基于深度学习算法预测燃爆的临界淬熄直径

聂仲恒; 汪丽; 高伟; 姜海鹏

doi:10.11883/bzycj-2025-0218

基于深度学习算法预测燃爆的临界淬熄直径

doi: 10.11883/bzycj-2025-0218

聂仲恒^{1, 2,},
汪丽²,
高伟¹,
姜海鹏^1, ,

1.
大连理工大学化工学院，辽宁省大连市 116024
2.
中国航天科技集团有限公司航天推进技术研究院北京航天试验技术研究所，北京 100074

基金项目: 国家重点研发计划青年科学家项目（2024YFC3082600）；国家自然科学基金（52574230）

详细信息

作者简介:
聂仲恒（2000－　），男，硕士，nzh2490823579@163.com

通讯作者:
姜海鹏（1990－　），男，博士，副教授，jhp@dlut.edu.cn

中图分类号: X932
计量
- 文章访问数: 286
- HTML全文浏览量: 31
- PDF下载量: 61
- 被引次数: 0
出版历程
- 收稿日期: 2025-07-14
- 修回日期: 2026-03-28
- 网络出版日期: 2026-04-09

Prediction of critical quenching diameter for deflagration and detonation based on deep learning algorithms

NIE Zhongheng^{1, 2
,},
WANG Li²,
GAO Wei¹,
JIANG Haipeng^{1
, ,}

1.
Dalian University of Technology, Dalian, Liaoning Province, 116024, China
2.
Beijing Institute of Aerospace Testing Technology, Beijing 100074, China

摘要

摘要: 聚焦密闭管道可燃气体燃爆火焰淬熄的安全防护需求，基于燃气组分、管道几何、初始条件及多孔介质结构参数构建了九维特征空间，建立了多孔介质临界淬熄直径预测模型。为突破传统经验公式的精度不足问题与应用局限性，通过系统性的超参数优化与模型验证，深入对比并证实了Transformer架构在临界淬熄直径预测问题上的显著优越性：其预测性能（平均绝对误差δ_MAE=0.068，均方误差δ_MSE=0.008，相关系数R²=0.928）显著超越了广泛应用的卷积神经网络（convolutional neural network，CNN）模型（δ_MAE=0.079，δ_MSE=0.012，R²=0.906），不仅整体误差更低，且对离群点的鲁棒性更强。深入分析发现，Transformer模型的核心优势源于其自注意力机制对淬熄过程中关键临界特征的精准捕获与高效建模能力。在数据归一化敏感性验证中，Transformer模型展现出优异的鲁棒性，这归功于其层归一化机制所赋予的特征解耦与稳定表示能力。基于上述系统性评估，最终确立Transformer模型为预测多孔介质临界淬熄直径的最优模型，为燃爆安全防控策略的量化制定及管道阻火器安全性能的精细化设计提供了强大的、可操作的决策支持工具，具有重要的理论指导意义。
- 管道爆炸 /
- 深度学习 /
- 临界淬熄直径预测 /
- 阻火器
Abstract: The research addresses the safety imperative of preventing flammable gas explosions in enclosed pipelines by establishing a predictive model for the critical quenching diameter within porous media flame arresters. A novel predictive framework was developed based on a comprehensive nine-dimensional feature space incorporating gas composition parameters (e.g., hydrogen equivalence ratio), pipeline geometry dimensions (length-to-diameter ratio), initial thermodynamic conditions (pressure), and porous medium structural characteristics (thickness, material thermal conductivity). A systematic investigation was conducted to identify the optimal hyperparameter configurations for both convolutional neural network (CNN) and Transformer architectures. Rigorous validation demonstrated that the Transformer model exhibits statistically significant superiority over the CNN model across all key performance metrics. Specifically, the model achieved a δ_MAE (mean absolute error, MAE) of 0.068, a δ_MSE (mean squared error, MSE) of 0.008, and an coefficient of determination R² of 0.928. The performance notably surpassed the CNN results (δ_MAE=0.079, δ_MSE=0.012, R²=0.906). Beyond evaluation indicators, detailed error distribution analysis confirmed the Transformer's enhanced predictive accuracy and reduced susceptibility to outliers. The superior performance is attributed to the Transformer’s intrinsic self-attention mechanism, which excels at dynamically identifying and weighting critical interdependencies among the diverse input features governing the complex quenching process. The capability enables more precise capture of the nonlinear phenomena defining the quenching limit. Furthermore, robustness testing involving diverse data normalization schemes revealed that the Transformer model exhibits greater stability. This resilience stems from its inherent layer normalization mechanism, which effectively decouples feature dependencies and mitigates sensitivity to input scaling variations. Consequently, the Transformer architecture is established as the definitive optimal model for this critical safety prediction task. Its significant advantage lies in the minimal data preprocessing required before deployment, a feature that greatly enhances its practical utility. The model provides robust quantitative decision-making support essential for formulating effective gas explosion mitigation strategies and optimizing the safety design parameters of pipeline flame arresters. By accurately predicting the critical quenching diameter across various scenarios, this work delivers a valuable tool for enhancing inherent safety in industries handling combustible gases within confined pipeline systems.
- Pipeline explosion /
- Deep learning /
- Critical quenching diameter prediction /
- Flame arrester

