This study addresses human factors in aviation maintenance by converting routine e-log text into computable communication-resilience indicators – closure-loop ratio, read-back adherence, ambiguity density, temporal/referential completeness, error-catch latency, and cross-shift continuity – and testing whether strengthening these signals reduces defects with minimal operational burden. An integrated design-and-validation pipeline was deployed in a Maintenance, Repair and Overhaul (MRO) setting using a phased rollout (Baseline → Assist → Nudge), and causal effects were estimated via interrupted time-series analysis and, where applicable, stepped-wedge Generalized Linear Mixed Model (GLMM). A Natural Language Processing (NLP) stack (Term Frequency–Inverse Document Frequency (TF-IDF) + regularized logistic regression, with an optional compact transformer) extracts linguistic cues; the predicted probabilities are calibrated to support reliable dashboard thresholds. Results show immediate reductions in level and sustained improvements in slope in sign-off error rates after Assist, with larger step-downs under Nudge. Mediation analyses indicate that gains operate through improved communication KPIs rather than generic attentional effects. Model diagnostics light-strong discrimination with low calibration error; robustness checks and a cross-shift/fleet evaluation show stable transfer with minimal recalibration. Governance emphasizes de-identification, advisory-only AI with human-in-the-loop, and transparent, non-punitive use. Findings operationalize Safety-II as quantifiable communication behavior and demonstrate a scalable, low-friction pathway – advisory Assist plus light User Interface (UI) nudges – that advances Air Transport Technologies & Development while improving safety and quality in maintenance operations.
Arthur C. Dela Peña (Fri,) studied this question.