A Novel Time Series Transformer Method for Hard Landing Early Warning for Civil Aviation

Hard landing, as one of the key risks during the flight phase of landing, has become a prominent topic in aviation safety research. Because such events are often accompanied by complex flight-control behaviors and environmental disturbances, there is an urgent need for statistical methods that are both predictive and interpretable to enable early warning. To this end, we propose TimeXer-ALD (Time Series Transformer with eXogenous variables for Aircraft Landing Detection), an improved temporal classification model specifically designed to predict hard-landing risk events. Built on flight Quick Access Recorder (QAR) data, the method defines a detection window around the suspected hard-landing segment and assigns a binary label by checking whether the maximum vertical acceleration VRTG_MAX exceeds a threshold θ. The model then ingests a sequence of historical flight measurements preceding this window as input features. In contrast to conventional approaches that incorporate endogenous variables for prediction, TimeXer-ALD constructs its input features solely from historical exogenous flight parameters. This design effectively avoids potential information leakage and redundant interference, thereby enhancing the model's stability and generalization capability. To enhance interpretability, we incorporate SHAP (SHapley Additive exPlanations) to quantify the importance of each input variable across the full time window, thereby identifying key contributing factors from a statistical perspective. Experiments configured under four different early-warning horizons show that TimeXer-ALD achieves consistently strong classification performance. The resulting feature-importance visualizations further highlight the pivotal roles of variables such as pitch angle and spoiler position in the mechanism underlying hard landings, corroborating the scientific validity and practical utility of the proposed approach for flight-safety early warning.