Enhanced Load Prediction for Aircraft Landing Gear Utilizing Graph Convolutional Neural Network

This study presents an accurate aircraft landing gear load estimation model leveraging graph convolutional neural networks (GCNs), which predicts loads from structural strain distribution data. A ground-based experimental system is established, deploying fiber-grating strain sensors at key landing gear points to gather strain data under various operating conditions for model training. The GCN model undergoes strain-to-load mapping training and testing, with prediction accuracy and stability evaluated using maximum relative error, average relative error, and standard deviation. Results showcase stable and precise predictions, with X, Y, and Z load predictions achieving maximum relative errors of 5.18%, 4.15%, and 3.57%, respectively, and average relative errors of 1.58%, 0.61%, and 0.75%, respectively, alongside low standard deviations of 0.59, 0.74, and 0.46 N. Comparative analyses against multiple linear regression and advanced neural networks [long short-term memory (LSTM), convolutional neural network (CNN), and multilayer perceptron (MLP)] underscore the GCN model’s superior prediction accuracy. This work holds significant potential for applications in aircraft structural health monitoring.

I. Introduction

During long-term airplane operation, the landing gear system, frequently subjected to complex and variable mechanical stresses, accumulates fatigue damage, impacting takeoff, landing safety, and flight performance. Load monitoring is pivotal in ensuring flight safety and reliability, offering real-time stress state feedback to enable preventive maintenance, mitigating structural failure and accelerated fatigue due to overload [1], [2], [3], [4], [5], [6]. In addition, load monitoring data validate design-stage load estimation accuracy, assessing in-service load capacity against design expectations [7], [8], thereby preventing safety hazards stemming from underestimated design loads. Advancing beyond traditional strain-based regression coefficient calculations and sensor installations [9], [10], [11], modern load monitoring technology evolves toward increased intelligence and efficiency.