Fault Detection Based on Vibration Measurements and Variational Autoencoder-Desirability Function

In the field of electrical machines maintenance, accurate and timely diagnosis plays a crucial role in ensuring reliability and efficiency. Variational autoencoder (VAE) techniques have emerged as a promising tool for fault classification due to their robustness in handling complex data. However, the inherent nondeterministic aspect of the VAE creates a significant challenge as it leads to varying cluster locations for identical health states across different machines. This variability complicates the creation of a standardized applicable diagnostic tool and challenges for the implementation of effective real-time health monitoring and prognostics. Addressing this issue, a novel approach is proposed wherein a desirability function-based term is integrated into the cost function of the VAE. The enhancement achieved by this approach arises from the standardization of classification, guaranteeing that analogous faults are assigned to identical geolocations within a 2-D user-friendly space. This method's efficacy is validated through two separate case studies: one analyzing vibration data from two diverse designs of large existing hydrogenerators, and the other utilizing vibration data sourced from an open-access dataset focused on bearing fault. The findings of both studies show that the model can cluster 97% of similar faults into preset zones, compared with 40% when the desirability term is excluded.