PaperlayerView exact revision on arXiv

Source-linked AI summary

Advanced modelling and data analytics in aviation

Aziida Nanyonga

arXiv:2608.14746v1cs.AI

TL;DR

Existing ML approaches to aviation safety leave gaps in extracting insights from incident narratives. This thesis applies NLP, deep learning, and topic modelling to predict probable causes, classify aircraft damage, and extract safety themes. sRNN led damage-classification accuracy and precision, while BERTopic achieved higher topic coherence than PLSA.

  • Problem

    Aviation incident investigation is resource-intensive, while existing ML approaches predominantly use structured data and leave gaps in extracting insights from unstructured narratives.

    Source: Chapter 1. Introduction · p. 23

  • Method

    The thesis applies NLP and deep learning to incident narratives for probable-cause prediction and aircraft-damage classification, and compares topic-modelling methods for aviation safety reports.

    Source: §4.2 · p. 79, §7.2 · p. 122, §17.6 · pp. 235–236

  • Results

    Across damage classification and topic modelling, sRNN led classification accuracy and precision, while BERTopic surpassed PLSA in topic coherence (0.41 versus 0.37).

    Source: §11.5.1 · p. 166, §17.6 · p. 235

  • Takeaways & Limitations

    Topic modelling yields interpretable aviation-safety themes, including pilot error, mechanical failure, weather conditions, and training deficiencies.

    Source: §17.6 · p. 236

    Military and Private classes remained difficult to classify, likely because of class imbalance and limited distinguishing features.

    Source: §13.5.4 · pp. 187–188

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 the submission to be their own work, with no unacknowledged material written or published by others or substantially accepted toward another degree or diploma, to the best of their knowledge and belief.

Certificate of Originality

  • The author declares that the submission is their own work.Source: Certificate of Originality · p. 4
  • To the best of the author’s knowledge and belief, material previously published or written by another person is excluded unless duly acknowledged in the thesis.Source: Certificate of Originality · p. 4
  • To the best of the author’s knowledge and belief, material substantially accepted for another degree or diploma at UNSW or elsewhere is excluded unless duly acknowledged.Source: Certificate of Originality · p. 4

List of Publications

The publication list covers aviation safety reviews, incident prediction, classification, and topic modelling. It includes submitted manuscripts and conference papers dated 2023–2025.

List of Publications

  • Submitted manuscripts investigate incident-cause prediction using multi-head attention transformers and NLP with variational auto-encoders, alongside explainable supervised aviation prediction in Australia.The transformer, variational auto-encoder, and explainable-learning manuscripts were submitted to Modelling, IEEE Access, and Journal of Aerospace, respectively.Source: List of Publications · p. 19
  • Review papers cover machine-learning methods for aviation accident research and NLP applications in aviation safety.The systematic machine-learning review was submitted to TRIP; the NLP review and qualitative analysis appears in AIAA SCITECH 2025 Forum.Source: List of Publications · p. 19
  • Classification publications address flight phases, sequential aviation safety occurrences, and operational records using NLP and deep learning.Flight-phase studies include LSTM, GRU, and BiLSTM models using ASN data, and NLP with deep learning using ATSB reports; operational-record classification appears in ICPCT 2025.Source: List of Publications · pp. 19–20
  • A SPIN 2024 paper compares deep-learning architectures for textual damage-level classification.The publication list supplies the study topic but no comparative performance results.Source: List of Publications · p. 19
  • A GCITC 2023 paper examines the impact of dataset size and data source on NLP-based aviation safety incident prediction models.No effect sizes or preferred data sources are reported in the publication list.Source: List of Publications · p. 19
  • Topic-modelling publications explore incident narratives through clustering and comparisons involving LDA, NMF, and PLSA, including studies using ATSB and Socrata data.The listed studies include narrative exploration, comparative modelling of ATSB narratives, an LDA–NMF comparison of accident reports, and a Socrata case study.Source: List of Publications · p. 20

Chapter 1. Introduction

The thesis applies transformers, variational autoencoders, and explainable AI to aviation safety analysis, focusing on incident-cause prediction from textual narratives and model transparency.

1.1 Background

  • Resource-intensive investigations and structured-data-focused ML leave narratives underused, while interpretability, predictive accuracy, and data imbalance concerns hinder adoption.Traditional approaches also face human biases, delayed risk identification, and large-scale data-management challenges. The research proposes predictive modelling and transparency improvements validated against real-world safety data.Source: Chapter 1. Introduction · pp. 22–23

1.2 Significance

  • Integrating ML and NLP targets underused incident narratives and interpretable predictions that could support safety-management systems, proactive risk mitigation, and stakeholder decision-making.Transformer-based models, VAEs, and XAI are proposed to automate incident-cause identification and improve aviation risk assessment accuracy.Source: Chapter 1. Introduction · p. 23

1.3 Motivations

  • Growing traffic, difficult-to-process safety data, AI advances elsewhere, and distrust of black-box models motivate scalable, explainable aviation safety analysis.SHAP and other explainability methods address transparency, regulatory-compliance, and accountability concerns; the broader motivation is accident reduction and improved investigation and policy.Source: Chapter 1. Introduction · p. 24

1.4 Research Aim and Objectives

  • The thesis aims to apply transformers, VAEs, and SHAP-based explainability to incident-cause prediction, assessing effectiveness on real-world NTSB and ATSB datasets.Its objectives address predictive accuracy, model interpretability, and underexplored applications of ML in aviation.Source: Chapter 1. Introduction · pp. 24–25

1.5 Research Questions

  • The research asks how ML and NLP can predict incident causes and improve interpretability, examining research gaps, transformer accuracy, VAE inference, and explainability.Secondary questions investigate current trends, limitations, and future directions alongside the predictive and interpretability roles of these methods.Source: Chapter 1. Introduction · p. 25

1.6 Thesis Structure

  • Twelve conference papers and four journal articles underpin the thesis, whose NLP review analyzes 34 Scopus-indexed publications available as of May 2024.The review identifies annotated-data and interpretability challenges, proposing active learning for annotation and explainable AI for transparency.Source: Chapter 1. Introduction · pp. 25–26
  • Core contributions examine VAE causal-factor inference, XAI transparency, and transformer prediction, alongside deep-learning incident classification and pLSA/BERTopic extraction of recurring themes and risks.Chapters 4–6 contain the core studies, Chapters 7–13 cover deep learning, and Chapters 14–18 address topic modelling.Source: Chapter 1. Introduction · p. 26
  • The concluding chapter synthesizes findings, implications, limitations, and future research, while intervening chapters retain individual-publication structures within the thesis’s shared aviation-safety focus.References are consolidated in a bibliography at the end of the thesis.Source: Chapter 1. Introduction · pp. 26–27

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

This systematic review synthesizes 87 studies of machine learning in aviation post-accident analysis, situating this work alongside prediction and real-time detection. It examines methods, outcomes, applications, and publication characteristics and outlines research priorities for transparent, trustworthy, and proactive safety analysis.

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

  • 87 studies selected from 3832 papers document machine learning applications in post-accident analysis, predictive modelling, and real-time incident detection.The review covers 25 years, using Scopus searches and backward reference searches through Google Scholar. It examines application contexts, learning types, tasks, algorithms, and data sources; ASRS and NTSB datasets were the most used.Source: Chapter 2. A Systematic Review of… · pp. 27, 29, 30
  • Transparency, fairness, and bias mitigation underpin credible ML-based aviation safety research, informing seven recommended directions for future work.The directions are interpretable AI, real-time prediction, hybrid models, handling unbalanced datasets, privacy and data security, human-machine interfaces for safety professionals, and regulatory implications.Source: Chapter 2. A Systematic Review of… · pp. 27–28

2.3.1 Artificial intelligence

  • AI studies intelligent agents that mimic human learning and problem-solving, enabling perception, natural language processing, and decision-making.Machine learning is presented as a prominent subset of AI with applications in aviation.Source: §2.3.1 · pp. 30–31

2.3.2 Machine Learning

  • ML, an AI subset, depends on high-quality, balanced data and problem-appropriate algorithms [37, 38].The review reports supervised, unsupervised, and semisupervised applications in aviation safety. Its background also describes reinforcement learning as interaction-based learning that maximizes cumulative rewards through trial and error.Source: Chapter 2. A Systematic Review of… · p. 27, §2.3.2 · pp. 31–32, §2.3.1 · p. 31

2.3.4 Machine Learning Tasks

  • ML tasks comprise classification, regression, clustering, anomaly detection, data reduction, and natural language processing.These assign categories, predict continuous values, group similar records, detect unusual cases, reduce complexity, and interpret or generate language, respectively. Data reduction limits computational overhead and overfitting, while NLP often enables other tasks as a sequential addition.Source: §2.3.4 · p. 32

2.3.5 Machine Learning Techniques in Aviation Safety

  • Neural-network decisions can be difficult to trace, despite applications to aviation risk classification and prediction of HFACS unsafe acts.Zhang & Mahadevan, combined random forests, deep neural networks, and SVM for NLP and classification of 11 years of ASRS reports. LSTM with variable selection, including LASSO, modelled temporal safety patterns across 10 years of ASRS data.Source: §2.3.5 · pp. 34–35
  • Naive Bayes was most effective against SVM and KNN for accident-cause classification in Koteeswaran et al.,.SVM classified ASRS accident data, while decision trees supported accident categorization and NLP analysis of 186,000 reports. KNN classified FAA accident data spanning 95 years in Koteeswaran et al.,.Source: §2.3.5 · pp. 33–34
  • LDA uncovers latent topics in accident narratives, while domain-specific NLP may require tailored dictionaries or lexicons.LDA represents documents as topic mixtures and topics as word distributions. One study combined LDA with other techniques to analyze 20,000 NTSB accidents spanning 59 years.Source: §2.3.4 · p. 33, §2.3.5 · p. 35
  • The methodology linked research questions to Scopus searching, study selection, and quality assessment.The preceding technique overview illustrates PCA-based dimensionality reduction with neural networks and decision trees for passenger-survival classification. Review analysis and reporting were manual, following Kitchenham et al.'s guidelines and findings in.Source: §2.3.5 · p. 36

