Rotary furnace combustion state prediction method and system based on flame image

Abstract

The present application belongs to the technical field of image analysis and processing, and particularly relates to a rotary furnace combustion state prediction method and system based on flame images. The method comprises: collecting video image sequences of the flame in the rotary furnace, and converting each image to the YUV color space; calculating the combustion contribution degree based on the luminance component and the chrominance component, and screening out the core flame pixel set; taking the combustion contribution degree as the weight, calculating the weighted covariance matrix, determining the confidence ellipse according to the eigenvalue and eigenvector of the matrix, the center of the confidence ellipse being the weighted centroid, the rotation angle being determined by the eigenvector, and the long and short semi-axis lengths being proportional to the square root of the eigenvalue; extracting the area, eccentricity, rotation angle and center position of the confidence ellipse as the combustion state feature vector at the current time; inputting the time sequence sequence composed of the combustion state feature vectors at multiple times into the pre-trained hidden Markov model, and outputting the rotary furnace combustion state. The present application improves the accuracy and anti-interference ability of the combustion state prediction.

Description

Technical Field

[0001] This invention belongs to the field of image analysis and processing technology, specifically relating to a method and system for predicting the combustion state of a rotary kiln based on flame images. Background Technology

[0002] As a core thermal equipment in industries such as cement, metallurgy, chemical engineering, and environmental protection, the stability of the internal combustion state of rotary kilns directly determines the accuracy of product quality control, production energy consumption levels, and equipment operational safety factors, making it a critical link requiring key management in industrial production. However, the combustion process is affected by multiple factors, including fluctuations in coal quality, dynamic changes in feed rate, kiln speed adjustment, and imbalances in air volume ratio, exhibiting strong time-varying and high uncertainty characteristics, posing a severe challenge to the accurate monitoring of the combustion state.

[0003] Traditional combustion status assessment mainly relies on manual observation of the flame's color, shape, and brightness through an observation hole. This method is not only labor-intensive but also easily affected by factors such as the operator's subjective experience, fatigue, and ambient lighting. This results in significant lag and individual differences in the assessment results, failing to meet the core needs of modern industry for real-time and precise control of the production process. It has become a prominent bottleneck restricting the improvement of production efficiency and the stability of product quality.

[0004] To overcome the limitations of manual monitoring, the industry has gradually adopted high-definition cameras at the kiln head to capture flame images. Image processing techniques are then used to extract feature parameters and establish a mapping relationship between these features and the combustion state, thus upgrading the monitoring method. However, existing technologies still have significant limitations: early methods focused on extracting global features such as average brightness, color moments, and overall area. These features are easily affected by background noise from kiln dust, material tumbling, and changes in shooting angle, and are insensitive to changes in the internal structure of the flame and the morphology of the core combustion area. Some improved solutions attempted to extract local features such as length, width, and oscillation frequency after separating the flame area using complex image segmentation algorithms, but these segmentation algorithms lack stability in the complex and variable kiln environment and are prone to segmentation errors. Furthermore, existing technologies generally ignore the differences in the contribution of different regions within the flame to combustion, failing to form integrated features that comprehensively reflect the size, shape, direction, and position of the core combustion area, directly limiting the accuracy and reliability of combustion state prediction models.

[0005] It is evident that existing combustion state monitoring technologies are either limited by the subjective inefficiency of manual processes or constrained by design flaws in image processing and feature extraction, making it difficult to accurately predict combustion states. Therefore, there is an urgent need to propose a monitoring method that can focus on the core combustion area, has strong anti-interference capabilities, and high feature integration, in order to address the pain points of existing technologies. Summary of the Invention

[0006] Therefore, the purpose of this invention is to propose a method and system for predicting the combustion state of a rotary kiln based on flame images, in order to solve the technical problems of existing technologies that ignore the differences in the contribution of combustion inside the flame, are greatly affected by noise such as dust, and thus have low reliability in predicting the combustion state.

[0007] To address the above problems, the technical solution of the rotary kiln combustion state prediction method based on flame images proposed in this invention is as follows:

[0008] A method for predicting the combustion state of a rotary kiln based on flame images includes the following steps:

[0009] The video image sequence of the flame inside the rotary kiln is acquired, and each frame is converted to the YUV color space. For the current frame, the combustion contribution of each pixel is calculated based on the luminance and chrominance components of the pixel, and pixels with a combustion contribution greater than the screening threshold are selected to form the core flame pixel set.

