A method, apparatus, and readable storage medium for predicting rotary kiln pulverized coal injection rate based on multimodal data dynamic gating.

Abstract

This invention proposes a method, device, and readable storage medium for predicting pulverized coal injection in rotary kilns based on dynamic gating of multimodal data. Addressing the issues of unstable quality and non-stationarity of multimodal data in industrial settings, the device employs a dynamic gating cross-attention mechanism. It achieves adaptive fusion of multimodal features by evaluating modal confidence levels. Combined with a hidden-state clustering deconstruction mechanism, it monitors the displacement of hidden-state cluster centers and identifies points of change in operating conditions using dynamic thresholds, achieving adaptive segmentation of time-series data. Furthermore, it constructs hierarchical features through local and global attention, ultimately outputting the predicted pulverized coal injection rate. This invention improves the accuracy, robustness, and real-time performance of predictions, providing support for intelligent industrial control.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction and control technology for industrial processes, and in particular to a method, apparatus and readable storage medium for predicting the pulverized coal injection rate of a rotary kiln based on dynamic gating of multimodal data. Background Technology

[0002] Rotary kilns are core heat treatment equipment in industries such as cement, metallurgy, and chemicals. They mainly achieve the physicochemical transformation of materials through high-temperature calcination. The amount of pulverized coal injected is a key process parameter that directly affects combustion efficiency and production stability.

[0003] Currently, rotary kiln pulverized coal injection control mainly relies on operator experience, which presents problems such as difficulty in grasping complex correlations of multiple parameters, susceptibility to judgment errors during long-term monitoring, and delayed response to sudden changes in operating conditions. Existing intelligent prediction methods face two major bottlenecks:

[0004] First, the quality of multimodal data (images, values, etc.) in industrial settings is unstable due to sensor failures, environmental interference, and other factors. Traditional fixed-weight fusion methods cannot dynamically adapt to data reliability.

[0005] Second, time series data is non-stationary, and traditional fixed sliding windows cannot accurately capture sudden changes in operating conditions, resulting in insufficient identification of key state changes.

[0006] Therefore, developing an intelligent prediction system that can dynamically integrate multimodal data and adaptively process temporal characteristics is of great significance for improving the operating efficiency of rotary kilns and achieving energy conservation and emission reduction. Summary of the Invention

[0007] This invention provides a method, device, and readable storage medium for predicting rotary kiln pulverized coal injection based on dynamic gating of multimodal data. It addresses the problems of current technologies, such as the inability to dynamically adapt to the quality fluctuations of multimodal data and the difficulty in effectively handling the non-stationarity of time-series data to accurately capture abrupt changes in operating conditions.

[0008] The core technology of this invention is to propose a dynamic gating cross-attention mechanism to achieve quality perception adaptive fusion of multimodal data, and to combine it with a hidden clustering deconstruction mechanism to achieve intelligent adaptive segmentation and hierarchical feature modeling of time series data, thereby constructing an end-to-end prediction device for rotary kiln pulverized coal injection.

[0009] In a first aspect, the present invention provides a method for predicting the pulverized coal injection rate of a rotary kiln based on dynamic gating of multimodal data, the method comprising the following steps:

[0010] Collect multimodal data during the rotary kiln process. The multimodal data includes at least combustion image data and process numerical data. The process numerical data includes at least combustion state parameters and historical pulverized coal injection data. The combustion state parameters include parameters corresponding to under-combustion, normal combustion, and over-combustion states.

[0011] Spatial features are extracted from combustion image data, and high-dimensional feature encoding is performed on process numerical data;

[0012] A dynamic gated cross-attention mechanism is used to fuse spatial features and encoded high-dimensional features. The dynamic gated cross-attention mechanism generates modality confidence by evaluating the quality of data from different modalities, and performs adaptive weighted fusion of features from different modalities based on the modality confidence.

[0013] A latent clustering deconstruction mechanism is used to perform time-series processing on the fused feature sequences. By monitoring the changes in the displacement of the latent cluster centers in the time-series model and combining the dynamic threshold algorithm to identify the points of change in operating conditions, adaptive segmentation of time-series data is achieved.

[0014] A local attention mechanism is used to extract fine-grained features within adaptive segments, while a global attention mechanism is used to integrate information across segments to model long-term dependencies.

[0015] The global temporal features after hierarchical attention processing are subjected to nonlinear transformation to output the predicted value of pulverized coal injection in the rotary kiln.

[0016] Furthermore, spatial feature extraction from the combustion image data includes:

[0017] Extract brightness features, sharpness features, smoke concentration features, and inter-frame consistency features from combustion images;

[0018] High-dimensional feature encoding of process numerical data includes:

[0019] Extract the process parameter change rate characteristics, short-term fluctuation characteristics, and combustion state coding characteristics of the process values.

[0020] Furthermore, in the dynamic gating cross-attention mechanism, modal confidence is generated by nonlinear mapping of the extracted features through a multilayer perceptron, and the confidence weights of different modalities are balanced by regularization constraints.

[0021] Furthermore, in the latent clustering deconstruction mechanism, the latent state of the time series model is extracted through a long short-term memory network. The identification of the working condition change point is determined based on whether the displacement between the current latent state and the historical cluster center exceeds a dynamic threshold. The dynamic threshold is the sum of the mean of the historical displacement and the preset sensitivity coefficient multiplied by the standard deviation of the historical displacement.

[0022] Secondly, the present invention provides a rotary kiln pulverized coal injection quantity prediction device based on multimodal data dynamic gating, comprising:

[0023] The data acquisition module is used to collect multimodal data during the rotary kiln process. The multimodal data includes at least combustion image data and process numerical data. The process numerical data includes at least combustion state parameters and historical pulverized coal injection data. The combustion state parameters include parameters corresponding to under-combustion, normal combustion, and over-combustion states.

[0024] The feature extraction module is used to extract spatial features from combustion image data and perform high-dimensional feature encoding on process numerical data.

[0025] The multimodal fusion module employs a dynamic gated cross-attention mechanism to fuse spatial features and high-dimensional features. This mechanism evaluates the quality of different modal data to generate modal confidence and then performs adaptive weighted fusion of features based on the modal confidence.

[0026] The time series processing module uses a hidden state clustering deconstruction mechanism to process the fused feature sequences. The hidden state clustering deconstruction mechanism monitors the displacement changes of the hidden state cluster centers of the time series model and combines them with a dynamic threshold algorithm to identify the points of change in operating conditions, thereby achieving adaptive segmentation of the time series data.

[0027] The hierarchical attention module includes a local attention mechanism and a global attention mechanism. The local attention mechanism is used to capture local mutation features and small state transitions within the adaptive segmentation, while the global attention mechanism is used to model long-term dependencies and global change trends through cross-segment information interaction.

