A computer vision-based coal mine fire identification method

Abstract

The application discloses a coal mine fire identification method based on computer vision, relates to the technical field of coal mine fire identification, and solves the technical problem of low accuracy of coal mine fire identification. The technical scheme of the coal mine fire identification method based on computer vision comprises the following steps: collecting data features of multiple characteristics such as color, shape and dynamic mode through a monitoring camera in a coal mine, improving the accuracy of fire detection; through a specially designed preprocessing step, the model can better adapt to the diversity and complexity of the coal mine environment; a unique local pixel-dependent transformation and morphological operation strengthens the special texture and morphological features related to fire in the image; by introducing the principle of elastic wave fluctuation, the optimization process of the deep learning network can effectively avoid falling into a local optimal solution, thereby improving the feature extraction effect; and the fire ant algorithm is adopted to optimize the hidden layer weight and bias of the extreme learning machine, thereby improving the stability and classification accuracy of the model.

Description

Technical Field

[0001] This application relates to the field of coal mine fire identification technology, and in particular to a coal mine fire identification method based on computer vision. Background Technology

[0002] Coal mines are important energy production sites, but they are also extremely dangerous working environments. Fire is one of the most serious risks to coal mine safety. Traditional fire detection methods rely heavily on temperature and smoke sensors, but these methods have several limitations, such as slow response times and high false alarm rates. With the rapid development of Industry 4.0 and smart manufacturing, the use of advanced computer vision and machine learning algorithms for real-time monitoring and fire early warning in coal mines is receiving increasing attention.

[0003] However, the coal mine environment is complex, with various interfering factors such as dust, changes in lighting, and mechanical vibrations, making automatic fire detection a highly challenging problem. Furthermore, fire detection algorithms need to possess high accuracy and real-time performance to issue timely warnings in the early stages of a fire, thereby reducing casualties and property damage.

[0004] In the prior art, Chinese invention patent CN202320592554.X proposes an automatic flame identification device for dangerous goods storage tank areas, relating to the field of ultraviolet flame detector technology. It includes a mounting pole, with a mounting base fixedly installed at the upper end of the pole, an ultraviolet flame detector fixedly installed at the upper end of the mounting base, an ultraviolet sensor fixedly installed at the front end of the ultraviolet flame detector, and a fixing base fixedly installed at the lower end of the ultraviolet flame detector. A conduit is fixedly installed on one side of the fixing base. A protective tube fixedly installed on one side of the conduit provides fixation and protection for the bent detector connection wire. The separate first and second arc-shaped connecting blocks facilitate assembly at one end of the protective tube and fix the protective tube to one side of the conduit. Simultaneously, the first and second arc-shaped mounting blocks installed at the lower end of the protective tube, along with multiple arc-shaped clips inside, limit and fix the outward end of the connection wire, preventing loosening of the connection joint due to continuous shaking.

[0005] Chinese invention patent CN202310239911.9 proposes a flame recognition method based on deep learning, which relates to the field of image recognition technology. It addresses the problem that existing technologies are prone to misdetecting slightly moving backgrounds as moving targets and, in some cases, misdetecting stationary backgrounds as moving targets, thus resulting in low flame recognition accuracy.

[0006] Chinese invention patent CN202310318217.6 proposes a flame identification and detection method for an oxidation furnace, which includes the following steps: acquiring images from inside the oxidation furnace; performing grayscale processing on the image data to obtain a grayscale image; performing global threshold segmentation on the grayscale image, aggregating the segmented pixels, and determining whether there are suspected areas; if a suspected area exists, proceed to the next step; if no suspected area exists, repeat the above steps; convert the image data into three-channel image data; extract the grayscale mean of the suspected areas from the three-channel image data and perform flame color screening; output an alarm signal based on the flame color screening result. In actual production, workers use flashlights or headlamps to shine into the oxidation furnace to check whether the yarn is traveling normally along the yarn path; in this invention, after selecting a suspected flame area, the ratio of the grayscale mean values ​​of the R, G, and B channels is used to further distinguish the flame from the flashlight light, preventing artificial light such as flashlights or headlamps from affecting the detection and causing false alarms.

[0007] The aforementioned existing technologies still have the following unresolved problems:

[0008] (1) Insufficient response speed and accuracy: Traditional coal mine fire detection methods, such as smoke and temperature sensors, often take several minutes or even longer to detect a fire after it occurs, which reduces the effectiveness of emergency response.

