Asphalt oil source classification method based on conditional generative adversarial network and soft voting integration

Abstract

The application discloses a kind of asphalt oil source classification methods based on conditional generative adversarial network and soft voting integration. Including the infrared spectrum data of different oil source asphalt samples is collected, the peak area of key characteristic peak is extracted as a characteristic vector to establish a characteristic fingerprint database;The original sample set is input into the conditional generative adversarial network, and the training strategy is automatically adjusted according to the sample amount difference of each category, to generate synthetic samples and mix output balanced class expansion feature data set;According to the standardization processing of a variety of wave band range reference calculation characteristic functional group index;The core discriminant feature is extracted by nonlinear dimension reduction to the standardized multi-reference feature set;The core discriminant feature is input into the integrated learning classification model optimized by hyperparameter, and the final oil source attribution category is output by the soft voting mechanism combined with the screened multiple base learners and meta-learners.This application establishes a logical closed loop with physical and chemical explainability between micro-chemical fingerprint and macro-tracing, and has strong generalization ability.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of asphalt material testing and artificial intelligence. Specifically, it relates to an asphalt oil source classification method based on conditional generative adversarial networks and soft voting integration, which can be used for oil source traceability and quality monitoring of asphalt materials in transportation infrastructure construction. Background Technology

[0002] The construction and long-term service of high-quality transportation infrastructure rely heavily on a rigorous material quality control and traceability system. Under complex multi-factor loads, asphalt materials of unknown origin or containing inferior components are highly susceptible to early spalling, fatigue cracking, and severe deterioration of the pavement. Traditional asphalt quality evaluation systems are mainly based on macroscopic physical and mechanical indicators such as penetration, softening point, and dynamic shear rheology. However, these macroscopic indicators are easily affected by the testing environment and human intervention, and cannot fundamentally reflect the structural evolution of materials at the molecular level. Therefore, modern road construction materials science is rapidly undergoing a paradigm shift from "macroscopic performance characterization" to "microscopic chemical fingerprint traceability."

[0003] Among numerous microscopic characterization methods, attenuated total reflectance Fourier transform infrared spectroscopy (ATR-FTIR) has become a core method for non-destructive determination of the chemical properties of asphalt due to its high sensitivity to the vibrations of molecular functional groups. Specific infrared absorption bands, such as 1700 cm⁻¹, are also relevant. -1 carbonyl group (C=O) at 1031 cm -1 The sulfoxide group (S=O) at the location and 1600 cm -1 The aromatic ring vibrations at the location provide a highly reliable microscopic chemical fingerprint for asphalt. Faced with exponentially increasing high-dimensional and complex spectral data, machine learning and deep learning techniques have demonstrated significantly superior feature extraction and target classification efficiency compared to traditional empirical methods.

[0004] Despite the immense potential of spectral intelligent identification technology in materials characterization, its practical application in the deep tracing of asphalt materials remains constrained by three major theoretical and technological bottlenecks that urgently need to be overcome. First, acquiring high-quality asphalt spectral data is costly, resulting in datasets that generally exhibit strong class imbalance and long-tailed distribution characteristics. Traditional classifiers inevitably shift towards the majority class when processing such data, significantly weakening the model's generalization ability. Second, asphalt spectral data itself has extremely high feature dimensionality and multicollinearity. Traditional linear dimensionality reduction methods, such as principal component analysis (PCA), struggle to effectively separate complex nonlinear chemical structures and are prone to losing important trace chemical features during feature mapping. Third, existing research lacks a unified standard for quantitative integration of functional groups, leading to inconsistencies in the input features of different algorithms. Finally, current advanced ensemble learning architectures generally exhibit strong "black box" characteristics. This lack of algorithmic transparency severs the scientific connection between mathematical weights and the physicochemical mechanisms of asphalt, severely limiting the reliability and credibility of data-driven models in practical engineering compliance assessments.

[0005] To overcome the aforementioned technical barriers, there is an urgent need for a method for classifying asphalt oil sources that can adaptively eliminate data imbalance, accurately decouple high-dimensional nonlinear features, and has a clear physicochemical mechanism explanation, so as to establish a rigorous logical closed loop from microscopic chemical fingerprints to macroscopic oil source tracing. Summary of the Invention

[0006] The purpose of this invention is to provide an asphalt oil source classification method based on conditional generative adversarial networks and soft voting integration, so as to solve the technical problems of unbalanced asphalt spectral data categories, difficulty in effectively decoupling high-dimensional nonlinear features, and lack of physicochemical interpretability of integrated models in the prior art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for classifying asphalt oil sources based on a conditional generative adversarial network and soft voting integration, comprising: Infrared spectral data of asphalt samples from different oil sources were collected, and the peak areas of key characteristic peaks were extracted. The key characteristic peak areas of each asphalt sample were used as feature vectors to establish an asphalt infrared spectral feature fingerprint database. The feature vectors of each oil source category in the database constitute the original sample set. The original sample set is input into the conditional generative adversarial network for training. The conditional generative adversarial network automatically adjusts its training strategy according to the difference in the sample size of each category, generates synthetic samples for categories with insufficient sample size, and mixes the original sample set with the synthetic samples to output a category-balanced augmented feature dataset. Based on the peak area data in the amplified feature dataset, feature intervals are divided according to at least two band range benchmarks, feature functional group indices covering at least two structural feature types are calculated, and standardization processing is performed to obtain a standardized multi-benchmark feature set. The standardized multi-benchmark feature set is subjected to nonlinear dimensionality reduction processing to extract core discriminative features; The core discriminative features are input into an ensemble learning classification model optimized by hyperparameters. The ensemble learning classification model includes multiple base learners that process features in parallel after screening, and a meta-learner that fuses the outputs of each base learner. The final oil source classification is output through a soft voting mechanism.

