Comparative Evaluation of Outlier Detection Methods for the Asphalt Fracture Tests

Asphalt mixture laboratory cracking evaluation methods often rely on the use of notched test specimens that are tested using a constant displacement rate to obtain fracture properties. Disk-shaped compact tension (DCT) test is one of the most widely adopted fracture tests for asphalt mixtures. It is standardized as ASTM D7313-20, Standard Test Method for Determining Fracture Energy of Asphalt Mixtures Using the Disk-Shaped Compact Tension Geometry. Variability in load-displacement curves among replicates of the same asphalt mixture type complicates the calculation of performance indices obtained from asphalt fracture tests for cracking evaluation. Abnormalities such as irregularities under the curve, unusual maximum loads, or slopes after the peak can occur. To improve the precision and dependability of cracking performance assessments, it is essential to detect and remove these atypical replicates. This research examines four outlier detection methodologies: the David, Hartley, and Pearson (DHP) method, the Tietjen-Moore (T-M) method, the uniform variance estimator (UVE), and the robust estimator (RE), focusing on their effectiveness and precision when applied to DCT performance indices. Both fracture energy and post-peak index (PPI) parameters were evaluated, with the RE method consistently demonstrating superior outlier detection capabilities. This was attributed to the RE method’s adaptability to both normally and nonnormally distributed data and its heightened sensitivity to outlier presence. As the nonnormality of the PPI datasets escalated, the T-M method tended to overestimate outliers, thereby further substantiating the robustness of the Mahalanobis distance (MD) methods, which include both UVE and RE. Even though the UVE method significantly surpassed the DHP method in outlier detection, it still showed sensitivity to existing outliers. Collectively, these findings underscore the significance of accounting for the nature of distribution for the dataset when selecting an outlier detection method and corroborate the potential of the RE method as a versatile and robust choice for outlier detection.