Application of AI Techniques for Asphalt Concrete Mix Production Optimization

The application of AI in asphalt mix production is an actively developing field. Botella et al. [1] investigate ML techniques for estimating bitumen activity in recycled asphalt pavement. Additional research by Liu et al. [2] also uses ML to optimize the asphalt mix design by predicting the effective asphalt content and absorbed asphalt content.

Rahman et al. [3] apply ML techniques to predict metrics related to the performance of asphalt mixes. This study is complemented by Sebaaly et al. [4] who use ANN and genetic algorithms to optimize the design process of asphalt mixtures.

The application of AI in predicting asphalt concrete mix properties is investigated by Androjić and Marović [7] using ANN and MLR models. Mousa et al. [5] apply these models to estimate the optimum asphalt content from aggregate gradation.

Le et al. [6] propose an AI-based model for predicting the dynamic modulus of stone mastic asphalt mix. The work by Karballaeezadeh et al. [9] represents a further development of this field by presenting hybrid prediction models for intelligent road inspection.

The importance of applying AI when building sustainable highway and road systems is highlighted by Arifuzzaman et al. [10]. Hoang [8] presents an AI method for detecting potholes in asphalt pavement using the least-squares method (LSM) and NN.

AI plays an important role in predicting rutting and fatigue parameters in modified asphalt binders as shown by Uwanuakwa et al. [11]. Hosseinzadeh et al. [12] use ML algorithms to predict Marshall asphalt stability.

ANNs have also been used to control the quality of asphalt compaction as shown by the study [13].

Previous research primarily focuses on traditional methods or early-stage AI application, lacking in predictive accuracy and optimization capabilities for asphalt mix design. Studies have shown that reinforcement learning can be effective when solving route optimization problems, resource allocation, and managing infrastructure, such as energy grids and transportation systems.

Reinforcement learning requires a large amount of data for training, which can be challenging in real-world environments. However, modern ML techniques and advances in computer processing power make it possible to solve reinforcement learning problems efficiently, making it increasingly popular in various industries.

Thus, reinforcement learning is a promising area in the field of AI and can be effectively applied to optimize management in various industries.

The ANN and LS-SVM methods are the ML techniques used in the automated process control line to produce asphalt concrete mixes. Both methods are used to analyze asphalt quality and optimize production parameters.

The data were collected from an automated process control line at an asphalt concrete mix production facility. Historical data from the production line were utilized to train the AI models, namely ANN and LS-SVM. This training involved using the data to predict asphalt quality and control the production process effectively.

ANN is an ML technique that simulates the human brain using NNs. ANN is used to analyze asphalt quality data that is collected during production. ANN is trained on a dataset that is used to determine the relationships between input and output data. ANN can predict asphalt quality based on the available data allowing operators to quickly and accurately determine production parameters.

LS-SVM is an ML technique, used to solve regression and classification problems. It is based on SVM but uses LSM to solve the optimization problem. LS-SVM is used to optimize production parameters, such as temperature and mixing time to achieve optimum asphalt quality.

The ANN and LS-SVM operation process in the automated process control line to produce asphalt concrete mixes consists of the following stages:

·Data acquisition: asphalt quality data are collected during production.

·Model training: the dataset is used to train ANN and LS-SVM.

·Parameter prediction: based on the available data, ANN predicts the asphalt quality and LS-SVM optimizes the production parameters.

·Parameter adjustment: the automated process control line to produce asphalt concrete mixes adjusts production parameters based on predicted data and optimized parameters.

·Monitoring: the automated process control line to produce asphalt concrete mixes monitors the production process and asphalt quality data to allow operators to respond quickly to any deviations from required parameters. For example, if the asphalt quality does not meet established standards, the automated system can automatically adjust production parameters to achieve the required quality.

Data processing: using ANN and LS-SVM, the automated system can process asphalt quality and other production data to determine the optimal parameters for producing high-quality asphalt mixes.

Production management: based on data resulting from monitoring and data processing, the automated system can manage the production process to optimize asphalt quality and minimize raw material and energy losses.

Quality control: the automated process control line to produce asphalt concrete mixes also includes a quality control system, which can automatically check the asphalt quality at various production stages to ensure that the product meets established standards.