[0028] The prediction module performs a nonlinear transformation on the global time-series features after hierarchical attention processing, and outputs the predicted value of the rotary kiln pulverized coal injection.

[0029] Furthermore, the feature extraction module includes:

[0030] The image feature extraction unit is used to extract brightness features, sharpness features, smoke concentration features, and inter-frame consistency features of combustion image data through a deep learning network.

[0031] The numerical feature extraction unit is used to extract process parameter change rate features, short-term fluctuation features, and combustion state coding features from process numerical data through a deep learning network.

[0032] Furthermore, the multimodal fusion module includes:

[0033] The confidence evaluation unit is used to generate image modal confidence based on the features output by the image feature extraction unit, and to generate numerical modal confidence based on the features output by the numerical feature extraction unit.

[0034] The dynamic weighting unit is used to construct a confidence weighting matrix based on image modal confidence and numerical modal confidence. The confidence weighting matrix is ​​used to adjust the weights of key features and value features in the feature fusion process to achieve adaptive enhancement of high-quality modal features.

[0035] Furthermore, the confidence evaluation unit performs nonlinear mapping on image features and numerical features through a multilayer perceptron to generate image modal confidence and numerical modal confidence; the multimodal fusion module also includes a regularization unit, which is used to balance image modal confidence and numerical modal confidence through regularization constraints to avoid the model from over-relying on a single modality.

[0036] Furthermore, the timing processing module includes:

[0037] The hidden state extraction unit is used to process the fused feature sequence through a long short-term memory network to generate temporal hidden states;

[0038] The cluster center calculation unit is used to calculate the cluster centers of the time-series hidden state and the displacement between the current hidden state and the historical cluster centers in real time.

[0039] The change point identification unit is used to compare the displacement with the dynamic threshold to identify the change point of the working condition in order to complete the adaptive segmentation of the time series data. The dynamic threshold is determined based on the statistical characteristics of the historical displacement.

[0040] The dynamic threshold is the sum of the mean of historical displacements and the preset sensitivity coefficient multiplied by the standard deviation of historical displacements.

[0041] Thirdly, the present invention provides a readable storage medium storing a computer program, the computer program including program code for controlling a process to execute the process, the process including the rotary kiln pulverized coal injection quantity prediction method based on the above-described multimodal data dynamic gating.

[0042] The main contributions and innovations of this invention are as follows:

[0043] 1. Effectively solves the problem of unstable multimodal data quality: The dynamic gating cross-attention mechanism evaluates the quality of modal data in real time and generates confidence scores, dynamically adjusts the weights of different modal features, automatically enhances the influence of high-quality modalities, suppresses interference from low-quality data, avoids the model's over-reliance on a single modality, balances the complementary value of image morphological features and process values, and significantly improves the robustness and adaptability of multimodal feature fusion.

[0044] 2. Breaking through the limitations of traditional time series modeling: By monitoring the displacement of hidden state cluster centers through a hidden state clustering deconstruction mechanism and combining it with dynamic thresholds, the system accurately identifies points of change in operating conditions, achieving adaptive segmentation of time series data. This overcomes the shortcomings of traditional fixed sliding windows in capturing abrupt changes. Combined with a hierarchical architecture of local and global attention, the system can capture local abrupt change features and model long-term dependencies, effectively solving the prediction lag problem of non-stationary time series data and improving the ability to identify and predict key operating condition changes.

[0045] 3. Overall Improvement in Predictive Performance: Through the synergistic effect of the above mechanisms, the accuracy, robustness, and real-time performance of rotary kiln pulverized coal injection prediction are significantly improved, providing reliable support for parameter optimization and intelligent decision-making in industrial production processes, and powerfully promoting the intelligent and refined management of rotary kiln operation.

[0046] Details of one or more embodiments of the present invention are set forth in the following drawings and description, so that other features, objects and advantages of the invention will be more readily understood. Attached Figure Description

[0047] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0048] Figure 1 This is a flowchart of a rotary kiln pulverized coal injection prediction method based on multimodal data dynamic gating according to an embodiment of the present invention;

[0049] Figure 2 This is an overall structural diagram according to an embodiment of the present invention;

[0050] Figure 3 This is a structural diagram of the dynamic gating cross-attention mechanism according to an embodiment of the present invention;

[0051] Figure 4 This is a scatter plot of predicted pulverized coal injection volume according to an embodiment of the present invention;

[0052] Figure 5 This is a graph showing the effect of time series prediction of pulverized coal injection volume according to an embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0054] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0055] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0056] As a core piece of equipment in the cement, metallurgy and other fields, the prediction of pulverized coal injection in rotary kilns is affected by multimodal data such as combustion images (flame state) and process values ​​(temperature, flow rate, etc.), and it is necessary to deal with dynamic operating conditions in complex industrial environments.

[0057] Based on this, the present invention addresses the problems existing in the prior art by implementing quality-aware adaptive fusion of multimodal data based on a dynamic gating cross-attention mechanism.

[0058] Example 1

[0059] This invention aims to propose a method for predicting the pulverized coal injection rate of a rotary kiln based on dynamic gating of multimodal data. Specifically, it refers to... Figure 1 The method includes the following steps:

[0060] Step 1: Collect multimodal data during the rotary kiln process. The multimodal data should include at least combustion image data and process numerical data. The process numerical data should include at least combustion state parameters and historical pulverized coal injection data. The combustion state parameters should include parameters corresponding to under-combustion, normal combustion, and over-combustion states.

[0061] In this embodiment, combustion images, combustion status data, and pulverized coal injection data during the rotary kiln process are acquired during the dataset construction phase to construct a multimodal dataset.

[0062] Step 2: Extract spatial features from combustion image data and encode high-dimensional features from process numerical data;

[0063] In this embodiment, as Figure 2As shown, in the feature extraction and representation learning stage, a deep convolutional neural network (CNN) is used to extract spatial features from the combustion image, capturing flame morphology, brightness distribution, and boundary features. Simultaneously, a multilayer perceptron (MLP) network is used to perform nonlinear mapping and representation learning on combustion state parameters and pulverized coal injection values, achieving high-dimensional semantic encoding of numerical features. For example:

[0064] (1) Extract the flame brightness, sharpness, dust concentration, and inter-frame consistency attributes from the image data as follows:

[0065] 1. Flame brightness:

[0066]

[0067] Where H and W represent the height and width of the combustion image, respectively, that is, the image contains 1 pixel; The values ​​are the three channels (red, green, and blue) of pixel (i,j), and the results are linearly normalized to the [0,1] interval.

