Sea-entering river pollution early warning and tracing system

Abstract

The invention discloses a sea-entering river pollution early warning and traceability system, and relates to the field of intelligent water environment supervision, and the system comprises an integrated monitoring network which is used for monitoring river monitoring stations, river-entering discharge port monitoring nodes and hydro-meteorological monitoring units which are arranged in a spatial distribution manner, collecting multi-parameter indexes of a water body, discharge characteristic data of a discharge port and environmental power data; the intelligent analysis and traceability engine is used for carrying out fusion processing on the monitoring data and executing dynamic early warning judgment and pollution source intelligent traceability analysis based on a fusion data set, and the traceability analysis is coupled with a mechanism driving analysis normal form and a data driving analysis normal form; and the early warning and decision support platform is used for issuing graded early warning information according to an output result of the engine, providing visual situation display based on an electronic map and automatically generating a structured traceability report. According to the scheme, the efficiency of monitoring the environment of the river entering the sea can be remarkably improved, and the system realizes early perception and accurate early warning of pollution events.

Description

Technical Field

[0001] This invention relates to the field of intelligent water environment monitoring, and in particular to a pollution early warning and source tracing system for rivers flowing into the sea. Background Technology

[0002] Rivers flowing into the sea serve as crucial channels for the transport of land-based pollutants, and their water quality has a decisive impact on the health and stability of nearshore marine ecosystems. Currently, the monitoring system for pollution in these rivers primarily relies on fixed automatic water quality monitoring stations deployed along the main river channels, along with regular manual inspections and sampling analyses. While this model has played a fundamental role in long-term environmental management, its ability to respond to pollution incidents in real-time and accurately trace their sources has become insufficient as pollution sources become increasingly complex and concealed. Under the existing technological framework, various monitoring data are typically scattered across different systems, creating "information silos" and lacking effective collaborative analysis mechanisms. Therefore, the environmental management field urgently needs a comprehensive technological solution capable of integrating multi-source heterogeneous data, enabling a shift from passive monitoring to proactive early warning, and intelligently identifying pollution sources.

[0003] Existing technologies and methods suffer from several inherent flaws that limit their regulatory effectiveness. First, they exhibit significant lag and passivity; the spatial coverage density of fixed monitoring stations is limited, and by the time downstream stations detect pollution incidents, the pollution plume has often migrated far, missing the optimal response time. Second, their source tracing capabilities are extremely weak; relying solely on downstream station data on emissions exceeding standards is insufficient to quickly and accurately identify the responsible party from numerous potential upstream emission sources, still requiring extensive manual investigation, which is inefficient and results in incomplete evidence chains. Third, existing systems lack continuous and effective monitoring of dispersed river discharge outlets, creating regulatory blind spots. Finally, there is a lack of deep integration and intelligent correlation analysis between various data types; early warning mechanisms largely rely on simple threshold exceedance judgments, leading to high false alarm and false negative rates, failing to support a modern environmental regulatory closed loop integrating real-time perception, intelligent diagnosis, and precise decision-making. Therefore, based on the above challenges, this invention proposes a pollution early warning and source tracing system for rivers flowing into the sea. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide a pollution early warning and source tracing system for rivers flowing into the sea. By constructing an integrated monitoring network combining "points, lines, and surfaces" and deeply integrating multi-source heterogeneous data, and utilizing big data analysis and artificial intelligence algorithms, the system achieves early detection, rapid early warning, and precise source tracing of pollution events, thereby comprehensively improving the initiative, accuracy, and intelligence of environmental supervision.

[0005] To achieve the above objectives, this invention provides a pollution early warning and source tracing system for rivers flowing into the sea. This system first establishes a spatiotemporally coordinated integrated monitoring network by deploying monitoring equipment along the main river channel, key discharge outlets, and surrounding environments, enabling comprehensive data collection. Then, using IoT technology, multi-source heterogeneous data is transmitted to a central processing platform for cleaning, fusion, and spatiotemporal alignment, forming a unified analysis dataset. Subsequently, an intelligent analysis engine, based on this dataset, performs dynamic early warning determination through parallel fusion threshold, trend, and correlation analysis. Upon triggering an early warning, it collaboratively initiates pollution diffusion inversion based on hydrodynamic models, discharge outlet contribution analysis based on spatiotemporal correlation, and source identification probability inference based on deep learning models, thereby accurately identifying suspected pollution sources. Finally, a decision support platform enables automatic dissemination of multi-level early warning information, visualization of the full-element situation, and one-click generation of structured source tracing reports, forming a closed-loop management system from perception and analysis to decision-making.

[0006] In a first aspect, the present invention provides a pollution early warning and source tracing system for rivers flowing into the sea, comprising: An integrated monitoring network is used to collect multiple parameters of water bodies, discharge characteristics data and environmental dynamic data through spatially distributed river monitoring stations, river discharge outlet monitoring nodes and hydro-meteorological monitoring units. The data acquisition and communication module is used to transmit monitoring data to the central processing platform in real time via a wireless communication network. The intelligent analysis and tracing engine is used to fuse the monitoring data and perform dynamic early warning judgment and intelligent pollution source tracing analysis based on the fused dataset, wherein the tracing analysis couples the mechanism-driven and data-driven analysis paradigms. The early warning and decision support platform is used to issue tiered early warning information based on the output of the engine, provide a visual situation display based on electronic maps, and automatically generate structured traceability reports.