By leveraging advanced sensors, real-time data analysis, and automated adjustments, the system ensures that each batch of asphalt concrete mix is consistent with the last, adhering to the precise specifications required for infrastructure projects. This not only enhances the durability and performance of roads and highways but also contributes to safer and more reliable transportation networks. Furthermore, the ability to monitor and adjust the production process in real time significantly reduces waste, improves efficiency, and minimizes environmental impact, demonstrating a forward-thinking approach to sustainable construction practices.

The mathematical model of the automated process control line to produce asphalt concrete mixes is trained on a set of previously obtained data and based on these data predicts asphalt quality and optimizes production parameters to maximize quality. Operators can monitor the production process using the data provided to the automated process control line to produce asphalt concrete mixes and make decisions based on the model predictions.

The model is implemented using an NN, trained on a large amount of asphalt production data, as well as asphalt quality data, resulting from laboratory testing. An NN represents a deep learning model, consisting of many layers of neurons connected by weighted links.

Input data to the model may include parameters, such as temperature, moisture, mixing time, number of components used, etc. Based on this data, the model predicts the quality of the produced asphalt mix.

Deep learning techniques, such as the backpropagation algorithm (or backward propagation of errors), are used to train an NN. During the training process, the model is adjusted to the available data and set to the optimal values of weights and biases.

Moreover, the model can be augmented with SVM or other ML algorithms to improve the accuracy of predictions.

Thus, the mathematical model of an automated process control line to produce asphalt concrete mixes using AI allows for increasing the efficiency of asphalt production, improving the quality of manufactured product, and reducing production costs.

The end result of the automated process control line to produce asphalt concrete mixes can be represented as a matrix (Table 1), where each row corresponds to a certain batch of mix produced, while columns show mix characteristics, such as temperature, density, viscosity, etc.

Table 1. Asphalt mix production data matrix

Various ML techniques, such as ANN and LS-SVM, are used to process the data and build the mathematical model.

ANN uses NNs to process and analyze data. NNs consist of many interconnected nodes called neurons, which accept input data and process them to produce output data [14-17]. ANN can be trained based on the available data to create a mathematical model, which can be used to predict the performance characteristics of the asphalt concrete mix.

LS-SVM uses SVM to build the mathematical model. SVM consists of finding a hyperplane, which separates data by classes to the utmost. In the case of asphalt mix quality analysis, LS-SVM can be trained based on the available data to create a mathematical model that can be used to predict mix performance characteristics.

Both ANN and LS-SVM models are used in the automated process control line to produce asphalt concrete mixes to predict asphalt quality characteristics based on data incoming in the real-time production process.

The mathematical model of ANN is based on ANNs, which are multilayer structures consisting of interconnected neurons. ANN uses data on previously received asphalt quality characteristics and other production parameters to train the NN for creating predictions. The formula for the direct propagation of the input signal through the NN layers is as follows [18]:

$a^{(l)}=f\left(z^{(l)}\right)=f\left(w^{(l)} a^{(l-1)}+b^{(l)}\right)$      (1)

where, l is the layer number, a^(l) is the activation vector of l layer, z^(l) is the weighted input signal of l layer, w^(l) is the matrix of weights between l–1 layers, b^(l) is the displacement vector of l layer, and f is the activation function. A backpropagation algorithm is used to update the weights, which minimizes the loss function and adjusts the weights of the NN to achieve the best prediction.

The LS-SVM mathematical model is based on SVM, which uses a training data set to construct a hyperplane separating different data classes. LS-SVM is an extension of SVM, which uses regularization to reduce noise and improve the generalization performance of the model. LS-SVM minimizes the error function based on the training data set, which is as follows [19, 20]:

$\min _{\mathrm{w}, \mathrm{b}, \xi} \frac{1}{2} w^T w+C \sum_{\mathrm{i}=1}^{\mathrm{n}} \xi_{\mathrm{i}}$       (2)

where, w is the weight vector, b is the bias, ξ is the error, C is the regularization parameter, and n is the amount of training data.

