Abstract

Maintenance has always been an essential and inseparable part of manufacturing and industrial sectors. Generally speaking, maintenance strategies aim to prevent asset failures/downtimes to protect investments and to provide a safe working environment. With the recent growth in sensor and data acquisition technologies, a rich amount of condition monitoring data has become available in manufacturing and industrial sectors. Consequently, there has been a recent surge of interest in using more advanced solutions, especially those based on Machine Learning (ML), Reinforcement Learning (RL), and Deep Learning (DL) models, to utilize such extensive and high-quality data in the maintenance domain. In this context, the thesis proposed different ML and hybrid models for prognostic and health management purposes to further advance the maintenance field. In particular, we conducted the following three studies: In the first work, a hybrid and semi-supervised framework is designed based on the hazard rate of the system. The proposed framework can extract the hidden state of the system without domain knowledge. To evaluate the efficacy of the proposed method, a real dataset is used where optimal maintenance policies are obtained based on the extracted states via RL. In the second study, a DL-based model is proposed to predict the hazard rate of the underlying system. As opposed to its statistical counterparts, the proposed predictive model does not assume any linear relationship between the sensors' measurements, and is capable of learning from censored data. In the last study, we investigated application of the proposed methods on high-dimensional data such as images. The proposed methods achieved promising results illustrating their great potential to be used in real-world applications.