Thermoelectric system boiler combustion state detection method based on dual-light fusion

Abstract

The invention relates to a thermoelectric system boiler combustion state detection method based on dual-light fusion, which comprises the following steps of: respectively acquiring boiler flame combustion images by using a visible light camera and an infrared camera, and fusing the images through a dual-light fusion algorithm after processing; taking the fused continuous images as a group of feature sequences, inputting the feature sequences into a three-dimensional convolutional neural network for classification, and obtaining a boiler flame combustion state through classification; the boiler airflow state is obtained through the collected infrared image; and the boiler flame combustion state and the boiler airflow state are combined to obtain the boiler combustion state. According to the method, high robustness is kept in a complex environment and different combustion states, the combustion state of the boiler can be reflected more comprehensively and accurately, and the detection accuracy is improved; the method is suitable for various types of thermoelectric system boilers and has high universality; the combustion state of the boiler can be evaluated from multiple levels, and the combustion condition in the boiler can be accurately reflected in real time; the method is high in identification accuracy, wide in application range and high in anti-interference capability.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control of a cogeneration system, and in particular to a method for detecting a combustion state of a boiler in a cogeneration system based on dual-light fusion. Background Art

[0002] With the continuous growth of industrial production and energy demand, boilers are key equipment in thermal power systems, and monitoring their flame combustion status is particularly important. The boiler combustion status is closely related to the operating efficiency, safety and environmental performance of the thermal power system. By obtaining the boiler combustion status, the operation of the thermal power system can be better controlled.

[0003] The traditional indirect boiler flame combustion state detection method mainly relies on visible light images or infrared images. Due to the limitations of technical means, it is often difficult to achieve ideal results in complex industrial environments. This is because the visible light flame combustion detection method can monitor the color and shape of the flame and provide intuitive combustion state information. Its advantage is high resolution and can clearly display the flame shape and changes. However, visible light images are difficult to provide accurate temperature information and are easily disturbed in high dust and high temperature environments; while the infrared image flame combustion detection method can provide flame temperature distribution information by capturing the infrared radiation of the flame. It is less affected by the high temperature environment, can penetrate smoke, and provide reliable boiler flame temperature data. However, the resolution of infrared images is usually low, and it is difficult to clearly display the detailed shape of the flame, and it is easily disturbed by the infrared radiation of slag, coke and fly ash. At the same time, the boiler combustion state is also affected by gas airflow factors. The stability and uniformity of the airflow are crucial to the combustion efficiency and stability of the boiler. Changes in gas composition directly affect the combustion reaction. Traditional airflow detection technology often relies on contact sensors, which are easily affected by environmental factors such as high temperature and dust, resulting in inaccurate measurement results.

[0004] Therefore, it is difficult for a single technical means to fully and accurately reflect the boiler flame combustion status.

[0005] With the advancement of image processing and computer vision technology, fusion algorithms have made great progress. Designing a fusion algorithm that can efficiently register and align visible light and infrared images and combine the features of the two images so that the acquired data has both high resolution and high temperature accuracy, will be able to provide more comprehensive and accurate combustion status information in complex industrial environments compared to traditional single image technology, thereby significantly improving the operating efficiency and safety of thermal power systems. Summary of the invention

[0006] The present invention solves the problems existing in the prior art and provides a method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion.

[0007] The technical solution adopted by the present invention is a method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion. The method uses a visible light camera and an infrared camera to collect boiler flame combustion images respectively, and after processing, the images are fused through a dual-light fusion algorithm. The fused continuous images are input as a set of feature sequences into a three-dimensional convolutional neural network for classification, and the boiler flame combustion state is obtained by classification.

[0008] Through the collected infrared images, the temperature distribution and radiation characteristics of the gas inside the boiler are detected to obtain the boiler airflow status;

[0009] The boiler combustion state is obtained by combining the boiler flame combustion state and the boiler airflow state.

