A framework for degradation start detection and rul prognosis of physical assets

Abstract

Data-driven approaches for Prognostics and Health Management (PHM) of physical assets have increased in popularity in the recent decade due to the large amounts of data produced by the asset sensor systems, the availability to high computational power and the open access to self-learning algorithms, like artificial neural networks (ANN). Among the various PHM objectives is the prediction of remaining useful lifetime (RUL), which with the development of data-driven approaches is causing a paradigm shift from standard corrective and preventive maintenance to a predictive one. Typical RUL prognostic approaches have been based on the asset's sensor data to perform predictions, as these should correlate to the current health state of the asset. ANNs implemented for these tasks have obtained remarkable results, proving its capabilities on the field. However, a typical problem found in these implementations is that sensor data in the early stages of operation does not show noticeable degradation pattern, resulting in predictions with significant errors in comparison to the true value. This problem has partially been avoided by arbitrarily setting a maximum RUL value that can be predicted or by setting a statistical threshold after which predictions can actually be made. These adaptations have made improvements in RUL accuracy, implying that added considerations and restrictions to the prognostic method can be a subject of study in itself. Considering this matter, for the present Thesis work is proposed a 2-stage RUL prognostic Framework in which the first stage performs a degradation assessment of the asset with a following stage to make RUL predictions after a certain threshold is crossed. The first stage builds a degradation index based on the Mahalanobis distance (MD) between current operational data and known healthy data, thus quantifying the advance of an on going degradative process, if any. The second stage uses an ANN to make the RUL predictions only after a degradative process is detected. This proposed Framework is tested on both a turbofan engine degradation dataset (C-MAPSS) and a rolling bearing degradation dataset (FEMTO). The implementation done with proposed Framework proves that healthy data does not provide meaningful information for accurate RUL predictions, and therefore, can be discarded without negative effects on RUL accuracy. Moreover, by starting with RUL predictions only after the degradation threshold is crossed, RUL inaccuracies at early stage of operation are avoided.