[0007] Furthermore, in the integrated monitoring network, the river discharge outlet monitoring nodes are miniaturized, low-power sensor devices, which are deployed at industrial, municipal, and stormwater discharge outlets to perform high-frequency monitoring of discharge flow and key pollutants.

[0008] Furthermore, the data acquisition and communication module supports multiple low-power wide-area network communication protocols and has data encryption and transmission interruption recovery mechanisms to adapt to reliable data backhaul in complex field environments.

[0009] Furthermore, the intelligent analysis and tracing engine includes: The data fusion unit is used to standardize, remove outliers, and perform spatiotemporal alignment on multi-source heterogeneous data from different monitoring units to build a unified spatiotemporal dataset. The dynamic early warning unit is used to perform in parallel the following tasks based on the unified spatiotemporal dataset: threshold comparison based on water quality standards, trend deviation detection based on historical behavior patterns, and abnormal propagation analysis based on the correlation between upstream and downstream monitoring points, so as to comprehensively trigger early warning signals. The intelligent source tracing unit is used to initiate a collaborative source tracing process in response to the warning signal. The process integrates pollution diffusion inversion based on hydrodynamics, discharge outlet contribution analysis based on spatiotemporal correlation, and pollution source probability inference based on deep learning models.

[0010] Furthermore, the trend deviation detection in the dynamic early warning unit identifies statistical anomalies relative to historical normal patterns by analyzing the dynamic characteristics of the time series of monitoring indicators; the correlation anomaly propagation analysis constructs a data correlation model between upstream and downstream monitoring points, and automatically triggers a targeted scan of potential emission sources in the intermediate area when an anomaly occurs downstream but no chain anomaly occurs upstream. This effectively narrows the scope of source tracing and investigation and improves the targeting of regulatory actions.

[0011] Furthermore, the intelligent source tracing unit utilizes a simplified hydrodynamic model to simulate the migration and diffusion process of pollutants in the upstream basin of the warning point, in order to determine the possible release time window and spatial range of pollutants. Within the time window and spatial range, the spatiotemporal matching degree of the emission data of each discharge monitoring node and the downstream water quality anomaly data is calculated to identify a set of discharge outlets with high correlation. The feature data, hydrological conditions, and historical emission patterns of the discharge outlet set are then input into a pre-trained deep learning network to output a probability estimate that each suspected discharge outlet is a real pollution source. The deep learning network introduces an attention mechanism to dynamically weight the importance of different features in the source tracing decision-making, thereby enhancing the model's generalization and reasoning capabilities in complex scenarios.

[0012] Furthermore, in the intelligent source tracing unit, the part that performs the spatiotemporal matching degree calculation is configured to run a collaborative source tracing algorithm based on spatiotemporal contribution entropy. This algorithm calculates the temporal consistency between the discharge behavior of each suspected discharge outlet within the pollution event time window and the downstream water quality anomaly, and combines the spatial distribution anomaly of its discharge intensity relative to its own historical level to construct a unified spatiotemporal contribution evaluation index. Then, based on this index, the comprehensive suspicion degree of each discharge outlet is calculated, and a source tracing list sorted by suspicion probability is generated accordingly.

[0013] Furthermore, the deep learning network is trained using historical pollution event samples to learn the propagation characteristics and source identification patterns of pollution events under complex hydrological and emission conditions; the probability estimation integrates factors such as emission time consistency, concentration change correlation, and historical behavior deviation.

[0014] Furthermore, the deep learning network is specifically an adaptive graph attention neural network; this network abstracts monitoring stations and discharge outlets as graph nodes, and abstracts water flow direction and hydraulic connections as directed edges. Through an attention mechanism that embeds real-time hydrological influencing factors, it dynamically calculates the influence weights between nodes; this mechanism enables the network to adapt to dynamic hydrological conditions such as changes in flow velocity and flow rate, thereby achieving the simulation of pollution propagation paths and robust inference of pollution source probabilities.

[0015] Furthermore, the system is equipped with a feedback learning mechanism, which can incrementally update and optimize the deep learning network using subsequently confirmed data on the truth of pollution events, thereby enabling the system's source tracing capabilities to have the characteristics of continuous self-evolution and improvement.

[0016] Furthermore, the early warning and decision support platform issues different levels of early warning information based on the severity and scope of the pollution incident, and pushes it to relevant management personnel through multiple human-computer interaction channels. On the electronic map, the status of monitoring points, dynamic changes in water quality parameters, simulated diffusion range of pollution plumes, and spatial distribution and probability ranking of suspected discharge outlets locked by the system are presented in real time. Based on the analysis results of the engine, a source tracing decision report is automatically generated, which includes event chain reconstruction, supporting evidence data, suspected source ranking, and disposal suggestions. The platform provides an interactive source tracing path verification tool, allowing management personnel to manually adjust parameters and observe the impact on the source tracing results in real time, thereby enhancing the transparency and credibility of the decision-making process.

