An application of Deep Neural Networks to the in-flight parameter identification for detection and characterization of aircraft icing

This paper applies the Deep Neural Networks to the in-flight parameter identification for detection and characterization of the aircraft icing. General dynamics of the aircraft are firstly presented, ice effects on the dynamics are characterized. Deep Neural Networks (DNNs) including Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) are briefly introduced. We propose a state-image approach for the pre-processing of the input flight state, then we design a DNN structure which models both local connectivity (using CNN) and temporal characteristics (using RNN) of the flight state. The identified parameters are exported from the DNN output layer directly. To fully evaluate the performance of the DNN-based approach, we conduct simulation tests for different cases which correspond to clean and aircraft icing at different locations (wing, tail, wing and tail) with different severities (moderate, severe). A comparison of the DNN-based approach with a baseline H-infinity-based identification algorithm (state-of-the-art for aircraft icing) is also delivered. Based on the test and comparison results, the DNN-based approach yields more accurate identification performance for more parameters, which shows promising applicability to the in-flight parameter identification problem. (C) 2018 Elsevier Masson SAS. All rights reserved.