[0009] (2) High false alarm rate: Traditional methods are easily affected by environmental factors (such as humidity, airflow, mechanical vibration, etc.), resulting in a relatively high false alarm rate, which may cause unnecessary panic and waste of resources.

[0010] (3) Poor environmental adaptability: The coal mine environment has a variety of complex and unstable factors (such as light, dust, etc.), which may affect the effectiveness of traditional detection methods.

[0011] (4) Limited information: Traditional sensors mainly provide one-dimensional or two-dimensional signals, such as temperature or smoke concentration. This information is relatively limited and insufficient to capture the multifaceted characteristics of a fire.

[0012] (5) Model optimization problem: Traditional fire identification algorithms usually do not take into account the diversity and imbalance of data in complex environments, which makes the model prone to getting trapped in local optima or overfitting.

[0013] Therefore, this application proposes a computer vision-based method for coal mine fire identification to solve the above problems. Summary of the Invention

[0014] This application provides a computer vision-based method for identifying coal mine fires. Its technical objective is to accurately and efficiently detect flames in videos, thereby improving the accuracy of coal mine fire identification.

[0015] The above-mentioned technical objective of this application is achieved through the following technical solution:

[0016] A computer vision-based method for coal mine fire identification includes:

[0017] S1: Label the image data of the coal mine interior acquired by the camera to obtain labeled image data; wherein, the image data is an RGB image with dimensions of H×W×C, where H represents the height of the image, W represents the width of the image, and C represents the number of channels of the image;

[0018] S2: Preprocess the labeled image data to obtain preprocessed image data;

[0019] S3: Perform data augmentation on the preprocessed image data to obtain augmented image data;

[0020] S4: Input the augmented image data into a deep learning network for feature extraction to obtain image features;

[0021] S5: Input the image features into the classifier for training to obtain the classifier model;

[0022] S6: Identify coal mine fires using the classifier model.

[0023] Further, step S1 includes:

[0024] S11: Convert the RGB image to a YCbCr space image, represented as:

[0025] Y = 0.299R + 0.587G + 0.114B;

[0026] Cb = -0.169R - 0.331G + 0.500B;

[0027] Cr = 0.500R - 0.419G - 0.081B;

[0028] S12: The YCbCr spatial image is judged to determine whether it is a flame region by a judgment function, and the judgment result is obtained; wherein, the judgment function is expressed as:

[0029]

[0030] Wherein, when f(Cr,Cb)=1, it means that the Y Cb Cr spatial image is a flame region; when f(Cr,Cb)=0, it means that the Y Cb Cr spatial image is not a flame region;

[0031] T1 and T2 both represent adaptive thresholds, expressed as follows:

[0032] T1 = μCr，non-fire +k×σ Cr，non-fire ;

[0033] T2=σ Cr，non-fire -k×σ Cb，non-fire ;

[0034] Where, μ Cr，non-fire σ represents the mean value of the non-fire zone in Cr space; Cr，non-fire σ represents the variance of the non-fire zone in Cr space; Cr，non-fire σ represents the mean value of the non-fire zone in Cb space; Cb，non-fire This represents the variance of the non-fire zone in Cb space;

[0035] S13: Based on the judgment result, label the corresponding RGB image as "fire" or "non-fire" to obtain the labeled image data.

[0036] Further, step S2 includes:

[0037] S21: Standardize the labeled image data and represent it as follows:

[0038]

[0039] Among them, I i，norm The image is denoted as I after standardization. i The average gray value, σ represents the image I i The grayscale standard deviation; i∈[1,N], represents that there are N labeled grayscale images, and image I i It is an m×n matrix;

[0040] S22: Detect whether the pixels (x, y) in the standardized image are abnormal pixels. If so, proceed to step S23; otherwise, proceed to step S24.

