New Inverse Model for Detecting Fire-Source Location and Intensity

Abstract

An inverse model and a procedure to detect fire location and determine fire intensity are developed. A time-dependent temperature database over the entire parameter space is generated using a fire simulation model. Then Markov chain Monte Carlo sampling based on Bayesian inferencing is used to determine parameters such as source location and strength. The probability distributions of source location and fire intensity are then calculated by inference using a Markov chain. Three test cases are used to evaluate the model. First, the model is validated using experimental data from the National Bureau of Standards multiroom test series for a simple setting involving a relatively small room and a long corridor. Second, a two-story office-building fire with 35 compartments is used to investigate the sensitivity and reliability of the model. Third, a high-rise building with a large space structure is used to improve the inverse model. It is shown that predicted fire source and intensity match the actual values for both constant- and varied-intensity fires. The effects of the sensors' sampling interval, intersensor spacing, measurement error, working range, and delay time on the sensitivity and reliability of the method are studied. The results indicate that a 50 s sampling interval generally results in high estimation performance with a relative error of 1%, but decreasing the intersensor spacing from 20 to 10 m does not significantly improve the accuracy of the inverse intensity if the sampling interval is small enough, such as 100 s. It is also found that using the sensor network with its lower upper limit less than 124°C leads to overestimation of the fire intensity. In addition, the accuracy of the predicted fire location is not affected by the accuracy of the forward fire model, while the accuracy of fire intensity predicted by the inverse model is sensitive to the systematic errors or the accuracy of the forward model.