Pollution traceability decision support method based on multi-source data fusion and agent collaboration

Abstract

This application discloses a pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration, relating to the field of ecological and environmental monitoring technology, and possessing the support capabilities of perception, learning, reasoning, analysis, generation, and scheduling. The method is implemented based on a cloud-edge-device collaborative architecture, including: adaptive perception and preprocessing of multi-source heterogeneous environmental data at the edge, and uploading the processed data to a cloud server; in the cloud, based on an environmental professional rule base, knowledge base, and data foundation, and utilizing retrieval-enhanced generation technology combined with a large language model to perform data fusion analysis and judgment, generating environmental situation analysis results and disposal strategies, further generating target data packages and sending them to preset target terminal devices, receiving on-site feedback signals, incrementally updating the cloud-based environmental professional rule base, environmental professional knowledge base, and model parameters, and feeding the updated rule packages and / or model parameters back to the edge-side nodes to achieve closed-loop optimization of the system.

Description

Technical Field

[0001] This application relates to the field of ecological and environmental monitoring technology. Specifically, it relates to a pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration, which is particularly suitable for application scenarios such as real-time monitoring of atmospheric environmental quality, anomaly identification, pollution source tracing, and auxiliary decision scheduling. Background Technology

[0002] With the deepening of ecological civilization construction, ecological and environmental governance is undergoing a profound transformation from digital management to intelligent decision-making support. In this process, data such as air quality monitoring, weather forecasting, online pollution source monitoring, satellite remote sensing, mobile monitoring, spatial geographic information, policies, standards and norms, governance case databases, and daily work reports are continuously accumulating, providing a rich data foundation for environmental management. However, in actual operation, this data is often scattered across different business platforms, belonging to different functional departments, or sealed in various unstructured files, forming serious "data silos" that are difficult to access, correlate, and quickly translate into concrete management actions.

[0003] Especially in management scenarios at the district and county levels and at the grassroots level, environmental management personnel face immense work pressure and complex task challenges. They are not only responsible for real-time data monitoring, anomaly identification, compliance assessment, and early warning information dissemination, but also for on-site pollution source tracing, cross-departmental coordination, policy consultation, and the preparation of various reports. Under traditional operating models, these tasks heavily rely on manual multi-system queries, manual data collection, experience-based situation assessment, and manual document writing. This approach not only limits response speed but also demands extremely high levels of professional expertise. Faced with massive amounts of monitoring data, complex pollutant transmission processes, and the heavy workload of multi-departmental collaboration, the contradiction between limited human and technical resources at the grassroots level and the ever-increasing demand for regulatory efficiency is becoming increasingly prominent, making it difficult to meet the urgent requirements of precise pollution control, scientific pollution control, and rapid closed-loop disposal.

[0004] Currently, various information systems have been built in the field of air quality monitoring and control. These mainly include: an online air quality monitoring platform, used to display real-time concentration data and trends of pollutants such as sulfur dioxide, nitrogen oxides, and fine particulate matter (PM2.5) at various monitoring stations; an online pollution source monitoring platform, used to connect the automatic monitoring equipment of key pollutant discharge units to realize real-time collection of pollutant emission data and alarms for exceeding standards; a meteorological forecasting and analysis platform, providing observation and numerical forecast results of meteorological elements such as wind field, temperature stratification, and precipitation, serving the analysis of pollution diffusion conditions; a Geographic Information System (GIS) display system, used to present spatial information such as the distribution of monitoring stations, spatial location of pollution sources, land use, and road networks; an environmental data reporting and statistics system, which generates various daily, weekly, monthly, and ranking reports based on preset templates; a general knowledge base and question-and-answer tools, as well as a dialogue system based on a general large language model, attempting to provide text retrieval and question-and-answer services for environmental protection policies, standards, and norms. While existing systems can perform single-dimensional data display, pollution source monitoring, weather trend forecasting, spatial distribution visualization, fixed-format report export, or general domain knowledge Q&A, most systems still primarily operate as single systems, handle single tasks, and analyze single data sources. This fragmented data management model leads to the following significant shortcomings in practical operation:

[0005] First, the data and knowledge infrastructure are fragmented, resulting in poor business synergy. Existing technologies lack a unified data and knowledge infrastructure, making it impossible to effectively link and integrate air quality, meteorological, pollution source, remote sensing, GIS spatial data, policy standards, and historical cases. When managers need to conduct pollution cause analysis, calculate compliance margins, or coordinate across departments, they still heavily rely on manually logging into multiple systems to query data and manually integrating and compiling statistics based on experience—a cumbersome and inefficient process.

[0006] Secondly, the rigid business rules and calculation methods make it difficult to adapt to dynamic management needs. Core indicators related to air quality assessment, such as real-time values, cumulative values, year-on-year changes, city rankings, and remaining allowable concentrations for compliance, are often calculated using fixed reports or manual spreadsheet software. Faced with dynamic adjustments to different cities, monitoring stations, and annual assessment targets, the existing fixed reporting model cannot flexibly configure calculation rules, making it difficult to quickly and accurately provide dynamic assessment results.

[0007] Third, the general-purpose model lacks environmental business knowledge constraints, resulting in insufficient professionalism and traceability in its answers. Although the general-purpose language model question-answering tool demonstrates certain natural language interaction capabilities, its knowledge sources are too broad, lacking constraints on local environmental protection policies, specific technical specifications, historical cases, and internal research findings. When directly applied to environmental management scenarios, it easily generates "illusionary" content detached from actual business realities, and the generated answers fail to meet the rigorous environmental management requirements in terms of basis, traceability, and business adaptability.

[0008] Fourth, the system's functionality emphasizes visual display, while its lightweight analysis and closed-loop response capabilities are weak. Most traditional platforms are designed with data display and chart generation in mind, failing to logically link key aspects such as anomaly identification, changes in meteorological conditions, pollution source emission inventories, spatial transmission trajectories, mobile monitoring, and on-site response feedback. When high pollution levels are detected, the system cannot automatically correlate multi-source data to provide possible priority directions for pollution source tracing and specific on-site investigation suggestions, resulting in management actions still relying on secondary analysis based on human experience.

[0009] Fifth, the system entry points are fragmented, resulting in high barriers to entry for grassroots users. For district and county-level managers, law enforcement personnel, monitoring staff, and grid workers, completing daily tasks such as real-time data monitoring, anomaly identification, compliance assessment, early warning dissemination, on-site source tracing, cross-departmental coordination, policy consultation, and document preparation requires repeatedly switching between multiple independent systems. The lack of a lightweight, unified entry point makes it difficult for users to quickly and conveniently access the necessary data, analytical models, knowledge bases, and report templates, increasing learning costs and operational burdens.

[0010] Sixth, there is a lack of continuous feedback and self-optimization mechanisms. Most existing systems operate on a one-way information output model, lacking effective capture of user interaction and learning feedback. User corrections and annotations to system-pushed alerts, Q&A responses, image recognition results, and generated report drafts cannot be automatically recorded and used to optimize the knowledge base, business rule base, predictive model parameters, and report generation template strategies. This makes it difficult for system performance to continuously improve through actual use. Summary of the Invention

[0011] This application provides a pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration, aiming to solve the technical problems in the prior art such as scattered multi-source environmental data, lack of accurate guidance for pollution source tracing, delayed abnormal response, and inability of the system to continuously optimize the closed loop.

[0012] The specific technical solution is as follows: In a first aspect, embodiments of this application provide a pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration. The method is implemented based on a cloud-edge-device collaborative architecture, including: Multi-source heterogeneous environmental data is acquired through edge-side nodes. The multi-source heterogeneous environmental data includes air quality monitoring data, meteorological data, online pollution source monitoring data, electricity consumption monitoring data, mobile monitoring data, satellite remote sensing data, video surveillance data, as well as unstructured regulatory text data and historical case document data. Adaptive sensing and preprocessing of the multi-source heterogeneous environment data are performed at the edge. The data collection frequency and caching strategy are dynamically adjusted according to the risk level of the environment. The processed multi-source heterogeneous environment data is then uploaded to the cloud server. Based on the constructed environmental professional rule base, environmental professional knowledge base and data foundation in the cloud, and using retrieval enhancement generation technology combined with a large language model, the processed multi-source heterogeneous environmental data is fused and analyzed to generate environmental situation analysis results and disposal strategies. Based on the environmental situation analysis results and the disposal strategy, a target data packet is generated and sent to a preset target terminal device via a wireless communication interface. The target terminal device then returns a field feedback signal. The target data packet includes at least one of the following: early warning data, pollution source location data, or disposal instruction data. Based on the on-site feedback signals, the environmental professional rule base, the environmental professional knowledge base, and the model parameters in the cloud are incrementally updated, and the updated rule package and / or model parameters are fed back to the edge-side nodes to achieve closed-loop optimization of the system.

[0013] In some embodiments of this application, the adaptive sensing and preprocessing of the multi-source heterogeneous environmental data at the edge specifically includes: When the multi-source heterogeneous environmental data meets the first preset condition, the data acquisition mode is switched from the normal acquisition mode to the abnormal acquisition mode. In the abnormal acquisition mode, the data acquisition frequency is adjusted from the first frequency to the second frequency, and / or the number of monitoring nodes per unit area in the monitoring area is adjusted from the first total number of nodes to the second total number of nodes. Wherein, the second frequency is greater than the first frequency, the second total number of nodes is greater than the first total number of nodes, and the first preset condition includes at least one of the following: the indicator value reaches a first percentage of the preset warning threshold, the meteorological diffusion index is lower than the first preset diffusion threshold, or the fluctuation range of the online monitoring data of the pollution source exceeds the first preset fluctuation range. When the multi-source heterogeneous environmental data meets the second preset condition, the data acquisition mode is switched to the early warning perception mode. In the early warning perception mode, the data retention time of the dynamic ring buffer set on the edge side is extended from the first time to the second time. When the multi-source heterogeneous environmental data meets the third preset condition, the data acquisition mode is switched to the anomaly perception mode. In the anomaly perception mode, based on the trigger time of the anomaly perception mode, the data in the dynamic ring buffer located between the third time before the benchmark and the fourth time after the benchmark is marked as non-coverable data. The non-coverable data is extracted to generate a pollution event tracing data packet. The second preset condition... The conditions include at least one of the following: the indicator value reaches a second percentage of the preset warning threshold, the meteorological diffusion index is lower than the second preset diffusion threshold, or the fluctuation range of the online monitoring data of the pollution source exceeds the second preset fluctuation range; the third preset condition includes at least one of the following: the rate of change of the monitoring data within a unit time window exceeds the first preset rate of change threshold, the monitoring data continuously exceeds the first preset high threshold in multiple consecutive sampling cycles, the instantaneous value of the monitoring data exceeds the first preset over threshold, the temporal logical relationship or numerical ratio relationship between multiple sets of associated monitoring data deviates from the preset benchmark interval, or the characteristic fingerprint in the monitoring data matches the abnormal spectrum in the preset pollution source feature library; Missing value detection is performed on the structured data in the multi-source heterogeneous environmental data, marking null values ​​in the data fields as missing values. Outlier detection is also performed on the structured data, calculating the deviation of each data item from the mean, and marking data items with deviation values ​​exceeding a preset deviation threshold as candidate outliers. The structured data is then processed by removing duplicate values, verifying timestamps, matching site codes, matching spatial locations, and converting pollutant indicator calibers. Simultaneously, the unstructured data in the multi-source heterogeneous environmental data is processed by text parsing, chapter segmentation, tag extraction, and vectorization. The preprocessed relevant outlier data and the uncoverable data are encapsulated into a high-priority data packet and sent to the cloud server first through a communication interface. The relevant outlier data includes at least one of the following: amplitude domain outlier features, temporal gradient features, threshold proximity values, pollution source outliers, or spatial hotspot data.

[0014] In some embodiments of this application, the step of using retrieval enhancement generation technology combined with a large language model to fuse and analyze the processed multi-source heterogeneous environmental data specifically includes: Based on the preset reliability parameters of each data source, the time deviation between the data acquisition timestamp and the reference time base, and the spatial proximity between the corresponding monitoring points of each data source, the dynamic reliability weight of each data source is calculated, and the data of the data source with a dynamic reliability weight lower than the first preset value in the processed multi-source heterogeneous environment data is removed, so as to realize the cross-validation of the processed multi-source heterogeneous environment data. The index values ​​in the verified data are compared with the preset index thresholds. If any index value exceeds the preset index threshold, a deep analysis is triggered. The verified data is then input into the spatiotemporal prediction model, the pollution source tracing model, and the knowledge graph to obtain pollution diffusion trend data and candidate pollution source data. Based on the feature vectors of the candidate pollution source data, the corresponding regulatory text data and similar historical case document data are retrieved from the environmental professional knowledge base. The retrieval results are combined with the pollution diffusion trend data to form an enhanced prompt context. The enhanced prompt context is then input into a large language model to output a judgment conclusion text containing a description of the pollution event and disposal suggestions.

[0015] In some embodiments of this application, after generating the environmental situation analysis results and handling strategies, the method further includes: generating a precise source tracing and collaborative scheduling scheme, specifically including: Based on the pollution source location information in the environmental situation analysis results, combined with meteorological wind field vector data, mobile mission planning data containing at least one path coordinate sequence is generated; wherein, based on the pollutant concentration value at each pollution source location and the preset sorting rules, the investigation priority weight value of each pollution source is calculated, and pollution source investigation sequence data is generated in descending order of the investigation priority weight values. Based on the mobile mission planning data and the preset spatial grid ownership data, determine the jurisdiction identifier and timeliness identifier corresponding to each grid; The event markers, time markers, pollution source identifiers, and mobile mission planning data from the environmental situation analysis results are encapsulated into a structured mission data package according to a preset data structure template, and the structured mission data package is sent to the target terminal device corresponding to the jurisdiction identifier through a communication interface.

[0016] In some embodiments of this application, the method further includes a scene-adaptive dynamic threshold determination step in the cloud, specifically including: The system obtains the scenario type identifier of the current environmental monitoring scenario and retrieves the corresponding benchmark index threshold from the preset benchmark threshold library based on the scenario type identifier. The scenario type identifier includes static and stable accumulation scenario, dust transport scenario, and major event support scenario. The benchmark index threshold is composed of a static benchmark layer, a dynamic benchmark layer, and a variable strategy layer. The static benchmark layer contains at least one set of national mandatory emission limits and corresponding local emission assessment characteristic values. The variable strategy layer temporarily adjusts the benchmark values ​​of the static benchmark layer by a multiplier based on a user-defined time window weighting factor and a local area scaling factor to obtain the benchmark index threshold. The dynamic benchmark layer is used to determine whether the residual exceeds a preset multiple standard deviation of the benchmark index threshold to calibrate statistical anomaly characteristics. Acquire current real-time meteorological data, and calculate the correction coefficient of each meteorological factor to the benchmark index threshold based on the current real-time meteorological data; the current real-time meteorological data includes meteorological condition parameters, boundary layer height, wind speed, wind direction and humidity data, historical level data for the same period, station background concentration data, pollution source operation status data, regional transmission characteristic data, time-specific dynamic warning line and expected compliance difficulty coefficient; Based on the benchmark threshold and the correction coefficient, a preset index threshold adapted to the current environmental monitoring scenario is calculated. Based on the preset index threshold, anomaly determination is made on the pollutant concentration monitoring values ​​in the processed multi-source heterogeneous environmental data to obtain anomaly marking results. In the static and stable accumulation scenario, the alarm sensitivity of the preset index threshold is higher than that of the benchmark index threshold; in the sand and dust transport scenario, the alarm sensitivity of the preset index threshold is lower than that of the benchmark index threshold.

