Method and system for identifying perfluorooctanoic acid pollutants in tap water

Abstract

The invention discloses a method and system for identifying perfluorooctanoic acid pollutants in tap water, and relates to the field of water environment new pollutant detection, and the method comprises the steps of SERS spectral signal collection, signal preprocessing, bimodal learning device construction and training verification, and concentration prediction result generation based on numerical value and image bimodal. The to-be-predicted data is put into the trained bimodal learning device for prediction and identification, and the concentration of the detection sample is output, so that batch rapid detection based on surface enhanced Raman spectroscopy and bimodal deep learning is realized, and high-precision identification and prediction based on small sample data are realized; the accurate detection result of the perfluorooctanoic acid pollutants in the tap water is obtained.

Description

Technical Field

[0001] The present invention relates to the field of detection of new pollutants in water environments, and in particular to a method and system for identifying perfluorooctanoic acid pollutants in tap water. Background Art

[0002] PFOA, a typical perfluoroalkyl substance (PFAS), has attracted considerable attention due to its strong hydrophobicity, chemical stability, and potential harm to human health. While traditional chromatography-mass spectrometry offers highly sensitive detection, and fluorescence "on / off" methods have been widely used due to their ease of operation, these methods generally rely on indirect signal changes, resulting in low sensitivity and insufficient specificity. In contrast, surface-enhanced Raman scattering (SERS) technology, with its molecular fingerprinting capabilities and single-molecule detection potential, is becoming a cutting-edge approach in PFOA detection. By designing functional nanosubstrates or introducing Raman markers, SERS technology can obtain molecular vibrational fingerprints to specifically extract the chemical bond vibration modes and molecular configuration information of the analyte. This overcomes the challenges of the low Raman cross-section of PFOA's carbon-fluorine bonds and weak binding to "hotspot" regions, enabling direct and rapid detection of PFOA and facilitating on-site monitoring.

[0003] However, existing SERS detection methods still have many limitations. Testing of natural aquatic environments is mostly focused on deionized water, and there is no effective data-driven monitoring method for PFOA. Data processing efficiency is insufficient, and manual analysis methods are time-consuming and labor-intensive, resulting in poor predictions of PFOA concentrations and an inability to accurately reflect actual conditions. Algorithm adaptability is limited, and existing machine learning solutions mostly rely on high-throughput training data. However, in actual field monitoring scenarios, only small sample data can be obtained. In such situations, the accuracy of existing machine learning models can drop significantly to below 50%. Moreover, most of these models lack integrated batch detection capabilities, making them unable to meet the needs of large-scale, rapid field testing. Summary of the Invention

[0004] In order to solve the technical problems in the prior art, the present invention provides a method and system for identifying perfluorooctanoic acid pollutants in tap water.

[0005] The present invention is achieved through the following technical solutions: A method for identifying perfluorooctanoic acid contaminants in tap water, comprising: S1, SERS spectral signal acquisition: obtain sample solutions in batches, and collect enhanced spectral signals through SERS technology, and save the collected spectral signals in groups and transmit them to the terminal; S2, signal preprocessing: performing signal correction and noise reduction on the collected raw SERS spectral signal, generating bimodal data of numerical modal samples and image modal samples, and establishing a unified category mapping logic; S3. Build a bimodal learner; generate concentration prediction results based on numerical and image bimodal data; S4, training and validating the bimodal learner; S5. Output the concentration of the detected sample; put the data to be predicted into the trained bimodal learner for prediction and recognition, and output the obtained sample concentration prediction value.

[0006] Furthermore, the bimodal data of the numerical modal samples and the image modal samples correspond to spectral intensity and spectral peak maps respectively.

[0007] Furthermore, the bimodal learner is divided into a numerical modality learning module, an image modality learning module, a cross-modality connection mechanism and a bimodal dynamic decision module.

