This website uses cookies to store information on your computer. Some of these cookies are used for visitor analysis, others are essential to making our site function properly and improve the user experience. By using this site, you consent to the placement of these cookies. Click Accept to consent and dismiss this message or Deny to leave this website. Read our Privacy Statement for more.
FPE Extra Issue 3, March 2016

The Fire Navigator: Forecasting the Spread of Building Fires on the Basis of Sensor Data

By Nahom Daniel and Dr. Guillermo Rein

The SFPE Foundation awarded the 2015 Chief Donald J. Burns Memorial Research Grant to the Department of Mechanical Engineering at the Imperial College of London. The grant honors the memory of FDNY Assistant Chief Donald Burns who perished while setting up his command post during the collapse of the World Trade Center Towers on September 11, 2001. The award aims to encourage the use of information modelling as a means to improve infrastructure safety and fire service preparedness. The grant was funded by Bentley Systems.

A novel cellular automata building fire model employing sensor data assimilation, inverse modelling and genetic algorithm techniques was implemented in order to indirectly uncover the governing parameters of a fire such as the flame spread rate, the smoke ceiling jet velocity and the outbreak location and time. These parameters were then used to produce real-time as well as forecast maps of the flame spread and smoke propagation. The research developed a tool called the Fire Navigator (Figure 1) which bridges the gap between fire safety systems and building information models (BIM) by making use of the mountain of data already produced by high-rise building sensors such as smoke and heat sensors.

It is envisioned that the forecasting of fire dynamics in buildings can lead to a paradigm shift in the response to fire emergencies, providing the fire service with essential local and global information about smoke propagation and flame spread ahead of time (i.e. seconds to minutes before it happens). Disposing of information on fire events before they actually happen can have a very positive effect on the fire service efficiency and safety, therefore saving human lives and mitigating equipment destruction. The idea is that fire fighters attending an emergency in a smart building can reach the fire location with as much information as possible about the past, present and future of the blaze using building sensor data feeds.

Smart buildings, which naturally developed from the concept of smart homes, will be able to anticipate the occupants’ needs with the help of various sensors. Control of heating, energy consumption and lighting are now common examples of how simple sensors can allow control over key aspects of homes and buildings. This can be extended to fire safety concerns via the transition from tradition-based fire fighting to smart fire fighting [1]. Already existing commercial smoke and heat sensors, as well as sprinklers, generate data that has yet to be fully harnessed and used in fire modelling tools in order to improve fire emergency responses. The advent of the BIM approach and the availability of computational power gives the opportunity to improve fire service efficiency by increasing situational awareness and by giving real-time feedback to first-responders on the optimal way to evacuate occupants and fight fires. BIM can be defined as a global digital representation storing key physical and functional aspects of a facility. Its purpose is to provide a coherent basis for all of the building inception and maintenance actors to make information-rich decisions. Only very recently, research such as that from the University of Edinburgh and the Imperial College of London [2] [3] [4], the Fire Research Group at the University of Texas at Austin [5] and the Department of Fire Protection Engineering at the University of Maryland [6] has started to develop the use of sensor data assimilation, inverse modelling and structural BIM in a global framework. The Fire Navigator’s ambition is to help these ideas coalesce in the very near-future.

Figure 2 illustrates the main components of the Fire Navigator and the flow of information into and out of the system’s modules. Three core modules process two types of input in order to produce real-time and forecast maps of flame spread and smoke propagation.

Building BIM files are the first input processed by the Fire Navigator via the BIM module interface. Any given floor’s digital representation is reduced to a two dimensional grid map as seen in Figure 3. This was done not only to simplify the problem but also to prepare the structural data input for the 2D fire model. Bentley Systems BIM software was used for the preliminary analysis of a case study and real-world Industry Foundation Classes (IFC) BIM file were provided by the project’s industrial advisor. A custom BIM file viewer and information extraction module was then implemented. The result is a stand-alone non-proprietary module which allows anyone to easily extract 2D floor cross sections from an IFC file.

The second input into the system are real-world sensor data feeds theoretically processed by the Flame-Smoke (FS) module which interfaces with Fire Alarm and Security systems. However, access to and processing of Building Management System (BMS) log files was not possible given the time frame of the project. Therefore, it was decided early on to implement a custom fire dynamics simulator based on Cellular Automata (CA) theory into the FS module itself.

The FS module starts with the 2D floor representation output by the BIM module and implements a novel flame spread and smoke propagation model based on CA theory [7] [8] [9] [10] [11]. While CA simulations simplify the problem at hand and do not produce results as accurate as computational fluid dynamics, they are computationally fast, reasonably accurate and therefore have the potential for use in real-world fire-fighting situations. Figure 4 illustrates the basic concept of CA modelling. Each cell of the floor grid is associated to a set of variables that reflect either the cell’s nature (wall, window, door etc.) or physical state (unburned fuel cell, burning fuel cell, boundary smoke cell etc.). Simple evolution rules are then applied at each time step and cell states are updated. For instance, smoke will propagate from one cell to its neighbouring cells after a certain amount of time which depends on the underlying fire dynamics parameters. The beauty of this approach is that the phenomena is processed at a high level which guarantees fast computation. Once a reasonable fire model (so-called forward model) was determined after literature review, the FS module successfully produced synthetic sensor data for the Inverse Modelling (IM) module.