ANN uses multiple processing of data through several hidden layers, which allows accounting for the complicated relationships between different factors and improving the prediction accuracy. The formula for calculating the NN output in the case of multilayer architecture is as follows:

$y_k=f\left(\sum_{j=1}^{N_h} w_{k j}^{(2)} f\left(\sum_{i=1}^{N i n} w_{j i}^{(1)} x_i+w_{j 0}^{(1)}\right)+w_{k 0}^{(2)}\right)$     (3)

where, $y_k$ is the output signal of the $k$-th neuron, $N_h$ is the number of neurons in the hidden layer, $w_{k j}^{(2)}$ is the weight of the synapse between the $j$-th neuron of the hidden layer and the $k$-th neuron of the output layer, $f$ is the activation function, $N_{i n}$ is the number of inputs, $w_{j i}^{(1)}$ is the weight of the synapses between the $i$-th input and the $j$-th neuron of the hidden layer, $w_{j 0}^{(1)}$ is the bias of the $j$-th neuron of the hidden layer, $w_{k 0}^{(2)}$ is the bias of the $k$-th neuron of the output layer.

On the other hand, LS-SVM uses SVM, which is a special case of NNs and can be implemented as a single-layer NN without hidden layers. The formula for calculating the LS-SVM output is as follows:

$y_k=\sum_{i=1}^n \alpha_i y_i K\left(x_i, x_k\right)+b$      (4)

where, y_k is the output signal; n is the number of training examples; α_i is the weight assigned to the i-th training example; y_i is the target value of the i-th training example; K(x_i, x_k) is the kernel, which defines the function mapping of the input data into a higher-dimensional space, where classification or regression is easier to perform. Typically, Gaussian or polynomial kernels are used for this purpose.

To train ANN and LS-SVM, it is necessary to have a training data set, which consists of the input parameters of the asphalt mix production process and corresponding values of the asphalt quality, which is target variable.

Once the models are trained based on the training dataset, they can be used to predict the asphalt quality based on the current input parameters of the production process. The prediction results can be displayed on the operator's screen or sent to the process control system to adjust the production parameters automatically.

A kernel function is used to transform the data into a higher-dimensional space, where the data become linearly separable. The following kernel functions are commonly used:

·Linear kernel: K(x, y)=x * y

·Polynomial kernel: K(x, y)=gamma*x*y+coef0)^degree

·Radial basis kernel (RBF): K(x, y)=exp (-gamma*||x-y||²)

Here gamma, coef0, and degree are parameters, which can be adjusted for optimal model performance.

Table 2. Comparison between ANN and LS-SVM

In general, the choice between ANN and LS-SVM depends on the specific task and the requirements to performance and interpretability of the results (Table 2). The following optimization algorithm is used to build the LS-SVM model, subject to conditions:

y_i-(w^Tφ(x_i)+b)≤ϵ+ξ_i

(w^Tφ(x_i)+b)-y_i≤ϵ+ξ_i

ξ_i≥0, i=1, …, n

where, w and b are the model parameters, φ(x_i) is the attribute vector, ξ_i are non-negative variables, ϵ is the allowable error, C is the regularization factor, and p is the degree of softness, which determines how heavily unimplemented constraints are weighted.

Both models, ANN and LS-SVM, can be used to predict asphalt quality, based on data obtained in the course of production. Using these models allows automating the quality control process and optimizing production processes, resulting in improved quality and resource savings.

The coefficients in these models can be determined using optimization techniques, such as the LSM or the gradient descent method. In the LSM for LS-SVM, it is necessary to find the solution of the system of equations:

$K \alpha=y$    (5)

where, K is the kernel function matrix, α is the coefficient vector, and y is the target value vector.

To train ANN, one can use the backpropagation algorithm. In this method, the network weights are updated by gradient descent method using the formula:

$w_{i, j}=w_{i, j}-\alpha \frac{\partial L}{\partial w_{i, j}}$    (6)

where, w_i,j is the weight between the i-th input node and the j-th hidden node, α is the training speed, and L is the loss function, defining the difference between the network output and the target values.

Thus, using AI-based mathematical models in the automated process control line to produce asphalt concrete mixes allows conducting automatic analysis of asphalt quality data and optimizing the production process, which contributes to improved product quality, as well as time and resources savings.

