Deformation of the bridge is an important index of bridge condition assessment. The direct measurement methods for the bridge deformation such as LVDT, RTK-GPS and LDV have a lot of limitations. Because it is difficult to obtain the real-time displacement modes and strain modes from the measured dynamic strains while the sampling frequency of strain measurement is low, the indirect bridge deformation reconstruction method based on strain still has some shortcomings in engineering practice. This paper proposed a data fusion method to reconstruct the deformation of bridge structure based on LSTM neural network using the measured strain and acceleration data. In this paper, the accuracy of the reconstruction results is verified by numerical simulation and experiment without strain modes, displacement modes and neutral axis. The results show that data fusion method based on LSTM can achieve high accuracy under different conditions.