Current and Vibration Signal Feature Engineering for Defect Classification in Rotary Machines

Use of current and vibration signal analysis for fault detection and condition monitoring of rotary machines is well understood in literature. However, to perform machine learning driven fault analysis in rotary machines, the entire machine learning pipeline of data collection, cleaning, de-noising, feature engineering, model training and hyper-parameter tuning has to be implemented. Among all these steps, the feature engineering becomes most critical because of multiple reasons. First being the fact that we do not directly incorporate the time series but instead translate it into encoded variables. This can lead to a large explosion of features which in presence of limited data may become an ill-posed problem. Therefore, reduction of features to the most appropriate ones is not only important for developing a simpler model but may become a necessity due to limited data. In this article, various methods of feature engineering have been deployed based on the underlying machine learning model. It has been found that there are several features that are common irrespective of the underlining model. The article provides a discussion on the possible reasons for the specific subset selection.