Towards bearing health prognosis using generative adversarial networks: Modeling bearing degradation

Abstract

Condition based maintenance of rotary machines is centered on bearings, as they are the leading source of breakdowns in induction motors used in the industry. The health prognosis of bearings primarily involves the estimation of its remaining useful life (RUL). The accurate modeling of a bearing's degradation is key to correctly estimating its RUL. This paper investigates using generative adversarial networks (GANs) for modeling the degradation behavior of a bearing. GANs are used to estimate generative models, which can be sampled directly to generate the future trajectory of a bearing's health indicator. In the GAN framework, two artificial neural networks, a generator network G and a discriminator network D, engage in a game, where the network G tries to fool the network D by generating samples of data that resemble real data. The training process of GANs finds the Nash equilibrium to this game. The proposed approach for generating future trajectories of a bearing's health indicator is tested using publicly available run-to-failure test data. The results of this preliminary study indicate that the GAN framework is effective in modeling the degradation behavior of bearings.