Water environment pollution source reasoning and tracing method, system, equipment and medium

Abstract

The invention relates to a water environment pollution source reasoning and tracing method, system and equipment and a medium. The method comprises the following steps: constructing a pollution source characteristic database containing spectrum fingerprints and spatio-temporal information; performing time sequence enhancement on the mixed spectral signal of the monitoring point to extract stable features; a hydrological model is coupled to dynamically simulate pollutant transport paths and probabilities, and a potential contribution source set with weights is generated; analyzing the mixed signal by adopting a deep unmixing network embedded with space-time constraint, and separating and quantifying known source contribution and unknown pollution components; and finally, generating a visual pollution contribution degree spatial distribution diagram and a quantitative traceability report through geographic information mapping. The method solves the problems that complex mixed signals are difficult to accurately analyze, multi-source contribution is difficult to dynamically quantify and unknown pollution components are difficult to effectively recognize in the prior art, and intelligent and accurate traceability of municipal water environment non-point source pollution is achieved.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, specifically to a method, system, equipment, and medium for inferring and tracing the sources of water pollution. Background Technology

[0002] In the field of water environment management, especially when facing non-point source pollution with complex composition and dispersed sources, accurately tracing the source of pollutant discharge is a key prerequisite for effective control and accountability. Water environment pollution source inference and tracing methods are a technical system designed to solve this problem. Its core objective is to infer and quantify the contribution of each potential pollution source from the mixed signals of polluted water bodies by comprehensively utilizing environmental monitoring, hydrological simulation, and data analysis. Traditional source tracing approaches often rely on pollution source inventory surveys, water quality model simulations, or simple statistical analyses. These methods often fail to clarify the specific responsibility of different sources when facing complex scenarios with mixed pollution from multiple sources and significant spatiotemporal dynamic changes.

[0003] The existing technologies used in this field have several shortcomings: First, for mixed signals acquired by monitoring equipment such as spectrometers and chromatograms, there is a lack of intelligent methods capable of deeply analyzing the superposition of multiple source features, making it impossible to effectively separate the contribution of a single source from the complex "pollution fingerprint" mixture. Second, under dynamic hydrological conditions, the migration path and probability of pollutants from the source to the monitoring point change constantly, while existing methods often use static or simplified models, making it difficult to accurately simulate and quantify the impact weight of such spatiotemporal dynamic changes on the source tracing results. Third, when unknown or unrecorded pollutant components appear in the monitoring signal, existing systems usually cannot effectively identify and separate the contribution of this "unknown source," leading to distortion or omission of the source tracing results. Summary of the Invention

[0004] Based on this, the purpose of the present invention is to provide a method, system, device and medium for inferring and tracing the source of water pollution, which can accurately analyze mixed pollution signals, dynamically quantify the contributions of multiple sources and effectively identify unknown components.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides a method for reasoning and tracing the sources of water pollution, comprising the following steps:

[0007] S1: Perform fluorescence spectroscopy analysis and spatiotemporal information acquisition on water samples from potential pollution sources within the target watershed, obtain the spectral fingerprint feature vector of each pollution source and its corresponding geographical location coordinates and sampling timestamp, and generate a pollution source feature database;

[0008] S2: Perform time-series feature enhancement on the raw water quality mixed spectral signal collected in real time at the monitoring point, and extract stable spectral features in the raw signal within a continuous time window based on a one-dimensional convolutional neural network to generate a mixed spectral signal;

[0009] S3: Based on the pollution source feature database, combined with the obtained monitoring point location information and the current sampling time, simulate and calculate the transport path and transport probability of each potential pollution source to the monitoring point under the current hydrological conditions, and generate a dynamic set of potential contributing sources with contribution probability weights.

[0010] S4: Perform deep unmixing network calculation based on mixed spectral signals and dynamic potential contributing source set, call encoder-decoder network with embedded spatiotemporal constraints to separate known source contributions and unknown source contributions in mixed signals, and generate quantitative contribution vectors of each known pollution source and spectral signals of unknown pollution components;

[0011] S5: Geographic information mapping is performed on the quantitative contribution vector and the spectral signal of unknown pollutant components. Combined with the geographical coordinates in the pollution source feature database, the contribution value is mapped to the geographic space and integrated into a graphic report, generating a visualized pollution source tracing result that includes a spatial distribution map of pollution contribution and a quantitative analysis report.

[0012] In one embodiment, S1 of the water environment pollution source inference and tracing method provided by the present invention specifically includes the following steps:

[0013] S11: Perform three-dimensional fluorescence spectroscopy scanning on the collected water samples from various pollution sources. The excitation wavelength and emission wavelength form a two-dimensional spectral matrix. Extract the position, intensity, and contour information of the characteristic fluorescence peaks in the two-dimensional spectral matrix to generate the original high-dimensional spectral data.

[0014] S12: Preprocess the raw high-dimensional spectral data to eliminate the fluorescence intensity distortion caused by water sample background scattering and concentration, and apply multivariate scattering correction to smooth the spectral baseline drift caused by suspended particulate matter, eliminate instrument noise and environmental interference, and generate a standardized spectral fingerprint feature vector.

[0015] S13: Perform spatiotemporal attribute binding on the spectral fingerprint feature vector, extract GPS coordinates and UTC timestamp metadata from the sampling records, convert the coordinates into vector points in the standard projected coordinate system, hash the timestamp and the vector unique identifier and write it into the data header to generate a pollution source feature database with spatiotemporal labels.

[0016] In one embodiment, S2 of the water environment pollution source inference and tracing method provided by the present invention specifically includes the following steps:

[0017] S21: Perform sliding window segmentation processing on the raw water quality mixed spectral signal output by the online spectrometer at the monitoring point, divide the continuous time-series signal into multiple overlapping signal segments according to a fixed duration, and generate a serialized signal unit to be enhanced.

[0018] S22: Perform multi-layer convolution and pooling feature learning processing on the serialized signal units. Use a one-dimensional convolutional neural network to slide and scan the serialized signal units in parallel using multiple convolutional kernels of different widths at different scales to capture the intensity change patterns and fluorescence peak morphology within the local wavelength range of the serialized signal units. Then, use pooling layers to downsample the generated feature maps of each layer, retain the significant features with the highest response values ​​in the local regions of each feature map, and extract deep abstract features.

[0019] S23: Dimensionality reduction and reconstruction of deep abstract features are performed. The deep abstract features are gradually upsampled and mapped back to the original signal space through a deconvolution layer with a symmetrical structure. The skip connections of the coding layer are combined to supplement the detailed information in the output of the corresponding convolution layer, and the final mixed spectral signal is generated.

[0020] In one embodiment, S3 of the water environment pollution source inference and tracing method provided by the present invention specifically includes the following steps:

[0021] S31: Perform hydrological spatiotemporal alignment processing on the geographical coordinates, sampling timestamps, and real-time monitoring point locations and current sampling times in the pollution source characteristic database. Combine the accessed real-time rainfall intensity and river flow data with the digital elevation model obtained through remote sensing satellite image interpretation to calculate the flow direction, confluence time, and the path decay coefficient based on the pollutant half-life, and generate a spatiotemporal parameter set to drive the simulation.

[0022] S32: Distributed simulation of pollutant transport paths based on spatiotemporal parameter sets. A distributed hydrodynamic model based on graph theory and Eulerian-Lagrange method is constructed and the random walk and diffusion process of tens of thousands of virtual pollutant particles released from various potential source locations in the river network topology driven by hydrodynamic conditions is simulated. Simulation results containing the set of transport paths of each potential source and the arrival time range are generated.

[0023] S33: Perform probability matching processing on the arrival time range in the simulation results and the current sampling time window. Calculate the transport probability of each potential source based on the spatiotemporal density of particles appearing at the monitoring point, and screen out high-probability pollution sources according to the preset probability threshold. At the same time, normalize the transport probability of the screened high-probability pollution sources and assign it as a contribution weight to generate a dynamic set of potential contributing sources containing contribution probability weights.

[0024] In one embodiment, S4 of the water environment pollution source inference and tracing method provided by the present invention specifically includes the following steps:

[0025] S41: Perform structured input construction processing on the spectral fingerprint feature vectors corresponding to the mixed spectral signal and the dynamic potential contributing source set. Use the mixed spectral signal as the target vector, arrange the spectral fingerprint feature vectors of all pollution sources in the dynamic potential contributing source set in order as the source feature matrix, and transform the contribution probability weights corresponding to the dynamic potential contributing source set into the attention mask matrix to generate the standard input tensor of the encoder-decoder network.

[0026] S42: The standard input tensor is encoded and de-entangled and decoded based on a multi-head spatiotemporal attention mechanism. The spectral dependence of the mixed signal is captured by the multi-head self-attention layer of the encoder. The contribution probability weight is injected as the prior knowledge of the encoder through cross-attention layer. The encoded features are separated into components related to the known source and independent residual components in the latent space of the decoder using the de-entanglement regularization term. The contribution distribution of the known source and the preliminary estimate of the unknown component are reconstructed respectively.