HTML全文

图 1 管道可燃气燃爆阻火实验平台

Figure 1. Pipeline combustible gas explosion and flame quenching test platform

下载: 全尺寸图片幻灯片

图 2 Transformer模型的核心架构

Figure 2. Core architecture of the Transformer model

下载: 全尺寸图片幻灯片

图 3 CNN模型的核心架构

Figure 3. Core architecture of the CNN model

下载: 全尺寸图片幻灯片

图 4 不同批量大小对Transformer模型性能的影响

Figure 4. The impact of different batch_size on Transformer model performance

下载: 全尺寸图片幻灯片

图 5 不同训练轮次对Transformer模型性能的影响

Figure 5. The impact of different epochs on the performance of Transformer models

下载: 全尺寸图片幻灯片

图 6 Transformer模型预测值与真实值比较

Figure 6. Comparison of Transformer model predictions and actual values

下载: 全尺寸图片幻灯片

图 7 CNN模型预测值与真实值比较

Figure 7. Comparison of CNN model predictions and actual values

下载: 全尺寸图片幻灯片

图 8 两种模型的评估指标分布

Figure 8. Distribution of evaluation indicators for the CNN and Transformer models

下载: 全尺寸图片幻灯片

图 9 两种模型的实验值和误差值比较

Figure 9. Comparison of experimental values and error values between the CNN and Transformer models

下载: 全尺寸图片幻灯片

图 10 归一化处理对两种模型性能的影响

Figure 10. The impact of normalization on the performance of the CNN and Transformer models

下载: 全尺寸图片幻灯片

表 1 燃气临界淬熄直径实验数据库摘要

Table 1. Abstract of gas critical quenching diameter experiment database

参考文献	燃气浓度/%	长径比	初始压力/ kPa	多孔介质厚度/mm	热导率/ (W·m⁻¹·K⁻¹)
[27]	14.6～38.7	14.188	101.325	40～120	16.2
[30]	17.4～42.4	7.143	20～100	20～100	238
[29]	27.4	30～40	101.325	30～120	16.2
[29]	4.2	50～70	101.325	50～150	16.2
[28]	1.2～2.5	14.188	101.325	40～120	16.2
[31]	5.0～9.5	50	30～60	38	16.2
[31]	4.2	50	70～80	38	16.2

下载: 导出CSV

表 2 Transformer模型最优超参数组合

Table 2. Optimal hyperparameter combinations for Transformer models

超参数	最优值
批量大小	32
训练轮次	500

下载: 导出CSV

表 3 CNN模型最优超参数组合

Table 3. Optimal hyperparameter combination for CNN models

超参数	最优值
批量大小	16
训练轮次	600

下载: 导出CSV

参考文献(37)

