Condition health monitoring of dynamic systems based on vibrationsignatures has generally relied upon Fourier based analysis as a meansof translating vibration signals in time domain into the frequencydomain. The wavelet packet transform is introduced as an alternativemeans of extracting time-frequency information from vibrationsignatures. Moreover, with the aid of statistical based featureselection criteria, many feature components containing littlediscriminant information could be discarded resulting in a featuresubset with reduced number of parameters. This significantly reduces thelong training time that is often associated with neural networkclassifier and increases the generalization ability of the neuralnetwork classifier. To validate the feature extraction algorithmproposed, the simulations have been performed with the benchmark dataknown as Westland vibration data set. The results show significantimprovement when the data is subjected to various white, colored andpink noise
展开▼
