Abstract
High-fidelity physics simulations such as finite element analyses (FEA) provide accurate solutions to complex physical problems that are encountered in equipment subjected to transient conditions. However, such simulations come with significant computational expense, which often prevents them from being used for condition monitoring as a ‘digital twin’. Physics-based reduced order models (ROMs) seek to reduce this computational cost by using classical equations, augmented with empirical factors, to predict some unknown physical quantity of interest over time. Such ROMs are fast to run and simple to tailor for different load cases, but often require significant bespoke development effort. Additionally, developing the formulation of such ROMs becomes increasingly difficult as the complexity of the modelled physical phenomena increases.
Advances in Artificial Intelligence and Machine Learning have resulted in techniques that are ideally placed to address the shortfalls of ROMs and maintain their benefits. A common problem that a physics-based ROM attempts to solve is predicting an unknown time-varying parameter (such as stress) based on a set of known time-varying parameters. Recurrent Neural Networks (RNNs) are well suited to tackle exactly this problem as an RNN-ROM.
An RNN-ROM has been developed and validated against finite element analyses for a piece of rotating machinery subjected to complex thermal transients. Transient thermal datasets were extracted from locations of interest in an FEA representation of a rotating component and used as inputs to the RNN-ROM to predict time-dependent transient thermal behavior. The results were compared to the high-fidelity finite element model from which the input data was extracted; these results showed significantly improved accuracy compared to physics-based ROMs.
