On the usage of hybrid 1-D convolutional network and long-short-term-memory network for constant-amplitude multiple-site fatigue damage prediction on aircraft lap joints

Multiple site fatigue damage is a problem that affects many operators of aging aircraft. The methods currently in place for prediction of such damage are conservative, sensitive to noise and cannot fully account for grain-level material variations, which results in aircrafts being more conservatively designed than they need to be. The authors augmented the dataset of the FAA AR-07/22 report into a sizable body of variable- and constant-amplitude multiple-site fatigue damage sequences. This was done implementing and tuning the algorithm from AFGROW, implementing the plastic zone linkup criteria for crack interaction, as well as adding Gaussian noise at different stages of the computation. The interim model used for predicting the damage is a hybrid 1-D convolution and bidirectional LSTM model, which achieved an average of 171.26% MAPE, and 4.028 MSLE on the interim version of the dataset. A detailed breakdown of the error characteristics and the hyperparameters that have salient effects on the performance of the model are also examined.