AI-based coal-fired flame intelligent monitoring and control method and system

Abstract

The application provides an AI-based intelligent monitoring and control method and system for coal combustion flame, and relates to the technical field of coal combustion flame monitoring, comprising the following steps: obtaining a multi-band image sequence of a combustion area, extracting spectral response and spatial gradient features, constructing a combustion state feature vector through wavelet transform, tensor fusion and dimension reduction processing; modeling the combustion time evolution relationship by using a self-attention mechanism and a causal convolution, identifying the combustion stability boundary by combining the kinetic trajectory and the Lyapunov index, and generating a combustion deviation state identifier by measuring the phase space distance; analyzing the dominant control variable according to the regulation response relationship data and the Jacobian matrix, and performing constraint optimization on the fuel supply and the combustion-supporting medium flow to generate executable regulation parameters and control the actuator to complete the combustion regulation.

Description

Technical Field

[0001] This invention relates to the field of coal combustion flame monitoring technology, and in particular to an AI-based intelligent monitoring and control method and system for coal combustion flames. Background Technology

[0002] Existing methods for monitoring and controlling coal-fired flames typically rely on visible light or infrared images acquired by a single sensor to extract basic features such as brightness, area, or flicker frequency. These features are then compared with preset stable combustion parameters to adjust the fuel-air ratio. This method depends on manually set empirical thresholds and employs PID control or simple fuzzy logic to achieve closed-loop regulation.

[0003] However, existing technologies using single-band images cannot fully reflect the multi-physics coupling information during combustion. The distribution of chemical reactions in the flame, temperature field gradients, and particulate radiation characteristics are easily masked in the visible light band, leading to a lag in the identification of early abnormal states such as combustion instability, localized flameout, or pollutant generation. Secondly, judgment methods based on fixed thresholds are difficult to adapt to dynamic operating conditions such as coal quality fluctuations and load changes, easily causing misjudgments or control lags, thereby reducing combustion efficiency and increasing emission risks. In addition, existing technologies lack effective modeling of the nonlinear coupling relationship between fuel supply and combustion-supporting media, making it impossible to identify the dominant control variables under complex operating conditions, resulting in insufficient response accuracy of control commands and difficulty in achieving rapid and stable combustion optimization. Summary of the Invention

[0004] This invention provides an AI-based intelligent monitoring and control method and system for coal-fired flames, which can at least solve some of the problems existing in the prior art.

[0005] A first aspect of this invention provides an AI-based intelligent monitoring and control method for coal-fired flames, comprising:

[0006] A multi-band image sequence of flames within the combustion area is acquired, and the spectral response curve and spatial gradient distribution features of the multi-band image sequence are extracted. Wavelet transform is performed on the spectral response curve to obtain multi-scale frequency domain features. Tensor fusion and dimensionality reduction are performed on the multi-scale frequency domain features and the spatial gradient distribution features to obtain a combustion state feature vector.

[0007] The combustion state feature vector is modeled using a self-attention mechanism to obtain a long-range correlation representation. The long-range correlation representation is then causally convolutionally encoded to obtain a temporal dependency sequence. Based on the temporal dependency sequence, a dynamic evolution trajectory is constructed and the Lyapunov exponent is calculated to obtain a stability boundary. The stability boundary is then compared with the pre-stored thermodynamic equilibrium state data using a phase space distance metric to obtain a deviation index. Based on the deviation index, combustion deviation modes are identified and a state identifier is generated.

[0008] According to the state identifier, query the Jacobian matrix corresponding to the pre-stored control response relationship data, solve the influence weights corresponding to the fuel supply rate and the combustion medium flow rate based on the Jacobian matrix, identify the dominant control variable based on the influence weights, and perform constraint optimization on the dominant control variable to obtain the executable control parameter set, convert the executable control parameter set into a control signal and send it to the actuator for execution.

[0009] In one alternative implementation,

[0010] Acquire a multi-band image sequence of the flame within the combustion area, extract the spectral response curve and spatial gradient distribution features of the multi-band image sequence, and perform wavelet transform on the spectral response curve to obtain multi-scale frequency domain features, including:

[0011] The image frame sequence of the combustion area is simultaneously acquired in different band ranges by a multispectral acquisition device. The image frame sequence is timestamped and arranged in band order to obtain a time-aligned image sequence. The radiation intensity value of each pixel position in the time-aligned image sequence in different bands is extracted to construct a spectral radiation intensity matrix and interpolation is performed along the band dimension to obtain the spectral response curve.

[0012] For each band image in the time-aligned image sequence, the horizontal and vertical gradients are calculated and vectors are synthesized to obtain the gradient magnitude distribution. The gradient magnitude distribution is then subjected to multi-scale Gaussian kernel convolution to obtain a smooth gradient field set. The difference response between the gradient fields at each scale in the smooth gradient field set is calculated and a gradient pyramid is constructed. The spatial gradient distribution features are obtained by performing spatial statistics on the gradient responses at each level in the gradient pyramid.

[0013] Initialize the scale parameters to construct a wavelet basis sequence and convolve it with the spectral response curve to obtain a set of scale wavelet coefficients. Arrange the set of scale wavelet coefficients in two dimensions according to the scale dimension and the frequency dimension to construct a wavelet transform matrix. Perform energy normalization processing on the wavelet transform matrix to obtain multi-scale frequency domain features.

[0014] In one alternative implementation,

[0015] The combustion state feature vector is obtained by tensor fusion and dimensionality reduction of the multi-scale frequency domain features and the spatial gradient distribution features, including:

[0016] The multi-scale frequency domain features are expanded into a frequency domain feature matrix according to the scale dimension and the frequency dimension. The spatial gradient distribution features are expanded into a gradient feature matrix according to the spatial position and gradient direction. The frequency domain feature rank and the gradient feature rank corresponding to the frequency domain feature matrix are calculated. The tensor decomposition target dimension is determined based on the frequency domain feature rank and the gradient feature rank. The frequency domain feature matrix and the gradient feature matrix are truncated based on the tensor decomposition target dimension to obtain a truncated frequency domain matrix and a truncated gradient matrix. The truncated frequency domain matrix and the truncated gradient matrix are stacked to construct a heterogeneous feature tensor.

[0017] The cross-modal correlation between the truncated frequency domain matrix and the truncated gradient matrix is ​​calculated to construct a modal coupling weight tensor. The heterogeneous feature tensor and the modal coupling weight tensor are subjected to Hadamard product operation and Tucker decomposition to obtain a fused feature tensor. The fused feature tensor is then subjected to dimensionality reduction projection and flattening to obtain a combustion state feature vector.

[0018] In one alternative implementation,

[0019] The combustion state feature vector is modeled using a self-attention mechanism to obtain a long-range correlation representation. This long-range correlation representation is then causally convolutionally encoded to obtain a temporal dependency sequence. Based on this temporal dependency sequence, a dynamic evolution trajectory is constructed, and the Lyapunov exponent is calculated to obtain the stability boundary, including:

[0020] The combustion state feature vectors are arranged into a state sequence matrix according to the time series. The state sequence matrix is ​​linearly projected to obtain a query matrix, a key matrix, and a value matrix. The key matrix is ​​transposed to obtain a transposed key matrix. The matrix product of the query matrix and the transposed key matrix is ​​calculated to obtain a similarity matrix. The similarity matrix is ​​softmax normalized to obtain an attention weight matrix. The long-range association representation is calculated based on the attention weight matrix and the value matrix.

[0021] Multiple sets of dilated causal convolution kernels with increasing dilation rates are set and multi-level dilated causal convolutions are performed on the long-range correlation representation in sequence. The output of each level is accumulated and flattened to obtain the temporal dependency sequence. The temporal dependency sequence is embedded with time delay to obtain the embedding state matrix. The displacement vectors of adjacent state points in the embedding state matrix are calculated and numerically integrated to obtain the dynamic evolution trajectory.

[0022] The trajectory segments are extracted from the dynamic evolution trajectory and the phase space distance between the starting point and the ending point is calculated. The phase space distance is fitted with logarithmic regression with respect to the time interval to obtain the Lyapunov exponent. Based on the Lyapunov exponent, the chaotic characteristics of the combustion state are obtained and the stability boundary is determined.

[0023] In one alternative implementation,

[0024] The deviation index is obtained by measuring the phase space distance between the stability boundary and the pre-stored thermodynamic equilibrium state data. Based on the deviation index, combustion deviation modes are identified and state identifiers are generated, including:

[0025] The equilibrium manifold is obtained by embedding the pre-stored thermodynamic equilibrium state data into a high-dimensional phase space. The Riemannian metric tensor of the equilibrium manifold is calculated and a geodesic coordinate system is constructed. The stability boundary is transformed and mapped to the geodesic coordinate system to obtain the boundary state trajectory. The normal distance field from the boundary state trajectory to the equilibrium manifold is calculated and integrated to obtain the Wasserstein distance. The deviation index is obtained by solving based on the Wasserstein distance and the pre-acquired Lyapunov exponent.

[0026] The deviation index is subjected to continuous homology decomposition to obtain a multi-level topological feature set. The Betti number sequence of each level in the multi-level topological feature set is calculated, and the topological phase transition mode is identified based on the jump point of the Betti number sequence. Multi-scale permutation entropy is extracted from the time-dependent sequence and an entropy rate time-frequency map is constructed. The entropy rate time-frequency map is subjected to multifractal spectrum analysis to obtain the Hausdorff dimension and multifractal spectrum width. The combustion dynamics complexity level is determined based on the Hausdorff dimension and multifractal spectrum width, and a high-dimensional decision hyperplane is constructed in combination with the topological phase transition mode. The deviation index is projected and classified in the high-dimensional decision hyperplane to identify combustion deviation modes and encode them to generate state identifiers.

[0027] In one alternative implementation,

[0028] The Jacobian matrix corresponding to the pre-stored control response relationship data is queried based on the state identifier. The influence weights corresponding to the fuel supply rate and combustion medium flow rate are then calculated based on the Jacobian matrix, including:

[0029] The state identifier is hashed to obtain the query index value, and the corresponding Jacobian matrix is ​​retrieved from the pre-stored control response relationship data. The Jacobian matrix is ​​decomposed into singular value diagonal matrix to obtain the singular value diagonal matrix and the corresponding condition number is calculated. When the condition number exceeds the preset condition number threshold, the maximum singular value in the singular value diagonal matrix is ​​extracted and the cutoff value is obtained by combining it with the preset regularization coefficient. The singular values ​​in the singular value diagonal matrix that are smaller than the cutoff value are replaced with the cutoff value and the replaced singular value diagonal matrix is ​​reconstructed to obtain the regularized Jacobian matrix.

