A pollution traceability method and system based on intelligent fingerprint spectrum library matching

Abstract

The application discloses a kind of pollution traceability method and system based on intelligent fingerprint library matching.Thereby obtaining the historical data of each process production area of target industrial agglomeration area, the fingerprint library containing pollutant type, concentration and change information is constructed;Using machine learning model to calculate the SHAP value of each characteristic pollutant as initial weight, and combining environmental persistence and detection frequency to construct correction coefficient, after fusion, specific weight vector is obtained;Based on this weight, a multi-modal mixed kernel density estimation model of each emission area under multiple processes is constructed using kernel density estimation method;By calculating the similarity score of the concentration distribution of pollutants in the sample to be traced and the fingerprint library, the matching degree of each pollution source and the sample is determined, thereby realizing the accurate and rapid traceability of pollution emission area.The application effectively improves the accuracy and adaptability of pollutant traceability in complex industrial environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of environmental monitoring, and in particular to a pollution source tracing method and system based on intelligent fingerprint library matching. BACKGROUND

[0002] Traditional pollution tracing methods have significant technical bottlenecks in the fingerprint matching link. Existing methods usually rely on simple comparison of pollutant concentrations (such as Euclidean distance or Pearson correlation coefficient), and the matching degree calculation dimension is single, which cannot effectively quantify and evaluate the specificity and significance difference of different pollutants as tracing diagnostic markers. This leads to the signal of key indicators being easily overwhelmed by high-concentration background interferents, reducing the sensitivity and accuracy of source tracing.

[0003] At the same time, in the face of complex mixed pollution scenarios in industrial clusters, existing methods lack effective decoupling ability for multi-source pollution contribution, and cannot accurately strip and quantify the respective contribution of multiple potential pollution sources to the receptor sample, which easily causes source misjudgment or omission. In addition, traditional methods are not adaptive to the evolution of pollution source characteristics over time (such as process modification and process fluctuation), and lack quantitative evaluation of the reliability of matching results.

[0004] These problems seriously restrict the accuracy, interpretability and engineering practicality of pollution tracing technology in actual complex environmental management. Therefore, it is urgent to develop an intelligent fingerprint matching and tracing method that can integrate multi-dimensional information, have self-adaptive learning ability and be able to evaluate uncertainty, in order to improve the precision decision support level of industrial pollution tracing. SUMMARY

[0005] To solve at least one of the above technical problems, the present application provides a pollution source tracing method and system based on intelligent fingerprint library matching.

[0006] The first aspect of the present application provides a pollution source tracing method based on intelligent fingerprint library matching, comprising:

[0007] Obtaining historical environmental sample monitoring data of each process production area in a target industrial cluster, and constructing a pollution fingerprint library of the target industrial cluster according to the historical environmental sample monitoring data, wherein the pollution fingerprint library includes the characteristic pollutant type and concentration distribution, concentration time series change information of each process production area;

[0008] Calculating the SHAP value of each characteristic pollutant for pollution source classification based on a machine learning model as an initial diagnostic weight, and introducing an environmental persistence parameter and a detection frequency to construct a correction coefficient, and weighting and fusing the initial diagnostic weight and the correction coefficient to obtain a specificity weight vector of the characteristic pollutant;

[0009] According to the specific weight vector of the feature pollutants, the concentration data of each feature pollutant is fitted with a probability density by using a kernel density estimation method, and a multi-modal mixed kernel density estimation model of each process production area under a multi-process operation mode is constructed;

[0010] According to the multi-modal mixed kernel density estimation model, a similarity score of the measured concentration of each type of pollutant in the sample to be traced and the concentration distribution of each pollutant in the pollutant fingerprint library is calculated, and the matching degree of the pollution source of each process production area and the sample to be traced is determined according to the similarity score;

[0011] According to the matching degree, the source area of the pollutant is determined.

[0012] In the scheme, the historical environmental sample monitoring data of each process production area in the target industrial agglomeration area is obtained, and a pollutant fingerprint library of the target industrial agglomeration area is constructed according to the historical environmental sample monitoring data, wherein the pollutant fingerprint library includes the type and concentration distribution of the feature pollutants of each process production area, and the concentration time sequence change information, specifically:

[0013] The environmental sample monitoring data of each process production area in the target industrial agglomeration area in a preset historical period is obtained, the environmental sample includes wastewater, underground water and soil, and solid waste sample, the type of the pollutant in the environmental sample monitoring data is identified, and the concentration data of each pollutant and the corresponding sampling time and spatial position information are determined;

[0014] The concentration data of all types of pollutants in the environmental sample monitoring data is standardized, the covariance matrix of the standardized concentration data is calculated based on the principal component analysis method, the covariance matrix is decomposed to obtain eigenvalues and corresponding eigenvectors, the loading coefficient of each pollutant on each principal component is calculated according to the eigenvectors, and the pollutants with absolute value of the loading coefficient greater than a preset loading threshold are determined as candidate feature pollutants;

[0015] The information gain algorithm is introduced to calculate the information gain value of each candidate feature pollutant for the classification of the process production area, and the candidate feature pollutants with the information gain value greater than a preset information gain threshold are determined as the feature pollutants of each process production area;

[0016] According to the concentration data, sampling time information and spatial position information of the feature pollutants in the preset historical period, the concentration distribution statistical characteristics and concentration time sequence change characteristics of the feature pollutants in each process production area are calculated;

[0017] The type, concentration distribution statistical characteristics and concentration time sequence change characteristics of the feature pollutants of each process production area are integrated to construct the pollutant fingerprint library of the target industrial agglomeration area.