[0027] S43: Joint optimization and normalization are performed on the contribution distribution of known sources and the preliminary estimate of unknown components. Based on the gradient descent algorithm, a joint loss function consisting of spectral reconstruction error, sparsity penalty of known source contribution, and unknown-known source spectral separation is constructed. The network parameters are iteratively updated through backpropagation to minimize the joint loss function, and the normalized quantitative contribution vector and the spectral signal of unknown pollutant components with independent spectral characteristics are output.

[0028] In one embodiment, the expression for the joint loss function of the water pollution source inference and tracing method provided by the present invention is as follows:

[0029]

[0030] in, For the joint loss function value, The input is a mixed spectral signal vector. The source feature matrix is ​​composed of known source spectral fingerprint feature vectors. This is the quantitative contribution vector of the known sources to be optimized. For the unknown pollutant component spectral signal vector to be optimized, , These are the positive coefficient hyperparameters used to balance the weights of various losses.

[0031] In one embodiment, S5 of the water environment pollution source inference and tracing method provided by the present invention specifically includes the following steps:

[0032] S51: Perform spatial matching and interpolation processing on the quantitative contribution vector and the geographical coordinates in the pollution source feature database, associate the contribution value of each pollution source with the corresponding geographical coordinate point, and perform spatial interpolation rendering on the watershed map to generate a heat map of the spatial distribution of pollution contribution.

[0033] S52: Perform comprehensive analytical processing on the spectral signals and quantitative contribution vectors of unknown pollutants, analyze the intensity, spectral characteristics and spatiotemporal variation of unknown signals, assess the potential sources and contribution levels of unknown pollutants, and generate a quantitative analysis summary of unknown pollutants.

[0034] S53: Synthesize the spatial distribution heat map of pollution contribution and the quantitative analysis summary into a report, integrate key data, charts and conclusions, and generate a complete graphic quantitative analysis report containing problem description, source tracing process, main findings and remediation recommendations according to a preset template.

[0035] Secondly, this invention provides a water environment pollution source inference and tracing system, which is configured with the following modules:

[0036] The pollution source feature database construction module is used to perform fluorescence spectral analysis and spatiotemporal information acquisition on water samples from potential pollution sources within the target watershed, obtain the spectral fingerprint feature vector of each pollution source and its corresponding geographical location coordinates and sampling timestamp, and generate a pollution source feature database.

[0037] The mixed spectral feature extraction module is used to enhance the temporal features of the raw water quality mixed spectral signals collected in real time from the monitoring points. It extracts stable spectral features from the raw signals within a continuous time window based on a one-dimensional convolutional neural network to generate mixed spectral signals.

[0038] The dynamic potential source analysis module is used to simulate and calculate the transport path and transport probability of each potential pollution source to the monitoring point under the current hydrological conditions based on the pollution source feature database, combined with the obtained monitoring point location information and the current sampling time, and generate a dynamic potential source set containing contribution probability weights.

[0039] The deep unmixing module for spectral signals is used to perform deep unmixing network calculations based on mixed spectral signals and a dynamic set of potential contributing sources. It calls an encoder-decoder network with embedded spatiotemporal constraints to separate known and unknown source contributions in the mixed signal, and generates quantitative contribution vectors of each known pollution source and spectral signals of unknown pollution components.

[0040] The pollution source tracing result visualization module is used to map the quantitative contribution vector to the spectral signal of unknown pollutants using geographic information. It combines the geographic coordinates in the pollution source feature database to map the contribution value to the geographic space and integrate it into a graphic report, generating a visualized pollution source tracing result that includes a spatial distribution map of pollution contribution and a quantitative analysis report.

[0041] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned methods for reasoning and tracing water pollution sources.

[0042] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-mentioned methods for reasoning and tracing water pollution sources.

[0043] In summary, the water pollution source inference and tracing method provided in this application establishes a queryable and comparable benchmark for accurate source tracing by constructing a pollution source feature database that integrates spectral fingerprints and spatiotemporal information; it enhances the temporal features of the mixed spectral signals of monitoring points based on a one-dimensional convolutional neural network, effectively extracting and strengthening stable pollution features hidden in dynamic noise, significantly improving signal quality; it generates a set of potential sources with contribution probability weights that fit real-time hydrological conditions by dynamically simulating the transport paths and probabilities of pollutants through coupled hydrological models; it transforms the uncertainty of the physical migration process into quantifiable prior knowledge; it calls an encoder-decoder network embedded with the above spatiotemporal constraints to perform deep demixing of the mixed signals, enabling efficient and high-precision separation and quantitative analysis of known and unknown source contributions in complex spectral signals; and it combines the quantified contribution degree with spatial location through geographic information mapping to generate a graphic report, which can intuitively present the spatial distribution and quantitative conclusions of pollution contributions. This method can systematically solve the shortcomings of traditional methods in analyzing complex mixed signals, quantifying dynamic multi-source contributions, and identifying unknown components, providing reliable technical support for the precise treatment and responsibility determination of non-point source pollution in municipal water environments.

[0044] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0045] Figure 1 A flowchart illustrating a method for reasoning and tracing water pollution sources, provided in an embodiment of this application;

[0046] Figure 2 A schematic diagram illustrating the process of generating spectral signals of unknown pollutants provided in an embodiment of this application;

[0047] Figure 3This is a schematic diagram of a water environment pollution source inference and tracing system provided in another embodiment of this application. Detailed Implementation

[0048] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0050] In one embodiment, such as Figure 1 As shown, a method for tracing and reasoning about water pollution sources is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0051] S1: Perform fluorescence spectroscopy analysis and spatiotemporal information acquisition on water samples from potential pollution sources within the target watershed, obtain the spectral fingerprint feature vector of each pollution source and its corresponding geographical location coordinates and sampling timestamp, and generate a pollution source feature database.

[0052] Specifically, the system conducts a potential pollution source investigation for the target watershed, identifying various locations within the watershed that may generate pollution emissions, and collecting water samples from each potential pollution source according to a pre-set investigation scope. After collection, the system imports the water samples into a fluorescence spectroscopy analysis device. The device excites the pollutants in the water samples to produce characteristic fluorescence responses, capturing fluorescence intensity data under different combinations of excitation and emission wavelengths. The acquired fluorescence intensity data is then systematically integrated according to wavelength to form a unique spectral fingerprint feature vector for each potential pollution source. This vector uniquely characterizes the pollutant composition and content distribution patterns of the corresponding pollution source.

[0053] Furthermore, the system activates the positioning and timing modules to collect the geographical location and sampling time information of each pollution source. The geographical location information is converted into standardized coordinates, and the sampling time information is recorded as a unified format timestamp. The system performs correlation processing on the spectral fingerprint feature vector, standardized geographical location coordinates, and unified format timestamps to establish a one-to-one correspondence among them. Following structured data storage specifications, all the correlated information is entered into a preset database framework, completing the construction of the pollution source characteristic database. This database provides basic data support for subsequent source tracing analysis, ensuring that the basic characteristic information of each potential pollution source can be quickly accessed during the source tracing process. During the database construction process, the system performs integrity verification on all entered data, removing invalid data to ensure the accuracy and usability of the data within the database.

[0054] S2: Perform time-series feature enhancement on the raw water quality mixed spectral signal collected in real time from the monitoring point, and extract stable spectral features from the raw signal within a continuous time window based on a one-dimensional convolutional neural network to generate a mixed spectral signal.

[0055] Specifically, the system acquires raw water quality mixed spectral signals of the target water body in real time through spectral acquisition equipment deployed at monitoring points. This signal contains superimposed spectral characteristics of pollutants emitted from various potential pollution sources within the watershed, but is subject to noise and baseline drift due to environmental interference. The system performs time-series feature enhancement processing on the real-time acquired raw spectral signals. First, a continuous sliding time window is constructed, and the raw signal within the time series is segmented to achieve orderly processing of the continuous signal. For the signal data within each time window, the system performs baseline correction processing to eliminate the influence of instrument baseline fluctuations on the signal. Then, a signal filtering algorithm is used to remove environmental noise interference and purify the effective spectral information in the raw signal.

[0056] After processing, the system invokes a one-dimensional convolutional neural network, inputting the preprocessed temporal signal into the network's input layer. The convolutional layer performs convolution operations on the input signal using preset kernels, extracting local temporal correlation features and capturing the variation patterns of spectral signals at different time points. After the convolution operation, the system uses pooling layers to reduce the dimensionality of the feature maps output by the convolutional layers, preserving key feature information while reducing data computation and improving feature extraction efficiency. Finally, the system integrates the local features output by the pooling layers through fully connected layers, constructing a complete feature system and generating a mixed spectral signal with high discriminative power and stability. This provides high-quality signal data support for subsequent deep unmixing operations, ensuring that the unmixing process can accurately separate the spectral features of each pollution source.

[0057] S3: Based on the pollution source feature database, combined with the obtained monitoring point location information and the current sampling time, simulate and calculate the transport path and transport probability of each potential pollution source to the monitoring point under the current hydrological conditions, and generate a dynamic set of potential contributing sources with contribution probability weights.

