A DA-BiGRU model based on sliding window: aircraft landing gear load prediction using static sensor data

Abstract

Landing gear load monitoring is crucial for early detection of structural hazards and prevention of safety accidents. As the core component of load assessment, the rationality of the monitoring method directly determines the accuracy of the assessment. However, existing methods have obvious limitations: on the one hand, they do not adequately model nonlinear characteristics; on the other hand, they overly rely on complete and continuous time-series data. In engineering practice, due to factors such as acquisition system failures and environmental interference, the above requirements are often difficult to meet, which restricts the effectiveness of monitoring. To address this gap, this article proposes a novel framework that transforms disordered static measurements into structured pseudo-time-series using a sliding window approach, enabling the capture of dynamic load patterns without any data interpolation. A dedicated dual-attention bidirectional gated recurrent unit network is designed to model these sequences, with a weighted pinball loss employed to balance prediction errors across multiple axes. Extensive experiments demonstrate that the proposed model achieves mean absolute errors of 0.11, 0.38, and 0.05 kN, and coefficient of determination values of 0.9981, 0.9949, and 0.9989 in the X (heading), Y (lateral), and Z (vertical/longitudinal) directions, respectively, outperforming all baseline and state-of-the-art methods on every metric. This work provides a robust and practical solution for high-precision landing gear load prediction under real-world data constraints.