[0041] The detection of abnormal pixels is represented as follows:

[0042]

[0043] When ΔI(x,y)>T Δ Then, I represents the pixel value of pixel (x, y). i,norm (x,y) is abnormal; T Δ Indicates the threshold for abnormal pixels;

[0044] S23: Perform dynamic thresholding on abnormal pixels, and adjust the threshold T of abnormal pixels using the dynamic threshold. Δ Adjustments are made to repair abnormal pixels, and after repair, proceed to step S24; wherein, the dynamic threshold segmentation is expressed as:

[0045] Tdynamic (x,y)=α1•mean W(x,y) (I i,norm )+β1·std W(x,y) (I i,norm );

[0046] Among them, T dynamic (x,y) represents the dynamic threshold, α1 and β1 are both weight parameters, and mean W(x,y) and std W(x,y) Let W(x,y) represent the mean and standard deviation within the window W(x,y), respectively.

[0047] S24: Repeat step S22 to continue anomaly detection on other pixels in the standardized image until anomaly pixel detection is completed on all standardized images, and preprocessed image data is obtained.

[0048] Further, step S3 includes:

[0049] S31: Perform an erosion operation on the preprocessed image data, represented as:

[0050]

[0051] Where I' represents the original image, and B1 represents a structuring element. Indicates an etching operation;

[0052] S32: Perform a dilation operation on the preprocessed image data, represented as:

[0053]

[0054] Where I' represents the original image, and B2 represents a structuring element. Indicates an etching operation;

[0055] S33: Apply a local pixel-dependent transformation P to each pixel of E(I') and D(I') to obtain two new images p(E(I'),p) and P(D(I'),p). The augmented image data is then represented as:

[0056] Aug={I',E(I'),D(I'),P(E(I'),p),P(D(I'),p)};

[0057] in, Let N(p) represent the gradient of pixel p, N(p) represent the local neighbor pixels of pixel p, α² and β² represent the weight coefficients, and w pq This represents the distance weight from pixel p to pixel q.

[0058] Furthermore, in step S4, the training process of the deep learning network includes:

[0059] (1) Initialize the weights w, bias b, elasticity constant κ and fluctuation velocity v of the deep learning network, and set the time step Δt;

[0060] (2) Calculate the initial loss function L(w,b) of the deep learning network;

[0061] (3) Train the deep learning network. For each iteration i, the training includes:

[0062] (3.1) Update e and b using the elastic fluctuation update rule. For the lth... th Layer weight W (l) and bias b (l) The elastic fluctuation update rule includes:

[0063]

[0064]

[0065]

[0066]

[0067] (3.2) κ and v are updated using an oscillation adjustment mechanism. The update of each layer l using the oscillation adjustment mechanism is expressed as follows:

[0068]

[0069]

[0070] in, and They represent the lth th The initial elastic constant and wave velocity of the layer; β z and γ z All represent adjustment factors;

[0071] (3.3) The loss function is recalculated;

[0072] (4) Repeat steps (3.1) to (3.3) until the training converges and the trained deep learning network is obtained.

[0073] The loss function is expressed as follows:

[0074]

[0075] y and These represent the input of features extracted by the deep learning network to the target output and the actual output of the preset classifier, respectively. N represents the number of input samples, t represents the current training iteration round, and α represents the input of features extracted by the deep learning network to the target output and the actual output of the preset classifier, respectively. L β represents the amplitude coefficient of a cosine wave. L φ represents the amplitude coefficient of the sine wave, ω represents the frequency of the cosine wave, and φ represents the frequency of the sine wave.

[0076] Furthermore, in step (4), the method for determining whether the training has reached convergence includes: convergence is achieved when the change in the loss value ΔL no longer changes within a preset number of iterations;

[0077] The change in the loss function, ΔL, is expressed as:

[0078] ΔL=L(w i+1 ,b i+1 )-L(w i ,b i );

[0079] Based on the elastic fluctuation update rule and Taylor series expansion, we obtain:

[0080]

[0081] Substituting the elastic fluctuation update rule into the above formula, we get:

[0082]

[0083] Furthermore, in step S5, during the training process, the classifier optimizes the hidden layer weight matrix Q and the hidden layer bias b' input to the classifier using the fire ant algorithm until the optimal weight matrix Q is obtained. * and optimal bias b' * Through the optimal weight matrix Q * and optimal bias b' * The hidden layer output matrix and output layer weights of the final classifier model are obtained as follows:

[0084] H opt =φ(XQ * +b' * );

[0085]

[0086] Among them, H opt β represents the hidden layer output matrix of the classifier model after training. opt This represents the output layer weights of the classifier model after training; X = [x1, x2, ..., x...]. N [] represents the image features after feature extraction, and N represents the number of samples; T = [t1, t2, ..., t N] represents the target label, t () This indicates the output label.

