Indoor fire source positioning method based on infrared thermogram of outer surface of glass curtain wall

Abstract

An indoor fire source positioning method based on an infrared thermogram of the outer surface of a glass curtain wall comprises the following steps: database construction, which mainly comprises construction of a fire data set, temperature distribution of the outer surface of the glass curtain wall, and image extraction and processing; a Resnet deep learning model is constructed and trained; and carrying out fire source positioning by using a Resnet deep learning model. In the early stage of fire disaster development, when fire behavior reconnaissance of a super high-rise building fire disaster is carried out, a reconnaissance unmanned aerial vehicle carrying an infrared thermal imager is controlled to fly out of a glass curtain wall of a fire floor, and the position of the reconnaissance unmanned aerial vehicle is adjusted to obtain a complete glass curtain wall outer surface infrared thermal image of the fire floor; an infrared thermal image of the outer surface of the glass curtain wall acquired by the thermal infrared imager is input into the trained Resnet-50 deep learning model, the model analyzes the infrared thermal image to give a position number of a fire source, the position of the fire source can be accurately recognized in the early stage of fire development, the real-time performance and reliability are high, and the fire early warning capacity and rescue efficiency of the super high-rise building are improved.

Description

Technical Field

[0001] The invention relates to an indoor fire source positioning method based on an infrared thermal image of the outer surface of a glass curtain wall, and belongs to the technical field of fire source positioning. Background Art

[0002] As the pace of urbanization and town construction accelerates, super-high-rise buildings have emerged. With their huge scale, complex structure and high density, super-high-rise buildings, while providing convenience, also bring huge fire safety hazards.

[0003] After a fire occurs in a super high-rise building, the smoke layer in the burning room accumulates and gradually settles, forming a certain thickness of smoke layer in a short period of time, which greatly reduces the visibility indoors and makes it impossible to accurately identify the location of the fire source. Existing super high-rise buildings mostly use glass curtain wall structures on the exterior surface. The glass curtain wall prevents outdoor infrared thermal imagers from accurately identifying the indoor temperature field, making it difficult to conduct outdoor fire reconnaissance, affecting the rescue efficiency of super high-rise building fires, causing the fire to expand and causing serious casualties and property losses.

[0004] At present, the initial fire source positioning technology relies on the equipment in the fire building for positioning. The room fire source positioning technology mainly includes visual fire positioning technology, wireless sensor network positioning technology and optical fiber sensor positioning technology. These positioning technologies all require fixed position sensors, which need to be installed during building construction. Fixed deployment has poor flexibility and is likely to fail during a fire. In addition, these positioning technologies have a high time cost and cannot meet the real-time requirements in actual application scenarios. Summary of the invention

[0005] The present invention provides an indoor fire source positioning method based on infrared thermal images of the outer surface of a glass curtain wall. The positioning method can accurately identify the location of the fire source in the early stage of fire development, has strong real-time reliability, and improves the fire warning capability and rescue efficiency of super high-rise buildings.

[0006] In order to achieve the above object, the present invention provides a method for locating indoor fire sources based on infrared thermal images of the outer surface of a glass curtain wall, comprising the following steps:

[0007] S1, database construction, mainly including the construction of fire data set, temperature distribution of the outer surface of glass curtain wall, image extraction and processing;

[0008] S2. Build and train the Resnet deep learning model;

[0009] S3. Use the Resnet deep learning model to locate the fire source.

[0010] Furthermore, the specific process of S1 is as follows:

[0011] S1.1. Use fire computational fluid dynamics simulation software FDS to simulate fire data and construct a fire data set. The specific process is as follows:

[0012] S1.1.1. Build a fire combustion model: set parameters, including grid parameters, geometric structure of the burning room (including glass curtain wall and ventilation openings), surface parameters, reaction parameters, initial ambient temperature T0, fire source location p, fire source heat release rate q, combustible smoke production rate f, ventilation opening opening time l; set a total of p×q×f×l groups of simulation conditions;

[0013] S1.1.2. Set n×n temperature measuring points on the inner surface of the fire room glass in a square array with equal rows and columns;

[0014] S1.1.3, set the simulation time t to simulate, and obtain the data set of n×n temperature measurement points {T nt}, T nt It represents the temperature data of the nth measuring point at the tth second. The time-temperature data sets of different temperature measuring points are as follows:

[0015]