2.4.1. Search Strategy / 2.4.2. Database Searches

  • 3832 initial articles became 68 after removing 3731 irrelevant records and 33 duplicates; supplementary citation searching added 26 articles.Scopus search terms referenced aviation-safety sources including NASA ASRS, ICAO, NTSB, and Australian transport safety. Reference searching supplemented the database search.Source: §2.4.1 · p. 36, §2.4.2 · pp. 37–38

2.4.4 Selection criteria

  • Selection excluded non-English studies, work outside aviation accident analysis, non-peer-reviewed sources, and articles with inaccessible full texts.Screening proceeded from titles and abstracts to full-text assessment, using inclusion and exclusion criteria.Source: §2.4.4 · pp. 38–39

2.4.5. Data Extraction, Coding, and Quality / 2.4.6. Data Synthesis

  • 87 papers remained after quality assessment excluded seven additional studies, forming the corpus for thematic synthesis of ML applications in aviation post-accident analysis.A 22-question Microsoft Forms tool captured study characteristics and quality assessments, with extracted data organized in Excel. Synthesis compared themes, methodologies, algorithms, and outcomes to identify trends, gaps, challenges, and future research directions.Source: §2.4.5 · pp. 39–40, §2.4.6 · pp. 40–41

2.5.1 Aviation Application

  • 44% of studies addressed all aviation applications, versus 31% regular passenger transport, 7% general aviation, 3% military aviation, and 1% cargo.The same passage later gives approximately 55% for the all-applications category, rather than 44%.Source: Figure 5 · p. 41, §2.5.1 · p. 41

2.5.2. Sources and Types of Publications

  • 52% of publications were journal articles, followed by conference papers at 44% and theses at 4%.Scopus supplied 61 qualifying articles, reported as 69%, while Google Scholar's “cited by” search supplied another 26. Conferences and theses were proportionally more common among Google Scholar results.Source: §2.5.2 · pp. 41–42, Figure 6 · p. 42

2.5.3 Duration of Studies

  • 24.7 years was the mean dataset duration across 67 studies reporting coverage; the median was 15 years.Examples ranged from 1 year to 101 years, with the distribution peaking at 10–19 years. Duration was calculated as final year minus initial year plus one, with stated annual uncertainty of ±0.5 years.Source: §2.5.3 · p. 42, Figure 7 · p. 43

2.5.4 Publication Trends

  • 12.5 papers per year characterized the third research generation, compared with 3.4 in the second and 0.4 in the first.Methods progressed from decision trees in 1998 and logistic regression in 2004 to support vector machines in 2010 and deep learning in 2017; publications peaked in 2020. The authors interpret growth as increasing ML adoption and suggest NLP's ability to infer aviation-safety ontology may contribute.Source: §2.5.4 · pp. 43–44

2.5.5 Software Packages Utilized for Analysis

This section examines software tools, data sources, and ML and NLP methods used in aviation safety analysis, including predictive applications and the identification and generation of probable incident causes from narratives. It also considers research gaps and deployment priorities involving interpretability, data availability, privacy, human factors, and regulatory frameworks.

2.5.5 Software Packages Utilized for Analysis

  • 40% of studies used Python, followed by WEKA at 20%, MATLAB at 16%, and R at 11%.Python’s appeal includes extensive ML libraries, versatility, and batch scripting; WEKA offers an accessible interface, MATLAB numerical computing, and R statistical analysis and visualization. Despite Python’s emerging dominance, the review finds no consensus on the best software.Source: §2.5.5 · p. 45, Figure 9 · p. 46

2.5.6 Data Source Utilized in the Studies; 2.6.2. Data Sources and Quality Assessment

  • 41% of studies used ASRS data and 25% used NTSB data, making them the leading sources for aviation safety analysis.Other sources accounted for 18%, FAA for 8%, ASN for 4%, surveys for 3%, and ATSB for 1%. The review emphasizes relevant, high-quality datasets and rigorous quality assessment as foundations for reliable findings and actionable decisions.Source: §2.5.6 · p. 46, §2.6.2 · pp. 53–54

2.5.7. Country of Publication Analysis; 2.5.8. Authors

  • 45 publications placed the United States ahead of India with 12 and China with 11; Khan, L and Mavris, DN led authorship with five papers each.Canada, Türkiye, Italy, and Saudi Arabia contributed three to four publications each, while Australia, the UK, and Germany contributed one each. Within the 1998–2023 review, Christopher, AA and Puranik, TG each authored four papers, and Alias Balamurugan, SA, Mahadevan, S, Rao, AH, Robinson, SD, and Zhang, X each authored three.Source: §2.5.7 · p. 46, §2.5.8 · p. 47, §2.5.10 · p. 53

2.5.9. Machine Learning Dimensions; 2.5.9.1 Machine Learning Approaches; 2.5.9.2 Machine Learning Types

  • 65% of studies used supervised learning, while classification was the leading task, reported inconsistently as 59% and 60%.Research approaches were quantitative in 62% of studies, mixed-methods in 30%, and qualitative-only in 8%; unsupervised learning accounted for 30% and semi-supervised learning for 5%. Data reduction represented 5% of tasks and regression 1%.Source: §2.5.9.2 · pp. 48–50, §2.5.9.1 · p. 48

2.5.9.2 Machine Learning Types; 2.5.10 Enhancements to Safety

  • 18% of studies used NLP, 9% clustering, and 8% anomaly detection, extracting textual insights, grouping similar records, and flagging potential hazards.Safety applications include forecasting risks with SVM, random forests, and Naive Bayes, detecting anomalies with decision trees and deep learning, and mining narratives with Latent Dirichlet Allocation and recurrent neural networks. K-Means, deep belief networks, and self-organizing maps identify latent patterns, while PCA and non-negative matrix factorization simplify complex data and expose important features.Source: §2.5.9.2 · p. 49, §2.5.10 · p. 52

2.5.9.2 Machine Learning Types; 2.6.4. Variability in Study Types and Purposes

  • 21% of studies used deep learning, the leading algorithm, but the review identifies no primary algorithmic focus.Reported usage was decision trees 16%, SVM 14%, Naive Bayes 12%, KNN 10%, random forests 9%, linear discriminant analysis 6%, logistic regression 6%, PCA 4%, and K-Means 2%. Studies often tested multiple approaches; the review emphasizes considering their varied types and objectives when interpreting this methodological diversity.Source: §2.5.9.2 · pp. 50–51, Figure 14 · p. 51, §2.6.4 · p. 54

2.6.5. Future Research Directions; 2.6.5.1. Interpretable and Explainable AI:; 2.6.5.2. Real-time Accident Prediction:; 2.6.5.3. Hybrid Models:

  • Future priorities include transparent, fair models, real-time accident prediction, and hybrid approaches combining ML with expert systems or physics-based models.Explainable decisions are important for stakeholder trust in safety-critical applications, while predictive analytics and real-time processing could provide early warnings and enable swift interventions. Hybrid approaches are proposed to improve accuracy and robustness, alongside broader attention to ethics and bias mitigation.Source: §2.6.5 · p. 54, §2.6.5.1 · p. 54, §2.6.5.2 · p. 54

2.6.5.4. Handling Unbalanced Data Sets:; 2.5.10 Enhancements to Safety; 2.6.1. Diverse Machine Learning Techniques and Applications; 2.6.3. Safety Enhancement and Insights Generation

  • Cost-sensitive learning addresses rare accident cases by prioritizing minority-class classification, potentially accepting more errors on non-accident instances.Human-factor analysis can guide targeted interventions, and evaluation measures should match aviation-specific contexts, including applications beyond accident investigation. ML’s pattern-discovery and predictive capabilities support preventive strategies and informed safety decisions.Source: §2.5.10 · pp. 52–53, §2.6.1 · p. 53, §2.6.3 · p. 54

2.6.5.5. Privacy and Data Security:

  • Future aviation accident research should explore differential privacy and secure multi-party computation to support analysis and collaboration while protecting individual privacy.Differential privacy adds carefully controlled noise to data or query responses; secure multi-party computation enables joint computation without revealing participants’ inputs.Source: §2.6.5.5 · p. 55

2.6.5.6. Human-Machine Interfaces for Safety Professionals:

  • Future ML-driven accident analysis should integrate cognitive psychology and human factors expertise to account for human errors and produce more actionable insights.The passage identifies human errors as a significant contributor to accidents.Source: §2.6.5.6 · p. 55
  • Future research should develop regulatory frameworks and adoption strategies to guide ethical, effective use of ML-driven accident analysis across aviation.The passage links the need for these frameworks to increasing use of ML techniques in accident analysis.Source: §2.6.5.6 · p. 55

2.7 Conclusion

  • The review highlights classification, anomaly detection, predictive modelling, and real-time incident identification as aviation safety applications, emphasizing transparency, fairness, and bias mitigation.It synthesizes a quarter-century of studies across seven safety areas and identifies seven future research directions. It also encourages safety-focused AI/ML adoption in healthcare, autonomous vehicles, and industrial manufacturing.Source: §2.6.5.6 · p. 56

2.6.5.6. Human-Machine Interfaces for Safety Professionals:

  • Surveyed studies examine temporal topic modelling with expert review, structural topic modelling, and NLP-based clustering and analysis of aviation safety narratives.Source: §2.6.5.6 · p. 58
  • Surveyed studies also address incident-accident classification in aviation safety management and prediction of injuries and fatalities in aviation accidents.Source: §2.6.5.6 · p. 58
  • The review identifies large annotated-dataset requirements and complex-model interpretability as implementation challenges, proposing active learning for annotation and explainable AI for interpretation.Qualitative and quantitative analyses examine research methodologies, motivations, findings, and implications, while case studies illustrate successful NLP applications in aviation safety.Source: §2.6.5.6 · p. 60

