A multi-source spectral composite pollutant identification method and system based on graph deep learning

Abstract

The present application relates to water quality detection technical field, disclose a kind of multi-source spectral composite pollutant identification method and system based on graph depth learning, method contains following steps: S1, obtain multiple single pollution sample, multiple multi-pollution sample and unknown sample to be predicted;S2, obtain the spectral topology feature of multiple single pollution sample, multiple multi-pollution sample;S3, according to the spectral topology feature of single pollution sample, estimate single pollution source prototype vector;According to the spectral topology feature of multi-pollution sample, estimate multi-pollution source prototype vector.It has beneficial effect: the method introduces graph depth learning into the field of pollutant identification of spectroscopy for the first time, and the complex relationship between nodes can be better handled relative to traditional identification model.In order to estimate the prototype vector of unknown mixed spectrum using the prototype vector of known single pollution spectrum in the mixed pollution label missing scene, prototype contrast learning is introduced, and the mixed pollution spectrum representation is effectively extracted.

Description

Technical Field

[0001] This invention relates to the field of water quality testing technology, and in particular to a method and system for identifying multi-source spectral composite pollutants based on graph deep learning. Background Technology

[0002] With the increasing global demand for water resources, the need for sustainable water resource management is growing, and pollutant identification, as a key measure for water resource regulation, has become an extremely important research direction. However, traditional spectroscopic water pollutant identification focuses on the identification of non-mixed multi-source pollutants, while in reality, water samples often face mixed pollution from multiple sources.

[0003] Identifying water pollutants, as a crucial means of monitoring and protecting the aquatic environment, has become a research hotspot. Spectroscopic methods, due to their non-contact and non-secondary pollution characteristics, are widely used in agriculture, industry, medicine, and many other fields. Among them, three-dimensional fluorescence spectroscopy (3DF fluorescence spectroscopy), also known as excitation emission matrix (EEM), provides detailed information about fluorescent substances in a sample by measuring the changes in fluorescence emission spectra with excitation and emission wavelengths; ultraviolet-visible absorption spectroscopy (UV-Vis) provides information about the chemical composition and structural characteristics of a sample by measuring the absorption of light by substances in the ultraviolet to visible light region. These are often used in water quality testing. However, in the past decade, with the industrial clusters in China's industrial parks, spectroscopic identification of water pollutants faces the dual challenges of diverse pollutant types and mixed pollution sources. Using multi-source spectroscopy can more effectively represent the characteristic information of different substances; for mixed label-laden scenarios, research on semi-supervised spectroscopic pollutant identification will enable spectroscopic pollutant identification to more reliably handle complex scenarios with mixed pollution sources.

[0004] The spectroscopic framework for pollutant identification mainly consists of two parts: feature extraction and identification model. Spectroscopic feature extraction methods, such as the 3D fluorescence parallel factor method, peak picking method, region integration method, principal component analysis, and wavelet decomposition, neglect the representation of relationships between different feature points. Therefore, some researchers have converted 3D spectral features into graph-structured data based on 3D topological theory and used SVM models to identify pollutants, experimentally verifying the value of graph structures in spectroscopic pollutant identification. However, spectroscopic identification models, such as SVM, neural networks, EVM, PLS-DA, LDA, and K-nearest neighbors, do not directly model on graph structures with node and edge sets, making it difficult to fully utilize graph structure features. Currently, most spectroscopic water pollutant identification methods lack research on semi-supervised scenarios, ignoring real-world scenarios of multi-source co-flow, making labeling co-flow pollutants relatively difficult. Considering the relative difficulty in labeling co-flow pollutants in reality, there is an urgent need to study the problem of spectroscopic pollutant identification in scenarios where co-flow labels are missing. Summary of the Invention

[0005] To overcome the difficulties in identifying mixed labels in scenarios where mixed labels are missing, and the inability to accurately identify water pollutants in existing technologies, this invention proposes a multi-source spectral composite pollutant identification method based on graph deep learning to address the difficulty in identifying mixed labels in scenarios where mixed labels are missing. Specifically, it includes the following technical solutions.

