Application of Machine Learning Techniques for Asphalt Pavement Performance Prediction

Traffic (AADT). According to the statistical evaluation results, all the ML models exhibited excellent prediction capabilities, as evidenced by their high coefficient of determination (R^2) values of 96.8%,96.6%,97.1%, and 97.4% and low Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Square Error values of 1.888%, 1.874%,1.830, and 1.556%, and 2.529%,2.613%,2.391%, and 2.545% and 6.348%,6.828%,5.716%, and 5.081% and 9.98%, respectively. Furthermore, the results indicate that the ML models demonstrated superior prediction accuracy compared to the (MLR) models developed under the same data.


Introduction
The Pavement Management System (PMS) is a system used to control, assess, and monitor pavements designed to minimize maintenance costs, reduce environmental impacts, and provide long-term performance [1].PMS utilizes a variety of data sources, such as traffic volume, weather, and pavement condition to track pavement performance over time [2][3][4].The most widely used techniques in the PMS are the Present Serviceability Index (PSI), International Roughness Index (IRI), and Pavement Condition Index (PCI).The PSI is a measure of pavement condition and is defined as the difference between a pavement's present serviceability and its initial serviceability [5].The PCI is a measure of a pavement's overall condition and is derived by combining the PSI and IRI.However, each pavement condition rating system is helpful for different purposes [6].For instance, the PSI is useful for determining pavement maintenance needs, and the IRI is useful for assessing ride quality and safety [7,8].
In recent years, there has been a noteworthy growth in the use of ML methods for predicting pavement performance [9].ML algorithms are well suited for this task, as they can effectively capture the complex interactions between pavement properties and performance indicators [10].Ali et al. offered a technique for evaluating the pavement performance of 19 roads in St. John's, Newfoundland, Canada, where the PCI and IRI were the main indicators in characterizing the overall pavement performance of asphalt pavement [11].Sagheer et al. developed a knowledge-based technique for pavement distress categorization using logic programming and the Prolog language to assess distress in flexible pavements [12].Relatively few studies have been conducted in recent years to predict the PCI of flexible pavements using ML approaches [13,14].
In the literature, several studies have highlighted the importance of assessment programs that include PCI testing and distress surveys to determine the structural conditions of the pavement [15][16][17].Nowadays, the assessment of pavement performance using PCI is a fundamental component of any PMS, and ML models have been used to predict pavement performance.In 2010, Bianchini and Bandini established a model to predict pavement performance using neuro-fuzzy.Hence, the outcomes and precision of the established model were superior to those of the linear regression one [18].Similarly, Terzi (2006) demonstrated the PSI of flexible pavement by using ANNs.Thus, the regression value of the ANN-developed model was higher than the AASHTO model [19].Moreover, in 2020, Yamany et al. offered individual performance models for each state based on performance data obtained from its own road network.On the other hand, the random parameter regression model was superior in some cases when considering individual states [20].
The motivation of this research is to integrate PMS with machine learning (ML), which could be used to predict the maintenance needs of a pavement over time.ML technologies integrated with artificial intelligence (A.I.) technologies, including Random Forest (R.F.), Decision Tree (D.T.), Gradient Boosting (B.G.), and Adaboost, can predict various situations in the PMS, which deals with the rehabilitation and maintenance of flexible pavement.
This study aims to employ ML techniques to predict the PCI to provide insights into the future performance of the pavement.Pavement management also helps to identify current deficiencies and distresses, such as cracking and rutting.In addition, base failure or subgrade instability were more serious structural issues.

PROPOSED METHODOLOGY
For evaluating and predicting the pavement condition index (PCI) of asphalt pavements, 61 major roads with various operational conditions were selected from three U.S. states., namely California, Hawaii, and New Mexico. .These pavement problems negatively impact travel times, accidents, and the environment.Moreover, accidents tend to increase in areas with longer travel times due to drivers attempting to avoid poor road conditions.Developing a predictive model to determine pavement performance would be extremely useful for the competent authorities in selecting the most appropriate pavement maintenance system.Table (1) presents the collected data.The LTPP can access the data for free on its website.https://infopave.fhwa.dot.gov.Performance evaluation metrics A model validation process was employed to assess the predictive capabilities of the MLR and ML models.This process involved evaluating the models' ability to make accurate predictions.To verify the effectiveness of the models, 15% of the data was reserved for testing purposes.This data was used to predict PCI values, which were then compared to the actual PCI values.Various validation and performance measures are typically used to evaluate the validity and performance of statistical models.This study used the  2 , mean absolute error (MAE), root mean square error (RMSE), and mean square error (MSE), to assess the validity and compare the performance of the MLR and ML models.