[0017] In some embodiments of this application, the step of incrementally updating the environmental professional rule base, the environmental professional knowledge base, and model parameters in the cloud based on the field feedback signal, and feeding the updated rule package and / or model parameters back to the edge node, specifically includes: Acquire the original monitoring data, edge preprocessing result data, abnormal triggering rule identifiers, model judgment result data, push record data, and on-site verification feedback data corresponding to the current pollution event; According to the preset association rules, establish the temporal and causal association relationships between the original monitoring data, the edge preprocessing result data, the anomaly triggering rule identifier, the model judgment result data, the push record data, and the on-site verification feedback data, and generate a pollution event association data record set; When the cumulative number of abnormal records belonging to the same pollution type or the same monitoring station in the pollution event associated data record set reaches a preset threshold, rule update data or model parameter increment data is generated based on the verified data in the pollution event associated data record set. The rule update data and / or the incremental data of the model parameters are distributed to the edge nodes through a distributed model aggregation method, so that the edge nodes update the decision rule data and / or model parameter data stored locally.

[0018] In some embodiments of this application, the method further includes: When the environmental situation analysis results generated in the cloud meet the preset supplementary sampling trigger conditions, a recall command is generated and sent to the designated edge node; wherein, the preset supplementary sampling trigger conditions include at least one of the following: the concentration change rate of the current monitoring data within a preset time window exceeds a second preset change rate threshold, the difference between the current monitoring data and the monitoring data at the previous moment exceeds a first preset difference threshold, or the pollutant concentration value in the remote sensing inversion data exceeds a preset high value threshold; In response to the call command, the edge node performs a supplementary data acquisition operation, which includes: adjusting the data acquisition frequency corresponding to the current monitoring data from a first frequency to a third frequency, extracting historical monitoring data within the time range specified by the call command from the local cache of the edge node, and simultaneously acquiring associated video surveillance data and / or mobile trajectory data within the time range, and sending the historical monitoring data, the associated video surveillance data, and the mobile trajectory data as a supplementary data acquisition data packet to the cloud server; wherein the third frequency is greater than the first frequency.

[0019] In some embodiments of this application, the method further includes: Each edge node obtains the overall upload weight of the current data packet through its built-in data priority quantization engine, and each edge node then performs the upload based on the overall upload weight. Compared with the preset first priority threshold Second priority threshold Based on the comparison results, hierarchical transmission scheduling is performed on the current data packet: when When the current data packet is marked as a high-priority data packet, a real-time interruption of the upload is triggered, and the transmission delay does not exceed a first preset duration; when When the current data packet is marked as a medium-priority data packet, it is stored in a local cache queue and uploaded using a time window aggregation method. The aggregation window width is a second preset duration, and the second preset duration is greater than the first preset duration. When this happens, the current data packet is marked as a low-priority data packet, stored in local memory, and then uploaded in batches during a preset off-peak communication period; wherein, The comprehensive upload weight The calculation formula is: ; In the above formula, This is a normalized risk coefficient based on the current monitored value relative to the predicted statutory limit; Priority encoding value for sensor data type; Quality score based on signal integrity verification; This represents the influence radius coefficient of the current edge node in the spatial topology. It is a monotonically decreasing function representing the remaining time from the preset target termination time window at the current time; These are the dynamic configuration coefficients input through an external configuration interface; This represents the current real-time load rate in the cloud. a, b, c, d, e, g, h These are the preset weight coefficients for each factor, and their sum is 1; When the cloud server receives anomaly confirmation signals triggered simultaneously by multiple edge nodes, it executes a multi-node concurrent resolution step: It constructs a spatiotemporal coupling sorting factor for each anomaly event, sorts them in descending order according to the value of the spatiotemporal coupling sorting factor, generates an anomaly event processing priority sequence list, and dynamically determines the number K of events to be processed in a single batch based on the current real-time load rate of the cloud server. It then allocates the top K anomaly data packets in the anomaly event processing priority sequence list to the cloud server's reserved computing power resource pool, triggering a fusion analysis process. Simultaneously, it stores the top K+1 anomaly data packets and subsequent packets in the anomaly event processing priority sequence list into a waiting queue and starts a cyclic monitoring process. Every preset time interval, it re-acquires the current real-time load rate of the cloud server and repeats the aforementioned steps until all anomaly data packets have been allocated to the reserved computing power resource pool. The spatiotemporal coupling sorting factor... The calculation formula is: ; In the above formula, For the first Monitoring values ​​uploaded by each edge node This indicates the normalization overshoot factor of the node; For the first The edge side node and the first The Euclidean spatial distance between each edge node in the Deloni triangulation. Indicates spatial proximity and density; The upwind source probability value is calculated based on real-time wind field vector data; For the first The remaining time between the current time of each edge node and the preset target termination time window. This indicates the time urgency of the target task; , , , These are the preset weight coefficients for each sub-item, and their sum is 1.

[0020] In some embodiments of this application, the method further includes: The cloud server maps the processed multi-source heterogeneous environmental data sent by multiple edge nodes to a unified high-dimensional feature space through a preset feature encoder, and calculates the multi-source fusion confidence of abnormal events. The calculation formula is as follows: ; In the above formula, For the first Anomaly confidence scores from each data source output, ; For the first The dynamic fusion weights of each data source, wherein the sum of all weights is 1, wherein the dynamic fusion weights It is determined by the product of data credibility factor, spatial proximity factor, time synchronization factor, anomaly persistence factor, historical similarity factor, transmission direction consistency factor and preset configuration coefficient; When the multi-source fusion confidence level When the confidence level exceeds a preset first confidence threshold, an anomaly confirmation signal is output; when the multi-source fusion confidence level... When the confidence level is less than the first confidence threshold and greater than the preset second confidence threshold, it is marked as pending review; when the multi-source fusion confidence level... If the value is less than the second confidence threshold, it is determined to be isolated noise and discarded.

[0021] Secondly, embodiments of this application provide a pollution source tracing decision support system based on multi-source data fusion and intelligent agent collaboration, used to implement the method described in the first aspect, the system comprising: Edge nodes, used to access multi-source heterogeneous environmental data, include a preprocessing module and a first communication module. The multi-source heterogeneous environmental data includes air quality monitoring data, meteorological data, online pollution source monitoring data, electricity consumption monitoring data, mobile surveillance data, satellite remote sensing data, video surveillance data, and unstructured regulatory text data and historical case document data. The preprocessing module is used to perform adaptive sensing and preprocessing on the multi-source heterogeneous environmental data, dynamically adjusting the data acquisition frequency and caching strategy according to the risk level of the environmental state. The first communication module is used to upload the processed multi-source heterogeneous environmental data to a cloud server. The cloud server includes an analysis module, a generation module, a second communication module, and an incremental update module. The analysis module is used to perform fusion analysis on the processed multi-source heterogeneous environmental data based on the constructed environmental professional rule base, environmental professional knowledge base, and data foundation, and using retrieval enhancement generation technology combined with a large language model to generate environmental situation analysis results and disposal strategies. The generation module is used to generate target data packets based on the environmental situation analysis results and the disposal strategies. The second communication module is used to send the target data packets to a preset target terminal device through a wireless communication interface and receive on-site feedback signals returned by the target terminal device. The incremental update module is used to incrementally update the environmental professional rule base, environmental professional knowledge base, and model parameters in the cloud based on the on-site feedback signals, and feed the updated rule packets and / or model parameters back to the edge-side nodes to achieve closed-loop optimization of the system. The target data packets include at least one of early warning prompt data, pollution source location data, or disposal instruction data.

[0022] In some embodiments of this application, the cloud server further includes: The environmental indicator calculation and anomaly judgment module is used to automatically calculate the statistical sequence within the current monitoring period according to a preset limit reference system, and to judge the statistical sequence as high value, exceeding threshold, inverted, equipment abnormal, exceeding standard risk or pollution risk based on an environmental professional rule base; wherein, the statistical sequence includes real-time value, cumulative value, comparison value, ranking value, remaining capacity, time series progress deviation and high value statistics; A task distribution arbitrator is used to forward the current task to the corresponding processing module according to a preset routing mapping table based on the judgment result label and the request type code sent by the terminal device. The processing module includes an air quality information formatting module, a compliance status monitoring module, an abnormal event alarm output module, a forecast and early warning generation module, an automatic map generation module, a semantic retrieval and generation module, a multi-source data association analysis module, and a distributed task coordination module. The air quality information formatting module and the compliance status monitoring module are used to generate structured data reports based on the judgment results of the environmental indicator calculation and abnormal judgment modules. The structured data reports include real-time data reports, cumulative statistical reports, and... The system includes reports on differences, remaining capacity, time-series progress deviations, and site anomalies. The automatic map generation module uses meteorological numerical forecast output, air quality numerical model prediction results, pollution source emission inventories, remote sensing image inversion data, and geographic information system spatial data to automatically generate air quality prediction maps, meteorological anomaly maps, concentration spatial distribution maps, transmission path maps, remote sensing inversion maps, source tracing direction maps, time-series trend maps, or multi-source comparison maps. The semantic retrieval and generation module uses a multimodal index based on a knowledge base to perform query encoding, semantic retrieval, full-text matching retrieval, tag filtering retrieval, multi-path fusion sorting, and query generation output processing on query statements input through the user command interface on the terminal device. The comprehensive analysis report generation module is used to integrate the statistical sequence, judgment result labels, retrieval knowledge text and structured instruction set, and output a structured analysis report containing anomaly attribution label sequence, execution action sequence, data statistical summary, spatial distribution summary and trend prediction summary.

[0023] In some embodiments of this application, the preprocessing module includes: The risk perception unit is used to generate a mode switching signal when the multi-source heterogeneous environmental data meets a first preset condition; the first preset condition includes at least one of the following: the indicator value reaches a first percentage of a preset warning threshold, the meteorological diffusion index is lower than a preset diffusion threshold, or the fluctuation range of the online monitoring data of the pollution source exceeds a preset fluctuation range. The parameter adjustment module is used to adjust the data acquisition frequency from a first frequency to a second frequency in response to the mode switching signal, and / or adjust the number of monitoring nodes per unit area in the monitoring area from the first total number of nodes to the second total number of nodes; wherein, the second frequency is greater than the first frequency, and the second total number of nodes is greater than the first total number of nodes. A dynamic circular buffer is used to cache environmental monitoring data. The data retention time is determined according to the current acquisition mode: when in the normal acquisition mode, the data retention time is a first duration; when in the abnormal acquisition mode, the data retention time is a second duration; the second duration is longer than the first duration; the dynamic circular buffer is also used to mark data in the dynamic circular buffer that is within the range of the third duration before the reference to the fourth duration after the reference when in the abnormal acquisition mode, based on the trigger time of the abnormal acquisition mode, as non-coverable data, extract the non-coverable data and generate a pollution event tracing data packet; The structured data verification unit is used to perform missing value detection on the structured data in the multi-source heterogeneous environment data, mark null values ​​in the data fields as missing values, and perform outlier detection on the structured data, calculate the deviation value of each data item from the mean, and mark data items with deviation values ​​exceeding a preset deviation threshold as candidate outliers. An unstructured data parsing unit is used to perform text parsing on the unstructured data in the multi-source heterogeneous environment data, and extract entity feature words and numerical feature parameters from the unstructured data; The priority marking unit is used to encapsulate the candidate outliers, the statistical summary of the unstructured data, the entity feature words, and the numerical feature parameters into a high-priority data packet.

[0024] In some embodiments of this application, the analysis module includes: The cross-validation unit is used to calculate the dynamic reliability weight of each data source based on the preset reliability parameters of each data source, the time deviation between the data acquisition timestamp and the reference time base, and the spatial proximity between the corresponding monitoring points of each data source, and to remove data from the data source whose dynamic reliability weight is lower than the first preset value in the processed multi-source heterogeneous environment data, so as to realize the cross-validation of the processed multi-source heterogeneous environment data. The first-level analysis unit is used to perform first-level analysis on the verified data, comparing the indicator values ​​in the verified data with preset indicator thresholds to obtain the first-level analysis results. The second-level analysis unit is used to perform a second-level analysis when the result of the first-level analysis meets a preset trigger condition. The verified data is input into the spatiotemporal prediction model, the pollution source tracing model, and the knowledge graph to obtain pollution diffusion trend data and candidate pollution source data. The preset trigger condition includes: the existence of an indicator value that exceeds the preset indicator threshold. The retrieval and conclusion output unit is used to retrieve the corresponding regulatory text data and similar historical case document data from the environmental professional knowledge base based on the feature vector of the candidate pollution source data, combine the retrieval results with the pollution diffusion trend data to form an enhanced prompt context, input the enhanced prompt context into the large language model, and output the judgment conclusion text containing the pollution event description and disposal suggestions.

[0025] The beneficial effects of the embodiments of this application are as follows: This application constructs a unified fusion framework for multi-source heterogeneous data, deeply integrating diverse information such as air quality, meteorology, pollution sources, remote sensing, GIS, policies and standards, governance cases, and historical reports. This effectively solves the inherent problems of existing systems, including data fragmentation, functional isolation, and insufficient comprehensive analysis capabilities. It not only breaks down data silos but also lays a solid data foundation for intelligent collaborative applications. Furthermore, in terms of improving environmental management efficiency and intelligence, this application achieves uninterrupted 24 / 7 automatic monitoring capabilities, enabling continuous real-time monitoring and anomaly identification of core environmental data 24 / 7, 365 days a year. The system automatically completes high-frequency repetitive tasks such as data tracking, data retrieval, cumulative calculation, year-on-year ranking, compliance margin analysis, anomaly alarms, initial report generation, and chart output, significantly reducing the workload of frontline personnel and greatly improving the timeliness and accuracy of management responses. Meanwhile, based on the deep coupling analysis of high values ​​at stations, pollution source emissions, meteorological conditions, spatial distribution, regional transmission, and historical cases, the system can accurately generate priority investigation directions, suggested navigation routes, and inspection lists, providing a clear entry point for on-site source tracing work and effectively avoiding the waste of resources caused by indiscriminate investigation.

