Combustion optimization control method and system for semi-coke furnace based on machine vision

Abstract

The invention discloses a semi-coke furnace combustion optimization control method and system based on machine vision. The method comprises the following steps: respectively generating a flame dynamic texture map and a hearth temperature field space-time evolution matrix by collecting a combustion chamber image sequence; and inputting the flame texture and the temperature field into a coupling analysis network, excavating a deep correlation between the flame texture and the temperature field evolution, and outputting a coupling feature tensor representing a combustion state. And querying a pre-generated combustion mode knowledge graph based on the tensor, and matching the most similar reference control strategy. And finally, in combination with the real-time operation data, calculating and outputting a control instruction through the multivariable coordinator. By means of the method, deep coupling analysis and intelligent decision making of flame vision and the temperature field are achieved, the sensing precision and control adaptability of the combustion state are improved, and the combustion efficiency and stability are optimized.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for industrial kilns, specifically to a combustion optimization control method and system for a semi-coke furnace based on machine vision. Background Technology

[0002] In combustion control during semi-coke production, existing technologies typically rely on independent monitoring methods. Flame visual monitoring systems capture changes in flame shape, color, and brightness using cameras to qualitatively assess combustion stability; infrared thermography systems measure furnace temperature by scanning or at fixed points to obtain local or overall temperature distribution information. These two systems often operate in parallel, with data only undergoing simple comparisons or threshold alarms at the control level, failing to achieve a deep fusion analysis of visual characteristics and the dynamic evolution of the temperature field from a mechanistic perspective. This separate monitoring results in judgments of the combustion state remaining superficial or localized, unable to accurately reflect the coupling state between fuel, airflow, and complex combustion chemical reactions.

[0003] At the level of control strategy generation, existing methods largely rely on establishing precise combustion mathematical models or setting numerous "if-then" rules based on expert experience. Mathematical models heavily depend on prior parameters such as fuel characteristics and furnace structure, resulting in poor adaptability to fluctuations in raw materials and frequent changes in operating conditions, and high maintenance costs. While expert rule bases offer some flexibility, they cannot exhaustively cover all complex operating conditions, and conflicts between rules are difficult to coordinate, often leading to lagging and rigid control actions, failing to achieve real-time, flexible optimization based on the essential characteristics of combustion states. How to automatically extract features that essentially characterize the combustion coupling state from multi-source heterogeneous data and quickly match near-optimal control actions accordingly is a current technological challenge. Summary of the Invention

[0004] The purpose of this invention is to provide a combustion optimization control method and system for semi-coke furnaces based on machine vision, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a combustion optimization control method for a semi-coke furnace based on machine vision, the method comprising:

[0006] S1. Based on the image sequence of the semi-coke furnace combustion chamber within a set time period, a dynamic flame texture map is generated using an image processing algorithm. S2. Separate the infrared thermal imaging image subsequence from the image sequence, perform regional grid segmentation on the infrared thermal imaging image subsequence, calculate the statistical temperature value in each grid cell and the gradient of the statistical temperature value over time, and construct the spatiotemporal evolution matrix of the temperature field of the furnace based on the statistical temperature value and the gradient. S3. Input the flame dynamic texture map and the temperature field spatiotemporal evolution matrix into a pre-trained coupling analysis network to analyze the implicit correlation pattern between the dynamic changes of flame texture and the spatiotemporal evolution of temperature field, and output the coupling feature tensor representing the current combustion coupling state. S4. Based on the coupling feature tensor, perform similarity matching in the combustion mode knowledge graph to obtain the reference control strategy most similar to the coupling feature tensor; the combustion mode knowledge graph stores the mapping relationship between coupling feature tensors and optimal control strategies under various typical operating conditions; S5. Based on the reference control strategy, the real-time fuel supply flow rate data, primary air volume data and secondary air volume data of the semi-coke furnace are calculated by the multivariate coordinator to obtain the adjustment amount of the fuel quantity and air volume set value to make the combustion state approach the target coupling characteristics. S6. Based on the adjustment of the fuel quantity and air volume set values, generate a set of control instructions including the execution timing to drive the semi-coke furnace actuator to achieve combustion optimization control.

[0007] Preferably, the image sequence includes a visible light image reflecting the flame morphology and an infrared thermal imaging image reflecting the temperature field; step S1 includes: S11. Perform synchronous spatiotemporal registration on the image sequence to align the visible light image with the infrared thermal imaging image in spatial pixel coordinates and timestamps, thereby generating a spatiotemporally synchronized image sequence. S12. Based on the spatiotemporal synchronized image sequence, a visible light image subsequence is separated, and then transferred to a specific color space to separate brightness and chromaticity information. Dynamic texture features related to combustion stability are extracted to form a flame dynamic texture map.

[0008] Preferably, step S11 includes: S111. Perform feature point detection on the visible light image and the infrared thermal imaging image respectively, and extract stable corner points or edge features; S112. Based on the stable corner point or edge features, obtain the spatial transformation model parameters between the visible light image and the infrared thermal imaging image through a feature matching algorithm; S113. Based on the spatial transformation model parameters, the infrared thermal imaging image is spatially resampled to align the viewpoint and scale of the infrared thermal imaging image with that of the visible light image. S114. Timestamp the image sequence and perform inter-frame interpolation alignment between the visible light image and the infrared thermal imaging image based on the timestamps, so that the images at the same timestamp reflect the combustion state at the corresponding timestamp.

[0009] Preferably, step S12 includes: S121. Perform differential operation on multiple consecutive frames of the visible light image subsequence in the chroma channel to obtain a dynamic differential image sequence that reflects the color change of the flame. S122. Perform three-dimensional wavelet transform on the dynamic difference image sequence to extract wavelet coefficient energy of a specific frequency band, wherein the specific frequency band corresponds to the characteristic frequency range of flame shaking. S123. Arrange the wavelet coefficient energy in chronological order and expand it along the spatial dimension to form the flame dynamic texture map, wherein the horizontal axis of the flame dynamic texture map is time, the vertical axis is spatial position, and the map value is texture activity.

