Method and device for analyzing new urban pollution sources based on multi-agent machine learning

Abstract

The application discloses a kind of city new pollution source analysis method and device based on multi-agent machine learning, the method includes: the original multi-medium data collected to multi-agent is carried out feature extraction, obtains fusion feature matrix, sends fusion feature matrix to multi-agent, to make multi-agent output source analysis result, based on source analysis result, the monitoring strategy of multi-agent is adjusted, and strategy configuration is carried out, and the multi-agent after strategy configuration is dynamically tracked and monitored to pollution source area, based on dynamic tracking monitoring result, multi-agent is iteratively optimized, and the multi-agent after iteratively optimized is carried out city new pollution source analysis, and data monitoring and pollution analysis are carried out by multi-agent, and dynamically optimize monitoring strategy and analysis model, greatly improve the monitoring data coverage and cross-region analysis precision, effectively improve the identification accuracy and identification efficiency of city new pollution.

Description

Technical Field

[0001] This invention relates to the field of intelligent environmental perception, and in particular to a method and apparatus for analyzing new urban pollutant sources based on multi-agent machine learning. Background Technology

[0002] With rapid urbanization and the emergence of new industries, a large number of new pollutants have appeared in urban environments, which are difficult to identify and control using traditional monitoring systems. These new pollutants include, but are not limited to, pharmaceuticals and personal care products (PPCPs), endocrine disrupting chemicals (EDCs), per- and polyfluoroalkyl substances (PFAS), and microplastics. They are characterized by strong environmental persistence, high bioaccumulation, and long-term low-dose exposure, posing a potential threat to the ecological environment and human health. Currently, in the field of source apportionment of new urban pollutants, traditional non-machine learning methods still dominate. Existing technologies mainly employ a combination of monitoring point deployment and chemical analysis. By setting up monitoring points in different areas of the city, samples are collected for chemical analysis to determine the possible sources of pollutants. This approach mainly includes three technical paths: first, on-site sampling, laboratory analysis, and receptor modeling; second, on-site investigation, statistical analysis, and source inventory; and third, pollution source monitoring, simulation analysis, and source modeling.

[0003] Traditional urban pollutant monitoring and analysis technologies suffer from several shortcomings: Regarding monitoring site selection, traditional methods combine site selection with chemical analysis. However, these methods are heavily reliant on human experience and suffer from significant subjectivity. In terms of chemical analysis, they depend on chromatography-mass spectrometry (GC-MS), making it difficult to cover potential targets with different functional groups and structural characteristics given the diverse physicochemical properties of millions of industrial chemicals. Furthermore, their ability to identify emerging pollutants is insufficient. In terms of effectiveness, traditional methods suffer from low spatial and temporal resolution. Finally, in terms of cost-effectiveness, traditional methods are expensive and resource-intensive, relying heavily on substantial human and equipment investment for environmental monitoring. Summary of the Invention

[0004] The main objective of this invention is to provide a method and apparatus for analyzing new urban pollutant sources based on multi-agent machine learning. This invention aims to address the shortcomings of existing traditional urban pollutant source analysis technologies, such as strong reliance on human experience, limited coverage of monitoring data, insufficient ability to identify new pollutants, and low analysis efficiency and accuracy. These shortcomings make it difficult to efficiently and accurately analyze new urban pollutants.

[0005] To achieve the above objectives, this invention provides a method for apportioning emerging urban pollutant sources based on multi-agent machine learning, the method comprising the following steps:

[0006] Based on the multi-media occurrence characteristics of new urban pollutants, a multi-agent monitoring strategy is configured, and data is collected through the multi-agents configured with the strategy to obtain raw multi-media data. The multi-media occurrence characteristics are the characteristics of new urban pollutants when they are present in different media carriers. The monitoring strategy includes the monitoring range, monitoring tracking path, monitoring frequency and monitoring indicators of the agent carrier.

[0007] Feature extraction is performed on the original multi-media data to obtain a fused feature matrix;

[0008] The fused feature matrix is ​​sent to a multi-agent system, which is configured to construct a source resolution model based on the fused feature matrix and output source resolution results through the source resolution model. The source resolution results include pollution source areas and the pollution contribution of each pollution source area.

[0009] The monitoring strategy is adjusted based on the source resolution results output by the multi-agent, the multi-agent is configured with a strategy based on the adjusted monitoring strategy, and the pollution source area is dynamically tracked and monitored through the multi-agent with the configured strategy.

[0010] The multi-agent system is iteratively optimized based on the dynamic tracking and monitoring results. The iterative optimization includes optimizing the model parameters of the source resolution model and iteratively optimizing the monitoring strategy.

[0011] The new urban pollutant sources are analyzed through a multi-agent system after iterative optimization.

[0012] Optionally, the step of extracting features from the original multi-media data to obtain a fused feature matrix includes:

[0013] The original multi-media data is standardized to obtain initial multi-media data;

[0014] The initial multi-media data is preprocessed to obtain candidate multi-media data. The preprocessing includes data cleaning and data augmentation.

[0015] The candidate multi-media data is spatiotemporally aligned, and the spatiotemporally aligned candidate multi-media data is feature fused based on the multi-media storage features to obtain a fused feature matrix.

[0016] Optionally, the step of performing spatiotemporal alignment on the candidate multi-media data and fusing features of the spatiotemporally aligned candidate multi-media data based on the multi-media assignment features to obtain a fused feature matrix includes:

[0017] Based on the monitoring frequency and monitoring range of the candidate multi-media data, determine the time alignment anchor point information and the spatial alignment grid information;

[0018] A spatiotemporal coordinate system is constructed based on the time alignment anchor point information and the spatial alignment grid information, and the candidate multi-media data is mapped to the spatiotemporal coordinate system to perform spatiotemporal alignment on the candidate multi-media data;

[0019] Based on the multi-media storage characteristics, the source resolution contribution of each medium is analyzed, and based on the source resolution contribution, the spatiotemporally aligned candidate multi-media data are subjected to feature weighted fusion to obtain an initial fusion matrix.

[0020] Calculate the mutual information value between each feature in the initial fusion matrix and the pollution source type label, and perform feature filtering on the initial fusion matrix based on the mutual information value to obtain the fusion feature matrix.

[0021] Optionally, the process of the multi-agent constructing a source resolution model based on the fused feature matrix and outputting the source resolution result through the source resolution model includes:

[0022] The received fusion feature matrix is ​​subjected to data dimensionality reduction and data partitioning to obtain the model input data;

[0023] An original machine learning model is constructed, which includes a CNN spatial feature extraction module, an RNN temporal feature extraction module, a random forest pollution source classification module, and a contribution quantification module. The CNN spatial feature extraction module is configured to use multi-layer convolutional kernels to capture spatial correlation features in the data. The RNN temporal feature extraction module is configured to use LSTM units to capture temporal fluctuations in pollution concentration. The random forest pollution source classification module is configured to classify pollution source types through voting by multiple decision trees. The contribution quantification module is configured to quantify the contribution of pollution sources.

[0024] Construct a joint loss function, the mathematical expression of which is:

[0025]

[0026] in, Denotes the joint loss function. Represents the classification loss weights. Indicates the regression loss weights. Represents the classification loss term. The true label indicating the type of pollution source Predictive labels indicating the type of pollution source Represents the regression loss term. This represents the true value of the contribution of pollution sources. The predicted value representing the contribution of pollution sources;

[0027] The original machine learning model is trained based on the model input data and the joint loss function, and the parameters of the original machine learning model are optimized by the Adam optimizer to obtain the source resolution model;

[0028] The source resolution model outputs the source resolution result.

[0029] Optionally, adjusting the monitoring strategy based on the source resolution results output by the multi-agent includes:

[0030] Based on the source resolution results output by the multi-agent, the monitoring and tracking paths in the monitoring strategy are optimized to obtain the path optimization results. The path optimization is performed according to the following formula:

[0031]

[0032] in, Indicates the state Next action The expected cumulative reward, Indicates the learning rate. Indicates the execution of an action The instant reward obtained afterward Indicates the discount factor. Indicates the execution of an action The new state afterward Indicates a new state All available actions Indicates a new state The maximum expected cumulative reward that can be obtained from all available actions;

[0033] Configure the contribution threshold of each pollution source region based on the functional area type of each pollution source region in the source apportionment results.

[0034] Based on meteorological data and the source apportionment results, pollution diffusion paths are predicted to obtain pollution diffusion path prediction results. The prediction formula for the pollution diffusion path is as follows:

[0035]

[0036] in, Indicates that in spatial coordinates The time is Predicted pollution concentration at that time This represents the source strength, which is calculated based on the pollution contribution of each pollution source region in the source apportionment results. This represents the average wind speed. This indicates the effective source height, which is the effective emission height of the pollution source. Indicates the horizontal diffusion parameter. Indicates the vertical diffusion parameter. Represents the horizontal coordinates perpendicular to the downwind direction. Represents vertical coordinates, Indicates the distance downwind;

[0037] The monitoring strategy is adjusted based on the path optimization results, the contribution threshold, and the pollution diffusion path prediction results.

[0038] Optionally, the iterative optimization of the multi-agent system based on dynamic tracking and monitoring results includes:

[0039] The measured pollution concentration in the pollution source area is obtained based on the dynamic tracking and monitoring results, and the concentration error is calculated based on the measured pollution concentration and the predicted pollution concentration.

[0040] The source apportionment results are compared with the apportionment results of a pre-constructed control model to obtain a consistency coefficient. The control model is a chemical mass balance model.

[0041] The current pollution pattern characteristics are determined based on the source apportionment results, and the historical pollution pattern characteristics are determined based on the historical source apportionment results.

