Multi-Label Classification of Aviation Safety and Reliability Reports using SafeAeroBERT Fine-Tuned with Low-Rank Adaptation (LoRA)

The Aviation Safety Reporting System (ASRS), maintained by the National Aeronautics and Space Administration (NASA), contains a large and continually expanding collection of incident narratives submitted by aviation safety professionals. Manual review of this corpus is increasingly impractical and time-consuming, motivating automated methods that can extract structured safety information from free-text reports. This research presents SafeAeroBERT-Prime, a domain-specific transformer model pre-trained on ASRS and National Transportation Safety Board (NTSB) narratives and fine-tuned with Low-Rank Adaptation (LoRA) for multilabel classification of primary safety factors from narrative scenarios. Cross-validation results indicate that SafeAeroBERT-Prime provides stable and reproducible performance across evaluated data partitions, demonstrating that parameter-efficient fine-tuning can identify meaningful safety-factor patterns even under constrained data and computational resources. Beyond classification, this work establishes a foundation for integrating natural-language processing (NLP) into aviation safety-analysis workflows. A continuously updating machine-learning (ML) pipeline and an interactive dashboard integrated with a chatbot interface are under development to connect model predictions with failure-mode information and component-reliability data. Containerized deployment will support reproducibility and integration into existing engineering environments. For aviation reliability, maintainability, and safety (RMS) engineers, these capabilities can reduce the effort required to identify recurring safety hazards, prioritize relevant reports, and detect emerging safety trends. For the broader aerospace community, SafeAeroBERT-Prime illustrates how domain-adapted, computationally accessible language models may support more proactive, evidence-informed safety risk management (SRM). The system is intended to augment, not replace, established safety processes and expert judgment; accordingly, operational use requires ongoing validation, performance monitoring, and human oversight.