Testing for Knowledge: Maximizing Information Obtained from Fire Tests, Using Machine Learning Techniques

目前对各种材料的防火测试程序主要是针对合规性。然而，当实际火灾暴露与测试情况不同时，无论是持续时间、最高温度、HRR梯度还是其他方面，这都会带来性能不足的风险。拟议的研究项目旨在通过提供针对性测试的程序，对材料或配置的性能进行更完整的表征。作为博士生，我将在爱丁堡大学的Grunde Jomaas博士的监督下工作，并在根特大学的Ruben货车Coile博士的指导下工作，以触及问题的工程方面以及主要方面这是消防安全。项目计划包括一个分阶段的方法，从具体的应用到一个通用的框架工作。在第一阶段，先进的回归技术被应用于开发替代模型，为既定的火灾测试，即锥形量热仪测试和角燃烧器测试。利用已有的试验数据（训练集）对代理模型进行训练，并通过交叉验证数据（交叉验证集）确定模型结构，然后利用代理模型预测未知样本（试验集）的火灾试验结果，第二阶段利用代理模型对火灾试验中获得的知识进行泛化。应用领域的这种扩展与对模拟结果的信心丧失有关。这种信心的丧失自然可以通过测试来减少。通过研究代理模型的基本数学，可以识别出其结果预期会导致代理模型的最大置信度收益的测试设置。替代模型将根据额外的测试数据进行迭代更新，这些测试数据专门用于其“知识受益”。第三阶段将制定一个“知识测试”的一般程序，与任何一系列的防火测试（包括创新材料或设计的测试）结合使用。这被认为是非常重要的，因为它允许使用尽可能少的测试次数来获得对火灾性能的全面了解。通过反复测试被认为能产生最大“知识效益”的设置，以最小的成本获得全面的消防工程理解。在第4阶段，通过应用于创新的消防测试（新材料，新配置或其他），验证了所提出的程序。试验的具体细节将根据当时的试验要求确定（即，当时出于商业目的，需要更好地了解其性能的特定材料或配置）。

Current fire testing procedures for various materials are targeted primarily at compliance. This however introduces a risk of inadequate performance when the actual fire exposure differs from the test situation, be it in duration, maximum temperature, the HRR gradient or other. The proposed research project aims to provide a more complete characterization of the performance of materials or configurations by providing a procedure for targeted testing.As a PhD student, I will be working under the supervision of Dr. Grunde Jomaas from the University of Edinburgh and with the guidance of Dr. Ruben Van Coile from the University of Ghent to touch upon the engineering aspects of the problem as well as the main aspect which is fire safety.The project plan consists of a phased approach, working from specific applications towards a generalized framework.In phase 1 advanced regression techniques are applied to develop surrogate models for established fire tests, i.e. the Cone Calorimeter test and the Corner Burner test. After training the surrogate model with available test data (training set), and determining model architecture through cross validation data (cross validation set), the surrogate model is applied to predict fire test results of unseen samples (test set).In phase 2 the surrogate models are applied to generalize the knowledge obtained from the fire tests. This expansion of the field of application is associated with a loss of confidence in the simulation results. This loss of confidence can, naturally, be reduced through testing. By investigating the underlying mathematics of the surrogate models, test setups can be identified whose result is expected to result in the greatest benefit in confidence for the surrogate model. The surrogate models will be updated iteratively in function of additional test data which are specifically performed for their 'knowledge benefit'. Validation of the approach will be done through the execution of the specific identified tests.Phase 3 derives a general procedure for 'Testing for Knowledge' to be applied in conjunction with any series of fire tests, including the testing of innovative materials or designs. This is considered of great importance as it allows to obtain a comprehensive understanding of fire performance using the least possible number of tests. By iteratively testing set ups which are considered to result in the greatest 'knowledge benefit', a comprehensive fire engineering understanding is obtained at minimal cost.In phase 4 the proposed procedure is validated by application to innovative fire testing (new material, new configuration or other). The specifics of the tests will be determined with respect to testing requirements at the time (i.e. specific material or configuration whose performance needs to be better understood for commercial purposes at the time).