[0016] S1.1.3. Integrate the time-temperature data set into a complete FDS simulation data set as input data for subsequent finite element analysis;

[0017] S1.2. Use transient thermal analysis of computational fluid dynamics software ANSYS and fire computational fluid dynamics simulation software FDS to simulate the temperature load on the inner surface of the glass curtain wall and perform heat conduction simulation to obtain the temperature distribution on the outer surface of the glass curtain wall. The specific process is as follows:

[0018] S1.2.1. Use SpaceCliam to build a glass model with the same geometric dimensions as the FDS fire combustion model;

[0019] S1.2.2. Import the constructed glass model into transient thermal to complete the import of the geometric structure;

[0020] S1.2.3. Input the key material parameters of the glass model for subsequent thermal conductivity analysis, including: density, thermal conductivity, specific heat, and thermal expansion coefficient;

[0021] S1.2.4. Perform meshing. The parameters of meshing include: physical preference of the unit, unit order, unit size, number of nodes and units;

[0022] S1.2.5. Set the initial temperature of the glass model to be consistent with the initial ambient temperature T0 in the FDS fire combustion model;

[0023] S1.2.6, set the duration of ANSYS finite element analysis to be consistent with the duration t of FDS simulation;

[0024] S1.2.7. Set the convection heat transfer coefficient on the outer surface of the glass model to simulate the convection heat transfer between the hot glass and the external environment in a real high-rise building fire;

[0025] S1.2.8. Input the inner surface temperature data of the glass curtain wall obtained by FDS simulation as the "temperature" load to the inner surface of the ANSYS glass model. In this process, special attention should be paid: according to the resolution of the grid division, each "temperature" load can cover multiple grid cells. The spatial position of each input "temperature" load is consistent with the spatial position of the temperature measurement point set in the FDS simulation, that is, the temperature measurement point data of the nth FDS simulation corresponds to the nth AYSYS temperature load input, and the nth temperature load input to the ANSYS glass model is:

[0026]

[0027] S1.2.9. Perform ANSYS finite element analysis to obtain the temperature change on the outer surface of the glass model. This process is based on the heat conduction equation, and the heat conduction process is represented by the following calculus equation:

[0028]

[0029] Where ρ is the density of the glass model; c is the specific heat capacity of the glass model; T is the function of temperature changing with time and space; t * is time; k is the thermal conductivity of the material; Q is the thermal power density of the internal heat source. Since the glass model has no internal heat source, Q = 0;

[0030] At the initial time t * = 0 is T((x, y, z, 0) = T0, and n×n temperature loads are applied on one side of the glass model. The boundary conditions are expressed as:

[0031] T(x i ,y i ,t)=T boundary (x i ,y i ,t);

[0032] Among them, T boundary represents the n×n temperature loads applied on the inner surface of the glass model; x i and iis the coordinate position of the temperature load on the inner surface of the glass model; the convection heat transfer boundary condition set on the outer surface of the glass model is expressed as:

[0033]

[0034] in, represents the temperature gradient of the glass model surface in the normal direction; h is the convection heat transfer coefficient; T * is the temperature of the outer surface of the glass model; T ∞ is the temperature of the ambient fluid at the outer surface of the glass;

[0035] During the finite element solution process, the convection boundary conditions will be discretized and the solution equation is as follows:

[0036]

[0037] Among them, [C conv ] represents the contribution of convective heat transfer, [M] is the mass matrix, which is proportional to the specific heat capacity of the glass material and the volume of the unit, and describes the change in the energy storage of the overall system when the temperature of each model unit changes; [K] is the stiffness matrix, which represents the thermal stiffness of the model system, that is, the resistance of the glass model to heat transfer under unit temperature change; {Q} is the load vector, which represents the applied temperature load; {T} is the temperature that changes with time and space;

[0038] S1.2.10, select "Inverse Grayscale" for the color scheme of the solution results of the outer surface temperature of the glass model;

[0039] S1.3, Image extraction and processing:

[0040] S1.3.1. Extract the temperature distribution diagram of the outer surface of the glass model of all frames according to the frame number of ANSYS finite element analysis results;

[0041] S1.3.2, extract the temperature distribution diagram of the outer surface of all glass models under the working conditions of p×q×f×l group, and construct an unprocessed original image database;

