Hydroxyl polycyclic aromatic hydrocarbon Raman spectrum detection method based on dynamic weighted voting fusion network

Abstract

The invention discloses a hydroxyl polycyclic aromatic hydrocarbon Raman spectrum detection method based on a dynamic weighted voting fusion network, relates to the technical field of hydroxyl polycyclic aromatic hydrocarbon detection, and provides an effective method which is rapid, accurate and free of chromatographic separation. Spectral information of OH-PAHs is obtained by using surface enhanced Raman spectroscopy SERS, and a Raman spectrum is obtained through density functional theory DFT calculation. A dynamic weighted voting fusion network DWVF-Net model is constructed, wherein the model comprises two core parts of cross-modal attention fusion (CMAF) and dynamic weighted voting (DWV). The CMAF dynamically weights key areas of SERS and DFT Raman spectrums through cosine similarity to realize feature fusion; and the DWV adaptively adjusts the weight of the model according to the real-time performance of the model, and optimizes an integration decision. Through the synergistic effect of CMAF and DWV, complementary information of SERS and DFT Raman spectrums is effectively fused, and the adaptability advantage of a dynamic weighting mechanism in the aspect of processing a complex data environment is verified.

Description

Technical Field

[0001] The invention relates to the technical field of detection of hydroxyl polycyclic aromatic hydrocarbons, in particular to a Raman spectrum detection method of hydroxyl polycyclic aromatic hydrocarbons based on a dynamic weighted voting fusion network. Background Art

[0002] Hydroxylated polycyclic aromatic hydrocarbons (OH-PAHs) are a widespread environmental carcinogen whose carcinogenicity threatens human health, so they need to be detected quickly and accurately for exposure monitoring. Traditional detection methods for OH-PAHs mainly rely on chromatographic technology, which has many limitations. For example, the sample pretreatment process is complicated and requires a lot of time and manpower; the detection throughput is low, making it difficult to meet the needs of rapid detection of large-scale samples; and due to the structural similarity and polarity of the analytes, the selectivity is very limited, especially the high carcinogenicity at trace levels, which makes direct detection challenging. These problems limit the application of traditional methods in actual environmental monitoring, and there is an urgent need to develop new detection technologies to overcome these shortcomings. Summary of the Invention

[0003] The purpose of the present invention is to provide a Raman spectroscopy detection method for hydroxyl polycyclic aromatic hydrocarbons (OH-PAHs) based on a dynamic weighted voting fusion network, combining surface-enhanced Raman spectroscopy (SERS) with density functional theory (DFT) to improve the classification accuracy and noise resistance of OH-PAHs, and to provide a fast, accurate and effective method for the detection of OH-PAHs in the environment without the need for chromatographic separation.

[0004] The technical solution adopted by the present invention to solve the technical problem is: a Raman spectrum detection method for hydroxyl polycyclic aromatic hydrocarbons based on a dynamic weighted voting fusion network, comprising the following steps:

[0005] S1. Data collection and preprocessing

[0006] S1.1 Spectral data acquisition The Raman spectral data corresponding to the four derivatives of OH-PAHs, namely 3-OH-BAA, 3-OH-BAP, NAP and 9-OH-PHE, were obtained through density functional theory DFT calculation.

[0007] S1.2 Data Preprocessing

[0008] Perform feature selection, noise suppression and data partitioning on spectral data.

[0009] S2. DWVF-Net model construction and training S2.1 Construction of CMAF module S2.1.1 Feature alignment

[0010] The normalized SERS intensity matrix And the preprocessed DFT Raman time-frequency characteristics Dimensionality matching is performed via zero-filling or principal component analysis (PCA).

[0011] S2.1.2 Calculating Attention Weights

[0012] Cosine similarity is used to calculate the similarity between cross-modal features.

[0013] S2.1.3 Feature Fusion

[0014] The weighted features are spliced ​​together to obtain the fusion feature F fused =a*F SERS +(1-a)*F DFT (1); where F fused represents the fused feature vector; a represents the modality-specific importance score obtained by the previous calculation; F SERS represents the SERS spectrum feature vector after preprocessing, feature extraction and alignment; F DFT represents the DFT Raman spectrum feature vector after corresponding processing and alignment.

