Investigating the Relationship of As-Constructed Asphalt Pavement Air Voids to Pavement Performance (2021)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

178 Table 18 Summary of All Analysis Methods Performance Characteristic Pavement Type Analysis Method 1 Analysis Method 2 Analysis Method 3 Rutting New 57% met expectation 21% contradict Approximate R2 = 0.31 nominally contradict Predicted R2 = 0.46 nominally met Rehabilitation 67% met expectation 0% contradict Approximate R2 = 0.16 minimally contradict Predicted R2 = 0.47 no influence Fatigue Cracking New 82% met expectation 9% contradict Approximate R2 = 0.41 significantly met Predicted R2 = 0.62 mixed expectation Rehabilitation 40% met expectation 30% contradict Approximate R2 = 0.35 nominally met Predicted R2 = 0.46 significantly met Thermal Cracking New 42% met expectation 50% contradict Approximate R2 = 0.38 nominally met Predicted R2 = 0.36 no influence Rehabilitation 50% met expectation 50% contradict Approximate R2 = 0.32 nominally met Predicted R2 = 0.30 nominally met Ride New 54% met expectation 38% contradict Approximate R2 = 0.24 nominally met Predicted R2 = 0.19 nominally met Rehabilitation 25% met expectation 42% contradict Approximate R2 = 0.21 nominally met Predicted R2 = 0.39 nominally met Met expectation – lower as-constructed air voids improved pavement performance Contradicted – higher as-constructed air voids improved pavement performance Approximate R2 – applied to models that used quintile regression to fit the median Predicted R2 – applied to ANN model prediction using test data Regional Analysis Approach This section of the User Guide provides an agency with recommendations on the process to perform this analysis with data from the agency’s dataset. Many of the recommendations reflect lessons learned by the research team as the NCHRP study progressed. In general, a significant amount of effort will be required to assemble a quality dataset in preparation for each analysis. The strength of each analysis method improves as the size of the dataset increases. This is particularly true for using an artificial neural network analysis method, which was limited to approximately 200 available LTPP sections for the new construction analysis in this NCHRP study. The minimum number of sections for Analysis Method 1 will be controlled by the number of subgroups assembled, but Analysis Methods 2 and 3 require a large set of data to improve the accuracy of the prediction models developed. An agency may elect to apply all three analysis methods and will need to assemble a dataset that will satisfy the minimum requirements for all three analyses. Further, agencies will need to assemble the types of data that are already available and some of the data will not match the units used in the NCHRP study. This is expected to be very true for the pavement performance characteristics measured by each agency. Each analysis approach will be discussed in more detail in the remainder of this section. Each pavement section has a unique combination of input variables, which results in a unique performance curve. To investigate the influences of as-constructed AV on pavement performance, examine the predicted performance of each pavement section applying Analysis Method 2 and/or 3 using the following steps. 1) assemble 10 to 15 years of the input data for each pavement section used in the study; 2) vary as-constructed AV model input values from 3% to 9% for each section; 3) predict pavement performance from one to ten years using the developed models; and 4) average the predicted pavement performance for each combination of air void and pavement age. Using this approach, the established relationship between as- constructed AV and pavement performance would be representative for most of the pavement sections.

179 Analysis Method 1 Subgroup Scatter Plots This analysis approach divides the agency’s pavement network into subgroups with common climate, traffic, and pavement structure such that the performance of the pavement sections within a subgroup is expected to be similar. The climate conditions within the agency’s boundaries will determine how many climate zones are needed. Agencies with significant elevation differences due to mountainous terrain should examine high elevation climate conditions separate from lower level valleys, plains, and coastlines. The differences in temperature (freeze, no freeze) and precipitation (wet, dry) are commonly applied to define climate zones. The traffic subgroup levels will reflect the amount of traffic within the agency’s highway network and more appropriately the amount of heavy truck traffic. The NCHRP research team limited the traffic to high and low categories, which may not be appropriate for some agencies. As a minimum, an agency should consider three categories representing low volume rural truck traffic, medium volume arterial network truck traffic, and high volume freeway truck traffic. The traffic should be quantified by annual ESALs as a common measure of the truck traffic’s influence on the pavement section’s performance. To accurately compute the ESALs, the agency should account for the axle load distribution and pavement structure, but some agencies will need to apply more general ESAL per truck class values to estimate the annual ESALs for each pavement section. The pavement structure subgroups may strongly correlate to the traffic level because an agency often applies the same pavement structure concept to all routes with similar traffic. It is possible that the agency would only establish two pavement structure levels to each traffic subgroup. For example, low truck traffic routes will be divided into thin and medium pavement structures while high truck traffic routes will be divided into medium and thick pavement structures. Similar to quantifying truck traffic, the pavement structure needs to be consolidated into a single value for each pavement section. The NCHRP research team elected to use the SN value for this analysis. An agency may determine that 12 subgroups (2 climates, 3 traffic, and 2 pavement structures) best divide the pavement network for this analysis. Once the number of subgroups is established, the agency needs to identify a minimum of 6 (ideally 10) pavement sections for each subgroup. Just as the NCHRP research team was bound by the use of the LTPP database, the agency will be bound by the pavements within their network. The pavement sections should be randomly selected, but as-constructed AV for selected pavement sections may, or may not, provide a sufficient range of values from 4% Gmm to 10% Gmm to achieve a good subgroup analysis. Most agencies will not have as-constructed AV in their pavement management database and will have to look into historic construction as-built records to collect this value. Unlike controlled LTPP research sections, the agency will be working with the lengths of each management database section. Assembling as-constructed AV requires information on each asphalt lift constructed so a weighted average AV value can be computed. The NCHRP research team examined four performance characteristics: rutting, fatigue cracking, thermal cracking, and ride, which are common long-term measures of performance. An agency should select only the performance characteristics that are critical to their network. The number of years of performance data needed will depend on the frequency that data was collected. A minimum of 6 performance data points are recommended to build a performance

