Probabilistic Machine Learning for Reliability Analysis of an Engineered System Case Study—Remaining Useful Life Estimation of Battery

Traditional machine learning models are deterministic and often provide point estimates or deterministic predictions. For example, a classification model might say “this engineered system, say an electronic component is not reliable due to its failure very often.” However, there always exists an uncertainty in deterministic machine learning model, which may arise in classification due to data uncertainty and in regression model due to the issue of overfitting and underfitting. To handle these issues, we generally apply an innovative concept of probabilistic machine learning (PML), as it goes beyond these point estimates. It aims to model the uncertainty inherent in predictions by providing a probability distribution over possible outcomes. The output of the PML model is a probability distribution over possible outcomes. This allows for a more nuanced and realistic understanding of complex systems. With a view to this fact, PML offers a powerful framework for reliability analysis of engineered systems, addressing the inherent uncertainties associated with component degradation and system performance. The present work researches the application of PML techniques such as Bayesian modelling to estimate system reliability and predict remaining useful life (RUL). With the help of prior knowledge and observed data, PML quantifies uncertainty in model parameters and predictions. A case study of estimating RUL of Li-ion battery is presented in detail using PML. The advantage of PML models is that it tracks the bias-variance tradeoffs that usually exist in deterministic machine learning (DML) models.