Anomaly detection for wind turbine pitch bearings via autoencoder enhanced nonlinear autoregressive model

Chao Zhang, Long Zhang

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

122 Downloads (Pure)

Abstract

The pitch bearing, as a safety-critical unit in wind
turbines, is prone to damage. To prevent severe accidents,
early anomaly detection for wind turbine pitch bearings is
highly desirable. The two main challenges need to be solved:
1) non-stationary signals under the condition of slow rotating
speed. 2) noise from electric and mechanical movement signals
contaminating the measured vibration signals. To address these
two challenges, a novel method, autoencoder (AE) enhanced
nonlinear autoregressive (NAR) model, namely AE-NAR, has
been proposed. Firstly, the autoencoder possesses the ability of
global feature extraction and can be utilized for the processing of
non-stationary signals. Secondly, the sparse representation of the
NAR model with Bayesian Augmented Lagrangian can deal with
the noise problem. The effectiveness of the model is first validated
using open bearing datasets, then verified using real signals
collected from industrial-scale wind turbine pitch bearings. The
results show that the AE-NAR method can effectively detect the
abnormal state of wind turbine pitch bearings.
Original languageEnglish
Title of host publicationThe third International Conference on Sensing, Measurement & Data Analytics in the era of Artificial Intelligence (ICSMD 2022)
Publication statusAccepted/In press - Oct 2022

Keywords

  • predictive maintenance
  • nonlinear autoregressive model
  • condition monitoring
  • anomaly detection
  • wind turbine

Fingerprint

Dive into the research topics of 'Anomaly detection for wind turbine pitch bearings via autoencoder enhanced nonlinear autoregressive model'. Together they form a unique fingerprint.

Cite this