[0008] Furthermore, the numerical modal learning module uses a Stacking ensemble learning architecture, which is divided into three core stages: integration and optimization of pre-learners, feature extraction and integration, meta-model training and final evaluation; five few-sample targeted learners are used as base learners for feature extraction, and after this step is completed, the output features of the above five base learners are orderly integrated into the Stacking ensemble learning architecture, and the GBDT meta-learner is used to perform the concentration grading prediction task to generate output module prediction results.

[0009] Furthermore, the image modality learning module is based on a dual-convolution architecture CNN model that can realize cascade feature extraction, which is composed of two cascaded convolution-pooling modules.

[0010] Furthermore, the cross-modal connection mechanism pre-sets a confidence threshold, screens the numerical modal prediction probability, extracts samples whose values ​​are below the threshold and whose numerical modal prediction concentration magnitude is a specific category, finds their matching image modal data and inputs them into the multi-task learning network.

[0011] Furthermore, the dual-modal dynamic decision module sets fusion weights to construct a decision fusion strategy based on the numerical modal learner. When the numerical modal learner outputs a high-concentration sample, the decision mechanism directly adopts the learning result. When the numerical modal learner outputs a low-concentration sample, a dynamic decision mechanism is set.

[0012] Furthermore, the model training process is divided into numerical and image modality training verification and cross-modality comprehensive verification.

[0013] The present invention also provides a system for identifying perfluorooctanoic acid contaminants in tap water, based on the above-mentioned method for identifying perfluorooctanoic acid contaminants in tap water, comprising: A spectral signal acquisition module, which is used to acquire SERS spectral signals; The signal preprocessing module is used to perform signal correction and noise reduction on the collected raw SERS spectral signal, generate bimodal data of numerical modal samples and image modal samples, and establish a unified category mapping logic; A dual-modal learner building block, which includes a numerical modality learning module, an image modality learning module, a cross-modality connection mechanism, and a dual-modality dynamic decision module, is used to generate concentration prediction results based on numerical and image dual modalities; The dual-modal learner training and verification module is used to train and verify the corresponding modal learning modules according to different modalities; The output module is used to put the data to be predicted into the trained bimodal learner for prediction and recognition, and output the obtained sample concentration prediction value.

[0014] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, on which program instructions of a method for identifying perfluorooctanoic acid contaminants in tap water are stored. The program instructions of the method for identifying perfluorooctanoic acid contaminants in tap water can be executed by one or more processors to implement the steps of the method for identifying perfluorooctanoic acid contaminants in tap water as described above.

[0015] Compared with the existing technology, the beneficial effects of the present invention are that it realizes batch rapid detection based on surface-enhanced Raman spectroscopy analysis and dual-modal deep learning, and realizes high-precision detection based on small sample data. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 This is a flow chart of a method for identifying perfluorooctanoic acid contaminants in tap water according to an embodiment of the present application; Figure 2 is a graph showing the relationship between the characteristic wavelength peak and the concentration of the target detection substance in the solution according to an embodiment of the present application; Figure 3 It is a schematic diagram of the process of constructing a dual-modal learner according to an embodiment of the present application. DETAILED DESCRIPTION

[0017] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0018] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0019] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention. The illustrations only show components related to the present invention and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be changed at will, and the component layout form may also be more complicated.

[0020] See also Figure 1 A method for identifying perfluorooctanoic acid contaminants in tap water comprises the following steps: S1. SERS spectral signal acquisition Under the same operating procedures and natural environment, sample solutions were obtained in batches using the same method, and enhanced spectral signals were collected using SERS technology. The collected spectral signals were grouped, saved, and transmitted to the terminal.

[0021] Preferably, in step S1, the signal acquisition method is specifically: using microstructured optical fibers (MOFs) modified with β-cyclodextrin functionalized silver nanoparticles (β-CD / AgNPs) as a SERS substrate, collecting SERS signals in batches and at different concentrations with similar time spans in solutions with PFOA concentrations of 0 mg / L, 4 mg / L, 40 mg / L, 400 mg / L, and 4×10³ mg / L, and controlling the quantity during the process so that the sample size of each concentration gradient is similar.

