Asphalt material rapid identification method based on infrared spectrum

Abstract

The invention provides an asphalt material rapid identification method based on an infrared spectrum, and belongs to the technical field of road engineering material detection and intelligent identification. According to the method, the infrared spectrum of asphalt serves as a chemical fingerprint, background interference is eliminated through standardized spectrum preprocessing, multi-scale features (including key functional group peak intensity / peak area, spectral form statistical features and PCA dimensionality reduction features) are extracted, and rapid identification and confidence output of asphalt categories are achieved in combination with machine learning models such as a random forest. Recognition categories cover matrix asphalt, SBS modified asphalt, rubber powder modified asphalt and the like, and can be expanded to aging grade judgment and mixing amount interval judgment. Compared with an existing method depending on solvent separation or experience discrimination, the method has the advantages of being small in sample demand quantity, simple in pretreatment, high in recognition speed, quantifiable in result, capable of supporting field deployment and the like, is suitable for field acceptance, mixing early warning and construction process quality control, and provides a reliable technical means for efficient and accurate discrimination of asphalt materials.

Description

Technical Field

[0001] This invention belongs to the field of road engineering material testing and intelligent identification technology, specifically involving a rapid identification method for asphalt materials based on infrared spectroscopy. Background Technology

[0002] In the production and construction of asphalt materials, problems such as confusion regarding modifier types, inaccurate dosage, or batch mixing often occur, seriously affecting the quality of pavement engineering. Currently, common identification methods mainly include traditional chemical testing methods (such as solvent separation and chromatographic analysis), conventional physical performance tests (such as penetration and softening point), and manual judgment. However, these methods or processes are cumbersome, cannot be quickly implemented on-site, or cannot accurately identify the type and dosage of modifiers, and the results are easily affected by subjective factors. In recent years, while machine learning-based oil source identification technology has improved identification efficiency, it mainly focuses on differentiating asphalt oil sources, and still lacks systematic and effective solutions for modifier type identification, dosage range judgment, and result confidence assessment.

[0003] Previous studies have utilized infrared spectral features and pattern recognition to distinguish between asphalt from different oil sources. However, this method still has the following significant limitations: first, it cannot effectively identify whether modifiers have been added to the asphalt and their specific types (such as SBS, rubber powder, etc.); second, it lacks the ability to discriminate between different modifier dosage ranges; and third, it does not introduce confidence level output and uncertainty sample indication mechanisms, making it difficult to handle complex and variable field samples in practical engineering. Therefore, there is an urgent need to develop a field-applicable method that can quickly and accurately identify asphalt type and modification state, and has the ability to assess the reliability of the results. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a rapid identification method for asphalt materials based on infrared spectroscopy.

[0005] To achieve the above objectives, the following steps are specifically included: S1. Based on the pre-defined categories of asphalt samples, infrared spectral data are collected using an attenuated total reflectance-Fourier transform infrared spectrometer (ATR-FTIR) to construct an infrared spectral fingerprint database. The preset categories specifically include base asphalt, SBS modified asphalt, rubber powder modified asphalt, and composite modified asphalt; the infrared spectral fingerprint database includes infrared spectral data and tag vectors; This invention employs an ATR-FTIR spectrometer, with specific parameters including a wavenumber range of 4000–600 cm⁻¹. -1 The resolution is set to 4 cm. -1The number of scans was set to 32, and the temperature control range was 25±1℃. Each asphalt sample was collected 3 times, and the average spectrum was calculated to construct infrared spectral data, which was then labeled with category tags. (0 represents base asphalt, 1 represents SBS modified asphalt, 2 represents rubber powder modified asphalt, and 3 represents composite modified asphalt), construct a label vector; S2. Based on the infrared spectral fingerprint database, the preprocessed spectrum is output through preprocessing operations; The preprocessing operations specifically include polynomial fitting baseline correction, derivative transformation to enhance features, and classification-oriented normalization. In this invention, a polynomial fitting baseline correction method is used to determine the spectral intensity after baseline correction. The expression is as follows: In the formula, The original spectral intensity at wavenumber v, Let be a polynomial function of degree k, representing the estimated baseline. In this embodiment, considering both fitting ability and stability, k is set to 3. In this embodiment, the fitting interval for infrared spectral data is selected as 2500-2000 cm⁻¹. -1 and 1800-1700 cm -1 .