[0008] Furthermore, the conditional generative adversarial network consists of an adversarial architecture with a generator and a discriminator, wherein: The generator receives a standard normal noise vector of a preset dimension and a class label as input, and generates the synthetic sample with the same dimension as the feature vector by using a multi-layer fully connected network in conjunction with an activation function and batch normalization. The discriminator adopts a network structure that is mirror-symmetric to the generator, with hidden layers combined with Dropout regularization, and outputs the true and false probabilities through the Sigmoid function. During training, the discriminator is first updated based on the samples in the original sample set and the synthetic samples generated by the generator. The generator is then updated with the goal of deceiving the discriminator.

[0009] Furthermore, the automatic training adjustment strategy based on the differences in sample size across different categories includes: For categories whose sample size in the original sample set is lower than a preset threshold, the training rounds of the generator are automatically increased, the training batch size is reduced, and the tolerance threshold of the early stopping mechanism is increased to force the generator to deeply fit the feature manifold distribution of the samples in that category.

[0010] Furthermore, the characteristic functional group index includes the aliphatic index I. B Aromatic Index I Ar Carbonyl index I C=O , sulfoxide index I S=O Asymmetric aliphatic index I B,a At least two of them, and the calculation methods for each index are as follows: I B =(A 2924 +A 2852 ) / ΣA; I Ar =A 1600 / ΣA; I C=O =A 1700 / ΣA; I S=O =A 1031 / ΣA; I B,a =(A 1377 +A 1461 ) / ΣA; Among them, A 2924 The wave number is 2924 cm⁻¹ -1 The peak area of ​​the antisymmetric stretching vibration of aliphatic CH, A 2852 The wave number is 2852 cm⁻¹ -1 The peak area of ​​the symmetric stretching vibration of aliphatic CH, A 1700 Wave number 1700cm -1 The peak area of ​​the carbonyl C=O stretching vibration, A 1600 Wavenumber 1600cm -1 The peak area of ​​the C=C skeleton vibration of the aromatic ring, A 1031 The wave number is 1031 cm⁻¹ -1 The peak area of ​​the S=O stretching vibration of the sulfoxide group, A 1461 The wave number is 1461 cm⁻¹ -1 The peak area of ​​the asymmetric bending vibration of aliphatic CH, A 1377 The wave number is 1377 cm⁻¹ -1 The peak area of ​​the symmetrical bending vibration of aliphatic CH; ΣA is the sum of the peak areas of each key characteristic peak within the corresponding band range reference; I B Characterizing the relative content of aliphatic compounds in asphalt, I Ar Characterizing the relative content of aromatic ring structures in bitumen, I C=O Characterizing the relative content of carbonyl groups in asphalt, I S=O Characterizing the relative content of sulfoxide groups in bitumen, I B,a Characterizes the relative content of asymmetric aliphatic functional groups in asphalt.

[0011] Furthermore, the band range reference includes at least two of the following: Benchmark I: 4000~700cm -1 The sum of the peak areas across the entire spectrum within the range is taken as ΣA; Standard II: Based on 4000~1400cm -1 The sum of the peak areas of the stretching vibrations within the range is taken as ΣA; Standard III: Based on 1400~700cm -1 The sum of the areas of the bending vibration peaks within the range is taken as ΣA; Standard IV: 3000~2800cm -1 The sum of the peak areas of CH stretching vibration within the range is taken as ΣA.

[0012] Furthermore, the nonlinear dimensionality reduction process employs a t-distributed random neighborhood embedding algorithm, specifically including: Conditional probability is used to measure the topological similarity distribution among samples in the standardized multi-benchmark feature set in a high-dimensional space. The similarity probability distribution between samples is reconstructed using a t-distribution with 1 degree of freedom in a low-dimensional mapping space; By minimizing the KL divergence between the probability distributions in high-dimensional and low-dimensional spaces, gradient descent is applied iteratively to optimize and output the core discriminative features.

[0013] Furthermore, the construction process of the hyperparameter-optimized ensemble learning classification model includes: A model library containing multiple candidate algorithms is constructed. The hyperparameter Bayesian optimization of the candidate algorithms is performed using a probabilistic surrogate model through a tree-structured Parzen estimator to minimize the cross-validation error. A predetermined number of heterogeneous models are selected from the optimized candidate algorithms as the multiple base learners. The multiple base learners process the core discriminative features in parallel and output the posterior probability distribution of each category as meta-features. The meta-learner fits the probability mapping relationship of the meta-features, and the local prediction bias of each base learner is dynamically calibrated in conjunction with the soft voting mechanism to output the oil source attribution probability of each category; the category with the highest attribution probability is determined as the final oil source attribution category.

[0014] Furthermore, the candidate algorithms include at least three of the following: random forest, support vector machine, extreme gradient boosting, K-nearest neighbors, artificial neural network, and Naive Bayes; the meta-learner is a logistic regression model with L2 regularization; the preset number is three, and the selection criterion is the accuracy of hierarchical cross-validation.

[0015] Furthermore, following the output of the final oil source attribution category, it also includes: The SHAP attribution algorithm is introduced to globally deconstruct the ensemble learning classification model. By calculating the marginal contribution of each feature functional group index to the model prediction results, a feature importance distribution is established, and the mathematical weights of the ensemble learning classification model are mapped to the differences in the microscopic chemical structure of asphalt.

[0016] Furthermore, the characteristic absorption band corresponding to the key characteristic peak includes a wavenumber of 2924 cm⁻¹. -1 2852cm -1 1700cm -1 1600cm -1 1461cm -1 1377cm -1 1031cm -1 868cm-1 812cm -1 747cm -1 and 722cm -1 .

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a closed-loop identification system from microscopic chemical fingerprinting to macroscopic oil source tracing, effectively addressing the inherent complexity challenges of real material data. To address the high cost of spectral data acquisition and class imbalance caused by long-tailed distributions, this invention introduces an adaptive conditional generative adversarial network (cGAN), achieving unbiased balance of the data feature layer through deep fitting of the real feature manifold. In the feature processing stage, a multi-benchmark functional group index quantitative system is established, unifying the feature calculation standards across different spectral bands and giving input features clear physicochemical meanings. Simultaneously, it abandons the traditional linear dimensionality reduction that easily obscures the classification boundaries of microscopic chemical features, instead employing the t-SNE algorithm for nonlinear mapping of the high-dimensional feature space, preserving the topological independence of core discriminative features. Based on this, the TPE algorithm is used to perform global Bayesian optimization of the heterogeneous basis learner in a non-convex high-dimensional space, and combined with soft voting and a meta-learner architecture, improving the classification robustness and generalization ability under complex spectral features. This invention uses the SHAP attribution algorithm to globally deconstruct the decision logic of the ensemble model, confirming the core driving role of key functional groups such as carbonyl, sulfoxide, and aromatic groups. It transforms the "black box" model of deep learning into a reliable decision system with clear materials science significance and physicochemical mechanisms, providing a new technical path for the digital assessment and intelligent quality supervision of transportation infrastructure, and has practical engineering application value. Attached Figure Description

[0018] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the present invention will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without any creative effort. In the drawings: Figure 1 This is a flowchart of the asphalt oil source classification method based on the integration of conditional generative adversarial networks and soft voting, as described in an embodiment of the present invention. Figure 2 A comparison and verification chart showing the consistency of the correlation coefficient (Pearson's r) between the features of the synthesized sample and the original sample set after data augmentation using the traditional SMOTE algorithm and the cGAN algorithm of this invention. Figure 3To visualize the spectral features of asphalt using different dimensionality reduction algorithms, (A) is the t-SNE projection of the original spectral data, (B) is the t-SNE projection of the normalized spectral data, (C) is the PCA projection of the original spectral data, and (D) is the PCA projection of the normalized spectral data. Figure 4 Heatmaps showing the test accuracy of each classification model under benchmarks I-IV; Figure 5 The overall performance diagram of the optimal ensemble model under different infrared spectral reference feature sets; Figure 6 A bee colony diagram showing the importance of infrared spectral features based on SHAP values; Figure 7 Thermogram of significant infrared spectral characteristic absorption peaks for the test set samples. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. The described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0020] Reference Figure 1 This embodiment provides a method for classifying asphalt oil sources based on a conditional generative adversarial network and soft voting integration, including the following steps: Step 1: Collect infrared spectral data of asphalt samples from different oil sources, extract the peak areas of key characteristic peaks, and use the key characteristic peak areas of each asphalt sample as feature vectors to establish a discrete asphalt infrared spectral fingerprint database. The feature vectors of each oil source category constitute the original sample set. Specifically, nine different typical base asphalts and SBS (styrene-butadiene-styrene block copolymer) modified asphalt samples were collected, covering 70#, 90#, and different oil source categories, numbered from 0 to 8. The original molecular structure information of the samples was obtained using an infrared spectrometer, and baseline correction and spectral smoothing and noise reduction were performed. The infrared spectra of different asphalt types were compared. Eleven typical characteristic absorption bands were selected, corresponding to wavenumbers of 2924 cm⁻¹. -1 2852cm -1 1700cm -1 1600cm -1 1461cm -1 1377cm -1 1031cm -1 868cm-1 812cm -1 747cm -1 and 722cm -1 By systematically testing 111 different asphalt samples, the characteristic peak areas were extracted and calculated, resulting in 1221 original peak area data points. A discretized asphalt infrared spectral characteristic fingerprint database was established, as shown in Table 1. Samples from each oil source category constitute the original sample set. The sample size distribution of each category in the original sample set is uneven, exhibiting a long-tail distribution characteristic. The category with the most samples (number 0) has 56 samples, while the category with the fewest samples (numbers 3, 4, 5, and 6) has only 5 samples, and number 7 has 4 samples. Taking samples numbered 0-1 and 1-1 as examples, the characteristic peak areas are: A 2924 The values ​​are 1161.715 and 1484.28 respectively, A 2852 The values ​​are 408.23 and 473.166 respectively, A 1700 The values ​​are 24.419 and 119.424 respectively, A 1600 The values ​​are 154.09 and 353.512 respectively, A 1461 The values ​​are 733.504 and 836.109 respectively, A 1377 The values ​​are 141.145 and 242.601 respectively, A 1031 The values ​​are 12.215 and 42.831 respectively, A 868 The values ​​are 86.874 and 69.66 respectively, A 812 The values ​​are 49.884 and 48.407 respectively, A 747 The values ​​are 26.93 and 23.547 respectively, A 722 The values ​​are 256.535 and 223.212 respectively.

[0021] Table 1. Asphalt Infrared Spectrum Database

[0022] In the table, A represents the area of ​​the characteristic peak, and 0-1, 1-1, 2-1, 7-1, 8-1, etc. represent the asphalt number.

