Aiming at the reconstruction of explosion shock wave signal， the deep convolutional neural network （DCNN） was introduced to capture the local information and higher-order features of the shock wave signal， and the bi-directional long-term and short-term memory network （Bi-LSTM） was introduced to capture the time series dependence of shock wave overpressure data， and then the reconstruction model of explosion shock wave signal based on deep learning is constructed. The experimental results show that the reconstruction model of explosion shock wave signal constructed in this paper comprehensively considers the characteristic information of signal such as time sequence relationship， spectral characteristics and data variation law. In the pressure distribution reconstruction experiment of shock wave field based on finite measuring point data， the average errors of simulated and measured overpressure peaks are 3.53% and 13.71%， the average errors of positive pressure time are 7.35% and 14.26%， and the average errors of specific impulse are 4.02% and 11.92%， respectively. In the reconstruction experiment of shock wave pressure curve based on incomplete data， the missing values of simulated and measured signals are basically consistent with the original values， and the deviations are all around zero. All meet the requirements of explosion shock wave pressure reconstruction index. The research results have important guiding significance for signal reconstruction of explosion shock wave.