[0042] S1.3.3. Perform Gaussian blur processing on all images in the original image database to eliminate the regionalization of the temperature distribution on the outer surface of the glass model caused by the input of multiple temperature loads in the S1.2 process. The Gaussian filter of the two-dimensional image of the temperature image on the outer surface of the glass model is expressed as:

[0043]

[0044] Among them, x and y are the horizontal and vertical distances between the image and the center pixel respectively; G(x,y) is the output value of the Gaussian filter, which is the discretization form of the two-dimensional Gaussian function; δ is the standard deviation; it is used to control the size of the Gaussian filter. The larger the δ, the larger the size of the Gaussian filter, and the more obvious the blurring effect will be; when performing Gaussian blurring, the Gaussian filter is used to continue convolution operations on the image to achieve the blurring effect. For a glass model outer surface temperature distribution image I(x,y) and a Gaussian filter G(x,y), the convolution result I'(x,y) is:

[0045]

[0046] In the formula, G(i,j) is a two-dimensional Gaussian filter kernel, which represents the weight of the distance to the center pixel;

[0047] By weighted averaging each pixel and its neighborhood, the sharp edges and details of the image are reduced; the image is made smoother, and the regionalization of the temperature distribution on the outer surface of the glass model caused by the input of multiple temperature loads in the S1.2 process is eliminated;

[0048] S1.4. Perform Laplace sharpening calculation on the temperature distribution image of the outer surface of the glass model after Gaussian blur processing to strengthen the image boundary between the high temperature area and the low temperature area in the temperature image. Laplace sharpening is an image enhancement technology used to enhance the edge details in the image and highlight the structure of the image. It combines the Laplace operator and the addition operation of the image. The Laplace operator is a second-order derivative operator and is defined in two-dimensional space as:

[0049]

[0050] Among them, f(x,y) is the pixel value of the image at the point (x,y), which represents the Laplace operator. In discrete images, the Laplace operator is approximated by a filter, and the common Laplace filter is used to approximate the Laplace operator:

[0051]

[0052] The Laplace filter calculates the neighborhood difference of each pixel in the image to highlight the edges and details of the image;

[0053] Laplace sharpening combines the original pixel values ​​of the image with the output of the Laplace operator. Its mathematical expression is as follows:

[0054] f sharp (x,y)=f(x,y)-α·Δf(x,y);

[0055] Where f(x,y) is the pixel value of the original image; Δf(x,y) is the result of applying the Laplacian operator to the image f(x,y);

[0056] α is a parameter that controls the sharpening intensity and is a constant;

[0057] The steps for Laplace sharpening of the image database are as follows:

[0058] S1.4.1. Calculate the Laplace operator result of the image f(x,y) to obtain the edge information in the image;

[0059] S1.4.2. Select the control parameter α according to the strength of sharpening; if α is too large, the image may become over-sharpened, resulting in enhanced noise; if α is too small, the sharpening effect is not obvious.

[0060] S1.4.3. Combine the original image with the Laplace operator result to obtain the sharpened image f sharp (x,y);

[0061] S1.4.4. Set a label of the fire source location for each image in the image database after image processing, that is, a number between 0 and p, to complete the construction of the glass model outer surface temperature image database.

[0062] Furthermore, the specific process of S2 is:

[0063] S2.1, using the Resnet-50 deep learning model, using the glass model outer surface temperature distribution image database built in S1 to train the Resnet-50 deep learning model, extracting deep features from the temperature distribution image. The model structure includes:

[0064] S2.1.1, Input layer: The images input from the image database are adjusted for suitability in size through this layer. The images in the image database are all generated using the "reverse grayscale" color scheme. The images used in training are all grayscale images and are all single-channel;

[0065] S2.1.2, convolution layer: This layer performs preliminary feature extraction and reduces the spatial size of the output feature map;

[0066] S2.1.3, Max Pooling Layer: Reduces the spatial dimension of the image, takes the maximum value in each local area, retains the most significant features, and makes feature extraction more robust to small changes in the input;

[0067] S2.1.4, multiple groups of residual blocks: Each residual block includes an identity block (the input is directly passed to the output and added to the result of the convolution) and a convolution block (to keep the dimensions of the input and output consistent. This model uses convolution of a specific size for dimensionality increase or decrease), which is used to avoid gradient vanishing and degradation in deep neural networks; for ideal feature mapping:

[0068] Φ(x * )=f(x * )+x * ;