[0023] Step 2: The original sample set is input into a Conditional Generative Adversarial Network (cGAN) for training. The cGAN automatically adjusts its training strategy based on the differences in sample size across categories, generating synthetic samples for categories with insufficient sample size. The original sample set and the synthetic samples are then mixed to output a category-balanced augmented feature dataset. Specifically, to address the long-tailed distribution and significant class imbalance in the original sample set, a cGAN model based on a fully connected neural network architecture is developed. The cGAN model is a fully connected layer-based conditional GAN ​​model, trained independently for each category to reduce the model's dependence on specific samples. The cGAN model incorporates regularization and early stopping mechanisms. The generator takes a 100-dimensional standard normal noise vector sampled from the latent space and its corresponding class label, concatenates them after embedding layer processing, and uses this as input. This input is then processed by three fully connected layers (containing 128, 256, and 512 nodes respectively), combined with Leaky ReLU activation function (negative slope α=0.2) and batch normalization (momentum parameter momentum=0.8, epsilon=1e-5), mapping it to an 11-dimensional spectral feature vector as the synthetic sample. The output layer uses the Tanh activation function to match the normalized feature range. The discriminator uses a mirror-symmetric structure (512-256-128 nodes). The hidden layers also use Leaky ReLU (α=0.2) activation with a Dropout rate of 0.3 to prevent overfitting. It receives original samples from the original sample set and synthetic samples output by the generator, and uses the Sigmoid function to distinguish between true and false samples. Both the generator and discriminator use the Adam optimizer (learning rate lr=2×10⁻⁶). -4The parameters are updated using the first-order moment estimation exponential decay rate β1=0.5 and the second-order moment estimation exponential decay rate β2=0.999. In each epoch, the discriminator is updated first, optimized by the binary cross-entropy (BCE) loss between the samples in the original sample set and the synthetic samples; the generator is then updated, aiming to minimize the BCE loss when the synthetic samples are judged as false, forcing the generator to deeply fit the true feature manifold distribution. For categories with fewer samples in the original sample set, the system automatically starts an adaptive learning strategy based on the difference in sample size among categories. In this embodiment, the preset threshold is set to 10, that is, the adaptive strategy is triggered when the sample size of a certain category is less than 10; number 0 (56 samples), number 2 (15 samples), and number 1 (11 samples) are trained according to the conventional strategy, while numbers 3 to 8 (all with ≤5 samples) all trigger the adaptive strategy. Specifically, for the categories triggering the adaptive strategy: the number of training rounds was increased from the default 500 rounds to a maximum of 2000 rounds, the batch size was reduced from the default 64 to 16, and the patience threshold of the early stopping mechanism was increased from the default 50 rounds to 200 rounds. The early stopping mechanism uses the generator loss not improving within consecutive patience rounds as the trigger condition for stopping training. All synthetic samples were de-standardized and then labeled with class labels for storage. After amplification, the number of samples in each category was uniformly increased from the original 4-56 to 72. The amplification results are shown in Table 2. The original sample set and the generated high-fidelity synthetic samples were mixed to output a class-balanced amplified feature dataset. Before amplification, the number of samples in each oil source category were: 56 for No. 0, 11 for No. 1, 15 for No. 2, 5 for No. 3, 5 for No. 4, 5 for No. 5, 5 for No. 6, 4 for No. 7, and 5 for No. 8, for a total of 111. After amplification, the number of samples in each category was uniformly increased to 72, for a total of 648. To verify the reliability of the synthesized samples, a correlation consistency check was performed between the synthesized samples and the samples in the original sample set. Figure 2 The results showed that the Pearson correlation coefficient between the cGAN-synthesized samples and the original sample set was as high as 0.9348 (p<0.001), and the linear fitting curve closely followed the ideal contour lines, effectively capturing the microstructure of asphalt materials at the feature level.

