Transfer learning in hard landing risk prediction

Abstract In flight safety monitoring, risk prediction demands high accuracy for rare but critical events such as engine failures and weather-related hazards. However, the limited sample size of such extreme cases within a single dataset poses a significant challenge to achieving precise predictions. Aggregating data across sources is a potential solution, but datasets are often not independently and identically distributed. Transfer learning can improve target performance by leveraging related sources, yet classical methods struggle with heavy-tailed data due to the lack of robust loss functions. This paper proposes a transfer learning algorithm based on composite quantile regression (CQR). It identifies and transfers information from similar sources and constructs pseudo response variables to linearise the CQR loss, enabling robust computation under heavy-tailed noise. Extensive simulation studies demonstrate the superior performance of the proposed method. The practical utility of the algorithm is further illustrated through its application to quick access recorder (QAR) flight data for hard landing risk prediction. This work aims to identify and quantify key controllable factors influencing landing safety through a data-driven approach. Our research provides quantitative analytical tools for multiple core safety aspects, including flight operation standardisation, aircraft condition monitoring, operational environment assessment and support resource allocation, thereby enabling more precise and proactive risk mitigation strategies.