[0010] Preferably, step S2 includes: S21. Divide each frame of the infrared thermal imaging image subsequence into a regular grid according to the physical structure of the furnace. S22. Calculate the average temperature value of all infrared pixels in each grid of each frame image; S23. Based on the infrared thermal imaging image subsequence, record the temperature time series of the average temperature value of each grid changing with time, and calculate the temperature change rate of each grid at adjacent time points based on the temperature time series to obtain the temperature gradient; S24. Arrange the average temperature values ​​of all grids at the same time according to the grid position to form a first matrix, and arrange the temperature gradients of all grids at the same time according to the grid position to form a second matrix. Stack the first matrix and the second matrix based on the time dimension to form a spatiotemporal evolution matrix.

[0011] Preferably, the coupling analysis network includes a texture feature encoding path and a temperature feature encoding path; In step S3, the step of inputting the flame dynamic texture map and the temperature field spatiotemporal evolution matrix into a pre-trained coupled analysis network to analyze the implicit correlation patterns between the dynamic changes of flame texture and the spatiotemporal evolution of temperature field includes: S31. The flame dynamic texture map is convolved through the texture feature encoding path to extract multi-scale spatiotemporal texture features; S32. The temperature field spatiotemporal evolution matrix is ​​convolved through the temperature feature encoding path to extract multi-scale spatiotemporal temperature distribution and change features. S33. In the fusion layer of the coupled analysis network, the multi-scale spatiotemporal texture features and the multi-scale spatiotemporal temperature distribution and change features are subjected to outer product interaction processing to obtain a high-order interaction feature map. S34. Perform global pooling and fully connected transformation on the high-order interaction feature map to obtain the coupling feature tensor representing the combustion coupling state.

[0012] Preferably, the combustion mode knowledge graph is constructed from historical operating data, and the nodes in the combustion mode knowledge graph include: coupled feature tensor samples extracted under various historical operating conditions and the control parameter combinations corresponding to the coupled feature tensor samples; Step S4 includes: S41. Perform similarity matching between the coupled feature tensor and the coupled feature tensor samples of all nodes in the combustion mode knowledge graph; S42. Filter all nearest neighbor nodes whose similarity exceeds the set threshold; S43. Generate a reference control strategy based on the combination of control parameters corresponding to all the neighboring nodes.

[0013] Preferably, the multivariate coordinator has a built-in simplified dynamic model, which is configured to describe the dynamic mapping relationship between fuel quantity, primary air volume, secondary air volume and coupled feature tensor; The multivariate coordinator also defines an objective function for minimizing the gap between the combustion state predicted by the simplified dynamic model and the desired target in the prediction time domain, and for penalizing excessive changes in the control quantity. The multivariable coordinator is also used to set process constraints, which include at least: upper and lower limits of air volume, fuel valve opening limit, and safe range of air-fuel ratio. Step S5 includes: S51. The combination of control parameters in the reference control strategy is taken as the desired target, and the current fuel supply flow rate data, primary air volume data and secondary air volume data are taken as the initial state and input into the simplified dynamic model. S52. Obtain the control quantity change sequence that optimizes the objective function under the process constraints, and use the instantaneous control quantity change value in the control quantity change sequence as the adjustment amount of the fuel quantity and air volume setpoint.

[0014] Preferably, step S6 includes: S61. Based on the analysis results of the adjustment amount of the fuel quantity and air volume setpoints, generate independent adjustment commands for the fuel supply valve, primary air damper and secondary air damper respectively. S62. Based on the sequential logical order, assign an execution start time and execution duration to each independent adjustment instruction; the sequential logical order is that primary air volume takes priority over fuel volume, and fuel volume takes priority over secondary air volume. S63. Divide the adjustment amount of each independent adjustment command into multiple adjustment steps, set an execution interval for each adjustment step, and form a smooth ramp control signal. S64. Package all independent adjustment instructions into the control instruction set.

[0015] The present invention also provides a combustion optimization control system for a semi-coke furnace based on machine vision. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the combustion optimization control method for a semi-coke furnace based on machine vision described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By designing a coupled analysis network, the dynamic texture map of the flame and the spatiotemporal evolution matrix of the temperature field are simultaneously input and subjected to deep feature learning. This network can automatically uncover the nonlinear correlation and spatiotemporal synergistic laws between flame morphological pulsations, texture changes, and internal furnace temperature distribution and gradient migration. This process goes beyond simple data overlay, generating a unified, high-dimensional coupled feature tensor. This tensor integrates the core information of the visual radiation field and the thermal field, forming a deep, integrated digital representation of the multi-physics interaction state within the combustion chamber. This allows the control system to perceive the combustion state from a separate, superficial level to a comprehensive, essential level.

