1. Introduction
Enclosure or confined fires are situations that firefighters are used to handling. The consequences of this type of emergency can involve people and structures. A confined fire is one that takes place inside structures such as blocks of houses, garages, single-family homes and commercial establishments. The last International Association of Fire and Rescue Services (CTIF) study [
1] suggests that 35% of fires (see
Figure 1) located in several cities around the world are structure fires. These kinds of emergencies caused civilian (non-fire service) deaths, civilian fire injuries and property damage. Currently, the reported data about this phenomenon in Europe is very poor. This is because there is not a common database which involves all European countries and most of the fire services do not document such situations. An example of this is the study carried out by the European Fire Safety Alliance (EFSA), which only involves nine countries.
Table 1 shows an overview of the results from research [
2] carried out for EFSA related to residential fires.
To compare these results with data from another continent the reader can refers to the National Fire Protection Association (NFPA [
3]) statistics. According to NFPA, during the five-year period between 2010–2014, USA fire departments responded to an estimated annual average of 358,300 home structure fires. These kinds of emergencies caused an annual average of 2520 civilian (non-fire service) deaths, 12,720 civilian fire injuries, and US
$ billion in direct property damage. On average, seven people died each day in the US in home fires during this period. Currently it is very difficult to obtain accurate data to evaluate confined space fires in Europe, because there is no database available for this dangerous type of emergency. Furthermore, there is no data available on how this type of enclosure fire affects emergency teams, materials or victims.
In order to get some idea of the risk for people involved in these situations,
Table 2 provides a more detailed breakdown of losses by occupancy in two types of buildings, family homes and buildings. 70% reported home structure fires and 84% of the fatal home fire injuries occurred in one- or two-bedroom family homes, including manufactured homes.
A compartment fire [
3] is one that takes place inside an enclosed area, such as a house, in which two circumstances can occur:
The enclosure is ventilated, the oxygen consumption is less than the available amount.
The room is under-ventilated (lack of inlet air or lack of an outlet for the smoke). During the combustion process, oxygen levels in the enclosure decrease. No open vents or oxygen displacement by combustion gases can be the causes.
To understand the combustion process in the enclosure it is necessary to introduce the reader to fuel-controlled (FC) and ventilation-controlled (VC) concepts. If there is sufficient oxygen to consume the fuel pyrolysis gases released, this is known as an FC scenario. Conversely, if there is lack of oxygen to continue the reaction, this is a VC situation. Taking into account both enclosure situations (ventilated and under-ventilated), once the ignition in a material has occurred, there a several ways in which the situation may evolve. In a ventilated enclosure, if the fire dies after ignition, this can be because the energy released is not sufficient to further pyrolysis in the same material or in nearby materials. For that reason it is not possible for the fire to spread (FC). Alternatively, if the enclosure is under-ventilated, the lack of oxygen in early stages can result in a non-adequate mix with the fuel gases impeding the combustion process (VC). Another possibility is when the heat released from the flames is enough to generate new pyrolysis gases from new unburned material and sufficient oxygen is present to maintain the combustion process. Subsequently, fire reaches the growing stage, where fire can spread over the same surface as flames spread or reach other surface materials by radiation. In a ventilated confined space, all fuel can be burned and fire dies or a transition from FC to VC can take place causing the fire to die out due to lack of oxygen. In some cases, the fire can grow to become a fully developed fire after a transition stage known as flashover.
Flashover phenomenon, or also called generalised sudden combustion, can be defined as a transition phase from growth to fully developed stage (see
Figure 2a). As a consequence, all combustible surfaces inside the enclosure, that were not involve in the fire, begin to burn. It occurs due to the radiation received from the smoke layer which can be up to 170 kW/m
[
4]. Once all fuels are involved in the fire the fully developed stage is reached. As a result, firefighters can’t stay inside the enclosure under flashover conditions, 80 kW/m
is the maximum supported radiation by their clothes according to NFPA 1971, Standard on Protective Ensemble for Structural Fire Fighting. In the
Table 3 different radiation values are shown, and they can be compared to the maximum heat flux for a post-flashover case in order to get an idea of this phenomenon’s power.
The NFPA [
4] also defines flashover as the transient phase in the development of an indoor fire in which surfaces exposed to thermal radiation reach their ignition temperature almost simultaneously and the fire spreads rapidly throughout the space available within the enclosure.
