Feature Measurement Invariance Under Morphological Damage Heterogeneity: A Cross-Dataset Validation Framework for Information-Theoretic Bearing Fault Diagnosis

Abstract

Vibration-based bearing fault diagnosis relies on features extracted from continuous sensor measurements; yet their reliability as diagnostic measurements degrades when damage morphology differs between training and deployment datasets. This article frames domain transfer as a measurement invariance problem: a feature is considered a reliable diagnostic measurement only if it is simultaneously discriminative within a domain and stable across domains of differing damage morphology. We propose a composite feature reliability index, the MI/DKL score, that quantifies this trade-off by dividing the mutual information of a feature with the fault class by its Kullback–Leibler divergence across source and target domains. From a 24-D information-theoretic and nonlinear dynamics feature set, a subset of 12 invariant features is selected via a threshold on the score median. Cross-dataset validation spanning three bearing datasets—MFPT (artificial electro-discharge machined (EDM) defects), Paderborn (natural fatigue progression), and Fraunhofer LBF (artificial EDM defects)—yields six transfer directions covering all morphological pair types. With joint distribution adaptation and 50-iteration stratified bootstrap validation, the invariant subset improves macro-F1 by up to +0.181 in artificial-to-artificial transfers, whereas natural-to-artificial transfers show consistent degradation ( $\Delta = -0.022 \pm 0.025$ ), revealing a conditional structure of feature invariance governed by damage morphology. The discriminability–transferability paradox is introduced as a theoretical contribution to measurement science: features that maximize within-domain discriminability systematically exhibit higher cross-domain distributional divergence, constituting an irreconcilable measurement uncertainty source under morphological heterogeneity.