[0017] Secondly, the present invention also provides a method for early warning and source tracing of pollution in rivers flowing into the sea, the method being based on the system described in the first aspect above, comprising: Continuously collect multi-source monitoring data on river water bodies, river outlets, and hydrological and meteorological data, and standardize and align the multi-source monitoring data in time and space to form a unified time and space dataset. Based on the unified spatiotemporal dataset, threshold comparison, trend deviation detection, and upstream and downstream correlation analysis are performed in parallel to comprehensively trigger dynamic early warnings. Based on hydrodynamic principles, the migration and diffusion process of pollutants is simulated to infer the possible release time and spatial range of pollutants. Within the specified time and space range, calculate the spatiotemporal correlation matching degree between the discharge data of each outlet and the downstream water quality anomaly data; The emission characteristics, hydrological conditions, and historical behavior patterns are input into a pre-trained deep learning model to obtain the source tracing probability of each suspected discharge outlet. Based on the results of dynamic early warning and intelligent source tracing, hierarchical early warning information, visual situation maps, and structured source tracing reports are generated and output.

[0018] This invention provides a pollution early warning and source tracing system for rivers flowing into the sea. Its core lies in constructing an integrated monitoring network that incorporates multi-dimensional sensing of "points, lines, and surfaces," achieving real-time transmission and fusion of multi-source heterogeneous data through Internet of Things (IoT) technology. The system employs an intelligent analysis engine that couples mechanistic and data-driven models. The dynamic early warning module achieves early pollution detection through threshold, trend, and correlation analysis, while the intelligent source tracing module collaboratively utilizes hydrodynamic inversion, spatiotemporal correlation calculation, and deep learning probabilistic inference to cross-verify and accurately pinpoint suspected pollution sources from multiple perspectives. Finally, a comprehensive decision-making platform enables tiered early warning, situation visualization, and automated generation of source tracing reports, forming a complete technological closed loop from perception and diagnosis to decision-making.

[0019] This solution enables early detection and accurate warning of pollution incidents, effectively overcoming regulatory lag and reducing the risks of false alarms and underreporting. Its integrated source tracing mechanism transforms traditional open-ended investigations relying on manual experience into intelligent diagnosis based on multi-source evidence chains, significantly improving the efficiency and accuracy of source tracing. Ultimately, this technological system constitutes an automated and intelligent regulatory closed loop, greatly enhancing the initiative and scientific rigor of environmental management and providing strong decision support for rapid response and precise enforcement.

[0020] Beneficial effects By implementing the pollution early warning and source tracing system for rivers flowing into the sea provided by the present invention, the following technical effects are achieved: (1) The dynamic early warning model based on multi-level correlation analysis achieves multi-level triggering and comprehensive judgment of early warning signals by performing threshold comparison, trend deviation detection and upstream and downstream correlation analysis in parallel. This enables the system to capture potential abnormal signs earlier and effectively distinguish between local discharge events and watershed water quality fluctuations by analyzing the spatial propagation relationship of anomalies. This reduces false alarm and false alarm rates while providing a more accurate time starting point and spatial investigation scope for subsequent source tracing analysis.

[0021] (2) The hybrid intelligent source tracing architecture, which integrates mechanism and data-driven approaches, performs collaborative computation and cross-validation of results between a mechanism model based on hydrodynamic equations and an intelligent model trained on historical data. This forms a complementary judgment mechanism. The mechanism model ensures the physical interpretability of the source tracing process and its basic reliability under extreme conditions, while the data-driven model provides intelligent inference capabilities learned from complex historical patterns. Together, they ensure the consistency of source tracing conclusions in terms of logical rigor and intelligent accuracy.

[0022] (3) The collaborative source tracing algorithm based on spatiotemporal contribution entropy significantly improves the system's diagnostic capabilities in complex pollution discharge scenarios by constructing a unified evaluation index that integrates temporal consistency and spatial emission anomaly. It can effectively distinguish between different pollution patterns caused by sudden emission events and background steady-state emissions, thereby greatly enhancing the accurate identification of real pollution sources in the presence of multiple potential interference sources, and significantly reducing the risk of misjudgment caused by differences in historical emission baselines or accidental synchronous emissions.

[0023] (4) The adaptive graph attention neural network for source tracing tasks endows the model with the inherent ability to dynamically perceive changes in the watershed state by embedding real-time hydrological dynamic factors into the graph attention mechanism. This enables the system to exhibit excellent robustness and generalization ability when facing local failures of the monitoring network or extreme hydrological conditions that it has never experienced. Its source tracing reasoning process is closer to physical laws, effectively overcoming the common problem of the sharp performance degradation of existing data-driven models when the application scenario deviates from the distribution of the training set.

[0024] (5) By introducing an online self-optimization mechanism based on dynamic feedback and incremental learning, the system is endowed with the ability to continuously learn from historical source tracing decision results, thereby effectively overcoming the performance degradation problem of existing artificial intelligence models caused by environmental changes and the evolution of pollution source characteristics. This mechanism uses the results of manual verification as high-quality labeled samples to drive the source tracing model to continuously and steadily optimize parameters, enabling it to adapt to the dynamic changes in the watershed pollution pattern. This significantly improves the source tracing accuracy, environmental adaptability, and robustness of the system in long-term operation scenarios, realizing a leap from a static analysis tool to an intelligent diagnostic system with autonomous evolution capabilities. Attached Figure Description

[0025] To make the above-described pollution early warning and source tracing system for rivers flowing into the sea more clear and understandable, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating the method described in this application. Detailed Implementation

[0027] Example 1: A system and method for monitoring, early warning, and tracing pollution sources in rivers flowing into the sea based on multi-source data fusion are provided. The method flow is as follows: Figure 1 As shown, the details are as follows.

