Baijiu park wastewater fluorescence tracing method based on double-input convolutional neural network

Abstract

The application discloses a liquor park wastewater fluorescence tracing method based on a double-input convolutional neural network, comprising the following steps: obtaining three-dimensional excitation-emission matrix fluorescence spectrum data of a wastewater sample to be traced, generating a standardized EEM image and extracting a fluorescence feature parameter vector; inputting the two into an image feature extraction branch and a parameter feature mapping branch of a double-input convolutional neural network model respectively, obtaining image features and parameter features, outputting probability distributions of various pollution source categories after cross-modal fusion, and determining a wastewater pollution source identification result. The method can fuse fluorescence image spatial features and fluorescence parameter information, and improve the accuracy and reliability of similar wastewater pollution source identification in a liquor park.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring and water pollution source tracing technology, specifically to a fluorescent source tracing method for wastewater from a liquor industrial park based on a dual-input convolutional neural network. Background Technology

[0002] The Baijiu Industrial Park is a comprehensive industrial cluster with Baijiu brewing as its core industry, complemented by diverse supporting industries such as printing and packaging, and machinery processing. The wastewater generated during Baijiu brewing has distinct characteristics: massive volume, extremely high organic matter concentration, and contains easily degradable substances such as alcohols, esters, and organic acids, as well as recalcitrant components such as phenols and long-chain organic compounds. Environmental protection departments require centralized wastewater treatment in industrial parks; however, in actual supervision, some enterprises still engage in illegal discharges, leaks, or excessive emissions to reduce pollution control costs. When sudden pollution incidents occur, the lack of rapid and accurate pollution source tracing technology makes it difficult to identify the responsible party in a timely manner, posing challenges to environmental supervision and enforcement.

[0003] Currently, industrial wastewater source tracing mainly relies on traditional chemical analysis methods. These methods include measuring traditional water quality indicators such as pH, TOC, COD, BOD, NH3-N, TN, and TP to reflect the overall level of pollutants; using gas chromatography-mass spectrometry (GC-MS) to analyze the organic composition of wastewater; using inductively coupled plasma mass spectrometry (ICP-MS) to analyze the distribution of metal elements; and using Fourier transform infrared spectroscopy (FTIR) to identify the structure of organic functional groups. While these methods can reveal the chemical fingerprint characteristics of wastewater at different levels and provide a basis for pollution source tracing, they suffer from problems such as long detection cycles, expensive equipment, complex operation, and inability to achieve rapid response, making it difficult to meet the minute-level emergency response requirements for sudden pollution events.

[0004] Three-dimensional fluorescence spectroscopy, by simultaneously scanning excitation and emission wavelengths, can acquire a complete fluorescent fingerprint of dissolved organic matter in wastewater. Compared with traditional methods, this technology is information-rich, requires no complex chemical pretreatment, and is simple to operate, and has been widely used in industrial wastewater source tracing in recent years. However, liquor brewing wastewater is mainly composed of microbial metabolites, with a highly complex composition of dissolved organic matter. Similar enterprises exhibit high fluorescence similarity, making it difficult for traditional appearance feature analysis to achieve precise enterprise-level identification. While parallel factor analysis can reveal essential distinguishability, it requires offline spectral interpretation using specialized software, and single-sample processing is time-consuming, contradicting the need for rapid response to sudden pollution events.

[0005] In recent years, researchers have begun to incorporate deep learning into the intelligent analysis of three-dimensional fluorescence spectra. Existing methods based on single-input convolutional neural networks typically use a preprocessed EEM grayscale image as a single input. A CNN automatically extracts spatial features from the fluorescence spectrum, and a Softmax classifier outputs class probabilities to achieve end-to-end wastewater category identification. However, a single EEM image input is insufficient to distinguish wastewater samples with highly similar fluorescence characteristics and lacks essential physicochemical information such as the source attributes of dissolved organic matter and the degree of humification. Traditional hard classification methods are prone to misclassification when faced with samples with extremely high fluorescence similarity and lack a coupling system between a fluorescence fingerprint database and an intelligent identification model. Therefore, there is an urgent need for a technical solution suitable for the rapid and accurate identification of wastewater pollution sources in liquor industrial parks. Summary of the Invention

[0006] The purpose of this invention is to address the above-mentioned technical problems existing in the prior art by providing a method for tracing the source of fluorescence in wastewater from a liquor industrial park based on a dual-input convolutional neural network and three-dimensional fluorescence spectroscopy.

[0007] To achieve the above objectives, this invention provides a method for tracing the source of fluorescence in wastewater from a liquor industrial park based on a dual-input convolutional neural network, comprising: Obtain three-dimensional excitation-emission matrix fluorescence spectral data of wastewater samples from the liquor industrial park to be traced; A standardized EEM image is generated based on the three-dimensional excitation-emission matrix fluorescence spectral data, and fluorescence feature parameters are extracted from the same three-dimensional excitation-emission matrix fluorescence spectral data to form a fluorescence feature parameter vector; The standardized EEM image is input into the image feature extraction branch of a trained dual-input convolutional neural network model to obtain image features; The fluorescence feature parameter vector is input into the parameter feature mapping branch of the dual-input convolutional neural network model to obtain the parameter features; The image features and the parameter features are fused through a cross-modal fusion layer to obtain fused features; The fused features are input into the probability output layer to obtain the probability distribution of each pollution source category, and the wastewater pollution source identification result is determined based on the probability distribution.

[0008] As a further improvement to the above technical solution, before obtaining the three-dimensional excitation-emission matrix fluorescence spectral data of the wastewater samples to be traced in the liquor industrial park, the method further includes: collecting wastewater samples from each enterprise in the liquor industrial park, obtaining the EEM data of each enterprise's wastewater samples through a three-dimensional fluorescence spectrometer, and associating the EEM data of each enterprise's wastewater samples with the corresponding pollution source category label; constructing a three-dimensional fluorescence fingerprint database based on the EEM data of each enterprise's wastewater samples, wherein the three-dimensional fluorescence fingerprint database includes an appearance feature layer, a feature parameter layer, and a specific component layer; the appearance feature layer is used to record the position and intensity information of the main fluorescence peaks of the wastewater samples, the feature parameter layer is used to record the fluorescence feature parameters extracted from the EEM data, and the specific component layer is used to record the enterprise-specific fluorescence components obtained through parallel factor analysis; The three-dimensional fluorescent fingerprint database is used to provide the EEM data, pollution source category labels and fluorescence feature parameters required for training the dual-input convolutional neural network model, and to establish the correspondence between the pollution source categories in the probability distribution and the enterprises in the liquor industrial park.

