A novel Lamb wave-based multi-damage dataset construction and quantification algorithm under the framework of multi-task deep learning

Abstract

Lamb wave-based damage quantification in large-scale composites has always been one of the concerning and intractable problems in aircraft structural health monitoring. In recent years, machine learning (ML) algorithms have been utilized to deeply explore the damage feature of Lamb wave signals, which aims to enhance the precision and accuracy of damage quantification. However, multi-damage quantification becomes one of the bottleneck problems because ML algorithms critically depend on the dataset. In this paper, a prioritizing selection and orderly permutation method is proposed to construct multi-damage dataset based on Born approximation principle, which shows the interaction between wave signals under multi- and single-damage conditions. Based on the multi-damage dataset, a multi-task deep learning algorithm is introduced to identify multiple damage, including the damage number, location, and size, in composite laminates. In the algorithm, a multi-branch 1D-convolution neural network framework, which includes a trunk network and branch networks is established to explore the damage features in Lamb wave scattering signals. Compared with single-task models, it has the ability to learn shared features for multiple tasks, effectively boosting the task results. The results show that the proposed multi-task learning (MTL) method saves 23.03% training time compared with the single-task learning method. In the task of quantifying multiple damage of composite laminate, the results of MTL are good for both the constructed test set and the measured test set, especially in the quantification of damage size, which shows the feasibility and reliability of this method.