[0015] S2.2 Building DWV Module S2.2.1 Model Training

[0016] Train a pool of heterogeneous models on data from a single SERS or Raman modality.

[0017] S2.2.2 Weight calculation

[0018] Calculate the F1 score of each model according to the formula Dynamically assign weights, give priority to high-performance models, and suppress models with poor performance. k Indicates the weight assigned to the k-th classification model in the integrated model; F1 k represents the F1 score of the k-th classification model; It represents the sum of the F1 scores of all N classification models participating in the ensemble.

[0019] S2.2.3 Result Fusion

[0020] The final category probability is obtained through weighted aggregation, and the formula is: Where, P final (y|x) represents the final category probability; N represents the number of basic models involved in the fusion; w k represents the weight of the kth basic model; P k (y|x) represents the predicted probability of the k-th base model that the sample belongs to category y given the input feature x.

[0021] S2.3 Overall model training

[0022] Taking SERS and DFT Raman data as dual input, the SERS function powers three parallel models and is simultaneously combined with the DFT function via a CMAF fusion module. The CMAF module generates a discriminative hybrid representation by aligning feature dimensions and computing cross-modal similarity scores.

[0023] S3. Model optimization and evaluation

[0024] Six machine learning algorithms, including K-nearest neighbor (KNN), gradient boosting machine (GBM), Adaboost, support vector machine (SVM), multi-layer perceptron (MLP) and ensemble learning (Ensemble), are used to form a basic model pool.

[0025] Furthermore, the feature selection refers to: for SERS data, independently analyzing Raman shift and intensity; for DFT Raman data, performing 5-level decomposition through Daubechies-4 wavelet to extract time domain statistical features and frequency domain wavelet coefficients, where the time domain statistical features include mean and variance.

[0026] Furthermore, the noise suppression refers to: using a Savitzky-Golay filter to smooth the signal to suppress high-frequency noise, and then using an averaging method to perform baseline correction. The intensity calculation formula after correction is: Where, I corrected Indicates the corrected intensity; I smooth represents the spectral intensity signal after smoothing by the Savitzky-Golay filter; N represents the number of data points involved in the average calculation when calculating the baseline correction; i represents the index of the data point currently being processed; w represents the window size of the Savitzky-Golay filter.

[0027] Furthermore, the data division is to divide the spectral data after noise suppression into a training set and a test set.

[0028] Furthermore, the formula for calculating the similarity between cross-modal features is: Where, represents the SERS spectrum feature vector after preprocessing and feature extraction; Represents the processed DFT Raman spectrum feature vector; Represents the dot product operation of two eigenvectors; yes The L2 norm of yes The L2 norm of .

[0029] Furthermore, the modality-specific importance score a is generated through softmax normalization. ij , the formula is Where aij represents the modality-specific importance score obtained after softmax normalization; Represents the calculated cross-modal feature similarity Perform exponential operation.

[0030] Furthermore, after obtaining the fusion features, classification is performed through a fully connected network. The classification formula is: P(y|x)=Softmax(WF fused +b)(7); where P(y|x) represents the value of the given input feature x=F fused In the case of , the probability distribution of the sample belonging to category y; Softmax represents an activation function; W represents the weight matrix of the fully connected network; F fused represents the fused feature vector obtained by the comprehensive feature fusion step of SERS and DFT; b represents the bias vector of the fully connected network.

[0031] Furthermore, in step S2.2.1, K-nearest neighbor (KNN), gradient boosting machine (GBM), and ensemble learning (Ensemble) are selected as the basic models.

[0032] The beneficial effects of the present invention are as follows: the DWVF-Net model effectively integrates the complementary information of SERS and DFT Raman spectra through the synergistic effect of CMAF and DWV, verifies the adaptability advantage of the dynamic weighting mechanism in processing complex data environments, and proves that it can provide a lightweight but effective multimodal integration solution. The final classification accuracy rate reaches 97.14%. Compared with the basic model, DWVF-Net has a significant improvement in accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1a Surface-enhanced Raman spectra of four common derivatives of OH-PAHs;

[0034] Figure 1b Density functional theory Raman spectra of four common derivatives of OH-PAHs;

[0035] Figure 2 Apply feature strategy graph for CMAF feature fusion;

[0036] Figure 3 Flowchart of the proposed DWVF-Net;

[0037] Figure 4 is the accuracy plot of all the base models in the SERS data classification of four OH-PAHs;

[0038] Figure 5 This is the confusion matrix diagram of DWVF-Net;

[0039] Figure 6 This is the AUC value graph of the ROC curve of DWVF-Net;

[0040] Figure 7 This is a system structure diagram of the present invention. DETAILED DESCRIPTION

[0041] The present invention combines spectral technology, theoretical calculations and advanced machine learning algorithms to achieve accurate detection of OH-PAHs by constructing and optimizing the dynamic weighted voting fusion network DWVF-Net model. (1) Surface enhanced Raman spectroscopy (SERS) is used to obtain the spectral information of OH-PAHs, enabling SERS to detect molecular fingerprints, thereby distinguishing structurally similar compounds at trace levels. At the same time, the Raman spectrum is obtained through density functional theory (DFT) calculations to verify the spectral source of the SERS signal and enhance the interpretability of the model. (2) A dynamic weighted voting fusion network DWVF-Net model is constructed, which includes two core parts: cross-modal attention fusion (CMAF) and dynamic weighted voting (DWV). CMAF dynamically weights the key areas of the SERS and DFT Raman spectra through cosine similarity to achieve feature fusion; DWV adaptively adjusts the model weights according to the real-time performance of the model to optimize the integrated decision. (3) The present invention was experimentally verified using experimentally collected Raman spectral data of four OH-PAHs derivatives. The experimental results demonstrated that the present invention has high accuracy and strong noise immunity in the classification and detection of OH-PAHs, and can be effectively applied to the detection of environmental pollutants. The following describes the present invention's Raman spectral detection method for hydroxylated polycyclic aromatic hydrocarbons based on a dynamic weighted voting fusion network.

[0042] S1. Data Collection and Preprocessing S1.1 Spectral Data Collection The SERS spectra of four common derivatives of OH-PAHs were obtained from the laboratory, and the corresponding Raman spectra of the four derivatives, namely 3-OH-BAA, 3-OH-BAP, NAP and 9-OH-PHE, were obtained by DFT calculation. Figure 1a 、 Figure 1b The following table shows the SERS spectra of the four derivatives and the average Raman data obtained by DFT calculation. Judging from the SERS peak shape alone, the differences between the four derivatives are not obvious. For DFT Raman, it is not easy to find significant similarity between DFT and the corresponding material in SERS.

[0043] S1.2 Data Preprocessing

[0044] S1.2.1 Feature Selection For SERS data, Raman shift and intensity were analyzed independently. For DFT Raman data, five-level decomposition was performed using Daubechies-4 wavelet to extract time-domain statistical features and frequency-domain wavelet coefficients to capture dynamic spectral features. The time-domain statistical features include mean and variance.

[0045] S1.2.2 Noise suppression

[0046] The signal was smoothed using a Savitzky-Golay filter with the following filter parameters: window = 15, polynomial order = 3. High-frequency noise was suppressed, and baseline correction was performed using the averaging method to eliminate baseline drift. The intensity after correction was calculated using the formula: Where, I corrected Indicates the corrected intensity; I smooth represents the spectral intensity signal after smoothing by the Savitzky-Golay filter; N represents the number of data points involved in the average calculation when calculating the baseline correction; i represents the index of the data point currently being processed; w represents the window size of the Savitzky-Golay filter.

[0047] S1.2.3 Data Division

[0048] A stratified sampling ratio of 7:3 was used to divide the preprocessed data into training and test sets to ensure consistent distribution of various types of data and avoid errors.

[0049] S2. DWVF-Net model construction and training DWVF-Net aims to integrate complementary information from different data sources to enhance the classification ability of the model through collaboration. The different data sources are spectra acquired by SERS and Raman spectra obtained by DFT calculations. Since the Raman spectrum obtained by DFT calculations can determine the ground state geometry of the molecule, if high similarity features can be resolved from SERS, then the Raman spectral features obtained by DFT calculations may be an effective indicator. The core goal of this framework is to address the limitations of single-peak data, such as noise sensitivity and feature ambiguity, and to improve the accuracy and robustness of classification. The joint architecture of CMAF and DMV here optimizes the fusion effect from the perspective of feature interaction and model integration.