180 curve, so if performance was only measured every 2 years, the performance period is 12 years. For Analysis Method 1, each performance characteristic is represented by a single value of condition (inches of rutting) or time (initial year of thermal cracking) but the entire performance curve should be determined to conclude that the pavement section performed normally. The agency should remove pavement sections from the analysis if the performance curve indicates that poor materials or construction caused a premature failure. The complete dataset for this subgroup analysis enters each pavement section as a single row with individual columns for climate zone, traffic level, pavement structure, as-constructed AV for the surface lift, as-constructed AV for the entire asphalt layer, and a column for each performance value as shown in Table 19. The assembled dataset is first sorted by climate, then each climate subgroup is sorted by traffic, and finally each traffic subgroup is sorted by pavement structure. Each subgroup now represents a common set of climate, traffic, and pavement conditions from which each pavement section in the subgroup is plotted as a single point using as-constructed AV (x-axis) and a specific performance value (y-axis) as shown previously in Figure 2. Each subgroup should have a minimum of six pavement sections but Table 19 is only showing three pavement sections per subgroup to limit the size of the example table. Table 19. Generic Dataset Format Sorted into Eight Analysis Subgroups Pvmt ID Climate Traffic Pvmt Struct. Air voids Surface Air voids Average Perform 1 Perform 2 Perform 3 Q No frz low Thin W No frz low Thin E No frz low Thin R No frz low thick T No frz low thick Y No frz low thick U No frz high Thin I No frz high Thin O No frz high Thin P No frz high thick A No frz high thick S No frz high thick D Freeze low Thin F Freeze low Thin G Freeze low Thin H Freeze low thick B Freeze low thick J Freeze low thick K Freeze high Thin L Freeze high Thin Z Freeze high Thin X Freeze high thick C Freeze high thick V Freeze high thick Using Analysis Method 2 Regression Models This section describes how to assess the effect of as-constructed AV on pavement performance by quantifying and interpreting the coefficients of statistical models. A separate model is developed for each combination of pavement type and performance characteristic of interest. The models include input for climate, traffic, pavement structure, and material properties along with accounting for potential interactions on the effect of as-constructed AV on pavement performance. Performance can be based on measured distress over time or the time

181 that distress reaches a critical value. For example, rutting measured over time or the year in which thermal cracking first appears. A dataset is created for each performance characteristic to be examined. Step 1 – Assemble the Dataset. The recommended list of input variables is given in Table 5. Identify and assemble specific climate parameters for each section. Combine pavement layer material attributes into section-level attributes for each construction event. Analyze the extent and type of missing values in the dataset to decide what observations to remove or impute (by inserting an approximate value) in order to obtain a complete section attributes dataset for analysis. Determine which construction events caused significant immediate changes in the surface condition or significant changes in the performance trend. These significant events define periods of analysis with consistent performance trends. Assemble the pavement performance data and time-dependent climate and traffic variables. Finally, merge the time-dependent dataset and section attributes dataset for each pavement section to obtain the final analysis dataset for each performance characteristic. Each row of data is one increment in time consisting of climate, traffic, pavement section, and performance. Step 2 – Determine Model Functional Form. The model functional forms considered for each performance model are selected from the patterns observed in the performance dataset and reported in related studies that balanced best model fit and lowest complexity. Step 3 – Pre-Process the Variables. First, consider alternative transformations to some variables, such as logarithmic or power transformations, to either reduce their skewness or to increase the predictive power. Given the very different scales of magnitude among variables, the second type of pre-processing consists of scaling all variables for better numerical stability and more direct comparison of effects. The scaling of each variable consists of subtracting its mean and dividing by its standard deviation. Step 4 – Apply a Regression Estimation Technique. Consider a quantile regression to reduce the effect of the outliers observed in the dataset on the estimated model parameters without the loss of information resulting from removing data with outliers. Use boxplots of scaled variables to identify the outliers. Step 5 – Select the Optimum Variables. The initial step consists of identifying combinations of variables that may cause issues in the estimation (such as pairs of variables presenting high correlation). The next step consists of running a stepwise algorithm that increases the number of variables and interaction variables in each step and stops when the addition of variables does not result in significant improvement in model fit. Once the optimum set of variables is found, the final step consists of removing any variables not statistically significant. Step 6 – Interpret the Regression Model. The final results of each regression analysis are assembled in a table format for making observations relative to the influence of as- constructed AV and other variables in the regression model. The table columns include variable name, estimated regression coefficient value, standard error, t value, and p-value. In general, higher absolute values for the estimated coefficient and t value are associated with greater influence on the regressed performance value. Refer to the final report for a more detailed discussion of the how these steps were accomplished for the LTPP dataset used in this study.

