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.
132 C H A P T E R 3 Findings and Applications The research team approached the project with an optimistic expectation that the proposed work plan was straightforward and would proceed without delay. However, as the work progressed, a number of tasks required more effort to accomplish than anticipated. Components of the dataset needed more processing effort due to inconsistencies between data collection methods (traffic AADTT and ESAL data did not agree), a lack of understanding of the data (proper use of project start date versus open to traffic date), or limitations of the data (no measured distress over the performance period). The research team specifically proposed three analysis methods recognizing that the diverse nature of the input and output data may not lend itself to one perspective. Among the analysis methods, the use of artificial neural networks proved to be particularly challenging. Other research teams proposing to re-examine this topic should pay attention to these lessons learned. A research team using LTPP data must acquire a proper understanding of each variable extracted from the LTPP database. The research team made a number of initial assumptions about the LTPP dataset that were later determined to be incorrect and required extra research time to re-analyze the data. A key early misstep was selecting an incorrect LTPP date parameter to initiate the performance time. LTPP measured and recorded cracking data describes the physical location of the crack but does not assess the cause of the crack as fatigue, thermal, related to poor construction (such as segregation or delamination), or other cause. The research team made a general assumption that all cracking located in the wheel path was related to fatigue cracking and all transverse cracking was related to thermal cracking. Calendar dates assigned to LTPP sections must be understood and applied correctly. Three different dates were applied to start time for output pavement performance measures. Open to traffic date was used for rutting, fatigue cracking, and ride. Pavement construction completion date was used for thermal cracking. The date a section was assigned to the LTPP was used for all rehabilitation sections when the other dates were not recorded. It was recognized that staged construction practices would create deviations for some sections, but that degree of section data extraction was not performed. Assembling a quality dataset for the analysis took time. While individual variables in the LTPP database are carefully reviewed through a LTPP quality control evaluation, there is no guarantee that two similar variables, such as measures of traffic, will have comparable values. When the research team recognized the discrepancy in the traffic data and began formulating a solution, it was discovered that the LTPP had already noted this issue and initiated a study to address these issues. The team took advantage of the LTPP study to apply the best available traffic data. Performance data for a number of LTPP sections, specifically thermal cracking data, recorded no measured distress over the entire performance period. This complicated the analysis. Some sections were used as a zero measured distress output value and other sections were omitted from the analysis as sections with insufficient data.
133 Analysis Method 1 examined the distribution of climate, traffic, and pavement structure data to select boundaries for establishing a matrix of subgroups. LTPP sections were assigned to the appropriate subgroup and 27 subgroups had a sufficient number of sections for further analysis. Each sectionâs measured performance was consolidated to a single value obtained from the sectionâs performance curve. A graph was created for each subgroup that plotted the performance value and as-constructed AV value for each LTPP section. The trend of the plotted values was subjectively examined and classified as meeting expectation (improved performance as AV decreased), no influence (similar influence on performance across a range of AV), or contradicting expectation (reduced performance as AV decreased). ⢠For rutting, 62% of the subgroups met expectation and 12% contradicted expectation. ⢠For fatigue cracking, 62% met expectation and 19% contradicted expectation. ⢠For thermal cracking, 46% met expectation and 50% contradicted expectation. ⢠For ride, 40% met expectation and 40% contradicted expectation. Analysis Method 2 examined the dataset using statistical regression methodologies. More detailed climate FI, layer material stiffness, and pavement structure layer thickness input variables were added to the dataset and all input variables were scaled so that each input was equally weighted in the analysis. A separate regression model was developed for each combination of pavement type (new construction and rehabilitation) and performance measure (rutting, fatigue cracking, thermal cracking, and ride). This analysis forced the regression models to fit a specific curve shape that was considered common for each performance measure. Only statistically significant input variables were retained in each model. The model development used quantile regression to fit the median, so an approximate R2 was applied, which is similar in concept to R2 but is not directly comparable. After the models were developed, a set of average input variables and increments of as-constructed AV were used to create a series of performance curves to assess the influence of AV. These curves were placed in the User Guide. ⢠For rutting of new construction, the regression model had a median absolute error (MAE) of 0.04 inches, an approximate R2 was 0.31, as-constructed AV had a very small contradicting statistically significant effect, and model prediction nominally contradicted the expectation. ⢠For rutting of rehabilitation, model MAE was 0.04 inches, an approximate R2 was 0.16, as- constructed AV had a very small contradicting statistically significant effect, and model prediction minimally contradicted the expectation. ⢠For fatigue of new construction, model MAE was 1.74%, an approximate R2 was 0.41, as- constructed AV had a very small contradicting statistically significant effect, but model prediction significantly met the expectation. ⢠For fatigue of rehabilitation, model MAE was 1.82%, an approximate R2 was 0.35, as-constructed AV had no statistically significant effect as an independent variable, and model prediction nominally met the expectation. ⢠For thermal cracking of new construction, model MAE was 164 ft/ln mi, an approximate R2 was 0.38, as-constructed AV had no statistically significant effect as an independent variable, and model prediction nominally met the expectation. ⢠For thermal cracking of rehabilitation, model MAE was 158 ft/ln mi, an approximate R2 was 0.32, as-constructed AV had no statistically significant effect as an independent variable, and model prediction nominally met the expectation. ⢠For ride of new construction, model MAE was 7.88 in/mile, an approximate R2 was 0.24, as- constructed AV had no statistically significant effect as an independent variable, and model prediction nominally met the expectation. ⢠For ride of rehabilitation, model MAE was 7.88 in/mile, an approximate R2 was 0.21, as- constructed AV had a very small statistically significant effect meeting the expectation, and model prediction nominally met the expectation.