[0030] The column vector corresponding to the fuel supply rate is extracted from the regularized Jacobian matrix as the fuel sensitivity vector, and its norm is calculated to obtain the fuel regulation intensity. The column vector corresponding to the combustion medium flow rate is extracted as the medium sensitivity vector, and its norm is calculated to obtain the medium regulation intensity. The temperature deviation component and the oxygen concentration deviation component are extracted from the deviation index. The inner product of the fuel sensitivity vector and the temperature deviation component is calculated to obtain the fuel temperature responsiveness. The inner product of the medium sensitivity vector and the oxygen concentration deviation component is calculated to obtain the medium oxygen concentration responsiveness. The fuel influence weight is calculated based on the fuel regulation intensity and the fuel temperature responsiveness. The medium influence weight is calculated based on the medium regulation intensity and the medium oxygen concentration responsiveness.

[0031] In one alternative implementation,

[0032] Based on the influence weights, the dominant control variables are identified, and constraint optimization is performed on the dominant control variables to obtain an executable control parameter set. The executable control parameter set is then converted into control signals and sent to the actuator for execution, including:

[0033] The comprehensive fuel evaluation index is obtained by multiplying the fuel influence weight and the fuel regulation intensity. The comprehensive medium evaluation index is obtained by multiplying the medium influence weight and the medium regulation intensity. The dominant control variable is determined by combining the comprehensive fuel evaluation index. The target deviation component corresponding to the dominant control variable is extracted from the deviation index. The column vector corresponding to the dominant control variable is extracted from the pre-acquired regularized Jacobian matrix and the ratio with the target deviation component is calculated to obtain the initial regulation increment.

[0034] Extract the upper and lower bounds of the control variable corresponding to the dominant control variable from the equipment operation constraint data, obtain the current operating value corresponding to the dominant control variable, take the target deviation component as the optimization target and the upper and lower bounds of the control variable as the constraint conditions, perform constrained quadratic programming based on the column vector to obtain the optimal control increment, sum the current operating value and the optimal control increment to obtain the dominant control parameter value, allocate the dominant control parameter value to the corresponding control parameter based on the control type of the dominant control variable to construct an executable control parameter group, convert the executable control parameter group into a control signal and send it to the corresponding actuator for execution.

[0035] A second aspect of this invention provides an AI-based intelligent monitoring and control system for coal-fired flames, comprising:

[0036] The feature extraction unit is used to acquire a multi-band image sequence of flames within the combustion area, extract the spectral response curve and spatial gradient distribution features of the multi-band image sequence, perform wavelet transform on the spectral response curve to obtain multi-scale frequency domain features, and perform tensor fusion and dimensionality reduction on the multi-scale frequency domain features and the spatial gradient distribution features to obtain a combustion state feature vector.

[0037] The state analysis unit is used to model the combustion state feature vector through a self-attention mechanism to obtain a long-range correlation representation, perform causal convolution encoding on the long-range correlation representation to obtain a time-dependent sequence, construct a dynamic evolution trajectory based on the time-dependent sequence and calculate the Lyapunov exponent to obtain a stability boundary, measure the phase space distance between the stability boundary and the pre-stored thermodynamic equilibrium state data to obtain a deviation index, identify combustion deviation modes based on the deviation index and generate a state identifier;

[0038] The control execution unit is used to query the Jacobian matrix corresponding to the pre-stored control response relationship data according to the state identifier, solve the influence weights corresponding to the fuel supply rate and the combustion medium flow rate based on the Jacobian matrix, identify the dominant control variable based on the influence weights and perform constraint optimization on the dominant control variable to obtain an executable control parameter set, convert the executable control parameter set into a control signal and send it to the actuator for execution.

[0039] A third aspect of the present invention provides an electronic device, comprising:

[0040] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0041] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0042] In this invention, by fusing spectral and spatial gradient features from multi-band image sequences, subtle changes in flame combustion can be accurately captured, avoiding the limitations of single-dimensional information. The combination of multi-scale frequency domain features extracted by wavelet transform and spatial gradient distribution effectively filters out noise interference, making the combustion state feature vector closer to the actual physical process and providing a reliable foundation for subsequent analysis. Utilizing self-attention mechanisms and causal convolutional coding, the long-range dependencies of flame dynamics can be automatically learned, constructing an accurate dynamic evolution trajectory. The calculation of the Lyapunov exponent directly quantifies the stability boundary of the combustion system. By measuring the phase space distance with thermodynamic equilibrium data, combustion deviation modes can be quickly identified and unique state identifiers generated, enabling early warning of abnormal states without manual intervention. The executable control parameter set obtained through constraint optimization reduces emission fluctuations while ensuring combustion efficiency, and the control signal is directly sent to the actuator to form a closed-loop regulation. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the AI-based intelligent monitoring and control method for coal-fired flames according to an embodiment of the present invention.

[0044] Figure 2 This is a flowchart illustrating the combustion state characteristic modeling and stability analysis of the AI-based intelligent monitoring and control method for coal-fired flames, as described in this embodiment of the invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0046] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0047] Figure 1 This is a flowchart illustrating the AI-based intelligent monitoring and control method for coal-fired flames according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0048] A multi-band image sequence of flames within the combustion area is acquired, and the spectral response curve and spatial gradient distribution features of the multi-band image sequence are extracted. Wavelet transform is performed on the spectral response curve to obtain multi-scale frequency domain features. Tensor fusion and dimensionality reduction are performed on the multi-scale frequency domain features and the spatial gradient distribution features to obtain a combustion state feature vector.

[0049] The combustion state feature vector is modeled using a self-attention mechanism to obtain a long-range correlation representation. The long-range correlation representation is then causally convolutionally encoded to obtain a temporal dependency sequence. Based on the temporal dependency sequence, a dynamic evolution trajectory is constructed and the Lyapunov exponent is calculated to obtain a stability boundary. The stability boundary is then compared with the pre-stored thermodynamic equilibrium state data using a phase space distance metric to obtain a deviation index. Based on the deviation index, combustion deviation modes are identified and a state identifier is generated.

[0050] According to the state identifier, query the Jacobian matrix corresponding to the pre-stored control response relationship data, solve the influence weights corresponding to the fuel supply rate and the combustion medium flow rate based on the Jacobian matrix, identify the dominant control variable based on the influence weights, and perform constraint optimization on the dominant control variable to obtain the executable control parameter set, convert the executable control parameter set into a control signal and send it to the actuator for execution.

[0051] In one alternative implementation,

[0052] Acquire a multi-band image sequence of the flame within the combustion area, extract the spectral response curve and spatial gradient distribution features of the multi-band image sequence, and perform wavelet transform on the spectral response curve to obtain multi-scale frequency domain features, including:

[0053] The image frame sequence of the combustion area is simultaneously acquired in different band ranges by a multispectral acquisition device. The image frame sequence is timestamped and arranged in band order to obtain a time-aligned image sequence. The radiation intensity value of each pixel position in the time-aligned image sequence in different bands is extracted to construct a spectral radiation intensity matrix and interpolation is performed along the band dimension to obtain the spectral response curve.

[0054] For each band image in the time-aligned image sequence, the horizontal and vertical gradients are calculated and vectors are synthesized to obtain the gradient magnitude distribution. The gradient magnitude distribution is then subjected to multi-scale Gaussian kernel convolution to obtain a smooth gradient field set. The difference response between the gradient fields at each scale in the smooth gradient field set is calculated and a gradient pyramid is constructed. The spatial gradient distribution features are obtained by performing spatial statistics on the gradient responses at each level in the gradient pyramid.

[0055] Initialize the scale parameters to construct a wavelet basis sequence and convolve it with the spectral response curve to obtain a set of scale wavelet coefficients. Arrange the set of scale wavelet coefficients in two dimensions according to the scale dimension and the frequency dimension to construct a wavelet transform matrix. Perform energy normalization processing on the wavelet transform matrix to obtain multi-scale frequency domain features.

[0056] Multispectral acquisition devices are typically equipped with multiple filter assemblies, each corresponding to a specific wavelength range, such as the visible light band (400-700nm), the near-infrared band (700-1100nm), and the mid-infrared band (1100-2500nm). The acquisition device simultaneously images the combustion area using a beam splitter or by rapidly switching filters, ensuring that image frames from different wavelengths maintain temporal consistency. The acquired raw image frame sequence needs to undergo timestamp alignment. This is done by reading the hardware timestamp of each frame, calculating the time deviation between images of different wavelengths, and when the deviation exceeds a preset threshold (e.g., 5 milliseconds), performing time correction on the image sequence using linear interpolation or nearest neighbor interpolation. The aligned images are arranged in ascending order of wavelength, forming a time-aligned image sequence. The data structure of this sequence can be represented as a three-dimensional tensor.

[0057] Spectral response curves are constructed by extracting the radiant intensity values ​​of each pixel location in different bands from a time-aligned image sequence. For spatial coordinates... For each pixel, the image is traversed across all bands, and the grayscale value or radiant intensity value at that location is read to form a one-dimensional vector. The length of the vector is equal to the number of bands, and each element in the vector corresponds to the radiant intensity of a band. The radiant intensity vectors of all pixel locations are arranged according to their spatial positions to construct a spectral dimensional radiant intensity matrix. Since the actual number of bands acquired is limited, interpolation fitting along the band dimension is required to obtain a continuous spectral response curve. Commonly used interpolation methods include cubic spline interpolation or piecewise linear interpolation. The interpolated spectral response curve can reflect the radiation characteristics of the flame within a continuous wavelength range, providing high-resolution input data for subsequent frequency domain analysis.