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

[0007] Furthermore, in S3, the prototype vector p of the single pollution source of type c pollution... c The estimation process is as follows:

[0008]

[0009] Where: p c f represents the prototype vector of a single pollution source of type c pollution. i This represents the feature of the network output for the i-th sample in the training set. The characteristic function y is given by the condition that it satisfies the given condition. i This represents the contamination label of the i-th sample;

[0010] The prototype vectors of multiple pollution sources p for pollution of type c and pollution of type d c,d The first estimation process is as follows:

[0011]

[0012] Among them 0 <k c,d <1 indicates a learnable parameter; p c,d This represents the prototype vector of mixed pollution of type c and type d.

[0013] Then, a pseudo-label is assigned to the i-th multi-contaminated sample. The pseudo-label for the i-th multi-contaminated sample is obtained through the following process:

[0014]

[0015] Where cos(·) denotes the cosine similarity function, k∈{1,2,…,K}&{1,2,…,K}∪{1,2,…,K}, and & denotes the calculation of the unordered product of sets. This represents the score of the i-th sample with respect to the c-th pollution category, where 0 < α < 1, and α defaults to 0.8. Initialize to 0; f i p represents the feature of the network output of the i-th sample in the training set. k Represents the prototype vector of the k-th type of pollution;

[0016] In the multi-source prototype vector p of pollution type c and pollution type d c,d A second estimation is performed, and the process for the second estimation is as follows:

[0017]

[0018] Furthermore, in S4, the process of comparing the prototype vector of the unknown sample with the prototype vector of a single pollution source and the prototype vector of multiple pollution sources is as follows: the cosine similarity between the network output features of the unknown sample and all prototype vectors is calculated, and the pollutant corresponding to the prototype vector with the highest similarity is taken as the pollutant of the unknown sample.

[0019] Further, S3 includes the following steps: 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 features and ultraviolet-visible absorption spectral 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, thus the three-dimensional fluorescence feature of the k-th contamination of sample n is obtained. The propagation rules are as follows:

[0021]

[0022] in It is an adjacency matrix that includes edge weights. This represents the weight of the edge from node i to node j. This indicates that the nth sample originates from the node embedding of the kth type of three-dimensional fluorescence spectrum; yes The degree matrix is ​​defined as This represents the weighted degree of each node; It is a trainable weight matrix.

[0023] UV-Vis absorption spectral characteristics of the kth contaminant in sample n The propagation rules are as follows:

[0024]

[0025] in It is an adjacency matrix that includes edge weights. This represents the weight of the edge from node i to node j. This indicates that the nth sample originates from the node embedding of the kth type of three-dimensional fluorescence spectrum; yes The degree matrix is ​​defined as This represents the weighted degree of each node; It is a trainable weight matrix.

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

[0027]

[0028] The Concat function concatenates feature vectors, the max function calculates the maximum value of the input node features in each dimension, and the avg function calculates the mean value of the input node features in each dimension. This represents the output feature of the graph convolutional network corresponding to the three-dimensional fluorescence spectrum of the nth sample and the pollutant class k. This represents the output feature of the graph convolutional network corresponding to the UV-Vis absorption spectrum of the nth sample and the kth type of pollutant. This represents the set of graph structure nodes corresponding to the k-th type of pollutant in the ultraviolet-visible absorption spectrum. Let |·| represent the set of graph structure nodes corresponding to the k-th type of pollutant in the three-dimensional fluorescence spectrum, and |·| represent the modulus of the computation 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 standard pollution sources; then it extracts the characteristic peak positions of the standard pollution spectra and uses the characteristic peak positions of each pollution source as nodes of each pollution topology feature graph; then it connects each node and uses the characteristic peak relationship function of the fluorescence intensity or absorbance of the two characteristic peaks of the standard pollutants to represent the edge weights.

[0030] Furthermore, the spectral topological characteristics of the k-th standard pollution source are as follows:

[0031]

[0032] Where k = 1, 2, ..., K, 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 k-th standard pollution source is...

[0034]

[0035] The UV-Vis characteristic peak relationship function of the k-th standard pollution source is:

[0036]

[0037] The spectral topological characteristics of the k-th standard pollution source in the n-th sample are as follows:

[0038]

[0039] This represents the set of spectral values ​​for sample n at the characteristic peak position of the k-th pollution source under the corresponding spectral type; This indicates the composition of ultraviolet-visible absorption spectra. The value of the i-th node in the set, This represents a spectrum composed of three-dimensional fluorescence spectra. The value of the i-th node in the set, σ(·) is the sigmoid function.