[0042] The current pollution pattern distribution characteristics and the historical pollution pattern characteristics are used to perform pollution drift detection to obtain pollution drift data;

[0043] The multi-agent system is iteratively optimized based on the concentration error, the consistency coefficient, and the pollution drift data.

[0044] Furthermore, to achieve the above objectives, the present invention also proposes an urban new pollutant source apportionment device based on multi-agent machine learning, the urban new pollutant source apportionment device based on multi-agent machine learning comprising:

[0045] The multi-agent data acquisition module is used to configure the monitoring strategy of the multi-agent based on the multi-media storage characteristics of new urban pollutants, and to collect data through the multi-agent after the strategy configuration to obtain the original multi-media data. The multi-media storage characteristics are the characteristics of new urban pollutants when they are stored in different media carriers. The monitoring strategy includes the monitoring range, monitoring tracking path, monitoring frequency and monitoring indicators of the agent carrier.

[0046] A multi-media feature extraction module is used to extract features from the original multi-media data to obtain a fused feature matrix;

[0047] A multi-agent machine learning source resolution module is used to send the fused feature matrix to a multi-agent, which is configured to build a source resolution model based on the fused feature matrix and output source resolution results through the source resolution model. The source resolution results include pollution source areas and the pollution contribution of each pollution source area.

[0048] The pollution dynamic tracking module is used to adjust the monitoring strategy based on the source analysis results output by the multi-agent, configure the multi-agent strategy based on the adjusted monitoring strategy, and dynamically track and monitor the pollution source area through the multi-agent strategy configured with the strategy.

[0049] A multi-agent optimization module is used to iteratively optimize the multi-agent based on dynamic tracking and monitoring results. The iterative optimization includes optimizing the model parameters of the source resolution model and iteratively optimizing the monitoring strategy.

[0050] The new pollutant analysis module is used to analyze new urban pollutant sources through a multi-agent system that has undergone iterative optimization.

[0051] Furthermore, to achieve the above objectives, this application also proposes an urban new pollutant source apportionment device based on multi-agent machine learning. The device includes: a memory, a processor, and an urban new pollutant source apportionment program based on multi-agent machine learning stored in the memory. The processor is used to run the urban new pollutant source apportionment program based on multi-agent machine learning. The computer program is configured to implement the steps of the urban new pollutant source apportionment method based on multi-agent machine learning as described above.

[0052] In addition, to achieve the above objectives, this application also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the urban new pollutant source apportionment method based on multi-agent machine learning as described above.

[0053] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the urban new pollutant source apportionment method based on multi-agent machine learning as described above.

[0054] This invention configures a multi-agent monitoring strategy based on the multi-media occurrence characteristics of emerging urban pollutants, and collects raw multi-media data through the configured multi-agents. The multi-media occurrence characteristics refer to the features of emerging urban pollutants when they are present in different media carriers. The monitoring strategy includes the monitoring range, monitoring tracking path, monitoring frequency, and monitoring indicators of the agent carriers. Feature extraction is performed on the raw multi-media data to obtain a fused feature matrix, which is then sent to the multi-agents. The multi-agents are configured to construct a source apportionment model based on the fused feature matrix and output source apportionment results. The source apportionment results include pollution source areas and the pollution contribution of each pollution source area. The monitoring strategy is adjusted based on the source apportionment results output by the multi-agents, and the adjusted monitoring strategy is then applied to the multi-agents. The system configures strategies and uses multiple agents to dynamically track and monitor pollution source areas. Based on the dynamic tracking and monitoring results, iterative optimization is performed on the multiple agents. This iterative optimization includes optimizing the model parameters of the source apportionment model and iteratively optimizing the monitoring strategy. After the iterative optimization, the multiple agents are used to apportion new urban pollutant sources. Because this invention uses multiple agents to collect multi-media data and apportion sources of new pollutants, it improves the efficiency and accuracy of new pollutant source apportionment, avoids the lag and limitations of traditional apportionment methods, optimizes the allocation of monitoring resources, reduces the cost losses caused by ineffective monitoring, enhances the adaptability to dynamic changes in pollution, and ensures the long-term effectiveness of the apportionment results. This provides precise targeting for the treatment of new pollutants, helps improve the efficiency of pollution control, and promotes the scientific and efficient governance of urban ecological environment. Attached Figure Description

[0055] 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a schematic diagram of the structure of an urban new pollutant source apportionment device based on multi-agent machine learning, which is part of the hardware operating environment of the embodiment of the present invention.

[0057] Figure 2This is a flowchart illustrating the first embodiment of the urban new pollutant source analysis method based on multi-agent machine learning of the present invention;

[0058] Figure 3 This is a flowchart illustrating the second embodiment of the urban new pollutant source analysis method based on multi-agent machine learning of the present invention;

[0059] Figure 4 This is a flowchart illustrating the third embodiment of the urban new pollutant source analysis method based on multi-agent machine learning of the present invention;

[0060] Figure 5 This is a flowchart illustrating the fourth embodiment of the urban new pollutant source analysis method based on multi-agent machine learning of the present invention;

[0061] Figure 6 This is a structural block diagram of the first embodiment of the urban new pollutant source apportionment device based on multi-agent machine learning of the present invention.

[0062] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0063] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0064] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of an urban new pollutant source apportionment device based on multi-agent machine learning, which is part of the hardware operating environment of the embodiment of the present invention.

[0065] like Figure 1As shown, the urban new pollutant source apportionment device based on multi-agent machine learning may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0066] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on urban new pollutant source apportionment devices based on multi-agent machine learning, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0067] like Figure 1 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and an urban new pollutant source analysis program based on multi-agent machine learning.

[0068] exist Figure 1 In the urban new pollutant source apportionment device based on multi-agent machine learning shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and memory 1005 in the urban new pollutant source apportionment device based on multi-agent machine learning of the present invention can be set in the urban new pollutant source apportionment device based on multi-agent machine learning. The urban new pollutant source apportionment device based on multi-agent machine learning calls the urban new pollutant source apportionment program based on multi-agent machine learning stored in the memory 1005 through the processor 1001 and executes the urban new pollutant source apportionment method based on multi-agent machine learning provided in the embodiments of the present invention.

[0069] This invention provides a method for apportioning emerging urban pollutant sources based on multi-agent machine learning, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the urban new pollutant source analysis method based on multi-agent machine learning of the present invention.

[0070] In this embodiment, the urban emerging pollutant source apportionment method based on multi-agent machine learning includes the following steps:

[0071] Step S10: Configure a multi-agent monitoring strategy based on the multi-media occurrence characteristics of new urban pollutants, and collect data through the multi-agent with the configured strategy to obtain raw multi-media data.

[0072] It should be noted that this embodiment is applied to the detection and analysis of emerging urban pollutants. The core technical problems to be solved in this embodiment are concentrated in five key dimensions: First, the problem of low level of intelligence. Traditional methods lack intelligent decision support, making it difficult to effectively process and analyze massive amounts of data, and thus unable to provide a scientific basis for environmental protection decisions. Second, the problem of insufficient identification ability of emerging pollutants. Existing methods are severely inadequate in identifying emerging pollutants such as pharmaceuticals and personal care products (PPCPs), endocrine disruptors (EDCs), and perfluorinated compounds (PFAS). These pollutants have the characteristics of low-concentration persistence and are widely detected in organisms and the environment. Third, the problem of lack of multi-media collaborative monitoring capability. Traditional methods mostly focus on the monitoring of single environmental media and lack the ability to conduct collaborative monitoring of multiple media such as air, water, and soil. Fourth, the problem of insufficient real-time dynamic analysis capability. Existing technologies mostly adopt offline analysis mode and lack the ability to track and analyze pollution sources in real time. Fifth, the problem of cost-effectiveness imbalance. Traditional methods cannot effectively control costs while ensuring analysis accuracy.

[0073] It should be understood that the executing entity of this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a terminal electronic device capable of realizing the above functions. The following description uses a multi-agent machine learning-based urban new pollutant source apportionment device (apportionment device) as an example to illustrate this embodiment and the following embodiments.

[0074] It should be noted that new urban pollutants can refer to novel chemical substances that pose potential risks to the ecological environment and human health, in addition to traditional pollutants. These include pharmaceuticals, personal care products (PPCPs), endocrine disruptors (EDCs), perfluorinated compounds (PFAS), microplastics, and novel flame retardants.

[0075] It should be noted that the medium carrier can be a carrier of pollutants, such as air, soil, and water. The multi-media occurrence characteristics refer to the characteristics of new urban pollutants when they are present in different medium carriers. For example, multi-media occurrence characteristics refer to the distribution state, concentration level, migration and transformation patterns of new urban pollutants in different medium carriers such as air, water, soil, and sediment.

[0076] It should be noted that a multi-agent system can be a system composed of multiple monitoring carriers with autonomous sensing, data transmission, and collaborative operation capabilities, including fixed ground sensors, drones, mobile ground monitoring vehicles, fixed monitoring stations, and underwater monitoring equipment. The raw multi-media data refers to the original monitoring data related to the new pollutant in each medium directly collected by the multi-agent system through its onboard sensors.

[0077] It should be noted that the monitoring strategy includes the monitoring range, monitoring and tracking path, monitoring frequency, and monitoring indicators of the intelligent agent carrier.

[0078] In some embodiments, the monitoring strategy is the core scheme to guide multiple agents to carry out monitoring work, covering the monitoring range of the agent carrier (such as specific areas such as urban industrial areas, residential areas, river basins, etc.), monitoring tracking path (such as downwind path based on pollution diffusion prediction), monitoring frequency (such as regular 15 minutes / time, 5 minutes / time in high-risk areas) and monitoring indicators (such as auxiliary indicators such as the concentration of specific new pollutants, medium pH value, temperature, etc.).