[0028] The system first constructs an integrated monitoring network encompassing "points, lines, and surfaces." Automatic river monitoring stations are deployed along the main river channels and important tributaries, forming a "linear" monitoring baseline. These stations continuously monitor multiple water quality indicators, including the five conventional parameters, permanganate index, ammonia nitrogen, total phosphorus, total nitrogen, heavy metals, and toxic organic matter, acting as the first "sentinels" to detect pollution events. At key river outfalls, including industrial, municipal, and stormwater outfalls, low-cost, miniaturized, and easy-to-maintain water quality sensor nodes are deployed, forming a "point" monitoring network. These nodes perform high-frequency, real-time monitoring of outfall flow and key water quality indicators, providing crucial "clues" for tracing pollution sources. Furthermore, the system integrates existing hydrological and meteorological stations to form auxiliary monitoring units, continuously collecting environmental dynamic data such as flow velocity, flow rate, water level, rainfall, wind speed, and wind direction, providing an accurate background field for subsequent pollution diffusion simulation and source tracing analysis.

[0029] Monitoring data is transmitted in real-time and remotely to the central data processing platform via the data acquisition and transmission module, utilizing IoT technology and various wireless communication methods. This module features data encryption and breakpoint resume capabilities, ensuring the integrity, security, and continuity of data during transmission. All data is ultimately integrated into the data center and intelligent analysis module. Here, the data cleaning and fusion unit first standardizes the multi-source heterogeneous data from different sources, removing obvious outliers, and then precisely aligns and deeply fuses water quality, hydrological, and meteorological data across temporal and spatial scales, forming a unified, clean dataset suitable for advanced analysis.

[0030] Based on this fused dataset, the dynamic early warning model unit begins to function. This unit employs a multi-strategy parallel early warning mechanism: threshold early warning sets static thresholds based on water quality standards to quickly respond to instantaneous exceedances; trend early warning uses time series analysis or machine learning algorithms to learn the normal change patterns of various indicators based on historical data and provides forward-looking warnings for abnormal fluctuations that deviate from these normal trends; in addition, correlation early warning analyzes the data correlation between upstream and downstream monitoring stations, as well as between discharge outlets and river stations. When an anomaly occurs at a downstream station but no chain reaction is observed at the directly upstream station, the system can automatically trigger a focused scan and data analysis of all discharge outlets in the intermediate river section, thereby accurately locating the possible area of ​​abnormal input.

[0031] Once the warning is triggered, the intelligent source tracing analysis unit is immediately activated. This unit integrates mechanistic models and data-driven models to form a combined analytical force. The pollution diffusion simulation subunit, based on mature hydrodynamic theory and combined with real-time acquired hydrological data, simulates the convection, diffusion, and transformation processes of pollutants in water bodies, thereby deducing the approximate time when the pollution plume reaches the downstream warning point and the possible release location range upstream, defining an initial spatiotemporal search window for source tracing. Based on this, the multi-source data comparison source tracing subunit begins operation, establishing a spatiotemporal correlation model between abnormal discharge from outlets and abnormal water quality downstream. The system automatically backtracks monitoring data from all upstream outlets within a specific time window calculated based on flow velocity before the warning time. By running a series of intelligent algorithms, it quickly filters out one or more suspected outlets with the highest temporal and spatial matching degree between discharge patterns and downstream pollution events. To further improve the accuracy and intelligence of source tracing, the intelligent source tracing model subunit utilizes the system's long-accumulated historical monitoring data and a confirmed pollution event case library to train a deep learning neural network. This model can learn and memorize complex pollution propagation dynamics and the unique emission characteristics of each outlet, thus enabling faster and smarter pollution source identification and probability prediction when faced with new pollution events.

[0032] Ultimately, all analysis results, early warning information, and source tracing conclusions are transmitted to the early warning, source tracing, and visualization platform module. Based on the severity and potential impact of the pollution incident, the platform activates a multi-level early warning mechanism and automatically notifies relevant environmental management personnel in a timely manner through various means, including the platform interface, SMS, email, and mobile application push notifications. The platform also provides a map-based visualization interface that displays the real-time geographical location and current status of all monitoring points, dynamic trends in water quality parameters, simulated diffusion range of the pollution plume, the current early warning area, and a list of suspected discharge outlets intelligently locked by the system. Furthermore, the platform features a one-click source tracing report generation function, automatically integrating event descriptions, early warning times, relevant monitoring data change curves, a list of suspected discharge outlets, detailed evidence analysis, and targeted disposal recommendations to form a well-structured and complete professional report. This provides clear direction and solid scientific basis for environmental enforcement personnel's on-site verification and disposal, thus forming a complete technological closed loop from perception, analysis, decision-making to action support.

[0033] Example 2: Building upon the aforementioned embodiments, this paper further elaborates and highlights how the intelligent analysis and source tracing engine enables more accurate early warning and more reliable source tracing of pollution events. Having completed the basic deployment of an integrated monitoring network and the fusion of multi-source data, the system's intelligent analysis module employs a more advanced collaborative computing strategy.