[1]	S. G S. Deflagration and detonation flame arresters [M]. 美国: John Wiley & Sons, 2010.
[2]	NIE Z H, GAO W, JIANG H P, et al. Predicting critical flame quenching thickness using machine learning approach with ResNet and ANN [J]. Journal of Loss Prevention in the Process Industries, 2024, 92: 105448. DOI: 10.1016/j.jlp.2024.105448.
[3]	LANGFORD B, PALMER K N, TONKIN P S. The performance of flame arresters against flames propagating in various fuel/air mixtures: Fire Research Note 486 [R]. Herts: Fire Research Station, 1961.
[4]	SONG X Z, ZUO X C, YANG Z K, et al. The explosion-suppression performance of mesh aluminum alloys and spherical nonmetallic materials on hydrogen-air mixtures [J]. International Journal of Hydrogen Energy, 2020, 45(56): 32686–32701. DOI: 10.1016/j.ijhydene.2020.08.197.
[5]	NIE B S, HE X Q, ZHANG R M, et al. The roles of foam ceramics in suppression of gas explosion overpressure and quenching of flame propagation [J]. Journal of Hazardous Materials, 2011, 192(2): 741–747. DOI: 10.1016/j.jhazmat.2011.05.083.
[6]	SUN S C, LIU G, LIU J X, et al. Effect of porosity and element thickness on flame quenching for in-line crimped-ribbon flame arresters [J]. Journal of Loss Prevention in the Process Industries, 2017, 50: 221–228. DOI: 10.1016/j.jlp.2017.09.017.
[7]	CAO X Y, WANG Z, WANG Y, et al. Effect of wire mesh structure parameter on the flame propagation characteristics of syngas explosion [J]. Fuel, 2023, 334: 126658. DOI: 10.1016/j.fuel.2022.126658.
[8]	LIU Z Y, GU Z Y, LIU P. Deep learning [M]//LIU Z Y, GU Z Y, LIU P. Transportation Big Data: Theory and Methods. Amsterdam: Elsevier, 2025: 383–430.
[9]	ZHAO F Y, GAO W, LU J X, et al. Real-time concentration detection of Al dust using GRU-based Kalman filtering approach [J]. Process Safety and Environmental Protection, 2024, 189: 154–163. DOI: 10.1016/j.psep.2024.06.052.
[10]	ZHANG K, ZHANG K, BAO R. Prediction of gas explosion pressures: A machine learning algorithm based on KPCA and an optimized LSSVM [J]. Journal of Loss Prevention in the Process Industries, 2023, 83: 105082. DOI: 10.1016/j.jlp.2023.105082.
[11]	CHEN J G, NIE Z H, ZHAO F Y, et al. Improving the stability of electrostatic induction dust concentration detection using kalman filtering algorithm aided by machine learning [J]. Process Safety and Environmental Protection, 2023, 174: 882–890. DOI: 10.1016/j.psep.2023.05.013.
[12]	GUIBERTI T F, BELHI M, DAMAZO J S, et al. Quenching distance of laminar methane-air flames at cryogenic temperatures and implications for flame arrester design [J]. Applications in Energy and Combustion Science, 2020, 1/2/3/4: 100001. DOI: 10.1016/j.jaecs.2020.100001.
[13]	FRIEDMAN R. The quenching of laminar oxyhydrogen flames by solid surfaces [J]. Symposium on Combustion and Flame, and Explosion Phenomena, 1948, 3(1): 110–120. DOI: 10.1016/S1062-2896(49)80014-6.
[14]	WEN X P, MA M, SUN W C, et al. The quenching characteristics of gas deflagration flame in narrow channel [J]. Advanced Materials Research, 2012, 455/456: 289–295. DOI: 10.4028/www.scientific.net/AMR.455-456.289.
[15]	WU Y, WEN X P, GUO Z D, et al. Experimental study on the propagation characteristics of hydrogen/methane/air premixed flames in a narrow channel [J]. International Journal of Hydrogen Energy, 2022, 47(9): 6377–6387. DOI: 10.1016/j.ijhydene.2021.12.019.
[16]	LONG F Y, DUAN Y L, YU S W, et al. Effect of porous materials on explosion characteristics of low ratio hydrogen/methane mixture in barrier tube [J]. Journal of Loss Prevention in the Process Industries, 2022, 80: 104875. DOI: 10.1016/j.jlp.2022.104875.
[17]	周凯元李宗芬, 周自金. 波纹板阻火器对爆燃火焰淬熄作用的实验研究 [J]. 中国科学技术大学学报, 1997, 27(4): 77–82. ZHOU K Y, LI Z F, ZHOU Z J. The quenching of deflagration by crimped-ribbon flame arresters [J]. Journal of China University of Science and Technology, 1997, 27(4): 77–82.
[18]	WANG L Q, MA H H, SHEN Z W. The quenching of propane deflagrations by crimped ribbon flame arrestors [J]. Journal of Loss Prevention in the Process Industries, 2016, 43: 567–574. DOI: 10.1016/j.jlp.2016.07.025.
[19]	HOLM J M. II. On the initiation of gaseous explosions by small flames [J]. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 1932, 14(89): 18–56. DOI: 10.1080/14786443209462032.
