A rotary kiln energy consumption optimization method and system combined with visual recognition

Abstract

The application provides a rotary kiln energy consumption optimization method and system combined with visual recognition, comprising: preprocessing kiln operation data, and establishing a unified fusion feature representation; adopting a stacked long short-term memory network to process long-term dependence relationship, and simultaneously processing local time sequence mode through a multi-layer expansion convolution, and weighting and fusing the two kinds of features through a self-attention mechanism; in order to improve prediction reliability, training a plurality of basic models with different initializations and architectures, and applying a Bayesian model average technology to obtain a point estimate value, a prediction interval and an uncertainty estimate; based on the prediction model, constructing a graph structure representation of a kiln process parameter space, calculating a predicted coal consumption value of each parameter combination, thereby identifying an optimal process parameter combination, and realizing kiln energy efficiency optimization. The application improves coal consumption prediction accuracy and process parameter optimization efficiency.

Description

Technical Field

[0001] This invention relates to the field of industrial kiln energy consumption optimization technology, specifically to a rotary kiln energy consumption optimization method and system that combines visual recognition, which is particularly suitable for the energy consumption optimization of rotary kilns in the phosphogypsum acid production process. Background Technology

[0002] In the industrial production of acid from phosphogypsum, the rotary kiln is a core piece of equipment, and optimizing its energy consumption is crucial for reducing production costs, improving resource utilization efficiency, and minimizing environmental pollution. Energy consumption control of the rotary kiln involves the coordinated optimization of multiple parameters, requiring precise understanding of the firing conditions within the kiln and adjustment of process parameters accordingly.

[0003] In existing technologies, rotary kiln energy consumption optimization mainly employs empirical rule-based control systems or simple single-model prediction methods. For example, temperature sensors are installed to monitor the temperature distribution inside the kiln, and a fuzzy control model is established by combining parameters such as fuel flow rate and kiln speed; or a single neural network model is used to analyze historical production data, predict energy consumption levels, and provide adjustment suggestions.

[0004] A more advanced approach employs multi-source data fusion technology, combining kiln head flame image data with process parameter data for analysis. This method extracts flame features, such as color, shape, and brightness, using common computer vision algorithms, and then inputs them along with time-series process data into a predictive model to generate energy consumption prediction results and optimize control parameters. This method represents an improvement over traditional techniques and enables a certain degree of refined management of kiln operating conditions.

[0005] However, this method has two main technical drawbacks: First, traditional computer vision algorithms are not accurate enough in extracting the complex and ever-changing flame features of the kiln head, and cannot effectively identify and classify the flame features under different firing conditions; second, due to the complex nonlinear characteristics and multi-parameter coupling relationship of the kiln system itself, the existing prediction model is difficult to efficiently process and analyze a large amount of heterogeneous data, resulting in insufficient accuracy of energy consumption optimization results and failure to achieve true closed-loop intelligent control. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for optimizing the energy consumption of rotary kilns by combining visual recognition. By applying the isolated maximum faction algorithm to the flame feature extraction of the kiln head and combining it with a time-series energy consumption prediction model optimized by dual norms, the invention achieves accurate analysis of the kiln's operating status and effective optimization of energy consumption, thus solving the problems of inaccurate flame feature extraction and insufficient energy consumption prediction accuracy in the prior art.

[0007] To achieve the above objectives, the present invention provides a method for optimizing the energy consumption of a rotary kiln by incorporating visual recognition, comprising:

[0008] Real-time video data and multi-dimensional process parameter data are collected during the operation of the kiln. The real-time video data and the multi-dimensional process parameter data are preprocessed to obtain a preprocessed kiln operation dataset. Based on the flame video data in the preprocessed kiln operation dataset, an image segmentation algorithm is applied to segment the flame region and extract flame features to obtain the kiln firing condition classification result. Based on the process parameter data in the preprocessed kiln operation dataset and the kiln firing condition classification results, a time series prediction model is constructed and trained to obtain a trained time series prediction model. The process parameter data in the preprocessed kiln operation dataset is then input into the trained time series prediction model for calculation to obtain the predicted unit coal consumption of the kiln. Based on the range of process parameter data in the preprocessed kiln operation dataset, a kiln process parameter space is constructed in conjunction with production constraints. The trained time-series prediction model is used to calculate the predicted coal consumption value of candidate parameter combinations in the process parameter space. An optimization algorithm is applied to search for the optimal solution with the minimum predicted coal consumption value as the objective, and the optimal process parameter combination is obtained. Based on the optimal combination of process parameters, parameter adjustment instructions are generated and sent to the control unit for adjustment. The kiln operation status after adjustment is monitored in real time to obtain the actual kiln unit coal consumption result. The actual kiln unit coal consumption result is compared and verified with the kiln unit coal consumption prediction result and fed back to the trained time series prediction model for model update, forming a closed-loop intelligent control system.

[0009] Further, the preprocessing of the real-time video data and the multi-dimensional process parameter data to obtain the preprocessed kiln operation dataset includes: Based on the real-time video data and the multi-dimensional process parameter data, a timestamp alignment algorithm is used to achieve time synchronization of data from different sources. The data is then integrated through a multi-level data fusion framework to generate a data stream in a unified format. Based on the unified format data stream, an adaptive anomaly detection algorithm is applied to identify outliers, missing values, and noise in the unified format data stream. Robust statistical methods are used to replace the outliers, temporal interpolation techniques are used to fill in the missing values, and wavelet transform is used to filter the noise, resulting in a cleaned dataset. Based on the cleaned dataset, parameters of different dimensions in the cleaned dataset are unified to the same numerical range. Based on domain knowledge, derived features are generated from the cleaned dataset to obtain the preprocessed kiln operation dataset.

[0010] Furthermore, the application of the adaptive anomaly detection algorithm to identify outliers, missing values, and noise in the uniformly formatted data stream includes: Based on the unified format data stream, combined with statistical methods and machine learning techniques, outliers, missing values, and noise in the unified format data stream are identified to obtain anomaly data identification results.

[0011] Furthermore, the application of the image segmentation algorithm to segment the flame region and extract flame features includes: Based on the flame video data in the preprocessed kiln operation dataset, the flame video data is enhanced by applying adaptive histogram equalization technology, noise is suppressed by applying nonlocal mean filtering algorithm, and adaptive brightness correction is achieved to obtain the enhanced flame image. Based on the enhanced flame image, a semantic segmentation network based on the DeepLabV3+ architecture is used to segment the flame region in the enhanced flame image, and the fusion result of the flame region mask and the original image is obtained.

[0012] Furthermore, the application of the image segmentation algorithm to segment the flame region and extract flame features includes: Based on the fusion result of the flame region mask and the original image, a superpixel segmentation algorithm is applied to divide the flame region in the fusion result of the flame region mask and the original image into multiple sub-regions. Each sub-region is defined as a node in a graph structure to obtain a node set. Extract the color features, texture features, position features, and dynamic features of each node in the node set, calculate the similarity score between any two nodes in the node set, and establish an edge between any two nodes when the similarity score exceeds a preset threshold to obtain a graph structure representation of the flame region.

[0013] Furthermore, the applied image segmentation algorithm segments the flame region and extracts flame features to obtain the kiln firing condition classification result, including: Based on the graph structure representation of the flame region, the Tomita recursive algorithm framework is applied, and a pruning heuristic method including edge density threshold pruning, average external connectivity pruning, feature variance analysis, and importance scoring is introduced to identify the largest faction in the graph structure representation of the flame region where there is an edge between any two nodes within the faction, the ratio of the number of edges within the faction to the number of nodes in the faction is greater than a first preset threshold, and the ratio of the number of edges between nodes within the faction and nodes outside the faction to the total number of nodes within the faction is less than a second preset threshold, thus obtaining the set of isolated largest factions; Based on the graph structure representation of the isolated maximum faction set and the flame region, the topological structure features and node attribute features of each faction in the isolated maximum faction set are extracted, and the feature vector of each faction in the isolated maximum faction set is constructed to obtain the faction feature vector set; Based on the set of faction feature vectors and the graph structure representation of the flame region, a graph neural network model is constructed. A graph convolutional layer is used to aggregate neighborhood information of the node features in the graph structure representation of the flame region. The set of faction feature vectors is weighted through an attention mechanism. After mapping through a fully connected layer, the probability distribution of the working condition category is obtained, and the classification result of the kiln firing working condition is obtained.

[0014] Further, the construction and training of the time series prediction model to obtain the trained time series prediction model includes: Based on the process parameter data in the preprocessed kiln operation dataset and the kiln firing condition classification results, the process parameter data in the preprocessed kiln operation dataset is encoded using a temporal convolutional network, and the kiln firing condition classification results are encoded using an embedding layer and a bidirectional long short-term memory network to obtain process parameter feature vectors and condition classification feature vectors. Based on the process parameter feature vector and the operating condition classification feature vector, a cross-modal attention mechanism is applied to interactively fuse the process parameter feature vector and the operating condition classification feature vector. The updated process parameter feature vector and operating condition classification feature vector are then fused through a multilayer perceptron to obtain a unified fused feature representation. Based on the unified fusion feature representation, a time series prediction model is constructed to obtain the initialized time series prediction model; Based on the initialized time series prediction model and training data, L1 regularization and L2 regularization terms are added to the loss function. An online load balancing strategy is used to dynamically adjust the L1 regularization coefficient corresponding to the L1 regularization term and the L2 regularization coefficient corresponding to the L2 regularization term. The initialized time series prediction model is trained until convergence, and the trained time series prediction model is obtained.

[0015] Furthermore, the temporal prediction model employs a hybrid architecture of long short-term memory networks and temporal convolutional networks, including: Based on the unified fusion feature representation, long-term dependencies in the unified fusion feature representation are processed by a stacked long short-term memory network, and local temporal patterns in the unified fusion feature representation are processed by a multi-layer dilated convolutional structure, resulting in the output features of the long short-term memory network and the output features of the temporal convolutional network. Based on the output features of the Long Short-Term Memory Network (LSTM) and the output features of the Temporal Convolutional Network (TCNN), a temporal smoothness index and a temporal volatility index of the LSM output features are calculated. A first weighting coefficient and a second weighting coefficient are calculated based on the temporal smoothness index and the temporal volatility index using a self-attention mechanism. The LSM output features and the TCNN output features are then weighted and fused to obtain a combined temporal feature representation.

[0016] Further, the process parameter data from the preprocessed kiln operation dataset is input into the trained time-series prediction model for calculation to obtain the predicted unit coal consumption of the kiln, including: Based on the completed time series prediction model, multiple base models are trained using different initializations and different architectures to obtain multiple base prediction models. Based on the multiple basic prediction models and the process parameter data in the preprocessed kiln operation dataset, the process parameter data in the preprocessed kiln operation dataset is input into the multiple basic prediction models for prediction calculation. The prediction results of the multiple basic prediction models are combined using the Bayesian model averaging technique to obtain the point estimate, prediction interval, and uncertainty estimate of coal consumption. The point estimate, prediction interval, and uncertainty estimate are used to construct the prediction result of the unit coal consumption of the kiln.

[0017] Further, the step of constructing a kiln process parameter space based on the value range of process parameter data in the preprocessed kiln operation dataset and combining it with production constraints, and then applying the trained time-series prediction model to calculate the predicted coal consumption value of candidate parameter combinations in the process parameter space, includes: Based on the value range of process parameter data in the preprocessed kiln operation dataset and the production constraints, a set of key process parameters is determined, a value range and adjustment step size are defined for each parameter in the set of key process parameters, and the constraint relationship between parameters in the set of key process parameters is established to obtain the kiln process parameter space. Based on the kiln process parameter space, each parameter combination that satisfies the constraints in the kiln process parameter space is taken as a node in the graph. Edges are established between node pairs that satisfy the conditions of adjacency, smoothness and constraint compatibility to obtain the graph structure representation of the kiln process parameter space. Based on the graph structure representation of the kiln process parameter space and the trained time-series prediction model, the parameter combination corresponding to each node in the graph structure representation of the kiln process parameter space is input into the trained time-series prediction model for prediction calculation to obtain the predicted coal consumption value corresponding to each node, and the predicted coal consumption value is used as the weight of each node.