[0068] Used to evaluate the quality of image modalities, providing fundamental features for confidence calculation in dynamic gating cross-attention mechanisms.

[0069] 2. Clarity:

[0070] The Laplacian operator is used to calculate the mean gradient of the image, quantifying the sharpness of the flame edges. The formula is as follows:

[0071]

[0072] Where H and W represent the height and width of the combustion image, respectively, that is, the image contains Each pixel; f(i,j) represents the grayscale value (or color channel value) of the image at pixel (i,j); It is the result of the Laplacian operator's calculation for this pixel. Its essence is the second derivative of the image, which reflects the curvature change of the pixel value in the local area. In the edge area, due to the sudden change in pixel value, the absolute value of the second derivative is larger. After normalization, the result takes the value [0,1]. The larger the value, the clearer the image.

[0073] The calculated result C, after normalization, takes a value range of [0,1]. The larger the value, the sharper the flame edge and the clearer the outline in the image. This feature is used for subsequent image modality quality assessment and provides a key basis for calculating image confidence in the dynamic gating cross-attention mechanism.

[0074] 3. Smoke and dust concentration:

[0075] The formula for calculating the proportion of gray pixels in the HSV color space is:

[0076]

[0077] Where H and W represent the height and width of the combustion image, respectively, that is, the image contains 1 pixel; This is a discriminant function used to determine whether pixel (i,j) is a gray pixel: when the pixel's parameters in the HSV color space satisfy "hue H is 0~180, saturation S < 30, and lightness V < 100", The value is 1 (determined as a gray pixel), otherwise it is 0 (non-gray pixel); H, S, and V are the hue, saturation, and brightness values ​​in the HSV space, respectively; the value of S ranges from [0,1]. The larger the value, the higher the proportion of gray pixels in the image, which corresponds to a higher concentration of smoke in the combustion image (because smoke often appears as a gray area in the image). This index is used to evaluate the quality of image data. Too high a smoke concentration will cause the flame features to become blurred, thereby reducing the reliability of the image modality and providing key feature basis for the subsequent dynamic gating cross-attention mechanism to calculate the image confidence.

[0078] This is equivalent to counting all gray pixels in the image (by summing them up) and then dividing by the total number of pixels. The gray pixel ratio SC is obtained.

[0079] 4. Inter-frame consistency:

[0080] The structural similarity index (SSIM) is used to calculate the similarity between adjacent frames, and the formula is as follows:

[0081]

[0082] Where x and y represent two adjacent frames (such as the current frame and the previous frame). These are the average pixel values ​​of the two frames, reflecting the overall brightness level. These represent the variance of pixel values ​​in two frames, reflecting contrast characteristics. The covariance of pixel values ​​between two frames reflects the structural correlation between pixels; It is a constant (usually set based on the image dynamic range) used to avoid the denominator being zero and to ensure calculation stability. The SSIM value ranges from [0,1]. The closer the value is to 1, the higher the similarity between adjacent frames and the better the consistency between frames (such as stable flame state, no severe shaking or noise interference); the lower the value, the greater the difference between frames (may be due to high temperature dust, sensor shaking causing image instability).

[0083] This metric, as a key feature of image modality, is used to evaluate the stability and quality of image data, and provides a basis for the subsequent calculation of image modality confidence using dynamic gating cross-attention mechanism. High inter-frame consistency indicates strong image reliability and improved confidence; conversely, low consistency reduces confidence, prompting the model to dynamically enhance its attention to high-quality numerical modalities and achieve multimodal adaptive fusion.

[0084] The extracted multi-dimensional image features are input into a deep convolutional neural network (CNN). Through operations such as convolution and pooling, spatial features are integrated and mapped to high dimensions, ultimately outputting a fused image feature vector. This vector contains comprehensive semantic information about the image modality.

[0085] Preferably, the method further includes nonlinearly mapping image features to image modal confidence scores using two layers of multilayer perceptrons (MLPs).

[0086]

[0087] in, The output image modal confidence score is used to quantify the "health" or reliability of combustion image data. The value range is usually [0,1] (the closer the value is to 1, the higher the data quality). The activation function (such as the sigmoid function) restricts the network output to the [0,1] interval, which conforms to the quantization characteristics of confidence. The input image feature set includes key features extracted earlier, such as flame brightness, sharpness, smoke concentration, and inter-frame consistency. These are the weight matrices for the two MLP layers, respectively. This is the corresponding bias term, used to capture the correlation between features through linear transformation.

[0088] (2) Extract the process parameter change rate, short-term fluctuation, and combustion state code values ​​from the numerical data, as follows:

[0089] 1. Rate of change of process parameters:

[0090] Reflects the adjustment range of process parameters between units:

[0091]

[0092] in, This represents the current time (e.g., time t) and the values ​​of process parameters (such as pulverized coal injection rate, temperature, pressure, and other key process indicators). This represents the value of the same process parameter at the previous time point (e.g., time t-1); calculation results The value represents the relative adjustment range of process parameters per unit time. The larger the value, the more drastic the parameter change (such as a sudden and significant increase / decrease in the amount of pulverized coal injected).

[0093] 2. Fluctuation of process parameters:

[0094] The stability of a pulverized coal injection system is quantified using the standard deviation of short-term data, and the formula is as follows:

[0095]

[0096] in, This represents the value of the i-th process parameter (such as pulverized coal injection rate, secondary air temperature, etc.) within a short time window (e.g., from time t-n+1 to time t); n is the window size (usually the number of data points within 10 minutes), representing the short-term data range used to evaluate stability. This is the mean of all process parameter values ​​within the window, reflecting the overall short-term level. Standard deviation This directly reflects the degree of fluctuation in short-term process parameters: The larger the value, the more significant the fluctuation of the parameter value in the short term, and the worse the stability of the pulverized coal injection system. The smaller the value, the more gradual the parameter changes, and the more stable the system operation.

[0097] 3. Combustion status code value:

[0098] The combustion state is transformed into a One-Hot vector, as follows:

[0099] When the combustion state is "underburned", the corresponding vector is [1,0,0];

[0100] When the combustion state is "normal combustion", the corresponding vector is [0,1,0].

[0101] When the combustion state is "overburning", the corresponding vector is [0,0,1].

[0102] The One-Hot vector has the following characteristics: each vector has only one position with a "1" and the rest with "0". The position of "1" uniquely corresponds to a combustion state. This encoding method avoids the misleading "implicit priority of numerical size" that may be introduced when directly using numbers (such as 1, 2, 3) to represent categories (for example, it will not make the model mistakenly think that "overburning" (3) is more "important" than "underburning" (1)). It distinguishes categories only by positional differences.