[0017] Control matching is performed based on a pre-generated combustion mode knowledge graph, which stores the correspondence between coupled feature tensors under various historical typical optimal operating conditions and verified optimal control strategies. The system performs rapid similarity retrieval and matching between the real-time generated coupled feature tensors and features in the knowledge graph, directly mapping them to historically proven optimal strategy references. This method bypasses the complex and time-consuming real-time physical modeling or rule reasoning process, transforming control decision-making into rapid pattern recognition and case retrieval of high-dimensional state features, thus improving decision-making speed. Furthermore, since the knowledge graph originates from accumulated actual optimal operating conditions, its recommended control strategies naturally possess stronger operating condition adaptability and practical reliability, realizing a shift from "model-driven" or "rule-driven" to "case data-driven" intelligent decision-making. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the machine vision-based combustion optimization control method for semi-coke furnaces described in this invention. Figure 2 A flowchart for generating a dynamic flame texture map; Figure 3 A flowchart for extracting dynamic texture features to form a dynamic flame texture map; Figure 4 This is a graph showing the similarity distribution of coupling features and the nearest neighbor selection. Figure 5 For multivariate coordinated optimization process and predictive control chart. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides a machine vision-based combustion optimization control method for a semi-coke furnace, comprising: generating a dynamic flame texture map based on an image processing algorithm using an image processing algorithm, based on an image sequence of the semi-coke furnace combustion chamber acquired within a set time period; separating an infrared thermal imaging image subsequence from the image sequence, performing regional grid segmentation on the infrared thermal imaging image subsequence, calculating the statistical temperature value and the gradient of the statistical temperature value over time within each grid cell, and constructing a spatiotemporal evolution matrix of the furnace temperature field based on the statistical temperature value and the gradient of the statistical temperature value over time; inputting the dynamic flame texture map and the spatiotemporal evolution matrix of the furnace temperature field into a coupling analysis network, which is used to mine the implicit correlation patterns between the dynamic changes of the flame texture and the spatiotemporal evolution of the temperature field, and outputting a coupling feature tensor representing the current combustion coupling state; and querying a pre-generated combustion mode knowledge graph based on the coupling feature tensor, which stores the mapping relationships between the coupling feature tensors and the optimal control strategies under various typical operating conditions, and matching the reference control strategy most similar to the current coupling feature tensor. Real-time fuel supply flow rate, primary air volume, and secondary air volume data of the semi-coke furnace are acquired. Combined with this reference control strategy, a multivariate coordinator is used to calculate the adjustment amounts of the fuel and air volume setpoints to approximate the target coupling characteristics of the combustion state, while satisfying process constraints. Based on these adjustments, a control instruction set including execution timing is generated to drive the semi-coke furnace actuators to achieve optimized combustion control.

[0021] Example 1: See Figure 2In specific implementation, a sequence of images of the combustion chamber of a semi-coke furnace within a set time period is collected to generate a dynamic flame texture map. The set time period is, for example, 1 second. This image sequence includes visible light images reflecting the flame morphology and infrared thermal images reflecting the temperature field. Both the visible light and infrared thermal images are acquired at a frequency of 30 frames per second, resulting in 30 frames of visible light images and 30 frames of infrared thermal images per second. The image sequence is then synchronized spatiotemporally and spatially registered to align the visible light and infrared thermal images in terms of spatial pixel coordinates and timestamps, generating a registered spatiotemporally synchronized image sequence. In unregistered examples, the pixel coordinates of the center point of the flame region in the visible light image deviate from the pixel coordinates of the corresponding highest temperature point in the infrared thermal image, with a deviation range of 20 to 50 pixels. After synchronization spatiotemporally and spatially registration, this deviation is reduced to within 2 pixels. Visible light image subsequences are separated from spatiotemporally synchronized image sequences and converted to a specific color space. Dynamic texture features related to combustion stability are extracted in this specific color space to form a flame dynamic texture map. The dimension of the flame dynamic texture map is time multiplied by spatial position, where the spatial position corresponds to the unfolded arrangement of pixels in the visible light image subsequence.

[0022] In some embodiments, synchronous spatiotemporal registration of the image sequence further includes feature point detection for both the visible light image and the infrared thermal imaging image, i.e., extracting stable corner points or edges from the images. Based on the extracted feature points, a feature matching algorithm is used to calculate the spatial transformation model parameters between the visible light image and the infrared thermal imaging image. The infrared thermal imaging image is spatially resampled using the spatial transformation model parameters to align its viewpoint and scale with that of the visible light image. In this embodiment, the spatial transformation model parameters are affine transformation matrices, including rotation, translation, and scaling components. For the time dimension, each frame in the image sequence is labeled with a precise timestamp, and inter-frame interpolation alignment is performed on the visible light image and the infrared thermal imaging image based on the timestamp to ensure that the image content corresponding to the same timestamp reflects the combustion state at the same instant. In a specific example, due to synchronization errors in the acquisition hardware of the visible light image and the infrared thermal imaging image, after timestamp annotation, they are aligned to a unified millisecond-level time grid using a linear interpolation method.

[0023] In the embodiments of this application, the feature matching algorithm uses descriptors based on scale-invariant feature transformation for matching. The matching process calculates the Euclidean distance between feature points in the visible light image and feature points in the infrared thermal imaging image; if the distance is less than a threshold, it is considered a matching pair. The spatial transformation model parameters are solved using a random sampling consensus algorithm to estimate the optimal transformation from the matching pairs. The transformation error function is defined as the sum of squares of the distances between matching feature point pairs, and the objective is to minimize the error function. It can be understood that the minimization of the error function is achieved through iterative optimization. In each iteration, the smallest subset of matching pairs is randomly selected to calculate the transformation parameters, and the number of interior points is evaluated.

[0024] In some embodiments, extracting dynamic texture features related to combustion stability includes converting the registered visible light image subsequence from the original color space to a color space that separates luminance and chromaticity information, using the YCbCr color space. Differential operations are performed on multiple consecutive frames of images in the chromaticity channel of this color space to obtain a dynamic difference image sequence reflecting color changes in the flame region. The differential operations are performed separately for the Cb and Cr channels, calculating the absolute difference between corresponding pixels in adjacent frames. A three-dimensional wavelet transform is performed on the dynamic difference image sequence to extract wavelet coefficient energy in specific frequency bands. These specific frequency bands correspond to the characteristic frequency range of flame jitter, determined through historical data analysis, for example, 5 Hz to 15 Hz. The extracted wavelet coefficient energy is arranged chronologically and expanded along the spatial dimension to construct a dynamic flame texture map, where the horizontal axis represents time, the vertical axis represents spatial location, and the map values ​​represent the texture activity at that spatiotemporal point. The texture activity values ​​are normalized between 0 and 1.

[0025] In this embodiment, the three-dimensional wavelet transform uses the Daubechies wavelet basis, with a decomposition layer of 3. The extracted specific frequency bands are the approximation coefficients or detail coefficients of the third layer, and the wavelet coefficient energy is obtained by calculating the sum of squared coefficients. It can be understood that the generation process of the flame dynamic texture map involves data dimensionality reduction. Spatial dimension expansion arranges the pixels of the two-dimensional image into a one-dimensional vector in row-major order, while the temporal dimension preserves the sequence order. The final map is in two-dimensional matrix form. In the example scenario, the resolution of the visible light image subsequence is 640x480 pixels. The dimension of the generated flame dynamic texture map is 30 (time frame) multiplied by 307200 (spatial location), reducing the data volume by approximately 40% compared to the original image sequence, retaining only the dynamic texture feature information related to combustion stability.