[0079] It is understandable that the goal of multi-agent data acquisition in this embodiment is to overcome the bottlenecks of traditional monitoring, such as "deployment based on experience, insufficient coverage of multiple media, and poor real-time performance". By using hierarchical distributed data acquisition agents (sensors, drones, mobile monitoring vehicles, etc.), it can achieve "real-time acquisition, edge transmission, and standardized access" of new pollutant data in multiple media such as air, water, and soil, providing a full-dimensional and high-quality data foundation for subsequent source analysis.

[0080] In some embodiments, the analysis device, based on the multi-media occurrence characteristics of new urban pollutants, clarifies the monitoring scope and core indicators of each intelligent agent, ensuring coverage of the entire chain of "atmosphere-water-soil-pollution source", as shown in the table below:

[0081] Table 1, Multi-media data source table

[0082]

[0083] Step S20: Extract features from the original multi-media data to obtain a fused feature matrix.

[0084] Understandably, analytical devices can preprocess raw multi-media data (such as normalization, data cleaning, and data augmentation) to remove redundant and noisy data and extract core information that reflects the pollution characteristics of new pollutants.

[0085] It should be noted that the fusion feature matrix is ​​a matrix-based data structure formed by integrating raw data collected from different media and different agents according to a unified dimension after feature extraction. It contains multi-dimensional key information such as spatiotemporal features, concentration features, and media environment features.

[0086] Step S30: Send the fused feature matrix to the multi-agent, which is configured to construct a source resolution model based on the fused feature matrix and output the source resolution result through the source resolution model.

[0087] It should be noted that the source apportionment results include the pollution source area (the specific spatial range of new pollutant emissions, such as the area surrounding a chemical industrial park or a wastewater treatment plant) and the pollution contribution (the proportion of each pollution source area's impact on the overall pollution). The pollution source area can be a specific spatial region that can be clearly traced and continuously or intermittently emits the target new pollutant; the pollution contribution can be the proportion of new pollutants emitted by each pollution source area in the overall pollution load, and is a core indicator for measuring the degree of pollution source impact.

[0088] In some embodiments, multiple agents collaborate to compute, using the fused feature matrix as input data, and combine the mass conservation principle of traditional receptor models with machine learning algorithms (such as random forests and neural networks) to construct a composite source resolution model. During model training, distributed computing power of multiple agents is used to improve training efficiency, and cross-validation is used to optimize model stability. After training, the fused feature matrix is ​​input, and the model identifies the correspondence between features and pollution sources, outputting the specific location of each pollution source area and its corresponding pollution contribution.

[0089] Step S40: Adjust the monitoring strategy based on the source resolution results output by the multi-agent, configure the multi-agent strategy based on the adjusted monitoring strategy, and dynamically track and monitor the pollution source area through the multi-agent strategy configured.

[0090] It is understood that this embodiment achieves dynamic tracking and detection of pollution sources by dynamically updating and adjusting the monitoring strategy of the multi-agent system. The monitoring strategy is adjusted in real time according to the source apportionment results, enabling the multi-agent system to carry out continuous and targeted monitoring focusing on the pollution source area and pollution diffusion path.

[0091] In some embodiments, based on source apportionment results, the analysis device prioritizes areas with high pollution contribution to the core monitoring scope, expanding the number of agents deployed in these areas; it re-plans monitoring and tracking paths using reinforcement learning algorithms, enabling agents to extend monitoring along the direction of pollution source diffusion (such as downwind of air pollutants or flow direction of water pollutants); it increases the monitoring frequency for high-contribution pollution source areas and supplements monitoring indicators that match the emission characteristics of the pollution source; and it reconfigures multiple agents according to the adjusted monitoring strategy, enabling them to dynamically track the pollution source area and surrounding diffusion areas, continuously collecting pollution change data, thereby achieving precise allocation of monitoring resources, avoiding resource waste, improving the ability to capture dynamic changes in pollution sources, ensuring that monitoring work matches the actual pollution situation in real time, and guaranteeing the timeliness and relevance of monitoring data.

[0092] Step S50: Iteratively optimize the multi-agent system based on the dynamic tracking and monitoring results.

[0093] It should be noted that the iterative optimization includes optimizing the model parameters of the source resolution model and iteratively optimizing the monitoring strategy.

[0094] Understandably, the parsing device can iteratively optimize the source resolution model of the multi-agent system based on the dynamic tracking and monitoring results, and iteratively optimize the monitoring strategy of the multi-agent system, thereby achieving continuous dynamic optimization of the dynamic tracking performance and source resolution accuracy of the multi-agent system.

[0095] In some embodiments, the analysis device collects new data generated by dynamic tracking and monitoring, and detects whether pollution patterns have drifted using KL divergence. If drift is detected, the parameters of the source apportionment model are updated and optimized using an incremental gradient descent algorithm to adapt the model to new pollution characteristics. At the same time, the rationality of the existing monitoring strategy is evaluated based on the new monitoring data, and the boundaries of the monitoring range, the planning logic of the tracking path, the gradient setting of the monitoring frequency, and the selection criteria of the monitoring indicators are adjusted. The above optimization process is repeated to achieve continuous iteration of model parameters and monitoring strategies, so that the source apportionment model can adapt to the dynamic changes of pollution characteristics, maintain long-term analysis accuracy, and avoid the problems of static fixation of the source apportionment model and inaccurate monitoring range coverage.

[0096] Furthermore, in order to continuously optimize the source resolution model and multi-agent strategy, solve the accuracy degradation problem caused by the static fixation of traditional models, and improve the accuracy of pollution resolution, the above step S50 may include:

[0097] Step S501: Obtain the measured pollution concentration in the pollution source area based on the dynamic tracking monitoring results, and calculate the concentration error based on the measured pollution concentration and the predicted pollution concentration;

[0098] Step S502: Compare the source apportionment results with the apportionment results of the pre-constructed control model to obtain the consistency coefficient. The control model is a chemical mass balance model.

[0099] Step S503: Determine the current pollution pattern characteristics based on the source apportionment results, and determine the historical pollution pattern characteristics based on the historical source apportionment results;

[0100] Step S504: Perform pollution drift detection on the current pollution pattern distribution characteristics and the historical pollution pattern characteristics to obtain pollution drift data;

[0101] Step S505: Iteratively optimize the multi-agent system based on the concentration error, the consistency coefficient, and the pollution drift data.

[0102] It is understandable that this embodiment continuously optimizes the source apportionment model and multi-agent strategy through "real-world verification + pollution pattern drift detection" to solve the problem of decreased accuracy caused by the "static solidification" of traditional models (such as increased error after adding pollution sources).

[0103] It should be understood that the analysis equipment can collect measured concentrations from known pollution sources (such as chemical plant emission outlets), calculate the relative error, and refer to the following formula:

[0104]

[0105] in, Indicates relative error. To be qualified, Marked as "parsing error", manual verification is required; This represents the measured concentration of a known pollution source. This represents the concentration of pollution sources predicted through source apportionment results.

[0106] It should be noted that the resolution device can perform cross-method validation to compare resolution results: compare with the results of traditional receptor models (such as the Chemical Mass Balance Model (CMB)) and calculate the consistency coefficient, as shown in the following formula:

[0107]

[0108] in, For the first The source resolution result, For traditional CMB results, The mean, Indicates the number of samples.

[0109] In some embodiments, if the consistency coefficient is greater than 0.85, the source resolution result is considered reliable.

[0110] It should be noted that the analysis equipment can identify pollution drift (such as changes in heavy metal pollution caused by the addition of new energy plants) by quantifying the differences between the current pollution pattern and historical patterns, as shown in the following formula:

[0111]

[0112] in, This represents the KL divergence (relative entropy). Historical distribution (joint distribution of "grid-pollution source" for the previous 30 days). Let x be the current distribution (distribution of the most recent day), and x be the feature-concentration sample; if (Empirical threshold) is determined to be "data distribution drift", triggering multi-agent optimization.

[0113] In practical implementation, the parsing device can fine-tune the model parameters based on drift data, avoiding full retraining, as shown in the following formula:

[0114]

[0115] in, These are the model parameters before and after the update (such as CNN weights and random forest decision trees). (Online learning rate is 0.001 lower than offline training rate to avoid parameter oscillation); The data after drift and the measured labels; The gradient of the loss function;

[0116] In some embodiments, in the edge-cloud collaborative division of labor system, the edge can be configured to perform local verification, calculate a simplified version of KL divergence (core features only), and report to the cloud when drift occurs; the cloud can be configured to perform global verification and online learning, update model parameters and distribute them, and record drift events (for source tracing analysis).

[0117] Step S60: After iterative optimization, analyze the new sources of urban pollutants using a multi-agent system.

[0118] In its implementation, the analysis equipment uses a multi-agent system that has undergone multiple rounds of iterative optimization, combined with an optimized source apportionment model, to conduct comprehensive and systematic source apportionment of new pollutants across the entire city. Through high-quality data collected collaboratively by multiple agents and accurate calculations by the model, it ultimately outputs clear and reliable results on the regional distribution of pollution sources and the pollution contribution of each region.