[0069] Among them, f(x * )=Φ(x * )-x * It is the difference between the expected learned features and the input features, i.e., the residual, generated by the stack of convolutional layers. Residual learning is better than directly learning the ideal feature map Φ(x * ) is easier:

[0070]

[0071] in, is the loss function, and the output of the residual block is y * , is the loss on the input x * The gradient of , through the residual connection, the gradient directly contains a direct path The gradient is directly transferred from the back layer to the front layer to avoid gradient disappearance;

[0072] S2.1.5, global average pooling layer: perform global dimensionality reduction on the feature map, map the high-dimensional spatial features to low-dimensional global features, output a single feature value, and provide input for the final fire source location classification task;

[0073] S2.1.6, Fully connected layer, softmax function: After the output of the fully connected layer passes through the softmax function, the output is the category of the fire source identification task, that is, the label of the fire source location, that is, a number between 0 and p; the Softmax function is a normalization function used to map a real vector to a probability distribution; it is used here as the last layer of the multi-classification task to convert the input of the fully connected layer into a probability value in the range of [0, 1] to represent the probability of the fire source belonging to each location. For the output vector z of the fully connected layer i , the Softmax function converts it into a probability distribution P i , and satisfy:

[0074]

[0075] The formula of the Softmax function is as follows:

[0076]

[0077] The gradient calculation and back propagation of the Softmax function are based on the cross entropy loss function:

[0078]

[0079] Among them, γ is the cross entropy loss, yi is the target category, if it is the correct fire source location, it is 1, if it is the other fire source location, it is 0;

[0080] The gradient of Softmax output is:

[0081]

[0082] The output gradient is used for back propagation to continuously optimize the model output results;

[0083] S2.2. Divide the image database into training set and validation set according to the ratio, train the Resnet-50 deep learning model, adjust the number of training times until the validation accuracy and loss reach the set values, and complete the Resnet-50 deep learning model training.

[0084] Furthermore, the specific process of S3 is as follows:

[0085] S3.1. Select grayscale color matching for the infrared thermal imager. Aim the infrared thermal imager lens at the outer surface of the glass in the fire room and keep facing the center of the glass to collect grayscale infrared thermal images. If the infrared thermal imager does not have grayscale color matching, convert the image to a grayscale image before performing subsequent operations. The color image pixel is composed of R, G, and B of the RGB channel. The grayscale value Y corresponding to the pixel is:

[0086] Y=w R ·R+w G ·G+w B B;

[0087] Among them, w R is the weighting coefficient of the R channel, w G is the weighting coefficient of the G channel, w B is the weighting coefficient of the B channel;

[0088] S3.2, input the grayscale infrared thermal image obtained by the infrared thermal imager into the Resnet-50 deep learning model trained in S2;

[0089] S3.3. The Resnet-50 deep learning model outputs the label of the fire source location, which is a number between 0 and p to determine the fire source location.

[0090] The present invention constructs an image database, corresponds each image in the image database to a specific "fire source location" label, and uses the constructed image database to train the residual neural network Resnet-50 deep learning model to complete the subsequent image recognition task. In the early stage of fire development, when conducting fire reconnaissance of super high-rise building fires, control the reconnaissance drone carrying an infrared thermal imager to fly to the outside of the glass curtain wall of the fire floor, adjust the position of the reconnaissance drone to obtain the complete infrared thermal image of the outer surface of the glass curtain wall of the fire floor, and input the infrared thermal image of the outer surface of the glass curtain wall obtained by the infrared thermal imager into the trained Resnet-50 deep learning model. The model gives the location number of the fire source after analysis, realizes the accurate identification of the fire source location in the early stage of fire development, has strong real-time reliability, and greatly improves the fire warning capability and rescue efficiency of super high-rise buildings. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 It is a schematic diagram of the Resnet-50 deep learning model workflow of the present invention;

[0092] Figure 2 This is a Resnet-50 deep learning model training result diagram of the present invention;

[0093] Figure 3 is an infrared thermal image of the outer surface of the glass curtain wall obtained by the infrared thermal imager facing the glass in the embodiment of the present invention;

[0094] Figure 4 It is an infrared thermal image of the outer surface of the glass curtain wall obtained by the infrared thermal imager facing the glass sideways in the embodiment of the present invention. DETAILED DESCRIPTION

[0095] The present invention will be further described below in conjunction with the accompanying drawings.