[0010] Preferably, the method comprises the following steps:

[0011] S1 sets up a visible light camera and an infrared camera to collect visible light images of the boiler flame at the same sampling frequency I v and infrared image I i ;

[0012] S2 in visible light image I v and infrared image I i The feature points are obtained and matched, and the infrared image I is transformed based on the feature points. i Aligned to visible light image I v superior;

[0013] S3 takes the aligned infrared image I i ' and visible light image I v Use Laplacian pyramid for multi-resolution weighted fusion to obtain the dual-light fusion image I fused ;

[0014] S4 segments a collection of several consecutive frames of dual-light fusion images into feature subsequences, and inputs the historically collected feature subsequences into a three-dimensional convolutional neural network for training; different from conventional convolutional neural networks that can only process two-dimensional images, the three-dimensional convolutional neural network here refers to replacing the layers in the conventional convolutional neural network with three-dimensional layers, which can then be used to process three-dimensional sequence data;

[0015] S5 inputs the real-time dual-light fusion image sequence into the trained three-dimensional convolutional neural network, classifies the flame combustion state, and outputs the boiler flame combustion state;

[0016] S6 through infrared image I i The sequence represents the rate of change of boiler temperature, obtains the airflow velocity v, and uses the optical flow method to track the continuous frame infrared image I i The change of gas temperature pixels in the image is used to obtain the velocity vector field of the airflow. Get boiler airflow status;

[0017] S7 is calculated by taking the boiler airflow state characteristic f airflow and the characteristics of the boiler flame combustion state f flame Perform weighted fusion to obtain the boiler combustion state feature f combined , and obtain the boiler combustion status.

[0018] Preferably, S2 comprises the following steps:

[0019] S2.1 Use SIFT algorithm to detect feature point sets in visible light image and infrared image respectively {P v}、{P i};

[0020] S2.2 Using the corresponding feature point set {P v}、{P i} feature descriptor collection {D v} and {D i}Calculate the similarity of feature points and obtain the most matching feature point pair;

[0021] S2.3 Based on the matched feature point pairs, perform affine transformation on the infrared image to transform the infrared image I i Align to visible light image I v superior.

[0022] Preferably, in S2.1, P v ={p v1 ,p v2 ,…,p vn}, P i ={p i1 ,p i2 ,…,p im}, where n and m are the feature point sets {P v}、{P i}The number of elements in

[0023] In S2.2, D v ={d v1 ,d v2 ,…,d vn}, D i ={d i1 ,d i2 ,…,d im}, get the matching pair set M, satisfying,

[0024]

[0025] Among them, ε is the similarity threshold, p vk and p ilRepresents the matching feature point pairs in the visible light image and the infrared image.

[0026] Preferably, S2.3 comprises the following steps:

[0027] S2.3.1 extract the coordinates of the matched feature point pairs, and obtain the coordinates of the feature points on the visible light image and the coordinates of the feature points on the infrared image; in the actual operation process, a large number of feature point pairs will be obtained, and the wrong feature point pairs will be screened out by comparing and filtering after obtaining the affine transformation of each pair;

[0028] S2.3.2 Define an affine transformation model T that satisfies,

[0029]

[0030] Where a and b are scaling factors in the x and y directions, c and d are rotation factors, and t x and t y are the translation factors in the x and y directions respectively;

[0031] Let the original coordinates of the matching point in the infrared image be (x, y), and the point after affine transformation by model T be (x′, y′), satisfying x′=ax+by+t x , y'=cx+dy+t y ;

[0032] S2.3.3 Combine at least three sets of matching point pairs and calculate a, b, c, d, t in model T x and t y ;

[0033] S2.3.4 Infrared image I i The model T after applying the calculated parameters is aligned to the visible light image I v Above, aligned infrared image I i 'The pixel value at each pixel position (x', y') is equal to the original infrared image I i The pixel value at the corresponding position (x, y) satisfies,

[0034]

[0035] Preferably, S3 comprises the following steps:

[0036] S3.1 For visible light image I v and the aligned infrared image I i 'Build a Gaussian pyramid to satisfy

[0037]

[0038] Among them, the superscript of G is the number of layers of the Gaussian pyramid image, the 0th layer is the original image, and pyrDown is the downsampling operation;

[0039] S3.2 Based on the obtained Gaussian pyramid, the visible light image I v and the aligned infrared image I i 'Construct a Laplace pyramid to satisfy,

[0040]

[0041]