[0050] S2.1 Construction of CMAF module CMAF addresses the limitations of traditional feature concatenation methods by dynamically weighting the SERS and DFT Raman modal discrimination spectral regions. SERS and DFT modal feature extraction is performed through two backbone networks, respectively. That is, SERS feature extraction is performed through the network used for SERS modal feature extraction, and DFT modal feature extraction is performed through the network used for DFT Raman modal feature extraction. The two extracted features are then fused into one feature for classification and recognition, as shown in the following example: Figure 2 shown.

[0051] S2.1.1 Feature Alignment

[0052] The normalized SERS intensity matrix And the preprocessed DFT Raman time-frequency characteristics Dimensionality matching is performed via zero-filling or principal component analysis (PCA).

[0053] S2.1.2 Calculating Attention Weights

[0054] The cosine similarity is used to calculate the similarity between cross-modal features. The formula is: Where, represents the SERS spectrum feature vector after preprocessing and feature extraction; Represents the processed DFT Raman spectrum feature vector; Represents the dot product operation of two eigenvectors; yes The L2 (Euclidean) norm of yes The L2 (Euclidean) norm of .

[0055] Then the modality-specific importance score a is generated through softmax normalization ij , the formula is Where a ij represents the modality-specific importance score obtained after softmax normalization; Represents the calculated cross-modal feature similarity Perform exponential operation.

[0056] S2.1.3 Feature Fusion

[0057] The weighted features are spliced ​​together to obtain the fusion feature F fused =a*F SERS +(1-a)*F DFT (4). Where, F fused represents the fused feature vector; a represents the modality-specific importance score obtained by the previous calculation; F SERS represents the SERS spectrum feature vector after preprocessing, feature extraction and alignment; F DFT It represents the DFT Raman spectrum feature vector after corresponding processing and alignment. And it is classified through the fully connected network. The classification formula is: P(y|x)=Softmax(WF fused +b)(5). In the formula, P(y|x) represents the input feature x (here the fusion feature F fused ), the probability distribution of the sample belonging to category y; Softmax represents an activation function; W represents the weight matrix of the fully connected network; F fused represents the fused feature vector obtained by the previous feature fusion step (comprehensive features of SERS and DFT); b represents the bias vector of the fully connected network, also called the deviation vector.

[0058] S2.2 Building the DWV Module

[0059] DWV adaptively adjusts model contributions based on real-time performance to overcome the rigidity of fixed-weight integration methods.

[0060] S2.2.1 Model Training

[0061] A pool of heterogeneous models is trained on single SERS or Raman modal data, such as K-nearest neighbor (KNN), gradient boosting machine (GBM), and ensemble learning (Ensemble) as the base models.

[0062] S2.2.2 Weight calculation

[0063] During the validation process, the F1 score of each model is calculated to reflect its discriminative ability. F1 is an indicator used to evaluate the performance of classification models. According to the formula Dynamically assign weights, give priority to high-performance models, and suppress models with poor performance. k Indicates the weight assigned to the k-th classification model in the integrated model; F1 k represents the F1 score of the k-th classification model; It represents the sum of the F1 scores of all N classification models participating in the ensemble.

[0064] S2.2.3 Result Fusion

[0065] The final category probability is obtained through weighted aggregation, and the formula is: Where, P final (y|x) represents the final category probability; N represents the number of basic models involved in the fusion; w k represents the weight of the kth basic model; P k (y|x) represents the predicted probability of the k-th base model that the sample belongs to category y given the input feature x.

[0066] S2.3 Overall model training

[0067] With SERS and DFT Raman data as dual inputs, the SERS function powers three parallel models and is combined with the DFT function through the CMAF fusion module. The CMAF module generates a discriminative hybrid representation by aligning feature dimensions and calculating cross-modal similarity scores. The prediction results of the basic model are evaluated for performance, and the evaluation results are used to guide the dynamic weight generator to assign voting weights to the model. At the same time, the dynamic weight generator combines the updated features of CMAF to form a dual-channel weighting system, ensuring that the system comprehensively considers model performance and feature quality, and obtains the final result through optimized aggregation, such as Figure 3 shown.

