Unsupervised Machine Learning Anomaly Detection for Multivariate Time-Series Data in Wind Turbine Converters
Yu, Chuang (2020)
Yu, Chuang
2020
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2020060316640
https://urn.fi/URN:NBN:fi:amk-2020060316640
Tiivistelmä
Because wind power is one the main clean energy sources, the demand for wind generated energy has been rapidly increasing all over the world. As wind turbine converter is one of the key components in wind turbine, it is critical to ensure the reliability of its operation without human monitoring in addition to cost efficiency. This thesis studies and experiments two unsupervised machine learning models to detect anomaly turbine converters: Hidden Markov Model and Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) model. The aim is to compare the results from the two selected models and cross validate the results with existing models and visualized graphs for data analytics. With Hidden Markov Model, three distance computation methods are applied: Discrete Frechet, Dynamic Time Warping (DTW) and Partial Curve Mapping (PCM). The experiments show that PCM is the fastest but also produces the worst result, while Discrete Frechet is the slowest, and produces the similar result as DTW. HDBSCAN is very intuitive to use and relatively fast to produce the clusters, and it works exceptional good on certain data Analysis Group. The experiment results show that both models do not provide satisfactory result compared to the existing models.