Evidence-Grounded Decision Support for Aircraft Line Maintenance Using Conformal Prediction and Retrieval-Augmented NLP from Technical Log Records

Aircraft line maintenance requires rapid triage of short defect narratives despite variable documentation quality, which can weaken coding consistency and traceability. This study develops and evaluates an evidence-grounded decision-support pipeline using secondary records from the Federal Aviation Administration Service Difficulty Reporting System. The corpus comprised 4,430 deduplicated events from 2021 to 2023 across 10 receiving regions and 290 aircraft make–model combinations, with a heavily imbalanced distribution of Joint Aircraft System/Component (JASC) codes as supervised labels; broader Air Transport Association chapter groupings were used only for post hoc aggregation. Validation used a chronological 70/10/20 train/calibration/test split, with calibration reserved for conformal tuning, plus a station hold-out protocol to assess cross-site transferability. The pipeline combined a lightweight text classifier for JASC triage, term frequency–inverse document frequency (TF-IDF) retrieval to identify label-consistent precedents, and split-conformal prediction with a review-required abstention rule. On the temporal test set, the TF-IDF representation with a linear support vector machine classifier achieved top-1 accuracy of 0.509 and top-3 accuracy of 0.683, while retrieval achieved Recall@10 = 0.688 and mean reciprocal rank = 0.466. Conformal prediction attained empirical coverage of 0.919 at a 90% target. Performance declined directionally under station shift, highlighting the need for monitoring, governance, and human oversight in deployment. Overall, the study contributes a deployment-aligned evaluation protocol integrating retrieval and conformal prediction for auditable line-maintenance triage.