基于神经网络的密集城市建筑群爆炸载荷快速预测

doi:10.12382/bgxb.2024.0987

摘要/Abstract

摘要：

城市建筑的交错布局提高了爆炸冲击波传播路径的复杂性,从而增加了对毁伤效应进行全面准确评估的难度。基于计算流体动力学的数值模拟方法虽然能够精确模拟爆炸产生的冲击载荷,但所需的计算量大且计算时间长。为实现在密集城市建筑群场景下爆炸冲击波载荷的快速预测,提出基于神经网络的快速预测方法。分析训练样本数量对模型预测精度的影响,研究区域划分对模型预测性能的作用。为满足神经网络模型训练的数据需求,使用爆炸模拟软件在一个典型密集城市建筑群场景下进行网格敏感性分析,在兼顾模拟速度和精度情况下产生了80组爆炸情况的数据;为确定合适的模型结构,构建不同层数的全连接神经网络进行对比实验与分析;通过对比实验,分析训练样本数量、区域划分以及构建双模型对模型预测精度的影响。实验结果表明:新方法在训练数据以外未见过的16组测试数据上预测误差达到10%以内,推理时间只需2s,为实现城市建筑群爆炸载荷快速预测提供了一种新的途径和视角。

关键词: 复杂城市环境, 爆炸载荷, 神经网络, 快速预测

Abstract:

The staggered layout of urban building complex makes the propagation path of explosion shock wave more complicated,thereby increasing the difficulty of evaluating the damage effect comprehensively and accurately.Numerical simulation methods based on computational fluid dynamics can accurately simulate the blast loading,but the calculated amount is large and the calculation time is long.In order to rapidly predict the blast loading in urban building complex,a fast blast loading prediction method based on neural network is proposed.The influence of the number of training samples on the prediction accuracy and the effect of regional division on the prediction performance of the model are analyzed.In order to meet the data requirements for training the neural network model,the explosion simulation software is used to analyze the mesh sensitivity in a typical dense urban building complex and generate a dataset for 80 sets of explosion scenarios while considering simulation speed and accuracy.In order to determine the appropriate model structure,the fully connected neural networks with different numbers of layers are constructed for comparative experiment and analysis.The effects of the number of training samples,the division of region and the construction of dual models on the prediction accuracy of the model are analyzed through comparative experiments.The results show that the prediction error of the proposed method is less than 10% on 16 sets of test data except the training data,and the inference time only takes 2 seconds.The proposed method has a balanced and good prediction ability for various ranges of peak overpressure,and provides a new approach and perspective for realizing the rapid prediction of blast loading in urban building complex.

Key words: complex urban environment, blast loading, neural network, rapid prediction

黄沛吉, 彭卫文, 冷春江, 张情, 钟巍. 基于神经网络的密集城市建筑群爆炸载荷快速预测[J]. 兵工学报, 2025, 46(8): 240987-.

HUANG Peiji, PENG Weiwen, LENG Chunjiang, ZHANG Qing, ZHONG Wei. Rapid Prediction of Blast Loading in Dense Urban Building Complex Based on Neural Networks[J]. Acta Armamentarii, 2025, 46(8): 240987-.

图/表 22

图1 密集城市建筑群研究场景[21]

Fig.1 Typical complex urban environment[21]

图2 密集城市建筑群模型[22]

Fig.2 Model of dense urban building complex[22]

表1 3个兴趣测点的位置

Table 1 Coordinates of three measuring points of interest m

测点编号	x	y	z
T1	0.30	1.1	0.105
T11	-0.30	1.0	0.075
T12	1.26	1.0	0.075

表2 经验常数

Table 2 Values of empirical constants

A/kPa	B/kPa	R₁	R₂
3.737×10⁸	3.737×10⁶	4.15	0.9
λ	E/(kJ·m^-3)	ρ₀/(kg·m^-3)
0.35	6.0×10⁶	1 630

图3 各网格尺寸下模拟波形和实验波形对比

Fig.3 Comparison of simulated and experimental waveforms for different mesh sizes