[0034] The dynamic early warning model unit is designed as a parallel processing architecture, no longer relying on a single early warning rule. The threshold early warning module, in addition to static standards, introduces a dynamic threshold adjustment mechanism based on historical statistical data, enabling the early warning benchmark to adapt to different hydrological seasons and water environment backgrounds. The trend early warning module integrates a more complex time-series prediction model, which not only identifies simple trend deviations but also captures potential risks that accumulate slowly or erupt suddenly by analyzing deeper characteristics such as fluctuation patterns and rates of change in the sequence. Crucially, the correlation early warning module constructs a virtual river network correlation diagram to quantify the influence weights between upstream and downstream stations. Once an anomaly is detected at a downstream station, the system immediately checks whether the anomaly has formed the expected propagation chain along the water flow direction. If the propagation chain is interrupted—that is, the downstream is abnormal while the upstream station is normal—the system intelligently infers that the pollution originates from a discharge outlet along the way and immediately initiates a synchronous in-depth analysis of all discharge outlets within the correlation area, greatly narrowing down the scope of suspicion and improving the targeting of the early warning.

[0035] The system's intelligent source tracing analysis unit deeply integrates and collaboratively performs calculations with physical mechanism models and data-driven artificial intelligence models. When an early warning is triggered, the pollution diffusion simulation subunit first uses simplified hydrodynamic equations, combined with real-time flow velocity and flow rate data, to quickly simulate the pollutant's trajectory, providing a preliminary spatiotemporal framework for the pollution source release. The results of this physical model provide a constraint consistent with physical laws for subsequent data analysis. Subsequently, the multi-source data comparison and source tracing subunit operates under this constraint. It does not simply perform data matching but executes a multi-indicator spatiotemporal correlation calculation. This calculation comprehensively considers the degree of agreement between the discharge outlet's peak emission and the model-inferred release time, the potential relationship between emission intensity and downstream pollution load, and a comparative analysis of the outlet's historical emission behavior. Through a comprehensive evaluation algorithm, the suspicion index of each discharge outlet in relation to this event is calculated.

[0036] Simultaneously, the intelligent source tracing model subunit is activated. This network can discern complex nonlinear relationships hidden within the data, such as the emission pattern of a particular discharge outlet under specific rainfall conditions, or the combined impact of minor emissions from multiple outlets on downstream areas. The model receives all relevant information from the data fusion unit, including real-time monitoring data, hydrological and meteorological conditions, and historical behavior records of each discharge outlet, independently outputting a set of probabilistic predictions for pollution sources. Finally, the system establishes an analysis center to cross-validate and weightedly fuse the data comparison results guided by the mechanism model with the model's probabilistic predictions. If both point to the same discharge outlet, the confidence level of the source tracing conclusion is greatly improved; if discrepancies exist, the system will comprehensively present evidence from all parties and provide an uncertainty assessment, guiding management personnel to focus their attention.

[0037] This dual-core traceability architecture enables the system to not only possess interpretability based on physical laws but also expert-level intuitive judgment capabilities learned from massive historical data, significantly improving the success rate and reliability of traceability in complex scenarios. The results of all this complex analysis are ultimately presented to users in a clear and intuitive way through a highly visualized decision support platform, generating detailed traceability analysis reports. This completes the full-process automation and intelligence from data perception to intelligent diagnosis to decision support, forming a self-verifying and continuously optimized technological closed loop.

[0038] Example 3: Building upon the two aforementioned embodiments, a comprehensive evaluation index, spatiotemporal contribution entropy, is constructed to simultaneously measure the temporal consistency and spatial anomaly of pollution discharge behavior. This addresses the insufficient discrimination of existing source tracing methods in complex emission scenarios. The model posits that a genuine pollution source's emission behavior should not only be highly synchronized with downstream pollution events in time, but its emission intensity should also exhibit significant anomalies relative to its historical levels. By incorporating the concept of information entropy, these two dimensions of information are merged into a single scalar. The lower the entropy value, the higher the certainty that the discharge outlet is a pollution source, thereby achieving more accurate and interference-resistant ranking of suspected sources.

[0039] When a pollution warning is triggered at a downstream monitoring point, the intelligent source tracing unit is activated. The system uses a hydrodynamic model to deduce the possible release time window of pollutants and determine the potential spatial range of impact.

[0040] For each suspected discharge outlet, its emission concentration time series is collected within a time window. The Pearson correlation coefficient is calculated between this series and the concentration changes before and after a pollution event at downstream monitoring points. The time fit factor is defined as the normalized correlation coefficient.

[0041] In the formula, For discharge outlet Time consistency factor; For discharge outlet The Pearson correlation coefficient between the emission concentration sequence and the concentration change sequence at downstream monitoring points.

[0042] This step normalizes the correlation coefficient to... The larger the value in the range, the better the time synchronization.

[0043] For each suspected emission outlet, extract its emission concentration data for the same period in a long historical period before the warning event occurs to form a historical emission set.

[0044] Calculate the average emission concentration of the outlet within the window during this event, along with the mean and standard deviation of its historical emission sets. Define the current emission intensity anomaly index of the outlet as:

[0045] In the formula, For discharge outlet The current emission intensity anomaly index; For this incident, the discharge outlet Average emission concentration within the time window; For discharge outlet The average of historical emission concentration datasets; For discharge outlet The standard deviation of the historical emission concentration dataset.

