Human Experts may diagnose the state of a rotating machine by listening to its vibration sounds. Variational Auto-encoder (VAE) is an alternative method to realize the diagnosis by AI learning. Previous studies have shown that when the normal state can be assumed to be a single Gaussian distribution, anomaly scores can be calculated as deviations from the center of the distribution in the VAE latent space. However, when the normal state consists of different modes corresponding to operating conditions, the calculation may not be a simple task. As a way to solve this problem, the use of Conditional Variational Auto-encoder (CVAE) which performs VAE learning including operating conditions, seems promising. In this study, we show verification results that the anomaly score based on the normalized Euclidian distance in the CVAE latent space can detect anomalous conditions using synthesized data and real acceleration measurement data with rotation speed changes.

19^th IMEKO TC10 Conference "Measurement for Diagnostics, Optimisation and Control to Support Sustainability and Resilience"