[0018] Furthermore, the application optimization algorithm searches for the optimal solution with the goal of minimizing the predicted coal consumption value, and obtains the optimal combination of process parameters, including: Based on the graph structure representation of the kiln process parameter space and the current process parameter value, taking the node corresponding to the current process parameter value as the starting point, an adaptive branching strategy is adopted to select the branch variable and branching order to obtain a set of candidate branches; Based on the candidate branch set and the predicted coal consumption value, a multi-level bounding strategy is applied to estimate the lower bound of energy consumption, perform reachability analysis and constraint propagation on the candidate branch set, and prune the branches in the candidate branch set whose predicted coal consumption value is higher than the current optimal solution to obtain the retained branch set. Based on the set of retained branches, a dynamic constraint strategy is introduced to adjust the bundle width. The number of nodes retained in each layer is dynamically controlled according to the search state. The optimal solution set is continuously updated with the goal of minimizing the predicted coal consumption value, thereby obtaining the optimal combination of process parameters.

[0019] Furthermore, the step of generating parameter adjustment instructions based on the optimal combination of process parameters and sending them to the control unit for adjustment includes: Based on the optimal combination of process parameters and the current process parameter value, find the optimal path from the node corresponding to the current process parameter value to the node corresponding to the optimal combination of process parameters in the graph structure representation of the kiln process parameter space, and design a timing adjustment curve for each process parameter in the optimal path to obtain the parameter adjustment scheme; Based on the parameter adjustment scheme, the parameter adjustment scheme is verified by a thermal process simulation model to check the transient response and parameter cross-influence of the parameter adjustment scheme, and the verified parameter adjustment scheme is obtained. Based on the verified parameter adjustment scheme, the parameter adjustment instruction is generated, and the parameter adjustment instruction is encoded into an industrial control protocol format to obtain an executable parameter adjustment instruction.

[0020] The present invention also provides a rotary kiln energy consumption optimization system incorporating visual recognition, comprising: The data acquisition and preprocessing module is used to acquire real-time video data and multi-dimensional process parameter data during the kiln operation process, and to preprocess the real-time video data and the multi-dimensional process parameter data to obtain a preprocessed kiln operation dataset. The working condition identification module is used to segment the flame area and extract flame features based on the flame video data in the preprocessed kiln operation dataset, and obtain the kiln firing working condition classification result. The model training and prediction module is used to construct and train a time series prediction model based on the process parameter data in the preprocessed kiln operation dataset and the kiln firing condition classification results, to obtain the trained time series prediction model, and input the process parameter data in the preprocessed kiln operation dataset into the trained time series prediction model for calculation to obtain the predicted unit coal consumption of the kiln. The parameter space construction and optimization module is used to construct the kiln process parameter space based on the value range of the process parameter data in the preprocessed kiln operation dataset and the production constraints, apply the trained time series prediction model to calculate the predicted coal consumption value of the candidate parameter combination in the process parameter space, and apply the optimization algorithm to search for the optimal solution with the minimum predicted coal consumption value to obtain the optimal process parameter combination. The closed-loop control module is used to generate parameter adjustment instructions based on the optimal combination of process parameters and send them to the control unit for adjustment. It monitors the kiln operation status after adjustment in real time, obtains the actual kiln unit coal consumption result, compares and verifies the actual kiln unit coal consumption result with the predicted kiln unit coal consumption result, and feeds it back to the trained time series prediction model for model update, thus forming a closed-loop intelligent control system.

[0021] The beneficial effects of this invention are: The isolated maximum faction algorithm was innovatively applied to kiln head flame feature extraction. By representing flame features through graph structures and identifying key patterns, it solved the problem that traditional computer vision methods could not accurately extract complex flame features. The dual-normalization optimization technique was applied to train the time-series energy consumption prediction model, and the regularization weights were dynamically adjusted to improve the accuracy of energy consumption prediction. The kiln parameter optimization problem was transformed into the maximum low-diameter dense subgraph problem and solved by applying an improved branch and bound algorithm, realizing multi-parameter collaborative optimization. A complete closed-loop control system architecture was designed to achieve full-process automation from data acquisition and feature extraction to parameter optimization, forming a continuously self-optimizing intelligent control system. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of a rotary kiln energy consumption optimization method combined with visual recognition provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of a rotary kiln energy consumption optimization system that combines visual recognition, provided in an embodiment of the present invention. Detailed Implementation

[0024] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0025] Example 1: like Figure 1 As shown, this invention provides a method for optimizing the energy consumption of a rotary kiln by incorporating visual recognition, comprising: Step S1: Collect real-time video data and multi-dimensional process parameter data during the kiln operation process, and preprocess the real-time video data and the multi-dimensional process parameter data to obtain the preprocessed kiln operation dataset. Step S2: Based on the flame video data in the preprocessed kiln operation dataset, apply an image segmentation algorithm to segment the flame region and extract flame features to obtain the kiln firing condition classification result. Step S3: Based on the process parameter data in the preprocessed kiln operation dataset and the kiln firing condition classification results, construct and train a time series prediction model to obtain a trained time series prediction model. Input the process parameter data in the preprocessed kiln operation dataset into the trained time series prediction model for calculation to obtain the predicted unit coal consumption of the kiln. Step S4: Based on the range of values ​​of process parameter data in the preprocessed kiln operation dataset, construct the kiln process parameter space in combination with production constraints, apply the trained time series prediction model to calculate the predicted coal consumption value of candidate parameter combinations in the process parameter space, and apply the optimization algorithm to search for the optimal solution with the minimum predicted coal consumption value to obtain the optimal process parameter combination. Step S5: Based on the optimal combination of process parameters, generate parameter adjustment instructions and send them to the control unit for adjustment. Monitor the kiln operation status after adjustment in real time to obtain the actual kiln unit coal consumption result. Compare and verify the actual kiln unit coal consumption result with the predicted kiln unit coal consumption result and feed it back to the trained time series prediction model for model update, forming a closed-loop intelligent control system.

[0026] Specifically, in step S1, real-time video data and multi-dimensional process parameter data during the kiln operation are first collected through a distributed sensor network and a high-definition camera system. The real-time video data primarily comes from a high-definition camera system installed at the kiln head to capture changes in the flame state. The multi-dimensional process parameter data comes from temperature sensors, pressure sensors, gas concentration sensors, etc., located at key positions in the kiln head, kiln tail, and kiln body. These sensors are connected to the edge computing unit via an industrial-grade data bus to achieve initial data integration.

[0027] Preprocessing the collected raw data is a crucial step in ensuring the accuracy of subsequent analysis. The preprocessing process first applies a timestamp alignment algorithm to synchronize the time series of data from different sources, resolving time discrepancies caused by varying sampling frequencies and transmission delays. The timestamp alignment algorithm is a synchronization technique for heterogeneous data sources. It establishes a unified time reference system by identifying timestamps in each data stream, thereby aligning data with different sampling rates and latency characteristics onto the same timeline.

[0028] The synchronized data stream is then processed using an adaptive anomaly detection algorithm to identify and handle outliers, missing values, and noise. This algorithm combines statistical methods and machine learning techniques, automatically adjusting the detection threshold based on the historical distribution characteristics of the data to adapt to data fluctuations under different operating conditions. For detected outliers, robust statistical methods are used for replacement, such as median filtering or estimations based on local trends. For missing values, temporal interpolation techniques are applied for completion, such as linear interpolation, spline interpolation, or interpolation based on similar temporal patterns. For noise, wavelet transform is used for filtering; wavelet decomposition breaks down the signal into different frequency components, filtering out high-frequency noise before reconstructing the signal and preserving the data's valid information.

[0029] Finally, the cleaned data undergoes standardization to unify parameters with different dimensions to the same numerical range. Standardization is a technique that transforms numerical features to a similar scale; common methods include min-max scaling and Z-score standardization. Simultaneously, feature engineering based on domain knowledge is performed to generate derived features, such as temperature gradients and pressure change rates, enhancing the data's representational capabilities. Through these preprocessing steps, a high-quality kiln operation dataset is obtained, laying the foundation for subsequent operating condition identification and energy consumption prediction.

[0030] In step S2, based on the flame video data in the preprocessed kiln operation dataset, this step applies an advanced image segmentation algorithm to accurately segment the flame region and extract features, thereby identifying the kiln firing conditions. Flame image segmentation refers to the process of separating the flame region from the background, and it is the foundation of flame feature extraction.

[0031] First, the flame image is enhanced by applying adaptive histogram equalization to improve contrast. This technique divides the image into multiple small regions and performs histogram equalization on each region separately. Then, bilinear interpolation is used to reconstruct the overall image, effectively enhancing the contrast between the flame and the background while avoiding excessive noise enhancement. Next, a nonlocal mean filtering algorithm is applied to suppress noise. This algorithm uses a weighted average of regions with similar structures in the image, effectively suppressing noise while preserving the detailed texture of the flame. Simultaneously, adaptive brightness correction is implemented by analyzing the brightness variation trend between consecutive frames and dynamically adjusting the brightness parameters to ensure consistent flame brightness.

[0032] Next, a semantic segmentation network based on the DeepLabV3+ architecture is used to segment the flame region in the enhanced flame image. DeepLabV3+ is an advanced semantic segmentation network architecture that combines dilated convolution and an encoder-decoder structure, enabling it to capture multi-scale contextual information and generate fine segmentation boundaries. In this method, the network is specifically optimized for flame characteristics. A pre-trained backbone network is used to extract multi-scale features, and an Atrous spatial pyramid pooling module is used to capture the multi-scale contextual information of the flame. Finally, the fused result of the flame region mask and the original image is output.

[0033] The segmented flame image is divided into multiple sub-regions using a superpixel segmentation algorithm, with each sub-region serving as a node in a graph structure. Superpixel segmentation is a technique that divides an image into a series of regions with similar visual characteristics. Unlike traditional regular grid segmentation, it can generate physically meaningful local regions based on the flame's color and spatial continuity. Color, texture, position, and dynamic features are extracted for each node. The similarity score between any two nodes is calculated, and when the similarity score exceeds a preset threshold, an edge is established between the nodes, thus constructing a graph representation of the flame region.

[0034] The innovation of this method lies in introducing the isolated maximum clique algorithm to extract key features of flames. In graph theory, a maximum clique refers to a complete subgraph, meaning that any two nodes in the subgraph are connected by an edge, and this completeness cannot be maintained by adding other nodes. Isolated maximum cliques are those maximal cliques with tight internal connections but relatively few external connections; these cliques are more likely to represent the characteristic structure of a flame. Based on the Tomita recursive algorithm framework, four pruning heuristics are introduced: edge density threshold pruning, average external connectivity pruning, feature variance analysis, and importance scoring, to identify isolated maximum cliques in the flame graph structure and obtain a set of isolated maximum cliques.

[0035] Finally, feature vectors for each faction in the isolated maximal faction set are constructed. These feature vectors are then input into a trained graph neural network for processing to obtain the classification results of the kiln firing conditions. Graph neural networks are deep learning models specifically designed for processing graph-structured data. They can effectively capture the relationships between nodes and global topological characteristics, mapping complex graph structure features to operating condition categories, such as "overfired," "moderate," and "underfired," providing crucial input for subsequent energy consumption prediction and optimization.

