WrenchBuddy: A Governance Framework for Human-Centered Industrial AI Fault Diagnostics

Abstract

Industrial artificial intelligence (AI) fault-diagnostic systems can identify plausible causes under uncertainty, but their outputs alone do not determine how diagnostic support should be delivered during maintenance. This paper presents WrenchBuddy, a human-centered governance framework that manages three decisions during a fault episode: how much diagnostic structure to expose, what assistance posture to provide, and which first confirmatory question to prioritize. The framework is method-agnostic. In this illustration, its roles are instantiated using capped Bayesian-network views, a Comprehensive Operational Cost burden proxy, Fault-Situation Difficulty, a scenario-level physiological readiness input, Data Envelopment Analysis, a Help/Escalate rule, and Value of Information query ranking. A maintenance corpus from approximately 300 remotely monitored uninterruptible power supply machines provides 429 tagged incidents, while an independent physiological dataset provides the scenario-level readiness input. The reconstructed cause–alarm graph has 35 nodes and 34 edges. The Bounded view retains approximately 85% of Baseline coverage with approximately 57% of its interpretability burden. At the selected operating point, three alarms are escalated; bootstrap analysis shows that the two highest-difficulty decisions are stable, whereas the third is a borderline result based on seven incidents. Because the two datasets are not synchronized, the study evaluates governance-policy behavior rather than operational-outcome improvement. WrenchBuddy therefore provides an auditable framework for governing diagnostic exposure, assistance, and first-query selection, while synchronized human-subject validation remains future work.