[0018] In this scheme, the SHAP value of each characteristic pollutant for pollution source classification is calculated based on the machine learning model as the initial diagnosis weight, and an environmental persistence parameter and a detection frequency are introduced to construct a correction coefficient. The initial diagnosis weight and the correction coefficient are weighted and fused to obtain a specific weight vector of the characteristic pollutant, specifically:

[0019] Based on the gradient boosting decision tree algorithm as the basic classifier, the concentration data of all pollutants in the environmental samples of each process production area in a preset historical period is taken as the input feature, and the corresponding process production area is taken as the output category for training to obtain a pollution source classification model;

[0020] Based on the SHAP interpreter, the average absolute SHAP value of the pollution source classification model in the process of classifying each characteristic pollutant into a process production area is calculated, and the average absolute SHAP value is taken as the initial diagnosis weight W1 of the characteristic pollutant for pollution source classification;

[0021] An environmental persistence parameter of each characteristic pollutant is obtained, including a compound half-life and an octanol-water partition coefficient. The environmental persistence parameter is normalized and weighted to obtain an environmental persistence correction coefficient W2;

[0022] The detection frequency f of each characteristic pollutant in the environmental samples of the corresponding process production area in a preset historical period is calculated, and the detection frequency is logarithmically transformed to obtain a detection frequency correction coefficient W3=log(1+f);

[0023] After normalization, W1, W2, and W3 are multiplied to obtain a specific weight vector of each characteristic pollutant .

[0024] In this scheme, according to the specific weight vector of the characteristic pollutant, a probability density fitting of the concentration data of each characteristic pollutant is performed by using a kernel density estimation method, and a multi-modal mixed kernel density estimation model of each process production area under a multi-process operation mode is constructed, specifically:

[0025] The concentration change data of each characteristic pollutant in the environmental samples of the target process production area in a preset historical period is subjected to kernel density estimation by using an Epanechnikov quadratic kernel function K(u) as a basic kernel function, and the kernel function is defined as:

[0026] ,

[0027] Wherein, u is the normalized concentration variable;

[0028] Based on the Silverman empirical rule, the reference bandwidth is calculated :

[0029] ,

[0030] wherein, sigma is the standard deviation of concentration data, N is the sample size of concentration variation data;

[0031] According to the local density, the reference bandwidth is dynamically adjusted, and the local bandwidth of each sample point is calculated :

[0032] ,

[0033] wherein, g is the geometric mean of the initial density estimate, is the initial density estimate value of the sample point ;

[0034] A time decay factor w(t) is introduced to weight the preset historical period environmental samples, reflecting the change of the emission characteristics of the pollution source with time, and the time decay factor is expressed as:

[0035] ,

[0036] wherein, Delta t is the number of months from the collection of the environmental samples, is the time constant;

[0037] Based on the weighted samples, the probability density estimate of the jth characteristic pollutant under the given pollution source condition is constructed:

[0038] ,

[0039] The mean shift clustering algorithm is used to identify the multi-modal of the concentration variation data, and the concentration distribution mode corresponding to different production processes is divided, and the multi-modal mixed kernel density estimation model is constructed according to the concentration distribution mode of each production process, and the multi-modal mixed kernel density estimation model is expressed as:

[0040] ,

[0041] wherein, M is the number of identified process modes, is the prior probability of the kth process, is the kernel density estimation function corresponding to the kth process.

[0042] In the scheme, the similarity score of the measured concentration of each type of pollutant in the sample to be traced and the concentration distribution of each pollutant in the pollutant fingerprint spectrum library is calculated according to the multi-modal mixed kernel density estimation model, and the matching degree of each process production area pollution source and the sample to be traced is determined according to the similarity score, specifically:

[0043] Obtain the type data of the contaminant substances in the sample to be traced and the measured concentration value c* of each contaminant substance, and calculate the conditional probability density of the measured concentration value of each contaminant substance according to the multi-modal mixed kernel density estimation model ;

[0044] Find a matching process production area of the contaminant substance with the same contaminant substance type as the sample to be traced, and extract the maximum value max of the probability density of each characteristic contaminant in the matching process production area of the contaminant substance according to the multi-modal mixed kernel density estimation model ;

[0045] Calculate the concentration similarity score S(c*) between each contaminant substance of the sample to be traced and each characteristic contaminant:

[0046] ,

[0047] According to the specificity weight vector of each characteristic contaminant of the process production area and the concentration similarity score, the comprehensive matching degree SS of the sample to be traced and each process emission area pollution source in the contaminant fingerprint library is calculated by weighted fusion, and the calculation formula is:

[0048] ,

[0049] Wherein, n is the total number of characteristic contaminants defined for the target process production area in the contaminant fingerprint library, is the specific weight of the i th characteristic contaminant, is the concentration similarity of the i th characteristic contaminant.

[0050] In the scheme, the contaminant source area is determined according to the matching degree, and specifically:

[0051] According to the comprehensive matching degree SS of the sample to be traced and each process production area pollution source in the contaminant fingerprint library, all process production areas are sorted in descending order of comprehensive matching degree to form a matching degree sorting list;

[0052] Set a matching degree threshold T, and mark the process production areas with a comprehensive matching degree SS greater than or equal to the threshold T in the matching degree sorting list as candidate pollution sources. When the number of candidate pollution sources is 1, the process production area corresponding to the candidate pollution source is directly determined as the final pollution source area;

[0053] When the number of candidate pollution sources is greater than 1, multiple candidate pollution sources are analyzed to construct a contribution rate calculation model based on constrained least squares:

[0054] ,

[0055] ,

[0056] wherein, is a pollutant concentration vector of a to-be-traced sample, is a characteristic pollutant concentration template vector of the i th candidate pollution source in the pollutant fingerprint library, is a contribution ratio of the i th candidate pollution source to the to-be-traced sample, and m is the number of candidate pollution sources, represents the square of the Euclidean norm;

[0057] solving the contribution rate calculation model to obtain the contribution ratio of each candidate pollution source determining the process production area corresponding to the candidate pollution source with the largest contribution ratio as the final pollution source area.

[0058] The second aspect of the present application also provides a pollutant tracing system based on intelligent fingerprint library matching, which comprises a memory and a processor, the memory comprises a pollutant tracing method based on intelligent fingerprint library matching, and the pollutant tracing method based on intelligent fingerprint library matching is executed by the processor to realize the following steps:

[0059] obtaining historical environmental sample monitoring data of each process production area in a target industrial agglomeration area, and constructing a pollutant fingerprint library of the target industrial agglomeration area according to the historical environmental sample monitoring data, wherein the pollutant fingerprint library comprises characteristic pollutant types and concentration distributions, concentration time sequence change information of each process production area;

[0060] calculating SHAP values of each characteristic pollutant for pollution source classification as initial diagnosis weights based on a machine learning model, introducing an environmental persistence parameter and a detection frequency to construct a correction coefficient, and weighting and fusing the initial diagnosis weights and the correction coefficient to obtain a specificity weight vector of the characteristic pollutant;

[0061] according to the specificity weight vector of the characteristic pollutant, using a kernel density estimation method to perform probability density fitting on concentration data of each characteristic pollutant, and constructing a multi-modal mixed kernel density estimation model of each process production area under a multi-process operation mode;

[0062] calculating similarity scores of measured concentrations of each type of pollutant in the to-be-traced sample and concentration distributions of pollutants in the pollutant fingerprint library according to the multi-modal mixed kernel density estimation model, and determining matching degrees of pollution sources of each process production area and the to-be-traced sample according to the similarity scores;

[0063] performing pollutant tracing according to the matching degrees to determine a pollution source area.