[0010] Using the combustion contribution of each core flame pixel as a weight, calculate the weighted covariance matrix of the position coordinates of the core flame pixel set; based on the eigenvalues ​​and eigenvectors of the weighted covariance matrix, determine a confidence ellipse, the center of which is the weighted centroid of the core flame pixel set, the rotation angle is determined by the eigenvectors, and the length of the major and minor semi-axis is proportional to the square root of the eigenvalues;

[0011] The area, eccentricity, rotation angle, and center position of the confidence ellipse are extracted as the combustion state feature vector at the current moment;

[0012] The time sequence consisting of the combustion state feature vectors at multiple times is input into a pre-trained Hidden Markov Model to output the combustion state of the rotary kiln.

[0013] Furthermore, the combustion contribution of each pixel is calculated as follows:

[0014] For a pixel at the current coordinate position, if the value of its luminance component Y is greater than 120, the value of its chrominance component U is between 70 and 110, and the value of its chrominance component V is greater than 140, the burning contribution of the pixel is set to its luminance component value; if the above conditions are not met, the burning contribution of the pixel is set to zero.

[0015] Furthermore, the method for obtaining the core flame pixel set is as follows:

[0016] The combustion contribution values ​​of all pixels in the current frame are used to form a single-channel grayscale image. The Otsu method is used to segment the single-channel grayscale image, and the segmented foreground pixels are used as the core flame pixel set.

[0017] Furthermore, the weighted covariance matrix of the core flame pixel set location coordinates is calculated as follows:

[0018] Assume the core flame pixels are concentrated in The first core flame pixel, the... The coordinates of each core flame pixel are , No. The combustion contribution of each core flame pixel is ;

[0019] Calculate the weighted centroid coordinates of the core flame pixel set. ,in, , ;

[0020] The weighted covariance matrix is ​​expressed as The matrix elements in the weighted covariance matrix are: ; ;

[0021] .

[0022] Furthermore, the method for determining the confidence ellipse is as follows:

[0023] The weighted covariance matrix is ​​subjected to eigenvalue decomposition to obtain two eigenvalues ​​of different sizes and two corresponding eigenvectors; the length of the major semi-axis of the confidence ellipse is equal to twice the square root of the larger eigenvalue, the length of the minor semi-axis of the confidence ellipse is equal to twice the square root of the smaller eigenvalue, and the rotation angle of the confidence ellipse is the angle between the eigenvector corresponding to the larger eigenvalue and the horizontal axis of the image.

[0024] Furthermore, the time sequence composed of the combustion state feature vectors at multiple times is constructed using a time sliding window of length 10, and the time sequence includes the combustion state feature vectors at the current time and the previous nine times.

[0025] Furthermore, the training process of the Hidden Markov Model is as follows:

[0026] The combustion state of the rotary kiln is pre-divided into three states: "normal combustion", "under-combustion" and "over-combustion" as the hidden states of the hidden Markov model.

[0027] Feature vectors were extracted from a large number of flame images under different conditions to form a training sample set with time sequence and state labels;

[0028] The state transition probability matrix, observation probability distribution, and initial state probability are calculated based on the training sample set using the maximum likelihood estimation method.

[0029] Furthermore, the output process of the combustion state of the rotary kiln is as follows:

[0030] For the time series input to the Hidden Markov Model, the Viterbi algorithm is used to calculate the hidden state sequence most likely to generate the time series, and the hidden state corresponding to the last moment of the hidden state sequence is output as the combustion state prediction result of the current rotary kiln.

[0031] The technical solution of the rotary kiln combustion state prediction system based on flame images proposed in this invention is as follows:

[0032] The rotary kiln combustion state prediction system based on flame images includes the following modules:

[0033] The filtering module is used to acquire video image sequences of flames inside the rotary kiln and convert each frame of the image to the YUV color space. For the current frame of the image, the combustion contribution of each pixel is calculated based on the luminance and chrominance components of the pixel, and pixels with combustion contribution greater than the filtering threshold are filtered out to form the core flame pixel set.

