Statistical methods and database splitting for hcf data analysis

Fatigue curves are receiving nowadays an increased level of attention in the wake of experimental campaigns showing that the original ASME III mean air curve, also known as the Langer curve [1], does not represent accurately part of the recently obtained laboratory data. EDF, VTT and E.ON have been working towards a relevant fatigue assessment strategy. The three organizations recently exchanged HCF databases, providing a common benchmark to test and compare the various analysis methods. Following the 2014 PVP paper [2], several statistical approaches are being investigated. A special focus is given to methods able to properly account for run-out data points, which do not have the same statistical significance as failed data points. Besides, it has to be noted that only a limited number of material grades are used in NPP primary loop components and in each case, the material batches are identified and specified in de-tail. Therefore, a more accurate and relevant fatigue assessment might be obtained by splitting large datasets that generally mix various testing conditions and material grades. A comparison is made between a "mixed" approach and a "separated" one, in which the fatigue assessment is performed successively for two or more subsets, e.g. associated with two testing temperature ranges and/or steel grades. Both "mixed" and "separated" strategies are applied to the EDF and the E.ON databases containing fatigue data in different temperatures for non-stabilized and stabilized austenitic stainless steels. The resulting data scatters are compared and the significance of these statistical approaches to fatigue assessment is discussed.