[0103] The transformed One-Hot vector serves as a key feature of the numerical mode and participates in the evaluation of the numerical mode's confidence. For example, if the combustion state frequently jumps between "underburning" and "overburning" (corresponding to rapid vector switching), it may indicate sensor malfunction or unstable operating conditions, thus reducing the reliability of the numerical mode. The dynamic gating cross-attention mechanism incorporates this feature to adjust the weights of the numerical mode, ensuring that the model responds appropriately to changes in the combustion state and improving the accuracy of pulverized coal injection prediction.

[0104] The extracted multidimensional numerical features are input into a multilayer perceptron (MLP), and high-dimensional semantic mapping and fusion are performed through nonlinear transformations (such as activation functions and fully connected layers), ultimately outputting a numerical feature vector. This vector contains comprehensive dynamic information about the numerical modes.

[0105] Preferably, the numerical modal features are nonlinearly mapped to numerical modal confidence levels using a two-layer multilayer perceptron (MLP).

[0106]

[0107] in, The output numerical mode confidence score is used to quantify the reliability of process numerical data (such as parameters such as pulverized coal injection rate and temperature). The value range is usually [0,1] (the closer the value is to 1, the higher the data quality). The activation function (such as the sigmoid function) restricts the network output to the [0,1] interval, which conforms to the quantization characteristics of confidence. The set of numerical features to be input includes key features such as the rate of change of process parameters, short-term volatility, and combustion state coding extracted above; These are the weight matrices for the two MLP layers, respectively. This is the corresponding bias term, used to capture the correlation between features through linear transformation.

[0108] Finally, for image feature vectors ( (Image feature dimension) and numerical feature vector ( (For numerical feature dimensions), calculate its Query, Key, and Value:

[0109]

[0110] in, The weight matrices corresponding to the image modalities are used to map image features to the "query", "key", and "value" spaces, respectively.

[0111] The weight matrices corresponding to the numerical modes are used to map numerical features to the "query", "key", and "value" spaces, respectively.

[0112] Among them, the generation of Query, Key, and Value for the image modality:

[0113] Image feature vector It is a high-dimensional representation of an image modality formed by feature extraction (such as fusing features like flame brightness and sharpness extracted by CNN). To enable it to participate in cross-modal attention calculations, it needs to undergo a linear transformation using three learnable weight matrices:

[0114] Image features are queried through the weight matrix. The mapping yields the degree of correlation between "active query" and numerical modal features.

[0115] Image features are obtained through a key weight matrix. The mapped value is used to match the numerical modality query and calculate the association weight.

[0116] Image features are weighted by a value weight matrix. The mapped information is the core feature information in the image modality used for final fusion.

[0117] Query, Key, and Value Generation in Numerical Modalities:

[0118] Numerical eigenvectors It is a high-dimensional representation of numerical modes formed by feature encoding (such as the fusion of features like process parameter change rate and volatility processed by MLP). Similarly, a linear transformation is performed through three dedicated learnable weight matrices:

[0119] Query ( ): Calculated by querying the weight matrix based on numerical features The mapping yields the degree of correlation between "active query" and image modal features;

[0120] Key ( ): (Based on numerical features through the key weight matrix) The mapped value is used to match the image modality query and calculate the association weight;

[0121] Value ): (Based on numerical features through a value weight matrix) The mapped data is the core feature information used for final fusion in the numerical modes.

[0122] This set of formulas is used in the dynamic gated cross-attention mechanism to map the original feature vectors of the image modality and the numerical modality into the "query," "key," and "value" vectors required by the attention mechanism, respectively. In short, these mappings are a "preprocessing" step of the attention mechanism, converting features from different modalities into vector forms suitable for calculating association weights, thus laying the foundation for the "quality-aware adaptive fusion" achieved by the dynamic gated cross-attention mechanism.

[0123] Step 3: A dynamic gated cross-attention mechanism is used to fuse spatial features and encoded high-dimensional features. The dynamic gated cross-attention mechanism generates modality confidence by evaluating the quality of different modality data, and performs adaptive weighted fusion of different modality features based on the modality confidence.

[0124] In this embodiment, during the multimodal dynamic gating fusion stage, a dynamic gating cross-attention mechanism is used to evaluate the quality and reliability of different modal data in real time, calculate modality confidence scores, and then adaptively weighted fusion of different modal features is performed accordingly. This effectively suppresses interference from low-quality data and significantly improves the robustness and adaptability of feature fusion. For example:

[0125] like Figure 3 As shown, in the dynamic gated cross-attention mechanism, the weight calculation integrates feature relevance and modality reliability, essentially coupling the influencing factors of both through a mathematical structure. In contrast, the weights in traditional attention mechanisms are determined solely by the semantic association between features, as shown in the formula:

[0126]

[0127] The dynamic gating cross-attention mechanism upgrades the cross-attention mechanism by introducing modality confidence g:

[0128]

[0129] in, This improved method optimizes attention weight allocation from two aspects: feature relevance and modal reliability, and represents the diagonal matrix corresponding to the confidence level.

[0130] 1. Feature correlation:

[0131] (Query vector after image feature mapping) and The dot product of the key vectors (after numerical feature mapping) measures the semantic association between image features and numerical features.

[0132] (Query vector after numerical feature mapping) and The dot product of the key vectors (after image feature mapping) captures the logical relationship between numerical features and image features.

[0133] 2. Modal reliability:

[0134] When calculating At that time, the value of the numerical mode quilt Weighted, if the numerical mode confidence level Low (e.g., data jumps caused by sensor malfunction), then The weights are attenuated, weakening the impact of unreliable numerical features;

[0135] Similarly, Image modality quilt Weighting, when the image is blurred due to dust When the value is low, its feature weights decrease accordingly.

[0136] By regularizing the modal confidence score To avoid the model becoming overly reliant on a single modality and losing the advantages of multimodal fusion.

[0137] Finally, the two attention outputs are concatenated along the channel dimension, and then feature mapping is performed using a non-linear activation function and a fully connected layer.

[0138]

[0139] in, The output mapping matrix is ​​denoted by ; Concat represents the feature concatenation operation along the channel dimension; ReLU is a non-linear activation function.

[0140] Final output It is the core achievement of the dynamic gating cross-attention mechanism, which not only contains the dominant information of high-quality modalities, but also balances the complementary value of bimodalities, laying a reliable multimodal feature foundation for subsequent processing.

[0141] Step 4: The fused feature sequence is processed by a hidden state clustering deconstruction mechanism. By monitoring the displacement changes of the hidden state cluster centers in the time series model and combining the dynamic threshold algorithm to identify the points of change in working conditions, the time series data is adaptively segmented.