[0087] Furthermore, the objective function of the fire ant algorithm is expressed as: F(Q,b')=||Hβ-T|| 2 Where H represents the hidden layer output matrix of the classifier, and β represents the output layer weights.

[0088] The beneficial effects of this application are as follows: The computer vision-based coal mine fire identification method described in this application collects data features that fuse multiple features such as color, shape, and dynamic patterns through monitoring cameras inside the coal mine, thereby improving the accuracy of fire detection; through specially designed preprocessing steps, the model can better adapt to the diversity and complexity of the coal mine environment; unique local pixel-dependent transformation and morphological operations enhance the special texture and morphological features related to fire in the image; by introducing the principle of elastic fluctuation, the optimization process of the deep learning network can effectively avoid getting trapped in local optima, thereby improving the effect of feature extraction; the fire ant algorithm is used to optimize the hidden layer weights and biases of the extreme learning machine, improving the stability and classification accuracy of the model. Attached Figure Description

[0089] Figure 1 This is a flowchart of the method described in this application;

[0090] Figure 2 This is an architecture diagram of a deep learning network;

[0091] Figure 3 A flowchart illustrating optimization strategies for deep learning networks. Detailed Implementation

[0092] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0093] like Figure 1 As shown, the computer vision-based coal mine fire identification method described in this application includes:

[0094] S1: Label the image data of the coal mine interior acquired by the camera to obtain labeled image data.

[0095] In this embodiment, the data source is a monitoring camera inside the coal mine. Due to the possibility of fire in the environment, the key attributes of the data include color, shape, and dynamic pattern. The data format is an image format, specifically an RGB image, with dimensions of H×W×C, where H represents the height of the image, W represents the width of the image, and C represents the number of channels of the image (in this embodiment, C=3 corresponds to RGB).

[0096] In a specific embodiment, let I be an image captured by a surveillance camera, with a size of 256×256×3. The color of each pixel p can be represented as (R,G,B), where R,G,B are the pixel values ​​of the red, green, and blue channels, respectively.

[0097] For each image I, it is labeled as a binary category: fire (1) or non-fire (0).

[0098] In the annotation process, this application proposes an adaptive annotation algorithm. Traditional annotation methods usually rely on manual work, but in the coal mine environment, manual annotation is not only time-consuming but may also be inaccurate. Therefore, this application proposes an automatic annotation algorithm based on adaptive threshold and color space transformation.

[0099] Specifically, the annotation process of this annotation algorithm includes:

[0100] S11: Convert the RGB image to a YCbCr color space image. This helps reduce the impact of changes in illumination, and is represented as:

[0101] Y = 0.299R + 0.587G + 0.114B;

[0102] Cb = -0.169R - 0.331G + 0.500B;

[0103] Cr = 0.500R - 0.419G - 0.081B;

[0104] S12: The YCbCr spatial image is judged to determine whether it is a flame region by a judgment function, and the judgment result is obtained; wherein, the judgment function is expressed as:

[0105]

[0106] Wherein, when f(Cr,Cb)=1, it means that the Y Cb Cr spatial image is a flame region; when f(Cr,Cb)=0, it means that the Y Cb Cr spatial image is not a flame region;

[0107] T1 and T2 both represent adaptive thresholds, expressed as follows:

[0108] T1 = μ Cr,non-fire +k×σ Cr,non-fire ;

[0109] T2=σ Cr,non-fire -k×σ Cb,non-fire ;

[0110] Where, μ Cr,non-fire σ represents the mean value of the non-fire zone in Cr space; Cr,non-fire σ represents the variance of the non-fire zone in Cr space;Cr,non-fire σ represents the mean value of the non-fire zone in Cb space; Cb,non-fire represents the variance of the non-fire zone in Cb space; k represents an adjustable parameter, usually k∈(0,1).

[0111] The aforementioned adaptive annotation algorithm enables more accurate automatic annotation of fire images, thereby improving the accuracy and reliability of the fire identification system.

[0112] S13: Based on the judgment result, label the corresponding RGB image as "fire" or "non-fire" to obtain the labeled image data.

[0113] S2: Preprocess the labeled image data to obtain preprocessed image data.

