FIRE DETECTION SYSTEM FOR URBAN STRUCTURES WITH INTEGRATED USE OF MULTISENSORS IN EMBODIED TECHNOLOGY

Objective: Create a sustainable, low-cost and effective fire detection system (FDS) for urban buildings, based on embedded technology, as an alternative to traditional detection systems. Method: experimental research that used the hypothetical-deductive method, with assemble hardware using the ATMEGA328P microcontroller and the MQ2 gas sensors, MLX90614 temperature and LM393 flame. When building the software, a qualitative model was created, based on everyday observations, associating it with a logical expression; Then, a quantitative model was created based on experimental data subjected to computational analysis using the decision tree method, defining a heuristic for early detection of fire patterns with precision. Results and conclusions: The suggested FDS had a shorter response time than the red ampoule sprinkler and an accuracy of 94%. This system proved to be more agile than common detectors, reducing water and energy consumption when fighting fires and the amount of resources used to replace affected assets. Furthermore, it has a low cost, making it accessible even at a residential. Implications of the research: It can be said that this research represents an important step in the construction of sustainable FDS, bringing social, economic and environmental benefits; and efficient due to the use of computational data analysis.


INTRODUCTION
The fire is a historical problem of the cities.Since the beginning, they have brought losses, motivating actions to control it (Silva, 2014).Ancient peoples, such as the Greeks and Romans, constituted lookouts and guards to inform citizens about the fire and to act in fighting the fire.These measures were in place until the early modern period (Pirenne, 2014;Ward, 2020).
In the 17th century, it began to look at prevention, creating laws that defined the constructive materials to the cities, like the English Law of 1667 (Ijeh, 2016).With the technological advance, there was growth of cities by urban density, increasing the quantity of fuels in the same place and the probability of fires (Negrissolo et al., 2019).
In the 18th and 19th centuries, SCI-related technologies and institutions emerged, such as sprinklers, standards bodies such as the National Fire Protection Association and certification laboratories such as the FM Approval (FMA, 2023;Grant, 2020;Richardson, 2003).Today, SCI aims at prevention and control by Fire Safety Systems -SSCI (Grant, 2020).
Despite the evolution of the SSCI, their absence or failure still causes fires, as there was in the Grenfell Tower in 2017 (Kirkpatrick et al, 2017), which caused human and economic losses; what occurred in the National Museum -Rio de Janeiro, in 2018 (Fioravanti, 2019), and in the Cathedral of Notre-Dame, published in 2020, (2) Such tragedies show that it is necessary to continue the search for technologies that improve such systems.
In addition to preventive and combat actions, the SCI must aim at sustainability in the use of resources (Gollner et al., 2012), since the pressure on them has grown in the face of the exponential increase in population between 1800 and 2021 (UN, 1995b apud U.S. Census Bureau, 2004); do mass consumption (Colombo et al., 2008);and & Roser, 2018).
Faced with the unsustainability of the current development model in the world context, the UN published in the Report "Our Common Future", the term Sustainable Development -DS, defining it as the one that seeks to ensure the provision to the needs of the current generation without affecting the care to the needs of future generations (UN, 1987).Of the actions to make it possible, one can highlight the promotion of technologies to optimize the use of resources and reduce consumption.
At the UNEP and the Conference on Sustainable Development (Rio +20), the concept of "Green Economy Initiative" (UN, 2012) was brought in place of "eco-development" (UN, 1972;UN, 1992).This Initiative is made up of productive processes that target the DS (Bang et al., 2023), and one of them is composed of Green Technologies, which are responsible for reducing environmental, economic and social impacts (Dastbaz et al., 2015).
Embedded technologies are used to improve usability and efficiency, and if they optimize resource usage, they are considered Green Technologies.Studies in the area of SCI have created methods of monitoring and detection using these technologies, avoiding accidents or reducing combat time due to rapid control, decreasing resource use (Silva et al., 2014).
Another issue is the high costs for acquiring and maintaining the SSCI in Brazil, for this reason, they are not adopted in homes and companies.The point is that from 2018 to 2020, structural fires accounted for more than 40% of the total (Morato & Motta, 2021).Faced with the need to adopt technologies in SDI in order to achieve sustainability and make SDI universal through cost reduction, the question arises: could a low-cost and effective SDI in fire detection be built?
The objective of this research is to constitute a low cost SDI, effective for urban buildings, based on embedded technology.This is done by: building an SDI with ATMEGA328P microprocessor and sensors of smoke, temperature and infrared; setting up a heuristic that perceives the occurrence or not of fire; and making burns with paper and MDF in a controlled environment to adjust sensors and parameters necessary for the operation of the SDI algorithm.