[0022] Preferably, the sample solution is collected in three batches; each batch contains about 200 samples, and the sample size for each concentration gradient is similar. The first two of the three groups of samples are used as training sets, and the last group is used as a test set.

[0023] S2. Signal preprocessing, including: S21: signal sample generation; The collected raw SERS spectral signal is processed in two steps: first, signal correction and noise reduction are performed, and then the corrected raw spectral signal is used to generate corresponding images and numerical data. That is, the spectral signal is converted into dual-modal data of numerical mode (spectral intensity) samples and image mode (spectral peak map) samples and stored separately.

[0024] Preferably, in step S21, the signal correction and noise reduction processing method is specifically: The original SERS signal in S1 was subjected to the adaptive iterative reweighted penalized least squares method (airPLS) with consistent parameters to achieve signal baseline correction and noise reduction. The peak value of the characteristic wavelength of the spectral signal in the training set was analyzed by simple average analysis. The relationship between the known characteristic wavelength peak value and the concentration of the target detection substance in the solution was as follows: Figure 2 .

[0025] Optionally, the program is implemented in Python.

[0026] S22: signal sample encoding; Based on the generated sample bimodal data, a unified category mapping logic is established, and the training set and test set are divided to obtain logically corresponding bimodal training set and test set.

[0027] The sample features are analyzed using a random forest based classifier, and a subsequent learning mechanism is constructed based on the sample features.

[0028] Preferably, the method for establishing the mapping logic is specifically as follows: Class mapping and logical alignment were performed on the image and numerical modal data of the original samples to ensure one-to-one correspondence between samples in different modalities. A unified class mapping was performed for samples in both numerical and image modalities. Samples extracted from solutions containing 0 mg / L, 4 mg / L, 40 mg / L, 400 mg / L, and 4×10³ mg / L in S1 were labeled as class 0, class 1, class 2, class 3, and class 4, respectively. This classification rule was enforced across both image and numerical modalities to ensure semantic consistency across modalities. Furthermore, the bimodal data, after classification, was logically bundled according to the unique index identifier generated by the time-series centralized acquisition (the sequential numbering of the centralized acquisition). While maintaining the correspondence between the same classes, the number and content of image and numerical samples were strictly aligned, with no mismatches. This resulted in a verifiable mapping path from physical entity (concentration solution signal) to abstract label (classes 0-4) to data entity (multimodal sample).

[0029] S23: Signal feature encoder The encoded samples are divided into modalities for feature encoding. Image enhancement and standardization conversion are performed on image modal data, and parallel normalization is performed on numerical modal data.

[0030] Preferably, in step S23, the feature coding processing method is specifically: The sample data encoding operation for the image modality performs feature enhancement based on the sample data of the image modality. The input spectrum is first mapped and scaled to a base size of 224×224 pixels. Augmentation operations such as flipping are then performed. The original PIL image is converted to the PyTorch tensor format and normalized by RGB channel (mean [0.485, 0.456, 0.406], standard deviation [0.229, 0.224, 0.225]).

[0031] The feature encoding method for numerical modes includes: first normalizing the data, and the spectral data are normalized to the [0,1] range to suppress background fluorescence interference; performing dual-modal data parallelism, and using standardized values ​​and intrinsic values ​​as training data to enhance the recognition ability of PFOA characteristic wavelength peaks.

[0032] S3. Building a bimodal learner Based on the feature analysis, an integrated learning mechanism for bimodal fusion is formulated.

[0033] The bimodal learner adopts a numerical-image bimodal input and dynamic prediction output architecture, which is divided into a numerical modal learning module, an image modal learning module, a cross-modal connection mechanism and a bimodal dynamic decision module.

[0034] The input data samples enter the numerical modality learning module and the image modality learning module respectively according to the different modalities to realize multi-scale feature extraction. After completing the feature extraction, the low-dimensional features output by the two learning modules will be input into the fixed cross-modal decision module, and the recognition results will be comprehensively processed according to the preset decision rules to generate the predicted concentration as the final output result, realizing accurate analysis and effective prediction of the data. The specific process is as follows Figure 3 shown.