[0036] In step S3, based on the process parameter data in the preprocessed kiln operation dataset and the kiln firing condition classification results obtained in the previous step, this step constructs and trains a time-series prediction model to achieve accurate prediction of the unit coal consumption of the kiln. A time-series prediction model is a mathematical model that can predict future values ​​based on historical data sequences; in kiln energy consumption optimization, it is used to predict energy consumption performance under different parameter combinations.

[0037] First, a feature fusion network is employed to effectively integrate numerical process parameters and categorized operating condition results. The process parameter data is encoded using a temporal convolutional network (TCNN), a special convolutional neural network structure designed for time-series data, which captures the temporal variation patterns of parameters through one-dimensional convolution operations. The operating condition classification results are encoded using an embedding layer and a bidirectional long short-term memory (LSTM) network. The embedding layer converts the discrete classification results into a dense vector representation, while the bidirectional LSM network simultaneously considers past and future sequence information, capturing the temporal dependencies of operating condition changes.

[0038] Next, a cross-modal attention mechanism is applied to interactively fuse the process parameter feature vector and the operating condition classification feature vector. Cross-modal attention is a technique that establishes associations between different types of features. By calculating the correlation weights between features, process parameters and operating condition classifications can mutually reinforce each other. For example, when the operating condition is displayed as "overheated," more attention will be automatically paid to related parameters such as temperature and airflow; conversely, when certain parameters are abnormal, more attention will be paid to the corresponding operating condition classification results. The updated feature vectors are then fused using a multilayer perceptron to obtain a unified fused feature representation.

[0039] A temporal prediction model is constructed based on a unified fusion feature representation. This method employs a hybrid architecture of Long Short-Term Memory (LSTM) networks and temporal convolutional networks, capable of simultaneously capturing long-term dependencies and local temporal patterns. LSM networks are a special type of recurrent neural network structure that addresses the difficulty of traditional recurrent networks in learning long-term dependencies through gating mechanisms; temporal convolutional networks effectively capture local temporal patterns through multi-layered dilated convolutional structures. The combination of these two networks enables the model to handle both long-term dynamic characteristics in kiln systems (such as changes in furnace lining thermal state) and short-term changes (such as fuel supply adjustments).

[0040] During training, this method innovatively applies a dual-regularization optimization technique, simultaneously incorporating L1 and L2 regularization terms into the loss function and dynamically adjusting the regularization coefficients using an online load balancing strategy. L1 regularization tends to produce a sparse model, which helps to automatically select important features; L2 regularization tends to distribute model complexity, preventing some parameter values ​​from becoming too large. Through the online load balancing strategy, the system can dynamically adjust the weights of the two regularizations according to the training state. In the early stages of training, L2 regularization is favored to encourage the model to learn a smooth parameter distribution, while the proportion of L1 regularization is gradually increased as training progresses, allowing the model to focus on important features, effectively balancing the sparsity and generalization ability of the model.

[0041] After training, the process parameter data from the preprocessed kiln operation dataset is input into the trained time-series prediction model for calculation, yielding the predicted unit coal consumption of the kiln. To improve prediction reliability, the system trains multiple base models using different initializations and architectures, and employs Bayesian model averaging to synthesize the prediction results from multiple models. Bayesian model averaging is a statistical method that calculates model weights based on each model's performance on historical data, and then performs a weighted average of the prediction results from multiple models. This not only provides more accurate point estimates but also gives prediction intervals and uncertainty estimates, providing comprehensive information support for subsequent decision-making.

[0042] In step S4, based on the range of process parameter values ​​in the preprocessed kiln operation dataset and considering production constraints, this step constructs the kiln process parameter space and applies an optimization algorithm to find the optimal parameter combination with the lowest energy consumption. The kiln process parameter space refers to a multi-dimensional space composed of all possible parameter combinations, where each point represents a specific set of parameter settings.

[0043] First, determine the set of key process parameters, including kiln speed, primary air volume, secondary air volume, fuel supply rate, kiln head negative pressure, feeding rate, and cooling air volume—parameters that significantly affect energy consumption. Define a reasonable value range and adjustment step size for each parameter. The value range is determined based on equipment physical limitations, process safety boundaries, and historical production experience; the adjustment step size considers the sensitivity of the parameter and the precision of the executing equipment. For example, the kiln speed adjustment step size is typically ±0.1 r / min, and the air volume adjustment step size is ±50 Nm³ / h. Simultaneously, establish constraints between parameters, including hard constraints (physical or safety limitations that must be strictly met) and soft constraints (constraints that affect the optimization objective but can be appropriately relaxed). Through these definitions and constraints, a structured kiln process parameter space is formed.

[0044] This method innovatively represents the parameter space as a graph structure, where nodes represent parameter combinations that satisfy constraints, and edges represent parameter adjustment operations. Edges are established between node pairs that satisfy adjacency (two parameter combinations differ by at most a few parameters with minimal differences), smoothness (parameter changes do not cause drastic system fluctuations), and constraint compatibility (connected nodes satisfy constraints or have acceptable levels of violation), thus constructing a graph structure representation of the kiln process parameter space.

[0045] Next, the trained time-series prediction model is used to calculate the predicted coal consumption value for each node in the parameter space, and the predicted coal consumption value is used as the weight of the node. In this way, the parameter optimization problem is transformed into an optimal path or subgraph search problem on the graph. Specifically, it is to find a subgraph such that the average energy consumption of the nodes in the subgraph is minimized and the transition between nodes is smooth.

[0046] To efficiently solve this optimization problem, an improved branch and bound algorithm is applied. Branch and bound is an algorithmic framework for solving combinatorial optimization problems. It generates candidate solutions through branching operations and prunes branches that cannot be better than the current optimal solution through bounding operations. The improvement of this method lies in the introduction of an adaptive branching strategy, a multi-level bounding strategy, and a dynamic constraint strategy.

[0047] The adaptive branching strategy dynamically selects branch variables and order based on the impact of parameters on energy consumption and the coupling relationship between parameters, prioritizing the exploration of parameter adjustment directions most likely to reduce energy consumption. The multi-level bounding strategy prunes branches that are unlikely to contain the optimal solution through energy consumption lower bound estimation, reachability analysis, and constraint propagation, significantly reducing the search space. The dynamic binding strategy dynamically adjusts the number of nodes retained at each level based on the search state, concentrating resources on exploring high-quality solution regions and expanding the exploration range when the search gets stuck in a local optimum, balancing search efficiency and solution quality.

[0048] This improved optimization algorithm can find a near-global optimal parameter combination within a reasonable computation time, thus obtaining the optimal process parameter combination and providing clear guidance for the next step of parameter adjustment.

[0049] In step S5, based on the optimal combination of process parameters obtained in the previous step, this step generates parameter adjustment instructions and sends them to the control unit for adjustment, thereby realizing closed-loop intelligent control of kiln energy consumption. Closed-loop control refers to the process by which the system automatically adjusts its control strategy based on feedback information. In this method, it manifests as continuously optimizing the prediction model and parameter settings based on actual energy consumption performance.

[0050] First, the optimal path from the current parameter state to the optimal parameter state is found in the parameter space graph structure, and timing adjustment curves are designed for each process parameter in the path. Optimal path planning considers factors such as path length, energy consumption smoothness, and safety margin, and typically employs an improved A / B algorithm. The algorithm is used to solve the problem. The design of the timing adjustment curve takes into account the equipment response model, process dynamic characteristics, and parameter coupling effects. Piecewise S-curves are usually used to ensure that the adjustment process is smooth and meets physical constraints.

[0051] Next, the parameter adjustment scheme is verified for safety using a thermal process simulation model. The thermal process simulation model is a mathematical model describing heat transfer, material movement, and chemical reactions within the kiln, capable of predicting the impact of parameter changes on the system state. Simulation analysis examines the transient response and parameter cross-influence of the parameter adjustment scheme. When potential safety risks or parameter conflicts are detected, the system automatically adjusts the scheme, such as modifying the adjustment sequence, reducing the adjustment rate, or inserting an intermediate stable point, ensuring the safety of the adjustment process.

[0052] The parameter adjustment scheme, after passing safety verification, is converted into executable parameter adjustment instructions, encoded into an industrial control protocol format (such as OPC UA, Modbus, etc.), and sent to the field control unit for execution. Upon receiving the instructions, the control unit adjusts each actuator, such as the frequency converter, fan, and feeder, according to the specified timing and target values, thus achieving the actual parameter adjustment.

[0053] During parameter adjustment, the kiln's operating status is monitored in real time, including key indicators such as temperature, pressure, and flame morphology. If any abnormal trend is detected, a safety intervention mechanism is immediately activated to pause or reverse the adjustment process, ensuring the system always operates within safe limits. After parameter adjustment is complete, the system continuously monitors the kiln's actual operating status and energy consumption indicators to obtain the actual unit coal consumption of the kiln.

[0054] Finally, the actual unit coal consumption of the kiln is compared and verified with the predicted coal consumption, and the prediction error and actual optimization effect are calculated. This feedback information is used to update the time series prediction model, and the model parameters are adjusted through incremental learning to improve prediction accuracy. At the same time, the system continuously accumulates operational data and optimization experience, and updates the knowledge base, including the mapping relationship between operating conditions, parameters, and energy consumption, anomaly handling strategies, and best practice cases.

[0055] Through this continuous self-improvement mechanism, a closed-loop intelligent control system is formed, which can adapt to long-term changing factors such as changes in raw material properties and equipment aging, continuously improve the energy utilization efficiency of the kiln, and achieve continuous optimization of energy consumption.

[0056] Example 2: In this embodiment, the preprocessing of the real-time video data and the multi-dimensional process parameter data to obtain the preprocessed kiln operation dataset includes: based on the real-time video data and the multi-dimensional process parameter data, using a timestamp alignment algorithm to achieve time-series synchronization of data from different sources, integrating them through a multi-level data fusion framework to generate a data stream in a unified format; according to the unified format data stream, applying an adaptive anomaly detection algorithm to identify outliers, missing values, and noise in the unified format data stream, replacing outliers with robust statistical methods, filling in missing values ​​with temporal interpolation techniques, and filtering noise with wavelet transform to obtain a cleaned dataset; based on the cleaned dataset, unifying parameters of different dimensions in the cleaned dataset to the same numerical range, and generating derived features from the cleaned dataset based on domain knowledge to obtain the preprocessed kiln operation dataset.

[0057] Specifically, firstly, based on the acquired real-time video data and multi-dimensional process parameter data, a timestamp alignment algorithm is used to achieve time synchronization of data from different sources. The timestamp alignment algorithm is a key technology for solving the problem of synchronizing multi-source heterogeneous data. Its core idea is to map events from each data source onto a common time axis based on a unified time reference system. In the specific implementation, a unified time base is first established for all data sources (usually using UTC time or the number of milliseconds since system startup), and then precise timestamp information is appended to each data point. For high-frequency data (such as video data, typically 25-30 frames / second), the system uses downsampling or keyframe extraction techniques to reduce the data volume; for low-frequency data (such as some slowly varying process parameters, which may be sampled at the minute level), interpolation methods are used to estimate the values ​​at intermediate moments. The timestamp alignment process also needs to consider data transmission latency and processing latency, which are compensated for through Network Time Protocol (NTP) or by estimating transmission latency to ensure that data from different sources can be accurately aligned in the time dimension.

[0058] Time-aligned data is integrated using a multi-level data fusion framework to generate a data stream in a unified format. This multi-level data fusion framework is a systematic approach to processing heterogeneous data, typically comprising three levels of fusion: data-level fusion, feature-level fusion, and decision-level fusion. In data-level fusion, different types of raw data are converted into a unified data format, including data type conversion, time-series resampling, and dimension alignment. In feature-level fusion, features from various data types are extracted and combined; for example, combining flame morphology features extracted from video data with temperature sensor data. In decision-level fusion, a comprehensive judgment is made based on the analysis results from different data sources, improving the reliability of system decisions. This multi-level fusion method integrates different types of information, such as video data, temperature data, and pressure data, into a structured, unified data stream, providing a complete data view for subsequent analysis.