[0026] Example 2: See Figure 3In the specific implementation, dynamic texture features related to combustion stability are extracted to form a flame dynamic texture map. The registered visible light image subsequence is converted from the original color space to a color space that separates brightness and chromaticity information. The YCbCr color space is selected for separating brightness and chromaticity information. In the example scenario, the original color space is sRGB. The conversion process applies standard linear and nonlinear transformation matrices to each 640x480 pixel visible light image frame. On the chromaticity channel of the color space, differential operations are performed on multiple consecutive frames of images to obtain a dynamic difference image sequence reflecting the color changes in the flame area. The differential operation is performed independently for the Cb and Cr channels, calculating the absolute difference of the chromaticity value at each pixel position between two adjacent frames. A visible light image subsequence containing 30 frames is converted into a sequence containing 29 dynamic difference images after differential operations. Each dynamic difference image contains two components: a Cb difference image and a Cr difference image.

[0027] A three-dimensional wavelet transform is performed on the dynamic difference image sequence to extract wavelet coefficient energy in a specific frequency band. The three-dimensional wavelet transform is performed in two spatial dimensions and a temporal dimension, using the 'db4' wavelet as the mother wavelet. The decomposition level is three layers. The specific frequency band corresponds to the characteristic frequency range of flame shaking, which is preset to 5 Hz to 15 Hz. This frequency band corresponds to the approximate coefficients of the third layer after wavelet decomposition. The extracted wavelet coefficient energy is arranged in chronological order and expanded along the spatial dimension to form a flame dynamic texture map. The horizontal axis of the map represents time, and its length is equal to the number of data points in the time dimension after the wavelet transform. The vertical axis represents spatial location, and its length is equal to the product of the image spatial resolution and the spatial dimension after wavelet transform downsampling. The value of each point in the map represents the wavelet coefficient energy at that spatiotemporal point, which is obtained by squaring the wavelet coefficient corresponding to that point.

[0028] A spatiotemporal evolution matrix of the temperature field in the furnace is constructed by dividing each frame of the infrared thermal imaging image subsequence into a regular row and column grid according to the physical structure of the furnace. In one example, an infrared image with a resolution of 320x240 pixels is divided into 20 rows and 15 columns, totaling 300 grid cells. The physical size of each grid cell corresponds to a fixed spatial region within the furnace. The average temperature value of all infrared pixels in each grid in each frame is calculated as the representative temperature of that grid at that moment. For a grid containing 256 pixels, the average temperature value is obtained by summing the temperature values ​​of these pixels and dividing by 256. All frames of the infrared thermal imaging image subsequence are traversed, and the sequence of representative temperature changes over time for each grid is recorded. A subsequence containing 30 frames will generate a representative temperature time series of length 30 for each of the 300 grids. For the temperature time series of each grid, the rate of temperature change between adjacent time points is calculated to form a temperature gradient sequence. The length of the temperature gradient sequence is 29, and the calculation formula is as follows:

[0029] in: This represents the temperature gradient of the grid in the i-th row and j-th column at time t. This indicates the representative temperature of the grid at time t. It is the time interval between adjacent frames. In an example of 30 frames per second, The time interval is 1 / 30 of a second. The representative temperatures of all grids at the same moment are arranged into a two-dimensional matrix according to their grid positions. The matrix has 20 rows and 15 columns. The temperature gradients of all grids at the same moment are arranged into another two-dimensional matrix with the same dimensions according to their positions. These two-dimensional matrices are stacked along the time dimension to form the spatiotemporal evolution matrix of the furnace temperature field. The final matrix is ​​a four-dimensional tensor with dimensions of 30 (time) x 2 (data type: temperature and gradient) x 20 (rows) x 15 (columns).

[0030] In some embodiments, the specific process of three-dimensional wavelet transform includes performing a one-dimensional discrete wavelet transform on the dynamic difference image sequence in the time dimension, and a two-dimensional discrete wavelet transform on the width and height spatial dimensions of the image. After three-level decomposition, the original data is decomposed into multiple sub-bands of different frequency bands. The wavelet coefficient energy extraction operation for specific frequency bands focuses on the low-frequency approximate sub-band generated by the third-level decomposition. The coefficients of this sub-band are considered to contain the main energy information of flame flickering within a preset characteristic frequency range. It can be understood that the flame dynamic texture map is ultimately a two-dimensional mapping representation of spatiotemporal-energy features, which compresses and transforms the original high-dimensional image sequence data into a feature form that better characterizes combustion instability.

[0031] In the embodiments of this application, the mesh division can be non-uniformly divided according to the physical characteristics of the combustion zone within the furnace. For example, a denser mesh can be used in the high-temperature region of the flame core, while a sparser mesh can be used in the edge region. The number of rows and columns of the mesh is determined based on the actual furnace structure dimensions and the resolution of the infrared thermal imaging camera. When calculating the representative temperature of each mesh, in addition to the arithmetic mean, the median or the average value after removing outliers can also be used as the statistical temperature value to resist interference from noise points that may exist in the image. The calculation of the temperature gradient sequence can use a higher-order difference format, such as central difference, to obtain a smoother estimate of the rate of temperature change. In some embodiments, the construction of the spatiotemporal evolution matrix of the furnace temperature field can use only the representative temperature matrix without including the temperature gradient matrix, or it can include higher-order statistical features, such as the temperature variance within each mesh. It is understood that an evolution matrix including the temperature gradient can simultaneously reflect the static information of the spatial distribution of the temperature field and the dynamic trend information of its evolution over time.

[0032] Example 3: In specific implementation, the implicit correlation pattern between the dynamic changes of flame texture and the spatiotemporal evolution of the temperature field is explored. The coupling analysis network includes parallel texture feature encoding pathways and temperature feature encoding pathways. The texture feature encoding pathway performs convolution processing on the input flame dynamic texture map to extract multi-scale spatiotemporal texture features. The dimension of the flame dynamic texture map is the time dimension T multiplied by the spatial dimension S. The texture feature encoding pathway adopts a three-dimensional convolutional neural network structure. Its first layer uses a filter with a convolution kernel size of 3x3x3 and a number of 16 to perform convolution and ReLU activation operations on the input map, followed by a 3x3x3 max pooling layer with a stride of 2. The temperature feature encoding pathway performs convolution processing on the input spatiotemporal evolution matrix of the furnace temperature field to extract multi-scale spatiotemporal temperature distribution and change features. The dimensions of the spatiotemporal evolution matrix of the furnace temperature field are time dimension T multiplied by feature type dimension C multiplied by spatial height H multiplied by spatial width W. The feature type dimension C contains two data types representing temperature and temperature gradient. The input layer of the temperature feature encoding pathway first merges the feature type dimension C with the spatial dimension to form a three-dimensional tensor Tx(CH)xW. Then, it is processed by a three-dimensional convolutional layer and pooling layer with independent parameters, which are similar in structure to the texture feature encoding pathway.

[0033] In the fusion layer of the coupling analysis network, multi-scale spatiotemporal texture features and multi-scale spatiotemporal temperature distribution and variation features are subjected to an outer product interaction operation to generate a high-order interaction feature map. Assuming the feature tensor output by the texture feature encoding path has dimensions [T1, S1, D1] and the feature tensor output by the temperature feature encoding path has dimensions [T1, S1, D2], the outer product interaction operation is performed on the last feature dimension, generating a high-order interaction feature map with dimensions [T1, S1, D1D2]. Global pooling and fully connected transformations are applied to the high-order interaction feature map to compress it and obtain a coupling feature tensor representing the current combustion coupling state. Global pooling is then applied along the time dimension T1 and the spatial dimension S1 to average the high-order interaction feature map, resulting in a tensor of length D1. The one-dimensional vector of D2 is then passed through a two-layer fully connected network with 256 neurons, ultimately outputting a coupled feature tensor of length 128.

[0034] The reference control strategy most similar to the current coupled feature tensor is obtained through matching. The combustion mode knowledge graph is constructed from historical operating data. Nodes in the graph include coupled feature tensor samples extracted from various historical operating conditions, and corresponding control parameter combinations that have been verified as efficient or stable for each sample. The historical operating data comes from the stable operation phase of the semi-coke furnace under different loads and fuel characteristics over the past six months, extracting approximately 5000 coupled feature tensor samples and their corresponding control parameter combinations to form the knowledge graph nodes. The similarity between the currently calculated coupled feature tensor and the coupled feature tensor samples of all nodes in the combustion mode knowledge graph is calculated using cosine similarity, with the formula:

[0035] in: This represents the currently calculated coupled feature tensor of length 128. This represents the historical coupling feature tensor sample of the k-th node in the combustion mode knowledge graph. This represents the dot product operation of vectors. This represents the L2 norm of the vector. Several nearest neighbor nodes whose similarity metric exceeds a set threshold (e.g., 0.92) are selected. If the number is less than five, the five nodes with the highest similarity are chosen as nearest neighbors. A preliminary reference control strategy is generated by combining the control parameter combinations corresponding to these nearest neighbors through a weighted average or voting mechanism. The control parameter combinations include fuel supply flow rate setpoints, primary air volume setpoints, and secondary air volume setpoints. The weights of the weighted average are determined by normalizing the similarity values ​​between each nearest neighbor node and the current coupled feature tensor.

[0036] In some embodiments, the network structure of the texture feature encoding pathway can include multiple cascaded three-dimensional convolutional blocks. Each convolutional block consists of a convolutional layer, a batch normalization layer, an activation function layer, and a pooling layer. Multi-scale spatiotemporal texture features are captured by gradually increasing the number of filters (e.g., 16, 32, 64) and expanding the receptive field of the convolutional kernels. Before merging the feature type dimension C with the spatial dimension, the temperature feature encoding pathway can perform independent convolutional preprocessing on the temperature data and temperature gradient data to extract static temperature distribution features and dynamic change features respectively. The two feature streams are then merged and input into the subsequent common encoding layer. It can be understood that the outer product interaction operation can explicitly model the pairwise correlation between texture features and temperature features in different channels. This relationship may correspond to combustion physics mechanisms, such as the correlation between flame flicker frequency and local temperature change rate.

[0037] In the embodiments of this application, the training of the coupling analysis network is a supervised process, using historical data pairs labeled with combustion state tags such as "efficient," "stable," or "poor." The training objective is to minimize the distance between the network's output coupling feature tensor and the corresponding tag in the feature space. The construction process of the combustion mode knowledge graph includes clustering operations on historical coupling feature tensor samples, grouping samples with similar features into the same typical operating condition mode, and using the feature centroid of the mode and the mean of its corresponding optimized control parameter combination as a node, thereby reducing the number of redundant nodes in the graph. The similarity measure can also use the reciprocal of Euclidean distance or other distance measures. The number of nearest neighbor nodes can be a fixed value K, for example, K=5, that is, directly selecting the 5 nodes with the highest similarity without setting a threshold. When generating the reference control strategy by weighted average, a secondary correction can be introduced for macroscopic parameters such as global load and fuel calorific value of historical and current operating conditions to make the generated strategy more targeted. In some embodiments, the voting mechanism is suitable for cases where the control parameters are discrete values ​​or categories. Each nearest neighbor node votes for its own control parameter combination, and the combination with the most votes is selected as the reference control strategy.