[0009] As a further improvement to the above technical solution, generating a standardized EEM image based on the three-dimensional excitation-emission matrix fluorescence spectral data includes: performing bilinear interpolation on the original EEM matrix corresponding to the three-dimensional excitation-emission matrix fluorescence spectral data to uniformly adjust it into a standardized grayscale image with a preset pixel size; and copying the standardized grayscale image into a three-channel image to obtain the standardized EEM image.

[0010] As a further improvement to the above technical solution, the steps for constructing the fluorescence feature parameter vector include: extracting the fluorescence index, humification index, biological index, freshness index, total fluorescence intensity, and main peak position from the three-dimensional excitation-emission matrix fluorescence spectral data, and then normalizing them to construct the fluorescence feature parameter vector.

[0011] As a further improvement to the above technical solution, the dual-input convolutional neural network model includes an image feature extraction branch, a parameter feature mapping branch, a cross-modal fusion layer, and a probability output layer. The image feature extraction branch is used to perform convolutional feature extraction on the standardized EEM image to obtain image features characterizing the fluorescence peak morphology and intensity distribution. The parameter feature mapping branch is used to perform nonlinear dimensionality upscaling mapping on the fluorescence feature parameter vector to obtain parameter features that match the image feature dimension. The cross-modal fusion layer is used to fuse the image features and the parameter features, and the probability output layer is used to output the probability distribution of each pollution source category based on the fused features.

[0012] As a further improvement to the above technical solution, the image feature extraction branch uses MobileNetV2 as the backbone network to perform depthwise separable convolution processing on the standardized EEM image to obtain image features; the parameter feature mapping branch includes at least two fully connected layers, and the fluorescence feature parameter vector is nonlinearly up-dimensionally mapped through at least two fully connected layers to obtain parameter features consistent with the image feature dimension.

[0013] As a further improvement to the above technical solution, the image features and the parameter features are fused through a cross-modal fusion layer to obtain fused features, including: concatenating the image features and the parameter features along the channel dimension to obtain joint features; and inputting the joint features into a fully connected fusion layer for adaptive fusion to obtain fused features. The probability output layer uses a Softmax classifier.

[0014] As a further improvement to the above technical solution, the training steps of the dual-input convolutional neural network model include: obtaining EEM data, pollution source category labels, and fluorescence feature parameters corresponding to wastewater samples with known pollution source categories from the three-dimensional fluorescence fingerprint database; performing data augmentation on the original EEM matrix corresponding to the EEM data to obtain an enhanced EEM matrix, wherein the data augmentation includes at least one of random scaling of fluorescence intensity, micro-shifting of excitation and emission wavelength directions, and addition of Gaussian noise; generating a standardized EEM image for training based on the original EEM matrix and the enhanced EEM matrix, and constructing a fluorescence feature parameter vector for training based on the fluorescence feature parameters; constructing training samples based on the standardized EEM image for training, the fluorescence feature parameter vector for training, and the pollution source category labels; and training the dual-input convolutional neural network model using the training samples.

[0015] As a further improvement to the above technical solution, determining the wastewater pollution source identification result based on the probability distribution includes: converting the probability of each pollution source category in the probability distribution into a category score, sorting the category scores of each pollution source in descending order, and taking the top K items to obtain the Top-K ranking result, where K≥3; calculating the first category score in the Top-K ranking result. Compared to the second category score The score interval ΔS between them, the score interval When ΔS≥10, the Top-1 result is output as the wastewater pollution source identification result; when 5≤ΔS<10, the Top-2 candidate result is output as the wastewater pollution source identification result; when ΔS<5, the Top-3 candidate result is output as the wastewater pollution source identification result.

[0016] As a further improvement to the above technical solution, the dual-input convolutional neural network model satisfies at least one of the following configurations: the image feature extraction branch adopts one of MobileNetV3, ShuffleNetV2, and EfficientNet-B0; the cross-modal fusion layer adopts one of attention-weighted fusion, bilinear fusion, and element-wise additive fusion; the input of the parameter feature mapping branch also includes at least one of parallel factor analysis component score vector, fluorescence region integral parameter, and ultraviolet absorption value at 254nm.

[0017] The beneficial effects of the fluorescent source tracing method for wastewater from liquor industrial parks provided by this invention are as follows: 1. Overcome the technical bottleneck of high fluorescence similarity (60%-90%) in wastewater from similar liquor brewing enterprises, making precise identification difficult. By integrating the physicochemical essence information of EEM image spatial features and fluorescence feature parameters, cross-modal feature complementarity is achieved, thereby improving the differentiation accuracy of similar wastewater. 2. To address the overfitting problem of deep learning models under small sample conditions, a high-precision recognition model suitable for limited samples is constructed through joint optimization of lightweight network architecture, transfer learning strategy and data augmentation technology; 3. Breaking through the limitations of the traditional hard classification model, a Top-K probabilistic decision-making mechanism is introduced, transforming the rigid output of unique identification into a flexible tracing of a suspect list, quantifying uncertainty and providing a hierarchical decision-making basis; 4. Establish a complete technical system from fluorescent fingerprint database construction to intelligent identification decision-making, and form a standardized rapid identification solution for wastewater pollution sources in liquor industrial parks. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the fluorescent source tracing method for wastewater from a liquor industrial park in this embodiment of the invention. Figure 2 is a schematic diagram of the training results of the single-input baseline model in an embodiment of the present invention, wherein, Figure 2a This is the accuracy change curve. Figure 2b The loss function decreases as shown by the curve. Figure 2c This is the confusion matrix; Figure 3 is a comparison diagram of typical wastewater EEM effects in an embodiment of the present invention, wherein... Figure 3a For non-alcoholic beverage companies, EEM Figure 3b For alcoholic beverage companies, EEM; Figure 4 is a graph showing the relevant fluorescence parameters of typical wastewater in an embodiment of the present invention. Figure 4a The fluorescence index, Figure 4b As a freshness index, Figure 4c The index is the self-originating index. Figure 4d The humification index; Figure 5This is a flowchart illustrating the dual-input fusion network structure and dimensional transformation in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of protection of this invention.