[0058] Specifically, the system calls upon the established pollution source characteristic database to extract basic characteristic information of each potential pollution source, and simultaneously obtains the geographical coordinates of the monitoring points and the current sampling time. This information is input into the hydrological and hydrodynamic simulation model as basic parameters. The system synchronously collects topographic data, hydrological runoff data, pollutant degradation characteristic data, and current hydrological conditions of the target watershed. After standardizing various data types, it imports them into the simulation model, providing comprehensive data support for model calculations. Based on the input parameters, the model simulates the migration and diffusion process of pollutants after discharge from each potential pollution source. Combining the laws of water flow, it derives the specific transport path of pollutants from the pollution source location to the monitoring point location, clarifying the diffusion range and trajectory of pollutants during the migration process.

[0059] Furthermore, based on water flow velocity, pollutant diffusion coefficient, and pollutant decay law, the system calculates the transport probability of pollutants emitted from each potential pollution source reaching the monitoring point. Simultaneously, it combines this with sampling time information recorded in the pollution source characteristic database to determine whether each potential pollution source is in a pollutant emission state during the current sampling time period, calculating the probability that each pollution source is in an emission state. Combining these two probability results, the system calculates the comprehensive contribution probability of each potential pollution source, sets a comprehensive contribution probability screening criterion, and selects pollution sources with the potential to contribute to pollution. The system assigns a corresponding contribution probability weight to each selected pollution source. The weight allocation is determined based on the comprehensive contribution probability, forming a dynamic set of potential contributing sources containing contribution probability weights, achieving accurate screening and priority ranking of potential contributing sources.

[0060] S4: Perform deep unmixing network calculations based on the mixed spectral signal and dynamic potential contributing source set, call the encoder-decoder network with embedded spatiotemporal constraints to separate the known source contributions and unknown source contributions in the mixed signal, and generate the quantitative contribution vector of each known pollution source and the spectral signal of the unknown pollution component.

[0061] Specifically, the system uses the generated mixed spectral signal and the dynamic potential contributing source set as input data, which are then fed into a deep demixing network for processing. The deep demixing network adopts an encoder-decoder network architecture with embedded spatiotemporal constraints, which enables accurate separation and quantitative analysis of the mixed signal. The encoder module performs multi-layer feature encoding on the input mixed spectral signal, extracts the dimensional features of the mixed spectral signal through convolutional layers, and introduces an attention mechanism to strengthen the signal components related to the pollution source features in the dynamic potential contributing source set. The spatiotemporal constraint information in the dynamic potential contributing source set is transformed into quantifiable constraint parameters and embedded into the feature encoding process. The constraint parameters limit the value range of the encoded features to ensure that the encoded features match the current spatiotemporal conditions.

[0062] The decoder module performs signal demixing based on the encoder's output encoding features and spatiotemporal constraint parameters. It separates known and unknown source contribution signals from the mixed spectral signal through reverse derivation. Known source contribution signals correspond to each pollution source in the dynamic potential contributing source set, while unknown source contribution signals correspond to pollution components not recorded in the pollution source feature database. Quantitative analysis is performed on the separated known source contribution signals. By matching signal intensity with pollution source characteristics, a quantitative contribution vector is obtained for each known pollution source. This vector reflects the degree of contribution of each known pollution source to the water pollution at the monitoring point. Simultaneously, the spectral features of the unknown source contribution signals are extracted to generate spectral signals of unknown pollution components. These signals provide characteristic evidence for subsequent investigation of unknown pollution sources. The entire demixing process optimizes the objective function to ensure the accuracy of the demixing results, balancing signal reconstruction errors with the rationality of contribution analysis, thus improving the reliability of the demixing results.

[0063] S5: Geographic information mapping is performed on the quantitative contribution vector and the spectral signal of unknown pollutant components. Combined with the geographical coordinates in the pollution source feature database, the contribution value is mapped to the geographic space and integrated into a graphic report, generating a visualized pollution source tracing result that includes a spatial distribution map of pollution contribution and a quantitative analysis report.

[0064] Specifically, the system performs geographic information mapping on the obtained quantitative contribution vector and the spectral signals of unknown pollutants. It retrieves the geographic coordinates of each known pollution source from the pollution source feature database and precisely associates the quantitative contribution value of each known pollution source with its corresponding geographic coordinates, ensuring a one-to-one correspondence between contribution values ​​and geographic locations. Based on the geographic information system, the associated contribution data undergoes spatial visualization processing. Spatial rendering technology transforms different contribution values ​​into corresponding visual presentation effects, intuitively displaying the distribution of pollution source contributions to monitoring points and surrounding water bodies in each region, generating a spatial distribution map of pollution contribution. This map clearly reflects the spatial differences in pollution contribution characteristics.

[0065] Statistical analysis is performed on the quantitative contribution vector to identify the distribution patterns, proportions, and variation characteristics of the contributions from known pollution sources. The wavelength characteristics and signal intensity of the spectral signals of unknown pollutants are analyzed to extract their core spectral information. A quantitative analysis report is written based on the results, covering the contribution analysis of known pollution sources, the description of the spectral characteristics of unknown pollutants, and an assessment of the reasonableness of the source tracing results. The generated spatial distribution map of pollution contributions is integrated with the quantitative analysis report and formatted according to patent document specifications to form a visualized pollution source tracing result. This result provides accurate data support and decision-making basis for water environment governance, pollution source control, and accountability, while also supporting the export and sharing of results to meet the needs of different application scenarios.

[0066] In summary, the water pollution source inference and tracing method provided in this application establishes a queryable and comparable benchmark for accurate source tracing by constructing a pollution source feature database that integrates spectral fingerprints and spatiotemporal information; it enhances the temporal features of the mixed spectral signals of monitoring points based on a one-dimensional convolutional neural network, effectively extracting and strengthening stable pollution features hidden in dynamic noise, significantly improving signal quality; it generates a set of potential sources with contribution probability weights that fit real-time hydrological conditions by dynamically simulating the transport paths and probabilities of pollutants through coupled hydrological models; it transforms the uncertainty of the physical migration process into quantifiable prior knowledge; it calls an encoder-decoder network embedded with the above spatiotemporal constraints to perform deep demixing of the mixed signals, enabling efficient and high-precision separation and quantitative analysis of known and unknown source contributions in complex spectral signals; and it combines the quantified contribution degree with spatial location through geographic information mapping to generate a graphic report, which can intuitively present the spatial distribution and quantitative conclusions of pollution contributions. This methodology can systematically address the shortcomings of traditional methods in analyzing complex mixed signals, quantifying dynamic multi-source contributions, and identifying unknown components, providing reliable technical support for the precise treatment and liability determination of non-point source pollution in municipal water environments.

[0067] In one embodiment, S1 of the water environment pollution source inference and tracing method provided by the present invention specifically includes the following steps:

[0068] S11: Perform three-dimensional fluorescence spectroscopy scanning on the collected water samples from various pollution sources. The excitation wavelength and emission wavelength form a two-dimensional spectral matrix. Extract the position, intensity, and contour information of the characteristic fluorescence peaks in the two-dimensional spectral matrix to generate the original high-dimensional spectral data.

[0069] Specifically, the system performs three-dimensional fluorescence spectroscopy scanning on the collected water samples from various pollution sources. By adjusting the range of excitation and emission wavelengths using a fluorescence spectrometer, the excitation and emission wavelengths are sequentially combined and scanned at fixed intervals within a set range, forming a two-dimensional spectral matrix covering the entire wavelength range. This matrix has the excitation wavelength as the horizontal axis and the emission wavelength as the vertical axis, with matrix elements corresponding to fluorescence intensity response values ​​under different wavelength combinations. During the scanning process, the system maintains consistency in the scanning parameters to ensure the comparability of spectral data from different water samples and avoid spectral distortion caused by parameter differences.

[0070] After scanning, the system extracts features from the two-dimensional spectral matrix, identifying the specific locations of characteristic fluorescence peaks. These locations are uniquely determined by the corresponding excitation and emission wavelengths. Simultaneously, the system captures the intensity information of the characteristic fluorescence peaks, reflecting the relative content of the corresponding pollutant components. The system also extracts the contour information of the characteristic fluorescence peaks, including peak width, symmetry, and peak shape variation trends. This information characterizes the structural properties of the pollutant components. The system integrates the extracted characteristic fluorescence peak locations, intensities, and contour information, arranging them in wavelength order to form original high-dimensional spectral data. This data fully preserves the fluorescence spectral characteristics of each pollutant component in the water sample.

[0071] S12: Preprocess the raw high-dimensional spectral data to eliminate background scattering of water samples and fluorescence intensity distortion caused by concentration, and apply multivariate scattering correction to smooth the spectral baseline drift caused by suspended particulate matter, eliminate instrument noise and environmental interference, and generate a standardized spectral fingerprint feature vector.