[0040] Furthermore, in S1, a standard pollutant sample spectral library is established, natural water samples from multiple locations in different regions are collected, and simulated wastewater from papermaking, metallurgy, pharmaceuticals, and pesticides are used as standard pollutants. Natural water samples are mixed with standard pollutants to obtain single pollutant samples. Standard pollutants are mixed in pairs and then mixed with natural water samples to obtain multi-pollutant samples.

[0041] A multi-source spectral composite pollutant identification system based on graph deep learning includes a standard pollutant sample spectral library, a pollutant topological feature extraction module, and a multi-source spectral identification module. The multi-source spectral identification module includes a spectral graph convolutional layer, a pollutant prototype estimation module, a spectral fusion module, and a pollutant 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 difficulty of identifying mixed labels in scenarios where mixed labels are missing. This method constructs a spectral graph structure based on peak-picking features of standard pollution source spectra, enhancing the connection between different peaks compared to traditional spectral feature extraction methods. It is the first to introduce graph deep learning into the field of spectral pollutant identification, and it can better handle complex relationships between nodes compared to traditional identification models. To estimate the prototype vector of unknown mixed spectra using the prototype vector of known single-pollutant spectra in scenarios where mixed pollution labels are missing, prototype contrastive learning is introduced to effectively extract the spectral representation of mixed pollution.

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

[0044] Figure 1 This is a flowchart of the method of the present invention.

[0045] Figure 2 This is an architecture diagram of a 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 This is a schematic diagram of the ultraviolet-visible spectroscopy and three-dimensional fluorescence spectroscopy measurement platform.

[0048] Figure 5 This is a block diagram of the multi-source spectral recognition module.

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

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

[0051] Figure 8 This is a schematic diagram illustrating the prediction performance of the five models on the test set. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

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

[0054] The identification method of this invention mainly performs identification through a corresponding system, the system architecture of which 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 recognition module.

[0055] like Figure 3 As shown, the standard pollutant sample spectral library collected 18 locations in two areas of the Yangtze River Basin in Chongqing as natural backgrounds. Four types of simulated wastewater from papermaking, metallurgy, pharmaceuticals, and pesticides were used as standard pollutants. 72 single-pollutant samples were obtained by mixing natural water samples. The standard pollutants were mixed in pairs and then mixed with natural water samples to obtain 60 multi-pollutant samples.

[0056] The ultraviolet absorption spectrum of the samples was acquired using a UV-Vis fiber optic spectrometer. More specifically, a Fuxiang DH-2000 UV-Vis light source was used, with a 10mm × 10mm quartz cuvette as the sample cell. The spectrometer was a Maya2000 Pro manufactured by Ocean Optics. The UV absorption spectroscopy detection system mainly consisted of a deuterium halogen lamp light source, attenuator, sample holder, spectrometer, and a PC with acquisition control software installed. A fluorescence spectrophotometer was used to acquire the three-dimensional fluorescence spectrum, using a 10mm × 10mm quartz cuvette as the sample cell. The fluorescence spectrometer was an F98 fluorescence spectrometer from Shanghai Lingguang Technology Co., Ltd., consisting of the spectrometer and a PC with acquisition control software installed. The UV-Vis spectroscopy and three-dimensional fluorescence spectroscopy measurement platform is as follows: Figure 4 As shown.

[0057] The UV-Vis fiber optic spectrometer was configured with an integration time of 100 ms and an average of 50 scans. To avoid the influence of ambient light on the test results and ensure the consistency of the acquisition environment, a specially designed sample holder was used for light shielding throughout the measurement process. After multiple acquisitions and experiments, UV-Vis absorption spectra were obtained. Simultaneous measurement of three-dimensional fluorescence data and UV-Vis spectral data was also performed, with the experimental scanning configuration: excitation wavelength 200–600 nm (step size 5 nm), emission wavelength 250–650 nm (step size 1 nm). The slit width for both excitation and emission light was 10 nm, the scan speed was 30,000 nm / min, the gain voltage was 650 V, and the blank water sample was ultrapure water. The corresponding excitation-emission matrix was obtained from the sample scanned by the fluorescence spectrometer.

