Reducing Uncertainty in Failure Prediction Using Singular Value Decomposition Feature Selection

Abstract

AbstractUncertainty is one of the indicators to evaluate the prediction of deep learning models. In real-world prognostic domain, predicting failure with uncertainty provides users with powerful argument when important decisions related to safety, security and monetary must be made. While quantifying uncertainty has slowly becoming a norm in deep learning, no work has been dedicated in formulating the appropriate feature selection framework that lessens the prediction uncertainties. Much of the available techniques have only been proven to improve point estimate predictions, without considering its effect on uncertainty. In this paper, a feature selection method based on elimination of noisy data is proposed to reduce the Aleatoric uncertainty in a Remaining Useful Life (RUL) prediction problem. Singular Value Decomposition (SVD) technique is employed to denoise sensor data in SVD matrix by filtering higher SVD modes susceptible to contain noise. Then, the “cleaned” Signal to Noise Ratio (SNR) of each feature is calculated and ranked. Features with low SNR, thus with higher noise, are eliminated. We compare the uncertainty level and behavior between a full feature dataset and different percentage of features selected using SVD and SNR. The same comparison is done between our approach and other feature selection methods such as Pearson and Spearman correlations as well as F Regression. The results show that our approach achieved a lower uncertainty degree with generally better prediction performance than the other mentioned methods.KeywordsUncertaintyFeature selectionPHMRULCMAPSS