[0072] Specifically, the system preprocesses the generated raw high-dimensional spectral data to eliminate fluorescence intensity distortion caused by background scattering and concentration. By establishing a background subtraction model, it extracts the spectrum of a blank water sample free of pollutants as a background reference. The system then performs a difference calculation between the raw high-dimensional spectral data and the background reference spectrum to eliminate the influence of background scattering on the spectral signal. Simultaneously, an intensity normalization method is used to correct fluorescence intensity distortion caused by differences in concentration among different water samples, ensuring that the spectral intensity accurately reflects the inherent characteristics of the pollutant components. To address the spectral baseline drift caused by suspended particulate matter, the system applies a multivariate scattering correction algorithm for smoothing. This algorithm fits and adjusts the spectral baseline, eliminating baseline shift caused by scattering from suspended particulate matter and restoring the original spectral characteristics.

[0073] Furthermore, the system uses filtering algorithms to eliminate noise generated during instrument operation and signal fluctuations caused by external environmental interference, filtering out abnormal interference components in the spectral data and retaining effective spectral signals related to pollutant components. After a series of preprocessing operations, interference factors in the original high-dimensional spectral data are effectively removed, and the accuracy and stability of spectral features are significantly improved. Based on the processed data, the system generates a standardized spectral fingerprint feature vector, which can accurately characterize the inherent spectral features of the corresponding pollution source.

[0074] S13: Perform spatiotemporal attribute binding on the spectral fingerprint feature vector, extract GPS coordinates and UTC timestamp metadata from the sampling records, convert the coordinates into vector points in the standard projected coordinate system, hash the timestamp and the vector unique identifier and write it into the data header to generate a pollution source feature database with spatiotemporal labels.

[0075] Specifically, for GPS coordinates, the system uses a coordinate transformation algorithm to convert the spectral fingerprint feature vector from the original geographic coordinate system to vector points in the standard projected coordinate system. This eliminates positional deviations caused by different coordinate systems, ensuring that the vector points accurately correspond to the actual sampling locations in geographic space and guaranteeing the consistency and accuracy of geographic location information. For UTC timestamps, the system hashes and associates them with the unique identifier of the spectral fingerprint feature vector. A unique association code is generated using a hash algorithm to ensure a one-to-one correspondence between the timestamp and the corresponding spectral fingerprint feature vector, avoiding confusion or mismatch during data association. The system writes the associated spatiotemporal information into the data header of the spectral fingerprint feature vector, forming a complete data unit containing spectral features, geographic location, and time information. Simultaneously, according to a preset data storage structure, all data units with spatiotemporal tags are organized and archived to establish a pollution source feature database with spatiotemporal tags. This database enables linked queries of spectral fingerprint features and spatiotemporal attributes.

[0076] In one embodiment, S2 of the water environment pollution source inference and tracing method provided by the present invention specifically includes the following steps:

[0077] S21: Perform sliding window segmentation processing on the raw water quality mixed spectral signal output by the online spectrometer at the monitoring point, divide the continuous time-series signal into multiple overlapping signal segments according to a fixed duration, and generate the serialized signal unit to be enhanced.

[0078] Specifically, the system acquires the raw mixed spectral signal of water quality continuously output by the online spectrometer at the monitoring point. This signal is a continuously generated spectral data sequence in the time dimension, containing the spectral superposition information of various pollutant components in the water body at the monitoring point, as well as interference signals caused by environmental fluctuations, instrument operation, and other factors. The system performs sliding window segmentation processing on this raw mixed spectral signal. The time span of the sliding window is set according to a fixed duration, and the continuous time-series signal is segmented in segments based on this window. During the segmentation process, some data overlap is maintained between adjacent sliding windows to avoid the continuity of the time-series signal caused by window segmentation, ensuring that the subsequent feature learning process can completely capture the changing trend and correlation features of the spectral signal in the time dimension. After each window segmentation is completed, the system organizes all the spectral data covered in the window according to the wavelength and time dimensions to form independent signal segments. All the segmented signal segments are arranged in an orderly manner according to the original time sequence to generate a serialized signal unit to be enhanced. This serialized signal unit retains the temporal continuity and local signal details of the raw mixed spectral signal, while decomposing the continuous long time-series signal into structured small unit data.

[0079] S22: Perform multi-layer convolution and pooling feature learning processing on the serialized signal units. Use a one-dimensional convolutional neural network to slide and scan the serialized signal units in parallel using multiple convolutional kernels of different widths at different scales to capture the intensity change patterns and fluorescence peak morphology within the local wavelength range of the serialized signal units. Then, use pooling layers to downsample the generated feature maps of each layer, retain the significant features with the highest response values ​​in the local regions of each feature map, and extract deep abstract features.

[0080] Specifically, the system inputs the generated serialized signal units to be enhanced into a one-dimensional convolutional neural network (CNN) to perform multi-layer convolution and pooling feature learning processing. In the convolutional feature extraction stage, the CNN is configured with multiple feature learning scales, deploying multiple convolutional kernels of different widths at different scales. All convolutional kernels slide and scan the serialized signal units in parallel. During the scanning process, each convolutional kernel performs convolution operations with the corresponding local signal segment of the serialized signal unit, achieving the capture and extraction of local signal features. Different widths of convolutional kernels correspond to different local wavelength coverage ranges. Narrower width kernels are used to capture the detailed fluorescence peak morphology within a fine local wavelength range in the serialized signal unit, while wider width kernels are used to capture the spectral intensity variation patterns within a large local wavelength range in the serialized signal unit. Through parallel scanning and computation of multi-width convolutional kernels, the effective feature information in the serialized signal unit is comprehensively covered, avoiding the omission of features by single-scale convolutional kernels.

[0081] After each convolutional operation, a feature map representing the learning result of that layer is generated. The system sequentially inputs these feature maps into the corresponding pooling layers for downsampling. The pooling layers filter and integrate local regions in the feature maps according to preset pooling rules, retaining the most significant features with the highest response values ​​in each local region while removing redundant low-response information and invalid interference data. This effectively reduces the dimensionality of data processing and subsequent computation while ensuring that core significant features are not lost. The system alternately executes multiple convolutional and pooling operations. With each round of convolution and pooling, the abstraction level of the feature maps is further improved. After multiple rounds of iterative processing, the system extracts deep abstract features from the serialized signal units that can accurately represent the core spectral characteristics of pollutants. These deep abstract features eliminate various interference information in the original signal and condense the effective feature content of the serialized signal units.

[0082] S23: Dimensionality reduction and reconstruction of deep abstract features are performed. The deep abstract features are gradually upsampled and mapped back to the original signal space through a deconvolution layer with a symmetrical structure. The skip connections of the coding layer are combined to supplement the detailed information in the output of the corresponding convolution layer, and the final mixed spectral signal is generated.

[0083] Specifically, the system performs dimensionality reduction and reconstruction on the extracted deep abstract features. It calls a deconvolutional layer with a symmetrical structure to perform feature upsampling. The network structure of this deconvolutional layer is symmetrically distributed with the convolutional layer structure used in the previous convolutional feature extraction, enabling feature dimension mapping and detail restoration that is the reverse of the convolutional feature extraction process. The deconvolutional layer progressively performs upsampling on the input deep abstract features. Each deconvolutional operation specifically enhances the dimension of the deep abstract features, gradually restoring the local details lost during convolution and pooling. Through multiple rounds of layer-by-layer deconvolutional upsampling, the highly abstract deep features are gradually mapped back to the signal space corresponding to the original water quality mixed spectral signal, ensuring that the reconstructed signal dimension remains consistent with the original signal dimension.

[0084] Throughout the deconvolution upsampling process, the system introduces a skip connection mechanism in the coding layer. This mechanism directly transmits the feature map details from each level of convolutional processing to the corresponding deconvolutional layer via skip connections. The feature data generated by the deconvolution upsampling at that layer is then fused with the data, supplementing local signal details that are difficult to reconstruct during the deconvolution upsampling process. This addresses the issues of missing details and feature distortion during feature reconstruction, ensuring that the reconstructed signal retains both the core value of deep abstract features and the fine details of the original spectral signal. Through the collaborative processing of deconvolution upsampling and skip connection detail supplementation, the system completes the dimensionality reduction and reconstruction of deep abstract features, transforming the abstract feature data into a spectral signal with temporal stability and feature integrity. This signal eliminates various noises and transient interferences in the original mixed spectral signal of the water quality, accurately preserving the core spectral superposition features of pollutants in the water body at the monitoring point, and ultimately generating a mixed spectral signal that meets the computational requirements of the subsequent deep unmixing network.

[0085] In one embodiment, S3 of the water environment pollution source inference and tracing method provided by the present invention specifically includes the following steps:

[0086] S31: Perform hydrological spatiotemporal alignment processing on the geographical coordinates, sampling timestamps, and real-time monitoring point locations and current sampling times in the pollution source characteristic database. Combine the accessed real-time rainfall intensity, river flow data, and digital elevation models obtained through remote sensing satellite image interpretation to calculate flow direction, confluence time, and flow rate decay coefficient based on pollutant half-life, and generate a spatiotemporal parameter set to drive the simulation.

