A method to enhance fire identification capabilities of smart construction sites using AI large models

Abstract

The present invention discloses a method for enhancing the fire identification capability of a smart construction site by utilizing a large AI model, which relates to the field of fire identification and solves the technical problems of the existing fire identification technology in that the detection distance is limited and the detection of hidden fires is insufficient. The method includes a dark fire identification model construction step and a dark fire identification analysis step. The dark fire identification model construction step includes, S11: collecting dark fire training samples, the dark fire training samples include first eigenvalues of hidden heat sources, air expansion and distortion phenomena, and light refraction phenomena; S12: training the dark fire training samples through a deep convolutional neural network model. Compared with traditional fire identification algorithms, the present invention reduces dependence on smoke alarms and temperature fire detectors, and the detection distance of fires has been significantly increased. Timely dark fire warnings can prevent the occurrence of some fires and ensure project progress and construction safety.

Description

Technical Field

[0001] The present invention relates to the field of fire identification, and more specifically, to a method for enhancing the fire identification capability of smart construction sites by utilizing an AI large model. Background Art

[0002] Construction sites typically have a large number of workers and management personnel. A fire can result in serious casualties and severely impact project progress and safety. Therefore, effective fire prevention measures can protect workers' lives and are crucial for ensuring construction safety. Currently, fire prevention measures on construction sites primarily include implementing fire management systems, rationally planning site layouts, equipping with adequate firefighting equipment, and strengthening safety education. However, early detection and prevention of fires remain relatively weak.

[0003] Currently, fire detection and identification at smart construction sites primarily relies on high-definition cameras and supporting flame recognition algorithms, working in conjunction with smoke alarms and heat detectors. Current flame detection and identification technology is highly dependent on camera quality and the efficiency and accuracy of the flame recognition algorithms. Furthermore, smoke alarms and heat detectors have limited detection ranges, making them incapable of long-distance detection.

[0004] Current smart construction site fire identification solutions are severely limited in detecting hidden fires, sometimes failing to detect them at all. Because hidden fires are concealed, lacking visible flames and smoke, they can go undetected for a long time. Furthermore, once a hidden fire develops to a certain extent, it has a high probability of becoming an open flame and eventually causing a fire. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the shortcomings of the existing technology and provide a method for enhancing the fire identification capability of smart construction sites by using a large AI model, which solves the technical problems of the existing fire identification technology with limited detection distance and insufficient detection of hidden fires.

[0006] The present invention discloses a method for enhancing the fire identification capability of a smart construction site by using an AI large model, which includes the following steps: constructing a dark fire identification model and analyzing the dark fire identification:

[0007] The dark fire identification model construction step includes:

[0008] S11: Collecting dark fire training samples, wherein the dark fire training samples include first eigenvalues of hidden heat sources, air expansion and distortion phenomena, and light refraction phenomena;

[0009] S12: Training the Dark Fire training sample using a deep convolutional neural network model to extract a first eigenvalue from the Dark Fire training sample, and evaluating the first eigenvalue using a first standard feature set to obtain a first optimal eigenvalue;

[0010] S13: Generate a dark fire recognition model using the first optimal eigenvalue as model data;

[0011] The dark fire identification and analysis step comprises:

[0012] S21: Collecting environmental video of the actual environment;

[0013] S22: The dark fire recognition model detects whether the video frame of the environmental video contains air distortion and expansion phenomenon and light refraction phenomenon according to the first optimal eigenvalue. If so, proceed to S23;

[0014] S23: Analyzing the air distortion and expansion phenomenon using the dark fire recognition model to obtain light parameters of the location where the air distortion and expansion phenomenon occurs in the video frame, and analyzing the heat source location and heat source temperature using the light parameters;

[0015] S24: Obtaining the size of the heat source according to the heat source position and the heat source temperature, and analyzing the danger level of the heat source according to the heat source temperature and the heat source size, and triggering a dark fire alarm when the danger level reaches a preset level threshold.

