Master/ Bachelor Thesis: Deep Learning and Uncertainty in Predictive Maintenance

Introduction Predictive Maintenance is considered part of the holistic approach Prognostics and Health Management (PHM). It consists of three main tasks: fault detection, fault diagnostics, and prognostics; the prediction of the remaining useful life (RUL) of a system or component. In recent years, practitioners and researchers have turned to Deep Learning (DL)-based methods to address these tasks. This trend is supported by the increasing availability of sensor data of industrial systems. The systems in question are often safety-critical and failures can lead to expensive or catastrophic events. In such contexts, the uncertainty about an information is often more valuable than the information itself. A multitude of models with different Deep Learning architectures (CCNs, Autoencoder, RBMs, etc.) has been published to address the above-mentioned tasks. However, the great majority of approaches does not incorporate ways to express uncertainties. Problem description Deep Learning models usually only deliver point estimates of their predictions. Monte Carlo dropout is a method that overcomes this limitation: An arbitrary neural network with dropout applied to every layer is mathematically equivalent to an approximation of Bayesian inference. It provides a relatively simple way of extracting uncertainties from Deep Learning models without impeding computational efficiency or model accuracy. The objective of this thesis is to evaluate the state-of-the-art and extend existing DL models to account for uncertainty.

Keywords: Machine Learning, Deep Learning, Predictive Maintenance, PHM, Prognostics and Health Management, Uncertainty, Bayesian Inference

Calendar

Location

Institute of Machine Tools and Manufacturing (ETHZ)

Labels

Topics

• Information, Computing and Communication Sciences
• Engineering and Technology

Documents

Name	Comment	Size	Actions
210112_MA-Auschreibung_Uncertainty.pdf		197KB	Download