Fire Risk Analysis Based on Artificial Neural Network Modelling Techniques

Fire phenomena involve complicated and highly non-linear interactions between parameters. Currently, computer simulations by use of numerical methods applying on mathematical models - zone and field modelling techniques in form of differential equations have been widely adopted for predicting results of fire phenomena. While the accuracy of the prediction is highly relying on the accuracy of the mathematical models applying on the fire phenomena, the latter also involves extremely expensive computations. This research explores the use of the artificial neural network (ANN) to simulate the behaviour of the enclosure fire growth phenomena. The ANN is an information processing system similar to the human judgment mechanism and has been considered effective in recognizing and judging a situation with “fluctuating” and “obscure” factors. In fact, it is a complicated mathematical tool for data regression. In this research study, in order to address to these issues, a novel and unique ANN model suitable for enclosure fire growth and fire risk analysis (FRA) will be developed based on General Regression Neural Network (GRNN) and Fuzzy ART (FA) Algorithms. The former is extremely powerful for the identification of highly non-linear characteristics of a system, while the latter will enhance the model by clustering of data into representative kernels. As the kernels formulated by FA cannot be directly used with the GRNN, a new compression technique will also be developed based on the knearestneigbourhood theory to provide a closure to this approach. For training and validations of the ANN fire model, the CFD models shall be used to provide simulation results including transient smoke layer height and temperature, time to flashover and flow through openings, etc. Also, experimental results from enclosure fire tests will be used for training and validations of the model. When applied to complicated studies of fire scenarios in buildings, the new ANN model will be relatively much faster and computationally economical than its counterparts (i.e. zone or field models). The results of this research study is expected to shed new lights to further enhance applications of ANN techniques for fire engineering predictions including flame and smoke spread and the fire risk assessment for buildings. The new GRNN-FA model will offer a much efficient and economic means for fire safety analysis in buildings which may be adopted in the Performance-Based Fire Engineering approach or Codes in Hong Kong and other countries internationally.