<p>Traditional approaches of damage localization such as data-driven approaches, model-driven approaches and hybrid approaches were frequently applied in structural health monitoring. However, these approaches usually have the problems of low efficiency owing to a large number of devices and information need to be preprocessed. To enhance the efficiency and accuracy, researchers introduce deep learning techniques in structure health monitoring, while for time series data, existing networks still have the disadvantages of not fully extracting local features and low efficiency. In view of this, this study presents a deep learning method integrating multi-scale 1D Convolutional Neural Network (MSCNN), Transformer and Bidirectional Gating Recurrent Network (BiGRU) to localize structural damage. First of all, local features of structural temporal signals were extracted through MSCNN. Then efficient parallel computing and long sequence information extraction were achieved through the combination of Transformer and BiGRU. Finally, the damage localization performances of the proposed algorithm with different combinations of MSCNN, BiGRU, and Transformer were compared by locating the damage of a steel frame structure. It is verified that the proposed hybrid model can locate structural damage with the characteristics of multi-scale temporal feature extraction, high precision and efficiency.</p>