PaperlayerView exact revision on arXiv

Source-linked AI summary

Advanced modelling and data analytics in aviation

Aziida Nanyonga

arXiv:2608.14746v1cs.AI

TL;DR

Aviation safety research seeks better use of accumulated safety data to improve incident analysis and prevention. This study applies machine learning and natural language processing to aviation safety reports, finding BLSTM most effective for the reported classification task at 72% accuracy.

  • Problem

    Aviation safety is critical for the global aviation industry, motivating efforts to minimize safety risks.

  • Method

    The study applies NLP and deep learning models to aviation safety report narratives and evaluates their classification performance.

  • Results

    BLSTM achieved the highest reported accuracy of 72% for the classification task.

  • Takeaways & Limitations

    BLSTM was the most effective model for this classification task, supporting its use for aviation safety report analysis.

  • Takeaways & Limitations

    The study identified challenges in accurately classifying the Private class, including zero precision, recall, and F1-score for that category.

Abstract

from arXiv · show

The aviation industry characterized by its stringent safety standards has seen a growing need for innovative approaches to enhance safety measures. Despite the vast accumulation of aviation safety data over time, its full potential in predicting and preventing incidents has not been fully realized. This research addresses this gap by applying machine learning (ML) and natural language processing (NLP) techniques to analyze aviation safety data from Socrata, the Australian Transport Safety Bureau (ATSB), the National Transportation Safety Board (NTSB), and the Aviation Safety Network (ASN). By leveraging existing ML models, including deep learning and transformer-based architectures alongside NLP methods for mining aviation incident narratives, this study uncovers patterns contributing to safety related incidents such as accidents and near-misses. Additionally, it employs various topic modelling techniques to extract meaningful themes from unstructured safety reports, enhancing the interpretability of incident analysis. Causal inference techniques and interpretable AI frameworks are further explored to improve model transparency and trustworthiness. A key contribution of this work is the deployment of advanced ML methodologies in a structured aviation safety context, assessing their effectiveness and providing insights into their practical implementation. The findings offer valuable insights for aviation stakeholders, including regulators, airlines, and policymakers, by providing data-driven solutions that enhance incident analysis and decision making. Ultimately, this research supports the industry s ongoing efforts to minimize risks, improve passenger and crew security, and integrate AI driven methodologies into aviation safety management.

Certificate of Originality

The author declares that the submission is original, has not been substantially submitted for another degree or diploma, and fully acknowledges external contributions.

  • The author declares that the submission is their own work, to the best of their knowledge and belief.
  • The submission contains no material previously published or written by another person, except where due acknowledgement is made.
  • The work has not been substantially accepted for another degree or diploma, and colleagues’ research contributions are fully acknowledged.

Chapter 2. A Systematic Review of Machine Learning Analytic Methods for Aviation Accident Research

This systematic review synthesizes 87 studies published over 25 years on machine-learning applications in aviation safety and post-accident analysis. It identifies effective applications, dominant data sources and methods, credibility requirements, and seven priorities for future research.

  • Findings: ML techniques supported post-accident analysis, predictive modelling, and real-time incident detection across diverse aviation scenarios.The review covered supervised, unsupervised, and semisupervised methods, including neural networks, decision trees, support vector machines, and deep learning models.
  • Findings: The Aviation Safety Reporting System and NTSB datasets were the most frequently used data sources, while transparency, fairness, and bias mitigation shaped ML credibility and acceptance.These contextual factors were identified as critical to trustworthy aviation safety analysis.
  • Future research: The review recommends seven research directions: interpretable AI, real-time prediction, hybrid models, imbalanced-data handling, privacy and security, human-machine interfaces, and regulatory implications.Together, these directions provide a blueprint for further ML-based aviation safety research and emphasize appropriate use of ML in data-driven safety practice.
  • Methods: The review identified 87 relevant papers after database, backward-reference, and quality-assessment screening, providing a foundation for understanding ML in aviation post-accident analysis.Backward searching added 26 articles, while quality assessment excluded 7 papers before the final selection.
  • Findings: Studies concentrated on ALL aviation applications (44%), Regular Passenger Transport (31%), and general aviation (7%), with smaller shares for military (3%) and cargo (1%).Figure 5 presents this distribution as a Pareto plot, showing broader coverage of ALL applications than other categories.

2.5.9.1 Machine Learning Approaches

The review finds that machine learning is widely applied in aviation safety, especially through supervised learning and classification, while deep learning is the most common algorithmic approach. Across related NLP applications, these methods support objective analysis of unstructured reports, risk detection, and proactive safety management.

  • Supervised learning accounted for 65% of studies, while classification was the most common ML task at 59%.
  • Deep learning was the most widespread algorithm, appearing in 21% of studies, although no single ML algorithm dominated aviation accident analysis.
  • ML applications span classification, anomaly detection, predictive modelling, and real-time incident identification across seven aviation safety areas.
  • NLP methods process large volumes of unstructured aviation safety reports to reveal latent risks, causal factors, hidden patterns, and emerging trends.
  • The reviewed applications improve safety-analysis efficiency and accuracy, support proactive risk mitigation, and inform evidence-based decisions, while requiring transparency, fairness, and bias mitigation.

Chapter 9. Phase of Flight Classification in Aviation Safety using LSTM, GRU,

This study uses NLP and recurrent deep-learning models to infer aircraft flight phases from unstructured ASN safety-occurrence narratives. Hybrid models improved classification, with LSTM+BiLSTM achieving the strongest joint-model performance, while limited dataset size constrained further gains.

