A Critical Review and Methodological Framework for Robust Tabular Learning in Industrial Predictive Maintenance

Data-driven predictive maintenance (PdM) relies increasingly on machine learning models trained in sensor and operational data. However, missing measurements, sensor corruption, and distribution shift can complicate the evaluation of predictive reliability under changing industrial conditions. This critical narrative review examines how these data imperfections are addressed in industrial maintenance research and related tabular-learning literature. It critically synthesizes evidence from industrial tabular studies, sequential sensor applications, general-purpose benchmarks, and methodological research, focusing on differences in data characteristics, prediction tasks, stressor mechanisms, and validation procedures. The synthesis identifies important limitations in the available evidence, particularly the lack of comparable controlled factorial evaluations of missingness, measurement corruption, and distribution shift in the industrial tabular PdM studies examined. Consequently, the reviewed findings do not provide a sufficient basis for general conclusions about model robustness under compound imperfections or universal rankings across model families. To support future investigations, this review develops a mechanism-aware conceptual framework distinguishing the three stressors and presents pairwise and three-way loss-interaction contrasts for analyzing potential non-additive effects. It also proposes a leakage-aware evaluation protocol incorporating realistic stressor configurations, appropriate validation strategies, uncertainty estimation, and maintenance-oriented performance measures. The review contributes a critical synthesis and a methodological proposal for future empirical research; neither the proposed framework nor its practical effectiveness is experimentally validated in this study.