[0114] Real-time identification of coal mine fires is a task with severe safety requirements. Image data, as a rich source of information, is one of the key elements in solving this problem. Due to the complexity of the field environment and the diversity of actual data, preprocessing becomes an important step for successful fire identification.

[0115] Specifically, step S2 includes:

[0116] S21: Standardize the labeled image data and represent it as follows:

[0117]

[0118] Among them, I i，norm The image is denoted as I after standardization. i The average gray value, σ represents the image I i The grayscale standard deviation; i∈[1,N], represents that there are N labeled grayscale images, and image I i It is an m×n matrix.

[0119] S22: In fire scene images, there are often abnormal pixels affected by environmental factors such as smoke and water mist. Therefore, it is necessary to detect whether the pixel (x,y) in the standardized image is an abnormal pixel. If it is, proceed to step S23; otherwise, proceed to step S24.

[0120] The detection of abnormal pixels employs a local first-order differential operator, expressed as:

[0121]

[0122] When ΔI(x,y)>T Δ Then, I represents the pixel value of pixel (x, y). i,norm (x,t) is abnormal; T Δ Indicates the threshold for abnormal pixels;

[0123] S23: Since traditional global thresholding methods perform poorly in handling complex environments (such as fire scenarios), this application performs dynamic thresholding segmentation on abnormal pixels, and sets the threshold T of abnormal pixels based on the dynamic threshold. Δ Adjustments are made to repair abnormal pixels, and after repair, proceed to step S24; wherein, the dynamic threshold segmentation is expressed as:

[0124] T dynamic (x,y)=α1·mean W(x,y) (I i,norm )+β1·std W(x,y) (I i,norm );

[0125] Among them, T dynamic (x,y) represents the dynamic threshold, α1 and β1 are both weight parameters, and mean W(x,y) and std W(x,y) Let w(x,y) represent the mean and standard deviation within the window w(x,y), respectively.

[0126] S24: Repeat step S22 to continue anomaly detection on other pixels in the standardized image until anomaly pixel detection is completed on all standardized images, and preprocessed image data is obtained.

[0127] The aforementioned preprocessing methods are better suited to various complex factors in fire scenarios, providing higher quality data input for subsequent data expansion and fire identification tasks.

[0128] S3: Perform data augmentation on the preprocessed image data to obtain augmented image data.

[0129] Data augmentation is an important method for increasing training data for models, especially crucial for coal mine fire identification. This application proposes a computer vision algorithm based on morphological features for data augmentation. Since coal mine fire images possess unique texture and morphological features (such as smoke and flames), the proposed local pixel-dependent transformation P can capture these local features and further enhance them through morphological operations, thereby providing more useful information for the identification algorithm.

[0130] Specifically, step S3 includes:

[0131] S31: Perform an erosion operation on the preprocessed image data, represented as:

[0132]

[0133] Where I' represents the original image, and B1 represents a structuring element. Indicates an etching operation;

[0134] S32: Perform a dilation operation on the preprocessed image data, represented as:

[0135]

[0136] Where I' represents the original image, and B2 represents a structuring element. Indicates an etching operation;

[0137] S33: Apply a local pixel-dependent transformation P to each pixel of E(I') and D(I'), resulting in two new images P(E(I'),p) and P(D(I'),p). The augmented image data is then represented as:

[0138] Aug={I',E(I'),D(I'),P(E(I'),p),P(D(I'),p)};

[0139] in, Let N(p) represent the gradient of pixel p, N(p) represent the local neighbor pixels of pixel p, α² and β² represent the weight coefficients, and w pq This represents the distance weight from pixel p to pixel q.

[0140] S4: Input the augmented image data into a deep learning network for feature extraction to obtain image features.

[0141] For feature extraction of the augmented image data, this application proposes using a deep learning network. Traditional neural networks typically employ gradient descent optimization during training, which is prone to getting trapped in local optima during feature extraction, leading to poor extraction results. Inspired by the principle of elastic wave motion—a natural phenomenon where elastic objects (such as springs and rubber bands) oscillate under external force and eventually reach a steady state—this principle can be used as a reference for parameter optimization. Therefore, this application proposes a deep learning network based on elastic wave motion optimization, which, unlike gradient-based optimization algorithms, effectively avoids the problem of deep learning networks getting trapped in local optima during training.

