Self-adaptive composite fire detection alarm model construction method for historic building

Abstract

The invention provides an ancient building adaptive composite fire detection alarm model construction method, and relates to the field of fire detection, and the method comprises the following steps: firstly, carrying out on-site survey to obtain an ancient building structure and internal decoration combustibles and layout, and employing a cone calorimeter to measure the combustion performance parameters of building materials and decorations; then, FDS software is adopted for conducting three-dimensional modeling and computational grid division, boundary and condition setting, fire point and fire source power setting, detection point position setting and the like on the internal space and display of the historic building, and data sets of CO concentration, CO2 concentration, O2 concentration, smoke shading rate and temperature parameter changes in the fire occurrence and development process are obtained through numerical simulation calculation; and finally, learning and training the initial multi-characteristic parameter data of the fire through a BP neural network, continuously correcting and optimizing in application, and constructing a self-adaptive fire alarm model, thereby more accurately and comprehensively sensing the fire of the ancient building, and realizing accurate early warning of the early fire.

Description

Technical Field

[0001] The present invention relates to the technical field of fire detection, and in particular to a method for constructing an adaptive composite fire detection alarm model for ancient buildings. Background Art

[0002] Since ancient buildings have many environmental interference sources, such as smoke, incense, oil smoke, water vapor, etc., directly using a single parameter fire detection product will have problems such as high false alarm rate and poor reliability. Especially in ancient buildings such as temples, there are many combustibles such as incense, fabrics, and wooden products. The combustion will produce smoke, gas, heat, light and other fire characteristics;

[0003] Dual-sensor detection devices that combine temperature and smoke detection are relatively common. In recent years, the market has launched a three-sensor detection device that integrates image, spectrum, and infrared temperature sensing. However, since fire detectors are usually composed of a microprocessor, a sensor, and some related circuits, integrating all kinds of fire sensors will make the device too complicated and reduce the speed at which the microprocessor processes data, making it difficult to complete signal processing inside the detector. With the integration of fire detection technology and Internet of Things technology, the collection, analysis, storage, and processing of fire information are developing in the direction of comprehensive perception, intelligent analysis, comprehensive integration, and smart Internet of Things. Therefore, the present invention proposes a method for constructing an adaptive composite fire detection alarm model for ancient buildings to solve the problems existing in the prior art. Summary of the invention

[0004] In view of the above problems, the present invention proposes a method for constructing an adaptive composite fire detection alarm model for ancient buildings. This method for constructing an adaptive composite fire detection alarm model for ancient buildings can not only fundamentally solve the false alarms caused by non-fire interference sources, thereby greatly reducing the false alarm rate of fires, but also overcome the underreporting caused by the use of a single fire parameter.

[0005] To achieve the purpose of the present invention, the present invention is implemented by the following technical scheme: a method for constructing an adaptive composite fire detection alarm model for ancient buildings, comprising the following steps:

[0006] S1: The architectural structure and internal layout of a certain ancient building are obtained through field survey and mapping, and the characteristic parameters of material combustion are measured by cone calorimeter test, including the fire parameters of the combustion performance of the building materials and interior decorations of the ancient building;

[0007] S2: The grid division, boundary and condition settings, ignition point and fire source power settings, and detection point settings of the internal space and statues of the ancient building were modeled through FDS software, and the data set of CO concentration, CO2 concentration, O2 concentration, smoke shading rate, temperature and other parameter changes during the fire development process were simulated;

[0008] S3: The fire detection data set is divided into training set and test set, and the BP neural network is used to train the data to build an alarm model, which is then applied to ancient buildings. The model is continuously adaptively optimized according to the actual alarm data to form a fire detector alarm model that integrates multiple fire parameters.

[0009] A further improvement is that: S1 comprises the following steps:

[0010] Conduct on-site surveys of target buildings, survey and map the architectural data of typical ancient buildings and the structural layout and combustible materials in the buildings;

[0011] Analyze the heat source and combustible materials in the building to set the fire point;

[0012] For the flammable materials in the building, the heat release rate of the materials shall be measured by cone calorimeter;

[0013] It provides parameter sources for numerical simulation of fire of building structures and material combustion performance for model construction.

[0014] A further improvement is that in S2, the grid division includes the following steps:

[0015] Mesh size calculation verification:

[0016]

[0017] Where: D * —characteristic diameter of the flame, m;

[0018] —Heat release rate, kW / m 2 ;

[0019] ρ ∞ —Density of air, kg / m 3 ;

[0020] c p —Specific heat of air at constant pressure, J / (kg·℃);

[0021] T ∞ —Ambient temperature, °C;

[0022] g—acceleration due to gravity; m / s 2 ;

[0023] ρ ∞ Take 1.1, c p Take 1014, T ∞ Take 20, g take 9.8;

[0024] According to the distribution of combustible materials in the building, multiple calculation areas with different grid sizes are divided.

[0025] A further improvement is that in S2, the boundary and condition setting includes the following steps:

[0026] Set environmental parameters, including starting temperature, atmospheric pressure, and gravitational acceleration;

[0027] Set ventilation conditions and simulate the opening and closing status of the doors and ventilation windows in the hall;

[0028] Set the working time and research scope, set a 2D open boundary slice of visibility at a building height of N, and all detectors to obtain corresponding data for research.