[0142] In this embodiment, an optimized sliding window strategy is applied to the fused feature sequence during the latent clustering and temporal decomposition stages. The latent state representation is extracted by LSTM, the displacement change of the latent cluster center is monitored, and a dynamic threshold algorithm is combined to achieve accurate identification of the working condition change point and adaptive segmentation of the time series data.

[0143] Specifically, addressing the non-stationary nature of industrial time-series data, this invention proposes a Latent State Clustering Decomposition Mechanism (LSCDM) for time-series data. This mechanism achieves adaptive segmentation by monitoring the displacement of LSTM latent state cluster centers, and combines this with dynamic thresholding to identify points of change in operating conditions, avoiding the failure to capture abrupt changes by traditional fixed windows. This mechanism employs a two-layer attention architecture: local attention enhances the extraction of abrupt change features within dynamic segmentation, while global attention models long-term dependencies across segments, effectively solving the prediction lag problem caused by the non-stationarity of industrial data. For example:

[0144] 1. Use the feature sequence output by the dynamic gating cross-attention mechanism as input:

[0145]

[0146] in, : Represents the fusion feature output by the dynamic gating cross-attention mechanism at time step t. It is a comprehensive feature of the bidirectional attention results of "image → numerical" and "numerical → image" after splicing and nonlinear transformation (containing complementary information of image mode and numerical mode, and has been filtered by the dynamic gating mechanism to remove low-quality data interference).

[0147] : The fused features at time step t Defined as the input characteristics of the timing processing module .

[0148] : t is the time step index, T is the total sequence length, indicating that the entire input is a feature sequence arranged in chronological order. That is, from all time steps A time-series feature sequence composed of time dimensions.

[0149] 2. Use LSTM to process the fused feature sequence and generate the hidden state at each time step:

[0150]

[0151] in, The LSTM hidden state at the current time step (time t) is a compressed representation of "current fused features + historical time series information", containing key temporal dynamic features up to time t;

[0152] LSTM(·): The computational unit of a Long Short-Term Memory network, which controls the inflow, retention and output of information through gating mechanisms (input gate, forget gate, output gate), and solves the long-range dependency forgetting problem of traditional recurrent neural networks (RNN);

[0153] : The LSTM hidden state at the previous time step (t-1), which contains historical time series information up to t-1.

[0154] 3. For a fixed segment at time t, calculate the cluster centers of that segment before time t:

[0155]

[0156] in, : Represents the cluster centers before time t within a fixed segment;

[0157] k: The starting time step of the current fixed segment (i.e., the segment starts from time k).

[0158] t: The current time step to be evaluated (it is necessary to determine whether time t still belongs to the current segment);

[0159] : The hidden state output by the LSTM at time j (containing the temporal dynamic information up to time j).

[0160] 4. Calculate the hidden state at the current time point t and Euclidean distance (displacement):

[0161]

[0162] in, The cluster center of the hidden state within the current fixed segment up to time t-1 is the "typical feature benchmark" of the historical hidden states within that segment.

[0163] It is a "quantitative indicator" for judging changes in operating conditions in the latent clustering deconstruction mechanism: by measuring the degree of deviation between the current latent state and the historical benchmark, it can accurately capture abrupt changes in time series data, providing a basis for adaptive segmentation of time series data. This dynamic monitoring method overcomes the shortcomings of traditional fixed windows in flexibly responding to abrupt changes in operating conditions, ensuring that the subsequent hierarchical attention model can perform targeted modeling for the operating condition characteristics of different segments, thereby improving the predictive performance of non-stationary time series data.

[0164] 5. Compare the current cluster center offset with the dynamic threshold to determine whether the current node is a segmentation point:

[0165]

[0166] in, The dynamic threshold at time t is used to determine the current offset. Whether a fluctuation is considered "abnormal" (i.e., a sudden change in operating conditions) is essentially based on the statistical characteristics of historical offsets.

[0167]

[0168] in : Represents the set of all historical offsets from time 2 to time t-1 (i.e. These offsets reflect the historical fluctuations in the deviation of the hidden state from the cluster center under normal operating conditions;

[0169] The mean of the historical offset set reflects the typical level (central trend) of the offset under normal operating conditions.

[0170] The standard deviation of the historical offset set reflects the fluctuation range (dispersion) of the offset under normal operating conditions.

[0171] Hyperparameter (usually 2~3) is used to adjust the contribution weight of the standard deviation to the threshold (the larger the weight, the stronger the tolerance of the threshold to historical fluctuations).

[0172] This dynamic threshold based on historical statistical characteristics overcomes the shortcomings of fixed thresholds (such as a preset constant) that cannot adapt to dynamic changes in operating conditions (for example, a fixed threshold may be too loose under stable operating conditions, leading to missed judgments, or too strict under fluctuating operating conditions, leading to misjudgments). It makes time series segmentation more accurate, provides a reliable basis for subsequent hierarchical attention models to model different segments, and ultimately improves the predictive robustness of non-stationary time series data.

[0173] Step 5: Use a local attention mechanism to extract fine-grained features within the adaptive segmentation, and use a global attention mechanism to integrate information across segments to model long-term dependencies.

[0174] In this embodiment, during the hierarchical attention feature modeling stage, a local attention mechanism is deployed within the identified dynamic segments to perform fine-grained feature extraction and enhancement, accurately capturing local mutation patterns. Simultaneously, a global attention mechanism is used to achieve cross-segment information interaction and integration, effectively modeling long-term dependencies and constructing a complete temporal representation. For example:

[0175] 1. For each dynamic segment S i First, a local attention (LocalAtt) mechanism is applied to extract fine-grained features within a segment:

[0176]

[0177] Among them, S i This represents the i-th local temporal segment output by the hidden clustering deconstruction mechanism, which is a subsequence obtained after adaptive segmentation of the temporal data. It originates from step four, based on the hidden state offset. With dynamic threshold In comparison, when Time-triggered segmentation divides time-series data into multiple consecutive subsequences. Each S i This corresponds to a relatively stable operating condition (such as "normal combustion stable period" or "pulverized coal injection quantity adjustment transition period").

[0178] 2. Perform global modeling of the representation of all segments:

[0179] Within a fixed window L, local features of all segments are considered. Perform global modeling:

[0180]

[0181] Global attention calculation:

[0182]

[0183] First, core features of each operating condition segment are extracted through local modeling. Then, the relationships between segments are integrated through global modeling. This enables the model to accurately capture local details while grasping the overall time series trend, significantly improving the modeling capability for non-stationary industrial time series data. Global attention calculation integrates information from multiple local segments within a window, capturing long-range dependencies across segments (such as causal relationships between different operating conditions), providing feature support from a global perspective for subsequent pulverized coal injection prediction.