Flashover can occurs in two different scenarios related to the enclosure configuration and it depends on the fire is located inside a structure with constant air supply or not. The first one known as a fuel-limited fire is shown in the
Figure 2a. In the growth stage there is enough oxygen to maintain the combustion process. If amount of fuel involved is sufficient therefore the energy level necessary for the occurrence of flashover can be reached followed by the full developed phase. Then, as the fuel is burned away, the energy level begins to decay. Once the fully developed phase is achieved, in most cases the fuel mass release rate becomes so high that the air supply rate becomes insufficient to consume all released pyrolysis gases (fuel-controlled to ventilation-controlled) [
5] and the energy level begins to decay. On the second one (see
Figure 2b), classified as under-ventilated situations, flashover stage can be reached after a fire dynamics change. It can be produced as a consequence of an induced ventilation or opening a vent in the structure such as door or window. Furthermore, due to the high quantity of radiation necessary to reach the flashover stage a minimum amount of fuel is necessary to produce it. In addition, the value of the energy generated by indoor fires with a single opening through which the air inlet and gas outflow are channelled can be approximate using the Kawagoe equation [
6].
Currently, fire services use thermal image cameras (TIC) in different types of emergencies related with fire and rescue [
7]. Furthermore it can provide value real-time information during a enclosure fire. However, the prediction of this phenomenon is not easy and on many occasions it could be a problem for firefighter teams who risk their lives in this situations. The focus of this review is on certain types of fires in confined spaces where flashover phenomenon is likely to occur [
4]. In these emergencies, firefighters try to anticipate this situation by reading enclosure fire dynamics (e.g., see
Figure 3) tracks. Sometimes this is not possible due to the rapid response multitasking nature of the emergency and the corresponding stress that these situations generate for the firefighter. Attempting a prevention of the flashover phenomenon described in this paper, the firefighter handling the TIC normally tries to monitor the temperature of the hot gas layer to detect changes in the environment. However, this method is not very accurate because flashover depends on multiple factors such as compartment configuration, fuel and ventilation, among others, and most importantly, the human factor. Another reason is that temperature acquisition with a TIC does not work in the same way when a smoke layer is present. This is because the temperature shown by the camera is the reflection temperature and not the inside temperature of the smoke layer. For instance, carbon particles or water vapour can reflect the visible light interfering in the image acquisition process. On the other hand due to the long wave of infrared light it is not easily reflected by smoke layer particles.
In view of the previous discussion, it is worthwhile mentioning the potential for using AI to predict this phenomenon. For this reason, this paper reviews work which has applied AI in similar situations. Through AI techniques, which focus on the study of intelligent agents, a device can take actions that maximise the chances of successfully completing a task based on environment perception. Nowadays AI techniques, such as flame detection [
8,
9], fire spread [
10], fire and smoke classification [
11], flashover occurrence [
12,
13], thermal interface location in a single compartment fire [
14] or temperature and velocity profiles [
15] among others, have been proven to predict certain fire-related situations. Different approaches have been used to research fire prediction, the most commonly adopted approach is artificial neural networks (ANN). ANN are computing systems usually used to find complex relationships between a source (input) and a target(output). Furthermore, research about fire science including this technology can be found. For example, in image recognition, they might learn to identify images containing fire by analysing example images and using the results to detect fire in other images [
16]. Related to image dataset, convolutional neural networks (CNN) seems to be more effective to find highly accuracy patterns. It is noteworthy to mention that it is very difficult to find a useful data set of real flashover emergencies. To address this interesting point, this paper reviews research that used synthetic data obtained from simulations.
The main contribution of this manuscript is to conduct a review of the literature related to the modelling and simulation of enclosure fires with respect to flashover phenomenon. This is very important to be aware of the transcendence of the phenomenon as well as to analyze how technological innovation can contribute to predict it.
The outline of this paper is as follows. Firstly, research related with flashover modelling and simulation techniques is reviewed. Secondly, the cutting edge technology for comparing synthetic images with real images from thermal imaging cameras is revised. Third, the models that have been developed to predict flashover are reviewed. Fourth, we include a results and discussion section, where the key points of each section are treated. Conclusions and future lines of work can be consulted in the last section.
4. Flashover Occurrence Prediction
Regarding the prediction of these phenomenon, most of the relevant studies shown below have been carried out using neural networks. Moreover, ANN has been used in research cited in this review to predict some physical aspects related to fires, even to predict the flashover phenomenon but not using TIC.