[0024] Table 2 Comparison of GAN samples before and after amplification

[0025] Step 3: Based on the peak area data in the amplified feature dataset, feature intervals are divided according to at least two band range benchmarks. Feature functional group indices covering at least two structural feature types are calculated, and standardization is performed to obtain a standardized multi-benchmark feature set. Specifically, feature functional group indices are calculated based on multiple band range benchmarks, including the aliphatic index I. B Aromatic Index IAr Carbonyl index I C=O , sulfoxide index I S=O Asymmetric aliphatic index I B,a The calculation methods for each index are as follows: I B =(A 2924 +A 2852 ) / ΣA;I Ar =A 1600 / ΣA;I C=O =A 1700 / ΣA;I S=O =A 1031 / ΣA;I B,a =(A 1377 +A 1461 ) / ΣA; where A 2924 The wave number is 2924 cm⁻¹ -1 The peak area of ​​the antisymmetric stretching vibration of aliphatic CH, A 2852 The wave number is 2852 cm⁻¹ -1 The peak area of ​​the symmetric stretching vibration of aliphatic CH, A 1700 Wave number 1700cm -1 The peak area of ​​the carbonyl C=O stretching vibration, A 1600 Wavenumber 1600cm -1 The peak area of ​​the C=C skeleton vibration of the aromatic ring, A 1031 The wave number is 1031 cm⁻¹ -1 The peak area of ​​the S=O stretching vibration of the sulfoxide group, A 1461 The wave number is 1461 cm⁻¹ -1 The peak area of ​​the asymmetric bending vibration of aliphatic CH, A 1377 The wave number is 1377 cm⁻¹ -1 The peak area of ​​the symmetrical bending vibration of aliphatic CH; ΣA is the sum of the peak areas of each key characteristic peak within the corresponding band range reference, using 4000~700cm. -1 The peak area ratio of characteristic functional groups is compared based on the peak area of ​​the range spectrum and the peak area of ​​the full spectrum, as defined below: ; Use 4000~1400cm -1 The peak areas of the range spectrum are used as a benchmark to compare the peak area ratios of characteristic functional groups, which are mainly absorptions caused by stretching vibrations, and are defined as follows: ; Use 1400~700cm -1 The peak areas of the range spectrum are used as a benchmark to compare the peak area ratios of characteristic functional groups, mainly for absorption caused by bending vibrations, and are defined as follows: ; using 3000~2800cm -1The peak areas at each point are compared with the peak area ratios of characteristic functional groups, using the sum of the peak areas at each point as a benchmark. The infrared spectrum of asphalt mainly exhibits the stretching vibrations of CH, defined as follows: ;I B Characterizing the relative content of aliphatic compounds in asphalt, I Ar Characterizing the relative content of aromatic ring structures in bitumen, I C=O Characterizing the relative content of carbonyl groups in asphalt, I S=O Characterizing the relative content of sulfoxide groups in bitumen, I B,a Characterizing the relative content of asymmetric aliphatic functional groups in asphalt. The evaluation criteria are specifically divided into: Criterion I: based on the full spectral range (4000~700cm). -1 The sum of the peak areas of the two peaks is ΣA; the reference II is based on the stretching vibration range (4000~1400cm). -1 The sum of the peak areas is ΣA; Reference III uses the bending vibration range (1400~700cm) -1 The sum of the peak areas is ΣA; the reference IV is based on the CH stretching vibration range (3000~2800cm). -1 The sum of the peak areas is ΣA. The five functional group indices were calculated under each benchmark, resulting in 20 feature dimensions, as shown in Table 3. Subsequently, a min-max scaler strategy was applied to all 20 features to linearly map each feature value to the [0,1] interval, eliminating dimensional differences between different functional group indices and obtaining a standardized multi-benchmark feature set.

[0026] Table 3 Characteristic Functional Group Index Table

[0027] in , , , , These represent the functional group indices of aromatic, carbonyl, sulfoxide, aliphatic, and asymmetric aliphatic groups under reference 1-4, respectively; 0-1, 1-1, 2-1, 7-1, 8-1, etc. represent asphalt numbers.

[0028] Step 4: Perform nonlinear dimensionality reduction on the standardized multi-benchmark feature set to extract core discriminative features. Specifically, since the infrared spectral features after mixing real and synthetic samples have a high input dimension, and traditional linear dimensionality reduction easily masks microscopic chemical features, traditional asphalt spectral classification studies generally use linear dimensionality reduction tools such as PCA. However, this essentially assumes that the variation in data can be fully expressed through linear combinations. However, asphalt is an extremely complex colloidal system, there are complex nonlinear interactions and multicollinearity problems among its various internal features. The t-SNE introduced in this step is not a simple application of general dimensionality reduction algorithms, but a specific adaptation precisely for the nonlinear manifold characteristics of asphalt spectral data. This effectively avoids the loss of trace chemical features during the linear mapping process, thus providing high-quality nonlinear decoupling features for subsequent extraction of functional groups and class boundary division. The t-distributed random neighborhood embedding (t-SNE) algorithm is applied to perform nonlinear dimensionality reduction on the standardized multi-benchmark feature set. In this embodiment, the key hyperparameters of t-SNE are set as follows: perplexity = 30, output dimension = 2, maximum number of iterations = 1000, learning rate = 200, initialization method is PCA initialization, Euclidean distance is used as the distance metric, and the random seed is set to 42 to ensure the reproducibility of the results. The specific algorithm process is as follows: conditional probability is used to measure the topological similarity distribution between samples in the standardized multi-benchmark feature set in the high-dimensional space. In the low-dimensional mapping space, the similarity probability distribution between samples is reconstructed using a t-distribution with 1 degree of freedom. The core discriminative features are output by iteratively optimizing the Kullback-Leibler (KL) divergence between the probability distributions in the high-dimensional and low-dimensional spaces. To address the issue that t-SNE, as a non-parametric dimensionality reduction algorithm, cannot directly generate low-dimensional mappings for new samples outside the training set, this embodiment employs the following processing strategy: The augmented feature dataset (648 samples) is divided into training and test sets in a 7:3 ratio. The training and test sets are then merged into a single dataset, and a joint t-SNE dimensionality reduction is performed to obtain the core discriminative features of all samples in the low-dimensional space. These core discriminative features are then separated into training and test subsets according to the original partition index for subsequent training and evaluation of the classification model. For entirely new test samples encountered during actual deployment, the standardized multi-benchmark feature vectors of the new samples are merged with the original augmented feature dataset, and t-SNE dimensionality reduction is performed again. The low-dimensional features corresponding to the new samples are used as input to the classification model to ensure the consistency of the low-dimensional space distribution. The comparison results are as follows: Figure 3As shown, compared to the silhouette score of only 0.279 after traditional principal component analysis (PCA) mapping, the silhouette score of the sample after t-SNE nonlinear decoupling is significantly improved to 0.541, and the Calinski-Harabasz index is as high as 2526.7, which greatly improves the cluster separation and topological independence of different oil source categories.