[0042] Among them, pyrUp is the upsampling operation;

[0043] S3.3 performs weighted fusion on each layer of Laplacian images to satisfy,

[0044]

[0045] Among them, w1 and w2 are weighted coefficients, satisfying w1+w2=1; obviously, w1,w2∈(0,1);

[0046] S3.4 Reconstruction of dual-light fusion image based on fused Laplacian pyramid I fused ,satisfy,

[0047]

[0048] Preferably, in S4, a collection of 50 consecutive frames of dual-light fusion images within 2 seconds is segmented into feature subsequences;

[0049] Here, 25 frames are sampled per second. In order to identify the flame combustion state, every 50 frames are set as a subsequence (a segment) to ensure data continuity. Through a long continuous image sequence, the accuracy of classifying the combustion state is improved.

[0050] Preferably, in S4, training the model comprises the following steps:

[0051] S4.1 obtains historically collected dual-light fusion images, obtains feature subsequences, and preprocesses the feature subsequences to meet input requirements;

[0052] S4.2 marking the characteristic subsequence according to the historical operating status of the boiler;

[0053] S4.3 inputs the feature subsequence into the three-dimensional convolutional neural network to obtain the forward propagation prediction result, and back-propagates to update the network weights until the training iteration ends.

[0054] Preferably, the three-dimensional convolutional neural network includes an input layer, a three-dimensional convolutional layer, a three-dimensional pooling layer, a residual block group, a global average pooling layer, a fully connected layer and a Softmax classifier arranged in sequence;

[0055] The residual block group includes a plurality of residual blocks, and any residual block includes two three-dimensional convolutional layers, two three-dimensional batch normalization layers and a ReLU activation function arranged in sequence.

[0056] Preferably, S6 comprises the following steps:

[0057] S6.1 Through continuous infrared images I i Sequence, get the airflow velocity v, satisfying,

[0058]

[0059] Where v(x,y) is the air flow velocity at the position (x,y), △T(x,y) is the temperature change at the position (x,y), △t is the time interval, and α is the thermal diffusion coefficient of the gas;

[0060] S6.2 uses the optical flow method based on continuous infrared light images I i Sequence to obtain the velocity vector field of airflow satisfy,

[0061]

[0062] in, is the airflow velocity vector field at the (x,y) position, and is the gradient of the pixel in the image at the (x, y) position in the spatial direction, is the temporal gradient of the image.

[0063] The present invention relates to a method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion. A visible light camera and an infrared camera are used to collect boiler flame combustion images respectively. After processing, the images are fused through a dual-light fusion algorithm. The fused continuous images are input into a three-dimensional convolutional neural network as a group of feature sequences for classification, and the boiler flame combustion state is obtained through classification; the boiler airflow state is obtained through the collected infrared images; and the boiler combustion state is obtained by combining the boiler flame combustion state and the boiler airflow state.

[0064] The beneficial effects of the present invention are:

[0065] (1) The combustion state detection is performed by using dual-light fusion images. The fused image not only has the detailed information of the visible light image, but also contains the temperature information of the infrared image. It makes full use of the complementary characteristics of the two images, can maintain high robustness in complex environments and under different combustion conditions, and can more comprehensively and accurately reflect the boiler combustion state, thereby improving the detection accuracy.

[0066] (2) Using non-contact machine vision detection methods, through different boiler flame data sets, it can handle a variety of fuel combustion modes, including layer combustion, chamber combustion and boiling combustion. It is suitable for various types of thermal power system boilers and has strong versatility;

[0067] (3) By using continuous multi-frame flame images as the input of the three-dimensional classification model, the changes in the flame combustion state can be dynamically captured and analyzed. By combining airflow state detection, the combustion state of the boiler can be evaluated from multiple levels. The dynamic changes of the airflow and the flame combustion state work together to enable the present invention to reflect the combustion conditions inside the boiler in real time and accurately.

[0068] The invention has the advantages of high recognition accuracy, wide application range, strong anti-interference ability, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a flow chart of the boiler combustion state detection method of the present invention;

[0070] Figure 2 It is a flow chart of the dual-light fusion algorithm of the present invention;

[0071] Figure 3 This is a structural diagram of the three-dimensional convolutional neural network model of the present invention. DETAILED DESCRIPTION

[0072] The present invention is further described in detail below in conjunction with embodiments, but the protection scope of the present invention is not limited thereto.