[0034] The determination module is used to calculate the weighted covariance matrix of the position coordinates of the core flame pixel set, with the combustion contribution of each core flame pixel as the weight; and to determine a confidence ellipse based on the eigenvalues ​​and eigenvectors of the weighted covariance matrix, wherein the center of the confidence ellipse is the weighted centroid of the core flame pixel set, the rotation angle is determined by the eigenvectors, and the length of the major and minor semi-axis is proportional to the square root of the eigenvalues.

[0035] The extraction module is used to extract the area, eccentricity, rotation angle, and center position of the confidence ellipse as the combustion state feature vector at the current moment;

[0036] The output module is used to input a time sequence consisting of the combustion state feature vectors at multiple times into a pre-trained hidden Markov model and output the combustion state of the rotary kiln.

[0037] Furthermore, the combustion contribution of each pixel is calculated as follows:

[0038] For a pixel at the current coordinate position, if the value of its luminance component Y is greater than 120, the value of its chrominance component U is between 70 and 110, and the value of its chrominance component V is greater than 140, the burning contribution of the pixel is set to its luminance component value; if the above conditions are not met, the burning contribution of the pixel is set to zero.

[0039] Furthermore, the method for obtaining the core flame pixel set is as follows:

[0040] The combustion contribution values ​​of all pixels in the current frame are used to form a single-channel grayscale image. The Otsu method is used to segment the single-channel grayscale image, and the segmented foreground pixels are used as the core flame pixel set.

[0041] Furthermore, the weighted covariance matrix of the core flame pixel set location coordinates is calculated as follows:

[0042] Assume the core flame pixels are concentrated in The first core flame pixel, the... The coordinates of each core flame pixel are , No. The combustion contribution of each core flame pixel is ;

[0043] Calculate the weighted centroid coordinates of the core flame pixel set. ,in, , ;

[0044] The weighted covariance matrix is ​​expressed as The matrix elements in the weighted covariance matrix are: ; ;

[0045] .

[0046] Before performing the above calculations, it must be ensured that the core flame pixel set is not empty. If the core flame pixel set is empty in some frames, the current combustion state can be determined as a preset special state such as "extinguished" or "no flame", and the calculation step can be skipped.

[0047] Furthermore, the method for determining the confidence ellipse is as follows:

[0048] The weighted covariance matrix is ​​subjected to eigenvalue decomposition to obtain two eigenvalues ​​of different sizes and two corresponding eigenvectors; the length of the major semi-axis of the confidence ellipse is equal to twice the square root of the larger eigenvalue, the length of the minor semi-axis of the confidence ellipse is equal to twice the square root of the smaller eigenvalue, and the rotation angle of the confidence ellipse is the angle between the eigenvector corresponding to the larger eigenvalue and the horizontal axis of the image.

[0049] Furthermore, the time sequence composed of the combustion state feature vectors at multiple times is constructed using a time sliding window of length 10, and the time sequence includes the combustion state feature vectors at the current time and the previous nine times.

[0050] Furthermore, the training process of the Hidden Markov Model is as follows:

[0051] The combustion state of the rotary kiln is pre-divided into three states: "normal combustion", "under-combustion" and "over-combustion" as the hidden states of the hidden Markov model.

[0052] Feature vectors were extracted from a large number of flame images under different conditions to form a training sample set with time sequence and state labels;

[0053] The state transition probability matrix, observation probability distribution, and initial state probability are calculated based on the training sample set using the maximum likelihood estimation method.

[0054] Furthermore, the output process of the combustion state of the rotary kiln is as follows:

[0055] For the time series input to the Hidden Markov Model, the Viterbi algorithm is used to calculate the hidden state sequence most likely to generate the time series, and the hidden state corresponding to the last moment of the hidden state sequence is output as the combustion state prediction result of the current rotary kiln.

