33th Congress of the International Council of the Aeronautical Sciences

04.1 - Aerospace Grade Materials, Structural Analysis, Fatigue and Damage Tolerance


Y. Yang¹, S. Lyu¹; ¹Aircraft Strength Research Institute, China

Deep learning can help to improve the guided-wave-based damage detection of composite structures, but it needs a large number of damage samples. Based on a great amount of simulated damage samples and a small number of real ones, a domain adaptive damage identification model is designed to realize the migration from simulated damage detection to real damage detection. Firstly, guided-wave signals of faked damage are collected extensively in the form of mass attachment on to the structure surface, and corresponding deep learning model based on convolutional-timing-sequential hybrid neural network is designed to achieve a high accuracy of damage detection. Secondly, a certain amount of guided-wave signals of real damage are collected, and a domain adaptive module is adopted by the model, which approximates the data distribution law of simulated damage and real dam-age in the feature space. With this framework, the model could detect the real damage without the labelling process in advance, and the experimental results demonstrate the detection accuracy of 85.7%, which is ahead of other traditional deep learning models.

View Paper