New approach for enhanced damage identification in frame structures with limited data: leveraging integrated convolutional neural network model with transfer learning

Abstract

To address the challenge of high-precision damage identification in frame structures with limited data availability, this article proposes an integrated HOLMSST-DAMIIM-2DCNN model by combing higher-order local maximum synchrosqueezing transform (HOLMSST), dual attention mechanism (DAM), improved inception module (IIM), and two-dimensional convolutional neural network (2DCNN).In this method, the HOLMSST algorithm is initially proposed to transform the one-dimensional response signals into time–frequency images. Subsequently, a new 2DCNN architecture is established by incorporating an IIM with a DAM, and then pretrained on a proportionally divided dataset. Furthermore, the pretrained model is fine-tuned to the target task through a transfer learning (TL) strategy, significantly bolstering the damage identification performance of the proposed model in scenarios characterized by data scarcity. Finally, the accuracy and generalization ability of the proposed HOLMSST-DAMIIM-2DCNN model and TL strategy are validated through four datasets. The results demonstrate that the proposed HOLMSST-DAMIIM-2DCNN model exhibits high accuracy and robust noise resistance in multicategory damage identification tasks for frame structures. Notably, even in the face of a limited number of samples, the integration of the HOLMSST-DAMIIM-2DCNN model with the TL strategy can still enhance the learning capability and damage identification accuracy significantly for frame structures.