Impact response of transparent ceramic sandwich structures and its prediction by BP neural network
-
摘要: 首先,选择以蓝宝石陶瓷为迎弹层、二氧化硅无机玻璃和聚碳酸酯有机玻璃为吸能层和聚氨酯为胶结材料的透明夹层结构为研究对象,采用一级轻气炮对试样进行了冲击实验,试样呈现陶瓷层弯曲失效破坏主导和冲击压缩破坏主导的2种破坏模式,通过高速摄像详细记录裂纹动态扩展过程。然后,采用Abaqus 有限元软件对多组结构层厚度配比的透明夹层结构进行120、150、180 m/s 速度的弹体冲击模拟,针对陶瓷材料引入了基于JH-2 本构模型的子程序,并结合单元删除法,对裂纹扩展和碎片飞溅过程进行了数值模拟。模拟结果与实验结果吻合较好。最后,采用BP 神经网络算法对冲击点后侧位移峰值进行了预测,单层和多层神经网络模型平均计算耗时分别为1和3 min,与位移峰值的数值模拟结果相比,2种神经网络模型预测结果的平均相对误差分别为7.6% 和3.2%。该BP 神经网络模型计算时效和精度都满足要求,相比传统消耗5 h的有限元计算,节省大量时间,可对透明夹层结构的设计提供指导。Abstract: Transparent sandwich structures can combine the advantages of various materials, thus avoiding secondary damage caused by brittle material fragments. Therefore, they are widely used in various impact protection fields. However, the impact resistance of the structure is influenced by different material thicknesses in a complex manner, and there is a lack of quick and convenient design guidance. Artificial neural network has good applicability to the nonlinear problems of multi-material structures, and provides a novel approach to structural design. In this study, a transparent sandwich structure consisting of sapphire (Al2O3) ceramic as the impact-absorbing layer, silica inorganic glass, polycarbonate plexiglass as the energy absorption layers, and polyurethane as the bonding material was selected as the research subject. Impact experiments were conducted on samples using a first-stage light-gas gun. Two failure modes were observed in the samples: ceramic layer bending failure and impact compression failure. The dynamic crack propagation process was meticulously captured using high-speed cameras. Subsequently, Abaqus finite element software was employed to simulate projectile impact on transparent sandwich structures with varying layers thickness ratios at 120, 150, and 180 m/s. For ceramic materials, a subroutine based on the JH-2 constitutive model was introduced. The numerical simulation of crack propagation and debris splashing process was performed using the element deletion method. The simulation results exhibited good agreement with the experimental results. Finally, the BP neural network algorithm was utilized to predict peak displacement behind the impact point. The average calculation time for single-layer and multi-layer neural network models was 1 minute and 3 minutes, respectively. Compared with the numerical simulation results of displacement peak, the average relative error of the predicted results of the two neural network models was 7.6% and 3.2%, respectively. The BP neural network model fulfills the requirement for calculation time and accuracy, saving a substantial amount of time compared to the traditional 5-hour finite element calculations. It can provide valuable guidance for the design and development of transparent sandwich structures.
