In the field of bridge structure condition monitoring, sensor systems often have abnormal or missing monitoring data due to environmental interference or equipment abnormalities. This situation may affect the accuracy of structural safety assessment. In this study, combined with deep learning, aiming at the problem of monitoring data completion, a prediction method based on BILSTM is proposed, and a CNN-BILSTM combined model is constructed for acceleration signal prediction. The prediction effects of the three algorithms, LSTM, BILSTM and CNN-BILSTM, were compared horizontally. Through the analysis of the MAX-MAE-RMSE-R2 index system and the fit degree of the prediction curve, it was confirmed that the combined model has better prediction performance. The experiment adopted the monitoring data of the accelerometer of the Hardanger Bridge in Norway as the training test set. Taking the H1E measurement point data as an example, the MAE index of CNN-BILSTM is 4.2% lower than that of BILSTM, and it has achieved a significant improvement of 20.0% compared with LSTM, verifying the accurate prediction ability of this model for real bridge acceleration data. For the concurrent failure scenarios of multiple sensors, a method for reconstructing the time series of missing signals based on the data correlation of adjacent sensor nodes is proposed. The remaining channel acceleration data were jointly trained through the CNN-BILSTM model. By comparing the prediction performance of different algorithms, it was found that this combined model could still maintain a high prediction accuracy in the case of the absence of multi-sensor data. Its error index was significantly better than that of the single LSTM and BILSTM models, showing good engineering applicability.