Fault Detection System for Bearings in Electric Motors using Variational Auto Encoders

Abstract

Electric motors play a fundamental role in essential industries such as energy, transport and aeronautics, which require efficient maintenance to ensure productivity. Bearings are the most common failure point, making Prognostics and Health Management of this component crucial for Industry 4.0. This paper introduces a Fault Detection System based on Variational Auto Encoders (VAEs) trained exclusively on healthy vibration data from two public datasets. By analysing the resultant Gaussian distributions the system identifies early indicators of faults. This approach overcomes the common challenge of requiring faulty data for training, while also making it applicable to any other dataset. The study reveals an initial degradation stage in the training datasets, a critical oversight in previous studies, providing a more accurate depiction of bearing degradation profiles.