Comparison of Machine Learning with Conventional Techniques
Several metrics, including  2 , MAE, RMSE, and MSE were employed to evaluate the effectiveness of machine learning techniques compared to conventional methods.The analysis of the results obtained from both approaches revealed that all models exhibited high accuracy.A comparative study between traditional and ML techniques is presented in Table (5), while Figures (7) to (10) depict the contrast between machine learning and conventional methods.Table (5) and Figures (7) to (10)

Figure ( 1 )
presents the research methodology used, which consists of the following: ▪ Data collection.▪ Data Preprocessing ▪ Model Development • Conventional Technique (using the Statistical Package for the Social Sciences (SPSS) software • Four Machine Learning (ML) algorithms: Random Forest (R.F.), Decision Tree (D.T.), Gradient Boosting (B.G.), and Adaboost.▪ Comparison and Validation.Data Description and Preprocessing The data set was collected from Long Term Pavement Performance (LTPP).The data set consists of 61 rows and eight columns.The data collected from 61 roads involved three states in the U.S. (California, Hawaii, and New Mexico).To achieve the aims of this research, the data obtained was employed for predicting the PCI model as a function of pavement distress and traffic volume.The inputs related to pavement distress and traffic volume variables' effects: pavement age, fatigue cracking, longitudinal cracking, transverse cracking, Cumulative Equivalent Single Axle Load (ESAL), Annual Average Daily Truck Traffic (AADTT), and Annual Average Daily Traffic (AADT)

Fig 1 .
Fig 1. Methodological framework depicted.Pavement Condition Index (PCI)The ASTM D6433-18 method was employed to determine the PCI values for the 61 road sections using the data acquired from the LTPP dataset.

Fig 2 .
Fig 2. PCI Measured versus PCI Predicted plot.Equation(1) shows the result of the regression analysis for PCI.The PCI had negatively correlated with age, longitudinal cracking, AADTT, and AADTT, while PCI had positively correlated with fatigue cracking, Transverse Cracking, and ESAL.The regression model developed using the mathematical method mentioned above was validated using statistical error measures, namely  2 , MAE, RMSE, and MSE.The results indicated that  2 was strong, while the values of MAE, RMSE, and MSE were deemed acceptable.Developing Machine Learning (ML) ModelsDeveloping ML Techniques Models four techniques were used in this research (R.F., D.T., G.B., and Adaboost) to predict PCI based on four types of pavement distress and three traffic volume variables for asphalt pavement.Table(4) presents the modelling results for this study's three machine-learning techniques.Figures(3), (4),(5), and (6) present the (R.F.), (D.T.), (G.B.), and (Adaboost) prediction results for PCI models.Table 4: Performance of PCI models ML techniques based on pavement distress and traffic volume.

Fig 10 .
Fig 10.Comparison among the ML and MLR techniques (MSE).Conclusions In this study, the prediction of the PCI is based on pavement distress and traffic volume for three U.S. states.California, Hawaii, and New Mexico were developed using the conventional Technique (MLR) and four ML techniques R.F., D.T., G.B., and Adaboost.Based on the light of the study's findings, the following conclusions can be drawn: • This study focused on predicting PCI value by analysing pavement distress and traffic volume using sixty-one road sections for flexible pavements selected in three U.S. states from the LTPP dataset.• The assessment of ML techniques indicated that Adaboost was more precise in predicting PCI values compared to R.F., D.T., and G.B. techniques, respectively.• Based on the study's findings, both ML and MLR techniques were viable models for predicting PCI values based on pavement distress and traffic volume values.• The study demonstrated that both techniques could accurately predict PCI values, decrease the need for visual examination of PCI values, and save budgets and time.

Table 2 . Mathematical representation of the performance metrics
Table (2) presents these measures are calculated.

Research Analysis Approaches Developing Conventional Techniques Models The
(1) technique has proven reliable for creating precise and efficient models.Its purpose is to forecast the PCI value for flexible pavement based on seven types of pavement damage.To evaluate the model's effectiveness, the  2 values, as well as the RMSE and MAE methods, were used.Table(1)shows the MLR technique for the PCI model.The results of the MLR prediction for the PCI models are