[0058] Spatial gradient distribution features are extracted to capture the sharpness of flame boundaries and the variation patterns of internal texture. For each band image in the temporally aligned image sequence, the horizontal and vertical gradients are calculated separately. The horizontal gradient is obtained by convolving the image using either the Sobel or Prewitt operator, while the vertical gradient is obtained by convolving the image vertically using the same operator. and vertical gradient Then, the gradient magnitude is calculated through vector synthesis, and the synthesis formula is as follows: The gradient magnitude distribution reflects the edge intensity at various locations in the image, but the original gradient is sensitive to noise, so multi-scale smoothing is required.

[0059] Gradient smoothing is achieved through multi-scale Gaussian kernel convolution. A set of Gaussian kernels with different standard deviations is selected, for example... , , Each of these is convolved with the gradient magnitude distribution. The two-dimensional form of the Gaussian kernel is... After convolution, smooth gradient fields at different scales are obtained. These smooth gradient fields are grouped into a set, with each element in the set corresponding to the gradient response at a specific scale. To further extract the difference information between scales, the difference response between gradient fields at adjacent scales is calculated. The difference response is defined as... ,in Scale indexing is used. The difference responses at each scale are stacked in ascending order of scale to construct a gradient pyramid structure. Each level of the gradient pyramid reflects the edge variation characteristics within a specific scale range.

[0060] Spatial statistics are performed on the gradient responses at each level of the gradient pyramid to extract statistical measures that characterize the spatial distribution of flames. Commonly used statistical measures include mean, variance, skewness, and kurtosis. The mean reflects the overall strength of the gradient response at that level, the variance reflects the dispersion of the gradient response, the skewness reflects the symmetry of the gradient distribution, and the kurtosis reflects the sharpness of the gradient distribution. For the ... Hierarchical gradient response Calculate the mean value over the entire image region. ,variance skewness and kurtosis All statistical measures at all levels are arranged in hierarchical order to form a spatial gradient distribution feature vector. This feature vector can comprehensively reflect the edge distribution pattern of the flame at different spatial scales, providing important spatial domain information for subsequent combustion state identification.

[0061] Wavelet transform is used to convert spectral response curves from the time or spatial domain to the frequency domain. Wavelet transform can simultaneously provide frequency and scale information, making it suitable for analyzing non-stationary signals. When initializing the scale parameters, a set of discrete scale values ​​is selected, for example... ,in Indicates the first Several scale parameters. A wavelet basis sequence is constructed based on the selected wavelet basis functions, such as the Morlet wavelet or the Daubechies wavelet. The expression for the Morlet wavelet is: ,in For the center frequency, These are time or space variables. For scale parameters... The wavelet basis function is scaled to .

[0062] The wavelet basis sequence and the spectral response curve are convolved to obtain the scaled wavelet coefficients. For the spectral response curve... ,in Representing wavelength, the convolution operation of wavelet transform is defined as follows: ,in This represents the conjugate of the wavelet basis functions. In the discrete case, the convolution operation is implemented using the Fast Fourier Transform (FFT). First, the spectral response curve and the wavelet basis functions are subjected to FFTs respectively, followed by dot multiplication in the frequency domain, and then the wavelet coefficients are obtained through the inverse Fourier transform. The above convolution operation is repeated for all scale parameters to obtain the set of scale wavelet coefficients.

[0063] The scale wavelet coefficient set contains wavelet coefficients at different scales and frequency positions. For ease of subsequent processing, these coefficients are arranged in a two-dimensional array according to scale and frequency dimensions to construct a wavelet transform matrix. The row indices of the wavelet transform matrix correspond to the scale parameter, and the column indices correspond to the frequency or wavelength position. Each element in the matrix represents the amplitude of a wavelet coefficient at a specific scale and frequency position. Since the amplitude range of wavelet coefficients at different scales may vary significantly, energy normalization is required. Energy normalization is achieved by calculating the sum of the energies of all wavelet coefficients at each scale. Then divide all wavelet coefficients at that scale by The normalized wavelet coefficients are obtained. The normalized wavelet transform matrix is ​​the multi-scale frequency domain feature, which can reflect the energy distribution law of the spectral response curve at different frequency components and different scales, providing a frequency domain feature representation for subsequent tensor fusion and combustion state identification.

[0064] In practical applications, the acquisition frequency of multi-band image sequences is typically set to 10-30 frames per second to ensure the capture of dynamic changes in the flame. The choice of the number of bands requires a trade-off between spectral resolution and computational complexity; a common configuration is 5-10 bands. The number of scale parameters for wavelet transform is typically set to 8-16, covering the full frequency range from high-frequency details to low-frequency trends. The number of levels in the gradient pyramid is generally set to 3-5 to capture edge information at different spatial scales. Through these steps, rich spectral, spatial, and frequency domain features can be extracted from the original multi-band image sequences, providing a comprehensive data foundation for subsequent combustion state modeling and control decisions.

[0065] In one alternative implementation,

[0066] The combustion state feature vector is obtained by tensor fusion and dimensionality reduction of the multi-scale frequency domain features and the spatial gradient distribution features, including:

[0067] The multi-scale frequency domain features are expanded into a frequency domain feature matrix according to the scale dimension and the frequency dimension. The spatial gradient distribution features are expanded into a gradient feature matrix according to the spatial position and gradient direction. The frequency domain feature rank and the gradient feature rank corresponding to the frequency domain feature matrix are calculated. The tensor decomposition target dimension is determined based on the frequency domain feature rank and the gradient feature rank. The frequency domain feature matrix and the gradient feature matrix are truncated based on the tensor decomposition target dimension to obtain a truncated frequency domain matrix and a truncated gradient matrix. The truncated frequency domain matrix and the truncated gradient matrix are stacked to construct a heterogeneous feature tensor.

[0068] The cross-modal correlation between the truncated frequency domain matrix and the truncated gradient matrix is ​​calculated to construct a modal coupling weight tensor. The heterogeneous feature tensor and the modal coupling weight tensor are subjected to Hadamard product operation and Tucker decomposition to obtain a fused feature tensor. The fused feature tensor is then subjected to dimensionality reduction projection and flattening to obtain a combustion state feature vector.

[0069] After obtaining multi-scale frequency domain features and spatial gradient distribution features, these two heterogeneous features need to be effectively fused to construct a unified combustion state characterization. Multi-scale frequency domain features reflect the energy distribution patterns of the flame spectrum at different frequency components, while spatial gradient distribution features characterize the texture variation patterns of the flame in the spatial domain. Since these two types of features originate from frequency domain analysis and spatial domain analysis respectively, their data structures and physical meanings differ significantly. Directly concatenating them would lead to feature dimensionality redundancy and fail to capture the intrinsic correlations across modes. Therefore, a tensor fusion method is adopted, which preserves the multi-dimensional correlations between features by constructing a high-order tensor structure and extracts a compact combustion state feature vector through decomposition and dimensionality reduction techniques.

[0070] The multi-scale frequency domain features are expanded into a frequency domain feature matrix according to the scale dimension and the frequency dimension. Assuming that after wavelet transform, the following is obtained... Frequency domain features at each scale level, each scale level containing If there are 1 frequency component, then the frequency domain characteristic matrix is... The dimension is The first in the matrix Line number The elements of the column represent the first... The scale in the first The response intensity at each frequency component is used to expand the spatial gradient distribution characteristics into a gradient feature matrix according to spatial location and gradient direction. In practical processing, the combustion zone is divided into... For each spatial grid cell, the gradient statistics in both the horizontal and vertical directions are calculated. Then the gradient characteristic matrix... The dimension is , of which The first column of the row stores the first... The first column stores the horizontal gradient statistics of each spatial unit, and the second column stores the vertical gradient statistics.

[0071] To determine a reasonable target dimension for tensor decomposition, it is necessary to calculate the effective rank of the frequency domain feature matrix and the gradient feature matrix, respectively. For the frequency domain feature matrix... Perform singular value decomposition to obtain a sequence of singular values. Sort in descending order. Calculate the cumulative energy ratio. , select to make minimum rank As the rank of the frequency domain features. Similarly, for the gradient feature matrix... Perform singular value decomposition to obtain a sequence of singular values. Calculate the cumulative energy ratio , select to make minimum rank The gradient feature rank. The target dimension of tensor decomposition is set to... This ensures that the fused tensor retains the main components of the feature with greater information content from the two types of features.

[0072] Based on the defined tensor decomposition target dimension, the frequency domain feature matrix and gradient feature matrix are truncated. The frequency domain feature matrix... Singular value decomposition yields , keep before Construct a truncated frequency domain matrix using the singular values ​​and their corresponding left and right singular vectors. Its dimensions are But the rank was reduced to For the gradient feature matrix Performing the same operation yields the truncated gradient matrix. , dimension And the rank is The truncation operation effectively removes noise components and redundant information from the original feature matrix, providing a more compact input for subsequent tensor fusion.

[0073] A heterogeneous feature tensor is constructed by stacking the truncated frequency domain matrix and the truncated gradient matrix. Since the two matrices have different dimensions, dimension alignment is required first. The truncated frequency domain matrix... Extending along the third dimension, we obtain a third-order tensor. The gradient matrix will be truncated. Extending along the third dimension, we obtain a third-order tensor. To achieve stacking, the two tensors are zero-padded to unify their first two dimensions. The unified dimension is set to... and ,right and Zero-padding is performed separately to obtain the dimension as follows: The tensor. Stack two padded tensors along the third dimension to construct a heterogeneous feature tensor. The first slice corresponds to the frequency domain feature, and the second slice corresponds to the gradient feature.

[0074] To capture the cross-modal correlation between frequency domain features and gradient features, the canonical correlation coefficient between the truncated frequency domain matrix and the truncated gradient matrix is ​​calculated. The two matrices are then centered to obtain... and Calculate the covariance matrix. and and cross-covariance matrix By solving the generalized eigenvalue problem The canonical correlation coefficient was obtained. Construct modally coupled weight tensors The element of the first slice is set to The elements of the second slice are set to If the number of canonical correlation coefficients is insufficient, the minimum correlation coefficient is used to fill the gap. This weight tensor reflects the correlation strength between the two types of features at different spatial locations and frequency components.