[0087] Specifically, the system retrieves the geographic coordinates and sampling timestamps of each potential pollution source from the pollution source characteristic database, simultaneously acquiring the real-time location and current sampling time of monitoring points, and performing hydrological spatiotemporal alignment processing on the two types of spatiotemporal data. A unified geographic coordinate system and time reference are established; a coordinate transformation algorithm is used to normalize geographic location information from different sources to the same projected coordinate system, and the sampling timestamps and current sampling times are unified to the UTC time standard, thereby eliminating analytical biases caused by reference differences. The system accesses real-time rainfall intensity and river flow data, and simultaneously retrieves the target watershed digital elevation model generated from remote sensing satellite image interpretation. Based on the topographic features of this model, the water flow direction is calculated, following the Laplace operator relation; the calculation expression is:

[0088]

[0089] in, Indicates the direction vector of water flow. This represents the terrain elevation value in the digital elevation model. The gradient operator is used to characterize the spatial rate of change of terrain slope. The system combines water flow direction and river flow rate to derive the velocity distribution of different river sections within the watershed, and calculates the confluence time based on the spatial distance between the pollution source and the monitoring point and the velocity parameters of the corresponding river section. Preferably, the system introduces the pollutant half-life parameter to calculate the attenuation coefficient along the flow path, using the following formula:

[0090]

[0091] in, Indicates the friction loss coefficient. This indicates the pollutant's half-life, used to quantify the natural concentration decay characteristics of pollutants during migration due to physicochemical processes. The system integrates the aforementioned water flow direction, confluence time, friction loss coefficient, and real-time hydrological data, arranging them in an ordered spatiotemporal parameter set according to a fixed data structure.

[0092] S32: Distributed simulation of pollutant transport paths based on spatiotemporal parameter sets. A distributed hydrodynamic model based on graph theory and Eulerian-Lagrange method is constructed and the random walk and diffusion process of tens of thousands of virtual pollutant particles released from various potential source locations in the river network topology driven by hydrodynamic conditions is simulated. Simulation results containing the set of transport paths of each potential source and the arrival time range are generated.

[0093] Specifically, the system conducts distributed simulations of pollutant transport paths based on the generated spatiotemporal parameter set, constructing a distributed hydrodynamic model integrating graph theory and the Eulerian-Lagrange method. Graph theory abstracts the river network topology as a set of nodes and edges, where nodes correspond to river confluences and key control sections, and edges correspond to river segments connecting nodes. Edges are also assigned properties such as flow resistance and cross-sectional area to recreate the actual connectivity characteristics of the river network. The Eulerian method describes the macroscopic motion of the entire water body, characterizing the spatial distribution of flow velocity and direction through the control volume equations. The Lagrange method tracks the microscopic migration trajectory of individual pollutant particles, achieving coupled simulation of the macroscopic flow field and microscopic particle motion. Preferably, the system releases virtual pollutant particles from the locations of potential pollution sources. The particle motion is jointly driven by flow velocity, flow direction, and diffusion coefficient, and its displacement changes satisfy the stochastic diffusion equation:

[0094]

[0095] in, This represents the spatial position of the particle at time t. This indicates the initial release location of the particles, i.e., the coordinates of the pollution source. express The water flow velocity of the river section at that time Indicates the pollutant diffusion coefficient. This represents a random term that follows a standard normal distribution, used to characterize the randomness of pollutant molecule diffusion and turbulent diffusion. The system records the migration trajectory and time node of each particle in real time, statistically analyzes the time intervals of particles released from each pollution source arriving at the monitoring point, integrates all particle trajectories of the same pollution source to form a transport path set, summarizes the arrival time intervals to determine the pollutant arrival time range of the corresponding pollution source, and generates simulation results containing complete information after simulating all potential pollution sources.

[0096] S33: Perform probability matching processing on the arrival time range in the simulation results and the current sampling time window. Calculate the transport probability of each potential source based on the spatiotemporal density of particles appearing at the monitoring point, and screen out high-probability pollution sources according to the preset probability threshold. At the same time, normalize the transport probability of the screened high-probability pollution sources and assign it as a contribution weight to generate a dynamic set of potential contributing sources containing contribution probability weights.

[0097] Specifically, the system performs probability matching between the arrival time range of pollutants from each potential pollution source and the current sampling time window. The current sampling time window is the continuous time interval of the raw water quality mixed spectral signal collected at the monitoring point. The matching process is based on the intersection range of the time intervals. The system counts the number of virtual pollutant particles released from each potential pollution source whose arrival time falls within this intersection range. Based on the ratio of this number to the total number of particles released by the pollution source, the spatiotemporal density is calculated. This density directly corresponds to the transport probability of each potential pollution source. The calculation expression is:

[0098]

[0099] in, This represents the transport probability of the i-th potential pollution source. This indicates the number of particles released from the pollution source whose arrival time falls within the sampling time window. This represents the total number of virtual pollutant particles released by the pollution source. The system calls a preset probability threshold to filter out high-probability pollution sources with a transport probability not lower than the threshold. The transport probability of these sources is then normalized to eliminate the influence of dimensions, using the following formula:

[0100]

[0101] in, This represents the contribution weight of the i-th high-probability pollution source. This represents a set of high-probability pollution sources. Normalization is applied to ensure the sum of the contribution weights of all high-probability pollution sources equals 1, guaranteeing the rationality of weight allocation. The system associates the spectral fingerprint feature vectors, geographic coordinates, and calculated contribution weights of each high-probability pollution source, sorting them chronologically and by weight to form a dynamic set of potential contributing sources containing contribution probability weights.

[0102] In one embodiment, S4 of the water environment pollution source inference and tracing method provided by the present invention specifically includes the following steps:

[0103] S41: Perform structured input construction processing on the spectral fingerprint feature vectors corresponding to the mixed spectral signal and the dynamic potential contributing source set. Use the mixed spectral signal as the target vector, arrange the spectral fingerprint feature vectors of all pollution sources in the dynamic potential contributing source set in order to form the source feature matrix, and transform the contribution probability weights corresponding to the dynamic potential contributing source set into the attention mask matrix to generate the standard input tensor of the encoder-decoder network.

[0104] Specifically, the system constructs a structured input of the spectral fingerprint feature vectors corresponding to the mixed spectral signal and the dynamic potential contributing source set. Using the wavelength scanning range of the mixed spectral signal as a benchmark, the system performs wavelength dimension normalization on all spectral fingerprint feature vectors in the dynamic potential contributing source set. A linear interpolation algorithm is used to complete missing wavelength dimension data, and redundant data exceeding the benchmark range is pruned through interval truncation. This ensures complete uniformity in the number of spectral dimensions, wavelength range, and sampling interval, avoiding matrix operation errors caused by dimensional differences. Furthermore, the system determines the mixed spectral signal as the target vector for model optimization, fully preserving its temporal stability characteristics and fluorescence response information at each wavelength. This vector serves as the core benchmark for subsequent signal reconstruction accuracy evaluation and error comparison. The system constructs a source feature matrix by sequentially arranging the corresponding spectral fingerprint feature vectors in descending order of the contribution probability weights of each pollution source in the dynamic potential contributing source set. Each row of the matrix corresponds to the complete spectral fingerprint feature of a single pollution source, and each column corresponds to the feature response value of each pollution source at the same wavelength dimension. The matrix structure is adjusted synchronously with the update of the dynamic potential contributing source set. Simultaneously, the contribution probability weights are converted into an attention mask matrix, following the formula:

[0105]

[0106] in, This is an attention mask matrix, where each element corresponds to the weight of each contamination source. The matrix represents the contribution probability weights of a dynamic set of potential contributing sources. The softmax function maps these weights to the 0-1 interval using exponential normalization, ensuring the matrix accurately represents the priority of each pollution source's potential contribution to the mixed signal. The system uses linear interpolation to complete the wavelength data.

[0107]

[0108] in, These are the interpolated eigenvalues. , The characteristic values ​​are those of adjacent known wavelengths. , For adjacent known wavelengths, The target wavelength to be completed is then used. Next, tensor transformations are performed on the target vector, source feature matrix, and attention mask matrix to supplement batch and channel dimension information. A dimension alignment algorithm is then used to adjust the dimension matching relationship among the three, eliminating dimension incompatibility issues and ensuring that the input specifications of the encoder-decoder network are met. Finally, a standard input tensor with a standardized structure and uniform dimensions is generated.

[0109] S42: The standard input tensor is encoded and de-entangled and decoded based on a multi-head spatiotemporal attention mechanism. The spectral dependence of the mixed signal is captured by the multi-head self-attention layer of the encoder. The contribution probability weights are injected as prior knowledge of the encoder through cross-attention layers. In the latent space of the decoder, the encoded features are separated into components related to the known source and independent residual components by using the de-entanglement regularization term. The contribution distribution of the known source and the preliminary estimate of the unknown component are reconstructed respectively.

[0110] Specifically, the encoder's multi-head self-attention layer is configured with multiple parallel attention heads. Each attention head independently focuses on different wavelength ranges of the mixed spectral signal. By calculating the correlation weights of features at different wavelength positions, it synchronously captures the short-term and long-term spectral dependencies within the signal, uncovering source feature correlation information hidden in the mixed signal. The feature outputs of each attention head are concatenated and linearly transformed to complete the deep feature encoding of the mixed spectral signal, improving the comprehensiveness of feature representation. Across the attention layer, the contribution probability weights in the attention mask matrix are injected into the encoding process as prior knowledge. Through a weight allocation mechanism, the correlation between the spectral features of high-probability pollution sources and the encoded features is strengthened, while the feature interference from low-weight pollution sources is weakened. This makes the encoded features more closely match the features of high-probability contributing sources under current hydrological conditions, reducing the impact of irrelevant features on subsequent demixing.