Research to perform flame detection using neural networks was conducted by Huseynov et al. [
8,
9]. Joint time-frequency analysis (JTFA) was used with a reduced time Fourier transform to extract relevant information. This study focused on the domain of the signal frequency, since the signals that a fire source emit can vary in amplitude and function depending on the distance, angle, presence of obstacles and other sources that are not flame generators (sun, wind, random modulation, bright lights, rain, fog, dust, etc.) that may look like flames in the eyes of an infrared sensor. In contrast to the time domain, the frequency of flickering of a flame remains relatively independent of the different environmental conditions. Due to the amount of several false positives that can occur, it was considered to use a multiple neural network instead of a large scale. This model was mounted on a digital signal processor (DSP, Texas Instruments (TI) F2812), reaching the conclusion that the response to the detection of a flame of this model is equal to or greater than some of the current detection systems. Another approach is the one given by Won-Ho Kim et al. [
52]. His research is based on prediction by studying the neighbouring pixels of a video, provided by a TIC. The author proposes a flame detection system based on an algorithm optimised with a digital signal processor to optimise detection time, between 5 and 20 s.
Predicting parameters in compartment fires is an extremely fast alternative approach. In this line, several ANN’s have been developed to do it instead of CFD techniques, which need more computational time to do the same work. In these cases, CFD techniques were used to validate ANN. In the following research, a novel ANN model GRNNFA, which is a fusion of the Fuzzy Adaptive Resonance Theory (FA) model and the General Regression Neural Network (GRNN) model, was developed to predict the locations of the thermal interfaces by Lee et al. [
14]. To train and evaluate the ANN, experimental data obtained from Steckler et al. [
24] were employed in a total of 55 experiments conducted by Steckler (with different fire locations, fire intensities and window and door sizes). Due to the diffusion and mixing effects of the fluid flow, the thermal height could not be determined precisely. For these cases, FDS was used to validate the ANN performing a simulation for the experimental data. When the simulations were compared with the experimental data, the authors noticed that FDS consistently under-predicted the experimental results of the thermal interface height. For this reason, the absolute difference was averaged, and this correction factor was applied for the neural network training. The predicted results by GRNNFA of the 55 cases were compared with the simulations results. It was found that the GRNNFA was able to predict the locations of the thermal interfaces well within the range envelope of the experimental results up to 94.5% accuracy (i.e., only 3 out of 55 samples was predicted outside the range envelope). The previous model was applied to five test cases that not appear in the experiments and the results were compared to FDS simulations. Subsequently, it was found that the difference between the predictions and simulations are well within the minimum error range. The key issue here is GRNNFA was able to capture efficiently the genuine, predominant characteristics of the fire phenomenon within seconds from limited samples. This ANN was applied to predict the flashover phenomenon in enclosure fires [
13] based on the compartment geometry.
Regarding flashover prediction, a new ANN model was developed by Lee et al. [
12] using probabilistic mapping with maximum entropy (PEmap). It was employed as a binary classifier for predicting this phenomenon occurrence in single compartment fire and results were compared and verified against data obtained by Fuzzy ARTMAP (FAM). A computer package (FAST [
53]) was used to introduce a flashover criterion (the temperature of the upper hot gas layer ≥600
C) to predict flashover. Due to the impossibility of getting enough data of real emergencies, the training of the model was carried out through computer software simulations. Specifically, 375 simulations (190 samples of flashover and 185 non-flashover) were carried out to train the PEmap model, varying randomly different parameters of the compartment:
Length of the compartment (varies randomly from 2 to 10 m)
Width of the compartment (varies randomly from 2 to 10 m)
Height of the compartment (varies randomly from 2 to 10 m)
Maximum heat release rate (varies randomly from 10 to 6000 kW)
Subsequently, for different combinations of these room dimensions and the maximum heat release rate the occurrence of flashover was determined by FAST. Albeit the prediction results show that PEmap is an efficient classification tool for determination of flashover with a high degree of accuracy (96.8%), the neural network model was not implemented in any device, such as a thermal camera, to be tested in a real case.