图4 各网格尺寸下模拟5ms所需计算时间

Fig.4 Calculation time for different mesh size

图5 爆炸模拟软件数值仿真压力云图

Fig.5 Numerically simulated pressure nephogram of explosion simulation software

图6 爆源和测点位置

Fig.6 Coordinates of explosion sources and measuring points

表3 爆源尺寸和位置选项

Table 3 Options for coordinates of explosion sources and measuring points

参数	数值
TNT爆源尺寸/g	4,8,12,16,20,24,28,32
爆源位置/m	爆源1(0.478,0.35),爆源2(0.678,0.35),爆源3(0.878,0.35),爆源4(0.478,0.75),爆源5 (0.278,0.75),爆源6 (0.078,0.75),爆源7(0.180,0.10),爆源8(0.800,1.10),爆源9(-0.150,1.10),爆源10(-0.150,0.30)

图7 全连接神经网络模型结构

Fig.7 Artificial neural networks structure

图8 数据集划分和5折交叉验证

Fig.8 Dataset partitioning and 5-fold cross validation

表4 模型参数选项总结

Table 4 Summary of model parameter options

参数	数值
隐藏层层数	[2,3,4,5]
隐藏层神经元个数	[500,1000]
激活函数	ReLU(输出层为Linear)
损失函数	MAE
优化器	AdaGrad
学习率	0.1 (每训练200次,衰减1/2)
训练迭代次数	1000
批量大小	100
Dropout率	0.3

表5 不同隐藏层架构的评估指标结果对比

Table 5 Comparison of evaluation metrics for different hidden layer architectures

隐藏层结构	训练 p_MAE/kPa	验证 p_MAE/kPa	验证 e/%	验证 R²
500-500	781.408	767.054	49.04	0.001 9
500-1000-500	672.423	683.698	27.09	0.0910
500-1000-1000-500	12.722	33.215	19.11	0.9977
500-1000-1000-1000-500	11.623	27.682	15.62	0.9996

表6 超压峰值是否取对数处理模型效果对比

Table 6 Comparison of evaluation metrics for logarithmic processing of peak overpressure

是否对超压峰值取对数	训练 p_MAE/kPa	验证 p_MAE/kPa	验证 e/%	验证 R²
否	12.722	33.215	19.11	0.9977
是	371.964	413.306	7.10	0.5387

图9 是否对数处理模型预测效果对比

Fig.9 Comparison of model effects with logarithmic processing

表7 单个算例不同训练测点数量模型预测效果对比

Table 7 Comparison of evaluation metrics for different numbers of training points in an example

单个算例训练测点数量	训练 p_MAE/kPa	验证 p_MAE/kPa	验证 e/%	验证 R²
300	739.596	746.616	12.79	0.0046
500	760.850	731.538	10.61	0.0100
800	647.787	696.221	8.95	0.0451
1000	644.602	638.197	8.18	0.1647
1587	371.964	413.306	7.10	0.5387

表8 禁区划分对模型预测效果的影响

Table 8 The impact of exclusion zone division on model prediction effect

训练数据集	验证数据集	p_MAE/kPa	e/%	R²
全部测点	全部测点	413.306	7.10	0.5387
	禁区外测点	3.638	6.95	0.9897
	禁区内测点	4746.238	8.77	0.5373
禁区外测点	禁区外测点	4.530	7.82	0.9803
禁区内测点	禁区内测点	2846.708	9.92	0.8213

图10 禁区划分与测点超压峰值数量级分布

Fig.10 Exclusion zone and distribution of peak overpressure

图11 模型测试预测效果

Fig.11 Model test prediction effect

表9 测试性能评估指标结果

Table 9 Test performance evaluation metrics results

训练测点数据	测试测点数据	p_MAE/kPa	e/%	R²
禁区外	禁区外	4.101	4.14	0.9842
禁区内	禁区内	1498.396	8.02	0.9450
全部测点数据	全部测点数据	154.166	4.53	0.9452

