On Data Preprocessing for an Improved Performance of the Sources Classification

Abstract

Data preprocessing is emphasized, as a major step to generate features for data analysis and classification. Data preprocessing is on the feedback loop of data analysis and is driven by experimented users with valuable software tools. In this work, data preprocessing is extended with decorrelation capacity, based on matched data transforms, e.g., discrete cosine transforms. The context is fixed by the availability of power spectra, as input to the classification stage. The preprocessing structure is evaluated with physical recorded signals, representing vibrations generated by faults in the bearings of rotating machines. The structure of the preprocessing is general and can be applied in many other paradigms as machine learning, for the generation of the training sets with independent features.