3.2. Introduction

  • Manual incident-report analysis is time-consuming and error-prone; NLP and ML offer more efficient, accurate analysis of large text-report collections.Manual review is also constrained by human capacity, subjectivity, and oversight. Anticipated conventional and urban traffic growth provides context for using AI/ML to exploit automation and available safety data.Source: §3.2 · pp. 60–61

3.2.1 Natural Language Processing

  • 70% accuracy was reported for BERT-based question answering that extracted relevant information from free-text ASRS narratives.NLP processes unstructured narratives for pattern identification, classification, and prediction; examples include SVD for dangerous-goods topics, LSA for causal-factor mapping in Rankin et al. (2015), and STM for technical-report themes. Dong et al. used LSTM networks to identify incident causal factors, reporting significant improvements over traditional methods.Source: §3.2.1 · p. 61

3.2.2 Air Transport System Safety

  • Predictive safety management uses forecasting and causal modelling to identify future hazards and guide mitigation, supporting planning across aviation sectors.NASA system safety seeks acceptable mishap risk throughout the system lifecycle within operational, time, and cost constraints. The System-Wide Safety project supports safe aviation transformation as demand increases.Source: §3.2.2 · pp. 61–62

3.2.3. Motivation and Significance

  • NLP supports qualitative predictive safety by tracking operational language for precursors and learning from near-miss reports to inform preventive measures.Potential sources include voice communications, ACARS messages, and check-and-training reports. Further motivations include scalable processing, reduced manual-categorization bias, and earlier detection of latent safety issues.Source: §3.2.3 · p. 62

3.2.4. Aim and Research Objectives

  • The study systematically reviews Scopus-indexed literature and uses qualitative analysis to evaluate NLP applications in aviation safety, identify gaps, and propose research directions.Research questions address NLP algorithms, critical safety issues, implementation challenges, predictive safety management, and implications for industry stakeholders.Source: §3.2.4 · pp. 62–63

3.3 Methodology; 3.3.1 Data Collection; 3.3.2 Statistical Analysis; 3.3.3 Study Design Considerations

  • 34 papers were reviewed, combining 17 relevant studies published through 2022 with additional papers from 2023 and 2024.The guiding systematic review identified 87 papers through Scopus and Google Scholar backward-reference searches, with inclusion emphasizing authoritative aviation safety data sources. Extraction covered bibliographic information, datasets, algorithms, objectives, designs, and outcomes; analysis examined study motivations and methods, counted keywords, generated WordItOut/Textalyser word clouds, and mapped objectives and abstracts using VOSviewer.Source: §3.3 · p. 63, §3.3.1 · p. 63, §3.3.2 · pp. 63–64

3.4.2 Accuracy and Objectivity; 3.7.1.3 Semantic Analysis and Word Embeddings

  • 44% unsupervised categorization accuracy was reported for Robinson’s LSA, while Kierszbaum & Lapasset et al.’s BERT achieved roughly 70% correct answers.These results concern causal-factor categorization within an existing taxonomy and question answering on ASRS reports, respectively, rather than a shared benchmark. LSA and word embeddings such as word2vec also captured semantic relationships and supported more interpretable narrative analysis.Source: §3.4.2 · p. 64, §3.5.1 · p. 66, §3.5.2 · p. 66

3.4.4 Enhancing Human Factors Analysis; 3.5.1 Methodologies and Machine Learning Algorithms; 3.7.1.2 Deep Learning Models

  • Dong et al.’s LSTM approach improved the accuracy and efficiency of identifying human-error and other causal factors compared with traditional analysis.The approach identified primary and contributing factors in incident narratives, while the broader review also reported promising LSTM and BLSTM results for incident classification and causal-factor identification.Source: §3.4.4 · p. 65, §3.5.1 · p. 66, §3.7.1.2 · p. 71

3.4.1 Efficiency and Scalability; 3.5.1 Methodologies and Machine Learning Algorithms; 3.5.2 Research Objectives and Study Designs

  • Walton and Marion used SVD to extract topics and trends from ASRS dangerous-goods narratives, supporting efficient analysis without extensive manual intervention.Their analysis covered ten years of incidents. More generally, NLP automated information extraction from unstructured safety narratives, addressing the limitations of structured-data methods and labour-intensive manual review.Source: §3.4.1 · p. 64, §3.3.3 · p. 64, §3.5.1 · p. 65

3.4.3 Early Detection and Proactive Safety Management; 3.5.1 Methodologies and Machine Learning Algorithms; 3.5.2 Research Objectives and Study Designs; 3.7.1.1 Topic Modeling Techniques

  • Topic modelling exposed latent safety themes, including Rose et al.’s metadata-capable STM analysis of ASRS and NTSB reports.Kuhn’s STM study revealed previously unreported connections and themes relevant to proactive safety management. Across the review, LDA and STM uncovered underlying patterns, while broader narrative analysis supported emerging-risk detection and preventive measures.Source: §3.4.3 · p. 65, §3.5.1 · p. 66, §3.5.2 · pp. 66–67

3.4.5 Contribution to Knowledge and Best Practices; 3.5.3 Key Outcomes and Implications; 3.7 Discussion; 3.7.1.4 Implications for the Industry

  • The review links NLP’s efficiency, analytical accuracy, and latent-issue detection to improved safety management, with studies informing best practices and policymaking.Method effectiveness varied with study objectives and aviation applications. The synthesis identifies implications for safety-management practices and accident-prevention strategies.Source: §3.4.5 · p. 65, §3.5.3 · p. 67, §3.7 · p. 71

3.5.3 Key Outcomes and Implications

  • 0.879 AUC was reported for discriminating relevant from irrelevant words across the 34 reviewed NLP papers.ASRS accounted for 57% of studies; application coverage comprised 27 all-aviation, six airline-specific, and one cargo-related study. DL and LDA each appeared in 23% of studies, versus 20% for BERT; statistical analysis used the Satisfy app.Source: Figure 15 · p. 67, §3.5.3 · p. 67, Figure 16 · p. 67
  • Over 17 mentions made “NLP” the most frequent term, followed by “ASRS” (12 mentions), “Text mining” (9 mentions), and “Aviation safety” (8 mentions).VOSviewer mapped keyword frequency and co-occurrence, with larger nodes indicating greater frequency and shorter distances indicating stronger co-occurrence. Natural language processing, aviation safety, and text mining also dominated the bibliometric analysis and word cloud.Source: §3.5.3 · p. 68, Figure 21 · p. 69

3.7.2.1 Research Gaps

  • Research gaps include limited multimodal integration, underused real-time data, insufficient human-factor analysis, geographic concentration, and inconsistent NLP evaluation metrics.Priorities include combining text, audio, video, and sensor data and developing streaming systems for immediate safety interventions. Deeper analysis of fatigue, decision-making, and communication, broader geographic and cultural coverage, and standardized evaluation would address further limitations.Source: §3.7.2.1 · pp. 72–73

3.7.3 Practical Recommendations

  • Integrating NLP into safety management systems can reveal recurring incident patterns and prioritize corrective actions through risk assessment.Standardized annotation guidelines, automated annotation tools, and expert review and feedback are recommended to improve training-data accuracy and relevance.Source: §3.7.2.1 · p. 74

3.7.3.3 Guidance on Model Integration, Deployment, and Capacity Building

  • NLP deployment requires integration frameworks, performance metrics, update and maintenance protocols, regulatory compliance, research partnerships, and internal capacity building.Organizations should account for operational requirements and constraints, collaborate on aviation-specific research, and train personnel in NLP, data analytics, and machine learning.Source: §3.7.3.3 · p. 74

3.8.2 Predictive Maintenance Using NLP

  • NLP case studies cover incident-risk discovery, predictive maintenance, and safety-culture assessment.ASRS narratives revealed latent risks and contributing factors, while maintenance records and aircraft-performance data supported early issue detection, optimized schedules, cost savings, and efficiency improvements. Employee reports and feedback revealed safety-related sentiments, attitudes, and perceptions to guide interventions and training.Source: §3.8.2 · p. 75, §3.7.3.3 · p. 75

3.9 Challenges and Solutions

  • Implementation challenges span variable language, privacy requirements, legacy-system scalability, specialist expertise, and opaque models, motivating technical and organizational remedies.Suggested remedies include deep learning and semantic analysis [157], federated learning and differential privacy, and modular interoperable platforms complemented by cloud services. Training and academic collaboration, alongside interpretable methods, visualizations, SHAP, and LIME, address expertise and transparency.Source: §3.8.2 · p. 76

3.7.1.4 Implications for the Industry

  • Seventeen Scopus-indexed studies indicate NLP’s potential for identifying safety issues, interpreting complex data, and supporting predictive safety management.Methods from SVD to BERT and LSTM are described as more objective and accurate than manual analysis; active learning and explainable AI are proposed for annotation and interpretability challenges. Industry implications include deeper causal insights, prioritized interventions, more efficient reporting, improved risk prediction, and evidence-based decisions.Source: §3.7.1.4 · p. 71, §3.8.2 · p. 77

Section II. Fundamentals of VAE, Explainable AI and Transformer.

