【作者】 郭超；

【导师】 闫治国；

【作者基本信息】 同济大学 ， 土木工程， 2023， 博士

【摘要】 火灾安全是隧道运营阶段需重点解决的关键问题。由于车辆碰撞和火灾蔓延等原因,隧道内极易发生双火源火灾事故,导致火灾规模和危险性显著增加,严重威胁隧道内被困人员的生命安全。隧道火灾应急处置不仅需要掌握隧道复杂火灾灾害的演化机理,同时需要实时准确的隧道现场火情信息作为决策依据。因此,深入研究隧道双火源火灾演化特性及火灾态势实时预测方法对于探明隧道复杂火灾致灾机理和发展隧道智慧消防技术具有重要意义。前人采用小尺寸模型试验和数值模拟方法对隧道双火源火灾开展了研究,但目前全尺寸隧道真实环境中双火源火灾燃烧行为,烟气流动模式和烟气温度特性尚不清楚,已有的隧道现场火情信息预测方法存在预测不准确,计算效率低下等问题,适用于真实运营隧道的实时火灾态势预测方法需要进一步研究。基于此,本论文将从知识驱动和数据驱动两个维度开展研究,第一部分采用经典隧道火灾科学研究思路,通过全尺寸火灾试验、理论分析和数值模拟方法对隧道双火源火灾演化特性开展研究。第二部分采用数据驱动研究思路,采用机器学习和迁移学习方法对隧道双火源火灾态势实时预测问题开展研究。主要工作内容和结论如下:(1)研究了全尺寸隧道真实环境中双火源火灾燃烧行为和烟气流动模式,揭示了全尺寸隧道双火源火灾遮挡卷吸效应和热反馈增强效应的耦合作用机制。通过在三条全尺寸隧道开展系列火灾试验,发现了通风状态和火源间距可以改变遮挡卷吸效应和热反馈增强效应耦合作用中的主导关系,从而使隧道双火源火灾燃烧特性和烟气流动模式呈现出不同的演化规律。当火源间距较小时(无量纲火源间距S/D(28)0.75),在自然通风情况下,遮挡卷吸效应占据主导,隧道双火源火灾规模较小且保持稳定;在纵向通风情况下,热反馈增强效应占据主导,隧道双火源火焰可能融合,火灾规模会显著增大。纵向通风双火源火焰融合时火灾热释放速率相比自然通风双火源火焰没有融合时增加了4倍。随着火源间距增大,双火源之间相互作用逐渐减弱,双火源火灾的燃烧特性与单火源火灾相似。此外,全尺寸隧道双火源烟气流动模式分析表明两阶段通风相比全程纵向通风更加适合隧道双火源火灾事故烟气控制,尤其可以保障双火源之间区域被困人员的安全。(2)研究了全尺寸隧道双火源火灾增长阶段和稳定阶段烟气温度演化特性。阐明了隧道双火源火灾稳定阶段不同区域烟气温度分布规律,基于全尺寸隧道火灾试验结果建立了隧道双火源下游区域纵向烟气温升分布模型。结合数值模拟方法,探究了火源间距和隧道坡度对隧道双火源火灾稳定阶段烟气最高温度的影响规律,并建立了相应的烟气最高温升预测模型。通过与全尺寸火灾试验和前人开展的小尺寸模型试验结果进行对比,验证了预测模型的有效性。此外,研究了全尺寸隧道弱羽流火灾增长阶段烟气最高温度演化特性,发现火灾增长阶段实时烟气最高温升与实时火灾总释放热量高度相关,采用理论分析和量纲分析方法建立了全尺寸隧道弱羽流火灾增长阶段实时烟气最高温升预测模型。通过与四条全尺寸隧道单火源和双火源火灾试验结果对比,验证了所提出模型的有效性。(3)针对全尺寸隧道单火源火灾关键信息实时预测问题,建立了一种结合隧道火灾理论和机器学习的隧道火灾态势实时预测和火灾场景重构方法。通过模块化隧道空间策略,实现了全尺寸隧道物理空间全域以及火灾增长和稳定阶段时间全过程的火情信息实时准确预测。将隧道单火源火灾场景的数值模拟结果和全尺寸火灾试验结果分别作为测试集,对比了不同火灾场景下两种隧道火灾理论模型和四种机器学习模型实时预测火灾态势的准确性。结果表明,隧道火灾理论模型可靠性不足,在真实隧道火灾场景适用性受限,而机器学习模型在不同隧道火灾场景测试集中均有较好的实时火灾态势预测能力,其中基于长短期记忆神经网络的隧道火灾态势实时预测模型在全尺寸火灾试验场景测试集和部分温度数据缺失的数值模拟场景测试集上的决定系数分别达到82.3%和93.7%。最后,基于全尺寸隧道火灾试验数据验证了隧道火灾温度场重构方法的有效性。(4)在隧道单火源火灾态势实时预测的基础上,提出了一种基于迁移学习的全尺寸隧道双火源火灾态势实时预测方法,解决了全尺寸隧道火灾态势预测通用模型与特定隧道差异化特征之间的矛盾,实现了近火源区域关键温度传感器损坏情况下隧道火灾态势的准确预测。以三条不同全尺寸隧道的火灾试验结果作为测试集,分析了训练数据集规模对隧道火灾态势预测模型准确性的影响,实现了小数据集下不同全尺寸隧道单火源和双火源火灾态势预测模型的迁移,相关模型在全尺寸隧道火灾试验测试数据集上的决定系数最高达到80.2%。随后,评估了基于迁移学习的隧道火灾态势预测模型的计算效率和可靠性。测试结果表明,模型完成六个火灾场景实时态势预测的耗时均在1 s以内,具有显著的计算效率优势,满足隧道火灾事故应急处置的时效性需求;全尺寸隧道单火源和双火源火灾试验场景下,当近火源区域内1个温度传感器完全损坏时,相关模型的准确性平均仅下降8%,表明该方法在隧道复杂火灾场景中具有较好的可靠性。更多还原

【Abstract】 Fire safety is a critical concern that must be addressed during the tunnel operation stage.Due to vehicle collisions and fire spread,complex tunnel fire accidents,which involve double-source fires,are highly likely to occur,leading to a significant escalation in both the scale and danger of fires,posing a severe threat to the safety of trapped personnel in tunnels.To effectively manage tunnel fires in emergencies,it is necessary to understand the evolution mechanisms of complex tunnel fires and to have access to real-time and accurate fire information for making effective decisions during emergency response.Therefore,it is of great significance to conduct in-depth studies on the fire evolution characteristics and fire state prediction methods of tunnel double-source fires for finding out the evolution mechanism of tunnel complex fires and developing tunnel intelligent fire-fighting technologies.Previous studies have focused on tunnel double-source fires through small-scale model tests and numerical simulations.However,the combustion behavior,gas temperature characteristics and smoke flow patterns of double-source fires in full-scale tunnels remain poorly understood.The existing methods for predicting fire-related information at tunnel sites suffer from issues such as large prediction error and insufficient computational efficiency.Further research is required to develop real-time fire information prediction methods suitable for operational