Anomaly detection in vibration signals for structural health monitoring of an offshore wind turbine
Bel-Hadj, Y.; Weijtjens, W. (2023). Anomaly detection in vibration signals for structural health monitoring of an offshore wind turbine, in: Rizzo, P. et al. European Workshop on Structural Health Monitoring. EWSHM 2022 - Volume 3. pp. 348-358. https://dx.doi.org/10.1007/978-3-031-07322-9_36
In: Rizzo, P.; Milazzo, A. (Ed.) (2023). European Workshop on Structural Health Monitoring. EWSHM 2022 - Volume 3. Springer: Cham. ISBN 978-3-031-07321-2; e-ISBN 978-3-031-07322-9. xxiii, 1069 pp. https://dx.doi.org/10.1007/978-3-031-07322-9, meer
|
| Beschikbaar in | Auteurs |
|
Documenttype: Congresbijdrage
|
| Trefwoord |
|
| Author keywords |
Acceleration measurements; Offshore wind energy; Novelty detection;Machine learning; Autoencoder |
| Abstract |
The current approach for detecting anomalies in acceleration signals relies extensively on feature engineering. Indeed, detecting rotor imbalances in wind turbines starts by first isolating and then assessing the energy of the 1P harmonic, leading to a feature that is efficient but not failure mode agnostic. While different engineered features can be used concurrently, some anomalies in the acceleration signal might remain undetected by the algorithm, even though they are visually noticeable to a human in the signal's spectrogram. Thus, this project aims to build an AI algorithm capable of detecting anomalies in spectrograms, agnostic of their origin, providing an early warning for potential structural issues. The proposed algorithm infers spectrograms of acceleration signals through a deep autoencoder. Anomalies are identified based on a custom reconstruction error. A sensitivity analysis is performed for two types of anomaly, in which waveforms with different energy levels are artificially added to an acceleration signal measured from an offshore wind turbine (OWT). For a 1P harmonic anomaly representing 20% of the total signal energy, the proposed approach yielded an efficiency (AUC) equal to 96% thanks to a novel reconstruction error, which significantly increased the performances. |
|