[0006] Subsequently, derivative transformation is performed to enhance the features, specifically including enhancing peak position differences with the first derivative and separating overlapping peaks with the second derivative to highlight hidden features. Finally, classification-guided normalization is performed using max-min normalization to preserve the relative peak intensity ratio, resulting in the preprocessed spectrum, as shown in the following expression: In the formula, The value is the spectral intensity value, and scale_factor is a preset range of dynamic weighting coefficients. In this embodiment, the dynamic weighting coefficient ranges from 0.8 to 1.5, and high weights are assigned to the characteristic peaks of the modifier based on known chemical knowledge.

[0007] S3. Based on the preprocessed spectrum, after extracting time-domain statistical features and spectral shape features, feature fusion is performed to output a spliced ​​feature vector, which is then spliced ​​with the basic spectral feature vector to output a dimension-reduced fingerprint feature. The fundamental spectral feature vector is the original Each spectral intensity value, The total number of spectral data points is 850 in this invention; Specifically, the following steps are included: S3.1 Input the preprocessed spectrum, extract time-domain statistical features, and output a time-domain statistical feature vector; This invention calculates (based on preprocessed spectra) using basic statistical characteristics. ) and extreme value characteristic calculation ( Output time-domain statistical feature vector There are 7 dimensions in total, among which, It is a characteristic of the mean. For variance characteristics, As a skewness feature, Kurtosis is a characteristic feature. The position of the maximum value. The position of the minimum value. The peak-to-valley difference; specifically includes the following steps: (1) The mean characteristic, representing the average intensity of the spectrum, reflects the overall absorbance level of the sample. Its calculation expression is as follows: In the formula, The total number of spectral data points. for index, For the spectral vector, in this embodiment, m=850.

[0008] (2) The variance characteristic represents the degree of fluctuation in spectral intensity and reflects the complexity of the spectral shape. Its calculation expression is as follows: .

[0009] (3) Skewness characteristic, representing the asymmetry of spectral distribution, with positive values ​​indicating a right-side (high wavenumber) tail and negative values ​​indicating a left-side tail. The calculation expression is as follows: In the formula, The standard deviation is cubed.

[0010] (4) Kurtosis, representing the asymmetry of the spectral distribution, indicates a right-side (high wavenumber) tail and a left-side tail. The calculation expression is as follows: In the formula, The standard deviation is the fourth power.

[0011] (5) The location of the maximum value is calculated using the following expression: In the formula, For wavenumber vectors, This is the spectral vector.

[0012] (6) The minimum value is located, and the calculation expression is as follows: .

[0013] (7) The peak-to-valley difference is calculated using the following expression: .

[0014] This invention constructs a time-domain statistical feature vector through the above calculations, thereby completing the extraction of time-domain statistical features.

[0015] S3.2 Input the preprocessed spectrum, extract spectral shape features, and output the spectral shape feature vector; In this embodiment, the wavenumber range is set to 4000–600 cm⁻¹. -1 Therefore, the corresponding wavenumber vector middle, , ; This invention calculates based on the preprocessed spectrum and the characteristics of the spectral band shape. ) and local peak characteristic calculation ( Output spectral eigenvectors: 5 dimensions in total For the center of gravity of the score, For spectral width, The number of local peaks, This represents the average value of the symmetry index of the spectral peaks. The standard deviation of the symmetry index of spectral peaks; Specifically, the following steps are included: (1) Calculate the centroid of the spectral band using the first moment of the spectrum The calculation expression is as follows: In the formula, This represents the corresponding wave value.

[0016] (2) Calculate the spectral band width using the second-order central moments The calculation expression is as follows: .

