P5.2 Gear-Oil Condition Classification by Means of Support Vector Machine, a first approach
- Event
- 16. GMA/ITG-Fachtagung Sensoren und Messsysteme 2012
2012-05-22 - 2012-05-23
Nürnberg, Germany - Chapter
- P5 Chemische Sensoren und Analysesysteme
- Author(s)
- D. Dorigo, B. Wiesent, A. Pérez Grassi, A. Koch - Technische Universität München
- Pages
- 780 - 787
- DOI
- 10.5162/sensoren2012/P5.2
- ISBN
- 978-3-9813484-0-8
- Price
- free
Abstract
During the last years the amount of energy gained from renewable energy sources was steadily increasing and this is still under development. A big role is played by offshore wind parks, which use the wind energy over the see for electricity generation. In order to reduce the high maintaining costs of these plants, different online condition monitoring systems are currently being developed. Especially condition monitoring of gear-oils promises to have an important impact on such maintaining costs. A well-known off-line method for gear-oil quality monitoring is based on the evaluation of their Infrared (IR) spectra.
The analysis of IR spectra, which is performed in specialized laboratories together with other chemical and mechanical tests, is a common tool for lubricant quality evaluation. Important oil parameters affect the IR spectra and can be therefore deduced and classified by spectral analysis. A significant parameter is the Total Acid Number (TAN). The importance of the TAN value lies in its relation with the oil’s “age”.
In this paper, a preliminary study of gear-oil classification by means of TAN in combination with Support Vector Machines (SVMs) is presented. The Mid-Infrared-spectra (MIR-spectra) of three commercially used gear-oils, a mineral and two synthetic ones, were analyzed and classified according to the range of their TAN value. The classification was performed using up to five classes and tested using full cross validation. The classification results were compared using different kernel functions. The robustness against noise is tested by adding different noise levels to the spectra before classification. This latter leads to a first feasibility check for classifying spectra gained in field application, under rough conditions.