Year: 2025
Type: conference-paper
Abstract: Wind turbine fault monitoring is crucial for reducing operational costs, yet remains challenging due to the scarcity of fault events and turbine-specific operational characteristics.This paper presents a novel two-stage approach combining turbine-specific anomaly detection with cross-turbine fault classification.We train individual autoencoders for each turbine to capture normal operational patterns and use their reconstruction errors as input for a shared InceptionTime classifier.This partitioning reduces the task of the classifier to classifying anomalies instead of general operational data while preserving turbine-specific particularities.Further, our approach can flexibly integrate existing methods for autoencoderbased fault detection and is, in principle, classifier agnostic.Using a comprehensive real-world dataset of 15 turbines with 4,899 repair events spanning at least four years of operation, we demonstrate that our approach consistently outperforms classification based directly on SCADA data.
Authors: Michael Eichelbeck, Max Adam, Kenneth Laureys, Milena Zurmühl, Matthias Althoff
Institution: Technical University of Munich
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Funder: Deutsche Forschungsgemeinschaft
Award: G6030229085
OpenAlex ID: W4411345872