[0035] The numerical modal learning module utilizes a stacking ensemble learning architecture, which can be divided into three core stages: integration and optimization of pre-learners, feature extraction and integration, and meta-model training and final evaluation. Five small-shot targeted learners—random forest (RF), gradient boosted decision tree (GBDT), support vector machine (SVM), Adaboost classifier, and Catboost classifier—are used as base learners for feature extraction. After this step, the output features of these five base learners are systematically integrated into the stacking ensemble learning architecture. The GBDT meta-learner is used to perform concentration grading predictions, generating outputtable module prediction results. A strict 5-fold cross-validation strategy is employed to generate input features for the meta-model.

[0036] The integration and optimization method of the pre-learner specifically includes: using multiple base learners with linear and nonlinear kernel functions to perform feature dimensionality reduction, and performing parameter optimization based on a cross-validation score structure.

[0037] After optimization, the following corresponding pre-learners were obtained: The Random Forest (RF) pre-learner enhances modeling capabilities by setting a large number of base decision trees and an entropy criterion, while limiting the maximum number of features per tree to 52 for increased robustness. AdaBoost uses shallow decision trees (max_depth=5) as its base classifier, employs the SAMME algorithm, and has a high learning rate (0.8). It iteratively adjusts sample weights to improve its ability to recognize difficult samples. The Gradient Boosted Tree (GBDT) uses a small learning rate (0.45) to balance model bias and variance during a gradual optimization process in the gradient direction. CatBoost utilizes a tree structure with a depth of 9 and GPU-accelerated training, with an early stopping mechanism to dynamically terminate invalid iterations. The Support Vector Machine (SVC) uses the RBF kernel function to handle nonlinear separability problems, and adjusts the regularization strength (C=0.2) to mitigate the impact of class imbalance.

[0038] The image modality learning module specifically initializes a convolutional neural network (CNN) model. Specifically, the CNN model employs a dual-convolution architecture that enables cascaded feature extraction. This model consists of two cascaded convolution-pooling modules. The first module processes a three-channel input image and extracts low-level visual features such as edges and spots using 16 3×3 convolution kernels. All features are processed using a ReLU activation function before being reduced to a 112×112 resolution using a 2×2 max pooling layer. The second module expands the number of channels to 32, feeding the initially extracted features into the second module to further capture mid-level semantic features such as color combinations and local shape patterns. After further processing through the pooling layer, a 56×56 resolution feature image is output. Finally, the input spatial features are flattened and fed into a 128-dimensional fully connected layer. A softmax function is used to map the features to a three-dimensional output space of classes 0, 1, and 2, completing the nonlinear conversion from raw pixels to semantic categories. The model training process has three core features: automatic selection of training devices based on GPU availability to accelerate calculations; optimization specifically for difficult-to-distinguish low-concentration training sample sets; the use of a cross-entropy loss function during training, dynamic selection of optimal parameters by real-time monitoring of the score value (weighted F1-score) of the iterative process, and the provision of an early stopping mechanism to actively suppress the risk of overfitting.

[0039] Optionally, the first convolutional layer processes the 3-channel input image, extracting low-level visual features (such as edges and color spots) using 16 3×3 convolutional kernels. This is then downsampled to a 112×112 resolution after ReLU activation and 2×2 max pooling. The second convolutional layer expands the number of channels to 32, further extracting mid-level features (such as texture combinations and local shape patterns), and pooling again to produce a 56×56 feature map. The flattened features are then mapped to a 3-class softmax output space via a 128-dimensional fully connected layer, achieving a nonlinear conversion from raw pixels to semantic categories.

[0040] Specifically, the cross-modal connection mechanism achieves cross-modal connection through confidence threshold screening and local feature mapping. A pre-set confidence threshold is used to screen the numerical modal prediction probabilities. Samples below the threshold q and with numerical modal predicted concentrations in categories 0, 1, and 2 in S2 are extracted. Exactly matching image modal data is then found and fed into a multi-task learning network (image feature learning module). Within this module, classification and spectral peak area regression tasks are performed simultaneously, calculating the classification loss and spectral peak area regression loss, respectively. These two tasks are then jointly optimized using an optimization algorithm.