THEORETICAL FRAME
Fire is a combustion reaction due to the oxidizing of the fuel by the oxidizer when receiving heat, with water and carbon dioxide or monoxide among its products, releasing heat and flame.By propagating in time and space, without control, fire is said to be a fire (Chang & Goldsby, 2013;Seito et al., 2008).

Conventional Fire Detection and Fire Alarm Systems (SDAI)
The smoke concentration and variations in temperature and flame intensity are linked to the fire, so detectors are based on sensitivity to one of these variables.The SDAI has as its functions: to monitor variation of a certain property by a heuristic that associates the measured value with a predefined situation; to detect the occurrence or not of fire depending on the measure exceeds or not a limit; and to control the loss by data transmission to the execution of protocols.
SSCI items are said to be passive, if they do not act directly in combat, but aim to reduce the speed of the propagation of fire, like compartmentalization (Brentano, 2015;Negrissolo et al., 2019).The items of ative protection, on the other hand, act directly in combat, preventing the propagation, like extinguishers, automatic showers, alarms and detectors.
SDAIs have central alarm elements, audio-visual flag, automatic trigger and detector, and manual triggers.The plant collects the data and, if it confirms the fire, activates the flags.However, the manual triggers serve the cases in which the fire is perceived before the activation of SDAI and sprinklers (ABNT NBR 17.240, 2010;NFPA 72, 2019).
The main detectors are smoke, temperature, multi-sensory and flame detectors.The smoke detector measures variations in the concentrations of combustion gases, so that if limit values are exceeded, the fire is allowed and the sensor is activated.The temperature detector is sensitive to its variation, being thermal, when the variation indicates fire, or thermovelocimetric, when the speed with which the temperature varies indicate it (Brentano, 2015;Negrissolo et al., 2019).
The multi-sensory detector has smoke and temperature sensors, sending the data to the fire station, which receive heuristic treatment described by algorithm to guide decision making.The flame detector is sensitive to radiation from the flame spectrum, and is used in wide or open locations, where there is no obstacle to preventing the reception of radiation, and the smoke and heat dissipate (Brentano, 2015;Negrissolo et al., 2019;Seito et al., 2008).

Embedded Technologies Employed in SDI
Conventional SDIs do not act according to distinctions between the environment and fuels, and most of them are reactive, as they only respond after a long time after the start of the accident.To refine them, technologies have been used that enable them to act under dynamic response strategy, bringing performance advancement, resource optimization, intelligence insertion, making them apprentices.