[0017] (3) The derivative spectrum is obtained by taking the first derivative of the infrared spectral data. Search By identifying local maxima at a given point, the number of local peaks can be determined. .

[0018] The local maximum is specifically defined as if the derivative curve If the curve is positive (rising) before crossing zero and negative (falling) after crossing zero, then this zero-crossing point corresponds to a local maximum in the infrared spectrum.

[0019] (4) For each peak, calculate the left and right half-width ratios. Statistical calculations yielded and The calculation expression is as follows: In the formula, for The left half of the peak is wide. for The right half of the peak is wide. This is an index for the number of local peaks.

[0020] The average value of the symmetry index of the spectral peaks The calculation expression is as follows: In the formula, This represents the number of local peaks.

[0021] Standard deviation of the symmetry index of spectral peaks The calculation expression is as follows: .

[0022] This invention constructs a spectral feature vector through the above calculations, thereby completing the spectral feature extraction.

[0023] S3.3 Input the time-domain statistical feature vector and the spectral feature vector, perform feature fusion through weighted feature concatenation, and output the concatenated feature vector; The expression for the feature concatenation of this invention is: Here, || represents a horizontal splicing operation. For time-domain statistical feature vectors, For spectral eigenvectors, To pre-determine the weights of the time-domain statistical features, Preset spectral feature weights.

[0024] This invention, based on feature engineering experience and previous feature importance analysis, weights time-domain statistical features. (Default is 1.0), weights for spectral features (Because it is highly correlated with local peak characteristics) an enhancement weight is assigned (1.2 in this invention).

[0025] S3.4 Input the spliced ​​feature vector and the basic spectral feature vector, perform feature splicing, and output the complete feature vector; The complete feature vector expression output by this invention is as follows: In the formula, The total number of spectral data points. The base spectral feature vector is denoted by ||, which represents the horizontal splicing operation.

[0026] S4. Based on the complete feature vector, the dimensionality-reduced fingerprint feature vector is output through principal component analysis (PCA) calculation and principal component selection. In this embodiment, all principal components are first calculated using PCA, and a cumulative variance contribution rate threshold (95%) is set to understand the overall information content of the data; such as Figure 4 As shown, the contribution of functional group regions to the infrared spectra of different types of asphalt is analyzed; for example... Figure 5 As shown, 123 principal components (K_cum=123) are required to reach this threshold. Subsequently, based on feature importance analysis or model performance optimization, this embodiment selects the top 25 principal components (K_used=25) with the highest discriminative power for subsequent modeling. The standardized data is projected onto these 25 principal components to obtain the dimensionality-reduced features. K is the preset principal component value; S5. Construct a dataset based on the dimensionality-reduced fingerprint feature vector and label vector, divide the dataset into a training set, a validation set and a test set, input the training set and validation set into the classification model for training, and obtain the asphalt material recognition model; The classification model uses a random forest classifier. The specific parameter settings for the random forest classifier in this invention include: 300 trees, 20 maximum depths, 5 minimum leaf node samples, default feature sampling ratio, and 42 random seeds. S6. Input the test set into the asphalt material recognition model. Through the confidence decision module, perform threshold judgment. When the confidence level... If the condition is met, proceed to the retest suggestion module; otherwise, directly output the asphalt material category to complete the rapid asphalt material identification method. In this invention, the confidence decision module specifically calculates the maximum probability and its corresponding category, and performs a threshold judgment. The expressions for the maximum probability and its corresponding category are as follows: in, This represents the probability vector output by the model, and its value ranges from [0, 1]. Execute threshold judgment: In this invention, the pre-set confidence threshold is 0.75. This value is determined based on cross-validation and engineering requirements: at a threshold of 0.75, the model accuracy is 96.2%, and it can directly provide reliable judgments for 82% of the samples. At the same time, low-confidence samples (18%) are classified as "uncertain / requires retesting," which ensures both identification accuracy and on-site testing efficiency, meeting engineering quality control requirements.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention acquires asphalt infrared spectral data using an ATR-FTIR spectrometer, performs standardized preprocessing through polynomial fitting baseline correction, derivative transformation, and classification-oriented normalization, extracts multi-scale spectral features including time-domain statistics, spectral shape features, and PCA dimensionality reduction, constructs a fused feature vector, and establishes an identification model based on a random forest classifier. Finally, it combines a confidence decision module to output asphalt category and confidence assessment, achieving rapid, accurate, and quantifiable asphalt material identification. Attached Figure Description