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

[0059] The multi-source spectral identification module of this invention simultaneously uses ultraviolet-visible absorption spectral features and three-dimensional fluorescence spectral multi-source map features as inputs to the model. Based on the characteristics of semi-supervised, graph-structured data, four modules are designed: a spectral map convolutional layer, a pollution prototype estimation module, a spectral fusion module, and a pollution prediction layer, as follows: Figure 5 As shown, during the training phase, a prototype vector is determined through a prototype estimation module, and then each unknown sample is labeled for training. During the prediction phase, the prototype vector determined during the training phase is used to predict pollutants.

[0060] Compared to fully labeled samples, we used mixed-label missing data to study spectral pollution identification in a semi-supervised scenario, which is closer to the real-world scenario of multi-source pollution mixing and mixed pollution with missing labels. In this scenario, the distribution of the training set and the concept set is severely inconsistent, directly leading to a significant decrease in the generalization ability of models based on traditional supervised learning methods.

[0061] To address this practical challenge, we introduce prototype contrastive learning. Compared to traditional contrastive learning, which directly compares the feature representations of individual samples, prototype contrastive learning introduces the concept of prototype vectors. Typically, the average feature representation of a set of samples is used as a prototype to characterize the feature centers of similar samples, allowing for the learning of more stable and aggregated class features. Inspired by this, for unknown input samples, pseudo-labels can be added by comparing them with the prototype vectors on the hypersphere. These pseudo-labels guide model convergence, and the sample features are approximated as uniformly distributed on the hypersphere. Furthermore, to reliably estimate the pseudo-labels, we introduce a weighted exponential mechanism.

[0062] The supervised losses for classification problems mainly consist of BE Loss and ET Loss. BE Loss guides the model to learn the marginal distributions of different classes, while ET Loss guides the model to learn the joint distribution of classes. Considering that the chemical mechanisms of multi-source pollution mixing may interfere with the spectral features of individual pollution sources, resulting in unclear marginal distribution features, we treat the feature vectors of single pollution samples and mixed pollution samples as originating 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 to ET Loss.

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

[0064]

[0065] in For a set of tags, The indicator function that satisfies the condition.

[0066] The prototype vector of multiple pollution sources needs to be estimated twice, for pollution type c and pollution type d. Multi-source prototype vector p c,d The initial estimation process can be described as follows:

[0067]

[0068] Among them 0 <k c,d <1 indicates a learnable parameter.

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

[0070]

[0071] Where cos(·) denotes the cosine similarity function, k∈{1,2,…,K}&{1,2,…,K}∪{1,2,…,K}, and & denotes the calculation of the unordered product of sets. This represents the score of the i-th sample with respect to the c-th pollution category, where 0 < α < 1, and α defaults to 0.8. Initialize to 0.

[0072] Class C pollution and Class D pollution Multi-source prototype vector p c,dThe second estimation process can be described as follows:

[0073]

[0074] We utilize the estimated prototypes to learn prototype features of mixed multi-source pollution samples through contrastive learning. Specifically, within the pollutant feature representation space, we possess pseudo-pollutant prototypes, denoted as p. k Let k∈S, and S:={1,2,…,K}&{1,2,…,K}∪{1,2,…,K}. Our method uses a prototype contrastive loss (PCL), which is formulated as follows:

[0075]

[0076] Based on the analysis in ProPosp, it can be determined that Equation (7) can be approximated by prototype alignment and prototype uniformity outlined in Equation (8). This theoretical verification supports obtaining uniform and class conflict-averse feature representations through our method. Furthermore, this strategy implicitly transforms the fundamental relationships between samples into relationships between prototypes. Since these learned prototypes provide greater discriminative power than the original sample features, this further enhances our ability to perform semi-supervised contaminant identification using pseudo-labels.

