Unsupervised Minimum Redundancy Maximum Relevance Feature Selection for Predictive Maintenance

Abstract

Identifying and selecting optimal prognostic health indicators in the context of predictive maintenance is essential to obtain a good model and make accurate predictions. Several metrics have been proposed in the past decade to quantify the relevance of those prognostic parameters. Other works have used the well-known minimum redundancy maximum relevance (mRMR) algorithm to select features that are both relevant and non-redundant. However, the relevance criterion is based on labelled machine malfunctions which are not always available in real life scenarios. In this paper, we develop a prognostic mRMR feature selection, an adaptation of the conventional mRMR algorithm, to a situation where class labels are a priori unknown, which we call unsupervised feature selection. In addition, this paper proposes new metrics for computing the relevance and compares different methods to estimate redundancy between features. We show that using unsupervised feature selection as well as adapting relevance metrics with the dynamic time warping algorithm help increase the effectiveness of the selection of health indicators for a rotating machine case study.