Degradation Modelling and Prognostics of Rotating Equipment with Automated Machine Learning

Abstract

Evaluating and predicting equipment degradation is essential to the reliability and safety of mechanical equipment systems. Condition monitoring based on sensors measuring vibration signals has emerged as the most well-established method for enabling detection and prediction of the health state of rotating equipment. One of the most well-established methods for degradation modelling based on vibration signals is Weibull distribution. Most recent approaches take advantage of Machine Learning (ML) and deep learning to generate prognostics insights. On the other hand, AutoML automates the configuration of ML pipelines to achieve a higher level of automation. Therefore, it has the credentials to enable predictive maintenance applications by optimizing the adopted predictive models; however, the exploitation of AutoML in rotating machine prognostics is an underexplored area. In this paper, we model the degradation procedure according to the Weibull distribution, we extract, combine, and compare various time-domain features, and we apply the H2O AutoML framework for providing predictions about the future health state of the rotating machinery. Our proposed approach is validated in a real-life scenario of cold rolling in the steel industry. Then, we implement and compare various pipelines with different combinations of features in terms of R-squared, MAE, and execution time.