Answers are grounded in this paper and cite their sources.
Aviation accident and incident investigations generate extensive unstructured textual information containing evidence relevant to the causes and contributing factors of safety occurrences. Automatically identifying such information is challenging because relevant evidence may be distributed across long and complex investigation narratives. This study investigates a variational-autoencoder-based framework for predicting representations of expert-reported probable causes from aviation investigation narratives, with the broader objective of supporting the identification of probable-cause information when the corresponding causal statement is unavailable to the prediction system. The framework uses pretrained GloVe embeddings reduced to single-channel numerical sequences, which are processed by a convolutional encoder with a 20-dimensional probabilistic latent representation. The decoder predicts numerical representations of probable-cause statements using a cosine-similarity-based objective with Kullback–Leibler divergence regularization. The framework was evaluated using 20,919 finalized U.S. National Transportation Safety Board investigation records from 2005 to 2020, including 3,138 independent test observations. The variational model achieved a mean test cosine similarity of 0.299585 (SD = 0.240962), compared with 0.296384 for a matched deterministic encoder–decoder. The mean paired improvement was 0.003202, with a 95% bootstrap confidence interval of [0.001577, 0.004838], indicating a statistically detectable but small advantage. Qualitative illustrations identified aviation-related terminology associated with selected investigation circumstances. These exploratory examples are presented separately from the quantitative evaluation of numerical probable-cause representations. The findings demonstrate the feasibility of learning numerical correspondences between investigation narratives and documented probable-cause representations while highlighting limitations in representation quality, prediction agreement, and textual interpretability. The framework is intended as an exploratory approach to aviation safety information analysis rather than an independent system for causal determination.
No takes yet. Share an insight, caveat, or question.
Nanyonga et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: