Asphalt material softening point prediction modeling method based on infrared spectrum

Abstract

The invention provides an asphalt material softening point prediction modeling method based on an infrared spectrum, and belongs to the technical field of road engineering material performance prediction. The method comprises the following steps: collecting infrared spectrum data of a multi-source asphalt sample and carrying out standardization pretreatment; fusing and extracting functional group quantitative features, wavelet multi-scale features and principal component dimension reduction features; constructing a high-stability feature subset by adopting a progressive feature optimization strategy guided by a physical mechanism; a self-adaptive Stacking fusion model is constructed and trained, the model integrates four base learners including PLSR, SVR, random forest and gradient boosting regression, and Bayesian ridge regression is used as a meta-learner to realize uncertainty quantification of a prediction result. The method can quickly output the predicted value of the softening point without a physical test, solves the technical problems of low efficiency and poor consistency of a traditional method, and is suitable for production quality control, on-site quick detection and formula optimization of asphalt materials.

Description

Technical Field

[0001] This invention belongs to the field of road engineering material performance prediction technology, specifically involving a method for predicting and modeling the softening point of asphalt materials based on infrared spectroscopy. Background Technology

[0002] Infrared spectroscopy has been used for qualitative or semi-quantitative analysis of the chemical composition of asphalt in aging evaluation and modification feature identification. However, this technology still has significant limitations: First, the spectral data is high-dimensional and subject to significant noise interference, making it difficult to establish a stable quantitative relationship with macroscopic performance indicators (such as softening point); second, it lacks a systematic feature extraction strategy, and the model is easily affected by different asphalt sources and modification systems, resulting in insufficient generalization ability; third, it cannot meet the needs of rapid and real-time detection in engineering fields and remains at the stage of offline laboratory analysis.

[0003] The softening point is a key parameter determining the high-temperature performance of asphalt. Traditional physical testing methods (such as the ring and ball method) are time-consuming and inefficient, making them unsuitable for the rapid, batch, and online testing requirements of modern production and construction. Accurately extracting effective features reflecting the softening point from infrared spectral data still faces significant technical obstacles, including complex spectral interference, unclear mapping relationships between chemical structure and macroscopic properties, and insufficient model robustness. These issues restrict the in-depth application of spectroscopic technology in asphalt quality control.

[0004] Therefore, there is an urgent need to construct a quantitative prediction model for asphalt softening point based on infrared spectroscopy. Through systematic data preprocessing, feature extraction, and machine learning algorithms, this model can overcome technical obstacles such as inaccurate feature extraction and weak model generalization ability in existing technologies, enabling rapid and accurate prediction of softening point. This will promote the development of asphalt quality testing towards high efficiency and intelligence, and has significant engineering application value. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for predicting and modeling the softening point of asphalt materials based on infrared spectroscopy. By combining deep learning, machine learning modeling, and feature extraction techniques, it achieves rapid and accurate prediction of the softening point.

[0006] To achieve the above objectives, the following steps are specifically included: S1. Collect datasets of three types of matrix asphalt samples. By setting the acquisition parameters of the infrared spectrometer and collecting the raw spectral data according to the preset number of acquisitions for each sample, construct the raw spectral data matrix, and then use the standard ring and ball method to collect softening point labels and output the softening point label vector. The three types of matrix asphalt samples specifically include: naphthenic asphalt, intermediate asphalt, and paraffinic asphalt; The ring and ball method uses glycerol as the medium and a heating rate of 5°C / min.

[0007] S2. Based on the original spectral data matrix, after spectral preprocessing, output the preprocessed spectral matrix; The spectral preprocessing operations include: Savitzky-Golay smoothing, asymmetric least squares baseline correction, standard normal variable transformation, and first derivative enhancement of peak shape characteristics.

[0008] S3. Input the preprocessed spectral matrix, and then perform quantitative extraction of feature functional groups, wavelet transform multi-scale feature extraction, and principal component analysis dimensionality reduction in sequence. Then perform feature fusion and output the fused feature matrix. Infrared spectral data is high in dimensionality and prone to collinearity. Directly inputting the full spectrum data into the model may lead to complex model calculations and an increased risk of overfitting. Therefore, this embodiment performs feature extraction and dimensionality reduction on the preprocessed spectral matrix.

[0009] Specifically, it includes the following steps: S3.1 Input the preprocessed spectral matrix, quantitatively extract chemical features through characteristic functional groups, and output the chemical feature matrix; The chemical characteristic matrix includes: carbonyl index, sulfoxide index, aromaticity index, aliphatic chain index, and spectral centroid.

[0010] S3.2 Input the preprocessed spectral matrix, extract multi-scale features through wavelet transform, and output the wavelet feature matrix; This invention selects the Daubechies 4 (db4) wavelet as the basis function, which has a fourth-order vanishing moment and is suitable for smooth feature extraction of spectral signals. A three-level discrete wavelet decomposition is performed on the spectral curve of each sample. For each of the three levels, the total energy of the detail coefficients, the total energy of the approximation coefficients, and the proportion of each level's energy to the total energy are calculated, outputting the wavelet feature matrix. Each row corresponds to 7 wavelet features of a sample.

[0011] The seven wavelet features include: three detail layer energies, one approximation layer energy, and three energy percentages.

[0012] S3.3 Input the preprocessed spectral matrix, perform dimensionality reduction using principal component analysis (PCA), and output the PCA feature matrix.

[0013] S3.4 Based on the chemical feature matrix, wavelet feature matrix and PCA feature matrix, feature fusion is performed by feature concatenation, and the fused feature matrix is ​​output. This invention horizontally concatenates the extracted chemical feature matrix, wavelet feature matrix, and PCA dimensionality-reduced feature matrix column-wise to form an original fused feature matrix. Subsequently, Z-score standardization is performed on each feature column, i.e., the mean of that column is subtracted and divided by the standard deviation, transforming all features into a standardized distribution with a mean of 0 and a standard deviation of 1. Finally, a standardized fused feature matrix suitable for modeling is obtained. Specifically, it includes the following formulas: The expression for feature concatenation is: Where | denotes the horizontal concatenation operation of matrices. , This represents the number of PCA features. After feature concatenation, Z-score normalization is performed to output a fused feature matrix. .

[0014] S4. Based on the fused feature matrix and softening point labels, the optimized feature matrix is ​​output through initial screening of physical correlation of chemical mechanisms, feature denoising based on stability assessment and feature supplementation driven by interaction effects. The pre-selected mandatory features for the initial screening of the physical correlation of the chemical mechanism are carbonyl index, aromaticity index, and spectral barycenter. The Spearman rank correlation coefficient between these features and the softening point is calculated. Remove Features; The feature denoising for stability assessment specifically involves performing a preset number of Bootstrap resampling operations on the initial feature set, while retaining the stability index. A stable set of features; This invention employs a preliminary screening method based on the physical correlation between chemical mechanisms and the softening point of asphalt, with pre-defined feature retention rules: (1) Based on the correlation mechanism between bitumen chemistry and softening point, the pre-selected essential features are carbonyl index, aromaticity index, and spectral centroid (because they are directly related to bitumen polarity and molecular weight distribution).

[0015] (2) Calculate the Spearman rank correlation coefficient between the remaining features and the softening point. Remove Its characteristics.

[0016] In asphalt softening point testing, ±2℃ is an acceptable error range for engineering applications. This has been verified through experiments. The characteristics of this feature have a weak explanatory power for changes in softening point, and including them will introduce noise rather than effective signals.

[0017] This invention uses the physical correlation of chemical mechanisms for initial screening to obtain a preliminary feature set. This feature set is then subjected to 1000 Bootstrap resampling cycles (70% of the samples each time), and the selection frequency of each feature is recorded. The stability index is then calculated. ,reserve A stable set of features; Features remain stably selected in more than 80% of resampling, reducing the model's sensitivity to sample fluctuations.

[0018] Based on a stable feature set and softening point labels, random forest is used to calculate the feature interaction importance. Identify the importance of interactions For each feature pair, interactive features are constructed and added to the feature set to obtain an interactive enhanced feature set. This set is then used for final feature optimization. Recursive Feature Elimination (RFE) combined with a random forest is employed, and 5-fold cross-validation is used to determine the optimal number of features, resulting in an optimized feature matrix. Set up 50% cross-validation.

