Session: 14-06-01: Developments in Design Theory for Component and System Safety and Reliability

Paper Number: 114086

114086 - Uncertainty Quantification in the Prediction of Remaining Useful Life Considering Multiple Failure Modes 

Despite the substantive literature on remaining useful life (RUL) prediction, there is less attention paid to the influence of epistemic uncertainty and/or aleatory uncertainty in multiple failure behaviors in the prediction accuracy of RUL. Multiple failure behaviors may influence the consistency of feature space, which makes it difficult to satisfy the hypotheses that the training data must have the same distribution as the testing data. This problem may lead to the performance degradation of the conventional deep learning model in the face of different failure behaviors. Besides the existing relevant research related to multiple failure behaviors, which mainly focuses on reliability estimation, prognostics approaches are scarce. Characterizing aleatory and epistemic uncertainty in the predicted RUL can help in designing and constructing more reliable systems.

 

The research question in this study was: can uncertainties be quantified in the prediction of RUL of systems with multiple failure modes? The first objective was to provide a qualitative analysis of epistemic and aleatory uncertainties of systems with more than one failure mode and how they affect RUL prediction. The second objective was to calculate the epistemic and aleatory uncertainty in the prediction of RUL considering multiple failure modes.

 

The epistemic and aleatory uncertainties were qualitative studied in systems with more than one failure mode. Then, it was quantitative analyzed how they affect the RUL prediction. The Bayesian Neural Network (BNN) was used to quantify epistemic and aleatory uncertainty while predicting RUL. When a probability distribution is placed on the unknown inputs, BNN can discover the underlying dimensionality of a data set. The initial distribution for the training and test data are calculated using the principle of maximum entropy, to make the consistency of feature space. After applying BNN, the RUL was calculated using a form of linear regression, then the epistemic and aleatory uncertainties in the signal were also involved in the calculation of the RUL. Finally, the performance of the proposed uncertainty framework was applied on two distinct datasets, namely bearing dataset and battery dataset.

 

The results of the qualitative epistemic and aleatory uncertainty on RUL in systems with more than one failure mode were presented and discussed in detail. Also, the study resulted in an RUL uncertainty quantification model for multiple failure modes. The result of RUL demonstrated much more probabilistic information, instead of only a common deterministic final value presented in literature. The performance of the proposed framework in the prediction of RUL was demonstrated o a bearing and battery dataset. Therefore, it was possible to extract the RUL uncertainty and the propagation of uncertainty over time/cycle on those datasets. It was demonstrated that the BNN parameters should be tuned to best capture the initial uncertainty of the data to use this model on other types of datasets.

 

The results reveal an analysis of different kinds of uncertainties. Also, the epistemic and aleatory uncertainty were properly quantified in the RUL of the system with multiple failure modes. Finally, the results demonstrated a superior RUL prediction performance. The generalization of the results in this paper may lead to the design of more reliable systems that exhibit multiple failure modes.

Presenting Author: Nazir Gandur Texas Tech University

Presenting Author Biography: Mr. Nazir Gandur is a PhD student at Texas Tech University.

Authors:

Nazir Gandur Texas Tech University
Stephen Ekwaro-Osire Texas Tech Univ