[0041] Specifically, the dual-modal dynamic decision-making module sets fusion weights to construct a decision fusion strategy based on the numerical modality learner. The weight of the CNN model prediction results based on image samples is 0.3, and the weight of the Stack model prediction results based on numerical samples is 0.7. For samples with high concentrations output by the numerical modality learner (category 3 and category 4), the decision mechanism directly adopts the learning results. For samples with low concentrations output by the numerical modality learner (category 0, category 1, category 2), a dynamic decision-making mechanism with a confidence threshold of 0.8 is set. For samples with conflicting dual-modality results, the predicted probability metric of the stack model is prioritized. If the confidence level is greater than 0.8, the numerical modality recognition concentration value is prioritized. Otherwise, the reference image modality prediction value is used as the output result.

[0042] The Stacking integrated learning architecture is specifically as follows: First, all base learners (bottom-level learners) are trained and their parameters optimized through rigorous 5-fold cross-validation (K-fold, K=5). This transforms each sample in the original training set into a learnable prediction probability vector (i.e., a predicted probability vector for each sample's concentration level) to prevent data leakage. Next, the probabilistic predictions of the base learners generated by these cross-validation steps are extracted as new meta-features. These are sequentially concatenated and integrated according to the base learner type and class probability, creating a meta-training set with dimensions equal to the number of base learners x the number of classes (5 x 3 = 15 dimensions), with the labels remaining unchanged from the original concentration levels. This meta-training set is then used to train the GBDT model (the upper-level learner in the architecture) to optimally combine the predictions from the base learners. Finally, the base models derived from the complete training set simultaneously generate new probabilistic predictions, which are then sequentially integrated into a set of meta-feature vectors using the same method. This is then fed into the trained GBDT meta-learner to output the final concentration classification result.

[0043] S4. Training and validating the bimodal learner The corresponding modal learning module is used for iterative learning according to the different modalities. During the training process, the training set data is divided into five groups for cross-validation. The optimal model parameters learned are fixed and saved to a file. Then, an evaluation mechanism is used to comprehensively verify and evaluate the performance of the bimodal learner based on the bimodal training data set.

[0044] The model training process is divided into two parts: numerical and image modal training verification and cross-modal comprehensive verification.

[0045] (1) Numerical modal training: Multiple categories of base learners are selected and differentiated parameter configuration strategies are adopted to optimize the performance of a single model.

[0046] The random forest algorithm uses 200 decision trees and an entropy splitting criterion, with a maximum feature count of 52 to control overfitting. AdaBoost uses a decision tree with a maximum depth of 5 as the base classifier, employing the SAMME algorithm and a high learning rate (0.8) to improve the ability to distinguish difficult samples. GBDT uses a learning rate of 0.45 to balance bias-variance. SVC uses the RBF kernel with a regularization strength adjusted to 0.2 to mitigate the effects of class imbalance. CatBoost uses a lightweight tree structure (depth 9) and an early stopping mechanism (terminating after 50 rounds of no improvement). The feature fusion stage generates metamodel inputs through 5-fold cross-validation, and predictions on the test set use a multi-fold averaging strategy. The metamodel uses a lightweight GBM algorithm with a learning rate of 0.27, using Accuracy and weighted F1-score as core metrics, and is validated across multiple dimensions using confusion matrix heatmaps and learning curves.