[0119] Understandably, given the differences in the characteristics of new pollutants across different functional areas of a city, this embodiment needs to address the technical challenges of scenario-based adaptation. New pollutants in residential areas mainly originate from fuel combustion, household waste, and sewage, resulting in high emissions of volatile organic compounds (VOCs), high population density, and a large amount of waste such as household waste and sewage. Commercial areas, including shopping malls, supermarkets, and hotels, experience high pedestrian traffic and traffic congestion, leading to high VOC emissions. These emissions primarily originate from fuel combustion, building decoration, and daily life, and are influenced by seasonality, weather, and holidays. Industrial areas are the main source of new pollutant emissions. Industrial production processes use large amounts of organic solvents, raw materials, and fuels, resulting in significant differences in the types and concentrations of VOCs emitted by different industrial sectors, with industries such as chemicals, pharmaceuticals, and printing exhibiting high emissions. New pollutants in transportation hub areas are dynamic and dispersed. When the proportion of heavy-duty vehicles increases from less than 4% to over 16%, the concentrations of road black carbon (BC) and PM1 increase by approximately 184% and 63%, respectively. The urban heat island effect exacerbates the local accumulation of pollutants, leading to persistently high pollution concentrations in specific areas (such as industrial zones and transportation hubs). Emerging industries, such as new energy vehicle manufacturing, exhibit pollution characteristics characterized by the risk of heavy metal and battery material leaks, requiring close monitoring.

[0120] Based on the above analysis of technical problems, this embodiment systematically solves the fundamental technical defects of traditional methods, such as strong reliance on human experience, limited coverage of monitoring data, insufficient ability to identify new pollutants, low analysis efficiency and accuracy, and high cost, by combining a distributed multi-agent architecture with intelligent machine learning algorithms. This enables efficient and accurate source analysis of new urban pollutants.

[0121] This embodiment configures a multi-agent monitoring strategy based on the multi-media occurrence characteristics of new urban pollutants, and collects data through the configured multi-agents to obtain raw multi-media data. The multi-media occurrence characteristics refer to the features of new urban pollutants when they are present in different media carriers. The monitoring strategy includes the monitoring range, monitoring tracking path, monitoring frequency, and monitoring indicators of the agent carrier. Feature extraction is performed on the raw multi-media data to obtain a fused feature matrix, which is then sent to the multi-agents. The multi-agents are configured to construct a source apportionment model based on the fused feature matrix and output source apportionment results through the model. The source apportionment results include pollution source areas and the pollution contribution of each source area. The monitoring strategy is adjusted based on the source apportionment results output by the multi-agents, and the multi-agents are then monitored based on the adjusted strategy. The system configures strategies and uses multiple agents to dynamically track and monitor pollution source areas. Based on the dynamic tracking and monitoring results, iterative optimization is performed on the multiple agents. This iterative optimization includes optimizing the model parameters of the source apportionment model and iteratively optimizing the monitoring strategy. After the iterative optimization, the multiple agents perform source apportionment of new urban pollutants. Because this embodiment uses multiple agents to collect multi-media data and perform source apportionment of new pollutants, it improves the efficiency and accuracy of new pollutant source apportionment, avoids the lag and limitations of traditional apportionment methods, optimizes the allocation of monitoring resources, reduces the cost losses caused by ineffective monitoring, enhances the adaptability to dynamic changes in pollution, and ensures the long-term effectiveness of the apportionment results. This provides precise targeting for the treatment of new pollutants, helps improve the efficiency of pollution control, and promotes the scientific and efficient governance of the urban ecological environment.

[0122] In some embodiments, the analysis device can address the problems of unexplainable and unclear risks in traditional "black box model" analysis results through interpretability analysis and uncertainty assessment. By quantifying uncertainty and identifying core driving factors, it can improve the interpretability of source analysis results and enhance the practicality of decision-making.

[0123] The analytical device can perform uncertainty assessment through Monte Carlo sampling. The uncertainty assessment process may include:

[0124] First, the error distribution assumption: assume the source resolution contribution error. ;

[0125] in, The mean of the error. The error standard deviation is calculated using historical validation data (such as the mean and standard deviation of E over the past 30 days).

[0126] Second, Monte Carlo sampling: The analytical result C of each grid is sampled M=1000 times, and the sampling error is... The formula is as follows:

[0127]

[0128] Third, uncertainty quantification:

[0129] Set a 95% confidence interval: for Sort, then remove the 2.5th percentile. and 97.5 percentile ,Right now .

[0130] The relative uncertainty is calculated using the following formula:

[0131]

[0132] Level Classification: (High reliability) (Reliable) (Low reliability, requires encrypted monitoring).

[0133] Interpretability analysis may include:

[0134] First, SHAP value decomposition: This involves analyzing the contribution of the solution. The decomposition is divided into "benchmark value + contribution of each feature", as shown in the following formula:

[0135]

[0136] in, The average contribution of all samples. For the first The SHAP value of each feature (positive indicates increased contribution, negative indicates inhibition). For feature values;

[0137] Second, core driving factor identification: Calculate the mean absolute SHAP value of each feature, sort them, and take the top 5 features, as shown in the following formula:

[0138]

[0139] For example, the core driving factor for industrial area grids is: perfluorooctane sulfonic acid concentration. Chemical emission intensity ”;

[0140] Third, pollution source tracing path: Based on spatial SHAP values, draw a path map of "pollution source → affected grid", such as "chemical enterprise (SHAP value = 1.2) → downstream water grid (SHAP value = 0.9) → surrounding soil grid (SHAP value = 0.6)", to clarify the pollution diffusion chain.

[0141] In the edge-cloud collaborative division of labor system, the edge can be configured to output a simplified version of the results (the top 3 driving factors of the local grid and the 95% confidence interval); the cloud can be configured to generate a full report (global driving factors and cross-regional tracing paths) for macro-level decision-making.

[0142] In some embodiments, the analysis device can transform source analysis results into "executable, scenario-specific" pollution prevention and control decision-making suggestions, realizing a closed loop of "analysis-decision-prevention" and solving the problem of the disconnect between traditional results and actual prevention and control.

[0143] First, generate a pollution source inventory report:

[0144] Pollution sources are sorted in descending order of contribution, generating a structured list. Example output:

[0145]

[0146] Secondly, functional zone-based prevention and control recommendations:

[0147] Based on the pollution characteristics and core driving factors of functional zones, targeted recommendations are generated, using a "feature-recommendation" mapping function:

[0148]

[0149] For example, a specific mapping may include:

[0150] Residential areas (Key Factor: PPCPs / EDCs): Recommendations include "Optimize domestic wastewater treatment processes (increase PPCP removal rate to 85%) and strengthen waste sorting";

[0151] Commercial Area (Key Factor: VOCs): Recommendations include "Restricting renovation hours (9:00-17:00) and promoting low-VOCs building materials (VOCs emission <100g / L)";

[0152] Industrial Zone (Key Factor: PFAS / Heavy Metals): Recommendations include "Installing waste gas treatment equipment (PFAS removal rate > 90%) and carrying out soil remediation (reducing heavy metal content to GB15618-2018 standard)".

[0153] Transportation hubs (Key Factor: PM2.5 / black carbon): The generated recommendations include "Promote new energy vehicles (increase the proportion to 30%) and optimize traffic management (reduce idling time by 20%)".

[0154] Finally, a tiered push notification mechanism:

[0155] The formula for matching push notification channels based on alert level and user role is as follows:

[0156]

[0157] For example, the push notification content specifications may include "warning level, affected area, core driving factors, recommended measures, and contact person".

[0158] In the edge-cloud collaborative division of labor system, the edge can be configured to generate localized suggestions (such as community sewage management) and push them to grassroots environmental protection stations and enterprises; the cloud can be configured to generate cross-regional macro suggestions and push them to environmental protection departments or public platforms.

[0159] In some embodiments, the analysis device can continuously improve the effectiveness of multi-agent data collection strategies, source analysis models, and decision recommendations through a closed loop of "decision feedback-process optimization," thereby reducing system costs and ensuring long-term adaptability to urban governance needs. Feedback analysis is achieved through prevention and control effectiveness analysis and cost-benefit analysis.

[0160] First, the analysis of the prevention and control effectiveness includes calculating the concentration reduction rate after the implementation of the decision recommendations, using the following formula:

[0161]

[0162] Secondly, the cost-benefit analysis includes calculating the ratio of multi-agent operation and maintenance costs to pollution control benefits, as shown in the following formula:

[0163]

[0164] Multi-agent end-to-end optimization strategies may include:

[0165] (1) Data acquisition optimization: Adjust the agent density based on monitoring efficiency, as shown in the following formula:

[0166]

[0167] in, (Low efficiency) → Increase agent density; (High efficiency) → Reduced density;

[0168] (2) Model optimization: If If the prevention and control effect is poor, online learning is triggered (see Part VI). The model feature weights are adjusted, such as adding the "prevention and control measures" feature with a weight of 0.1.

[0169] (3) Decision optimization: Based on the user adoption rate, the suggestions are revised, and the formula is as follows:

[0170]

[0171] Among them, if If so, a simplification suggestion will be made.

[0172] refer to Figure 3 , Figure 3 This is a flowchart illustrating the second embodiment of the urban new pollutant source analysis method based on multi-agent machine learning of the present invention.

[0173] Based on the first embodiment described above, in this embodiment, step S20 further includes:

[0174] Step S201: Standardize the original multi-media data to obtain initial multi-media data.

[0175] It is understandable that this embodiment establishes a unified standard for "time-space-unit-format" for heterogeneous data collected by multiple agents, eliminating data barriers, specifically including:

[0176] 1. Time unification algorithm: All data is marked with a UTC+8 timestamp and aligned to a 15-minute / hourly anchor point. A deviation greater than 3 minutes triggers data retransmission. The formula is as follows:

[0177]

[0178] in, This is the original timestamp (unit: seconds). The target time granularity is (15 minutes = 900 seconds, 1 hour = 3600 seconds). For the aligned timestamp;

[0179] Logical explanation: Rounding down ensures that all data is anchored to a unified time point, avoiding analysis errors caused by time deviations.

