The data substrate of predictive maintenance: a review of benchmark provenance, label derivation and evidence coverage

Abstract

Predictive maintenance research is evaluated almost entirely on a small stock of public benchmark datasets, yet the properties of that stock are rarely examined as objects of study. This review characterizes sixteen widely used predictive-maintenance benchmarks along five provenance attributes: the origin of the record, the derivation of the label, the density of genuine end-of-life events, the span of operating conditions, and the coverage of channels that describe the maintenance decision rather than the machine. Coding these attributes reveals a corpus that is structurally narrow. Every benchmark carries at least one condition-monitoring channel, but only a quarter carry any channel describing the maintenance action, the parts consumed, the energy used or the disposition of the replaced component. Seven of the sixteen carry labels that were never observed on a real in-service failure, being derived instead from a simulation model or from a fault deliberately machined into a test specimen. Not one benchmark in the corpus records energy consumption, and only one records what happened to the component after removal. We then introduce a claim-evidence admissibility mapping that states, for each class of claim made in the predictive-maintenance literature, which data properties are necessary to support it, and we show that claims about energy saving, carbon reduction, lifecycle extension and human-machine collaboration have no admissible substrate in the current stock. This suggests a complementary explanation for the persistent sustainability deficit reported in framework-level gap analyses: the field may be unable to demonstrate environmental benefit not because its architectures exclude it, but because nothing it measures would reveal it. We set out the minimum contents of a sustainability-capable maintenance dataset and a corresponding research agenda.