[0026] Furthermore, this application leverages environmental expertise bases, knowledge graphs, and large language model-driven retrieval enhancement generation technology to enable grassroots personnel to conveniently access policy basis and governance measures through natural language interaction, thus democratizing complex assessment techniques. Based on this, the system connects the entire business chain from anomaly identification, problem assessment, responsibility matching, task assignment, to result verification, forming a cross-departmental closed-loop management model of "assessment – ​​task assignment – ​​handling – feedback – verification," ensuring efficient flow and substantive resolution of environmental issues. Moreover, this application possesses high security, controllability, and cost-effectiveness. Through multiple security mechanisms such as data isolation, hierarchical authorization, encrypted transmission, and content compliance verification, the system's reliable operation within the authorized scope is guaranteed. The system operates on existing monitoring infrastructure, requiring no large-scale new hardware investment, effectively revitalizing existing data assets and transforming them into executable and pushable business leads. This application adopts a flexible architecture of "thick middle platform, thin front-end," which can not only be rapidly replicated and promoted in the field of air pollution prevention and control but also expanded as needed to diverse ecological and environmental governance scenarios such as water environment, soil, and solid waste, demonstrating excellent scalability and broad application prospects. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A flowchart illustrating the pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration provided in Embodiment 1 of this application; Figure 2 A schematic diagram of the composition of the pollution source tracing decision support system based on multi-source data fusion and intelligent agent collaboration provided in Embodiment 2 of this application; Figure 3 This is a closed-loop schematic diagram of the pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration provided in Embodiment 3 of this application. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The terms "comprising" and "having," and any variations thereof, in the embodiments and drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0031] Example 1 This application, in its first embodiment, addresses the technical challenges of multi-source heterogeneous ecological and environmental monitoring data, complex spatiotemporal correlations, and nonlinear evolution of pollution processes. It discloses a pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration. This method is implemented based on a cloud-edge-device collaborative system architecture and possesses the capabilities of perception, learning, reasoning, judgment, generation, and scheduling.

[0032] The system architecture comprises three physical layers: cloud, edge, and terminal, with each layer operating collaboratively. Edge nodes are deployed at various air quality monitoring stations, emission outlets of key pollution sources, mobile monitoring vehicles, and grid-based monitoring areas, possessing data acquisition, preliminary processing, and communication capabilities. Each edge node integrates a data acquisition interface and a lightweight computing unit, responsible for accessing raw physical quantity data from local sensors (including gas concentration sensors, meteorological parameter sensors, and video cameras), performing preprocessing operations such as data cleaning, missing value imputation, and outlier removal, and transmitting standardized data to the cloud via an encrypted communication protocol. Simultaneously, each edge node runs a lightweight inference engine locally, performing sliding window detection and dynamic threshold comparison on real-time data streams to achieve millisecond-level preliminary identification, local verification, and early warning of anomalies. The cloud server, deployed in a central computer room or government cloud platform, serves as the system's core computing and global decision-making center, running a multi-source data fusion engine, a large-scale pre-trained model cluster, a knowledge graph construction and evolution module, and various deep learning inference models. It undertakes computationally intensive tasks such as global situation modeling, cross-regional pollution propagation simulation, incremental model training, and dynamic updates of the knowledge base. The target terminal devices can be mobile terminals of ecological and environmental law enforcement personnel (such as law enforcement PDAs and smartphones), government WeChat accounts, enterprise WeChat accounts, or business terminals of grid workers, etc. They receive warning signals, judgment conclusions, navigation route suggestions, and report summaries from the cloud and edge terminals through mobile networks or dedicated lines, and present them in the form of text messages, visual charts, map layer overlays, etc. for human-computer interaction to support on-site verification and disposal decisions.

[0033] The core functions of the system are achieved collaboratively by the data fusion layer, the knowledge construction layer, and the intelligent reasoning layer.

[0034] The data fusion layer establishes a pipeline for spatiotemporal alignment and standardization of multi-source heterogeneous data. Furthermore, after processing the multi-source data through spatiotemporal interpolation and data assimilation algorithms, a unified gridded ecological environment status tensor is generated, which serves as the unified mathematical input for subsequent model inference.

[0035] Based on the aforementioned situation tensor and historically accumulated environmental data, the knowledge construction layer builds a structured and traceable environmental professional knowledge base. Specifically, it employs knowledge graph technology to construct a semantic network linking pollution sources, pollutants, meteorological conditions, and spatiotemporal locations, supporting explicit modeling and multi-hop reasoning of relationships between entities (such as emission sources, monitoring stations, and administrative regions). A Retrieval-Augmented Generation (RAG) framework is used to segment, vectorize, and index unstructured text data (including environmental protection laws and regulations, policy documents, standards and specifications, historical pollution event records, remediation cases, and assessment reports), supporting semantic retrieval and knowledge recall by pollution type, geographical region, time scale, and other dimensions. A dynamic update mechanism for the knowledge base is established, where local verification data returned from the edge and the processing results fed back from the terminal serve as new labeled samples. After cloud review, this triggers updates to the relation weights of the knowledge graph and incremental reconstruction of the vector index.

[0036] The intelligent inference layer deploys multiple deep learning models, covering four core inference tasks: prediction, parsing, question answering, and generation. The spatiotemporal prediction model, based on the Transformer architecture, takes historical trend tensor sequences as input and outputs predicted pollutant concentrations for each monitoring point and grid cell over the next 72 hours, along with confidence intervals. The pollution source tracing model, based on a graph neural network, uses the spatial topology of pollution sources and the meteorological transport matrix as a graph structure to quantitatively calculate the contribution weight of each potential emission source to high-value events at the target site, outputting a ranking of contributions. The semantic understanding and intelligent question answering model, based on a large language model, identifies the intent and fills in slots for ecological and environmental queries input by end-users in natural language, and uses knowledge graph retrieval and RAG recall results for joint inference to generate structured answer text and supporting citations. The automatic report and suggestion generation model, based on the text generation capabilities of the large language model, takes prediction results, source tracing conclusions, and related policy standards and cases from the knowledge base as input, and automatically generates formatted report text, data visualization charts, and descriptions of suggested remediation measures.

[0037] This system architecture supports centralized storage and version management of models, knowledge, algorithms, and rules within a unified algorithm service platform at the system level. Simultaneously, it enables independent deployment and functional iteration at the front end through lightweight intelligent interaction units. Each deployment unit reuses the reasoning and decision-making capabilities provided by the unified algorithm service platform and establishes independent data spaces at the logical level to achieve cross-node calls under data isolation conditions, forming a distributed architecture with multi-node collaborative operation.

[0038] Figure 1 This illustrates a pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration, according to Embodiment 1 of this application. Figure 1As shown, the method includes the following steps: Step S110: Obtain multi-source heterogeneous environment data through edge-side nodes.

[0039] In this application, the multi-source heterogeneous environmental data includes air quality monitoring data, meteorological data, online pollution source monitoring data, electricity consumption monitoring data, mobile surveillance data, satellite remote sensing data, video surveillance data, as well as unstructured regulatory text data and historical case document data.

[0040] Specifically, air quality monitoring data includes 5-minute, hourly, daily, monthly, and yearly air quality data for urban stations, assessment stations, benchmark cities, and target cities. Air quality data includes physical quantities such as CO, NO2, O3, PM2.5, PM10, and SO2 concentrations. Meteorological data includes meteorological observation and forecast data, including meteorological parameters such as wind speed, wind direction, temperature, humidity, and boundary layer height. Online pollution source monitoring data includes online monitoring data from the emission outlets of various air-polluting enterprises, including emission rates, emission volumes, and outlet operating conditions. Electricity consumption monitoring data reflects the intensity of enterprise production activities. Mobile monitoring data includes pollutant concentration data acquired by mobile monitoring vehicles along their routes. Satellite remote sensing data includes pollutant concentration grid data and hotspot grid data retrieved from remote sensing. Video surveillance data includes video footage of key areas and the vicinity of key enterprises. Regulatory text data includes national and local environmental protection laws, regulations, policy documents, and standards. Historical case document data includes records of historical pollution events, remediation cases, and assessment reports.

[0041] Step S120: Perform adaptive sensing and preprocessing on multi-source heterogeneous environmental data at the edge, dynamically adjust the data collection frequency and caching strategy according to the risk level of the environmental state, and upload the processed multi-source heterogeneous environmental data to the cloud server.

[0042] In this application, adaptive perception specifically refers to the edge side determining the environmental risk level based on real-time monitoring data and dynamically adjusting the data acquisition strategy accordingly. Preprocessing includes at least data cleaning and format standardization. Furthermore, this method determines the environmental risk level and its corresponding processing level based on the degree of anomaly of at least one parameter in multi-source heterogeneous environmental data.

[0043] In a specific embodiment, adaptive sensing and preprocessing of multi-source heterogeneous environmental data at the edge side are performed, specifically including: When multi-source heterogeneous environmental data meets the first preset condition, the data acquisition mode is switched from the normal acquisition mode to the abnormal acquisition mode. In the abnormal acquisition mode, the data acquisition frequency is adjusted from the first frequency to the second frequency, and / or the number of monitoring nodes per unit area in the monitoring area is adjusted from the first total number of nodes to the second total number of nodes.

[0044] The second frequency is greater than the first frequency, and the total number of the second nodes is greater than the total number of the first nodes. The first preset conditions include: the indicator value reaching a first percentage of a preset warning threshold (e.g., 80%), the meteorological diffusion index being lower than a preset diffusion threshold, or the fluctuation amplitude of the online monitoring data of the pollution source exceeding at least one of the preset fluctuation ranges. In this embodiment, the edge-side nodes determine the environmental status in real time by setting a risk perception unit. When the indicator approaches the warning threshold, diffusion conditions deteriorate, or increased fluctuations in pollution sources are detected, the system automatically switches from the normal acquisition mode to the abnormal acquisition mode to increase the data acquisition density of the relevant areas and indicators.

[0045] When the multi-source heterogeneous environmental data meets the second preset condition, the data acquisition mode is switched to the early warning perception mode. In the early warning perception mode, the data retention time of the dynamic ring buffer set on the edge side is extended from the first time to the second time. When the multi-source heterogeneous environmental data meets the third preset condition, the data acquisition mode is switched to the anomaly perception mode. In the anomaly perception mode, based on the trigger time of the anomaly perception mode, the data in the dynamic ring buffer located in the range from the third time before the benchmark to the fourth time after the benchmark is marked as non-coverable data, the non-coverable data is extracted, and a pollution event tracing data package is generated.

[0046] The second preset condition includes at least one of the following warning determination events: The indicator value reaches the second percentage of the preset warning threshold; The meteorological diffusion index is below the second preset diffusion threshold; The fluctuation range of the online monitoring data of the pollution source exceeds the second preset fluctuation range.

[0047] In addition, the third preset condition includes at least one of the following abnormal judgment events: The rate of change of the monitored data within a unit time window exceeds the first preset rate of change threshold. The monitoring data continuously exceeded the first preset high threshold for multiple consecutive sampling periods; The instantaneous value of the monitored data exceeds the first preset threshold. The temporal logical relationship or numerical ratio relationship between multiple sets of related monitoring data deviates from the preset benchmark range; The characteristic fingerprints in the monitoring data are matched with the abnormal spectra in the preset pollution source characteristic database.

[0048] Traditional environmental monitoring systems typically employ fixed-frequency, fixed-channel, and fixed-process data acquisition methods, failing to adapt data acquisition strategies to changes in pollution risk. To address this issue, this application establishes risk sensing units at the edge of the system to perform real-time assessments of air quality concentration, pollution source emissions, electricity load, meteorological diffusion conditions, high-value patches from remote sensing, mobile monitoring anomalies, and historical similar pollution events. When environmental conditions are stable, the system collects and accesses data at the normal frequency. When indicators approach preset warning thresholds, diffusion conditions deteriorate, pollution source fluctuations increase, or high-value patches appear in the remote sensing grid, the system automatically increases the data acquisition density for relevant areas, stations, pollution sources, and indicators. When sudden increases, continuous high values, exceeding thresholds, inversions, or suspected pollution source anomalies occur, the system enters anomaly detection mode, prioritizing the capture of high-temporal-resolution data before, during, and after the anomaly, providing a complete data foundation for subsequent pollution cause analysis and on-site source tracing. This adaptive sensing mechanism driven by environmental risks can transform data collection from "passive collection at equal intervals" to "risk-triggered active sensing". In stable environmental conditions, it maintains a low collection frequency to reduce redundant data transmission and storage costs, while automatically increasing the collection frequency or density when preset abnormal conditions are met, thereby improving data density and timeliness during risky periods.

[0049] This application employs a risk-level caching mechanism for pollution event precursors, addressing the technical problem of traditional systems that only trace data after the pollution process has occurred, easily resulting in the loss of precursor information. The system sets different data caching windows according to stable, early warning, and abnormal states. In a stable state, only historical data within the regular time window is retained; in an early warning state, the retention period for historical data of relevant stations, pollution sources, grid areas, and meteorological elements is automatically extended; in an abnormal state, the system completely locks key data segments within the time window before, during, and after the anomaly, and constructs a multi-dimensional heterogeneous data association index for that time window: it associates static threshold comparison records and dynamic residual exceedance records that trigger the abnormal state; it associates the probability distribution and intermediate layer activation features output by the anomaly identification model (including classifiers or time-series prediction networks); it associates the results of geographic information map analysis triggered synchronously during that period, including concentration contour maps or spatial hotspot distribution matrices; and it associates the interruption signals or alarm push protocol logs distributed externally by the system. Through the association of the above time axis slices with the logical axis, the complete pollution process evolution chain is preserved to support the digital reproduction and feature backtracking of subsequent pollution evolution processes. Therefore, by employing time window slicing and multi-level index binding techniques, the recall rate and association accuracy of abnormal data can be improved.

[0050] Furthermore, this application employs a risk-level caching mechanism based on a dynamic circular buffer. The system sets up a dynamically adjustable circular buffer (i.e., a dynamic circular buffer) at the edge side to store historical data under different risk levels. In a stable state, the circular buffer retains only short-cycle data. Upon entering an early warning state, the system automatically expands the logical length of the buffer, extending the data retention time for relevant stations, pollution sources, grids, and meteorological elements. Upon entering an abnormal state, the system automatically extends the historical data retention time and locks all data segments in the buffer before, during, and after the anomaly, forming a pollution event precursor data package. For example, when an anomaly detection mode is triggered, the system automatically extends the historical data retention time for 30 minutes before and 60 minutes after the anomaly, locking in the complete data chain of pollution event precursors and evolution. This mechanism can preserve a continuous chain of evidence for pollution process review, on-site source tracing, and responsibility determination.