-
近年来,对于云雾燃爆的实验研究主要集中在气-液两相混合物的起爆能量[1]、爆燃转爆轰及爆轰波结构[2-7]等方面。云雾燃爆涉及燃料的物化性能、液滴大小、云雾质量浓度、环境等多种影响因素,当前对密闭容器云雾等粒径条件下,不同质量浓度的云雾燃爆性能的研究还较少,尤其是乙醚云雾此方面的研究更少。乙醚自燃温度为160 ℃,闪点为-45 ℃,可在无火焰或火花的热表面环境下点燃。乙醚在气相状态下与空气混合的燃爆极限为(1.9~36.0)%(体积分数)。由于乙醚易于生产、成本低,同时具有高挥发、低闪点等特性,在军事、燃油替代等领域得到了广泛应用。本文中,通过自行研发的20 L二次脉冲气动多相爆炸测试系统和全散射粒径测量系统,对索特平均直径相同、质量浓度不同的乙醚云雾燃爆参数进行实验研究。
1. 实验装置
1.1 概述
多相爆炸测试系统包括:20 L球型爆炸罐体、双喷头二次脉冲气动雾化子系统、100 J无级可调火花放电点火子系统[8]、高速数据采集处理存储子系统[9-10]、时序触发控制中心。如图 1所示,20 L球型爆炸罐体内径为337 mm, 罐体包括一对透明光学检测窗口。时序触发控制中心可同时控制气动雾化、点火、数据采集、全散射粒径测量等子系统按实验要求,不同时间精确触发,精度级别1 ms。全散射粒径测量系统光路通过罐体透明光学检测窗口,并穿过球罐中心;激光发射单元光学波长分别为447、543、638 nm,功率为50 mW。
1.2 二次脉冲雾化系统
为实现瞬态大张角雾化效果,自行研制了双喷头二次脉冲气动雾化子系统。其组成包括:空气泵、电磁阀、高压气室、储液室、气液输送管段(内直径为20 mm)、球型腔体喷头(内直径为40mm,前半球表面上布置11圈直径为1.0~1.5 mm的小孔119个)、多孔均布空心小球(小球直径为25 mm; 开孔42个, 孔径为3 mm)。气动雾化喷腔结构如图 2所示。
由于在气液输送管段高速高压瞬态脉动喷射作用下,形成以环状流为主的流型结构,即气相在管道中心形成高速流动的气芯,液相以液膜形式沿管壁周围向前运动,如图 3所示。在喷雾腔室加装均布带孔小球,上述流型进入气液参混扩散区后(喷雾腔室),一部分气体进入小球形成高压气腔,并由小球前半部分小孔排出产生二次气动;另部分气体沿小球外部形成漩涡,加强气液掺混;最终由多孔小球排出的带压气体推动气液掺混区的液体沿喷头喷出, 形成大张角圆锥体结构的悬浮云雾。
1.3 粒径测量系统
为实时测量云雾粒径及质量浓度,自行研发了全散射粒径测量系统,并同步使用实时喷雾激光粒度仪HELOS-VARIO进行了实验数据比较测试,结果见表 1,表中n为测量次数。2套系统所测得特征直径D10、D50、D90和索特平均直径D32的误差分别为2.85%、4.65%、1.27%、0.24%。结果证明,研发的全散射粒径测量系统能够满足实验测量要求。
表 1 2种测量系统的测量结果Table 1. Measured results by two measuring systems测量系统 n 测量结果 D10/μm D50/μm D90/μm D32/μm 实时喷雾激光粒度仪 30 最大值 26.50 52.33 78.88 22.45 最小值 22.52 43.63 65.24 18.45 平均值 24.51 47.98 72.06 20.45 全散射粒径测量系统 30 最大值 28.68 55.63 79.89 23.12 最小值 21.74 44.81 66.07 17.88 平均值 25.21 50.22 72.98 20.50 2. 实验过程
实验工况为:初始环境温度, 21 ℃; 喷雾时长, 50 ms; 点火时刻, 100 ms; 点火能量, 40.32 J。首先, 通过调节气动压力和设计喷雾剂量,在实验环境恒定的条件下进行等粒径、不同质量浓度的实验数据收集;然后,通过统计获得一组等粒径、不同质量浓度的实验数据,进行云雾燃爆参数测试实验,即同步触发气动喷雾、高速摄像机及点火系统;最后,记录燃爆超压、温度、点火延迟等实验数据。
3. 结果与分析
3.1 粒径分布与质量浓度
索特平均直径为22.90 μm,由乙醚云雾不同质量浓度ρ数据统计结果(见表 2)可知:(1)在喷雾时长50 ms作用下,400~800 kPa不同气动压力pp雾化过程完成后100 ms,罐体内部压力pv仅上升2~4 kPa, 满足后期常压燃爆实验环境要求;(2)由实验测量发现,喷雾损耗剂量ml不可忽视,其占设计喷雾剂量md的15%~25%;(3)索特平均直径22.90 μm为不同质量浓度下实测粒径D32的平均值,其偏差小于5%。