Embedded technology detection system
Projects for fire monitoring, detection and control with such technologies have used the ATMEGA328P microcontroller to analyze certain quantities as parameters, each measured by a certain sensor.The literature shows several practical cases for fire detection, based on several heuristics and strategies, and among them can be cited the use of binary logic (Rashid & Rafid, 2018), up to the use of neural networks (Muhammad et al., 2018).
In Kenya, an SDI was made with the ATMEGA328P and sensors under a fuzzy logic, with efficiency of 83% (Obanda, 2017).In Senegal, an SDI was created that had ATMEGA328P as a control unit, with four flame sensors connected to it, and through the use of binary logic a detection dynamic was instituted by which only after simultaneous sensitization of the sensors the solenoid that linked the sprinklers was triggered (Rashid & Rafid, 2018).
In Brazil, an IoT-based SDI was created, linked to an ESP32, sending data over the internet to centers and users (Muenchen, 2018).In Indonesia and Malaysia, an SDI was made under fuzzy logic, with a MQ-2 sensor connected to ATMEGA328P (Labellapansa et al., 2019).In Tanzania, an SDI was created for detection with smoke, temperature and humidity sensors linked to ATMEGA328P, sending data via Wi-Fi (Lutakamale & Kaijage, 2017).Such studies show that the theme is usual, and such detectors are already predicted in standard (NFPA 72, 2019).2.2.2 Proposed detection system 2.2.2.1 Hardware Related Concepts ATMEGA328P Microcontroller -consists of a control unit that has high performance and low cost, having memory for storage and pinouts for data input and output.Presents a large number of users due to the high degree of maturation that possesses Mendonça & Zelenovsky, 2019).The pinouts that make up its physical structure are represented in Figure 1, and its attributes, in its datasheet (Datasheet, 2023).Sensor LM393 -It is sensitive to radiation contained in the infrared range, wavelength between 760nm and 1100nm, detecting heat sources and flames.In their presence, the digital pin indicates state 0, and in the absence, state 1.The electrical layout of the sensor is described in Figure 2a and its and properties are contained in its datasheet.
MQ-2 Gas Sensor -Measures the concentration of smoke, hydrocarbons and alcohols, such as LPG and butane, with optimal sensitivity.It works on the basis of the action of an internal heater next to the strange electrochemical sensor (SnO2), of low and constant conductivity in contact with the air.When detecting such gases, it varies in proportion to the concentration.The electrical circuit of the sensor is present in Figure 2b, and its characteristics, in its datasheet (Datasheet, 2023).
Temperature Sensor MLX 90614 -It is a remote and punctual reading thermometer made by capturing infrared radiation.As for the limit angulation for sensory perception, the value is 60º, so that the front of the sensor the reading accuracy is 100%, while the 30º to the left or right of the axis perpendicular to the plane of the sensor, the accuracy is 50%.The electrical circuit is explained in Figure 2c, and its characteristics, in its datasheet (Datasheet, 2023).When building software you need to use good strategies to solve the problem.One of them is binary logic, which admits two conditions to an event: if it does not occur, the logical value is (0); and if it does occur, it is (1) (Silva et al., 2017).One defines a model with such logic by simple proposition, analyzing a single variable; or composed, evaluating more than one, and by means of expression found by logical operations, deliberate whether the event occurs or not (Alencar Filho, 2017).
To model an event, qualitative or quantitative analysis is used.The first is based on assumptions based on the observation of daily practice, and from them, it is defined whether the occurrence of the event is confirmed (1) or denied (0).The second is based on numerical and dimensionable data, obtained by standardized methods, such as questionnaires and experimental tests, using systematic observations to establish the results (Silva et al., 2017).
In this study, two models were created to confirm the occurrence or not of the fire, the first being founded under qualitative analysis, and from it, the second was made, under a rule fixed by computer learning process, which has been used in the various areas of science, to indicate data patterns and analyze the process itself, by a set of actions, identifying a model to the causal relationship between input and output data (Alpaydin, 2020).
In machine learning, one seeks to compose tactics to give the object a language of learning, so that it assimilates how to act in each experience, according to distinctions made in the software, from a computational intelligence (Russell & Norvig, 2021;Ted et al., 2023¹).Of the strategies for effecting such language, the tree method is one of the most employed, being such tool capable of visualizing the structure of the problem by systematic analysis of the variables involved in it (Silveira & Bullock, 2017).
The tree method classifies data, subjecting them to a series of questions, with subsequent questions depending on previous answers, and at the end of processing, classifies the data with a high degree of accuracy (Okada et al., 2019).There are programs used to analyze variables from this method, which are high-performance software for solving mathematical problems, from calculus and numerical analysis (Vieira & Moraes, 2013).