表10 模型在不同超压峰值幅度范围内的误差分布

Table 10 Error distributions in different overpressure ranges

超压峰值幅度范围/kPa	样本数量	位于相应误差e范围的样本百分比/%
超压峰值幅度范围/kPa	样本数量	[0,5)	[5,10)	[10,30)	[30,∞)	[0,10]
[0,25)	9 724	75.85	18.05	5.70	0.40	93.90
[25,50)	5647	78.57	15.99	4.82	0.62	94.56
[50,100)	3 671	76.14	16.44	6.59	0.63	92.58
[100,200)	2 157	66.11	23.55	9.36	0.97	89.66
[200,∞)	4 193	48.82	29.12	19.94	2.12	77.94

图12 模型在不同超压范围内的误差分布

Fig.12 Error distributions in different overpressure ranges

参考文献 27

[1]	RIGBY S E, LODGE T J, ALOTAIBI S, et al. Preliminary yield estimation of the 2020 Beirut explosion using video footage from social media[J]. Shock Waves, 2020, 30(6):671-675.
[2]	曹涛, 孙浩, 周游, 等. 近地爆炸冲击波传播特性数值模拟与应用[J]. 兵器装备工程学报, 2020, 41(12):187-191.
	CAO T, SUN H, ZHOU Y, et al. Numerical simulation and application of propagation characteristics of shock wave near ground explosion[J]. Journal of Ordnance Equipment Engineering, 2020, 41(12):187-191. (in Chinese)
[3]	张云峰, 陈博, 魏欣, 等. 空气自由场爆炸冲击波数值建模及应用[J]. 爆炸与冲击, 2023, 43(11):114202.
	ZHANG Y F, CHEN B, WEI X, et al. Numerical modeling and application of shock wave of free-field air explosion[J]. Explosion and Shock Waves, 2023, 43(11):114202. (in Chinese)
[4]	张晓颖, 李胜杰, 李志强. 爆炸载荷作用下夹层玻璃动态响应的数值模拟[J]. 兵工学报, 2018, 39(7):1379-1388. doi: 10.3969/j.issn.1000-1093.2018.07.016
	ZHANG X Y, LI S J, LI Z Q. Numerical simulation of dynamic response of laminated glass subjected to blast load[J]. Acta Armamentarii, 2018, 39(7):1379-1388. (in Chinese)
[5]	赵海涛, 王成. 空中爆炸问题的高精度数值模拟研究[J]. 兵工学报, 2013, 34(12):1536-1546. doi: 10.3969/j.issn.1000-1093.2013.12.008
	ZHAO H T, WANG C. High resolution numerical simulation of air explosion[J]. Acta Armamentarii, 2013, 34(12):1536-1546. (in Chinese)
[6]	SHI Y C, LIU S Z, LI Z X, et al. Review on quick safety assessment of building structures in complex urban environment after extreme explosion events[J]. International Journal of Protective Structures, 2023, 14(3):438-458.
[7]	REMENNIKOV A M, MENDIS P A. Prediction of airblast loads in complex environments using artificial neural networks[J]. WIT Transactions on The Built Environment, 2006, 87:269-278.
[8]	FLOOD I, BEWICK B T, RAUCH E. Rapid simulation of blast wave propagation in built environments using coarse-grain simulation[J]. International Journal of Protective Structures, 2012, 3(4):431-448.
[9]	PANNELL J J, RIGBY S E, PANOUTSOS G. Physics-informed regularisation procedure in neural networks:an application in blast protection engineering[J]. International Journal of Protective Structures, 2022, 13(3):555-578.
[10]	PANNELL J J, RIGBY S E, PANOUTSOS G. Application of transfer learning for the prediction of blast impulse[J]. International Journal of Protective Structures, 2023, 14(2):242-262.
[11]	DENNIS A A, RIGBY S E. The direction-encoded neural network:a machine learning approach to rapidly predict blast loading in obstructed environments[J]. International Journal of Protective Structures, 2024, 15(3):455-483.
[12]	LI Q L, WANG Y, LI L, et al. Prediction of BLEVE loads on structures using machine learning and CFD[J]. Process Safety and Environmental Protection, 2023, 171:914-925.