[0029] Step 5: Input the core discriminative features into the hyperparameter-optimized ensemble learning classification model. This model includes multiple base learners that process features in parallel after selection, and a meta-learner that fuses the outputs of each base learner. A soft voting mechanism is then used to output the final oil source classification. Specifically, under the feature metrics of a multi-band benchmark, accuracy, precision, recall, and F1 score are selected as performance evaluation metrics for the model. Accuracy refers to the proportion of correctly classified samples out of the total sample size; precision refers to the proportion of predicted positive examples that are actually positive; recall refers to the proportion of actually positive examples that are correctly predicted as positive; and the F1 score is the harmonic mean of precision and recall. A model library containing multiple candidate algorithms was constructed, including Random Forest (RF), K-nearest neighbor (KNN), support vector machine (SVM), artificial neural network (ANN), Naive Bayes (NB), and eXtremeGradient Boosting (XGBoost). Using a tree-structured Parzen estimator (TPE) within the Hyperopt framework, a probabilistic surrogate model was employed to optimize the hyperparameters of each candidate algorithm using Bayesian optimization, minimizing the five-fold cross-validation error. From the optimized candidate algorithms, three heterogeneous models (XGBoost, SVM, and RF) with the highest hierarchical cross-validation accuracy were selected as Level-0 base learners to process the core discriminative features in parallel. Each base learner independently outputs the posterior probability distribution of each category as meta-features. The Level-1 layer employs L2-regularized logistic regression as the meta-learner, using the posterior probabilities (meta-features) output by each base learner as input. By fitting the probability mapping relationship between meta-features and true labels, it automatically learns the optimal weighting coefficients for each base learner across all categories. Compared to traditional soft voting with fixed weights (where each base learner has equal weights or is manually set), this invention achieves an "adaptive soft voting" effect by learning weighting coefficients through data-driven meta-learning. Essentially, the weighting mechanism of the meta-learner is a learnable weighted soft voting mechanism that dynamically calibrates the contribution of each base learner based on its local predictive ability across different categories, suppressing local prediction bias. Finally, the meta-learner outputs the oil source attribution probability for each category, and the category with the highest probability is determined as the final oil source attribution category. The test results are as follows... Figure 4 , Figure 5 As shown, when using the global optimal strategy of combining baseline I with soft voting ensemble, the model achieves an accuracy of 95.92% and an F1 score of 95.94% on the independent test set, with a total computation time of only 46.98 seconds, achieving a good balance between recognition accuracy and computational efficiency.

[0030] Step Six: Introduce the SHAP (Shapley Additive exPlanations) attribution algorithm to globally deconstruct the ensemble learning classification model. By calculating the marginal contribution of each feature functional group index to the model's prediction results, a feature importance distribution is established, mapping the mathematical weights of the ensemble learning classification model to the differences in the microscopic chemical structure of asphalt. Specifically, after outputting the final oil source classification, the game theory-based SHAP attribution algorithm is introduced to globally deconstruct the optimal soft-voting ensemble model. A feature importance distribution is established by calculating the marginal contribution of each feature functional group index to the model's prediction results. The results show that the asphalt functional group index I... C=O I S=O and I Ar It occupies a dominant position in the decision chain of the model (e.g.) Figure 6 , Figure 7 (As shown).

[0031] This step, through the SHAP algorithm, not only established the long-tailed distribution of feature importance but also confirmed the specific chemical functional group characteristics, namely the carbonyl index ( ), sulfoxide index ( ) and aromatic index ( This plays a dominant role in the decision-making chain of the integrated model. In-depth attribution analysis shows that the model does not rely on meaningless background noise, but accurately captures the true differences in key microscopic chemical bonds in asphalt materials. When core features... When the intensity is high, the corresponding SHAP value exhibits a significantly positive or negative polarization distribution; simultaneously, in the significant spectral feature thermogram, the original spectra of different asphalt oil sources show distinctly different "barcode-like" clustered color bands in the vertical direction. The thermodynamic distribution of the original sample set is highly consistent with the bipolarity phenomenon of the SHAP value.

[0032] SHAP analysis shows that aromatic characteristics have a very high marginal contribution rate to the establishment of classification boundaries, which essentially reflects the "genetic" differences in asphalt crude oil. As the bottom residue of crude oil refining, asphalt's microstructure highly inherits the characteristics of the parent crude oil. Crude oils from different geological basins, different sedimentary environments (such as marine and terrestrial sediments), and different geological evolution maturity exhibit significant natural barriers at the molecular level in terms of aromatic ring abundance and polycyclic aromatic hydrocarbon condensation degree. The ensemble model is highly dependent on... The process of classifying oil sources is precisely a keen capture of the fundamental skeletal fingerprint of oil sources, which is determined by geological origin and is difficult to erase through conventional physical refining methods in the later stages.

[0033] In the SHAP swarm diagram, the eigenvalues ​​of carbonyl and sulfoxide groups exhibit a significant positive and negative polarization distribution, profoundly reflecting the technological differences in processing techniques among different refineries. Carbonyl and sulfoxide groups are typical polar functional groups, and their relative abundance in the matrix bitumen is directly controlled by the thermal history and oxidation degree during crude oil refining. On the one hand, differences in temperature control and residence time at the bottom of atmospheric and vacuum distillation columns among different refineries lead to varying degrees of microscopic thermal condensation and mild thermo-oxidative aging reactions in heavy oil components, thereby causing sulfur to oxidize into sulfoxide groups (…). ), carbon chain oxidation to carbonyl ( The rates of oxidation and oxidation differ. On the other hand, some refineries use a mild oxidation air-blowing process to blend soft asphalt to specific road grades (such as 70# or 90#). This artificial intervention process causes the asphalt molecules to undergo violent dehydrogenation condensation and oxidation reactions, resulting in abnormally prominent absorption peaks for carbonyl and sulfoxide groups.

