Human Factors in Aviation Safety using NLP

📄 Research Paper

Human Factors in Aviation Safety: NLP on Pilot Reports

• 📌 Published in IEEE
• 📊 Indexed in Scopus

🔗 DOI: https://doi.org/10.1109/ICAAIC64647.2025.11329454 🔗 IEEE Link: https://ieeexplore.ieee.org/document/11329454

---

🧠 Overview

This research focuses on analyzing aviation safety by detecting human factors from pilot incident reports using advanced Natural Language Processing (NLP) and deep learning techniques. The system processes unstructured ASRS (Aviation Safety Reporting System) narratives to identify critical issues such as situational awareness failures, communication breakdowns, and cognitive overload.

---

⚙️ Methodology

📊 Data Source

• NASA ASRS (Aviation Safety Reporting System)
• ~2000 balanced pilot incident reports

🧹 Preprocessing Pipeline

• Text cleaning and normalization

• Lemmatization (WordNet)

• Tokenization and padding (max length = 600)

• Data augmentation:

  • Synonym replacement
  • Random word swapping

---

🤖 Model Architectures

Three hybrid deep learning models were designed and compared:

🔹 Model 1 (Best Performing)

CNN + BiLSTM + BiGRU + Attention

🔹 Model 2

CNN + BiLSTM + Transformer + Multi-Head Self-Attention

🔹 Model 3

CNN + BiGRU + Transformer + Multi-Head Self-Attention

• Word embeddings: Word2Vec (300 dimensions)
• Loss Function: Focal Loss
• Optimizer: Adam

---

📈 Results

Model	Accuracy	Precision	Recall	F1 Score
Model 1	99.62%	100%	99.25%	99.62%
Model 2	99.25%	99.70%	98.80%	99.25%
Model 3	98.89%	99.10%	98.20%	98.89%

👉 Model 1 outperformed transformer-based models by effectively capturing both local and sequential dependencies.

---

🎯 Objective

To automate the detection of human-factor-related risks in aviation reports and support predictive safety analysis.

---

🚀 Tech Stack

• Python
• TensorFlow / Keras
• NLTK
• Word2Vec
• Deep Learning (CNN, BiLSTM, GRU, Attention, Transformer)

---

📊 Key Contributions

• Developed hybrid deep learning architectures for aviation NLP
• Designed a robust preprocessing + data augmentation pipeline
• Achieved state-of-the-art accuracy (99.62%)
• Improved interpretability using attention mechanisms

---

📌 Note

This repository contains the author's preprint version. The final published version is available via IEEE and linked above.