Enhancing real-time remaining useful life prediction with information entropy & uncertainty quantified deep learning models

Abstract

Accurate prediction of the Remaining Useful Life (RUL) of critical components in industrial applications is essential for optimizing maintenance strategies and ensuring operational safety. Traditional methods often struggle with the complexities of real-time data integration and fail to provide uncertainty measures critical for high-stakes decision-making. This study introduces a novel deep learning framework that not only predicts RUL in real-time but also quantifies the uncertainty of these predictions, enhancing the reliability of the prognosis. To demonstrate this approach, neural network architectures were developed to understand real-time inputs of time-series type acoustic emission nondestructive evaluation data. The key innovation of the approach is the use of information entropy with window-analysis to analyze streaming data. The combination of window analysis and deep learning allows for autonomous analysis for the user. Furthermore, a digital thread was developed to predict RUL as data was streamed. To address the challenge of uncertainty, the model architectures incorporate Monte Carlo dropout, providing a probabilistic interpretation of the predictions. Experimental results demonstrate that the approach not only improves the accuracy of RUL estimates compared to traditional models but also offers meaningful uncertainty bounds which are vital for risk-averse industries.