Research on small-sample aviation accidents prediction using a coordinate attention-based Bi-LSTM model

Abstract

Aviation accidents are distinguished by an extremely low probability of occurrence coupled with severe consequences, rendering the achievement of high-precision accident prevention paramount. While existing aviation accident prediction approaches a rely on large sample sizes for forecasting purposes, making it difficult to extract meaningful insights from limited and small datasets. To address the challenges, we propose a Coordinate Attention-based Bidirectional Long Short-Term Memory (CA-based Bi-LSTM) model to forecast these trends. This paper utilizes the publicly available Aviation Safety Reporting System (ASRS) dataset, focusing on U.S. domestic flight accidents that occurred between 2012 and 2024. Monte Carlo simulation method is first employed to expand the original datasets, thereby addressing the bias introduced by a small amount of accidental data. The simulation was applied again to estimate the Value at Risk (VaR) for six accident types, forecasting the maximum number of accidents that may occur in the next 5 years at a 95% confidence level. By integrating coordinate attention mechanism, the CA-based Bi-LSTM model more effectively captures temporal dependencies. The VaR analysis aligns closely with the most recent annual data, and the comparative analysis across five models demonstrates that the CA-based Bi-LSTM model achieves competitive performance across multiple accident types. This research contributes to the analysis of aviation accident patterns and provides insights to support strategic prediction and the future development of proactive early warning strategies in aviation safety.