Indoor fire source location method based on infrared thermal images of the exterior surface of a glass curtain wall

Abstract

A method for identifying the location of an indoor fire source based on infrared thermal images of the exterior surface of a glass curtain wall is provided. The solution includes the following steps: data structure construction, which mainly involves building a fire dataset, temperature distribution of the glass curtain wall exterior surface, and image extraction and processing; building and training a Resnet deep learning model; and identifying the fire source location using the Resnet deep learning model. In the early stages of a fire, a reconnaissance drone equipped with an infrared thermal imaging camera is operated to fly outside the glass curtain wall of the fire floor, and the position of the reconnaissance drone is adjusted to capture a complete infrared thermal image of the glass curtain wall exterior surface of the fire floor. The infrared thermal image of the glass curtain wall exterior surface captured by the infrared thermal imaging camera is input into the trained Resnet-50 deep learning model, which then analyzes it and outputs the location number of the fire source.

Description

[Technical Field]

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

[0002] As urbanization and town development accelerate, super-tall buildings are on the rise. While providing convenience, super-tall buildings also pose serious fire safety risks due to their enormous scale, complex structures, and high density.

[0003] After a fire breaks out in a high-rise building, a layer of smoke accumulates inside the fire chamber and gradually settles. A certain thickness of smoke layer forms in a short time, significantly reducing visibility inside the room and making it impossible to accurately identify the source of the fire. Most existing high-rise buildings use glass curtain walls for their exteriors, which prevent outdoor infrared thermal imaging cameras from accurately identifying the temperature field inside the room, making it difficult to conduct outdoor fire reconnaissance. This affects the efficiency of fire rescue in high-rise buildings, causing the fire to spread and resulting in serious casualties and property damage.

[0004] Currently, early fire localization technologies rely on equipment inside a building during a fire. Indoor fire localization technologies primarily include visual fire localization, wireless sensor network localization, and fiber optic sensor localization. These localization technologies all require fixed location sensors, which must be installed during building construction. Fixed placement is inflexible and has a high probability of malfunction during the fire process. Furthermore, these localization technologies are time-consuming and cannot meet the real-time requirements of practical application scenarios. Summary of the Invention [Problem to be solved by the invention]

[0005] The present invention provides a method for locating an indoor fire source based on infrared thermal images of the exterior surface of a glass curtain wall, which can accurately identify the fire source location in the early stage of fire development, has excellent real-time performance and reliability, and improves the fire warning capability and rescue efficiency of high-rise buildings. [Means for solving the problem]

[0006] In order to achieve the above object, the present invention provides a method for identifying an indoor fire source location based on an infrared thermal image of the outer surface of a glass curtain wall, S1, Data structure construction: This step mainly includes the construction of fire datasets, temperature distribution on the outer surface of the glass curtain wall, and image extraction and processing; S2, the construction and training step of the Resnet deep learning model; S3, a fire source localization step using the Resnet deep learning model.

[0007] Furthermore, the specific steps of S1 are as follows: S1.1. Fire science computational fluid dynamics simulation software FDS is used to simulate fire data and build a fire dataset. The specific process is as follows: S1.1.1. Construct a fire combustion model: Set parameters including mesh parameters, geometric structure of the fire room (including glass curtain wall and ventilation openings), surface parameters, reaction parameters, initial ambient temperature T0, fire source position p, fire source heat release rate q, combustible smoke generation rate f, and ventilation opening opening time l, and set a total of p × q × f × l sets of simulation conditions. S1.1.2. Set n x n temperature measurement points on the inner surface of the fire chamber glass in a square matrix with equal rows and columns. S1.1.3: Set the simulation time t and perform the simulation. nt} and T nt represents the temperature data at the nth measurement point at the tth second, and the time-temperature data set for different temperature measurement points is as follows:

