A machine learning approach to nonlinear ultrasonics for classifying annealing conditions in austenitic stainless steel

Abstract

This paper explores the feasibility of machine learning algorithms on nonlinear ultrasonics for classification of the austenitic stainless-steel material subjected to different annealing conditions. The material that is isothermally annealed at 1323 K for different soaking times showed a marginal variation in its nonlinearity parameter at larger mean grain sizes. The grain growth during annealing followed the Arrhenius type equation fairly well, which has been verified using a genetic algorithm approach. The machine learning algorithms are trained using features such as the ratio of the harmonic amplitudes, root-mean-square value, and the phase difference between the fundamental and second harmonic components derived from the nonlinear ultrasonic response. Upon evaluating the performance of decision tree and ensemble learning algorithms in the classification of annealed materials, it was observed that the LPBoost classifier has the highest accuracy of 97%. According to the results, it is concluded that a machine learning strategy based on a minimal number of features can effectively classify specimens that are otherwise indistinguishable in their nonlinear response. This research takes a step forward to the automation of non-destructive testing toward Industrial Revolution 4.0. The results also pointed out the necessity of parameter fusion in non-destructive decision making.