Uncertainty is a vital indicator in evaluating the prediction of real-world deep learning models. Without uncertainty, important decision-making process linked to safety, security and investment cannot be executed correctly by the users. While probabilistic deep learning methods have shown promise in quantifying prediction uncertainty, very few works in managing these uncertainties can be found in the literature. In failure prediction especially, it is favourable that the model produces the lowest level of prediction uncertainty, i.e., the highest level of confidence in order to initiate any degree of problem-solving mechanism. In this work, we present a technique to reduce Aleatoric uncertainty associated with noisy data using Singular Value Decomposition (SVD). Sensor data from industrial assets are presented in SVD matrix form and higher SVD modes more susceptible to contamination are eliminated to denoise the data. We compare the uncertainty level between the original dataset with the ones denoised by SVD and several known denoising methods in a Remaining Useful Life (RUL) prediction problem. Our results show that SVD denoising treatment outperforms the other denoising methods in reducing prediction uncertainty and improving prediction performance. © 2022, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.