Multivariate calibration for machine health monitoring: kernel partial least squares combined with variable selection

Abstract

This work proposes and demonstrates the use of data mining techniques for machine health monitoring through a multivariate calibration model. It utilizes a genetic algorithm (GA)-based variable selection combined with a preprocessing technique of orthogonal signal correction (OSC) for constructing reliable calibration models of shaft misalignment conditions. Improper aligning of shafts often leads to severe problems in many rotating machines. Thus the prediction of shaft alignment conditions is quite essential in making decisions on when to perform alignment maintenance. The main goal of this calibration model is to predict misalignment conditions from historical data. A case study using real misalignment data showed that the prediction results of the proposed calibration models improved significantly compared to existing calibration models. As an extension of linear calibration models, a nonlinear kernel calibration model was also presented. It turned out that linear and nonlinear calibration models of shaft misalignment conditions produced better prediction performance through the use of GA-based variable selection combined with OSC.