Multi-source spectrum combined pollutant identification method and system based on graph deep learning

Abstract

The invention relates to the technical field of water quality detection, and discloses a multi-source spectrum composite pollutant identification method and system based on graph deep learning, and the method comprises the following steps: S1, obtaining a plurality of single-pollution samples, a plurality of multi-pollution samples, and a to-be-predicted unknown sample; s2, acquiring spectral topological characteristics of a plurality of single-pollution samples and a plurality of multi-pollution samples; s3, estimating a single pollution source prototype vector according to the spectral topological characteristics of the single pollution sample; and estimating a multi-pollution source prototype vector according to the spectral topological characteristics of the multi-pollution sample. The method has the beneficial effects that image deep learning is introduced into the field of spectrum pollutant identification for the first time, and compared with a traditional identification model, the method can better process a complex relation between nodes. In order to estimate a prototype vector of an unknown mixed spectrum by using a prototype vector of a known single pollution spectrum in a mixed pollution label missing scene, prototype contrast learning is introduced, and mixed pollution spectrum characterization is effectively extracted.

Description

Technical Field

[0001] The present invention relates to the technical field of water quality detection, and in particular to a multi-source spectral composite pollutant identification method and system based on graph deep learning. Background Art

[0002] As the global demand for water resources grows, the need for sustainable water resource management is increasing. Pollutant identification has become an extremely important research direction as a major measure for water resource regulation. However, traditional spectral water pollutant identification focuses on the identification of non-mixed multi-source pollutants. In practice, the real prediction of water samples faces the situation of multi-source pollution mixing.

[0003] Identification of water pollutants has become a research hotspot as an important means of supervising and protecting the water environment. Spectroscopy is widely used in many fields such as agriculture, industry, and medicine due to its non-contact and non-secondary pollution characteristics. Among them, three-dimensional fluorescence spectroscopy (3D Fluorescence Spectroscopy), also known as Excitation Emission Matrix (EEM), provides detailed information about fluorescent substances in samples by measuring the changes in fluorescence emission spectra with excitation wavelengths and emission wavelengths; Ultraviolet-Visible Absorption Spectroscopy (UV-Vis) provides information about the chemical composition and structural characteristics of samples by measuring the absorption of light by substances in the ultraviolet to visible light region. They are often used in the field of water quality testing. However, in the past decade, China's industrial parks have been agglomerated, and the identification of water pollutants by spectroscopy faces the dual challenges of a wide variety of pollution types and mixed sources of pollution. The use of multi-source spectra can more effectively represent the characteristic information of different substances; for mixed label missing scenarios, the study of semi-supervised spectral pollutant identification will make spectral pollutant identification more reliable in dealing with complex scenarios with mixed sources of pollution.

[0004] The pollutant identification framework of spectroscopy mainly consists of two parts: feature extraction and identification model. Spectral feature extraction schemes such as three-dimensional fluorescence parallel factor method, peak picking method, regional integration method, principal component analysis method, wavelet decomposition and other methods ignore the representation of the relationship between different feature points. Therefore, some scholars converted three-dimensional spectral features into graph structure data based on three-dimensional topological theory, used SVM model to identify pollutants, and experimentally verified the value of graph structure in spectral pollutant identification. However, spectral identification models such as SVM, neural network, EVM, PLS-DA, LDA, K nearest neighbor and other methods do not directly model on the graph structure with node set and edge set, and it is difficult to make full use of graph structure features. At present, most spectral water pollutant identification methods lack research on semi-supervised scenarios, ignore the real scene of mixed flow of multiple pollution sources, and it is relatively difficult to label the mixed flow pollution label. Considering that it is relatively difficult to label the mixed flow pollution label in reality, it is urgent to study the problem of spectral pollutant identification in the mixed label missing scenario. Summary of the invention

[0005] In order to overcome the difficulty in identifying mixed labels in scenarios where mixed labels are missing and the problem that water source pollutants cannot be accurately identified in the prior art, the present invention proposes a multi-source spectral composite pollutant identification method based on graph deep learning to address the problem of difficulty in identifying mixed labels in scenarios where mixed labels are missing, and specifically includes the following technical solutions.

[0006] A multi-source spectral composite pollutant identification method based on graph deep learning includes the following steps: S1, obtaining multiple single-pollution samples, multiple multi-pollution samples, and unknown samples to be predicted; S2, obtaining spectral topological features of multiple single-pollution samples and multiple multi-pollution samples; S3, estimating single-pollution source prototype vectors according to spectral topological features of single-pollution samples; estimating multi-pollution source prototype vectors according to spectral topological features of multi-pollution samples; S4, calculating prototype vectors of unknown samples, comparing single-pollution source prototype vectors and multi-pollution source prototype vectors with prototype vectors of unknown samples, and predicting pollutants in unknown samples.