tunnels.Hence,this thesis conducts research from two perspectives:the knowledge-driven and the data-driven.In the first part,the classical tunnel fire science research paradigm is adopted to investigate the fire evolution characteristics of tunnel double-source fires using full-scale fire tests,theoretical analysis,and numerical simulation methods.In the second part,the data-driven research paradigm is adopted for studying the real-time prediction of the fire state in tunnel double-source fire scenarios using machine learning and transfer learning methods to address the emergency response requirements for tunnel fires.The main work contents and conclusions are as follows:(1)The combustion characteristics and smoke flow patterns of the entire process of double-source fires in full-scale tunnels are studied,revealing the coupling effects of shielding entrainment and thermal feedback enhancement on tunnel double-source fires.Through a series of fire tests conducted in three full-scale tunnels,it is found that the ventilation state and the fire source spacing can alter the dominant relationship between the shielding entrainment effect and the thermal feedback enhancement effect,resulting in different evolution patterns of combustion characteristics and smoke flow patterns of tunnel double-source fires.When the fire source spacing is small(dimensionless fire source spacing S/D(28)0.75),the shielding entrainment effect is dominant under natural ventilation,leading to lower and stable fire scale.And the thermal feedback enhancement effect becomes dominant under longitudinal ventilation.The double fire source flames may merge and the fire scale will increase significantly.The fire heat release rate increases by 4 times when the double fire source flames merged under the longitudinal ventilation compared with the double fire source flames without fusion under the natural ventilation.With the increase of fire source spacing,the interaction between double fire sources gradually weakens,and the combustion characteristics of double-source fires become similar to those of single-source fires.Additionally,the analysis of the smoke flow patterns shows that the two-stage ventilation is more suitable for smoke control compared to full longitudinal ventilation in tunnel double-source fires accidents,especially for ensuring the safety of individuals trapped in the area between the two fire sources.(2)The characteristics of gas temperature in the growth and steady stages of double-source fires in full-scale tunnels are investigated,elucidating the gas temperature distribution in different regions during the steady stage of tunnel double-source fires.The longitudinal gas temperature rise distribution models in the downstream region of tunnel double-source fires are established based on the data from full-scale tunnel fire tests.By combining numerical simulation methods,the effects of fire source spacing and tunnel slope on the maximum gas temperature rise during the steady stage of tunnel double-source fires are explored,and the corresponding prediction model for maximum gas temperature rise is established.The effectiveness of the prediction model is validated by comparing it with the results of full-scale fire tests and previous small-scale model tests.Additionally,the evolution characteristics of maximum gas temperature rise during the growth stage with weak plume in full-scale tunnels are studied.It is found that the real-time maximum gas temperature rise is related to the real-time total heat release during the fire growth stage.A prediction model for real-time maximum gas temperature rise during the growth stage of full-scale tunnel fires with a weak plume is established using theoretical analysis and dimensional analysis methods.The effectiveness of the proposed model is validated by comparing it with the results of four full-scale tunnel fire tests with single-source fires and double-source fires.