Deep Transfer Learning for Guided-Wave Damage Localization in Nonuniform Thermal Fields

Accurate damage localization in nonuniform temperature environments is crucial yet challenging due to temperature-induced disturbances in guided waves. In this paper, a novel structural damage localization method based on deep transfer learning in nonuniform temperature fields is proposed. To address the scarcity of experimental data under nonuniform temperature fields and reduce the training cost of data-driven models, a hybrid approach combining finite element modeling and machine learning is employed to establish a relationship between uniform and nonuniform temperature compensation models, followed by the application of transfer learning. This hybrid approach enhances the credibility of simulation data through model updating and validation, while mitigating data scarcity in nonuniform temperature environments by transferring parameters from finite element models and neural networks. At the same time, a piecewise interpolation method is used to extract path temperature information between sensors in the nonuniform temperature field. The effectiveness of this method was validated through two nonuniform temperature fields. The research results demonstrate that the proposed method successfully identifies damage in nonuniform temperature environments. Compared to nontransfer learning methods, this approach achieves higher training accuracy, faster iterative convergence, and, to some extent, overcomes the issue of sparse data for temperature compensation in nonuniform environments.