[0096] A method for locating indoor fire sources based on infrared thermal images of the outer surface of a glass curtain wall comprises the following steps:

[0097] S1, database construction, mainly including the construction of fire data set, temperature distribution of the outer surface of glass curtain wall, image extraction and processing;

[0098] S2, build and train the Resnet deep learning model; Figure 1 As shown in the figure, the Resnet-50 deep learning model includes an input layer, a convolutional layer, a maximum pooling layer, multiple groups of residual blocks, a global average pooling layer, and a fully connected layer including a softmax function. After the above layers, the Resnet deep learning model is continuously optimized and divided into a training set and a validation set in a ratio of 7:3. The training is completed until the validation accuracy and loss reach the set value, as shown in the figure. Figure 2 As shown, Figure 2 (a) to (f) are respectively the loss value result graphs and the training and verification accuracy value result graphs of three batches (each batch contains three epochs) of the Resnet-50 deep learning model trained using the method of the present invention, a and b are the result graphs of one batch, c and d are the result graphs of one batch, and e and f are the result graphs of one batch;

[0099] S3, using the Resnet deep learning model to locate the fire source, such as Figure 3 and Figure 4 As shown;

[0100] Embodiment: A flashover prediction experiment was carried out based on an existing full-scale fire experimental device. The FDS fire simulation model was consistent with the actual full-scale fire experimental layout. A total of 9 groups of fire source positions were set up, and multiple groups of full-scale fire experiments were carried out. 20 grayscale thermal images of the glass outer surface were selected under experimental conditions at different fire source positions and input into the trained Resnet-50 deep learning model. The experimental data table is as follows:

[0101]

[0102]

[0103] Through the above examples, the following conclusions are drawn: in the test of 20 grayscale thermal images of the outer surface of glass, the prediction accuracy reaches 95%. Therefore, the fire source location prediction technology of the present invention, which integrates computational fluid dynamics, finite element analysis simulation technology, deep learning technology, and infrared thermal imaging technology, is suitable for fire reconnaissance tasks in high-rise building fires. Compared with other fire source location methods, the present invention has higher universality, timeliness, and efficiency, and also has a higher level of accuracy.

Claims

1. A method for locating indoor fire sources based on infrared thermal images of the outer surface of a glass curtain wall, characterized in that: The steps include: S1, database construction, mainly including the construction of fire data set, temperature distribution of the outer surface of glass curtain wall, image extraction and processing; S2. Build and train the Resnet deep learning model; S3. Use the Resnet deep learning model to locate the fire source.

2. The method for locating indoor fire sources based on the infrared thermal image of the outer surface of the glass curtain wall according to claim 1 is characterized in that: The specific process of S1 is as follows: S1.

1. Use fire computational fluid dynamics simulation software FDS to simulate fire data and construct a fire data set. The specific process is as follows: S1.1.

1. Build a fire combustion model: set parameters, including grid parameters, geometric structure of the burning room, surface parameters, reaction parameters, initial ambient temperature T0, fire source location p, fire source heat release rate q, combustible smoke production rate f, ventilation opening opening time l; set a total of p×q×f×l groups of simulation conditions; S1.1.

2. Set n×n temperature measuring points on the inner surface of the fire room glass in a square array with equal rows and columns; S1.1.3, set the simulation time t to simulate, and obtain the data set of n×n temperature measurement points {T nt }, T nt It represents the temperature data of the nth measuring point at the tth second. The time-temperature data sets of different temperature measuring points are as follows: S1.1.

3. Integrate the time-temperature data set into a complete FDS simulation data set as input data for subsequent finite element analysis; S1.

2. Use transient thermal analysis of computational fluid dynamics software ANSYS and fire computational fluid dynamics simulation software FDS to simulate the temperature load on the inner surface of the glass curtain wall and perform heat conduction simulation to obtain the temperature distribution on the outer surface of the glass curtain wall. The specific process is as follows: S1.2.

1. Use SpaceCliam to build a glass model with the same geometric dimensions as the FDS fire combustion model; S1.2.

2. Import the constructed glass model into transient thermal to complete the import of the geometric structure; S1.2.

3. Input the key material parameters of the glass model for subsequent thermal conductivity analysis, including: density, thermal conductivity, specific heat, and thermal expansion coefficient; S1.2.