[0016] As a further improvement, in S12, evaluating the first eigenvalue using the first standard feature set to obtain a first optimal eigenvalue specifically includes:

[0017] The first eigenvalue is compared one by one with the first standard eigenvalues in the first standard feature set. If there is a first standard eigenvalue that is consistent with the first eigenvalue, the first eigenvalue is used as the first optimal eigenvalue.

[0018] Furthermore, it is detected whether the video frame contains air distortion and expansion phenomena and light refraction phenomena, specifically: the video frame is traversed according to the air expansion and distortion phenomena and light refraction phenomena in the first optimal eigenvalue. If the video frame contains characteristic phenomena that are consistent with the air expansion and distortion phenomena and light refraction phenomena in the first optimal eigenvalue, it is determined that air distortion and expansion phenomena and light refraction phenomena exist.

[0019] Furthermore, in S23, light parameters are obtained by analyzing the phenomenon of air distortion and expansion, specifically: the temperature and pressure in the actual environment are obtained, and the refractive index of the light is calculated according to the temperature and pressure; and the refractive angle of the light is calculated according to the refractive index.

[0020] Furthermore, in S23, the heat source position and the heat source temperature are analyzed by the light parameters, specifically: the light parameters are compared with the light parameter threshold of the light refraction phenomenon in the dark fire identification model. If the light parameters are consistent with the light parameter threshold, or the light parameters exceed the light parameter threshold, it is judged that the position corresponding to the light parameters is the heat source position, and the heat source temperature of the heat source position is obtained at the same time.

[0021] Furthermore, the method further comprises an open flame identification model construction step and an open flame identification analysis step;

[0022] The open flame recognition model construction step specifically includes:

[0023] S31: Collecting open fire training samples, where the open fire training samples include second eigenvalues of flame and smoke;

[0024] S32: training the open fire training sample using a deep convolutional neural network model to extract a second eigenvalue from the open fire training sample, and evaluating the second eigenvalue using a second standard feature set to obtain a second optimal eigenvalue;

[0025] S33: Generate an open flame recognition model using the second optimal characteristic value as model data;

[0026] The open flame identification and analysis step specifically includes:

[0027] The open fire recognition model detects whether there is flame and / or smoke in the video frame of the environmental video based on the second optimal eigenvalue; if so, the smoke eigenvalue, motion eigenvalue and texture eigenvalue are obtained, and it is determined whether an open fire alarm needs to be triggered based on the smoke eigenvalue, motion eigenvalue and texture eigenvalue.

[0028] Furthermore, the detection expression of the smoke characteristic value is:

[0029]

[0030] Among them, U (χ,n) Refers to the chromaticity value of the pixel at position χ in the n-th frame image I, V (χ,n) Refers to the concentration value of the pixel at position χ in the n-th frame image I, P(χ,n) is the smoke characteristic value of the pixel at position χ in the n-th frame image I,

[0031] Set a brightness threshold T1, where the brightness Y of the pixel at position x in the n-th frame image I is obtained as follows: (χ,n) The expression:

[0032]

[0033] When the brightness Y of the pixel at position x in the n-th frame image I is (χ,n)When the brightness threshold T1 is less than the value of the smoke characteristic value P(χ,n) is -1; when the brightness Y of the pixel at position χ in the n-th frame image I is (χ,n) When the brightness threshold T1 is greater than the value of the smoke characteristic value P(χ,n)

[0034] Furthermore, the detection expression of the motion feature value is:

[0035]

[0036] Where I(χ,n) refers to the intensity value of the pixel at position χ in the n-th frame image I, a is a smoothing constant between (0,1), B(χ,n+1) refers to the motion feature value of the n+1-th frame image, and B(χ,n) refers to the motion feature value of the n-th frame image;

[0037] When the motion feature value B(χ,n) of the nth frame image is greater than the preset motion feature threshold, B(χ,n+1)=B(χ,n); when the motion feature value B(χ,n) of the nth frame image is less than or equal to the motion feature threshold, then

[0038] B(χ,n+1)=aB(χ,n)+(1-a)I(χ,n);

[0039] The detection formula of the texture feature value is:

[0040]