[0038] See Figure 4This chart illustrates the similarity matching process between the coupled feature tensor and historical samples. The blue bars represent the distribution of all historical samples sorted by similarity, the red bars represent selected nearest neighbor nodes, and the green dashed line represents the similarity threshold. The chart visually demonstrates the degree of matching between the coupled feature tensor corresponding to the current combustion state and the historical samples stored in the knowledge graph. Higher similarity values ​​indicate a greater similarity between the current operating condition and historical operating conditions. The selected nearest neighbor nodes will be used to generate a reference control strategy, and the control parameter combinations corresponding to these nodes will be fused using a weighted average. The similarity distribution in the chart exhibits some fluctuation, reflecting the feature diversity under different operating conditions. The threshold line ensures that only highly similar operating conditions are used to generate the control strategy, thus guaranteeing the reliability of the control recommendations.

[0039] Example 4: In specific implementation, the adjustment amounts of fuel and air volume setpoints that make the combustion state approximate the target coupling characteristics are solved. The multivariate coordinator has a built-in simplified dynamic model of the semi-coke furnace combustion process. The simplified dynamic model describes the approximate mathematical relationship between fuel quantity, primary air volume, secondary air volume, and coupling characteristic tensor. Taking the combination of control parameters in the reference control strategy as the desired target, and the currently measured fuel supply flow rate data, primary air volume data, and secondary air volume data as the initial state input to the simplified dynamic model, in an example scenario, the desired target values ​​given by the reference control strategy are a fuel supply flow rate setpoint of 1200 cubic meters per hour, a primary air volume setpoint of 8000 standard cubic meters per hour, and a secondary air volume setpoint of 2000 standard cubic meters per hour. The currently measured fuel supply flow rate data is 1180 cubic meters per hour, the primary air volume data is 7900 standard cubic meters per hour, and the secondary air volume data is 2050 standard cubic meters per hour. In the multivariate coordinator, an objective function is defined to minimize the difference between the combustion state predicted by the model and the desired target over a future prediction time domain, while penalizing excessive changes in control variables. The mathematical expression of the objective function is as follows:

[0040] in: It is the objective function value. This refers to the prediction time domain length, for example, set to 10 control cycles. It is the coupled feature tensor after the nth step of the simplified dynamic model prediction. It is the expected coupled feature tensor obtained by mapping the setpoint of the reference control strategy through the model. Represented by matrix The weighted Euclidean norm squared for the weights. This controls the length of the time domain, for example, setting it to 5 control cycles. This is the control quantity change vector for step n, which includes adjustments to fuel flow rate, primary air volume, and secondary air volume. Represented by matrix The weighted Euclidean norm squared is the weight of the weights.

[0041] Referring to Table 1, process constraints are set in the multivariate coordinator. These constraints include upper and lower limits of airflow, fuel valve opening limits, and the safe range of the air-fuel ratio. The optimal sequence of control variables that satisfies the process constraints is then determined. The instantaneous control variable changes in this sequence are used as the adjustment amounts for the fuel and airflow setpoints; that is, the optimized sequence is adopted. This is the adjustment that needs to be performed at the current moment.

[0042] Table 1. Process Constraints of the Multivariable Coordinator

[0043] In some embodiments, the simplified dynamic model can be established using a system identification method based on linear time-varying or Hammerstein-Wiener structures. Its inputs are three control variables: fuel supply flow rate, primary air volume, and secondary air volume. The output is several principal components of a coupled feature tensor. The model parameters are obtained by fitting step responses or pseudo-random sequence test data from historical operating data. (Prediction time domain) and control time domain The length of the curve is determined based on the dominant time constant of the semi-coke furnace combustion process, which can be obtained by analyzing the response curve of historical temperature fields or flame textures. It can be understood that the weight matrix in the objective function... and The selection balances tracking accuracy and control smoothness. The diagonal elements of the matrix are typically set to positive numbers that relate to the importance of each component of the coupled feature tensor. The diagonal elements of the matrix are used to limit the range of motion of each actuator.

[0044] Optionally, the safe range of the wind-fuel ratio in the process constraints needs to be dynamically adjusted based on real-time or periodic test data of the fuel calorific value. When the fuel calorific value fluctuates beyond the set range, the upper and lower limits of the safe range of the wind-fuel ratio are scaled proportionally accordingly. Solving the constrained optimization problem can employ numerical optimization algorithms such as sequential quadratic programming, interior-point methods, or effective set methods. Considering the real-time requirements of online computation, the optimization problem can also be transformed into a quadratic programming problem and a dedicated solver can be called for computation. In some embodiments, if the real-time solution fails or times out, the multivariate coordinator can enable a rule-based backoff strategy. This strategy directly outputs a predefined small fixed adjustment amount based on the sign of the deviation between the current state and the desired target to ensure the basic safe operation of the system.

[0045] See Figure 5 This diagram illustrates the optimization control process of a multivariate coordinator. The graph shows the trajectories of fuel flow rate, primary air volume, and secondary air volume in the prediction time domain, along with the corresponding control adjustments. The blue solid line represents the predicted change path of fuel flow rate, the green solid line represents the change in primary air volume (scaled), and the red solid line represents the change in secondary air volume (scaled). Dashed lines represent the target setpoints for each parameter. Purple bars represent the total adjustment for each control step, reflecting the magnitude of the control action. The graph shows how the multivariate coordinator, through a model predictive control algorithm, gradually adjusts the current combustion state to the target state while meeting process constraints. The optimization process considers a balance between control accuracy and smoothness to ensure a stable transition in the combustion process. It can be observed from the graph that the control adjustments are relatively large initially, then gradually converge to the target value, which is consistent with the dynamic characteristics of the combustion process. Air volume adjustments are usually prioritized over fuel quantity adjustments to maintain a suitable air-fuel ratio.

