Prediction model for civil aircraft hard landing Informer neural network

基于Informer的民用飞机硬着陆预测模型

¹ Key Laboratory of Civil Aircraft Airworthiness Technology, Civil Aviation University of China, Tianjin 300300, China
² Department of Science and Technology, Civil Aviation University of China, Tianjin 300300, China
³ CETC Avionics Company Limited, Chengdu 611731, China
⁴ Equipment Management and UAV Engineering College, Air Force Engineering University, Xi'an 710051, China
⁵ Sino-European Institute of Aviation Engineering, Civil Aviation University of China, Tianjin 300300, China

Received: 9 September 2024

Abstract

The prediction of hard landing is of great significance to support the pilot's landing decision and ensure flight safety. Aiming at the short prediction time and poor explainability of the existing hard landing prediction models, the Informer hard landing prediction model based on QAR data is proposed, and the explainability analysis for the output of the model prediction is conducted. Firstly, the QAR data are processed by using a series of appropriate methods such as forward-backward filtering and Granger causality test to establish a dataset that meets the prediction demand. Secondly, according to the characteristics of multivariate time series discontinuity, the Informer network is localized to extend the range of accepted data. Then, the hyper parameters of the model are optimized in training and testing to ensure that the model has high prediction accuracy and good generalization. Finally, the output attention weight matrix improves the transparency of the model decision. Experiments show that the localization-improved Informer neural network improves the prediction accuracy by 23.5% and enhances the learning ability of discontinuous multivariate time series; compared with the LSTM network, the prediction accuracy improves by 18.83% overall, and it has a better ability to predict long time series.

摘要

硬着陆的预测对支持飞行员着陆决策、保障飞行安全具有重要的意义。针对现有硬着陆预测模型预测时间较短以及可解释性差等问题, 提出基于QAR数据的Informer硬着陆预测模型, 并针对模型预测的输出进行可解释性分析。利用前向后向滤波、Granger因果检验等一系列合适的方法处理QAR数据, 建立满足预测需求的数据集。根据多元时间序列不连续的特点, 将Informer网络做本地化处理, 扩展接受数据的范围。在训练和测试时优化模型的超参数, 确保模型具有较高预测精度和良好泛化性。输出注意力权重矩阵提高模型运行逻辑的透明度。实验表明, 本地化改进后的Informer神经网络预测精度提升了23.5%, 增强了对不连续多元时间序列的学习能力; 相比LSTM网络, 预测精度整体提升了18.83%, 且具备更好的长时间序列预测能力。