The Role of Reliability Knowledge in Unsupervised Fault Detection of Wind Turbines

Unsupervised learning techniques have become essential tools for data-driven fault detection in complex industrial systems. However, their effectiveness depends strongly on how the input space represents the underlying physical and causal relationships of degradation. This study investigates the role of incorporating reliability knowledge, derived from analyses such as Failure Mode and Effects Analysis, Failure Mode and Symptoms Analysis, or Failure Mode and Observability Analysis, in shaping the data representation underlying the effectiveness and interpretability of anomaly detection models. The analysis focuses on the failure-mode level, comparing generalist models, trained with all available monitoring variables, to knowledge-guided specialist models, trained only with variables causally associated with specific failure mechanisms. Two unsupervised algorithms, including One-Class Support Vector Machine and Isolation Forest, are employed as neutral demonstration vectors, so that the observed effects can be attributed primarily to the structure of the feature space rather than to algorithmic characteristics. Performance is assessed using open SCADA data from wind turbines through metrics such as false alarm rate, early warning time, and alarm persistence. Additionally, geometric distances between the subspaces learned by generalist and specialist models are analysed to quantify structural divergence in feature representation. The study aims to assess whether knowledge-guided specialization can facilitate the emergence of fault-related deviations in unsupervised settings and support more interpretable model behaviour. By examining the consistency of these effects across different learning algorithms, the study provides quantitative insights into how reliability knowledge shapes feature spaces toward physically meaningful dimensions of system behaviour.

Keywords: Unsupervised fault detection, reliability knowledge, feature selection, failure modes, nominal space representation, anomaly detection, SCADA data, wind turbines.