Answers are grounded in this paper and cite their sources.
Voluntary safety-reporting systems contain detailed descriptions of emerging uncrewed aircraft system (UAS) hazards, but their narrative and operational fields may not support one stable taxonomy. This study develops and internally evaluates a purpose-separated framework for neural representation, operating-ecosystem segmentation, and operational-oversight configuration discovery. The data comprised 612 unique UAS reports submitted to the National Aeronautics and Space Administration Aviation Safety Reporting System, with occurrence dates from February 2019 through June 2026. A multimodal variational autoencoder (VAE) integrated sentence-transformer narrative embeddings with structured variables and was benchmarked against conventional clustering. Latent class analysis (LCA) was applied to broad operating-context variables and, more restrictively, to 15 indicators of flight context, control, crew size, visual-observer use, and authorization. The VAE produced a two-group solution with weak repeatability (median adjusted Rand index = 0.264; median normalized mutual information = 0.180) and did not outperform structured K-means, so the evidence did not support a robust neural taxonomy. In contrast, the operational LCA identified four well-separated ecosystem classes (bootstrap median adjusted Rand index = 0.889), and the restricted LCA identified two oversight configurations under the original specification (N = 409 and 203; entropy = 0.916; mean maximum posterior probability = 0.978; bootstrap median adjusted Rand index = 0.938). These configurations distinguished predominantly solo/no-visual-observer operations from crew/visual-observer-supported operations. Additional validation showed near-equivalence to crew size and sensitivity to omitting either staffing indicator. Correcting a flying-context grouping error retained two classes (411 and 201 reports) with 99.7% agreement with the original assignments. Withheld narrative language from the same reports provided internal convergent evidence, and reports coded “Ambiguous” occurred in both configurations. The study contributes a validation-centered approach for determining when unsupervised UAS safety patterns warrant interpretation and shows that descriptive operational profiles can be estimated even when a global AI taxonomy is unsupported. The configurations provide a defensible basis for prospective research on authorization support, role clarity, communication, and handoffs, while the observational design does not establish causal risk differences or intervention effectiveness.
No takes yet. Share an insight, caveat, or question.
Dothang Truong (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: