COMPARTMENT FIRE AND MACHINE LEARNING BASED PREDICTION OF FIRE RESISTIVE PROPERTIES

Abstract

Modern structures are common and highly complex; examples include residential complexes, office buildings, educational institutions, and industrial facilities. Several fire incidents in the past have resulted in the loss of life and expensive assets, raising concerns about building fire safety regulations. A thorough understanding of various fuels under various ventilation settings and the interaction between fire and its surroundings are necessary when designing a fire protection system. Extensive experimentation and better measurement methods can improve the complexity of understanding fire behavior. Many experiments with solid, liquid, and gaseous fuels have been carried out in full-scale and reduced-scale compartments throughout the past few decades. Numerous studies have examined the characteristics of compartment fires and connections between flame height, plume centerline temperature, velocity, and hot gas layer temperature. Compared to other combustible materials, the burning behavior of the wood cribs is more like the actual fire development in the compartment. Wood cribs are frequently used as ignition sources in room fire tests (e.g., UL 1715 and ISO 9705 test standards) and for other tests requiring repeatable heat release rates, such as performance evaluation of fire extinguishers (ANSI/UL 711). This is because of their simple structure and burning behavior, which is similar to the actual fire development in the compartment. To better understand wood crib fires, many studies have been conducted on the factors affecting the combustion process of wood cribs. Generally, factors influencing the burning of wood cribs could be divided into two aspects, namely internal properties and the surrounding environment. The internal properties include the species of wood, the size of wood cribs, the layer arrangement of the wood sticks, etc. Besides its characteristics, the external environments also significantly impact the burning behavior of wood cribs, including burning rate, radiation heat flux, and compartment ventilation. Although much research and investigations have been done based on the burning of wood cribs, the burning characteristics of wood cribs under different ventilation conditions in a confined space are seldom studied. The characteristics of wood crib burning are considered the basis for further studies about fire science and safety engineering, such as the fire safety assessment and the performance-based design of buildings. This study has three specific objectives. The first objective is focused on studying the burning characteristics of wood cribs in a confined space with different ventilation conditions. For the experiments, a full-scale compartment setup was built in the fire research laboratory of the Mechanical and Industrial Department, IIT Roorkee, with the dimensions of 4 m (L) x 4 m (W) x 4 m (H). For ventilation, a doorway opening of measurement 1 m width by 2 m height was provided in the middle of the front wall. A sliding door was installed to differentiate the ventilation conditions in the compartment. The mango wood cribs were used as the solid fuel in the compartment. The wood crib comprises 49 sticks of 50 cm length and a square cross-section of 5 cm in size, arranged in seven layers with seven sticks in each layer for all experiments. Various parameters have been measured in these experiments, such as heat release rate, mass loss rate, exhaust gas concentration, flame temperature profiles, flame height, upper zone temperature, and heat flux at different wall locations. The heat release rate has been measured by the large-scale heat release calorimeter, which works on the principle of oxygen depletion. Analyzing results, it has been found that the whole combustion process can be divided into the five stages of burning: Ignition, growth, steady burning, recess, and collapse. The ventilation situation significantly impacted the mass loss rate, and the heat release rate was remarkably affected by the vent areas; with the maximum vent, the heat release rate was the highest. An attempt has also been made to calculate the flame height, and for that, images taken from the wood crib burning were processed in MATLAB. The second objective is to perform numerical simulations for the wood crib compartment fire and elevated diesel compartment fire. Fire Dynamic Simulator (FDS), developed by NIST, USA, was used for the simulation. FDS numerically solves a large eddy simulation form of Navier Stokes equations, which are based on the assumptions of low speed and thermally driven flow, focusing on smoke flow and heat transport to describe the fire behavior. The heat release rate measured by the oxygen depletion calorimeter was specified as the input in FDS. The computational domain has been extended to the size of 5 m (L) x 5 m (W) x 5 m (H) for the wood crib compartment fire and 6 m (L) x 6 m (W) x 6 m (H) for the elevated diesel pool fire simulation. Simulations have been carried out for full door openings, and the total number of cells in the model is in the range of from 175616 to 2097152 for wood crib models and from 274625 to 1815848 for elevated diesel pool fires. The effect of empirical constants sub-grid scale coefficient, Sc, turbulent Prandtl number, Pr, and the sub-grid scale coefficient, Cs, has been studied. Simulation results show that with refinement in mesh size, predicted results are close to the experiment values, and the average temperature difference between experimental and predicted results during the steady state are 17, 14, and 14 °C, corresponding to simulations 1, 2, and 3, respectively. Simulation results for compartment gas temperature are underpredicted compared to experimental results for the burning stage, and after this stage, the results are close to the experimental values. In the case of the wood crib compartment fire, no significant effect was observed with the variation of empirical constants, and default values of FDS gave better accuracy for the simulation results. From the numerical simulation of elevated diesel pool fire, the value of D*/dx can be taken in the range of 10 to 15 to accurately resolve the events of compartment fire. The effect of different values of empirical constants on 6 pan elevations has been observed. Lesser Pr/Sc (Pr = Sc = 0.3) values gave better accuracy than the default values of these constants, and higher values overpredict the centerline flame temperature and doorway temperature. The third objective is to predict the fire-resistive properties of the concrete-filled steel tubular (CFST) columns with machine learning (ML) modeling. Artificial intelligence (AI) and machine learning are two novel techniques that have emerged due to the expansion of the scientific and technology community. The ultimate goal of artificial intelligence (AI) is to make computers intelligent like humans. Complex civil engineering problems have also been demonstrated to be resolved with the use of ML tools and AI methodologies. This study utilized individual and supervised machine learning algorithms to predict the RSI of circular CFST columns after exposure to high temperatures. Artificial neural network (ANN), Ensemble, Gaussian Process Regressor with different kernels, and Support Vector Machines models were used to forecast the RSI of CCFST columns. Using the literature, a database of 56 experimental tests was generated. The database was then divided into a test set and a training set, which were utilized for training and evaluating the performance of the trained ML models. The ReliefF feature ranking method was used for the initial weightage to consider the input parameters while modeling. ANN Bi-layered and Quadratic SVM outperformed the other ML models from the perspective of four typical performance metrics. To model the fire resistance rate of the CFST columns, 182 data points were gathered From the literature. Gene expression programming was used to model the data. An equation was given from the expression trees to predict the fire resistance rate of the CFST columns. Due to significant variation in the dataset, the regression was observed as 0.802. This is because all the data was extracted from the existing literature and comprised different concrete strength and testing conditions. The present study will be helpful for making decisions regarding the safe handling of fuels considered in the study in the chemical process plants and warehouses. These studies will also be useful for designing fire safety methods and, consequently, for fire safety standards.