[0029] A further improvement is that in S2, the setting of the ignition point and the ignition source power includes the following steps:

[0030] Set working conditions according to the cause of the disaster;

[0031] When combustibles are ignited artificially, simulate the ignition of combustibles;

[0032] According to the area of ​​the fire source, select the fire speed type and calculate the heat release rate per unit area.

[0033] A further improvement is that in S2, the detection point setting includes the following steps:

[0034] Install several detectors in the building;

[0035] Detector selection: composite adaptive detector;

[0036] Detect temperature, flue gas, CO, CO2, and O2 content.

[0037] Further improvement is that: in said S3, using BP neural network to train data to construct alarm model specifically includes fire alarm model training based on fusion of 5 detection data of specific building, and fire alarm model training based on fusion of 3 independent detection data of specific building;

[0038] The fire alarm model training based on the fusion of five detection data of a specific building includes the following steps:

[0039] Five detection parameters of the target building detection points: CO, CO2, O2, visibility, and temperature were selected as learning data. The detection parameters of the target building during the entire simulated combustion process were used as original data. The data were screened and the boundary data were removed. Representative no-fire data, smoldering data, open fire data, and flashover data were selected from the data set. 1 / 2 was selected as the training set, 1 / 4 as the training set, and 1 / 4 as the validation set.

[0040] A four-layer BP neural network with two hidden layers, one input layer and one output layer was used to train and test the alarm model using the fire detection data set obtained by simulation;

[0041] After the model training is completed, the generalization ability test is carried out on ancient buildings with the same architectural structure and similar interior decoration materials. The data is selected according to the above process, and the accuracy is determined after multiple rounds of training iterations.

[0042] Further improvements are: Fire alarm model training based on the fusion of three independent detection data for a specific building, including the following steps:

[0043] The detection parameters CO, visibility, and temperature of the target building detection points during the entire simulated combustion process were used as the original data. The data were screened and the fuzzy data at the boundaries were removed. Representative data of no fire, smoldering, open flame, and data with obvious flashover characteristics were selected from the data set. 1 / 2 was selected as the training set, 1 / 4 as the training set, and 1 / 4 as the validation set.

[0044] A four-layer BP neural network with two hidden layers, one input layer and one output layer is used to train and test the alarm model using the fire detection data set obtained by simulation.

[0045] Further improvements are: testing and applying the training model in actual ancient building scenes, and adaptively optimizing the model, including the following steps:

[0046] Apply the trained fire alarm model to the fire alarm control of the target building for real-time data collection;

[0047] Collect and record the actual fire alarm data of the target building detection points, confirm the fire status category according to the actual situation, and use it as input data to optimize the alarm model;

[0048] The optimized model is evaluated for model transfer and generalization capabilities on ancient buildings with the same architectural structure and similar interior decoration materials. If the accuracy is lower than 90%, it will not be applicable. For new ancient buildings, data collection and selection are carried out according to the above process, and the model is determined after multiple rounds of training iterations to verify the accuracy.

[0049] The beneficial effects of the present invention are:

[0050] 1. The present invention uses multi-parameter composite detection technology to fuse the data of the collected initial characteristic parameters of the fire, and constructs a fire alarm model that is adaptive to specific ancient buildings, so as to perceive the fire in the ancient building more accurately and comprehensively, and realize the precise detection of the early fire in the ancient building. It can not only fundamentally solve the false alarms caused by non-fire interference sources, greatly reducing the false alarm rate of fire, but also overcome the omissions caused by the use of a single fire parameter.

[0051] 2. The present invention selects different types of sensors according to the detection environment, the formation and development laws of the initial fire in the detection area, the causes of false alarms, etc., and uses the Internet of Things technology to combine single-type fire detectors into a fire detector system, and then uses a data fusion algorithm to achieve information sharing and performance complementarity among multiple detectors. The redundant data between detectors can enhance the reliability of the detector system, and the complementary data between detectors can also expand the performance of a single detector, overcome the uncertainty and limitations of a single type of detector, and improve the effective performance of the entire fire detector system. At the same time, multi-detection and data fusion algorithms are used to give full play to the advantages of various types of detectors and perceive fire characteristic parameters more accurately and comprehensively.

