Abstract: To address the problem that existing bridge dynamic displacement reconstruction methods based on physical information neural network (PINN) require huge response data of target points, an unsupervised reconstruction method based on PINN is proposed. Firstly, a PINN for dynamic displacement response reconstruction is constructed based on structural frequency response function and RNN. Secondly, the adaptive recurrent neural network (AdaRNN) is constructed based on the temporal drift theory, and the unsupervised response reconstruction network is constructed by combining PINN. Then, the temporal similarity of strain data is quantified and pre-training method is used to learn the changing characteristics of strain data. Finally, PINN network is used to estimate the displacement response training values of target points and non-target points, which are inputted into AdaRNN network for training, and the displacement response reconstructed by PINN is modified through pre-training learning results. Numerical examples and laboratory model experiments verify that the proposed method can reconstruct the bridge dynamic response with high accuracy in the absence of response at target points in network training. The effectiveness of the proposed method, and the effects of sensor number, position and measurement noise on the reconstruction accuracy are systematically discussed.