[0051] This application organizes the data, rules, models, maps, knowledge base, and disposal results involved in an anomaly identification or pollution process handling into a traceable pollution event chain. This pollution event chain includes raw monitoring data, edge preprocessing results, anomaly triggering rules, dynamic threshold parameters, model judgment results, spatial maps, remote sensing patches, pollution source association results, knowledge base references, group push records, dispatch objects, on-site verification feedback, rectification suggestions, and final verification results. Through a pollution event chain construction mechanism oriented towards on-site source tracing, this application establishes a Directed Acyclic Graph (DAG) lineage index between the output results after anomaly triggering and the raw monitoring data, judgment rules, inference models, and historical cases. This achieves lossless backtracking from each output item to the input, overcoming the technical deficiency of existing technologies that lack traceability. It achieves the technical effects of traceable decision paths, offline reproducibility of evidence chains, and simulation verification of disposal processes.

[0052] Missing value detection is performed on structured data in multi-source heterogeneous environmental data, marking null values ​​in data fields as missing values. Outlier detection is also performed on the structured data, calculating the deviation of each data item from the mean, and marking data items with deviations exceeding a preset deviation threshold as candidate outliers. Duplicate value removal, timestamp verification, site code matching, spatial location matching, and pollutant indicator caliber conversion are also performed on the structured data. Simultaneously, unstructured data in the multi-source heterogeneous environmental data undergoes text parsing, chapter segmentation, tag extraction, and vectorization. Preprocessed relevant outlier and uncoverable data are encapsulated into high-priority data packets and sent to the cloud server via a communication interface.

[0053] The relevant abnormal data includes at least one of the following: amplitude domain outlier characteristics, temporal gradient characteristics, threshold proximity values, pollution source anomalies, or spatial hotspot data. Amplitude domain outlier characteristics refer to detected values ​​that are significantly higher in absolute value than normal state data values ​​or historical averages. Temporal gradient characteristics refer to monitored values ​​that experience sharp increases or decreases within a very short time interval. Threshold proximity values ​​refer to values ​​that have not yet exceeded the set warning line but have exceeded the preset warning line. Pollution source anomalies refer to monitoring data from a specific emission outlet or process stage that contradicts the material balance and production process logic of that source. Spatial hotspot data refer to data from a specific geographical area that is significantly higher than data from its surrounding areas.

[0054] In this embodiment, during edge-side intelligent preprocessing, structured data undergoes missing value identification, outlier screening, duplicate value removal, timestamp verification, site code matching, spatial location matching, and pollutant indicator caliber conversion. Unstructured data undergoes text parsing, chapter segmentation, tag extraction, and vectorization. Relevant abnormal and uncoverable data are prioritized for uploading to the cloud. Further, the text parsing of unstructured data specifically includes: word segmentation to obtain word sequences; matching the word sequences with a pre-defined environmental monitoring domain feature word library, extracting successfully matched words as entity feature words; and identifying the numerical values ​​and their associated units in the word sequences, extracting them as numerical feature parameters.

[0055] This application employs an edge-side intelligent preprocessing mechanism, setting up a preprocessing unit at the edge to perform preliminary cleaning, noise reduction, alignment, filtering, and labeling of raw data. Specifically, for structured data, the system performs missing value identification, outlier screening, duplicate value removal, timestamp verification, site code matching, spatial location matching, and pollutant indicator caliber conversion. For unstructured documents, it performs parsing, segmentation, labeling, and vectorization. Furthermore, for video, mobile monitoring, and remote sensing data, the system can extract anomalous segments, hotspot grids, spatial patches, trajectory changes, and key feature information. For stable, low-risk, and highly repetitive data, the edge side only uploads summary statistics, feature vectors, or standardized indicators; for amplitude domain outliers, temporal gradient features, near-threshold values, pollution source anomalies, or spatial hotspot data, the system marks them as high-value feature data and prioritizes uploading them to the cloud along with non-coverable data. This mechanism reduces the repeated transmission of low-value raw data, lowers network bandwidth and cloud computing power pressure, and improves the efficiency of anomalous data capture.

[0056] In addition, each edge node is equipped with a data priority quantization engine. Each edge node obtains the comprehensive upload weight of the current data packet through its built-in data priority quantization engine, and each edge node then calculates the comprehensive upload weight. Compared with the preset first priority threshold Second priority threshold Based on the comparison results, hierarchical transmission scheduling is performed on the current data packet: when When the current data packet is marked as a high-priority data packet, a real-time interruption of the upload is triggered, and the transmission delay does not exceed the first preset duration; when When the current data packet is marked as a medium-priority data packet, it is stored in the local cache queue and uploaded using a time window aggregation method. The aggregation window width is the second preset duration, and the second preset duration is greater than the first preset duration. When this happens, the current data packet is marked as a low-priority data packet, stored in local memory, and then uploaded in batches during preset off-peak communication periods; among which, Overall upload weight The calculation formula is as follows: ; In the above formula, This is a normalized risk coefficient based on the current monitored value relative to the predicted statutory limit; This is a priority coding value for sensor data types, used to distinguish between routine monitoring parameters and characteristic pollutants. The coding value of characteristic pollutants is greater than that of routine monitoring parameters. For example, the coding values ​​of routine monitoring parameters such as temperature, humidity, atmospheric pressure, wind speed, and wind direction are set in the first coding interval, while the coding values ​​of characteristic pollutants such as PM2.5, PM10, SO2, NO2, and CO are set in the second coding interval. The lower limit of the second coding interval is greater than the upper limit of the first coding interval. The quality score is based on signal integrity verification, which includes packet cyclic redundancy check and timestamp continuity check. This is the influence radius coefficient of the current edge node in the spatial topology, which is constructed based on the geographic coordinates of each edge node through Deloni triangulation or inverse distance weighting. It is a monotonically decreasing function representing the remaining time from the preset target termination time window at the current time; These are dynamic configuration coefficients input through an external configuration interface, used to characterize the level of attention that users or management policies place on the current data packet; The current real-time load rate in the cloud is defined as the percentage of currently used computing resources to total computing resources. a, b, c, d, e, g, h These are the preset weight coefficients for each factor, and their sum is 1.

[0057] When the cloud server receives anomaly confirmation signals triggered simultaneously by multiple edge nodes, it executes a multi-node concurrent resolution step: It constructs a spatiotemporal coupling sorting factor for each anomaly event, sorts them in descending order according to the value of the spatiotemporal coupling sorting factor, generates an anomaly event processing priority sequence list, and dynamically determines the number K of events to be processed in a single batch based on the current real-time load rate of the cloud server. The top K anomaly data packets in the anomaly event processing priority sequence list are allocated to the cloud server's reserved computing power resource pool, triggering a fusion analysis process. Simultaneously, the top K+1 anomaly data packets and subsequent packets in the anomaly event processing priority sequence list are stored in a waiting queue, and a cyclic monitoring process is started. The current real-time load rate of the cloud server is re-acquired at preset time intervals, and the aforementioned steps are repeated until all anomaly data packets have been allocated to the reserved computing power resource pool. The spatiotemporal coupling sorting factor... The calculation formula is: ; In the above formula, For the first Monitoring values ​​uploaded by each edge node This indicates the normalization overshoot factor of the node; For the first The edge side node and the first The Euclidean spatial distance between each edge node in the Deloni triangulation. Indicates spatial proximity and density; The upwind source probability value is calculated based on real-time wind field vector data, which includes wind direction and wind speed. For the first The remaining time between the current time of each edge node and the preset target termination time window. This indicates the time urgency of the target task; , , , These are the preset weight coefficients for each sub-item, and their sum is 1.

[0058] In some specific implementation processes, the formula for calculating the number of events K processed in a single batch is as follows: ; in, This represents the maximum concurrent processing capacity threshold for a single batch on a cloud server. This represents the current real-time load rate of the cloud server. This is the preset lower limit threshold for the idle load rate of cloud servers; This is the preset maximum threshold for cloud server load rate; when hour, ;when hour, .

[0059] Quality score based on signal integrity verification The calculation formula is as follows: ; in, This is the result value of the cyclic redundancy check (CRC) of the data packet. When the check passes, When the verification fails, ; This is the timestamp continuity verification result value. When the time interval between adjacent data packets is less than the preset maximum allowable interval, ,otherwise, ; , These are the preset weighting coefficients corresponding to the verification result values.

[0060] The influence radius coefficient of the current edge node in the spatial topology The calculation method is as follows: When the spatial topology is constructed based on Deloni triangulation ,in, For the first The area of ​​Voronoi cells occupied by each edge node in the Dironoi triangulation. The sum of the areas of all Voronoi cells at the edge nodes; When the spatial topology is constructed based on the inverse distance weighting method ,in, For the first The edge side node and its first Spatial distance between edge nodes For the first The edge side node and its first Spatial distance between edge nodes The distance is a preset power parameter. This is the preset number of neighboring nodes. This represents the total number of all edge-side nodes.

[0061] A monotonically decreasing function representing the remaining time until the preset target termination time window at the current moment. The calculation formula is as follows: ; in, This represents the remaining time until the preset target termination time; This is a preset positive offset to prevent division by zero errors; if the current time has exceeded the preset target termination time, The value is negative. The value of satisfies ,in, for The minimum value that can be achieved throughout the entire operating cycle, to ensure Heng was established.

[0062] in addition, a, b, c, d, e, g, h as well as , , , It is stored in a read-write register on the edge node or cloud server and can be overwritten via an external configuration interface.

[0063] This application automatically generates a data upload priority queue based on data risk level, data type, data quality, spatial impact range, assessment target pressure, user attention level, and current cloud load when multiple sites, pollution sources, grids, or edge nodes operate simultaneously. Data related to exceeding standards alarms, high values ​​at sites, pollution source anomalies, meteorological transmission risks, remote sensing hotspots, and mobile monitoring anomaly clues are uploaded first; routine low-risk data is uploaded with low priority, in batches, or with delays. When multiple nodes trigger anomalies simultaneously, the system sorts them according to the degree of pollution impact, spatial proximity, upwind correlation, and the urgency of the target task. This ensures that critical anomaly data is prioritized for cloud-based fusion analysis, avoiding network congestion or delays caused by simultaneous uploads from multiple nodes. This achieves intelligent transmission scheduling of "critical data priority, low-risk data postponement, and cloud resource collaboration."

[0064] Furthermore, to enhance real-time processing capabilities at the edge, this application can deploy lightweight neural network models or lightweight anomaly recognition models on edge nodes, processed through channel pruning, parameter quantization, knowledge distillation, or low-rank compression, to locally perform data denoising, initial anomaly screening, feature extraction, dimensionality compression, and risk labeling. Instead of directly uploading all raw data at the edge, the system prioritizes uploading anomaly features, statistical summaries, key segments, spatial hotspots, and risk labels. Through this lightweight model pre-processing, the system can reduce transmission bandwidth usage and cloud computing pressure without significantly increasing edge computing power burden, and improve the efficiency of anomaly data entering the cloud for analysis.

[0065] Step S130: Based on the constructed environmental professional rule base, environmental professional knowledge base and data foundation in the cloud, and using retrieval enhancement generation technology combined with a large language model, the processed multi-source heterogeneous environmental data is fused and analyzed to generate environmental situation analysis results and disposal strategies.

[0066] In this application, the environmental professional rule base includes assessment rules, threshold rules, alarm rules, compliance margin rules, push rules, and dispatch trigger conditions. The environmental professional knowledge base includes an environmental general knowledge base, a legal, standard, policy and solution base, a technical method base, a multi-scenario pollution problem and measure base, a multi-scenario control solution / PPT base, a multi-scenario governance effect evaluation base, a multi-scenario multi-city governance practice and case base, and an industry dynamic research base.

[0067] This application establishes a data foundation and a knowledge foundation. The data foundation includes structured environmental data and spatial data, while the knowledge foundation contains multiple knowledge and rule bases, such as regulations and standards, technical methods, measures and solutions, effect evaluations, and practical cases. The system processes unstructured knowledge documents through hierarchical structure parsing, semantic paragraph segmentation, multi-dimensional tag annotation, and dense vector embedding mapping. It also constructs a triple indexing system consisting of a full-text inverted index, a semantic vector index, and a scene tag index, enabling the intelligent agent to accurately retrieve, call, and trace the source of knowledge. In this embodiment, the data foundation and knowledge foundation collaborate via a unified data bus, allowing the intelligent agent's inference engine to simultaneously call real-time monitoring data and historical knowledge text, generate a structured processing instruction set, and distribute it through a communication interface. This achieves cross-domain association, joint inference, and automated output of historical monitoring data and knowledge assets.

[0068] In a specific embodiment, the retrieval enhancement generation technique combined with a large language model is used to fuse and analyze the processed multi-source heterogeneous environmental data, specifically including:

[0069] Step S131: Calculate the dynamic reliability weight of each data source based on the preset reliability parameters of each data source, the time deviation between the data collection timestamp and the reference time base, and the spatial proximity between the corresponding monitoring points of each data source. Then, remove the data from the data source with a dynamic reliability weight lower than the first preset value in the processed multi-source heterogeneous environment data to achieve cross-validation of the processed multi-source heterogeneous environment data.

[0070] In this embodiment, the method constructs a multi-source heterogeneous data cross-validation mechanism to avoid misjudgments caused by single sensor anomalies, single site fluctuations, or isolated data changes, thereby improving the accuracy of judgment and the reliability of on-site command in complex pollution scenarios. Before generating high-value alarms, compliance pressures, pollution source anomalies, or on-site source tracing suggestions, the system correlates and compares air quality monitoring data, online pollution source data, electricity consumption data, meteorological wind fields, remote sensing grids, mobile monitoring, video surveillance, historical events, and changes at surrounding sites. Dynamic credibility weights are assigned to different data sources based on data source reliability, time synchronization, spatial proximity, historical consistency, duration of anomalies, and multi-source consistency. When multiple data sources point to the same anomaly area or the same pollution source type, the warning level and source tracing confidence are increased; when only a single data source is abnormal and lacks surrounding support, it is marked as a clue to be verified or a suspected equipment anomaly event.

[0071] Specifically, this application sets up a multi-source heterogeneous data fusion and analysis engine. This engine maps input data from air quality monitoring stations, online pollution source monitoring equipment, electricity load monitoring modules, meteorological observation stations, satellite remote sensing grids, mobile monitoring equipment, video surveillance equipment, and historical case databases to a unified high-dimensional feature space through a preset feature encoder, and calculates the multi-source fusion confidence of abnormal events. The calculation formula is as follows: ; in, For the first Anomaly confidence scores from each data source output, ; For the first The dynamic fusion weights of each data source are set, and the sum of all weights is 1. Simultaneously, this dynamic fusion weight... The data reliability factor is determined by multiplying the data credibility factor, spatial proximity factor, temporal synchronization factor, anomaly persistence factor, historical similarity factor, transmission direction consistency factor, and preset configuration coefficient. The data credibility factor is calculated based on the historical false alarm rate of the data source and the current signal integrity verification value. The spatial proximity factor is the reciprocal normalized Euclidean distance between the data source's acquisition point and the anomaly event's location coordinates. The temporal synchronization factor is the reciprocal normalized time difference between the data source's timestamp and the anomaly event's trigger time. The anomaly persistence factor is the ratio of the data source's continuous exceeding duration to a preset duration threshold. The historical similarity factor is the reciprocal normalized Mahalanobis distance between the data source's current value and the historical average for the same period. The transmission direction consistency factor is determined based on the correlation coefficient or gradient direction consistency between the data source and the upwind monitoring point. The preset configuration coefficient is a scenario adaptation weighted value input through an external interface.