METHODOLOGY
This research adopted the hypothetical-deductive method, with inquiries coming from deficiencies in the use of existing SDIs, basing the problem, hypotheses with expectations that will be corroborated or frustrated, depending on the results.It is bibliographic in nature, because it is based on technical-scientific productions; it has a quantitative aspect because it aims at quantifiable results, based on data that need to be measured to be fully understood; and it is experimental in making use of laboratory tests to survey the data and establish the results (Gil, 2022).
The tests of the SDI were carried out at the premises of the Materials and Structures Test Laboratory of the Federal University of Paraíba (LABEME -UFPB), and the experiment was divided into three stages: (i) assembly of the hardware that will form the structure of the SDI; (ii) construction of the software responsible for the implementation of the heuristics; and (iii) instrumentation and definition of methods to be adopted for carrying out the calibration of the SDI and validation of results.

Hardware Assembly
In the Hardware assembly, the following devices were used: 01 (one) temperature sensor MLX 90614, 01 (one) gas sensor MQ-2, 01 (one) flame sensor LM 393, 01 (one) led light of 5 mm red color, 01 (one) plate with Microcontroller ATMEGA328P, 01 (one) protoboard of 400 points, 201 a) USB cable (A/B) -50 cm.The sensors and the lamp were connected to protoboard, and this one, connected to ATMEGA328P.

Building the Software
The software was written in C to be used on a ATMEGA328PMicrocontroller. Initially, the programming was done for preliminary tests, with random limits of sensitivity to the sensors, and for values above the thresholds, it was considered to be fire.The sensors' responses to natural and artificial light radiation (led light bulb, gasoline burning, paraffin and paper sheet) were analyzed, and in the face of inconsistencies, by misreading or interference, means were created to remedy them.
The flame sensor has a photodiode that captures the radiation and generates current that varies according to intensity, which reaches the ATMEGA328P as analog signal, being converted into digital, and can take on 1024 values, for having signal-input ratio of 10 bits¹, which take on values 0 or 1.A dimensionless unit scale (a.u.) ranging from discrete values of 0 to 1023 has been adopted, with the value 0 attached to the lowest intensity, which occurs at wavelength 1100nm, and the value 1023 is associated with the greatest intensity that happens when the wavelength is 760nm.
The gas sensor also has a 10-bit signal-input ratio, and the digital pin can take 1024 values, and a dimensionless unit scale (a.u.) ranging from 0 to 1023 is adopted, with each value associated with a gas concentration value, 0 being associated with the lowest concentration value read by the sensor -200ppm -and 1023, associated with the highest value, which is 10,000ppm.
For the MLX 90614 sensor, it was not necessary to establish a parallel scale, because it has a processor that, when connected to the ATMEGA328P, converts the digital signal received into analog, sending the information already converted into continuous value within the original temperature scale.
The modeling of SDI was done under binary logic, where the logical value (0) indicates no fire and (1) the occurrence.The values obtained by each sensor were classified by binary states, 0 being the non-activation of the sensor and 1 indicating that the measured value exceeded the parameter value, triggering the sensor.From the possible composite propositions, it was defined whether or not there was a fire.In Table 1, the binary states of each sensor are described.Establishing a qualitative model based on assumptions derived from daily observations, a truth table was made containing in the first three columns the logical values of the sensors in the possible cases and in the last, showing when there is fire (1) or not (0), from the combination of the logical values of the lines, only being admitted the condition of fire, when no possible situation is found with these values in which there is no fire (Table 2).In the first and last line are the trivial solutions.In the first line, all sensors assume value 0, which means that the measured quantities are at lower levels than the cut-off level, deducing that there is no fire in this situation (0).In the eighth line the opposite occurs, the quantities passed the cutting levels, and therefore indicates that there is fire (1).
In line 2 only the temperature exceeds the cut-off limit, but such behavior may be due to heating of an ironing or welding iron, but fire is unlikely (0) as the flame and gas levels are below the critical value indicating the occurrence of the event.
In line 3, there is an indication that the level of smoke is above the limit, but with values of temperature and intensity of flame below the thresholds, an attribute inherent to the fire.Therefore, the increase in the level of gas or smoke may be associated with other factors, such as a stove mouth with the gas flow open, but possibly not due to fire (0).
The flame sensor is below the cut level in line 4, and although the gas and temperature sensors are above the threshold, it does not mean that there is fire in these conditions, because it is possible that in this scenario is occurring, for example, heating a welding apparatus tip, with the release of gases or smoke, even if there is no fire (0).
In line 5, the intensity of the flames is above the cut-off level, but there is no elevation in temperature and concentration of gas or smoke to levels peculiar to the fire, so the flame sensor may be detecting a light that emits a large amount of radiation in the infrared range, for example, but it is unlikely that there will be a fire (0).
The flame and temperature sensors on line 6 indicate that the measured values are beyond the established limit.In this situation there is likely to be a fire, as variations in high levels of thermal and light energy are inherent during this event.Thus, the configuration can be classified as a fire (1) and the low level of smoke can be justified either by the fact that the burning is slow, or due to the combustion releasing little smoke, the environment not being saturated.
Finally, on line 7, the luminous intensity and gaseous levels are beyond the cut-off limits.However, without elevating the temperature above the threshold values, it is not possible to characterize this situation as fire (0), since the elevation of the temperature to these levels is intrinsic to fire.
With all the data gathered from the possible combinations of the simple propositions, it is possible to perceive that there will only be confirmation of fire in two cases of the eight possible: in the 6th line and in the 8th line (101 or 111).Using a Boolean simplification method, one can write the logical expression representing table 2 as being I = C˄T.
The model found, despite having a certain degree of precision, as a result of qualitative analysis, can be improved.This is why a quantitative model was made, motivated by experimental data, using the machine learning decision tree technique to obtain the rules that make the SDI respond in the simplest and most assertive way if there is fire (1) or not (0).

