategy, the model parameters pre-trained on pipeline network data in other regions are used as initial values, and the model is quickly adapted by fine-tuning with local data. Step S35: Set up the online learning module for the model. When a discrepancy is detected between the traceability results and the manual verification, the incremental training process will be automatically started and the model parameters will be updated.

[0029] The attention mechanism in the feature extraction process employs a time-weighted function based on contamination evolution dynamics to assign differentiated attention levels to the hidden states output at each time step of the LSTM network. The attention weight vector is defined as follows: ; in Let represent the hidden state at time step t. Let ψ be the context vector, and ψ be a scoring function constructed based on a bidirectional GRU to characterize the correlation between temporal variation features and the pollution fingerprint response. This mechanism is used to enhance the model's response sensitivity to pollution outbreak nodes. The online learning module is triggered by an error-driven incremental update strategy. Let the current source tracing prediction result be... The manual verification label is y, and its loss function is... Exceeding the dynamic threshold The fine-tuning process is initiated at that time, in which The standard deviation of recent sample loss is γ, which is an adjustment factor. This mechanism ensures that the model maintains continuous adaptability and source tracing accuracy under pollution type drift and spectral changes.

[0030] Example 4 Based on Example 3, the method for calculating the pollution source probability distribution map in step S4 further includes: Step S41: Simulate the transmission and diffusion process of pollutants based on the pipeline hydraulic model, and correct the propagation velocity parameters by combining real-time flow data; Step S42: Generate a candidate set of pollution sources using the Monte Carlo method, and optimize the convergence efficiency of candidate solutions through Markov Chain Monte Carlo (MCMC) sampling; Step S43: Establish a multi-objective optimization function to establish consistency constraints among candidate pollution source locations, occurrence times, and spectral fingerprint matching results, so as to avoid bias in the source tracing results caused by minimizing a single error; Furthermore, in this embodiment, the multi-objective optimization function constructs a joint loss function by fusing spatial error, temporal error, and spectral feature error. ; in Indicates the predicted location of the pollution source The Euclidean distance error between the actual position P and the actual position P. Indicates the predicted time of the event. The absolute difference from the actual time T, This represents the cosine similarity error between the predicted spectral feature vector and the best-matching fingerprint vector in the database. These are the error term weight coefficients; this optimization function guides MCMC sampling to cluster in high-confidence regions, while also incorporating the support functions of each node in DS evidence fusion. Conflict coefficient Confidence distribution correction is performed to make the final pollution source confidence area show a stable convergence trend, significantly reducing the interference of outliers on the shape and path prediction of the distribution map.

[0031] Step S44: Introduce evidence theory (DS theory) to fuse multi-sensor data and reduce the impact of single-node data anomalies on the source tracing results; Step S45: The output results include the confidence ellipse region for locating the pollution source, the topological sequence of the most likely confluence paths, and the timeline prediction of the pollution event development.

[0032] Example 5 Based on Example 4, the visualization interface and early warning mechanism in step S5 are specifically implemented as follows: Step S51: Develop a WebGL-based 3D pipeline visualization engine to support dynamic rendering of pollution propagation paths and historical event backtracking. Step S52: The early warning level is divided into a four-level response mechanism, which corresponds to different confidence thresholds and response time requirements; Step S53: The disposal suggestion generation module integrates the expert system rule base, including pipeline blocking priority assessment, diversion scheme optimization algorithm and emergency resource scheduling model; Furthermore, in this embodiment, the diversion scheme optimization algorithm in the disposal suggestion generation module is constructed based on a multi-objective constrained programming model, assuming the pollutant concentration is... The capacity of the pipe section is The state of the control valve is a binary variable. The objective function to be minimized is the weighted sum of total pollutant exposure and emergency response costs, i.e., minimizing... ; in These are the weighting factors for response costs and environmental risks, respectively. The function is solved under constraints, including the processing capacity limitations of each node and the flow conservation condition. The final output is the optimal diversion path set that satisfies the controllability of the pipeline network, thereby improving the spatial coordination efficiency of emergency response and suppressing the spread of pollutants to highly sensitive areas.

