A heterogeneous decision voting-based transfer domain adaptation method for damage localization of CFRP composite structures

Abstract

To achieve accurate damage localization for CFRP structures in various scenarios with limited sample sizes, this paper proposes a novel transfer learning strategy called the heterogeneous decision voting-based transfer domain adaptation method for damage localization (HDV-TDADL). This method can achieve high-precision localization without the need for model fine-tuning. Firstly, this paper presents an HDV method, which eliminates irrelevant Lamb wave signals through three different voting principles, extracts and reconstructs key sub-signals from the Lamb wave signals. Subsequently, a double transfer domain adaptation (DTDA) damage feature extraction method is introduced, which is based on the fusion of linear and nonlinear features to achieve adaptive damage mapping. The core of method proposed in this paper lies in finding shared damage diagnostic features between the source and target domains to address the domain mismatch issue, thereby reducing the domain shift phenomenon. Finally, an adaptive enhanced damage localization (ADEL) method is proposed, which effectively integrates multiple weak learners by adaptively adjusting feature weights, thereby constructing a stronger learner with better performance for precise damage localization. This paper designed twelve experimental scenarios covering a variety of damage transfer conditions and compared them with the current state-of-the-art methods. The experimental results demonstrate the significant advantages of the proposed method in terms of generalization ability and robustness.