Early anomaly detection in wind turbine bolts breaking problem — Methodology and application
Chung‐Wei Wu, Mei Chen · 2018
Early anomaly detection plays an important role in many fields, such as fraud detection in financial data, a signal indicating machine unhealthy status, etc. It leads to fault diagnostics and even prognostics according to data analytics need. In this paper, we apply an early anomaly detection model in wind turbine bolts breaking problem with good detection result, which can be used to substitute prognostic model in a solid manner. Wind turbine bolts breaking is a common problem in wind turbine, which may cause serious impact but hard to detect, especially when no extra sensor on total 50 bolts for each blade. To solve this problem, a novel process is designed to answer following questions: (1) whether there are parameters whose behavior is influenced by the bolts breaking, (2) can we detect such parameters reflecting the breaking problem, and (3) find out how the impacts happen and how far the impacts grow by comparing different number of blots are broken. We develop a conditional detection model, which is able to detect only 2 bolts breaking by automatically extracting key features with 92% accuracy.