(3)Aiming at the tunnel single-source fires,a method combining tunnel fire theory and machine learning is developed to address the real-time prediction of key information of tunnel fires,as well as the reconstruction of tunnel fire scenarios.Through the modular tunnel space strategy,the real-time prediction of tunnel fire information has been achieved for the entire full-scale tunnel physical space,covering both the fire growth and steady stages throughout the entire duration.The test datasets consist of numerical simulation results and full-scale fire test results of tunnel single-source fire scenarios,which are used to compare the real-time heat release rate inversion performance of four machine learning methods and two tunnel fire theoretical model methods.The results demonstrate that the applicability and reliability of tunnel fire theory models are limited,making them inadequate for the prediction in real tunnel fire scenarios.On the other hand,machine learning models exhibit good real-time fire state prediction capabilities across different tunnel fire scenarios in the test datasets.Among these models,the real-time tunnel fire state prediction model based on long short-term memory networks achieves determination coefficients of 82.3%and 93.7%on the full-scale fire tests dataset and the numerical simulation scenarios dataset with partial temperature data missing,respectively.Lastly,the effectiveness of the tunnel fire temperature field reconstruction method is validated based on the data from full-scale tunnel fire experiments.(4)On the basis of the real-time fire state prediction of tunnel single-source fires,the transfer learning strategy is adopted to predict the real-time fire state of tunnel double-source fires.This method resolves the contradiction between the general model for predicting full-scale tunnel fire states and the differentiated characteristics of specific tunnels,achieving accurate prediction of real-time tunnel fire state in the presence of missing key temperature data near the fire source.By taking three different full-scale tunnel fire tests data as the test datasets,the influence of training dataset size on the prediction accuracy of tunnel fire state is analyzed.The transfer of fire state prediction models for single-source fire and double-source fire scenarios in different full-scale tunnels with small training datasets is achieved,and the determination coefficient of the relevant models reach up to 80.2%on the full-scale tunnel fire test dataset.Then,the efficiency and reliability of the tunnel fire state prediction models based on transfer learning are evaluated.The results indicate that the time taken by the models to complete real-time tunnel fire state prediction for six fire scenarios is within 1 second,which has a significant advantage in computational efficiency and meets the timeliness requirements for emergency response to tunnel fire accidents.When one temperature sensor near the fire source is completely damaged,the average decrease in prediction accuracy is only 8%in the full-scale tunnel single-source fire and double-source fire scenarios,demonstrating good reliability in tunnel complex fires scenarios.更多还原