[0007] Furthermore, in S3, the prototype vector p of the c-th type of pollution source is c The estimation process is:

[0008]

[0009] Where: p c represents the prototype vector of the c-th type of pollution source, f i represents the characteristics of the network output of the i-th sample in the training set, is the indicative function when the condition is met, y i Represents the contamination label of the i-th sample;

[0010] The prototype vector p of multiple pollution sources of type c and type d pollution c,d The first estimation process of is,

[0011]

[0012] Where 0 <k c,d <1, is a learnable parameter; p c,d The prototype vector representing the mixed pollution of type c and type d;

[0013] Then the pseudo label is annotated for the i-th multi-contaminated sample. The pseudo label of the i-th multi-contaminated sample is obtained through the following process:

[0014]

[0015] where cos(·) represents the cosine similarity function, k∈{1,2,…,K}&{1,2,…,K}∪{1,2,…,K}, & represents the calculation of the unordered product of the set, It represents the score of the ith sample for the cth pollution category, 0<α<1, and the default value of α is 0.8. Initialized to 0; f i represents the characteristics of the network output of the i-th sample in the training set, p k Represents the k-th type of pollution prototype vector;

[0016] In the prototype vector p of multiple pollution sources for type C pollution and type D pollution c,d Make a second estimate. The process of the second estimate is:

[0017]

[0018] Furthermore, in S4, the process of comparing the prototype vector of the unknown sample with the single pollution source prototype vector and the multi-pollution source prototype vector is as follows: using the network output features of the unknown sample and all prototype vectors to calculate the cosine similarity, and taking the pollutant corresponding to the prototype vector with the largest similarity as the pollutant of the unknown sample.

[0019] Furthermore, in S3, the following steps are included: S31, inputting the spectral topological features into the graph convolutional network to obtain three-dimensional fluorescence features and ultraviolet-visible absorption spectrum features; S32, inputting the three-dimensional fluorescence features and ultraviolet-visible absorption spectrum features into the spectral fusion layer to obtain a feature vector; S33, inputting the feature vector into the pollution prototype estimation module to obtain a prototype vector.

[0020] Furthermore, in S31, a single-layer graph convolutional network is assigned to each spectral topological feature, then the three-dimensional fluorescence feature of the k-th contamination of sample n The propagation rules are:

[0021]

[0022] in is the adjacency matrix containing edge weights, represents the weight of the edge from node i to node j, The node embedding representing the nth sample originating from the kth three-dimensional fluorescence spectrum; yes The degree matrix of Represents the weighted degree of each node; is a trainable weight matrix.

[0023] UV-visible absorption spectrum characteristics of the kth contamination of sample n The propagation rules are:

[0024]

[0025] in is the adjacency matrix containing edge weights, represents the weight of the edge from node i to node j, The node embedding representing the nth sample originating from the kth three-dimensional fluorescence spectrum; yes The degree matrix of Represents the weighted degree of each node; is a trainable weight matrix.

[0026] Further, in S32, the feature vector H output by the spectral fusion layer is n The calculation is:

[0027]

[0028] The Concat function represents the concatenation of feature vectors, the max function represents the maximum value of the input node features in each dimension, and the avg function represents the average value of the input node features in each dimension; Represents the output features of the graph convolutional network corresponding to the k-th type of pollutant of the three-dimensional fluorescence spectrum of the n-th sample, The output features of the graph convolutional network representing the UV-visible absorption spectrum of the nth sample corresponding to the kth type of pollutant, Represents the graph structure node set constructed by the UV-visible absorption spectrum corresponding to the k-th type of pollutant, represents the set of graph structure nodes constructed for the k-th type of pollutant corresponding to the three-dimensional fluorescence spectrum, and |·| represents the modulus of the calculation set.

[0029] Furthermore, in S2, the pollution feature extraction module first converts the multi-source spectra of a single pollution sample or multiple pollution samples into four undirected graphs related to the standard pollution sources; then extracts the characteristic peak positions of the standard pollution spectra, and corresponds the characteristic peak positions of each pollution source as the nodes of each pollution topological feature graph; then connects each node, and represents the edge weight using the characteristic peak relationship function of the fluorescence intensity or absorbance of the standard pollutant between the two characteristic peaks.

[0030] Furthermore, the spectral topological characteristics of the kth standard pollution source are:

[0031]

[0032] Where k = 1, 2, ..., K, is the characteristic peak relationship function;

[0033] The characteristic peak relationship function is defined as follows: the three-dimensional fluorescence characteristic peak relationship function of the kth standard pollution source is

[0034]

[0035] The UV-visible characteristic peak relationship function of the kth standard pollution source is:

[0036]

[0037] The spectral topological characteristics of the kth standard pollution source of the nth sample are:

[0038]

[0039] It represents the set of spectral values ​​of sample n at the characteristic peak position of the kth pollution source under the corresponding spectral type; It means that the ultraviolet-visible absorption spectrum The value of the i-th node in the collection, Represents a three-dimensional fluorescence spectrum is the value of the ith node in the set, and σ(·) is the sigmoid function.