[0019] S5. Data partitioning and adaptive standardization based on spatiotemporal consistency: Based on the optimized feature matrix and soft point labels, training set, validation set and test set are output through temporal hierarchical partitioning, adaptive standardization parameter calculation and dataset standardization transformation. The time-series stratification is specifically divided into three consecutive time periods based on the collection time: the training period (first 70% of the time samples), the validation period (middle 15%), and the testing period (last 15%).

[0020] S6. Input the training set, and through base learner design and training, and Bayesian Stacking fusion, obtain the Bayesian Stacking model to complete the method for predicting the softening point of asphalt materials. Specifically, the following steps are included: S6.1 Base Learner Configuration and Training: Input the training set, use temporal cross-validation to train and validate the model within the training set, and generate the meta-feature matrix; the validation set is used for hyperparameter tuning and early stopping control, and the output is a set of base learners; The four base learners include: PLSR, SVR, random forest, and gradient boosting regression; The PLSR used in this invention is chemically confined ( The model's parameters are set as follows: number of latent variables = 10, functional group interval weighting factor λ = 0.5, and weighting constraint: for carbonyl groups (1700cm²...). - ¹), sulfoxide (1030cm) - ¹), Fragrance (1600cm) - ¹) Apply a weight of 1.5 to the three feature intervals and train the output. Model and Predicted value.

[0021] SVR is an adaptive kernel model. In this embodiment, the parameter settings are as follows: The expression for the composite kernel is as follows: in, For radial basis kernel functions, It is a linear kernel function. To preset the radial basis function weights, Preset linear kernel function weights; training output Model and Predicted value.

[0022] Random forest is a stability-enhanced model. In this invention, the parameters are specifically set as follows: n_estimators=150, max_depth=12, min_samples_split=5. The feature sampling strategy in this invention is based on feature importance weighting, with the top 30% of features being selected with a 2x probability. The training output model... Model and Predicted value.

[0023] Gradient boosting regression is physically regularized. In this invention, the parameters are specifically set as follows: n_estimators=120, learning_rate=0.08, max_depth=6, and physical constraints are added by including a softpoint range penalty term in the loss function, as shown in the following expression: in, The softening point predicted by the model; training output model Model and Predicted value.

[0024] This invention trains four base learners using temporal cross-validation, specifically by training on five training folds and predicting on the corresponding validation folds, then concatenating the results to obtain the meta-feature columns of the complete training set.

[0025] Output base learner set Sum of characteristic matrices .

[0026] S6.2, Bayesian Stacking Fusion: Input meta-feature matrix and softening point labels, train the model and obtain the posterior distribution through the preset configuration of the meta-learner, and output a Bayesian Stacking model; In this invention, the pre-configured meta-learner is trained using Bayesian ridge regression and the method of maximizing log-marginal likelihood inference. The expression for maximizing log-marginal likelihood is as follows: In the formula, To soften the label, Let n be the meta-feature matrix, and n be the number of samples. For weighted prior accuracy, For noise accuracy, Let be the covariance matrix.

[0027] The Bayesian Stacking model is obtained by using the posterior distribution.

[0028] S7. Input the test set into the Bayesian Stacking model, perform model prediction, and output the accuracy index; The accuracy indicators include: coefficient of determination, root mean square error, and mean absolute error. Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively solves the technical problems of traditional physical testing methods, such as long cycle time, low efficiency, and significant human error, as well as the insufficient prediction ability of existing spectral analysis technology for macroscopic performance indicators and weak model generalization ability, by constructing a quantitative prediction model for asphalt softening point based on the deep integration of infrared spectral data and machine learning. This invention innovatively proposes a multi-stage feature extraction and optimization framework oriented towards the spectral characteristics of asphalt, systematically overcoming the interference of spectral noise and source differences on modeling; further, it adopts an adaptive stacking fusion modeling strategy, integrating the advantages of multiple machine learning algorithms, to achieve high-precision and robust mapping between spectral features and softening point; finally, it establishes a rapid prediction system that can output prediction uncertainty without physical testing. This invention significantly improves detection efficiency and enhances result consistency, providing a reliable and efficient solution for real-time monitoring of asphalt material production processes, rapid on-site assessment in engineering projects, and quality consistency control, and has outstanding engineering application value and industrialization prospects. Attached Figure Description

[0029] Figure 1 This is a flowchart of a method for predicting and modeling the softening point of asphalt materials based on infrared spectroscopy, according to the present invention. Figure 2 This is a comparison diagram of the original infrared spectrum and the pretreated spectrum of the asphalt sample of the present invention; Figure 3 This is a scatter plot showing the correlation between the predicted values ​​and the measured values ​​of the model in this invention. Detailed Implementation

[0030] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0031] like Figure 1 As shown, a method for predicting and modeling the softening point of asphalt materials based on infrared spectroscopy specifically includes the following steps: S1. Collect datasets of three types of matrix asphalt samples. By setting the acquisition parameters of the infrared spectrometer and collecting the raw spectral data according to the preset number of acquisitions for each sample, construct the raw spectral data matrix, and then use the standard ring and ball method to collect softening point labels and output the softening point label vector. The three types of matrix asphalt samples specifically include: naphthenic asphalt, intermediate asphalt, and paraffinic asphalt; The ring and ball method uses glycerol as a medium and a heating rate of 5°C / min. This invention selects a dataset of asphalt samples from different crude oil producing regions in China, specifically including asphalt samples from three major categories of crude oil sources: naphthenic asphalt, intermediate-based asphalt, and paraffinic asphalt. Since asphalt from different crude oil types has fundamentally different chemical compositions, it directly affects the mapping relationship between infrared spectral characteristics and softening points. In this embodiment, each category has 170 sample groups, for a total of 510 asphalt sample groups. For each sample, 15 infrared spectral data points and a softening point label are collected.

[0032] like Figure 2 As shown, this embodiment uses a Fourier transform infrared spectrometer (FTIR-ATR) to acquire full-spectrum data in the wavenumber range of 600–4000 cm⁻¹. The acquisition parameters are set to a resolution of 2 cm⁻¹ and a temperature control of 25 ± 0.5 °C. After acquiring the raw spectral data, the raw spectral data is saved in the form of a digital matrix, with matrix rows corresponding to sample numbers and matrix columns corresponding to the absorption intensity at wavenumber points. The raw spectral data matrix is ​​shown below. , where n is the number of samples and m is the absorption intensity at the corresponding wavenumber point.

[0033] Simultaneously, the softening point of the same batch of asphalt samples was determined using the ring and ball method. Specifically, a standard ring and ball softening point tester was used, with glycerol as the medium and a heating rate of 5℃ / min, to obtain the softening point label for each sample. Each sample was measured in parallel five times. After removing obvious outliers, the arithmetic mean was taken as the softening point label for that sample. This creates a one-to-one correspondence between spectral data and softening point label.

[0034] S2. Based on the original spectral data matrix, after spectral preprocessing, output the preprocessed spectral matrix; The spectral preprocessing operations include: Savitzky-Golay smoothing, asymmetric least squares baseline correction, standard normal variable transformation, and first derivative enhancement of peak shape characteristics; Since the raw infrared spectrum may be affected by instrument noise, environmental background, baseline drift, and differences in sample thickness / scattering, this invention performs image enhancement operations on the raw spectral data matrix through spectral preprocessing to improve model stability and prediction accuracy. The spectral preprocessing operation includes the following steps: (1) Denoising / smoothing: Savitzky-Golay smoothing was used, with a window size of 15 and a polynomial order of 3. (2) Baseline correction: Asymmetric least squares method is used for baseline correction; (3) Standardization process: Standard normal variable transformation (SNV) is used to reduce sample preparation differences and scattering effects; (4) Derivative processing: First derivative (SG derivative, window set to 11, order set to 2) is used to enhance peak shape characteristics.

[0035] After spectral preprocessing, the preprocessed spectral matrix is ​​output. This provides input for subsequent feature extraction and modeling.

