Deep digital twins for detection, diagnostics and prognostics

Abstract

A generic framework for prognostics and health monitoring (PHM) which is rapidly deployable to heterogeneous fleets of assets would allow for the automation of predictive maintenance scheduling directly from operational data. Deep learning based PHM implementations provide part of the solution, but their main benefits are lost when predictions still rely on historical failure data and case-by-case feature engineering. We propose a solution to these challenges in the form of a Deep Digital Twin (DDT). The DDT is constructed from deep generative models which learn the distribution of healthy data directly from operational data at the beginning of an asset’s life-cycle. As the DDT learns the distribution of healthy data it does not rely on historical failure data in order to produce an estimation of asset health. This article presents an overview of the DDT framework and investigates its performance on a number of datasets. Based on these investigations, it is demonstrated that the DDT is able to detect incipient faults, track asset degradation and differentiate between failure modes in both stationary and non-stationary operating conditions when trained on only healthy operating data.