Aircraft landing gear load prediction based on LSTM-KAN network

Abstract

A long short-term memory kolmogorov-arnold network (LSTM-KAN) computational model that can accurately estimate the loads of an aircraft landing gear is proposed, which can accurately estimate the loads based on monitoring the strain distribution of the landing gear structure. First, the ground strain load test system based on the landing gear structure is built. Fiber-optic grating strain sensors were installed at the monitoring positions of the landing gear, and the strain data under various loading conditions were collected using the loading system. Experimental datasets for model training and testing were generated based on the obtained strain and load data, the strain-load calculation model was trained and tested, and its prediction results were compared with the traditional linear regression method and other algorithms using the same experimental datasets. The load prediction results show that the maximum absolute percentage errors of the loads in the three directions are 2.74 %, 2.41 % and 3.29 %, respectively, and the corresponding root-mean-square errors are 51.31 N, 202.29 N and 39.86 N, respectively, and the overall average absolute percentage errors are reduced to less than 1 %, and the three-direction ones are 0.81 %, 0.73 % and 0.88 %, respectively, which proves that the LSTM-KAN model has better performance than the multiple linear regression method and other neural network algorithms, and can be effectively and accurately predicted in the field of health monitoring of aircraft landing gears and other aircraft structures.