[0052] 3. The present invention obtains fire characteristic parameter data such as CO concentration, CO2 concentration, O2 concentration, temperature, smoke visibility, etc. through fire numerical simulation, adopts a 4-layer BP neural network, applies 5 and 3 detection data fusion respectively, and adopts a variety of training and test set combination schemes to train the fire alarm model. The research results show that the accuracy of the trained alarm model fused with 5 and 3 detection data is close, CO concentration, temperature, and smoke visibility are sensitive characteristic parameters, the fire alarm model has a strong generalization ability for ancient buildings with similar architectural structures, and the alarm accuracy is high. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of the FDS numerical simulation data acquisition technology route of the present invention;

[0054] Figure 2 This is a schematic diagram of the detection point distribution of Hall A of the present invention;

[0055] Figure 3 This is a schematic diagram of the detection point distribution of Hall B of the present invention;

[0056] Figure 4 This is a schematic diagram of the detection point distribution of Hall C of the present invention;

[0057] Figure 5 This is a diagram of a fire alarm model training solution based on the fusion of five detection data of the three main halls of the present invention;

[0058] Figure 6 It is a curve diagram of accuracy and loss of the five detection data fusion training model based on the three main halls of the present invention;

[0059] Figure 7 This is a diagram of a fire alarm model training solution based on the fusion of three detection data of the three main halls of the present invention;

[0060] Figure 8 This is a graph showing the accuracy and loss of the three detection data fusion training model based on the three main halls of the present invention;

[0061] Fig. 9 This is a graph showing the accuracy and loss of the fusion training model of the three detection data of Hall A of the present invention;

[0062] Fig.10 This is a graph showing the accuracy and loss of the three detection data fusion training model of Hall B of the present invention;

[0063] Fig.11 This is a curve diagram of the accuracy and loss of the three detection data fusion training model of Hall C of the present invention. DETAILED DESCRIPTION

[0064] In order to deepen the understanding of the present invention, the present invention will be further described in detail below in conjunction with examples. The examples are only used to explain the present invention and do not constitute a limitation on the protection scope of the present invention.

[0065] Embodiment 1

[0066] according to Figure 1-11 As shown, this embodiment proposes a method for constructing an adaptive composite fire detection alarm model for ancient buildings, comprising the following steps:

[0067] The architectural structure and internal layout of a certain ancient building are obtained through field survey and mapping, and the characteristic parameters of material combustion are determined by cone calorimeter test, including the fire parameters of the combustion performance of the building materials and interior decorations of the ancient building; the specific steps include: field survey of the target building, mapping the architectural data of typical ancient buildings and the structural layout and combustible conditions in the building; analyzing the heat source and combustible conditions in the building to set the ignition point; using cone calorimeter to determine the heat release rate of the flammable materials in the building; providing the parameter source of fire numerical simulation of the building structure and material combustion performance for model construction;

[0068] The FDS software is used to model the grid division, boundary and condition settings, ignition point and fire source power settings, and detection point settings of the internal space and statues of the ancient building, and simulate the data set of the changes in parameters such as CO concentration, CO2 concentration, O2 concentration, smoke shading rate, and temperature during the fire development process; the grid division includes the following steps:

[0069] Mesh size calculation verification:

[0070]

[0071] Where: D * —characteristic diameter of the flame, m;

[0072] —Heat release rate, kW / m 2 ;

[0073] ρ ∞—Density of air, kg / m 3 ;

[0074] c p —Specific heat of air at constant pressure, J / (kg·℃);

[0075] T ∞ —Ambient temperature, °C;

[0076] g—acceleration due to gravity; m / s 2 ;

[0077] ρ ∞ Take 1.1, c p Take 1014, T ∞ Take 20, g take 9.8;

[0078] According to the distribution of combustible materials in the building, multiple calculation areas with different grid sizes are divided.

[0079] The boundary and condition settings include the following steps:

[0080] Set environmental parameters, including starting temperature, atmospheric pressure, and gravitational acceleration;

[0081] Set ventilation conditions and simulate the opening and closing status of the doors and ventilation windows in the hall;

[0082] Set the working time and research scope, set a 2D open boundary slice of visibility at a building height of N, and all detectors to obtain corresponding data for research.

[0083] The ignition point and ignition source power setting includes the following steps:

[0084] Set working conditions according to the cause of the disaster;

[0085] When combustibles are ignited artificially, simulate the ignition of combustibles;

[0086] According to the area of ​​the fire source, select the fire speed type and calculate the heat release rate per unit area.

[0087] The detection point setting includes the following steps:

[0088] Install several detectors in the building;

[0089] Detector selection: composite adaptive detector;

[0090] Detect temperature, flue gas, CO, CO2, O2 content;

[0091] The fire detection data set is divided into training set and test set, and the BP neural network is used to train the data to build an alarm model. The model is applied in ancient buildings, and the model is continuously optimized according to the actual alarm data to form a fire detector alarm model that integrates multiple fire parameters.

[0092] The BP neural network is used to train the data to build the alarm model, including the fire alarm model training based on the fusion of 5 detection data of a specific building, and the fire alarm model training based on the fusion of 3 independent detection data of a specific building;

[0093] The fire alarm model training based on the fusion of five detection data of a specific building includes the following steps:

[0094] Five detection parameters of the target building detection points: CO, CO2, O2, visibility, and temperature were selected as learning data. The detection parameters of the target building during the entire simulated combustion process were used as original data. The data were screened and the boundary data were removed. Representative no-fire data, smoldering data, open fire data, and flashover data were selected from the data set. 1 / 2 was selected as the training set, 1 / 4 as the training set, and 1 / 4 as the validation set.