[0059] Next, based on the generated unified format data stream, an adaptive anomaly detection algorithm is applied to identify outliers, missing values, and noise. Data quality issues are common in industrial environments, and anomalous data may originate from sensor malfunctions, communication interruptions, or external interference. Effective identification of these anomalous data is crucial to ensuring the reliability of subsequent analysis. The adaptive anomaly detection algorithm can automatically adjust detection parameters based on the historical distribution characteristics of the data and current operating conditions, offering greater flexibility and accuracy compared to fixed threshold methods. After the anomaly detection process is completed, corresponding processing strategies are adopted for different types of data problems: for outliers, robust statistical methods are used for replacement, such as using the median or estimates based on local trends to replace outliers, avoiding the impact of extreme values ​​on subsequent analysis; for missing values, time-series interpolation techniques are used for completion, selecting appropriate interpolation methods based on the temporal continuity characteristics of the data, such as linear interpolation, polynomial interpolation, or interpolation based on similar time-series patterns; for noise, wavelet transform is used for filtering. Wavelet transform can analyze signals at different scales, effectively separating signal and noise components, preserving valuable patterns in the data while removing random disturbances.

[0060] Robust statistical methods are a class of statistical techniques that are insensitive to outliers. Unlike traditional statistics such as the mean, which are easily affected by extreme values, they maintain stability for a small number of outlier observations through specially designed algorithms. Commonly used robust statistical methods include the Huber method, Tukey's double-weighted method, and median absolute deviation (MAD). In this system, robust statistical methods are used to identify outliers and calculate surrogate values. For example, when an abnormally high temperature sensor reading is detected, a robust estimator is used to calculate a reasonable surrogate value based on the parameter's normal fluctuation range and recent trend, instead of simply using the average or deleting the value.

[0061] Temporal interpolation is a specialized method for handling missing values ​​in time series data. It utilizes the temporal continuity of the data to estimate the values ​​at the missing points. Unlike ordinary interpolation methods, temporal interpolation specifically considers the time dependence and seasonality of the data. In kiln parameter processing, appropriate temporal interpolation methods are selected based on the characteristics of different parameters: spline interpolation can be used for parameters that change steadily (such as furnace wall temperature); Fourier interpolation can be used for parameters that change periodically (such as certain pressure fluctuations); and predictive imputation based on ARIMA models can be used for parameters with a clear trend.

[0062] Wavelet transform is a powerful signal processing tool that can simultaneously provide information in both the time and frequency domains, making it particularly suitable for processing non-stationary signals. Compared to the traditional Fourier transform, wavelet transform uses time-localized basis functions (wavelets), which can better capture abrupt changes and local features of the signal. In noise filtering applications, the system decomposes the original signal into wavelet coefficients at multiple scales, performs thresholding (soft or hard thresholding) on ​​the high-frequency coefficients representing noise, and then reconstructs the signal through inverse transform, achieving noise removal while preserving the key features of the signal. For parameters such as kiln temperature and pressure, wavelet transform can effectively remove random fluctuations and retain the true trend of process changes.

[0063] Finally, the cleaned dataset is standardized to generate derived features. Standardization is the process of unifying parameters with different dimensions to the same numerical range, which helps prevent certain features from dominating model training due to their large values ​​when the model processes different features. Commonly used standardization methods include min-max scaling (linearly transforming the data to the [0,1] interval) and Z-score standardization (converting the data to a distribution with a mean of 0 and a standard deviation of 1). In kiln parameter processing, an appropriate standardization method is selected based on the distribution characteristics of different parameters. For example, Z-score method can be used for parameters with an approximately normal distribution, while min-max scaling is used for parameters with clear physical boundaries.

[0064] Generating derived features from cleaned datasets based on domain knowledge is a crucial method for enhancing data expressiveness. Derived features are new features generated by transforming or combining original features, and they typically reflect the underlying physical processes or technological laws more directly. In kiln systems, typical derived features include: temperature gradients (the temperature difference between adjacent measuring points, reflecting heat conduction), parameter change rates (reflecting dynamic characteristics of the process), parameter fluctuation amplitudes (reflecting system stability), and specific parameter ratios (such as the air-fuel ratio, reflecting combustion efficiency). These derived features, combined with kiln process expert knowledge, can capture the complex relationships implicit in the original data, significantly improving the predictive capabilities of subsequent models.

[0065] Through the above series of preprocessing steps, a high-quality kiln operation dataset is finally obtained, providing a reliable data foundation for subsequent operating condition identification and energy consumption prediction. Compared with the original data, the preprocessed dataset has the characteristics of temporal consistency, numerical integrity, low noise level, and rich feature representation, which can effectively improve the performance and reliability of the entire energy consumption optimization system.

[0066] Example 3: In this embodiment, the application of the adaptive anomaly detection algorithm to identify outliers, missing values, and noise in the unified format data stream includes: based on the unified format data stream, combining statistical methods and machine learning techniques to identify outliers, missing values, and noise in the unified format data stream, and obtaining anomaly data identification results.

[0067] Specifically, in industrial data processing, anomaly detection is a crucial step in ensuring data quality, but traditional fixed threshold methods are ill-suited to the highly variable operating conditions of kiln systems. The adaptive anomaly detection algorithm proposed in this embodiment combines statistical methods and machine learning techniques, enabling it to dynamically adjust the detection strategy based on data distribution characteristics, significantly improving the accuracy of anomaly identification.

[0068] The core idea of ​​adaptive anomaly detection algorithms is to treat anomaly detection as a dynamic learning problem, rather than simply applying fixed rules. Based on a unified data stream, it is first necessary to understand the distribution characteristics of normal data. This process combines various techniques from statistics and machine learning, including distribution fitting, density estimation, cluster analysis, and sequence pattern learning. For different types of data and anomaly patterns, the system automatically selects the most suitable detection method and adjusts the detection parameters based on real-time data feedback.

[0069] In terms of statistical methods, two main categories are parametric and non-parametric methods. Parametric methods assume that the data follows a specific distribution (such as a normal distribution) and define the normal range by estimating the distribution parameters. For example, for physical parameters such as temperature and pressure, the mean-standard deviation method can be used to mark values ​​that deviate from the mean by more than a certain multiple of the standard deviation (usually 3σ or an adjusted dynamic threshold) as outliers. Unlike traditional methods, the parameter estimation process in this system considers time windows and operating condition changes, dynamically adjusting the calculation methods of the mean and standard deviation. Non-parametric methods do not assume a specific distribution form but directly learn the distribution characteristics from the data. For example, kernel density estimation (KDE) is used to construct the probability density function of the parameters, and observations falling in low probability density regions (such as below the 0.5% quantile or above the 99.5% quantile) are marked as outliers. The advantage of kernel density estimation is that it can handle multimodal and asymmetric distributions and adapt to the complexity of parameter distributions under different operating conditions of the kiln.

[0070] In machine learning techniques, two main paradigms are employed: supervised learning and unsupervised learning. Supervised learning methods utilize historical labeled data to train classification models, such as random forests, support vector machines, or gradient boosting trees, classifying new data points as normal or anomalous. The advantage of these methods is their ability to learn complex anomaly patterns using expert-labeled anomaly cases, but they require a sufficient number of labeled samples. Unsupervised learning methods, on the other hand, do not rely on pre-labeled data but instead discover anomalies by identifying the inherent structure of the data. For example, the system might use the Isolation Forest algorithm, which randomly partitions the feature space based on decision trees, typically isolating anomalies quickly; or it might use the Local Anomaly Factor (LOF) algorithm, which identifies anomalies by comparing the local density of a data point with its neighbors, making it particularly suitable for discovering local anomalies. In time-series data processing, the temporal characteristics of the data are also considered, using autoregressive models, long short-term memory networks, etc., to predict normal time-series patterns, and identifying observations that significantly deviate from the prediction as anomalies.

[0071] Another key characteristic of adaptive anomaly detection algorithms is their adaptive mechanism, which includes three main aspects: parameter adaptation, method adaptation, and context adaptation. Parameter adaptation refers to the algorithm's ability to dynamically adjust detection parameters based on data characteristics. For example, when process parameters fluctuate significantly, the anomaly threshold is automatically relaxed to avoid excessive false alarms. Method adaptation means the system can automatically select the most suitable detection method based on the data type and anomaly type. For example, time-frequency analysis-based methods are used for parameters with strong periodicity, while correlation structure-based methods are used for parameter sets with strong multivariate correlations. Context adaptation considers the current operating conditions and mode, adjusting the anomaly judgment criteria accordingly. For example, a rapid temperature rise is normal during kiln start-up, but may indicate an anomaly during stable operation.

[0072] For the identification of missing values, two methods are used: timestamp checking and data integrity verification. Timestamp checking identifies missing segments that exceed the normal sampling interval by analyzing the time interval of data points; data integrity verification identifies data missing due to sensor failure or communication interruption by using predefined data quality rules, such as sensor status code checking and communication quality index analysis.

[0073] For noise identification, techniques such as spectrum analysis, wavelet analysis, and signal-to-noise ratio (SNR) evaluation are used. Spectrum analysis decomposes the signal into different frequency components through Fourier transform to identify abnormal high-frequency components; wavelet analysis provides local characteristics in the time and frequency domain, enabling the location of noise occurrence within specific time periods and frequency bands; and SNR evaluation quantifies the noise level by calculating the ratio of signal to noise power, providing a basis for subsequent filtering.

[0074] By comprehensively applying the aforementioned statistical methods and machine learning techniques, the adaptive anomaly detection algorithm can accurately identify outliers, missing values, and noise in the kiln data stream, generating detailed anomaly data labeling results. These labeling results include information such as anomaly type (e.g., peak anomaly, trend anomaly, seasonal anomaly), anomaly severity (e.g., mild, moderate, severe), and reliability score, providing precise guidance for subsequent data processing and ensuring that the quality of the preprocessed dataset meets the requirements for model training and prediction.

[0075] Compared with traditional fixed rule methods, the adaptive anomaly detection algorithm shows significant advantages in complex systems such as kilns with variable operating conditions and coupled parameters. It not only improves the accuracy of anomaly identification but also reduces false alarms and false negatives, providing a solid data quality guarantee for the reliable operation of the entire energy consumption optimization system.

[0076] Example 4: In this embodiment, the application of the image segmentation algorithm to segment the flame region and extract flame features includes: based on the flame video data in the preprocessed kiln operation dataset, applying adaptive histogram equalization technology to enhance the contrast of the flame video data, applying a nonlocal mean filtering algorithm to suppress noise in the flame video data, and implementing adaptive brightness correction on the flame video data to obtain an enhanced flame image; based on the enhanced flame image, using a semantic segmentation network based on the DeepLabV3+ architecture to segment the flame region in the enhanced flame image to obtain the fusion result of the flame region mask and the original image.

[0077] Specifically, firstly, based on the flame video data in the preprocessed kiln operation dataset, a series of image enhancement techniques are applied to improve the recognizability of flame features. The primary image enhancement technique is Adaptive Histogram Equalization (AHE), an improved histogram equalization method that enhances local contrast while avoiding noise amplification. Traditional histogram equalization techniques apply the same transformation to the entire image, often leading to over-enhancement or loss of detail in certain areas. AHE, however, divides the image into multiple small regions, performs histogram equalization independently on each region, and then reconstructs the overall image through bilinear interpolation, achieving local adaptive contrast enhancement. In flame image processing, the image is typically divided into an 8×8 or 16×16 grid. A local histogram is calculated for each grid and equalization is applied. Interpolation is then used to ensure smooth transitions between regions. To further prevent excessive noise enhancement, Contrast-Limited Adaptive Histogram Equalization (CLAHE) is employed. This method suppresses noise amplification by limiting the histogram height, making it particularly suitable for processing flame edges and detailed areas.

[0078] Next, a non-local means filtering algorithm is applied to suppress noise in the flame video data. Non-local means (NLM) filtering is an advanced image denoising technique. Its core idea is to use a weighted average of regions with similar structures in the image, rather than relying solely on the spatial proximity of pixels. Compared to traditional local filters (such as Gaussian filtering and median filtering), NLM can better preserve image details and texture while effectively suppressing noise. NLM is particularly useful in flame image processing because flames typically contain rich texture information, which is crucial for subsequent feature extraction. In the algorithm implementation, firstly, image patches similar to the surrounding area of ​​the current pixel are searched in the image. Then, weights are calculated based on similarity, and a weighted average is performed on the similar regions to generate the denoised pixel value. To adapt to the characteristics of flame images, the search range and similarity calculation parameters are dynamically adjusted according to the flame brightness and texture complexity, preserving more details in high-texture areas and applying stronger filtering to smooth areas.

[0079] The third step is to implement adaptive brightness correction to address the common overexposure or underexposure problems in flame images. The brightness of flames inside a kiln varies drastically, making it difficult for the automatic exposure systems of conventional cameras to adapt, resulting in unstable image quality. Adaptive brightness correction technology analyzes the image's brightness distribution and time-series changes to dynamically adjust brightness parameters, ensuring that key flame features remain well-visible. Specifically, first, the image's brightness histogram is calculated to identify overexposed areas (pixel sets close to the maximum brightness value) and underexposed areas (pixel sets close to the minimum brightness value). Then, based on the proportion and distribution of these areas, a brightness adjustment curve is calculated, typically using a piecewise linear or gamma correction function. For continuous video frames, the temporal trend of brightness changes is also analyzed, and brightness parameters are adjusted through a smooth transition to avoid visual inconsistencies caused by abrupt brightness changes. In cases of extremely uneven flame brightness, High Dynamic Range (HDR) technology may be used, expanding the image's dynamic range through the synthesis of multiple exposures to ensure that details in both bright and dark areas of the flame are simultaneously visible. Through these image enhancement steps, an enhanced flame image is finally obtained, laying the foundation for subsequent accurate segmentation.

[0080] Based on the enhanced flame image, a semantic segmentation network based on the DeepLabV3+ architecture is used to accurately segment the flame region. Semantic segmentation is an important task in computer vision, aiming to classify each pixel in an image into a predefined category. DeepLabV3+ is one of the most advanced semantic segmentation architectures available, combining dilated convolutions and an encoder-decoder structure to effectively capture multi-scale contextual information and generate fine-grained segmentation boundaries.

[0081] The DeepLabV3+ architecture consists of three main parts: a backbone network, an Atrous Spatial Pyramid Pooling (ASPP) module, and a decoder module. The backbone network is responsible for extracting basic image features and typically employs a pre-trained deep convolutional network, such as ResNet or Xception. In the flame segmentation application, the system uses EfficientNet-B4 as the backbone network. This network balances network depth, width, and resolution through a composite scaling method, achieving a good balance between computational efficiency and performance, making it particularly suitable for real-time processing needs in edge computing environments. The ASPP module is a core innovation of DeepLabV3+. It captures multi-scale contextual information by using atrous convolutions with different dilation rates. Atrous convolutions increase the receptive field without increasing the number of parameters or computational complexity by inserting "holes" into the convolution kernel. In flame segmentation, the system uses atrous convolutions with dilation rates of 1, 6, 12, and 18 to effectively capture flame features at different scales, from fine textures to overall morphology. The decoder module is responsible for restoring low-resolution feature maps to the original image resolution, generating refined segmentation results. DeepLabV3+ employs a simple yet effective decoder architecture, combining low-level features and encoder output to recover spatial details through upsampling. In flame segmentation applications, the system features a specially designed boundary refinement module to enhance the decoder's ability to process flame edges and resolve the issue of blurred flame boundaries.

[0082] To adapt to the specific needs of flame segmentation, several optimizations were made to the DeepLabV3+ architecture: First, temporal dimension information was introduced, using multiple consecutive frames as input to capture the temporal dynamic characteristics of the flame through 3D convolution or temporal convolution; second, a special loss function was designed, such as a combination of weighted cross-entropy and boundary-aware loss, to pay more attention to the accuracy of flame boundaries; finally, a transfer learning strategy was adopted, using a synthetic flame dataset for pre-training, and then fine-tuning on actual kiln flame data to solve the problem of insufficient labeled data.

[0083] During training, a hybrid data strategy was employed, combining real labeled data, pseudo-labels generated by expert rules, and synthetic data generated by the physical model. Training data augmentation techniques included random cropping, rotation, brightness variation, and contrast adjustment to enhance the model's adaptability to different flame morphologies and imaging conditions. Training adopted a phased strategy: first, the backbone network was frozen to train other modules, then each network was fine-tuned using the Adam optimizer with cosine annealing, an initial learning rate of 0.0001, and a batch size of 8.

[0084] This specially designed deep learning architecture enables accurate segmentation of flame regions, maintaining stable segmentation performance even in complex industrial environments with varying flame shapes. The segmentation results include a flame region mask (a binary image indicating whether each pixel belongs to a flame region) and a fusion result with the original image (used for visualization and subsequent feature extraction). This precise flame region segmentation lays a solid foundation for subsequent graph structure representation and feature extraction, and is a key step in achieving accurate kiln condition identification.

[0085] Compared with traditional segmentation methods based on thresholds or edge detection, the deep learning method based on DeepLabV3+ has significant advantages in processing semi-transparent and irregularly shaped flames, with segmentation accuracy typically improved by 20-30%. It is particularly stable under complex backgrounds and changing lighting conditions, providing high-quality input for subsequent flame feature extraction and working condition classification.

[0086] Example 5: In this embodiment, the application of the image segmentation algorithm to segment the flame region and extract flame features includes: based on the fusion result of the flame region mask and the original image, applying a superpixel segmentation algorithm to divide the flame region in the fusion result of the flame region mask and the original image into multiple sub-regions, defining each sub-region as a node in a graph structure to obtain a node set; extracting the color features, texture features, position features, and dynamic features of each node in the node set, calculating the similarity score between any two nodes in the node set, and establishing an edge between any two nodes when the similarity score exceeds a preset threshold to obtain a graph structure representation of the flame region.

[0087] Specifically, firstly, based on the fusion result of the flame region mask and the original image, a superpixel segmentation algorithm is applied to divide the flame region into multiple sub-regions. Superpixel segmentation is an image preprocessing technique that divides an image into multiple small regions (superpixels) with similar visual characteristics. Each superpixel typically consists of tens to hundreds of neighboring pixels. Compared to directly processing pixels, superpixel segmentation can significantly reduce the computational complexity of subsequent processing while preserving important structural information of the image. Superpixels are particularly useful in flame analysis because they can naturally divide different burning areas of a flame based on color, brightness, and texture, such as the flame root, flame core, and flame tail.

[0088] This system employs an improved Simple Linear Iterative Clustering (SLIC) algorithm for superpixel segmentation. SLIC is a superpixel generation algorithm based on K-means clustering, which performs local clustering in a five-dimensional space (the three channels of the color space plus the x and y coordinates of the pixel) to generate superpixels of similar size. For flame characteristics, the system incorporates three key improvements: First, it uses the LAB color space instead of the RGB space because LAB better aligns with human visual perception and can better distinguish subtle color differences in flames; second, it introduces an adaptive compactness factor, using smaller compactness constraints in the flame edge region, allowing superpixels to adapt more flexibly to irregular boundaries; and finally, it integrates temporal information, considering pixel changes over multiple consecutive frames, ensuring temporal consistency in the generated superpixels and reducing random flickering.

[0089] During segmentation, the number of superpixels is automatically adjusted based on the size and complexity of the flame, typically between 100 and 300. More superpixels are used for large or complex flames to capture details, while fewer are used for small or simple flames to improve processing efficiency. The size and shape of the superpixels are adaptively adjusted, generating smaller superpixels in areas rich in flame detail (such as the combustion front) and larger superpixels in relatively uniform areas (such as the flame center). Furthermore, boundary optimization techniques are applied, adjusting the superpixel boundaries through an energy minimization method to better align them with the natural boundaries and structural lines within the flame.

[0090] After superpixel segmentation, each sub-region is defined as a node in a graph structure, forming a set of nodes. A graph structure is a mathematical model composed of nodes and edges, capable of describing the relationships and connection patterns between elements. In flame representation, the advantage of a graph structure lies in its ability to flexibly express the topological relationships within the flame's interior regions, unrestricted by regular grids, and more closely resembling the natural structure of a flame. Each node (i.e., a superpixel) represents a local region of the flame, containing information about its location, shape, and visual characteristics.

[0091] Next, multidimensional features of each node in the node set are extracted, including color features, texture features, positional features, and dynamic features. Color features are the most basic and important features in flame analysis, directly reflecting the combustion state and temperature distribution. Color statistics are extracted in both HSV and LAB color spaces, including the mean, standard deviation, skewness, and kurtosis of each channel, as well as hue histogram features. Flame colors show significant differences under different combustion states: high-temperature flames of complete combustion are bluish-white, medium-temperature flames are yellow, and low-temperature or incomplete combustion flames are orange-red. By analyzing these color patterns, the system can preliminarily determine the local combustion state.

[0092] Texture features describe the surface structure and internal patterns of the flame, reflecting the degree of turbulence and combustion stability. The system uses various texture descriptors to extract these features, including gray-level co-occurrence matrix (GLCM) features (such as energy, contrast, correlation, and entropy), local binary mode (LBP) features, and Gabor filter responses. Under normal combustion conditions, the flame typically exhibits a uniform texture and smooth edges; however, under abnormal conditions, such as unstable fuel supply or uneven air distribution, the flame texture becomes rough and irregular, with drastic edge fluctuations.

[0093] Positional features record the spatial location and geometric characteristics of nodes within the overall flame structure, including center coordinates, area, perimeter, eccentricity, and orientation. These features are crucial for understanding the overall morphology and structure of the flame. For example, the root, middle, and tail of a flame typically exhibit different combustion characteristics, which can be distinguished by positional features. Furthermore, positional features reflect the degree of flame deflection and symmetry, important indicators for assessing air distribution uniformity and combustion stability.

[0094] Dynamic features describe the changes in flame over time and are key information for distinguishing different operating conditions. The system extracts dynamic features by analyzing the motion trajectory, brightness changes, and shape deformation of nodes in multiple consecutive frames. The main dynamic features include optical flow vector statistics (describing the local motion direction and velocity), temporal brightness fluctuation amplitude and frequency, and shape deformation rate. These features enable the identification of flame oscillation modes, pulsation frequencies, and propagation characteristics, all of which are closely related to the kiln's operating state.

[0095] After extracting all node features, a similarity score is calculated between any two nodes to establish edges in the graph structure. The similarity calculation employs a multi-feature weighted fusion method, comprehensively considering the distance between nodes across various feature dimensions. Specifically, for any two nodes i and j, their standardized distances in color space, texture space, position space, and dynamic feature space are calculated. These distances are then weighted and summed to obtain a comprehensive distance, which is then converted into a similarity score using a kernel function (usually a Gaussian or exponential kernel). The weight parameters are optimized and determined using offline validation data, reflecting the importance of different features in a specific kiln environment.

[0096] The similarity is calculated using the weighted cosine similarity method, and the formula is as follows: Sim(i,j) = ∑(w k × cos(f ik , f jk )) Among them, f ik and f jk These are the k-th eigenvectors of nodes i and j, respectively, w kThis is the weight coefficient for that feature dimension. When the similarity between two nodes exceeds a preset threshold (usually set to 0.7-0.8), an edge is established between the two nodes, and the weight of the edge is proportional to the similarity value.

