Intelligent approach to architectural design for fire safety

The Computational Fluid Dynamics (CFD) techniques are currently widely adopted to simulate the behaviour of fire. The major shortcoming of the CFD is the requirements of extensive computer storage and lengthy computational time. In actual applications, although comprehensive field information of velocities, temperature, pressure, fraction of different constitutes etc. can be obtained from the CFD simulation, the user may be only interested in few important parameters which index the performance of the compartment design in the event of fire. Height of thermal interface (HTI) is one of the key indices. It is the average height above the floor level inside the fire compartment at which the temperature gradient is the highest. In practice, the fire compartment is considered untenable when the HTI descends lower than the respiratory level of the occupants. In the course of fire system design optimization, if the resultant HTI of a fire compartment design evaluated by the CFD is too low, another set of the design parameters (e.g. width of the door opening) are required to be tried. This trial and error exercise continues until a close optimum set of the design parameter achieved. This approach is theoretically feasible but requires lengthy computational time. This paper proposes to apply Artificial Neural Network (ANN) approach as a fast alternative to the CFD models to simulate the behaviour of the compartment fire. A novel ANN model denoted as GRNNFA has been developed particular for fire studies. It is a hybrid ANN model combining the General Regression Neural Network (GRNN) and the Fuzzy ART (FA). The GRNNFA model owns the features of incremental growth of network structure, stable learning and removal of the noise embedded in the experimental fire data. It has been employed to establish a system response surface based on the knowledge of the available training samples. Since the available training samples may not be sufficient to describe the system behaviour especially for fire data, it is proposed to acquire extra knowledge of the system from human expert knowledge. The human expert intervened network training was developed to remedy the established system response surface. After the transformation of the remedied system response surface to the problem domain, Genetic Algorithm (GA) is applied to evaluate the close optimum set of the design parameters.