Analyzing Unmanned Aircraft System (UAS) Incidents from NASA ASRS Data Using Unsupervised Machine Learning - Research Poster Competition 2025

Aug 4, 2026

Upload by Riley

The poster presents an efficient workflow for analyzing Unmanned Aircraft System (UAS) safety incident reports from NASA’s Aviation Safety Reporting System (ASRS) using unsupervised machine learning. Instead of manual parsing that is time-consuming and potentially subjective, the approach filters reports from 2013–2023 under specified regulations, preprocesses text with bag-of-words and TF-IDF/LSA, and applies K-Means clustering with silhouette evaluation and topic modeling via LDA. The results identify meaningful topic structure and recurrent n-grams, supporting faster pattern discovery for safety insights.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages2

Reading Time5 min

AccessFree to download

Introduction

Literature Review

Results

Research Methods

Conclusions and Further Work

References and Information

University of Arkansas, Fayetteville ScholarWorks@UARK | 2025 Research Poster Competition | Research Poster Competition | | --- | --- | | 2025
Analyzing Unmanned Aircraft System (UAS) Incidents from NASA ASRS Data Using Unsupervised Machine Learning
Kacey Haws
University of Arkansas, Fayetteville
Neelakshi Majumdar
University of Arkansas, Fayetteville
Follow this and additional works at: [https://scholarworks.uark.edu/hnrcsturpc25](https://scholarworks.uark.edu/hnrcsturpc25)
Part of the Engineering Commons
Click here to let us know how this document benefits you. | | Citation Haws, K., & Majumdar, N. (2025) . Analyzing Unmanned Aircraft System (UAS) Incidents from NASA ASRS Data Using Unsupervised Machine Learning. 2025 Research Poster Competition. Retrieved from [https://scholarworks.uark.edu/hnrcsturpc25/37](https://scholarworks.uark.edu/hnrcsturpc25/37) This Book is brought to you for free and open access by the Research Poster Competition at ScholarWorks@UARK. It has been accepted for inclusion in 2025 Research Poster Competition by an authorized administrator of ScholarWorks@UARK. For more information, please contact [uarepos@uark.edu](uarepos@uark.edu). Analyzing Unmanned Aircraft System (UAS) Incidents from NASA ASRS Data Using Unsupervised Machine Learning Kacey Haws, Neelakshi Majumdar Ph.D. University of Arkansas, Fayetteville, AR Introduction NASA Aviation Safety Reporting System (ASRS): voluntarily submitted aviation safety incident reports in their database [1] . Previous work in the ASYST lab involved manually parsing UAS reports: time consuming, technically challenging, and can lead to subjectivity. My research analyzes UAS incident using unsupervised machine learning to discern patterns quicker than manual analysis. This aligns with Dr. Majumdar’s research by further expanding on her work with natural language processing (NLP) to analysis using unsupervised machine learning methods [2] . Literature Review ▪ Topic Modeling: Effective in parsing ASRS narratives, but work did not focus on UAS incidents [3] . ▪ Supervised Learning: Works well on varied amounts of data but requires dataset be labeled before processing [4] . ▪ Large Language Models: Can assist in identifying near-miss incidents in reports but is “guessing” based on its interpretation of the rest of the sentence [5] . ▪ Sparse Machine Learning: Can address user interpretability, but unsupervised methods are better when the data is unlabeled [6] . Results Figure 1 : Silhouette Score for Different Cluster Counts The silhouette score analyzes the density of clusters. The highest density was ten clusters with a silhouette score of~0.039 [Fig. 1] . After setting it to allocate for 10 different topics, Topic 7 computed its top seven terms as shown [Fig. 2] . Figure 2: Top Terms Discovered Using Topic Modeling Figure 3: Top 10 Most Common N-grams I found that 4-grams best allow for data clarity and varied frequencies. This data is sorted by frequency found in the text [Fig. 3] . This research has been supported by a grant from the Honors College. Research Methods ▪ Dataset: UAS reports from 2013 to 2023 filtered by operations under Federal Aviation Regulations (FAR)“Public Aircraft Operations (UAS) or Recreational Operations / Section 44809 (UAS)”. ▪ Used bag-of-words to eliminate common terms. ▪ Clustering: ▪ Term Frequency-Inverse Document Frequency (TF-IDF) method for term weight. ▪ Reduced by latent semantic analysis (LSA) ▪ Grouped with K-Means and silhouette score. ▪ Topic Modeling: ▪ TF-IDF for term weight. ▪ Latent Dirichlet Allocation (LDA) distribution using Gibbs formula [Fig. 4] . ▪ Computed top n terms for each topic. Figure 4: Gibbs sampling formula [7] . ▪ N-Grams: ▪ Converted the text to lowercase. ▪ Computed a list of n consecutive words. Conclusions and Further Work Unsupervised machine learning techniques offer a more efficient w

FAQs

Why does the research move away from manual parsing of ASRS UAS reports?

Manual parsing is time-consuming and technically challenging, and it may introduce subjectivity. The study uses unsupervised machine learning to find patterns more efficiently.

What dataset and time range are used in the analysis?

The dataset includes UAS reports from 2013 to 2023 filtered by operations under specified Federal Aviation Regulations for Public Aircraft Operations (UAS) or Recreational Operations (Section 44809).

Which unsupervised techniques drive the results, and how are topics evaluated?

The method uses TF-IDF weighting with latent semantic analysis, K-Means clustering evaluated by silhouette score, and topic modeling with LDA (via Gibbs sampling). It also extracts top terms and n-grams to characterize discovered topics and recurring phrases.

Aug 4, 2026

Upload by Riley