Density based Anomaly Detection for Wind Turbine Condition Monitoring

Johannes Bernhard, Johannes Bernhard, Jonas Schmidt, Jonas Schmidt, Mark Schutera, Mark Schutera

2021

Abstract

Unsupervised and explainable approaches are critical in anomaly detection for mechanical systems. This work proposes a density-based k-nearest neighbor method to combine an unsupervised learning setup with the added value of explainability. The algorithm is applied to detect anomalies in vibration data from acceleration sensors or microphones. In a training phase, we transform healthy vibration data into mel-spectrograms and extract feature patches representing healthy turbines’ vibration energy distribution. We determine anomaly scores by calculating a k-nearest neighbor similarity between operational feature patches and healthy feature patches. Hence, we use basic statistical methods with interpretable results, which contrasts with deep learning techniques. The evaluation paradigm is data from damaged and healthy wind turbines and a secondary machine audio data set. This work introduces and explores a novel sensor-level anomaly score. The model identified all damaged sequences as anomalies on the wind turbine sequences. Furthermore, the method achieved competitive results on the more complex DCASE sound anomaly dataset. Concluding, our anomaly score lays the foundations for an interpretable condition monitoring system.

Paper Citation

in Harvard Style

Bernhard J., Schmidt J. and Schutera M. (2021). Density based Anomaly Detection for Wind Turbine Condition Monitoring. In Proceedings of the 1st International Joint Conference on Energy and Environmental Engineering - Volume 1: CoEEE, ISBN 978-989-758-599-9, pages 87-93. DOI: 10.5220/0011358600003355

in Bibtex Style

@conference{coeee21,
author={Johannes Bernhard and Jonas Schmidt and Mark Schutera},
title={Density based Anomaly Detection for Wind Turbine Condition Monitoring},
booktitle={Proceedings of the 1st International Joint Conference on Energy and Environmental Engineering - Volume 1: CoEEE,},
year={2021},
pages={87-93},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0011358600003355},
isbn={978-989-758-599-9},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 1st International Joint Conference on Energy and Environmental Engineering - Volume 1: CoEEE,
TI - Density based Anomaly Detection for Wind Turbine Condition Monitoring
SN - 978-989-758-599-9
AU - Bernhard J.
AU - Schmidt J.
AU - Schutera M.
PY - 2021
SP - 87
EP - 93
DO - 10.5220/0011358600003355