REPORTED OUTCOMES OF PREDICTIVE MAINTENANCE AND INDUSTRIAL AI DEPLOYMENTS

Abstract

Published predictive-maintenance outcomes are expressed through technical accuracy, equipment reliability, production performance and financial consequences. These measures describe different stages of an intervention and often use incompatible denominators. This paper develops a structured approach to evaluating and synthesising such evidence. A purposively selected evidence set comprises a US manufacturing survey reported by NIST and three published industrial case studies. The review preserves study design, operational boundary, observation period and source location, and recalculates selected reported changes. The audit identifies arithmetic and reporting inconsistencies that materially affect interpretation: percentagepoint changes are sometimes presented as percentages, one reported mean-time-between-failures improvement differs from the change implied by its own values, and one steel case uses different downtime and cost bases across analytical sections. These findings motivate an estimand ledger linking every effect to a population, intervention, comparator, outcome, horizon and uncertainty measure. The paper specifies when rate ratios, mean differences and response ratios are meaningful and when a structured synthesis is preferable to pooling. It presents explicitly hypothetical meta-analytic calculations to demonstrate weighting, heterogeneity and sensitivity. The evidence supports study-specific reported outcomes and a programme for more comparable evaluation; it does not establish a pooled causal effect or a GCC-wide benefit rate.