[0064] The application discloses a pollution source tracing method and system based on intelligent fingerprint spectrum library matching. The application discloses a pollution source tracing method and system based on intelligent fingerprint spectrum library matching. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 A flow chart of a pollution source tracing method based on intelligent fingerprint spectrum library matching is shown;

[0066] Figure 2 A flow chart of constructing a pollution fingerprint spectrum library of a target industrial agglomeration area is shown;

[0067] Figure 3 A flow chart of obtaining a specific weight vector of a characteristic pollutant is shown;

[0068] Figure 4 A block diagram of a pollution source tracing system based on intelligent fingerprint spectrum library matching is shown. DETAILED DESCRIPTION

[0069] In order to more clearly understand the above-mentioned purposes, features and advantages of the present application, the present application will be further described in detail below with reference to the drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0070] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0071] Figure 1 A flow chart of a pollution source tracing method based on intelligent fingerprint spectrum library matching is shown.

[0072] As Figure 1 shown, the first aspect of the present application provides a pollution source tracing method based on intelligent fingerprint spectrum library matching, comprising:

[0073] S102, obtain historical environmental sample monitoring data of each process production area in the target industrial agglomeration area, and construct a pollutant fingerprint atlas library of the target industrial agglomeration area according to the historical environmental sample monitoring data, wherein the pollutant fingerprint atlas library includes characteristic pollutant types and concentration distribution, concentration time sequence change information of each process production area;

[0074] S104, calculating SHAP value of each characteristic pollutant for pollution source classification as initial diagnosis weight based on the machine learning model, introducing environmental persistence parameter and detection frequency to construct correction coefficient, weighting and fusing the initial diagnosis weight and the correction coefficient to obtain specificity weight vector of the characteristic pollutant;

[0075] S106, fitting the concentration data of each characteristic pollutant by probability density according to the specificity weight vector of the characteristic pollutant, and constructing a multi-modal mixed kernel density estimation model of each process production area under multi-process operation mode;

[0076] S108, calculating similarity score of measured concentration of each type of pollutant in the sample to be traced and concentration distribution of each pollutant in the pollutant fingerprint atlas library according to the multi-modal mixed kernel density estimation model, and determining matching degree of each process production area pollution source and the sample to be traced according to the similarity score;

[0077] S110, performing pollutant tracing according to the matching degree to determine the source area of the pollutant.

[0078] It should be noted that the historical environmental sample data of each process production area in the target industrial agglomeration area is collected by the system, and the characteristic pollutants capable of significantly representing different pollution sources are intelligently screened therefrom to construct a comprehensive fingerprint library including the type, concentration distribution and time sequence change information of the pollutants. Then, in order to scientifically evaluate the importance of different characteristic pollutants in the source tracing diagnosis, the SHAP value of each pollutant to the pollution source classification is calculated as the initial weight by using the machine learning model, and a correction coefficient is further constructed by combining the persistence parameter and the actual detection frequency of the pollutant in the environment, and a weight vector capable of dynamically reflecting the specific contribution of the pollutant is formed by weighted fusion. Then, based on the weight vector, the probability density fitting of the historical concentration data of each characteristic pollutant is carried out by using the kernel density estimation technology, and the multi-modal mixed distribution model capable of characterizing the production characteristics of the pollution source under multiple processes is accurately constructed by adaptive bandwidth optimization and time decay weighting. When the sample to be traced needs to be analyzed, the similarity score of the measured concentration of the pollutant in the sample to the concentration distribution of the corresponding pollutant of each source in the fingerprint library is calculated, and the foregoing specific weight is used for weighted fusion, and finally the comprehensive matching degree of the sample to be traced to each potential pollution source is calculated. According to the preset matching degree judgment threshold and the contribution rate analysis model, the main pollution influence area can be accurately identified, so as to realize accurate source tracing.

[0079] Figure 2 A flow chart for constructing a pollutant fingerprint library of a target industrial agglomeration area is shown.

[0080] According to the embodiment of the present application, the historical environmental sample monitoring data of each process production area in the target industrial agglomeration area is obtained, and the pollutant fingerprint library of the target industrial agglomeration area is constructed according to the historical environmental sample monitoring data. The pollutant fingerprint library includes the type and concentration distribution of the characteristic pollutants of each process production area, and the concentration time sequence change information, specifically:

[0081] The environmental sample monitoring data of each process production area in the target industrial agglomeration area in a preset historical period is obtained, the environmental sample includes wastewater, groundwater and soil, solid waste sample, the type of pollutants in the environmental sample monitoring data is identified, and the concentration data and corresponding sampling time and spatial position information of each pollutant are determined;

[0082] The concentration data of all types of pollutants in the environmental sample monitoring data is standardized, the covariance matrix of the standardized concentration data is calculated based on the principal component analysis method, the covariance matrix is decomposed to obtain the eigenvalue and the corresponding eigenvector, the load coefficient of each pollutant on each principal component is calculated according to the eigenvector, and the pollutant with an absolute value of the load coefficient greater than a preset load threshold is taken as a candidate characteristic pollutant;

[0083] introducing an information gain algorithm to calculate an information gain value of each candidate characteristic pollutant for process production area classification, determining a candidate characteristic pollutant with an information gain value greater than a preset information gain threshold as a characteristic pollutant of each process production area;

[0084] According to the concentration data, sampling time information and spatial position information of the characteristic pollutant in a preset historical period, the concentration distribution statistical characteristics and concentration time sequence change characteristics of the characteristic pollutant in each process production area are calculated.