[0056] The beneficial effects of this invention are as follows: This invention identifies the core flame region by utilizing pixel combustion contribution, replacing traditional image segmentation methods. This effectively suppresses interference from complex background noise such as kiln dust, material tumbling, and changes in shooting angle, focusing on the region that plays a decisive role in the combustion state, thus improving the anti-interference ability and environmental adaptability of feature extraction. Furthermore, this invention uses confidence ellipses to provide an integrated and parameterized representation of the core flame morphology, converting information on the flame's size, shape, direction, and position into a feature vector composed of area, eccentricity, rotation angle, and center position. Compared to traditional fragmented global or local features, this feature vector more comprehensively and accurately represents the overall state of the core combustion region. The time series of this feature vector is processed using a Hidden Markov Model, fully considering the temporal continuity and changing patterns of the combustion state, enabling prediction of the combustion state and overcoming the lag and subjectivity of traditional monitoring methods. Attached Figure Description

[0057] Figure 1 This is a flowchart of the steps in the rotary kiln combustion state prediction method based on flame images of the present invention. Detailed Implementation

[0058] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0059] Specific embodiments of the rotary kiln combustion state prediction method based on flame images proposed in this invention are as follows:

[0060] like Figure 1 As shown, the method for predicting the combustion state of a rotary kiln based on flame images includes the following steps:

[0061] S1. Acquire video image sequences of flames inside the rotary kiln and convert each frame to the YUV color space. For the current frame, calculate the combustion contribution of each pixel based on the luminance and chrominance components, and select pixels with combustion contributions greater than the selection threshold to form the core flame pixel set.

[0062] The purpose of this step is to remove environmental noise interference and accurately locate the areas that play a decisive role in combustion. Converting the image from RGB to YUV space separates luminance (Y) and chromaticity (UV), facilitating analysis using the high brightness and specific color characteristics of the flame. Combustion contribution is determined by combining both color and luminance indicators, distinguishing the true flame core from kiln reflections, high-temperature materials, or dust. Thresholding filters remove background and blurred edge areas, retaining only the most intensely burning parts of the flame, thereby reducing data processing load and improving the signal-to-noise ratio of the features.

[0063] Specifically, a high-temperature resistant camera is installed on the kiln head hood of the rotary kiln to continuously acquire RGB color video streams of the flames inside the kiln. For each acquired RGB image frame, the RGB image is converted to a YUV color space image using a linear transformation formula. The conversion formulas are: Y = 0.299 × R component + 0.587 × G component + 0.114 × B component; U = -0.147 × R component - 0.289 × G component + 0.436 × B component; V = 0.615 × R component - 0.515 × G component - 0.100 × B component. Here, Y represents the luminance component, and U and V represent the chrominance components.

[0064] The core region of a flame typically exhibits high brightness and a reddish-yellow hue, corresponding to high Y and V values ​​in the YUV space. In this embodiment, the combustion contribution of each pixel is calculated as follows:

[0065] For a pixel at the current coordinate position, if the value of its luminance component Y is greater than 120, the value of its chrominance component U is between 70 and 110, and the value of its chrominance component V is greater than 140, the burning contribution of the pixel is set to its luminance component value; if the above conditions are not met, the burning contribution of the pixel is set to zero.

[0066] This process is performed on all pixels in the image, creating a single-channel grayscale image from the combustion contribution values ​​of all pixels in the current frame. The Otsu method is then used to segment this single-channel grayscale image, and the segmented foreground pixels are used as the core flame pixel set. Here, the selection threshold is a dynamically generated optimal segmentation value. When analyzing the single-channel grayscale image, the Otsu method iterates through all possible thresholds (0-255), calculating the inter-class variance after dividing the image pixels into foreground and background categories at each threshold, and finding a value that maximizes the inter-class variance, for example, 115. 115 becomes the selection threshold. Using this selection threshold, the combustion contribution image is binarized. All pixels with a combustion contribution value greater than 115 are marked as foreground pixels, and the set of foreground pixels constitutes the core flame pixel set.

[0067] Of course, in other embodiments, the combustion contribution of a pixel can be calculated in other ways, treating the combustion contribution C of each pixel as a weighted combination of the luminance component Y and the chrominance component V. For example, the combustion contribution C = 0.6Y + 0.4V. In this case, the screening threshold is a fixed value determined by statistical analysis of the flame images under combustion conditions; it is a preset value, for example, set to 180. If the combustion contribution C of a pixel is greater than 180, the coordinates of that pixel are recorded, and the coordinates of all pixels that meet the condition constitute the core flame pixel set.

[0068] S2, using the combustion contribution of each core flame pixel as a weight, calculate the weighted covariance matrix of the position coordinates of the core flame pixel set; based on the eigenvalues ​​and eigenvectors of the weighted covariance matrix, determine a confidence ellipse, the center of which is the weighted centroid of the core flame pixel set, the rotation angle is determined by the eigenvectors, and the length of the major and minor semi-axis is proportional to the square root of the eigenvalues.