[0097] When the similarity score exceeds a preset threshold, an edge is established between the two corresponding nodes, indicating that the two flame regions have similar visual or physical characteristics, may belong to the same combustion mode, or be affected by the same factors. The threshold is set using an adaptive method, dynamically adjusted according to the overall graph connectivity requirements, typically maintaining the average degree of the graph (average number of edges per node) between 4 and 8, ensuring sufficient connectivity while avoiding excessive density.

[0098] Through the above steps, a graph structure representation of the flame region is finally obtained. This representation method not only preserves the local features of the flame but also captures the topological relationships between regions, providing an ideal data structure for subsequent factional analysis and condition identification. Compared with traditional pixel-based or global feature-based methods, graph structure representation can more effectively capture the hierarchical structure and local variation patterns of the flame, significantly improving the accuracy and robustness of condition identification.

[0099] Example 6: In this embodiment, the application of image segmentation algorithm to segment the flame region and extract flame features to obtain the kiln firing condition classification result includes: based on the graph structure representation of the flame region, applying the Tomita recursive algorithm framework, and introducing a pruning heuristic method including edge density threshold pruning, average external connectivity pruning, feature variance analysis, and importance scoring, to identify the largest faction in the graph structure representation of the flame region where there is an edge between any two nodes within the faction, the ratio of the number of edges within the faction to the number of nodes in the faction is greater than a first preset threshold, and the ratio of the number of edges between nodes within the faction and nodes outside the faction to the total number of nodes within the faction is less than a second preset threshold, thus obtaining the largest isolated faction. A set of factions is formed; based on the isolated maximum faction set and the graph structure representation of the flame region, the topological structure features and node attribute features of each faction in the isolated maximum faction set are extracted, and the feature vector of each faction in the isolated maximum faction set is constructed to obtain a set of faction feature vectors; based on the set of faction feature vectors and the graph structure representation of the flame region, a graph neural network model is constructed, and a graph convolutional layer is used to aggregate the neighborhood information of the node features in the graph structure representation of the flame region. The set of faction feature vectors is weighted through an attention mechanism, and the working condition category probability distribution is obtained through a fully connected layer mapping to obtain the kiln firing working condition classification result.

[0100] Specifically, firstly, based on the graph structure representation of the flame region, the Tomita recursive algorithm framework is applied to identify the largest isolated clique set. In this step, the graph structure representation of the flame region refers to the graphical representation transformed after segmenting the flame image, where each pixel block or region is considered a node in the graph, and the relationships between nodes (such as spatial adjacency, brightness similarity, etc.) are represented as edges. The Tomita recursive algorithm is a classic clique enumeration algorithm that can effectively identify all cliques in a graph (a clique is a fully connected subgraph, i.e., there is a direct connection between any two nodes). However, in practical applications, the original Tomita algorithm may generate a large number of unnecessary small cliques. Therefore, this system introduces several pruning heuristics for optimization.

[0101] Specifically, the isolation degree of faction C is defined as follows: I(C) = |E(C)| / |E(C,V\C)|, Where |E(C)| is the number of edges within a faction, and |E(C,V\C)| is the number of edges connecting the faction to external nodes. An isolation threshold τ is set, typically between 2 and 3, to retain only factions that satisfy I(C)≥τ.

[0102] The first preset threshold is usually set between 0.7 and 0.9 to ensure high connectivity within the faction; the second preset threshold is usually set between 0.2 and 0.3 to ensure relative independence of the faction.

[0103] Specifically, the first step involves edge density threshold pruning, which calculates the edge density (the ratio of the number of edges within a faction to the theoretical maximum number of edges) within each candidate faction, retaining only those factions with edge densities exceeding a specific threshold. This ensures tight connections between nodes within a faction. Next, average external connectivity pruning is applied, calculating the average number of connections between nodes within and outside a faction, and eliminating factions with excessive external connections. This step helps identify relatively independent flame region structures in the graph. Subsequently, feature variance analysis is used to evaluate the consistency of node features (such as brightness, color, and texture) within each faction, retaining factions with high feature similarity. Finally, each candidate faction is scored for importance, considering faction size, internal connectivity tightness, external connectivity looseness, and feature consistency, selecting the set of factions with the highest scores.

[0104] Through the above pruning process, the largest cluster set satisfying two key conditions is finally identified: first, there must be an edge between any two nodes within the cluster, and the ratio of the number of edges within the cluster to the number of nodes in the cluster is greater than a first preset threshold (usually set between 0.7 and 0.9), ensuring high connectivity within the cluster; second, the ratio of the number of edges between nodes within and outside the cluster to the total number of nodes within the cluster is less than a second preset threshold (usually set between 0.2 and 0.3), ensuring the relative independence of the cluster. These largest clusters actually represent regions with similar characteristics in the flame, such as the core high-temperature zone, the complete combustion zone, and the mixing zone, which are of great significance for judging the kiln operating conditions.

[0105] Next, based on the graph structure representation of the isolated largest faction set and the flame region, feature vectors for each faction are extracted. This step involves the extraction of two types of key features: topological features and node attribute features. Topological features describe the connection patterns within and outside the faction, including faction size (number of nodes), density (ratio of actual edges to maximum possible edges), clustering coefficient (describing the degree of connection between node neighbors), centrality indices (such as degree centrality, proximity centrality, betweenness centrality, etc., used to identify key nodes in the faction), and the connection patterns between the faction and other factions. Node attribute features describe the flame characteristics of nodes within the faction, including brightness distribution (statistics such as mean, variance, skewness, kurtosis, etc.), color features (statistics of each channel in RGB or HSV space), texture features (features based on gray-level co-occurrence matrix or wavelet transform), and temporal variation features (if continuous frame data is available).

[0106] By calculating the aforementioned features for each faction and combining them into a high-dimensional feature vector, a set of faction feature vectors is constructed. These feature vectors not only contain the visual features of the flame region but also the topological relationship information between regions, providing rich input for subsequent condition classification. In practical applications, the system typically standardizes these features to eliminate the influence of different feature dimensions and may use techniques such as principal component analysis to reduce dimensionality, retaining the most discriminative feature combinations.

[0107] Finally, based on the faction feature vector set and the graph structure representation of the flame region, a graph neural network model is constructed to classify the kiln firing conditions. Graph neural networks are a type of deep learning model specifically designed for processing graph structure data, capable of simultaneously utilizing node features and graph topology for learning and inference. In this embodiment, a multi-layer graph convolutional network is first used to aggregate neighborhood information of node features in the graph structure of the flame region. Unlike traditional convolutional layers, graph convolutional layers do not rely on a regular grid structure but rather on feature propagation and aggregation based on node adjacency relationships. Specifically, each node collects feature information from its direct neighbors, combines it with its own features, and transforms it using a learnable weight matrix to generate a new node representation. By stacking multiple layers of graph convolutions, nodes can gradually acquire information from a wider neighborhood, forming a comprehensive feature representation that includes both local structure and global context.

[0108] Subsequently, the set of factional feature vectors is weighted using an attention mechanism. This attention mechanism allows the model to dynamically focus on the importance of different factions, especially those key regions that have a decisive impact on the condition judgment. In implementation, an attention score is calculated for each factional feature vector, reflecting the importance of that faction to the current condition judgment. Then, the factional features are weighted and summed based on these scores to obtain a fused global representation. This attention-based weighting method enables the model to adaptively adjust the contributions of different factions, improving classification accuracy and robustness.

[0109] The graph attention layer implements a graph attention mechanism based on the Transformer architecture, and the calculation formula is as follows: Attention(Ci, Cj) = softmax(Q(fi)·K(fj) T / d)·V(fj) Here, Q, K, and V are the query, key, and value transformation matrices, d is the feature dimension, and fi and fj are the feature vectors of factions Ci and Cj, respectively. This enables the network to capture long-distance dependencies and importance differences between factions.

[0110] It should be noted that the training configuration for the graph neural network model is as follows: Node embedding layer: A multilayer perceptron (MLP) is used to convert the feature vector of each faction into a fixed-dimensional (usually 128 or 256-dimensional) embedding representation; Graph attention layer: Stack 2-3 layers, each followed by a normalization and residual connection layer; Loss function: Cross-entropy loss function is used; Optimizer: Adam optimizer, learning rate set to 1e -4 ; Number of training rounds: 500-1000 rounds; Data augmentation: random rotation, scaling, brightness variation, etc.; Regularization: Set the Dropout rate to 0.3-0.5 to prevent overfitting; Finally, the weighted features are mapped to the operating condition category space through a fully connected layer to obtain the probability distribution of the operating condition categories, i.e., the classification result of the kiln firing conditions. This mapping process may include one or more fully connected layers, each followed by a nonlinear activation function (such as ReLU) and possible regularization operations (such as Dropout). Finally, the output is converted into a probability distribution through the Softmax function. The classification results may include, but are not limited to, different operating condition categories such as "normal combustion", "incomplete combustion", "excessive combustion", "uneven fire", and "unstable flame". Each category corresponds to a probability value, reflecting the system's confidence in the current operating condition judgment.

[0111] Through the three key steps described above, highly discriminative features can be extracted from kiln flame images, and the powerful representation learning capabilities of graph neural networks can be used to accurately classify the kiln firing conditions. This method, based on graph structure and factional analysis, fully utilizes the spatial structural information and visual features of the flame region, resulting in higher accuracy and better interpretability compared to traditional methods based on purely visual features.

[0112] Example 7: In this embodiment, the construction and training of the time-series prediction model to obtain the trained time-series prediction model includes: based on the process parameter data in the preprocessed kiln operation dataset and the kiln firing condition classification results, using a temporal convolutional network to encode the process parameter data in the preprocessed kiln operation dataset, and using an embedding layer and a bidirectional long short-term memory network to encode the kiln firing condition classification results, to obtain process parameter feature vectors and condition classification feature vectors; and applying a cross-modal attention mechanism to process the process parameter feature vectors and condition classification feature vectors according to the process parameter feature vectors and condition classification feature vectors. The process parameter feature vector and the operating condition classification feature vector are fused and updated through a multilayer perceptron to obtain a unified fused feature representation. Based on the unified fused feature representation, a time series prediction model is constructed to obtain an initialized time series prediction model. Based on the initialized time series prediction model and training data, L1 regularization and L2 regularization terms are added to the loss function. An online load balancing strategy is used to dynamically adjust the L1 regularization coefficient corresponding to the L1 regularization term and the L2 regularization coefficient corresponding to the L2 regularization term. The initialized time series prediction model is trained until convergence to obtain the trained time series prediction model.

[0113] Specifically, firstly, based on the process parameter data and kiln firing condition classification results in the preprocessed kiln operation dataset, different encoding methods are used to process these two types of heterogeneous data. For the process parameter data, a Temporal Convolutional Network (TCN) is used for encoding. A TCN is a deep learning architecture specifically designed for time series data, its core feature being the combination of causal convolution and dilated convolution. Causal convolution ensures that the model does not use future information for prediction, maintaining the causality of the time prediction task; dilated convolution, by inserting "holes" into the convolution kernel, enables the network to significantly expand its receptive field without increasing the number of parameters, effectively capturing long-range dependencies.

[0114] The TCN network structure comprises multiple residual blocks, each containing two one-dimensional convolutional layers, layer normalization, weight regularization, ReLU activation function, and skip connections. The dilation coefficient increases exponentially with the number of layers (e.g., 1, 2, 4, 8...), allowing shallow layers to focus on short-term patterns and deeper layers to capture long-term trends. In kiln parameter encoding, the system typically uses 3-5 residual blocks with an input window length of 24-48 time steps (corresponding to 6-12 hours of data), effectively capturing both short-term dynamics and medium-to-long-term trends in parameter changes.