[0184] Among them, the fixed window L refers to a time range containing multiple continuous local segments (such as a window composed of the three most recent stable operating conditions), which is used to limit the time range of global modeling, avoid information redundancy caused by excessively long time series, and at the same time ensure that the focus is on the correlation of recent operating conditions.

[0185] : Represents n local segment features within window L, each It is through local attention (LocalAtt) on the i-th stable operating condition segment ( The feature vector obtained after processing (i.e.) This includes key time step information within the segment (such as the core characteristics of the "normal combustion segment" and the key parameter changes in the "adjustment segment").

[0186] The n local features within the window are concatenated along the sequence dimension to form a global feature sequence, integrating the information of all local segments within the window.

[0187] The learnable global weight matrix is ​​used to map the concatenated global feature sequence to the "query", "key", and "value" spaces, respectively, to achieve the unification of feature dimensions and the extraction of global semantics.

[0188] The generated global query, key, and value vectors are used for subsequent global attention calculations to measure the degree of correlation between different local segments (such as the dependency between "normal combustion segment" and "over-combustion segment").

[0189] 3. Compress the output of global attention into a fixed-dimensional feature representation through pooling operations:

[0190]

[0191] Where GlobalAtt: the output of global attention, is the sum of all local segment features within a fixed window L ( The feature sequence after cross-segment association modeling. Its dimension is usually related to the number of segments n in the window (e.g., n segments correspond to n feature vectors, the dimension is [B,n,d], where B is the batch size and d is the feature dimension), so it will change with the number of segments in the window (i.e., "variable dimension").

[0192] Pool(·): Pooling operations (commonly average pooling, max pooling, or self-attention pooling, etc.) compress the sequence dimension of GlobalAtt, eliminating dimensional differences caused by varying numbers of segments within the window. For example, average pooling calculates the average of all feature vectors in the GlobalAtt sequence to obtain a comprehensive feature; max pooling selects the most significant feature vector in the sequence (e.g., the largest dimension), retaining key information.

[0193] The fixed-dimensional feature representation obtained after pooling has a dimension that is independent of the number of segments within the window (e.g., fixed as [B,d]). It is a "condensed version" of the global features within the window and contains the core information of cross-segment associations within the window.

[0194] Step 6: Perform a nonlinear transformation on the global time-series features after hierarchical attention processing to output the predicted value of the rotary kiln pulverized coal injection.

[0195] In this embodiment, during the end-to-end prediction and decision support stage, global temporal features processed with hierarchical attention are input into the prediction layer. A high-precision prediction of the rotary kiln's pulverized coal injection rate is generated through nonlinear transformation, providing real-time and accurate process parameter optimization suggestions for the production process, thus achieving intelligent control and optimization of the rotary kiln's combustion process. For example:

[0196] The end-to-end prediction and decision support stage can be broken down into three core components: "prediction layer modeling," "high-precision prediction," and "decision support generation," ultimately achieving intelligent control of the rotary kiln combustion process.

[0197] 1. Input global temporal features

[0198] The input features are fixed-dimensional features obtained by pooling operations after hierarchical attention processing. (Right now This feature integrates the following key information:

[0199] Local core features of a single stable operating condition segment (through) Extract key values ​​such as temperature and pressure during the "normal combustion phase".

[0200] Cross-segment correlation of multiple continuous operating conditions (captured through GlobalAtt, such as the impact of "coal injection rate adjustment segment" on the subsequent "combustion stabilization segment");

[0201] Global trends within the window (preserved through pooling, such as the overall combustion efficiency changes over the last three operating conditions).

[0202] 2. Nonlinear transformation of the prediction layer

[0203] The prediction layer typically employs a multilayer perceptron (MLP) or a fully connected network with activation functions (such as ReLU or Swish). Perform nonlinear transformation:

[0204] in This is the predicted value of the pulverized coal injection rate. The purpose of the nonlinear transformation is:

[0205] Explore the complex mapping relationship between global features and pulverized coal injection rate (e.g., the pattern of "combustion temperature decrease + pressure increase" may correspond to "the need to increase pulverized coal injection rate").

[0206] Fit nonlinear and strongly coupled process characteristics in industrial scenarios (such as the nonlinear relationship between pulverized coal injection rate and parameters such as rotary kiln speed and material moisture content).

[0207] 3. High-precision forecasting and decision support

[0208] High-precision prediction: Through model training (e.g., adjusting the weights of the MLP using historical data), The goal is to approximate the actual pulverized coal injection rate y as closely as possible (through loss functions such as MSE optimization), ultimately achieving high-precision prediction with errors within the industrially permissible range (e.g., error <2%).

[0209] Decision support generation: based on predicted values Based on production goals (such as "ensuring stable combustion temperature" and "reducing energy consumption"), specific process optimization suggestions are generated:

[0210] If the predicted pulverized coal injection rate is too low (which may lead to insufficient temperature), it is recommended to "increase the pulverized coal injection rate by X kg / h"; if the predicted pulverized coal injection rate is too high (which may lead to energy waste or overburning), it is recommended to "decrease the pulverized coal injection rate by Y kg / h".

[0211] Real-time performance guarantee: Since the previous feature extraction (hierarchical attention) and prediction layer calculation are both end-to-end neural network forward propagation, they can be completed in milliseconds, thus meeting the real-time requirements of industrial production (such as updating predictions and suggestions every 5 seconds).

[0212] 4. Intelligent control and optimization

[0213] Predicted value The decision-making suggestions will be fed back to the rotary kiln's control system (such as PLC or DCS) to achieve two optimization modes:

[0214] Passive optimization: Operators manually adjust the opening of the pulverized coal injection valve based on recommendations;

[0215] Proactive optimization: The system automatically adjusts the coal injection rate dynamically based on the predicted value (e.g., when the predicted "combustion efficiency declines", the coal injection rate is finely adjusted in advance to maintain stability), ultimately achieving the goal of "energy saving and consumption reduction, and reducing failures" (e.g., reducing coal consumption per ton of clinker and reducing the risk of kiln shutdown due to unstable combustion).

[0216] like Figure 4 As shown in the figure, the fitted curves demonstrate that the model exhibits excellent performance in time series data prediction tasks. The scatter plots of the true and predicted values ​​closely follow the ideal prediction line. The low RMSE (0.0729) and MAE (0.0262) values ​​indicate small prediction bias. The R² reaches 0.9600, indicating that the model can explain 96% of the data fluctuations and effectively capture the changing trends of time series data. It has high reliability in modeling and predicting non-stationary time series data of rotary kiln systems. Although there may be room for optimization in a few sparse points, the overall model is well-suited to practical application needs and provides strong support for prediction in complex time series scenarios.