[0075] For heterogeneous feature tensors Modal coupling weight tensor Perform the Hadamard product operation to obtain the weighted heterogeneous feature tensor. ,in This represents element-wise multiplication. This operation gives higher weights to highly correlated feature components and suppresses weakly correlated components. The weighted heterogeneous feature tensor is decomposed using Tucker decomposition, resulting in a product of the core tensor and three factor matrices. ,in For the core tensor, , , For factor matrix, Indicates along the first Tensor products of modalities. Target dimension of Tucker decomposition. , , By minimizing the reconstruction error Determine, while constraining , , A high-order orthogonal iterative algorithm is used to solve the Tucker decomposition, iteratively updating the factor matrix until convergence. The core tensor obtained from the decomposition... It contains a compact representation that combines frequency domain features and gradient features, and its dimension is much smaller than the original heterogeneous feature tensor.

[0076] For fusion feature tensor Dimensional reduction projection and flattening are performed to obtain the combustion state feature vector, and the core tensor is then... Expanding along the three modes into a matrix form, the matrix obtained by expanding along mode 1 is... Principal component analysis is performed on the expanded matrix to calculate the covariance matrix. Solving eigenvalue problems , to obtain eigenvalues and the corresponding feature vectors. Before selection The eigenvectors corresponding to the principal components constitute the projection matrix. ,in The selection of [a specific factor] makes the cumulative variance contribution rate [a certain percentage]. Projecting the expanded matrix onto a lower-dimensional space yields... , dimension Flattening the projected matrix row by row yields a matrix of length [length missing]. The one-dimensional vector is the combustion state feature vector. This feature vector integrates information from multi-scale frequency domain features and spatial gradient distribution features, significantly reducing dimensionality and facilitating subsequent time series modeling and control decisions.

[0077] In practical applications, the parameters involved in tensor fusion need to be adjusted according to the specific combustion monitoring scenario. For large boilers, the spatial resolution of the combustion zone is high, and the number of spatial grid cells is limited. The number of features can reach hundreds, at which point the dimensionality of the gradient feature matrix is ​​large, requiring a higher truncation rank to retain sufficient spatial detail. For small combustion devices, the scale hierarchy of frequency domain features... The rank of the frequency domain feature matrix is ​​likely lower, resulting in a corresponding reduction in the target dimension of the tensor decomposition. The core tensor dimension of Tucker decomposition is... , , The choice of dimensionality requires a trade-off between reconstruction accuracy and computational efficiency. Too small a dimensionality leads to information loss, while too large a dimensionality increases the computational burden and may introduce the risk of overfitting. By performing cross-validation on historical combustion data, the optimal combination of decomposition dimensions can be determined, minimizing reconstruction error while maintaining the discriminative power of the feature vectors.

[0078] Figure 2 This is a flowchart illustrating the combustion state characteristic modeling and stability analysis of the AI-based intelligent monitoring and control method for coal-fired flames, as described in this embodiment of the invention.

[0079] In one alternative implementation,

[0080] The combustion state feature vector is modeled using a self-attention mechanism to obtain a long-range correlation representation. This long-range correlation representation is then causally convolutionally encoded to obtain a temporal dependency sequence. Based on this temporal dependency sequence, a dynamic evolution trajectory is constructed, and the Lyapunov exponent is calculated to obtain the stability boundary, including:

[0081] The combustion state feature vectors are arranged into a state sequence matrix according to the time series. The state sequence matrix is ​​linearly projected to obtain a query matrix, a key matrix, and a value matrix. The key matrix is ​​transposed to obtain a transposed key matrix. The matrix product of the query matrix and the transposed key matrix is ​​calculated to obtain a similarity matrix. The similarity matrix is ​​softmax normalized to obtain an attention weight matrix. The long-range association representation is calculated based on the attention weight matrix and the value matrix.

[0082] Multiple sets of dilated causal convolution kernels with increasing dilation rates are set and multi-level dilated causal convolutions are performed on the long-range correlation representation in sequence. The output of each level is accumulated and flattened to obtain the temporal dependency sequence. The temporal dependency sequence is embedded with time delay to obtain the embedding state matrix. The displacement vectors of adjacent state points in the embedding state matrix are calculated and numerically integrated to obtain the dynamic evolution trajectory.

[0083] The trajectory segments are extracted from the dynamic evolution trajectory and the phase space distance between the starting point and the ending point is calculated. The phase space distance is fitted with logarithmic regression with respect to the time interval to obtain the Lyapunov exponent. Based on the Lyapunov exponent, the chaotic characteristics of the combustion state are obtained and the stability boundary is determined.

[0084] After obtaining the combustion state feature vectors, the feature vectors from multiple consecutively collected moments are arranged in chronological order to form a state sequence matrix. ,in Indicates the total number of time steps. The dimension is the feature vector. The query matrix is ​​obtained by performing a cubic linear projection transformation on the state sequence matrix. Key matrix Sum matrix ,in The projected weight matrix is ​​a learnable matrix. This represents the hidden dimension after projection. The transpose of the key matrix yields the transposed key matrix. The similarity matrix is ​​obtained by calculating the matrix product of the query matrix and the transpose key matrix. , of which This is to prevent the gradient from vanishing due to an excessively large dot product. A softmax normalization operation is performed on each row of the similarity matrix to obtain the attention weight matrix. , No. Line number Column elements are This weight matrix reflects the correlation strength between combustion states at different times. A long-range correlation characterization is calculated based on the attention weight matrix and the value matrix. This characterization captures the time-step dependencies during combustion, enabling the identification of periodic patterns of flame fluctuations and propagation paths of abrupt events.

[0085] To extract temporal dependency features from long-range association representations, multiple sets of dilated causal convolution kernels with increasing dilation rates are set. Specifically, the following settings are implemented: Layer dilated causal convolutional layer, the first The expansion rate of the layer is The kernel size is . No. The output of the layer dilated causal convolution is ,in For the first Layer Each convolution weight, Use this as the initial input. Causal convolution ensures that the first... The output at any given time depends only on the input at the current and historical times, without introducing future information, thus meeting the causality requirement of real-time monitoring. The exponential growth of the expansion rate allows the network to cover a longer historical window with fewer layers. The receptive field length of the layer is Residual connections and nonlinear activations are applied to the output of each layer, and the fused features are obtained by summing the outputs of all layers. ,in For the first Layer fusion weights. The fusion features are flattened along the time dimension to obtain the temporal dependency sequence. This sequence encodes the evolution of combustion states across multiple time scales.

[0086] Time-delayed embedding is performed on time-dependent sequences to construct an embedding state matrix for reconstructing the phase space trajectory of the combustion system. Embedding dimensions are selected. and delay time The first time-dependent sequence Each time step is mapped to an embedding vector. ,in Indicates the sequence at the 1st The scalar value at time. Embedded dimension. The delay time is usually determined using the pseudo-nearest neighbor method. The embedding state is determined by the first zero crossing of the autocorrelation function or the minimum of the mutual information. The embedding vectors at all valid time points are arranged row-wise to obtain the embedding state matrix. ,in This represents the number of effective embedding points. Calculate the displacement vectors of adjacent state points in the embedding state matrix. The displacement vector reflects the instantaneous evolution direction of the combustion state in phase space. Numerical integration of the displacement vector, using a fourth-order Runge-Kutta method for stepwise accumulation, yields the dynamic evolution trajectory. This trajectory depicts the motion of the combustion system in a high-dimensional phase space, revealing the attractor structure and chaotic characteristics of the system.

[0087] Multiple trajectory segments are extracted from the dynamic evolution trajectory to calculate the Lyapunov exponent, thereby quantifying the system's stability. The length of each trajectory segment is selected. Randomly selected from the trajectory There are several starting points, and for each starting point... Extraction length is trajectory segment Calculate the starting point. and the end point phase space distance ,in This represents the Euclidean norm. For different time intervals... Calculate the corresponding average phase space distance A logarithmic regression was performed on the average phase spatial distance with respect to the time interval to establish a linear relationship. ,in The maximum Lyapunov index, Let be the intercept constant. The Lyapunov exponent is obtained by solving the regression coefficients using the least squares method. The estimated value.

[0088] The chaotic characteristics of the combustion state are determined based on the Lyapunov exponent, and the stability boundary is established. At this point, the system is in a chaotic state, with adjacent trajectories diverging exponentially. The combustion process exhibits irregular fluctuations and unpredictable abrupt changes. The system is extremely sensitive to initial conditions and disturbances, requiring timely intervention to prevent instability. At this point, the system is in a critical state, and the trajectory neither converges nor diverges; the combustion process may exhibit quasi-periodic oscillations or boundary layer separation. At this point, the system is in a stable state, adjacent trajectories converge to the attractor, and the combustion process is smooth and controllable. The stability boundary is defined as the threshold of the Lyapunov exponent. , usually take ,when The system is then determined to have exceeded the stability boundary. Further calculations are performed on the Lyapunov spectrum, i.e., all... Lyapunov index The dynamic type of the system is identified by analyzing the distribution characteristics of the spectrum. The Kaplan-Yorke dimension is calculated. ,in To meet and The largest integer of the dimension quantifies the fractal dimension of the attractor, reflecting the complexity of the combustion system. The Lyapunov exponent, stability boundary, and Kaplan-Yorke dimension are used as dynamic characteristics of the combustion state for subsequent deviation calculations and control decisions.

[0089] In practical applications, the multi-head extension of the self-attention mechanism can simultaneously capture multiple correlation patterns during the combustion process. Each attention head independently calculates attention weights and generates sub-representations. Finally, all sub-representations are concatenated and fused through a linear layer to obtain a multi-head long-range association representation. ,in The output projection matrix is ​​used. The residual connections of the dilated causal convolution employ a gating mechanism, using a sigmoid activation function to control information flow and avoid the gradient vanishing problem. The parameter selection for the time-delay embedding is crucial to the trajectory reconstruction quality; too small an embedding dimension leads to trajectory folding, while too large an embedding dimension introduces redundant noise. Too short a delay time results in highly correlated adjacent points, while too long a delay time leads to the loss of dynamic information. The calculation of the Lyapunov exponent requires a sufficiently long trajectory segment to ensure statistical stability, while also eliminating the influence of the initial transient process; typically, the first 20% of trajectory points are discarded. The stability boundary threshold can be adaptively adjusted according to the specific operating characteristics of the coal-fired boiler. The distribution of the Lyapunov exponent during normal operation is statistically analyzed using historical data, and the 95th percentile is used as the dynamic threshold.