[0033] Step S54: Establish an interpretability report generation system for the source tracing results, and automatically label the database entries and model feature contributions of key decision-making criteria; Furthermore, in this embodiment of the interpretability report generation system, the contribution of model features is evaluated using a quantitative method based on SHAP values, and a feature input vector is provided. The model outputs the prediction result as follows: Then each feature The contribution value is defined by the SHAP function as: ; Where S is a subset of features and n is the total number of features, this mechanism enables the quantitative interpretation of key pollution features for source tracing decisions, and generates an evidence chain for tracing paths by combining database entry indexes, thereby improving the transparency and verifiability of early warning results.

[0034] Step S55: Standardize the data interface with the municipal GIS platform to support cross-verification of traceability results with pipeline maintenance records and sewage discharge permit information.

[0035] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0036] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for tracing the source of pollution in stormwater pipe networks by combining three-dimensional fluorescent fingerprinting and graph neural networks, characterized in that, Includes the following steps: Step S1: By deploying fluorescence spectral sensor arrays at key nodes of the rainwater pipe network, fluorescence spectral data of water bodies in the pipe network are collected in real time and the data is synchronously transmitted to the cloud processing platform. Step S2: Construct and incrementally update the dynamic fluorescence database to store and index pollutant characteristic fingerprint spectra. The incremental update is based on the distribution offset between real-time data and historical data and is triggered when a new pollution characteristic is detected. Step S3: Use a machine learning model to extract features and recognize patterns from the fluorescence spectral data acquired in real time in step S1. Step S4: Through the synergistic effect of dynamic fluorescence database and machine learning model, combined with pipeline topology constraints, calculate the probability distribution of pollution sources in real time, optimize the source tracing results based on Bayesian inference algorithm, and output the confidence assessment of pollution source type, location and mixed connection path. Step S5: Feed back the source tracing results of step S4 to the regulatory terminal through the visualization interface, trigger the pipeline pollution early warning mechanism, and generate a pollution incident handling suggestion plan simultaneously.

2. The method for tracing the source of pollution in rainwater pipe networks combining three-dimensional fluorescent fingerprinting and graph neural networks according to claim 1, characterized in that, In step S2, a dynamic fluorescence database for non-steady-state pollution events in stormwater pipe networks is constructed. This database is used to characterize the time-varying evolution of pollutant spectral characteristics under rainfall-induced hydraulic condition abrupt changes, including but not limited to rainfall-induced hydraulic condition abrupt changes and source tracing under normal sunny conditions. By judging the degree of distribution deviation between real-time collected data and historical stable operating condition data, an incremental update process is triggered only when the pollution characteristics are determined to exceed the range of the existing stable model. Specifically, this includes: Step S21: Perform preprocessing operations on the newly added fluorescence spectral data for event detection to separate the background changes and pollution fingerprint background characteristics caused by the inflow and infiltration of external water, thereby providing a stable reference for subsequent pollution event identification; Step S22: Dynamically classify the pollution source feature fingerprint spectrum based on the unsupervised clustering analysis algorithm, and perform cluster center expansion operation when the following conditions are met simultaneously within the same pollution event cycle: The fingerprint appears repeatedly within multiple non-continuous time windows; Its distance from existing cluster centers in the feature space remains consistently higher than a set threshold; Its occurrence is accompanied by abnormal changes in the hydraulic state of the pipeline network; Step S23: Combining the pipeline network topology and water flow dynamics model, weights are assigned to the spatial correlation of pollution events, and a database index with spatiotemporal attributes is established.