[0046] This index reflects the degree to which current emissions deviate from their historical normal levels. Subsequently, its spatial anomaly entropy is calculated:

[0047] In the formula, For discharge outlet Spatial anomaly entropy; The attenuation coefficient is an empirical parameter greater than 0, used to adjust the intensity of the influence of the anomaly index on the entropy value.

[0048] The larger the emission intensity anomaly index, the smaller its entropy value, indicating that the emission event is more abnormal and significant in spatial distribution.

[0049] Combining temporal fit and spatial anomaly entropy, a formula for spatiotemporal contribution entropy is constructed:

[0050] In the formula, For discharge outlet The spatiotemporal contribution entropy; It is a very small positive number used to ensure that the parameter of the logarithmic function is positive, thus preventing mathematical calculation errors.

[0051] Ultimately, the overall suspicion level is defined as:

[0052] In the formula, For discharge outlet It represents the final overall suspicion level of the actual pollution source; the higher the value, the higher the suspicion.

[0053] Calculate the overall suspicion value for all suspected outlets and sort them from highest to lowest according to this value. This will give you the accurate list of suspected sources output by the system.

[0054] Verification showed that in a river section with 50 potential discharge outlets, moderately complex hydrological conditions, and significant tidal disturbances and flow velocity variations, the algorithm improved the accuracy of identifying the true pollution source as the primary suspect from approximately 70% using traditional correlation analysis methods to 92%. Furthermore, by incorporating historical emission baselines for normalization, the system significantly enhanced its ability to distinguish between outlets with consistently high emission standards and those experiencing sudden emission events, reducing the false alarm rate by approximately 35% compared to source tracing methods based solely on instantaneous concentration peak time matching. This algorithm, through dual spatiotemporal constraints, greatly improves the anti-interference capability and accuracy of intelligent source tracing analysis. Results indicate that the collaborative source tracing algorithm based on spatiotemporal contribution entropy, by coupling temporal consistency and spatial emission anomaly, significantly enhances the system's diagnostic capabilities in complex pollution discharge scenarios. Compared to traditional methods that rely on single-dimensional correlation, this algorithm can effectively distinguish between different pollution patterns caused by sudden emission events and background steady-state emissions. In the presence of multiple potential sources of interference, it can significantly improve the accuracy of identifying real pollution sources and significantly reduce the risk of misjudgment caused by differences in historical emission baselines or accidental synchronous emissions, making the source tracing conclusions more confident and practical.

[0055] Example 4: Building upon the aforementioned embodiments, a watershed graph structure is constructed, with monitoring stations and discharge outlets as nodes and water flow direction and hydraulic connections as edges. An adaptive graph attention network is introduced to address the insufficient generalization ability of existing deep learning source tracing models when facing dynamic changes in the monitoring network and complex hydrological conditions. This network not only learns the emission and water quality characteristics of each node, but more importantly, it dynamically learns the mutual influence weights between different nodes under specific hydrological conditions. By using real-time hydrological factors as regulators for the attention mechanism, the model can simulate the propagation process of pollutants in the dynamically changing river network and adaptively adjust the source tracing inference path based on real-time hydraulic data, thereby achieving highly robust source tracing of pollution events under different operating conditions.

[0056] The entire monitoring area is abstracted as a directed graph. The node set contains all automatic water quality monitoring stations and river discharge outlet monitoring nodes. The edge set is determined by the water flow direction, pointing from upstream nodes to their downstream neighbors. Each edge is accompanied by a feature vector containing the real-time acquired information about the slave node. To the node The river section's flow velocity, flow rate, and distance.

[0057] For each node, initialize its feature vector. For discharge outlet nodes, the features include time-series data of discharge concentration and flow rate within the retrospective time window; for monitoring station nodes, the features are time-series data of water quality parameters.

[0058] For each directed edge in the graph compute nodes For nodes The attention coefficient, which changes dynamically with hydrological conditions, is calculated using the following formula:

[0059] In the formula; For nodes For nodes The adaptive attention weights represent the values ​​under specific hydrological conditions. right The degree of impact; It is a learnable attention vector used to calculate the correlation between nodes; It is a learnable shared weight matrix used to perform linear transformations on the features of all nodes; For nodes The initial feature vector; For nodes The initial feature vector; This is the vector concatenation operator; This is a scaling parameter used to adjust the magnitude of hydrological influence factors; For the edge The corresponding real-time flow velocity of the river section; For the edge Real-time flow rate of the corresponding river section; For the edge The corresponding river section distance; It is a smoothing factor, a small positive number used to prevent the denominator from being zero and to enhance numerical stability.

[0060] Each node updates its own state by aggregating information from all its incoming edges. The updated feature vector of a node is:

[0061] In the formula, For nodes The feature vector is updated after aggregating neighbor information through the graph attention layer; It is a non-linear activation function; For nodes The set of all upstream neighbor nodes.

[0062] After multi-layer graph attention propagation, each node obtains a feature representation containing global watershed information. Finally, the feature representations of all discharge outlet nodes are input into a fully connected layer and a Softmax function, outputting the probability that each discharge outlet is the source of this pollution event.