[0040] Furthermore, in S1, a standard pollution sample spectral library is established, natural water samples are collected from multiple points in different areas, and papermaking simulated wastewater, metallurgical simulated wastewater, pharmaceutical simulated wastewater, and pesticide simulated wastewater are used as standard pollutants; natural water samples are mixed with standard pollutants to obtain single pollution samples; standard pollutants are mixed in pairs, and then mixed with natural water samples to obtain multi-pollution samples.

[0041] A multi-source spectral composite pollutant identification system based on graph deep learning, including a standard pollution sample spectral library, a pollution topological feature extraction module and a multi-source spectral identification module; the multi-source spectral identification module includes a spectral graph convolution layer, a pollution prototype estimation module, a spectral fusion module layer, and a pollution prediction layer; the multi-pollutant identification system is used to implement the above method.

[0042] According to any of the above embodiments, the present invention has at least the following beneficial effects: The present invention proposes a multi-source spectral composite pollutant identification method based on graph deep learning to address the problem of difficulty in mixed label identification in mixed label missing scenarios. This method constructs a spectral graph structure based on the peak-picking features of the standard pollution source spectrum, which enhances the connection between different peaks compared to traditional spectral feature extraction methods. For the first time, graph deep learning is introduced into the field of spectral pollutant identification, which can better handle the complex relationship between nodes compared to traditional recognition models. In order to achieve the estimation of the prototype vector of the unknown mixed spectrum using the prototype vector of the known single pollution spectrum in the mixed pollution label missing scenario, prototype contrast learning is introduced to effectively extract the mixed pollution spectrum representation.

[0043] Experiments show that in semi-supervised scenario comparison experiments, compared with the SVM model, the accuracy of this method is improved from 55% to 93%, and the macro f1 index is improved from 18% to 75%. Experiments show that the performance of this model in recognition tasks is better than that of multiple machine learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of the method of the present invention.

[0045] Figure 2 It is the architecture diagram of the multi-source spectral pollutant identification system.

[0046] Figure 3 This is a schematic diagram of the natural background water sample collection points.

[0047] Figure 4 It is a schematic diagram of the UV-visible spectrum and three-dimensional fluorescence spectrum measurement platform.

[0048] Figure 5 It is the structural block diagram of the multi-source spectrum recognition module.

[0049] Figure 6 It is a standard pollution source spectrum.

[0050] Figure 7 It is a set of spectral topological characteristic maps obtained by mixing pharmaceutical wastewater, pesticide wastewater and natural water samples.

[0051] Figure 8 This is a schematic diagram of the prediction results of the five models on the test set. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] like Figure 1 As shown, a multi-source spectral composite pollutant identification method based on graph deep learning includes the following steps: S1, obtaining multiple single-pollution samples, multiple multi-pollution samples, and unknown samples to be predicted; S2, obtaining spectral topological features of multiple single-pollution samples and multiple multi-pollution samples; S3, estimating the single pollution source prototype vector according to the spectral topological features of the single pollution sample; estimating the multi-pollution source prototype vector according to the spectral topological features of the multi-pollution samples; S4, calculating the prototype vector of the unknown sample, comparing the single pollution source prototype vector and the multi-pollution source prototype vector with the prototype vector of the unknown sample, and predicting the pollutants of the unknown sample.

[0054] The identification method of the present invention is mainly identified by the corresponding system, and the system architecture is as follows Figure 2 As shown, it includes a standard pollution sample spectral library, a pollution topological feature extraction module, and a multi-source spectral identification module.

[0055] like Figure 3 As shown in the figure, the standard pollution sample spectral library collected 18 points in two areas of the Yangtze River Basin in Chongqing as the natural background, used four simulated wastewaters such as papermaking, metallurgy, pharmaceuticals, and pesticides as standard pollutants, mixed natural water samples to obtain 72 single pollution samples, and mixed the standard pollutants in pairs, and then mixed with natural water samples to obtain 60 multi-pollution samples.

[0056] The UV absorption spectrum of the sample is collected using a UV-visible fiber optic spectrometer. More specifically, a Fuxiang DH-2000 UV-visible light source is used, and a 10mm×10mm quartz cuvette is used as the sample cell; the spectrometer uses the Maya2000 pro produced by Ocean Optics. The UV absorption spectrum detection system mainly consists of a deuterium halogen light source, an attenuator, a sample holder, a spectrometer, and a PC with acquisition control software installed. A fluorescence spectrophotometer is used to collect the three-dimensional fluorescence spectrum, and a 10mm×10mm quartz cuvette is used as the sample cell; the fluorescence spectrometer is the F98 fluorescence spectrometer from Shanghai Lingguang Technology Co., Ltd., which consists of a spectrometer and a PC with acquisition control software installed. The UV-visible spectrum and three-dimensional fluorescence spectrum measurement platform is as follows: Figure 4 shown.

