<![CDATA[Prediction accuracy does not necessarily indicate whether a machine-learning model warns of an impending equipment event at an actionable time. This study evaluates a data-driven, explainable framework using 15 accelerated XJTU-SY bearing tests. We audited 9,216 dual-axis vibration files, constructed 16 leakage-checked features from 9,201 preterminal observations, and compared an RMS score, logistic regression, and random forest using five bearing-disjoint folds. For the primary ten-minute recorded-end-of-test horizon, mean fold average precision was 0.6491, 0.4194, and 0.3098, respectively. Nevertheless, new actionable alarm episodes occurred in only 4, 4, and 6 of 15 held-out bearings, with 9, 32, and 27 premature episodes. An exploratory operating-condition holdout reduced actionable coverage to 0, 3, and 3 bearings. TreeSHAP explanations on 449 observations showed heterogeneous between-bearing feature rankings (median pairwise Spearman correlation 0.488). Feature-group ablation further demonstrated divergence between predictive ranking and alarm timing. These findings support decision-oriented evaluation alongside conventional metrics, but concern laboratory test termination rather than independently verified first damage or demonstrated industrial savings.]]>
Copyrights © 2023
