Dimensionality reduction via variables selection – Linear and nonlinear approaches with application to vibration-based condition monitoring of planetary gearbox

摘要: Abstract Feature extraction and variable selection are two important issues in monitoring diagnosing a planetary gearbox. The preparation of data sets for final classification decision making is usually multi-stage process. We consider from gearboxes, one healthy the other faulty state. First, gathered raw vibration time domain have been segmented transformed to frequency using power spectral density. Next, 15 variables denoting amplitudes calculated spectra were extracted; these further examined with respect their diagnostic ability. applied here novel hybrid approach: all subset search by multivariate linear regression (MLR) shrinkage least absolute operator (Lasso) performing non-linear approach. Both methods gave consistent results yielded subsets or properties.