[0047] (2) Image modality training: A convolutional network mechanism is established, and the parameters of each layer are shown in the following table. The Adam optimizer and cross entropy loss are used. The convolution layer uniformly uses a 3×3 kernel size and padding=1 to maintain spatial resolution. The pooling layer uses a 2×2 window and a stride of 2 to achieve feature dimensionality reduction. During training, random flip / rotation data augmentation is applied, and an early stopping mechanism is set (termination is performed if the loss does not decrease after three consecutive rounds of verification). The fully connected layer uses ReLU activation and finally outputs three-category classification probabilities. The following table shows the names of the various layers of the image convolutional network and their corresponding parameters:

[0048] (3) Cross-modal comprehensive verification: After the numerical and image modal models have been trained independently in stages, the parameters are frozen for fusion tuning. A fixed ratio fusion of 0.7 weight for the numerical modal Stack model and 0.3 weight for the image CNN is set. For samples whose outputs are categories 3 and 4 of the numerical modal learner, the learning result decision is directly adopted. For samples whose outputs are categories 0, 1, and 2 of the numerical modal learner, a dynamic decision mechanism with a confidence threshold of 0.8 is set. For samples with conflicting bimodal results, the predicted probability metric of the stack model is considered. If the confidence is greater than 0.8, the numerical modal recognition concentration value is adopted first. Otherwise, the reference image modal prediction value is used as the output result.

[0049] The evaluation mechanism uses Acc, Pre, and weighted F1-score as model performance evaluation indicators to evaluate the models in the single-modal output and cross-modal decision-making stages, and uses them to make a comprehensive evaluation of the models, as shown in the following table:

[0050] S5. Output the concentration of the detected sample The data to be predicted is put into the trained bimodal learner for prediction and recognition, the obtained sample concentration prediction value is returned and saved, and the analysis result is output.

[0051] Specifically, the input numerical data (normalized numerical samples and intrinsic numerical samples) are trained in the Stacking ensemble learning framework. The input image modality data enters the preset convolutional neural network training. The cross-entropy loss function (CrossEntropyLoss) is used to measure the difference between the model prediction results and the true labels. The dynamic stochastic gradient descent (SGD) optimizer is used to update the model parameters (learning rate is 0.001, momentum is 0.9). Each training cycle includes a training phase and a cross-validation phase.

[0052] By isolating the target concentration value and signal data, the dual-modal learning model obtained the predicted concentration value for the test group signal. Comparing the predicted concentration value with the original true concentration value, the obtained model classification result has high accuracy and good performance. Under the condition of the same training group and test group data, the test group prediction performance of the model implemented by the present invention and other commonly used single-modal models is compared in the following table:

[0053] In this embodiment, rapid batch detection based on surface-enhanced Raman spectroscopy and dual-modal deep learning is realized, and high-precision detection is achieved based on small sample data.

[0054] The embodiment of the present invention further provides a system for identifying perfluorooctanoic acid contaminants in tap water, based on the above-mentioned method for identifying perfluorooctanoic acid contaminants in tap water, comprising: A spectral signal acquisition module, which is used to acquire SERS spectral signals; The signal preprocessing module is used to perform signal correction and noise reduction on the collected raw SERS spectral signal and generate bimodal data of numerical mode samples and image mode samples and store them separately; based on the generated sample bimodal data, a unified category mapping logic is established; A dual-modal learner building block, which includes a numerical modality learning module, an image modality learning module, a cross-modality connection mechanism, and a dual-modality dynamic decision module, is used to generate concentration prediction results based on numerical and image dual modalities; The bimodal learner training and verification module is used to iteratively learn the corresponding modal learning modules according to different modalities. During the training process, the training set data is divided into five groups for cross-validation, and the learned optimal model parameters are fixed and saved to a file; the evaluation mechanism is used to verify and evaluate the bimodal learner of the bimodal training data set.

[0055] The output module is used to put the data to be predicted into the trained bimodal learner for prediction and recognition, and output the obtained sample concentration prediction value.

[0056] In addition, an embodiment of the present invention further proposes a computer-readable storage medium, on which program instructions for a method for identifying perfluorooctanoic acid contaminants in tap water are stored. The program instructions for the method for identifying perfluorooctanoic acid contaminants in tap water can be executed by one or more processors to implement the steps of the method for identifying perfluorooctanoic acid contaminants in tap water as described above.