[0217] like Figure 5 As shown in the figure, the time series prediction effect of pulverized coal injection demonstrates the good performance of the model of this invention. In the time series prediction curve above, the blue true value and the red predicted value have a high overall trend. The pink prediction interval effectively covers the fluctuations. The RMSE (0.0729) and MAE (0.0262) are low and the R² reaches 0.96, indicating that the model accurately captures data changes. In the error analysis figure below, the error mostly fluctuates slightly around the 0 line, with only a few peaks. This indicates that the model prediction is stable. Although there may be errors at extreme times, the overall prediction adaptability of pulverized coal injection in rotary kilns is strong, which can provide reliable support for process monitoring and decision-making.

[0218] In summary, this invention proposes a dynamic gating cross-attention mechanism that uses a lightweight gating module to assess the quality of data across different modalities. For image modalities, the mechanism extracts key features such as flame brightness, sharpness, smoke concentration, and inter-frame consistency; for numerical modalities, it extracts process-related indicators such as rate of change, volatility, and combustion state encoding. These features are input into their respective confidence assessment modules, quantifying and outputting the confidence scores of the modal data to effectively reflect the "health status" of the data.

[0219] The innovation of this scheme lies in deeply embedding modal confidence scores into a cross-attention mechanism, enabling weight calculation to simultaneously integrate feature correlation and modal reliability. Specifically, when the system detects a decrease in the confidence of a combustion image due to high-temperature dust, or data anomalies due to aging of the numerical sensor, it dynamically adjusts the attention allocation based on the confidence score—weighting the Key and Value using a confidence diagonal matrix, automatically enhancing attention to another high-quality modality, and achieving adaptive fusion based on data quality perception. The confidence assessment process is optimized through an end-to-end loss function, while introducing regularization constraints to avoid over-reliance on a single modality, ensuring a balance between the complementary value of image morphological features and process numerical values ​​in pulverized coal injection prediction.

[0220] Example 2

[0221] Based on the same concept, this invention also proposes a rotary kiln pulverized coal injection prediction device based on multimodal data dynamic gating, comprising:

[0222] The data acquisition module is used to collect multimodal data during the rotary kiln process. The multimodal data includes at least combustion image data and process numerical data. The process numerical data includes at least combustion state parameters and historical pulverized coal injection data. The combustion state parameters include parameters corresponding to under-combustion, normal combustion, and over-combustion states.

[0223] The feature extraction module is used to extract spatial features from combustion image data and perform high-dimensional feature encoding on process numerical data.

[0224] The multimodal fusion module employs a dynamic gated cross-attention mechanism to fuse spatial features and high-dimensional features. This mechanism evaluates the quality of different modal data to generate modal confidence and then performs adaptive weighted fusion of features based on the modal confidence.

[0225] The time series processing module uses a hidden state clustering deconstruction mechanism to process the fused feature sequences. The hidden state clustering deconstruction mechanism monitors the displacement changes of the hidden state cluster centers of the time series model and combines them with a dynamic threshold algorithm to identify the points of change in operating conditions, thereby achieving adaptive segmentation of the time series data.

[0226] The hierarchical attention module includes a local attention mechanism and a global attention mechanism. The local attention mechanism is used to capture local mutation features and small state transitions within the adaptive segmentation, while the global attention mechanism is used to model long-term dependencies and global change trends through cross-segment information interaction.

[0227] The prediction module performs a nonlinear transformation on the global time-series features after hierarchical attention processing, and outputs the predicted value of the rotary kiln pulverized coal injection.

[0228] In this embodiment, the feature extraction module includes:

[0229] The image feature extraction unit is used to extract brightness features, sharpness features, smoke concentration features, and inter-frame consistency features of combustion image data through a deep learning network.

[0230] The numerical feature extraction unit is used to extract process parameter change rate features, short-term fluctuation features, and combustion state coding features from process numerical data through a deep learning network.

[0231] In this embodiment, the multimodal fusion module includes:

[0232] The confidence evaluation unit is used to generate image modal confidence based on the features output by the image feature extraction unit, and to generate numerical modal confidence based on the features output by the numerical feature extraction unit.

[0233] The dynamic weighting unit is used to construct a confidence weighting matrix based on image modal confidence and numerical modal confidence. The confidence weighting matrix is ​​used to adjust the weights of key features and value features in the feature fusion process to achieve adaptive enhancement of high-quality modal features.

[0234] In this embodiment, the confidence evaluation unit performs nonlinear mapping on image features and numerical features using a multilayer perceptron to generate image modal confidence and numerical modal confidence. The multimodal fusion module also includes a regularization unit, which balances image modal confidence and numerical modal confidence through regularization constraints to avoid the model from over-relying on a single modality.

[0235] In this embodiment, the timing processing module includes:

[0236] The hidden state extraction unit is used to process the fused feature sequence through a long short-term memory network to generate temporal hidden states;

[0237] The cluster center calculation unit is used to calculate the cluster centers of the time-series hidden state and the displacement between the current hidden state and the historical cluster centers in real time.

[0238] The change point identification unit is used to compare the displacement with the dynamic threshold to identify the change point of the working condition in order to complete the adaptive segmentation of the time series data. The dynamic threshold is determined based on the statistical characteristics of the historical displacement.

[0239] The dynamic threshold is the sum of the mean of historical displacements and the preset sensitivity coefficient multiplied by the standard deviation of historical displacements.

[0240] Example 3

[0241] This embodiment also provides an electronic device, see reference. Figure 6 It includes a memory 404 and a processor 402, wherein the memory 404 stores a computer program and the processor 402 is configured to run the computer program to perform the steps in any of the above method embodiments.

[0242] Specifically, the processor 402 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement embodiments of the present invention.

[0243] Memory 404 may include a mass storage device for data or instructions. For example, and not limitingly, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to a data processing device. In a particular embodiment, memory 404 is non-volatile memory. In a particular embodiment, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random-Access Memory (FPMDRAM), Extended Data Out Dynamic Random-Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0244] The memory 404 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 402.

[0245] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any of the rotary kiln pulverized coal injection prediction methods based on multimodal data dynamic gating in the above embodiments.

[0246] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408, wherein the transmission device 406 is connected to the processor 402, and the input / output device 408 is connected to the processor 402.

[0247] The transmission device 406 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 406 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0248] Input / output device 408 is used to input or output information.

[0249] Example 4

[0250] This embodiment also provides a readable storage medium storing a computer program, the computer program including program code for controlling a process to execute the process, the process including the rotary kiln pulverized coal injection quantity prediction method based on multimodal data dynamic gating according to Embodiment 1.