[0046] Example 5: In specific implementation, a control instruction set including execution timing is generated based on the fuel and air volume setpoint adjustment amount. The fuel and air volume setpoint adjustment amount is parsed and decomposed into independent adjustment instructions for the fuel supply valve, primary air damper, and secondary air damper. Assuming that the fuel and air volume setpoint adjustment amount obtained from the multivariate coordinator is an increase of 20 cubic meters per hour in fuel flow, an increase of 100 standard cubic meters per hour in primary air volume, and a decrease of 30 standard cubic meters per hour in secondary air volume, the parsing process generates three independent adjustment instructions, namely "adjust the fuel supply valve, with a target increase in flow of 20 cubic meters per hour", "adjust the primary air damper, with a target increase in air volume of 100 standard cubic meters per hour", and "adjust the secondary air damper, with a target decrease in air volume of 30 standard cubic meters per hour". Each independent adjustment command is assigned an execution start time and execution duration to ensure that the actions of multiple commands follow a preset logical order. The adjustment of primary air volume takes precedence over the adjustment of fuel volume, and the adjustment of fuel volume takes precedence over the adjustment of secondary air volume. In an example with a control cycle of 1 minute, the execution start time of the primary air volume adjustment command is set to T0, and the execution duration is 10 seconds. The execution start time of the fuel volume adjustment command is set to T0+2 seconds, and the execution duration is 8 seconds. The execution start time of the secondary air volume adjustment command is set to T0+5 seconds, and the execution duration is 5 seconds. This logical order ensures that air volume changes precede fuel volume changes, and secondary air volume changes follow after fuel volume changes are initiated.

[0047] In each independent adjustment command, the total adjustment amount is subdivided into multiple small adjustment steps, and an execution interval is set for each step, forming a smooth ramp-like control signal. For a command to increase the primary air volume by 100 standard cubic meters per hour, the total adjustment amount is subdivided into 5 steps, each step increasing by 20 standard cubic meters per hour, with an execution interval of 2 seconds. The calculation formula is as follows:

[0048] in: This represents the change in each adjustment step. This indicates the total adjustment amount of the instruction. This represents the total number of subdivision steps, for fuel flow adjustment. cubic meters per hour, set Then each step size The flow rate is cubic meters per hour, with an execution interval of 2 seconds. All independent adjustment instructions with precise timing labels are packaged and encapsulated into a control instruction set. The control instruction set is a structured data list, and each instruction contains the actuator identifier, target adjustment amount, step size, step execution interval, absolute timestamp of instruction start execution, and total duration.

[0049] In some embodiments, the allocation of the execution time start point relies on a unified system clock to ensure that all control commands are strictly synchronized in the time dimension. The start timestamp of the command is typically set after a fixed network transmission and scheduling delay, such as 500 milliseconds after the current time. The setting of the execution duration needs to consider the mechanical response characteristics of different actuators (such as valves and dampers). Longer durations are allocated to actuators with high inertia, and shorter durations are allocated to fast-response actuators to match their physical action capabilities. It can be understood that subdividing the total adjustment into multiple small steps and executing them at intervals can avoid abrupt shocks to the combustion system. This ramp-like control signal allows key parameters such as fuel flow and air volume to smoothly transition to the new set values, which is beneficial for maintaining the stability of the combustion process.

[0050] In the embodiments of this application, the number of adjustment steps Based on the total adjustment amount The size is dynamically determined, following a preset mapping relationship, for example when When it is less than the threshold A, ;when When the thresholds are between A and B ;when When it is greater than the threshold B, The execution interval can also be non-fixed, arranged in a denser-to-sparser or sparser-to-dense manner to adapt to different dynamic response requirements. In some embodiments, the control instruction set undergoes a logical consistency check before encapsulation. The check includes confirming whether the execution time intervals of each instruction overlap or conflict, whether the cumulative adjustment of each step is equal to the total adjustment, and whether the timestamp of the instruction is within a reasonable future time window. It can be understood that the control instruction set containing precise execution timing is the bridge connecting the optimization decision layer and the underlying physical execution layer. It transforms the abstract adjustment value into a series of specific action commands that are precisely arranged on the time axis and finely divided in magnitude, thereby driving the actual control system of the semi-coke furnace to complete the optimization action.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A combustion optimization control method for a semi-coke furnace based on machine vision, characterized in that, The method includes: S1. Based on the image sequence of the semi-coke furnace combustion chamber within a set time period, a dynamic flame texture map is generated using an image processing algorithm. S2. Separate the infrared thermal imaging image subsequence from the image sequence, perform regional grid segmentation on the infrared thermal imaging image subsequence, calculate the statistical temperature value in each grid cell and the gradient of the statistical temperature value over time, and construct the spatiotemporal evolution matrix of the temperature field of the furnace based on the statistical temperature value and the gradient. S3. Input the flame dynamic texture map and the temperature field spatiotemporal evolution matrix into a pre-trained coupling analysis network to analyze the implicit correlation pattern between the dynamic changes of flame texture and the spatiotemporal evolution of temperature field, and output the coupling feature tensor representing the current combustion coupling state. S4. Based on the coupling feature tensor, perform similarity matching in the combustion mode knowledge graph to obtain the reference control strategy most similar to the coupling feature tensor; the combustion mode knowledge graph stores the mapping relationship between coupling feature tensors and optimal control strategies under various typical operating conditions; S5. Based on the reference control strategy, the real-time fuel supply flow rate data, primary air volume data and secondary air volume data of the semi-coke furnace are calculated by the multivariate coordinator to obtain the adjustment amount of the fuel quantity and air volume set value to make the combustion state approach the target coupling characteristics. S6. Based on the adjustment of the fuel quantity and air volume set values, generate a set of control instructions including the execution timing to drive the semi-coke furnace actuator to achieve combustion optimization control.

2. The combustion optimization control method for a semi-coke furnace based on machine vision according to claim 1, characterized in that, The image sequence includes visible light images reflecting the shape of the flame and infrared thermal imaging images reflecting the temperature field. Step S1 includes: S11. Perform synchronous spatiotemporal registration on the image sequence to align the visible light image with the infrared thermal imaging image in spatial pixel coordinates and timestamps, thereby generating a spatiotemporally synchronized image sequence. S12. Based on the spatiotemporal synchronized image sequence, a visible light image subsequence is separated, and then transferred to a specific color space to separate brightness and chromaticity information. Dynamic texture features related to combustion stability are extracted to form a flame dynamic texture map.

