Abstract: In view of the nonlinear and time-series characteristics of structural dynamic response data, a convolutional neural network (CNN) and a bidirectional long short-term memory (BiLSTM) combined structural nonlinear model updating method are proposed based on attention mechanism (AM). Firstly, the one-dimensional convolutional layer is employed in CNN for the nonlinear local feature extraction, and subsequently compressed by the maximum pooling operation. Then, time-series characteristics of the reduced-dimensional information conveyed by a CNN layer is extracted by forward and backward chain-connected LSTM units in a BiLSTM layer. Meanwhile, the AM mechanism is introduced to distinguish the importance degree of time information extracted from the BiLSTM hidden layer by weightings and mining deep time-series characteristics of dynamic response data. Finally, the outputs of the attention layer are adopted as the inputs of the fully-connection layer to obtain the final model prediction results. Numerical simulation on a bridge tower structure subjected to seismic loading is conducted. Furthermore, the shake table test of an actual scaled bridge tower is applied to verify the accuracy of the method proposed. Both numerical and experimental results have shown that the hybrid neural network method proposed is applicable to update the nonlinear model with high-dimensional spatial parameters. Meanwhile, comparing with conventional neural network approaches, this methodology exhibits stronger robustness and network generalization ability.