Missing Data Repairing for Bearing Vibrations using Generative Adversarial Networks

Abstract

A vibration data repair method based on Generative Adversarial Networks (GAN) is proposed to resolve the problem of incomplete data acquisition of bearing vibration data under certain circumstances (sensor failure, extreme environments, etc.), which leads to errors in data analysis. We use a GAN framework, combined with an Auto Encoder (AE), to become an Auto Encoder-Generative Adversarial Networks (AE-GAN) to generate synthetic data related to data interpolation. First, an Auto Encoder is introduced in the generator of the GAN to reconstruct the input with missing data by encoding and decoding. Then, the reconstructed data is continuously trained adversarially with the original data in the discriminator of the Generative Adversarial Networks. Finally, enabling the proposed model to generate interpolated data close to the actual data. The algorithm validity with the bearing vibration dataset from the IEEE PHM 2012 Predictive Challenge was verified, and the results showed that: for missing vibration datasets, the AE-GAN algorithm has better repair accuracy and convergence speed than traditional algorithms; the model is more stable for GAN training because of the addition of Auto Encoder; providing new ideas for deep learning research on industrial data.