[0095] A four-layer BP neural network with two hidden layers, one input layer and one output layer was used to train and test the alarm model using the fire detection data set obtained by simulation;

[0096] After the model training is completed, the generalization ability test is carried out on ancient buildings with the same architectural structure and similar interior decoration materials. The data is selected according to the above process, and the accuracy is determined after multiple rounds of training iterations.

[0097] The fire alarm model training based on the fusion of three independent detection data of a specific building includes the following steps:

[0098] The detection parameters CO, visibility, and temperature of the target building detection points during the entire simulated combustion process were used as the original data. The data were screened and the fuzzy data at the boundaries were removed. Representative data of no fire, smoldering, open flame, and data with obvious flashover characteristics were selected from the data set. 1 / 2 was selected as the training set, 1 / 4 as the training set, and 1 / 4 as the validation set.

[0099] A four-layer BP neural network with two hidden layers, one input layer and one output layer was used to train and test the alarm model using the fire detection data set obtained by simulation;

[0100] The training model is tested and applied in the actual ancient building scene to perform adaptive optimization of the model, including the following steps:

[0101] Apply the trained fire alarm model to the fire alarm control of the target building for real-time data collection;

[0102] Collect and record the actual fire alarm data of the target building detection points, confirm the fire status category according to the actual situation, and use it as input data to optimize the alarm model;

[0103] The optimized model is evaluated for model transfer and generalization capabilities on ancient buildings with the same architectural structure and similar interior decoration materials. If the accuracy is lower than 90%, it will not be applicable. For new ancient buildings, data collection and selection are carried out according to the above process, and the model is determined after multiple rounds of training iterations to verify the accuracy.

[0104] Embodiment 2

[0105] according to Figure 1-11 As shown, this embodiment proposes a method for constructing an adaptive composite fire detection alarm model for ancient buildings, comprising the following steps:

[0106] Data collection

[0107] Fire data acquisition method based on FDS numerical simulation

[0108] In view of the difficulty in conducting physical combustion tests on ancient buildings, obtaining the characteristic parameters of the early stage of fire in ancient buildings through FDS software simulation is a more feasible method of obtaining data. In order to obtain the fire data of the building, a field survey was conducted using the target palace complex as an example, and the architectural data of Hall A, Hall B, and Hall C, as well as the structural layout and combustible conditions in the building were mapped. The ignition point was set according to the heat source and combustible conditions in the building, and the heat release rate of the flammable materials in the building was measured using a cone calorimeter, providing a parameter source for the numerical simulation of fire of the building structure and material combustion performance for model construction. See the data acquisition path based on the FDS numerical simulation method. Figure 1 .

[0109] Basic information of the building

[0110] The target palace complex covers an area of ​​66,400m 2 The target palace complex includes Hall A, Hall B, and Hall C. This data collection focuses on Hall A, Hall B, and Hall C. The three buildings have their own unique architectural structures, usage functions, and internal furnishings.

[0111] Hall A

[0112] The main body of Hall A is painted red on the outside. It is located on a 1m high hall step. There is an annex on each side connected by a wall. There is a large incense burner 3m in front of the hall, and the smoke is everywhere. The outer contour of the target palace complex is measured by a laser rangefinder. The total length is 25.5m and the width is 11m. There are three double-opening arches on the front, each door is 2.1m wide and 2.5m high, and three double-opening arches on the back are 1.9m wide and 2.3m high. There are also two windows on both sides that are 1.3m wide and 1.4m high, 1m above the ground. The interior is divided into the main hall and the side rooms surrounded by fences on both sides, with two statues in each side room. The main hall is 15.4m long and the side rooms are 3.8m long. There is a wall in the center of the main hall with gold-painted wooden chairs. There are two 3m long and 1m wide altars in front of the fences on both sides. There are 8 load-bearing beams in the hall, which are stacked up and down and cross the side rooms and walls. There is a mahogany chair on the wall with a statue sitting on it, an altar in front, and tower-shaped decorations surrounded by fences on both sides.

[0113] Hall B

[0114] The outer layer of the main body of Hall B is similar to the architectural shape of Hall A. There is a 1m high temple step in front of the door, and a small incense burner 3m in front of the hall, which is filled with smoke. According to the measurement of the laser rangefinder, the overall length of the outer contour of Hall B is 25.7m and the width is 11.2m. There are three double-opening square doors on the front, each door is 1.6m wide and 2.7m high, and there are also three double-opening large arches on the back, which are 1.6m wide and 2m high. In addition, there are small openings on both sides of the front and back, which are 1m wide and 2m high. The two windows on the front are 2m wide, 3.2m high, and 1.6m above the ground. The interior is divided into the main hall and the side rooms surrounded by fences on both sides, with two statues in each side room. The main hall is 17.5m long and the side rooms are 3.2m long. In the center of the main hall is a wall with statues and two niche display cabinets on both sides. There are two 3m long and 1m wide altars in front of the fences on both sides. What is quite different from Hall A is that Hall B has 16 small pillars in the center and 6 silk fabrics hanging symmetrically from the ceiling.