[0111] Preferably, the decoder performs deentanglement decoding on the encoded features in the latent space, introducing a deentanglement regularization term to constrain the feature separation process. The constraint relationship is realized through the formula:

[0112]

[0113] in, To de-entangle, a regularization term is used to quantify the degree of feature separation. The covariance matrix of the encoded features is calculated from the variance and covariance of the encoded features. This is a diagonal matrix, where the diagonal elements correspond to the variance values ​​of the encoded features. This regularization term achieves accurate separation by maximizing the independence between features. Covariance matrix The calculation follows the formula:

[0114]

[0115] in, Let $n$ be the element in the i-th row and j-th column of the covariance matrix, $f_{ki}$ be the number of encoded features, $f_{ki}$ be the value of the i-th dimension of the k-th encoded feature, and $\mu_i$ be the mean of the i-th dimension. The system uses this regularization term to separate the encoded features into two types of components: one is the correlation component, which highly matches the spectral features of known sources, and the other is the residual component, which is independent of all known source features. The contribution distribution of each known source is reconstructed based on the correlation component, and a preliminary estimate of the unknown components is generated based on the residual component.

[0116] S43: Joint optimization and normalization are performed on the contribution distribution of known sources and the preliminary estimate of unknown components. Based on the gradient descent algorithm, a joint loss function consisting of spectral reconstruction error, sparsity penalty of known source contribution, and unknown-known source spectral separation is constructed. The network parameters are iteratively updated through backpropagation to minimize the joint loss function, and the normalized quantitative contribution vector and the spectral signal of unknown pollutant components with independent spectral characteristics are output.

[0117] Specifically, the system performs joint optimization and normalization on the contribution distribution of known sources and the preliminary estimate of unknown components. A gradient descent algorithm is used to construct a joint loss function, and iterative optimization is employed to improve the accuracy of the results, ensuring that the output data meets the requirements of source tracing analysis. The expression for the joint loss function is:

[0118]

[0119] in, The joint loss function value is used to comprehensively evaluate the model optimization effect. The input is a mixed spectral signal vector. The source feature matrix is ​​composed of known source spectral fingerprint feature vectors. This is the quantitative contribution vector of the known sources to be optimized. For the unknown pollutant component spectral signal vector to be optimized, , These are the positive coefficient hyperparameters used to balance the weights of various losses, and they respectively balance the weights of spectral reconstruction error, sparsity penalty for known source contributions, and spectral separation between unknown and known sources. , These are the L2 norms of U and E, respectively, used to standardize the correlation calculation results.

[0120] Furthermore, the system can update the network parameters using the gradient descent algorithm, with the parameter update formula as follows:

[0121]

[0122] in, For the updated network parameters, For the current parameter, For learning rate, This represents the gradient of the loss function with respect to the parameters. The loss value is backpropagated to each layer of the network using the backpropagation algorithm. The gradient step size is adaptively adjusted, and the convolutional kernel weights, bias parameters, and attention mechanism parameters of the encoder-decoder network are iteratively updated until the joint loss function value reaches a stable minimum, at which point the iteration stops. After optimization, L2 normalization is performed on the quantitative contribution vector to ensure that the sum of the contributions of all known sources is 1, eliminating the influence of dimensions. Simultaneously, the correlation between the spectral signals of the unknown pollutant components and the spectral feature matrices of known sources is calculated to verify their independence, ensuring that the spectral signals of the unknown pollutant components do not overlap with the spectral features of known sources. Finally, the normalized quantitative contribution vector and the spectral signals of the unknown pollutant components with independent spectral features are output.

[0123] In one embodiment, S5 of the water environment pollution source inference and tracing method provided by the present invention specifically includes the following steps:

[0124] S51: Perform spatial matching and interpolation processing on the quantitative contribution vector and the geographical coordinates in the pollution source feature database, associate the contribution value of each pollution source with the corresponding geographical coordinate point, and perform spatial interpolation rendering on the watershed map to generate a heat map of the spatial distribution of pollution contribution.

[0125] Specifically, the system retrieves the quantitative contribution vector and performs spatial matching and interpolation with the geographic coordinates in the pollution source feature database. Using the unique identifier of each pollution source as the core association, the system binds the quantitative contribution value of each pollution source to its corresponding geographic coordinate point, simultaneously verifying the binding relationship. By comparing the data integrity corresponding to the identifier, entries with missing coordinates or abnormal contribution values ​​are removed, avoiding spatial distribution distortion caused by data mismatch. To address the spatial data sparsity problem caused by the uneven distribution of potential pollution sources within the watershed, the system uses the Kriging interpolation algorithm to complete the data. The interpolation formula is:

[0126]

[0127] in, The estimated pollution contribution of the point to be interpolated. The spatial location to be interpolated on the watershed map. The number of known pollution sources around the interpolation point. The interpolation weights for the i-th known pollution source are... This represents the quantitative contribution value of the i-th known pollution source. Interpolation weights. The semivariance function is determined through calculation, and its formula is as follows:

[0128]

[0129] in, It is the semivariance of the sample points at a distance of h. This is the number of sample point pairs with a distance of h, where h is the spatial distance between sample points. This is used to quantify the correlation of spatial data and assign appropriate weights. After interpolation, the system uses the watershed digital elevation model as the base map, renders it hierarchically according to the contribution value interval, presents the pollution contribution intensity through color gradients, marks the location of known pollution sources and their contribution values, and generates a heat map of the spatial distribution of pollution contribution.

[0130] S52: Perform comprehensive analytical processing on the spectral signals and quantitative contribution vectors of unknown pollutants, analyze the intensity, spectral characteristics and spatiotemporal variation of unknown signals, assess the potential sources and contribution levels of unknown pollutants, and generate a quantitative analysis summary of unknown pollutants.

[0131] Specifically, the system performs comprehensive analytical processing on the spectral signals and quantitative contribution vectors of unknown pollutants to construct a multi-dimensional analytical system for mining core information. First, the intensity parameters of the spectral signals of unknown pollutants are extracted, and then the integral intensity of the signal is calculated using the following formula:

[0132]

[0133] in, The total intensity of the spectral signal of the unknown pollutant component. and These correspond to the minimum and maximum wavelengths of the spectral signal, respectively. For unknown pollutants at wavelength The spectral response values ​​at the monitoring points are used to quantify the relative content of unknown pollutants. Simultaneously, the positions, contours, and wavelength response patterns of characteristic fluorescence peaks of the unknown signals are extracted and compared with known spectral features in the pollution source characteristic database. Known spectral features with high similarity are selected as references. Combining the spatiotemporal distribution patterns of the quantitative contribution vector with the sampling time series of monitoring points, the spatiotemporal variation trends of the spectral signals of unknown pollutants are analyzed to determine the occurrence period, duration, and correlation with changes in the contribution of known pollution sources. Based on the comparison results and spatiotemporal variation patterns, the potential source range of unknown pollution is assessed, emission types and characteristics are inferred, and the contribution level is quantified by combining signal intensity calculation results to clarify its proportion in the overall pollution. The system integrates the results of intensity analysis, spectral feature comparison, spatiotemporal pattern judgment, and potential source assessment to extract core conclusions and generate a quantitative analysis summary of unknown pollutants.

[0134] S53: Synthesize the spatial distribution heat map of pollution contribution and the quantitative analysis summary into a report, integrate key data, charts and conclusions, and generate a complete graphic quantitative analysis report containing problem description, source tracing process, main findings and remediation recommendations according to a preset template.

[0135] Specifically, the system performs a report synthesis operation on the spatial distribution heatmap of pollution contribution and the quantitative analysis summary. It integrates all source tracing data and analysis results according to a preset template to generate a complete graphical quantitative analysis report. Before synthesis, the core data is organized and sorted, and the quantitative contribution vector statistics, spectral parameters of unknown pollutants, and spatial distribution data of the heatmap are classified and archived according to the report module to ensure accurate and efficient data retrieval. The spatial distribution heatmap of pollution contribution is embedded into the corresponding chapter of the report, supplemented with legends, coordinate systems, and data sources, clarifying the standards for heatmap interpretation. Following the preset template structure, four core contents are filled in sequentially. The problem description module outlines the current pollution status of the target watershed, the layout of monitoring points, and the core needs of source tracing; the source tracing process module describes the entire technical path from data acquisition, feature extraction, simulation calculation to signal demixing, listing key algorithms and core formulas; the main findings module integrates the spatial distribution patterns of the heatmap, the ranking of known pollution source contributions, and the analysis conclusions of unknown pollutants; and the governance recommendations module proposes control measures for high-contribution pollution sources and unknown potential pollution sources based on the source tracing results. The system performs logical verification and sentence standardization, optimizes the graphic and text layout format, and ensures that the report conforms to the presentation standards of patent documents and technical reports. It generates a complete graphic and text-based quantitative analysis report that can be directly output, providing data support for water environment management and pollution source control.