4. Perform meshing. The parameters of meshing include: physical preference of the unit, unit order, unit size, number of nodes and units; S1.2.

5. Set the initial temperature of the glass model to be consistent with the initial ambient temperature T0 in the FDS fire combustion model; S1.2.6, set the duration of ANSYS finite element analysis to be consistent with the duration t of FDS simulation; S1.2.

7. Set the convection heat transfer coefficient on the outer surface of the glass model; S1.2.

8. Input the inner surface temperature data of the glass curtain wall obtained by FDS simulation as the "temperature" load into the inner surface of the ANSYS glass model. The spatial position of each input "temperature" load is consistent with the spatial position of the temperature measurement point set in the FDS simulation. That is, the temperature measurement point data of the nth FDS simulation corresponds to the nth AYSYS temperature load input. The nth temperature load input into the ANSYS glass model is: S1.2.

9. Perform ANSYS finite element analysis to obtain the temperature change on the outer surface of the glass model. This process is based on the heat conduction equation, and the heat conduction process is represented by the following calculus equation: Where ρ is the density of the glass model; c is the specific heat capacity of the glass model; T is the function of temperature changing with time and space; t * is time; k is the thermal conductivity of the material; Q is the thermal power density of the internal heat source; At the initial time t * = 0 is T(x, y, z, 0) = T0, and n×n temperature loads are applied on one side of the glass model. The boundary conditions are expressed as: T(x i ,y i ,t)=T boundary (x i ,y i ,t); Among them, T boundary represents the n×n temperature loads applied on the inner surface of the glass model; x i and i is the coordinate position of the temperature load on the inner surface of the glass model; the convection heat transfer boundary condition set on the outer surface of the glass model is expressed as: in, represents the temperature gradient of the glass model surface in the normal direction; h is the convection heat transfer coefficient; T * is the temperature of the outer surface of the glass model; T ∞ is the temperature of the ambient fluid at the outer surface of the glass; During the finite element solution process, the convection boundary conditions will be discretized and the solution equation is as follows: Among them, [C conv ] represents the contribution of convective heat transfer, [M] is the mass matrix, which is proportional to the specific heat capacity of the glass material and the volume of the unit; [K] is the stiffness matrix, which represents the thermal stiffness of the model system, that is, the resistance of the glass model to heat transfer under unit temperature change; {Q} is the load vector, which represents the applied temperature load; {T} is the temperature that changes with time and space; S1.2.

10. Select "Inverse Grayscale" as the color scheme for the temperature solution of the glass model's outer surface. S1.3, Image extraction and processing: S1.3.

1. Extract the temperature distribution diagram of the outer surface of the glass model of all frames according to the frame number of ANSYS finite element analysis results; S1.3.2, extract the temperature distribution diagram of the outer surface of all glass models under the working conditions of p×q×f×l group, and construct an unprocessed original image database; S1.3.

3. Perform Gaussian blur processing on all images in the original image database. The Gaussian filter of the two-dimensional image of the outer surface temperature image of the glass model is expressed as: Among them, x and y are the horizontal and vertical distances between the image and the center pixel respectively; G(x, y) is the output value of the Gaussian filter, which is the discretization form of the two-dimensional Gaussian function; δ is the standard deviation; when performing Gaussian blur processing, the Gaussian filter is used to continue the convolution operation on the image. For a glass model outer surface temperature distribution image I(x, y) and a Gaussian filter G(x, y), the convolution result I'(x, y) is: In the formula, G(i,j) is a two-dimensional Gaussian filter kernel, which represents the weight of the distance to the center pixel; Reduce the sharp edges and details of the image by weighted averaging each pixel and its neighborhood; S1.

4. Perform Laplace sharpening calculation on the temperature distribution image of the outer surface of the glass model after Gaussian blur processing. Laplace sharpening combines the Laplace operator and the addition operation of the image. The Laplace operator is a second-order derivative operator and is defined in two-dimensional space as: Among them, f(x,y) is the pixel value of the image at the point (x,y), which represents the Laplace operator. In discrete images, the Laplace operator is approximated by a filter, and the common Laplace filter is used to approximate the Laplace operator: The Laplace filter calculates the neighborhood difference of each pixel in the image to highlight the edges and details of the image; Laplace sharpening combines the original pixel values ​​of the image with the output of the Laplace operator. Its mathematical expression is as follows: f sharp (x,y)=f(x,y)-α·Δf(x,y); Where f(x,y) is the pixel value of the original image; Δf(x,y) is the result of applying the Laplacian operator to the image f(x,y); α is a parameter that controls the sharpening intensity and is a constant; The steps for Laplace sharpening of the image database are as follows: S1.4.