[0090] In one alternative implementation,

[0091] The deviation index is obtained by measuring the phase space distance between the stability boundary and the pre-stored thermodynamic equilibrium state data. Based on the deviation index, combustion deviation modes are identified and state identifiers are generated, including:

[0092] The equilibrium manifold is obtained by embedding the pre-stored thermodynamic equilibrium state data into a high-dimensional phase space. The Riemannian metric tensor of the equilibrium manifold is calculated and a geodesic coordinate system is constructed. The stability boundary is transformed and mapped to the geodesic coordinate system to obtain the boundary state trajectory. The normal distance field from the boundary state trajectory to the equilibrium manifold is calculated and integrated to obtain the Wasserstein distance. The deviation index is obtained by solving based on the Wasserstein distance and the pre-acquired Lyapunov exponent.

[0093] The deviation index is subjected to continuous homology decomposition to obtain a multi-level topological feature set. The Betti number sequence of each level in the multi-level topological feature set is calculated, and the topological phase transition mode is identified based on the jump point of the Betti number sequence. Multi-scale permutation entropy is extracted from the time-dependent sequence and an entropy rate time-frequency map is constructed. The entropy rate time-frequency map is subjected to multifractal spectrum analysis to obtain the Hausdorff dimension and multifractal spectrum width. The combustion dynamics complexity level is determined based on the Hausdorff dimension and multifractal spectrum width, and a high-dimensional decision hyperplane is constructed in combination with the topological phase transition mode. The deviation index is projected and classified in the high-dimensional decision hyperplane to identify combustion deviation modes and encode them to generate state identifiers.

[0094] After acquiring pre-stored thermodynamic equilibrium data, it is represented as a high-dimensional set of state vectors, each containing thermodynamic parameters such as temperature, pressure, oxygen concentration, and fuel concentration. Manifold learning is then performed on this dataset, using a locally linear embedding method to map discrete state points to a continuous phase space. Specifically, for each state point in the dataset, its nearest neighbors in Euclidean space are searched, and local linear weights are determined by minimizing the reconstruction error, thus preserving the local geometry in the low-dimensional embedding space. After iterative optimization, an embedded representation of the equilibrium manifold is obtained, which appears as a smooth surface or hypersurface structure in the phase space.

[0095] A Riemannian geometric framework is established on the equilibrium manifold. For any point on the manifold, the basis vectors of the tangent space at that point are calculated, and the tangent vector field is obtained by taking the partial derivatives with respect to the embedded coordinates. The Riemannian metric tensor is defined as the inner product on the tangent space, and its components are obtained by calculating the dot product of the tangent vectors. Let the local coordinates of the manifold be... ,in If the dimension of the manifold is given, then the metric tensor is... The amount satisfy ,in For the position vector in the embedding space, and This is a coordinate index. Using this metric tensor, the geodesic distance between any two points on a manifold can be calculated. The geodesic equation is derived through the variational principle and is of the form: ,in For arc length parameters, Let Christofel notation be used, and let the metric tensor and its derivative be expressed as follows: By numerically solving the geodesic equations, a geodesic coordinate system is established, such that the metric tensor in this coordinate system is locally approximated as an identity matrix.

[0096] The stability boundary data undergoes coordinate transformation. The stability boundary is determined by the Lyapunov exponent threshold calculated in the preceding steps, corresponding to the set of state points in the dynamic evolution trajectory that satisfy specific stability conditions. The original eigenvectors of these state points are projected into the phase space through the same embedding mapping as the equilibrium manifold, yielding the boundary state trajectory. This trajectory is represented in phase space as one or more curves, the geometric shape of which reflects the dynamic characteristics of the system deviating from equilibrium. For each point on the boundary state trajectory, its coordinate representation in the geodesic coordinate system is calculated, and the mapping from the original eigenspace to the geodesic coordinate system is achieved through a coordinate transformation matrix.

[0097] Calculate the normal distance field from the boundary state trajectory to the equilibrium manifold. For any point on the boundary state trajectory... Find the point closest to it on the equilibrium manifold. The nearest point is determined by minimizing the geodesic distance. The normal distance is defined as... and The projection of the geodesic distance between them onto the direction of the manifold normal vector. The manifold in... The normal vector at a given point is obtained through the orthogonal complement of the tangent space. The entire boundary state trajectory is discretized and sampled, and the normal distance is calculated at each sampling point to form a normal distance field. Integrating this distance field along the trajectory yields the overall deviation metric. To more accurately characterize the distributional differences, the Wasserstein distance is used. Treating the boundary state trajectory and the equilibrium manifold as probability distributions, the Wasserstein distance between them is calculated by solving the optimal transport problem. The distance is defined as ,in Let be the set of all joint distributions. The transmission plan is given. The optimization problem is solved using linear programming or entropy regularization to obtain the Wasserstein distance.

[0098] A deviation index is constructed by combining the Wassstein distance and the Lyapunov exponent. The Lyapunov exponent reflects the stability of the system dynamics, while the Wassstein distance reflects the degree of deviation from the state distribution; together, they determine the deviation characteristics of the combustion system. Deviation Index Defined as ,in and The weighting coefficient is determined based on historical data statistics. This index integrates information from both geometric deviation and kinetic instability, and can comprehensively characterize the degree of abnormality in combustion state.

[0099] Continuous homology analysis is performed on the deviation index. The time series of the deviation index is treated as a one-dimensional signal, and a point cloud dataset is constructed using a sliding window. Continuous homology calculation is performed on the point cloud to track the generation and disappearance of topological features at different scales. Continuous homology decomposition generates multi-level topological feature sets, with each level corresponding to a scale parameter. At each scale, the sequence of Betti numbers is calculated, where the zeroth-order Betti number is... The first-order Betti number represents the number of connected components. The number of rings is represented by the Betti number, and higher-order Betti numbers represent higher-dimensional topological holes. The curve of the Betti number sequence changing with the scale parameter is called a persistence barcode or persistence map. Topological phase transition modes are identified by detecting jump points in the Betti number sequence. Jump points correspond to abrupt changes in the topological structure, reflecting the bifurcation or transition of the combustion state in phase space. The position and magnitude of the jump points are recorded to construct a topological phase transition feature vector.

[0100] Extracting multi-scale permutation entropy from time-dependent sequences. Permutation entropy is a complexity measure based on symbolic dynamics, quantifying irregularity by analyzing the permutation patterns of time series. Phase space reconstruction is performed on the time-dependent sequences, and the embedding dimension is selected. and delay time The sequence is divided into multiple embedding vectors. The elements in each embedding vector are sorted to obtain permutation patterns. The frequency of occurrence of all possible permutation patterns is counted, and the permutation entropy is calculated. ,in For the first The probability of various permutation patterns is calculated. By changing the embedding dimension and delay time, a multi-scale permutation entropy sequence is obtained. The permutation entropy sequence is then expanded in two dimensions along time and scale to construct an entropy rate time-frequency plot. This time-frequency plot reflects the evolution trend of entropy on the time axis and the distribution of complexity at different time scales on the scale axis.

[0101] Multifractal spectrum analysis was performed on the entropy rate time-frequency plot. Multifractal spectrum describes the differences in the scaling behavior of a signal in different local regions. A multifractal detrending fluctuation analysis method was employed, dividing the time-frequency plot into multiple non-overlapping time windows. Within each window, a polynomial trend was fitted and removed, and the detrended fluctuation function was calculated. Different orders of... Scaling analysis of the fluctuation function yields the generalized Hearst exponent. Through the Legendre transformation, Converted to multifractal spectrum ,in The singularity index. The width of the multifractal spectrum. This reflects the multifractal intensity of the signal. Hausdorff dimension. Using the box-counting method, the time-frequency plot was overlaid on grids of different scales. The scaling relationship between the number of non-empty grids and the grid size was statistically analyzed, and the result was fitted. .

[0102] Combustion dynamics complexity levels are determined based on Hausdorff dimension and multifractal spectrum width. A complexity level threshold is then set. and The combination of these factors maps to different levels. Low complexity corresponds to regular combustion states, while high complexity corresponds to turbulent or unstable combustion states. A high-dimensional decision hyperplane is constructed by combining topological phase transition modes. The decision hyperplane separates different combustion deviation modes in the feature space. The hyperplane parameters are determined by maximizing the classification margin using a support vector machine method. The hyperplane equation is as follows: ,in It is the normal vector. This is a comprehensive feature vector, including deviation index, topological features, complexity index, etc. This is a bias term.

[0103] The deviation index and its associated features are projected and classified onto a high-dimensional decision hyperplane. The signed distance from the feature vector to each hyperplane is calculated, and the category is determined based on the sign and magnitude of the distance. Combustion deviation modes include various types such as normal combustion, insufficient fuel, excess oxygen, partial flameout, and slagging tendency. Each mode corresponds to a unique state identifier, generated using binary encoding or hash encoding. The state identifier not only identifies the deviation mode type but also contains information on the degree of deviation and the level of urgency. The classification results are mapped to state identifiers using a lookup table or decision tree structure and output to the subsequent control module.

[0104] In one alternative implementation,

[0105] The Jacobian matrix corresponding to the pre-stored control response relationship data is queried based on the state identifier. The influence weights corresponding to the fuel supply rate and combustion medium flow rate are then calculated based on the Jacobian matrix, including:

[0106] The state identifier is hashed to obtain the query index value, and the corresponding Jacobian matrix is ​​retrieved from the pre-stored control response relationship data. The Jacobian matrix is ​​decomposed into singular value diagonal matrix to obtain the singular value diagonal matrix and the corresponding condition number is calculated. When the condition number exceeds the preset condition number threshold, the maximum singular value in the singular value diagonal matrix is ​​extracted and the cutoff value is obtained by combining it with the preset regularization coefficient. The singular values ​​in the singular value diagonal matrix that are smaller than the cutoff value are replaced with the cutoff value and the replaced singular value diagonal matrix is ​​reconstructed to obtain the regularized Jacobian matrix.