Physical simulator
Structured in masonry, in the shape of a cube, with 1m³, simulating a building room, on a scale, using cellulosic material as fuel.It has a door of 0.27m x 0.70m and window of 0.33m x 0.33m, both closed with zinc plate.The door is cut quadrangularly at half height, so that the burning is seen during testing.On the ceiling, there is a cut in the structure to position the SDI.Finally, a ½-diameter steel pipe was placed to secure the sprinkler.
To size the structure, it was observed specifications of the Technical Instructions of the Fire Department of the Military Police of the State of São Paulo (CBPMESP, 2019).For this reason, the simulator was made with 1m 3 , designed for burns of small burdens of fire, with height of 1m, corresponding to one third of the right foot of the majority of urban buildings (Figure 3).

Sensor calibration
Phase composed of 06 burns: 03 with A4 paper as fuel, and the others with MDF.To measure temperature and mass, 02 analog thermometers (0° to 350°C; 5°C) and 01 digital scales (5g to 8200g; 0.1g) were used.As they are predominant in urban buildings, cellulosic materials were preferred.The mass for burning was defined according to IT 14/2020 (CBPMESP, 2019).
In the firing phase, the cycle started, putting the fuel in the center of the simulator.Before it, the processing of the programming began, if any incoherence was found, the connections were checked, and if the problem was not resolved, the program was closed, the programming revised and tests were carried out with the sensors.Once the coherence of the data was confirmed, the SDI kept on capturing it and combustion was started.The ignition source was a small alcohol wipe lit with phosphorus.
When combustion started, the data was captured every five seconds.Burning continued until temperature saturation, between 80ºC and 100ºC, since the fire charge was not sufficient to reach higher temperatures, and burning was interrupted after reaching its maximum stage of development.The temperature was monitored by the use of analog thermometers installed in the ceiling and at the bottom of the simulator, ending the cycle when it reached room temperature.
Once the burns were closed, the acquired information was stored in .txtformat and imported from the serial monitor to .csvformat, to package the data in Excel.For each set of burning data, a total of four (04) graphs were constructed, with each burn represented by a color.
The first graphs bring the variation in concentration of smoke and luminous intensity with the passage of time, with the quantities ranging from 0 to 1023, on a scale of dimensionless unity and time in seconds.The third, brings a variation in temperature with time, with temperature in Celsius and time in seconds.The last chart shows the state of fire (1) and not fire (0).
Data on these states was collected by observing the development of combustion.Burning was only understood to be a fire when there was the propagation of fire in time and space.With the inchoate flames and smoke and temperature at low levels and without successive variations, before or after the full development of the fire, no fire was considered.
From the graphs, the speed of burning and the response of each sensor as to the material were analyzed, whether it was the response time or the level of sensitivity with which the sensor received the data.The occurrence or not of the fire according to visual inspection and the data obtained in the burns were submitted to computer analysis, so that after processing, by the use of the tree method, the best cut limits and solution strategies were fixed for detecting the fire.

Validation of results
Among the common detectors, the sprinkler of red ampoule was used to be compared with the proposed SDI, being connected to a wet tube Automatic Shower System, with breakage expected to succeed at 68°C.In this phase, two burns occurred: (i) the preliminary one, to remove the humidity from the environment and to test the level of reading of the sensors; and (ii) the final burning, to validate the results, using in the programming the parameters obtained in the computational analysis.
Based on these new parameters, using the decision tree method on the data provided by the SDI and observing the boundary conditions of this study, a high degree of assertiveness was obtained.The last burning of this phase was carried out to compare the response time to the fire given by the sprinkler and the response time to the same situation, offered by the SDI in the particular practical case studied in this research.

Preliminary Phase
This is the phase in which adjustments were made to make the sensors capable of capturing the data with the minimum of interference.The flame and temperature sensors needed to undergo certain adaptations in order to be able to resolve some issues related to the accuracy and interference suffered by them during the reading process.
Certain non-fire radiation sources (solar radiation and artificial light), indoors and outdoors, interfered with the response of the flame sensor.To mitigate these noises, a plastic tube was placed wrapped in black insulating tape so that radiation would not pass through the side walls, and at one end of the tube red adhesive tape was placed in front of the sensor as a filter for the radiation of the red and infrared spectral bands.
As for sensor MLX90614, its datasheet shows that for a 90º viewing angle, there is loss of sensitivity as the heat source distances from the orthogonal axis r to the sensor plane.At a distance of 1.0 m from the heat source, this sensor receives the average temperature in the circular region of radius 1 m² in the plane perpendicular to the r axis, causing discrepancy between the measured value and the temperature of the heat point (Figure 4a).
To solve the problem mentioned above, a waveguide of 7 cm in length and 1 cm in diameter, made of aluminum, was put on top of it by a polytetrafluoroethylene resin tape, which is a thermal insulator (Wang et al., 2017).This layer of tape was arranged with the purpose of thermally insulating the metal from the ambient air, so that this heat exchange did not interfere with the accuracy of the reading performed by the sensor (Figure 4b).Thus, despite the reduction of the sensor's scanning area, the sensor had sensitivity to heat in a region of 15 cm in diameter (Figure 4b), which proved sufficient for any heat source in the region to be perceived, being remedied the inconsistencies in the measurements.

Calibration Phase
At this stage, the data was captured by sensors in a situation similar to fire, and by the decision tree method, ideal cut-off thresholds are constituted to obtain a solution strategy that, by simplified heuristics, exhibited a high degree of effectiveness.The cut-off values used in the calibration tests responded well to the empirical conditions of the preliminary phase, with these values in the gas, flame and temperature sensors equal to 80 au, 500 u.a. and 60 ºC, in this order.
Three burns were made with each fuel, collecting the data from each burn and constructing a figure with four graphs alluding to the burns of each material, and having each graph, three curves: blue (1st burn), red (2nd burn) and green (3rd burn).The graphs show the variation in the values measured by the sensors with the passage of time, except for the fourth, in which the value 0 was set for visual non-perception of the fire and 1 for its visualization.
The graphs in figures 5 and 6 show, in this order, the burning curves with paper and MDF, showing the values of the readings made by the sensors and the evolution of the fire as a function of time.The fire develops (1) and then the environment returns to normal (0), with increased gases, decreased temperature and absence of flame.Fire development and sensor responses have similar behaviors at the 1st, 2nd, and 3rd burn, respectively.

Results of paper burns
It is noted in Figure 5 that after the time lapse for preparation of burning, it started, there being little smoke, with subtle change in MQ-2.Due to the low flame, LM 393 was barely sensitized.But the least sensitive sensor was the MLX 90614, with almost no variation.When viewing Figure 5, it is noted that at the first burn (blue curve), 10s from the start of the fire, the flame sensor reaches 800 u.a. and the temperature, at 60 ºC; at 15s, the MQ-2 passes 900 u.a., being the fire detected.In the second burn (red curve), the LM 393 and the MQ-2, after 10 seconds of the start of the fire, reach the thresholds, while the MLX 90614 only passes 60°C after 35 seconds.In the third burn (green curve), the sensors pass the thresholds after 10s.So it can be said that in the burning of the paper, LM 393 first detected the MQ-2, and finally the MLX 90614.

Results of MDF burns
According to Figure 6, before the start of each burn there was a time lapse due to the preparation for the tests.Then the combustion began, a phase in which there was little smoke and flame, besides an extremely low variation in temperature.Figure 6 shows that 25 seconds passed from the beginning of the first burn (blue curve) without significant variation in the levels of measurements made by the sensors, when the fire is then characterized.After 20s of this moment, the flame sensor passed 800 u.a.Within 30 to 40 seconds of the start of the fire, the temperature sensor exceeded 60 °C, and only 40 seconds after that, was the gas sensor recording the cut-off value.In the second burn (red curve), after 10 seconds of the start of the fire the MQ-2 had crossed its threshold, and with more 5s the other sensors passed their limit values.Finally, on the third burn (green curve), the temperature sensor exceeded 60ºC after 10s of the fire had started.After another 5s, the LM 393 passed its threshold, and after another 10s, the MQ-2 exceeded 80 u.a.In general, for MDF burns, the flame sensor was the first to reach the threshold, followed by the temperature sensor and finally the gas sensor.
Under the binary logic for determining the occurrence or not of the fire, a result was arrived at based on the visual inspection carried out by an observer and another arising from each numerical tuple formed by the values measured by the sensors at a certain moment.The data was submitted to computational analysis, and new cut parameters were obtained for each sensor, optimizing the percentage of agreement between input and output data and reducing false positives and false negatives.Of the solution strategies expressed in the computational analysis, it was indicated to give a high level of assertiveness to the SDI and to have a simple programming logic, made on the basis of a decision tree with the minimum of leaves.Thus, the programming was redone from the new cut-off points, obeying the logic expressed in Figure 7.

Validation Phase
At this stage, two burns were carried out: (i) the first of them called preliminary, which had as its main objetives the reduction of the humidity of the environment and the checking of the calibration parameters of the sensors; and (ii) the second was done for consummation of the test, already implemented in the SDI, programming done on the basis of the parameters established from computer analysis.
After verifying the sensitivity and accuracy of the sensors and confirming the proper functioning of the SDI at the first burn, the values of the measured variables were re-established at the levels of normal conditions, and then the second burn started.The fuel used was sheets of A4 paper and MDF, totaling 500 grams of material.Combustion was done in the presence of the SDI and sprinkler, in order to compare the fire response times given by both.
In this comparison, it was observed that, in the specific case tested, the SDI detected the fire according to the expected parameters, being triggered 20 seconds after the visual observation of the start of the fire.The sprinkler, in turn, even at a temperature of 68ºC, did not have the bulb ruptured, and was for more than 2 minutes submitted to temperatures between 70ºC and 80ºC, only being activated after more than 3 minutes since the visual perception of the start of the fire.

