쌍방향 장단기 기억 네트워크를 이용한 주기별 저주기 피로 응력-변형률 거동 예측

초록

Bi-Long Short-Term Memory (Bi-LSTM) network is a long short-term memory network with end-to-end learning capabilities and it is widely applied in various fields such as machine translations and image predictions. If the low cycle fatigue stress-strain behavior of a material could be predicted by cycle using Bi-LSTM, the analysis on the life and characteristics of the material is possible by applying diverse machine learning methods based on the predicted results. In this study, partial fraction of stress and strain data acquired from low cycle fatigue test on stainless steels were inputted as the training set of Bi-LSTM network. Subsequently, low cycle fatigue stress-strain behavior of testing set was predicted based on the trained prediction model. Bi-LSTM network configured which was set to contain 2,000 hidden nodes and to use adaptive moment estimation as optimization function. After training on 50% of low cycle fatigue data, the accuracy between predicted stress- strain and experimental values was evaluated using coefficient of determination. Analysis showed that model using Bi-LSTM network had prediction accuracy of coefficient of determination R 2=0.9833. Therefore, the proposed methodology of low cycle fatigue stress-strain behavior prediction model configuration would be utilized in the fields of real-time self-diagnosis of fatigue life.

키워드