[0036] S3. Input the preprocessed spectral matrix, and then perform quantitative extraction of feature functional groups, wavelet transform multi-scale feature extraction, and principal component analysis dimensionality reduction in sequence. Then perform feature fusion and output the fused feature matrix. Infrared spectral data is high-dimensional and prone to collinearity. Directly inputting the full spectrum data into the model may lead to computational complexity and increased risk of overfitting. Therefore, this embodiment performs feature extraction and dimensionality reduction on the preprocessed spectral matrix, specifically including the following steps: S3.1 Input the preprocessed spectral matrix, quantitatively extract chemical features through characteristic functional groups, and output the chemical feature matrix; The chemical characteristic matrix includes: carbonyl index, sulfoxide index, aromaticity index, aliphatic chain index, and spectral centroid; This embodiment calculates the peak area / peak height by identifying characteristic peak regions (carbonyl, sulfoxide, aromatic, aliphatic chain). The characteristic peak regions include: carbonyl area. (Range: 1690-1740 cm) - ¹) Sulfoxide area (Range: 1020-1070 cm) - ¹) Aromatic area (Range: 1580-1620 cm) - ¹) and the peak area of ​​adipose chains (Range: 1450-1470 cm) - ¹).

[0037] Based on the characteristic peak area selected in this embodiment, the standardization index is calculated as follows: The expression for the carbonyl index is: The expression for the sulfoxide index is: The expression for the aroma index is: The expression for the fatty acid chain index is: The spectral centroids are calculated based on the carbonyl index, sulfoxide index, aromaticity index, and aliphatic chain index, and the chemical characteristic matrix is ​​output. ; S3.2 Input the preprocessed spectral matrix, extract multi-scale features through wavelet transform, and output the wavelet feature matrix; This invention selects the Daubechies 4 (db4) wavelet as the basis function, which has a fourth-order vanishing moment and is suitable for smooth feature extraction of spectral signals. A three-level discrete wavelet decomposition is performed on the spectral curve of each sample. For each of the three levels, the total energy of the detail coefficients, the total energy of the approximation coefficients, and the proportion of each level's energy to the total energy are calculated, outputting the wavelet feature matrix. Each row corresponds to 7 wavelet features of a sample.

[0038] The seven wavelet features include: three detail layer energies, one approximation layer energy, and three energy percentages.

[0039] S3.3 Input the preprocessed spectral matrix, perform dimensionality reduction using principal component analysis (PCA), and output the PCA feature matrix; This embodiment centers the spectral matrix to eliminate baseline shift, then calculates its covariance matrix and performs eigenvalue decomposition, extracting the top k principal components with the largest eigenvalues ​​as the dimensionality-reduced feature variables. The value of k is determined based on the principle that the cumulative variance contribution rate is not less than 95%. This embodiment compresses the preprocessed spectral data to an 8-12 dimensional principal component feature space, significantly reducing data dimensionality and redundancy while retaining most of the spectral information. In this invention, k is set to 11. , The number of PCA features. The range of values ​​is .

[0040] S3.4 Based on the chemical feature matrix, wavelet feature matrix and PCA feature matrix, feature fusion is performed by feature concatenation, and the fused feature matrix is ​​output. This invention horizontally concatenates the extracted chemical feature matrix, wavelet feature matrix, and PCA dimensionality-reduced feature matrix column-wise to form an original fused feature matrix. Subsequently, Z-score standardization is performed on each feature column, i.e., the mean of that column is subtracted and divided by the standard deviation, transforming all features into a standardized distribution with a mean of 0 and a standard deviation of 1. Finally, a standardized fused feature matrix suitable for modeling is obtained. Specifically, it includes the following formulas: The expression for feature concatenation is: Where | denotes the horizontal concatenation operation of matrices. , This represents the number of PCA features. After feature concatenation, Z-score normalization is performed to output a fused feature matrix. .