[0057] The UV-visible fiber optic spectrometer is configured with an integration time of 100ms and an average scanning number of 50. In order to avoid the influence of ambient light on the test results and ensure the consistency of the collection environment, a special sample holder is used for light shielding throughout the measurement process. After multiple collections and experiments, the UV-visible absorption spectrum is obtained. The three-dimensional fluorescence data and UV-visible spectrum data are simultaneously measured with an experimental scanning excitation wavelength of 200-600nm (step length of 5nm) and an emission wavelength of 250-650nm (step length of 1nm). The slit width of the excitation light and the emission light is 10nm, the scanning speed is 30000nm / min, the gain voltage is 650V, and the blank water sample is ultrapure water. The sample scanned by the fluorescence spectrometer obtains the corresponding excitation-emission matrix.

[0058] During data preprocessing, EEM was processed as follows: (1) blank water was subtracted from the measured samples; (2) the scattering peaks were eliminated using the Delaunay triangulation method; and UV-Vis was processed as follows: SG smoothing was used to eliminate noise.

[0059] The multi-source spectrum recognition module of the present invention uses both the ultraviolet-visible absorption spectrum characteristics and the three-dimensional fluorescence spectrum multi-source graph characteristics as the input of the model. Based on the semi-supervised and graph-structured data characteristics, four modules are designed, namely, the spectrum graph convolution layer, the pollution prototype estimation module, the spectrum fusion module layer, and the pollution prediction layer. Figure 5 As shown in Figure 2. In the training phase, the prototype vector is determined by the prototype estimation module, and then each unknown sample is labeled for training. In the prediction phase, the prototype vector determined in the training phase is used to predict pollutants.

[0060] Compared with fully labeled samples, we use data with missing mixed labels to study spectral pollution identification in semi-supervised scenarios, which is closer to the real-life scenarios of multi-source pollution mixing and mixed pollution labeling missing. In this scenario, the distribution of the training set and the concept set is seriously inconsistent, which directly leads to a serious decline in the generalization ability of the model based on traditional supervised learning methods.

[0061] To meet this practical challenge, we introduced prototype contrastive learning. Compared with traditional contrastive learning that directly compares the feature representations of individual samples, prototype contrastive learning introduces the concept of prototype vectors. Usually, the average feature representation of a group of samples is used as the prototype to characterize the feature center of similar samples, which can learn more stable and aggregated class features. Inspired by this, for unknown input samples, pseudo-labels can be added to them by matching the prototype vectors on the hypersphere, and the pseudo-labels can be used to guide the model convergence, and the sample features can be approximated as uniformly distributed on the hypersphere. At the same time, in order to estimate the pseudo-labels relatively reliably, we introduce a weighted exponential mechanism.

[0062] The supervision loss of classification problems is mainly BE Loss and ET Loss. Among them, BE Loss guides the model to learn the marginal distribution of different categories, while ET Loss guides the model to learn the joint distribution of categories. Considering that the chemical mechanism of the mixture of multiple pollution sources may interfere with the characteristics of a single pollution spectrum, resulting in unclear marginal distribution characteristics. Therefore, we regard the feature vector of a single pollution sample and the feature vector of a mixed pollution sample as derived from different prototype feature vectors, and use ET Loss to learn the joint distribution of pollution, aiming to obtain more mixed pollution features. The prediction layer uses the cosine similarity with each prototype vector as the input of ET Loss.

[0063] To achieve the above goals, we first calculate the prototype vector of a single pollution source based on a single pollution label. Single pollution source prototype vector p c The estimation process can be described as follows

[0064]

[0065] in is a tag collection, is the indicative function when the condition is met,

[0066] The prototype vector of multiple pollution sources needs to be estimated twice, the c-th pollution and the d-th pollution Multiple pollution source prototype vector p c,d The first estimation process of can be described as follows:

[0067]

[0068] Where 0 <k c,d <1, which is a learnable parameter.

[0069] Subsequently, we use the weighted exponential nearest source prototype method to obtain pseudo labels for unlabeled data. Specifically, the pseudo label of the i-th sample is obtained by the following process:

[0070]

[0071] where cos(·) represents the cosine similarity function, k∈{1,2,…,K}&{1,2,…,K}∪{1,2,…,K}, & represents the calculation of the unordered product of the set, It represents the score of the ith sample for the cth pollution category, 0<α<1, and the default value of α is 0.8. Initialized to 0.

[0072] Type C pollution and Type D pollution Multiple pollution source prototype vector p c,dThe second estimation process of can be described as follows,

[0073]

[0074] We use the estimated prototypes to learn the prototype features of mixed multi-source pollution samples through contrastive learning. Specifically, in the pollutant feature representation space, we have a pseudo-pollutant prototype, denoted as p k , k∈S,S:={1,2,…,K}&{1,2,…,K}∪{1,2,…,K}. Our method uses a prototype contrast loss (PCL) with the following formula:

[0075]

[0076] According to the analysis in ProPosp, it can be determined that Equation (7) can be approximated by prototype alignment and prototype uniformity as outlined in Equation (8). This theoretical validation supports the acquisition of unified and class conflict-averse feature representations by our approach. In addition, the strategy implicitly transforms the basic relationships between samples into relationships between prototypes. Since these learned prototypes provide greater discriminative power than the original sample features, they further strengthen our ability to use pseudo-labels for semi-supervised contaminant identification.

[0077] GCN can learn directly from the topological structure of the graph and has significant advantages in processing graph data. The graph convolution operation propagates information from one node to its neighbors. By recursively aggregating the information of neighboring nodes, GCN can effectively capture the correlation features between nodes. To control the number of parameters, only a single layer of GCN is assigned to each contamination feature. For example, the propagation rule can be expressed as:

[0078]

[0079] in is the adjacency matrix containing edge weights, represents the weight of the edge from node i to node j, Node embedding representing the nth sample originating from the kth 3D fluorescence spectrum. yes The degree matrix of Represents the weighted degree of each node. is a trainable weight matrix.

[0080] UV-visible absorption spectrum characteristics of the kth contamination of sample n The propagation rules are:

[0081]

[0082] in is the adjacency matrix containing edge weights, represents the weight of the edge from node i to node j, The node embedding representing the nth sample originating from the kth three-dimensional fluorescence spectrum; yes The degree matrix of Represents the weighted degree of each node; is a trainable weight matrix.

[0083] The spectral fusion layer outputs the feature vector H n The calculation method is as follows:

[0084]

[0085] The Concat function represents the concatenation of feature vectors, the max function represents the maximum value of the input node features in each dimension, and the avg function represents the average value of the input node features in each dimension.

[0086] A graph is a topological structure used to describe objects and their relationships, usually composed of nodes (representing objects) and edges (representing relationships between nodes). It can intuitively capture the associations between objects and retain rich structural information, so that machine learning models can effectively use this information for prediction and reasoning, providing higher accuracy and interpretability. First, a standard pollution spectral library was established, which included UV-visible absorption spectra and three-dimensional fluorescence spectra of four simulated wastewaters, including papermaking, pharmaceuticals, pesticides, and metallurgy.

[0087] In order to extract the topological structure related to the pollution source between different characteristic peaks, the pollution feature extraction module converts the multi-source spectrum of the sample into 4 undirected graphs related to the standard pollution source. First, the peak picking method is used to extract the characteristic peak positions of the standard pollution spectrum. These characteristic peak positions reveal the chemical information of the pollutants in the sample. The positions of the characteristic peaks of the three-dimensional fluorescence spectrum usually indicate the fluorescence characteristics of specific molecules or substances in the sample, and the positions of the characteristic peaks of the ultraviolet-visible absorption spectrum usually indicate specific types of chemical bonds and functional groups in the molecule. The characteristic peak positions of each pollution source correspond to the nodes of each pollution topological feature graph. The ratio of the fluorescence intensity or absorbance of each characteristic peak of the standard pollutant reflects the relationship between these peaks. Connect these nodes and represent the edge weights with the characteristic peak relationship function of the fluorescence intensity or absorbance of the standard pollutant of the two characteristic peaks. Intuitively, the closer the fluorescence intensity or absorbance of the two characteristic peaks, the larger the characteristic peak relationship function, and the more the edge weight should be 1. Experiments have found that if the difference between the values ​​of the two characteristic peaks is greater, the characteristic peak relationship function is larger, and the former will achieve better results. The spectral topological characteristics of the kth standard pollution source are

[0088]

[0089] Where k = 1, 2, ..., K, is the characteristic peak relationship function. The characteristic peak relationship function in this paper is defined as follows: the three-dimensional fluorescence characteristic peak relationship function of the kth standard pollution source is

[0090]

[0091] The UV-visible characteristic peak relationship function of the kth standard pollution source is:

[0092]

[0093] The spectral topological characteristics of the kth standard pollution source of the nth sample are:

[0094]

[0095] Indicates the spectral value of sample n at the characteristic peak position of the kth pollution source under the corresponding spectral type.