表 2 乙醚云雾质量浓度和粒径实验数据Table 2. Experimental data of particle size and mass concentration of diethyl ether mistpp/MPa pv/MPa md/(g·m-3) ml/(g·m-3) 平均值 D32/μm ρ/(g·m-3) 0.80 0.104 713.40 142.68 23.16 570.72 0.75 0.104 642.06 96.31 23.30 545.75 0.70 0.103 570.72 114.14 22.07 456.58 0.60 0.103 535.05 117.71 22.10 417.34 0.50 0.103 499.38 124.85 21.73 374.54 0.50 0.103 428.04 107.01 23.72 321.03 0.50 0.103 356.70 78.47 21.52 278.23 0.50 0.103 285.36 57.07 23.35 228.29 0.50 0.103 214.02 53.51 22.90 160.52 0.45 0.102 142.68 35.67 24.04 107.01 0.45 0.102 107.01 26.75 23.55 80.26 0.40 0.102 71.34 14.27 23.36 57.07 3.2 燃爆参数
在点火能为40.32 J、乙醚云雾索特平均直径为22.90 μm的条件下,由图 4可知,乙醚云雾燃爆极限范围为80.26~417.34 g/m3。
乙醚气相燃爆极限浓度范围在1.9%~36%(体积分数)[11]。在本实验研究中,按标准状态气相换算云雾体积浓度为2.42%~12.60%(体积分数)。可见实验获得的云雾燃爆下限80.26 g/m3(2.42%)比文献中乙醚纯气相燃爆下限1.9%高;同时,云雾燃爆上限417.34 g/m3(12.60%)比文献中乙醚气相燃爆上限36%低,因此可以得出:(1)在乙醚云雾索特平均直径为22.90 μm的条件下,云雾浓度对乙醚云雾燃爆上、下限有较大影响;(2)燃爆上限的下降是由于随云雾液滴浓度的增大,液滴群吸热汽化并燃烧所需的能量逐渐增加,发生了淬火现象所致。
乙醚气相与空气混合(乙醚当量体积分数Cst≈3.39%)化学反应方程如下:
C2H5OC2H5+6(O2+7921N2)=4H2O+5CO2+6×7921N2 理论上,上述方程说明乙醚气相与空气混合燃爆超压峰值p的最大值应出现在乙醚当量体积分数为3.99%[12-13]的位置。而实验获得最大超压出现在乙醚云雾质量浓度为278.23 g/m3(6.89%(体积分数))时,这说明由于云雾液相颗粒群主要以扩散燃烧模式并促使了对应于最大燃爆超压的乙醚云雾质量浓度推迟来临。
气相可燃物与空气混合燃爆超压和温度θ变化趋势同步[14], 但在本实验中,最大火焰扩散温度出现在乙醚云雾质量浓度为228.29 g/m3时;而最大燃爆超压出现在乙醚云雾质量浓度为278.23g/m3时;这表明乙醚云雾液相颗粒群蒸发扩散燃烧过程对温度、压力同步发生了影响,即最大燃爆温度对应的乙醚云雾质量浓度低于最大燃爆超压对应的乙醚云雾质量浓度。
3.3 点火延迟时间
对于云雾场燃爆通常由点火时刻到火焰传播发生,存在不同的点火延迟时间ti, d,如图 5所示云雾点火超压过程示意图。这是由于:(1)放电火花点火过程发生的放热过程及液滴群吸热过程存在一定的时间;(2)不同物化性能的云雾也是影响点火延迟时间的重要因素;(3)对于不同浓度的云雾场,点火延迟时长也不尽相同。通过高速摄像机以10 000 fps的速度拍摄不同乙醚云雾点火延迟时间及火焰发展过程,如图 6所示,由不同质量浓度下点火延迟时间及趋势可知,在乙醚云雾质量浓度为228.29 g/m3时其点火延迟时长最短,约为15 ms;最长点火延迟时长26 ms发生在乙醚云雾燃爆上限417.34 g/m3。乙醚云雾与空气混合点火延迟时长在燃爆极限范围内呈U型分布。
如图 7所示,实验观测得到火焰由球罐下半部向整个罐体传播,分析可得:(1)在雾化完成后,液滴的沉降效应导致球罐下半部浓度随时间而升高;(2)双喷头二次脉冲气动雾化方法,减缓了气动对罐体内部湍流强度,从而使湍流效应弱于液滴群沉降效应。
4. 结论
建立了20 L二次脉冲气动喷雾多相爆炸测试系统和全散射粒径测量系统,通过不同气动压力与设计喷雾剂量的协调,获得了乙醚云雾在索特平均直径为22.90μm的条件下不同质量浓度实验数据,并在点火能为40.32 J、常温常压条件下进行了燃爆超压、温度及点火延迟时间等燃爆参数的实验研究,得到结论如下:
(1) 乙醚云雾与空气混合物燃爆质量浓度极限范围为80.26~417.34 g/m3。
(2) 乙醚云雾与空气混合物燃爆最大超压0.78 MPa, 出现在乙醚云雾质量浓度为278.34 g/m3时。乙醚云雾燃爆最大温度1 260 ℃,出现在乙醚云雾质量浓度为228.29 g/m3时。
(3) 乙醚云雾与空气混合物点火延迟时长在燃爆极限范围内呈U型分布,最短点火延迟时长15 ms发生在乙醚云雾质量浓度为228.29 g/m3时;最长点火延迟时长26 ms发生在乙醚云雾燃爆质量浓度上限417.34 g/m3时。
-
表 1 实验条件及结果
Table 1. Experimental conditions and the corresponding results
试样编号 弹体速度/(m·s−1) 弹体质量/g 厚度/mm 破坏模式 陶瓷层 无机玻璃层 有机玻璃层 1 164.79 16.7 8 6 4 弯曲失效 2 193.38 16.7 8 6 4 压缩失效 3 186.28 16.7 8 6 10 弯曲失效 4 183.84 16.7 8 8 4 压缩失效 5 183.18 16.7 6 4 10 压缩失效 6 184.22 16.7 6 3 5 压缩失效 表 2 模拟工况
