Reconstruction Method of Incomplete Shock Wave Signals Based on Deep Learning

The parameters of explosion shock waves are one of the main criteria for evaluating the power of ammunition. However, during the actual testing process, the testing system may be damaged by fragments or other factors, making it unable to capture the complete signal and thus affecting subsequent damage assessment. Therefore, this article proposed a method based on bidirectional long short-term memory network（BiLSTM） and multi head self-attention module fusion to construct the integrity of incomplete shockwave signals. BiLSTM was used to analyze the local temporal dependencies of shockwave signals, and the multi head self-attention module captured frequency information in the signal. Finally, the fusion of temporal signals and frequency information was achieved, resulting in a complete shockwave signal. In the process of information collection, the measured signal data is usually only dozens of sets, which leads to the problem of small sample size. This article established a GAN network with LSTM units as the generator to expand the complete shock wave signal and enhance the dataset capacity. The experimental results based on the expanded dataset show that the MSE and MAE between the complete signal constructed by the method proposed in this paper and the original signal are 0. 006 8 and 0. 146 2, respectively, which are superior to the LSTM, BiLSTM, and CNN+BiLSTM methods. The method proposed in this paper meets the practical requirements for constructing incomplete shock wave signals.