[0136] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0137] Based on the same inventive concept, this application also provides a water environment pollution source inference and tracing system for implementing the above-mentioned water environment pollution source inference and tracing method. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more water environment pollution source inference and tracing system embodiments provided below can be found in the limitations of the water environment pollution source inference and tracing method above, and will not be repeated here.

[0138] Preferably, such as Figure 3 As shown, the present invention provides a water environment pollution source inference and tracing system 600, which is configured with the following modules:

[0139] The pollution source feature database construction module 610 is used to perform fluorescence spectral analysis and spatiotemporal information acquisition on water samples from potential pollution sources within the target watershed, obtain the spectral fingerprint feature vector of each pollution source and its corresponding geographical location coordinates and sampling timestamp, and generate a pollution source feature database.

[0140] The mixed spectral feature extraction module 620 is used to enhance the temporal features of the raw water quality mixed spectral signals collected in real time at the monitoring points. It extracts stable spectral features from the raw signals within a continuous time window based on a one-dimensional convolutional neural network to generate mixed spectral signals.

[0141] The dynamic potential contribution source analysis module 630 is used to simulate and calculate the transport path and transport probability of each potential pollution source to the monitoring point under the current hydrological conditions based on the pollution source feature database, combined with the obtained monitoring point location information and the current sampling time, and generate a dynamic potential contribution source set containing contribution probability weights.

[0142] The spectral signal deep unmixing module 640 is used to perform deep unmixing network calculations based on the mixed spectral signal and the dynamic potential contributing source set. It calls an encoder-decoder network with embedded spatiotemporal constraints to separate the known source contributions and unknown source contributions in the mixed signal, and generates the quantitative contribution vector of each known pollution source and the spectral signal of the unknown pollution component.

[0143] The pollution source tracing result visualization module 650 is used to map the quantitative contribution vector to the spectral signal of unknown pollution components using geographic information. It combines the geographic coordinates in the pollution source feature database to map the contribution value to the geographic space and integrate them into a graphic report, generating a visualized pollution source tracing result that includes a spatial distribution map of pollution contribution and a quantitative analysis report.

[0144] Preferably, the pollution source characteristic database construction module 610 provided in this application is configured with the following units:

[0145] The pollution source water sample spectral scanning unit is used to perform three-dimensional fluorescence spectral scanning on the collected pollution source water samples. The excitation wavelength and emission wavelength form a two-dimensional spectral matrix. The position, intensity and contour information of the characteristic fluorescence peaks in the two-dimensional spectral matrix are extracted to generate the original high-dimensional spectral data.

[0146] The spectral data standardization processing unit is used to preprocess the raw high-dimensional spectral data, eliminate the fluorescence intensity distortion caused by water sample background scattering and concentration, and apply multivariate scattering correction to smooth the spectral baseline drift caused by suspended particulate matter, eliminate instrument noise and environmental interference, and generate a standardized spectral fingerprint feature vector.

[0147] The spectral feature spatiotemporal label binding unit is used to bind the spatiotemporal attributes of the spectral fingerprint feature vector, extract the GPS coordinates and UTC timestamp metadata from the sampling record, convert the coordinates into vector points in the standard projected coordinate system, hash the timestamp and the vector unique identifier and write it into the data header to generate a pollution source feature database with spatiotemporal labels.

[0148] Preferably, the hybrid spectral feature extraction module 620 provided in this application is configured with the following units:

[0149] The spectral signal sliding window segmentation unit is used to perform sliding window segmentation processing on the raw water quality mixed spectral signal output by the online spectrometer at the monitoring point. It divides the continuous time-series signal into multiple overlapping signal segments according to a fixed duration, and generates a serialized signal unit to be enhanced.

[0150] The serialized signal convolutional feature extraction unit is used to perform multi-layer convolution and pooling feature learning processing on the serialized signal units. It uses a one-dimensional convolutional neural network to slide and scan the serialized signal units in parallel using multiple convolutional kernels of different widths at different scales to capture the intensity change patterns and fluorescence peak morphology of the serialized signal units within the local wavelength range. The pooling layer downsamples the generated feature maps of each layer, retains the significant features with the highest response values ​​in the local regions of each feature map, and extracts deep abstract features.

[0151] The spectral feature dimensionality reduction and reconstruction unit is used to perform dimensionality reduction and reconstruction processing on deep abstract features. It uses a deconvolution layer with a symmetrical structure to progressively upsample the deep abstract features and map them back to the original signal space. Combined with the skip connections of the coding layer, it supplements the detailed information in the corresponding level convolution output to generate the final mixed spectral signal.

[0152] Preferably, the dynamic potential contribution source analysis module 630 provided in this application is configured with the following units:

[0153] The pollution source hydrological spatiotemporal parameter calculation unit is used to perform hydrological spatiotemporal alignment processing on the geographical location coordinates, sampling timestamps and real-time monitoring point locations and current sampling times in the pollution source feature database. Combined with the accessed real-time rainfall intensity, river flow data and digital elevation model obtained through remote sensing satellite image interpretation, it calculates the flow direction, confluence time and the path decay coefficient based on the pollutant half-life to generate a spatiotemporal parameter set for driving the simulation.

[0154] The distributed simulation unit for pollutant transport paths is used to perform distributed simulation of pollutant transport paths based on spatiotemporal parameter sets. It constructs a distributed hydrodynamic model based on graph theory and the Eulerian-Lagrange method and simulates the random walk and diffusion process of tens of thousands of virtual pollutant particles released from various potential source locations in the river network topology driven by hydrodynamic conditions. It generates simulation results that include the set of transport paths for each potential source and the arrival time range.

[0155] The dynamic potential contributing source weight assignment unit is used to perform probability matching processing between the arrival time range in the simulation results and the current sampling time window. It calculates the transport probability of each potential source based on the spatiotemporal density of particles appearing at the monitoring point, and selects high-probability pollution sources according to the preset probability threshold. At the same time, it normalizes the transport probability of the selected high-probability pollution sources and assigns it as the contribution weight to generate a dynamic potential contributing source set containing contribution probability weights.

[0156] Preferably, the spectral signal deep demixing module 640 provided in this application is configured with the following units:

[0157] The input tensor construction unit of the spectral unmixing network is used to perform structured input construction processing on the spectral fingerprint feature vectors corresponding to the mixed spectral signal and the dynamic potential contributing source set. The mixed spectral signal is used as the target vector, the spectral fingerprint feature vectors of all pollution sources in the dynamic potential contributing source set are arranged in order to form the source feature matrix, and the contribution probability weights corresponding to the dynamic potential contributing source set are transformed into the attention mask matrix to generate the standard input tensor of the encoder-decoder network.

[0158] The multi-head spatiotemporal attention spectral deentanglement encoding and decoding unit is used to encode and deentangle the standard input tensor based on the multi-head spatiotemporal attention mechanism. It captures the spectral dependence of the mixed signal through the multi-head self-attention layer of the encoder, and injects the contribution probability weight as the prior knowledge of the encoder through cross-attention layer. In the latent space of the decoder, the coded features are separated into components related to the known source and independent residual components using the deentanglement regularization term, and the contribution distribution of the known source and the preliminary estimate of the unknown component are reconstructed respectively.

[0159] The joint optimization and normalization unit for spectral unmixing results is used to jointly optimize and normalize the contribution distribution of known sources and the preliminary estimate of unknown components. Based on the gradient descent algorithm, a joint loss function consisting of spectral reconstruction error, sparsity penalty of known source contribution, and unknown-known source spectral separation is constructed. The network parameters are iteratively updated through backpropagation to minimize the joint loss function, and the output is a normalized quantitative contribution vector and spectral signals of unknown pollutant components with independent spectral characteristics.

[0160] Preferably, the pollution source tracing result visualization module 650 provided in this application is configured with the following units:

[0161] The pollution contribution spatial mapping rendering unit is used to perform spatial matching and interpolation processing between the quantitative contribution vector and the geographical coordinates in the pollution source feature database. It associates the contribution value of each pollution source with the corresponding geographical coordinate point and performs spatial interpolation rendering on the watershed map to generate a heat map of the spatial distribution of pollution contribution.

[0162] The unknown pollutant component comprehensive analysis unit is used to comprehensively analyze and process the spectral signals and quantitative contribution vectors of unknown pollutants, analyze the intensity, spectral characteristics and spatiotemporal variation of unknown signals, assess the potential sources and contribution levels of unknown pollutants, and generate a quantitative analysis summary of unknown pollutants.

[0163] The pollution source tracing report synthesis unit is used to synthesize a report from a heat map of the spatial distribution of pollution contribution and a summary of quantitative analysis. It integrates key data, charts and conclusions, and generates a complete graphic quantitative analysis report containing a problem description, source tracing process, main findings and remediation recommendations according to a preset template.

[0164] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method for reasoning and tracing the sources of water pollution.

[0165] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for reasoning and tracing water pollution sources.