[0041] S4. Progressive feature screening based on asphalt spectral properties: Based on the fused feature matrix and softening point label, the system performs initial screening through physical correlation of chemical mechanisms, feature denoising based on stability assessment, and feature supplementation driven by interaction effects, and outputs an optimized feature matrix. The preset mandatory features for the initial screening of the physical correlation of the chemical mechanism are carbonyl index, aromaticity index, and spectral centroid. The feature denoising for stability assessment specifically involves performing a preset number of Bootstrap resampling operations on the initial feature set, while retaining the stability index. A stable set of features; This invention employs a preliminary screening method based on the physical correlation between chemical mechanisms and the softening point of asphalt, with pre-defined feature retention rules: (1) Based on the correlation mechanism between bitumen chemistry and softening point, the pre-selected essential features are carbonyl index, aromaticity index, and spectral centroid (because they are directly related to bitumen polarity and molecular weight distribution).

[0042] (2) Calculate the Spearman rank correlation coefficient between the remaining features and the softening point. Remove Its characteristics.

[0043] In asphalt softening point testing, ±2℃ is an acceptable error range for engineering applications. This has been verified through experiments. The characteristics of this feature have a weak explanatory power for changes in softening point, and including them will introduce noise rather than effective signals.

[0044] This invention uses the physical correlation of chemical mechanisms for initial screening to obtain a preliminary feature set. This feature set is then subjected to 1000 Bootstrap resampling cycles (70% of the samples each time), and the selection frequency of each feature is recorded. The stability index is then calculated. ,reserve A stable set of features; Features remain stably selected in more than 80% of resampling, reducing the model's sensitivity to sample fluctuations.

[0045] Based on a stable feature set and softening point labels, random forest is used to calculate the feature interaction importance. Identify the importance of interactions For each feature pair, interactive features are constructed and added to the feature set to obtain an interactive enhanced feature set. This set is then used for final feature optimization. Recursive Feature Elimination (RFE) combined with a random forest is employed, and 5-fold cross-validation is used to determine the optimal number of features, resulting in an optimized feature matrix. A 5-fold cross-validation was implemented to avoid evaluation bias caused by uneven sample distribution. Based on a systematic experiment conducted on 510 asphalt samples, when the characteristic number... Reducing the number of trees from 12 to 11 resulted in an RMSE increase of only 0.01°C, a significant reduction in model complexity, an 8.3% decrease in the number of parameters, and a 12% increase in inference speed. This embodiment was tested with 50-300 trees, and the OOB error tended to stabilize at 150 trees. Therefore, when using random forest to calculate the importance of feature interactions, n_estimators=150.

[0046] S5. Data partitioning and adaptive standardization based on spatiotemporal consistency: Based on the optimized feature matrix and soft point labels, training set, validation set and test set are output through temporal hierarchical partitioning, adaptive standardization parameter calculation and dataset standardization transformation. The time-series stratification is specifically divided into three consecutive time periods based on the collection time: training period (first 70% of the time samples), validation period (middle 15%), and testing period (last 15%). This invention uses the sliding window statistical method to calculate adaptive standardization parameters based on the feature data of the training set after time-series partitioning: for each feature dimension, a sliding window is formed by taking 50 consecutive samples, and the mean and standard deviation of the features within the window are calculated sequentially as the standardization parameters of the window, which are used as the training set.

[0047] The validation and test sets are transformed using the same statistics from the nearest neighbor window.

[0048] S6. Input the training set, and through base learner design and training, and Bayesian Stacking fusion, obtain the Bayesian Stacking model to complete the method for predicting the softening point of asphalt materials. Specifically, the following steps are included: S6.1 Base Learner Configuration and Training: Input the training set, use temporal cross-validation to train and validate the model within the training set, and generate the meta-feature matrix; the validation set is used for hyperparameter tuning and early stopping control, and the output is a set of base learners; The four base learners include: PLSR, SVR, random forest, and gradient boosting regression; The PLSR used in this invention is chemically confined ( The model's parameters are set as follows: number of latent variables = 10, functional group interval weighting factor λ = 0.5, and weighting constraint: for carbonyl groups (1700cm²...). -¹), sulfoxide (1030cm) - ¹), Fragrance (1600cm) - ¹) Apply a weight of 1.5 to the three feature intervals and train the output. Model and Predicted value.

