The time span of the civil engineering structural damage state usually accounts for a small part of the total life cycle. In order to solve the problem that traditional data-driven structural damage identification methods lack enough damage state data for training, a structural damage identification method based on digital twin and deep learning is proposed in this paper for practical application. Firstly, the digital twin is constructed by using the numerical simulation model and online monitoring data to obtain the "big data" of the structural dynamic response under different damage conditions. Secondly, in order to get rid of the dependence on the external excitation, the empirical mode decomposition method and transmissibility function are used to preprocess the obtained data. Then, the damage identification is realized by using deep learning. To verify the effectiveness of this method, untrained monitoring data of structures are analyzed. The results show that the method has good generalization ability and can identify the structural damage condition effectively. The problem of data hunger is solved by digital twin technology, and the deep neural network trained by the intrinsic mode vibration transmissibility function data sets can still maintain a high accuracy of damage identification without any seismic information. The combination of the two methods can make structural health monitoring more active, reliable and efficient.