On the Capacity of Deep Autoencoder-Based Normal Behaviour Models in Wind Turbine Condition Monitoring

This study compares Deep Autencoder (AE)-based Normal Behaviour Models (NBM) for anomaly detection in Wind Turbine SCADA data. Using a proprietary industrial dataset, we evaluate performance under realworld challenges, like class imbalance and unseen anomalies. We conduct a systematic comparison across unsupervised, semi-supervised, and supervised approaches, and examine the impact of auxiliary loss functions. Our results show that unsupervised Vanilla AEs struggle to separate normal and abnormal data, different from what the NBM literature claims about the effectiveness of unsupervised setups, as we obtain a significant increase in the Area Under the Curve (AUC) via its classification head. We propose an Adversarial Robust AE (ARAE) to improve detection in data-scarce scenarios. In settings with limited abnormal data, ARAE maintains stable recall at a 1:4 abnormal-to-normal ratio, outperforming other models under severe class imbalance. Based on these results, we recommend a Bottleneck architecture for scenarios with abundant labelled data and ARAE for those with scarce abnormal examples.