134 Analysis Method 3 examined the dataset using ANN methodology. More detailed climate FI, layer material stiffness, and pavement structure layer thickness input variables were added to the dataset and all input variables were normalized so that each input was equally weighted in the analysis. A separate ANN model was developed for each combination of pavement type (new construction and rehabilitation) and performance measure (rutting, fatigue cracking, thermal cracking, and ride). All input variables were retained in each model. Each model was tested with 10% of the data and an R2 was computed on the prediction power of the model using the test data. After the eight models were developed, increments of as- constructed AV were applied with the climate, traffic, and other material inputs for each LTPP section to create a series of performance curves to assess the influence of as-constructed AV. The results for all LTPP sections at each AV increment were averaged together to present a global influence of AV on each performance measure. ⢠For rutting of new construction, the ANN training and validation model R2 was 0.96-0.90, the prediction accuracy using test dataset was R2 of 0.46, and model prediction curves nominally met the expectation. ⢠For rutting of rehabilitation, the ANN training and validation R2 was 0.96-0.91, test dataset R2 was 0.47, and model prediction showed no practical influence. ⢠For fatigue of new construction, the ANN training and validation R2 was 0.98-0.93, test dataset R2 was 0.62, but model prediction only met the expectation at high AV. ⢠For fatigue of rehabilitation, the ANN training and validation R2 was 0.96-0.85, test dataset R2 was 0.46, and model prediction significantly met the expectation. ⢠For thermal cracking of new construction, the ANN training and validation R2 was 0.997-0.95, test dataset R2 was 0.36, and model prediction showed no practical influence. ⢠For thermal cracking of rehabilitation, the ANN training and validation R2 was 0.98-0.92, test dataset R2 was 0.30, and model prediction nominally met the expectation. ⢠For ride of new construction, the ANN training and validation R2 was 0.98-0.93, test dataset R2 was 0.19, and model prediction nominally met the expectation. ⢠For ride of rehabilitation, the ANN training and validation R2 was 0.99-0.96, test dataset R2 was 0.39, and model prediction nominally met the expectation. Two methods were used to validate the results of the three analysis methods. The first validation method compared the predicted performance using AASHTO Pavement ME to results in Analysis Method 1. Nine LTPP sections from two subgroups were selected and processed to obtain Pavement ME predicted performance for rutting and fatigue cracking over a range of as-constructed AV. The rutting and fatigue trends agreed (met expectation) with the study for the climate zone 2 subgroup but differed for the climate zone 4 subgroup. The second validation method compared the actual performance of research test sites at MnROAD and NCAT to the predicted model performance of those research test sites for each analysis method. Due to the nature of MnROAD and NCATâs test facilities, there were several limitations on the available datasets. For instance, almost 50% of the NCAT sections were characterized by having high SN (above 6.4) and heavy traffic (over 10 million ESALs). These conditions were not found in the LTPP dataset for climate zone 4 (wet â no freeze). Based on these and other limitations, the external validation pavement sections were restricted (filtered) to the range of pavement structures and traffic conditions found in the LTPP dataset. After performing data filtering, predicted values tended to be closer to observed results based on plots of observed versus predicted values. Some of these observations were statistically verified after conducting an F-test of the mean square errors of both model and validation data. Analysis Method 2 models were found to be a fair fit for the LTPP dataset; and for several cases, these models were applicable to external validation data. On the other hand, Analysis Method 3 artificial neural network models were a significantly better fit for the LTPP dataset, but these models were not applicable to the external validation data.