阅读说明：本技术 一种基于机器学习的沥青动态黏弹特性预测方法 (Asphalt dynamic viscoelastic characteristic prediction method based on machine learning ) 是由 单丽岩 王亚杰 杨金龙 王建杰 于 2021-08-24 设计创作，主要内容包括：本发明公开了一种基于机器学习的沥青动态黏弹特性预测方法,所述方法采用动态剪切流变仪测试了样本的动态黏弹特性,采用元素分析仪和红外光谱仪测试了样本的微观组成结构；运用主成分分析法对微观组成结构参数进行了特征提取；通过距离相关系数法确定了动态黏弹特性参数与微观组成结构参数间的映射关系；建立了基于支持向量机机器学习算法的沥青动态黏弹特性预测模型。该方法揭示了沥青材料动态黏弹特性与微观组成结构间的对应关系,有利于快速获取沥青的动态黏弹力学响应,为沥青材料的设计生产提供了基础理论支撑,对于指导沥青及其他材料的按需设计具有积极意义。(The invention discloses a method for predicting asphalt dynamic viscoelastic characteristics based on machine learning, which adopts a dynamic shear rheometer to test the dynamic viscoelastic characteristics of a sample, and adopts an element analyzer and an infrared spectrometer to test the microstructure of the sample; extracting the characteristics of the microstructure parameters by using a principal component analysis method; determining a mapping relation between the dynamic viscoelastic characteristic parameters and the microconstituent structure parameters by a distance correlation coefficient method; and establishing an asphalt dynamic viscoelastic characteristic prediction model based on a support vector machine learning algorithm. The method reveals the corresponding relation between the dynamic viscoelastic property and the microcosmic composition structure of the asphalt material, is beneficial to rapidly acquiring the dynamic viscoelastic mechanical response of the asphalt, provides basic theoretical support for the design and production of the asphalt material, and has positive significance for guiding the design of the asphalt and other materials according to requirements.)

1. A pitch dynamic viscoelastic characteristic prediction method based on machine learning is characterized by comprising the following steps:

the method comprises the following steps: testing different asphalt samples by adopting a dynamic shear rheometer, an element analyzer and an infrared spectrometer, establishing a dynamic modulus main curve based on a CAM model, and obtaining the element content and an infrared spectrum curve of the samples;

step two: selecting a dynamic modulus main curve shift factor parameter C of each sample₁、C₂Shape parameters k, m_eGlass state dynamic modulus G_gAnd a crossover frequency f_cConstructing a macroscopic characteristic parameter data set, selecting C, H, N, S, O element content, C/H and C/N of each sample and infrared spectrum characteristic peak height and peak area to construct a microscopic characteristic parameter data set;

step three: carrying out dimensionless processing on the data set data by using a mean method, and calculating the variance of the data after the mean processing;

step four: calculating the Pearson correlation coefficient of the microscopic characteristic parameters obtained in the third step, reducing the dimension of the selected microscopic characteristic parameters with larger Pearson correlation coefficient based on principal component analysis, and replacing the selected microscopic characteristic parameters with fewer new characteristic parameters V₁、V₂…；

Step five: calculating a distance correlation coefficient between the dynamic modulus main curve fitting parameter selected in the step two and the microcosmic characteristic parameter processed in the step four, determining the microcosmic characteristic parameter with larger correlation with the dynamic modulus main curve fitting parameter as a model input variable, and establishing a new macro-microcosmic data set in one-to-one correspondence with the microcosmic characteristic parameter;

step six: dividing the data set constructed in the fifth step into a training set and a testing set according to the proportion of 9:1, selecting a proper kernel function type to construct an asphalt dynamic viscoelastic characteristic prediction model based on a support vector machine, training the asphalt dynamic viscoelastic characteristic prediction model based on the data of the training set, and optimizing model parameters by means of a grid search method and a cross validation method;

step seven: and (4) testing the trained asphalt dynamic viscoelastic characteristic prediction model in the sixth step based on the test set data, and testing the accuracy of the asphalt dynamic viscoelastic characteristic prediction model by using the average absolute percentage error.