[0073] The present invention relates to a method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion. The method uses a visible light camera and an infrared camera to respectively collect boiler flame combustion images, and after processing, the images are fused through a dual-light fusion algorithm. The fused continuous images are input into a three-dimensional convolutional neural network as a set of feature sequences for classification, and the boiler flame combustion state is obtained by classification.

[0074] Obtain boiler airflow status through collected infrared images;

[0075] The boiler combustion state is obtained by combining the boiler flame combustion state and the boiler airflow state.

[0076] Specifically, the method comprises the following steps:

[0077] S1 sets up a visible light camera and an infrared camera to collect visible light images of the boiler flame at the same sampling frequency I v and infrared image I i ;

[0078] S2 in visible light image I v and infrared image I i The feature points are obtained and matched, and the infrared image I is transformed based on the feature points. i Aligned to visible light image I v superior;

[0079] S3 takes the aligned infrared image I i ' and visible light image I v Use Laplacian pyramid for multi-resolution weighted fusion to obtain the dual-light fusion image I fused ;

[0080] S4 segments a collection of several consecutive frames of dual-light fusion images into feature subsequences, and inputs the historically collected feature subsequences into a three-dimensional convolutional neural network for training;

[0081] S5 inputs the real-time dual-light fusion image sequence into the trained three-dimensional convolutional neural network, classifies the flame combustion state, and outputs the boiler flame combustion state.

[0082] S6 through infrared image I i The sequence represents the rate of change of boiler temperature, obtains the airflow velocity v, and uses the optical flow method to track the continuous frame infrared image I i The change of gas temperature pixels in the image is used to obtain the velocity vector field of the airflow. Get boiler airflow status;

[0083] S7 is calculated by taking the boiler airflow state characteristic f airflow and the characteristics of the boiler flame combustion state f flame Perform weighted fusion to obtain the boiler combustion state feature f combined , and obtain the boiler combustion status.

[0084] The method is further described below in conjunction with specific implementation methods.

[0085] S1 sets up a visible light camera and an infrared camera to collect visible light images of the boiler flame at the same sampling frequency I v and infrared image I i ;

[0086] The visible light image of boiler flame combustion is collected by the visible light camera of the binocular camera I v ;

[0087] The infrared image of boiler flame combustion is collected by the infrared camera of the binocular camera. i ;

[0088] In practical applications, the visible light camera and the infrared camera are integrated into a binocular camera to facilitate the equipotential acquisition of images.

[0089] S2 in visible light image I v and infrared image I i The feature points are obtained and matched, and the infrared image I is transformed based on the feature points. i Aligned to visible light image I v superior.

[0090] (2-1) Using SIFT algorithm to detect feature point set in visible light image {P v}、The feature point set of infrared image {P i},include:

[0091] Using SIFT algorithm in visible light image I v The detection feature point set {P v};

[0092] P v ={p v1 ,p v2 ,…,p vn}

[0093] Using SIFT algorithm in infrared image I i The detection feature point set {P i};

[0094] P i ={p i1 ,p i2 ,…,p im}

[0095] (2-2) Using the corresponding feature descriptor set {D v} and {D i}Calculate the similarity of feature points and find the best match, including:

[0096] For visible light image I v The feature point set {P v}, and its feature descriptor is {D v};

[0097] D v ={d v1 ,d v2 ,…,d vn}

[0098] For infrared image I i The feature point set {P i}, and its feature descriptor is {D i};

[0099] D i ={d i1 ,d i2 ,…,d im}

[0100] By comparing the similarity of feature descriptors, the best match for each pair of feature points is found, and the similarity is calculated using Euclidean distance. The matching function is expressed as:

[0101] M={(p vk ,p il )|||d vk -d il ||≤ε}

[0102] Where M is the set of matching pairs, ε is the similarity threshold; here p vk and p il Represents the matching feature point pairs in the visible light image and the infrared image;