In a previous study [
14] the GRNNFA model was applied to predict the thermal interface location in a single compartment fire. The performance of this ANN was proven to be comparable of the CFD model, moreover, computational speed of the GRNNFA was faster than CFD model. GRNNFA presented some limitations as it restricts the results in the application area of the model. In addition, Yuen et al. [
15] presented a modification of the original GRNNFA model for multi-dimensional prediction problems and it was used to predict the velocity and temperature profiles at the centre of the doorway in a single compartment fire. GRNNFA model was successfully applied to predict the velocity and temperature at the centre of the doorway in a single compartment fire and multi-dimensional environment. There were some problems with the application of this technique in the local region because the training samples in that region may not be sufficient to describe the general behaviour of the system. Finally, this model was applied to predict the flashover under the given fire parameters (length, width, height and maximum heat release rate). It is concluded that this model is statistically superior than the Fuzzy ARTMAP and PEmap models [
54]. In a real flashover situation, it is not possible to introduce parameters of a confined space in a device like TIC. The reason is that not only is the speed of events one of the crucial factors, but also the impossibility to know the exact dimensions of the enclosure. Image recognition techniques could be a useful technology to solve this problem. Another approach to predict the location of the thermal interface in a fire compartment (HTI) is studied in the work of Lee et al. [
55]. A neural network model based on probabilistic entropy (PENN) was used as an alternative to the CFD technique (computer simulation). Using this technology the network requires less computational time for a case study. PENN model is applied to the prediction of the HTI within a confined fire using real experimental data. The statistical analysis of the results positively demonstrates the accuracy of PENN model in a confined fire. Moreover, the computational time incurred in the prediction of the HTI is much better than CFD techniques. On the other hand, authors clarify that although the use of this technique is encouraging, new lines of research are needed to deepen the determination of the initial kernel radius, because these have been determined empirically for this study.
In a fire emergency time is crucial to prevent a dangerous situation. As a consequence, predictions should be done in real time. The following researches include real time applications techniques to predict certain situations related to confined fires. A small world network model to predict real-time fire spread on board naval vessels was developed by Kacem et al. [
10]. Despite the fact that the purpose of this research is not the flashover prediction, it was taken in account to evaluate the fire compartment and decide if it will be crucial for fire spread to other enclosure. To determinate the occurrence of flashover the pyrolysis rate in the compartment was evaluated and the duration of the fully developed phase is calculated using empirical formulas for a fuel limited combustion and ventilated limited combustion. Finally, full-scale experiments in a multi-compartment fire container with mechanical ventilation and CFD simulation were conducted to validate the model and determinate the probability of spreading a fire in a specified configuration naval vessel. Despite pyrolysis rate is considered to predict the occurrence of flashover it cannot be considered for predicting these situations with TIC. Related to TIC Kim et al., proposed a method [
56] for fire heading in smoke-filled indoor environments using thermal imaginary designed for SAFFiR (Shipboard Autonomous Firefighting Robot). It is based on a previous study [
11] in which it was used a probabilistic classification for fire, smoke, their thermal reflections and hot objects using Bayesian theory. The key point is the fire and smoke classification, a clustered-based image technique is used to discriminate the two classes foreground (objects) and background, resulting in fast computation during real-time implementation. Despite this method being used inside an enclosure, another approach classifying smoke from an outside compartment with a thermal camera could provide some relevant information about the fire development. In a recent study [
57] Conditional Generative Adversarial Networks were used to predict rapid fire growth (flashover) in real time. There are some visual signals that can warn firefighters about a possible flashover, like dark smoke, high heat and rollover. A standard body camera was used to analyse the colour video stream. Very dark fire and smoke patterns were enhanced applying generative adversarial neural networks. Previously image-to-image conversion techniques were applied using conditional adversarial networks in order to obtain a thermal image from a regular colour image. The neural network training and test was carried out using 30 and 10 videos provided by two different fire departments. As a result flashover was predicted as early as 55 s before it occurred. While this technology was developed from a colour image authors remarks it can be applied over thermal images from TIC’s to predict flashover situations. Currently training and test datasets from real situations are limited. Only is possible analysed a few cases with simple configurations like fire containers. CFD software could be an interesting tool to generate datasets with different configurations and fuels in order to train neural networks from TIC’s images.
Table 6 provides an overview of the references related with fire prediction cited in this section. The reader can detect at a glance if the reference is related with flashover phenomenon, which are the most important techniques used in the research, the software used for the simulations, infrared sensors and TIC employment.
5. Challenges An Opportunities
Novel techniques used for modelling, simulation and prediction of flashover phenomenon in confined fires have been reviewed in this paper, identifying future opportunities and challenges. As far as modelling and simulation is concerned, the most recent studies of flashover have been reviewed. While the articles analysed describe a deep level of knowledge about this phenomenon, there are some issues that require further exploration.