number

number

number

number

number

number

number

number

number

number

number

[0008] Furthermore, the specific process of S2 is as follows: S2.1. The Resnet-50 deep learning model is adopted, and the glass model outer surface temperature distribution image database constructed in S1 is used to train the Resnet-50 deep learning model to extract deep features from the temperature distribution image. The model structure is as follows: S2.1.1: This layer adjusts the image size adaptively based on the input image from the image database. All images in the image database are generated using the "inverted grayscale" color scheme. All images input to the training are grayscale images and are all single-channel. S2.1.2, a convolutional layer that performs initial feature extraction and reduces the spatial size of the output feature map; S2.1.3, a max pooling layer that reduces the spatial dimension of the image, taking the maximum value in each local region to retain the most salient features while making feature extraction more robust to small changes in the input; S2.1.4, each residual block includes an identity block (directly propagating the input to the output and adding it with the convolution result) and a convolution block (matching the dimensions of the input and output, this model uses a certain size of convolution to perform dimension expansion or dimension reduction), which is used to avoid gradient vanishing and degeneration in deep neural networks, and for the ideal feature mapping:

number

number

number

number

number

number

number

[0009] Furthermore, the specific process of S3 is as follows: S3.1. Select a grayscale color scheme for the infrared thermal imager, and aim the infrared thermal imager lens at the outer surface of the glass of the fire room and directly at the center of the glass to collect a grayscale infrared thermal image. If the infrared thermal imager does not have a grayscale color scheme, the image should first be converted to a grayscale image before further operations. The pixel points of the color image are composed of R, G, and B of the RGB channels, and the grayscale value Y corresponding to this pixel point is expressed as follows:

number

[0010] This invention constructs an image database, associates each image in the image database with a specific "fire source location" label, and uses the constructed image database to train a residual neural network (Resnet-50) deep learning model to complete subsequent image recognition tasks. To conduct fire situation reconnaissance in a high-rise building during the early stages of fire development, a reconnaissance drone equipped with an infrared thermal imaging camera is operated to fly outside the glass curtain wall of the fire floor and adjust the drone's position to capture a complete infrared thermal image of the exterior surface of the glass curtain wall of the fire floor. The infrared thermal image of the exterior surface of the glass curtain wall captured by the infrared thermal imaging camera is input into the trained Resnet-50 deep learning model, which then analyzes and outputs a fire source location number. This enables accurate identification of the fire source location during the early stages of fire development, with excellent real-time and reliability, significantly improving the fire warning capability and rescue efficiency of high-rise buildings. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a schematic diagram of the workflow of the Resnet-50 deep learning model of the present invention. [Figures 2(a)-2(f)] FIG. 1 is a training result diagram of the Resnet-50 deep learning model of the present invention. [Figure 3] 1 is an infrared thermal image of the outer surface of a glass curtain wall taken by an infrared thermal imaging camera facing directly at the glass in an embodiment of the present invention. [Figure 4] 1 is an infrared thermal image of the outer surface of a glass curtain wall taken by an infrared thermal imaging camera directed obliquely at the glass in an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0012] The present invention will now be further described with reference to the drawings.

[0013] The method for identifying the location of an indoor fire source based on infrared thermal images of the exterior surface of a glass curtain wall is as follows: S1, Data structure construction: This step mainly includes the construction of fire datasets, temperature distribution on the outer surface of the glass curtain wall, and image extraction and processing; S2. Construction and training of Resnet deep learning model: As shown in Figure 1, the Resnet-50 deep learning model includes an input layer, a convolutional layer, a max pooling layer, multiple sets of residual blocks, a global average pooling layer, and a fully connected layer including a softmax function. The Resnet deep learning model is continuously optimized through each of the above layers, and divided into a training set and a validation set in a 7:3 ratio. Training is continued until the validation accuracy and loss reach the set values. The training of the Resnet-50 deep learning model is completed, as shown in Figure 2. Here, Figures 2(a) to 2(f) show the loss value result chart and the training and validation accuracy result chart for three batches (each batch contains three epochs) of training the Resnet-50 deep learning model using the method of the present invention. a, b are the result charts of one batch, c, d are the result charts of one batch, and e, f are the result charts of one batch. S3, a fire source localization step using the Resnet deep learning model, as shown in Figures 3 and 4.