[0180] 2. Spatial Unification Algorithm: Using the WGS84 coordinate system, all monitoring point data are mapped to a 500m×500m grid, and a unique grid number is assigned. The formula is as follows:

[0181]

[0182] in, The coordinates of the monitoring point after latitude and longitude conversion (unit: meters). for Total number of grid cells in each direction A unique identifier for the grid;

[0183] Logical explanation: Converting discrete point data into continuous grid data lays the foundation for subsequent multi-media spatial fusion.

[0184] 3. Standardized Unit Algorithm: The concentration of new pollutants is standardized to a uniform unit to avoid the impact of differences in magnitude on the analysis. The formula is as follows:

[0185]

[0186] in, This refers to the original concentration (e.g., ng / L, nmol / L). The unit conversion factor is 0.001 (e.g., ng / L → μg / L). Target concentration (μg / m³ for atmosphere, μg / L for water / soil);

[0187] Logical explanation: Ensure cross-media data comparability, for example, convert PFOA concentration in water from ng / L to μg / L to be on the same order of magnitude as PFOA concentration in soil.

[0188] In practical implementation, the analysis device can build an edge-cloud collaborative division of labor system, which includes an edge terminal and a cloud terminal. The edge terminal is used for data acquisition agents to directly access edge nodes, complete format conversion, timestamp calibration and preliminary filtering (removing obviously invalid data, such as negative concentration values), and control the data transmission delay to <10s. The cloud terminal is used to receive standardized data from each edge node, establish a multi-media data library, and synchronously distribute acquisition strategies (such as encrypting the soil monitoring frequency in industrial areas from 12 hours / time to 6 hours / time).

[0189] Step S202: Preprocess the initial multi-media data to obtain candidate multi-media data.

[0190] It should be noted that the preprocessing includes data cleaning and data augmentation to ensure the reliability of the source parsing data.

[0191] In practice, the analysis device removes outliers and fills in missing values ​​in the multi-agent data through data cleaning, and expands the training sample size through data augmentation to ensure the "completeness (missing value rate <3%), accuracy (outlier misclassification rate <0.2%), and usability (adapting to machine learning model input)" of the preprocessed data.

[0192] Data cleaning specifically includes:

[0193] 1. Outlier Removal: Multi-agent collaborative verification method, combined with traditional 3D modeling... The principle is to correlate data from multiple media sources to avoid misjudgments caused by the failure of a single sensor. The steps are as follows:

[0194] The first step is to calculate the statistics: for single-agent time series data, calculate the mean. with standard deviation :

[0195]

[0196] in, For the data sample size, For the first The monitoring values ​​at each moment;

[0197] The second step is to mark suspected anomalies: if the data meets the following criteria... It is marked as a suspected anomaly.

[0198] The third step is multi-media collaborative verification: Data from adjacent grid cells within the same medium or across media is retrieved. If there are no anomalies in the adjacent data and no collaborative fluctuations in the cross-media data, it is determined to be a genuine anomaly, and a weighted average replacement is used.

[0199]

[0200] in, , These are the two normal data points preceding the abnormal time. The data is the first normal data point after the abnormal event, and its weight decays over time to ensure that recent data has a greater impact.

[0201] 2. Missing Measure Completion: Multi-media collaborative Kalman filtering is employed to address missing measurements caused by sensor offline operation or severe weather. Multi-media data is introduced to assist in completion. The core is to establish a state-space model and iteratively optimize it. The steps are as follows:

[0202] The first step, the state equation and the observation equation:

[0203] Equation of state (describing the law of concentration change):

[0204] Observation equation (correlated cross-media observations):

[0205] in, Let A be the state vector at time k (the pollutant concentration to be completed), A=0.98 (short-term continuity coefficient of concentration), and B=0.02 (contribution weight of cross-media data). This is the mean of cross-media observations. (Process noise, Q=0.01) For the observed value, H=1 (the observation directly corresponds to the state). (Observation noise, R=0.04);

[0206] The second step is the Kalman filter iteration:

[0207] Prediction step: Infer the current state and covariance based on historical states, referring to the following formula:

[0208]

[0209] Update step: Correct the prediction results based on the observed values, referring to the following formula:

[0210]

[0211] in, To predict the state, To predict covariance, Kalman gain (balancing prediction and observation weights). This is the updated state (i.e., missing test completion value).

[0212] Data augmentation specifically includes: spatiotemporal feature expansion, which addresses the issue of limited sample sizes for new pollutants by expanding the data in two ways:

[0213] Time series sliding window: This method splits 24 hours of 15-minute data (96 points in total) into multiple 6-hour subsequences, using the following formula:

[0214]

[0215] Logical explanation: Each subsequence contains 6 data points at the 15-minute level, the sample size is expanded by 4 times, and the continuity of the time dimension is preserved;

[0216] Feature normalization: Min-Max normalization (mapping to the [0,1] interval) is performed on multi-media features to avoid the impact of magnitude differences on model training. The formula is as follows:

[0217]

[0218] in, These are the original eigenvalues. , These are the minimum and maximum values ​​of the feature, respectively.

[0219] In practical implementation, the parsing device can construct an edge-cloud collaborative system, which includes an edge terminal and a cloud terminal. The edge terminal performs outlier removal and short-term missing data completion (<2 hours of missing data) on locally collected data, and outputs preprocessed data to the cloud terminal. The cloud terminal performs multi-media collaborative completion for long-term missing data across regions (such as soil monitoring missing data caused by large-scale rainfall), generates a global preprocessing report, and sends outlier judgment thresholds (such as those for industrial areas) to the edge terminal. Adjusted to 1.2 times the normal value).

[0220] Step S203: Perform spatiotemporal alignment on the candidate multi-media data, and fuse the spatiotemporally aligned candidate multi-media data based on the multi-media storage features to obtain a fused feature matrix.

[0221] It is understandable that, in order to solve the problem of "inconsistent spatiotemporal scales and weak feature correlation" in multi-media data of atmosphere, water and soil, this embodiment unifies the preprocessed data to a spatiotemporal coordinate system of "500m grid + 15-minute level", and integrates multi-media features to generate a model input matrix of "high correlation and low redundancy".

[0222] Furthermore, to improve the accuracy and efficiency of feature extraction, step S203 may include:

[0223] Step S2031: Determine the time alignment anchor point information and spatial alignment grid information based on the monitoring frequency and monitoring range of the candidate multi-media data;

[0224] Step S2032: Construct a spatiotemporal coordinate system based on the time alignment anchor information and the spatial alignment grid information, and map the candidate multi-media data to the spatiotemporal coordinate system to perform spatiotemporal alignment on the candidate multi-media data;

[0225] Step S2033: Analyze the source resolution contribution of each medium based on the multi-media storage characteristics, and perform feature weighted fusion on the spatiotemporally aligned candidate multi-media data based on the source resolution contribution to obtain an initial fusion matrix;

[0226] Step S2034: Calculate the mutual information value between each feature in the initial fusion matrix and the pollution source type label, and perform feature filtering on the initial fusion matrix based on the mutual information value to obtain the fusion feature matrix.

[0227] In the specific implementation, time alignment is performed: using 15-minute anchor points for atmospheric data as a benchmark, low-frequency data (water bodies every hour / time, soil every 12 hours / time) are interpolated and resampled.

[0228] Hourly to 15-minute intervals (linear interpolation):

[0229]

[0230] in, , For adjacent hourly data, The time points are in 15-minute increments, and 900 represents seconds.

[0231] 12-hour timescale → 15-minute timescale (cubic spline interpolation): Let... Let the pollutant concentration be a time function that is twice continuously differentiable. The interpolation formula is:

[0232]

[0233] Logical explanation: Preserve the long-term trend of soil data and avoid distortion caused by interpolation.

[0234] Spatial alignment: Multi-media point data is converted into 500m grid data, and inverse distance weighted (IDW) interpolation is used, as shown in the following formula:

[0235]

[0236] in, For grid center concentration value, Let i be the concentration at the i-th monitoring point. (Euclidean distance from the grid center to the monitoring point) (Distance attenuation coefficient, closer points have higher weight);

[0237] Logical explanation: Ensure that the grid data can reflect the spatial gradient of the monitoring points, such as the concentration of the grid around the chemical industrial park decreasing with distance.

[0238] It is understood that this embodiment can achieve multi-media feature fusion by constructing a weighted feature matrix. The feature fusion specifically includes: allocating weights based on the differences in the contribution of multiple media to source resolution using the entropy weight method, and constructing a fused feature matrix. ( =Time step, =Number of grids, =Fusion feature number), the steps are as follows:

[0239] (1) Feature classification, including:

[0240] Atmospheric characteristics VOCs concentration, PM2.5-related microplastic content ( );

[0241] Water characteristics PPCPs concentration, EDCs concentration ( );

[0242] Soil characteristics Microplastic abundance, heavy metal content ( );

[0243] auxiliary features Meteorological data (wind speed / wind direction), functional zone identification (residential area = 1 / industrial area = 2) );

[0244] (2) Calculate the weights using the entropy weight method:

[0245] Step 1: Feature normalization (dimension elimination), refer to the following formula:

[0246]

[0247] in, For the first The first sample The proportion of each feature;

[0248] Step 2: Calculate the feature entropy value (the smaller the entropy value, the richer the information), referring to the following formula:

[0249]

[0250] Step 3: Calculate the feature weights (the smaller the entropy value, the larger the weight), referring to the following formula:

[0251]

[0252] (3) Fusion formula:

[0253]

[0254] Weight values: Optimized based on measured data from the Yangtze River Delta region. (atmosphere), (Water bodies) (Increased weighting of soil and industrial zones) (Auxiliary features), satisfying .