2. The method of claim 1, wherein in the first step, the different asphalt samples tested comprise: asphalt with different labels, asphalt with different oil sources, asphalt mixed with different modifiers and asphalt subjected to aging or not.

3. The asphalt dynamic viscoelastic property prediction method based on machine learning as claimed in claim 1, wherein in the third step, the formula of the averaging method is as follows:

wherein, x'_iDe-dimensioned value, x, for a characteristic parameter_iIn order to be the original value of the characteristic parameter,the characteristic parameter is the mean value under all samples.

4. The asphalt dynamic viscoelastic characteristic prediction method based on machine learning as claimed in claim 1, wherein in the sixth step, the support vector machine maps variables x, y of the input space into a high-dimensional feature space, and finds the non-linear relationship between the variables in the high-dimensional feature space, so as to find an optimal function f (x), where f (x) has the following formula:

wherein x is_iX is a kernel function,α_iand b is a model parameter.

5. The machine-learning-based asphalt dynamic viscoelastic property prediction method according to claim 4, wherein the kernel function type comprises: linear kernel function, polynomial kernel function, gaussian kernel function, sigmoid kernel function.

6. The asphalt dynamic viscoelastic property prediction method based on machine learning as claimed in claim 1, wherein in the seventh step, the formula of the average absolute percentage error is:

wherein, y_iTo test the value, f (x)_i) Is a predicted value.

Technical Field

The invention belongs to the field of asphalt material performance prediction, and relates to a machine learning-based asphalt dynamic viscoelastic characteristic prediction method.

Background

The viscoelastic property is the essential attribute of the asphalt material, the mechanical behavior of the viscoelastic material has direct influence on road performance, the mechanical response rule of the asphalt is researched and analyzed by adopting the viscoelastic theory, and positive reference is provided for solving the problems existing in the evaluation and use of the asphalt and asphalt mixture. The dynamic modulus is one of the key characteristics reflecting the dynamic viscoelastic characteristics, which are closely related to the damages such as permanent deformation, cracks and rutting of the pavement, and the key point of the structural design system of the asphalt pavement changing from static state to dynamic state. The microstructure of the asphalt has important influence on the dynamic viscoelastic property, and the research on the correlation between the microstructure and the macroscopic performance is carried out, which is a key step in the research and development process of new materials.

The establishment of the relationship between the dynamic viscoelastic property of the asphalt and the microstructure is the core and key for predicting the viscoelastic property of the asphalt and realizing the rheological design of the asphalt. At present, the relationship between the viscoelastic property of asphalt and the microstructure of the micro-composition is mainly divided into a multivariable fitting functional relationship between the composition and the rheology and a micro-microscopic-macroscopic multi-scale corresponding relationship. However, due to the complexity of the asphalt microstructure, the existing relational models mainly have the following problems: (1) the performance of model prediction is single; (2) changes in materials, environment, etc. can severely reduce the accuracy of the model; (3) the multi-scale corresponding relation model is extremely complex in establishing process, difficult to popularize and apply and the like.

Compared with the traditional statistical calculation method, machine learning has the characteristics of high prediction precision, high operation speed and the like, is gradually a powerful research tool for workers in the scientific research field, and has the core that unknown data is predicted by learning data, analyzing the data and training a model by using an algorithm. At present, machine learning is widely applied to the construction of material macroscopic performance quantitative structure-activity relationship, and is vigorously developed in the fields of biology, chemistry, automation and the like. In recent years, machine learning methods are also gradually applied to the road field, and establishing an asphalt dynamic viscoelastic characteristic prediction model based on a machine learning theory is the most effective way for establishing the relationship between the asphalt micro-composition structure and the viscoelastic performance at present.

Disclosure of Invention

The invention provides a machine learning-based asphalt dynamic viscoelastic characteristic prediction method, aiming at the problem that correlation is difficult to establish due to complex macro-micro properties of asphalt. The method reveals the corresponding relation between the dynamic viscoelastic property and the microcosmic composition structure of the asphalt material, is beneficial to rapidly acquiring the dynamic viscoelastic mechanical response of the asphalt, provides basic theoretical support for the design and production of the asphalt material, and has positive significance for guiding the design of the asphalt and other materials according to requirements.