[0085] The characteristic pollutant type, concentration distribution statistical characteristics and concentration time sequence change characteristics of each process production area are integrated to construct a pollutant fingerprint spectrum library of the target industrial agglomeration area.

[0086] It should be noted that the concentration distribution statistical characteristics include concentration mean, concentration variance and concentration skewness, and the concentration time sequence change characteristics include the trend slope of concentration change over time and the periodic fluctuation amplitude. Through the double screening mechanism of principal component analysis and information gain algorithm, those pollutants which can not only represent the main variance characteristics of different pollution sources but also have significant contribution to source classification discrimination are accurately identified as characteristic pollutants, thereby significantly improving the representativeness and specificity of the fingerprint spectrum. Furthermore, by quantifying the concentration distribution statistical characteristics and time sequence change law of the characteristic pollutants, the constructed fingerprint spectrum not only contains static concentration information, but also dynamically reflects the emission mode of the pollution source at multiple spatiotemporal scales.

[0087] Figure 3 A flowchart for obtaining the specificity weight vector of the characteristic pollutant is shown.

[0088] According to an embodiment of the present application, the SHAP value of each characteristic pollutant for pollution source classification is calculated based on a machine learning model as an initial diagnosis weight, and an environmental persistence parameter and a detection frequency are introduced to construct a correction coefficient, and the initial diagnosis weight and the correction coefficient are weighted and fused to obtain a specificity weight vector of the characteristic pollutant, specifically:

[0089] Based on the gradient boosting decision tree algorithm as the basic classifier, the concentration data of all pollutants in each process production area environmental sample in a preset historical period are taken as input features, and the corresponding process production area is taken as the output category for training to obtain a pollution source classification model.

[0090] Based on the SHAP interpreter, the average absolute SHAP value of the pollution source classification model in the process of classifying each characteristic pollutant into a process production area is calculated, and the average absolute SHAP value is taken as the initial diagnosis weight W1 of the characteristic pollutant for pollution source classification.

[0091] An environmental persistence parameter of each characteristic pollutant is acquired, including a compound half-life and an octanol-water partition coefficient, the environmental persistence parameter is normalized and weighted summation is performed to construct an environmental persistence correction coefficient W2;

[0092] A detection frequency f of each characteristic pollutant in a preset historical period environmental sample of a production area of the process is calculated, logarithmic transformation is performed on the detection frequency to obtain a detection frequency correction coefficient W3=log(1+f);

[0093] After normalization, W1, W2 and W3 are multiplied to obtain a specificity weight vector of each characteristic pollutant .

[0094] It should be noted that the SHAP value can objectively quantify the actual contribution of each characteristic pollutant in distinguishing different process production pollution sources from a data-driven perspective based on the trained gradient boosting decision tree model, thereby avoiding the deviation of subjective experience judgment. However, only considering the model discrimination ability is not enough, the actual behavior of the pollutant in the environment is also crucial; therefore, the environmental persistence correction coefficient W2 is introduced to reflect the signal attenuation characteristics of the pollutant in the migration process, to ensure that the pollutants that stably exist in the environment obtain a higher weight, and the detection frequency correction coefficient W3 is introduced to enhance the weight of the pollutants that continuously appear in the historical samples and represent the stable pollution characteristics of the pollution source. By weighting and fusing the three, the final specificity weight vector Wi not only contains the discrimination importance of the pollutant in the machine learning model, but also comprehensively considers the retention ability of the pollutant in the environment and the pollution stability of the pollution source, so that the weight distribution is more scientific, comprehensive and consistent with the actual logic of environmental management. By constructing a multi-dimensional weighted evaluation system, the rationality and robustness of the weight distribution of the characteristic pollutants are significantly improved, so that the subsequent traceability matching process can focus more on the pollutants with strong discrimination ability, stable environmental signal and typical pollution characteristics, and finally effectively improve the accuracy and reliability of the pollutant traceability.

[0095] According to the embodiment of the application, according to the specificity weight vector of the characteristic pollutant, a probability density fitting of the concentration data of each characteristic pollutant is performed by using a kernel density estimation method, and a multi-modal mixed kernel density estimation model of each process production area under a multi-process operation mode is constructed, specifically:

[0096] The concentration change data of each characteristic pollutant in the environmental sample of the target process production area in a preset historical period is subjected to kernel density estimation by using an Epanechnikov quadratic kernel function K(u) as a basic kernel function, and the kernel function is defined as:

[0097] ,

[0098] wherein u is the normalized concentration variable;

[0099] Calculate the reference bandwidth based on Silverman's empirical rule :

[0100] ,

[0101] wherein σ is the standard deviation of the concentration data, and N is the sample size of the concentration change data;

[0102] According to the local density, the reference bandwidth is dynamically adjusted, and the local bandwidth of each sample point is calculated :

[0103] ,

[0104] wherein g is the geometric mean of the initial density estimate, is the initial density estimate value of the sample point ;

[0105] A time decay factor w(t) is introduced to weight the preset historical period environmental samples, reflecting the change of the emission characteristics of the pollution source with time, and the time decay factor is expressed as:

[0106] ,

[0107] wherein Δt is the number of months from the collection of the environmental samples, is the time constant;

[0108] Based on the weighted samples, the probability density estimate of the jth characteristic pollutant under the given pollution source condition is constructed:

[0109] ,

[0110] The mean shift clustering algorithm is used to identify the multi-modal of the concentration change data, and the concentration distribution mode corresponding to different production processes is divided, and a multi-modal mixed kernel density estimation model is constructed according to the concentration distribution mode of each production process, and the multi-modal mixed kernel density estimation model is expressed as:

[0111] ,

[0112] wherein M is the number of identified process modes, is the prior probability of the kth process, is the kernel density estimation function corresponding to the kth process.