[0255] Understandably, feature selection involves variance thresholding combined with mutual information to eliminate redundant features and retain core, relevant features. Specifically, this includes:

[0256] (1) Variance threshold method: Set a threshold of 0.01 to remove low-information features with variance < 0.01 (such as background value features that have not fluctuated for a long time).

[0257] (2) Mutual Information Method: Calculate the mutual information value between the remaining features and the "Pollution Source Type Label", and retain the highly correlated features with mutual information > 0.2. The formula is as follows:

[0258]

[0259] in, Features With tags The joint probability, Marginal probability;

[0260] Logical Explanation: To ensure that the selected features have strong explanatory power for pollution source identification, the final feature dimensions are... .

[0261] In some embodiments, in the edge-cloud collaborative division of labor, the edge is used to complete the spatiotemporal alignment and local feature fusion of a small region (a single administrative region), outputting a simplified fusion matrix (F=12) with a delay control of <30s; the cloud is used to perform cross-regional global fusion, optimize feature weights (such as increasing the water feature weights for PFAS pollution events), generate a full fusion matrix (F=20), and send the weight parameters to the edge.

[0262] This embodiment standardizes the original multi-media data to obtain initial multi-media data, preprocesses the initial multi-media data to obtain candidate multi-media data (including data cleaning and data augmentation), performs spatiotemporal alignment on the candidate multi-media data, and fuses the spatiotemporally aligned candidate multi-media data based on the multi-media storage characteristics to obtain a fused feature matrix. Because this embodiment achieves efficient integration and quality improvement of multi-media, multi-source raw data through a full-process data processing approach encompassing standardization, preprocessing, spatiotemporal alignment, and feature fusion, its macroscopic effects are: improved data processing efficiency and quality; elimination of analytical obstacles caused by data heterogeneity and spatiotemporal misalignment; and enhanced correlation of multi-media contamination features, providing high-quality, high-value feature input for subsequent core applications such as source analysis.

[0263] refer to Figure 4 , Figure 4 This is a flowchart illustrating the third embodiment of the urban new pollutant source analysis method based on multi-agent machine learning of the present invention.

[0264] Based on the above embodiments, in this embodiment, the multi-agent constructs a source resolution model based on the fused feature matrix, and outputs source resolution results through the source resolution model, including:

[0265] Step S31: Perform data dimensionality reduction and data partitioning on the received fusion feature matrix to obtain model input data.

[0266] It should be noted that, in this embodiment, the multiple agents may include: a data processing agent, a source resolution agent, a decision support agent, and a feedback optimization agent;

[0267] The data processing agent receives the fused feature matrix, performs dimensionality reduction (retaining 95% of the information) and data partitioning (70% for training set, 20% for validation set, and 10% for test set), and outputs the input data adapted to the model.

[0268] Source resolution agents are used to integrate multi-model collaborative resolution. The core algorithms include CNN (spatial features), RNN (temporal features), random forest (type classification), and reinforcement learning (policy optimization).

[0269] The decision support agent receives source analysis results and generates a three-dimensional report of "grid - pollution source - contribution".

[0270] The feedback optimization agent is used to dynamically adjust model parameters based on the validation results.

[0271] Step S32: Construct the original machine learning model.

[0272] It should be noted that the original machine learning model includes a CNN spatial feature extraction module, an RNN temporal feature extraction module, a random forest pollution source classification module, and a contribution quantification module. The CNN spatial feature extraction module is configured to use multi-layer convolutional kernels to capture spatial correlation features in the data. The RNN temporal feature extraction module is configured to use LSTM units to capture temporal fluctuations in pollution concentration. The random forest pollution source classification module is configured to classify pollution source types through voting by multiple decision trees. The contribution quantification module is configured to quantify the contribution of pollution sources.

[0273] In its implementation, the CNN spatial feature extraction module uses 3×3 convolutional kernels to capture the spatial correlation of grid data (such as the pollution gradient between an industrial area and its surroundings), as shown in the following formula:

[0274]

[0275] in, This represents the spatial feature map of the output. For the first Each convolutional kernel weight, This is the feature map of the previous layer. For bias, for Activation function ( ), (Number of convolution kernels);

[0276] Output: Spatial feature map ( (spatial feature dimension).

[0277] The RNN time feature extraction module uses LSTM units to avoid gradient vanishing and captures temporal fluctuations in pollution concentration (such as VOCs peak values ​​during morning traffic rush hours).

[0278] The random forest pollution source classification module classifies pollution sources (industrial / domestic / transportation emissions) through voting using multiple decision trees, as shown in the following formula:

[0279]

[0280] in, This indicates the final classification result (predicted pollution source type, such as industrial / residential / transportation emissions). Indicates the pollution source type category index (maximizing the voting results): (Number of decision trees) For the first The predicted results for each tree. This is an indicator function (it takes the value 1 if the condition is met, and 0 otherwise). This is the final classification result.

[0281] The contribution quantification module can be constructed based on the CMB-PMF fusion algorithm, combining the advantages of the Chemical Mass Balance (CMB) method and Positive Matrix Factorization (PMF), as shown in the following formula:

[0282]

[0283] in, (Spatiotemporal joint characteristics) The CMB weights are based on the known pollution source composition profiles. PMF weights (unsupervised identification of unknown sources). (Integration coefficient), C represents the contribution of pollution sources (%).

[0284] Step S33: Construct the joint loss function.

[0285] It should be noted that, taking into account both the classification task (pollution source type) and the regression task (contribution), the mathematical expression of the joint loss function is as follows:

[0286]

[0287] in, Denotes the joint loss function. This represents the classification loss weight (e.g., 0.6). This represents the regression loss weight (e.g., 0.4). This represents the classification loss term (used to optimize pollution source type identification). The true label indicating the type of pollution source Predictive labels indicating the type of pollution source This represents the regression loss term (used to optimize the calculation of pollution source contribution). This represents the true value of the contribution of pollution sources. This represents the predicted value of the contribution of pollution sources.

[0288] Step S34: Train the original machine learning model based on the model input data and the joint loss function, and optimize the parameters of the original machine learning model using the Adam optimizer to obtain the source resolution model.

[0289] It should be noted that the Adam optimizer is used for adaptive learning rate and converges quickly, as shown in the following formula:

[0290]

[0291] in, , (Momentum parameter) The gradient of the loss function. (Initial learning rate, decaying by 10% every 10 rounds) (To avoid a denominator of 0) These are the model parameters (convolution kernels, decision tree nodes, etc.).

[0292] Step S35: Output the source resolution result through the source resolution model.

[0293] In some embodiments, in the edge-cloud collaborative division of labor system, the edge can be configured to deploy a lightweight model (simplified CNN + random forest) to complete local grid fast parsing (latency < 30s) and focus on outputting high-contribution pollution sources (C > 50%); the cloud can be configured to deploy a full model (CNN + RNN + reinforcement learning) to perform cross-regional source parsing (such as the Yangtze River Delta urban agglomeration), update model parameters with full data daily, and distribute them to the edge.

[0294] In this embodiment, multiple agents perform data dimensionality reduction and data partitioning on the received fused feature matrix to obtain model input data, construct an original machine learning model, construct a joint loss function, train the original machine learning model based on the model input data and the joint loss function, and optimize the parameters of the original machine learning model through the Adam optimizer to obtain a source resolution model. The source resolution model outputs source resolution results, thereby realizing multi-agent collaboration, integrating deep learning, machine learning and reinforcement learning algorithms to achieve accurate identification of new urban pollutant types, effectively trace pollution sources, and significantly reduce resolution time and improve source resolution efficiency.

[0295] refer to Figure 5 , Figure 5 This is a flowchart illustrating the fourth embodiment of the urban new pollutant source analysis method based on multi-agent machine learning of the present invention.

[0296] Based on the above embodiments, in this embodiment, step S40 may include:

[0297] Step S401: Based on the source resolution results output by the multi-agent, optimize the monitoring and tracking path in the monitoring strategy to obtain the path optimization result.

[0298] It is understood that this embodiment, based on source apportionment results, uses multi-agent collaborative adjustment of monitoring strategies to achieve real-time tracking (100m accuracy) and dynamic early warning of mobile pollution sources (industrial leaks, traffic pollution), thus solving the problem of lag in traditional static monitoring.

[0299] In its implementation, this embodiment employs reinforcement learning-driven path planning, using drones and mobile monitoring vehicles as physical carriers. It optimizes the tracking path through Q-learning, with the core being the iterative updating of the Q-value (state-action value), as shown in the following path optimization formula:

[0300]

[0301] in, Indicates the state Next action The expected cumulative reward, Indicates the learning rate. Indicates the execution of an action The instant reward obtained afterward Indicates the discount factor. Indicates the execution of an action The new state afterward Indicates a new state All available actions Indicates a new state The maximum expected cumulative reward that can be obtained from all optional actions.

[0302] In some embodiments, the analysis device can update the Q value and tracking target every 15 minutes to ensure that the agent prioritizes tracking high-contribution pollution sources, such as chemical leaks.

[0303] Step S402: Configure the contribution threshold of each pollution source region based on the functional area type of each pollution source region in the source apportionment results.

[0304] It is understood that this embodiment can set a contribution threshold based on the functional area type to trigger a tiered warning, as shown in the following formula:

[0305]

[0306] in, Functional zone threshold (industrial zone) Residential areas Transportation hubs , Contribution to the grid.

[0307] In some embodiments, a yellow alert notifies the local environmental protection bureau, an orange alert dispatches multi-agent encrypted monitoring (the frequency is increased from 15 minutes / time to 5 minutes / time), and a red alert initiates an emergency response (such as enterprise shutdown or vehicle restriction).

[0308] Step S403: Based on meteorological data and the source apportionment results, predict the pollution diffusion path and obtain the pollution diffusion path prediction results.

