To monitor the health status of highway bridges and ensure the safety of vehicle driving decks, with the combination of the bridge deflection data monitored by millimeter wave radar and deep learning theory, a bridge deflection prediction model based on the combination of convolutional neural network (CNN) and gate recurrent unit (GRU) is proposed. Firstly, the high-precision deflection data of the expressway bridge was monitored. Through data preprocessing, some noise data was repaired on the basis of retaining the original data characteristics. Secondly, the processed sample data, time step and three-dimensional data of feature number were taken as the input matrix, and the input matrix constructed by the bridge deflection data sequence was taken as the input layer. After passing through the dense connection layer of the CNN-GRU combined model, the predicted bridge deflection values became output. Finally, the root mean square error (RMSE), mean absolute error(MAE) and mean absolute percentage error(MAPE) were used to verify the prediction effect. The results show that the CNN-GRU model has higher accuracy: compared with the traditional long short-term memory (LSTM) model, RMSE is improved by 59.65 %, and MAE is improved by 61.30 %. Compared with the CNN-LSTM model, RMSE and MAE are increased by 2.48 % and 4.87 %. Its judgment on the extreme value and trend of bridge deflection is basically accurate and can be used as a bridge health.