Toward safer flight training: The data-driven modeling of accident risk network using text mining based on deep learning

Abstract The flight training, a critical component of the general aviation industry, exhibits a relatively high severity of risk due to its complexity and the uncertainty inherent in risk interactions. To mine the risk factors and dynamic evolution characteristics affecting flight safety, a data-driven network modeling methodology that integrates text mining with domain knowledge in accident analysis is proposed for the analysis of accident risks specific to flight training. Firstly, flight training accident reports are labeled using domain knowledge gained from accident causation theory to provide basic data for subsequent study. Secondly, the adversarial training algorithm is introduced to enhance the generalization capability of BERT model in processing imbalanced accident textual data. The fine-tuned BERT, Bi-directional Long Short-Term Memory (Bi-LSTM) Conditional Random Field (CRF) algorithm is fused to construct an ensemble algorithm for risk identification, which accomplishes the joint entity-relationship extraction of accident reports. Thirdly, based on the risk identification results, data-driven modeling of the Flight Training Risk Network (FTRN) is performed to quantify the accident evolution characteristics. Then, the aforementioned tasks are meticulously optimized and integrated, subsequently applied to a case study focusing on loss of control in flight (LOCI) accidents. The findings suggest that the identification algorithm effectively and efficiently extracts risk information and interrelationships. Additionally, the network analysis results reveal the key insights into flight training accidents, facilitating the development of holistic risk control strategies. This study provides offers a powerful and innovative analytical tool for safety management departments, enhancing safety and reliability in flight training operations.