[0096] Four types of simulated industrial wastewater, namely metallurgical wastewater, papermaking wastewater, pharmaceutical wastewater, and pesticide wastewater, were used in the experiment of the present invention. Metallurgical wastewater is usually composed of metal ions and acidic substances. In this experiment, copper sulfate was used to simulate the metal ions, and then sulfuric acid was used to adjust the pH value to 3-4. Papermaking wastewater contains suspended solids and organic pollutants. In this experiment, phenol was used to simulate lignin and organic pollutants, and kaolin was used to simulate suspended solids. Pharmaceutical wastewater usually contains drug residues and nitrogen pollution. In this experiment, acetaminophen was used to simulate drug residues, and ammonia water was used to simulate ammonia nitrogen pollution. Pesticide wastewater uses glyphosate isopropylamine salt to simulate pollution. The three-dimensional fluorescence spectra and ultraviolet-visible absorption spectra of the four simulated wastewaters were used, and the peak picking method was used to extract the characteristic peak positions of the standard pollution and the corresponding physical values ​​of the positions, such as Figure 6 As shown, the graph group of each group of contaminated samples is constructed according to formula (13-17).

[0097] Figure 7 The spectral topological features extracted from the spectra of polluted samples mixed with natural water samples, papermaking wastewater, and metallurgical wastewater are shown in Figure 2. Figure 7It can be seen that in this experiment, after the two pollution sources, pharmaceutical wastewater and pesticide wastewater, were mixed with natural water samples, the three-dimensional fluorescence secondary characteristic peaks of pharmaceutical wastewater and pesticide wastewater were weakened or disappeared under the two spectra used in the experiment. The ultraviolet-visible absorption spectrum of pharmaceutical wastewater disappeared in the fine structure of 200nm-380nm and the characteristic peak disappeared in 420nm-590nm, showing the nonlinear change of the spectrum under the mutual influence of different pollution sources. In order to make full use of the standard pollutant information, ensure the completeness of the pollutant information of the sample, and comprehensively consider the relationship between multi-source pollution spectra, this study uses the characteristics of multiple pollution sources to independently construct the graph structure, merge to obtain the pollution information graph group, and extract the pollution representation in the way of in-process fusion.

[0098] We verified the proposed multi-source spectral recognition model in two scenarios, supervised and semi-supervised. In the supervised scenario, the training set and the test set have the same distribution, which requires that each pollution situation in the training set appears in the test set with a similar frequency. Most models in this scenario have obtained good experimental results. In the semi-supervised scenario, the distribution of the training set and the test set is seriously inconsistent. In this experiment, only the single pollution source samples in the training set have labels, while the test set needs to predict all unlabeled mixed pollution samples and single pollution samples. In this scenario, except for the model we proposed, other comparison models have serious performance degradation and cannot perform prediction tasks.

[0099] Due to the high dimension of spectral data features, using spectral data features to directly establish a model requires a large number of parameters, which weakens the generalization ability of the model. Therefore, it is usually necessary to perform feature extraction to reduce the feature dimension. In this experiment, PCA and PeakPicking, two common spectral data dimension reduction methods, were selected as feature extraction methods for the comparison model. PCA converts a set of possibly related variables into a set of linearly unrelated variables through orthogonal transformation, which can extract linear features in spectral data; Peak Picking is based on identifying the key wavelength positions in the spectrum, that is, peaks, which usually correspond to specific chemical components or physical properties of substances. MPL, SVM, KNN, LDA, and PLS-DA are common spectral recognition models, so they are included in the comparison experiment. LR and SVM are both linear classifiers. They distinguish different categories through a decision boundary, but they differ in loss function and sensitivity to data. RF builds multiple decision trees and integrates the results. Its advantages are that it can handle high-dimensional data, has good control over overfitting, and has high accuracy in multiple fields. Therefore, LR and RF are also included in the comparison experiment, and the experimental results are shown in Tables 1 and 2.

[0100]

[0101]

[0102] Table 1

[0103] Table 1 is a comparison experiment of multiple models in the supervised scenario. It can be seen from Table 1 that in the supervised scenario, RF, MPL, SVM, LDA, KNN, and PLS-DA can achieve an accuracy greater than 0.95 and a Micro-F1 score higher than 0.78 under appropriate feature extraction methods. At the same time, in this scenario, our proposed model is basically ahead of other traditional models, and ranks first among the above models together with the PLS-DA model.

[0104]

[0105]