[0113] It should be noted that the Epanechnikov kernel function is used as a basic estimation tool, which has optimal efficiency in the sense of mean square error and can guarantee the statistical effectiveness of density estimation; then the benchmark bandwidth is calculated based on the Silverman empirical rule, which provides an initial scale benchmark for density estimation, but the fixed bandwidth is difficult to adapt to the multi-scale characteristics of the concentration distribution, so a local density adaptive mechanism is further introduced to dynamically adjust the bandwidth parameter by calculating the density value in the neighborhood of each sample point, so that the bandwidth is reduced in high-density areas to improve the distribution detail resolution, and the bandwidth is expanded in low-density areas to smooth random fluctuations and abnormal value interference, which enhances the fitting ability of the model to complex distribution patterns. By introducing a time decay factor, different weights are given according to the nearness of the sample collection time, so that recent data contribute more to model construction, thereby effectively capturing the evolution law of the characteristics of pollutants from pollution sources due to production process transformation and governance measures upgrading. Finally, the weighted concentration data are subjected to multi-modal recognition by the mean shift clustering algorithm, and the concentration distribution sub-mode corresponding to different production processes (such as normal production, equipment maintenance, accident discharge, etc.) is automatically divided, and the kernel density estimation function of each sub-mode is weighted and superimposed according to its occurrence probability, and finally a comprehensive probability model is formed, which can fully reflect the multi-state emission characteristics of pollution sources. The kernel density estimation method is used to depict the probability of the occurrence of pollutant concentration in each numerical range.

[0114] According to the embodiment of the present application, the similarity score of the measured concentration of each type of pollutant in the sample to be traced and the concentration distribution of each pollutant in the pollutant fingerprint library is calculated according to the multi-modal mixed kernel density estimation model, and the matching degree of each process production area pollution source and the sample to be traced is determined according to the similarity score, specifically:

[0115] The type of pollutant data and the measured concentration value c* of each pollutant in the sample to be traced are obtained, and the conditional probability density of each pollutant measured concentration value is calculated according to the multi-modal mixed kernel density estimation model ;

[0116] The matching process production area of the pollutant with the same pollutant type as the sample to be traced is found, and the maximum value max( ) of the probability density of each characteristic pollutant in the matching process production area of the pollutant is extracted according to the multi-modal mixed kernel density estimation model;

[0117] The concentration similarity score S(c*) between each pollutant in the sample to be traced and each characteristic pollutant is calculated:

[0118] ,

[0119] It should be noted that dividing the measured concentration conditional probability density of the pollutant in the sample by its maximum probability density in the corresponding pollution source fingerprint spectrum aims to standardize the similarity score to a comparable range between 0 and 1. Without standardization, the conditional probability density values ​​of different pollutants may differ by several orders of magnitude due to differences in concentration and distribution range, causing the probability differences of high-concentration pollutants to overwhelm the discrimination signal of low-concentration pollutants. By dividing by the maximum probability density of the pollutant in the source spectrum, it is equivalent to comparing the probability contribution of the measured concentration value with the most likely concentration value (i.e., the distribution peak), thus unifying the calculation result into a relative index representing the degree of proximity to the most typical pollution concentration of the source.

[0120] Based on the specific weight vector of each characteristic pollutant in the process production area Based on the concentration similarity score, the comprehensive matching degree SS between the sample to be traced and each industrial emission source in the pollutant fingerprint database is calculated through weighted fusion. The calculation formula is as follows:

[0121] ,

[0122] Where n represents the total number of characteristic pollutants defined for the target process production area in the pollutant fingerprint database. The specific weight of the i-th characteristic pollutant is... Let represent the similarity of the concentration of the i-th characteristic pollutant.

[0123] It should be noted that by using a multimodal hybrid kernel density estimation model to calculate the conditional probability density of the measured concentration of pollutants in the sample to be traced, the degree of agreement between the concentration value and the complete emission characteristics of the target pollution source under various production processes can be accurately characterized and quantified. This effectively captures the complex concentration distribution patterns of the pollution source under different states such as normal production, equipment maintenance, or accidental emissions. Furthermore, by standardizing the ratio of the conditional probability density to the maximum probability density of the pollutant in the corresponding pollution source map, the absolute probability value is successfully transformed into a relative similarity score that characterizes the degree to which the measured concentration approaches the most typical emission pattern of the pollution source. This process not only eliminates the comparability barrier caused by differences in concentration magnitude and distribution range of different pollutants, but also highlights the advantages of the multimodal model in identifying the matching degree between the sample and a specific emission process. Finally, by weighting and fusing the standardized similarity scores of each pollutant with specific weights that integrate the judgment importance, environmental persistence, and emission stability, a comprehensive matching degree is generated.

[0124] According to an embodiment of the present invention, the step of tracing the source of pollutants based on the matching degree to determine the source area of ​​the pollutants specifically includes:

[0125] According to the comprehensive matching degree SS of the to-be-traced sample and each process production area pollution source in the pollution fingerprint library, all process production areas are sorted from high to low according to the comprehensive matching degree to form a matching degree sorting list;

[0126] A matching degree judgment threshold T is set, process production areas with a comprehensive matching degree SS greater than or equal to the threshold T in the matching degree sorting list are marked as candidate pollution sources, and when the number of candidate pollution sources is 1, the process production area corresponding to the candidate pollution source is directly determined as the final pollution source area;

[0127] When the number of candidate pollution sources is greater than 1, mixed pollution source analysis is performed on the multiple candidate pollution sources, and a contribution rate calculation model based on constrained least squares is constructed:

[0128] ,

[0129] ,

[0130] Among them, is the pollution concentration vector of the to-be-traced sample, is the feature pollution concentration template vector of the i th candidate pollution source in the pollution fingerprint library, is the contribution proportion of the i th candidate pollution source to the to-be-traced sample, and m is the number of candidate pollution sources, represents the square of the Euclidean norm;

[0131] The contribution rate calculation model is solved to obtain the contribution proportion of each candidate pollution source The process production area corresponding to the candidate pollution source with the largest contribution proportion is determined as the final pollution source area.

[0132] It should be noted that after the comprehensive matching degree of the to-be-traced sample and the multiple process production area pollution sources in the pollution fingerprint library has been calculated, how to accurately determine the final pollution emission area from multiple matching degree results. When the comprehensive matching degrees of multiple pollution sources are all high and exceed the preset threshold, an accurate decision mechanism is needed to distinguish the primary and secondary contributions of these potential pollution sources to avoid the ambiguity of the tracing result. Therefore, the candidate pollution source set is quickly locked by matching degree sorting and threshold screening, and the contribution rate calculation model based on constrained least squares is further used to mathematically optimally analyze the specific contribution proportion of each candidate source to the unknown sample, so that the main responsible emission area is uniquely determined from multiple high-matching-degree pollution sources.