[0069] This step transforms irregular, discrete flame pixels into a parametric geometric model, achieving a mathematical description of the flame morphology. The weighted covariance matrix incorporates combustion contribution as a weight, meaning that points with more intense combustion have a greater impact on the model. This ensures that the calculated model not only reflects the flame's outline but also implicitly contains spatial distribution information about the flame's internal intensity. Through confidence ellipse modeling, eigenvalue decomposition transforms the complex weighted covariance matrix into elliptical geometric parameters, achieving dimensionality reduction of the data. This allows a standard ellipse to approximate the irregular flame cloud, laying the foundation for subsequent extraction of standardized features.

[0070] Specifically, the weighted covariance matrix of the core flame pixel set location coordinates is calculated as follows:

[0071] Suppose there are N core flame pixels in the core flame pixel set, the th The coordinates of each core flame pixel are , No. The combustion contribution of each core flame pixel is ;

[0072] Calculate the weighted centroid coordinates of the core flame pixel set. ,in, , ;

[0073] The weighted covariance matrix is ​​expressed as The matrix elements in the weighted covariance matrix are: ; ;

[0074] ; and These represent the degree of dispersion of the flame in the horizontal and vertical directions, respectively; , It represents the distribution correlation of flames in the horizontal and vertical directions. The weighted covariance matrix contains representations of the flame's size, extension direction, and shape.

[0075] The weighted covariance matrix obtained above is subjected to eigenvalue decomposition to obtain two eigenvalues. and and the corresponding two orthogonal eigenvectors and ,in, > The magnitudes of the two eigenvalues ​​correspond to the dispersion of the flame in the principal and secondary directions. The center of the confidence ellipse is the weighted centroid calculated above. The rotation angle of the confidence ellipse From the principal eigenvector The direction is determined. equal of Directional components and The arctangent value of the directional component reflects the main direction of the flame jet within the rotary furnace. The length of the semi-major axis of the confidence ellipse... With larger eigenvalues It is proportional to the square root, and the length of the minor semi-axis is... With smaller eigenvalues It is proportional to the square root.

[0076] In this embodiment, the length of the semi-major axis of the confidence ellipse is... Length of the short half-axis The rotation angle of the confidence ellipse For the angle between the eigenvector corresponding to the largest eigenvalue and the horizontal axis of the image, determine a confidence ellipse centered on the flame centroid that can represent the size and orientation of the flame body.

[0077] S3, extract the area, eccentricity, rotation angle and center position of the confidence ellipse as the combustion state feature vector at the current moment.

[0078] Specifically, the area of ​​the confidence ellipse is defined as... The eccentricity is The combustion state feature vector is .in, This value reflects the overall size of the flame; This value represents the shape of the flame; the closer it is to 1, the thinner and longer the flame. The center position is calculated in step S2. and Indicates the rotation angle This has already been calculated in step S2. Therefore, the combustion state characteristic vector... Represented as: .

[0079] S4, the time sequence consisting of the combustion state feature vectors at multiple times is input into the pre-trained Hidden Markov Model, and the combustion state of the rotary kiln is output.

[0080] The combustion state of the rotary kiln is pre-divided into three states: "normal combustion", "under-combustion" and "over-combustion" as the hidden states of the Hidden Markov Model. A large number of flame images under different states are collected to extract feature vectors and form a training sample set with time series and state labels. The state transition probability matrix, observation probability distribution and initial state probability are calculated based on the above training sample set using the maximum likelihood estimation method.

[0081] Specifically, experts categorized the combustion process in the rotary kiln, defining three states to be identified: normal combustion, under-combustion, and over-combustion. These states were used as hidden states in the Hidden Markov Model (HMM). A large number of rotary kiln flame videos recorded under different states were collected as raw data. For each frame of each video segment, a five-dimensional combustion state feature vector was extracted using the aforementioned method. Each video segment with a state label was converted into a sequence of feature vectors with time sequence and state labels. All feature vector sequences together formed the training sample set for the HMM. The labeled feature vector sequences were input into the training algorithm, and the data was statistically analyzed to learn the three core parameter sets of the HMM: 1) the state transition probability matrix, representing the probability of transitioning from one combustion state to another; 2) the observation probability distribution, representing the probability of a specific flame feature vector appearing in each combustion state, which was represented by a Gaussian mixture model; and 3) the initial state probability. After training, the HMM possessed the ability to infer the current combustion state based on new flame feature vector sequences.