[0107] The column vector corresponding to the fuel supply rate is extracted from the regularized Jacobian matrix as the fuel sensitivity vector, and its norm is calculated to obtain the fuel regulation intensity. The column vector corresponding to the combustion medium flow rate is extracted as the medium sensitivity vector, and its norm is calculated to obtain the medium regulation intensity. The temperature deviation component and the oxygen concentration deviation component are extracted from the deviation index. The inner product of the fuel sensitivity vector and the temperature deviation component is calculated to obtain the fuel temperature responsiveness. The inner product of the medium sensitivity vector and the oxygen concentration deviation component is calculated to obtain the medium oxygen concentration responsiveness. The fuel influence weight is calculated based on the fuel regulation intensity and the fuel temperature responsiveness. The medium influence weight is calculated based on the medium regulation intensity and the medium oxygen concentration responsiveness.

[0108] After obtaining the state identifier, the corresponding Jacobian matrix needs to be retrieved from a pre-built regulatory response relation database. The state identifier is typically in string or symbolic form; to achieve fast retrieval, a hash mapping mechanism is used to convert it into a numerical query index value. Specifically, the SHA-256 hash algorithm is used to calculate the digest of the state identifier. The first 64 bits of the generated 256-bit hash value are truncated and converted to a decimal integer. Then, the modulo operation is performed with the total number of Jacobian matrix entries in the database to obtain the query index value. This index value directly points to the storage location in the regulatory response relational database, from which the corresponding Jacobian matrix is ​​extracted. The rows of this matrix correspond to system output variables, including thermodynamic and chemical state quantities such as combustion temperature, oxygen concentration, carbon monoxide concentration, and nitrogen oxide concentration; the columns correspond to control input variables, including adjustable parameters such as fuel supply rate, primary air flow rate, secondary air flow rate, and air distribution angle. Matrix elements Indicates the first The control input pair for the first The local sensitivity of an output variable, that is, the partial derivative value near the current operating point.

[0109] Due to the strong nonlinear characteristics of the combustion system and the presence of measurement noise, the Jacobian matrix... Often exhibiting ill-conditioned characteristics, i.e., the condition number of the matrix is ​​large, leading to a significant amplification of small perturbations during numerical solution. To assess the degree of ill-conditioning of the matrix, [the following is a process / method] is used. Perform singular value decomposition to obtain ,in and These are the left singular vector matrix and the right singular vector matrix, respectively. For a singular value diagonal matrix, its diagonal elements Sort in descending order To control the total number of input variables, the condition number of a matrix is ​​defined as follows: ,when As the condition number approaches zero, it approaches infinity, indicating that the matrix is ​​approaching singularity. A condition number threshold is set. for ,when In this case, the decision matrix needs to be regularized.

[0110] Regularization is performed using a truncated singular value decomposition method. The maximum singular value is extracted. Combined with preset regularization coefficients Calculate the cutoff value Regularization coefficient The typical value range is: to Adjustments are made based on the system noise level and control accuracy requirements. Traverse the singular value diagonal matrix. All diagonal elements in the set, for satisfying Replace the singular values ​​with truncated values. The regularized singular value diagonal matrix is ​​obtained. The Jacobian matrix is ​​reconstructed using the corrected singular value matrix to obtain the regularized Jacobian matrix. The regularization process is equivalent to applying Tikhonov regularization to the original matrix, which suppresses high-frequency noise components corresponding to small singular values ​​and improves the stability of subsequent numerical calculations.

[0111] In the regularized Jacobian matrix In the middle, the first The column corresponds to the control variable of fuel supply rate, and the column vector is extracted and denoted as the fuel sensitivity vector. ,in This represents the total number of system output variables. The th vector... Each component Indicates the effect of changes in fuel supply rate on the first The degree of influence of each output variable is calculated. The Euclidean norm of the fuel sensitivity vector is then used to obtain the fuel regulation intensity. This indicator comprehensively reflects the overall impact of fuel supply rate adjustments on the overall system state. Similarly, the first value is extracted from the regularized Jacobian matrix. The column vector corresponding to the flow rate of the combustion-supporting medium is denoted as the medium sensitivity vector. The norm is calculated to obtain the medium modulation intensity. .

[0112] To further quantify the targeted response of fuel supply rate and combustion medium flow rate to the current deviation, it is necessary to analyze the specific components of the deviation index. Deviation Index It is a multidimensional vector, whose components correspond to different degrees of deviation of thermodynamic and chemical state quantities from the equilibrium state. The temperature deviation component is extracted from it. and oxygen concentration deviation component The temperature deviation component reflects the deviation of the combustion zone temperature field from the target temperature distribution, while the oxygen concentration deviation component reflects the deviation of the oxygen concentration field from the ideal air distribution. The fuel sensitivity vector... Temperature deviation component Perform inner product calculation to obtain the fuel temperature response. The inner product value is a scalar, and its sign and magnitude reflect the direction and magnitude of the correction to temperature deviation caused by increasing the fuel supply rate. If This indicates that increasing fuel supply helps reduce temperature deviation; if This indicates that fuel supply should be reduced. Similarly, the medium sensitivity vector is calculated. Deviation component from oxygen concentration The inner product is used to obtain the oxygen concentration response of the medium. .

[0113] Based on the aforementioned calculations, the fuel regulation intensity is obtained. and fuel temperature response We construct the fuel impact weights. These weights need to consider both the overall impact of the control variables and their targeted response to current deviations. The fuel impact weights are calculated using a product form. Taking the absolute value is to focus only on the response intensity and not the direction, since the control direction will be determined by the sign in the subsequent optimization solution. To avoid numerical overflow and facilitate normalization, a logarithmic transformation can be performed on the influence weights to obtain... Similarly, based on the intensity of medium regulation. and medium oxygen concentration response Calculate the influence weight of the medium .

[0114] In practical applications, to ensure the comparability of influence weights, it is necessary to normalize the influence weights of fuel and medium. The total weights are then calculated. The normalized fuel influence weights were obtained. And normalized medium influence weight ,satisfy The normalized impact weights directly reflect the relative importance of the two control variables in the current deviation state. If... This indicates that the fuel supply rate is the dominant control variable and should be adjusted first; if If the flow rate of the combustion-supporting medium is close to 0.5, it indicates that the two control variables need to be adjusted in a coordinated manner.

[0115] To improve the robustness of the influence weight calculation, a multi-objective response evaluation mechanism can be introduced. In addition to the temperature deviation component and the oxygen concentration deviation component, the carbon monoxide concentration deviation component can also be extracted from the deviation index. and nitrogen oxide concentration deviation component The fuel's responsiveness to carbon monoxide is obtained by calculating the inner product of the fuel sensitivity vector and these deviation components. and fuel responsiveness to nitrogen oxides A weighted summation method was used to construct the comprehensive fuel response. ,in , , The weighting coefficients for each deviation component are set according to environmental protection requirements and economic objectives. Similarly, a comprehensive media responsiveness is constructed. Recalculating the influence weights based on the comprehensive responsiveness can more comprehensively reflect the control variables' ability to regulate the system's multi-objective state.

[0116] Under certain conditions, the Jacobian matrix may contain multiple near-singular singular values, leading to numerical instability even after regularization. In such cases, an adaptive regularization strategy can be employed, dynamically adjusting the regularization coefficients based on the condition number. Specifically, we define the adaptive regularization coefficient. ,in As the baseline regularization coefficient, To adjust the index, it is typically set between 0.5 and 1.0. As the condition number increases, the regularization coefficient increases accordingly, the cutoff value increases, and the suppression of small singular values ​​is enhanced, thereby improving numerical stability.

[0117] In one alternative implementation,

[0118] Based on the influence weights, the dominant control variables are identified, and constraint optimization is performed on the dominant control variables to obtain an executable control parameter set. The executable control parameter set is then converted into control signals and sent to the actuator for execution, including:

[0119] The comprehensive fuel evaluation index is obtained by multiplying the fuel influence weight and the fuel regulation intensity. The comprehensive medium evaluation index is obtained by multiplying the medium influence weight and the medium regulation intensity. The dominant control variable is determined by combining the comprehensive fuel evaluation index. The target deviation component corresponding to the dominant control variable is extracted from the deviation index. The column vector corresponding to the dominant control variable is extracted from the pre-acquired regularized Jacobian matrix and the ratio with the target deviation component is calculated to obtain the initial regulation increment.

[0120] Extract the upper and lower bounds of the control variable corresponding to the dominant control variable from the equipment operation constraint data, obtain the current operating value corresponding to the dominant control variable, take the target deviation component as the optimization target and the upper and lower bounds of the control variable as the constraint conditions, perform constrained quadratic programming based on the column vector to obtain the optimal control increment, sum the current operating value and the optimal control increment to obtain the dominant control parameter value, allocate the dominant control parameter value to the corresponding control parameter based on the control type of the dominant control variable to construct an executable control parameter group, convert the executable control parameter group into a control signal and send it to the corresponding actuator for execution.

[0121] Obtain fuel impact weights Weight of influence of medium Subsequently, it is necessary to further determine the dominant control variable by considering the actual control capabilities of each control variable. Fuel control intensity. This reflects the overall impact of the fuel supply rate on the system state, and the intensity of medium regulation. This reflects the overall impact of the combustion medium flow rate on the system state. To comprehensively evaluate the control priority of each control variable, a comprehensive fuel evaluation index is introduced. Comprehensive evaluation index of media The calculation method is as follows as well as These two comprehensive evaluation indicators multiply the influence weight by the control intensity, taking into account both the response sensitivity of each control variable under the current deviation state and the physical control capability of each control variable, thus more accurately reflecting the contribution potential of each control variable to the system's restoration of equilibrium.

[0122] The maximum value criterion was used to determine the dominant control variable. Comparative fuel comprehensive evaluation indicators were used. Comprehensive evaluation index of media The size, if If so, then the fuel supply rate is determined to be the dominant control variable, denoted as . Conversely, if If the flow rate of the combustion-supporting medium is determined to be the dominant control variable, it is denoted as... In very rare cases, if the two values ​​are close and the difference is less than a preset threshold, (Typically a value of 0.05) then both are considered as dominant control variables for subsequent optimization. Identifying the dominant control variable allows regulatory resources to be focused on the most effective control channel for the current deviation, avoiding coupling interference and control conflicts that may be caused by simultaneous regulation of multiple variables.