[0077] Graph Convolutional Networks (GCNs) can learn directly from the topology of a graph, offering significant advantages when processing graph data. Graph convolution operations propagate information from one node to its neighbors; by recursively aggregating information from neighboring nodes, GCNs can effectively capture the correlation features between nodes. To control the number of parameters, a single-layer GCN is assigned to each contamination feature. The three-dimensional fluorescence feature of the k-th contamination in sample n is used as an example. For example, the propagation rule can be expressed as:

[0078]

[0079] in It is an adjacency matrix that includes edge weights. This represents the weight of the edge from node i to node j. This indicates that the nth sample originates from the node embedding of the kth type of three-dimensional fluorescence spectrum. yes The degree matrix is ​​defined as This represents the weighted degree of each node. It is a trainable weight matrix.

[0080] UV-Vis absorption spectral characteristics of the kth contaminant in sample n The propagation rules are as follows:

[0081]

[0082] in It is an adjacency matrix that includes edge weights. This represents the weight of the edge from node i to node j. This indicates that the nth sample originates from the node embedding of the kth type of three-dimensional fluorescence spectrum; yes The degree matrix is ​​defined as This represents the weighted degree of each node; It 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 concatenates feature vectors, the max function calculates the maximum value of the input node features in each dimension, and the avg function calculates the mean of the input node features in each dimension.

[0086] A graph is a topological structure used to describe objects and their relationships, typically composed of nodes (representing objects) and edges (representing relationships between nodes). It visually captures the connections between objects, preserving rich structural information, allowing machine learning models to effectively utilize this information for prediction and reasoning, providing higher accuracy and interpretability. First, a standard pollution spectral library was established, providing UV-Vis absorption spectra and three-dimensional fluorescence spectra of four simulated wastewaters from papermaking, pharmaceuticals, pesticides, and metallurgy.

[0087] To extract the topological structure associated with pollution sources among different characteristic peaks, the pollution feature extraction module converts the multi-source spectrum of the sample into four undirected graphs associated with standard pollution sources. First, the peak-picking method is used to extract the positions of the characteristic peaks in the standard pollution spectra. These peak positions reveal the chemical information of the pollutants in the sample. The positions of characteristic peaks in three-dimensional fluorescence spectroscopy typically indicate the fluorescence properties of specific molecules or substances in the sample, while the positions of characteristic peaks in UV-Vis absorption spectroscopy typically indicate specific types of chemical bonds and functional groups in the molecule. The characteristic peak positions of each pollution source correspond to nodes in 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. These nodes are connected, and the edge weights are represented by the characteristic peak relationship function of the fluorescence intensity or absorbance of two characteristic peaks of the standard pollutant. Intuitively, the closer the fluorescence intensity or absorbance of two characteristic peaks, the larger the characteristic peak relationship function, and the higher the edge weight should be. Experiments show that the greater the difference between the values ​​of two characteristic peaks, the larger the characteristic peak relationship function, and comparing the former yields better results. The spectral topological feature of the k-th standard pollution source is...

[0088]

[0089] Where k = 1, 2, ..., K, Let be the characteristic peak relationship function. The characteristic peak relationship function is defined as follows: the three-dimensional fluorescence characteristic peak relationship function for the k-th standard pollution source is...

[0090]

[0091] The UV-Vis characteristic peak relationship function of the k-th standard pollution source is:

[0092]

[0093] The spectral topological characteristics of the k-th standard pollution source in the n-th sample are as follows:

[0094]

[0095] This represents the spectral value of sample n at the characteristic peak position of the k-th pollution source under the corresponding spectral type.

[0096] In the experiments of this invention, four types of simulated industrial wastewater were used: metallurgical wastewater, papermaking wastewater, pharmaceutical wastewater, and pesticide wastewater. Metallurgical wastewater typically consists of metal ions and acidic substances; in this experiment, copper sulfate was used to simulate metal ions, and then sulfuric acid was used to adjust the pH 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 typically 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. Glyphosate isopropylamine salt pesticide was used to simulate pollution in pesticide wastewater. The three-dimensional fluorescence spectra and UV-Vis absorption spectra of the four simulated wastewaters were analyzed. Peak picking was used to extract the characteristic peak positions of the standard pollutants and their corresponding physical values, such as... Figure 6 As shown, the groups of contaminated samples are constructed according to formulas (13-17).