[13]	HUANG Y, ZHU S J, CHEN S W. Deep learning-driven super-resolution reconstruction of two-dimensional explosion pressure fields[J]. Journal of Building Engineering, 2023, 78:107620.
[14]	杨森. 城市建筑群爆炸荷载预测及灾害效应快速评估[D]. 天津: 天津大学, 2021.
	YANG S. Prediction of blast loads and rapid evaluation of disaster effects in complex urban environments[D]. Tianjin: Tianjin University, 2021. (in Chinese)
[15]	陈梓薇, 王仲琦, 曾令辉. 基于BP神经网络的爆炸用激波管峰值压力预测方法[J]. 爆炸与冲击, 2024, 44(5):054101.
	CHEN Z W, WANG Z Q, ZENG L H. A method for predicting peak pressure in an explosion shock tube based on BP neural network[J]. Explosion and Shock Waves, 2024, 44(5):054101. (in Chinese)
[16]	LI Q L, LI L, SHAO Y D, et al. A multi-task machine learning approach for data efficient prediction of blast loading[J]. Engineering Structures, 2025, 326:119577.
[17]	SHI J, LI J, TAM W C, et al. Physics_GNN:Towards physics-informed graph neural network for the real-time simulation of obstructed gas explosion[J]. Reliability Engineering & System Safety, 2025, 256:110777.
[18]	徐永康, 薛琨. 基于人工神经网络算法的多相云雾爆轰毁伤效应预测模型[J]. 兵工学报, 2024, 45(6):1889-1905. doi: 10.12382/bgxb.2023.0094
	XU Y K, XUE K. Artificial neural network-based prediction model for damage effect of fuel-air explosive[J]. Acta Armamentarii, 2024, 45(6):1889-1905. (in Chinese) doi: 10.12382/bgxb.2023.0094
[19]	SHENG K K, LI Y, XU X, et al. Stator and rotor temperature prediction of permanent magnet synchronous motor based on BP neural network[C]// Proceedings of the 2024 43rd Chinese Control Conference. Kunming,China: IEEE, 2024:8832-8837.
[20]	ZUO K J, YE Z Y, BU S H, et al. Fast simulation of airfoil flow field via deep neural network[J]. Aerospace Science and Technology, 2024, 150:109207.
[21]	BRITTLE M A. Blast propagation in a geometrically complex environment[D]. Cranfield,UK: Cranfield University, 2004.
[22]	VALGER S A, FEDOROVA N N, FEDOROVA A V. Numerical simulation of blast action on civil structures in urban environment[J]. IOP Conference Series:Materials Science and Engineering. IOP Publishing, 2017, 245(6):062018.
[23]	HEYLMUN J, VONK P, BREWER T. Blastfoam theory and user guide[Z/OL]. Brazos Street Austin,TX,US: Synthetik Applied Technologies, 2019(2019-10-30). www.blastfoam.org/blastFoam_User_Guide.pdf.
[24]	JASAK H. OpenFOAM:open source CFD in research and industry[J]. International Journal of Naval Architecture and Ocean Engineering, 2009, 1(2):89-94.
[25]	LI J, HAO H. Numerical and analytical prediction of pressure and impulse from vented gas explosion in large cylindrical tanks[J]. Process Safety and Environmental Protection, 2019, 127:226-244.
[26]	REMENNIKOV A, GAN E C J, TAN S S, et al. Methodology for predicting explosion risk around underground coal mine openings towards developing exclusion zones[C]// Proceedings of the 2022 Resource Operators Conference. Wollongong,Australia: University of Wollongong Press, 2022.
[27]	DENNIS A A, PANNELL J J, SMYL D J, et al. Prediction of blast loading in an internal environment using artificial neural networks[J]. International Journal of Protective Structures, 2021, 12(3):287-314.