Table 2. Simulated conditions
工况 厚度/mm 弹体速度/(m·s−1) 陶瓷 无机玻璃 有机玻璃 1 6 3 5 120, 150, 180 2 6 4 4 120, 150, 180 3 6 5 5 120, 150, 180 4 8 4 4 120, 150, 180 5 6 8 4 120, 150, 180 6 8 6 4 120, 150, 180 7 6 4 10 120, 150, 180 8 8 8 4 120, 150, 180 9 6 6 10 120, 150, 180 10 8 10 4 120, 150, 180 11 6 6 10 120, 150, 180 12 8 6 10 120, 150, 180 13 6 6 6 120, 150, 180 表 3 陶瓷JH-2本构参数
Table 3. Constitutive parameters of ceramic JH-2
A B C M N T/GPa G/GPa E/GPa K1/GPa K2/GPa 0.889 0.29 0.0045 0.53 0.764 0.2 120.34 295 184.560 185.870 K3/GPa σHEL/GPa pHEL/GPa D1 D2 β σ∗i,max σ∗f,max ρ/(kg·m−3) 157.540 6 3.268 0.005 1.0 1.0 0.2 1.0044 3700 表 4 材料参数
Table 4. Material parameters
材料 密度/(kg·m−3) 弹性模量/GPa 泊松比 钨钢弹体 7800 210 0.30 二氧化硅无机玻璃 2450 68 0.23 聚碳酸酯有机玻璃 1200 23 0.35 表 5 单层、双层神经网络模型计算效率的对比
Table 5. Comparison of computational efficiencies of single-layer and double-layer neural network models
计算方式 峰值位移/mm 峰值位移平均相对误差/% 平均计算时长 有限元模拟 0.41 5 h 单层BP神经网络 0.44 7.6 1 min 双层BP神经网络 0.42 3.2 3 min -
[1] RAMISETTY M, SASTRI S, KASHALIKAR U, et al. Transparent polycrystalline cubic spinels protect and defend [J]. American Ceramic Society Bulletin, 2013, 92(2): 20–25. DOI: 10.1007/s10971-012-2957-6. [2] MALLIK A, BASUMAJUMDAR A, KUNDU P, et al. Some studies on nucleation, crystallization, microstructure and mechanical properties of mica glass-ceramics in the system 0.2BaO·0.8K2O·4MgO·Al2O3·6SiO2·2MgF2 [J]. Ceramics International, 2013, 39(3): 2551–2559. DOI: 10.1016/j.ceramint.2012.09.015. [3] 王笛. 汽车风挡玻璃冲击破坏现象高精度仿真方法开发及应用研究[D]. 广州: 华南理工大学, 2020.WANG D. Development of high-precison numerical algorithms for the impact fracture analysis of automotive windshield glass and the related engineering applications [D]. Guangzhou, Guangdong, China: South China University of Technology, 2020. [4] CHEN S W, CHEN X, WU X Q. The mechanical behaviour of polyvinyl butyral at intermediate strain rates and different temperatures [J]. Construction and Building Materials, 2018, 182: 66–79. DOI: 10.1016/j.conbuildmat.2018.06.080. [5] GALUPPI L, ROYER-CARFAGNI G F. Effective thickness of laminated glass beams: new expression via a variational approach [J]. Engineering Structures, 2012, 38: 53–67. DOI: 10.1016/j.engstruct.2011.12.039. [6] 尹怡林. 