[0068] S3. Model optimization and evaluation

[0069] Six machine learning algorithms are applied to form a basic model pool, namely K-nearest neighbor KNN, gradient boosting machine GBM, Adaboost, support vector machine SVM, multi-layer perceptron MLP and ensemble learning Ensemble, Figure 4 The accuracy of all the base models in the SERS data classification of four OH-PAHs is shown. Since KNN, GBM, and Ensemble have significantly higher accuracy than the other three, they are selected as the base models in DWVF-Net.

[0070] The detection method of the present invention is experimentally verified below.

[0071] 1. Experimental Design

[0072] Dataset selection: The experimental dataset used SERS spectral data of four OH-PAHs derivatives collected in our laboratory, as well as the corresponding Raman spectral data obtained through DFT calculations, to ensure that the data covered the spectral characteristics of different OH-PAHs and that the data quality was reliable, without significant noise or incorrect annotations.

[0073] Data preprocessing: According to the data preprocessing method in S2, feature selection, noise suppression and data partitioning are performed on the experimental data.

[0074] Model selection: The DWVF-Net model is used as the core research model, and KNN, GBM, and Ensemble are selected as comparison models to compare and analyze the advantages of this model in classification accuracy and noise resistance.

[0075] Evaluation indicator setting: Determine the classification accuracy, F1 score, noise robustness, confusion matrix, and ROC curve AUC value as indicators for evaluating model performance.

[0076] 2. Experimental results

[0077] The CMAF framework dynamically calculates the correlation weights between SERS and DFT Raman spectral features to integrate complementary information and suppress noise artifacts. First, the two spectral modes are normalized and dimensionally aligned: the SERS feature focuses on the Raman shift bond peak, such as 1600 cm -1 The benzene ring vibrations are extracted, while the DFT Raman features extract the time domain average and frequency domain wavelet energy. A cross-modal attention matrix is ​​then calculated based on the cosine similarity metric. Softmax activation generates a normalized weight distribution to amplify discriminative features, such as SERS-specific peaks, and suppress redundant or noise artifacts, such as DFT Raman high-frequency fluctuations. Finally, the enhanced features undergo weighted fusion before being fed into dynamic weighting.

[0078] Model training: During training, we monitored model performance metrics, such as loss function value and accuracy, over time. We observed that the DWVF-Net model converged quickly during training, and its performance on the validation set gradually improved and stabilized.

[0079] Model evaluation: The DWV process uses KNN, GBM, and Ensemble as basic models to implement weighted fusion of prediction probabilities. The evaluation results on the test set show that the classification accuracy of the DWVF-Net model reached 97.14%, significantly higher than KNN's 91%, GBM's 86%, and Ensemble's 90%. Under 20dB Gaussian noise conditions, the accuracy of the DWVF-Net model only dropped by 1.8%, significantly better than the equal-weight strategy, which dropped by 3.2%. The confusion matrix shows that the DWVF-Net model has significantly reduced the classification error rate for various OH-PAHs, such as Figure 5 As shown in , the misclassification rate of class B is reduced from 7% to 3%. Figure 6 As shown in the figure, the AUC value of the ROC curve is always stable above 0.98.

[0080] Analysis of experimental results: The experimental results fully demonstrate that the DWVF-Net model effectively integrates the complementary information of SERS and DFT Raman spectra through the synergistic effect of CMAF and DWV. These results verify the adaptive advantages of the dynamic weighting mechanism in processing complex data environments and prove that it can provide a lightweight but effective multimodal integration solution. The final classification accuracy reached 97.14%. It can be seen from the results that compared with the basic model, DWVF-Net has a significant improvement in accuracy and robustness.

[0081] The system architecture and detection method of the present invention are described below with reference to specific examples.

[0082] 1. System architecture design:

[0083] Construct an OH-PAHs detection system based on DWVF-Net, such as Figure 7 As shown in the figure, the system architecture mainly includes data acquisition module, data preprocessing module, model construction and training module, and model evaluation and application module.