[0115] Hall C

[0116] Hall C is 25 meters high and magnificent, which is quite different from the previous two buildings. There is a statue inside, which is 18 meters above the ground and 8 meters underground. The statue is 8 meters wide and made of white sandalwood. Hall C has a three-story corridor structure surrounding the statue. Studying the spread of fire inside it is of great significance to the fire detection of ancient buildings.

[0117] Model setup

[0118] Meshing

[0119] The wooden roof structure on the top of the research object is complex. Under the premise of ensuring the authenticity and effectiveness of the simulation results, the modeling process can be simplified as much as possible, and the grid size calculation and verification are shown in Equation 1 and Equation 2:

[0120]

[0121] Where: D * —characteristic diameter of the flame, m;

[0122] —Heat release rate, kW / m 2 ;

[0123] ρ ∞ —Density of air, kg / m 3 ;

[0124] c p —Specific heat of air at constant pressure, J / (kg·℃);

[0125] T ∞ —Ambient temperature, °C;

[0126] g—acceleration due to gravity; m / s 2 ;

[0127] ρ ∞ Take 1.1, c p Take 1014, T ∞ Take 20, and g take 9.8. After calculation, the grid size is 0.2m×0.2m×0.2m. In order to take into account both calculation efficiency and quality, we can divide the calculation areas into multiple grids with different sizes according to the distribution of combustibles in the hall. The divided areas are as follows: Figure 1 shown.

[0128] Boundary and condition settings

[0129] Environmental parameters

[0130] In the model settings, the starting temperature is set to 20°C, the atmospheric pressure is 101.32KPa, and the gravity acceleration is set to 9.8m / s 2 ;

[0131] Ventilation

[0132] During the simulation, the six 2.1m×2.5m arched doors in front of Hall A and the rear door were set to open, with a ventilation window on each side, and a ventilation area of ​​1.3m×1.4m.

[0133] The six 2.1m×2.5m rectangular doors in front and at the back of Hall B are set to open. There is a 1.5m×2.5m side door on each side of the front protrusion. There is a ventilation window on each side, with a ventilation area of ​​1.3m×1.5m, so that the high-temperature smoke in the hall can convect with the natural air outside the hall under the action of atmospheric pressure.

[0134] The six 2.6m×2.5m arched doors at the front and back doors of Hall C are set to open. There is a ventilation window on each side, with a ventilation area of ​​1.3m×1.4m. In addition to the ventilation on the first floor, the air in the corridor structure space on the second and third floors circulates with each other. The corridor is also equipped with ventilation windows, and the outside air can also enter from the second and third floors.

[0135] Duration of the work and scope of the study

[0136] The simulation time of the working condition was set to 1000s. A 2D open boundary slice of visibility and all detectors were set at a height of 6.88m (the underground foundation was 1m above the ground) for the target palace complex to obtain corresponding data for research.

[0137] Ignition point and ignition power setting

[0138] Assuming that the disaster is caused by human factors, the target palace complex is equipped with a unified incense burner outside each building, and pilgrims can light the incense they bought. For Hall A, Hall B, and Hall C, it is assumed that pilgrims illegally bring incense into the hall, and accidentally drop it on the silk fabric on the ground, igniting the combustibles. To simulate the real situation, the silk fabric is simulated to be ignited, and the fire source area is 0.5m 2 , select t2 slow fire (α is 0.00293), the heat release rate per unit area is 500KW / m 2 .

[0139] Detection point setting

[0140] At present, single-parameter fire detectors are installed in all the halls, with a total of four detectors located in the southeast, northeast, southwest, and northwest of the top beam. In order to study composite adaptation, it is necessary to replace their positions with composite adaptive detectors. Composite detectors have the functions of detecting temperature, smoke, CO, CO2, and O2 content. Since the fire point is located in the center of the building, the four detection points are symmetrical in reality, so the detection results on the left and right symmetrical sides may be closer.

[0141] There are two detection points in Hall A: the left room, the right room and the front door. Figure 4 , two detection points behind the back door wall ( Figure 2 There are 6 detection points in total.

[0142] The structure of Hall B is similar, with two detection points at the left room, right room and front door. Figure 3 As shown, there are two detection points behind the back door wall ( Figure 3 There are 6 detection points in total.

[0143] Hall C is a three-story structure, with the bottom floor being the main hall and the second and third floors being corridors. Each corridor is equipped with windows for ventilation. The statue stands tall, from the center of the bottom floor, straight through the second and third floors. The top of the statue is the top of the main hall. There are 6 detection points on each floor, for a total of 18 detection points ( Figure 4 The second and third floor corridor structures are hidden in the figure to show the spatial location of the detection points. In addition, a group of 4 circular detection points are set around the statue, and 3 detection points are set from bottom to top on the statue body.

[0144] Adaptive fire detection algorithm model

[0145] Based on the above simulation results, the fire detection data set is divided into training set and test set, and the BP neural network is used to train the data, and finally the multi-data fusion fire detector alarm model is determined.

[0146] BP neural network is a widely used artificial neural network. From the original input data to the final output result, the model is built through multiple iterations of training through the error back propagation mechanism. BP neural network consists of three or more layers of neurons, namely input layer, hidden layer and output layer. Among them, the input layer receives the input signal, the output layer outputs the network training results, and the hidden layer is the intermediate layer used to process the input signal. Each layer of neurons is interconnected, and the weights and thresholds are continuously adjusted according to the training process.