[0309] It should be understood that this embodiment can predict pollution diffusion paths by combining meteorological data (wind speed / wind direction), referring to the following pollution diffusion path prediction formula:

[0310]

[0311] in, Indicates that in spatial coordinates The time is Predicted pollution concentration at that time Indicates source strength ( The source strength is calculated based on the pollution contribution of each pollution source region in the source apportionment results. Indicates average wind speed ( ), This indicates the effective source height, which is the effective emission height of the pollution source. Indicates the horizontal diffusion parameter. This represents the vertical diffusion parameter (calculated based on atmospheric stability, under unstable conditions). (increase) For spatial coordinates, Represents the horizontal coordinates perpendicular to the downwind direction. Represents vertical coordinates, Indicates the distance downwind;

[0312] In some embodiments, the analysis device can generate a heat map of pollution concentration distribution for the next 24 hours based on the pollution diffusion path prediction results output by the pollution diffusion path prediction formula, such as a chemical leak → downwind direction. Grid → Peak concentration after hours ³.

[0313] Step S404: Adjust the monitoring strategy based on the path optimization results, the contribution threshold, and the pollution diffusion path prediction results.

[0314] In some embodiments, in the edge-cloud collaborative division of labor system, the edge can be configured to calculate the contribution of the local grid in real time, trigger yellow and orange warnings, and schedule surrounding mobile intelligent agents to track; the cloud can be configured to aggregate cross-regional warning information, trigger red warnings, generate a global pollution spread map, and coordinate prevention and control resources in multiple cities.

[0315] This embodiment optimizes the monitoring and tracking paths in the monitoring strategy based on the source apportionment results output by the multi-agent, obtaining path optimization results. It configures the contribution thresholds for each pollution source region based on the functional zone type in the source apportionment results, predicts pollution diffusion paths based on meteorological data and the source apportionment results, obtaining pollution diffusion path prediction results. The monitoring strategy is then adjusted based on the path optimization results, the contribution thresholds, and the pollution diffusion path prediction results. Because this embodiment optimizes the monitoring and tracking paths to dynamically adjust monitoring routes, reduce ineffective path consumption, and improve the utilization efficiency of monitoring resources, and configures the contribution thresholds for each pollution source region to achieve differentiated and accurate determination of pollution contributions from different functional zones, it improves the reliability of source apportionment results. By predicting pollution diffusion paths, it enables advance prediction of the direction and scope of pollution propagation, enhancing the timeliness and proactivity of monitoring, thereby significantly improving the accuracy, efficiency, and adaptability of the entire pollution monitoring and source apportionment system.

[0316] Furthermore, embodiments of the present invention also propose a computer-readable storage medium storing a program for analyzing new urban pollutant sources based on multi-agent machine learning. When the program is executed by a processor, it implements the steps of the method for analyzing new urban pollutant sources based on multi-agent machine learning as described above.

[0317] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0318] The aforementioned computer-readable storage medium may be included in an urban new pollutant source apportionment device based on multi-agent machine learning; or it may exist independently and not be assembled into an urban new pollutant source apportionment device based on multi-agent machine learning.

[0319] Furthermore, this invention also proposes a computer program product, including a multi-agent machine learning-based urban new pollutant source analysis program, which, when executed by a processor, implements the steps of the multi-agent machine learning-based urban new pollutant source analysis method described above.

[0320] The specific implementation of the computer program product of the present invention is basically the same as the embodiments of the above-described method for analyzing new urban pollutant sources based on multi-agent machine learning, and will not be repeated here.

[0321] Reference Figure 6 , Figure 6 This is a structural block diagram of the first embodiment of the urban new pollutant source apportionment device based on multi-agent machine learning of the present invention.

[0322] like Figure 6 As shown, the urban new pollutant source apportionment device based on multi-agent machine learning proposed in this embodiment of the invention includes:

[0323] The multi-agent data acquisition module 10 is used to configure the monitoring strategy of the multi-agent based on the multi-media storage characteristics of new urban pollutants, and to collect data through the multi-agent after the strategy configuration to obtain the original multi-media data. The multi-media storage characteristics are the characteristics of new urban pollutants when they are stored in different media carriers. The monitoring strategy includes the monitoring range, monitoring tracking path, monitoring frequency and monitoring indicators of the agent carrier.

[0324] Multi-media feature extraction module 20 is used to extract features from the original multi-media data to obtain a fused feature matrix;

[0325] The multi-agent machine learning source resolution module 30 is used to send the fused feature matrix to the multi-agent, which is configured to build a source resolution model based on the fused feature matrix and output source resolution results through the source resolution model. The source resolution results include pollution source areas and the pollution contribution of each pollution source area.

[0326] The pollution dynamic tracking module 40 is used to adjust the monitoring strategy based on the source analysis results output by the multi-agent, configure the multi-agent strategy based on the adjusted monitoring strategy, and dynamically track and monitor the pollution source area through the multi-agent strategy configured with the strategy.

[0327] The multi-agent optimization module 50 is used to iteratively optimize the multi-agent based on the dynamic tracking and monitoring results. The iterative optimization includes optimizing the model parameters of the source resolution model and iteratively optimizing the monitoring strategy.

[0328] The new pollutant analysis module 60 is used to analyze new urban pollutant sources through a multi-agent system after iterative optimization.

[0329] This embodiment configures a multi-agent monitoring strategy based on the multi-media occurrence characteristics of new urban pollutants, and collects data through the configured multi-agents to obtain raw multi-media data. The multi-media occurrence characteristics refer to the features of new urban pollutants when they are present in different media carriers. The monitoring strategy includes the monitoring range, monitoring tracking path, monitoring frequency, and monitoring indicators of the agent carrier. Feature extraction is performed on the raw multi-media data to obtain a fused feature matrix, which is then sent to the multi-agents. The multi-agents are configured to construct a source apportionment model based on the fused feature matrix and output source apportionment results through the model. The source apportionment results include pollution source areas and the pollution contribution of each source area. The monitoring strategy is adjusted based on the source apportionment results output by the multi-agents, and the multi-agents are then monitored based on the adjusted strategy. The system configures strategies and uses multiple agents to dynamically track and monitor pollution source areas. Based on the dynamic tracking and monitoring results, iterative optimization is performed on the multiple agents. This iterative optimization includes optimizing the model parameters of the source apportionment model and iteratively optimizing the monitoring strategy. After the iterative optimization, the multiple agents perform source apportionment of new urban pollutants. Because this embodiment uses multiple agents to collect multi-media data and perform source apportionment of new pollutants, it improves the efficiency and accuracy of new pollutant source apportionment, avoids the lag and limitations of traditional apportionment methods, optimizes the allocation of monitoring resources, reduces the cost losses caused by ineffective monitoring, enhances the adaptability to dynamic changes in pollution, and ensures the long-term effectiveness of the apportionment results. This provides precise targeting for the treatment of new pollutants, helps improve the efficiency of pollution control, and promotes the scientific and efficient governance of the urban ecological environment.

[0330] The urban new pollutant source apportionment device based on multi-agent machine learning provided in this application employs the urban new pollutant source apportionment method based on multi-agent machine learning in the above embodiments, and can solve the technical problems of urban new pollutant source apportionment based on multi-agent machine learning. Compared with the prior art, the beneficial effects of the urban new pollutant source apportionment device based on multi-agent machine learning provided in this application are the same as the beneficial effects of the urban new pollutant source apportionment method based on multi-agent machine learning provided in the above embodiments, and other technical features in the urban new pollutant source apportionment device based on multi-agent machine learning are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0331] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0332] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0333] In addition, for technical details not described in detail in this embodiment, please refer to the urban new pollutant source analysis method based on multi-agent machine learning provided in any embodiment of the present invention, which will not be repeated here.

[0334] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0335] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0336] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0337] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for resolving urban new pollutant sources based on multi-agent machine learning, characterized in that, The urban new pollutant source resolution method based on multi-agent machine learning comprises: a monitoring strategy of the multi-agent is configured based on multi-medium occurrence characteristics of the urban new pollutant, and original multi-medium data are obtained through the multi-agent after the strategy configuration, the multi-medium occurrence characteristics are distribution states, concentration levels and migration and transformation rules of the urban new pollutant in different medium carriers such as air, water, soil and sediment, the monitoring strategy comprises a monitoring range, a monitoring tracking path, a monitoring frequency and a monitoring index of the agent carrier, and the original multi-medium data are original monitoring data related to the new pollutant in each medium directly collected by the multi-agent through a sensor carried by the multi-agent; feature extraction is performed on the original multi-medium data to obtain a fusion feature matrix; the fusion feature matrix is sent to the multi-agent, the multi-agent is configured to construct a source resolution model based on the fusion feature matrix, and a source resolution result is output through the source resolution model, the source resolution result comprises a pollution source area and a pollution contribution degree of each pollution source area; the monitoring strategy is adjusted based on the source resolution result output by the multi-agent, the multi-agent is configured based on the adjusted monitoring strategy, and dynamic tracking monitoring is performed on the pollution source area through the multi-agent after the strategy configuration; the multi-agent is iteratively optimized based on the dynamic tracking monitoring result, the iterative optimization comprises model parameter optimization of the source resolution model and iterative optimization of the monitoring strategy; the multi-agent after the iterative optimization is used for urban new pollutant source resolution; the process that the multi-agent constructs a source resolution model based on the fusion feature matrix and outputs a source resolution result through the source resolution model comprises: data dimension reduction and data division are performed on the received fusion feature matrix to obtain model input data; an original machine learning model is constructed, the original machine learning model comprises a CNN spatial feature extraction module, an RNN time feature extraction module, a random forest pollution source classification module and a contribution quantification module, the CNN spatial feature extraction module is configured to capture spatial correlation features in data by using multiple convolution kernels, the RNN time feature extraction module is configured to capture time fluctuations of pollution concentration by using LSTM units, the random forest pollution source classification module is configured to realize pollution source type classification through voting of multiple decision trees, and the contribution quantification module is configured to quantify pollution source contribution degrees; a joint loss function is constructed, a mathematical expression of the joint loss function is: wherein, denotes a joint loss function, denotes a classification loss weight, denotes a regression loss weight, denotes a classification loss term, denotes a true label of a pollution source type, denotes a predicted label of a pollution source type, denotes a regression loss term, denotes a true value of a pollution source contribution, denotes a predicted value of a pollution source contribution; the original machine learning model is trained based on the model input data and the joint loss function, and parameter optimization of the original machine learning model is performed through an Adam optimizer to obtain a source resolution model; a source resolution result is output through the source resolution model; the process that the monitoring strategy is adjusted based on the source resolution result output by the multi-agent comprises: a pollution diffusion path is predicted based on meteorological data and the source resolution result to obtain a pollution diffusion path prediction result, and a prediction formula of the pollution diffusion path is: wherein, represents a predicted pollution concentration at a spatial coordinate , and a time , represents a source strength, which is calculated based on pollution contribution degrees of each pollution source region in source resolution results, represents an average wind speed, represents an effective source height, which is an effective emission height of a pollution source, represents a horizontal diffusion parameter, represents a vertical diffusion parameter, represents a horizontal coordinate perpendicular to a downwind direction, represents a vertical coordinate, represents a downwind distance; Adjust the monitoring strategy based on the pollution diffusion path prediction result.