[0082] In practical applications, a time sliding window of length 10 is preferably used. The combustion state feature vectors of the current time and the previous nine time steps are used to form a feature vector sequence. This feature vector sequence is input into the pre-trained Hidden Markov Model. The Viterbi algorithm is used to calculate the hidden state sequence most likely to generate the time sequence. The hidden state corresponding to the last time step of the hidden state sequence, such as stable combustion, is output as the combustion state prediction result of the current rotary kiln.

[0083] This invention effectively filters out complex background noise such as dust and material tumbling within the kiln by introducing pixel combustion contribution calculation and Otsu's method for screening, ensuring that feature extraction focuses on the core combustion area that plays a decisive role. Secondly, it uses a confidence ellipse model to parametrically model irregular flames, integrating the size, shape, direction, and position information of the flames into a standardized five-dimensional feature vector, overcoming the shortcomings of traditional methods that suffer from scattered features and incomplete representation. Finally, it combines a hidden Markov model to process the time series of feature vectors, making full use of the time-varying and continuous nature of combustion states, eliminating the random fluctuations and lags in single-frame image judgment, thereby achieving continuous and reliable prediction of rotary kiln combustion conditions and improving the accuracy and stability of rotary kiln combustion state prediction.

[0084] Specific embodiments of the rotary kiln combustion state prediction system based on flame images proposed in this invention are as follows:

[0085] The rotary kiln combustion state prediction system based on flame images includes the following modules:

[0086] The filtering module is used to acquire video image sequences of flames inside the rotary kiln and convert each frame of the image to the YUV color space. For the current frame of the image, the combustion contribution of each pixel is calculated based on the luminance and chrominance components of the pixel, and pixels with combustion contribution greater than the filtering threshold are filtered out to form the core flame pixel set.

[0087] The determination module is used to calculate the weighted covariance matrix of the position coordinates of the core flame pixel set, with the combustion contribution of each core flame pixel as the weight; and to determine a confidence ellipse based on the eigenvalues ​​and eigenvectors of the weighted covariance matrix, wherein the center of the confidence ellipse is the weighted centroid of the core flame pixel set, the rotation angle is determined by the eigenvectors, and the length of the major and minor semi-axis is proportional to the square root of the eigenvalues.

[0088] The extraction module is used to extract the area, eccentricity, rotation angle, and center position of the confidence ellipse as the combustion state feature vector at the current moment;

[0089] The output module is used to input a time sequence consisting of the combustion state feature vectors at multiple times into a pre-trained hidden Markov model and output the combustion state of the rotary kiln.

[0090] It should be noted that the processing procedures of the filtering module, the determining module, the extracting module, and the output module are the same as those of steps S1, S2, S3, and S4 in the above embodiment of the rotary kiln combustion state prediction method based on flame images, and will not be described in detail here.

[0091] In the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

[0092] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A rotary furnace combustion state prediction method based on flame images, characterized by, The method comprises the following steps: a video image sequence of a flame in a rotary furnace is collected, and each frame image is converted to a YUV color space; for a current frame image, a burning contribution degree of each pixel is calculated based on a luminance component and a chroma component of the pixel, and pixels with a burning contribution degree greater than a screening threshold value are screened out to form a core flame pixel set; The calculation method of the weighted covariance matrix of the position coordinates of the core flame pixel set is as follows: assuming that there are core flame pixels in the core flame pixel set, the coordinates of the th core flame pixel are , and the combustion contribution degree of the th core flame pixel is ; the weighted centroid coordinates of the core flame pixel set are calculated, wherein , . The weighted covariance matrix is represented as The matrix elements in the weighted covariance matrix are: ; ; ； a confidence ellipse is determined according to an eigenvalue and an eigenvector of the weighted covariance matrix, comprising: the weighted covariance matrix is subjected to eigenvalue decomposition to obtain two eigenvalues of different sizes and two corresponding eigenvectors; a long semi-axis length of the confidence ellipse is equal to twice a square root of the larger eigenvalue, a short semi-axis length of the confidence ellipse is equal to twice a square root of the smaller eigenvalue, a rotation angle of the confidence ellipse is an included angle between the eigenvector corresponding to the larger eigenvalue and a horizontal axis of the image, and a center of the confidence ellipse is a weighted barycenter of the core flame pixel set; the rotation angle is determined by the eigenvector, and the long and short semi-axis lengths are directly proportional to square roots of the eigenvalues; an area, an eccentricity, a rotation angle and a center position of the confidence ellipse are extracted as a burning state feature vector at a current time; a time sequence sequence formed by the burning state feature vectors at multiple times is input into a pre-trained hidden Markov model, and a burning state of the rotary furnace is output.

