Machine learning investigation for wind turbine diagnostics
Location: INSA Rouen Normandy
Salary: approximately 1200€
Duration: 6 months
Starting date:early february or march 2019 (at best)
Gilles Gasso, LITIS – INSA Rouen Normandy, Avenue de l’Université 76 801 Saint Etienne du RouvrayCedex
firstname.lastname@example.org, tel: 02 32 95 98 96
Machine learning, signal processing, failure detection, diagnostics, wind turbine
This internship takes place in the general context of the wind energy development. The development of diagnostic tools and models are some of the cost reduction drivers of the wind applications. The internship aims at applying machine learning algorithms for failure detection and prediction on wind turbine. Intern will be hosted and will be supervised by LITIS, a computer science and information technology laboratory in Rouen. The subject was proposed by a major wind turbine actor through a subcontracting agreement with LITIS.
Description of the subject
The current subject is proposed in the context of R&D effort to design monitoring and diagnostic models for wind turbine surveillance. For this sake, we are interested in investigating the area of machine learning for detecting and identifying abnormal events in times series representative of operating wind turbines. The abnormal events may correspond to damages or failures that will alter or impede the normal behavior of the turbines. Early detection or prediction of these situations may prevent severe impacts on the materials and may help planning maintenance periods.Hence, the pursued goal of the internship is to develop machine learning tools for abnormal behaviour detection.
The signals of interest are collected on operating wind farms over several years and represent frequency structural oscillations. The project intends to survey, to implement and to test relevant online event detection algorithms on these time series. The detection algorithms may range from approaches based on kernel methods , , rank statistics  to those relying on deep learning , . Training with regards to the wind turbine operation and functionalities , ,  will be provided as well as assistance for further understanding of the data and the information collected from the wind farms. Tutorials on machine learning based event detection may also be offered.
1) Apply various machine learning algorithms and methodologies on the provided wind turbine data,
2) Drive conclusions and directions for the development of diagnostic models,
3) Develop diagnostic models when possible.
-Master or Engineer in Computer Science, Electrical Engineering or related fields
-Knowledge in Machine Learning, signal processing
-Programming in Python or Matlab
Email: email@example.com, tel: 02 32 95 98 96
 An online support vector machine for abnormal events detection, Manuel Davy, Frédéric Desobry, Arthur Gretton, Christian Doncarli. Signal Processing, Volume 86, Issue 8, 2006, Pages 2009-2025
 Kernel change-point analysis. Harchaoui, Z., Bach, F. et Moulines, E. In Koller, D., Schuurmans, D., Bengio, Y. et Bottou, L., éditeurs : Advances in Neural Information Processing Systems 21, pages 609–616. MIT Press.
 Homogeneity and change-point detection tests for multivariate data using rank statistics.Alexandre Lung Yut Fong, Celine Levy Leduc, Olivier Cappe. Journal de la SFdS, 2015, 156 (4), pp.133-162
 Deep One-Class Classification. Ruff, L., Vandermeulen, R., Goernitz, N., Deecke, L., Siddiqui, S.A., Binder, A., Müller, E. & Kloft, M.. (2018). Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:4393-4402
 Long Short Term Memory Networks for Anomaly Detection in Time Series. Malhotra, Pankaj & Vig, Lovekesh & Shroff, Gautam & Agarwal, Puneet. (2015). ESANN 2015 proceedings, Bruges (Belgium)
 Wind Energy Explained. J.F. Manwell, J.G. McGowan & A.L. Rogers,Wiley second edition 2012
 Wind Energy Handbook, 2nd Edition, Tony Burton, Nick Jenkins, David Sharpe, Ervin Bossanyi, Willey 2011
 Wind Turbine Control and Monitoring, Luo Ningsu, Vidal Yolanda, Acho Leonardo, Springer 2014
(c) GdR 720 ISIS - CNRS - 2011-2018.