[0142] First, the structure of the deep learning network is defined. In this application, a multi-layer neural network with L hidden layers is used for feature extraction; in a specific embodiment, L is 3. For the l-th... th There is a hidden layer, and the weights and biases of this layer are represented by W. (l) and b (l) ,like Figure 2 As shown.

[0143] Furthermore, we define the parameter space of the neural network. Let the parameter space of the neural network be... Any point Pj It can be determined by a set of weights w j and a set of bias b j By definition, then, can be expressed as:

[0144]

[0145] P j =(w j ,b j );

[0146] Where M represents the number of points in the parameter space.

[0147] Furthermore, the loss function of the neural network is defined as L(), which is a fluctuation loss function, and can be expressed as:

[0148]

[0149] Where y and These represent the input of features extracted by the deep learning network to the target output and the actual output of the preset classifier, respectively; N represents the number of input samples; and t represents the current training iteration round. L β represents the amplitude coefficient of a cosine wave. L α represents the amplitude coefficient of a sine wave. L and β L Both ω and φ are preset by humans; ω represents the frequency of the cosine wave, and φ represents the frequency of the sine wave. Unlike the traditional cross-entropy loss function, when the fluctuation loss function replaces the conventional loss function in the elastic fluctuation optimization algorithm, the network will experience different "pressure" and "tension" states during training, which helps the network explore the parameter space more effectively and avoid local optima.

[0150] Furthermore, the elasticity constant and the fluctuation velocity are defined. This application introduces the elasticity constant κ and the fluctuation velocity v, which are associated with the weights w and biases b of the deep learning network.

[0151] Furthermore, we define elastic fluctuation update rules, where the elastic fluctuation update process is the same as the parameter update process of the deep learning network. A set of elastic fluctuation update rules is defined as follows:

[0152]

[0153]

[0154] w i+1 =w i +Δw;

[0155] b i+1 =b i +Δb;

[0156] Where Δt represents the time step.

[0157] Furthermore, we define an oscillation adjustment mechanism, which is the adjustment strategy for the elasticity constant κ and the fluctuation velocity v, expressed as:

[0158]

[0159]

[0160] Where κ0 and v0 represent the initial elastic constant and initial wave velocity, respectively, and β z and γ z All of these represent adjustment factors.

[0161] like Figure 3 As shown, the training process of a deep learning network includes:

[0162] (1) Initialize the weights w, bias b, elasticity constant κ and fluctuation velocity v of the deep learning network, and set the time step Δt;

[0163] (2) Calculate the initial loss function L(w,b) of the deep learning network;

[0164] (3) Train the deep learning network. For each iteration i, the training includes:

[0165] (3.1) Update w and b using the elastic fluctuation update rule. For the l-th... th Layer weight W (l) and bias b (l) The elastic fluctuation update rule includes:

[0166]

[0167]

[0168]

[0169]

[0170] (3.2) κ and v are updated using an oscillation adjustment mechanism. The update of each layer l using the oscillation adjustment mechanism is expressed as follows:

[0171]

[0172]

[0173] in, and They represent the lth thThe initial elastic constant and fluctuation velocity of the layer;

[0174] (3.3) The loss function is recalculated;

[0175] (4) Repeat steps (3.1) to (3.3) until the training converges and the trained deep learning network is obtained.

[0176] In step (4) above, the method for determining whether training has reached convergence includes: convergence is achieved when the change in the loss value ΔL no longer changes within a preset number of iterations. The calculation basis for the change in the loss value ΔL is as follows: during the iteration process, considering the state transition of the algorithm between any iteration i and i+1, the change in the loss function ΔL is expressed as:

[0177] ΔL=L(w i+1 ,b i+1 )-L(w i ,b i );

[0178] Based on the elastic fluctuation update rule and Taylor series expansion, we obtain:

[0179]

[0180] Substituting the elastic fluctuation update rule into the above formula, we get:

[0181]

[0182] Since ΔL is negative or zero, the algorithm will guarantee that the loss function L(w,b) decreases or remains unchanged.

[0183] Once the algorithm terminates, it means that the training of the feature extraction deep learning network model is complete.

[0184] S5: Input the image features into the classifier for training to obtain the classifier model.

