Selection of Input Parameters for Multivariate Classifiers in Proactive Machine Health Monitoring by Clustering Envelope Spectrum Harmonics

In condition monitoring (CM) signal analysis the inherent problem of key characteristics being masked by noise can be addressed by analysis of the signal envelope. Envelope analysis of vibration signals is effective in extracting useful information for diagnosing different faults. However, the number of envelope features is generally too large to be effectively incorporated in system models. In this paper a novel method of extracting the pertinent information from such signals based on multivariate statistical techniques is developed which substantialy reduces the number of input parameters required for data classification models. This was achieved by clustering possible model variables into a number of homogeneous groups to assertain levels of interdependency. Representatives from each of the groups were selected for their power to discriminate between the categorical classes. The techniques established were applied to a reciprocating compressor rig wherein the target was identifying machine states with respect to operational health through comparison of signal outputs for healthy and faulty systems. The technique allowed near perfect fault classification. In addition methods for identifying seperable classes are investigated through profiling techniques, illustrated using Andrew’s Fourier curves.