1. Calculate the Laplace operator result of the image f(x,y) to obtain the edge information in the image; S1.4.2, select the control parameter α according to the intensity of sharpening; S1.4.

3. Combine the original image with the Laplace operator result to obtain the sharpened image f sharp (x,y); S1.4.

4. Set a label of the fire source location for each image in the image database after image processing, that is, a number between 0 and p, to complete the construction of the glass model outer surface temperature image database.

3. The method for locating indoor fire sources based on the infrared thermal image of the outer surface of the glass curtain wall according to claim 1 or 2, characterized in that: The specific process of S2 is: S2.1, using the Resnet-50 deep learning model, using the glass model outer surface temperature distribution image database built in S1 to train the Resnet-50 deep learning model, extracting deep features from the temperature distribution image. The model structure includes: S2.1.1, Input layer: The images input from the image database are adjusted for suitability in size through this layer. The images in the image database are all generated using the "reverse grayscale" color scheme. The images used in training are all grayscale images and are all single-channel; S2.1.2, convolution layer: This layer performs preliminary feature extraction and reduces the spatial size of the output feature map; S2.1.3, Max Pooling Layer: Reduces the spatial dimension of the image, takes the maximum value in each local area, retains the most significant features, and makes feature extraction more robust to small changes in the input; S2.1.4, multiple groups of residual blocks: Each residual block includes an identity block and a convolution block to avoid gradient vanishing and degradation in deep neural networks; for feature mapping: Φ(x * )=f(x * )+x * ; Among them, f(x * )=Φ(x * )-x * It is the difference between the expected learned features and the input features, i.e., the residual, generated by the stacking of convolutional layers. Residual learning is faster than directly learning feature mapping Φ(x * ) is easier: in, is the loss function, and the output of the residual block is y * , is the loss on the input x * The gradient of , through the residual connection, the gradient directly contains a direct path The gradient is directly transferred from the back layer to the front layer to avoid gradient disappearance; S2.1.5, global average pooling layer: perform global dimensionality reduction on the feature map, map the high-dimensional spatial features to low-dimensional global features, output a single feature value, and provide input for the final fire source location classification task; S2.1.6, Fully connected layer, softmax function: After the output of the fully connected layer passes through the softmax function, the output is the category of the fire source identification task, that is, the label of the fire source location, that is, a number between 0 and p; the Softmax function is a normalization function used to map a real number vector to a probability distribution; for the output vector z of the fully connected layer i , the Softmax function converts it into a probability distribution P i , and satisfy: The formula of the Softmax function is as follows: The gradient calculation and back propagation of the Softmax function are based on the cross entropy loss function: Among them, γ is the cross entropy loss, y i is the target category, which is 1 if it is the correct fire source location and 0 if it is any other fire source location; The gradient of Softmax output is: The output gradient is used for back propagation to continuously optimize the model output results; S2.

2. Divide the image database into training set and validation set according to the ratio, train the Resnet-50 deep learning model, adjust the number of training times until the validation accuracy and loss reach the set values, and complete the Resnet-50 deep learning model training.

4. The method for locating indoor fire sources based on infrared thermal images of the outer surface of a glass curtain wall according to claim 3, characterized in that: The specific process of S3 is as follows: S3.

1. Select grayscale color matching for the infrared thermal imager. Aim the infrared thermal imager lens at the outer surface of the glass in the fire room and keep facing the center of the glass to collect grayscale infrared thermal images. If the infrared thermal imager does not have grayscale color matching, convert the image to a grayscale image before performing subsequent operations. The color image pixel is composed of R, G, and B of the RGB channel. The grayscale value Y corresponding to the pixel is: Y=w R ·R+w G ·G+w B ·B; Among them, w R is the weighting coefficient of the R channel, w G is the weighting coefficient of the G channel, w B is the weighting coefficient of the B channel; S3.2, input the grayscale infrared thermal image acquired by the infrared thermal imager into the Resnet-50 deep learning model trained in S2; S3.3, the Resnet-50 deep learning model outputs the label of the fire source location, which is a number between 0 and p to determine the fire source location.

Images