An Introduction to Fire Performance Monitoring for Buildings

Wang, Honggang

Etd

An Introduction to Fire Performance Monitoring for Buildings

Public

Performance gaps exist in all building disciplines. Whether the performance (gap) in a building area can be monitored indicates the degree of maturity of this engineering area. Compared to the relatively high measurability of performance (gap) in building energy efficiency and structural health, it is much harder to identify and measure the fire performance (gap) of buildings in use, indicating less maturity of the fire protection discipline. This dissertation documents a study on building fire performance monitoring. The key idea is that although a fire accident is an acute phenomenon that is rare and unpredictable and makes observing a building’s fire performance very difficult, the evolution of underlying factors that could determine the building fire performance in a future fire accident is in fact a chronic phenomenon that is frequent, observable and predictable. A conceptual design of building fire performance monitoring (FPM) is proposed, which includes an input module, building fire performance gap (BFPG) checking module, and measure refining module. Due to the computational resource costs, CFD tools like FDS commonly used in performance-based fire protection design are inappropriate to be called frequently to estimate the dynamic fire performance of buildings in use. The development of substitute models is needed to check for BFPGs. Building fire performance includes various aspects like life safety, property loss, business continuity, and environmental damage, etc. This dissertation only focuses on the life safety aspect described by the available safe egress time (ASET) and/or required safe egress time (RSET) as well as the ratio of ASET to RSET which is the egress safety ratio (ESR). To set the fire scenarios, a small three-story apartment building is employed which includes eight 100m2 apartments in every story. A propane gas-burner fire is put close to the southeast corner of the southeast apartment at the first floor. There are two corridor doors set close to each end of the main corridor. In order to quickly calculate or estimate the building fire egress performance gap caused by changes in six input variables including peak heat release rate (HRR), the width of the apartment corridor door, the soot yield of the fuel, the width of the apartment window, the width of the apartment door, and the corridor smoke exhaust flow rate, three substitute methods are developed: sensitivity matrix method (SMM), response surface method (RSM), and artificial neural network (ANN). These methods are then applied and compared based on their applicability in terms of either model uncertainties including system bias and relative standard deviation (RSD), or percentages of model predictions falling in a preset acceptable error range. Based on Taylor’s linear approximation which only keeps the first order derivatives in Taylor’s series expansion, two SMMs are proposed: SMM-Center which uses center difference, and SMM-B/F which uses a combination of forward difference and backward difference depending on the value of an input variable in relation to its baseline value. The application results show that the SMM-B/F has slightly higher applicability, but the SMM-Center is more convenient. A prototype of fire performance monitoring (FPM) is demonstrated by tracking the building fire egress performance calculated by SMM-Center in the small three-story apartment building. Unlike the linear SMM, both RSM and ANN are commonly used non-linear function fitting methods. Different from the traditional RSMs where the necessary cases increase exponentially with the number of input variables, the two RSMs, i.e., the RSM-1 and RSM-2, introduced in this dissertation are based on a novel two-phase power function fitting process where the necessary cases rise linearly with the number of input variables, thanks to the two assumptions adopted: the power function relationship between an output quantity and input variables, and the independence among input variables. RSM-1 is developed from a specially designed dataset, whereas the RSM-2 is developed from a random dataset. The results of applying both RSMs to the same dataset shows that RSM-1 has higher applicability and lower cost. While both SMM and RSM are to some extent physics-based methods, the development of an ANN does not rely on any physical knowledge about how a system responses to input changes: it only depends on the training/validation dataset. In this dissertation MATLAB’s feedforward neural networks with error backpropagating algorithm is employed to approximate the FDS response. The optimizing process shows that a hidden layer size of 2 and training/validation dataset size of 80 are the best options as to the problem specified in this dissertation. The pros and cons of the three kinds of substitute methods are compared by applying them to the same group of data, which can be summarized as: the SMMs have the lowest cost and lowest applicability, whereas the RSMs and ANNs have comparable higher applicability; the ANNs work much better when the available safe egress time (ASET) is far away from the baseline ASET value, but fail to catch the changing directions of the ASET as in comparison to the other two methods. It is suggested from the comparative analysis of the methods that to better understand the BFPG dynamically it is a good choice to start with the SMMs and then move to RSMs and ANNs when enough cases are accumulated during the application of the SMMs. As an initial exploration in the area of FPM, this dissertation leaves many practical issues to be solved in the future such as: how to collect the data available in the current building management system and transform these data into what are directly adopted in the fire effect or egress models, how to integrate the fire frequency related factors into the process of FPM, and how to refine the current version of FPM tool

Creator