Domain-adapted deep learning for aviation incident classification with multiple labels and risk assessment

Aviation safety increasingly relies on narrative incident reports to derive actionable risk intelligence. However, reports from the Aviation Safety Reporting System (ASRS) feature long texts, imbalanced multi-label outcomes, and domain-specific terminology, posing challenges for conventional natural language processing methods. To address this, we introduce a domain-adapted deep learning model based on the Robustly Optimized Bidirectional Encoder Representations from Transformers Pretraining Approach (RoBERTa), tailored for fine-grained classification of 33 ASRS event result categories and their associated consequence-severity levels. The model integrates (1) a long-text segmentation and aggregation mechanism with representation-level fusion of metadata and narrative features; (2) a composite training objective combining binary cross-entropy (BCE), label smoothing, and focal loss to improve robustness under class imbalance; and (3) domain-adaptive pretraining with an extended aviation-specific vocabulary. Extensive experiments on ASRS data (2001–2023) demonstrate that the proposed model consistently outperforms seven baseline architectures, achieving improvements in Micro F1 and Recall, especially for rare yet operationally critical outcomes. Ablation results confirm the distinct and complementary contributions of fusion, loss design, and domain pretraining. Additionally, attention visualization supports interpretability by revealing how the model allocates representational focus across semantically relevant segments. Mapping predicted event results to five consequence-severity levels enables aggregated severity profiling and early warning. Overall, this study presents a reproducible pipeline from unstructured narratives to structured risk assessments, highlighting the effectiveness of domain-adapted Transformer models for proactive aviation safety monitoring and data-driven risk management.

Text 123984.pdf - Published Version Restricted to Repository staff only Download (10MB) Official URL or Download Paper: https://www.sciencedirect.com/science/article/pii/...

Actions (login required)

View Item