[0034] Therefore, the prominence of these two features in the decision chain indicates that the model has successfully extracted the irreversible "process fingerprint" imprinted on the asphalt molecular chain by complex thermodynamic processing. This invention, through cross-validation of SHAP attribution and the original feature thermodynamic distribution, establishes for the first time a quantitative scientific link between the black-box data model and the "macroscopic absorption peak differences determined by the geological origin and processing technology of crude oil." This closed-loop explanation, which leaps from purely data-driven approaches to verification of physicochemical mechanisms, completely breaks the black-box nature of machine learning algorithms, greatly enhancing the credibility and practical value of this intelligent classification model in the traceability and quality inspection of raw materials in actual road engineering.

[0035] Comparative Example 1: To verify the advancement of the technical solution of this invention, the same initial feature set was used to compare with traditional single classification models (such as Linear Discriminant Analysis (LDA) and Logistic Regression (LR). When dealing with complex chemical feature sets (such as baseline III), the test accuracy of traditional LDA and LR models dropped sharply to 0.571 and 0.566, respectively. This indicates that traditional linear architectures or single algorithms have extremely poor generalization ability when facing the multidimensional nonlinear characteristics of asphalt; while the intelligent classification method of this invention based on conditional generative adversarial networks and adaptive soft voting not only completely overcomes this defect (accuracy of 95.92%), but also provides a transparent and reliable physicochemical mechanism explanation for the traceability system through SHAP.

[0036] Comparative Example 2: When using the traditional SMOTE (Synthetic Minority Over-sampling Technique) algorithm to process extremely scarce minority class samples, it can only perform local linear interpolation between existing samples and cannot learn the manifold distribution structure of the original samples. (Comparison with experimental data...) Figure 2 The results show that the Pearson correlation coefficient between the SMOTE synthetic samples and the sample features in the original sample set is 0.9237 (p<0.001), and its linear fitting curve (green line) deviates downward from the ideal contour line. This indicates that traditional interpolation methods produce distorted noise and systematic bias at feature boundaries.

[0037] In contrast, the cGAN of this invention, through an adversarial game between a three-layer fully connected network and the discriminator, deeply fits the true manifold distribution of the spectral data. The Pearson correlation coefficient between its synthesized samples and the samples in the original sample set is as high as 0.9348 (p<0.001), and the linear fitting curve (red line) closely follows the ideal contour lines. This direct quantitative and visual comparison statistically confirms that the dynamic adaptive cGAN of this invention outperforms the traditional SMOTE algorithm in terms of high fidelity and feature restoration when dealing with the complex long-tailed spectral distribution of asphalt.

[0038] Comparative Example 3: The existing methods for identifying asphalt oil sources based on a single machine learning algorithm were compared using the same initial feature set. The classification performance of each model in this method is shown in Table 4 below: Table 4. Classification Performance Analysis of Comparative Models

[0039] The best model in the comparison was ANN, with an accuracy of only 81%. The ensemble learning model of this invention achieves an accuracy of 95.92% and an F1 score of 95.94%, which not only significantly improves the classification accuracy compared to the above comparison, but also reduces the sensitivity to single samples through the multi-model fusion architecture, resulting in stronger generalization ability and universality.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for classifying asphalt oil sources based on a conditional generative adversarial network and soft voting integration, characterized in that, include: Infrared spectral data of asphalt samples from different oil sources were collected, and the peak areas of key characteristic peaks were extracted. The key characteristic peak areas of each asphalt sample were used as feature vectors to establish an asphalt infrared spectral feature fingerprint database. The feature vectors of each oil source category in the database constitute the original sample set. The original sample set is input into the conditional generative adversarial network for training. The conditional generative adversarial network automatically adjusts its training strategy according to the difference in the sample size of each category, generates synthetic samples for categories with insufficient sample size, and mixes the original sample set with the synthetic samples to output a category-balanced augmented feature dataset. Based on the peak area data in the amplified feature dataset, feature intervals are divided according to at least two band range benchmarks, feature functional group indices covering at least two structural feature types are calculated, and standardization processing is performed to obtain a standardized multi-benchmark feature set. The standardized multi-benchmark feature set is subjected to nonlinear dimensionality reduction processing to extract core discriminative features; The core discriminative features are input into an ensemble learning classification model optimized by hyperparameters. The ensemble learning classification model includes multiple base learners that process features in parallel after screening, and a meta-learner that fuses the outputs of each base learner. The final oil source classification is output through a soft voting mechanism.

2. The asphalt oil source classification method based on conditional generative adversarial network and soft voting integration as described in claim 1, characterized in that, The conditional generative adversarial network consists of an opposing architecture of a generator and a discriminator, wherein: The generator receives a standard normal noise vector of a preset dimension and a class label as input, and generates the synthetic sample with the same dimension as the feature vector by using a multi-layer fully connected network in conjunction with an activation function and batch normalization. The discriminator adopts a network structure that is mirror-symmetric to the generator, with hidden layers combined with Dropout regularization, and outputs the true and false probabilities through the Sigmoid function. During training, the discriminator is first updated based on the samples in the original sample set and the synthetic samples generated by the generator. The generator is then updated with the goal of deceiving the discriminator.

3. The asphalt oil source classification method based on conditional generative adversarial network and soft voting integration as described in claim 2, characterized in that, The automatic training adjustment strategy based on the differences in sample size across different categories includes: For categories whose sample size in the original sample set is lower than a preset threshold, the training rounds of the generator are automatically increased, the training batch size is reduced, and the tolerance threshold of the early stopping mechanism is increased to force the generator to deeply fit the feature manifold distribution of the samples in that category.