[0041] Among them, B fast (χ,n) and B slow (χ,n) refers to the way of updating at different rates, B fast (χ,n) is updated once per frame, B slow (χ,n) is updated once per second, T high and T low Refers to the different brightness values set in the image attributes, A(χ,n) is the texture feature value; set a texture feature threshold range, when the rate difference B1=|B fast (χ,n)-B slow When (χ, n)| is greater than the texture feature threshold range, the texture feature value A(χ, n) is equal to 1; when the rate difference B1 is less than the texture feature threshold range, the texture feature value A(χ, n) is equal to -1; when the rate difference B1 is within the texture feature threshold range, the rate difference B1 is substituted into The texture feature value A(χ,n) is obtained.

[0042] Furthermore, whether it is necessary to trigger an open fire alarm is determined based on the smoke feature value, motion feature value and texture feature value. Specifically, the smoke feature value, motion feature value and texture feature value in the environmental video are compared one by one with the smoke standard feature value, motion standard feature value and texture standard feature value in the second standard feature set. If there are standard features that are inconsistent with the second standard feature set, an open fire alarm is triggered.

[0043] Beneficial effects

[0044] The advantages of the present invention are:

[0045] The present invention provides an open flame recognition model construction step, an open flame recognition analysis step, a dark fire recognition model construction step and a dark fire recognition analysis step, and repeatedly trains the collected training samples to generate a neural network model to obtain the location and range of the fire, thereby enhancing the flame recognition capability of the camera. Compared with traditional fire recognition algorithms, the present invention reduces the dependence on smoke alarms and temperature fire detectors, and significantly increases the detection distance of fire. Timely dark fire warnings can prevent the occurrence of some fires, thereby ensuring project progress and construction safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of the dark fire identification steps of the present invention;

[0047] Figure 2 This is a flow chart of the open flame identification steps of the present invention;

[0048] Figure 3 This is a schematic diagram of the derivation of the light path calculation formula in the present invention. DETAILED DESCRIPTION

[0049] The present invention will be further described below in conjunction with the embodiments, but this does not constitute any limitation to the present invention. Any limited number of modifications made by anyone within the scope of the claims of the present invention are still within the scope of the claims of the present invention.

[0050] See Figure 1-Figure 3 The present invention provides a method for enhancing the fire identification capability of a smart construction site by using an AI large model. The method includes a dark fire identification model construction step and a dark fire identification analysis step.

[0051] like Figure 1 As shown, the dark fire identification steps include:

[0052] The method includes a dark fire identification model construction step and a dark fire identification analysis step.

[0053] The steps to build the dark fire recognition model include:

[0054] S11: Collect dark fire training samples, where the dark fire training samples include first eigenvalues of hidden heat sources, air expansion and distortion phenomena, and light refraction phenomena.

[0055] S12: Training the Dark Fire training sample through a deep convolutional neural network model to extract a first eigenvalue from the Dark Fire training sample, and evaluating the first eigenvalue through a first standard feature set to obtain a first optimal eigenvalue.

[0056] In S12, the first eigenvalue is evaluated by the first standard feature set to obtain a first optimal eigenvalue, specifically including comparing the first eigenvalue with the first standard eigenvalues in the first standard feature set one by one. If there is a first standard eigenvalue that is consistent with the first eigenvalue, the first eigenvalue is used as the first optimal eigenvalue.

[0057] S13: Generate a dark fire recognition model using the first optimal eigenvalue as model data.

[0058] The dark fire identification and analysis steps include:

[0059] S21: Collect environmental video of the actual environment.

[0060] S22: The dark fire recognition model detects whether the video frame of the environmental video contains air distortion and expansion phenomena and light refraction phenomena based on the first optimal eigenvalue. If so, proceed to S23.