[0014] Example: A flashover prediction experiment was conducted based on an existing full-scale fire experiment device. The FDS fire simulation model was matched to the actual full-scale fire experiment layout, with a total of nine sets of fire source locations set up to conduct multiple full-scale fire experiments. Twenty grayscale thermal images of the outer glass surface under experimental conditions with different fire source locations were selected and input into the trained Resnet-50 deep learning model, resulting in the following experimental data table:

[0015] [Table 1]

[0016] Through the above example, we can conclude that the prediction accuracy reached 95% in the test of 20 grayscale thermal images of the outer surface of glass. Therefore, the fire source location prediction technology of this invention, which combines computational fluid dynamics, finite element analysis simulation technology, deep learning technology, and infrared thermal imaging technology, is applicable to the fire situation reconnaissance mission in high-rise building fires. Compared with other fire source location methods, this invention has high versatility, speed, and efficiency, while also achieving a high level of accuracy.

Claims

1. A method for identifying an indoor fire source location based on an infrared thermal image of the outer surface of a glass curtain wall, comprising: S1, database construction: a step including construction of fire dataset, temperature distribution on the outer surface of the glass curtain wall, image extraction and processing; S2, Resnet deep learning model construction and training step; S3, a fire source localization step using a Resnet deep learning model; S1.

2. Using the transient thermal of the computational fluid dynamics software ANSYS, the temperature load on the inner surface of the glass curtain wall obtained by the simulation of the fire science computational fluid dynamics simulation software FDS is analyzed, and a heat conduction simulation is performed to obtain the temperature distribution on the outer surface of the glass curtain wall. The specific process is as follows: S1.2.1, SpaceCliam was used to construct 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 and complete the import of the geometric structure; S1.2.3, input the key material parameters of the glass model, including density, thermal conductivity, specific heat, and thermal expansion coefficient, to be used in the subsequent heat transfer analysis; S1.2.4: Mesh division is performed. The parameters to be considered in mesh division include the physical preference of the element, the order of the element, the element size, the number of nodes, and the number of elements. S1.2.5, set the initial temperature of the glass model, and the initial environmental temperature T ０ and S1.2.6, set the time length of the ANSYS finite element analysis to match the time length t of the FDS simulation; S1.2.7, set the convection heat transfer coefficient on the outer surface of the glass model; S1.2.

8. The glass curtain wall inner surface temperature data obtained by the FDS simulation is input as a "temperature" load onto the inner surface of the ANSYS glass model, and the spatial position of each input "temperature" load is matched 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 ANSYS temperature load input, and the nth temperature load input into the ANSYS glass model is as follows: [Equation 1] T ｎｔ represents the temperature data at the nth measurement point at the tth second, S1.2.9, ANSYS finite element analysis is performed 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 expressed by the following differential and integral equation: [Equation 2] where ρ is the density of the glass model, c is the specific heat capacity of the glass model, T is a function of temperature change over time and space, and t ＊ is the time, k is the thermal conductivity of the material, and Q is the thermal power density of the internal heat source. initial time t ＊ The temperature distribution at 0 is T(x, y, z, 0) = T ０ and n x n temperature loads are applied to one side of the glass model, and the boundary conditions are expressed as follows: [Equation 3] Here, T 境界 represents the n × n temperature loads applied on the inner surface of the glass model, and x ｉ and y ｉ is the coordinate position of the temperature load on the inner surface of the glass model, and the convective heat transfer boundary condition set on the outer surface of the glass model is expressed as follows: [Equation 4] where ∂T / ∂n represents the temperature gradient in the normal direction of the glass model surface, h is the convection heat transfer coefficient, and T ＊ is the temperature of the outer surface of the glass model, T ∞ is the temperature of the environmental fluid on the outer surface of the glass, In the finite element solution process, the convection boundary conditions are discretized and the solution equation is: [Equation 5] Here, [C ｃｏｎｖ ] represents the contribution term of convective heat transfer, [M] is a mass matrix proportional to the specific heat capacity of the glass material and the volume of the element, [K] is a stiffness matrix representing the thermal stiffness of the model system, i.e., the ability of the glass model to resist heat transfer to a unit temperature change, {Q} is a load vector representing the applied temperature load, and {T} is the temperature that changes with time and space. S1.2.10: The method for locating an indoor fire source based on an infrared thermal image of the outer surface of a glass curtain wall, characterized in that the color scheme of the glass model outer surface temperature solution result is selected to be "inverted grayscale."