[0084] Data Acquisition Module: Acquire spectral data of samples containing OH-PAHs from laboratory experiments, environmental monitoring sites, and other sources. For SERS spectral data, refer to the laboratory's established sample processing and spectral acquisition procedures for collection. For DFT Raman spectral data, calculate and acquire it using appropriate theoretical calculation methods and software tools. During the data collection process, ensure data accuracy and record key information such as the data source and sample information.

[0085] Data preprocessing module: performs feature selection, noise suppression and data segmentation on the collected spectral data to provide high-quality data for subsequent model training.

[0086] Model building and training module: In this module, the DWVF-Net model is constructed, including the construction of the CMAF and DWV modules. The model is trained using preprocessed data, and the model hyperparameters are adjusted to optimize model performance.

[0087] Model Evaluation and Application Module: The trained model is evaluated using a test set, with metrics including classification accuracy, F1 score, and noise immunity. Once the evaluation passes, the model is applied to the detection of OH-PAHs in real-world environmental samples.

[0088] 2. Implementation steps

[0089] S1. Data Collection and Preprocessing S1.1 Data Collection and Transmission

[0090] In actual detection scenarios, spectral data of samples containing OH-PAHs is obtained from the corresponding sample source. The collected SERS spectral data and calculated DFT Raman spectral data are transmitted to the data preprocessing module according to established data transmission specifications. At the same time, during the data transmission process, it is necessary to ensure the integrity and accuracy of the data and verify it using appropriate data verification methods.

[0091] S1.2 Data Preprocessing

[0092] In the data preprocessing module, the collected spectral data is processed according to a predefined preprocessing process, including feature selection, noise suppression, and data partitioning. Once processed, the preprocessed data is transferred to the model building and training module. During the data collection process, the parameters of the acquisition equipment and environmental conditions must be strictly controlled to ensure the accuracy and reliability of the collected spectral data. During the data preprocessing process, the results must be quality-checked at each step to avoid data loss or the introduction of new errors.

[0093] Model construction and training: In the model construction and training module, a DWVF-Net model is constructed based on data characteristics and detection requirements. The model is trained, and during the training process, model performance indicators such as loss function value and accuracy are monitored in real time. Model hyperparameters are adjusted based on the monitoring results until the model performance reaches the optimal level. As new OH-PAHs sample data accumulates or the detection environment changes, the DWVF-Net model needs to be regularly updated and optimized. Operators using this detection system need to be trained to familiarize themselves with the system's operating procedures, data processing methods, and model evaluation indicators, so that they can correctly interpret the test results and ensure the effective operation of the system.

[0094] Model Evaluation and Application: The trained model is evaluated using a test set. If the evaluation results meet the preset performance indicators, the model is applied to actual detection tasks. In actual applications, new sample data is preprocessed and then input into the model. The model outputs the OH-PAHs classification results, enabling the detection of OH-PAHs in environmental samples.

[0095] Traditional OH-PAHs detection methods mainly rely on chromatographic techniques, such as high-performance liquid chromatography (HPLC) and gas chromatography-mass spectrometry (GC-MS). Although these methods can detect OH-PAHs to a certain extent, they have the following drawbacks:

[0096] Sample pretreatment is complex: it requires multiple tedious steps of extraction, concentration, and purification, which is not only time-consuming and labor-intensive, but also prone to errors.

[0097] Low detection throughput: Only a small number of samples can be detected at a time, which is difficult to meet the needs of large-scale environmental monitoring.

[0098] Limited sensitivity: Detection of trace levels of OH-PAHs is poor, and missed detections are common. Expensive equipment and complex operation: Requires specialized equipment and operators, resulting in high maintenance costs.

[0099] Compared with traditional methods, the present invention has the following significant advantages:

[0100] Rapid detection: No complicated sample pretreatment is required. Spectral technology is directly used to acquire data, and combined with rapid model analysis, the test results can be obtained in a short time.

[0101] High accuracy: By fusing multiple information through the DWVF-Net model, the classification accuracy is as high as 97.14%, significantly higher than traditional methods.

[0102] Strong noise resistance: Under 20dB Gaussian noise conditions, the accuracy rate only drops by 1.8%, and it can accurately detect OH-PAHs in complex environmental backgrounds.