[0020] Example 1 This embodiment provides a fluorescence source tracing method for wastewater from a liquor industrial park based on a dual-input convolutional neural network. For example... Figure 1 As shown, the method includes the following steps: acquiring three-dimensional excitation-emission matrix fluorescence spectral data of wastewater samples from a liquor industrial park; generating a standardized EEM image based on the three-dimensional excitation-emission matrix fluorescence spectral data, and extracting fluorescence feature parameters from the same three-dimensional excitation-emission matrix fluorescence spectral data to form a fluorescence feature parameter vector; inputting the standardized EEM image into the image feature extraction branch of a trained dual-input convolutional neural network model to obtain image features; inputting the fluorescence feature parameter vector into the parameter feature mapping branch of the dual-input convolutional neural network model to obtain parameter features; fusing the image features and parameter features through a cross-modal fusion layer to obtain fused features; inputting the fused features into a probability output layer to obtain the probability distribution of each pollution source category, and determining the wastewater pollution source identification result based on the probability distribution.

[0021] Here, EEM stands for three-dimensional excitation-emission matrix fluorescence spectroscopy, which is a three-dimensional data matrix of fluorescence intensity obtained by simultaneously scanning the excitation and emission wavelengths; DOM stands for dissolved organic matter, which refers to dissolved organic substances in water and is also the main target of three-dimensional fluorescence spectroscopy. The overall workflow uses the same EEM data to form both image and parameter inputs, enabling the model to utilize the physicochemical information of both the spatial features of the EEM image and the fluorescence characteristic parameters.

[0022] Example 2 Before obtaining the three-dimensional excitation-emission matrix fluorescence spectral data of the wastewater samples to be traced in the liquor industrial park, wastewater samples from each enterprise within the park are collected. EEM data of each enterprise's wastewater samples are obtained using a three-dimensional fluorescence spectrometer, and the EEM data of each enterprise's wastewater samples are associated with corresponding pollution source category labels. A three-dimensional fluorescence fingerprint database is constructed based on the EEM data of each enterprise's wastewater samples. The three-dimensional fluorescence fingerprint database provides the EEM data, pollution source category labels, and fluorescence feature parameters required for training a dual-input convolutional neural network model, and is used to establish the correspondence between pollution source categories in the probability distribution and each enterprise within the liquor industrial park. In one specific implementation, the sampling objects include wastewater from washing, machinery, printing, liquor brewing enterprises, and the total influent from the liquor wastewater treatment plant; sampling points are set at the enterprise's pipeline discharge outlets, and for enterprises equipped with pretreatment devices, untreated water samples can also be collected at the front end of the pretreatment device. Samples can be collected from multiple production batches within a year, for example, four batches can be sampled in March, June, September, and December. Within each batch, multiple subsamples can be collected from the same sampling point at preset intervals on the same sampling day, for example, once every 2 hours, for a total of 5 collections, which are then combined as the batch sample for that sampling point. The pollution source category label is jointly determined by the enterprise to which the sampling point belongs, the sampling record, the enterprise's drainage ledger, and the confirmation results of on-site personnel, ensuring a one-to-one correspondence between EEM data and pollution source categories.

[0023] In this embodiment, the three-dimensional fluorescent fingerprint database includes an appearance feature layer, a feature parameter layer, and a specific component layer. Through this three-layer structure, the database can provide a standardized data foundation and verification basis for intelligent recognition models at three levels: industry-level appearance differentiation, parameterized feature expression, and enterprise-level specific component identification.

[0024] In the three-dimensional fluorescence fingerprint database, each record is based on a wastewater sample from an enterprise, and the corresponding EEM data, pollution source category label, appearance feature layer information, feature parameter layer information, and specific component layer information are associated and saved. During the model training phase, a supervision label is constructed based on the pollution source category label. During the identification phase, the pollution source category in the probability output layer is mapped to the corresponding enterprise in the liquor industrial park based on this association.

[0025] The surface feature layer is used to record the location (Ex / Em) and intensity information of the main fluorescence peaks in the wastewater sample. As shown in Figure 3, Figure 3a For non-alcoholic beverage companies, EEM Figure 3b The wastewater from liquor enterprises is characterized by predominantly soluble microbial byproducts (Region IV) and humic acid-like substances (Region V), while non-liquor enterprises are dominated by tyrosine-like substances (Region I) and soluble microbial byproducts (Region IV), exhibiting industry-specific differences.

[0026] The feature parameter layer is used to record the fluorescence feature parameters extracted from the EEM data. In this layer, six fluorescence feature parameters are extracted from the EEM data to construct a one-dimensional feature vector. :

[0027] Wherein, FI (fluorescence index) = F(Ex=370nm, Em=450nm) / F(Ex=370nm, Em=500nm) is used to indicate the source attribute of DOM. When FI>1.9, it is mainly due to endogenous microbial metabolism. HIX (humification index) = (sum of F from Em=435nm to 480nm at Ex=254nm) / (sum of F from Em=300nm to 345nm at Ex=254nm), which characterizes the degree of humification of organic matter; BIX (Origin Index) = F(Ex=310nm, Em=380nm) / F(Ex=310nm, Em=430nm), reflects the proportion of DOM (Matter of Degradable Origin) contribution from origin. BIX > 0.8 indicates that it is rich in fresh biodegradable organic matter. beta:alpha (freshness index) = F(Ex=310 nm, Em=380 nm) / integral(F from Em=420 nm to 435 nm at Ex=310 nm), representing the proportion of newly formed DOM; T intensity is the total fluorescence intensity, and PeakPos is the position of the main peak (Ex / Em).

[0028] In the above formula for calculating fluorescence characteristic parameters, F(Ex=…,Em=…) represents the fluorescence intensity at the corresponding excitation wavelength Ex and emission wavelength Em, sum represents the summation of fluorescence intensity over the listed emission wavelength range, and integral represents the integration of fluorescence intensity over the listed emission wavelength range.

[0029] Wherein, FI is the fluorescence index, used to indicate the source attribute of DOM; HIX is the humification index, used to characterize the degree of organic matter humification; and BIX is the biogenic index, used to reflect the proportion of DOM contribution from its own source. Figure 4 shows the relevant fluorescence parameters of typical wastewater, where... Figure 4a The fluorescence index, Figure 4b As a freshness index, Figure 4c The index is the self-originating index. Figure 4d This is the humification index.

[0030] As shown in Figure 4, the FI of non-alcoholic wastewater mainly ranges from 1.4 to 1.7, indicating a mixed source of external input and microbial degradation. Furthermore, the BIX, β:α, and HIX values ​​are generally lower than those of baijiu (Chinese liquor) brewing wastewater. Baijiu brewing wastewater generally has a FI greater than 1.9, with higher BIX and β:α values, and a HIX value mostly above 2, indicating that its DOM (dominant morphology) is primarily driven by endogenous microbial metabolism, exhibiting high biological activity and a high degree of humification. For baijiu brewing enterprises and different process stages, further differentiation can be made based on the differences in the combination of parameters such as HIX, BIX, and β:α.

[0031] Specific component layer: The PARAFAC model was used to deconstruct single-enterprise samples, and the three-dimensional fluorescence data were analyzed. Decomposed into the product of three low-dimensional matrices:

[0032] in Let be the fluorescence intensity of sample i at emission wavelength j and excitation wavelength k; It is directly proportional to the concentration of the f-th analyte in sample i; It has a linear relationship with fluorescence quantum efficiency; The specific absorption coefficient is linearly proportional to the specific fluorescence. PARAFAC decomposition is used to identify the specific fluorescent components of each company, establishing a precise enterprise-level identification basis.

[0033] Parallel factor analysis (PARAFAC) is a multivariate statistical analysis method used to decompose three-dimensional fluorescence data into independent fluorescent components with physicochemical significance. The specific component layer records enterprise-specific fluorescent components obtained through parallel factor analysis to establish a basis for enterprise-level fine-grained identification.

[0034] Example 3 In this embodiment, when generating a standardized EEM image based on three-dimensional excitation-emission matrix fluorescence spectral data, the original EEM matrix corresponding to the three-dimensional excitation-emission matrix fluorescence spectral data is subjected to bilinear interpolation processing and uniformly adjusted to a standardized grayscale image with a preset pixel size; the standardized grayscale image is copied into a three-channel image to obtain a standardized EEM image.

[0035] In one specific implementation, the original EEM matrix has a dimension of 252×82 (excitation wavelength 200-600nm, step size 5nm; emission wavelength 200-700nm, step size 2nm), and is uniformly adjusted to a standardized grayscale image of 224×224 pixels using bilinear interpolation. .

[0036]

[0037] The single-channel grayscale image is copied to a three-channel RGB format to meet the input requirements of the MobileNetV2 network. The normalized EEM image is used as input to the image feature extraction branch to represent spatial features such as fluorescence peak morphology, peak contour, and intensity distribution.

[0038] The steps for constructing the fluorescence feature parameter vector include: extracting the fluorescence index, humification index, biological index, freshness index, total fluorescence intensity, and main peak position from the same three-dimensional excitation-emission matrix fluorescence spectral data, and then normalizing them to form the fluorescence feature parameter vector. The normalized fluorescence feature parameter vector serves as the second input mode, which, along with the standardized EEM image, is input into a dual-input convolutional neural network model.

[0039] Example 4 like Figure 5 As shown, the dual-input convolutional neural network model includes an image feature extraction branch, a parameter feature mapping branch, a cross-modal fusion layer, and a probability output layer. CNN, or Convolutional Neural Network, automatically extracts hierarchical features from data through local receptive fields and weight sharing mechanisms.

[0040] To address the issue of high fluorescence similarity among various liquor brewing wastewaters, this embodiment employs a dual-input fusion network. This network jointly fuses depth features extracted from EEM images by a CNN with fluorescence feature parameters calculated from the same EEM data. The network utilizes a jointly trainable dual-branch feature fusion framework. Fluorescence feature parameters characterize attributes such as DOM origin, biological activity, and humification level, while CNN depth features characterize spatial information such as fluorescence peak position, peak shape, and intensity distribution. The two types of features adaptively learn their combination relationship through a fully connected fusion layer, eliminating the need for manually preset fixed fusion weights.

[0041] The image feature extraction branch is used to perform convolutional feature extraction on the standardized EEM image to obtain image features characterizing the morphology and intensity distribution of fluorescence peaks. In one specific implementation, the image feature extraction branch uses MobileNetV2 as the backbone network to perform depthwise separable convolution processing on the standardized EEM image to obtain image features. MobileNetV2 is a lightweight convolutional neural network architecture that uses depthwise separable convolution and inverse residual structures, with approximately 4.2 million parameters.

[0042] Depthwise separable convolution decomposes standard convolution into two steps: depthwise convolution and pointwise convolution.

[0043]

[0044] in, The convolution kernel for the k-th channel, The kernel is 1×1, where x represents the input feature map, y represents the intermediate feature map after depthwise convolution, and z represents the output feature map after pointwise convolution. i and j represent the spatial indexes of the feature maps, m and n represent the offset indices of the convolution kernel in the spatial dimension, k represents the input channel index, and c represents the output channel index. This design reduces the number of parameters and computation to 1 / 8 to 1 / 9 of traditional convolution. MobileNetV2 introduces a linear bottleneck and inverted residual structure: first, dimensionality is increased through 1×1 convolution (expansion coefficient t=6), features are extracted through depthwise convolution, and then dimensionality is reduced through 1×1 convolution, with residual connections used. The input 224×224×3 EEM image is initially scanned by 32 3×3 convolution kernels, and then progressively downsampled to a 7×7×320 feature map through 17 inverted residual blocks.

[0045] The parametric feature mapping branch is used to perform nonlinear dimensionality upscaling on the fluorescence feature parameter vector to obtain parametric features that match the image feature dimension. Fluorescence feature parameters Since the 1×6 one-dimensional scalar vector lacks spatial structure, convolution is not used. Instead, a nonlinear dimensionality-upgrading mapping is performed through fully connected layers. In one specific implementation, the parameter feature mapping branch includes at least two fully connected layers. The fluorescence feature parameter vector undergoes nonlinear dimensionality-upgrading mapping through at least two fully connected layers to obtain parameter features consistent with the image feature dimension.

[0046] in It is a 6×64 dimensional weight matrix. It is a 64×320 dimensional weight matrix. , Here, is the bias vector, and ReLU is the activation function. The mapped feature dimension is 1×320, consistent with the output dimension of the image feature extraction branch.

[0047] Cross-modal fusion layers are used to fuse image features and parametric features. By fusing image features and parametric features through a cross-modal fusion layer to obtain fused features, the image features... (320-dimensional) and parametric features (320-dimensional) concatenated along the channel dimension to form a 640-dimensional joint feature vector:

[0048] Learnable adaptive fusion via fully connected layers:

[0049] in, This represents the fusion features after cross-modal fusion. This represents the weight matrix of the fully connected fusion layer. This represents the bias vector of the fully connected fusion layer.

[0050] The weight allocation of the two types of features is automatically optimized through backpropagation, achieving synergistic optimization of domain knowledge (fluorescence parameters) and data-driven features (CNN deep features). The probability output layer uses a Softmax classifier to convert the fused features into a normalized probability distribution of each pollution source category. The probability output represents the model's confidence in classifying the sample.

[0051] Example 5 In this embodiment, the training steps of the dual-input convolutional neural network model include: obtaining EEM data, pollution source category labels, and fluorescence feature parameters corresponding to wastewater samples with known pollution source categories from a three-dimensional fluorescence fingerprint database; performing data augmentation on the original EEM matrix corresponding to the EEM data to obtain an augmented EEM matrix; generating a standardized EEM image for training based on the original EEM matrix and the augmented EEM matrix, and constructing a fluorescence feature parameter vector for training based on the fluorescence feature parameters; constructing training samples based on the standardized EEM image for training, the fluorescence feature parameter vector for training, and the pollution source category labels; and training the dual-input convolutional neural network model using the training samples.

[0052] The data augmentation includes at least one of random scaling of fluorescence intensity, micro-shifting of excitation and emission wavelength directions, and addition of Gaussian noise. In one specific implementation, the original EEM matrix is ​​subjected to random scaling of intensity (scaling factor 0.8-1.2), micro-shifting of Ex / Em (±2 pixels), and addition of Gaussian noise (sigma=0.01), which expands the training set sample size to 5 times, effectively mitigating the risk of overfitting with small samples; the sample size after augmentation is expanded from 301 to 1505.

[0053] During training, the total sample is divided into training, validation, and test sets in a 7:2:1 ratio, using stratified sampling to ensure consistent sample proportions for each category. Initial learning rate. The Adam optimizer was used, with a batch size of 16 and an epoch of 100. To address the fine-grained classification characteristics of small samples, a dropout rate of 0.5 was set (higher than the usual 0.2-0.3 to enhance regularization).

[0054] Transfer learning is used to alleviate model overfitting by transferring model weights pre-trained on large-scale datasets to the target task. This is achieved by freezing low-level parameters and fine-tuning high-level parameters to mitigate overfitting on small samples. In one specific implementation, ImageNet pre-trained weights are loaded to initialize network parameters. Based on the domain similarity between pre-trained features and EEM low-level features (edges, peaks), the parameters of the first 130 layers are frozen, and only the last 24 layers are fine-tuned. This reduces the proportion of trainable parameters to approximately 12%, significantly reducing the dependence on sample size.

[0055] The cross-entropy loss function is used during training:

[0056] Where N represents the number of training samples, K represents the number of pollution source categories, and L represents the cross-entropy loss value. The true label for sample n (category c is 1, and the rest are 0). This represents the predicted class probability by the model. For comparison, as shown in Figure 2, the training results of the single-input baseline model include the accuracy change curve, the loss function decrease curve, and the confusion matrix. In this embodiment, the dual-input model, after fusing the physicochemical essence information of EEM image spatial features and fluorescence feature parameters, achieves a validation set accuracy of 90% and a test set accuracy of 86.7%, representing improvements of 16% and 16.7% respectively compared to the single-input baseline model; the independent validation set recognition accuracy reaches 91.3%, achieving zero misclassification among non-alcoholic beverage companies.

[0057] To further explain the reasons for the above performance improvement, Figure 2c In the confusion matrix shown, the blue area along the main diagonal represents the predicted category that matches the true category, while the orange area off-diagonal represents the misclassified category. The color intensity represents the number of corresponding samples. The overall classification accuracy of the single-input baseline model is 74.4%. Among them, the test samples of machinery A, liquor A, and liquor F were all correctly classified; printing A and washing A each had one sample misclassified as the other. Misclassification within the liquor brewing wastewater was more obvious: two samples of mixed liquor B were misclassified as liquor B and liquor D respectively; one sample of liquor B was misclassified as mixed liquor B; one sample of liquor C was misclassified as liquor D; and three samples of liquor D were misclassified, one as liquor C and two as liquor E; two samples of liquor E were misclassified as liquor D. The above results indicate that relying solely on the spatial structure and intensity distribution of EEM images is insufficient to fully characterize the subtle differences between highly similar liquor wastewater samples. Therefore, introducing feature parameters such as FI, BIX, HIX, β:α, total fluorescence intensity, and peak position, and fusing them with EEM images as dual inputs, helps improve the model's comprehensive ability to discriminate DOM attributes and spectral spatial features.

[0058] Example 6 When determining the wastewater pollution source identification results based on probability distribution, the probability P of each pollution source category in the probability distribution is used as follows: i Converted to category score The calculation formula is as follows:

[0059] in, This represents the original score of the i-th input sample after being identified by the convolutional network; W represents the parameter matrix. If the input feature dimension is D and the number of categories is K, then W should be K×D dimensional. This represents the feature vector after global average pooling in MobileNetV2; This represents the 1×K-dimensional threshold matrix obtained from training the convolutional network.

[0060]

[0061] in, —The final score of the i-th input sample after being identified by the convolutional network; This represents the maximum original score of all input samples; This represents the minimum original score of all input samples. The values ​​are dimensionless. The scores for each pollution source category are sorted in descending order, and the top K items are selected to obtain the Top-K ranking result, where K ≥ 3. The score interval ΔS between the first-ranked category score S1 and the second-ranked category score S2 in the Top-K ranking result is calculated. When ΔS ≥ 10, the Top-1 result is output as the wastewater pollution source identification result; when 5 ≤ ΔS < 10, the Top-2 candidate result is output as the wastewater pollution source identification result; when ΔS < 5, the Top-3 candidate result is output as the wastewater pollution source identification result. In this invention, the preset maximum number of candidates K can be flexibly set according to regulatory needs; when K is greater than 3, the actual number of official outputs is still dynamically determined by the distance ΔS between the first and second choice scores, specifically: when ΔS ≥ 10, Top-1 is output; when 5 ≤ ΔS < 10, Top-2 is output; when ΔS < 5, Top-3 is output; the categories from Top-4 to Top-K are only used as background sorting information and are not used as the official candidate pollution source identification results in this embodiment.

[0062] Top-K decision-making is a probabilistic ranking mechanism that retains the top K most probable candidate categories output by the model, rather than forcing a single category to be output. For industrial wastewater with extremely high fluorescence similarity (e.g., the Pearson similarity between alcoholic beverages D and E is >85%), traditional hard classification methods are prone to misclassification. This embodiment introduces a Top-K ranking mechanism, transforming rigid classification into flexible source tracing.

[0063] When performing Top-K ranking, the scores of each category output by the model are considered. Sort in descending order and take the first K items:

[0064] Where x represents the model output object corresponding to the wastewater sample to be traced, N represents the total number of pollution source categories, and K represents the number of candidate pollution source categories retained. This represents the category score of the Kth candidate category.

[0065] in As the primary suspect company, to The candidates are selected sequentially up to the Kth candidate. (Introduction) As a confidence criterion: when Delta S>=10, the Top-1 results can be directly adopted even if K=1; when 5<=Delta S<10, the list needs to be expanded to Top-2 candidates; when Delta S<5, all Top-3 candidates are included in the mandatory inspection list. Supervisory personnel conduct manual verification based on the ranking of the suspected list, combined with the company's drainage time period or pipeline network information.

[0066] Table 1. Wastewater source identification results based on Top-K ranking (K=3)

[0067] As shown in Table 1, Top-K flexible tracing transforms the hard classification of unique identification into a flexible tracing based on a suspect list. Simulations show that even after the wastewater is diluted and mixed with surface water, the Top-1 identification accuracy still reaches 85.7%, and the accuracy reaches 100% after combining Top-3 ranking.

[0068] Example 7 Wastewater from various enterprises was mixed with surface water from the surrounding industrial park in specific proportions to simulate illegal discharge scenarios. The wastewater volume fractions were set to 5%, 10%, 20%, and 50% to cover different conditions ranging from low-concentration dilution to high-concentration discharge. After mixing, the mixture was left to stand for approximately two days under room temperature and natural light before undergoing 3D fluorescence scanning. Background subtraction preprocessing used the EEM matrix of the surrounding surface water from the same batch as the background matrix, and subtraction was performed according to the mixing ratio. Then, scattering subtraction, interpolation standardization, and parameter normalization were performed before inputting the mixture into a trained dual-input model. A list of suspected enterprises was output through Top-K ranking. Evaluation indicators included Top-1 identification accuracy, Top-3 investigation hit rate, whether the actual enterprise was included in the mandatory investigation list, and the triggering result of the ΔS stratification criterion; as long as the actual enterprise fell into the top three, the source tracing was considered successful. This strategy avoided source tracing failures caused by a single misjudgment and effectively coupled data-driven approaches with engineering practice.

[0069] In practical applications, a three-layer database structure supports standardized identification: a three-layer fluorescent fingerprint database containing appearance features, feature parameters, and specific components is constructed, achieving multi-level coverage from industry-level to enterprise-level, providing a standardized data foundation and verification basis for intelligent identification models. Model inference can be completed in just seconds, improving efficiency by three orders of magnitude compared to traditional PARAFAC analysis (30-60 minutes), meeting the regulatory requirements for minute-level response to sudden pollution events.

[0070] Without departing from the inventive concept, the dual-input convolutional neural network model satisfies at least one of the following configurations: the image feature extraction branch uses one of MobileNetV3, ShuffleNetV2, or EfficientNet-B0; the cross-modal fusion layer uses one of attention-weighted fusion, bilinear fusion, or element-wise additive fusion; the input to the parameter feature mapping branch also includes at least one of parallel factor analysis component score vectors, fluorescence region integral parameters, and UV absorbance at 254nm. The final output feature dimensions of all branches must be consistent before fusion operations such as splicing and addition can be performed; and the transfer learning strategy should be maintained after substitution to alleviate overfitting on small samples. Substitutes with larger parameter counts require stronger regularization measures; the structure and interface of the probability output layer remain unchanged, and substitution only affects the feature extraction and fusion stages; any substitution scheme must be validated through ablation experiments to quantify the contribution of the substitution components to the overall performance.

[0071] Besides MobileNetV2, other lightweight network architectures can also be used for image feature extraction, including MobileNetV3, ShuffleNetV2, and EfficientNet-B0. These networks all feature low parameter counts and high inference efficiency, making them suitable for wastewater fluorescence recognition tasks with small sample sizes.

[0072] Besides concatenation, cross-modal feature fusion can also employ attention-weighted fusion, bilinear fusion, or element-wise additive fusion. Attention-weighted fusion introduces a channel attention mechanism (such as the SE module) to automatically learn the importance weights of each channel of image features and parametric features; bilinear fusion models the interaction relationship between the two types of feature channels through outer product operations; element-wise additive fusion uses element-wise addition after dimensional alignment and is suitable for scenarios where the feature semantic levels are similar.

[0073] Besides Top-K ranking, uncertainty quantification can also employ Monte Carlo Dropout, ensemble learning, or Bayesian neural networks. Monte Carlo Dropout keeps Dropout active during the inference phase, obtaining the variance of the prediction distribution through multiple forward propagations to quantify model uncertainty; ensemble learning trains multiple independent models and integrates the prediction results through voting or averaging to reduce the risk of misjudgment by a single model; Bayesian neural networks model the weight parameters as a probability distribution and infer the natural output of prediction uncertainty through posterior inference.

[0074] In addition to the six fluorescence characteristic parameters, the second input mode can also be expanded to include the component score vector after PARAFAC decomposition, the volume percentage of the five regions in the fluorescence region integral (FRI), and the ultraviolet absorbance at 254 nm (SUVA254). These parameters can all characterize the origin and chemical features of the DOM, effectively complementing the EEM image.

[0075] When using the above-mentioned alternative parameters as input to the parameter feature mapping branch, the selected alternative parameters and the basic fluorescence feature parameters are normalized together and organized into a one-dimensional parameter vector, which is still mapped to the parameter features consistent with the image feature dimension through the parameter feature mapping branch.

[0076] Example 8 This embodiment combines Figure 5 The dual-input fusion network structure and dimensionality transformation process shown illustrate the actual identification process of a wastewater sample to be traced. This embodiment further explains how the present invention inputs standardized EEM images and fluorescence feature parameters into a dual-input convolutional neural network model, and completes pollution source identification through feature extraction, feature mapping, feature concatenation, adaptive fusion, and probability output.

