JOURNAL OF MECHANICAL SCIENCE AND TECHNOLOGY, v.37, no.5, pp.2353 - 2362
Abstract
We developed a fatigue residual useful life (RUL) prediction model using the available time-series fatigue data of Ni-base alloy welds via a long short-term memory (LSTM) network. The effects of some LSTM network hyperparameters on model performance were investigated through sensitivity studies. The LSTM network model outperformed multiple regression models when the LSTM model hyperparameters were appropriately tuned. However, the additional gain was insignificant, considering that the LSTM network was much more complex than multiple regression models. The best performance of the LSTM network model was achieved when the number of hidden units, input window size, and batch size were small and the number of LSTM layers was large.