[0072] Multi-source fusion confidence When the confidence level exceeds a preset first confidence threshold, an anomaly confirmation signal is output; multi-source fusion confidence level When the confidence level is below the first confidence threshold but above the preset second confidence threshold, it is marked as pending review; multi-source fusion confidence level When the value is below the second confidence threshold, it is identified as isolated noise and discarded, thereby reducing false triggering caused by instantaneous fluctuations of a single sensor.

[0073] Step S132: Compare the indicator values ​​in the verified data with the preset indicator thresholds. If any indicator value exceeds the preset indicator threshold, trigger in-depth analysis and input the verified data into the spatiotemporal prediction model, pollution source tracing model, and knowledge graph to obtain pollution diffusion trend data and candidate pollution source data.

[0074] In this embodiment, the cloud server employs a two-level anomaly assessment mechanism. It performs basic judgment through a rule engine, followed by a secondary comprehensive assessment using a spatiotemporal prediction model, a pollution source tracing model, and a knowledge graph. Alternatively, basic rule judgment, initial anomaly screening, data quality verification, and risk labeling are performed at the edge to achieve rapid response. Then, in the cloud, a secondary comprehensive assessment of suspected anomalies is conducted based on an environmental professional rule base, a spatiotemporal prediction model, a knowledge graph, a pollution source tracing model, a RAG knowledge base, and a large language model. The method in this embodiment adopts a two-level anomaly judgment mechanism that combines cloud and edge collaboration. Based on a two-level architecture of rapid initial screening at the edge and in-depth judgment in the cloud, for high values ​​or anomalies identified at the edge, the cloud further integrates historical data from the same period, meteorological wind fields, diffusion conditions, pollution source lists, online emission data, electricity consumption changes, remote sensing grids, mobile data, and historical cases to determine whether the anomaly is due to equipment malfunction, local pollution, regional transmission, fluctuations in pollution source emissions, cumulative meteorological impact, or a comprehensive pollution process. Based on this, early warning information, source tracing suggestions, mobile routes, inspection lists, dispatch suggestions, or report materials are generated, realizing the collaborative advantage of "fast at the edge and accurate at the cloud".

[0075] Step S133: Based on the feature vectors of the candidate pollution source data, retrieve the corresponding regulatory text data and similar historical case document data from the environmental professional knowledge base, combine the retrieval results with the pollution diffusion trend data to form an enhanced prompt context, and input the enhanced prompt context into the large language model to output the judgment conclusion text containing the pollution event description and disposal suggestions.

[0076] This application establishes at least two levels of response strategies for exceeding or anomalies. For ordinary high values, slight exceedances of thresholds, or single-point short-term anomalies, the edge system can directly trigger local early warnings, data locking, and preliminary alerts. For continuous high values, severe exceedances, multi-source consistent anomalies, significantly increased risk of compliance issues, or suspected pollution source impact events, the system simultaneously links with the cloud server to initiate an in-depth analysis process, generating pollution cause analysis, priority investigation areas, navigation routes, inspection lists, recommendations for responsible parties, and disposal plans. This forms a layered processing model of rapid edge response, in-depth cloud analysis, and precise on-site disposal, improving the system's response efficiency and accuracy for events of varying severity.

[0077] In addition to generating environmental situation analysis results and response strategies, this method further includes generating a precise source tracing and collaborative scheduling scheme. Specifically, this involves generating mobile task planning data containing at least one path coordinate sequence based on the pollution source location information in the environmental situation analysis results and meteorological wind field vector data. Specifically, based on the pollutant concentration values ​​at each pollution source location and a preset sorting rule, the investigation priority weight value of each pollution source is calculated, and pollution source investigation sequence data is generated in descending order of investigation priority weight values. Based on the mobile task planning data and preset spatial grid ownership data, the jurisdiction identifier and timeliness identifier corresponding to each grid are determined. The event markers, time markers, pollution source identifiers, and mobile task planning data from the environmental situation analysis results are encapsulated into a structured task data package according to a preset data structure template, and the structured task data package is sent to the target terminal device corresponding to the jurisdiction identifier through a communication interface. In this embodiment, the structured task data package includes a data field and an identifier field. The data field carries the pollution source identification results and the mobile path coordinate sequence, and the identifier field carries the timeliness identifier, which is the response deadline timestamp or remaining valid duration value of the structured task data package. Furthermore, after receiving the structured task data packet, the edge node parses the identifier field in the data packet to obtain the expiration flag. Until the current system time reaches the deadline indicated by the expiration flag, the edge node maintains the structured task data packet in a pending state. If the current system time exceeds the expiration flag, the structured task data packet is marked as an expired data packet, and a retransmission request is generated and sent to the cloud.

[0078] In addition, in the cloud, this method also includes a scene-adaptive dynamic threshold determination step, specifically including: The system obtains the scene type identifier of the current environmental monitoring scenario and, based on the scene type identifier, retrieves the corresponding benchmark index threshold from a pre-defined benchmark threshold library. Scene type identifiers include static and stable accumulation scenarios, dust transport scenarios, and major event support scenarios.

[0079] Acquire current real-time meteorological data and calculate the correction coefficients for each meteorological factor to the benchmark threshold based on this data. The current real-time meteorological data includes meteorological condition parameters, boundary layer height, wind speed, direction, and humidity data, historical data for the same period, station background concentration data, pollution source operational status data, regional transport characteristic data, time-specific dynamic warning lines, and expected compliance difficulty coefficients. More specifically, a first correction coefficient is determined based on the ratio of the boundary layer height to the preset standard boundary layer height; a second correction coefficient is determined based on the difference between the wind speed value in the wind speed and direction data and the preset standard wind speed value; a third correction coefficient is determined based on the difference between the humidity value in the humidity data and the preset standard humidity value; and a fourth correction coefficient is determined based on the upwind pollutant concentration contribution rate in the regional transport characteristic data. The product of the first, second, third, and fourth correction coefficients is used as the final correction coefficient.

[0080] Based on the baseline threshold and correction coefficient, a preset threshold suitable for the current environmental monitoring scenario is calculated. This preset threshold is then used to determine anomalies in pollutant concentration monitoring values ​​from processed multi-source heterogeneous environmental data, resulting in anomaly marking. More specifically, when a pollutant concentration monitoring value exceeds a dynamic judgment threshold, the current monitoring time is marked as an anomaly. Specifically, in a static and stable accumulation scenario, the alarm sensitivity of the preset threshold is higher than the baseline threshold, meaning the dynamic judgment threshold is lower than the baseline threshold. In a dust transport scenario, the alarm sensitivity of the preset threshold is lower than the baseline threshold, meaning the dynamic judgment threshold is higher than the baseline threshold.

[0081] Traditional environmental regulation mostly relies on fixed thresholds or single standard rules, which are difficult to adapt to complex meteorological conditions, different regional background values, different site attributes, and different management objectives. This application introduces a scenario-adaptive dynamic threshold mechanism based on a basic standard rule base. The system adopts a heterogeneous benchmark coupling architecture. Its judgment logic's benchmark (i.e., benchmark threshold) is derived from a static benchmark layer, a dynamic benchmark layer, and a variable strategy layer. The static benchmark layer contains at least one set of national mandatory emission limits and corresponding local emission assessment characteristic values. This layer is used to set the system's absolute physical dead zone. When the monitored value reaches the limit of this layer, the system unconditionally triggers a first-level interruption alarm signal, which has higher priority than the output results of all subsequent layers. The dynamic benchmark layer is used to calculate the expected noise baseline value corresponding to the current moment based on monitoring data within the historical same-period window, using a median absolute deviation or local weighted regression algorithm. This layer is not used to determine exceedances, but only to determine whether the residual (measured value minus benchmark threshold) exceeds a preset multiple of the benchmark's standard deviation. If it does, it is marked as a "statistical anomaly." The variable strategy layer receives user-defined time window weighting factors and local area scaling coefficients. The system temporarily adjusts the benchmark value of the static benchmark layer based on the received external input. It should be noted that this adjustment only changes the system's sensitivity trigger slope and does not change the original stored value of the static benchmark layer. Furthermore, by combining real-time meteorological conditions, boundary layer height, wind speed and direction, humidity, historical levels for the same period, station background concentration, pollution source operating status, regional transmission characteristics, key period requirements, and target completion pressures, dynamic thresholds and early warning sensitivities adapted to the current scenario are generated. Additionally, for scenarios such as static accumulation, dust transport, ozone peak seasons, autumn and winter particulate matter processes, major event support, and key area control, the system can automatically load corresponding rule packages. For different project sites, a localized rule threshold library can be formed based on historical data and local needs. This application adopts a scenario-adaptive dynamic threshold and rule package generation mechanism, solving the problems of false alarms, missed alarms, and delayed responses in complex atmospheric governance scenarios with fixed thresholds. Simultaneously, by setting physical isolation between the static layer and the variable strategy layer, the system solves the technical problem in existing technologies where adjusting emergency control limits necessitates rewriting underlying firmware or database fields. Moreover, sensitivity can be steplessly adjusted using only external weighting coefficients. While ensuring the absolute immutability of legal limits, the system achieves dynamic adaptation to special periods, significantly improving its robustness and generalization capabilities.

[0082] Step S140: Based on the environmental situation analysis results and the handling strategy, generate a target data packet, send the target data packet to the preset target terminal device through the wireless communication interface, and receive the on-site feedback signal returned by the target terminal device.

[0083] In this application, the target data package includes at least one of environmental situation analysis data, early warning data, pollution source location data, or disposal instruction data. The environmental situation analysis data includes real-time reports, cumulative reports, comparison reports, and pollution source high value statistics. The early warning data includes today's remaining amount reminders, annual target achievement reminders, site high value alarms, indicator high value alarms, and equipment anomaly reminders.

[0084] Based on the environmental situation analysis results and response strategies, this embodiment pushes customized early warning information, source tracing suggestions or dispatch instructions to relevant personnel through the terminal interaction interface, and receives on-site feedback information.

[0085] Specifically, this method includes: automatically calculating the statistical sequence within the current monitoring period according to a preset limit reference system; determining high value, exceeding threshold, inversion, equipment malfunction, or risk of exceeding standards or pollution based on an environmental professional rule base; and forwarding the current task to at least one of the following processing modules according to a preset routing mapping table based on the label of the determination result and the request type encoding sent by the terminal device, in order to achieve functional scenario distribution.

[0086] The statistical sequence includes: real-time value, cumulative value, comparison value, ranking value, remaining capacity, time-series progress deviation, and high-value statistics. The real-time value is the instantaneous reading of the monitored value at the current moment; the cumulative value is the arithmetic mean or integral value of the monitored values ​​within the current monitoring period; the comparison value is the difference between the current monitored value and a preset reference benchmark value (including historical averages or measurements from parallel stations); the ranking value is the ascending or descending position of the current monitored value within the set of monitoring stations; the remaining capacity is the difference between a preset limit and the current cumulative value; the time-series progress deviation is the ratio of the remaining capacity to the remaining time length from the current moment to the end of the monitoring period; and the high-value statistics are the frequency and peak value of each pollution source exceeding a first preset threshold within the current monitoring period.

[0087] The air quality information formatting module and compliance status monitoring module can generate structured data reports such as real-time data reports, cumulative statistical reports, comparison difference reports, remaining capacity reports, time series progress deviation reports, and site anomaly reports based on the judgment results. This application's system can also call upon meteorological numerical forecast outputs, air quality numerical model prediction results, pollution source emission inventories, remote sensing image inversion data, and geographic information system spatial data to automatically generate air quality prediction maps, meteorological anomaly maps, concentration spatial distribution maps, transmission path maps, remote sensing inversion maps, source tracing direction maps, time series trend maps, or multi-source comparison maps.

[0088] Furthermore, the system in this application can also utilize the multimodal indexes of the knowledge base (including full-text inverted indexes, semantic vector indexes, and scene tag indexes) to perform query encoding, semantic retrieval, full-text matching retrieval, tag filtering retrieval, multi-path fusion sorting, and query output generation processing on query statements input through user command interfaces via terminal devices, thereby achieving intelligent question answering for the knowledge base. Simultaneously, the system can integrate the aforementioned statistical sequences, judgment result tags, retrieval knowledge text, and structured instruction sets to output a structured analysis report containing anomaly attribution tag sequences, recommended execution action sequences, data statistical summaries, spatial distribution summaries, and trend prediction summaries, enabling intelligent parsing of real-time environmental data and full-scene information.

[0089] Furthermore, this application includes a structured processing instruction generation module, which generates instructions based on multi-source fusion confidence levels. When the first confidence threshold is exceeded, the module automatically generates a structured disposal instruction set based on the pollution type label of the abnormal event, spatial coordinate location, pollution source contribution ranking results, preset administrative jurisdiction boundary vector layer, and preset disposal time limit mapping table. This structured disposal instruction set includes a unique event identifier code, anomaly timestamp, anomaly indicator vector, spatial impact range, list of suspected pollution source identifiers, evidence map storage address, recommended mobile route coordinate sequence, patrol point coordinate list, pre-configured responsible entity identifier list, standardized disposal operation code sequence, and feedback time window. Among them, the unique event identifier is generated based on the trigger time and spatial coordinates according to preset coding rules; the abnormal timestamp is the trigger time and the start and end time window of the abnormality; the abnormal indicator vector is the type of pollutant exceeding the standard and the multiple of exceeding the standard; the spatial impact range is the concentration isopleth boundary coordinate sequence calculated based on spatial interpolation; the list of suspected pollution source identifiers is the pollution source IDs arranged in descending order of contribution; the evidence map storage address is the storage path of the spatial distribution map and wind field vector map; the recommended mobile route coordinate sequence is generated by the path planning algorithm; the inspection point coordinate list is the spatial point sequence of priority investigation; the pre-configured responsible entity identifier list is the department code extracted based on the spatial matching of administrative jurisdiction boundaries; the standardized disposal operation code sequence is converted from the disposal measures matched in the knowledge base; and the feedback time window is the preset time limit for the return of verification results.