[0166] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0167] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0168] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for reasoning and tracing the sources of water pollution, characterized in that, Includes the following steps: S1: Perform fluorescence spectroscopy analysis and spatiotemporal information acquisition on water samples from potential pollution sources within the target watershed, obtain the spectral fingerprint feature vector of each pollution source and its corresponding geographical location coordinates and sampling timestamp, and generate a pollution source feature database; S2: Perform time-series feature enhancement on the raw water quality mixed spectral signal collected in real time at the monitoring point, and extract stable spectral features in the raw signal within a continuous time window based on a one-dimensional convolutional neural network to generate a mixed spectral signal; S3: Based on the pollution source feature database, combined with the acquired monitoring point location information and the current sampling time, simulate and calculate the transport path and transport probability of each potential pollution source reaching the monitoring point under the current hydrological conditions, and generate a dynamic set of potential contributing sources containing contribution probability weights. S4: Perform deep unmixing network calculation based on the mixed spectral signal and the set of dynamic potential contributing sources, call the encoder-decoder network with embedded spatiotemporal constraints to separate the known source contributions and unknown source contributions in the mixed signal, and generate the quantitative contribution vector of each known pollution source and the spectral signal of the unknown pollution component; S5: Map the quantitative contribution vector to the spectral signal of the unknown pollutant component using geographic information, combine the geographic coordinates in the pollution source feature database, map the contribution value to geographic space and integrate it into a graphic report, and generate a visualized pollution source tracing result that includes a spatial distribution map of pollution contribution and a quantitative analysis report.

2. The method according to claim 1, characterized in that, S1 includes: S11: Perform three-dimensional fluorescence spectroscopy scanning on the collected water samples from each pollution source. The excitation wavelength and emission wavelength form a two-dimensional spectral matrix. Extract the position, intensity, and contour information of the characteristic fluorescence peaks in the two-dimensional spectral matrix to generate the original high-dimensional spectral data. S12: Preprocess the original high-dimensional spectral data to eliminate the fluorescence intensity distortion caused by water sample background scattering and concentration, and apply multivariate scattering correction to smooth the spectral baseline drift caused by suspended particulate matter, eliminate instrument noise and environmental interference, and generate a standardized spectral fingerprint feature vector. S13: Perform spatiotemporal attribute binding on the spectral fingerprint feature vector, extract GPS coordinates and UTC timestamp metadata from the sampling record, convert the coordinates into vector points in the standard projected coordinate system, hash the timestamp and the vector unique identifier and write them into the data header to generate a pollution source feature database with spatiotemporal labels.

3. The method according to claim 1, characterized in that, S2 includes: S21: Perform sliding window segmentation processing on the raw water quality mixed spectral signal output by the online spectrometer at the monitoring point, divide the continuous time-series signal into multiple overlapping signal segments according to a fixed duration, and generate a serialized signal unit to be enhanced. S22: Perform multi-layer convolution and pooling feature learning processing on the serialized signal unit. Use a one-dimensional convolutional neural network to slide and scan the serialized signal unit in parallel using multiple convolutional kernels of different widths at different scales to capture the intensity change pattern and fluorescence peak shape of the serialized signal unit within the local wavelength range. Then, downsample the generated feature maps of each layer through pooling layers to retain the significant features with the highest response values ​​in the local regions of each feature map and extract deep abstract features. S23: Perform dimensionality reduction and reconstruction on the deep abstract features. The deep abstract features are gradually upsampled and mapped back to the original signal space through a deconvolution layer with a symmetrical structure. The skip connections of the coding layer are combined to supplement the detailed information in the corresponding level convolution output to generate the final mixed spectral signal.

4. The method according to claim 1, characterized in that, S3 includes: S31: Perform hydrological spatiotemporal alignment processing on the geographical location coordinates, sampling timestamps, real-time monitoring point locations, and current sampling time in the pollution source feature database. Combine the accessed real-time rainfall intensity, river flow data, and digital elevation model obtained through remote sensing satellite image interpretation to calculate the flow direction, confluence time, and the path decay coefficient based on the pollutant half-life to generate a spatiotemporal parameter set for driving the simulation. S32: Based on the spatiotemporal parameter set, perform distributed simulation of pollutant transport paths, construct a distributed hydrodynamic model based on graph theory and Eulerian-Lagrange method, and simulate the random walk and diffusion process of tens of thousands of virtual pollutant particles released from each potential source location in the river network topology driven by hydrodynamic conditions, generating simulation results that include the set of transport paths of each potential source and the arrival time range. S33: Perform probability matching processing on the arrival time range in the simulation results and the current sampling time window, calculate the transport probability of each potential source based on the spatiotemporal density of particles appearing at the monitoring point, and screen out high-probability pollution sources according to the preset probability threshold. At the same time, normalize the transport probability of the screened high-probability pollution sources and assign it as a contribution weight to generate a dynamic potential contributing source set containing contribution probability weights.

5. The method according to claim 1, characterized in that, S4 includes: S41: Perform structured input construction processing on the spectral fingerprint feature vectors corresponding to the mixed spectral signal and the dynamic potential contributing source set, take the mixed spectral signal as the target vector, arrange the spectral fingerprint feature vectors of all pollution sources in the dynamic potential contributing source set in order as the source feature matrix, and convert the contribution probability weights corresponding to the dynamic potential contributing source set into the attention mask matrix to generate the standard input tensor of the encoder-decoder network. S42: The standard input tensor is encoded and de-entangled and decoded based on a multi-head spatiotemporal attention mechanism. The spectral dependence of the mixed signal is captured by the multi-head self-attention layer of the encoder, and the contribution probability weight is injected as prior knowledge of the encoder through cross-attention layer. In the latent space of the decoder, the encoded features are separated into components related to the known source and independent residual components using the de-entanglement regularization term, and the contribution distribution of the known source and the preliminary estimate of the unknown component are reconstructed respectively. S43: Perform joint optimization and normalization on the contribution distribution of known sources and the preliminary estimate of unknown components. Construct a joint loss function based on gradient descent algorithm, which consists of spectral reconstruction error, sparsity penalty of known source contribution, and unknown-known source spectral separation. Iteratively update network parameters through backpropagation to minimize the joint loss function, and output a normalized quantitative contribution vector and spectral signals of unknown pollutant components with independent spectral features.

6. The method according to claim 5, characterized in that, The expression for the joint loss function is: in, For the joint loss function value, The input is a mixed spectral signal vector. The source feature matrix is ​​composed of known source spectral fingerprint feature vectors. This is the quantitative contribution vector of the known sources to be optimized. For the unknown pollutant component spectral signal vector to be optimized, , These are the positive coefficient hyperparameters used to balance the weights of various losses.

7. The method according to any one of claims 1-6, characterized in that, S5 includes: S51: Perform spatial matching and interpolation processing on the quantitative contribution vector and the geographical coordinates in the pollution source feature database, associate the contribution value of each pollution source with the corresponding geographical coordinate point, and perform spatial interpolation rendering on the watershed map to generate a heat map of the spatial distribution of pollution contribution. S52: Perform comprehensive analytical processing on the spectral signal of the unknown pollutant component and the quantitative contribution vector, analyze the intensity, spectral characteristics and spatiotemporal variation of the unknown signal, assess the potential sources and contribution levels of the unknown pollutant, and generate a quantitative analysis summary of the unknown pollutant component. S53: Combine the spatial distribution heat map of pollution contribution with the quantitative analysis summary to form a report, integrate key data, charts and conclusions, and generate a complete graphic quantitative analysis report containing problem description, source tracing process, main findings and governance recommendations according to a preset template.

8. A source tracing system for water pollution sources, characterized in that, The system includes: The pollution source feature database construction module is used to perform fluorescence spectral analysis and spatiotemporal information acquisition on water samples from potential pollution sources within the target watershed, obtain the spectral fingerprint feature vector of each pollution source and its corresponding geographical location coordinates and sampling timestamp, and generate a pollution source feature database. The mixed spectral feature extraction module is used to enhance the temporal features of the raw water quality mixed spectral signals collected in real time from the monitoring points. It extracts stable spectral features from the raw signals within a continuous time window based on a one-dimensional convolutional neural network to generate mixed spectral signals. The dynamic potential contribution source analysis module is used to simulate and calculate the transport path and transport probability of each potential pollution source to the monitoring point under the current hydrological conditions based on the pollution source feature database, combined with the acquired monitoring point location information and the current sampling time, and generate a dynamic potential contribution source set containing contribution probability weights. The deep unmixing module for spectral signals is used to perform deep unmixing network calculations based on the mixed spectral signals and the set of dynamic potential contributing sources. It calls an encoder-decoder network with embedded spatiotemporal constraints to separate the known source contributions and unknown source contributions in the mixed signals, and generates quantitative contribution vectors of each known pollution source and spectral signals of unknown pollution components. The pollution source tracing result visualization module is used to perform geographic information mapping between the quantitative contribution vector and the spectral signal of the unknown pollution component. Combined with the geographical coordinates in the pollution source feature database, the contribution value is mapped to the geographic space and integrated into a graphic report, generating a visualized pollution source tracing result that includes a spatial distribution map of pollution contribution and a quantitative analysis report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images