  • Results: BiLSTM outperformed the single RNN models with 64% accuracy, 63% precision, and 64% recall, compared with LSTM’s 63% accuracy.GRU achieved 60% accuracy and recall with 63% precision.
  • Results: 67% accuracy was achieved by the LSTM-BiLSTM joint model, the strongest reported joint-model result for classifying flight phases from ASN narratives.The study reports that LSTM+BiLSTM exhibited the highest accuracy and precision among joint models.
  • Objective and method: The study evaluates NLP and deep-learning methods for categorizing flight phases from unstructured safety-occurrence reports in the ASN database.Its objective is to assess whether raw accident narratives contain sufficient information to infer flight-phase labels for broader safety-occurrence analysis.
  • Implications: The results show that NLP and deep-learning models can extract flight-phase information from complex pre- and postaccident narratives, supporting proactive aviation-safety decision-making.The models are presented as useful for classifying phases from raw text and extracting patterns from unstructured reports.
  • Limitations: The seven-class results surpassed the 14% random-guess threshold, but limited dataset size indicates that larger and combined datasets are needed for further improvement.Model selection should also reflect application-specific trade-offs between precision and recall.

Chapter 10. Aviation Safety Enhancement via NLP & Deep Learning: Classifying Flight Phases

NLP and deep learning models accurately classify flight phases from unstructured aviation safety narratives, with LSTM and BLSTM showing the strongest accuracy and precision. The approach supports contextual, scalable safety occurrence analysis and more targeted safety measures.

  • 10.5 Results and Discussion: 87%, 88%, 87%, and 88% were achieved for the reported performance metrics, with LSTM achieving the highest performance.The models demonstrated high accuracy, precision, recall, and F1 scores in automating safety occurrence analysis.
  • Implications: Associating incidents with flight phases supports context-aware analysis because take-off and landing incidents have different causation factors than cruising incidents.This contextual classification can inform targeted safety measures and proactive safety protocols.
  • Study Design: 50,778 records were assembled for flight-phase analysis, enabling classification of flight phases from unstructured safety occurrence narratives.The study evaluates deep learning architectures including LSTM, CNN, BLSTM, and sRNN.
  • 10.5 Results and Discussion: LSTM and BLSTM exhibited the highest accuracy and precision for classifying flight phases from unstructured aviation safety narratives.Table 24 reports the key model performance results, while Figures 60 and 61 compare validation accuracy and loss across training epochs.
  • 10.6 Conclusion: NLP and deep learning can streamline report handling by helping authorities process more reports efficiently without overlooking critical safety information.The study identifies future work in model refinement, multimodal data, real-time analysis, cross-dataset validation, interpretability, and human-in-the-loop systems.

Chapter 13. Classification of Operational Records in Aviation Using Deep

The study evaluates deep-learning models for multi-class aviation operational-record classification and compares LDA with NMF for extracting themes from aviation accident reports. BLSTM performs best overall, while LDA offers greater topic coherence and NMF produces more distinct, granular topics.

  • Research Contribution: The study addresses limited comparative research by evaluating CNN, RNN, LSTM, and BLSTM models for aviation operational-record classification, including the Socrata dataset.It examines model strengths and limitations to assess deep learning’s potential for aviation safety analysis, operational efficiency, and automated classification.
  • Classification Results: CNN and sRNN classified Commercial records relatively well but struggled with Military and Private records, with sRNN producing zero precision, recall, and F1-score for Private.All models showed reduced Private-class performance, likely reflecting class imbalance or insufficient distinguishing features.
  • Classification Results: BLSTM achieved the highest overall accuracy at 72%, followed by LSTM at 71%, sRNN at 69%, and CNN at 67%.BLSTM and LSTM also showed lower validation loss and greater stability, indicating stronger generalization.
  • Topic Modeling: LDA produced a C_v coherence score of 0.497, while LDA showed higher topic coherence and NMF generated more distinct, granular topics.The comparison targets latent-theme extraction and interpretable analysis of aviation accident reports using the NTSB dataset.

1. Aircraft Mechanical Failures … 9. Aircraft Control and Runway Operations

Topic modelling of aviation incident narratives identified recurring safety themes and showed that model choice affects coherence, interpretability, granularity, and predictive performance. Across comparisons, BERTopic and LDA achieved the strongest coherence in their respective evaluations, while alternative methods offered complementary detail or probabilistic structure.

  • 9. Aircraft Control and Runway Operations: LDA produced broader, semantically coherent topics, while NMF generated more specific and granular themes useful for targeted accident analysis.Extracted themes included operations, mechanical issues, fuel systems, engine performance, weather, flight instruction, mechanical failures, and weather conditions.
  • 9. Aircraft Control and Runway Operations: LDA achieved the highest coherence among four methods, with 0.597 versus pLSA 0.583, LSA 0.542, and NMF 0.437.The comparison covered LDA, pLSA, LSA, and NMF applied to aviation incident narratives.
  • 17.5.6 Comparison of Topics and Themes: BERTopic vs. PLSA: BERTopic organized aviation safety terms into broad, context-driven themes, whereas PLSA emphasized narrower technical and problem-oriented topics through statistical co-occurrence.BERTopic captured themes such as Flight Operations and Takeoff & Landing, while PLSA emphasized Engine Issues, Flight Procedures, fuel shortages, and reported incidents.
  • 18.4.3 Topic Modeling Procedure: The models revealed recurring aviation safety factors including mechanical failures, pilot actions, weather, engine emergencies, terrain, runway events, and training-related issues.These thematic structures supported categorization by operator type and enabled targeted safety interventions and resource allocation.
Loading 2608.14746v1…