2. The specific steps of S1 are as follows: S1.

1. Fire science computational fluid dynamics simulation software FDS is used to simulate fire data and build a fire dataset. The specific process is as follows: S1.1.

1. Construct a fire combustion model: mesh parameters, fire chamber geometry, surface parameters, reaction parameters, initial environmental temperature T ０ , p fire source positions, q fire source heat release rates, f combustible smoke generation rates, and l ventilation opening opening times are set, and a total of p x q x f x l sets of simulation conditions are set, S1.1.

2. Set n x n temperature measurement points on the inner surface of the fire chamber glass in a square matrix with equal rows and columns; S1.1.3: Set a simulation time t and perform a simulation to obtain a data set of n × n temperature measurement points {T ｎｔ } and T ｎｔ represents the temperature data at the nth measurement point at the tth second, and the time-temperature data sets for different temperature measurement points are as follows: [Equation 6] S1.1.4, combining the time-temperature data set into a complete FDS simulation data set as input to subsequent finite element analysis; S1.

3. Image extraction and processing: S1.3.1, based on the frame number of the ANSYS finite element analysis result, extract the glass model outer surface temperature distribution map for all frame numbers; S1.3.2 Extract all the glass model outer surface temperature distribution maps under the p x q x f x l set of conditions, and construct an unprocessed original image database; S1.3.3, Gaussian blurring is performed on all images in the original image database, and the Gaussian filter of the two-dimensional image of the glass model outer surface temperature image is expressed as follows: [Equation 7] where 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 discretized form of the two-dimensional Gaussian function; and σ is the standard deviation. When performing Gaussian blurring, the Gaussian filter is used to continuously perform convolution 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 given by the following equation: [Equation 8] where G(i,j) is a two-dimensional Gaussian filter kernel, representing the weight for the distance from the center pixel; Reduces sharp edges and details in an image through a weighted average of each pixel and its neighboring regions. S1.4: Laplace sharpening is performed on the glass model outer surface temperature distribution image after Gaussian blurring. Laplace sharpening combines the Laplace operator with the image addition operation. The Laplace operator is a second-order differential operator, which is defined in two-dimensional space as follows: [Equation 9] where f(x, y) is the pixel value of the image at point (x, y) and represents the Laplacian operator. In a discrete image, the Laplacian operator is approximated by a filter. A general Laplacian filter is used to approximate the Laplacian operator: [Equation 10] The Laplace filter calculates the difference between neighboring pixels in an image, highlighting edges and details in the image. Laplacian sharpening combines the original pixel values ​​of an image with the output of a Laplacian operator, and its mathematical expression is as follows: [0011] where f(x,y) is the pixel value of the original image, and Δf(x,y) is the result of applying the Laplace operator to the image f(x,y), α is a parameter that controls the strength of sharpening and is a constant, The steps for Laplacian sharpening on an image database are as follows: S1.4.1, calculate the Laplace operator result of the image f(x, y) to obtain edge information in the image; S1.4.2, select a control parameter α based on the strength of sharpening; S1.4.3, the original image and the result of the Laplace operator are combined to produce the sharpened image f 鋭い (x, y) is obtained, S1.4.4: For each image in the image database after image processing, set a fire source location label, i.e., a number from 0 to p, to complete the construction of the glass model outer surface temperature image database.

