Constant-amplitude fatigue crack growth sequence regression on an aircraft lap joint using a 1-D convolutional network

Abstract

A difficult issue that faces the operators of aging aircraft is to assure the regulators that their aircraft is structurally sound. From a structural perspective, aircrafts are complex and susceptible to fatigue damage due to the usage of aluminium and the proliferation of rivets. Multiple site fatigue damage (MSD) refers to the presence of multiple fatigue cracks in the same structural element. These cracks may interact with each other, increasing their collective growth rates, which may result in a rapid degradation of the structural integrity of the structure. This paper describes this group's experiences related to the usage of 1-D convolutional neural networks (1D ConvNets), in conjunction to fracture mechanics-borne material fatigue damage models in the task of modeling the damage progression of an aircraft lap joint with multiple cracks, subjected to a specific range of possible load sequences. This multivariate regression task is a proxy to the crucial task of estimating the point at which a structural element's residual strength levels have been lowered enough that it cannot sustain the loads it is required to sustain anymore. Current methods of doing the above still require empirically determined correction factors to account for material-related scatter, which require expensive physical tests to determine. The 1D convolutional network is trained on a combined dataset, which consists of the fatigue test data from the FAA (Federal Aviation Administration) AR-07/22 report, augmented by data generated by a crack growth analysis program, with initial conditions and load sequences akin to those in the FAA's fatigue test. This initial version of the model obtained an average MSE of 1.05, MAE of 0.85 and an average MAPE of 125.26 over the 110-sequence testing set.