[0185] Specifically, this application optimizes the Extreme Learning Machine (ELM) algorithm (classifier) ​​using the fire ant algorithm, a swarm intelligence optimization algorithm that simulates the foraging behavior of fire ants to find the optimal solution to a problem. In traditional Extreme Learning Machines (ELM), the hidden layer weights and biases are randomly initialized, which leads to model instability. To address this issue, this application employs the fire ant algorithm to optimize these parameters.

[0186] Specifically, let the image features after feature extraction be X = [x1, x2, ..., x...]. N ], N represents the number of samples, each x is a d-dimensional feature vector; the target label is T = [t1, t2, ..., t N ], t () This indicates the output label.

[0187] In Extreme Learning Machine (ELM), the hidden layer output matrix H can be represented as:

[0188]

[0189] Where Q represents the weight matrix input to the hidden layer, with a size of d×m; b' represents the bias of the hidden layer, with a size of 1×m; represents the activation function; m represents the number of hidden nodes.

[0190] The output layer weights β can be obtained by solving the following optimization problem:

[0191] First, initialize: Set the number of fire ants to n, and randomly initialize the bit output layer weights of n fire ants in the parameter space (Q, b'). Then:

[0192] β = arg min||Hβ-T||;

[0193] Furthermore, this problem can be solved using the Moore-Penrose pseudoinverse, expressed as:

[0194]

[0195] in, It is the Moore-Penrose pseudoreverse of H.

[0196] The objective function of the fire ant algorithm is defined as: F(Q,b')=||Hβ-T|| 2 .

[0197] The Fire Ant algorithm optimizes Q and b' in the following ways:

[0198] (1) Randomly place n fire ants in the parameter space;

[0199] (2) Foraging and pheromone renewal: Each fire ant evaluates the objective function F(Q,b') and releases pheromones along its path according to the magnitude of F. The concentration of pheromones is inversely proportional to the value of the objective function.

[0200] (3) Transfer probability: Calculate the probability of each fire ant transferring to a nearby location based on the concentration of pheromones;

[0201] (4) Renewal and Evaporation: Renew the location of all fire ants and allow the pheromones to evaporate naturally;

[0202] (5) Convergence check: If the preset maximum number of iterations is reached or the objective function F is not significantly improved, then stop.

[0203] The optimized Q and b' are obtained using the fire ant algorithm. These parameters are then used to train the ELM model. The final ELM model is derived from the optimal weight matrix Q found by the fire ant algorithm. * and optimal bias b' * A decision is expressed as:

[0204] H opt =φ(XQ * +b' * );

[0205]

[0206] Among them, H opt β represents the hidden layer output matrix of the classifier after training. opt This represents the output layer weights of the classifier after training.

[0207] Based on this, an accurate and stable classifier model can be obtained, which is particularly suitable for application scenarios such as coal mine fire identification that require high sensitivity and accuracy.

[0208] S6: Identify coal mine fires using the classifier model.

[0209] The above are exemplary embodiments of this application, and the scope of protection of this application is defined by the claims and their equivalents.

Claims

1. A computer vision-based method for identifying coal mine fires, characterized in that, include: S1: Annotate the image data of the coal mine interior acquired by the camera to obtain annotated image data; wherein, the image data is an RGB image, and its size is , Indicates the height of the image. Indicates the width of the image. Indicates the number of channels in an image; S2: Preprocess the labeled image data to obtain preprocessed image data; S3: Perform data augmentation on the preprocessed image data to obtain augmented image data; S4: Input the augmented image data into a deep learning network for feature extraction to obtain image features; S5: Input the image features into the classifier for training to obtain the classifier model; S6: Identify coal mine fires using the classifier model; Step S3 includes: S31: Perform an erosion operation on the preprocessed image data, represented as: ； in, Represents the original image. Represents a structural element. Indicates an etching operation; S32: Perform a dilation operation on the preprocessed image data, represented as: ； in, Represents the original image. Represents a structural element. Indicates an etching operation; S33: Yes and Apply local pixel-dependent transformation to each pixel Two new images were obtained. and The expanded image data is then represented as: ； in, ; Represents pixels gradient, Represents pixels Local neighbor pixels, and All represent weighting coefficients. Indicates from pixel To pixel Distance weights.

