The reliability of gas turbine diagnostics clearly relies on reliable measurements. However, raw data reliability can be corrupted by label noise issues, as for instance an erroneous association between data and the respective unit of measure. Such issue, rarely investigated in the literature, is named Unit of Measure Inconsistency (UMI). Machine Learning classifiers are suitable tools to tackle the challenge of UMI detection. Thus, this paper investigates the capability of four Support Vector Machine approaches to detect UMIs. All approaches are tested on a dataset composed of field data taken on a fleet of Siemens gas turbines. The results of this study demonstrate that the Radial Basis Function with One-vs-One decomposition allows higher diagnostic accuracy.
Keywords: Anomaly detection; Data reliability; Gas turbine; Label noise; Support Vector Machine; Unit of Measure Inconsistency.
Copyright © 2021 ISA. Published by Elsevier Ltd. All rights reserved.