Aircraft Ice Accretion Prediction Using Neural Network and Wavelet Packet Transform

Purpose - This paper presents a new technique based on the combination of wavelet packet transform (WPT) and artificial neural networks (ANNs) for predicting the ice accretion on the surface of an airfoil. Design/methodology/approach - Wavelet packet decomposition is used to reduce the number of input vectors to ANN and improves the training convergence. An artificial neural network is developed with five variables (velocity, temperature, liquid water content, median volumetric diameter and exposure time) taken as input data and one dependent variable (the decomposed ice shape) given as the output. For the purpose of comparison, three different artificial neural networks; back-propagation network (BP), radial basis function network (RBF) and generalized regression neural network (GRNN), are respectively trained to simulate the wavelet packet coefficients as a function of the in-flight icing conditions. Findings - The predicted ice accretion shapes are compared with the corresponding results from previously published NASA experimentation, LEWICE and the Fourier-expansion-based method. It is found that the BP network has an advantage on predicting the rime ice, and the RBF network is relatively suitable for the glaze ice, while the GRNN can be applied for both without classifying the specimens. The results also show an advantage of WPT in performing the analysis of ice accretion information and the prediction accuracy is improved as well. Practical implications - The proposed method is open to further improvement and investment due to its small computational resource requirement and efficient performance. Originality/value - The simulation method combining ANN and WPT outlined here can lay the foundation for further research relating to ice accretion prediction under different ice cloud conditions.