On Training Data Selection in Condition Monitoring Applications—Case Azimuth Thrusters

Abstract

Machine learning techniques are commonly used in the vibration-based condition monitoring of rotating machines. However, few research studies have focused on model training from a practical viewpoint, namely, how to select representative training samples and operating areas for monitoring applications. We focus on these aspects by studying training sets with varying sizes and distributions, including their effects on the models to be identified. The analysis is based on acceleration and shaft speed data available from an azimuth thruster of a catamaran crane vessel. The considered machine learning algorithm was previously introduced in another study suggesting it could detect defects on the thruster driveline components. In this work, practical guidance is provided to facilitate its implementation, and furthermore, an adaptive method for training subset selection is proposed. Results show that the proposed method enabled the identification of usable training subsets in general, while the success of the previous approach was case-dependent. In addition, the use of Kolmogorov–Smirnov or Anderson–Darling tests for normal distribution, as a part of the method, enabled selections that covered the operating area broadly, while other tests were unfavorable in this regard. Overall, the study demonstrates that reconfigurable and automated model implementations could be achievable with minor effort.