Hard landing is one of the typical risk events that affect the operational safety of civil aviation. However, existing studies remain insufficient in jointly modeling and synergistically mining massive QAR flight parameters together with safety information texts. Focusing on the hard-landing scenario, this paper proposes a QAR time-series data–safety text fusion framework, Hard Landing oriented Multimodal Prototype-guided Transformer (HL-MPT), for multimodal representation learning to identify causal indicators of hard landing and enable automatic generation of safety information. First, the safety information texts are tokenized and cleaned, and a hierarchical fine-grained labeling system is constructed along the dimensions of human operation, environment, system, and management/crew factors. An encoder network based on the self-attention mechanism is then adopted to learn semantic representations of the texts. Meanwhile, for the aligned QAR time-series data, a temporal encoding network is employed to extract dynamic features of flight parameters. Subsequently, a cross-modal interaction and fusion mechanism is designed to obtain joint text–parameter representations, and a multi-label classification model is built for hard-landing cause identification, outputting cause labels and their associated key parameter indicators. On this basis, a conditional generation module is introduced to generate structured safety information texts conditioned on QAR features, realizing an automated mapping from flight parameters to report narratives. Experiments on 478 real-world aviation records demonstrate that, compared with unimodal baselines, the proposed method improves Micro-F1/Macro-F1 by more than 5% on average for multi-label cause identification. For report generation, it increases ROUGE-L by over 3% and raises the human factual-consistency score by \(\ge 0.3\) points on a 5-point scale.