[0077] First, a wastewater sample from the liquor industrial park was collected for source tracing. The sample was then filtered, diluted, and subjected to three-dimensional fluorescence scanning to obtain its three-dimensional excitation-emission matrix fluorescence spectral data. After blank subtraction, scattering removal, Raman normalization, and noise reduction, standardized three-dimensional fluorescence data suitable for model input was obtained from the raw EEM data.

[0078] Subsequently, two input data streams were generated based on the same three-dimensional excitation-emission matrix fluorescence spectroscopy data. The first input stream was... Figure 5 The EEM image input shown on the left converts the preprocessed EEM matrix into a standardized 224×224 pixel image through interpolation, and then copies it into a three-channel format to form an EEM image with a size of 224×224×3. This EEM image is used to represent the spatial structure information of the fluorescence peaks in the wastewater sample, such as their location, peak shape, regional distribution, and intensity variations.

[0079] The second input is Figure 5 The fluorescence index input shown on the right extracts six fluorescence characteristic parameters from the same EEM data: fluorescence index FI, humification index HIX, autogenous origin index BIX, freshness index β:α, total fluorescence intensity (TotalIntensity), and peak position (PeakPos). After normalization, these parameters are combined into a 1×6 fluorescence characteristic parameter vector. This vector is used to characterize the source attributes of dissolved organic matter in wastewater, the degree of humification, the proportion of fresh organic matter, the overall fluorescence response intensity, and the position of the main fluorescence peak.

[0080] Then, input an EEM image with a size of 224×224×3. Figure 5 The MobileNet-V2 feature extractor on the left. This feature extractor extracts features from the EEM image layer by layer through depthwise separable convolution, inverted residual structure, and linear bottleneck structure, and compresses the image features into a 320-dimensional image feature vector at the end through global average pooling. This 320-dimensional image feature vector mainly reflects the spatial distribution of fluorescence peaks, peak topology, and intensity texture differences in the EEM image.

[0081] Simultaneously, input a 1×6 fluorescence feature parameter vector. Figure 5 The fully connected layer FC-1 on the right. Since the fluorescence feature parameters are one-dimensional scalars and do not have the two-dimensional spatial structure of an image, convolution is not used. Instead, a nonlinear dimensionality-upgrading mapping is performed through the fully connected layer. In this embodiment, FC-1 maps the fluorescence feature parameter vector according to a 6→64→320 structure, converting the original 1×6 parameter vector into a 320-dimensional parameter feature vector. This 320-dimensional parameter feature vector maintains the same dimensionality as the 320-dimensional image feature vector output by MobileNet-V2 on the left, facilitating subsequent fusion.

[0082] Next, the 320-dimensional image feature vector output by MobileNet-V2 and the 320-dimensional parameter feature vector output by FC-1 are input into... Figure 5 The central feature stitching layer stitches together the two types of features along the channel dimension to obtain a 640-dimensional joint feature vector. This joint feature vector simultaneously contains the spatial structure information of the EEM image and the source information of dissolved organic matter represented by the fluorescence index, achieving cross-modal complementarity between image features and parametric features.

[0083] Then, the 640-dimensional joint feature vector is input. Figure 5The lower layer is the fully connected layer FC-2. FC-2 adaptively fuses joint features according to a 640→320 structure, outputting a 320-dimensional fused feature vector. During model training, this fully connected fusion layer automatically learns the weight relationship between image features and parameter features through backpropagation, enabling the model to automatically adjust the contribution of the two types of input information in the final discrimination based on the feature differences of different types of wastewater. For example, for non-alcoholic beverage wastewater, the model can make greater use of fluorescence index and peak position to distinguish exogenous mixed DOM; for baijiu wastewater such as alcoholic beverage D and alcoholic beverage E, which have high similarity in fluorescence images, the model can combine parameters such as HIX, BIX, and β:α to further amplify the subtle differences between the two in terms of humification degree and fresh organic matter contribution.

[0084] Finally, the 320-dimensional fused features output by FC-2 are input into the Softmax classifier. The Softmax classifier outputs the probability distribution of each pollution source category based on the fused features; the higher the probability value, the greater the likelihood that the wastewater sample belongs to the corresponding pollution source category. The model can further convert the probabilities of each category into category scores and generate a Top-K ranking result based on the scores from highest to lowest.

[0085] In a specific example, if a wastewater sample to be traced is identified by the above model, and the top three candidate categories corresponding to the Softmax output are alcoholic beverage E, alcoholic beverage D, and mixed alcoholic beverage B, with category scores of 90.9, 89.7, and 78.6 respectively, then the score interval between the first and second categories is ΔS = 90.9 - 89.7 = 1.2. Since ΔS < 5, it indicates that the model's confidence in distinguishing between the first and second categories is insufficient, and the two may have high fluorescence similarity. In this case, instead of directly outputting a single pollution source, alcoholic beverage E, alcoholic beverage D, and mixed alcoholic beverage B are simultaneously included in the Top-3 candidate pollution source list, allowing regulatory personnel to further confirm the source based on the company's discharge time, pipeline location, and on-site investigation results.

[0086] Therefore, this embodiment follows Figure 5 The illustrated process sequentially completes the entire operation of "EEM image input—MobileNet-V2 image feature extraction—fluorescence index input—FC-1 parameter feature mapping—feature stitching—FC-2 adaptive fusion—Softmax probability output—Top-K candidate source tracing." This process can simultaneously utilize the spatial structural features of EEM images and the physicochemical meaning of fluorescence feature parameters, avoiding the problem of misjudgment that easily occurs in the identification of highly similar liquor wastewater using a single EEM image input, thereby improving the accuracy, stability, and engineering applicability of wastewater pollution source identification in liquor industrial parks.

[0087] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the implementation methods of the present invention, and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the present invention.

Claims

1. A method for tracing the source of fluorescence in wastewater from a liquor industrial park based on a dual-input convolutional neural network, characterized in that, include: Obtain three-dimensional excitation-emission matrix fluorescence spectral data of wastewater samples from the liquor industrial park to be traced; A standardized EEM image is generated based on the three-dimensional excitation-emission matrix fluorescence spectral data, and fluorescence feature parameters are extracted from the same three-dimensional excitation-emission matrix fluorescence spectral data to form a fluorescence feature parameter vector; The standardized EEM image is input into the image feature extraction branch of a trained dual-input convolutional neural network model to obtain image features; The fluorescence feature parameter vector is input into the parameter feature mapping branch of the dual-input convolutional neural network model to obtain the parameter features; The image features and the parameter features are fused through a cross-modal fusion layer to obtain fused features; The fused features are input into the probability output layer to obtain the probability distribution of each pollution source category, and the wastewater pollution source identification result is determined based on the probability distribution.

2. The method for tracing the source of fluorescence in wastewater from a liquor industrial park based on a dual-input convolutional neural network according to claim 1, characterized in that, Before obtaining the three-dimensional excitation-emission matrix fluorescence spectral data of the wastewater samples to be traced in the liquor industrial park, the process also includes: collecting wastewater samples from each enterprise in the liquor industrial park, obtaining EEM data of each enterprise's wastewater samples through a three-dimensional fluorescence spectrometer, and associating the EEM data of each enterprise's wastewater samples with the corresponding pollution source category label; constructing a three-dimensional fluorescence fingerprint database based on the EEM data of each enterprise's wastewater samples, the three-dimensional fluorescence fingerprint database including an appearance feature layer, a feature parameter layer, and a specific component layer; the appearance feature layer is used to record the position and intensity information of the main fluorescence peaks of the wastewater samples, the feature parameter layer is used to record the fluorescence feature parameters extracted from the EEM data, and the specific component layer is used to record the enterprise-specific fluorescence components obtained through parallel factor analysis. The three-dimensional fluorescent fingerprint database is used to provide the EEM data, pollution source category labels and fluorescence feature parameters required for training the dual-input convolutional neural network model, and to establish the correspondence between the pollution source categories in the probability distribution and the enterprises in the liquor industrial park.

3. The method for fluorescence source tracing of wastewater from a liquor industrial park based on a dual-input convolutional neural network according to claim 1 or 2, characterized in that, Generating a standardized EEM image based on the three-dimensional excitation-emission matrix fluorescence spectral data includes: performing bilinear interpolation on the original EEM matrix corresponding to the three-dimensional excitation-emission matrix fluorescence spectral data to uniformly adjust it into a standardized grayscale image with a preset pixel size; and copying the standardized grayscale image into a three-channel image to obtain the standardized EEM image.

4. The method for fluorescence source tracing of wastewater from a liquor industrial park based on a dual-input convolutional neural network according to claim 1, characterized in that, The steps for constructing the fluorescence feature parameter vector include: extracting the fluorescence index, humification index, biological index, freshness index, total fluorescence intensity, and main peak position from the three-dimensional excitation-emission matrix fluorescence spectral data, and then normalizing them to construct the fluorescence feature parameter vector.

5. The method for tracing the source of fluorescence in wastewater from a liquor industrial park based on a dual-input convolutional neural network according to claim 1, characterized in that, The dual-input convolutional neural network model includes an image feature extraction branch, a parameter feature mapping branch, a cross-modal fusion layer, and a probability output layer. The image feature extraction branch is used to perform convolutional feature extraction on the standardized EEM image to obtain image features characterizing the morphology and intensity distribution of fluorescence peaks. The parameter feature mapping branch is used to perform nonlinear dimensionality upscaling mapping on the fluorescence feature parameter vector to obtain parameter features that match the dimension of the image features. The cross-modal fusion layer is used to fuse the image features and the parameter features, and the probability output layer is used to output the probability distribution of each pollution source category based on the fused features.

6. The method for tracing the source of fluorescence in liquor industrial park wastewater based on a dual-input convolutional neural network according to claim 5, characterized in that, The image feature extraction branch uses MobileNetV2 as the backbone network to perform depthwise separable convolution processing on the standardized EEM image to obtain image features; the parameter feature mapping branch includes at least two fully connected layers, and the fluorescence feature parameter vector is nonlinearly up-dimensionally mapped through at least two fully connected layers to obtain parameter features consistent with the image feature dimension.

7. The method for fluorescence source tracing of wastewater from a liquor industrial park based on a dual-input convolutional neural network according to claim 5 or 6, characterized in that, The image features and the parameter features are fused through a cross-modal fusion layer to obtain fused features, including: concatenating the image features and the parameter features along the channel dimension to obtain joint features; and inputting the joint features into a fully connected fusion layer for adaptive fusion to obtain fused features. The probability output layer uses a Softmax classifier.

8. The method for fluorescence source tracing of wastewater from a liquor industrial park based on a dual-input convolutional neural network according to claim 2, characterized in that, The training steps of the dual-input convolutional neural network model include: obtaining EEM data, pollution source category labels, and fluorescence feature parameters corresponding to wastewater samples with known pollution source categories from the three-dimensional fluorescence fingerprint database; performing data augmentation on the original EEM matrix corresponding to the EEM data to obtain an enhanced EEM matrix, wherein the data augmentation includes at least one of random scaling of fluorescence intensity, micro-shifting of excitation and emission wavelength directions, and addition of Gaussian noise; generating a standardized EEM image for training based on the original EEM matrix and the enhanced EEM matrix, and constructing a fluorescence feature parameter vector for training based on the fluorescence feature parameters; constructing training samples based on the standardized EEM image for training, the fluorescence feature parameter vector for training, and the pollution source category labels; and training the dual-input convolutional neural network model using the training samples.

9. The method for fluorescence source tracing of wastewater from a liquor industrial park based on a dual-input convolutional neural network according to claim 1, characterized in that, Determining the wastewater pollution source identification result based on the probability distribution includes: converting the probability of each pollution source category in the probability distribution into a category score, sorting the category scores of each pollution source in descending order, and taking the top K items to obtain the Top-K ranking result, where K≥3; calculating the first category score in the Top-K ranking result. Compared to the second category score The score interval ΔS between them, the score interval When ΔS≥10, the Top-1 result is output as the wastewater pollution source identification result; when 5≤ΔS<10, the Top-2 candidate result is output as the wastewater pollution source identification result; when ΔS<5, the Top-3 candidate result is output as the wastewater pollution source identification result.

10. The method for fluorescence source tracing of wastewater from a liquor industrial park based on a dual-input convolutional neural network according to claim 5, characterized in that, The dual-input convolutional neural network model satisfies at least one of the following configurations: the image feature extraction branch adopts one of MobileNetV3, ShuffleNetV2, and EfficientNet-B0; the cross-modal fusion layer adopts one of attention-weighted fusion, bilinear fusion, and element-wise additive fusion; the input of the parameter feature mapping branch further includes at least one of parallel factor analysis component score vector, fluorescence region integral parameter, and ultraviolet absorption value at 254nm.

Images