Application of machine learning and statistical analysis to creep and creep-fatigue damage evaluation for austenitic stainless steel

Abstract

The effectiveness of damage evaluation methods was surveyed among the master curve method, the parametrical statistics method and machine learning method by applying those methods to damage evaluation using KAM(Kernel Average Misorientation) parameters obtained from EBSD(Electron BackScateer Diffraction pattern) observations for the interrupted creep and creep fatigue tests of SUS304HTB equivalent heat resistant steel for boiler tube use. As for the parametric statistics method, the log-normal distribution was judged as the best fit distribution type among normal, lognormal and Weibul distributions. Being the algorithm of machine learning effective for pattern recognition, neural network was adopted for the analysis. As a result, it was found that the accuracy was higher in the order of neural network method, parametic statistics method, and master curve method. The reason why the neural network method was more accurate than the parametric statistics method was the latter method could not approximate the frequency distribution shape of KAM accurately. If the frequency distribution profile is unknown as in this case, the method like neural network independent on the distribution profile is considered to be very effective.