3. The specific steps of S2 are as follows: S2.1: The Resnet-50 deep learning model is adopted, and the glass model outer surface temperature distribution image database constructed in S1 is used to train the Resnet-50 deep learning model, and deep features are extracted from the temperature distribution image. The model structure is as follows: S2.1.1: This layer adjusts the image size adaptively based on the input image from the image database. All images in the image database are generated using the "inverted grayscale" color scheme. All images input to the training are grayscale images and are all single-channel. S2.1.2, a convolutional layer that performs initial feature extraction and reduces the spatial size of the output feature map; S2.1.3, a max pooling layer that reduces the spatial dimension of the image, taking the maximum value in each local region to retain the most salient features while making feature extraction more robust to small changes in the input; S2.1.4, each residual block includes an identity block and a convolution block, which are used to avoid gradient vanishing and degeneration in deep neural networks, for feature mapping: [0012] where: [0013] is the difference between the expected feature and the input feature, i.e., the residual, which is generated by stacking convolutional layers. Residual learning is done by the feature mapping Φ(x ＊ ) is easier than learning it directly, [0014] where ζ is the loss function, and the output of the residual block is y ＊ , ∂ζ / ∂x ＊ is the input x ＊ is the gradient of the loss with respect to , and through the residual connection, the gradient is the direct path ∂ζ / ∂y ＊ A set of residual blocks directly including the gradients, and the gradients are directly transmitted from the back layer to the front layer, avoiding gradient vanishing; S2.1.5, a global average pooling layer that performs global dimensionality reduction on the feature map, mapping high-dimensional spatial features to low-dimensional global features, outputting a single feature value and providing input to the final fire source location classification task; S2.1.6, a fully connected layer, including a softmax function. The output result of the fully connected layer passes through the softmax function, which outputs the category of the fire source identification task, i.e., the label of the fire source location, a number from 0 to p. The softmax function is a normalization function used to map a real vector to a probability distribution. The output vector z of the fully connected layer is ｉ The Softmax function maps it to the probability distribution P ｉ and satisfy the following equation: [Equation 15] The gradient calculation and backpropagation of the Softmax function is based on the cross-entropy loss function: [0016] where γ is the cross entropy loss, y ｉ is the target category, which is 1 for the correct fire source location and 0 for other fire source locations, The gradient of the Softmax output is: [Equation 17] The output gradients are used in backpropagation to achieve continuous optimization of the model output. S2.

2. The method for identifying an indoor fire source location based on infrared thermal images of the outer surface of a glass curtain wall according to claim 1 or 2, characterized in that: the image database is proportionally divided into a training set and a validation set, and the Resnet-50 deep learning model is trained; and the number of training rounds is adjusted until the validation accuracy and loss both reach the set values, thereby completing the training of the Resnet-50 deep learning model.

4. The specific steps of S3 are as follows: S3.

1. Select a grayscale color scheme for the infrared thermal imager, and aim the lens of the infrared thermal imager at the outer surface of the glass of the fire room and directly face the center of the glass to collect a grayscale infrared thermal image. If the infrared thermal imager does not have a grayscale color scheme, the image must first be converted to a grayscale image before further operations. The pixel points of the color image are composed of R, G, and B of the RGB channels, and the grayscale value Y corresponding to this pixel point is expressed as follows: [Equation 18] Here, w Ｒ is the weighting coefficient for the R channel, w Ｇ is the weighting coefficient for the G channel, w Ｂ is the weighting factor for the B channel, S3.2, the grayscale infrared thermal image acquired by the infrared thermal imaging camera is input into the Resnet-50 deep learning model trained in S2; S3.3: The method for identifying an indoor fire source location based on an infrared thermal image of the outer surface of a glass curtain wall according to claim 3, characterized in that the Resnet-50 deep learning model outputs a label for the fire source location, i.e., a number from 0 to p, to determine the fire source location.

Images