[0115] Compared to traditional recurrent neural networks (such as LSTM and GRU), TCN has several advantages: high parallel computing efficiency and fast training speed; its hierarchical receptive field design enables it to capture multi-scale temporal patterns simultaneously; and fixed convolutional windows and causal constraints reduce the risk of overfitting. These characteristics make TCN particularly suitable for processing data such as kiln process parameters that contain multi-scale dynamics and complex temporal patterns.

[0116] For the classification results of kiln firing conditions, an embedding layer and a bidirectional long short-term memory network (BiLSTM) are used for encoding. The classification results are essentially discrete categorical variables, which need to be converted into continuous vector representations before they can be effectively integrated with other features. The embedding layer is a standard method for handling categorical variables; it maps each category to a dense vector of fixed dimensions. These vectors are automatically learned during training and can capture the semantic relationships between categories. For example, the distance between "overfired" and "slightly overfired" states in the embedding space may be smaller than the distance between "overfired" and "underfired," reflecting their physical similarity. Typically, 32-64 dimensional embedding vectors are used, which can fully express categorical information while avoiding the computational burden of excessively high dimensionality.

[0117] The embedded work condition sequence is further encoded with temporal dependencies using a BiLSTM. The BiLSTM consists of two LSTMs, a forward LSTM and a backward LSTM, capable of simultaneously considering past and future contextual information. LSTM is an improvement on recurrent neural networks, addressing the difficulty of traditional RNNs in learning long-term dependencies by introducing gating mechanisms (input gate, forget gate, and output gate). The forward LSTM processes the sequence from past to present, capturing how historical work conditions affect the current state; the backward LSTM processes the sequence from future to present, capturing the inverse indication of how subsequent work condition changes affect the current state. The outputs of both are concatenated to form a complete contextual representation.

[0118] BiLSTM is particularly well-suited for processing operating condition sequences because kiln operating conditions typically exhibit distinct transition patterns and evolutionary patterns. For instance, the transition from "moderate" to "over-fired" conditions is usually accompanied by a series of predictable intermediate states. Bidirectional processing can capture these patterns, providing richer contextual information. The system typically uses 2-3 layers of BiLSTM with hidden layer dimensions of 128-256, which is sufficient to capture the complex temporal patterns of operating condition sequences.

[0119] Through TCN and BiLSTM encoding, process parameter feature vectors and operating condition classification feature vectors are obtained, respectively. These two types of features come from different sources and express different aspects of information, requiring effective fusion to comprehensively represent the kiln state. Therefore, a cross-modal attention mechanism is applied for feature fusion. Cross-modal attention is a technique that can learn the relationships between different feature types. Its core idea is to calculate the correlation weights between different modal features using a learnable attention matrix, enabling features to mutually reinforce each other.

[0120] In practice, the attention score matrix between the process parameter feature vectors and the operating condition classification feature vectors is first calculated. For each process parameter feature vector, its matching degree with all operating condition classification feature vectors is calculated to obtain the attention distribution; similarly, for each operating condition classification feature vector, its matching degree with all process parameter feature vectors is calculated. These attention scores are calculated using a scaled dot product attention mechanism, i.e., the dot product of two feature vectors is divided by a scaling factor (usually the square root of the feature dimension), and then normalized using softmax.

[0121] After obtaining attention scores, these scores are used as weights to perform weighted aggregation of features. Process parameter features are enhanced by focusing on relevant operating condition information, while operating condition features are supplemented by focusing on relevant parameter information. For example, when the operating condition is displayed as "overheated," more attention is automatically paid to key parameters such as temperature and air volume; and when these parameters fluctuate abnormally, more attention is paid to the changes in operating conditions at the corresponding time points. This two-way enhancement mechanism allows the two types of features to complement each other, forming a more comprehensive state representation.

[0122] The fused features are further integrated using a multilayer perceptron (MLP) to obtain a unified fused feature representation. An MLP is a fully connected feedforward neural network capable of learning nonlinear combinational relationships between features. The system typically uses 2-3 layers of MLP, with each layer followed by batch normalization and ReLU activation functions. The output dimension of the last layer is typically 128-256, serving as a unified representation of the kiln state.

[0123] Based on a unified fusion feature representation, a time series prediction model is constructed, resulting in an initialized time series prediction model. This model employs a multi-layer feedforward network structure, mapping the fusion features to the target variable (unit coal consumption). The output layer uses a linear activation function, suitable for regression tasks. Initialization uses Xavier or He initialization to ensure stable propagation of gradients across layers. This initialized model provides the basic framework for prediction, but its parameters require optimization through effective training strategies.

[0124] During the model training phase, an innovative approach was adopted: simultaneously incorporating L1 and L2 regularization terms into the loss function, and dynamically adjusting the regularization coefficients using an online load balancing strategy. Regularization is a crucial technique for preventing overfitting. By adding penalty terms to the loss function, it limits the size and complexity of model parameters. L1 regularization (Lasso) adds the sum of the absolute values ​​of the parameters as a penalty term, tending to produce sparse models and automatically performing feature selection, compressing the weights of unimportant features to zero. L2 regularization (Ridge) adds the sum of squared parameters as a penalty term, tending to produce models with evenly distributed weights, reducing extreme weight values, and improving the model's generalization ability.

[0125] Traditional methods typically employ only one regularization technique or use fixed-weight ElasticNet (a linear combination of L1 and L2 regularization). The innovation of this system lies in its online load balancing strategy, dynamically adjusting the L1 and L2 regularization coefficients based on the training process. In the early stages of training, the system favors a larger L2 regularization coefficient and a smaller L1 regularization coefficient, encouraging the model to learn a smooth parameter distribution. As training progresses, the weight of the L1 regularization coefficient is gradually increased, prompting the model to begin feature selection, focusing on the most relevant features. This dynamic adjustment strategy combines the advantages of both regularization methods, ensuring model stability while improving feature utilization efficiency.

[0126] In practice, the system evaluates the model's performance and overfitting trend on the validation set at each training cycle and adjusts the regularization coefficient accordingly. When signs of overfitting are detected (such as a continuous decrease in training loss while the validation loss begins to rise), the regularization strength is increased; when the model's learning stagnates, the regularization strength is appropriately decreased to allow the model to further explore the parameter space. Furthermore, the system monitors the distribution of feature importance; when certain features are found to dominate the model's output, the L1 regularization ratio is increased to promote a more balanced use of features.

[0127] The training process employs mini-batch gradient descent with the Adam optimizer, an initial learning rate of 0.001, and batch sizes ranging from 64 to 128. Learning rate scheduling utilizes cosine annealing, gradually reducing the learning rate as training progresses and adding periodic learning rate restarts later to help the model escape local optima. To prevent overfitting, the system also implements an early stopping strategy, terminating training when the validation loss no longer improves after several consecutive epochs.

[0128] Furthermore, the specific implementation of the cross-modal attention mechanism includes the following steps: First, calculate the attention weight of operating condition characteristics on process parameter characteristics: A cp = softmax(F c ·W cp ·F p T ) Calculate the attention weight of process parameter features on operating condition features: A pc = softmax(F p ·W pc ·F c T ) Among them, F c It is a characteristic of the operating condition, F p It is a characteristic of process parameters, W cp and W pc It is a learnable parameter matrix.

[0129] Then, the respective features are updated based on the attention weights: F c' = F c + A cp ·F p F p' = F p + A pc ·F c Finally, the updated features are fused using a multilayer perceptron: F fused = MLP([F c' ; F p']), where [;] represents the feature splicing operation.

[0130] This training method, combining feature encoding, cross-modal fusion, and dual-normative optimization, ultimately yields a trained time-series prediction model. Compared to traditional single-model prediction methods, this approach demonstrates significant advantages in handling multi-source heterogeneous data, capturing complex temporal dependencies, and preventing overfitting. In practical applications, this method can control the average relative error of kiln unit coal consumption prediction within 3-5%, providing a reliable decision-making basis for subsequent parameter optimization.

[0131] Example 8: In this embodiment, the temporal prediction model adopts a hybrid architecture of Long Short-Term Memory (LSTM) network and Temporal Convolutional Network (TCNN), including: based on the unified fused feature representation, processing long-term dependencies in the unified fused feature representation through a stacked LSM network, and processing local temporal patterns in the unified fused feature representation through a multi-layer dilated convolutional structure to obtain LSM network output features and TCNN output features; based on the LSM network output features and the TCNN output features, calculating the temporal smoothness index of the LSM network output features and the temporal volatility index of the TCNN output features; calculating a first weight coefficient and a second weight coefficient based on the temporal smoothness index and the temporal volatility index through a self-attention mechanism, and performing weighted fusion of the LSM network output features and the TCNN output features to obtain the combined temporal feature representation.

[0132] Specifically, kiln energy consumption prediction faces two major challenges: on the one hand, it needs to capture long-term dependencies, such as changes in the thermal state of the furnace lining and seasonal effects; on the other hand, it needs to respond to short-term changing patterns, such as fuel supply adjustments and air volume fluctuations. A single model cannot meet both of these needs simultaneously. Therefore, this embodiment proposes a hybrid architecture of Long Short-Term Memory Network (LSTM) and Temporal Convolutional Network (TCN), which adaptively integrates the advantages of the two models to achieve comprehensive and accurate energy consumption prediction.

[0133] First, based on a unified fusion feature representation, long-term dependencies are processed through a cascaded Long Short-Term Memory (LSTM) network. LSTM is an improved variant of recurrent neural networks, specifically designed to address the vanishing gradient problem in long sequence learning. The core innovation of LSTM lies in the introduction of memory cells and three gating mechanisms: the input gate controls the extent to which new information enters the memory cell, the forget gate controls the extent to which historical information is retained, and the output gate controls the degree to which the information from the memory cell influences the current output. This design enables LSTM to selectively remember important information and forget irrelevant information in long sequences, thereby effectively capturing long-distance dependencies.

[0134] In kiln energy consumption forecasting, long-term dependence manifests in several aspects: First, the gradual change in equipment condition, such as refractory brick aging and ring formation, which are slow but continuously affect efficiency; second, batch differences in raw material characteristics, which may affect the production process for days or even weeks; and third, seasonal factors, such as the impact of changes in ambient temperature and humidity on the kiln's thermal balance. These factors constitute long-term memory elements that cannot be ignored in energy consumption forecasting.

[0135] Stacked LSTM refers to a multi-layered LSTM structure, where the output of each layer serves as the input to the next. This particular system employs a 3-layer LSTM architecture, with each layer containing 256 hidden units. This multi-layered structure enables the network to learn more abstract and higher-level temporal feature representations; the first layer focuses on basic patterns, while higher layers capture complex long-term dependencies. To enhance long-term memory capabilities, the system introduces three improvements over the standard LSTM: First, residual connections are applied, allowing information to jump directly from lower to higher layers, alleviating the training difficulty of deep networks; second, peephole connections are added, enabling gating units to access the state of memory units, enhancing state monitoring capabilities; and finally, an attention-enhanced memory update mechanism is employed, dynamically adjusting the memory update ratio based on the correlation between the current input and historical states.

[0136] Simultaneously, local temporal patterns are processed through a multi-layered dilated convolutional structure. Temporal convolutional networks are convolutional architectures specifically designed for temporal data, with their key innovation lying in the application of dilated convolution. Traditional convolution slides continuously along the temporal dimension, only capturing relationships between immediately adjacent time steps. Dilated convolution, however, inserts "holes" into the convolutional kernel, causing the receptive field to grow exponentially, enabling efficient capture of patterns at different time scales while maintaining a reasonable level of parameter count and computational complexity.

[0137] In kiln systems, local time-series patterns are also rich and diverse: short-term adjustments to process parameters (such as changes in air volume and fuel rate) immediately affect combustion efficiency; periodic fluctuations in batch feeding create hourly fluctuation patterns; shift changes may lead to minor changes in parameter settings, forming specific time-period patterns. These short- to medium-term patterns are crucial for predicting energy consumption over the next few hours or day.