[0103] No need for chromatographic separation: avoids the limitations of chromatographic technology, simplifies the detection process, and reduces detection costs.

[0104] Potential portability: It is expected to be developed into a portable detection device to achieve on-site real-time detection and be suitable for different environmental scenarios.

[0105] Experimental comparisons of the present invention and traditional methods revealed that, for the same sample testing, the traditional chromatography method had an accuracy rate of approximately 80%, and its detection performance was unstable for trace samples. However, the DWVF-Net model of the present invention demonstrated high accuracy and strong noise immunity across a wide range of sample types, accurately identifying OH-PAHs even in complex sample backgrounds. These experimental results fully validate the advantages of this patented technology.

[0106] The present invention has broad application prospects in many fields:

[0107] (1) Environmental monitoring

[0108] It can be used to monitor OH-PAHs pollution in the atmosphere, water, and soil in real time. By deploying detection equipment at various environmental monitoring points, timely pollutant information can be obtained, providing data support for environmental quality assessment and pollution control. For example, real-time monitoring of OH-PAHs in the air in key areas such as industrial pollution zones and urban residential areas can promptly identify potential health risks.

[0109] (2) Food Safety Testing: OH-PAHs may be present during food processing and storage, posing a threat to food safety. This invention is used to detect OH-PAH residues in food to ensure food safety. For example, smoked foods, fried foods, and other foods containing high concentrations of OH-PAHs can be tested to ensure consumer health.

[0110] (3) Biomedical Research: The metabolism and accumulation of OH-PAHs in organisms are associated with the development and progression of various diseases. This invention is used to detect OH-PAHs in biological samples, such as urine and blood, providing a technical means for biomedical research and contributing to a deeper understanding of the pathogenesis and preventive measures of diseases.

[0111] This paper demonstrates the superior performance of a hybrid machine learning model in classifying hydroxylated PAHs using Raman spectroscopy, achieving high accuracy and robustness in complex spectral analysis. The proposed method has important practical value, especially in the deployment of portable Raman devices for on-site monitoring, which may revolutionize real-time environmental monitoring. In addition, the system has the potential to be expanded to a multi-pollutant monitoring platform, capable of simultaneously identifying a variety of pollutants in heterogeneous matrices. Future work will focus on extending this method to nitro PAHs and other derivatives to broaden its applicability, while exploring multimodal fusion strategies that integrate terahertz spectroscopy to enhance molecular specificity. This is expected to provide a comprehensive analytical tool for addressing emerging environmental health challenges, bridging the gap between laboratory accuracy and field practicality.

Claims

1. A Raman spectroscopy detection method for hydroxylated polycyclic aromatic hydrocarbons based on a dynamic weighted voting fusion network, characterized in that: The following steps are involved: S1. Data collection and preprocessing S1.1 Spectral data acquisition The Raman spectra data corresponding to the four derivatives of OH-PAHs, namely 3-OH-BAA, 3-OH-BAP, NAP and 9-OH-PHE, were obtained by density functional theory DFT calculations; S1.2 Data Preprocessing Perform feature selection, noise suppression and data partitioning on spectral data; S2. DWVF-Net model construction and training S2.1 Building the CMAF Module S2.1.1 Feature Alignment The normalized SERS intensity matrix And the preprocessed DFT Raman time-frequency characteristics Dimension matching via zero-filling or principal component analysis (PCA); S2.1.2 Calculating Attention Weights Use cosine similarity to calculate the similarity between cross-modal features; S2.1.3 Feature Fusion The weighted features are spliced ​​together to obtain the fusion feature F fused =a*F SERS +(1-a)*F DFT (1); where F fused represents the fused feature vector; a represents the modality-specific importance score obtained by the previous calculation; F SERS represents the SERS spectrum feature vector after preprocessing, feature extraction and alignment; F DFT represents the DFT Raman spectrum feature vector after corresponding processing and alignment; S2.2 Building the DWV Module S2.2.1 Model Training Training a pool of heterogeneous models on a single SERS or Raman modality data; S2.2.2 Weight calculation Calculate the F1 score of each model according to the formula Dynamically assign weights, give priority to high-performance models, and suppress models with poor performance; where w k Indicates the weight assigned to the k-th classification model in the integrated model; F1 k represents the F1 score of the k-th classification model; Indicates the sum of the F1 scores of all N classification models participating in the integration; S2.2.3 Result Fusion The final category probability is obtained through weighted aggregation, and the formula is: Where, P final (y|x) represents the final category probability; N represents the number of basic models involved in the fusion; w k represents the weight of the kth basic model; P k (y|x) represents the predicted probability of the sample belonging to category y given the input feature x of the k-th base model; S2.3 Overall model training Taking SERS and DFT Raman data as dual inputs, the SERS function powers three parallel models and is combined with the DFT function through the CMAF fusion module; the CMAF module generates a discriminative hybrid representation by aligning feature dimensions and calculating cross-modal similarity scores; S3. Model optimization and evaluation Six machine learning algorithms, including K-nearest neighbor (KNN), gradient boosting machine (GBM), Adaboost, support vector machine (SVM), multi-layer perceptron (MLP) and ensemble learning (Ensemble), are used to form a basic model pool.