3. The method for tracing the source of pollution in rainwater pipe networks combining three-dimensional fluorescent fingerprinting and graph neural networks according to claim 2, characterized in that, In step S21, the data cleaning stage of the preprocessing operation adopts a dual denoising strategy combining wavelet transform and Savitzky-Golay filtering, which targets the energy threshold function for high-frequency noise and the adaptive window smoothing algorithm for baseline drift. The wavelet transform is performed using the Daubechies-4 wavelet basis for multi-scale decomposition. Let the original fluorescence spectral signal be... The high-frequency noise threshold uses a threshold function based on energy distribution: ； in These are wavelet high-frequency coefficients. The coefficient is the empirical adjustment factor, and N is the number of coefficients. In baseline drift filtering, the Savitzky-Golay filter window width is adaptively adjusted according to signal stationarity to ensure curve shape preservation and smoothness; the pollution source feature anomaly detection mechanism combines local density ratio to calculate the isolation degree of the pollution source fingerprint vector, defining the frequency anomaly scoring function as follows: ； in Let i be the frequency of occurrence of the i-th type of contaminated fingerprint within the current time window. For its local density estimation in the feature space, when If the threshold is exceeded, it is automatically marked as a high-risk pollution source and manual review is triggered; the graph embedding process is based on the node vectors trained by the GNN model. By using cosine similarity for fast path matching and dynamically updating path retrieval weights, the accuracy of identifying contaminated paths can be improved. In step S22, the dynamic classification of the pollution source feature fingerprint spectrum adopts the K-means clustering algorithm, and a frequency anomaly detection mechanism is introduced in the process: when a certain type of pollution source feature appears abnormally frequently under low rainfall background conditions, and its frequency growth trend is inconsistent with the changes in the hydraulic parameters of the pipeline network, the system determines that the feature does not belong to the hydraulically driven background disturbance, and marks it as a high-risk new pollution source to trigger the manual review process. After the review is passed, the feature is included in the standard database. In step S23, a graph neural network model based on the pipeline network topology is used to map the pollution propagation process into a path state space constrained by the direction of water flow and the connectivity of nodes. This is used to exclude propagation paths that are physically inaccessible under the current hydraulic conditions when multiple candidate source tracing paths exist. Step S2 also includes establishing a database version control mechanism, retaining historical version data after each update, and automatically triggering a rollback command to the previous stable version and recalculating when the confidence level of the tracing result is lower than a preset threshold.

4. The method for tracing the source of pollution in rainwater pipe networks combining three-dimensional fluorescent fingerprinting and graph neural networks according to claim 1, characterized in that, In step S3, the machine learning model adopts a hybrid architecture of convolutional neural network and long short-term memory network. The CNN module is activated only when a spectral morphological change segment is detected to capture the local fingerprint structure at the moment of pollution event triggering. The LSTM module is used to model the evolution trajectory of pollution fingerprint with hydraulic conditions within the time window before and after the event to distinguish between transient noise disturbances and real pollution propagation behavior. The feature extraction process of the machine learning model includes the following steps: Step S31: Standardize the fluorescence spectral data, use Z-score normalization to eliminate dimensional differences between sensors, and remove data redundancy through principal component analysis to map the high-dimensional spectral signal to a compact feature space. Step S32: Construct a feature extraction layer containing a multi-branch parallel convolution kernel group, including three sizes: 1×3, 1×5, and 1×7, to simultaneously capture the fine morphological features of local peaks and global trend features in spectral data. Step S33: Introduce an attention mechanism into the LSTM network layer to strengthen the response weights at key time points before and after the pollution event, so as to avoid the dilution effect of historical data under long-term stable flow conditions on the characteristics of sudden pollution events. Step S34: Using a transfer learning strategy, the model parameters pre-trained on pipeline network data in other regions are used as initial values ​​and fine-tuned using local measured data to achieve rapid model adaptation. Step S35: Set up the online incremental learning module for the model and set the traceability deviation monitoring threshold. When the deviation between the traceability result and the manual verification exceeds the threshold, the incremental training process will be started automatically and the model parameters will be updated.

5. The method for tracing the source of pollution in rainwater pipe networks combining three-dimensional fluorescent fingerprinting and graph neural networks according to claim 4, characterized in that, In step S33, the attention mechanism employs a time-weighted function based on the dynamics of contamination evolution to assign differentiated attention levels to the hidden states output at each time step of the LSTM network, defining the attention weight vector as follows: ； in Let represent the hidden state at time step t. ψ is the context vector, and ψ is the scoring function constructed based on bidirectional GRU to characterize the correlation between temporal variation features and pollution fingerprint response. This mechanism is used to enhance the model's response sensitivity to pollution outbreak nodes. The online learning module is triggered by an error-driven incremental update strategy. Let the current source tracing prediction result be... The manual verification label is y, and its loss function is... Exceeding the dynamic threshold The fine-tuning process is initiated at that time, in which The standard deviation of recent sample loss is γ, which is an adjustment factor. This mechanism ensures that the model maintains continuous adaptability and source tracing accuracy under pollution type drift and spectral changes.

