The task of predicting how long a certain industrial asset will be able to operate within its nominal specifications is called Remaining Useful Life (RUL) estimation. Efficient methods of performing this task promise to drastically transform the world of industrial maintenance, paving the way for the so-called Industry 4.0 revolution. Given the abundance of data resulting from the advent of the digitalization era, Machine Learning (ML) models are the ideal candidates for tackling the RUL estimation problem in a fully data-driven fashion. However, given the safety-critical nature of maintenance operations on industrial assets, it's crucial that such ML-based methods be designed such that their levels of transparency and reliability are maximized. Modern ML algorithms, however, are often employed as black-box methods, which do not provide any clue regarding the confidence level associated with their output. In this paper, we address this limitation by investigating the performance of a recently proposed class of algorithms, Deep Gaussian Processes, which provide uncertainty estimates associated with their RUL prediction, yet retain the expressive power of modern ML techniques. Contrary to standard approaches to uncertainty quantification, such methods scale favourably with the size of the available datasets, allowing their usage in the "big data" setting. We perform a thorough evaluation and comparison of several variants of DGPs applied to RUL predictions. The performance of the algorithms is evaluated on the NASA N-CMAPSS (New Commercial Modular Aero-Propulsion System Simulation) dataset for aircraft engines. The results show that the proposed methods are able to yield very accurate RUL predictions along with sensible uncertainty estimates, providing more reliable solutions for (safety-critical) real-life industrial applications.
Uncertainty-Aware Prognosis via Deep Gaussian Process
Biggio, Luca;
2021
Abstract
The task of predicting how long a certain industrial asset will be able to operate within its nominal specifications is called Remaining Useful Life (RUL) estimation. Efficient methods of performing this task promise to drastically transform the world of industrial maintenance, paving the way for the so-called Industry 4.0 revolution. Given the abundance of data resulting from the advent of the digitalization era, Machine Learning (ML) models are the ideal candidates for tackling the RUL estimation problem in a fully data-driven fashion. However, given the safety-critical nature of maintenance operations on industrial assets, it's crucial that such ML-based methods be designed such that their levels of transparency and reliability are maximized. Modern ML algorithms, however, are often employed as black-box methods, which do not provide any clue regarding the confidence level associated with their output. In this paper, we address this limitation by investigating the performance of a recently proposed class of algorithms, Deep Gaussian Processes, which provide uncertainty estimates associated with their RUL prediction, yet retain the expressive power of modern ML techniques. Contrary to standard approaches to uncertainty quantification, such methods scale favourably with the size of the available datasets, allowing their usage in the "big data" setting. We perform a thorough evaluation and comparison of several variants of DGPs applied to RUL predictions. The performance of the algorithms is evaluated on the NASA N-CMAPSS (New Commercial Modular Aero-Propulsion System Simulation) dataset for aircraft engines. The results show that the proposed methods are able to yield very accurate RUL predictions along with sensible uncertainty estimates, providing more reliable solutions for (safety-critical) real-life industrial applications.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
