Feature selection for a real-world learning task

Abstract

Machine learning algorithms used in early fault detection for centrifugal pumps make it possible to better exploit the information content of measured signals, making machine monitoring more economical and application-oriented. The total amount of sensors is reduced by exhausting the information derived from the sensors far beyond the scope of traditional engineering through the application of various features and high-dimensional decision-making. The feature selection plays a crucial role in modelling an early fault detection system. Due to presence of noisy features with outliers and correlations between features a correctly determined subset of features will distinctly improve the classification rate. In addition the requirements for the hardware to monitor the pump decrease therefore its price. Wrappers and filters, the two major approaches for feature selection described in literature [4] will be investigated and compared using real-world data. © Springer-Verlag Berlin Heidelberg 2001.