This research investigates Variational Autoencoders (VAEs) for predictive maintenance (PdM) in buildings, aiming to facilitate their mainstream adoption and address data and technical gaps in existing analytical AI methods. The study developed, trained, and validated a VAE model on raw sensor data from an air handling unit, to assess its efficacy for anomaly detection. Findings revealed the capability of the VAE to autonomously extract features and generate representative data for anomaly detection without data labels. The model is not only capable of data augmentation for PdM but also effectively identifies anomalies and calculates loss values that can be used for fault prioritization. The latter provides promise for integrating criticality analysis into the fault detection process.