  • 0.786 was the average similarity score between VAE-generated probable causes and actual probable causes in the NTSB test dataset, evaluated using cosine similarity.NLP encodes incident narratives numerically, and the generative model uses these representations to generate probable causes.Source: Section II. Fundamentals of VAE,… · p. 78

4.2 Introduction

  • The proposed NLP–VAE approach predicts probable causes from initial incident narratives, addressing the time-consuming process of establishing incident causes.Its reconstruction objective maximizes cosine similarity between predicted and actual probable-cause vectors instead of minimizing conventional mean square error. Earlier topic-modelling approaches grouped related aviation terms to assist investigations.Source: §4.2 · pp. 78–79
  • Prior aviation applications used iVAE for uncertain engine-prognosis data and CCVAE for learning unseen target-domain classes with limited annotations.iVAE incorporated a family of priors and specialized neurons, outperforming ICA in engine-deterioration and remaining-useful-life case studies involving moderate-sized, uncertain datasets. CCVAE generated synthetic unseen-class features from source-domain data and showed strong performance on datasets including Baggage Xray.Source: §4.2 · pp. 80–81
  • Prior studies improved icing-flow reconstruction with VAE, detected helicopter climb anomalies with convolutional VAE, and favored LSTM for flight-phase classification.Latent-space temporal discontinuity hindered SINDy equation identification despite improved reconstruction; helicopter detection combined trajectory-pattern mining and shape analysis. For ATSB reports, LSTM outperformed CNN, BLSTM, and sRNN in accuracy, precision, recall, and F1.Source: §4.2 · p. 81
  • The proposed NLP–VAE approach encodes incident narratives into a learned latent distribution and decodes vectors representing likely probable causes.The redefined VAE loss optimizes cosine similarity between generated and actual probable causes. The authors describe this as a first aviation-safety application and suggest it could shorten investigations and improve safety.Source: §4.2 · pp. 81–82, §4.4.7 · pp. 89–90

4.4.1 Dataset

  • 20919 completed NTSB investigations from 2005 to 2020 supplied the analysis-narrative and probable-cause fields used for model training and validation.The reports were downloaded as JSON. Vectors were randomly split into 85% training and 15% testing, with 10% of the training set reserved for validation during each epoch.Source: §4.4.1 · p. 82, §4.4.7 · p. 87

4.4.2 DATA PRE-PROCESSING

  • 1200 and 200 were the maximum sequence lengths for analysis narratives and probable causes after spacy/nltk preprocessing and GloVe vectorization.Cleaning included tokenization, stopword and punctuation removal, URL/HTML removal, stemming, and lemmatization. Analysis narratives averaged 1,116 words with standard deviation 858.36; probable causes averaged 165 words with standard deviation 93.12.Source: §4.4.2 · p. 83

4.4.3 DATA STANDARDIZATION

  • 30-dimensional latent representations linked narrative encoding to probable-cause decoding, with min-max scaling normalizing input features between 0 and 1.scikit-learn MinMaxScaler prepared inputs reshaped as (1200,1) and (200,1); the encoder used 11 Conv2D hidden layers and the decoder 10 Conv2DTranspose hidden layers, with LeakyReLU and batch normalization. The VAE minimized reconstruction error and KL divergence while constraining its latent distribution toward normality.Source: §4.4.3 · pp. 83, 84, 86

4.4.6. Experimental environment

  • 256 CPU cores, four NVIDIA A100-SXM4-80GB GPUs, and 512GB RAM supported experiments on an Ubuntu server.Experiments were implemented in Python using TensorFlow, nltk, sklearn, NumPy, and pandas.Source: §4.4.6 · p. 87

4.4.7. Model Training and Evaluation

  • 0.786 mean cosine similarity, with standard deviation ±0.120, measured agreement between predicted and recorded probable causes on unseen test narratives.Scores concentrated toward higher similarity, with very few below 0.5. Generated vectors were converted back into words and displayed as word clouds for human experts.Source: §4.4.7 · p. 87, Figure 24 · p. 88, Figure 26 · p. 89

Chapter 5. Explainable Supervised Learning Models for Aviation Predictions in Australia.

This section examines supervised learning for three-class classification of Australian aviation occurrences and NLP-based prediction of probable incident causes, aircraft damage, and flight phase from NTSB narratives. It covers VAE-based class balancing, SHAP-based model explanations, and the evaluation of transformer and recurrent models for narrative-based prediction.

Chapter 5. Explainable Supervised Learning Models for Aviation Predictions in Australia.

  • Prior aviation work offered an XAI taxonomy without quantitative validation and demonstrated LIME-based explanations for aircraft trajectory anomalies.The anomaly-detection prototype provided human-understandable explanations intended to support subject-matter experts and acceptance of AI decision-support tools.Source: Chapter 5. Explainable Supervised… · pp. 92, 94
  • The study addresses a gap in explainable ATSB incident classification by examining balanced and imbalanced learning and using SHAP and LIME to explain feature influences.Feature-impact visualizations aim to make classifications more transparent and trustworthy for aviation stakeholders.Source: Chapter 5. Explainable Supervised… · pp. 92, 95

5.4.1 Dataset

  • 26,262 ATSB records covered Accidents, Incidents, and Serious Incidents recorded in Australia over 20 years.The data came directly from ATSB investigation authorities.Source: §5.4.1 · p. 95

5.4.2 Data preprocessing

  • Preprocessing removed outliers, missing values, and irrelevant features, one-hot encoded categorical variables, and augmented minority classes using a VAE.Encoded variables included State, FuelType, RegistrationType, ActivityType, and ActivitySubType.Source: §5.4.2 · p. 95

5.4.3 Variational Autoencoder for class imbalance

  • 16000 Accident and 17000 Serious Incident instances were generated using separate class-specific VAEs to balance the occurrence classes.Incidents constituted almost 80% of the original dataset, making accuracy potentially misleading for minority-class recognition. Each VAE learned from its respective minority class before its decoder generated additional samples.Source: §5.4.3 · p. 96

5.4.4 Model performance Evaluation.

  • Multiclass performance was assessed using precision, recall, F1-score, and accuracy to evaluate how predictions were distributed across classes.F1-score combines precision and recall and ranges from 0 to 1, with higher values preferred.Source: §5.4.4 · p. 97

i) Confusion matrix

  • Confusion matrices visualize test-set performance by distinguishing correct predictions from false positives and false negatives.The illustrative binary matrix identifies true positives and true negatives as correct predictions.Source: i) Confusion matrix · p. 98

5.4.5 Model Training and validation

  • 4:1 random train–test splits supported evaluation of LR, RF, SVM, and a five-hidden-layer DNN on both balanced and imbalanced datasets.RF used maximum depth 30 and SVM a linear kernel; other classical-model hyperparameters used defaults. The DNN used ReLU hidden activations, softmax outputs, and 20% dropout after its first three hidden layers.Source: §5.4.5 · p. 98
  • SHAP provided global explanations for LR, RF, SVM, and DNN, enabling visual comparison beyond local explanations such as LIME.Explanations were generated, visualized, and discussed for every model to support transparency and stakeholder trust.Source: §5.4.5 · p. 99, PRECISION RECALL F1-SCORE · p. 103

5.5.1 SHapley Additive exPlanations (SHAP)

  • SHAP assigns feature-level shap_value scores using additive importance to generate and visualize global explanations for complex models.The method supports explanations for ensembles and deep neural networks whose prediction processes are difficult to interpret.Source: §5.5.1 · p. 99

5.6.1 Machine learning model performance

  • VAE balancing improved classification accuracy across all four models and increased accident and serious-incident precision, recall, and F1-score.Without balancing, every model recorded serious-incident precision, recall, and F1-scores of 0.0 despite accuracy exceeding the 33.33% random-guess baseline.Source: §5.6.1 · pp. 99–100, Table 10 · p. 100, Table 12 · p. 100

PRECISION RECALL F1-SCORE

6.1 Abstract

  • 0.727 mean BLEU and 0.696 mean LSA similarity were reported for transformer-generated probable causes from NTSB investigation narratives.BLEU SD was 0.33 and LSA similarity SD was 0.152. ROUGE F-measure means were 0.618 for rouge-1 (SD=0.192), 0.452 for rouge-2 (SD=0.248), and 0.5560 for rouge-L (SD=0.220).Source: §6.1 · p. 104

6.2. Introduction

  • A multi-head attention transformer addresses a cause-generation gap by accepting short or long narratives and potentially retaining reports without probable-cause labels.The source dataset contains 29,676 concluded NTSB cases from 2001–2020, and decoder training uses target probable causes instead of a target language. Preprocessing removes HTML and URLs, decodes characters, and discards narratives exceeding 10,000 words or causes exceeding 1,000 words.Source: §6.2 · pp. 106, 108, 109
  • Over 87.9% accuracy was reported for aircraft-damage classification, while occurrence classification achieved 80-93% accuracy [124].Flight-phase classifiers exceeded 68% accuracy on a seven-class problem. Topic-model comparisons found greater coherence with LDA but more unique, detailed topics with NMF.Source: §6.2 · pp. 107–108

6.4.3.The Transformer

  • The Transformer uses multi-head self-attention for parallel processing and dependency capture, weighting values through softmax applied to scaled query–key dot products.This architecture avoids the sequential training constraint of conventional recurrent neural networks.Source: 6.4.3.The Transformer · p. 110

6.4.4.Experimental setup

  • 8 encoder and decoder layers, 8 attention heads, and a 1024-token input length configured the transformer for probable-cause generation.The model used 1024-dimensional embeddings and dropout of 0.1; training used a 90%/10% training/test split, 50 epochs, and Adam with learning rate 0.001. Trained on NTSB reports, it accepts short and long narratives and is intended to expedite investigations.Source: 6.4.4.Experimental setup · p. 111, §6.5.4 · pp. 119–120

6.4.5.Performance metrics

  • Evaluation combines precision-oriented BLEU, recall-oriented ROUGE, and semantic LSA to assess generated probable causes.ROUGE evaluates unigram, bigram, and longest-common-subsequence overlap through precision, recall, and F-measure. LSA uses vector similarity and tolerates word-choice differences that lexical overlap metrics may penalize.Source: 6.4.5.Performance metrics · p. 112, §6.4.5.2 · pp. 112–113

6.5.1. Model performance based on the BLEU score

  • 0.727 mean BLEU score, with standard deviation ±0.330, measured agreement between predicted and reference probable causes across the test set.The evaluation used an n-gram weight vector of (0.1,0.1,0,0), selected after evaluating random test samples of size 500.Source: §6.5.1 · pp. 114–115, Figure 37 · p. 115

6.5.2. Model performance based on the LSA Similarity score

  • 0.697 mean LSA similarity, with standard deviation ±0.153, measured semantic agreement between generated and reference probable causes.The evaluation converted both sentences into vectors using Google's pretrained UniversalSentence-Encoder Version 4 and compared their semantic similarity.Source: §6.5.2 · p. 115, Figure 38 · p. 116