[0103] (2-3) Using the matched feature points, perform affine transformation on the infrared image and transform the infrared image I i Aligned to visible light image I v Above, including:

[0104] (2-3-1) Extract the coordinates of the matching point pair, where the coordinates of the matching point on the visible light image S v , the coordinates of the matching point on the infrared image S i ,

[0105] S v ={(x vk ,y vk )|(p vk ,p il )∈M}

[0106] S i ={(x il ,y il )|(p vk ,p il )∈M}

[0107] (2-3-2) Calculate the affine transformation model T using the matching point pairs;

[0108] In affine transformation, the transformation matrix T is used to describe the linear transformation and displacement from one plane to another. The affine transformation matrix T is usually expressed as a 3×3 matrix, but one row is kept as [0,0,1], because the affine transformation is a linear transformation plus a translation, which is specifically expressed as:

[0109]

[0110] Where a and b are scaling factors, representing the scaling in the x and y directions respectively; c and d are rotation factors, representing the degree of rotation of the image; t x and t y is the translation factor, which represents the translation of the image in the x and y directions;

[0111] The matching point in the original infrared image is (x, y), and the point after affine transformation is (x′, y′), which can be calculated by the following formula:

[0112]

[0113] Expand the calculation to get:

[0114] x'=ax+by+t x

[0115] y'=cx+dy+t y

[0116] (2-3-3) In order to calculate the matrix parameters in the affine transformation model T, at least three matching point pairs are required. Let the matching point pairs be Matching points, where n = 1, 2, 3, these matching points satisfy the following equation:

[0117]

[0118] These equations can be expressed as a system of linear equations:

[0119]

[0120] By solving the linear equations, we can get the parameters a, b, c, d, t of the affine transformation matrix. x and t y ;

[0121] (2-3-4) The infrared image I i Apply the calculated transformation model T to align the visible light image I v This requires the aligned infrared image I i 'The pixel value at each pixel position (x', y') is equal to the original infrared image I i The pixel value at the corresponding position (x, y) is as follows:

[0122] I i '(x',y') T =T×I i (x,y) T

[0123] Through the above formula, the infrared image I i Each pixel position of is aligned to the visible light image I through affine transformationv On, I i ' is the infrared image after alignment.

[0124] S3 takes the aligned infrared image I i ' and visible light image I v Use Laplacian pyramid for multi-resolution weighted fusion to obtain the dual-light fusion image I fused .

[0125] (3-1) For visible light image I v and the aligned infrared image I i 'Construct a Gaussian pyramid, the formula is as follows:

[0126]

[0127]

[0128] The superscript G indicates the number of layers of the Gaussian pyramid image. The 0th layer is the original image. pyrDown indicates the downsampling operation. The downsampling operation reduces the image size by half. Each layer of the Gaussian pyramid image is the downsampling result of the previous layer.

[0129] (3-2) Based on the obtained Gaussian pyramid, the visible light image I v and the aligned infrared image I i 'Construct the Laplace pyramid, the formula is as follows:

[0130]

[0131]

[0132] (3-3) Where pyrUp represents the upsampling operation, and then the Laplacian image of each layer is weighted fused, and the formula is as follows:

[0133]

[0134] Wherein, w1 and w2 are weighting coefficients, satisfying w1+w2=1;

[0135] (3-4) Then reconstruct the bi-light fusion image I from the fused Laplacian pyramid fused :

[0136]

[0137] S4 segments a collection of several consecutive frames of dual-light fusion images into feature subsequences, and inputs the historically collected feature subsequences into a three-dimensional convolutional neural network for training.

[0138] Specifically, a collection of 50 consecutive dual-light fusion images within two seconds {Ifused The historically collected feature subsequences are input into a three-dimensional convolutional classification network for training.

[0139] In the present invention, boiler flame combustion is a dynamic process, and the boiler combustion state cannot be measured only from the features of a single dual-light fusion image. Therefore, in order to make the detection model pay attention to the dynamic information of boiler combustion, the historically sampled dual-light fusion image data is segmented into subsequences of 50 frames of images within two seconds; at this time, the input data has changed from two-dimensional image data to three-dimensional sequence data with time series information. The traditional two-dimensional image classification network is difficult to process this type of data well, and the three-dimensional convolution classification network expands the two-dimensional convolution into a three-dimensional convolution to capture the features on the time series, thereby better capturing the dynamic characteristics of the boiler flame combustion state.

