• A systematic review of 175 studies on AI and LLM in aviation safety is presented. • Most studies focus on human factors, accident analysis, and operational safety. • The use of LLM in accident analysis and virtual copilots is surging. • Trustworthy and certified AI is the cornerstone for AI in aviation safety. • Hybrid intelligence design shall be considered for better human-AI teaming. The global air traffic is projected to grow significantly in the coming decades, leading to denser airspace and higher operational complexities. Therefore, academic and practitioners are now unleashing the potential of artificial intelligence (AI), particularly the recent advances in large language models (LLM), computer vision, and speech recognition in enhancing aviation safety through advanced cockpit design, AI assistants, human performance monitoring, and supporting air accident investigations. These applications demonstrate a significant promise in enhancing aviation safety. Nevertheless, there are still challenges in applying safe and reliable AI in supporting these safety–critical domains. Indeed, many aviation safety issues, such as accident analysis, human factors, and preventive system designs, are interconnected instead of standalone issues. This systematic literature review explores the recent advances, challenges, and future perspectives on leveraging AI to enhance aviation safety from a macro perspective. Therefore, a framework is established to review relevant studies. First, we identify the relevant literature from initial search, inspection, and screening. After that, we analyse the domains applied and the models leveraged in aviation safety enhancement on the 175 selected studies using content analysis. Then, thematic analysis is applied to reveal the challenges of applying safe and reliable AI in aviation safety. Given the challenges identified, this review recommends future work to incorporate explainable AI, develop AI certification frameworks, design based on hybrid intelligence, and adopt diversified dataset for generalisation.

The development of time-based flow management has significantly enhanced the safety, reliability, and predictability of air traffic control (ATC). Actual flight paths often deviate from these standard terminal arrival routes due to pilots requesting shortcut arrivals or ATC officers implementing holding procedures to alleviate congestion. These deviations exacerbate the dynamic complexity of air traffic management (ATM). To address these challenges, we propose a novel online learning Transformer-bidirectional gated recurrent unit (Transformer-BiGRU) framework for tactical spatiotemporal flight trajectory prediction. BiGRU further obtains bidirectional sequence information to improve the Transformer’s spatiotemporal prediction. The proposed research utilises image processing techniques to produce ATC aeronautical holding instructions from historical automatic dependent surveillance-broadcast data. The framework significantly improves real-time prediction ability and environment adaptability by integrating holding instructions and online learning. Experiment results demonstrate that incorporating holding instructions with the proposed Transformer-BiGRU reduces the mean absolute error by approximately 10% in latitude, 8.9% in longitude, and 13.1% in flight level compared to the best baseline model. Furthermore, the mean deviation error of horizontal distance decreases from 0.49 to 0.42 nautical miles (a 13% improvement). These results confirm that the methodology benefits real-time ATC decision-making in various ATM scenarios and provides valuable insights to assure airspace safety.

Automation has been extensively used in flight operations, so pilots are less involved in actual flight control. With the long idle time during cruising, pilots may have their vigilance level reduced and eventually become out-of-the-loop. This research proposes a two-stage explainable adaptive automation approach to keep pilots in the loop based on Convolutional Neural Networks, Long Short-Term Memory, and EEG data collected from 24 participants in a one-hour simulator-based flight task in each level of automation. Our proposed spatiotemporal model yields test accuracy of 0.9918 and 0.9907 in the first and second stages, respectively, outperforming other benchmarking models by 30.79% and 10.73%, respectively. Furthermore, the Shapley additive explanations are adopted to strengthen the model interpretability and trustworthiness for safety-critical applications. Our model successfully identified that high delta and theta waves with low beta and gamma waves contribute positively to the out-of-the-loop state. It indicates that the classification aligns with the theoretical background and is trustworthy. The trustworthy adaptive deep learning model supports the dynamical automation configuration for improving human-automation collaboration in cruising flights.