The conducted study showed that using ML techniques allows more accurately determining asphalt quality and optimal parameters for asphalt production compared to conventional analysis methods.

The following results were obtained in the conducted research:

·A mathematical model of the automated process control line to produce asphalt concrete mixes using two AI-based ML techniques, ANN and SVM with linear kernel (LS-SVM), was developed and implemented.

·The quality of asphalt produced using the model-based quality control technique was investigated. The results showed that using the automated process control line to produce asphalt concrete mixes significantly improved the quality of produced asphalt.

·Data collected over several months of production line operation were used to train the model. These data were divided into training and validation samples, and the optimal model parameters were derived from the training sample.

·Using ANN and LS-SVM allowed achieving high accuracy in predicting asphalt quality.

·The developed model can be used to optimize the production of asphalt concrete mixes, which allows reducing the production time and improving the quality of the produced asphalt.

·The results obtained can be used by manufacturing companies engaged in the production of asphalt concrete mixes to improve product quality and increase production efficiency.

·The conducted study showed the scientific novelty of using SVM with linear kernel (LS-SVM) in the automated process control line to produce asphalt concrete mixes.

Comparing the obtained results with other recent research, several significant aspects were revealed. Botella et al. [1] investigate application of ML techniques to estimate binder activity in reclaimed asphalt crumb. Their results also confirm the benefits of ML techniques in analyzing asphalt materials.

Liu et al. [2] also examine asphalt mixture optimization by predicting the effective asphalt content and absorbed asphalt content using ML techniques. The results of the present study support their findings on the significance of using ML techniques in the asphalt mix design process.

Thus, the results of the present study showed that ML techniques can be effectively used to analyze the quality of asphalt and improve the quality of asphalt pavement. This opens new opportunities for improving road infrastructure and safety.

The proposed approach includes the following main stages:

(1) Data collection: this stage involves collecting data on asphalt mix production, including information on mix composition, production parameters, material quality, and other relevant variables. Various data sources can be used for this purpose, including laboratory tests, sensors, and production monitoring systems.

(2) Data preprocessing: this stage involves preprocessing and scrubbing of the collected data. This may include removing outliers, filling in missing values, normalizing the data, and converting variables into a format convenient to process.

(3) Model selection and training: this stage involves selecting and training an ML technique used to analyze the data and predict the quality parameters of asphalt concrete mixtures. Different models can be used, such as ANN, LS-SVM, or genetic algorithms. The model is trained based on the preliminary prepared data. The model is adjusted based on the known relationships between the input data and the target quality parameters.

(4) Model validation and testing: this stage involves assessment of model performance using selected validation data. The model outcomes are compared with known target quality parameters to analyze accuracy, efficiency, and stability of the model.

(5) Integration into an automated process line: at this stage, the model is integrated into the automated process control line to produce asphalt concrete mixes. The model can be used to make real-time decisions, manage and optimize production processes, and provide recommendations and predictions to operators.

The overall structure of the proposed approach includes the data acquisition, data preprocessing, model selection and training, model validation and testing, and model integration into process control line. This provides a comprehensive approach to analyzing data and optimizing production processes for quality results.

Analyzing asphalt quality data using ML techniques allowed us identifying hidden patterns in the data which were not detected using statistical methods. This allowed determining the effect of each parameter on asphalt quality more accurately and highlighting the most significant factors.

Determining the optimal parameters for producing quality asphalt mixes also showed that ML techniques can produce more accurate results than traditional optimization methods, such as least-squares.

Comparing the obtained results with existing methods for analyzing asphalt quality showed that ML techniques can produce more accurate results, especially when analyzing big data. Moreover, using ML techniques can significantly reduce the time needed for analysis and improve the accuracy of the results.

However, using ML techniques requires a large amount of data and skilled specialists to process data and interpret the results. Models need to be carefully validated and tested to avoid errors and flaws in the outcomes.

As for the practical implementations, the adoption of LS-SVM technology can streamline production processes, allowing for real-time adjustments that cater to varying environmental conditions and material properties. This level of adaptability ensures that the final product not only meets but exceeds quality standards, potentially extending the lifespan of road infrastructure and reducing the need for frequent repairs.