Discussion of the Results
To start with, the results obtained in the calibration were highlighted, based on characteristics expressed in the graphs and physical and chemical properties of the combustible materials used in the tests.The first is that the burning of the MDF was slower than that of the paper, since it has a smaller surface in contact with the flames, for being less fragmented (Chang & Goldsby, 2013).In addition, the temperature variation over time formed a similar graph to the traditional fire curves (Costa & Silva, 2006), although the fire load only allowed to reach a temperature of 90ºC, showing that these curves are independent of the amount of fuel used.
It should also be pointed out that, according to the literature, the classical detector that first detects fire are those of gases, then those of flames, and finally those of temperature (Abinee, 2013).However, in our study, the flame sensor was the one that first detected the fire.Furthermore, when the MDF was being burned, the temperature sensor detected the fire before the gas sensor, whereas when the A4 paper, which is faster, was being burned, the smoke sensor detected the fire before the temperature sensor did.
The difference in responses given by the sensors of the proposed SDI compared to the common detectors is justified by the fact that in these, there is analysis of a single variable, and in the multi-sensory, only the trivial solution in which the gas or temperature sensor reaches the threshold, the fire is confirmed.In the suggested SDI, there is a simultaneous and integrated computational analysis of three quantities, which reduces the incidence of false positives and false negatives as to the event (Gavira, 2003).
In recent studies, a detector was made whose activation only occurs if there is simultaneous sensitization of 04 flame sensors (Rashid & Rafid, 2018).But in this case, there is observation of only one variable, as in common detectors, and the simultaneous sensitization of the sensors is used as a differential to reduce wrong responses.There are studies that use other logical principles to reduce response errors, which classify with a high degree of accuracy not only well-defined situations of fire or not, but intermediate ones, which have characteristics of both, such as the study done under fuzzy logic and multi-sensory bias, which shows optimal sensitivity, achieving a hit rate of 83%, so that in six test cases, it failed only in one (Obanda, 2017).
Comparing the result of these studies with the one found in the proposed SDI, it is noted that in this there is analysis of more than one variable in an integrated way, and although this considers only two states defined for each variable, obtained a lower percentage of misleading evaluations -6%, if compared to Obanda's research (2017), having used the decision tree method, as machine learning strategies, to improve the degree of assertiveness of the SDI, which was 94%.
It can be stated that the response time of the SDI studied is excellent, because when this time was compared to the response time of the sprinkler of red ampoule, whose breakage temperature for release breakage is 68º (ABNT NBR 16.400, 2018), it was noted that this only detected the fire after more than three minutes of the start of it and at a temperature of more than 75ºC, while that detected the fire after 20s of its start.
The common detectors and the sprinkler are usually activated 60s after the start of the fire (Abinee, 2013), which allows the destruction of part of the goods, besides requiring a large amount of extinguishing agent and energy to contain the accident.The proposed SDI, in this sense, can be said to be sustainable, since it is economically viable and brings great social benefit, with affordable value for use in residences, businesses and places that keep goods of inestimable value, such as libraries, museums and archives, besides bringing environmental benefit, since it detects the fire more quickly, reducing the consumption of water and energy in combat and the use of resources to replace the damaged goods, since the degradation of them is less the faster the detection occurs.

CONCLUSION
Based on the results obtained from the computational analysis of the experimental data, it can be stated that, for the particular case studied, the system was able to identify the fire 20s after its onset was observed, ten times less than the time that the sprinkler spent to detect it.As to the level of precision, a concise and precise decision tree structure was defined, whose accuracy was 94%, that is, every 100 analyzes made, the results of only 6 of them contradict what happens in reality, showing in these cases false positives or false negatives.
As next steps, it is suggested to increase the number of tests and materials tested, to see the sensitivity of the sensors against the main products of combustion coming from the burning of the materials most commonly used in buildings.It is also important to try to optimize the programming by which the data captured by the sensors is analyzed, including with the implementation of fizer logic and neural network, as the process becomes denser and more robust.
More accurate and compact prototyping needs to be developed to optimize electronic items, and to implement computer vision and IoT, so that the data obtained is transferred by the sensors to a control center and to users of the service, via the Internet, and by means of thermal and visual images, the environment is monitored remotely in real time.
The area under study is usual in the scientific community and involves two issues intensively discussed within society: the optimization of resources and the use of IoT for process automation.Accordingly, the continuation of this research is necessary, since it fosters the development of effective and sustainable products in the area of fire safety in urban buildings.

Figure 3 :
Figure 3: Physical simulator built to conduct the experiments Source: Authors' own elaboration, 2023.

Table 1 :
Sensor Logical State Description

Table 2 :
Sensor Logical State Description