6.5.4. Analysis narrative length Vs BLEU/LSA scores

  • Narrative length showed no direct BLEU correlation; LSA showed no short-input correlation but converged toward its mean for longer inputs.The authors interpret this convergence as more stable performance with longer inputs, while identifying training-data requirements as a major limitation. Planned work would combine NTSB narratives with reports from other investigation bureaus.Source: §6.5.4 · pp. 116, 119, 120, Figure 39 · p. 117, Figure 40 · p. 117
  • 0.757 LSA similarity contrasted with rouge-2 Fmeasure=0.164 for a semantically similar example, illustrating lexical metrics’ sensitivity to wording.The authors favor semantic evaluation because differently worded causes can convey similar meanings. BLEU additionally depends on manually selected weights, with scores degrading when trigram and quadgram weights were included.Source: §6.5.4 · pp. 118–119

Section III. Deep Learning

  • Over 87.9% accuracy was achieved by every evaluated recurrent model for four-class aircraft-damage prediction, exceeding the 25% random-guess baseline.Across 27,000 NTSB reports, precision, recall, and F1 exceeded 80%, 88%, and 85%, respectively. sRNN achieved 90% recall and 90% accuracy, LSTM achieved 87% precision, and joint GRU+LSTM and sRNN+BLSTM+GRU achieved 90% recall and 90% accuracy.Source: Section III. Deep Learning · p. 120

7.2 Introduction

  • NLP and recurrent networks target aircraft-damage classification from narratives, addressing slow manual review and gaps in structured-data and topic-modelling approaches.The target categories are destroyed, substantial, minor, and none. Earlier aviation topic-modelling studies used LDA, but overlapping or ambiguous topics can undermine classification reliability.Source: §7.2 · pp. 121–122

7.3 Related Work

  • Below 80% performance was reported for accident-narrative models, while Adam-trained sRNN and combined sRNN–GRU led comparisons in another study.BLSTM, CNN, and RNN improved accident-narrative classification in, but further models and parameter changes were recommended. The study investigated F-16 modelling and fault-sensor prediction.Source: §7.3 · pp. 123–124

data sets used in this study.

  • 16,919 completed NTSB investigations from 2005–2020 supplied analysis narratives and four aircraft-damage labels for model training and validation.The dataset contained 1,409 destroyed, 15,163 substantial, 195 minor, and 152 no-damage records.Source: data sets used in this study · p. 124

7.4.2 Text Processing

  • 2,000-element narrative sequences were generated using Keras tokenization and spaCy cleaning and lemmatization, with shorter sequences zero-padded and longer sequences truncated.The vocabulary was capped at 100,000, and damage labels were numerically or one-hot encoded.Source: §7.4.2 · p. 125

7.4.3 Text Classification

  • 80% of records were allocated to training and 20% to testing, with 10% of training data held out for validation each epoch.CNN, LSTM, BLSTM, GRU, SimpleRNN, and combined architectures were trained using Adam; optimizer selection was not the study’s focus.Source: §7.4.3 · pp. 125–126

7.4.4 Deep learning model Architecture

  • Models shared embedding, hidden, and output layers, using ReLU hidden activations, SoftMax output probabilities, and argmax class selection.Combined models introduced slight architectural variations while preserving the common design for comparison.Source: §7.4.4 · p. 126

7.5.1 Detection performance of single RNN-based deep learning models

  • Among single RNN models, sRNN slightly led recall and accuracy, while LSTM achieved slightly better precision.Appropriately processed pre-accident narratives supported aircraft-damage prediction with accuracy far above random guessing.Source: Table 15 · p. 127, §7.5.1 · p. 127

7.5.2 Detection performance of joint RNN-based deep learning models / 7.5.3 Single RNN-based versus Joint RNN-based deep learning models

  • Joint RNN models generally outperformed single models in aircraft-damage classification, with sRNNLSTM achieving the best overall performance across metrics.Six joint combinations were evaluated on the same dataset. Joint architectures passed one RNN layer’s output into a subsequent layer using a different RNN architecture.Source: §7.5.1 · p. 127, §7.5.2 · p. 128, Table 17 · p. 129

8.1 Abstract / 8.2 Introduction

  • Over 68% accuracy was achieved by both flight-phase classifiers, with sRNN greatly outperforming the simplified ResNet architecture.The subsequent study used an initial 27,000 NTSB reports for seven-class classification, whose random-guess rate was 14%, and reported high precision, recall, and F1 scores. Inferring flight phases from narratives could help identify component-failure stressors, reduce unplanned maintenance, and locate hazards during critical flight phases.Source: §7.5.3 · pp. 130–131

7.5.3 Single RNN-based versus Joint RNN-based deep learning models

  • Chanen used word2vec to clarify report semantics, while Mersha et al. found sRNN best for selected F-16 prediction parameters.Chanen examined only one NLP tool, leaving alternative approaches unexplored. Song et al. applied BLSTM, CNN, and RNN to railroad narratives, but classification performance remained below 80%.Source: §7.5.3 · pp. 132–133

8.4.1 Data Acquisition

  • 16,919 cleaned records covered completed investigations from 2005 to 2020, with analysis narratives and flight-phase fields extracted for model training and validation.The records came from NTSB aviation incident and accident investigation reports downloaded in JSON format.Source: §8.4.1 · pp. 134–135

8.4.2 Text Processing

  • 2000-element sequences represented narratives through zero-padding or truncation, using a corpus vocabulary size of 100,000.Tokenizer converted text into numerical sequences; spaCy removed unwanted characters, punctuation, and stop-words and performed lemmatization.Source: §8.4.2 · p. 135

8.4.3 Text Classification

  • 80% of the data was randomly allocated to training and 20% to testing, with 10% of training data reserved for validation each epoch.ResNet and sRNN were optimized using Adam; identifying the best optimizer was outside the study's scope.Source: §8.4.3 · p. 135

8.4.4 Deep learning model Architecture

  • Two residual dense blocks, each with three skip connections, formed the simplified ResNet, supporting gradient flow and preserving earlier learned information.The models shared embedding, hidden, and output layers, using ReLU in hidden layers, SoftMax at output, and argmax for class selection.Source: §8.4.4 · p. 136

8.4.4.2 Simple Recurrent Neural Networks (RNNs).

  • sRNN outperformed ResNet on unseen test samples across accuracy, precision, recall, and F1 measure.The models predicted flight phase from unstructured event narratives. The authors characterize the experiments as preliminary.Source: Table 18 · p. 138, §8.4.4.2 · pp. 138–139
  • Over 87.9% accuracy was reported for earlier aircraft-damage classification; the current work extends narrative-based analysis to flight-phase prediction.Damage-level and flight-phase classifications are linked to maintenance decisions about return to service, repairs, budgets, and resources.Source: §8.4.4.2 · p. 139
  • Deep learning's black-box nature motivates further interpretability and explainability research so stakeholders can understand classification decisions and develop trust.Strong predictive performance does not remove concerns about how these models reach their decisions.Source: §8.4.4.2 · p. 139

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

NLP and recurrent deep-learning models inferred flight phases from unstructured ASN safety narratives, with LSTM+BiLSTM achieving the strongest reported performance. Limited dataset size motivates further evaluation using larger, multisource datasets.

9.2.Introduction

  • The study evaluates single and combined LSTM, GRU, and BiLSTM models for inferring flight phases from unstructured ASN safety occurrence narratives.Its contribution extends these architectures to ASN datasets that had not been extensively studied for this task.Source: 9.2.Introduction · p. 141, 9.3.Related Work · p. 143

9.3.Related Work.

  • 68% accuracy was exceeded by both ResNet and sRNN in seven-class NTSB narrative classification, with sRNN outperforming simplified ResNet.Puranik et al used offline-online Random Forest regression to predict landing true airspeed and ground speed from approach data. Fala et al found dimensionality-reduction combinations had no significant impact on phase identification, while K-means and Gaussian mixture models were suitable.Source: 9.3.Related Work · pp. 142–143

9.4.1 Data Collection

  • 4,372 cleaned ASN reports supplied narratives and phase-of-flight labels for a pipeline comprising data acquisition, text processing, and classification.Reports covered completed investigations from 01/01/2000 to 12/31/2020. The concluding account instead reports 4,327 records, leaving a dataset-size discrepancy.Source: §9.4.1 · pp. 143–144, 9.4.Methodology · p. 143, §9.5.2 · p. 148

9.4.2 Text Pre-processing

  • 2,000-position narrative sequences were produced through Keras tokenization, zero-padding or truncation, with vocabulary limited to 100,000 terms.spaCy handled special characters, punctuation, stop-words, and lemmatization; phase-of-flight labels were one-hot encoded.Source: §9.4.2 · p. 144

9.4.3 Text Classification

  • 80% of records formed the training set and 20% the test set, with 10% of training records reserved for validation.Adam optimized single and combined recurrent architectures, including LSTM+BiLSTM+GRU.Source: §9.4.3 · p. 144

9.4.4 Deep Learning Model Architecture

  • Models shared an embedding–hidden–output architecture with ReLU hidden activations, SoftMax output, and argmax prediction, allowing minor variations for combined models.Multiclass performance was evaluated using accuracy, precision, recall, and F1-score.Source: §9.4.4 · pp. 144–145

9.5.1 Performance of Single RNN-Based Deep Learning Models

9.5.2 Performance of Joint RNN-Based Models

  • 67% accuracy placed LSTM+BiLSTM ahead of GRU-LSTM (62%) and GRU-BiLSTM (60%); it also achieved the highest precision among evaluated models.Model selection should consider precision–recall trade-offs. Seven-class results exceeded the 14% random-guess threshold, but limited data motivate larger datasets incorporating NTSB and ATSB reports.Source: Chapter 9. Phase of Flight… · p. 140, Table 21 · p. 146, §9.5.2 · pp. 147–148

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

This section examines NLP and deep learning for classifying flight phases and aircraft damage from ATSB, ASN, and NTSB aviation safety narratives. It compares neural architectures and assesses how dataset size, source, and NLP techniques influence predictive accuracy and robustness.

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

  • 53,275 ATSB records from 2013–2022 yielded 50,778 cleaned reports, with Summary narratives and flight-phase labels extracted for model training and validation.Prior work found machine learning more accurate and efficient than rule-based classification, while phase-specific causation differences motivated contextual analysis. NTSB-based ResNet and sRNN models exceeded 68% accuracy against a 14% seven-class random baseline, with sRNN outperforming simplified ResNet.Source: Chapter 10. Aviation Safety… · pp. 150–152

10.4.2 Text Processing

  • 2,000-position narrative sequences were padded or truncated using a 100,000-word vocabulary, with an 80% training and 20% testing split.Keras tokenized narratives and one-hot encoded flight-phase labels. spaCy supported special-character, punctuation, and stop-word handling alongside lemmatization.Source: §10.4.2 · pp. 152–153

10.4.3 Text Classification

  • 10% of the training set supported validation during each epoch while LSTM, BLSTM, CNN, and sRNN were trained using Adam.Validation was intended to monitor robustness and prevent overfitting; identifying the best optimizer was outside the study’s scope.Source: §10.4.3 · p. 153

10.4.4 Deep Learning Model Architecture

  • A shared embedding–hidden–output architecture supported model comparison, using ReLU hidden activations, SoftMax multiclass outputs, and argmax to select the predicted class.The common architecture received slight adjustments for each model.Source: §10.4.4 · pp. 153–154

10.4.5 Model Performance Evaluation

  • Multiclass performance was assessed using accuracy, precision, recall, and F1-score to characterize prediction quality across flight-phase classes.Table 22 identifies the evaluation metrics, but its supplied caption contains no metric definitions or numerical results.Source: §10.4.5 · p. 154, Table 22 · p. 154

10.4.6 Confusion matrix

  • Confusion matrices distinguished correct diagonal predictions from off-diagonal errors; LSTM diagnostics also showed classification metrics and test-instance support across flight phases.The evaluated flight-phase dataset contained 50,778 ATSB reports.Source: §10.4.6 · p. 155, §10.5.1 · p. 156

10.5.1 Model Performance

  • 87% accuracy, 88% precision, 87% recall, and 88% F1 were achieved by LSTM; LSTM and BLSTM ranked highest in accuracy and precision.Automated flight-phase classification could support targeted safety measures, contextual analysis, and higher report throughput. Proposed future work includes model refinement, multimodal data, real-time analysis, cross-dataset validation, interpretability, and human-in-the-loop systems.Source: Chapter 10. Aviation Safety… · p. 149, Table 24 · p. 155, §10.5.1 · pp. 155, 157

Chapter 11. Comparative Study of Deep Learning Architectures for Textual Damage Level Classification

  • 89% accuracy and recall made sRNN the leading aircraft-damage classifier; all evaluated models exceeded 88% accuracy against a 25% four-class random-guess threshold.This subsequent study compares LSTM, BLSTM, GRU, and sRNN on ATSB textual narratives to automate damage-level classification. Its intended applications include incident-risk assessment, maintenance improvements, and targeted safety measures.Source: §10.5.1 · pp. 158–159

10.5.1 Model Performance

  • The study addresses scarce ATSB-focused comparisons by evaluating LSTM, sRNN, BLSTM, and GRU for damage-level classification from incident narratives.Its aim is to identify the most effective architecture and assess these models’ applicability to aviation incident analysis.Source: §10.5.1 · p. 160

11.4.1 Data Collection

  • 50,778 cleaned reports remained from 53,275 ATSB records covering 2013–2022, with Summary text and damageLevel labels extracted for modelling.The records were sourced directly from Australian investigation authorities.Source: §11.4.1 · p. 161

11.4.2 Text Processing

  • 2000-position sequences and a 100,000-word vocabulary standardized narratives, with an 80%/20% training–testing split.Keras tokenization represented text, while one-hot encoding represented four damage labels: destroyed, Substantial, Minor, and None. spaCy handled special characters, punctuation, stop words, and lemmatization; sequences were zero-padded or truncated.Source: §11.4.2 · p. 161

11.4.3 Text Classification

  • 10% of training data supported validation each epoch while sRNN, LSTM, BLSTM, and GRU were trained with Adam.Validation was intended to prevent overfitting; identifying the best optimizer was outside the study’s scope.Source: §11.4.3 · p. 162

11.4.4 Deep Learning Model Architecture

  • sRNN struggles with long-range dependencies, GRU uses two gates, LSTM uses gated memory cells, and BLSTM processes sequences bidirectionally.All models shared embedding, hidden, and output layers, with ReLU hidden activations, SoftMax outputs, and argmax class selection.Source: §11.4.4 · pp. 162–165

11.4.5 Model Performance Evaluation

  • Models were evaluated using accuracy, precision, recall, and F1-score, with validation curves, classification reports, and confusion matrices providing complementary diagnostics.Validation curves tracked accuracy and loss across epochs; the sRNN classification report included test-instance support by damage category.Source: §11.4.5 · p. 165, §11.5.1 · p. 167

11.5.1 Model Performance

  • sRNN outperformed LSTM, BLSTM, and GRU in accuracy and recall for aircraft damage-level classification, with concluding precision and F1 scores described as comparable.The prose differs on precision: an earlier passage describes sRNN as having the highest precision, whereas the concluding comparison describes parity.Source: Table 25 · p. 165, Figure 66 · p. 166, Figure 67 · p. 166
  • Future work proposes advanced textual features, semantic analysis, and incident-report images and videos to improve contextual understanding and damage-severity analysis.These are proposed extensions rather than evaluated components of the reported models.Source: §11.5.1 · p. 168
  • The study examines dataset-size and source effects on LSTM, GRU, CNN, and RNN aviation-narrative predictions, without data augmentation or transfer learning.It extends earlier flight-phase and damage-level narrative classification by explicitly varying dataset size, addressing limited comparisons between ASN and NTSB.Source: §11.5.1 · pp. 169–173

12.4.1 Data Acquisition

  • 24,137 cleaned NTSB records and 4,526 ASN records from 2000–2020 supplied narratives and labels for destroyed, substantial, minor, and none damage classes.The extracted report fields were 'analysis Narrative' and 'damage level'.Source: §12.4.1 · p. 174

12.4.2 Text Processing

  • Text preprocessing combined lowercasing, punctuation removal, tokenization, stop-word removal, and lemmatization, followed by TF-IDF and word-embedding feature extraction.Narratives were padded or truncated to length 2,000, with vocabulary size set to 100,000.Source: §12.4.2 · p. 175

12.4.3 Text Classification, Training, and Validation

  • 80% of records were allocated to training and 20% to testing, with 10% of training data reserved for validation.Training minimized categorical cross-entropy using Adam, with validation-loss early stopping to counter overfitting. Optimizer selection was not evaluated.Source: §12.4.3 · p. 175

12.4.4 Neural Network Model Architecture

  • Models used embedding, hidden, and output layers, with ReLU hidden activations, SoftMax output probabilities, and argmax class-label selection.The common architectural framework supported comparison across the neural models.Source: §12.4.4 · p. 176

12.5.2 Confusion Matrices Analysis

  • CNN confusion matrices for equally sized NTSB and ASN datasets exposed category-specific strengths and classification challenges requiring further refinement.The supplied discussion does not identify particular confused categories or quantify their errors.Source: Figure 76 · p. 177, §12.5.2 · p. 177

12.5.3 Performance Across Different Dataset Sizes

  • 0.84 accuracy was maintained by CNN on the smallest NTSB dataset (4,526 instances), whereas GRU performance declined significantly.At 24,137 instances, models performed consistently, with CNN slightly ahead in accuracy and F1 score; LSTM and sRNN remained stable across sizes. Equally sized ASN data yielded moderate LSTM, GRU, and CNN accuracy but relatively lower precision and recall, underscoring source differences and the need for models adaptable to data volume.Source: §12.5.3 · pp. 177–178

12.5.1 Performance Evaluation of Neural Network Models.

  • CNN outperformed LSTM, GRU, and RNN on unseen test samples across accuracy, precision, recall, and F1 measure.The study reports higher predictive accuracy with larger datasets and better NTSB performance than ASN at equal sample counts.Source: §11.5.1 · p. 168, §12.5.1 · p. 176, Table 26 · p. 178

12.5.3 Performance Across Different Dataset Sizes

  • Neural network architectures and NLP techniques influence predictive accuracy and robustness, reinforcing the importance of advanced NLP for analyzing textual aviation safety reports.Incident and accident narratives provide essential insights into events that compromise aviation safety.Source: §12.5.3 · p. 179
  • Larger datasets yield higher accuracy and effectiveness, while even the smallest NTSB dataset outperforms the ASN dataset.The study advocates investment in data collection and access to expansive datasets to strengthen safety incident prediction.Source: §12.5.3 · p. 179
  • Future model evaluations will explore training-epoch variations, parameter fine-tuning, and integration of state-of-the-art BERT models.These are proposed avenues for improving performance, rather than demonstrated gains in this part.Source: §12.5.3 · p. 179
  • ASN's modest performance highlights the need for improved data collection, while future work will examine combining ASN and NTSB datasets.The intended direction is a more robust and comprehensive aviation safety dataset.Source: §12.5.3 · p. 179

Chapter 13. Classification of Operational Records in Aviation Using Deep

This section compares four deep learning models for classifying Socrata aviation narratives into Commercial, Military, and Private operator categories and examines LDA and NMF for topic modelling of NTSB accident reports.

13.4.1 Data Annotation and Preprocessing

  • 4,863 records remained after cleaning the original 4,995-record dataset, with operator labels manually consolidated into Military, Private, and Commercial categories.Hundreds of original operator categories were cross-checked using additional sources and domain knowledge. Incident narratives supplied the text used to predict operator categories.Source: §13.4.1 · p. 183, Chapter 13. Classification of… · p. 183

13.4.2 Text Processing

  • 200-word input sequences were created using Keras tokenization, zero-padding shorter narratives, and truncating longer narratives.Preprocessing removed stop words, punctuation, and special characters using NLTK and SpaCy. Data were split into 80% training, 10% testing, and 10% validation.Source: §13.4.2 · p. 184

13.4.3 DL Model Architecture

  • All models used a shared embedding–hidden–output architecture, with model-specific modifications, ReLU hidden activations, SoftMax outputs, and argmax class selection.RNNs face vanishing or exploding gradients; LSTM manages longer dependencies, while BLSTM captures forward and backward context. CNNs capture local textual features.Source: Chapter 13. Classification of… · p. 181, §13.4.3 · p. 184

13.5.1 Model Performance: Classification Report

  • 85% Commercial recall for sRNN contrasted with zero Private precision, recall, and F1-score.CNN achieved 81% precision and 79% recall for Commercial records, but substantially weaker precision and recall for Military and Private records.Source: Figure 78 · p. 186, §13.5.1 · p. 186

13.5.2 Confusion Matrix Analysis

  • BLSTM showed relatively fewer Commercial and Military misclassifications, while LSTM classified Private records better than sRNN but worse than BLSTM.CNN classified Commercial records strongly but frequently confused Military and Private records; sRNN also substantially misclassified Private records.Source: Figure 79 · p. 186, §13.5.2 · p. 187

13.5.3 Model Accuracy Comparison

  • 72% accuracy placed BLSTM first, followed by LSTM at 71%, sRNN at 69%, and CNN at 67%.Evaluation included precision, recall, F1-score, confusion matrices, and validation metrics. The authors suggest that capturing temporal dependencies contributed to BLSTM and LSTM outperforming the other models.Source: §13.4.4 · p. 186, §13.5 · p. 186, Table 27 · p. 187

13.5.4 Macro Average Comparison

  • BLSTM and LSTM outperformed CNN and sRNN in macro-average precision, recall, and F1-score, but all models struggled with Private records.Possible causes included class imbalance or insufficient distinguishing features. Proposed remedies included data augmentation, oversampling, and feature engineering.Source: §13.5.4 · p. 187, Figure 80 · p. 188

13.5.4 Validation metrics

  • BLSTM achieved the lowest validation loss and highest validation accuracy, indicating stable learning and stronger generalization.LSTM followed closely with slightly higher validation loss. CNN and sRNN had higher validation losses, suggesting weaker generalization and potential overfitting.Source: Figure 81 · p. 188, §13.5.4 · p. 188
  • BLSTM was the most effective classification model, but further refinements were needed to address limitations in multiclass performance.Proposed improvements included augmentation and oversampling for class imbalance, attention mechanisms for underrepresented categories, ensemble learning, and domain-specific feature engineering for minority classes.Source: §13.5.4 · p. 189

Section IV. Topic modelling

  • The study compares LDA and NMF to extract interpretable accident-report themes, addressing the relative scarcity of comparative aviation topic-modelling studies.The intended safety applications include identifying recurring themes and contributing factors to inform regulatory enhancements, training programs, and aviation technology.Source: Section IV. Topic modelling · pp. 189, 190, 192

14.4.1 Data Acquisition

  • More than 36,000 JSON records from the NTSB dataset covered 2000–2020, including investigation narratives, findings, and recommendations.The study exclusively sourced its aviation incident and accident investigation reports from this dataset.Source: §14.4.1 · p. 192

14.4.2 Text Processing

  • Narrative preprocessing combined tokenization, lowercasing, cleaning, stopword removal, and lemmatization with TF-IDF and Word2Vec feature extraction.Cleaning removed punctuation, URLs, HTML tags, and non-alphanumeric characters. TF-IDF weighted terms, while Word2Vec encoded semantic relationships through dense vectors.Source: §14.4.2 · p. 193

14.4.3 Topic Modeling Procedure

  • LDA models documents as topic mixtures, while NMF factorizes the document-term frequency matrix into topic and term-distribution matrices.LDA uses Dirichlet sampling and multinomial word generation. NMF decomposes the original matrix V into lower-dimensional matrices W and H.Source: §14.4.3 · pp. 194–195

14.5.1 Topic Extraction and Coherence Evaluation

  • 0.497 for LDA and 0.437 for NMF were the C_v coherence scores, indicating greater semantic coherence for LDA-generated topics.Higher C_v values indicate more coherent topics.Source: §14.5.1 · p. 195

14.5.2 Topic Distribution

  • Ten topics per model showed distinct NMF clusters but some overlapping LDA themes, with varying numbers of narratives assigned to each topic.NMF produced distinct, granular topics supporting focused analysis of accident aspects. Topic distributions, a pyLDAvis intertopic-distance map, and word clouds supported interpretation.Source: Section IV. Topic modelling · p. 190, §14.5.2 · pp. 195, 197, Figure 85 · p. 196

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

This section compares topic-modelling approaches applied to NTSB, ATSB, and Socrata aviation incident narratives, complemented by clustering, to identify recurring safety themes and assess coherence, interpretability, and predictive performance. The analyses cover mechanical, operational, human, and environmental themes, alongside automated cause prediction and model explainability.

9. Aircraft Control and Runway Operations

  • The earlier LDA–NMF comparison found stronger topic coherence and broader themes with LDA, but more specific, granular accident topics with NMF.Both models covered operations, mechanical issues, fuel systems, engine performance, weather, and flight instruction; co-occurrence matrices examined topic overlaps. Model selection depends on broad coverage versus detailed analysis, and hybrid approaches were proposed for future research.Source: §9 · pp. 197–200

15.5.1 Topic Extraction and Coherence Evaluation. / 15.4.6 Evaluation Metrics

  • 0.597 coherence for LDA exceeded pLSA (0.583), LSA (0.542), and NMF (0.437) in the incident-narrative comparison.LDA’s score was identified as C_v, which measures semantic coherence. Evaluation also included interpretability metrics assessing how understandable the generated topics were to humans.Source: §9 · p. 200, §15.5.1 · pp. 206–207, §15.4.6 · p. 206

15.4.5 K-means Clustering

  • K-means grouped incidents by shared characteristics, revealing commonalities and distinctive insights within incident narratives.The algorithm assigned observations to their nearest cluster mean and iteratively refined assignments to minimize within-cluster variance.Source: §9 · p. 200, §15.4.5 · p. 206

15.4.1 Data Acquisition

  • More than 36,000 NTSB investigation records from 2000–2020 supplied textual narratives, findings, and recommendations in JSON format.The study sourced its incident and accident investigation reports exclusively from the NTSB dataset.Source: §15.4.1 · p. 202

15.4.2 Text Processing

  • Tokenization, lowercasing, punctuation removal, stopword elimination, and URL removal prepared narratives for analysis.TF-IDF represented term importance, while Word2Vec supplied dense word vectors capturing semantic relationships. Additional cleaning removed HTML and non-alphanumeric characters, and lemmatization standardized words to their base forms.Source: §15.4.2 · p. 203

15.4.3 Topic Modeling Techniques / 15.4.7 Implementation

  • LDA represented documents as topic mixtures, while NMF decomposed the document-term matrix into lower-dimensional topic and term-distribution matrices [17].Implementation used Python and Jupyter Notebook with NLTK, Gensim, and Scikit-Learn; techniques were applied iteratively with parameter fine-tuning.Source: §15.4.3 · pp. 203–204, §15.4.7 · p. 206

15.4.4 Semantic Analysis

  • LSA reduced dimensionality to uncover semantic relationships, whereas pLSA modelled probabilistic document-word relationships through latent topics.pLSA’s document-topic and topic-word distributions were trained using expectation-maximization to estimate parameters involving unobserved topics.Source: §15.4.4 · pp. 205–206

15.5.1 Topic Extraction and Coherence Evaluation.

  • LDA distributes attention across all topics, whereas NMF primarily emphasizes topics 4, 6, and 8 in the illustrated document.The supplied passages reference coherence evaluation but provide no coherence values or supported model ranking.Source: Figure 97 · p. 207, §15.5.1 · p. 207

15.5.3 Top Words for Each Topic and Word Clouds

  • Topic three’s words receive elevated scores in pLSA’s visualization of ten selected words per topic, indicating their prominence within the document.The analysis also presents topic distributions and thematic word clouds.Source: §15.5.3 · p. 207, Figure 98 · p. 208

15.5.4 pLSA and LSA models performance

  • LSA visualizes term weights for the first four topics, alongside explained variance and word clouds, to inform topic-number selection.These representations characterize the underlying thematic structure of aviation incident narratives.Source: §15.5.4 · p. 208, Figure 99 · p. 209

15.5.7 K-means Clustering of Incident Narratives

  • 10 distinctive K-means clusters organize incident narratives by shared characteristics, providing an interpretable complement to topic modelling.The NTSB analysis visualizes clusters using t-SNE, including themes such as runway incidents; proposed extensions include temporal analysis, additional datasets, and predictive safety models.Source: §15.5.7 · pp. 210–211, Figure 100

15.5.5 Common Themes Across Models:

  • Mechanical failures, fuel-system issues, and engine/power loss recur across topic models, revealing shared themes in aviation incident narratives.NMF, LSA, pLSA, and LDA offer distinct perspectives, but convergent keywords provide a foundation for further safety analysis.Source: Table 31 · p. 209, §15.5.5 · pp. 209–210

15.5.6 Unique Insights from Each Model:

  • NMF highlights helicopter operation and flight conditions, while LSA reveals relationships between wind/weather impacts and rudder/brake issues.These model-specific insights complement the recurring themes found across algorithms.Source: §15.5.6 · p. 210

16.4 Methodology

  • 50,778 records remain after cleaning 53,275 ATSB records from 2013–2022 for a comparison of pLSA, LSA, LDA, and NMF.The subsequent study aims to compare the techniques’ ability to uncover insights supporting aviation safety and risk assessment; these passages provide scope rather than comparative results.Source: §15.5.7 · pp. 212, 215
  • The ATSB pipeline cleans and tokenizes narratives, applies lemmatization, and extracts TF-IDF and Word2Vec features before implementing each topic model separately.Cleaning includes lowercasing and removal of punctuation, stopwords, URLs, HTML tags, and irrelevant characters; the procedure also describes a document-term frequency matrix.Source: §15.5.7 · p. 216