[0090] After the generated structured processing instruction set is distributed to the target terminal device corresponding to the pre-configured responsible entity identifier via the external communication interface, the feedback write-back module starts the timing monitoring process. If no verification result data packet is received within the feedback time window, a timeout alarm is triggered. If a verification result data packet is received within the feedback time window, format verification and content parsing are performed, the parsing result is written to the event chain storage area, and the multi-source fusion confidence of the multi-source heterogeneous data fusion analysis engine is triggered. Iterative updates.

[0091] This application also includes a multi-dimensional data security access control engine. This engine constructs a permission determination matrix composed of data variable type dimension, time dimension, node level dimension, cluster role dimension, sensitivity level dimension, and rule entry dimension. Specifically, the data variable type dimension represents the monitoring parameter type code of the currently accessed object; the time dimension represents the timestamp and time window of the data involved in the current access request; the node level dimension represents the edge node level code of the source of the current access request; the cluster role dimension represents the cluster role identifier code of the initiator of the current access request, including system administrator, data analyst, operations operator, read-only auditor, etc.; the sensitivity level dimension represents the preset sensitivity classification code of the currently requested data or model output, from level 1 to 5, with higher levels indicating greater sensitivity; and the rule entry dimension represents a preset permission matrix table based on scenario and user identity, with rule entries in increments of 2. 6 The bitmask is stored in the system configuration register.

[0092] The multi-dimensional data security access control engine's decision logic is as follows: Access is authorized only if the six-dimensional attribute combination of the current access request matches the rule entry mask; otherwise, it is rejected and an audit log is recorded. Furthermore, for access objects with sensitivity levels exceeding level three (including real-time enterprise emission data, raw video surveillance streams, personal information, and law enforcement inspection records), the engine can also perform de-identification, encryption, access auditing, and least privilege authorization according to sensitivity level, ensuring compliant data use during cross-regional, cross-departmental, and cross-node collaboration processes.

[0093] Furthermore, this application establishes a full-link traceability record for each anomaly confirmation signal output event to achieve full-link traceability and reliable evidence storage. This full-link traceability record includes a copy of the original monitoring data packet and data quality identifier at the trigger time, preprocessing results and cached fragments of each edge node, the version number and configuration parameters of the static threshold for triggering the exceedance judgment and the dynamic rule package, the model version number of the anomaly identification model, hyperparameter configuration and output probability distribution matrix, unique identifiers and similarity scores of knowledge text blocks retrieved from the knowledge base, the output structured handling instruction set and associated spatial distribution map, task records and distribution timestamps distributed through external communication interfaces, verification result data packets received through the feedback write-back interface, and the final verification conclusion and event status update record. Simultaneously, this application uses hash storage, trusted logs, timestamp signatures, or blockchain storage methods to save key node records and can automatically generate standardized handling work orders based on the event chain. This achieves the establishment of a fully verifiable evidence chain from original input to final output and on-site feedback.

[0094] Step S150: Based on the on-site feedback signals, incrementally update the environmental professional rule base, environmental professional knowledge base, and model parameters in the cloud, and feed the updated rule package and / or model parameters back to the edge node to achieve closed-loop optimization of the system.

[0095] In a specific embodiment, this step specifically includes: acquiring the original monitoring data, edge preprocessing result data, anomaly triggering rule identifiers, model judgment result data, push record data, and on-site verification feedback data corresponding to the current pollution event. According to preset association rules, establish temporal and causal relationships among the original monitoring data, edge preprocessing result data, anomaly triggering rule identifiers, model judgment result data, push record data, and on-site verification feedback data to generate a pollution event-related data record set. When the cumulative number of anomaly records belonging to the same pollution type or the same monitoring station in the pollution event-related data record set reaches a preset threshold, generate rule update data or model parameter increment data based on the verified data in the pollution event-related data record set. Distribute the rule update data and / or model parameter increment data to the edge nodes through a distributed model aggregation method, so that the edge nodes update the locally stored judgment rule data and / or model parameter data. More specifically, acquire the local model parameter update volume uploaded by each edge node, perform secure aggregation calculation on the local model parameter update volume of each edge node to generate a global model parameter update volume, and distribute the global model parameter update volume to each edge node. The edge-side nodes adjust the judgment parameters in the data acquisition mode switching conditions based on the updated judgment rule data and / or model parameter data, so that the edge-side nodes can trigger the abnormal acquisition mode with higher sensitivity in subsequent scenarios of the same type.

[0096] This application employs a cloud-edge collaborative bidirectional iterative and update mechanism. The edge side is responsible for quickly identifying suspected anomalies and transmitting feature data back, while the cloud server performs secondary verification and comprehensive analysis based on higher-precision models, knowledge graphs, historical case libraries, and RAG retrieval enhancement capabilities. For anomaly samples, rule parameters, feature weights, and threshold correction results verified as valid in the cloud, the system can generate updated feature rule packages or lightweight model update packages, and distribute them to relevant edge nodes through encrypted distribution, secure aggregation, or federated learning. This allows for incremental optimization of the edge model and rule base without directly exposing the original sensitive data.

[0097] In the implementation process, the cloud server receives on-site feedback signals from the target terminal device and parses these signals to obtain feedback data. This feedback data is compared with existing data in the cloud-stored environmental professional rule base and environmental professional knowledge base to determine the target data to be updated and the update type. Based on the update type, an incremental update package is generated, which is used to update at least one of the environmental professional rule base, environmental professional knowledge base, and model parameters. Furthermore, the incremental update package is distributed to edge nodes to enable them to update the locally stored rule data and / or model parameters.

[0098] In Example 1, edge nodes continuously collect local sensor data and perform real-time preprocessing. When the statistical characteristics (such as mean or rate of change) of a pollutant concentration within the sliding window exceed the dynamic threshold, the edge inference engine generates a preliminary warning signal within milliseconds. This signal is pushed to terminal devices for on-site personnel to refer to, and simultaneously, the warning event and raw data are packaged and uploaded to the cloud. After receiving the data uploaded by the edge nodes, the cloud triggers a global inference process, calls a spatiotemporal prediction model to assess the evolution trend of the pollution situation, and calls a pollution source tracing model to calculate the contribution ratio of local and external transmissions, thus arriving at a comprehensive judgment. The cloud then pushes the judgment, suggested navigation routes (including route coordinate sequences), and source tracing results to the corresponding target terminal devices. The terminal devices receive the above information via government WeChat / enterprise WeChat messages, mobile application push notifications, or business system pop-ups, and display it in text, chart, or map layer overlay formats. Based on this, law enforcement officers or grid workers conduct verification and handling. The handling results (such as on-site evidence photos and descriptions of handling measures) are transmitted back to the cloud through terminal devices for incremental training of subsequent models and knowledge base updates, forming a complete technical closed loop of "perception-reasoning-verification-update".

[0099] During operation, this application continuously receives user feedback, on-site verification results, manual verification results, and report adoption results. The cloud server updates rule parameters, abnormal samples, pollution cases, knowledge entries, and model prompt templates. When a certain type of pollution process, a certain type of site anomaly, or a certain type of pollution source behavior is verified multiple times, the cloud can generate updated feature rule packages, threshold correction rules, abnormal sample labels, or knowledge fragments, and distribute them to the corresponding edge nodes, enabling the edge to identify and respond faster in subsequent similar scenarios. Thus, the system forms a continuous evolution mechanism of "edge discovery – cloud learning – rule feedback – edge enhancement".

[0100] In addition, after anomaly identification is triggered, this application automatically performs pollution type classification, spatial impact range definition, responsible entity identifier matching, and emergency level classification. Based on spatial distribution data, meteorological wind field data, historical cases, and knowledge base, it generates a structured instruction set including the coordinates of the investigation area, the navigation route, the inspection list, the responsible entity identifier, and the disposal operation code. After being distributed through the communication interface, it receives the feedback results and updates the event chain, knowledge base, and rule base simultaneously, forming a closed-loop control loop from anomaly input to disposal output and then to rule iteration.

[0101] This application presents a pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration. This method organically couples the high-frequency sensing capabilities of the edge, the deep computing capabilities of the cloud, the professional constraints of the knowledge base, the intent understanding and generation capabilities of the intelligent agent, and the closed-loop execution capabilities of on-site scheduling. It solves industry technical bottlenecks in traditional systems, such as passive data acquisition, lack of anomaly precursors, rigid fixed thresholds, false alarms from single-source data, lack of direction in on-site source tracing, and difficulty in closing the loop for handling feedback. This application improves the timeliness of anomaly detection, reduces invalid data transmission and cloud computing burden, enhances the accuracy of exceeding standards determination, strengthens the interpretability of pollution source tracing, preserves a complete chain of evidence for pollution events, and precisely guides limited on-site resources to key areas, key time periods, key objects, and key source types, significantly improving the intelligence, precision, and closed-loop level of ecological and environmental management.

[0102] Furthermore, this method also includes a one-click recall and anomaly re-collection mechanism, specifically including: When the environmental situation analysis results generated in the cloud meet the preset supplementary data collection trigger conditions, a recall command is generated and sent to the designated edge node. The preset supplementary data collection trigger conditions include at least one of the following: the concentration change rate of the current monitoring data within a preset time window exceeds a first preset change rate threshold; the difference between the current monitoring data and the monitoring data at the previous moment exceeds a first preset difference threshold; or the pollutant concentration value in the remote sensing inversion data exceeds a preset high value threshold. The recall command includes a target node identifier, a supplementary data collection time range identifier, and a supplementary data collection data type identifier. The supplementary data collection time range identifier indicates the time window for the edge node to extract historical monitoring data, and the supplementary data collection data type identifier indicates the type of associated video data and / or mobile trajectory data that the edge node acquires.

[0103] In response to the call command, the edge node executes a supplementary data acquisition operation, which includes: adjusting the data acquisition frequency corresponding to the current monitoring data from a first frequency to a third frequency; retrieving historical monitoring data within the time range specified by the call command from the edge node's local cache; simultaneously acquiring associated video surveillance data and / or mobile trajectory data within the time range; and sending the historical monitoring data, associated video surveillance data, and mobile trajectory data as a supplementary data acquisition data packet to the cloud server. The third frequency is higher than the first frequency.

[0104] In addition to conventional automatic data collection, this application also supports one-click recall of data from cloud servers or edge devices to designated edge nodes, sites, pollution sources, mobile monitoring equipment, video points, or remote sensing grids. When the system detects a sudden increase in site concentration, indicators approaching thresholds, abnormal pollution sources, high values ​​in remote sensing patches, or the need for supplementary evidence during on-site verification, it can proactively issue recall commands to relevant edge nodes. That is, the cloud server or edge device can proactively initiate control commands to designated edge nodes, monitoring equipment, or data interfaces to collect, supplement, or transmit data instantly. This temporarily increases the sampling frequency, supplements data for key time periods, retrieves related video clips, or obtains mobile trajectory information, enabling the system to quickly supplement key evidence during the anomaly occurrence window and avoid missing pollution precursors or instantaneous anomalies due to relying solely on a fixed sampling frequency.

[0105] Furthermore, the system of this application also includes a scenario configuration module, which stores a preset scenario configuration template library. Each scenario template in this library is stored as an independent configuration file. Each scenario template includes data source combination configuration, threshold rule configuration, model link configuration, knowledge base retrieval configuration, and output strategy configuration. Specifically, the data source combination configuration specifies the type of data source to be called in this scenario (including site monitoring data, pollution source data, meteorological data, satellite data, etc.) and the data time window width; the threshold rule configuration specifies the static limit set, dynamic drift baseline calculation parameters, and early warning triggering conditions for this scenario; the model link configuration specifies the list of judgment models to be called sequentially in this scenario and the input / output interface mapping of each model; the knowledge base retrieval configuration specifies the priority knowledge base categories to be retrieved in this scenario and the limit on the number of retrieved results; and the output strategy configuration specifies the output format of the structured instruction set and the distribution target address for this scenario. By reading and loading different scenario configuration files, the system can switch to the corresponding business processing logic without modifying the underlying system code, achieving rapid adaptation to multiple scenarios.

[0106] In summary, this application constructs a cloud-edge-device collaborative sensing, edge pre-processing, cloud-to-ground fusion analysis, reliable traceability, and closed-loop feedback optimization framework for determining atmospheric pollution risk. Based on the pollution risk level, data anomaly characteristics, meteorological diffusion conditions, pollution source operating status, and a set of configurable local and external strategy parameters, it dynamically adjusts the data collection frequency, cache length, upload priority, threshold rules, model call path, and on-site handling strategies, thereby forming an intelligent closed loop of risk-driven sensing – edge initial screening and compression – cloud-to-ground fusion analysis – dynamic threshold determination – on-site precise source tracing – feedback iterative optimization. Furthermore, this application solves the problem of information fragmentation caused by the inconsistency of modalities in multi-source heterogeneous data by constructing a unified gridded situation tensor. It achieves quantitative source tracing and positioning through pollution source contribution analysis driven by graph neural networks, with a source tracing error significantly lower than existing empirical methods based on backward trajectories. Through a cloud-edge-device collaborative inference architecture, it achieves early warning response latency at the millisecond level, meeting the timeliness requirements of real-time ecological and environmental monitoring. At the same time, through joint retrieval and inference of knowledge graphs and RAG, the question-answering results have traceable source citations, solving the problem of factual illusion in professional fields caused by general large language models.

[0107] Example 2 like Figure 2 As shown in Embodiment 2 of this application, a pollution source tracing decision support system based on multi-source data fusion and intelligent agent collaboration is disclosed to implement the method described in Embodiment 1. The system includes: an edge node 210, a cloud server 220, and a target terminal device 230.

[0108] Edge node 210 is used to access multi-source heterogeneous environmental data, including a preprocessing module 211 and a first communication module 212. The multi-source heterogeneous environmental data includes air quality monitoring data, meteorological data, online pollution source monitoring data, electricity consumption monitoring data, mobile surveillance data, satellite remote sensing data, video surveillance data, and unstructured regulatory text data and historical case document data. Preprocessing module 211 is used for adaptive sensing and preprocessing of the multi-source heterogeneous environmental data, dynamically adjusting the data acquisition frequency and caching strategy according to the risk level of the environmental state. First communication module 212 is used to upload the processed multi-source heterogeneous environmental data to cloud server 220.

[0109] The cloud server 220 includes an analysis module 221, a generation module 222, a second communication module 223, and an incremental update module 224. The analysis module 221, based on the constructed environmental professional rule base, environmental professional knowledge base, and data foundation, utilizes retrieval-enhanced generation technology combined with a large language model to perform fusion analysis on processed multi-source heterogeneous environmental data, generating environmental situation analysis results and response strategies. The generation module 222 generates target data packets based on the environmental situation analysis results and response strategies. The second communication module 223 sends the target data packets to a preset target terminal device 230 via a wireless communication interface and receives on-site feedback signals returned by the target terminal device 230. The incremental update module 224 incrementally updates the cloud-based environmental professional rule base, environmental professional knowledge base, and model parameters based on the on-site feedback signals, and feeds the updated rule packets and / or model parameters back to the edge node 210, achieving closed-loop optimization of the system. The target data packets include at least one of early warning data, pollution source location data, or response instruction data.

