Page 25 - 《爆炸与冲击》2026年第9期
P. 25

第 46 卷    第 9 期                   爆    炸    与    冲    击                       Vol. 46, No. 9
                2026 年 9 月                    EXPLOSION AND SHOCK WAVES                          Sept., 2026

               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                                                           *


                                              聂仲恒 ,汪    丽 ,高    伟 ,姜海鹏     1
                                                    1,2
                                                            2
                                                                   1
                                        (1. 大连理工大学化工学院,辽宁省 大连市 116024;
                                            2. 北京航天试验技术研究所,北京 100074)

                  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
                                                                            2
               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  R =0.928)显著超越了广泛应用的卷积神
                                                                      2
               经网络(convolutional neural network,CNN)模型(δ MAE =0.079,δ MSE =0.012,R =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,可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强大的、可操作的决策支持工
               具,具有重要的理论指导意义。
                  关键词: 管道爆炸;深度学习;临界淬熄直径预测;阻火器
                  中图分类号: O381; X932   国标学科代码: 13035; 083700   文献标志码: A

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

                                                            2
                                                                      1
                                                 1,2
                                    NIE Zhongheng , WANG Li , GAO Wei , JIANG Haipeng 1
                                   (1. Dalian University of Technology, Dalian 116024, Liaoning, China;
                                2. Beijing Institute of Aerospace Testing Technology, Beijing 100074, China)

               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    (mean  squared  error,  MSE)  of  0.008,  and  an  coefficient  of  determination  R   of  0.928.  The
                                                                                               2
                                   MSE
                                                                       2
               performance notably surpassed the CNN results (δ MAE =0.079, δ MSE =0.012, R =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



                 *   收稿日期: 2025-07-14;修回日期: 2026-03-28
                   基金项目: 国家重点研发计划(2024YFC3082600);国家自然科学基金(52574230)
                   第一作者: 聂仲恒(2000- ),男,硕士,nzh2490823579@163.com
                   通信作者: 姜海鹏(1990- ),男,博士,副教授,jhp@dlut.edu.cn


                                                         091502-1
   20   21   22   23   24   25   26   27   28   29   30