2. The flame image-based rotary furnace combustion state prediction method according to claim 1, characterized by, The burning contribution degree of each pixel is calculated in the following manner: for a pixel at a current coordinate position, when a value of the luminance component Y of the pixel is greater than 120, a value of the chroma component U of the pixel is within 70-110, and a value of the chroma component V of the pixel is greater than 140, the burning contribution degree of the pixel is set as the value of the luminance component; if the above conditions are not met, the burning contribution degree of the pixel is set as zero.

3. The flame image-based rotary furnace combustion state prediction method according to claim 1, characterized by, The core flame pixel set is obtained in the following manner: a single-channel gray image is formed by the burning contribution degree values of all pixels of the current frame, the single-channel gray image is segmented by using the Otsu method, and foreground pixels segmented out are taken as the core flame pixel set.

4. The flame image-based rotary furnace combustion state prediction method according to claim 1, characterized by, The time sequence sequence formed by the burning state feature vectors at multiple times is constructed by using a time sliding window with a length of 10, and the time sequence sequence contains the burning state feature vectors at the current time and the previous nine times.

5. The flame image-based rotary furnace combustion state prediction method according to claim 1, characterized by, The training process of the hidden Markov model is as follows: the burning state of the rotary furnace is pre-divided into three states of "normal combustion", "under-combustion" and "over-combustion" as hidden states of the hidden Markov model; a large number of flame images under different states are collected to extract feature vectors, and a training sample set with time sequence and state labels is formed; a state transition probability matrix, an observation probability distribution and an initial state probability are calculated according to the training sample set by using a maximum likelihood estimation method.

6. The flame image-based rotary furnace combustion state prediction method according to claim 5, characterized by, The output process of the burning state of the rotary furnace is as follows: for the time sequence sequence input into the hidden Markov model, a hidden state sequence most likely to produce the time sequence sequence is calculated by using a Viterbi algorithm, and a hidden state corresponding to a last time of the hidden state sequence is taken as a prediction result of the burning state of the current rotary furnace and is output.

7. A system for implementing the flame image-based rotary furnace combustion state prediction method according to any one of claims 1 to 6, characterized in that, The method comprises the following modules: The screening module is configured to collect a video image sequence of a flame in the rotary furnace and convert each frame image to a YUV color space; for a current frame image, calculate a burning contribution degree of each pixel based on a luminance component and a chroma component of the pixel, and screen out pixels with a burning contribution degree greater than a screening threshold to form a core flame pixel set; The determining module is configured to calculate a weighted covariance matrix of position coordinates of the core flame pixel set with the burning contribution degrees of the core flame pixels as weights; determine a confidence ellipse based on eigenvalues and eigenvectors of the weighted covariance matrix, wherein a center of the confidence ellipse is a weighted centroid of the core flame pixel set, a rotation angle is determined by the eigenvectors, and lengths of major and minor semi-axes are proportional to square roots of the eigenvalues; The extracting module is configured to extract an area, an eccentricity, a rotation angle and a center position of the confidence ellipse as a burning state feature vector of a current time; The output module is configured to input a time sequence sequence composed of the burning state feature vectors of multiple time points to a pre-trained hidden Markov model and output a burning state of the rotary furnace.

8. The flame image-based rotary furnace combustion state prediction system according to claim 7, characterized by, The burning contribution degree of each pixel is calculated in the following manner: For a pixel at a current coordinate position, if a value of a luminance component Y of the pixel is greater than 120, a value of a chroma component U of the pixel is within 70-110, and a value of a chroma component V of the pixel is greater than 140, the burning contribution degree of the pixel is set as the value of the luminance component; if the above conditions are not met, the burning contribution degree of the pixel is set as zero.

Images