[0147] Model Training

[0148] A four-layer BP neural network with two hidden layers, one input layer and one output layer was used. The fire detection data simulated in the three main halls of the target palace complex was used to select multiple schemes for data sets for experimental training and testing of alarm model.

[0149] Fire alarm model training based on fusion of five detection data of the three main halls

[0150] All five detection parameters (CO, CO2, O2, visibility, and temperature) of the Device02 detection point in the center of Hall B were selected as the training set data. After the model training was completed, the generalization ability test was carried out in Hall A, Hall C, and the niche. Figure 5 shown.

[0151] Data processing

[0152] The detection parameters of the Device 2 detector in the center of Hall B during the entire simulated combustion process were used as the original data, and the data was screened. FDS outputs a set of data per second within 1000 seconds of the entire simulated combustion process. However, due to the continuity of the process, there are blurred boundaries, detection delays, and different response times of different modules. In order to achieve higher accuracy, it is necessary to remove boundary data. From the data set, 48 groups of representative no-fire data, 46 groups of smoldering data, 421 groups of open flame data, and 107 groups of flashover data were selected, totaling 750 groups as training sets and validation sets. Device 2 in Hall A and Device 2 in Hall C were used as test sets. Data selection is also based on the above principles. Some examples of training set data are shown in Table 1.

[0153] Table 1B Part of the training data

[0154] Serial number Temperature(℃) <![CDATA[Visibility loss (%·m -1 )]]> <![CDATA[CO concentration (10 -5 ppm)]]> <![CDATA[CO2 concentration (10 -5 ppm)]]> <![CDATA[O2 concentration (10 -5 ppm)]]> result 1 20.00 0 0 0.0003 2075.5882 2 20.01 0.0031 0.0061 0.0051 2075.5882 1 3 20.05 0.1069 0.0426 0.5332 2071.5773 1 … … … … … … … 749 324.39 22.94 1.7396 467.7973 746.5065 3 750 324.44 18.06 1.6739 450.1040 705.5291 3

[0155] Note: There are four result states: 1 No fire, 2 Smoldering, 3 Open flame, 4 Flashover.

[0156] Training and testing results

[0157] The number of iterations reached about 0.9 after about 30 rounds and gradually stabilized. After 85 rounds of iterations, the accuracy reached 0.9535. After 100 rounds of iterations, the accuracy finally reached 0.952. The test results of Hall A were 0.957 and the test results of Hall C were 0.724. The loss curve is shown in the figure below. Figure 6 As shown. It can be seen that the model trained using the simulated data of Hall B as the training set has a better generalization ability for Hall A, which has a similar structure to Hall B, but a weaker generalization ability for Hall C, which has a much different building structure. The main reason is that Hall C is a tall building, and its air volume and ventilation area are quite different from those of Hall A and Hall B. The laws of fire spread and development also show great differences, making the training model based on the simulated data of Hall B not applicable to Hall C.

[0158] Fire alarm model training based on fusion of three detection data of the three main halls

[0159] The detection parameters of Device02 detection point in the center of Hall B were selected as the training set data. After the model training was completed, the generalization ability test was carried out in Hall A, Hall C, and the niche. In order to achieve a more ideal fitting process, the CO2 and O2 data with large differences were no longer selected as the input layer. The relatively late and accurate temperature module data was selected as X1, the CO concentration module data with relatively medium-term warning function was selected as X2, and the smoke module data with relatively early rapid warning function was selected as X3. The training plan is as follows Figure 7 shown.

[0160] Data processing

[0161] The detection parameters of the Device 2 detector in the center of Hall B during the entire simulated combustion process were used as the original data, and the data were screened. A total of 750 groups were used as training sets. The selection method was the same as that of Scheme 1, eliminating the data with more blurred boundaries and selecting the data with more obvious features.

[0162] Training Results

[0163] After training the model based on the 750 layers of data from Hall B, which is more obvious, we selected the simulated combustion conditions of other halls for evaluation and generalization research. The results show that after 100 layers of training data iterations, the test accuracy of Hall B reached 0.9498, which is similar to the result of the fusion of 5 detection data. The test results of Hall A were 0.956, and the test results of Hall C were 0.725. The overall test data accuracy of the three halls could only reach 0.8, which is still not ideal. The loss curve is shown in Figure 1. Figure 8 shown.

[0164] Fire alarm model training based on fusion of three independent detection data for each hall

[0165] Training and test set division

[0166] The detection parameters of the Device 2 detection point in the center of Hall B during the entire simulated combustion process were used as the original data (CO, visibility, temperature), and the data were screened. A total of 375 groups were used as training sets. The selection method was the same as the principle of Scheme 1. The data with more blurred boundaries were eliminated, and the data with more obvious features were selected, leaving 375 groups as test sets. For Hall A, the Device 2 detection point in the center of the hall was selected. After the same screening, 375 groups were selected as training sets, and 375 groups were selected as test sets. For Hall C, the Device 2 detection point in the center of the first floor was selected. After the same screening, 375 groups were selected as training sets, and 375 groups were selected as test sets.