[0028] Figure 1 This is a flowchart of a rapid identification method for asphalt materials based on infrared spectroscopy according to the present invention; Figure 2 This is a comparison of the original infrared spectra of the database portion of the present invention after treatment of the matrix asphalt base. Figure 3 This is the ANOVA F-score diagram showing the discriminative band distribution of the infrared spectra in the entire database (first derivative) of this invention. Figure 4 A comparison diagram of the functional group region contributions of this invention; Figure 5 This is a weighted importance distribution diagram of the infrared spectrum PCA in the database of this invention; Figure 6 Figure (a) shows the two-dimensional scatter plot of the first two principal components of the single-matrix asphalt fingerprint database of this invention, where Figure (b) is the complete database and Figure (a) is a partial database. Figure 7 Figure (a) shows the three-dimensional scatter plot of the first three principal components of the single-matrix bitumen fingerprint database of this invention, and Figure (b) shows the complete database plot. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0030] like Figure 1 As shown, a rapid identification method for asphalt materials based on infrared spectroscopy specifically includes the following steps: Example 1: S1. Based on the pre-defined categories of asphalt samples, infrared spectral data are collected using an attenuated total reflectance-Fourier transform infrared spectrometer (ATR-FTIR) to construct an infrared spectral fingerprint database. The preset categories specifically include base asphalt, SBS modified asphalt, rubber powder modified asphalt, and composite modified asphalt; the infrared spectral fingerprint database includes infrared spectral data and tag vectors; This invention employs an ATR-FTIR spectrometer, with specific parameters including a wavenumber range of 4000–600 cm⁻¹. -1 The resolution is set to 4 cm. -1 The number of scans was set to 32, and the temperature control range was 25±1℃. Each asphalt sample was collected 3 times, and the average spectrum was calculated to construct infrared spectral data, which was then labeled with category tags. (0 represents base asphalt, 1 represents SBS modified asphalt, 2 represents rubber powder modified asphalt, and 3 represents composite modified asphalt), construct a label vector; such as Figure 2 The image shows a comparison of the original infrared spectra of base asphalt from different sources after treatment. The key spectral bands (e.g., 2920 cm⁻¹) show the base asphalt from different sources. -1 2850 cm -1 Commonalities and differences among the nearby CH stretching vibration peaks.

[0031] The base asphalt (category label: 0) is an unmodified base asphalt derived from different crude oil types (cycloalkyl, intermediate, paraffinic); the SBS modified asphalt (category label: 1) uses styrene-butadiene-styrene block copolymer as a modifier and is classified by dosage as low dosage (SBS content < 3%), medium dosage (SBS content 3%-4%), and high dosage (SBS content > 4%); the rubber powder modified asphalt (category label: 2) uses waste tire rubber powder as a modifier and is classified by fineness as coarse powder (20-40 mesh), medium powder (40-60 mesh), and fine powder (60-80 mesh); in this embodiment, there are no fewer than 50 samples of base asphalt, covering at least 5 crude oil sources; SBS modified asphalt covers 3 dosage levels; and rubber powder modified asphalt has no fewer than 70 samples, covering 3 particle size specifications.

[0032] In this embodiment, the total sample data is 500, and the spectral data points are 850.