[0057] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A method for identifying perfluorooctanoic acid pollutants in tap water, characterized in that: include: S1, SERS spectrum signal acquisition; Obtain sample solutions in batches, collect enhanced spectral signals using SERS technology, and save the collected spectral signals in groups and transmit them to the terminal; S2, signal preprocessing: performing signal correction and noise reduction on the collected raw SERS spectral signal, generating bimodal data of numerical modal samples and image modal samples, and establishing a unified category mapping logic; S3, build a bimodal learner; Generate concentration prediction results based on numerical and image dual modalities; S4, training and validating the bimodal learner; S5. Output the concentration of the detected sample; put the data to be predicted into the trained bimodal learner for prediction and recognition, and output the obtained sample concentration prediction value.

2. The method for identifying perfluorooctanoic acid pollutants in tap water according to claim 1, characterized in that: The bimodal data of the numerical modal samples and the image modal samples correspond to spectral intensity and spectral peak maps respectively.

3. The method for identifying perfluorooctanoic acid pollutants in tap water according to claim 1, characterized in that: The bimodal learner is divided into a numerical modality learning module, an image modality learning module, a cross-modality connection mechanism and a bimodal dynamic decision module.

4. The method for identifying perfluorooctanoic acid pollutants in tap water according to claim 1, characterized in that: In the numerical modal learning module, the Stacking ensemble learning architecture is used, which is divided into three core stages: integration and optimization of pre-learners, feature extraction and integration, meta-model training and final evaluation; five few-shot targeted learners are used as base learners for feature extraction, and after this step is completed, the output features of the above five base learners are orderly integrated into the Stacking ensemble learning architecture, and the GBDT meta-learner is used to perform the concentration grading prediction task to generate output module prediction results.

5. The method for identifying perfluorooctanoic acid pollutants in tap water according to claim 3, characterized in that: The image modality learning module is based on a dual-convolution architecture CNN model that can realize cascade feature extraction. The model consists of two cascaded convolution-pooling modules.

6. The method for identifying perfluorooctanoic acid pollutants in tap water according to claim 3, characterized in that: The cross-modal connection mechanism pre-sets a confidence threshold, screens the numerical modal prediction probability, extracts samples below the threshold and with a numerical modal prediction concentration magnitude of a specific category, finds their matching image modal data and inputs them into the multi-task learning network.

7. The method for identifying perfluorooctanoic acid pollutants in tap water according to claim 3, characterized in that: The dual-modal dynamic decision module sets fusion weights to construct a decision fusion strategy based on the numerical modal learner. When the numerical modal learner outputs a high-concentration sample, the decision mechanism directly adopts the learning result. When the numerical modal learner outputs a low-concentration sample, a dynamic decision mechanism is set.

8. The method for identifying perfluorooctanoic acid pollutants in tap water according to claim 1, characterized in that: The model training process is divided into numerical and image modal training verification and cross-modal comprehensive verification.

9. A system for identifying perfluorooctanoic acid contaminants in tap water, based on the method for identifying perfluorooctanoic acid contaminants in tap water according to any one of claims 1 to 8, comprising: A spectral signal acquisition module, which is used to acquire SERS spectral signals; The signal preprocessing module is used to perform signal correction and noise reduction on the collected raw SERS spectral signal, generate bimodal data of numerical modal samples and image modal samples, and establish a unified category mapping logic; A dual-modal learner building block, which includes a numerical modality learning module, an image modality learning module, a cross-modality connection mechanism, and a dual-modality dynamic decision module, is used to generate concentration prediction results based on numerical and image dual modalities; The dual-modal learner training and verification module is used to train and verify the corresponding modal learning modules according to different modalities; The output module is used to put the data to be predicted into the trained bimodal learner for prediction and recognition, and output the obtained sample concentration prediction value.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores program instructions for a method for identifying perfluorooctanoic acid contaminants in tap water. The program instructions for the method for identifying perfluorooctanoic acid contaminants in tap water can be executed by one or more processors to implement the steps of the method for identifying perfluorooctanoic acid contaminants in tap water as described in any one of claims 1 to 8.

Images