A Data Mining Method for Potential Fire Hazard Analysis of Urban Buildings based on Bayesian Network

Abstract

At present, with rapid development of China's urbanization, the population density increases, the structure of buildings become more complexity, and building materials and techniques emerge endlessly. Frequent unsafe personal behavior and complex external unsafe factors bring more uncontrollable influences on preventing and controlling fire hazard of buildings in urban area. Traditional methods of fire hazard analysis have limitations on fire hazards forecasting in complex urban areas. This paper presents a data mining method based on Bayesian Network for fire hazard analysis of urban buildings. Based on the historical records of fire incidents in a city of China in past three years, from 2014 to 2016, we analyze the potential fire risk according to building properties and outside influences of buildings. We process and analyze the data, and construct a Bayesian Network based on the analytic results and the actual fire extinguishing situation. After that, we train the model with positive samples and negative samples. At last, we use the Bayesian Network model to assess the risks of building fire hazards. By using ROC curve to analyze the accuracy of the model, we get accurate and stable results. Based on Bayesian Network model with building property and external influence, the building fire risk probability is about 1.0×10-9 to 1.0×10-12. We also introduce another machine learning method, Logistic Regression algorithm to evaluate the performance of Bayesian Network model. The results show that our Bayesian Network model can achieve better performance.