[0049] SVR is an adaptive kernel model. In this embodiment, the parameter settings are as follows: The expression for the composite kernel is as follows: in, For radial basis kernel functions, It is a linear kernel function. To preset the radial basis function weights, The preset linear kernel function weights are 0.8 (an empirically optimized value, selected based on cross-validation) and 0.2 (used to balance fitting and extrapolation); training output Model and Predicted value.

[0050] Random forest is a stability-enhanced model. In this embodiment, the parameters are set as follows: n_estimators=150, max_depth=12, min_samples_split=5. The feature sampling strategy in this embodiment is to weight features according to their importance, with the top 30% of features being selected with a probability of 2. The training output model is then used. Model and Predicted value.

[0051] Gradient boosting regression is physically regularized. In this embodiment, the parameters are specifically set as follows: n_estimators=120, learning_rate=0.08, max_depth=6, and physical constraints are added by including a softpoint range penalty term in the loss function, as shown in the following expression: in, The softening point predicted by the model; training output model Model and Predicted value.

[0052] This invention trains four base learners using temporal cross-validation, specifically by training on five training folds and predicting on the corresponding validation folds, then concatenating the results to obtain the meta-feature columns of the complete training set.

[0053] Output base learner set Sum of characteristic matrices .

[0054] S6.2, Bayesian Stacking Fusion: Input meta-feature matrix and softening point labels, train the model and obtain the posterior distribution through the preset configuration of the meta-learner, and output a Bayesian Stacking model; In this invention, the pre-configured meta-learner is trained using Bayesian ridge regression and the method of maximizing log-marginal likelihood inference. The expression for maximizing log-marginal likelihood is as follows: In the formula, To soften the label, Let n be the meta-feature matrix, and n be the number of samples. For weighted prior accuracy, For noise accuracy, Let be the covariance matrix.

[0055] The Bayesian Stacking model is obtained by using the posterior distribution. S7. Input the test set into the Bayesian Stacking model, perform model prediction, and output the accuracy index; This invention inputs the test set into a Bayesian Stacking model to obtain the predicted softening point value and performs a coefficient of determination. Calculation of accuracy indices: root mean square error (RMSE) and mean absolute error (MAE); This embodiment is based on an adaptive Bayesian Stacking model trained with over 500 sets of samples, specifically 510 sets, and implemented on the test set. , The prediction accuracy is ℃, and 94.6% of the prediction error is within the engineering tolerance range (±2℃), which meets the accuracy requirements for rapid on-site testing of asphalt materials.

[0056] like Figure 3 As shown, the scatter plot illustrates the correlation between the model's predicted values ​​and the measured values ​​obtained using the ring and ball method. All sample points are closely distributed on both sides of the reference line, indicating a uniform distribution of prediction errors without any obvious "underestimation of high values" or "overestimation of low values." This figure visually demonstrates that the model constructed in this invention possesses high accuracy and good generalization ability.

[0057] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting and modeling the softening point of asphalt materials based on infrared spectroscopy, characterized in that, Includes the following steps: S1. Collect datasets of three types of matrix asphalt samples. By setting the acquisition parameters of the infrared spectrometer and collecting the raw spectral data according to the preset number of acquisitions for each sample, construct the raw spectral data matrix, and then use the standard ring and ball method to collect softening point labels and output the softening point label vector. S2. Based on the original spectral data matrix, after spectral preprocessing, output the preprocessed spectral matrix; S3. Input the preprocessed spectral matrix, and then perform quantitative extraction of feature functional groups, wavelet transform multi-scale feature extraction, and principal component analysis dimensionality reduction in sequence. Then perform feature fusion and output the fused feature matrix. S4. Based on the fused feature matrix and softening point labels, the optimized feature matrix is ​​output through initial screening of physical correlation of chemical mechanisms, feature denoising based on stability assessment and feature supplementation driven by interaction effects. S5. Data partitioning and adaptive standardization based on spatiotemporal consistency: Based on the optimized feature matrix and soft point labels, training set, validation set and test set are output through temporal hierarchical partitioning, adaptive standardization parameter calculation and dataset standardization transformation. S6. Input the training set, and through base learner design and training, and Bayesian Stacking fusion, obtain the Bayesian Stacking model to complete the method for predicting the softening point of asphalt materials. S7. Input the test set into the Bayesian Stacking model, perform model prediction, and output the accuracy index; The accuracy indicators include: coefficient of determination, root mean square error, and mean absolute error.