4. The asphalt oil source classification method based on conditional generative adversarial network and soft voting integration as described in claim 1, characterized in that, The characteristic functional group index includes the aliphatic index I. B Aromatic Index I Ar Carbonyl index I C=O , sulfoxide index I S=O Asymmetric aliphatic index I B,a At least two of them, and the calculation methods for each index are as follows: I B =(A 2924 +A 2852 ) / ΣA; I Ar =A 1600 / ΣA; I C=O =A 1700 / ΣA; I S=O =A 1031 / ΣA; I B,a =(A 1377 +A 1461 ) / ΣA; Among them, A 2924 The wave number is 2924 cm⁻¹ -1 The peak area of ​​the antisymmetric stretching vibration of aliphatic CH, A 2852 The wave number is 2852 cm⁻¹ -1 The peak area of ​​the symmetric stretching vibration of aliphatic CH, A 1700 Wavenumber 1700cm -1 The peak area of ​​the carbonyl C=O stretching vibration, A 1600 Wavenumber 1600cm -1 The peak area of ​​the C=C skeleton vibration of the aromatic ring, A 1031 The wave number is 1031 cm⁻¹ -1 The peak area of ​​the S=O stretching vibration of the sulfoxide group, A 1461 The wave number is 1461 cm⁻¹ -1 The peak area of ​​the asymmetric bending vibration of aliphatic CH, A 1377 The wave number is 1377 cm⁻¹ -1 The peak area of ​​the symmetrical bending vibration of aliphatic CH; ΣA is the sum of the peak areas of each key characteristic peak within the corresponding band range reference; I B Characterizing the relative content of aliphatic compounds in asphalt, I Ar Characterizing the relative content of aromatic ring structures in bitumen, I C=O Characterizing the relative content of carbonyl groups in asphalt, I S=O Characterizing the relative content of sulfoxide groups in bitumen, I B,a Characterizes the relative content of asymmetric aliphatic functional groups in asphalt.

5. The asphalt oil source classification method based on conditional generative adversarial network and soft voting integration as described in claim 4, characterized in that, The band range reference includes at least two of the following: Benchmark I: 4000~700cm -1 The sum of the peak areas across the entire spectrum within the range is taken as ΣA; Standard II: Based on 4000~1400cm -1 The sum of the peak areas of the stretching vibrations within the range is taken as ΣA; Standard III: Based on 1400~700cm -1 The sum of the areas of the bending vibration peaks within the range is taken as ΣA; Standard IV: 3000~2800cm -1 The sum of the peak areas of CH stretching vibrations within the range is taken as ΣA.

6. The asphalt oil source classification method based on conditional generative adversarial network and soft voting integration as described in claim 1, characterized in that, The nonlinear dimensionality reduction process employs a t-distributed random neighborhood embedding algorithm, specifically including: Conditional probability is used to measure the topological similarity distribution among samples in the standardized multi-benchmark feature set in a high-dimensional space. The similarity probability distribution between samples is reconstructed using a t-distribution with 1 degree of freedom in a low-dimensional mapping space; By minimizing the KL divergence between the probability distributions in high-dimensional and low-dimensional spaces, gradient descent is applied iteratively to optimize and output the core discriminative features.

7. The asphalt oil source classification method based on conditional generative adversarial network and soft voting integration as described in claim 1, characterized in that, The process of constructing the hyperparameter-optimized ensemble learning classification model includes: A model library containing multiple candidate algorithms is constructed. The hyperparameter Bayesian optimization of the candidate algorithms is performed using a probabilistic surrogate model through a tree-structured Parzen estimator to minimize the cross-validation error. A predetermined number of heterogeneous models are selected from the optimized candidate algorithms as the multiple base learners. The multiple base learners process the core discriminative features in parallel and output the posterior probability distribution of each category as meta-features. The meta-learner fits the probability mapping relationship of the meta-features, and the local prediction bias of each base learner is dynamically calibrated in conjunction with the soft voting mechanism to output the oil source attribution probability of each category; the category with the highest attribution probability is determined as the final oil source attribution category.

8. The asphalt oil source classification method based on conditional generative adversarial network and soft voting integration as described in claim 7, characterized in that, The candidate algorithms include at least three of the following: random forest, support vector machine, extreme gradient boosting, K-nearest neighbors, artificial neural network, and Naive Bayes; the meta-learner is a logistic regression model with L2 regularization; the preset number is three, and the selection criterion is the accuracy of hierarchical cross-validation.

9. The asphalt oil source classification method based on conditional generative adversarial network and soft voting integration as described in claim 1, characterized in that, Following the output of the final oil source attribution category, the following is also included: The SHAP attribution algorithm is introduced to globally deconstruct the ensemble learning classification model. By calculating the marginal contribution of each feature functional group index to the model prediction results, a feature importance distribution is established, and the mathematical weights of the ensemble learning classification model are mapped to the differences in the microscopic chemical structure of asphalt.

10. The asphalt oil source classification method based on conditional generative adversarial network and soft voting integration according to claim 1, characterized in that, The characteristic absorption band corresponding to the key characteristic peak includes a wavenumber of 2924 cm⁻¹. -1 2852cm -1 1700cm -1 1600cm -1 1461cm -1 1377cm -1 1031cm -1 868cm -1 812cm -1 747cm -1 and 722cm -1 .

Images