[0106] Table 2

[0107] Table 2 is a comparative experiment of multiple models in semi-supervised scenarios. From Table 2, it can be seen that in semi-supervised scenarios, the accuracy of RF, MPL, SVM, LDA, KNN, and PLS-DA is less than 0.562, and the Micro-F1 score is less than 0.2. Among them, the accuracy of PLS-DA relative to the supervised scenario is reduced by more than 45.9%, and the Micro-F1 score is reduced by more than 82.9%. The accuracy of SVM relative to the supervised scenario is reduced by more than 42.2%, and the Micro-F1 score is reduced by more than 77.1%. When mixed labels are missing, the accuracy of the model we proposed reaches 0.939, and the Micro-F1 score reaches 0.751. Figure 7 As shown in the figure, due to the mutual influence of different pollutants, the spectral characteristics of mixed polluted samples are difficult to simply represent by the linear superposition of spectra of single pollution. Therefore, linear classifiers commonly used in spectral identification such as SVM, PLS-DA, and LDA cannot use the feature extraction method of Peak Picking. When the training set lacks mixed pollutant labels, traditional machine learning models lack learning methods for mixed samples and reasonable representation of mixtures, resulting in the inability of the model to make effective predictions. When mixed labels are missing, the model we proposed embeds the original representation into a high-dimensional space, and through adaptive calculation of mixed prototype representations, different pollution conditions are evenly represented in the hypersphere space. The prototype contrast loss keeps different mixed classes away from each other, and an effective representation of mixed samples can be extracted when labels are missing. Figure 8The prediction results of the five models, MSPI, LDA, SVM, PLS-DA, and LR, on the test set are shown in more detail. As can be seen from the figure, SVM, PLS-DA, and MSPI all show amazing accuracy in the prediction of single pollutants. However, in the scenario where there is a lack of mixed pollution sample labels, the detection rates of other models in the prediction of mixed pollution samples drop significantly, and the predictions for mixed situations are far lower than those of the MSPI model we extracted.

[0108] In this invention, we propose a multi-source spectral composite pollutant identification method based on graph deep learning, and design a feature extraction method for extracting edge weights and node weights of topological graphs for UV-visible absorption spectra and three-dimensional fluorescence spectra respectively. This method avoids the simple consideration of mathematical matrices in existing methods, and extracts the required pollution features in a targeted manner based on the characteristic knowledge of standard pollutants in the database. This study evaluates the prediction of the model when mixed pollution occurs in four pollution sources and labels are missing through controlled laboratory experiments and simulations, and compares it with seven machine learning models, among which SVM, PLS-DA, LDA, and KNN are included in the experiment as common spectral pollutant identification models, and LR and FR are included in the experiment as common classification models in various fields. The results show that the performance of other models has seriously declined in this scenario, among which PLS-DA has a relative decrease of more than 45.9% in the supervised scenario accuracy, and the Micro-F1 score has a decrease of more than 82.9%, SVM has a relative decrease of more than 42.2% in the supervised scenario accuracy, and the Micro-F1 score has a decrease of more than 77.1%. The accuracy of the model we proposed reached 0.939, and the Micro-F1 score reached 0.751. The accuracy of the supervised scene decreased by more than 0.06%, and the Micro-F1 score decreased by more than 24.9%, which is better than other methods. The current experiments show that as long as the positions or shapes of the pollution sources and standard pollutants are different, the method can still achieve a high accuracy rate when mixed pollution occurs, which can provide good technical support for human water resources management. At the same time, each module of the system independently collects spectral information for easy maintenance and management.

[0109] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-source spectral composite pollutant identification method based on graph deep learning, characterized in that: The following steps are involved: S1, obtaining multiple single-contaminated samples, multiple multi-contaminated samples, and unknown samples to be predicted; S2, obtaining the spectral topological features of multiple single-contaminated samples and multiple multi-contaminated samples; S3, estimating the prototype vector of a single pollution source according to the spectral topological characteristics of a single pollution sample; estimating the prototype vector of multiple pollution sources according to the spectral topological characteristics of multiple pollution samples; S4, calculate the prototype vector of the unknown sample, compare the single pollution source prototype vector and the multi-pollution source prototype vector with the prototype vector of the unknown sample, and predict the pollutants of the unknown sample.

2. The multi-source spectral composite pollutant identification method based on graph deep learning according to claim 1 is characterized in that: In S3, the prototype vector p of the c-th type of pollution source c The estimation process is: Where: p c represents the prototype vector of the c-th type of pollution source, f i represents the characteristics of the network output of the i-th sample in the training set, is the indicative function when the condition is met, y i Represents the contamination label of the i-th sample; The prototype vector p of multiple pollution sources of type c and type d pollution c,d The first estimation process of is, Where 0 <k c,d <1, is a learnable parameter; p c,d The prototype vector representing the mixed pollution of type c and type d; Then the pseudo label is annotated for the i-th multi-contaminated sample. The pseudo label of the i-th multi-contaminated sample is obtained through the following process: where cos(·) represents the cosine similarity function, k∈{1,2,…,K}&{1,2,…,K}∪{1,2,…,K}, & represents the calculation of the unordered product of the set, It represents the score of the ith sample for the cth pollution category, 0<α<1, and the default value of α is 0.

8. Initialized to 0; f i represents the characteristics of the network output of the i-th sample in the training set, p k Represents the k-th type of pollution prototype vector; In the prototype vector p of multiple pollution sources for type C pollution and type D pollution c,d Make a second estimate. The process of the second estimate is:

3. The multi-source spectral composite pollutant identification method based on graph deep learning according to claim 2 is characterized in that: In S4, the process of comparing the prototype vector of the unknown sample with the single-pollution source prototype vector and the multi-pollution source prototype vector is as follows: the cosine similarity is calculated using the network output features of the unknown sample and all prototype vectors, and the pollutant corresponding to the prototype vector with the largest similarity is taken as the pollutant of the unknown sample.