[0061] In S23, light parameters are obtained by analyzing the phenomenon of air distortion and expansion, specifically:

[0062] Obtain the temperature and pressure in the actual environment and calculate the refractive index of light based on the temperature and pressure;

[0063] Calculate the refraction angle of light based on the refractive index;

[0064] The light path of light in a continuously changing medium is obtained by the following formula:

[0065]

[0066] like Figure 3 As shown, in a continuously changing medium, the light path can be described in integral form. The above formula is derived from Monte Carlo integration:

[0067] exist Figure 3 The middle ab interval (i.e. Figure 3Take any point x in the gray area and get its corresponding f(x). Then, consider the function graph as a horizontal line, that is, its area becomes a rectangle with a base length of (ba) and a height of f(x). Use the area of such a rectangle to approximate the area of the irregular figure. Then repeat this process. We can get the areas of many rectangles, add them up and average them, and we can approximate the area of the function graph. Each time we get a rectangle is a sampling process, that is, if Figure 3 If this function value is sampled uniformly multiple times, it is actually equivalent to cutting the entire integral area into many rectangles and then adding up the areas of these small rectangles.

[0068] This leads to the definition of the definite integral function:

[0069]

[0070] Uniform random variable:

[0071]

[0072] This leads to the Monte Carlo integral equation:

[0073]

[0074] In order to obtain a smaller error value, importance sampling should also be considered. Of course, even if uniform distribution sampling is used, as long as there are enough sampling points, a relatively accurate result can still be obtained, but the performance overhead may increase. Therefore, uniform distribution is generally not used for sampling. Instead, a sampling distribution function that fits the overall shape and the integrand is used, which leads to a more general formula:

[0075]

[0076] The ability to identify dark fires is enhanced. The principle is mainly based on the expansion and distortion of air. Although dark fires do not have flames or smoke, there must be a heat source. As the temperature of the heat source rises, the air nearby expands due to the heat and the density becomes smaller, while the cold air outside the heat source has a higher density. This density difference will cause light to refract during propagation, thereby producing a visual effect of air expansion and distortion.

[0077] The calculation formula of air expansion coefficient is:

[0078] a=(1 / v)*(dV / dT);

[0079] Where a represents the thermal expansion coefficient, V represents the volume of the gas, T represents the temperature, dV represents the change in volume, and dT represents the change in temperature.

[0080] Refractive index calculation formula:

[0081] n=c / v;

[0082] Where n is the refractive index, c is the speed of light, and v is the speed of light in the medium.

[0083] The approximate calculation formula for atmospheric refraction under experimental conditions is:

[0084]

[0085] where n s is the refractive index of air in standard state, and n (T,p) It is the refractive index at a certain temperature and pressure, where pressure p is in torr (1 torr = 1 / 760 atm) and temperature T is in degrees Celsius.

[0086] Snell's law calculation formula:

[0087] nsin(θ1)=n′sin(θ2);

[0088] Where n is the refractive index of the incident medium, n' is the refractive index of the exit medium, θ1 is the incident angle, and θ2 is the refractive angle.

[0089] In S23, the heat source position and heat source temperature are analyzed by light parameters, specifically: the light parameters are compared with the light parameter threshold of the light refraction phenomenon in the dark fire recognition model. If the light parameters are consistent with the light parameter threshold, or the light parameters exceed the light parameter threshold, the position corresponding to the light parameters is judged to be the heat source position, and the heat source temperature of the heat source position is obtained at the same time.

[0090] A pre-trained convolutional neural network (CNN) model was trained on a large amount of video data showing heat sources and air expansion and distortion. During the training process, the large model extracts information such as air expansion and distortion, light refraction, and other information from the video frames to form model data that characterizes dark fire heat sources.

[0091] S23: Analyze the air distortion and expansion phenomenon through the dark fire recognition model to obtain light parameters of the position where the air distortion and expansion phenomenon occurs in the video frame, and analyze the heat source position and heat source temperature through the light parameters.

[0092] Detect whether the video frame contains air distortion and expansion phenomena and light refraction phenomena, specifically: traverse the video frame according to the air expansion and distortion phenomena and light refraction phenomena in the first optimal eigenvalue; if there are characteristic phenomena in the video frame that are consistent with the air expansion and distortion phenomena and light refraction phenomena in the first optimal eigenvalue, it is determined that the air distortion and expansion phenomena and light refraction phenomena exist.

[0093] S24: Obtain the size of the heat source according to the location and temperature of the heat source, and analyze the danger level of the heat source according to the temperature and size of the heat source. When the danger level reaches a preset level threshold, a dark fire alarm is triggered.

[0094] like Figure 2 As shown, the method further includes an open flame recognition model construction step and an open flame recognition analysis step;

[0095] The steps to build the open flame recognition model include:

[0096] S31: Collecting open fire training samples, where the open fire training samples include second eigenvalues of flame and smoke;

[0097] S32: Training the open flame training sample through a deep convolutional neural network model to extract a second eigenvalue from the open flame training sample, and evaluating the second eigenvalue through a second standard feature set to obtain a second optimal eigenvalue.

[0098] The enhanced fireworks recognition capability is mainly based on deep learning and video processing technology of large AI models. As a representative algorithm of deep learning, convolutional neural networks have representational learning capabilities and can extract high-order features from input information. In the fireworks recognition model, convolutional neural networks extract representative high-frequency and low-frequency features in the image and extract more general and complete features from the hidden layer signals.

[0099] The convolution operation is defined as:

[0100] [(I*K)(x,y)=sum{i=-infty}^{infty}sum{j=-infty}^{infty}I(i,j)K(xi,yi)];

[0101] Where (I) is the input image, (K) is the convolution kernel, and (x, y) represents the location of the output feature map.

[0102] Convolutional neural networks differ from conventional neural networks in that they contain multiple feature extractors composed of convolutional and pooling layers. In a convolutional layer of a convolutional neural network, a neuron is connected only to a subset of neurons in adjacent layers. A convolutional layer in a CNN typically contains several feature planes (feature maps), each composed of a rectangular array of neurons. Neurons in the same feature plane share weights, known as the convolution kernel. The convolution kernel is typically initialized with a random decimal matrix. During network training, the kernel learns reasonable weights. The direct benefit of sharing weights (convolution kernels) is that it reduces the number of connections between network layers and mitigates the risk of overfitting. Subsampling, also known as pooling, typically comes in two forms: mean pooling and max pooling. Subsampling can be considered a special form of convolution. Convolution and subsampling significantly simplify model complexity and reduce model parameters.

[0103] Formula derivation of neural network:

[0104]

[0105] This unit can also be called a logistic regression model. When multiple units are combined into a layered structure, a neural network model is formed. The training method for convolutional neural networks is similar to that of logistic regression, but due to its multi-layer nature, the chain rule is also used to derive the hidden layer nodes, i.e., gradient descent + chain rule.

[0106] The expression of convolution is:

[0107] S(t)=∫x(ta)w(a)da;

[0108] The discrete form is:

[0109]

[0110] The matrix form is:

[0111] s(t)=(X*W)(t);

[0112] Convolutional Neural Network (CNN) defines two-dimensional convolution as follows:

[0113]

[0114] Where W is our convolution kernel and X is our input. If X is a two-dimensional input matrix, W is also a two-dimensional matrix. If X is a multi-dimensional tensor, W is also a multi-dimensional tensor.

[0115] A pre-trained convolutional neural network (CNN) model (such as ResNet and VGG) is used as a feature extractor to train on large amounts of video data containing fireworks. During training, the large model extracts characteristic information about the fireworks in the video frames, such as color, motion, texture, and shape, to form model data for the fireworks' characteristics.

[0116] S33: Generate an open flame recognition model using the second optimal eigenvalue as model data.

[0117] Open flame identification and analysis steps:

[0118] The open fire recognition model detects whether there is flame and / or smoke in the video frame of the environmental video based on the second optimal eigenvalue; if so, the smoke eigenvalue, motion eigenvalue and texture eigenvalue are obtained, and it is determined whether an open fire alarm needs to be triggered based on the smoke eigenvalue, motion eigenvalue and texture eigenvalue.

[0119] The detection expression of motion feature value is:

[0120]

[0121] Where I(χ,n) refers to the intensity value of the pixel at position χ in the n-th frame image I, a is a smoothing constant between (0,1), B(χ,n+1) refers to the n+1-th frame image, and B(χ,n) refers to the n-th frame image.

[0122] When the motion feature value B(χ,n) of the nth frame image is greater than the preset motion feature threshold, B(χ,n+1)=B(χ,n); when the motion feature value B(χ,n) of the nth frame image is less than or equal to the motion feature threshold, B(χ,n+1)=aB(χ,n)+(1-a)I(χ,n).

[0123] The n+1th frame image B(χ,n+1) is estimated by a recursive formula involving frame I(χ,n) and image B(χ,n). The main algorithm used is the linear prediction algorithm. The following is the derivation principle:

[0124] Let B(χ,n) be the observed data of the nth frame, I(χ,n) be the intensity value of the pixel at χ in the nth frame, and I(χ,n+1) be the intensity prediction value of the pixel at χ in the n+1th frame. Assume that

[0125] I(x,n+1)=aB(x,n)+(1-a)I(x,n);

[0126] It can be proved

[0127] I(χ,n+1)=aB(χ,n)+a(1-a)B(χ,n-1)+a(1-a) 2B(χ, n-2)+...+a(1-a) n-1 B(χ, 1)+(1-a) n-1 l0;

[0128] Here, l0=I(χ, 1) is regarded as the initial value.

[0129] Detection formula of texture feature value:

[0130]

[0131] Among them, B fast (χ, n) and B slow (χ, n) refers to the way of updating at different rates, B fast (χ, n) is updated once per frame, B slow (χ, n) is updated once per second, T high and T low Refers to the different brightness values set in the image properties.

[0132] Among them, B fast (χ, n) and B slow (χ, n) refers to the way of updating at different rates, B fast (χ, n) is updated once per frame, B slow (χ, n) is updated once per second, T high and T low Refers to the different brightness values set in the image attributes, A(χ, n) is the texture feature value; set a texture feature threshold range, when the rate difference B1=|B fast (χ, n)-B slow When (χ, n)| is greater than the texture feature threshold range, the texture feature value A(χ, n) is equal to 1; when the rate difference B1 is less than the texture feature threshold range, the texture feature value A(χ, n) is equal to -1; when the rate difference B1 is within the texture feature threshold range, the rate difference B1 is substituted into The texture feature value A(χ, n) is obtained.

[0133] 0<T low <T high The threshold value is determined by experimental analysis. In the derivation process, we take T low =10, T high =30, then the texture feature threshold range is [10, 30].

[0134] By comparing two different images B fast (χ, n) and B slow(χ, n) is used to detect objects with slowly changing shapes within the camera's monitoring range. If objects with slowly changing shapes are detected continuously within a certain period of time, an alarm will be generated, and that area will be marked for further detection.

[0135] Determine whether there is smoke according to the following formula:

[0136]

[0137] The main components of smoke are carbon dioxide, water vapor, carbon monoxide, trace minerals, and dust, which makes it appear grayish black. These smoke areas can be detected by setting the threshold in the YUV color space. (χ,n) Refers to the brightness value of the pixel at position x in the n-th frame image I, U (χ,n) Refers to the chromaticity value of the pixel at position χ in the n-th frame image I, V (χ,n) Refers to the density value of the pixel at position x in the n-th frame image I.

[0138] Among them, U (χ,n) Refers to the chromaticity value of the pixel at position χ in the n-th frame image I, V (χ,n) Refers to the concentration value of the pixel at position χ in the n-th frame image I. Set a brightness threshold T1. The following formula is used to obtain the brightness Y of the pixel at position χ in the n-th frame image I. (χ,n) The expression:

[0139]

[0140] If the brightness Y of the pixel at position x in the n-th frame image I is (χ,n) When the brightness threshold T1 is less than the value of the smoke characteristic value P (χ,n) The value of is -1; when the brightness Y of the pixel at position x in the n-th frame image I is (χ,n) When T1 is greater than, the chromaticity value U of the pixel at position χ in the nth frame image I is (χ,n) and the concentration value V in the pixel at position x in the nth frame image I (χ,n) are both 128, then the smoke characteristic value P (χ,n) The value is 1. The brightness threshold T1 is set to exclude dark areas with low chromaticity values, because the colors in the area where fireworks exist may not have clear contrast. For the case where the brightness value is higher than T1, the chromaticity value U of the pixel at position χ in the nth frame image I is (χ,n) and the density value V of the pixel at position χ in the nth frame image I (χ,n) When the mean is around 128, the chromaticity value U of the pixel at position χ in the nth frame image I is (χ,n) and the density value V of the pixel at position χ in the nth frame image I (χ,n) Substitution In the case of (χ,n) The value is close to 1.

