Early Fault Detection in Rotating Machinery via Multivariate Autoencoder-Based Indicator Fusion

Abstract

This research proposes a normal behaviour multivariate autoencoder, which fuses multiple condition indicators into a single high-level health indicator to provide a comprehensive overview of the mechanical component's health. The model is trained exclusively on healthy data to learn normal behaviour and detect faults by observing deviations from this learned normal behaviour. The proposed method is validated in real time by monitoring run-to-failure bearing experiments on an FE8 bearing test rig.  It is employed to detect blind faults in real time during ongoing experiments, resulting in the termination of the experiment to analyse the cause of fault initiation.  Furthermore, historical data is utilised to quantify the lead time between the proposed method's detection and the final termination triggered by traditional condition monitoring methods. A comparative analysis of the physical bearing damage after the completion of tests demonstrates the capability of the proposed method to detect blind faults at an early stage. The results suggest that this approach identifies fault at an earlier damage stage as compared to traditional methods. It improves the conditions for studying bearing fatigue initiation by avoiding the interference of secondary damage or extensive spalling. In addition, the real-time blind fault detection capability demonstrates the practical application of the proposed framework in preventing catastrophic failures within high-value industrial assets.