[0133] According to the embodiment of the present application, further comprising:

[0134] Acquire environmental basic data of the target industrial agglomeration area, the environmental basic data including high-precision geographic information data, historical meteorological statistical data and historical hydrological statistical data;

[0135] Based on the environmental basic data, simulate the migration path of the pollutant from each process production area to the preset sampling point by using a Lagrangian particle diffusion model or an Euler convection diffusion equation;

[0136] For each characteristic pollutant in the pollutant fingerprint spectrum library, the time-varying concentration distribution on each migration path is calculated according to physical and chemical parameters, so as to construct a migration path fingerprint correction library, the physical and chemical parameters including half-life, Henry constant, octanol-water partition coefficient, vapor pressure and deposition rate in different media, and the migration path fingerprint correction library stores the concentration attenuation correction coefficient matrix of each characteristic pollutant from the source to the sampling point under different environmental conditions;

[0137] When a sample to be traced needs to be analyzed, real-time actual meteorological and hydrological monitoring data at the sampling time are acquired, the closest simulation scenario in the migration path fingerprint correction library is matched according to the actual meteorological and hydrological monitoring data, and the corresponding concentration attenuation correction coefficient matrix is extracted;

[0138] The source concentration distribution statistical characteristics of the corresponding characteristic pollutant in the pollutant fingerprint spectrum library are inversely corrected by using the extracted concentration attenuation correction coefficient matrix, so as to obtain a group of virtual source fingerprint spectrums after migration path correction, which are closer to the actual situation of the sampling point.

[0139] According to the embodiment of the present application, the source concentration distribution statistical characteristics of the corresponding characteristic pollutant in the pollutant fingerprint spectrum library are inversely corrected by using the extracted concentration attenuation correction coefficient matrix, specifically as follows:

[0140] Read the concentration distribution statistical characteristics of the characteristic pollutant in the target process production area in the pollutant fingerprint spectrum library, the concentration distribution statistical characteristics including concentration mean value, concentration variance and concentration skewness;

[0141] From the concentration attenuation correction coefficient matrix extracted from the migration path fingerprint correction library, the average attenuation coefficient and the variance disturbance coefficient of the corresponding migration path of the characteristic pollutant are acquired, the concentration mean value is amplified and corrected according to the average attenuation coefficient, so as to compensate the overall attenuation effect of the pollutant in the migration process;

[0142] Meanwhile, the concentration variance is corrected according to the variance disturbance coefficient, so as to represent the amplification or reduction effect of the migration and diffusion process on the dispersion degree of the pollutant distribution;

[0143] Further, based on the corrected concentration mean and concentration variance, the probability density function of the feature pollutant in the corrected virtual source fingerprint is reconstructed by using the matrix estimation method, the probability density function obtained by the reconstruction is substituted into the multi-modal mixed kernel density estimation model to replace the original kernel density estimation model which does not consider the migration path, and is used to calculate the similarity score of the measured concentration of the pollutant in the sample to be traced and the fingerprint of the pollution source, so that the interference caused by the environmental migration is effectively offset in the subsequent matching degree calculation.

[0144] It should be noted that when the pollutants discharged by the pollution source migrate to the sampling point under complex and variable meteorological and hydrological conditions, the original component proportion and concentration distribution characteristics will be distorted and attenuated to different degrees due to the differences in the physical and chemical properties of each pollutant. The environmental interference caused by the migration path makes the similarity matching method based on the static fingerprint of the source decrease the accuracy of the tracing result in actual application, especially in the industrial agglomeration area with complex terrain or variable meteorological conditions. By introducing a pre-constructed migration path fingerprint correction library, the physical mechanism of the environmental migration process is creatively incorporated into the tracing model. Instead of simply comparing the sampling point data with the static fingerprint of the source, a high-precision environmental model and the specific physical and chemical parameters of the pollutant are used to simulate the dynamic propagation process of the pollutant from each potential emission source to the sampling point, and a correction coefficient matrix containing concentration attenuation and distribution deformation information is generated accordingly. In actual tracing, the system calls the corresponding correction coefficient according to the real environmental data at the sampling time to reconstruct the source fingerprint in reverse, generating a "virtual source fingerprint" that is more close to the actual situation of the sampling point. The interference effect of environmental migration is effectively removed from the sample data to be measured, or the source fingerprint is "deduced" to the expected state under the sampling point environment, so that the subsequent similarity calculation is carried out on a more fair basis that eliminates the interference of the path.

[0145] Figure 4 A block diagram of a pollutant tracing system based on intelligent fingerprint library matching according to the present application is shown.

[0146] The second aspect of the present application also provides a pollutant tracing system based on intelligent fingerprint library matching, which comprises a memory 401, a processor 402 and a communication interface 403, the memory comprises a pollutant tracing method program based on intelligent fingerprint library matching, the communication interface is used for data connection and communication between the memory and the processor, and when the pollutant tracing method program based on intelligent fingerprint library matching is executed by the processor, the following steps are realized:

[0147] Acquire historical environmental sample monitoring data of each process production area in a target industrial agglomeration area, and construct a pollutant fingerprint library of the target industrial agglomeration area according to the historical environmental sample monitoring data, wherein the pollutant fingerprint library comprises characteristic pollutant types and concentration distribution, concentration time sequence change information of each process production area;

[0148] Calculate SHAP values of each characteristic pollutant for pollution source classification as initial diagnostic weights based on a machine learning model, introduce an environmental persistence parameter and a detection frequency to construct a correction coefficient, and perform weighted fusion on the initial diagnostic weights and the correction coefficient to obtain a specificity weight vector of the characteristic pollutant;

[0149] According to the specificity weight vector of the characteristic pollutant, a probability density fitting is performed on the concentration data of each characteristic pollutant by using a kernel density estimation method, and a multi-modal mixed kernel density estimation model of each process production area under a multi-process operation mode is constructed.

[0150] According to the multi-modal mixed kernel density estimation model, a similarity score of the measured concentration of each type of pollutant in the sample to be traced and the concentration distribution of the pollutant in the pollutant fingerprint library is calculated, and the matching degree of each process production area pollution source and the sample to be traced is determined according to the similarity score.

[0151] According to the matching degree, the pollutant source area is determined.