[0141] Determine whether it is necessary to trigger an open fire alarm based on the smoke feature value, motion feature value and texture feature value. Specifically, compare the smoke feature value, motion feature value and texture feature value in the environmental video with the smoke standard feature value, motion standard feature value and texture standard feature value in the second standard feature set one by one. If there are standard features that are inconsistent with the second standard feature set, an open fire alarm is triggered.

[0142] Through the above steps, we can build a training model for open flame recognition based on convolutional neural networks, achieve efficient and accurate open flame recognition, and deploy the trained model to actual application scenarios.

[0143] The above is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the structure of the present invention. These modifications and improvements will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A method for enhancing fire identification capabilities at smart construction sites using a large AI model, characterized in that: The method includes a dark fire identification model construction step and a dark fire identification analysis step: The dark fire identification model construction step includes: S11: Collecting dark fire training samples, wherein the dark fire training samples include first eigenvalues of hidden heat sources, air expansion and distortion phenomena, and light refraction phenomena; S12: Training the Dark Fire training sample using a deep convolutional neural network model to extract a first eigenvalue from the Dark Fire training sample, and evaluating the first eigenvalue using a first standard feature set to obtain a first optimal eigenvalue; S13: Generate a dark fire recognition model using the first optimal eigenvalue as model data; The dark fire identification and analysis step comprises: S21: Collecting environmental video of the actual environment; S22: The dark fire recognition model detects whether the video frame of the environmental video contains air distortion and expansion phenomenon and light refraction phenomenon according to the first optimal eigenvalue. If so, proceed to S23; S23: Analyzing the air distortion and expansion phenomenon using the dark fire recognition model to obtain light parameters of the location where the air distortion and expansion phenomenon occurs in the video frame, and analyzing the heat source location and heat source temperature using the light parameters; S24: Obtaining the size of the heat source according to the heat source position and the heat source temperature, and analyzing the danger level of the heat source according to the heat source temperature and the heat source size, and triggering a dark fire alarm when the danger level reaches a preset level threshold.

2. The method of using an AI large model to enhance the fire identification capability of a smart construction site according to claim 1 is characterized in that: In S12, evaluating the first eigenvalue using the first standard feature set to obtain a first optimal eigenvalue specifically includes: The first eigenvalue is compared one by one with the first standard eigenvalues in the first standard feature set. If there is a first standard eigenvalue that is consistent with the first eigenvalue, the first eigenvalue is used as the first optimal eigenvalue.

3. The method of using an AI large model to enhance the fire identification capability of a smart construction site according to claim 1 is characterized in that: Detect whether the video frame contains air distortion and expansion phenomena and light refraction phenomena, specifically: traverse the video frame according to the air expansion and distortion phenomena and light refraction phenomena in the first optimal eigenvalue; if the video frame has characteristic phenomena consistent with the air expansion and distortion phenomena and light refraction phenomena in the first optimal eigenvalue, it is determined that the air distortion and expansion phenomena and light refraction phenomena exist.

4. The method of using an AI large model to enhance the fire identification capability of a smart construction site according to claim 3 is characterized in that: In S23, light parameters are obtained by analyzing the phenomenon of air distortion and expansion, specifically: the temperature and pressure in the actual environment are obtained, and the refractive index of the light is calculated according to the temperature and pressure; and the refractive angle of the light is calculated according to the refractive index.

