A Study on Ice Accretion Prediction and Flight Envelope Analysis Based on Neural Network

Abstract

Wing icing can lead to a flight accident by affects the aerodynamic characteristics of aircraft and destroys flght performance, maneuverability and stability. The ice shape of airfoil is the base of aerodynamic characteristics analysis and icing system.A new method for prediction of ice shape and flight envelope is presented in the paper, by using the data dining and learning ability of neural network. The ice prediction model id established and the generalization of neural network is improved. The flight envelope of an iced airplane was evaluated and the effect of atmospheric parameters on flight envelope is described. The icing data of Business Jet airfoil is used in paper for ice accretion prediction. Using implicit formula to empress the connection between factors and ice shape, the new ice shape can be predicted by learning the icing data and establishing the BP neural network model which reflects the relation between factors and ice shape. The method which combined the regularization and neural network ensembles is used in the paper for enhance the gereralization performance of ice-predicted network by established the neural network model. The ice-shape curves caused by varied parameters were predicted and analysed. The result of predict is adjacent to the test one. There will be preciser and more satisfying as long as it has enough ice-swatch. The lift coefficient and drag coefficient of an airplane are predicted in the paper. Comparing engineering estimation with neural network to find the difference, the flight envelopes of iced airplane for different ice and atmospheric parameters are simulated and the effect of varied parameters on flight envelope is analysed. By testing the predicted data of ice and analyzing the flight envelopes of iced airplane, it is proved that the neural network method can avoid the substantive computation and some presupposition in numeric methods and can simulate engineering precisely, widely and easily. Therefore, the method has a bright future in practical engineering.

Alternate abstract:

飞机机翼结冰会影响飞机气动特性，导致飞行性能、操纵性以及稳定性下降，从而引发飞行事故。翼面冰型的获取是分析飞机动力学特性，改进飞机防冰系统的基础。本文以现有的冰型数据库为基础，对冰型数据进行充分挖掘，尝试提出基于人工神经网络的翼面冰型预测及飞行包线预估新方法，建立冰型预测网络模型仿真预测冰型，采用混合算法改善网络泛化性能。预估某飞机机翼结冰的飞行包线，分析了不同参数对飞行包线的影响。 本文使用Business Jet翼型的冰型数据，将各影响因素同冰型形状的关系用神经网络隐式表达，建立反应大气和飞行条件与冰型形状之间映射关系的BP网络预测模型。通过已有冰型数据资料对网络进行学习训练，仿真预测出新的冰型曲线。 为提高冰型预测网络的泛化能力，本文尝试使用不同的BP算法改善网络泛化性能，采用正则化和网络集成混合算法建立网络模型，仿真预测出不同飞行迎角、液态水含量、液态水滴直径、结冰时间、环境温度、翼面温度等参数变化下的冰型曲线，分析了冰型曲线的变化。仿真结果表明预测出的冰型与试验结果较吻合，若能增加样本容量，会有更高的精度和满意度。 以某飞机为例，对飞机升阻力系数进行仿真预测，比较了神经网络法与工程估算法应用于飞行包线的差异。预估出不同冰型和大气参数变化的结冰飞机飞行包线，并分析其对飞行包线的影响。 通过对预测冰型结果的检验以及飞行包线的分析，证明神经网络法能避免数值算法繁琐的计算过程和一些假设，计算精度高，泛化性能好，简便易行，具有广阔的工程应用前景。