Data-Driven Joint Response Prediction for Aircraft Landing Gear Systems

Predicting aircraft landing gear responses under extreme conditions is important for condition monitoring and safety assessment. In oleo-pneumatic landing gear systems, the vertical load and strut stroke are intrinsically coupled through the shock absorber. The vertical load exhibits bimodal or multimodal peaks with large amplitude variations, whereas the strut stroke varies more smoothly with smaller magnitudes. Joint prediction remains challenging since the two responses have distinct waveforms and scales, especially under extreme conditions with limited training data. Existing methods often focus on a single response or a single system, with limited generalization to different configurations. This paper proposes a unified Kolmogorov–Arnold Network (KAN)-Transformer hybrid network framework to jointly predict vertical load and strut stroke responses for two oleo-pneumatic strut landing gear configurations. A time-aware feature enhancement module augments inputs using force rate, stroke velocity, and sliding integral features. The network combines KAN for nonlinear mapping and Transformer for long-range dependency modeling. Multiscale convolutions, enhanced positional encoding, and a time-derivative consistency loss further improve local feature extraction and training stability. Results show accurate extrapolation to unseen extreme conditions and superior performance over baseline deep learning methods. Ablation studies validate that the combination of feature enhancement, hybrid architecture, and consistency loss yields optimal performance.