The purpose of the invention is realized by the following technical scheme:

a pitch dynamic viscoelastic characteristic prediction method based on machine learning comprises the following steps:

the method comprises the following steps: testing different asphalt samples by adopting a dynamic shear rheometer, an element analyzer and an infrared spectrometer, establishing a dynamic modulus main curve based on a CAM model, and obtaining the element content and an infrared spectrum curve of the samples;

step two: selecting a dynamic modulus main curve shift factor parameter C of each sample₁、C₂Shape parameters k, m_eGlass state dynamic modulus G_gAnd a crossover frequency f_cConstructing a macroscopic characteristic parameter data set, selecting C, H, N, S, O element content, C/H and C/N of each sample and infrared spectrum characteristic peak height and peak area to construct a microscopic characteristic parameter data set;

step three: carrying out dimensionless processing on the data set data by using a mean method, and calculating the variance of the data after the mean processing;

step four: calculating the Pearson correlation coefficient of the microscopic characteristic parameters obtained in the third step, reducing the dimension of the selected microscopic characteristic parameters with larger Pearson correlation coefficient based on principal component analysis, and replacing the selected microscopic characteristic parameters with fewer new characteristic parameters V₁、V₂…；

Step five: calculating a distance correlation coefficient between the dynamic modulus main curve fitting parameter selected in the step two and the microcosmic characteristic parameter processed in the step four, determining the microcosmic characteristic parameter with larger correlation with the dynamic modulus main curve fitting parameter as a model input variable, and establishing a new macro-microcosmic data set in one-to-one correspondence with the microcosmic characteristic parameter;

step six: dividing the data set constructed in the fifth step into a training set and a testing set according to the proportion of 9:1, selecting a proper kernel function type to construct an asphalt dynamic viscoelastic characteristic prediction model based on a support vector machine, training the asphalt dynamic viscoelastic characteristic prediction model based on the data of the training set, and optimizing model parameters by means of a grid search method and a cross validation method;

step seven: and (5) testing the prediction model of the dynamic viscoelastic characteristics of the asphalt trained in the step six based on the test set data, and testing the accuracy of the prediction model by Mean Absolute Percentage Error (MAPE).

Compared with the prior art, the invention has the following advantages:

the invention establishes the correlation between the dynamic viscoelastic property of the asphalt material and the microcosmic composition structure, realizes the purpose of predicting the macroscopic dynamic modulus through the microscopic characteristic parameters of the asphalt, enriches the evaluation method of the dynamic modulus of the asphalt, and provides technical support for realizing the design of the asphalt material according to the requirements.

Drawings

FIG. 1 is a flowchart illustrating a method for predicting dynamic viscoelastic properties of asphalt based on machine learning according to an embodiment;

FIG. 2 is a diagram illustrating the dynamic viscoelastic characteristics prediction results in an embodiment.

Detailed Description

The technical solution of the present invention is further described below with reference to the accompanying drawings, but not limited thereto, and any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention shall be covered by the protection scope of the present invention.

The invention provides a method for predicting asphalt dynamic viscoelastic characteristics based on machine learning, which adopts a dynamic shear rheometer to test the dynamic viscoelastic characteristics of a sample, and adopts an element analyzer and an infrared spectrometer to test the microstructure of the sample; extracting the characteristics of the microstructure parameters by using a principal component analysis method; determining a mapping relation between the dynamic viscoelastic characteristic parameters and the microconstituent structure parameters by a distance correlation coefficient method; and establishing an asphalt dynamic viscoelastic characteristic prediction model based on a support vector machine learning algorithm. The method comprises the following specific steps:

the method comprises the following steps: and testing different asphalt samples by adopting a dynamic shear rheometer, an element analyzer and an infrared spectrometer, establishing a dynamic modulus main curve based on a CAM model, and obtaining the element content and the infrared spectrum curve of the samples.

In this step, the different asphalt samples tested included: asphalt with different labels, asphalt with different oil sources, asphalt mixed with different modifiers, asphalt subjected to aging or not and the like.

Step two: selecting a dynamic modulus main curve shift factor parameter C of each sample₁、C₂Shape parameters k, m_eGlass state dynamic modulus G_gAnd a crossover frequency f_cAnd (3) constructing a macroscopic characteristic parameter data set, and selecting C, H, N, S, O element content, C/H and C/N of each sample and infrared spectrum characteristic peak height and peak area to construct a microscopic characteristic parameter data set.

Step three: and carrying out dimensionless processing on the data set data by using a mean method so as to solve the problem that the data set data does not have comparability due to different dimensions, calculating the variance of the data after the mean processing, and removing the micro characteristic parameters with smaller variance.

In this step, the formula of the averaging method is as follows:

wherein, x'_iDe-dimensioned value, x, for a characteristic parameter_iIn order to be the original value of the characteristic parameter,the characteristic parameter is the mean value under all samples.

Step four: calculating the Pearson correlation coefficient and the base of the microscopic characteristic parameters obtained in the third stepReducing the dimension of the selected microscopic characteristic parameter with larger Pearson correlation coefficient in principal component analysis, and replacing the microscopic characteristic parameter with fewer new characteristic parameters V₁、V₂…。

In this step, the Pearson correlation coefficient is used to measure the correlation between the microscopic characteristic parameters, that is, redundancy exists between the characteristics with greater correlation, and the characteristic parameters with greater correlation are processed by using principal component analysis to obtain fewer new parameters, so that the purposes of removing data redundancy and reducing dimensionality can be achieved.

Step five: and (4) calculating a distance correlation coefficient between the dynamic modulus main curve fitting parameter selected in the step two and the microcosmic characteristic parameter processed in the step four, determining the microcosmic characteristic parameter with larger correlation with the dynamic modulus main curve fitting parameter as a model input variable, and establishing a new macro-microcosmic data set in one-to-one correspondence with the microcosmic characteristic parameter.

Step six: dividing the data set constructed in the fifth step into a training set and a testing set according to the proportion of 9:1, selecting a proper kernel function type to construct an asphalt dynamic viscoelastic characteristic prediction model based on a Support Vector Machine (SVM), training the asphalt dynamic viscoelastic characteristic prediction model based on the training set data, and optimizing model parameters by means of a grid search method and a cross validation method.

In this step, the Support Vector Machine (SVM) may map the variables x, y of the input space into a high-dimensional feature space, and find the nonlinear relationship between the variables in the high-dimensional feature space, so as to find an optimal function f (x), where the formula of f (x) is as follows:

wherein x is_iX is a kernel function, the types of which include: linear kernel functions, polynomial kernel functions, gaussian kernel functions, sigmoid kernel functions,α_iand b is a model parameter.

Step seven: and (5) testing the prediction model of the dynamic viscoelastic characteristics of the asphalt trained in the step six based on the test set data, and testing the accuracy of the prediction model by Mean Absolute Percentage Error (MAPE).

In this step, the formula of the Mean Absolute Percentage Error (MAPE) is:

wherein, y_iTo test the value, f (x)_i) Is a predicted value.

Example (b):

the embodiment provides a method for predicting dynamic viscoelastic characteristics of asphalt based on machine learning, as shown in fig. 1, the specific implementation steps are as follows:

the method comprises the following steps: frequency scanning is carried out on 50 asphalt samples in the table 1 by adopting a DHR-2 type dynamic shear rheometer, the change condition of the dynamic modulus of the material along with the frequency is measured, and a dynamic modulus main curve with the reference temperature of 20 ℃ is established according to a time-temperature equivalent principle and a CAM model theory; collecting asphalt C, H, N, S element data by using a Vario EL Cube element analyzer, and acquiring O element data by using a difference method; and acquiring infrared data of the asphalt by using a Nicolet iS5 type infrared spectrometer.

TABLE 1 asphalt sample summary sheet

Note: in the asphalt numbers, L represents low grade, B represents binder, M represents modified, and A represents aging.

Step two: selecting 6 dynamic modulus main curve parameters C of all samples obtained in the step one₁、C₂、k、m_e、G_g、f_cConstructing a macroscopic characteristic parameter data set, and selecting C, H, N, S, O element content, C/H and C/N and infrared of all samplesAnd (4) constructing a micro characteristic parameter data set by using 31 parameters of the spectral characteristic peak height and the peak area. The characteristic peaks of the selected infrared spectra are summarized in Table 2, wherein I represents the peak height of each characteristic peak, and A represents the peak area of each characteristic peak.

TABLE 2 initial selection of characteristic peak height and peak area characteristic parameters

Step three: the data set data are subjected to dimensionless processing by using a mean-value method so as to solve the problem that the data are not comparable due to different dimensions, the variance of the data subjected to the mean-value processing is calculated, the variance of the C element content data is found to be 0.0002, the distinguishing degree of the sample is extremely weak, and therefore the influence of the microscopic characteristic parameter is ignored in the subsequent model establishment.

Step four: calculating the Pearson correlation coefficient of the microscopic characteristic parameters obtained in the third step, extracting the microscopic characteristic parameters with larger Pearson correlation coefficient, summarizing the microscopic characteristic parameters as shown in the table 3, reducing the dimension of the microscopic characteristic parameters as shown in the table 3 based on principal component analysis, and replacing the microscopic characteristic parameters with new characteristic parameters V in the table 4₁、V₂At this time, the microscopic characteristic parameter was reduced to 25.

TABLE 3 microscopic characteristic parameter correlation coefficient table with large correlation

TABLE 4V₁、V₂Data sheet

Step five: calculating the distance correlation coefficient between the dynamic modulus main curve characteristic parameter selected in the second step and the microscopic characteristic parameter processed in the fourth step, and analyzing the result as shown in Table 5 to obtain the microscopic characteristic parameter A₁₆₀₀、A₉₆₆、A₈₇₀、A₇₄₅、I₁₆₀₀、I₇₄₅、I₉₆₆And the correlation coefficient between the C/N and some dynamic modulus main curve characteristic parameters is larger, so that the dynamic modulus main curve is well predicted, and the 8 parameters are selected as model input variables from subsequent machine learning modeling data. Final data set collectively containing C₁、C₂、k、m_e、G_g、f_c6 dynamic modulus principal curve characteristic parameters, A₁₆₀₀、A₉₆₆、A₈₇₀、A₇₄₅、I₁₆₀₀、I₇₄₅、I₉₆₆And C/N8 microstructural characteristic parameters.

TABLE 5 distance correlation coefficient table

Note: a distance correlation coefficient above 0.4 indicates a greater correlation between the two features, and data with a distance correlation coefficient greater than 0.4 has been bolded in the table.

Step six: dividing the data set obtained in the fifth step into a training set and a testing set according to the proportion of 9:1, selecting a Gaussian kernel function to build an asphalt dynamic viscoelastic characteristic prediction model based on a support vector machine, training the asphalt dynamic viscoelastic characteristic prediction model based on the data of the training set, and optimizing model parameters by means of a grid search method and a cross validation method.

Step seven: based on the asphalt dynamic viscoelastic property prediction model trained in the test set data inspection step six, the average absolute percentage error (MAPE) evaluation result is shown in Table 6, and a partial asphalt dynamic modulus main curve is drawn according to the model prediction result and compared with the measured value (shown in FIG. 2).

TABLE 6 evaluation results of dynamic viscoelastic behavior prediction model

In conclusion, the asphalt dynamic viscoelastic property prediction method based on the invention provides main microcosmic structural indexes which have obvious influence on asphalt viscoelastic property, establishes the corresponding relation of asphalt macro-microcosmic characteristics, and simultaneously, MAPE of predicted dynamic modulus main curve parameters is less than 10%, so that the drawn dynamic modulus main curve is well fitted with measured values. The invention provides a new means for evaluating the dynamic viscoelastic property of the asphalt and provides a new idea for designing, producing and even researching asphalt materials.