Maintaining model efficiency, avoiding bias and reducing input parameter volume in compressor fault classification

With the exponential growth in data collection and storage and the necessity for timely prognostic health monitoring of industrial processes traditional methods of data analysis are becoming redundant. Big data sets and huge volumes of inputs give rise to equally massive computational requirements. In this paper the differences in input parameter selection using a subset of the original variables and using data reduction techniques are compared. Each selection procedure being illustrated by both statistical methods and machine learning techniques. It is shown that the subsequent classification models are highly comparable. Finally the merits of a combined multivariate statistical and wavelet decomposition approach is considered. Techniques are applied to output signals from an experimental compressor rig.