KISNet: Knowledge-enhanced independent subnetworks for robust prediction of flight data

Robust prediction is essential for ensuring the reliability of aircraft operations and mitigating safety risks in complex, dynamic environments, as required by condition-based maintenance. However, maintaining generalization across dynamic phase transitions and efficiently quantifying uncertainty remain significant challenges. To bridge this gap, this paper proposes a knowledge-enhanced independent subnetwork (KISNet) model for robust flight data prediction. Based on a multi-input multi-output independent subnetwork architecture, this model efficiently produces predictions with uncertainty estimates in a single forward pass and integrates aviation domain knowledge in two aspects. In data representation, a variable-scale flight phase subdivision method with directional augmentation is proposed to explicitly incorporate flight dynamic characteristics into input features, effectively overcoming cross-phase data distribution discrepancies and training sample imbalance issues. In model optimization, we construct a risk-aware optimization mechanism. By embedding phase-dependent dynamics and risk sensitivity into the learning objective, this mechanism steers the model towards physically consistent and robust updates. Extensive experiments based on real flight data from fixed-wing unmanned aerial vehicles (UAVs) demonstrate that KISNet reduces prediction errors while providing reliable uncertainty bounds and interpretability, showcasing superior robustness and efficiency in high-risk aviation scenarios. While currently validated on UAVs, this model offers a promising baseline for broader applications.