[0167] Training and testing results

[0168] After training the model based on 375 sets of data from Hall A, the remaining 375 sets of data were selected for testing. The model accuracy performed well, with the training set accuracy reaching 0.9653. The test set accuracy reached 0.9527. The accuracy and loss curves are shown in Figure 2. Fig. 9 shown.

[0169] After training the model based on 375 sets of data from Hall B, the remaining 375 sets of data were selected for testing. The model accuracy performed well, with the training set accuracy reaching 0.9520. The test set accuracy reached 0.9345. The accuracy and loss curves are shown in Figure 2. Fig.10 shown.

[0170] After training the model based on 375 sets of data from Hall C, the remaining 375 sets of data were selected for testing. The model accuracy performed well, with the training set accuracy reaching 0.9493. The test set accuracy reached 0.9309. The accuracy and loss curves are shown in Figure 2. Fig.11 shown.

[0171] in conclusion

[0172] By applying a 4-layer BP neural network and using the fire simulation data of the target palace complex A, B, and C as the training set, the results of the multi-data fusion fire detection alarm model training experiment show that:

[0173] The accuracy of the fire detection alarm model based on the fusion training of five detection data items, CO, CO2, O2, smoke visibility, and temperature, is close to that of the fire alarm model based on the fusion training of three detection data items, CO, smoke visibility, and temperature. The reason is that the fire simulation data shows that CO2 and O2 data change relatively slowly, which are relatively insensitive variables and have little impact on the accuracy of the model. Therefore, limited by the actual situation of the ancient building, the composite fire detector with three sensors, CO, smoke, and temperature, also has an ideal alarm effect.

[0174] The fire alarm model trained based on a certain building fire simulation data has a strong generalization ability and an accuracy rate of up to 0.95 when the building structure and decorations are similar to those of the training data source. However, for buildings with large differences in structure and interior decoration from the training buildings, the accuracy rate drops significantly to only 0.72.

[0175] The alarm model training using the building's own fire simulation data has a high accuracy rate for the training building, all above 0.93. Therefore, from the perspective of protecting the refined management of cultural relics, it is recommended to collect important building environmental data and fire simulation data in the building fire IoT system, and use it as a training data set to train the alarm model of multi-detection data fusion that is suitable for the ancient building itself, so as to enhance the accuracy of each independent ancient building fire alarm, rather than pursuing the generalization ability of the fire alarm model.

[0176] The method uses a multi-parameter composite detection technology to perform data fusion on the collected initial characteristic parameters of the fire, and constructs a fire alarm model that is adaptive to a specific ancient building, so as to more accurately and comprehensively perceive the fire of the ancient building, and realize the precise detection of the early fire of the ancient building. It can not only fundamentally solve the false alarm caused by non-fire interference sources, so that the false alarm rate of the fire is greatly reduced, but also overcome the omission caused by the use of a single fire parameter; and the present invention selects different types of sensors according to the detection environment, the formation and development law of the initial fire in the detection area, the cause of the false alarm, etc., and uses the Internet of Things technology to combine single-type fire detectors into a fire detector system, and then uses a data fusion algorithm to realize information sharing and performance complementarity among multiple detectors. The redundant data among the detectors can enhance the reliability of the detector system, and the complementary data among the detectors can also expand the performance of a single detector, overcome the uncertainty and limitation of a single type of detector, and improve the effective performance of the entire fire detector system. At the same time, the multi-detection and data fusion algorithms are used to give full play to the advantages of various types of detectors, and more accurately and comprehensively perceive the fire characteristic parameters, which is an inevitable trend for realizing the precise detection of the early fire of ancient buildings (groups). In summary, the present invention obtains fire characteristic parameter data such as CO concentration, CO2 concentration, O2 concentration, temperature, smoke visibility, etc. through fire numerical simulation, adopts a 4-layer BP neural network, applies 5 and 3 detection data fusion respectively, and adopts a variety of training and test set combination schemes to train the fire alarm model. The research results show that the training alarm models fused with 5 and 3 detection data have similar accuracy, CO concentration, temperature, and smoke visibility are sensitive characteristic parameters, and the fire alarm model has a strong generalization ability for ancient buildings with similar architectural structures, with an alarm accuracy of up to 0.95, and the alarm accuracy of non-similar structures is significantly lower, only 0.72. The training model based on the fire simulation parameters of each independent building has a high alarm accuracy for the building itself, which is above 0.93. The research process and results provide a theoretical basis for the selection of composite fire characteristic parameters, and provide a guiding method and implementation path for constructing an adaptive fire alarm model with multi-detection data fusion for important ancient buildings.