[0251] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0252] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0253] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets, and / or macros can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. The computer program product may include one or more computer-executable components configured to perform the embodiments when the program is run. The one or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted in this respect that, as Figure 1 Any box in the logical flow can represent a program step, or interconnected logic circuits, boxes and functions, or a combination of program steps and logic circuits, boxes and functions. Software can be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0254] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0255] The above embodiments are merely illustrative of several implementations of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A method for predicting the pulverized coal injection rate of a rotary kiln based on dynamic gating of multimodal data, characterized in that, The following steps are involved: Multimodal data is collected during the rotary kiln process. The multimodal data includes at least combustion image data and process numerical data. The process numerical data includes at least combustion state parameters and historical pulverized coal injection data. The combustion state parameters include parameters corresponding to under-combustion, normal combustion, and over-combustion states. Spatial features are extracted from the combustion image data, and high-dimensional feature encoding is performed on the process numerical data; A dynamic gated cross-attention mechanism is used to fuse the spatial features and the encoded high-dimensional features. The dynamic gated cross-attention mechanism generates modality confidence by evaluating the quality of different modality data, and performs adaptive weighted fusion of different modality features based on the modality confidence. A latent clustering deconstruction mechanism is used to perform time-series processing on the fused feature sequences. By monitoring the changes in the displacement of the latent cluster centers in the time-series model and combining the dynamic threshold algorithm to identify the points of change in operating conditions, adaptive segmentation of time-series data is achieved. A local attention mechanism is used to extract fine-grained features within the adaptive segments, while a global attention mechanism is used to integrate information across segments to model long-term dependencies. The global temporal features after hierarchical attention processing are subjected to nonlinear transformation to output the predicted value of pulverized coal injection in the rotary kiln.

2. The method for predicting rotary kiln pulverized coal injection rate based on multimodal data dynamic gating as described in claim 1, characterized in that, Spatial feature extraction from combustion image data includes: Extract brightness features, sharpness features, smoke concentration features, and inter-frame consistency features from combustion images; High-dimensional feature encoding of process numerical data includes: Extract the process parameter change rate characteristics, short-term fluctuation characteristics, and combustion state coding characteristics of the process values.

3. The rotary kiln pulverized coal injection prediction method based on multimodal data dynamic gating as described in claim 1, characterized in that, In the dynamic gating cross-attention mechanism, modal confidence is generated by nonlinear mapping of the extracted features through a multilayer perceptron, and the confidence weights of different modalities are balanced by regularization constraints.

4. A method for predicting rotary kiln pulverized coal injection rate based on dynamic gating of multimodal data as described in any one of claims 1 to 3, characterized in that, In the latent clustering deconstruction mechanism, the latent state of the time series model is extracted through a long short-term memory network. The identification of the working condition change point is determined based on whether the displacement between the current latent state and the historical cluster center exceeds a dynamic threshold. The dynamic threshold is the sum of the mean of the historical displacement and the preset sensitivity coefficient multiplied by the standard deviation of the historical displacement.

5. A rotary kiln pulverized coal injection quantity end-to-end prediction device, characterized in that, include: The data acquisition module is used to collect multimodal data during the rotary kiln process, and the multimodal data includes at least combustion image data and process numerical data; The process numerical data includes at least combustion state parameters and historical pulverized coal injection data. The combustion state parameters include parameters corresponding to under-combustion, normal combustion, and over-combustion states. The feature extraction module is used to extract spatial features from the combustion image data and to encode high-dimensional features from the process numerical data. The multimodal fusion module employs a dynamic gated cross-attention mechanism to fuse the spatial features and high-dimensional features. The dynamic gated cross-attention mechanism generates modal confidence by evaluating the quality of different modal data, and performs adaptive weighted fusion of different modal features based on the modal confidence. The time series processing module uses a hidden state clustering deconstruction mechanism to perform time series processing on the fused feature sequences. The hidden state clustering deconstruction mechanism monitors the displacement changes of the hidden state cluster centers of the time series model and combines a dynamic threshold algorithm to identify the points of change in working conditions, thereby realizing adaptive segmentation of the time series data. The hierarchical attention module includes a local attention mechanism and a global attention mechanism. The local attention mechanism is used to capture local mutation features and minor state transitions within the adaptive segmentation, while the global attention mechanism is used to model long-term dependencies and global change trends through cross-segment information interaction. The prediction module performs a nonlinear transformation on the global time-series features after hierarchical attention processing, and outputs the predicted value of the rotary kiln pulverized coal injection.

6. The rotary kiln pulverized coal injection quantity end-to-end prediction device as described in claim 5, characterized in that, The feature extraction module includes: The image feature extraction unit is used to extract the brightness features, sharpness features, smoke concentration features, and inter-frame consistency features of the combustion image data through a deep learning network. The numerical feature extraction unit is used to extract the process parameter change rate features, short-term fluctuation features, and combustion state coding features of the process numerical data through a deep learning network.

7. The rotary kiln pulverized coal injection quantity end-to-end prediction device as described in claim 6, characterized in that, The multimodal fusion module includes: The confidence evaluation unit is used to generate image modal confidence based on the features output by the image feature extraction unit, and to generate numerical modal confidence based on the features output by the numerical feature extraction unit. The dynamic weighting unit is used to construct a confidence weighting matrix based on the image modality confidence and the numerical modality confidence. The key features and value features in the feature fusion process are weighted and adjusted through the confidence weighting matrix to achieve adaptive enhancement of high-quality modality features.

8. The rotary kiln pulverized coal injection quantity end-to-end prediction device as described in claim 7, characterized in that, The confidence evaluation unit performs nonlinear mapping on the image features and numerical features using a multilayer perceptron to generate the image modal confidence and the numerical modal confidence. The multimodal fusion module also includes a regularization unit, which balances the image modal confidence and the numerical modal confidence through regularization constraints to avoid the model from over-relying on a single modality.

9. The rotary kiln pulverized coal injection quantity end-to-end prediction device as described in claim 5, characterized in that, The timing processing module includes: The hidden state extraction unit is used to process the fused feature sequence through a long short-term memory network to generate temporal hidden states; The cluster center calculation unit is used to calculate the cluster centers of the time-series hidden state and the displacement between the current hidden state and the historical cluster centers in real time. The change point identification unit is used to compare the displacement with the dynamic threshold to identify the change point of the working condition in order to complete the adaptive segmentation of the time series data. The dynamic threshold is determined based on the statistical characteristics of the historical displacement. The dynamic threshold is the sum of the mean of historical displacements and the preset sensitivity coefficient multiplied by the standard deviation of the historical displacements.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, the computer program including program code for controlling a process to execute the process, the process including the rotary kiln pulverized coal injection prediction method based on multimodal data dynamic gating according to any one of claims 1 to 4.

Images