16.5.1 Topic Extraction Using pLSA

  • 10 pLSA topics captured safety themes including bird strikes, engine issues, inspections, cockpit dynamics, communication, air traffic control, landing gear problems, and takeoff incidents.Associated words and probabilities characterized each topic, providing a foundation for further investigation, risk assessment, and accident prevention.Source: Table 33 · p. 218, §16.5.1 · pp. 218–220

16.5.2 Topic Extraction Using LSA

  • 10 LSA topics summarized safety-related events in aviation incident narratives through associated thematic keywords.Explained variance was examined across topic counts to help determine an optimal number of topics.Source: Figure 107 · p. 219, Table 34 · p. 220, §16.5.2 · p. 220

16.5.3 Topic Extraction Using LDA

  • Topics 1 and 4 were the most significant among ten LDA topics extracted from aviation incident narratives.The themes included engine issues, flight operations, pilot incidents, and bird strikes; word clouds summarized frequently occurring terms.Source: §16.5.3 · pp. 220–221, Figure 109 · p. 221

16.5.4 Topic Extraction Using NMF

  • 10 NMF topics covered engine and aircraft operations, pilot incidents, bird strikes, post-flight inspections, landing gear, runway safety, and aircraft damage.Topics 1, 4, 7, and 8 were selected for further investigation because of their prominence and relevance to aviation safety.Source: §16.5.4 · pp. 221–222, Figure 111 · p. 222, Figure 112 · p. 222

16.5.5 Comparative Analysis of Topic Modelling Techniques

  • 0.41 C_v for BERTopic exceeded 0.37 C_v for PLSA, with aviation safety experts also validating BERTopic’s superior interpretability.The ATSB comparison identified tradeoffs: pLSA offered interpretable discovery but scalability challenges; LSA reduced dimensionality and noise but depended on preprocessing; LDA provided topic distributions but required careful hyperparameters and interpretation; NMF offered interpretable, scalable topics but struggled with sparse data and required nonnegative inputs. Proposed extensions included advanced NLP, ensembles, aviation-specific models, real-time incident streams, and predictive risk assessment.Source: §16.5.5 · pp. 222–224

17.4.3 Topic Modeling Procedure

  • BERTopic used transformer-based embeddings and hierarchical clustering, whereas PLSA used probabilistic modelling with expectation-maximization; both were implemented separately on preprocessed text.PLSA’s bag-of-words representation limits semantic context; prior studies also report overfitting and the lack of a well-defined generative model for new documents.Source: §16.5.5 · pp. 224, 226, §17.4.3 · p. 228

17.4.1 Data Collection

  • Over 36,000 NTSB investigation records from 2000–2020 supplied narratives, findings, and recommendations for the BERTopic–PLSA comparison.The reports were sourced exclusively from the NTSB and obtained in JSON format.Source: §16.5.5 · p. 224, §17.4.1 · p. 227

17.4.2 Text Processing

  • Shared preprocessing applied NLTK tokenization, lowercasing, stopword removal, WordNet lemmatization, and filtering of punctuation, numbers, special characters, and other non-alphabetic content.The same cleaned corpus supported a consistent comparison between BERTopic and PLSA.Source: §17.4.2 · p. 227

17.5.5 Topic Word Distributions

  • BERTopic emphasized focused domain-specific vocabulary, assigned higher weights to those terms, and produced clearer six-topic themes than PLSA.Word clouds, word-score distributions, and top-ten-word summaries supported these observations. Experimental settings used six topics for PLSA and a minimum cluster size of 15 for BERTopic.Source: §17.4.7 · p. 228, §17.5.1 · p. 229, §17.5.4 · p. 230

17.5.6 Comparison of Topics and Themes: BERTopic vs. PLSA

  • BERTopic organized shared aviation themes contextually, while PLSA emphasized specific technical problems and statistical co-occurrence.Transformer embeddings and HDBSCAN supported BERTopic's semantically related clusters and interpretable hierarchy. PLSA used Gensim and Expectation-Maximization to estimate topic-word and document-topic distributions, emphasizing discrete issues such as fuel shortages.Source: §17.4.5 · p. 228, §17.4.4 · p. 228, Figure 116 · p. 229

17.6 Conclusion

  • 0.41 coherence for BERTopic exceeded PLSA's 0.37, alongside better interpretability and lower computational demands on large datasets.Contextual embeddings supported BERTopic's semantic consistency, whereas PLSA's bag-of-words representation produced overlapping topics. Evaluation included C_v coherence, aviation-expert inspection, computational time, and memory usage.Source: §17.4.6 · p. 228, §17.5.3 · p. 229, Table 38 · p. 232

17.5.8 Intertopic Distance Map for BERTopic: / 17.5.9 Word Distribution of Topics for PLSA:

  • BERTopic's distance map showed distinct semantic topics, whereas PLSA's dominant Topic 4 spanned most documents, limiting thematic diversity and granularity.The visual analysis contrasted BERTopic's contextually coherent separation with PLSA's emphasis on high-frequency patterns.Source: §17.5.8 · pp. 232–233, Figure 119 · p. 234, §17.5.9 · p. 234

17.6 Conclusion

  • 0.36 coherence and 38.2 perplexity were reported for LDA, versus NMF's 0.34 and 37.1 and PLSA's 0.32 and -4.6, respectively.The subsequent Socrata study identified pilot error, mechanical failure, weather conditions, and training deficiencies. LDA uncovered overlapping themes, NMF produced distinct interpretable topics, and PLSA offered probabilistic insights with interpretative complexity.Source: §17.6 · p. 236
  • Operator-specific modelling addresses overlapping narrative language and the underexplored distinction between military, commercial, and private aviation, supporting targeted interventions and policy development.Prior FAA-report LDA analyses and regional-airline PLSA research did not distinguish operator categories, while aviation applications of NMF remained underexplored. PLSA's computational complexity and overfitting tendency also limited adoption.Source: §17.6 · pp. 237–239

18.4.1 Data Preprocessing / 18.4.2 Text Preprocessing

  • 4,995 Socrata records, spanning September 17, 1908, to June 8, 2009, supplied narratives explicitly identifying military, commercial-airline, or private operators.Cleaning removed special characters, numbers, and extraneous symbols; tokenization, aviation-specific stopword removal, and lemmatization standardized the text. Incomplete or ambiguous records were excluded, yielding three operator-specific subsets using NLTK and Pandas.Source: §18.4.1 · p. 240, §18.4.2 · p. 240

18.4.3 Topic Modeling Procedure

  • The operator-focused procedure transformed narratives into document-term frequencies and separately implemented LDA, PLSA, and NMF, evaluating coherence and perplexity.Gensim LDA treated documents as topic mixtures and optimized topic counts by coherence; custom-package PLSA estimated probabilistic distributions with cross-validation against overfitting. Scikit-learn NMF factorized the frequency matrix into non-negative topic and word-distribution matrices [383], tuning topic counts for coherence and interpretability.Source: §18.4.3 · pp. 241–242
  • LDA, NMF, and PLSA were evaluated on Socrata data using coherence for topic interpretability and perplexity for predictive performance.Word clouds and topic distributions supported thematic interpretation; higher coherence and lower perplexity were treated as preferable.Source: §18.4.3 · p. 243, §18.5.3 · pp. 246–247

18.5.1 Topic Distribution Across the Dataset

  • Topic 0 dominated LDA’s relatively balanced distribution, whereas NMF concentrated on Topic 3 and PLSA on Topic 6.These distributions reflected differences in how the models represented the dataset’s underlying themes.Source: Figure 122 · p. 244, Figure 123 · p. 244, §18.5.1 · p. 244

18.5.3 Performance Metrics

  • 0.36 coherence and 38.2 perplexity made LDA most interpretable but least predictive, whereas PLSA’s 0.32 coherence and -4.6 perplexity favored prediction.NMF achieved 0.34 coherence and 37.1 perplexity, placing it between LDA and PLSA on both reported metrics.Source: Figure 128 · p. 247, §18.5.3 · p. 247
  • 1908–2009 aviation records spanning military, commercial, and private operators revealed pilot error, mechanical failures, weather effects, and military-specific training issues.Operator-based categorization exposed sector-specific patterns that could guide safety protocols, policy, targeted interventions, and resource allocation.Source: §18.5.3 · pp. 248–249

18.5.2 Topic Extraction Using LDA

  • LDA captured broader overlapping themes, NMF produced distinct operational and environmental topics, and PLSA emphasized specific accident scenarios with interpretability challenges.LDA covered flight conditions, pilot decisions, mechanical failures, runway incidents, and crew roles; NMF highlighted engine emergencies, weather, and mountainous terrain. PLSA captured nuanced probabilistic term relationships involving mechanical failures, pilot actions, and environmental accident factors.Source: §18.5.2 · pp. 245–246, §18.5.3 · p. 248

18.5.3 Performance Metrics

  • Future topic-modelling work proposes richer contextual variables, neural topic models, and sentiment-analysis integration to support automated incident reporting and risk assessment.Additional variables include environmental factors, aircraft types, and operational contexts; neural models could better handle large datasets and complex term-topic relationships.Source: §18.5.3 · p. 249
  • 0.786 average cosine similarity was achieved by the VAE for automated likely-cause generation, supporting its potential investigative use.Recurrent, convolutional, residual, and hybrid architectures supported classification and causal-factor prediction, while topic models extracted latent incident patterns. SHAP explained feature contributions, and transformer findings emphasized long-input NLP for incident-cause prediction.Source: §18.5.3 · p. 250
  • Future research proposes multimodal inputs, domain-adapted transformers, real-time warning systems, improved imbalance handling, reinforcement learning, and regulatory, ethical, and human–AI evaluation.Suggested inputs include telemetry, pilot communications, and sensor readings; larger aviation-specific datasets and GAN-based augmentation could improve robustness. User-centered evaluations would assess practical adoption in investigations and safety decisions.Source: §18.5.3 · p. 251

Related papers

Loading 2608.14746v1…