[0097] Figure 7 The image shows the spectral topological features extracted from a pollutant sample consisting of a mixture of natural water samples, papermaking wastewater, and metallurgical wastewater. From... Figure 7As can be seen from the data, in this experiment, after mixing pharmaceutical wastewater and pesticide wastewater with natural water samples from two pollution sources, the secondary characteristic peaks of three-dimensional fluorescence in the pharmaceutical and pesticide wastewater decreased or disappeared under the two spectra used in the experiment. The UV-Vis absorption spectrum of the pharmaceutical wastewater showed the disappearance of its fine structure in the 200nm-380nm range and the disappearance of its characteristic peak in the 420nm-590nm range, demonstrating the nonlinear changes in the spectrum under the mutual influence of different pollution sources. To fully utilize standard pollutant information and ensure the completeness of pollutant information in the samples, this study comprehensively considered the spectral relationships of multi-source pollution. It adopted a multi-source feature independent construction graph structure, merged them to obtain a pollution information graph group, and extracted pollution characterization through in-process fusion.

[0098] We validated the proposed multi-source spectral recognition model in both supervised and semi-supervised scenarios. In the supervised scenario, the training and test sets have the same distribution, requiring that each contamination type in the training set appears with similar frequencies in the test set. In this scenario, most models achieved good experimental results. In the semi-supervised scenario, the training and test sets have severely inconsistent distributions. In this experiment, only single-source contamination samples in the training set were labeled, while the test set required prediction of all unlabeled mixed-contamination samples and single-contamination samples. In this scenario, apart from our proposed model, other comparative models exhibited severe performance degradation and could not perform the prediction task.

[0099] Due to the high dimensionality of spectral data features, directly building a model using these features requires a large number of parameters, weakening the model's generalization ability. Therefore, feature extraction is usually necessary to reduce the feature dimensionality. In this experiment, PCA and Peak Picking, two common spectral data dimensionality reduction methods, were selected as feature extraction methods for comparison models. PCA transforms a set of potentially related variables into a set of linearly unrelated variables through orthogonal transformation, extracting linear features from spectral data. Peak Picking is based on identifying key wavelength positions in the spectrum, i.e., peaks, which usually correspond to specific chemical components or physical properties of substances. MPL, SVM, KNN, LDA, and PLS-DA, as common spectral identification models, were included in the comparison experiment. LR and SVM are both linear classifiers that distinguish different categories through a decision boundary, but they differ in their loss functions and sensitivity to data. RF, by constructing multiple decision trees and integrating the results, has the advantage of handling high-dimensional data, good control over overfitting, and high accuracy in multiple domains. Therefore, LR and RF were 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 shows the comparative experiments of multiple models in supervised scenarios. As can be seen from Table 1, in supervised scenarios, RF, MPL, SVM, LDA, KNN, and PLS-DA can all 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 also generally outperforms other traditional models and ranks first among the aforementioned models, along with the PLS-DA model.

[0104]

[0105]