[0110] Specifically, the preprocessing module 211 includes a risk perception unit 2111, a data acquisition parameter adjustment module 2112, a dynamic ring buffer 2113, a structured data verification unit 2114, an unstructured data parsing unit 2115, and a priority marking unit 2116. The risk perception unit 2111 generates a mode switching signal when multi-source heterogeneous environmental data meets a first preset condition. The first preset condition includes at least one of the following: the indicator value reaches a first percentage of a preset warning threshold, the meteorological diffusion index is lower than a preset diffusion threshold, or the fluctuation amplitude of the online monitoring data of the pollution source exceeds a preset fluctuation range. The data acquisition parameter adjustment module 2112 is used to adjust the data acquisition frequency from a first frequency to a second frequency in response to the mode switching signal, and / or adjust the number of monitoring nodes per unit area within the monitoring area from the first total number of nodes to the second total number of nodes. Wherein, the second frequency is greater than the first frequency, and the second total number of nodes is greater than the first total number of nodes. The dynamic circular buffer 2113 is used to cache environmental monitoring data. The data retention time is determined based on the current acquisition mode: a first retention time is used in the normal acquisition mode, and a second retention time is used in the abnormal acquisition mode; the second retention time is longer than the first retention time. The dynamic circular buffer 2113 is also used to mark data within the range of the third retention time before to the fourth retention time after the abnormal acquisition mode as non-coverable data when the abnormal acquisition mode is triggered, based on the trigger time of the abnormal acquisition mode. This non-coverable data is then extracted and a pollution event tracing data package is generated. The structured data verification unit 2114 is used to perform missing value detection on the structured data in the multi-source heterogeneous environmental data, marking null values ​​in data fields as missing values. It also performs outlier detection on the structured data, calculating the deviation of each data item from the mean, and marking data items with deviation values ​​exceeding a preset deviation threshold as candidate outliers. The unstructured data parsing unit 2115 is used to perform text parsing on the unstructured data in the multi-source heterogeneous environmental data, extracting entity feature words and numerical feature parameters from the unstructured data. Priority marking unit 2116 is used to encapsulate candidate outliers, statistical summaries of unstructured data, entity feature words, and numerical feature parameters into high-priority data packets.

[0111] The analysis module 221 includes a cross-validation unit 2211, a first-level analysis unit 2212, a second-level analysis unit 2213, and a retrieval and conclusion output unit 2214. The cross-validation unit 2211 calculates the dynamic reliability weight of each data source based on preset reliability parameters, the time deviation between the data acquisition timestamp and the reference time base, and the spatial proximity between the corresponding monitoring points of each data source. It then removes data from the processed multi-source heterogeneous environmental data whose dynamic reliability weight is lower than a first preset value, thus achieving cross-validation of the processed multi-source heterogeneous environmental data. The first-level analysis unit 2212 performs a first-level analysis on the validated data, comparing the indicator values ​​in the validated data with preset indicator thresholds to obtain the first-level analysis result. The second-level analysis unit 2213 executes a second-level analysis when the first-level analysis result meets preset trigger conditions, inputting the validated data into a spatiotemporal prediction model, a pollution source tracing model, and a knowledge graph to obtain pollution diffusion trend data and candidate pollution source data. The preset triggering conditions include: the existence of an indicator value exceeding a preset indicator threshold. The retrieval and conclusion output unit 2214 is used to retrieve corresponding regulatory text data and similar historical case document data from the environmental professional knowledge base based on the feature vector of the candidate pollution source data, combine the retrieval results with pollution diffusion trend data to form an enhanced prompt context, and input the enhanced prompt context into the large language model to output a judgment conclusion text containing a description of the pollution event and disposal suggestions.

[0112] In addition, the cloud server also includes: an environmental indicator calculation and anomaly detection module, a task distribution arbitrator, and a comprehensive analysis report generation module.

[0113] The environmental indicator calculation and anomaly detection module automatically calculates the statistical sequence for the current monitoring period according to a preset limit reference system, and determines whether the statistical sequence is high, exceeds the threshold, is inverted, has equipment malfunction, or has a risk of exceeding standards or pollution risk based on an environmental professional rule base. The statistical sequence includes real-time values, cumulative values, comparison values, ranking values, remaining capacity, time series progress deviation, and high-value statistics.

[0114] The task distribution arbitrator forwards the current task to the corresponding processing module according to a preset routing table, based on the judgment result label and the request type code sent by the terminal device. The processing modules include an air quality information formatting module, a compliance status monitoring module, an abnormal event alarm output module, a forecast and warning generation module, an automatic map generation module, a semantic retrieval and generation module, a multi-source data association analysis module, and a distributed task coordination module. The air quality information formatting module and the compliance status monitoring module generate structured data reports based on the environmental indicator calculation and anomaly judgment results. These structured data reports include real-time data reports, cumulative statistical reports, comparison difference reports, remaining capacity reports, time-series progress deviation reports, and site anomaly reports. Maps... The automatic generation module is used to call meteorological numerical forecast output, air quality numerical model prediction results, pollution source emission inventories, remote sensing image inversion data, and geographic information system spatial data to automatically generate air quality prediction maps, meteorological anomaly maps, concentration spatial distribution maps, transmission path maps, remote sensing inversion maps, source tracing direction maps, time series trend maps, or multi-source comparison maps; the semantic retrieval and generation module is used to perform query encoding, semantic retrieval, full-text matching retrieval, tag filtering retrieval, multi-path fusion sorting, and query generation output processing on query statements input through user command interfaces via terminal devices, using the multimodal index of the knowledge base.

[0115] The comprehensive analysis report generation module integrates statistical sequences, judgment result labels, retrieval knowledge text, and structured instruction sets to output a structured analysis report containing anomaly attribution label sequences, execution action sequences, data statistical summaries, spatial distribution summaries, and trend prediction summaries.

[0116] The above-described embodiments of the pollution source tracing decision support system based on multi-source data fusion and intelligent agent collaboration correspond to the method embodiments and have the same technical effects as the method embodiments. For details, please refer to the method embodiments, which will not be repeated here.

[0117] Example 3 This third embodiment uses a scenario of a sudden increase in PM2.5 concentration at an automatic air quality monitoring station in a certain district as an example to illustrate the specific implementation process of the pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration provided in this application.

[0118] like Figure 3As shown, under normal operating conditions, the system continuously receives air quality station data, online pollution source data, electricity consumption data, meteorological data, remote sensing grid data, mobile monitoring data, video location data, and GIS spatial data through edge nodes. Each edge node collects data at a regular sampling frequency and performs timestamp verification, station code matching, missing value identification, outlier screening, and data summary generation locally. At this time, the cloud server only receives routine statistical features, real-time concentration values, and necessary status identifiers to reduce network transmission and cloud computing pressure.

[0119] When the system detects a rapid increase in PM2.5 concentration at an air quality monitoring station within a short period, approaching or exceeding the local rule threshold, the edge-side risk perception unit identifies this change as a "sudden increase in station particulate matter event." The system then marks air quality, meteorological wind field, online pollution source data, electricity consumption, remote sensing grid data, and mobile monitoring data within a certain radius of the station as high-priority data and automatically increases the frequency of related data collection. Simultaneously, the edge-side dynamic ring buffer automatically extends the historical data retention window, locking in continuous data segments before, during, and after the anomaly to prevent precursors of pollution formation from being overwritten.

[0120] Upon triggering an anomaly, the system automatically sends a one-click recall command from the cloud server to relevant edge nodes. This command requests designated sites, online pollution source equipment, video points, mobile monitoring equipment, or remote sensing grid services to immediately supplement the collection or transmission of critical data. For example, the system can recall online pollution source data and electricity consumption fluctuation data of key air-polluting enterprises in the upwind area of ​​the site, retrieve video clips from surrounding video points during the abnormal period, obtain meteorological elements such as current wind direction and speed, boundary layer height, and humidity, and simultaneously retrieve the most recent high-value remote sensing patch and historical records of similar pollution events in the area.

[0121] After initial screening of abnormal data at the edge, suspected anomaly characteristics, anomaly occurrence time windows, station concentration change curves, online pollution source change characteristics, electricity consumption anomalies, meteorological wind field characteristics, and remote sensing hotspot grid information are uploaded to the cloud server. Upon receiving high-priority data, the cloud initiates a multi-model collaborative analysis process: First, the rule engine determines whether the sudden increase in PM2.5 triggers local high-value rules, continuous increase rules, compliance risk rules, or alarm push rules; second, the time-series model determines whether the station's concentration change significantly deviates from historical levels for the same period; third, the spatial analysis model compares the spatial relationships between surrounding stations, surrounding grids, and upwind pollution sources; then, the pollution source tracing model, combined with wind direction and speed, pollution source location, emission data, and remote sensing patches, determines the potentially affected areas; finally, the knowledge graph and retrieval-enhanced generation model retrieves historical similar cases, control experiences, and governance measures to provide professional evidence for the analysis conclusions.

[0122] In the process of determining whether a pollution level exceeds the limit, the system does not rely solely on fixed thresholds. Instead, it combines national standards, local assessment targets, station background values, historical concentrations for the same period, real-time meteorological conditions, pollution source operational status, changes at surrounding stations, and regional transmission conditions to generate dynamic judgment results. If only a single station experiences a short-term increase, and there is no supporting evidence from pollution sources, meteorology, remote sensing, or surrounding stations, the system marks it as "suspected equipment anomaly or clue awaiting verification." If the sudden increase in concentration at a station shows a high degree of consistency with upwind pollution source emission fluctuations, abnormal power consumption, meteorological transmission direction, high-value remote sensing patches, and similar historical events, the system increases the event confidence level and classifies it as a "local pollution impact event" or a "suspected pollution source impact event."

[0123] During the on-site source tracing phase, the system automatically generates pollution hotspots, priority investigation directions, suggested patrol routes, and inspection lists based on GIS spatial data, pollution source lists, meteorological wind fields, station concentration changes, remote sensing grid high-value distribution, and historical cases. Output results may include the grid number of the suspected affected area, the latitude and longitude of its center point, the name of the upwind road or area, the suggested patrol start and end points, a list of key inspection enterprises, the type of suspected pollution source, suggested verification items, and on-site evidence collection requirements. For example, the system may prompt on-site personnel to prioritize patrols along roads upwind of the stations, focusing on verifying the boundaries of an industrial park, the emission outlets of a certain air-polluting enterprise, a construction dust point, or a mobile source accumulation area along a road, while simultaneously explaining that this suggestion is derived from a comprehensive matching result of station high values, wind direction and speed, remote sensing patches, and historical problem events.

[0124] In the collaborative dispatch phase, the system automatically generates standardized disposal work orders or dispatch suggestions based on pollution type, administrative boundaries, responsible parties, disposal time limits, and historical disposal rules. Work order content includes event number, abnormal time, abnormal site, abnormal indicator, concentration change, triggering rules, suspected affected area, suggested navigation route, key verification targets, responsible unit, rectification suggestions, and feedback requirements. The system pushes work orders to relevant business departments, law enforcement personnel, local towns / streets, or industry authorities via government WeChat, enterprise WeChat, or business terminals, forming a closed-loop disposal process in work groups: "assessment – ​​precise dispatch – on-site verification – result feedback – intelligent verification."

[0125] During the feedback and optimization phase, on-site personnel upload the verification results, photos of problems, rectification status, and handling feedback to the system. The system automatically updates the pollution event chain and writes the verification conclusions into the historical case database, abnormal sample database, and rule optimization database. If such events occur repeatedly, the cloud server incrementally optimizes the relevant thresholds, feature weights, rule packages, navigation route recommendation strategies, and report templates, and distributes the updated rule packages or lightweight model parameters to the corresponding edge nodes, enabling the system to identify, respond, and push more accurately in subsequent similar scenarios.

[0126] As can be seen from the above specific application process, this application does not only complete data display or alarm notification in this scenario, but also realizes a complete technical closed loop from multi-source data adaptive collection, edge-side anomaly screening, cloud-based multi-model fusion analysis, dynamic threshold exceeding judgment, multi-source cross-validation, on-site navigation route generation, responsible object dispatch, handling result feedback to model rule reinjection.

[0127] It will be understood by those skilled in the art that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. The modules in the system of the embodiments may be distributed throughout the system of the embodiments as described in the embodiments, or may be located in one or more systems different from this embodiment with corresponding changes. The modules of the above embodiments may be combined into one module, or may be further divided into multiple sub-modules. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the above embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration, characterized in that, The method is implemented based on a cloud-edge-device collaborative architecture, including: Multi-source heterogeneous environmental data is acquired through edge-side nodes. The multi-source heterogeneous environmental data includes air quality monitoring data, meteorological data, online pollution source monitoring data, electricity consumption monitoring data, mobile monitoring data, satellite remote sensing data, video surveillance data, as well as unstructured regulatory text data and historical case document data. Adaptive sensing and preprocessing of the multi-source heterogeneous environment data are performed at the edge. The data collection frequency and caching strategy are dynamically adjusted according to the risk level of the environment. The processed multi-source heterogeneous environment data is then uploaded to the cloud server. Based on the constructed environmental professional rule base, environmental professional knowledge base and data foundation in the cloud, and using retrieval enhancement generation technology combined with a large language model, the processed multi-source heterogeneous environmental data is fused and analyzed to generate environmental situation analysis results and disposal strategies. Based on the environmental situation analysis results and the disposal strategy, a target data packet is generated and sent to a preset target terminal device via a wireless communication interface. The target terminal device then returns a field feedback signal. The target data packet includes at least one of the following: early warning data, pollution source location data, or disposal instruction data. Based on the on-site feedback signals, the environmental professional rule base, the environmental professional knowledge base, and the model parameters in the cloud are incrementally updated, and the updated rule package and / or model parameters are fed back to the edge-side nodes to achieve closed-loop optimization of the system.