2. The coal mine fire identification method as described in claim 1, characterized in that... Step S1 includes: S11: Convert RGB image to Spatial image, represented as: ； ； ； S12: By judging the function pair The spatial image is used to determine whether it is a flame region, and the determination result is obtained; wherein, the determination function is expressed as: ； Among them, when When, it indicates that The spatial image is the area of ​​flame; when When, it indicates that The spatial image is not of the fire region; and All represent adaptive thresholds, expressed as follows: ； ； in, Indicates in The mean value of the non-fire zone in the space; Indicates in Variance of non-fire zone in space; Indicates in The mean value of the non-fire zone in the space; Indicates in Variance of non-fire zone in space; Indicates an adjustable parameter; S13: Based on the judgment result, label the corresponding RGB image as "fire" or "non-fire" to obtain the labeled image data.

3. The coal mine fire identification method as described in claim 2, characterized in that, Step S2 includes: S21: Standardize the labeled image data and represent it as follows: ； in, This represents the standardized image. Representing an image Average gray value, Representing an image The standard deviation of gray level; , indicating shared ownership A labeled grayscale image, and the image for Matrix; S22: Pixels in the standardized image If the pixel is abnormal, proceed to step S23; otherwise, proceed to step S24. The detection of abnormal pixels is represented as follows: ； when , then represents a pixel pixel values abnormal; Indicates the threshold for abnormal pixels; S23: Perform dynamic thresholding on abnormal pixels, and apply dynamic thresholding to abnormal pixels. Adjustments are made to repair abnormal pixels, and after repair, proceed to step S24; wherein, the dynamic threshold segmentation is expressed as: ； in, Indicates dynamic threshold. and All represent weight parameters. and Representing windows respectively Mean and standard deviation within; S24: Repeat step S22 to continue anomaly detection on other pixels in the standardized image until anomaly pixel detection is completed on all standardized images, and preprocessed image data is obtained.

4. The coal mine fire identification method as described in claim 3, characterized in that, In step S4, the training process of the deep learning network includes: (1) Weights of deep learning networks Bias elastic constant and fluctuation speed Perform initialization and set the time step. ; (2) Initial loss function for deep learning networks Perform calculations; (3) Train the deep learning network for each iteration. ,include: (3.1) Update via elastic fluctuation update rules and For the first Layer weights and bias The elastic fluctuation update rule includes: ； ； ； ； (3.2) Updated through oscillation adjustment mechanism and For each layer The update representation of applying the oscillation adjustment mechanism is as follows: ； ； in, and They represent the first The initial elastic constant and fluctuation velocity of the layer; and All represent adjustment factors; (3.3) Recalculate the loss function; (4) Repeat steps (3.1) to (3.3) until the training converges and the trained deep learning network is obtained; The loss function is expressed as follows: ； and These represent the input of features extracted by the deep learning network to the target output and the actual output of the preset classifier, respectively. Indicates the number of input samples. Indicates the current training iteration round number. This represents the amplitude coefficient of a cosine wave. This represents the amplitude coefficient of a sine wave. Represents the frequency of a cosine wave. It represents the frequency of a sine wave.

5. The coal mine fire identification method as described in claim 4, characterized in that, In step (4), the method for determining whether training has reached convergence includes: the change in the loss value. Convergence is achieved when the value remains unchanged within a preset number of iterations. Among them, the change in the loss function Represented as: ； Based on the elastic fluctuation update rule and Taylor series expansion, we obtain: ； Substituting the elastic fluctuation update rule into the above formula, we get: 。 6. The coal mine fire identification method as described in claim 5, characterized in that, In step S5, the classifier uses the fire ant algorithm to adjust the hidden layer weight matrix input to the classifier during training. and hidden layer bias Optimize until the optimal weight matrix is ​​obtained. and optimal bias Through the optimal weight matrix and optimal bias The hidden layer output matrix and output layer weights of the final classifier model are obtained as follows: ； ； in, This represents the hidden layer output matrix of the classifier model after training. This represents the output layer weights of the classifier model after training. This represents the image features after feature extraction. Indicates the number of samples; Indicates the target label, This indicates the output label.

7. The coal mine fire identification method as described in claim 6, characterized in that, The objective function of the Fire Ant algorithm is expressed as: ;in, This represents the output matrix of the hidden layer of the classifier. This represents the output layer weights.

Images