Data-driven fatigue crack quantification and prognosis using nonlinear ultrasonic modulation

In this study, an online monitoring technique for continuous fatigue crack quantification and remaining fatigue life estimation is developed for plate-like structures using nonlinear ultrasonic modulation and an artificial neural network (ANN). First, multiple aluminum plates of different thicknesses are subjected to cyclic loading tests at a constant amplitude, and the ultrasonic responses are obtained from the piezoelectric transducers attached to each specimen. Second, an ANN is constructed by defining (1) the specimen thickness; the elapsed fatigue cycles; and two features extracted from the ultrasonic responses, namely, the cumulative increase and decrease in the nonlinear beta parameter, as inputs and (2) the crack length and remaining fatigue life as outputs. Then, the architecture and learning parameter of the ANN are optimized using the data obtained from the specimen tests. Finally, the performance of the trained ANN is examined using the blind test data obtained from additional specimens. The results of the blind tests indicate that the proposed technique can estimate the crack length and remaining fatigue life with a maximum error of 2 mm and 3 k cycles, respectively, for the tested aluminum plates. The uniqueness of this technique lies in (1) the fatigue crack quantification and remaining fatigue life estimation using nonlinear ultrasonic modulation and (2) the data-driven crack quantification and prognosis using an ANN for online monitoring.