Deep Learning for predictive maintenance

Abstract

Recently, with the appearance of Industry 4.0 (I4.0), machine learning (ML) within artificial intelligence (AI), industrial Internet of things (IIoT) and cyber-physical system (CPS) have accelerated the development of a data-orientated applications such as predictive maintenance (PdM). PdM applied to asset-dependent industries has led to operational cost savings, productivity improvements and enhanced safety management capabilities. In addition, predictive maintenance strategies provide useful information concerning the source of the failure or malfunction, reducing unnecessary maintenance operations. The concept of prognostics and health management (PHM) has appeared as a predictive maintenance process. PHM has become an unavoidable tendency in smart manufacturing to offer a reliable solution for handling industrial equipment’s health status. This later requires efficient and effective system health monitoring methods, including processing and analysing massive machinery data to detect anomalies and perform diagnosis and prognosis. Prognostics is considered a key PHM process with capabilities for predicting future states, mainly based on predicting the residual lifetime during which a machine can perform its intended function, i.e., estimating the remaining useful life (RUL) of a system. The prognostic research domain is far from being mature, which is still new and explains the various challenges that must be addressed. Therefore, the work presented in this thesis will mainly focus on the prognostic of monitored machinery from an RUL estimation point of view using Deep Learning (DL) algorithms. Capitalising on the recent success of the DL, this dissertation introduces methods and algorithms dedicated to predictive maintenance. We focused on improving the performance of aero-engine prognostic, particularly in estimating an accurate RUL using ensemble learning and deep learning. To this end, two contributions have been proposed, and the results obtained were validated by an extensive comparative analysis using public C-MAPSS turbofan engine benchmark datasets. The first contribution, for RUL predictions, we proposed two-hybrid methods based on the promising DL architectures to leverage the power of the multimodal and hybrid deep neural network in order to capture various information at different time intervals and ultimately achieve more accurate RUL predictions. The proposed end-to-end deep architectures jointly optimise the feature reduction and RUL prediction steps in a hierarchical manner, intending to achieve data representation in low dimensionality and minimal variable redundancy while preserving critical asset degradation information with minimal preprocessing effort. The second contribution, in a practical situation, RUL is usually affected by uncertainty. Therefore, we proposed an innovative RUL estimation strategy that assesses degrading machinery’s health status (provides the probabilities of system failure in different time windows) and provides the prediction of RUL window. Keywords: Prognostics and Health Management (PHM), Remaining useful life (RUL), Predictive Maintenance (PdM), C-MAPSS dataset, Ensemble learning, Deep learning