5. The method of using an AI large model to enhance the fire identification capability of a smart construction site according to claim 1 is characterized in that: In S23, the heat source position and heat source temperature are analyzed by the light parameters, specifically: the light parameters are compared with the light parameter threshold of the light refraction phenomenon in the dark fire recognition model; if the light parameters are consistent with the light parameter threshold, or the light parameters exceed the light parameter threshold, then the position corresponding to the light parameters is determined to be the heat source position, and the heat source temperature of the heat source position is obtained at the same time.

6. The method of using an AI large model to enhance the fire identification capability of a smart construction site according to claim 1 is characterized in that: The method further comprises an open flame identification model construction step and an open flame identification analysis step; The open flame recognition model construction step specifically includes: S31: Collecting open fire training samples, where the open fire training samples include second eigenvalues of flame and smoke; S32: training the open fire training sample using a deep convolutional neural network model to extract a second eigenvalue from the open fire training sample, and evaluating the second eigenvalue using a second standard feature set to obtain a second optimal eigenvalue; S33: Generate an open flame recognition model using the second optimal characteristic value as model data; The open flame identification and analysis step specifically includes: The open flame recognition model detects whether flames and / or smoke exist in the video frame of the environmental video according to the second optimal eigenvalue; If so, obtain the smoke feature value, motion feature value and texture feature value, and determine whether to trigger an open fire alarm based on the smoke feature value, motion feature value and texture feature value.

7. The method of using an AI large model to enhance the fire identification capability of a smart construction site according to claim 6 is characterized in that: The detection expression of the smoke characteristic value is: Among them, U (χ,n) Refers to the chromaticity value of the pixel at position χ in the n-th frame image I, V (χ,n) Refers to the concentration value of the pixel at position χ in the n-th frame image I, P(χ,n) is the smoke characteristic value of the pixel at position χ in the n-th frame image I, Set a brightness threshold T1, where the brightness Y of the pixel at position x in the n-th frame image I is obtained as follows: (χ,n) The expression: When the brightness Y of the pixel at position x in the n-th frame image I is (χ,n) When the brightness threshold T1 is less than the value of the smoke characteristic value P(χ,n) is -1; when the brightness Y of the pixel at position χ in the n-th frame image I is (χ,n) When the brightness threshold T1 is greater than the value of the smoke characteristic value P(χ,n) 8. The method of using an AI large model to enhance the fire identification capability of a smart construction site according to claim 6 is characterized in that: The detection expression of the motion characteristic value is: Where I(χ,n) refers to the intensity value of the pixel at position χ in the n-th frame image I, a is a smoothing constant between (0,1), B(χ,n+1) refers to the motion feature value of the n+1-th frame image, and B(χ,n) refers to the motion feature value of the n-th frame image; When the motion feature value B(χ,n) of the nth frame image is greater than the preset motion feature threshold, B(χ,n+1)=B(χ,n); when the motion feature value B(χ,n) of the nth frame image is less than or equal to the motion feature threshold, then B(χ,n+1)=aB(χ,n)+(1-a)I(χ,n); The detection formula of the texture feature value is: Among them, B fast (χ,n) and B slow (χ,n) refers to the way of updating at different rates, B fast (χ,n) is updated once per frame, B slow (χ,n) is updated once per second, T high and T low Refers to the different brightness values set in the image attributes, A(χ,n) is the texture feature value; set a texture feature threshold range, when the rate difference B1=|B fast (χ,n)-B slow When (χ, n)| is greater than the texture feature threshold range, the texture feature value A(χ, n) is equal to 1; when the rate difference B1 is less than the texture feature threshold range, the texture feature value A(χ, n) is equal to -1; when the rate difference B1 is within the texture feature threshold range, the rate difference B1 is substituted into The texture feature value A(χ,n) is obtained in .

9. The method of using an AI large model to enhance the fire identification capability of a smart construction site according to claim 6 is characterized in that: Whether it is necessary to trigger an open fire alarm is determined based on the smoke feature value, motion feature value and texture feature value, specifically: the smoke feature value, motion feature value and texture feature value in the environmental video are compared one by one with the smoke standard feature value, motion standard feature value and texture standard feature value in the second standard feature set. If there are standard features that are inconsistent with the second standard feature set, an open fire alarm is triggered.

Images