[20]	PIOTROWSKI T C. Specification of flame arresting devices for manifolded low pressure storage tanks [J]. Plant/operations Progress, 1991, 10(2): 102–106. DOI: 10.1002/prsb.720100210.
[21]	Chen D G, Ma H H, Shen Z W, et al. Investigation of the quenching of methane‐air deflagration in narrow parallel channels [J]. Process Safety Progress, 2020, 39(3): e12126. DOI: 10.1002/prs.12126.
[22]	LV X S, YU J L, HOU Y J, et al. The quenching and attenuation of hydrogen-air detonation after passing across capillaries [J]. Fuel, 2022, 324: 124535. DOI: 10.1016/j.fuel.2022.124535.
[23]	SHI J H, ZHANG H, LI J J, et al. Real-time natural gas explosion modeling of offshore platforms by using deep learning probability approach [J]. Ocean Engineering, 2023, 276: 114244. DOI: 10.1016/j.oceaneng.2023.114244.
[24]	SUZUKI Y, NAKAYAMA J, SUZUKI T, et al. Consequence analysis of a small-scale hydrogen leakage from the overhead hydrogen piping based on machine learning and physical modeling [J]. Journal of Loss Prevention in the Process Industries, 2024, 90: 105328. DOI: 10.1016/j.jlp.2024.105328.
[25]	QIN G J, XIA A L, LU H F, et al. A hybrid machine learning model for predicting crater width formed by explosions of natural gas pipelines [J]. Journal of Loss Prevention in the Process Industries, 2023, 82: 104994. DOI: 10.1016/j.jlp.2023.104994.
[26]	XU Q M, CHEN G H, SU S, et al. Prediction of venting gas explosion overpressure based on a combination of explosive theory and machine learning [J]. Expert Systems With Applications, 2023, 234: 121044. DOI: 10.1016/j.eswa.2023.121044.
[27]	NIE Z H, GAO W, JIANG H P, et al. Quenching characteristics and mechanism of hydrogen-air mixtures by corrugated plate flame arrester under inert conditions [J]. Fuel, 2024, 362: 130822. DOI: 10.1016/j.fuel.2023.130822.
[28]	NIE Z H, GAO W, JIANG H P, et al. Explosion quenching characteristics of corrugated plate microchannels in ethanol-gasoline/air mixtures [J]. International Journal of Thermal Sciences, 2024, 204: 109220. DOI: 10.1016/j.ijthermalsci.2024.109220.
[29]	孙少辰. 波纹管道阻火器系统的阻爆特性研究 [D]. 大连: 大连理工大学, 2018. SUN S C. Research of explosion resistant characteristic in in-line crimped-ribbon flame arrester [D]. Dalian: Dalian University of Technology, 2018.
[30]	万阳. 管道内针对氢-空气火焰的多孔板阻火特性研究 [D]. 合肥: 合肥工业大学, 2020. DOI: 10.27101/d.cnki.ghfgu.2020.001840. WAN Y. Study on fire resistance characteristics of perforated plate for hydrogen-air flame in a channel [D]. Hefei: Hefei University of Technology, 2020. DOI: 10.27101/d.cnki.ghfgu.2020.001840.
[31]	张炼卓, 闫兴清, 吕先舒, 等. 不同初始压力下可爆预混气体通过波纹板阻火器的淬熄特性研究 [J]. 爆炸与冲击, 2023, 43(1): 012202. DOI: 10.11883/bzycj-2021-4058. ZHANG L Z, YAN X Q, LV X S, et al. On quenching characteristics of combustible premixed gas through a crimped-ribbon flame arrester at different initial pressures [J]. Explosion and Shock Waves, 2023, 43(1): 012202. DOI: 10.11883/bzycj-2021-4058.
[32]	ZHONG X T, GALLAGHER B, EVES K, et al. A study of real-world micrograph data quality and machine learning model robustness [J]. npj Computational Materials, 2021, 7(1): 161. DOI: 10.1038/s41524-021-00616-3.
[33]	PEI X L, ZHAO Y H, CHEN L W, et al. Robustness of machine learning to color, size change, normalization, and image enhancement on micrograph datasets with large sample differences [J]. Materials & Design, 2023, 232: 112086. DOI: 10.1016/j.matdes.2023.112086.
[34]	VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need [C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach: Curran Associates Inc. , 2017: 6000–6010.
[35]	WAIBEL A, HANAZAWA T, HINTON G, et al. Phoneme recognition using time-delay neural networks [J]. IEEE Transactions on Acoustics, Speech, and Signal Processing, 1989, 37(3): 328–339. DOI: 10.1109/29.21701.
[36]	卢祖文. 基于X光图像与深度学习的耐张线夹缺陷识别 [D]. 南昌: 南昌大学, 2022. DOI: 10.27232/d.cnki.gnchu.2022.002699. LU Z W. Recognition of tension clamp defects based on X-ray image and deep learning [D]. Nanchang: Nanchang University, 2022. DOI: 10.27232/d.cnki.gnchu.2022.002699.
[37]	BI Y B, WANG Y B, WANG S L, et al. A hybrid expert neural network for predicting hydrogen concentration under the ceiling in underground garage [J]. Building and Environment, 2025, 268: 112325. DOI: 10.1016/j.buildenv.2024.112325.