夹层复合建筑玻璃抗低速冲击性能研究[D]. 哈尔滨: 哈尔滨工业大学, 2020. DOI: 10.27061/d.cnki.ghgdu.2020.006729.YIN Y L. Study on low-speed impact property of laminated glass for construction [D]. Harbin, Heilongjiang, China: Harbin Institute of Technology, 2020. DOI: 10.27061/d.cnki.ghgdu.2020.006729. [7] DENG G X, MA W, PENG Y, et al. Experimental study on laminated glass responses of high-speed trains subject to windblown sand particles loading [J]. Construction and Building Materials, 2021, 300: 124332. DOI: 10.1016/j.conbuildmat.2021.124332. [8] ZHANG X H, HAO H, WANG Z Q. Experimental study of laminated glass window responses under impulsive and blast loading [J]. International Journal of Impact Engineering, 2015, 78: 1–19. DOI: 10.1016/j.ijimpeng.2014.11.020. [9] ZHANG X H, HAO H. Experimental and numerical study of boundary and anchorage effect on laminated glass windows under blast loading [J]. Engineering Structures, 2015, 90: 96–116. DOI: 10.1016/j.engstruct.2015.02.022. [10] STRABBURGER E. Ballistic testing of transparent armour ceramics [J]. Journal of the European Ceramic Society, 2009, 29(2): 267–273. DOI: 10.1016/j.jeurceramsoc.2008.03.049. [11] GIOVANNA R, ZULLI F, ANDREOZZI L, et al. Test methods for the determination of interlayer properties in laminated glass [J]. Journal of Materials in Civil Engineering, 2017, 29(4): 04016268. DOI: 10.1061/(ASCE)MT.1943-5533.0001802. [12] STEINBERGER D, SANDFELD S. Machine learning-based classification of dislocation microstructures [J]. Frontiers in Materials, 2019, 6: 10. DOI: 10.3389/fmats.2019.00141. [13] BREIMAN L. Random Forests [J]. Machine Learning, 2001, 45(1): 5–32. DOI: 10.1023/A:1010933404324. [14] JAIN A K, MAO J, MOHIUDDIN K M. Artificial neural networks: a tutorial [J]. Computer, 1996, 29(3): 31–44. DOI: 10.1109/2.485891. [15] BASHEER I A, HAJMEER M. Artificial neural networks: fundamentals, computing, design, and application [J]. Journal of Microbiological Methods, 2000, 43(1): 3–31. DOI: 10.1016/S0167-7012(00)00201-3. [16] GUO K, YANG Z, YU C H, et al. Artificial intelligence and machine learning in design of mechanical materials [J]. Materials Horizons, 2021, 8(4): 