[0140] Input the dual-light fusion image subsequence into the 3D convolution classification network for training, including:

[0141] (4-1) Data preprocessing: The historically collected dual-light fusion images are segmented into subsequences of 50 frames in two seconds, and each subsequence is used as the input of the three-dimensional convolution classification network. At the same time, in order to improve the training effect, the input data needs to be preprocessed, and the size of each image is uniformly adjusted to H×W. Then the shape of the network input layer is (N, C, H, W), where N is the number of frames (the number of frames in this embodiment is 50), C is the number of channels (the image after dual-light fusion has four channels), and H and W are the height and width of the image respectively;

[0142] (4-2) Data labeling: labeling the collected data subsequences according to the historical operation status of the boiler. In this embodiment, the boiler flame combustion status is divided into seven statuses: low-load stable combustion, low-load unstable combustion, high-load stable combustion, high-load unstable combustion, load increase combustion process, load decrease combustion process and flame extinction;

[0143] (4-3) Model construction: The three-dimensional convolution classification network of this embodiment uses a 3D ResNet neural network. The network structure includes a network input of (N, C, H, W) shape, a three-dimensional convolution layer and a three-dimensional pooling layer, a stacked residual block, a global average pooling layer, a fully connected layer, and a Softmax classifier, and outputs a combustion state category. The three-dimensional convolution layer applies a three-dimensional convolution operation to the input data and outputs a feature map:

[0144] Conv3D(N,C,H,W)→(N,C',H',W')

[0145] Where Conv3D represents the three-dimensional convolution operation, C' is the number of output channels, H' and W' are the height and width of the feature map after convolution, and the three-dimensional pooling layer downsamples the feature map to reduce the data dimension while retaining important features:

[0146] MaxPool3D(N,C',H',W')→(N,C',H”,W”)

[0147] Where MaxPool3D represents the three-dimensional maximum pooling operation, H' and W' are the height and width of the feature map after pooling;

[0148] After the first layer of 3D convolutional layer and 3D pooling layer, there are several residual blocks. Each residual block contains two 3D convolutional layers, two 3D batch normalization layers, and a ReLU activation function:

[0149] Residual Block=Conv3D(C',C')→BatchNorm3D→ReLU

[0150] →Conv3D(C',C')→BatchNorm3D

[0151] Residual Block represents the residual block, and BatchNorm3D represents the three-dimensional batch normalization layer. By stacking residual blocks, a deeper network structure can be constructed to achieve better performance in classification tasks.

[0152] The global average pooling layer performs global average pooling on the feature map and outputs a fixed-size feature vector, expressed as:

[0153] GlobalAvgPool3D(N,C',H',W')→(N,C')

[0154] Among them, GlobalAvgPool3D represents the global average pooling operation. The fully connected layer maps the feature vector to the classifier output, and uses the Softmax function to calculate the probability of each category. The category with the highest probability is the category predicted by the model, and the final output is the category of the boiler flame combustion state.

[0155] FC(N,C')→(N,num_classes)

[0156] Softmax(N,num_classes)→(num_classes)

[0157] Among them, FC represents the fully connected layer, and num_classes represents the classification category.

[0158] The training process includes the following steps:

[0159] (4-3-1) Forward propagation is the process of passing the input data through the network layer by layer to finally obtain the predicted output, that is, the process of obtaining the category output after the image enters the three-dimensional convolutional network;

[0160] (4-3-2) Loss function calculation: The output of the three-dimensional convolution classification network is the predicted classification of the boiler flame combustion state. The cross entropy loss function can be used to measure the classification error;

[0161] (4-3-3) Back propagation to update parameters: The back propagation algorithm is used to calculate the gradient of the loss function to the model parameters, thereby guiding the update of the model parameters. Then the optimization algorithm (such as stochastic gradient descent and Adam optimizer) updates the model parameters according to the calculated gradient to minimize the loss function;

[0162] (4-3-4) Training iteration: Repeat the above process for several rounds (called epochs) until the loss function converges or the predetermined number of training times is reached.

