Structural damage identification method based on transfer learning and heterogeneous data alignment

Abstract. Structural damage identification is crucial for ensuring building safety, as it helps safeguard both lives and property. Recently, deep learning has become a prominent approach for damage detection. However, effective training of deep learning models requires abundant structural response data with accurate damage labels, which are often scarce in practical engineering applications. While existing models based on simulated data perform well in simulation environments, they typically struggle to achieve satisfactory performance on real-world measured data. To address this issue, this paper proposes a novel method for structural damage identification that combines transfer learning with the alignment of heterogeneous data (simulated and measured data). This method enables the transfer of damage detection capabilities from simulated data to measured data. First, finite element software is used to simulate various structural damage scenarios, generating a large volume of simulated data with precise damage labels. This data is then used to pre-train an initial model for damage identification. Next, transfer learning is applied, where the model is further trained using a combination of a small amount of measured data and the large simulated dataset. To address the distributional differences between the heterogeneous data sources, the Jensen-Shannon (JS) divergence is employed to quantify the discrepancy in the high-dimensional feature space. This is combined with the cross-entropy classification loss to form a composite loss function for model training. The resulting model is capable of accurately identifying structural damage in measured data. Experimental results show that this approach improves damage identification accuracy by 6% compared to traditional methods, demonstrating its effectiveness.