Machine Learning Based Prognostics of Fatigue Crack Growth in Notch Pre-cracked Aluminum 7075-T6 Rivet Hole

Constant stress amplitude fatigue tests were conducted on the notch pre-cracked Aluminum 7075-T6 rivet hole dog-bone coupons. Monitoring of visible surface crack length by special surface engraving using digital microscope images and by ultrasonic sensors signals was carried out to yield fatigue crack length measurements in relation to number of fatigue cycles applied. The experimental results provide ultrasonic sensor validation for fatigue crack length measurements. Fracto-graphic examination of failed fatigue surfaces has provided further confirmation of notch pre-crack length, crack initiation process, and crack growth marker bands. These experimental inputs were used in NASGRO and AFGROW software fatigue crack growth simulations. The simulation results did not match the crack initiation fatigue life measured by experiments. However, there was good agreement with crack growth simulations of larger cracks. Hence, we plan to develop a machine learning application that will learn the fatigue crack initiation and crack growth processes from data obtained from our own experiments and other fatigue data available from AFGROW databases. Nonlinear AutoRegressive models with eXogenous input (NARX) artificial neural network were used to predict crack growth longer than 5.0-mm. Particle filtering modeling with Bayesian updating was applied to these experimental data for prognostics of fatigue crack growth. A concept design and preliminary implementation results will be presented.