[0123] After determining the dominant control variable, it is necessary to use the deviation index. Extract the target deviation component corresponding to the dominant control variable. Deviation index It typically contains multiple sub-components, such as a temperature deviation component. Oxygen concentration deviation component carbon monoxide concentration deviation from the component and nitrogen oxide concentration deviation component Etc. If the dominant control variable is the fuel supply rate, then the target deviation component... It mainly consists of a temperature deviation component and a carbon monoxide concentration deviation component, and is calculated as follows: ,in and The weighting coefficients for each component are typically set based on the priority of the combustion conditions. If the dominant control variable is the flow rate of the combustion medium, the target deviation component mainly consists of the oxygen concentration deviation component and the nitrogen oxide concentration deviation component, calculated as follows: ,in and These are the corresponding weighting coefficients. Extracting the target deviation component ensures that subsequent optimization processes focus on the system state dimension most significantly influenced by the dominant control variable, improving the targeting and effectiveness of the control.

[0124] From the regularized Jacobian matrix Extract the column vector corresponding to the dominant control variable. If the dominant control variable is the fuel supply rate, then extract the column vector corresponding to the dominant control variable. Columns become column vectors If the dominant control variable is the flow rate of the combustion-supporting medium, then extract the first... Columns become column vectors This column vector describes the sensitivity relationship between the dominant control variable and each system output variable. To obtain the initial control increment... Calculate the ratio of the target deviation component to the corresponding component of the column vector. Specifically, if the target deviation component corresponds to the first [value] of the system output variable... If the dimension is specified, then the initial adjustment increment is calculated as follows: ,in column vector The The component has a negative sign, indicating that the control direction is opposite to the deviation direction to achieve state regression. The initial control increment provides a fast estimate based on a linearized model, providing an initial reference for subsequent constraint optimization.

[0125] Extract the upper bound of the control variable corresponding to the dominant control variable from the equipment operation constraint data. and the lower bound of regulation These constraint data are typically stored in equipment configuration files or real-time monitoring systems, reflecting the physical limitations and safe operating range of the actuators. For example, the upper limit of fuel supply rate regulation may be limited by the maximum flow rate of the fuel delivery pump, while the lower limit is limited by the minimum amount of fuel required to maintain stable combustion; the upper limit of combustion medium flow rate regulation may be limited by the maximum airflow of the blower, while the lower limit is limited by the minimum airflow required to prevent burner flameout. Simultaneously, the current operating values ​​corresponding to the dominant control variables are obtained. This value, obtained through real-time sensor measurements, reflects the current operating status of the actuator. The current operating value, along with the upper and lower bounds of the control, constitutes the feasible region boundary of the optimization problem.

[0126] The optimal control increment is obtained by solving a constrained quadratic programming problem based on the target deviation component and column vector. The objective function of the constrained quadratic programming problem is set as minimizing the quadratic form of the regulation increment and the weighted combination of the objective deviation components, specifically in the form of: ,in This is a penalty coefficient used to balance the cost of regulation with the effect of deviation elimination. Constraints include upper and lower bounds on the regulation increment. This ensures that the adjusted parameter values ​​do not exceed the equipment's operating range. Furthermore, a control rate constraint can be introduced. To prevent rapid changes in regulation from causing system oscillations or damage to actuators, constrained quadratic programming problems can be solved using the interior-point method or the effective set method, utilizing column vectors in the solution process. The gradient information accelerates convergence. The optimal control increment obtained by solving the problem... Under the premise of satisfying all constraints, it can eliminate the target deviation component to the greatest extent, while avoiding secondary problems caused by over-tuning.

[0127] Current running value With optimal control increment Summing yields the values ​​of the dominant control parameters. This parameter value represents the target setpoint that the dominant control variable should reach in the next control cycle. To ensure the validity of the parameter value, a boundary check is required. Then it will be corrected to ;like Then it will be corrected to Boundary correction ensures that the values ​​of the dominant control parameters always remain within the safe operating range of the equipment, avoiding the risk of exceeding the limits due to numerical errors in the optimization algorithm or uncertainties in the model.

[0128] Based on the control type of the dominant control variable, the dominant control parameter values ​​are assigned to the corresponding control parameters to construct an executable control parameter set. If the dominant control variable is the fuel supply rate, then the dominant control parameter values ​​are... The flow rate setting parameter allocated to the fuel supply system is denoted as... Simultaneously, to maintain a stable air-fuel ratio in the combustion system, the flow rate of the combustion medium needs to be adjusted in a coordinated manner according to changes in the fuel supply rate. The calculation method is as follows: ,in This is the air-fuel ratio coordination coefficient, typically determined based on fuel type and burner design. If the dominant control variable is the combustion medium flow rate, then the dominant control parameter value is allocated to the flow rate setting parameter of the combustion medium supply system, denoted as . Meanwhile, the fuel supply rate remains at its current value or is slightly adjusted, calculated as follows: ,in This is a compensation coefficient used to prevent combustion instability caused by controlling a single variable. It allows for the execution of a set of controllable parameters. It consists of the set values ​​of all control parameters, in the form of... .

[0129] The executable control parameter set is converted into control signals and sent to the corresponding actuators for execution. The conversion of control signals needs to be adapted according to the actuator's interface protocol and signal type. For actuators with analog control, the control parameter values ​​are converted into voltage or current signals through linear mapping, with the mapping relationship being as follows: ,in For output voltage signal, and This refers to the voltage range accepted by the actuator. For digitally controlled actuators, the control parameter values ​​are converted into digital commands via a digital-to-analog converter and then transmitted to the actuator's controller via fieldbus or industrial Ethernet protocol. During the control signal transmission process, timestamps and checksum calculations are required to ensure the real-time performance and reliability of the signal transmission. After receiving the control signal, the actuator adjusts physical quantities such as valve opening, pump speed, or fan frequency according to its internal control logic, thereby achieving precise control of the fuel supply rate or the flow rate of the combustion medium.

[0130] To improve the robustness and safety of control, multi-level safety checks are required before issuing control signals. First, it checks whether the values ​​of each parameter in the executable control parameter group meet the equipment's operating constraints; if any exceedances are found, an alarm is triggered and the signal is rejected. Second, it calculates the rate of change of the control increment. ,in This is the incremental adjustment from the previous control cycle. To control the cycle duration, if the rate of change exceeds a preset threshold... Smoothing is then performed using a first-order filter. Smoothing, where The smoothing coefficient is used. Next, the health and communication status of the actuators are checked. If an actuator is faulty or communication is interrupted, the system switches to a backup actuator or activates emergency control mode. Finally, the timestamps of control signal issuance, parameter values, and actuator response status are recorded to create a control log for subsequent analysis and auditing.

[0131] In practical applications, the generation and distribution of executable control parameter sets need to match the dynamic response characteristics of the combustion system. The response times of the fuel supply system and the combustion medium supply system typically differ; the fuel supply system's response time may be several seconds to tens of seconds, while the combustion medium supply system's response time may be hundreds of milliseconds to several seconds. To avoid instantaneous air-fuel ratio mismatch caused by asynchronous responses, a feedforward compensation strategy can be adopted. This strategy predicts the response delay of non-dominant control variables based on the control increment of the dominant control variable and adjusts the setpoints of the non-dominant control variables in advance. The feedforward compensation amount is calculated as follows: ,in For delay compensation function, To address the response time difference, feedforward compensation strategies can significantly improve the coordination and stability of the control process and reduce excessive oscillations during control.

[0132] A second aspect of this invention provides an AI-based intelligent monitoring and control system for coal-fired flames, comprising:

[0133] The feature extraction unit is used to acquire a multi-band image sequence of flames within the combustion area, extract the spectral response curve and spatial gradient distribution features of the multi-band image sequence, perform wavelet transform on the spectral response curve to obtain multi-scale frequency domain features, and perform tensor fusion and dimensionality reduction on the multi-scale frequency domain features and the spatial gradient distribution features to obtain a combustion state feature vector.

[0134] The state analysis unit is used to model the combustion state feature vector through a self-attention mechanism to obtain a long-range correlation representation, perform causal convolution encoding on the long-range correlation representation to obtain a time-dependent sequence, construct a dynamic evolution trajectory based on the time-dependent sequence and calculate the Lyapunov exponent to obtain a stability boundary, measure the phase space distance between the stability boundary and the pre-stored thermodynamic equilibrium state data to obtain a deviation index, identify combustion deviation modes based on the deviation index and generate a state identifier;

[0135] The control execution unit is used to query the Jacobian matrix corresponding to the pre-stored control response relationship data according to the state identifier, solve the influence weights corresponding to the fuel supply rate and the combustion medium flow rate based on the Jacobian matrix, identify the dominant control variable based on the influence weights and perform constraint optimization on the dominant control variable to obtain an executable control parameter set, convert the executable control parameter set into a control signal and send it to the actuator for execution.

[0136] A third aspect of the present invention provides an electronic device, comprising:

[0137] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0138] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0139] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based intelligent monitoring and control method for coal-fired flames, characterized in that, include: A multi-band image sequence of flames within the combustion area is acquired, and the spectral response curve and spatial gradient distribution features of the multi-band image sequence are extracted. Wavelet transform is performed on the spectral response curve to obtain multi-scale frequency domain features. Tensor fusion and dimensionality reduction are performed on the multi-scale frequency domain features and the spatial gradient distribution features to obtain a combustion state feature vector. The combustion state feature vector is modeled using a self-attention mechanism to obtain a long-range correlation representation. The long-range correlation representation is then causally convolutionally encoded to obtain a temporal dependency sequence. Based on the temporal dependency sequence, a dynamic evolution trajectory is constructed and the Lyapunov exponent is calculated to obtain a stability boundary. The stability boundary is then compared with the pre-stored thermodynamic equilibrium state data using a phase space distance metric to obtain a deviation index. Based on the deviation index, combustion deviation modes are identified and a state identifier is generated. According to the state identifier, query the Jacobian matrix corresponding to the pre-stored control response relationship data, solve the influence weights corresponding to the fuel supply rate and the combustion medium flow rate based on the Jacobian matrix, identify the dominant control variable based on the influence weights, and perform constraint optimization on the dominant control variable to obtain the executable control parameter set, convert the executable control parameter set into a control signal and send it to the actuator for execution.