6. The method for tracing the source of pollution in rainwater pipe networks combining three-dimensional fluorescent fingerprinting and graph neural networks according to claim 1, characterized in that, The method for calculating the pollution source probability distribution map in step S4 further includes: Step S41: Simulate the transmission and diffusion process of pollutants based on the pipeline hydraulic model, and correct the propagation velocity parameters by combining real-time flow data; Step S42: Generate a candidate set of pollution sources using the Monte Carlo method, and optimize the convergence efficiency of the candidate solutions through Markov chain Monte Carlo sampling; Step S43: Construct a multi-objective optimization function to establish consistency constraints among candidate pollution source locations, occurrence times, and spectral fingerprint matching results, so as to avoid bias in the source tracing results caused by minimizing a single error; Step S44: Introduce evidence theory to fuse multi-sensor data and reduce the impact of single-node data anomalies on the source tracing results; Step S45: The output results include the confidence ellipse region for locating the pollution source, the topological sequence of the most likely confluence paths, and the timeline prediction of the pollution event development.

7. The method for tracing the source of pollution in rainwater pipe networks combining three-dimensional fluorescent fingerprinting and graph neural networks according to claim 6, characterized in that, In step S43, the multi-objective optimization function constructs a joint loss function by fusing spatial error, temporal error, and spectral feature error. ； in Indicates the predicted location of the pollution source The Euclidean distance error between the actual position P and the actual position P. Indicates the predicted time of the event. The absolute difference from the actual time T, This represents the cosine similarity error between the predicted spectral feature vector and the best-matching fingerprint vector in the database. These are the error term weight coefficients; this optimization function guides MCMC sampling to cluster in high-confidence regions, while also incorporating the support functions of each node in DS evidence fusion. Conflict coefficient Confidence distribution correction is performed to make the final pollution source confidence area show a stable convergence trend, significantly reducing the interference of outliers on the shape and path prediction of the distribution map.

8. The method for tracing the source of pollution in rainwater pipe networks combining three-dimensional fluorescent fingerprinting and graph neural networks according to claim 1, characterized in that, The visualization interface and early warning mechanism in step S5 include the following steps: Step S51: Develop a 3D pipeline visualization engine based on the WebGL standard to support dynamic rendering of pollution propagation paths and interactive retrospective analysis of the entire process of historical events. Step S52: Establish a multi-dimensional hierarchical early warning and response mechanism. Based on the confidence interval of the source tracing results, the early warning level is divided into four levels, and a differentiated emergency response time limit and standardized handling plan are set for each level. Step S53: The disposal suggestion generation module integrates the expert system rule base, including pipeline blocking priority assessment, diversion scheme optimization algorithm and emergency resource scheduling model; Step S54: Establish an interpretability report generation system for the source tracing results, and automatically label the database entries and model feature contributions of key decision-making criteria; Step S55: Standardize the data interface with the municipal GIS platform to support cross-verification of traceability results with pipeline maintenance records and sewage discharge permit information.

9. The method for tracing the source of pollution in rainwater pipe networks combining three-dimensional fluorescent fingerprinting and graph neural networks according to claim 8, characterized in that, In step S53, the diversion scheme optimization algorithm in the treatment suggestion generation module is based on a multi-objective constrained programming model, assuming the pollutant concentration is... The capacity of the pipe section is The state of the control valve is a binary variable. The objective function to be minimized is the weighted sum of total pollutant exposure and emergency response costs. Minimize ; in These are the weighting factors for response costs and environmental risks, respectively. The function is solved under constraints, including the processing capacity limitations of each node and the flow conservation condition. The final output is the optimal diversion path set that satisfies the controllability of the pipeline network, thereby improving the spatial coordination efficiency of emergency response and suppressing the spread of pollutants to highly sensitive areas.

10. The method for tracing the source of pollution in rainwater pipe networks combining three-dimensional fluorescent fingerprinting and graph neural networks according to claim 8, characterized in that, In the interpretability report generation system of step S54, the model feature contribution is evaluated using a quantitative method based on SHAP values, and a feature input vector is provided. The model outputs the prediction result as follows: Then each feature The contribution value is defined by the SHAP function as: ； Where S is a subset of features and n is the total number of features, this mechanism enables the quantitative interpretation of key pollution features for source tracing decisions, and generates an evidence chain for tracing paths by combining database entry indexes, thereby improving the transparency and verifiability of early warning results.

Images