4. The multi-source spectral composite pollutant identification method based on graph deep learning according to claim 1 is characterized in that: In S3, the following steps are involved: S31, inputting the spectral topological features into the graph convolutional network to obtain three-dimensional fluorescence features and ultraviolet-visible absorption spectral features; S32, inputting the three-dimensional fluorescence feature and the ultraviolet visible absorption spectrum feature into the spectrum fusion layer to obtain a feature vector; S33, input the feature vector into the pollution prototype estimation module to obtain the prototype vector.

5. The multi-source spectral composite pollutant identification method based on graph deep learning according to claim 4 is characterized in that: In S31, a single-layer graph convolutional network is assigned to each spectral topological feature, then the three-dimensional fluorescence feature of the k-th contamination of sample n The propagation rules are: in is the adjacency matrix containing edge weights, represents the weight of the edge from node i to node j, The node embedding representing the nth sample originating from the kth three-dimensional fluorescence spectrum; yes The degree matrix of Represents the weighted degree of each node; is a trainable weight matrix. UV-visible absorption spectrum characteristics of the kth contamination of sample n The propagation rules are: in is the adjacency matrix containing edge weights, represents the weight of the edge from node i to node j, The node embedding representing the nth sample originating from the kth three-dimensional fluorescence spectrum; yes The degree matrix of Represents the weighted degree of each node; is a trainable weight matrix.

6. The multi-source spectral composite pollutant identification method based on graph deep learning according to claim 4 is characterized in that: In S32, the feature vector H output by the spectral fusion layer n The calculation is: The Concat function represents the concatenation of feature vectors, the max function represents the maximum value of the input node features in each dimension, and the avg function represents the average value of the input node features in each dimension; Represents the output features of the graph convolutional network corresponding to the k-th type of pollutant of the three-dimensional fluorescence spectrum of the n-th sample, The output features of the graph convolutional network representing the UV-visible absorption spectrum of the nth sample corresponding to the kth type of pollutant, Represents the graph structure node set constructed by the UV-visible absorption spectrum corresponding to the k-th type of pollutant, represents the set of graph structure nodes constructed for the k-th type of pollutant corresponding to the three-dimensional fluorescence spectrum, and |·| represents the modulus of the calculation set.

7. The multi-source spectral composite pollutant identification method based on graph deep learning according to claim 1 is characterized in that: In S2, the pollution feature extraction module first converts the multi-source spectra of single pollution samples or multiple pollution samples into four undirected graphs related to standard pollution sources; then extracts the characteristic peak positions of the standard pollution spectra, and corresponds the characteristic peak positions of each pollution source as the nodes of each pollution topological feature graph; then connects each node, and represents the edge weight using the characteristic peak relationship function of the fluorescence intensity or absorbance of the two characteristic peaks of the standard pollutant.

8. The multi-source spectral composite pollutant identification method based on graph deep learning according to claim 7 is characterized in that: The spectral topological characteristics of the kth standard pollution source are: Where k = 1, 2, ..., K, is the characteristic peak relationship function; The characteristic peak relationship function is defined as follows: the three-dimensional fluorescence characteristic peak relationship function of the kth standard pollution source is The UV-visible characteristic peak relationship function of the kth standard pollution source is: The spectral topological characteristics of the kth standard pollution source of the nth sample are: It represents the set of spectral values ​​of sample n at the characteristic peak position of the kth pollution source under the corresponding spectral type; It means that the ultraviolet-visible absorption spectrum The value of the i-th node in the collection, Represents a three-dimensional fluorescence spectrum is the value of the ith node in the set, and σ(·) is the sigmoid function.

9. The multi-source spectral composite pollutant identification method based on graph deep learning according to claim 1 is characterized in that: In S1, a standard pollution sample spectral library was established, natural water samples were collected from multiple points in different regions, and papermaking simulated wastewater, metallurgical simulated wastewater, pharmaceutical simulated wastewater, and pesticide simulated wastewater were used as standard pollutants; Mix natural water samples with standard pollutants separately to obtain single-polluted samples; mix standard pollutants in pairs and then mix them with natural water samples to obtain multi-polluted samples.

10. A multi-source spectral composite pollutant identification system based on graph deep learning, characterized in that: It includes a standard pollution sample spectral library, a pollution topological feature extraction module and a multi-source spectral identification module; the multi-source spectral identification module includes a spectral graph convolution layer, a pollution prototype estimation module, a spectral fusion module layer, and a pollution prediction layer; the multi-pollutant identification system is used to implement the multi-source spectral composite pollutant identification method based on graph deep learning as described in any one of claims 1-9.

Images