[0152] The application discloses a pollutant source tracing method and system based on intelligent fingerprint library matching. By acquiring historical data of each emission area in a target industrial agglomeration area, a fingerprint library containing pollutant types, concentrations and change information is constructed. Using a machine learning model, SHAP values of each characteristic pollutant are calculated as initial weights, and a correction coefficient is constructed by combining environmental persistence and detection frequency. Based on this weight, a multi-modal mixed kernel density estimation model of each emission area under a multi-process is constructed by using a kernel density estimation method. By calculating the similarity score of the sample to be traced and the concentration distribution of the pollutant in the fingerprint library, the matching degree of each pollution source and the sample is determined, thereby realizing accurate and rapid source tracing of the pollution emission area. The application effectively improves the accuracy and adaptability of pollutant source tracing in complex industrial environments.

[0153] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware, the foregoing program can be stored in a computer readable storage medium, and the program executes the steps of the method embodiments when executed; and the foregoing storage medium includes a mobile storage device, a read-only memory (ROM), a random access memory (RAM), a magnetic disc or an optical disc, and various storage medium capable of storing program codes.

[0154] Alternatively, the integrated unit of the present application can be stored in a computer readable storage medium if it is realized in the form of a software function module and sold or used as an independent product. Based on such understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, and the computer software product is stored in a storage medium, includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the embodiments of the present application. The foregoing storage medium includes a mobile storage device, a ROM, a RAM, a magnetic disc or an optical disc, and various storage medium capable of storing program codes.

[0155] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A pollutant source tracing method based on intelligent fingerprint database matching, characterized in that, Includes the following steps: Historical environmental sample monitoring data of each process production area within the target industrial cluster are obtained. Based on the historical environmental sample monitoring data, a pollutant fingerprint spectrum library of the target industrial cluster is constructed. The pollutant fingerprint spectrum library includes characteristic pollutant types and concentration distribution, and concentration time-series change information of each process production area. The SHAP value of each characteristic pollutant for pollution source classification is calculated based on the machine learning model as the initial diagnostic weight, and the environmental persistence parameter and detection frequency are introduced to construct the correction coefficient. The initial diagnostic weight and the correction coefficient are weighted and fused to obtain the specific weight vector of the characteristic pollutant. Based on the specific weight vector of the characteristic pollutant, the kernel density estimation method is used to fit the probability density of the concentration data of each characteristic pollutant, and a multimodal mixed kernel density estimation model under the multi-process operation mode of each process production area is constructed. The similarity score between the measured concentration of each type of pollutant in the sample to be traced and the concentration distribution of each pollutant in the pollutant fingerprint database is calculated based on the multimodal hybrid kernel density estimation model. The matching degree between the pollution source of each process production area and the sample to be traced is determined based on the similarity score. Based on the matching degree, pollutant sources are traced to determine the source area of ​​the pollutants.

2. The pollutant source tracing method based on intelligent fingerprint database matching according to claim 1, characterized in that, The process involves acquiring historical environmental sample monitoring data for each process production area within the target industrial cluster, and constructing a pollutant fingerprint database for the target industrial cluster based on the historical environmental sample monitoring data. The pollutant fingerprint database includes characteristic pollutant types and concentration distributions, as well as temporal concentration variation information for each process production area. Specifically: The environmental sample monitoring data of each process production area in the target industrial cluster area within a preset historical period is obtained. The environmental samples include wastewater, groundwater, soil, and solid waste samples. The pollutant type is identified from the environmental sample monitoring data, and the concentration data of each pollutant and the corresponding sampling time and spatial location information are determined. The concentration data of all types of pollutants in the environmental sample monitoring data are standardized. The covariance matrix of the standardized concentration data is calculated based on the principal component analysis method. The covariance matrix is ​​decomposed into eigenvalues ​​and corresponding eigenvectors. The loading coefficient of each pollutant on each principal component is calculated based on the eigenvectors. Pollutants with an absolute value of the loading coefficient greater than a preset loading threshold are selected as candidate feature pollutants. An information gain algorithm is introduced to calculate the information gain value of each candidate characteristic pollutant for classifying the process production area. Candidate characteristic pollutants with information gain values ​​greater than a preset information gain threshold are identified as characteristic pollutants of each process production area. Based on the concentration data, sampling time information, and spatial location information of the characteristic pollutants within a preset historical period, calculate the statistical characteristics of the concentration distribution and the temporal variation characteristics of the concentration of the characteristic pollutants in each process production area; By integrating the characteristic pollutant types, concentration distribution statistics, and concentration temporal variation characteristics of each process production area, a pollutant fingerprint database for the target industrial cluster is constructed.

3. The pollutant source tracing method based on intelligent fingerprint database matching according to claim 1, characterized in that, The SHAP value of each characteristic pollutant for pollution source classification, calculated based on a machine learning model, is used as the initial diagnostic weight. An environmental persistence parameter and detection frequency are introduced to construct a correction coefficient. The initial diagnostic weight and the correction coefficient are then weighted and fused to obtain a specific weight vector for the characteristic pollutant. Specifically: The gradient boosting decision tree algorithm is used as the basic classifier. The concentration data of all pollutants in the environmental samples of each process production area within a preset historical period are used as input features, and the corresponding process production area is used as the output category for training to obtain a pollution source classification model. The average absolute SHAP value of the pollution source classification model is calculated based on the SHAP interpreter during the process of classifying each characteristic pollutant into a process production area. The average absolute SHAP value is then used as the initial diagnostic weight W1 for the crop characteristic pollutant classification of the pollution source. The environmental persistence parameters of each characteristic pollutant are obtained, including the compound half-life and octanol-water partition coefficient. The environmental persistence parameters are normalized and weighted summed to construct the environmental persistence correction coefficient W2. Calculate the detection frequency f of each characteristic pollutant in the environmental samples of the preset historical period in the production area of ​​its respective process, and perform logarithmic transformation on the detection frequency to obtain the detection frequency correction coefficient W3=log(1+f); After normalizing W1, W2, and W3, multiply them to obtain the specific weight vector for each characteristic pollutant. .