2. The Raman spectrum detection method for hydroxylated polycyclic aromatic hydrocarbons based on a dynamic weighted voting fusion network according to claim 1, characterized in that: The feature selection includes: independently analyzing Raman shift and intensity for SERS data; and extracting time domain statistical features and frequency domain wavelet coefficients by performing 5-level decomposition using Daubechies-4 wavelet for DFT Raman data. The time domain statistical features include mean and variance.

3. The Raman spectrum detection method for hydroxylated polycyclic aromatic hydrocarbons based on a dynamic weighted voting fusion network according to claim 2, characterized in that: The noise suppression refers to: using Savitzky-Golay filter to smooth the signal to suppress high-frequency noise, and then using the averaging method to perform baseline correction. The intensity calculation formula after correction is: Where, I corrected Indicates the corrected intensity; I smooth represents the spectral intensity signal after smoothing by the Savitzky-Golay filter; N represents the number of data points involved in the average calculation when calculating the baseline correction; i represents the index of the data point currently being processed; w represents the window size of the Savitzky-Golay filter.

4. The Raman spectrum detection method for hydroxylated polycyclic aromatic hydrocarbons based on a dynamic weighted voting fusion network according to claim 3, characterized in that: The data division is to divide the spectral data after noise suppression into a training set and a test set.

5. The Raman spectrum detection method for hydroxylated polycyclic aromatic hydrocarbons based on a dynamic weighted voting fusion network according to claim 4, characterized in that: The formula for calculating the similarity between cross-modal features is: Where, represents the SERS spectrum feature vector after preprocessing and feature extraction; Represents the processed DFT Raman spectrum feature vector; Represents the dot product operation of two eigenvectors; yes The L2 norm of yes The L2 norm of .

6. The Raman spectrum detection method for hydroxylated polycyclic aromatic hydrocarbons based on a dynamic weighted voting fusion network according to claim 5, characterized in that: Then the modality-specific importance score a is generated through softmax normalization ij , the formula is Where a ij represents the modality-specific importance score obtained after softmax normalization; Represents the calculated cross-modal feature similarity Perform exponential operation.

7. The Raman spectrum detection method for hydroxylated polycyclic aromatic hydrocarbons based on a dynamic weighted voting fusion network according to claim 6, characterized in that: After obtaining the fusion features, classification is performed through the fully connected network. The classification formula is: P(y|x)=Softmax(WF fused +b)(7); where P(y|x) represents the value of the given input feature x=F fused In the case of , the probability distribution of the sample belonging to category y; Softmax represents an activation function; W represents the weight matrix of the fully connected network; F fused represents the fused feature vector obtained by the comprehensive feature fusion step of SERS and DFT; b represents the bias vector of the fully connected network.

8. The Raman spectrum detection method for hydroxylated polycyclic aromatic hydrocarbons based on a dynamic weighted voting fusion network according to claim 7, characterized in that: In step S2.2.1, K-nearest neighbor (KNN), gradient boosting machine (GBM), and ensemble learning (Ensemble) are selected as the base models.

Images