[0106] Table 2

[0107] Table 2 presents a comparative experiment of multiple models in a semi-supervised scenario. As shown in Table 2, in the semi-supervised scenario, 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. Specifically, PLS-DA shows a greater than 45.9% decrease in accuracy and a greater than 82.9% decrease in Micro-F1 score compared to the supervised scenario; SVM shows a greater than 42.2% decrease in accuracy and a greater than 77.1% decrease in Micro-F1 score compared to the supervised scenario. When mixed labels are missing, our proposed model achieves an accuracy of 0.939 and a Micro-F1 score of 0.751. Figure 7 As shown, due to the mutual influence of different pollutants, the spectral characteristics of mixed contamination samples are difficult to represent simply as a linear superposition of the spectra of a single pollutant. Therefore, linear classifiers commonly used in spectral identification methods such as SVM, PLS-DA, and LDA cannot effectively predict mixed samples when the training set lacks mixed pollutant labels, as traditional machine learning models lack methods for learning mixed samples and reasonable representations of mixtures. Our proposed model, when mixed labels are missing, embeds the original representation into a high-dimensional space. Through adaptive computation of mixed prototype representations, it uniformly represents different contamination conditions in a hypersphere. The prototype contrastive loss keeps different mixed classes far apart, enabling the extraction of effective representations of mixed samples even when labels are missing. Figure 8The results show the prediction performance of five models—MSPI, LDA, SVM, PLS-DA, and LR—on the test set in more detail. As can be seen from the figure, SVM, PLS-DA, and MSPI all showed amazing accuracy in predicting single pollutants. However, in scenarios where there are no labels for mixed pollutant samples, the detection rate of other models dropped significantly, and their prediction of mixed situations was far lower than that 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. We designed a feature extraction method for extracting edge weights and node weights from the topological graph, specifically for UV-Vis absorption spectroscopy and three-dimensional fluorescence spectroscopy. This method avoids the simplistic consideration of mathematical matrices in existing methods and extracts the required pollution features specifically based on standard pollutant feature knowledge in the database. This study evaluated the model's prediction performance in a controlled laboratory setting with mixed pollution from four sources and missing labels. The results were compared with seven machine learning models, including SVM, PLS-DA, LDA, and KNN, which are common spectral pollutant identification models, and LR and FR, which are common classification models in various fields. The results show that other models experienced a significant performance degradation in this scenario. Specifically, PLS-DA's accuracy decreased by more than 45.9% and its Micro-F1 score by more than 82.9% compared to supervised scenarios; SVM's accuracy decreased by more than 42.2% and its Micro-F1 score by more than 77.1% compared to supervised scenarios. Our proposed model achieves an accuracy of 0.939 and a Micro-F1 score of 0.751, representing a greater than 0.06% decrease in accuracy and a greater than 24.9% decrease in Micro-F1 score compared to supervised scenarios, outperforming other methods. Current experiments demonstrate that by establishing a database of pollutants from pollution sources as standards, this method can maintain high accuracy even when mixed pollution occurs, provided the pollutants differ in location or shape. This provides strong technical support for water resource management. Furthermore, the system's independent acquisition of spectral information by each module facilitates maintenance and management.

[0109] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying multi-source spectral composite pollutants based on graph deep learning, characterized in that, Includes the following steps: S1, obtain multiple single-contaminated samples, multiple multi-contaminated samples, and unknown samples to be predicted; S2, to obtain the spectral topological features of multiple single-contaminated samples and multiple multi-contaminated samples; S3, estimate the prototype vector of a single pollution source based on the spectral topological features of a single pollution sample; estimate the prototype vector of multiple pollution sources based on the spectral topological features of multiple pollution samples; S4, calculate the prototype vector of the unknown sample, compare the prototype vector of the single pollution source and the prototype vector of the multiple pollution sources with the prototype vector of the unknown sample, and predict the pollutants of the unknown sample. In S3, the first Class of pollution single pollution source prototype vector The estimation process is as follows: (1) in: Indicates the first Class-specific single-source prototype vectors for pollution. This represents the feature of the network output for the i-th sample in the training set. The indicator function that satisfies the condition, This represents the contamination label of the i-th sample; No. Multi-source prototype vectors of pollution type d and pollution type d The first estimation process is as follows: (2) in , which are learnable parameters; Indicates the first Class and First The prototype vector of mixed pollution types; Indicates the first Class-specific single-source prototype vector; Then, a pseudo-label is assigned to the i-th multi-contaminated sample. The pseudo-label for the i-th multi-contaminated sample is obtained through the following process: (3) (4) (5) in Represents the cosine similarity function. , This represents calculating the unordered product of sets. Let represent the score of the i-th sample with respect to the c-th pollution category. , The default value is 0.