[0163] After S5 training is completed, the real-time dual-light fusion image sequence is input into the trained three-dimensional convolutional neural network to classify the flame combustion state and output the boiler flame combustion state.

[0164] S6 through infrared image I i The sequence represents the rate of change of boiler temperature, obtains the airflow velocity v, and uses the optical flow method to track the continuous frame infrared image I i The change of gas temperature pixels in the image is used to obtain the velocity vector field of the airflow. Get boiler airflow status;

[0165] (6-1) Through continuous infrared light images I i The infrared imaging temperature change in the sequence estimates the airflow velocity v, which satisfies,

[0166]

[0167] Where v(x,y) is the air flow velocity at the position (x,y), △T(x,y) is the temperature change at the position (x,y), △t is the time interval, and α is the thermal diffusion coefficient of the gas;

[0168] (6-2) Using the optical flow method to analyze the continuous infrared image I i The sequence is estimated to estimate the velocity vector field of the airflow satisfy,

[0169]

[0170] in for The airflow velocity vector field at the position, is the gradient of the pixel at the image position (x, y) in the spatial direction, is the temporal gradient of the image.

[0171] S7 is calculated by taking the boiler airflow state characteristic f airflow and the characteristics of the boiler flame combustion state f flame Perform weighted fusion to obtain the boiler combustion state feature f combined , get the boiler combustion state;

[0172] In practical applications,

[0173] f combined =w1·f airflow +w2·f flame

[0174] Among them, w1 and w2 are weights, which indicate the importance of airflow features and flame image classification results. By adjusting the weights, the contribution of different features can be optimized according to the actual application.

[0175] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion, characterized in that: The method uses a visible light camera and an infrared camera to collect boiler flame combustion images respectively, fuses the images through a dual-light fusion algorithm after processing, and uses the fused continuous images as a set of feature sequences to input into a three-dimensional convolutional neural network for classification, and classifies the boiler flame combustion state; Through the collected infrared images, the temperature distribution and radiation characteristics of the gas inside the boiler are detected to obtain the boiler airflow status; The boiler combustion state is obtained by combining the boiler flame combustion state and the boiler airflow state.

2. According to claim 1, a method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion is characterized in that: The method comprises the following steps: S1 sets up a visible light camera and an infrared camera to collect visible light images of the boiler flame at the same sampling frequency I v and infrared image I i ; S2 in visible light image I v and infrared image I i The feature points are obtained and matched, and the infrared image I is transformed based on the feature points. i Aligned to visible light image I v superior; S3 takes the aligned infrared image I i ′ and visible light image I v Use Laplacian pyramid for multi-resolution weighted fusion to obtain the dual-light fusion image I fused ; S4 segments a collection of several consecutive frames of dual-light fusion images into feature subsequences, and inputs the historically collected feature subsequences into a three-dimensional convolutional neural network for training; S5 inputs the real-time dual-light fusion image sequence into the trained three-dimensional convolutional neural network, classifies the flame combustion state, and outputs the boiler flame combustion state; S6 through infrared image I i The sequence represents the rate of change of boiler temperature, obtains the airflow velocity v, and uses the optical flow method to track the continuous frame infrared image I i The change of gas temperature pixels in the image is used to obtain the velocity vector field of the airflow. Get boiler airflow status; S7 is calculated by taking the boiler airflow state characteristic f airflow and the characteristics of the boiler flame combustion state f flame Perform weighted fusion to obtain the boiler combustion state feature f combined , and obtain the boiler combustion status.

3. According to claim 2, a method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion is characterized in that: S2 includes the following steps: S2.1 Use SIFT algorithm to detect feature point sets in visible light image and infrared image respectively {P v }、{P i }; S2.2 Using the corresponding feature point set {P v }、{P i } feature descriptor collection {D v } and {D i }Calculate the similarity of feature points and obtain the most matching feature point pair; S2.3 Based on the matched feature point pairs, perform affine transformation on the infrared image and transform the infrared image I i Align to visible light image I v superior.

