時 間：112年12月26日（星期二下午14:10~15:00）
地 點：管理學院新大樓M240
主講人：許湘伶 教授兼所長 (高雄大學 統計學研究所)
講 題：A Comparative Study of Deep Learning-Based Predictive Methods for Remaining Useful Life
Abstract
Estimating the remaining useful life (RUL) is a crucial predictive activity in industrial applications. Two main research directions for RUL prediction have been identified in the literature: the first is to use deep learning models to estimate RUL, while the second involves similarity-based health state matching, which combines deep learning tools with a similarity measure based on the health indicator (HI) curve. This study proposes three methods to estimate RUL based on these two research directions. The first method adopts a Convolutional LSTM Network (ConvLSTM) model to estimate RUL. In contrast, the second method modifies the HI curve matching technique by introducing a CNN-based approach to construct the HI information and the cosine similarity method. Finally, the individual prediction results of these two methods are integrated to predict the RUL. The application results on the NASA C-MAPSS aircraft turbofan engine dataset show that the three proposed methods have better predictive performance regarding three evaluation metrics, ACC, SCORE, and RMSE, compared to the original literature method and that the ensemble of our methods produces the best average prediction performance. This is joint work with Yi-Zong Ji.