3. The combustion optimization control method for a semi-coke furnace based on machine vision according to claim 2, characterized in that, Step S11 includes: S111. Perform feature point detection on the visible light image and the infrared thermal imaging image respectively, and extract stable corner points or edge features; S112. Based on the stable corner point or edge features, obtain the spatial transformation model parameters between the visible light image and the infrared thermal imaging image through a feature matching algorithm; S113. Based on the spatial transformation model parameters, the infrared thermal imaging image is spatially resampled to align the viewpoint and scale of the infrared thermal imaging image with that of the visible light image. S114. Timestamp the image sequence and perform inter-frame interpolation alignment between the visible light image and the infrared thermal imaging image based on the timestamps, so that the images at the same timestamp reflect the combustion state at the corresponding timestamp.

4. The combustion optimization control method for a semi-coke furnace based on machine vision according to claim 2, characterized in that, Step S12 includes: S121. Perform differential operation on multiple consecutive frames of the visible light image subsequence in the chroma channel to obtain a dynamic differential image sequence that reflects the color change of the flame. S122. Perform three-dimensional wavelet transform on the dynamic difference image sequence to extract wavelet coefficient energy of a specific frequency band, wherein the specific frequency band corresponds to the characteristic frequency range of flame shaking. S123. Arrange the wavelet coefficient energy in chronological order and expand it along the spatial dimension to form the flame dynamic texture map, wherein the horizontal axis of the flame dynamic texture map is time, the vertical axis is spatial position, and the map value is texture activity.

5. The combustion optimization control method for a semi-coke furnace based on machine vision according to claim 1, characterized in that, Step S2 includes: S21. Divide each frame of the infrared thermal imaging image subsequence into a regular grid according to the physical structure of the furnace. S22. Calculate the average temperature value of all infrared pixels in each grid of each frame image; S23. Based on the infrared thermal imaging image subsequence, record the temperature time series of the average temperature value of each grid changing with time, and calculate the temperature change rate of each grid at adjacent time points based on the temperature time series to obtain the temperature gradient; S24. Arrange the average temperature values ​​of all grids at the same time according to the grid position to form a first matrix, and arrange the temperature gradients of all grids at the same time according to the grid position to form a second matrix. Stack the first matrix and the second matrix based on the time dimension to form a spatiotemporal evolution matrix.

6. The combustion optimization control method for a semi-coke furnace based on machine vision according to claim 1, characterized in that, The coupled analysis network includes a texture feature encoding path and a temperature feature encoding path; In step S3, the step of inputting the flame dynamic texture map and the temperature field spatiotemporal evolution matrix into a pre-trained coupled analysis network to analyze the implicit correlation patterns between the dynamic changes of flame texture and the spatiotemporal evolution of temperature field includes: S31. The flame dynamic texture map is convolved through the texture feature encoding path to extract multi-scale spatiotemporal texture features; S32. The temperature field spatiotemporal evolution matrix is ​​convolved through the temperature feature encoding path to extract multi-scale spatiotemporal temperature distribution and change features. S33. In the fusion layer of the coupled analysis network, the multi-scale spatiotemporal texture features and the multi-scale spatiotemporal temperature distribution and change features are subjected to outer product interaction processing to obtain a high-order interaction feature map. S34. Perform global pooling and fully connected transformation on the high-order interaction feature map to obtain the coupling feature tensor representing the combustion coupling state.

7. The combustion optimization control method for a semi-coke furnace based on machine vision according to claim 1, characterized in that, The combustion mode knowledge graph is constructed from historical operating data. The nodes in the combustion mode knowledge graph include: coupled feature tensor samples extracted under various historical operating conditions and the control parameter combinations corresponding to the coupled feature tensor samples. Step S4 includes: S41. Perform similarity matching between the coupled feature tensor and the coupled feature tensor samples of all nodes in the combustion mode knowledge graph; S42. Filter all nearest neighbor nodes whose similarity exceeds the set threshold; S43. Generate a reference control strategy based on the combination of control parameters corresponding to all the neighboring nodes.

8. The combustion optimization control method for a semi-coke furnace based on machine vision according to claim 7, characterized in that, The multivariate coordinator has a built-in simplified dynamic model, which is configured to describe the dynamic mapping relationship between fuel quantity, primary air volume, secondary air volume and coupled feature tensor. The multivariate coordinator also defines an objective function for minimizing the gap between the combustion state predicted by the simplified dynamic model and the desired target in the prediction time domain, and for penalizing excessive changes in the control quantity. The multivariable coordinator is also used to set process constraints, which include at least: upper and lower limits of air volume, fuel valve opening limit, and safe range of air-fuel ratio. Step S5 includes: S51. The combination of control parameters in the reference control strategy is taken as the desired target, and the current fuel supply flow rate data, primary air volume data and secondary air volume data are taken as the initial state and input into the simplified dynamic model. S52. Obtain the control quantity change sequence that optimizes the objective function under the process constraints, and use the instantaneous control quantity change value in the control quantity change sequence as the adjustment amount of the fuel quantity and air volume setpoint.

9. The combustion optimization control method for a semi-coke furnace based on machine vision according to claim 1, characterized in that, Step S6 includes: S61. Based on the analysis results of the adjustment amount of the fuel quantity and air volume setpoints, generate independent adjustment commands for the fuel supply valve, primary air damper and secondary air damper respectively. S62. Based on the sequential logical order, assign an execution start time and execution duration to each independent adjustment instruction; the sequential logical order is that primary air volume takes priority over fuel volume, and fuel volume takes priority over secondary air volume. S63. Divide the adjustment amount of each independent adjustment command into multiple adjustment steps, set an execution interval for each adjustment step, and form a smooth ramp control signal. S64. Package all independent adjustment instructions into the control instruction set.

10. A machine vision-based combustion optimization control system for a semi-coke furnace, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the machine vision-based combustion optimization control method for semi-coke furnaces as described in any one of claims 1 to 9.

Images