[0033] S2. Based on the infrared spectral fingerprint database, the preprocessed spectrum is output through preprocessing operations; The preprocessing operations specifically include polynomial fitting baseline correction, derivative transformation to enhance features, and classification-oriented normalization. In this invention, a polynomial fitting baseline correction method is used to determine the spectral intensity after baseline correction. The expression is as follows: In the formula, The original spectral intensity at wavenumber v, Let be a polynomial function of degree k, representing the estimated baseline. In this embodiment, considering both fitting ability and stability, k is set to 3. In this embodiment, the fitting interval for infrared spectral data is selected as 2500-2000 cm⁻¹. -1 and 1800-1700 cm -1 .

[0034] Subsequently, derivative transformation is performed to enhance the features, specifically including enhancing peak position differences with the first derivative and separating overlapping peaks with the second derivative to highlight hidden features. Finally, classification-guided normalization is performed using max-min normalization to preserve the relative peak intensity ratio, resulting in the preprocessed spectrum, as shown in the following expression: In the formula, is the spectral intensity value, and scale_factor is the dynamic weighting coefficient. In this embodiment, the dynamic weighting coefficient ranges from 0.8 to 1.5, and the characteristic peaks of the modifier are given high weights based on known chemical knowledge.

[0035] like Figure 3 As shown, ANOVA analysis reveals the discriminative band distribution (F-score) of the first derivative spectra of the entire database of this invention.

[0036] S3. Based on the preprocessed spectrum, after extracting time-domain statistical features and spectral shape features, feature fusion is performed to output a spliced ​​feature vector, which is then spliced ​​with the basic spectral feature vector to output a dimension-reduced fingerprint feature. The fundamental spectral feature vector is the original Each spectral intensity value, This represents the total number of spectral data points, which is 850 in this embodiment. Specifically, the following steps are included: S3.1 Input the preprocessed spectrum, extract time-domain statistical features, and output a time-domain statistical feature vector; This invention calculates (based on preprocessed spectra) using basic statistical characteristics. ) and extreme value characteristic calculation ( Output time-domain statistical feature vector There are 7 dimensions in total, among which, It is a characteristic of the mean. For variance characteristics, As a skewness feature, Kurtosis is a characteristic feature. The position of the maximum value. The position of the minimum value. The peak-to-valley difference; specifically includes the following steps: (1) The mean characteristic, representing the average intensity of the spectrum, reflects the overall absorbance level of the sample. Its calculation expression is as follows: In the formula, The total number of spectral data points. for index, For the spectral vector, in this embodiment, m=850. (2) The variance characteristic represents the degree of fluctuation in spectral intensity and reflects the complexity of the spectral shape. Its calculation expression is as follows: .

[0037] (3) Skewness characteristic, representing the asymmetry of spectral distribution, with positive values ​​indicating a right-side (high wavenumber) tail and negative values ​​indicating a left-side tail. The calculation expression is as follows: In the formula, The standard deviation is cubed.

[0038] (4) Kurtosis, representing the asymmetry of the spectral distribution, indicates a right-side (high wavenumber) tail and a left-side tail. The calculation expression is as follows: In the formula, The standard deviation is the fourth power.

[0039] (5) The location of the maximum value is calculated using the following expression: In the formula, For wavenumber vectors, This is the spectral vector.

[0040] (6) The minimum value is located, and the calculation expression is as follows: .

[0041] (7) The peak-to-valley difference is calculated using the following expression: .

[0042] This invention constructs a time-domain statistical feature vector through the above calculations, thereby completing the extraction of time-domain statistical features.

[0043] S3.2 Input the preprocessed spectrum, extract spectral shape features, and output the spectral shape feature vector; In this embodiment, the wavenumber range is set to 4000–600 cm⁻¹. -1 Therefore, the corresponding wavenumber vector middle, , ; This invention calculates based on the preprocessed spectrum and the characteristics of the spectral band shape. ) and local peak characteristic calculation ( Output spectral eigenvectors: 5 dimensions in total For the center of gravity of the score, For spectral width, The number of local peaks, This represents the average value of the symmetry index of the spectral peaks. The standard deviation of the symmetry index of spectral peaks; Specifically, the following steps are included: (1) Calculate the centroid of the spectral band using the first moment of the spectrum The calculation expression is as follows: In the formula, This represents the corresponding wave value.

[0044] (2) Calculate the spectral band width using the second-order central moments The calculation expression is as follows: .

[0045] (3) The derivative spectrum is obtained by taking the first derivative of the infrared spectral data. Search By identifying local maxima at a given point, the number of local peaks can be determined. ; The local maximum is specifically defined as if the derivative curve If the value is positive (curve rises) before the zero-crossing point and negative (curve falls) afterward, then this zero-crossing point corresponds to a local maximum in the infrared spectrum. (4) For each peak, calculate the left and right half-width ratios. Statistical calculations yielded and The calculation expression is as follows: In the formula, for The left half of the peak is wide. for The right half of the peak is wide. Index for the number of local peaks; The average value of the symmetry index of the spectral peaks The calculation expression is as follows: In the formula, The number of local peaks; Standard deviation of the symmetry index of spectral peaks The calculation expression is as follows: .

[0046] This invention constructs a spectral feature vector through the above calculations, thereby completing the spectral feature extraction.

[0047] S3.3 Input the time-domain statistical feature vector and the spectral feature vector, perform feature fusion through weighted feature concatenation, and output the concatenated feature vector; The expression for weighted feature concatenation in this invention is: Here, || represents a horizontal splicing operation. For time-domain statistical feature vectors, For spectral eigenvectors, To pre-determine the weights of the time-domain statistical features, Preset spectral feature weights.

[0048] This invention, based on feature engineering experience and previous feature importance analysis, weights time-domain statistical features. (Default is 1.0), weights for spectral features (Because it is highly correlated with local peak characteristics) an enhancement weight is assigned (1.2 in this embodiment).

[0049] S3.4 Input the spliced ​​feature vector and the basic spectral feature vector, perform feature splicing, and output the complete feature vector; The complete feature vector expression output by this invention is as follows: In the formula, The total number of spectral data points. The base spectral feature vector is denoted by ||, which represents the horizontal splicing operation.

[0050] S4. Based on the complete feature vector, the dimensionality-reduced fingerprint feature vector is output through principal component analysis (PCA) calculation and principal component selection. In this embodiment, all principal components are first calculated using PCA, and a cumulative variance contribution rate threshold (95%) is set to understand the overall information content of the data; such as Figure 4As shown, the contribution of functional group regions to the infrared spectra of different types of asphalt is analyzed; for example... Figure 5 As shown, 123 principal components (K_cum=123) are required to reach this threshold. Subsequently, based on feature importance analysis or model performance optimization, this embodiment selects the top 25 principal components (K_used=25) with the highest discriminative power for subsequent modeling. The standardized data is projected onto these 25 principal components to obtain the dimensionality-reduced features. K is the preset principal component value; like Figure 6 As shown, the distribution of all samples in the fingerprint database after dimensionality reduction by principal component analysis is a scatter plot in the two-dimensional space formed by the first two principal components (PC1 and PC2). The sample points of different categories (such as base asphalt, SBS modified asphalt, and rubber powder modified asphalt) in the figure show obvious clustering trends, which intuitively verifies the ability of the features extracted by this method to distinguish different types of asphalt materials. Figure 7 The distribution of samples in the three-dimensional space formed by the first three principal components (PC1, PC2, and PC3) is further demonstrated. This three-dimensional view reveals the intrinsic structure of the data more fully, and it can be observed that categories that may overlap in the two-dimensional plane are separated in the third dimension, further confirming the discriminative power and reliability of the feature space constructed by our method.

[0051] S5. Construct a dataset based on the dimensionality-reduced fingerprint feature vector and label vector, divide the dataset into a training set, a validation set and a test set, input the training set and validation set into the classification model for training, and obtain the asphalt material recognition model; The classification model uses a random forest classifier. In this embodiment, the specific parameter settings for the random forest classifier include: 300 trees, a maximum depth of 20, a minimum number of leaf node samples of 5, a feature sampling ratio of the default value, and a random seed of 42. For a medium-sized dataset with 25-dimensional features and 500 samples, 300 trees can ensure sufficient diversity. Experiments show that when the number of trees is greater than 300, the accuracy improvement is less than 0.5%, but the training time increases significantly. Balancing performance and computational efficiency, this method is suitable for rapid on-site identification needs. S6. Input the test set into the asphalt material recognition model. Through the confidence decision module, perform threshold judgment. When the confidence level... If the condition is met, proceed to the retest suggestion module; otherwise, directly output the asphalt material category to complete the rapid asphalt material identification method. In this invention, the confidence decision module specifically calculates the maximum probability and its corresponding category, as shown in the following expression: in, This represents the probability vector output by the model, and its value ranges from [0, 1]. Execute threshold judgment: In this invention, the pre-set confidence threshold is 0.75. This value is determined based on cross-validation and engineering requirements: at a threshold of 0.75, the model accuracy is 96.2%, and it can directly provide reliable judgments for 82% of the samples. At the same time, low-confidence samples (18%) are classified as "uncertain / requires retesting," which ensures both identification accuracy and on-site testing efficiency, meeting engineering quality control requirements.

[0052] In this embodiment, ATR-FTIR was selected as the spectral acquisition device, and the wavenumber range was uniformly set to 4000–600 cm⁻¹. -1 Resolution 4 cm -1 The process involved 32 scans. The asphalt sample was fully bonded to the ATR crystal, and spectra were acquired under constant contact pressure. The acquired spectra were processed using: rubber band / polynomial baseline correction → Savitzky-Golay smoothing → vector normalization → first / second derivative differentiation. The first K principal components of the PCA were extracted as combined features, and peak intensity, peak area, and peak intensity ratio of key peak intervals were superimposed. If necessary, time-domain and frequency-domain fusion features of the spectrum were extracted. These features were input into a trained classification model, which outputs the material category and confidence score. If the confidence score was below a threshold, a prompt was made to repeat sampling three times for consistency assessment or to increase the number of scans for retesting. This embodiment is suitable for on-site acceptance and rapid sampling inspection at construction sites.

[0053] Example 2 (Admixture Warning): The research steps are the same as in Example 1; A mixed sample set is added to the fingerprint database to train a primary classification model that outputs the category and confidence score. A dual-threshold discrimination strategy is set: when the confidence score is higher than 0.85, the sample is judged as pure; when the confidence score is between 0.75 and 0.85, the Euclidean distance between the spectrum to be tested and the typical spectrum of the corresponding category is further calculated. If the distance exceeds the threshold of 0.15, "suspected mixing" is output. This method achieves an accuracy of 88% for detecting mixing ratios of 10% or higher.

[0054] In the above embodiments, the fingerprint database can be continuously and incrementally updated: when new source materials or new admixture systems appear, new models are trained on the new categories of materials according to the above-mentioned spectral collection and training rules, thereby continuously improving the generalization ability and engineering applicability of the recognition system.

[0055] 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 rapid identification method for asphalt materials based on infrared spectroscopy, characterized in that, Includes the following steps: S1. Based on the pre-defined categories of asphalt samples, infrared spectral data are collected using an attenuated total reflectance-Fourier transform infrared spectrometer (ATR-FTIR) to construct an infrared spectral fingerprint database. The preset categories specifically include base asphalt, SBS modified asphalt, rubber powder modified asphalt, and composite modified asphalt; the infrared spectral fingerprint database includes infrared spectral data and tag vectors; S2. Based on the infrared spectral fingerprint database, the preprocessed spectrum is output through preprocessing operations; S3. Based on the preprocessed spectrum, after extracting time-domain statistical features and spectral shape features, feature fusion is performed to output a spliced ​​feature vector, which is then spliced ​​with the basic spectral feature vector to output a dimension-reduced fingerprint feature. The fundamental spectral feature vector is the original Each spectral intensity value, This represents the total number of spectral data points, specifically 850. S4. Based on the complete feature vector, the dimensionality-reduced fingerprint feature vector is output through principal component analysis (PCA) calculation and principal component selection. S5. Construct a dataset based on the dimensionality-reduced fingerprint feature vector and label vector, divide the dataset into a training set, a validation set and a test set, input the training set and validation set into the classification model for training, and obtain the asphalt material recognition model; The classification model uses a random forest classifier. S6. Input the test set into the asphalt material recognition model. Through the confidence decision module, perform threshold judgment. When the confidence level... If the condition is met, proceed to the retest suggestion module; otherwise, directly output the asphalt material category to complete the rapid asphalt material identification method.

