TY - THES ID - 134815230 TI - Application of deep Learning for predictive maintenance AU - Brahy, Pierre AU - Boute, Robert AU - Van Staden, Heletje AU - KU Leuven. Faculteit Economie en Bedrijfswetenschappen. Opleiding Master of Business Engineering PY - 2019 PB - Leuven KU Leuven. Faculteit Economie en Bedrijfswetenschappen DB - UniCat UR - https://www.unicat.be/uniCat?func=search&query=sysid:134815230 AB - Prognostics accurately predicts the remaining useful lifetime of components in order to proactively perform maintenance. In recent years, deep learning algorithms have proved superior performance thanks to their ability to automatically extract degradation features from normalized data. However, recent scientific literature indicates that this trait is slightly eroded for convolutional neural networks when applied to non-stationary data. Indeed, researches in the field usually involve extracting time-frequency features from normalized data before implementing the convolutional neural network (signal processing techniques), but, as of today, it has not been empirically corroborated that non-stationarity undermines the performance of a convolutional neural network in prognostics. The present thesis therefore pursues a twofold objective. First, it aims to fill the research gap by conducting an extensive literature review of convolutional neural networks applied to prognostics and discovering patterns. Second, it aims to examine the viability of the direct implementation of a fundamental convolutional neural network on non-stationary bearing data without signal processing techniques. It is built on an incremental approach, where the capacity of the convolutional neural network structure is gradually increased. The results demonstrate that the generalization performance of the convolutional neural network applied on non-stationary bearing test data is consistently poor, regardless of the network capacity. ER -