2. The method for predicting and modeling the softening point of asphalt materials based on infrared spectroscopy according to claim 1, characterized in that, S3 specifically includes the following steps: S3.1 Input the preprocessed spectral matrix, quantitatively extract chemical features through characteristic functional groups, and output the chemical feature matrix; The chemical characteristic matrix includes: carbonyl index, sulfoxide index, aromaticity index, aliphatic chain index, and spectral centroid; S3.2 Input the preprocessed spectral matrix, extract multi-scale features through wavelet transform, and output the wavelet feature matrix; The wavelet features are specifically seven, including: three detail layer energies, one approximation layer energy, and three energy percentages. S3.3 Input the preprocessed spectral matrix, perform dimensionality reduction using principal component analysis (PCA), and output the PCA feature matrix; S3.4 Based on the chemical feature matrix, wavelet feature matrix and PCA feature matrix, feature fusion is performed by feature concatenation, and the fused feature matrix is ​​output. The expression for feature concatenation is: 。 3. Among them, | indicates the horizontal concatenation operation of matrices. For chemical characteristic matrix, The wavelet feature matrix, Here is the PCA feature matrix, and n is the number of samples. , Preset the number of PCA features.

4. The method for predicting and modeling the softening point of asphalt materials based on infrared spectroscopy according to claim 1, characterized in that, In S2, the spectral preprocessing operations include: Savitzky-Golay smoothing, asymmetric least squares baseline correction, standard normal variable transformation, and first derivative enhancement of peak shape characteristics.

5. The method for predicting and modeling the softening point of asphalt materials based on infrared spectroscopy according to claim 1, characterized in that, In step S4, the preset mandatory features for the initial screening of the physical correlation of the chemical mechanism are carbonyl index, aromaticity index, and spectral barycenter. The Spearman rank correlation coefficient between the features and the softening point is calculated. Remove Features; The feature denoising for stability assessment specifically involves performing a preset number of Bootstrap resampling operations on the initial feature set, while retaining the stability index. The stable feature set.

6. The method for predicting and modeling the softening point of asphalt materials based on infrared spectroscopy according to claim 1, characterized in that, In S5, the validation set and the test set are transformed using the same statistics of the nearest neighbor window.

7. The method for predicting and modeling the softening point of asphalt materials based on infrared spectroscopy according to claim 1, characterized in that, S6 specifically includes the following steps: S6.1 Base Learner Configuration and Training: Input the training set, use temporal cross-validation to train and validate the model within the training set, and generate the meta-feature matrix; the validation set is used for hyperparameter tuning and early stopping control, and the output is a set of base learners; The four base learners include: PLSR, SVR, random forest, and gradient boosting regression; The SVR is an adaptive kernel model, where the expression for the SVR composite kernel is as follows: 。 8. Among them, For radial basis kernel functions, It is a linear kernel function. To preset the radial basis function weights, Preset the weights for the linear kernel function; S6.2 Bayesian Stacking Fusion: Input the meta-feature matrix and softening point labels, train the model and obtain the posterior distribution through the preset configuration of the meta-learner, and output the Bayesian Stacking model. The preset configuration of the meta-learner specifically uses Bayesian ridge regression and the maximum log-marginal likelihood inference method.

9. The method for predicting and modeling the softening point of asphalt materials based on infrared spectroscopy according to claim 6, characterized in that, In S6.2, the expression for maximizing the log-marginal likelihood is: 。 10. In the formula, To soften the label, Let n be the meta-feature matrix, and n be the number of samples. For weighted prior accuracy, For noise accuracy, Let be the covariance matrix.

11. The method for predicting and modeling the softening point of asphalt materials based on infrared spectroscopy according to claim 1, characterized in that, In S1, the three types of matrix asphalt samples specifically include: naphthenic asphalt, intermediate asphalt, and paraffinic asphalt; The ring and ball method uses glycerol as the medium and a heating rate of 5°C / min.

Images