Information Generative Bayesian Adversarial Networks: A Representation Learning Model for Transmission Gear Parameters

Abstract

Representation learning of transmission gear parameters has been a challenging industrial problem due to its difficulty in learning the inherence of gear data with high dimension and coupling characteristics. To solve this issue, this article presents information generative Bayesian adversarial networks (IGBAN) to learn the interpretable representation of gear parameters with insufficient data and improve the effectiveness of evaluating corresponding gear reliability with the learned representation. In particular, to offset the assumption of Gaussian distribution in the generative adversarial network (GAN), a Bayesian discriminator is designed to distinguish original from generated samples. Then, we establish a Q-net to obtain information gain measuring the difference between original samples and learned representations. Furthermore, in IGBAN, the performance on gear reliability by the variation of key parameters is discovered through maximizing mutual information between latent representations and the reliability. Experimental results on real-world transmission gear data validate the effectiveness of our proposed model.