2. The multi-agent machine learning based urban new pollutant source resolution method according to claim 1, wherein, The feature extraction on the original multi-medium data includes: The original multi-medium data is standardized to obtain initial multi-medium data. The initial multi-medium data is preprocessed to obtain candidate multi-medium data, and the preprocessing includes data cleaning and data enhancement. The candidate multi-medium data is spatio-temporally aligned, and the spatio-temporally aligned candidate multi-medium data is feature fused based on the multi-medium occurrence characteristics to obtain a fusion feature matrix. 3.The multi-agent machine learning based urban new pollutant source resolution method according to claim 2, wherein, The spatio-temporal alignment of the candidate multi-medium data and the feature fusion of the spatio-temporally aligned candidate multi-medium data based on the multi-medium occurrence characteristics to obtain a fusion feature matrix includes: Determine the time alignment anchor point information and the spatial alignment grid information based on the monitoring frequency and monitoring range of the candidate multi-medium data; Construct a spatio-temporal coordinate system based on the time alignment anchor point information and the spatial alignment grid information, and map the candidate multi-medium data to the spatio-temporal coordinate system to align the candidate multi-medium data in space and time; Analyze the source analysis contribution degree of each medium based on the multi-medium occurrence characteristics, and perform feature weighted fusion on the spatio-temporally aligned candidate multi-medium data based on the source analysis contribution degree to obtain an initial fusion matrix; Calculate the mutual information value between each feature in the initial fusion matrix and the pollution source type label, and perform feature screening on the initial fusion matrix based on the mutual information value to obtain a fusion feature matrix.

4. The multi-agent machine learning based urban new pollutant source resolution method according to any one of claims 1 to 3, characterized in that, The adjustment of the monitoring strategy based on the source analysis result output by the multi-agent includes: Path optimization is performed on the monitoring tracking path in the monitoring strategy based on the source analysis result output by the multi-agent to obtain a path optimization result, and the path optimization refers to the following formula: wherein, represents the expected cumulative reward of performing an action in a state , represents the learning rate, represents the immediate reward obtained after performing an action , represents the discount factor, represents the new state transitioned to after performing an action , represents all actions available in the new state , represents the maximum expected cumulative reward obtainable in the new state from all available actions; Configure the contribution degree threshold of each pollution source area based on the functional area type of each pollution source area in the source analysis result; Adjust the monitoring strategy based on the path optimization result, the contribution degree threshold, and the pollution diffusion path prediction result.

5. The multi-agent machine learning based urban novel pollutant source resolution method according to any one of claims 1 to 3, characterized in that, The iterative optimization of the multi-agent based on the dynamic tracking monitoring result includes: Obtain the measured pollution concentration of the pollution source area based on the dynamic tracking monitoring result, and calculate the concentration error based on the measured pollution concentration and the predicted pollution concentration; Compare the source analysis result with the analysis result of a pre-constructed control model to obtain a consistency coefficient, and the control model is a chemical mass balance model; Determine the current pollution pattern feature based on the source analysis result, and determine the historical pollution pattern feature based on the historical source analysis result; Perform pollution drift detection on the current pollution pattern feature and the historical pollution pattern feature to obtain pollution drift data; Iteratively optimize the multi-agent based on the concentration error, the consistency coefficient, and the pollution drift data.

6. A multi-agent machine learning-based urban new pollutant source resolution device, characterized by, The urban new pollutant source analysis device based on multi-agent machine learning includes: The multi-agent data acquisition module is configured to configure a monitoring strategy of the multi-agent based on multi-medium occurrence characteristics of the urban new pollutants, and to acquire original multi-medium data by the multi-agent after the strategy configuration, wherein the multi-medium occurrence characteristics refer to distribution states, concentration levels and migration and transformation rules of the urban new pollutants in different medium carriers such as air, water, soil and sediments, and the monitoring strategy includes a monitoring range, a monitoring tracking path, a monitoring frequency and a monitoring index of the agent carrier; and the original multi-medium data refer to original monitoring data related to the new pollutants in each medium directly collected by the multi-agent through a sensor. The multi-medium feature extraction module is configured to extract features from the original multi-medium data to obtain a fusion feature matrix. The multi-agent machine learning source analysis module is configured to send the fusion feature matrix to the multi-agent, and the multi-agent is configured to construct a source analysis model based on the fusion feature matrix and output a source analysis result through the source analysis model, wherein the source analysis result includes a pollution source area and a pollution contribution degree of each pollution source area. The pollution dynamic tracking module is configured to adjust the monitoring strategy based on the source analysis result output by the multi-agent, to configure the multi-agent based on the adjusted monitoring strategy, and to dynamically track and monitor the pollution source area through the multi-agent after the strategy configuration. The multi-agent optimization module is configured to iteratively optimize the multi-agent based on the dynamic tracking and monitoring result, wherein the iterative optimization includes model parameter optimization of the source analysis model and iterative optimization of the monitoring strategy. The new pollutant analysis module is configured to analyze the urban new pollutants through the multi-agent after the iterative optimization. The process in which the multi-agent constructs a source analysis model based on the fusion feature matrix and outputs a source analysis result through the source analysis model includes: data dimension reduction and data division are performed on the received fusion feature matrix to obtain model input data; an original machine learning model is constructed, wherein the original machine learning model includes a CNN spatial feature extraction module, an RNN time feature extraction module, a random forest pollution source classification module and a contribution quantification module, the CNN spatial feature extraction module is configured to capture spatial correlation features in the data by using multiple convolution kernels, the RNN time feature extraction module is configured to capture time fluctuations of pollution concentration by using LSTM units, the random forest pollution source classification module is configured to realize pollution source type classification by voting of multiple decision trees, and the contribution quantification module is configured to quantify the pollution source contribution degree; a joint loss function is constructed, and a mathematical expression of the joint loss function is: wherein, denotes a joint loss function, denotes a classification loss weight, denotes a regression loss weight, denotes a classification loss term, denotes a true label of a pollution source type, denotes a predicted label of a pollution source type, denotes a regression loss term, denotes a true value of a pollution source contribution, denotes a predicted value of a pollution source contribution; the original machine learning model is trained based on the model input data and the joint loss function, and the original machine learning model is parameter optimized through an Adam optimizer to obtain a source analysis model; and a source analysis result is output through the source analysis model. The pollution dynamic tracking module is further configured to predict a pollution diffusion path based on the meteorological data and the source resolution result, and obtain a pollution diffusion path prediction result, wherein a prediction formula of the pollution diffusion path is: wherein, represents a predicted pollution concentration at a spatial coordinate , and a time , represents a source strength, which is calculated based on pollution contribution degrees of each pollution source region in source resolution results, represents an average wind speed, represents an effective source height, which is an effective emission height of a pollution source, represents a horizontal diffusion parameter, represents a vertical diffusion parameter, represents a horizontal coordinate perpendicular to a downwind direction, represents a vertical coordinate, represents a downwind distance; The monitoring strategy is adjusted based on the pollution diffusion path prediction result.

7. A multi-agent machine learning-based urban new pollutant source resolution device, characterized by, The urban new pollution source resolution device based on multi-agent machine learning comprises a memory, a processor, and an urban new pollution source resolution program based on multi-agent machine learning stored in the memory, wherein the processor is configured to run the urban new pollution source resolution program based on multi-agent machine learning, and the urban new pollution source resolution program based on multi-agent machine learning is configured to implement the urban new pollution source resolution method based on multi-agent machine learning according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores an urban new pollution source resolution program based on multi-agent machine learning, and the urban new pollution source resolution program based on multi-agent machine learning implements the urban new pollution source resolution method based on multi-agent machine learning according to any one of claims 1 to 5 when executed by a processor.

9. A computer program product, characterised in that, The computer program product comprises an urban new pollution source resolution program based on multi-agent machine learning, and the urban new pollution source resolution program based on multi-agent machine learning implements the steps of the urban new pollution source resolution method based on multi-agent machine learning according to any one of claims 1 to 5 when executed by a processor.

Images