Mitigating Information Overload in Aviation Safety: AI-Driven Hierarchical Tagging and Summarization of NOTAM for Pre-flight Information Bulletin

<p>To address the challenges of aviation safety information overload in Pre-flight Information Bulletin (PIB) systems, this study proposes an intelligent classification framework (ERNIE-DPCNN) that integrates knowledge-enhanced semantic representation with a Deep Pyramid Convolutional Network. Traditional systems relying on rule-based filtering mechanisms suffer from inefficiencies in critical information identification and high risks of human misjudgment. The proposed framework achieves breakthroughs through three technical innovations: (1) An aviation domain-adapted ERNIE model is constructed, leveraging phrase-level masking strategies to enhance semantic representation of compound identifiers; (2) A Deep Pyramid Convolutional Network (DPCNN) is designed to extract multi-granularity features via hierarchical convolution-pooling architecture, optimized with residual connections for long-text gradient propagation; (3) The AdamW optimizer is introduced to dynamically adjust learning rates, improving model convergence efficiency. Evaluated on real-world NOTAM records from airlines, the framework achieves a weighted F1-score of 98.8% in binary classification (Flight Advisory/Restriction) and 91.5% in multi fine-grained classification, outperforming baseline models such as ERNIE-CNN and ERNIELSTM. Ablation studies demonstrated the effectiveness of the domain adaptive masking strategy and dynamic learning rate mechanism. The framework provides an interpretable and scalable technical approach for aviation safety information processing, and its hierarchical feature extraction mechanism facilitates the subsequent simplified deployment scenarios of PIB.</p>