基于串并行双分支网络的冲击波信号重构方法

Shockwave signal reconstruction method based on a serial-parallel double branch network

通过有限测点数据重建冲击波场内压力分布、通过残缺数据重构完整的冲击波压力曲线,对武器威力以及目标毁伤评估具有重要价值。针对爆炸冲击波信号重构问题,建立Res-GRU分支以串行方式捕捉冲击波超压信号局部时序依赖关系;建立Transformer分支以并行方式分析信号全局潜在特征;建立特征融合单元进行高阶特征融合,实现不同阶段信息逐层互补;进而构建了基于门控循环单元(gated recurrent unit,GRU)和Transformer模型的串并行双分支网络(G-TNet)。试验研究表明:G-TNet综合考量了信号的时序关系、数据变化规律等特征信息;在基于有限测点数据的冲击波场压力分布重构试验中,重建的模拟、实测超压数据与原始值之间均方误差(mean square error,MSE)分别为5.0×10^(-6)、1.2×10-3,平均峰值误差分别为0.49%、27.01%,平均正压时间误差分别为15.62%、15.91%,平均比冲量误差分别为17.66%、19.33%;在基于残缺数据的冲击波压力曲线重构试验中,重构的模拟、实测信号的缺失值与原始值之间MSE分别为5.0×10^(-6)和5.0×10^(-4),平均绝对误差(mean absolute error,MAE)分别为0.0010和0.0171;G-TNet重构结果优于主流方法,满足爆炸冲击波压力重构指标要求。

Reconstructing the pressure distribution in the shock wave field by using limited measurement point data and the complete shock wave pressure curve by using missing data are of great value for weapon power and target damage assessment.To address the problem of reconstruction of the shockwave signal,it was proposed to establish a Res-GRU branch to capture the local timing dependence of the shockwave overpressure signal in a serial manner;to establish a Transformer branch to analyze the global potential features of the signal in a parallel manner;and to establish a feature merging unit for higher-order feature integration to realize the complementary information of different stages layer by layer.Then,a serial-parallel double branch network(denoted as G-TNet)based on the gated recurrent unit(GRU)and Transformer model was constructed.The experimental study shows that the G-TNet integrates the signal timing relationship,data variation pattern and other characteristic information.In the reconstruction experiment of pressure distribution of shockwave field based on limited measurement point data,the mean square errors(MSEs)between the reconstructed simulated and measured overpressure data and the original value were 5.0×10^(-6) and 1.2×10-3,the average peak errors were 0.49% and 27.01%,the average positive pressure time errors were 15.62% and 15.91%,and the average specific impulse errors were 17.66%and 19.33%,respectively.In the reconstruction experiment of the shockwave pressure curve based on missing data,the MSEs between the reconstructed simulated and measured signals and the original values were 5.0×10^(-6) and 5.0×10^(-4),and the mean absolute errors(MAE)were 0.0010 and 0.0171,respectively.The G-TNet reconstruction results are better than the mainstream methods and fulfill the requirements of explosion shockwave pressure reconstruction index.