4. The pollutant source tracing method based on intelligent fingerprint database matching according to claim 1, characterized in that, The method involves using the kernel density estimation method to fit the probability density of the concentration data for each characteristic pollutant based on its specific weight vector, thereby constructing a multimodal hybrid kernel density estimation model for each process production area under multiple process operation modes. Specifically: Kernel density estimation was performed using the Epanechnikov quadratic kernel function K(u) as the basic kernel function for the concentration variation data of each characteristic pollutant in environmental samples from the target process production area over a pre-defined historical period. The kernel function is defined as follows: K(u)= , Where u is the standardized concentration variable; The baseline bandwidth h0 was calculated based on Silverman's rule of thumb. h0 = 1.06 × σ̂ × N⁻¹ / 5 ; Where σ̂ is the standard deviation of the concentration data, and N is the sample size of the concentration change data; The baseline bandwidth is dynamically adjusted based on the local density, and the local bandwidth hᵢ of each sample point cᵢ is calculated: hᵢ = h0 × λᵢ , λᵢ = (g / p̂(cᵢ)) 0 · 5 ; Where g is the geometric mean of the initial density estimate, and p̂(cᵢ) is the initial density estimate of the sample point cᵢ; A time decay factor w(t) is introduced to weight environmental samples over a preset historical period to reflect the changes in pollution source emission characteristics over time. The time decay factor is expressed as: w(t) = exp(-Δt / τ); Where Δt is the number of environmental samples collected in the months since the last collection, and τ is the time constant; Based on the weighted samples, construct the probability density estimate of the i-th characteristic pollutant under given pollution source conditions: ； Mean-shift clustering algorithm is used to identify multimodalities in concentration change data, classifying concentration distribution patterns corresponding to different production processes. A multimodal mixed kernel density estimation model is constructed based on the concentration distribution pattern of each production process. The multimodal mixed kernel density estimation model is expressed as follows: ； Where M represents the number of identified process modes. Let be the prior probability of the k-th process. This is the kernel density estimation function corresponding to the k-th process.

5. The pollutant source tracing method based on intelligent fingerprint database matching according to claim 1, characterized in that, The process involves calculating the similarity score between the measured concentration of each type of pollutant in the sample to be traced and the concentration distribution of each pollutant in the pollutant fingerprint database based on the multimodal hybrid kernel density estimation model, and determining the matching degree between the pollution source in each process production area and the sample to be traced based on the similarity score. Specifically: Obtain the pollutant type data and the measured concentration value c* of each pollutant in the sample to be traced, and calculate the conditional probability density of the measured concentration value of each pollutant according to the multimodal mixture kernel density estimation model. ; Find the production area of ​​the process that matches the pollutant type of the sample to be traced, and extract the maximum probability density (max) of each characteristic pollutant in the matching production area according to the multimodal mixed kernel density estimation model. ); Calculate the concentration similarity score S(c*) between each pollutant and each characteristic pollutant in the sample to be traced: S(c*) = / max( ); Based on the specific weight vector of each characteristic pollutant in the process production area Based on the concentration similarity score, the comprehensive matching degree SS between the sample to be traced and each industrial emission source in the pollutant fingerprint database is calculated through weighted fusion. The calculation formula is as follows: ； Where n represents the total number of characteristic pollutants defined for the target process production area in the pollutant fingerprint database. The specific weight of the i-th characteristic pollutant is... Let represent the similarity of the concentration of the i-th characteristic pollutant.

6. The pollutant source tracing method based on intelligent fingerprint database matching according to claim 1, characterized in that, The step of tracing pollutants based on the matching degree to determine the source area of ​​the pollutants specifically involves: Based on the comprehensive matching degree SS between the sample to be traced and the pollution source of each process production area in the pollutant fingerprint spectrum database, all process production areas are sorted from high to low according to the comprehensive matching degree to form a matching degree ranking list. Set a matching degree judgment threshold T, and mark the process production areas in the matching degree ranking list with a comprehensive matching degree SS greater than or equal to the threshold T as candidate pollution sources. When the number of candidate pollution sources is 1, directly determine the process production area corresponding to the candidate pollution source as the final pollution source area. When the number of candidate pollution sources is greater than 1, a mixed pollution source analysis is performed on multiple candidate pollution sources, and a contribution rate calculation model based on constrained least squares is constructed: ； subject to , ; in, Let be the pollutant concentration vector of the sample to be traced. Let i be the characteristic pollutant concentration template vector of the i-th candidate pollution source in the pollutant fingerprint database. Let m represent the contribution ratio of the i-th candidate pollution source to the samples to be traced, and m be the number of candidate pollution sources. Represents the square of the Euclidean norm; Solve the contribution rate calculation model to obtain the contribution ratio of each candidate pollution source. The process production area corresponding to the candidate pollution source with the largest contribution ratio is determined as the final pollution source area.

7. A pollutant source tracing system based on intelligent fingerprint database matching, characterized in that, The pollutant tracing system based on intelligent fingerprint database matching includes a storage device and a processor. The storage device includes a pollutant tracing method program based on intelligent fingerprint database matching. When the processor executes the pollutant tracing method program based on intelligent fingerprint database matching, it performs the following steps: Historical environmental sample monitoring data of each process production area within the target industrial cluster are obtained. Based on the historical environmental sample monitoring data, a pollutant fingerprint spectrum library of the target industrial cluster is constructed. The pollutant fingerprint spectrum library includes characteristic pollutant types and concentration distribution, and concentration time-series change information of each process production area. The SHAP value of each characteristic pollutant for pollution source classification is calculated based on the machine learning model as the initial diagnostic weight, and the environmental persistence parameter and detection frequency are introduced to construct the correction coefficient. The initial diagnostic weight and the correction coefficient are weighted and fused to obtain the specific weight vector of the characteristic pollutant. Based on the specific weight vector of the characteristic pollutant, the kernel density estimation method is used to fit the probability density of the concentration data of each characteristic pollutant, and a multimodal mixed kernel density estimation model under the multi-process operation mode of each process production area is constructed. The similarity score between the measured concentration of each type of pollutant in the sample to be traced and the concentration distribution of each pollutant in the pollutant fingerprint database is calculated based on the multimodal hybrid kernel density estimation model. The matching degree between the pollution source of each process production area and the sample to be traced is determined based on the similarity score. Based on the matching degree, pollutant sources are traced to determine the source area of ​​the pollutants.

Images