[0063] Verification showed that, compared to static graph neural network models or standard time-series models that do not consider hydrological dynamics, this model maintained a 95% Top-3 hit rate in source tracing scenarios with missing data, while the comparative model dropped to 78%. Furthermore, when dealing with extreme rainfall conditions not present in the training data, the source tracing accuracy of this model fluctuated within ±5%, significantly better than the comparative model's fluctuation exceeding ±15%, demonstrating its superior generalization ability and strong adaptability to complex hydrological conditions. The results indicate that the adaptive graph attention neural network for source tracing tasks, by embedding real-time hydrological dynamic factors into a graph attention mechanism, endows the model with the ability to dynamically perceive changes in the watershed state. This structure enables the model to exhibit excellent robustness and generalization ability when facing local failures in the monitoring network, dynamic additions or removals of nodes, or extreme hydrological conditions not previously experienced. Compared to static graph models or traditional time-series models, its source tracing inference process not only more closely resembles the physical migration and diffusion patterns of water pollutants, but also significantly enhances the stability of the output results, effectively overcoming the common problem of rapid performance degradation in existing data-driven models when the application scenario deviates from the training set distribution.

[0064] Example 5: Building upon the aforementioned embodiments, the subsequent manual verification results or authoritative rulings of each source tracing conclusion are treated as a highly confident "newly labeled sample." This sample, along with the real-time multi-source data that triggered the source tracing, forms a feedback loop, addressing the issue of model performance "aging" or "deterioration" caused by environmental changes and alterations in emission source characteristics after deployment. The system utilizes this continuously generated feedback data to perform online, incremental learning and optimization of the existing intelligent source tracing model, rather than periodically and in batches retraining. This mechanism allows the model to continuously accumulate and correct its knowledge while processing new cases, thereby dynamically adapting to changes in watershed pollution characteristics and achieving self-evolution and continuous improvement of source tracing capabilities.

[0065] Once the system completes intelligent source tracing of a pollution incident and outputs a list of suspected discharge outlets, environmental management personnel will conduct on-site verification or enforcement according to the system's guidance, and enter the final verification results into the system through the interactive interface of the early warning and decision support platform. The system automatically associates and stores the complete data package of the incident with the verification results. This data package includes: the unified spatiotemporal dataset fused when the early warning is triggered; all intermediate data and features relied upon by the intelligent source tracing unit when performing source tracing analysis; the source tracing probability of each suspected discharge outlet initially output by the system; and the finally confirmed pollution source identification.

[0066] The system maintains a fixed-capacity incremental learning sample queue. When new event feedback data packets are added, the queue follows a first-in, first-out (FIFO) principle to maintain the total number and timeliness of samples. This design ensures that the model can learn the latest contamination patterns, avoids the computational burden caused by the infinite expansion of the dataset, and maintains the stability of the model by retaining a certain number of historical samples, preventing the contradiction between recent individual events and traditional concepts.

[0067] Incremental learning can be triggered in two modes: periodic triggering: for example, the system automatically starts a model optimization process whenever 10 valid feedback samples are added to the sample queue; performance monitoring triggering: the system continuously monitors the model's performance on the feedback dataset, and immediately triggers optimization when it finds that the accuracy continues to drop beyond a preset threshold.

[0068] After the optimization process is initiated, the system extracts batch data from the sample queue and incrementally trains the existing intelligent traceability model. The training process employs the following key strategies: low learning rate and fine-tuning strategy: a very low learning rate is used to fine-tune the model parameters rather than reshape them, in order to avoid catastrophic forgetting; weighted loss function: in the loss function, newly added feedback samples are given higher weights, while regularization constraints on changes in old model parameters are added to the loss function to ensure that the model adapts to new knowledge without deviating excessively from its robust cognition built on massive amounts of historical data.

[0069] After each model update, the system generates a new model version and quickly evaluates it in an isolated test environment using historical validation sets and the latest feedback sample queue. Only when the performance evaluation results of the new model version are consistently better than or equal to the current online master model will the system automatically deploy it as the new master model. If the new model's performance fails to meet the standards, it is automatically rolled back to the previous stable version, and this optimization failure is recorded to provide a basis for subsequent algorithm adjustments, ensuring the stability of the online service.

[0070] Compared to traditional systems that use static models and undergo comprehensive retraining only once a year, our system maintained a consistently high quarterly source tracing accuracy of 92% to 95% with minimal fluctuations during the pilot period. In contrast, the accuracy of the comparison system gradually declined from an initial 90% to 78%, with particularly noticeable performance degradation after seasonal transitions and changes in rainfall patterns. When a new industrial enterprise with specific emission characteristics was added to the watershed, our system identified it as a high-suspect source in subsequent pollution events with similar characteristics after receiving two confirmations of the enterprise's illegal emissions. The comparison system, however, failed to effectively identify this new pollution source until the next annual model refactoring.