2. The method according to claim 1, characterized in that, Acquire a multi-band image sequence of the flame within the combustion area, extract the spectral response curve and spatial gradient distribution features of the multi-band image sequence, and perform wavelet transform on the spectral response curve to obtain multi-scale frequency domain features, including: The image frame sequence of the combustion area is simultaneously acquired in different band ranges by a multispectral acquisition device. The image frame sequence is timestamped and arranged in band order to obtain a time-aligned image sequence. The radiation intensity value of each pixel position in the time-aligned image sequence in different bands is extracted to construct a spectral radiation intensity matrix and interpolation is performed along the band dimension to obtain the spectral response curve. For each band image in the time-aligned image sequence, the horizontal and vertical gradients are calculated and vectors are synthesized to obtain the gradient magnitude distribution. The gradient magnitude distribution is then subjected to multi-scale Gaussian kernel convolution to obtain a smooth gradient field set. The difference response between the gradient fields at each scale in the smooth gradient field set is calculated and a gradient pyramid is constructed. The spatial gradient distribution features are obtained by performing spatial statistics on the gradient responses at each level in the gradient pyramid. Initialize the scale parameters to construct a wavelet basis sequence and convolve it with the spectral response curve to obtain a set of scale wavelet coefficients. Arrange the set of scale wavelet coefficients in two dimensions according to the scale dimension and the frequency dimension to construct a wavelet transform matrix. Perform energy normalization processing on the wavelet transform matrix to obtain multi-scale frequency domain features.

3. The method according to claim 1, characterized in that, The combustion state feature vector is obtained by tensor fusion and dimensionality reduction of the multi-scale frequency domain features and the spatial gradient distribution features, including: The multi-scale frequency domain features are expanded into a frequency domain feature matrix according to the scale dimension and the frequency dimension. The spatial gradient distribution features are expanded into a gradient feature matrix according to the spatial position and gradient direction. The frequency domain feature rank and the gradient feature rank corresponding to the frequency domain feature matrix are calculated. The tensor decomposition target dimension is determined based on the frequency domain feature rank and the gradient feature rank. The frequency domain feature matrix and the gradient feature matrix are truncated based on the tensor decomposition target dimension to obtain a truncated frequency domain matrix and a truncated gradient matrix. The truncated frequency domain matrix and the truncated gradient matrix are stacked to construct a heterogeneous feature tensor. The cross-modal correlation between the truncated frequency domain matrix and the truncated gradient matrix is ​​calculated to construct a modal coupling weight tensor. The heterogeneous feature tensor and the modal coupling weight tensor are subjected to Hadamard product operation and Tucker decomposition to obtain a fused feature tensor. The fused feature tensor is then subjected to dimensionality reduction projection and flattening to obtain a combustion state feature vector.

4. The method according to claim 1, characterized in that, The combustion state feature vector is modeled using a self-attention mechanism to obtain a long-range correlation representation. This long-range correlation representation is then causally convolutionally encoded to obtain a temporal dependency sequence. Based on this temporal dependency sequence, a dynamic evolution trajectory is constructed, and the Lyapunov exponent is calculated to obtain the stability boundary, including: The combustion state feature vectors are arranged into a state sequence matrix according to the time series. The state sequence matrix is ​​linearly projected to obtain a query matrix, a key matrix, and a value matrix. The key matrix is ​​transposed to obtain a transposed key matrix. The matrix product of the query matrix and the transposed key matrix is ​​calculated to obtain a similarity matrix. The similarity matrix is ​​softmax normalized to obtain an attention weight matrix. The long-range association representation is calculated based on the attention weight matrix and the value matrix. Multiple sets of dilated causal convolution kernels with increasing dilation rates are set and multi-level dilated causal convolutions are performed on the long-range correlation representation in sequence. The output of each level is accumulated and flattened to obtain the temporal dependency sequence. The temporal dependency sequence is embedded with time delay to obtain the embedding state matrix. The displacement vectors of adjacent state points in the embedding state matrix are calculated and numerically integrated to obtain the dynamic evolution trajectory. The trajectory segments are extracted from the dynamic evolution trajectory and the phase space distance between the starting point and the ending point is calculated. The phase space distance is fitted with logarithmic regression with respect to the time interval to obtain the Lyapunov exponent. Based on the Lyapunov exponent, the chaotic characteristics of the combustion state are obtained and the stability boundary is determined.

5. The method according to claim 1, characterized in that, The deviation index is obtained by measuring the phase space distance between the stability boundary and the pre-stored thermodynamic equilibrium state data. Based on the deviation index, combustion deviation modes are identified and state identifiers are generated, including: The equilibrium manifold is obtained by embedding the pre-stored thermodynamic equilibrium state data into a high-dimensional phase space. The Riemannian metric tensor of the equilibrium manifold is calculated and a geodesic coordinate system is constructed. The stability boundary is transformed and mapped to the geodesic coordinate system to obtain the boundary state trajectory. The normal distance field from the boundary state trajectory to the equilibrium manifold is calculated and integrated to obtain the Wasserstein distance. The deviation index is obtained by solving based on the Wasserstein distance and the pre-acquired Lyapunov exponent. The deviation index is subjected to continuous homology decomposition to obtain a multi-level topological feature set. The Betti number sequence of each level in the multi-level topological feature set is calculated, and the topological phase transition mode is identified based on the jump point of the Betti number sequence. Multi-scale permutation entropy is extracted from the time-dependent sequence and an entropy rate time-frequency map is constructed. The entropy rate time-frequency map is subjected to multifractal spectrum analysis to obtain the Hausdorff dimension and multifractal spectrum width. The combustion dynamics complexity level is determined based on the Hausdorff dimension and multifractal spectrum width, and a high-dimensional decision hyperplane is constructed in combination with the topological phase transition mode. The deviation index is projected and classified in the high-dimensional decision hyperplane to identify combustion deviation modes and encode them to generate state identifiers.

6. The method according to claim 1, characterized in that, The Jacobian matrix corresponding to the pre-stored control response relationship data is queried based on the state identifier. The influence weights corresponding to the fuel supply rate and combustion medium flow rate are then calculated based on the Jacobian matrix, including: The state identifier is hashed to obtain the query index value, and the corresponding Jacobian matrix is ​​retrieved from the pre-stored control response relationship data. The Jacobian matrix is ​​decomposed into singular value diagonal matrix to obtain the singular value diagonal matrix and the corresponding condition number is calculated. When the condition number exceeds the preset condition number threshold, the maximum singular value in the singular value diagonal matrix is ​​extracted and the cutoff value is obtained by combining it with the preset regularization coefficient. The singular values ​​in the singular value diagonal matrix that are smaller than the cutoff value are replaced with the cutoff value and the replaced singular value diagonal matrix is ​​reconstructed to obtain the regularized Jacobian matrix. The column vector corresponding to the fuel supply rate is extracted from the regularized Jacobian matrix as the fuel sensitivity vector, and its norm is calculated to obtain the fuel regulation intensity. The column vector corresponding to the combustion medium flow rate is extracted as the medium sensitivity vector, and its norm is calculated to obtain the medium regulation intensity. The temperature deviation component and the oxygen concentration deviation component are extracted from the deviation index. The inner product of the fuel sensitivity vector and the temperature deviation component is calculated to obtain the fuel temperature responsiveness. The inner product of the medium sensitivity vector and the oxygen concentration deviation component is calculated to obtain the medium oxygen concentration responsiveness. The fuel influence weight is calculated based on the fuel regulation intensity and the fuel temperature responsiveness. The medium influence weight is calculated based on the medium regulation intensity and the medium oxygen concentration responsiveness.

7. The method according to claim 1, characterized in that, Based on the influence weights, the dominant control variables are identified, and constraint optimization is performed on the dominant control variables to obtain an executable control parameter set. The executable control parameter set is then converted into control signals and sent to the actuator for execution, including: The comprehensive fuel evaluation index is obtained by multiplying the fuel influence weight and the fuel regulation intensity. The comprehensive medium evaluation index is obtained by multiplying the medium influence weight and the medium regulation intensity. The dominant control variable is determined by combining the comprehensive fuel evaluation index. The target deviation component corresponding to the dominant control variable is extracted from the deviation index. The column vector corresponding to the dominant control variable is extracted from the pre-acquired regularized Jacobian matrix and the ratio with the target deviation component is calculated to obtain the initial regulation increment. Extract the upper and lower bounds of the control variable corresponding to the dominant control variable from the equipment operation constraint data, obtain the current operating value corresponding to the dominant control variable, take the target deviation component as the optimization target and the upper and lower bounds of the control variable as the constraint conditions, perform constrained quadratic programming based on the column vector to obtain the optimal control increment, sum the current operating value and the optimal control increment to obtain the dominant control parameter value, allocate the dominant control parameter value to the corresponding control parameter based on the control type of the dominant control variable to construct an executable control parameter group, convert the executable control parameter group into a control signal and send it to the corresponding actuator for execution.

8. An AI-based intelligent monitoring and control system for coal-fired flames, used to implement the method described in any one of claims 1-7, characterized in that, include: The feature extraction unit is used to acquire a multi-band image sequence of flames within the combustion area, extract the spectral response curve and spatial gradient distribution features of the multi-band image sequence, perform wavelet transform on the spectral response curve to obtain multi-scale frequency domain features, and perform tensor fusion and dimensionality reduction on the multi-scale frequency domain features and the spatial gradient distribution features to obtain a combustion state feature vector. The state analysis unit is used to model the combustion state feature vector through a self-attention mechanism to obtain a long-range correlation representation, perform causal convolution encoding on the long-range correlation representation to obtain a time-dependent sequence, construct a dynamic evolution trajectory based on the time-dependent sequence and calculate the Lyapunov exponent to obtain a stability boundary, measure the phase space distance between the stability boundary and the pre-stored thermodynamic equilibrium state data to obtain a deviation index, identify combustion deviation modes based on the deviation index and generate a state identifier; The control execution unit is used to query the Jacobian matrix corresponding to the pre-stored control response relationship data according to the state identifier, solve the influence weights corresponding to the fuel supply rate and the combustion medium flow rate based on the Jacobian matrix, identify the dominant control variable based on the influence weights and perform constraint optimization on the dominant control variable to obtain an executable control parameter set, convert the executable control parameter set into a control signal and send it to the actuator for execution.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images