Predictive maintenance models often lack reliable fault labels. An unsupervised adaptation of minimum redundancy maximum relevance ranks health indicators using degradation behavior, then accounts for overlap between features to help identify a compact, useful subset.