Claims

1. A pollution early warning and tracing system for a river flowing into the sea, characterized in that, The system comprises: an integrated monitoring network for collecting water multi-parameter indicators, discharge characteristics data of discharge outlets, and environmental dynamic data through spatially distributed river monitoring stations, river inlet monitoring nodes, and hydro-meteorological monitoring units; a data acquisition and communication module for transmitting monitoring data to a central processing platform in real time through a wireless communication network; an intelligent analysis and tracing engine for fusion processing of the monitoring data and performing dynamic early warning judgment and intelligent pollution source tracing analysis based on the fusion data set, wherein the tracing analysis is coupled with mechanism-driven and data-driven analysis paradigms; an early warning and decision support platform for publishing graded early warning information, providing visualized situation display based on an electronic map, and automatically generating a structured tracing report according to the output of the engine.

2. The system of claim 1, wherein: in the integrated monitoring network, the river inlet monitoring nodes are miniaturized, low-power sensor devices arranged at industrial, municipal, and rain flood discharge outlets for high-frequency monitoring of discharge flow and key pollution factors.

3. The system of claim 1, wherein, the intelligent analysis and tracing engine comprises: a data fusion unit for standardizing, removing outliers, and spatio-temporally aligning multi-source heterogeneous data from different monitoring units to construct a unified spatio-temporal data set; a dynamic early warning unit for performing threshold comparison based on water quality standards, trend deviation detection based on historical behavior patterns, and abnormal propagation analysis based on the correlation between upstream and downstream monitoring points in parallel based on the unified spatio-temporal data set to trigger early warning signals comprehensively; an intelligent tracing unit for starting a collaborative tracing process in response to the early warning signals, which integrates pollution diffusion inversion based on hydrodynamics, discharge contribution analysis based on spatio-temporal correlation degree, and pollution source probability inference based on a deep learning model.

4. The system of claim 3, wherein: the trend deviation detection in the dynamic early warning unit identifies statistical anomalies of a monitoring indicator time series relative to historical normal patterns by analyzing the dynamic characteristics of the time series; the correlation abnormal propagation analysis automatically triggers a directional scan of potential discharge sources in the middle region when an abnormality appears downstream without a chain abnormality upstream by constructing a data correlation model between upstream and downstream monitoring points.

5. The system of claim 3, wherein: the intelligent tracing unit uses a simplified hydrodynamic model to simulate the migration and diffusion process of pollutants in the upstream watershed of the early warning point to determine the possible release time window and spatial range of the pollutants; within the time window and spatial range, spatio-temporal matching degree calculation is performed on the discharge data of each discharge monitoring node and the downstream water quality abnormal data to identify a high-suspected discharge outlet set; and the feature data, hydrological conditions, and historical discharge patterns of the discharge outlet set are input into a pre-trained deep learning network to output the probability estimate of each suspected discharge outlet being a real pollution source.

6. The system of claim 5, wherein: The part of the intelligent traceability unit that performs the spatiotemporal matching degree calculation is configured to run a spatiotemporal contribution entropy-based collaborative traceability algorithm; the algorithm calculates the time coincidence degree of the emission behavior of each suspected outlet within the pollution event time window and the downstream water quality anomaly, and combines the spatial distribution anomaly degree of the emission intensity relative to the historical level of the outlet, to construct a unified spatiotemporal contribution evaluation index; then, the algorithm calculates the comprehensive suspicion degree of each outlet according to the index, and generates a traceability list sorted by suspicion probability.

7. The system of claim 5, wherein: The deep learning network is trained by historical pollution event samples to learn the propagation characteristics and source identification mode of pollution events under complex hydrological and emission conditions; the probability estimate comprehensively considers the emission time coincidence degree, concentration change correlation, and historical behavior deviation degree factors.

8. The system of claim 7, wherein: The deep learning network is specifically an adaptive graph attention neural network; the network abstracts the monitoring sites and outlets as graph nodes, and the water flow direction and hydraulic connection as directed edges, and dynamically calculates the influence weight between nodes by an attention mechanism embedded with real-time hydrological influence factors.

9. The system of claim 1, wherein: The early warning and decision support platform issues early warning information of different levels according to the severity and influence range of the pollution event, and pushes the information to relevant management personnel through various human-computer interaction channels; On the electronic map, the monitoring point state, water quality parameter dynamic change, pollution plume simulation diffusion range, and the spatial distribution and probability sorting of the suspected outlet locked by the system are presented in real time; and based on the analysis results of the engine, a traceability decision report including event chain restoration, evidence support data, suspected source sorting, and disposal suggestions is automatically generated.

10. A pollution early warning and traceability method for an estuary river, wherein: The implementation of the method is based on the system of any one of claims 1-9: The method comprises: Continuously collecting multi-source monitoring data of the river water body, river-inlet outlet, and hydrological and meteorological conditions, and performing standardization and spatiotemporal alignment processing on the multi-source monitoring data to form a unified spatiotemporal data set; Based on the unified spatiotemporal data set, threshold comparison, trend deviation detection, and upstream and downstream correlation analysis are performed in parallel to trigger a dynamic early warning; Based on the principle of hydrodynamics, the migration and diffusion process of the pollutant is simulated to deduce the possible release time and spatial range of the pollutant; Within the time and spatial range, the spatiotemporal correlation matching degree of the emission data of each outlet and the downstream water quality anomaly data is calculated; The emission characteristics, hydrological conditions, and historical behavior mode are input into a pre-trained deep learning model to obtain the traceability probability of each suspected outlet; Based on the results of the dynamic early warning and intelligent traceability, graded early warning information, a visual situation map, and a structured traceability report are generated and output.

Images