2. The rapid identification method for asphalt materials based on infrared spectroscopy according to claim 1, characterized in that, In S2, the preprocessing operation specifically includes polynomial fitting baseline correction, derivative transformation to enhance features, and classification-oriented normalization. The expression for polynomial-fit baseline correction is as follows: In the formula, The original spectral intensity at wavenumber v, It is a polynomial function of degree k; The expression for classification-oriented normalization is as follows: In the formula, The value is the spectral intensity, and scale_factor is a preset range of dynamic weighting coefficients.

3. The rapid identification method for asphalt materials based on infrared spectroscopy according to claim 1, characterized in that, S3 specifically includes the following steps: S3.1 Input the preprocessed spectrum, extract time-domain statistical features, and output a time-domain statistical feature vector; S3.2 Input the preprocessed spectrum, extract spectral shape features, and output the spectral shape feature vector; S3.3 Input the time-domain statistical feature vector and the spectral feature vector, perform feature fusion through weighted feature concatenation, and output the concatenated feature vector; S3.4 Input the spliced ​​feature vector and the basic spectral feature vector, perform feature splicing, and output the complete feature vector.

4. The rapid identification method for asphalt materials based on infrared spectroscopy according to claim 3, characterized in that, In step S3.1, the time-domain statistical feature vector includes: mean feature, variance feature, skewness feature, kurtosis feature, maximum value location, minimum value location, and peak-to-valley difference, specifically including the following calculation formulas: Mean characteristics The calculation expression is as follows: In the formula, The total number of spectral data points. for index, It is a spectral vector; Variance characteristics The calculation expression is as follows: ； Skewness characteristics The calculation expression is as follows: In the formula, The standard deviation is cubed. Kurtosis characteristics The calculation expression is as follows: In the formula, The standard deviation is the fourth power. Maximum value position The calculation expression is as follows: In the formula, For wavenumber vectors, It is a spectral vector; Minimum position The calculation expression is as follows: Peak-valley difference The calculation expression is as follows: 。 5. The rapid identification method for asphalt materials based on infrared spectroscopy according to claim 3, characterized in that, In S3.2, the spectral feature vector includes: the centroid of the spectral band, the width of the spectral band, the number of local peaks, the average value of the symmetry index of the spectral peaks, and the standard deviation of the symmetry index of the spectral peaks.

6. The rapid identification method for asphalt materials based on infrared spectroscopy according to claim 1, characterized in that, In S3.3, the expression for weighted feature concatenation is: Here, || represents a horizontal splicing operation. For time-domain statistical feature vectors, For spectral eigenvectors, To pre-determine the weights of the time-domain statistical features, Preset spectral feature weights.

7. The rapid identification method for asphalt materials based on infrared spectroscopy according to claim 1, characterized in that, In S5, the specific parameter settings for the random forest classifier include: 300 trees, 20 maximum depths, 5 minimum leaf node samples, default feature sampling ratio, and 42 random seeds.

8. The rapid identification method for asphalt materials based on infrared spectroscopy according to claim 1, characterized in that, In step S6, the confidence decision module specifically calculates the maximum probability and its corresponding category, and performs a threshold judgment. The expressions for the maximum probability and its corresponding category are as follows: in, This represents the probability vector output by the model, and its value ranges from [0, 1]. Execute threshold judgment: 。

Images