[0177] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for constructing an adaptive composite fire detection alarm model for ancient buildings, characterized in that: The following steps are involved: S1: The architectural structure and internal layout of a certain ancient building are obtained through field survey and mapping, and the characteristic parameters of material combustion are measured by cone calorimeter test, including the fire parameters of the combustion performance of the building materials and interior decorations of the ancient building; S2: The grid division, boundary and condition settings, ignition point and fire source power settings, and detection point settings of the internal space and statues of the ancient building were modeled through FDS software, and the data set of CO concentration, CO2 concentration, O2 concentration, smoke shading rate, temperature and other parameter changes during the fire development process were simulated; S3: The fire detection data set is divided into training set and test set, and the BP neural network is used to train the data to build an alarm model, which is then applied to ancient buildings. The model is continuously adaptively optimized according to the actual alarm data to form a fire detector alarm model that integrates multiple fire parameters.

2. The method for constructing an adaptive composite fire detection and alarm model for ancient buildings according to claim 1 is characterized in that: The S1 comprises the following steps: Conduct on-site surveys of target buildings, survey and map the architectural data of typical ancient buildings and the structural layout and combustible materials in the buildings; Analyze the heat source and combustible materials in the building to set the fire point; For the flammable materials in the building, the heat release rate of the materials shall be measured by cone calorimeter; It provides parameter sources for numerical simulation of fire of building structures and material combustion performance for model construction.

3. The method for constructing an adaptive composite fire detection alarm model for ancient buildings according to claim 2, characterized in that: In S2, the grid division includes the following steps: Mesh size calculation verification: Where: D * —characteristic diameter of the flame, m; —Heat release rate, kW / m 2 ; ρ ∞ —Density of air, kg / m 3 ; c p —Specific heat of air at constant pressure, J / (kg·℃); T ∞ —Ambient temperature, °C; g—acceleration due to gravity; m / s 2 ; ρ ∞ Take 1.1, c p Take 1014, T ∞ Take 20, g take 9.8; According to the distribution of combustible materials in the building, multiple calculation areas with different grid sizes are divided.

4. The method for constructing an adaptive composite fire detection and alarm model for ancient buildings according to claim 3 is characterized by: In S2, the boundary and condition setting includes the following steps: Set environmental parameters, including starting temperature, atmospheric pressure, and gravitational acceleration; Set ventilation conditions and simulate the opening and closing status of the doors and ventilation windows in the hall; Set the working time and research scope, set a 2D open boundary slice of visibility at a building height of N, and all detectors to obtain corresponding data for research.

5. The method for constructing an adaptive composite fire detection and alarm model for ancient buildings according to claim 4, characterized in that: In S2, the setting of the ignition point and the ignition source power includes the following steps: Set working conditions according to the cause of the disaster; When combustibles are ignited artificially, simulate the ignition of combustibles; According to the area of ​​the fire source, select the fire speed type and calculate the heat release rate per unit area.

6. The method for constructing an adaptive composite fire detection and alarm model for ancient buildings according to claim 5, characterized in that: In S2, the detection point setting includes the following steps: Install several detectors in the building; Detector selection: composite adaptive detector; Detect temperature, flue gas, CO, CO2, and O2 content.

7. The method for constructing an adaptive composite fire detection and alarm model for ancient buildings according to claim 6, characterized in that: In S3, the BP neural network is used to train the data to construct the alarm model, which specifically includes the fire alarm model training based on the fusion of five detection data of a specific building and the fire alarm model training based on the fusion of three independent detection data of a specific building; The fire alarm model training based on the fusion of five detection data of a specific building includes the following steps: Five detection parameters of the target building detection points: CO, CO2, O2, visibility, and temperature were selected as learning data. The detection parameters of the target building during the entire simulated combustion process were used as original data. Representative no-fire data, smoldering data, open fire data, and flashover data were selected from the data set, and 1 / 2 was selected as the training set and 1 / 2 as the validation set. A four-layer BP neural network with two hidden layers, one input layer and one output layer is used to train and test the alarm model using the fire detection data set obtained by simulation.

8. The method for constructing an adaptive composite fire detection and alarm model for ancient buildings according to claim 7, characterized in that: The fire alarm model training based on the fusion of three independent detection data of a specific building includes the following steps: The detection parameters CO, visibility, and temperature of the target building detection point during the entire simulated combustion process were used as the original data. Representative no-fire data, smoldering data, open flame data, and data with obvious flashover characteristics were selected from the data set, 1 / 2 of which was selected as the training set and 1 / 2 as the validation set. A four-layer BP neural network with two hidden layers, one input layer and one output layer is used to train and test the alarm model using the fire detection data set obtained by simulation.

9. A method for constructing an adaptive composite fire detection alarm model for ancient buildings according to claim 8, characterized in that: The training model is tested and applied in the actual ancient building scene to perform adaptive optimization of the model, including the following steps: Apply the trained fire alarm model to the fire alarm control of the target building for real-time data collection; Collect and record the actual fire alarm data of the target building detection points, confirm the fire status category according to the actual situation, and use it as input data to optimize the alarm model; The optimized model is evaluated for model transfer and generalization capabilities on ancient buildings with the same architectural structure and similar interior decoration materials. If the accuracy is lower than 90%, it will not be applicable. For new ancient buildings, data collection and selection are carried out according to the above process, and the model is determined after multiple rounds of training iterations to verify the accuracy.

Images