The IM module receives as input, the processed structural CA grid and allows the user to place smoke and heat sensors on the map. The fire simulation conducted by the FS module is then used to create synthetic data feeds reflecting sensor activation times. This so-called data assimilation phase corresponds to the real-world situation where sensors send signals to the BMS when events occur. The ambition of the Fire Navigator is to uncover the underlying fire parameters by solving an inverse problem which relies solely on these mere sensor activation signals. In general terms, solving an inverse problem consists of measuring the effects of a physical phenomenon (i.e. via sensor data) and indirectly finding the unmeasured causal parameters that best describe them. These parameters can then be used to make forecasts of the phenomenon. Indeed, predicting events before they occur is the final objective of the Fire Navigator. The capacity to make forecasts is determined by what is called the lead time. It is the delay between the time of a predicted event and the time of its prediction. In other words, it quantifies the forecasting capacity of the algorithm. Achieving positive lead times has motivated all of the choices that were made with regard to which physics models and algorithms to implement.

Synthetic smoke sensor data for a generic office building floor with a regular array of smoke sensors and varying noise was produced and exploited in order to investigate the model’s efficiency and robustness. Figure 5 shows that after only 5 sensor activations (sensor IDs 1 to 5), all of the ensuing detection events are correctly predicted (3% error rate). Not only are the events predicted about a minute or so after the outbreak, the computation is fast enough -- seconds to solve the inverse problem, that only, the underlying governing parameters of the fire model have already converged for the remainder of the simulation. In other words, positive lead times of several minutes can be achieved and thus the predictions are actually forecasts.

Of course, these performance results are theoretical because the inverse algorithms were not tested on actual sensor data feeds. Nonetheless, the feasibility of the project’s ambition has been confirmed. This new technology of fire forecasting is now ready for a proof of concept inside a real building and extension to the protection of other key infrastructures like tunnels and power-plants.

Nahom Daniel and Dr. Guillermo Rein are with the Department of Mechanical Engineering, Imperial College London, UK


The authors thank Arup, specially Judith Schulz, for sharing their expertise in BIM and fire protection systems, and thank KPF for permitting the use of their architectural BIM models.


1Anthony Hamins, Casey Grant, P.E., FSFPE, Albert Jones, and Nelson Bryner, "Smart Fire Fighting and Fire Protection", Fire Protection Engineering, URL

2W. Jahn, G. Rein, J. Torero, Forecasting fire dynamics using inverse computational fluid dynamics and tangent linearisation, Advances in Engineering Software 47 (1) (2012) 114 – 126. doi: URL

W. Jahn, G. Rein, J. Torero, Forecasting fire growth using an inverse zone modelling approach, Fire Safety Journal 46 (3) (2011) 81 – 88. doi: URL

3W. Jahn, Inverse modelling to forecast enclosure fire dynamics, Ph.D. thesis, University of Edinburgh (2012). URL

4K. Overholt, O. Ezekoye, Characterizing heat release rates using an inverse fire modeling technique, Fire Technology 48 (4) (2012) 893–909. doi: 10.1007/s10694-011-0250-9.

5M. Price, "Using Inverse Fire modeling With Multiple Input Signals to Obtain Heat Release Rates in Compartment Fire Scenarios,” University of Maryland, Department of Fire Protection Engineering (2014).

6R. M. Almeida, E. E. N. Macau, Stochastic cellular automata model for wildland fire spread dynamics, Journal of Physics: Conference Series 285 (1) (2011) 012038. URL

7L. H. Encinas, S. H. White, A. M. del Rey, G. R. Sanchez, modeling forest fire spread using hexagonal cellular automata, Applied Mathematical modeling 31 (6) (2007) 1213 – 1227. doi: http://dx.doi. org/10.1016/j.apm.2006.04.001. URL

8C. Li, J. Li, L. Hu, D. Hou, Visualization and simulation model of underground mine fire disaster based on cellular automata, Applied Mathematical modeling 39 (15) (2015) 4351 – 4364. doi: http: // URL

9S. Wolfram, Statistical mechanics of cellular automata, Rev. Mod. Phys. 55 (1983) 601–644. doi: 10.1103/RevModPhys.55.601. URL

10A. Cowlard, W. Jahn, C. Abecassis-Empis, G. Rein, J. Torero, Sensor assisted fire fighting, Fire Technology 46 (3) (2010) 719–741. doi: 10.1007/s10694-008-0069-1. URL

© SFPE® | All Rights Reserved
Privacy Policy