Fault Detection of Rolling Bearings using Real NVP Transformation

Abstract The scarcity of rolling bearing fault data and the difficulty in annotating fault types prompt us to employ unsupervised fault detection methods. However, existing unsupervised fault detection methods rely on manual feature extraction or use indirect approaches such as reconstruction error or density approximation for fault detection. To address this issue, this paper proposes an unsupervised rolling bearing fault detection method based on a Normalizing flow model, namely the Real-Valued Non-Volume Preserving (Real NVP) model. Compared to existing methods, this approach utilizes vibration data from normal bearings for model training. It transforms the complex distribution of normal bearing data into a simple prior distribution through the Real NVP transformation. Subsequently, fault detection indicators are constructed within this simple prior distribution to achieve rolling bearing fault detection. Experiments on the benchmark bearing dataset demonstrate that the proposed method can effectively identify the fault states of rolling bearings and achieve fault detection results superior to other fault detection methods.