1153–1172. DOI: 10.1039/D0MH01451F. [17] XU Y J, YOU T, DU C J. An integrated micromechanical model and BP neural network for predicting elastic modulus of 3-D multi-phase and multi-layer braided composite [J]. Composite Structures, 2015, 122: 308–315. DOI: 10.1016/j.compstruct.2014.11.052. [18] MAURO J C, TANDIA A, VARGHEESE K D, et al. Accelerating the design of functional glasses through modeling [J]. Chemistry of Materials , 2016, 28(12): 4267–4277. DOI: 10.1021/acs.chemmater.6b01054. [19] 刘源, 庞宝君. 基于贝叶斯正则化BP神经网络的铝平板超高速撞击损伤模式识别 [J]. 振动与冲击, 2016, 35(12): 22-27. DOI: 10.13465/j.cnki.jvs.2016.12.004.LIU Y, PANG B J. Hypervelocity impact damage pattern recognition on aluminum plates based on bayesian regularization BP neural network [J]. Journal of Vibration and Shock, 2016, 35(12): 22–27. DOI: 10.13465/j.cnki.jvs.2016.12.004. [20] 孟嫣然, 王星尔, 杨健, 等. 基于机器学习算法的夹层玻璃冲击破坏预测模型研究 [J]. 无机材料学报, 2021, 36(1): 61–68. DOI: 10.15541/jim20200187.MENG Y R, WANG X E, YANG J, et al. Research on machine learning based model for predicting the impact status of laminated glass [J]. Journal of Inorganic Materials, 2021, 36(1): 61–68. DOI: 10.15541/jim20200187. [21] CHEN J, XU J, YAO X, et al. Experimental investigation on the radial and circular crack propagation of PVB laminated glass subject to dynamic out-of-plane loading [J]. Engineering Fracture Mechanics, 2013, 112/113: 26–40. DOI: 10.1016/j.engfracmech.2013.09.010. [22] COBLE R L. Transparent alumina and method of preparation: US 3026210-A [P]. 1962-03-20. [23] MORGAN W L. Manufacture and characteristics of laminated glass [J]. Industrial and Engineering Chemistry, 1931, 23(5): 505–508. DOI: 10.1021/ie50257a008. [24] 杨震琦, 庞宝君, 王立闻, 等. JH-2模型及其在Al2O3陶瓷低速撞击数值模拟中的应用 [J]. 爆炸与冲击, 2010, 30(5): 463–471. DOI: 10.11883/1001-1455(2010)05-0463-09.YANG Z Q, PANG B J, WANG L W, et al. JH-2 model and its application to numerical simulation on Al2O3 ceramic under low-velocity impact [J]. Explosion and Shock Waves, 2010, 30(5): 463–471. DOI: 10.11883/1001-1455(2010)05-0463-09. [25] VOGL T P, MANGIS J K, RIGLER A K, et al. Accelerating the convergence of the back-propagation method [J]. Biological Cybernetics, 1988, 59(4/5): 257–263. DOI: 10.1007/BF00332914. -