182 Analysis Method 3 ANN Models This section describes how to assess the effect of as-constructed AV on pavement performance by ANN models. The ANN approach is a computer based adaptive information processing technique that allows for establishing correlations between the input variables Xi and the output variables Yj through the inter-connected neurons (i.e., weight factors, wji). A separate model is developed for each combination of pavement type and performance characteristic of interest. The models include input for climate, traffic, pavement structure, and material properties. Performance is based on measured distress over time. A dataset is created for each performance characteristic to be examined. Step 1 – Assemble the Dataset. For the objective of this study, the ANN approach used to develop prediction models requires a large number of data (typically over 1000 data points) collected from historical databases. The pavement sections selected must have comprehensive information of traffic, climate, layer material, pavement structure, and long-term performance as identified in Table 8. Each pavement section must contain multiple data points for historical climate, traffic, and long-term performance. If 200 pavement sections are selected and each has 6 increments of measured performance data, then the ANN model has 1200 data points to apply to the analysis. The newly constructed and rehabilitated pavement sections are separated for analysis because they required some different model inputs and are expected to have different performance curves. Step 2 – Pre-process the Variables. The input variables Xi and the output variables Yj are normalized using a minimum-maximum strategy to xi and yj respectively. Then, the identified pavement sections are randomly divided into the two clusters, namely a training-validation cluster and a test cluster, with a ratio of 9:1 in terms of the number of pavement sections. The training-validation cluster is used to develop the ANN models and the test cluster is used to examine the prediction accuracy of the models. The training-validation cluster is further divided into a training subset and validation subset with a ratio of 7:3 or 8:2 by randomly selecting the individual data increments. Step 3 – Determine the Model Architecture. A four-layered ANN architecture consisting of one input layer, two hidden layers, and one output layer is constructed to develop a model for each measured performance characteristic. The input variables contain the traffic, climate, layer material, and pavement structure characteristics, and the output variable is one of the pavement performance datasets. Each ANN performance model can have slightly different input parameters and number of neurons in hidden layers. Table 20 presents the ANN architecture for each performance model developed in this NCHRP study. Table 20. Architecture of ANN Performance Models for The NCHRP Study ANN Performance Models Model Architecture Rutting – New Construction 14-25-5-1 Rutting – Rehabilitation 15-15-5-1 Fatigue Cracking – New Construction 14-25-5-1 Fatigue Cracking – Rehabilitation 15-15-5-1 Thermal Cracking – New Construction 14-25-5-1 Thermal Cracking – Rehabilitation 15-20-3-1 Ride – New Construction 14-25-5-1 Ride – Rehabilitation 15-15-5-1

183 Step 4 – Apply the ANN Processing Technique. Once the training and validation datasets are assembled and the ANN architecture is defined, the ANN process uses transfer functions, a back propagation method to minimize the mean squared error with the training dataset, and a learning algorithm to adjust the neuron weight factors all programmed using the Matlab software. Once the weighting factors are established, the validation dataset is applied to the ANN model and the mean squared error for the validation dataset is measured. The mean squared errors of the training and validation predictions are compared to determine if the ANN model requires further adjustment. If the comparison is not satisfactory, the training and validation datasets are resorted (Step 2), the ANN architecture is revised (Step 3), and ANN processing is repeated (Step 4). Step 5 – Determine the ANN Model Prediction Accuracy. The test dataset is not used for the development of ANN model. The test data over the first ten years of measured performance is used to demonstrate the pavement performance prediction accuracy of the developed ANN model. The ANN model will likely show lower prediction accuracy for the test set than the training-validation set. The prediction accuracy of the ANN models is often attributed to the relatively small amount of test data representing the diverse parameters of a broad training and validation dataset and the models should improve with a narrower dataset, such as a specific class of pavements. Refer to the final report for a more detailed discussion of the how these steps were accomplished for the LTPP dataset used in this NCHRP study.