Flashover research reveals some physical events which determine the beginning of this phenomenon, especially upper gas temperature ≥ 600
C and heat flux at floor level ≥ 20 kW/m
. Both factors have to be considered together to define the initial point of this transition phase to fully developed fire. Furthermore, ventilation factor has a vital impact on flashover occurrence. The location and size of the vents in the enclosure maintains a close relation to the development of this kind of phenomenon. Vent identification using computer vision techniques [
7] in real time could provide valuable information to predict flashover. Furthermore, flashover is clearly determined by an HRR/time relation, which could be a key issue for prediction with ANN models.
The simulations reviewed from mathematical models reveal that nowadays computational fluid dynamics (CFD) simulation predominates over other analysed models. The detailed analysis of thermal environments, which can be used to train and validate ANN with certain accuracy, is one of the many advantages of this type of simulation. Various CFD software packages have been compared in different articles shown in this review. Notice in most cases FDS have been used to conduct research simulations. Not only have been FDS validated with acceptable scores but also ANSYS CFX software obtained similar results. In particular, the use of FDS is widely extended because it is a free and open-source software tool. Furthermore, flashover phenomenon has been widely simulated in most of the articles reviewed.
With regard to TIC, it is important to remark that according to the reviewed articles it is not possible to compare images from FDS simulations with thermal images obtained directly from TIC. Some transformations need to be done before this step, since the final image depends on the technical thermal camera parameters and conditions of the medium (scattering or non-scattering medium). This technique could provides the opportunity to compare images taken from emergencies real cases, where there are not sensors inside the enclosure, with images from simulations. A better approximation of simulation to the real case can be achieve which can be of great help for the fire investigation or firefighters training.
As for predictions using AI, this review shows that there are different lines of research that cover flashover. Different prediction models can be observed, with the predominant lines oriented to AI and neural networks. ANN models have been developed to predict fire, the location of the thermal interface in a fire compartment (HTI) or, when the flashover occurs. Furthermore, to the best of our knowledge, no document has been found that discusses AI in TIC to predict cases of flashover using datasets from CFD software. In fact, data sets used to predict flashover occurrence were based on simulation sensors information located inside the enclosure. For this reason, the identification of characteristic parameters which define a flashover situation using ANN trained with data from CFD software could be a big challenge. This technique could be used in other combustion processes in which real cases are very difficult to reproduce and could be interesting to predict a specific situation.
6. Conclusions and Future Work
Regarding to flashover phenomenon modelling, real complex scenarios with different vents and fuels have not been evaluated. Reduced-scale experiments with a simple configuration are used to study this kind of situation. However, to determine the occurrence of flashover, not only do gas layer temperature and heat flux radiation at floor level need to be considered but also ventilation factor and HRR/time. For this reason, other configurations of the enclosure should be analysed. In conclusion, flashover is well simulated using different CFD software.
The articles reviewed suggest that it is possible to obtain images from CFD software that are similar to a specific thermal imaging camera using its technical parameters. This should be taken into account for developing artificial intelligence technology based on TIC if CFD data want to be used to train ANN’s. Programming and computer vision techniques can be used to conduct the mathematical calculations to achieve the final result.
Some research regarding flashover prediction using ANN was identified but to the best of our knowledge, research that predicts this phenomenon with TIC’s using datasets from CFD software in order to train ANN’s have not been found. Furthermore, thermal image cameras may be useful devices to predict flashover by fire services using ANN. Considering the research analysed, new prediction models can be developed focusing on real-time thermal images.
Taking into account the different areas studied and analysed in this review we conclude that flashover phenomenon is sufficient studied to be modelled and simulated using simple configurations. Furthermore, AI techniques have been used to predict its occurrence using sensors placed inside the enclosure. However, there are not any research which use data from CFD software to train ANN with flashover prediction purposes in TIC’s.
We envisage future research on predicting flashover phenomenon with TIC’s using datasets made from CFD software to train ANN’s. This technology can prevent dangerous situations by warning firefighters, who handle thermal cameras, in real-time. Gas layer temperature, heat flux radiation at floor level, ventilation factor and HRR/time in the enclosure should be factored in. CFD techniques may be used to train the ANN model, and real experiments in a fire container may be a useful way to validate the final model. Furthermore, the study of this phenomenon using AI can help to delve into certain aspects that may not have been considered previously.
As future work, we plan to use other parameters from field models in order to create a dataset with thermal vision characteristics to train ANN models based on TIC. This will allow a breakthrough for the simulation and prediction of the flashover phenomenon.