Abstract

Large plants in the process industry are monitored and maintained at regular intervals and repeatedly maintenance is either too early or too late. This causes unnecessary costs due to technicians, spare parts procurement as well as delivery issues and to high downtime costs due to unexpected shutdowns. In this context, the Remaining Useful Life (RUL) plays a major role, as it is an indicator of how long a machine or component can run without breakdown, repair or replacement. By predicting RUL using predictive maintenance, maintenance can be better planned, operational efficiency optimized, and unplanned downtime avoided. Optimizing the prediction accuracy should therefore always be in the foreground and is therefore the topic of this paper.

Abstract

With the increasing availability of data for Prognostics and Health Management (PHM), Deep Learning (DL) techniques are now the subject of considerable attention in Prognostics for Predictive Maintenance, achieving more accurate Remaining Useful Life (RUL) predictions. However, one of the major challenges for DL techniques resides in the difficulty of obtaining large amounts of labeled data on industrial systems. To overcome this lack of labeled data, an emerging learning technique is considered in this work : Self-Supervised Learning, a sub-category of unsupervised learning approaches. This paper aims to investigate whether pre-training DL models in a self-supervised way on unlabeled sensors data can be useful for downstream tasks in PHM (i.e. RUL estimation) with only limited amount of labelled data. A synthetic dataset composed of strain data is used. Results show that the self-supervised pretrained models significantly outperform the non pre-trained models in downstream Remaining Useful Life (RUL) prediction task, showing promising results in prognostic tasks when only limited labeled data is available.