4. According to claim 3, a method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion is characterized in that: In S2.1, P v ={p v1 ,p v2 ,…,p vn }, P i ={p i1 ,p i2 ,…,p im }, where n and m are the feature point sets {P v }、{P i }The number of elements in In S2.2, D v ={d v1 ,d v2 ,…,d vn }, D i ={d i1 ,d i2 ,…,d im }, get the matching pair set M, satisfying M={(p vk ,p il )||||d vk -d il ||≤ε} Among them, ε is the similarity threshold, p vk and p il Represents the matching feature point pairs in the visible light image and the infrared image.

5. According to claim 4, a method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion is characterized in that: S2.3 includes the following steps: S2.3.1 extract the coordinates of the matched feature point pairs to obtain the coordinates of the feature points on the visible light image and the coordinates of the feature points on the infrared image; S2.3.2 Define an affine transformation model T that satisfies, Where a and b are scaling factors in the x and y directions, c and d are rotation factors, and t x and t y are the translation factors in the x and y directions respectively; Let the original coordinates of the matching point in the infrared image be (x, y), and the point after affine transformation by model T be (x′, y′), satisfying x′=ax+by+t x , y'=cx+dy+t y ; S2.3.3 Combine at least three sets of matching point pairs and calculate a, b, c, d, t in model T x and t y ; S2.3.4 Infrared image I i The model T after applying the calculated parameters is aligned to the visible light image I v Above, aligned infrared image I i 'The pixel value at each pixel position (x', y') is equal to the original infrared image I i The pixel value at the corresponding position (x, y) satisfies, Yo i '(x',y') T =T×I i (x,y) T 。 6. The method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion according to claim 2 is characterized in that: S3 includes the following steps: S3.1 For visible light image I v and the aligned infrared image I i 'Build a Gaussian pyramid to satisfy The superscript of G is the number of layers of the Gaussian pyramid image, with the 0th layer as the original image. pyrDown is the downsampling operation; S3.2 Based on the obtained Gaussian pyramid, the visible light image I v and the aligned infrared image I i 'Construct a Laplace pyramid to satisfy, Among them, pyrUp is the upsampling operation; S3.3 performs weighted fusion on each layer of Laplacian images to satisfy, Wherein, w1 and w2 are weighting coefficients, satisfying w1+w2=1; S3.4 Reconstruction of dual-light fusion image based on fused Laplacian pyramid I fused ,satisfy, 7. The method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion according to claim 2 is characterized in that: In S4, a collection of 50 consecutive dual-light fusion images within 2 seconds is segmented into feature subsequences.

8. The method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion according to claim 7 is characterized in that: In S4, training the model comprises the following steps: S4.1 obtains historically collected dual-light fusion images, obtains feature subsequences, and preprocesses the feature subsequences to meet input requirements; S4.2 marking the characteristic subsequence according to the historical operating status of the boiler; S4.3 inputs the feature subsequence into the three-dimensional convolutional neural network to obtain the forward propagation prediction result, and back-propagates to update the network weights until the training iteration ends.

9. A method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion according to claim 1 or 8, characterized in that: The three-dimensional convolutional neural network includes an input layer, a three-dimensional convolutional layer, a three-dimensional pooling layer, a residual block group, a global average pooling layer, a fully connected layer and a Softmax classifier which are arranged in sequence; The residual block group includes a plurality of residual blocks, and any residual block includes two three-dimensional convolutional layers, two three-dimensional batch normalization layers and a ReLU activation function arranged in sequence.

10. The method for detecting the combustion state of a boiler in a thermoelectric system based on dual-light fusion according to claim 2 is characterized in that: S6 includes the following steps: S6.1 Through continuous infrared images I i Sequence, get the airflow velocity v, satisfying, Where v(x,y) is the air flow velocity at the position (x,y), △T(x,y) is the temperature change at the position (x,y), △t is the time interval, and α is the thermal diffusion coefficient of the gas; S6.2 uses the optical flow method based on continuous infrared light images I i Sequence to obtain the velocity vector field of airflow satisfy, in, is the airflow velocity vector field at the (x,y) position, and is the gradient of the pixel in the image at the (x, y) position in the spatial direction, is the temporal gradient of the image.

Images