2. The pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration according to claim 1, characterized in that, The adaptive sensing and preprocessing of the multi-source heterogeneous environmental data at the edge specifically includes: When the multi-source heterogeneous environmental data meets the first preset condition, the data acquisition mode is switched from the normal acquisition mode to the abnormal acquisition mode. In the abnormal acquisition mode, the data acquisition frequency is adjusted from the first frequency to the second frequency, and / or the number of monitoring nodes per unit area in the monitoring area is adjusted from the first total number of nodes to the second total number of nodes. Wherein, the second frequency is greater than the first frequency, the second total number of nodes is greater than the first total number of nodes, and the first preset condition includes at least one of the following: the indicator value reaches a first percentage of the preset warning threshold, the meteorological diffusion index is lower than the first preset diffusion threshold, or the fluctuation range of the online monitoring data of the pollution source exceeds the first preset fluctuation range. When the multi-source heterogeneous environmental data meets the second preset condition, the data acquisition mode is switched to the early warning perception mode. In the early warning perception mode, the data retention time of the dynamic ring buffer set on the edge side is extended from the first time to the second time. When the multi-source heterogeneous environmental data meets the third preset condition, the data acquisition mode is switched to the anomaly perception mode. In the anomaly perception mode, based on the trigger time of the anomaly perception mode, the data in the dynamic ring buffer located between the third time before the benchmark and the fourth time after the benchmark is marked as non-coverable data. The non-coverable data is extracted to generate a pollution event tracing data packet. The second preset condition... The conditions include at least one of the following: the indicator value reaches a second percentage of the preset warning threshold, the meteorological diffusion index is lower than the second preset diffusion threshold, or the fluctuation range of the online monitoring data of the pollution source exceeds the second preset fluctuation range; the third preset condition includes at least one of the following: the rate of change of the monitoring data within a unit time window exceeds the first preset rate of change threshold, the monitoring data continuously exceeds the first preset high threshold in multiple consecutive sampling cycles, the instantaneous value of the monitoring data exceeds the first preset over threshold, the temporal logical relationship or numerical ratio relationship between multiple sets of associated monitoring data deviates from the preset benchmark interval, or the characteristic fingerprint in the monitoring data matches the abnormal spectrum in the preset pollution source feature library; Missing value detection is performed on the structured data in the multi-source heterogeneous environmental data, marking null values ​​in the data fields as missing values. Outlier detection is also performed on the structured data, calculating the deviation of each data item from the mean, and marking data items with deviation values ​​exceeding a preset deviation threshold as candidate outliers. The structured data is then processed by removing duplicate values, verifying timestamps, matching site codes, matching spatial locations, and converting pollutant indicator calibers. Simultaneously, the unstructured data in the multi-source heterogeneous environmental data is processed by text parsing, chapter segmentation, tag extraction, and vectorization. The preprocessed relevant outlier data and the uncoverable data are encapsulated into a high-priority data packet and sent to the cloud server first through a communication interface. The relevant outlier data includes at least one of the following: amplitude domain outlier features, temporal gradient features, threshold proximity values, pollution source outliers, or spatial hotspot data.

3. The pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration according to claim 1, characterized in that, The process of fusing and analyzing the processed multi-source heterogeneous environmental data using retrieval enhancement generation technology combined with a large language model specifically includes: Based on the preset reliability parameters of each data source, the time deviation between the data acquisition timestamp and the reference time base, and the spatial proximity between the corresponding monitoring points of each data source, the dynamic reliability weight of each data source is calculated, and the data of the data source with a dynamic reliability weight lower than the first preset value in the processed multi-source heterogeneous environment data is removed, so as to realize the cross-validation of the processed multi-source heterogeneous environment data. The index values ​​in the verified data are compared with the preset index thresholds. If any index value exceeds the preset index threshold, a deep analysis is triggered. The verified data is then input into the spatiotemporal prediction model, the pollution source tracing model, and the knowledge graph to obtain pollution diffusion trend data and candidate pollution source data. Based on the feature vectors of the candidate pollution source data, the corresponding regulatory text data and similar historical case document data are retrieved from the environmental professional knowledge base. The retrieval results are combined with the pollution diffusion trend data to form an enhanced prompt context. The enhanced prompt context is then input into a large language model to output a judgment conclusion text containing a description of the pollution event and disposal suggestions.

4. The pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration according to claim 1, characterized in that, After generating the environmental situation analysis results and response strategies, the method further includes: generating a precise source tracing and collaborative scheduling scheme, specifically including: Based on the pollution source location information in the environmental situation analysis results, combined with meteorological wind field vector data, mobile mission planning data containing at least one path coordinate sequence is generated; wherein, based on the pollutant concentration value at each pollution source location and the preset sorting rules, the investigation priority weight value of each pollution source is calculated, and pollution source investigation sequence data is generated in descending order of the investigation priority weight values. Based on the mobile mission planning data and the preset spatial grid ownership data, determine the jurisdiction identifier and timeliness identifier corresponding to each grid; The event markers, time markers, pollution source identifiers, and mobile mission planning data from the environmental situation analysis results are encapsulated into a structured mission data package according to a preset data structure template, and the structured mission data package is sent to the target terminal device corresponding to the jurisdiction identifier through a communication interface.

5. The pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration according to claim 1, characterized in that, In the cloud, the method further includes: a scene-adaptive dynamic threshold determination step, specifically including: The system obtains the scenario type identifier of the current environmental monitoring scenario and retrieves the corresponding benchmark index threshold from the preset benchmark threshold library based on the scenario type identifier. The scenario type identifier includes static and stable accumulation scenario, dust transport scenario, and major event support scenario. The benchmark index threshold is composed of a static benchmark layer, a dynamic benchmark layer, and a variable strategy layer. The static benchmark layer contains at least one set of national mandatory emission limits and corresponding local emission assessment characteristic values. The variable strategy layer temporarily adjusts the benchmark values ​​of the static benchmark layer by a multiplier based on a user-defined time window weighting factor and a local area scaling factor to obtain the benchmark index threshold. The dynamic benchmark layer is used to determine whether the residual exceeds a preset multiple standard deviation of the benchmark index threshold to calibrate statistical anomaly characteristics. Acquire current real-time meteorological data, and calculate the correction coefficient of each meteorological factor to the benchmark index threshold based on the current real-time meteorological data; the current real-time meteorological data includes meteorological condition parameters, boundary layer height, wind speed, wind direction and humidity data, historical level data for the same period, station background concentration data, pollution source operation status data, regional transmission characteristic data, time-specific dynamic warning line and expected compliance difficulty coefficient; Based on the benchmark threshold and the correction coefficient, a preset index threshold adapted to the current environmental monitoring scenario is calculated. Based on the preset index threshold, anomaly determination is made on the pollutant concentration monitoring values ​​in the processed multi-source heterogeneous environmental data to obtain anomaly marking results. Specifically, in the static and stable accumulation scenario, the alarm sensitivity of the preset index threshold is higher than that of the benchmark index threshold; in the sand and dust transport scenario, the alarm sensitivity of the preset index threshold is lower than that of the benchmark index threshold.

6. The pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration according to claim 1, characterized in that, The step of incrementally updating the environmental professional rule base, the environmental professional knowledge base, and model parameters in the cloud based on the on-site feedback signal, and then feeding the updated rule package and / or model parameters back to the edge node, specifically includes: Acquire the original monitoring data, edge preprocessing result data, abnormal triggering rule identifiers, model judgment result data, push record data, and on-site verification feedback data corresponding to the current pollution event; According to the preset association rules, establish the temporal and causal association relationships between the original monitoring data, the edge preprocessing result data, the anomaly triggering rule identifier, the model judgment result data, the push record data, and the on-site verification feedback data, and generate a pollution event association data record set; When the cumulative number of abnormal records belonging to the same pollution type or the same monitoring station in the pollution event associated data record set reaches a preset threshold, rule update data or model parameter increment data is generated based on the verified data in the pollution event associated data record set. The rule update data and / or the model parameter increment data are distributed to the edge nodes through a distributed model aggregation method, so that the edge nodes update the decision rule data and / or model parameter data stored locally.

7. The pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration according to claim 1, characterized in that, The method further includes: When the environmental situation analysis results generated in the cloud meet the preset supplementary sampling trigger conditions, a recall command is generated and sent to the designated edge node; wherein, the preset supplementary sampling trigger conditions include at least one of the following: the concentration change rate of the current monitoring data within a preset time window exceeds a second preset change rate threshold, the difference between the current monitoring data and the monitoring data at the previous moment exceeds a first preset difference threshold, or the pollutant concentration value in the remote sensing inversion data exceeds a preset high value threshold; In response to the call command, the edge node performs a supplementary data acquisition operation, which includes: adjusting the data acquisition frequency corresponding to the current monitoring data from a first frequency to a third frequency, extracting historical monitoring data within the time range specified by the call command from the local cache of the edge node, and simultaneously acquiring associated video surveillance data and / or mobile trajectory data within the time range, and sending the historical monitoring data, the associated video surveillance data, and the mobile trajectory data as a supplementary data acquisition data packet to the cloud server; wherein the third frequency is greater than the first frequency.

8. The pollution source tracing decision support method based on multi-source data fusion and intelligent agent collaboration according to claim 1, characterized in that, The method further includes: Each edge node obtains the overall upload weight of the current data packet through its built-in data priority quantization engine, and each edge node then performs the upload based on the overall upload weight. Compared with the preset first priority threshold Second priority threshold Based on the comparison results, hierarchical transmission scheduling is performed on the current data packet: when When the current data packet is marked as a high-priority data packet, a real-time interruption of the upload is triggered, and the transmission delay does not exceed a first preset duration; when When the current data packet is marked as a medium-priority data packet, it is stored in a local cache queue and uploaded using a time window aggregation method. The aggregation window width is a second preset duration, and the second preset duration is greater than the first preset duration. When this happens, the current data packet is marked as a low-priority data packet, stored in local memory, and then uploaded in batches during a preset off-peak communication period; wherein, The comprehensive upload weight The calculation formula is: ； In the above formula, This is a normalized risk coefficient based on the current monitored value relative to the predicted statutory limit; Priority encoding value for sensor data type; Quality score based on signal integrity verification; This represents the influence radius coefficient of the current edge node in the spatial topology. It is a monotonically decreasing function representing the remaining time from the preset target termination time window at the current time; These are the dynamic configuration coefficients input through an external configuration interface; This represents the current real-time load rate in the cloud. a, b, c, d, e, g, h These are the preset weight coefficients for each factor, and their sum is 1; When the cloud server receives anomaly confirmation signals triggered simultaneously by multiple edge nodes, it executes a multi-node concurrent resolution step: It constructs a spatiotemporal coupling sorting factor for each anomaly event, sorts them in descending order according to the value of the spatiotemporal coupling sorting factor, generates an anomaly event processing priority sequence list, and dynamically determines the number K of events to be processed in a single batch based on the current real-time load rate of the cloud server. It then allocates the top K anomaly data packets in the anomaly event processing priority sequence list to the cloud server's reserved computing power resource pool, triggering a fusion analysis process. Simultaneously, it stores the top K+1 anomaly data packets and subsequent packets in the anomaly event processing priority sequence list into a waiting queue and starts a cyclic monitoring process. Every preset time interval, it re-acquires the current real-time load rate of the cloud server and repeats the aforementioned steps until all anomaly data packets have been allocated to the reserved computing power resource pool. The spatiotemporal coupling sorting factor... The calculation formula is: ； In the above formula, For the first Monitoring values ​​uploaded by each edge node This indicates the normalization overshoot factor of the node; For the first The edge side node and the first The Euclidean spatial distance between each edge node in the Deloni triangulation. Indicates spatial proximity and density; The upwind source probability value is calculated based on real-time wind field vector data; For the first The remaining time between the current time of each edge node and the preset target termination time window. This indicates the time urgency of the target task; , , , These are the preset weight coefficients for each sub-item, and their sum is 1.

9. A pollution source tracing decision support system based on multi-source data fusion and intelligent agent collaboration, characterized in that, The system for implementing the method as described in any one of claims 1–8 comprises: Edge nodes, used to access multi-source heterogeneous environmental data, include a preprocessing module and a first communication module. The multi-source heterogeneous environmental data includes air quality monitoring data, meteorological data, online pollution source monitoring data, electricity consumption monitoring data, mobile surveillance data, satellite remote sensing data, video surveillance data, and unstructured regulatory text data and historical case document data. The preprocessing module is used to perform adaptive sensing and preprocessing on the multi-source heterogeneous environmental data, dynamically adjusting the data acquisition frequency and caching strategy according to the risk level of the environmental state. The first communication module is used to upload the processed multi-source heterogeneous environmental data to a cloud server. The cloud server includes an analysis module, a generation module, a second communication module, and an incremental update module. The analysis module is used to perform fusion analysis on the processed multi-source heterogeneous environmental data based on the constructed environmental professional rule base, environmental professional knowledge base, and data foundation, and using retrieval enhancement generation technology combined with a large language model to generate environmental situation analysis results and disposal strategies. The generation module is used to generate target data packets based on the environmental situation analysis results and the disposal strategies. The second communication module is used to send the target data packets to a preset target terminal device through a wireless communication interface and receive on-site feedback signals returned by the target terminal device. The incremental update module is used to incrementally update the environmental professional rule base, environmental professional knowledge base, and model parameters in the cloud based on the on-site feedback signals, and feed the updated rule packets and / or model parameters back to the edge-side nodes to achieve closed-loop optimization of the system. The target data packets include at least one of early warning prompt data, pollution source location data, or disposal instruction data.

10. The pollution source tracing decision support system based on multi-source data fusion and intelligent agent collaboration according to claim 9, characterized in that, The cloud server also includes: The environmental indicator calculation and anomaly judgment module is used to automatically calculate the statistical sequence within the current monitoring period according to a preset limit reference system, and to judge the statistical sequence as high value, exceeding threshold, inverted, equipment abnormal, exceeding standard risk or pollution risk based on an environmental professional rule base; wherein, the statistical sequence includes real-time value, cumulative value, comparison value, ranking value, remaining capacity, time series progress deviation and high value statistics; A task distribution arbitrator is used to forward the current task to the corresponding processing module according to a preset routing mapping table based on the judgment result label and the request type code sent by the terminal device. The processing module includes an air quality information formatting module, a compliance status monitoring module, an abnormal event alarm output module, a forecast and early warning generation module, an automatic map generation module, a semantic retrieval and generation module, a multi-source data association analysis module, and a distributed task coordination module. The air quality information formatting module and the compliance status monitoring module are used to generate structured data reports based on the judgment results of the environmental indicator calculation and abnormal judgment modules. The structured data reports include real-time data reports, cumulative statistical reports, and... The system includes reports on differences, remaining capacity, time-series progress deviations, and site anomalies. The automatic map generation module uses meteorological numerical forecast output, air quality numerical model prediction results, pollution source emission inventories, remote sensing image inversion data, and geographic information system spatial data to automatically generate air quality prediction maps, meteorological anomaly maps, concentration spatial distribution maps, transmission path maps, remote sensing inversion maps, source tracing direction maps, time-series trend maps, or multi-source comparison maps. The semantic retrieval and generation module uses a multimodal index based on a knowledge base to perform query encoding, semantic retrieval, full-text matching retrieval, tag filtering retrieval, multi-path fusion sorting, and query generation output processing on query statements input through the user command interface on the terminal device. The comprehensive analysis report generation module is used to integrate the statistical sequence, judgment result labels, retrieval knowledge text and structured instruction set, and output a structured analysis report containing anomaly attribution label sequence, execution action sequence, data statistical summary, spatial distribution summary and trend prediction summary.

Images