8. Initialize to 0; This represents the feature of the network output for the i-th sample in the training set. Indicates the first Class-based pollution prototype vector; In the Multi-source prototype vectors of pollution type d and pollution type d A second estimation is performed, and the process for the second estimation is as follows: (6)； S3 includes the following steps: S31, input the spectral topological features into the graph convolutional network to obtain three-dimensional fluorescence features and ultraviolet-visible absorption spectral features; S32, input the three-dimensional fluorescence features and the ultraviolet-visible absorption spectral features into the spectral fusion layer to obtain the feature vector; S33, Input the feature vector into the pollution prototype estimation module to obtain the prototype vector; In S31, a single-layer graph convolutional network is assigned to each spectral topological feature, then the nth sample... Three-dimensional fluorescence characteristics of the contaminant The propagation rules are as follows: (9) in , It is an adjacency matrix that includes edge weights. Indicates from node To the node The weight of the edge. This indicates that the nth sample originates from the node embedding of the kth type of three-dimensional fluorescence spectrum; yes The degree matrix is ​​defined as , representing the weighted degree of each node; It is a trainable weight matrix; The first sample n The UV-Vis absorption spectral characteristics of the pollution The propagation rules are as follows: (18) in , It is an adjacency matrix that includes edge weights. Indicates from node To the node The weight of the edge. This indicates that the nth sample originates from the node embedding of the kth type of three-dimensional fluorescence spectrum; yes The degree matrix is ​​defined as , representing the weighted degree of each node; It is a trainable weight matrix; In S32, the feature vector output by the spectral fusion layer The calculation is as follows: (10) (11) (12) The Concat function concatenates feature vectors, the max function calculates the maximum value of the input node features in each dimension, and the avg function calculates the mean value of the input node features in each dimension. Indicates the first The three-dimensional fluorescence spectrum of the sample corresponds to the first... Output characteristics of graph convolutional networks for pollutants. Indicates the first The UV-Vis absorption spectrum of the sample corresponds to the first... Output characteristics of graph convolutional networks for pollutants. This indicates the UV-Vis absorption spectrum corresponding to the first... A graph structure set of nodes constructed from pollutants. Indicates the three-dimensional fluorescence spectrum corresponding to the first A graph structure set of nodes constructed from pollutants. This represents the modulus of the set being computed.

2. The method for identifying multi-source spectral composite pollutants based on graph deep learning according to claim 1, characterized in that, In S4, the process of comparing the prototype vector of the unknown sample with the prototype vector of a single pollution source and the prototype vector of multiple pollution sources is as follows: the cosine similarity between the network output features of the unknown sample and all prototype vectors is calculated, and the pollutant corresponding to the prototype vector with the largest similarity is taken as the pollutant of the unknown sample.

3. The method for identifying multi-source spectral composite pollutants based on graph deep learning according to claim 1, characterized in that, 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 standard pollution sources; then it extracts the characteristic peak positions of the standard pollution spectra and uses the characteristic peak positions of each pollution source as nodes of each pollution topology feature graph; then it connects each node and uses the characteristic peak relationship function of the fluorescence intensity or absorbance of the two characteristic peaks of the standard pollutants to represent the edge weights.

4. The method for identifying multi-source spectral composite pollutants based on graph deep learning according to claim 3, characterized in that, No. The spectral topological characteristics of the standard pollution sources are as follows: (13) (14) in, , , , The characteristic peak relationship function; The characteristic peak relation function is defined as follows, the first... The three-dimensional fluorescence characteristic peak relationship function of the standard pollution sources is as follows: (15) No. The UV-Vis characteristic peak relationship function of a standard pollution source is as follows: (16) No. The first sample The spectral topological characteristics of the standard pollution sources are as follows: (17) This represents the set of spectral values ​​for sample n at the characteristic peak position of the k-th pollution source under the corresponding spectral type; This indicates the composition of ultraviolet-visible absorption spectra. The set of The value of each node, This represents a spectrum composed of three-dimensional fluorescence spectra. The set of The value of each node, It is the sigmoid function.

5. The method for identifying multi-source spectral composite pollutants based on graph deep learning according to claim 1, characterized in that, In S1, a standard pollutant sample spectral library was established, natural water samples were collected from multiple locations in different regions, and simulated wastewater from papermaking, metallurgy, pharmaceuticals, and pesticides were used as standard pollutants. Single-pollutant samples are obtained by mixing natural water samples with standard pollutants separately; multi-pollutant samples are obtained by mixing standard pollutants in pairs and then mixing them with natural water samples.

6. A multi-source spectral composite pollutant identification system based on graph deep learning, characterized in that, The system includes a standard pollution sample spectral library, a pollution topology feature extraction module, and a multi-source spectral identification module. The multi-source spectral identification module includes a spectral graph convolutional layer, a pollution prototype estimation module, a spectral fusion module, and a pollution prediction layer. The multi-source spectral composite 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 1-5.

Images