Using Feature Selection to Determine Optimal Depth for Wavelet Packet Decomposition of Vibration Signals for Ocean System Reliability

Abstract

Vibration signals are an important source of information for machine condition monitoring/prognostic health monitoring to ensure the reliability of ocean systems. Because they are waveforms, vibration data must be transformed into the frequency domain before they can be used to build classification and prediction models. One popular transformation is wavelet packet decomposition, a higher resolution variant of wavelet transformation. For wavelet packet decomposition, depth is an important parameter to control the maximum level of detail while minimizing the computational time when constructing and using the decomposition tree. Little guidance exists in the literature to assist researchers in choosing a depth, however. In this paper, we present a feature selection-based approach to determining the optimum depth for wavelet packet decomposition. First, the data is transformed using a very high depth, and all of the features are ordered based on their importance for predicting the class. Then, a depth which captures the most important features is chosen. Finally, a model is built using that depth. We show that a classification model built according to this procedure retains almost all of the accuracy of models built using a much deeper transform, while allowing for smaller depths and vastly fewer features.