[0138] The designed TCN comprises six dilated convolutional layers with dilation rates of 1, 2, 4, 8, 16, and 32, enabling the network to effectively cover time ranges from minutes to hours. Each convolutional layer uses 128 filters with a kernel size of 3 to ensure sufficient feature extraction capability. Each dilated convolutional layer is followed by batch normalization and parameterized ReLU activation functions to improve training stability and expressive power. To enhance local pattern recognition capabilities, the system introduces a multi-resolution analysis framework, adding skip connections between different convolutional layers, allowing the network to simultaneously process features at different time scales and comprehensively consider them in the final prediction.

[0139] Through the two parallel processing branches described above, the output features of LSTM and TCN are obtained, respectively. LSTM features excel at capturing stable, long-term trends, while TCN features are better at expressing local fluctuations and short-term patterns. To optimally combine the advantages of both features, their weights need to be dynamically adjusted based on the characteristics of the current data.

[0140] First, the temporal smoothness index of the LSTM output features and the temporal volatility index of the TCN output features are calculated. The temporal smoothness index quantifies the stationarity of the feature sequence and is evaluated by calculating the mean square value of the first difference or the autocorrelation coefficient of the sequence. A sequence with high smoothness indicates that the LSTM has effectively captured a stable trend and should be given higher weight in the fusion. The temporal volatility index quantifies the intensity of local variations in the feature sequence and is evaluated by calculating the energy ratio of high-frequency components or the local rate of change. A sequence with high volatility indicates that the TCN has successfully identified important short-term patterns and should be given higher weight in the corresponding time period.

[0141] Based on these two metrics, fusion weights are dynamically calculated using a self-attention mechanism. Self-attention is a technique that adaptively allocates attention based on data content and is widely used in sequence models. Specifically, LSTM and TCN features are concatenated and linearly transformed to obtain a query vector, a key vector, and a value vector. The similarity between the query and the key is calculated, and then normalized using softmax to obtain an attention score. These scores reflect the importance of different features in the current prediction task and are used to weight and aggregate the value vectors, achieving dynamic and selective feature fusion.

[0142] In this system, the self-attention mechanism is extended to conditional self-attention, incorporating temporal smoothness and volatility indices as conditional information into the attention calculation process. Specifically, the system constructs a smoothness-volatility feature matrix, where each time step contains the smoothness index of the LSTM feature and the volatility index of the TCN feature. This matrix, after a nonlinear transformation, is multiplied by the original attention score to adjust the attention allocation. For example, when the energy consumption curve is detected to enter a stable phase, the system increases the weight of the LSTM feature; conversely, when sudden changes or periodic fluctuations are detected, the weight of the TCN feature is increased.

[0143] This dynamic weighted fusion mechanism yields a combined temporal feature representation that retains both the long-term dependencies captured by LSTM and the local temporal patterns identified by TCN. The final prediction layer is typically a fully connected network that maps the fused features to predicted values ​​of coal consumption per unit of kiln.

[0144] A key advantage of the hybrid architecture is its adaptability and robustness. When the kiln enters a stable operating phase, the LSTM branch contributes more, ensuring long-term predictive consistency. When kiln parameters are adjusted or external conditions change, the TCN branch can quickly respond to these changes and adjust the predictions accordingly. This dynamic balance allows the model to adapt to different stages and states of kiln operation, maintaining high-accuracy predictive capabilities.

[0145] In practical applications, the LSTM-TCN hybrid architecture improves accuracy by approximately 25-30% in capturing short-term fluctuations compared to a single LSTM model, and improves stability by approximately 15-20% in long-term trend prediction compared to a single TCN model. For responding to sudden events, the hybrid model can adjust predictions within 3-5 time steps, which is 2-3 times faster than a single model. These performance improvements directly translate into energy consumption optimization benefits, enabling the system to more accurately predict energy consumption performance under different parameter combinations, providing a reliable basis for subsequent optimization decisions.

[0146] Example 9: In this embodiment, the step of inputting the process parameter data from the preprocessed kiln operation dataset into the trained time-series prediction model for calculation to obtain the kiln unit coal consumption prediction result includes: training multiple basic models using different initializations and different architectures based on the trained time-series prediction model to obtain multiple basic prediction models; inputting the process parameter data from the preprocessed kiln operation dataset into the multiple basic prediction models for prediction calculation according to the multiple basic prediction models and the process parameter data from the preprocessed kiln operation dataset; using Bayesian model averaging to synthesize the prediction results of the multiple basic prediction models to obtain a point estimate, prediction interval, and uncertainty estimate of coal consumption; and using the point estimate, the prediction interval, and the uncertainty estimate to construct the kiln unit coal consumption prediction result.

[0147] Specifically, firstly, during the model training phase, multiple base models with different initialization parameters and structural variations are created based on the already trained time-series prediction model architecture. Specifically, the system retains the core structure of the original time-series prediction model but uses different random seeds for parameter initialization, resulting in a different starting point for each model's training and allowing exploration of different regions of the parameter space. Simultaneously, the system fine-tunes the model architecture, such as adjusting the number of network layers (e.g., using a 2-4 layer LSTM structure), the number of hidden units (e.g., 64, 128, or 256 neurons), the activation function type (e.g., ReLU, Tanh, or Sigmoid), the regularization method (e.g., L1, L2 regularization, or Dropout), and optimizer parameters (e.g., learning rate, momentum value). This diverse model design enables different models to capture different patterns and features in the data, creating a complementary effect. Typically, 8-15 different base prediction models are trained, each using the same training data, but cross-validation is used for evaluation and selection, retaining the set of models with the best prediction performance.

[0148] During the predictive calculation phase, the process parameters from the preprocessed kiln operation dataset are input into each basic predictive model. These process parameters include, but are not limited to, key parameters such as kiln head temperature, kiln tail temperature, kiln speed, pulverized coal injection rate, primary air volume, secondary air volume, tertiary air volume, and oxygen content within the kiln. Each basic model independently calculates a predicted coal consumption value based on these parameters. Due to the diversity of the models, these predicted values ​​will exhibit different characteristics: some models may perform better under specific operating conditions but worse under others; some models may be more sensitive to changes in certain parameters but less sensitive to changes in others. This diversity allows model ensemble to more comprehensively capture the predictive patterns under different conditions.

[0149] In the results synthesis stage, Bayesian model averaging is used to fuse the results of multiple base prediction models. Bayesian model averaging is an ensemble learning method based on Bayesian theory. It treats each model as an approximation of the true model and performs a weighted average of these approximations using posterior probabilities. Unlike simple averaging, Bayesian model averaging considers the reliability and uncertainty of each model, enabling more intelligent weight allocation. Specifically, first, the performance metrics (such as mean squared error, mean absolute error, etc.) of each base model on the validation set are calculated. Then, based on these metrics and Bayesian theory, the posterior probability of each model is calculated. The better the model performs, the higher its corresponding posterior probability, and the greater its weight in the final prediction. In this way, a point estimate of coal consumption, i.e., the most probable predicted coal consumption value, is obtained.

[0150] In addition to point estimates, prediction intervals and uncertainty estimates are also calculated, which is a significant advantage of Bayesian model averaging. The prediction interval is typically a 95% confidence interval, indicating that the actual coal consumption value has a 95% probability of falling within this interval. By analyzing the distribution characteristics of the prediction results from each model and combining Bayesian statistical methods, the upper and lower bounds of this interval are calculated. The width of the prediction interval reflects the degree of uncertainty in the prediction: the narrower the interval, the more certain the prediction; the wider the interval, the more uncertain the prediction. The uncertainty estimate directly quantifies the reliability of the prediction result; it can be represented by the variance or standard deviation of the prediction distribution. Higher uncertainty means that the model lacks confidence in the current prediction, which may be due to abnormal input data, insufficient model coverage, or encountering unprecedented operating conditions.

[0151] Finally, the point estimates, prediction intervals, and uncertainty estimates are combined to form a complete prediction of the unit coal consumption of the kiln. This multi-dimensional prediction result not only provides specific predicted values ​​but also information on the reliability of the prediction, offering kiln operators and managers a more comprehensive basis for decision-making. For example, when the prediction result shows low coal consumption but high uncertainty, operators may need to adjust process parameters more cautiously; while when the prediction result shows high coal consumption but a narrow prediction interval (high certainty), immediate measures may be needed to reduce coal consumption. This comprehensive prediction method, which combines point prediction and uncertainty, significantly improves the application value of the prediction system in actual production.

[0152] Example 10: In this embodiment, the step of constructing a kiln process parameter space based on the value range of process parameter data in the preprocessed kiln operation dataset and production constraints, and then applying the trained time-series prediction model to calculate the predicted coal consumption value of candidate parameter combinations in the process parameter space, includes: determining a set of key process parameters based on the value range of process parameter data in the preprocessed kiln operation dataset and the production constraints; defining a value range and adjustment step size for each parameter in the set of key process parameters; establishing constraint relationships between parameters in the set of key process parameters; and obtaining the kiln process parameter space. Based on the kiln process parameter space, each parameter combination that satisfies the constraints in the kiln process parameter space is taken as a node in the graph. Edges are established between node pairs that satisfy the conditions of adjacency, smoothness, and constraint compatibility to obtain the graph structure representation of the kiln process parameter space. Based on the graph structure representation of the kiln process parameter space and the trained time series prediction model, the parameter combination corresponding to each node in the graph structure representation of the kiln process parameter space is input into the trained time series prediction model for prediction calculation to obtain the predicted coal consumption value corresponding to each node. The predicted coal consumption value is used as the weight of each node.

[0153] Specifically, the core of kiln energy consumption optimization is finding the optimal parameter combination. However, traditional parameter optimization methods often neglect the complex constraints between parameters and process safety, leading to difficulties in applying the optimization results in practical applications. This embodiment innovatively proposes a parameter space representation method based on graph structure, which not only considers the feasible range of the parameters themselves but also simulates the process characteristics of parameter adjustment, providing a more realistic search space for subsequent optimization algorithms.

[0154] First, based on the range of process parameter values ​​and production constraints in the pre-processed kiln operation dataset, a set of key process parameters is determined. Key process parameters refer to operational variables that significantly impact kiln energy consumption, determined through parameter importance analysis and expert knowledge. For a typical cement kiln system, key process parameters typically include: kiln speed (kiln cylinder rotation speed), primary air volume (air volume injected through the burner), secondary air volume (air volume injected through the kiln head), fuel supply rate, kiln head negative pressure, feeding rate, and cooling air volume. These parameters directly affect combustion efficiency, material transport, and heat distribution, thus influencing unit coal consumption.

[0155] Once the parameters are determined, a reasonable range and adjustment step size are defined for each parameter. The range is determined based on three aspects: physical limitations of the equipment (such as the maximum output capacity of the blower, mechanical limitations of the kiln speed, etc.), process safety boundaries (such as the minimum kiln head negative pressure requirement, the maximum kiln temperature limit, etc.), and historical production experience (such as the normal operating range). For example, the kiln speed is typically in the range of 3.0-4.5 r / min, the fuel supply rate is within ±15% of the base value calculated based on the production capacity, and the kiln head negative pressure is between -5Pa and -30Pa. Determining the range must ensure sufficient adjustment space for optimization while also ensuring the safe and stable operation of the system.

[0156] The adjustment step size refers to the smallest unit of parameter change, directly affecting the discretization accuracy of the parameter space and the search efficiency of the optimization algorithm. Setting the adjustment step size considers two factors: the sensitivity of the parameter and the precision of the executing device. Highly sensitive parameters (such as fuel supply rate) require smaller step sizes for fine-tuning; precision-limited actuators (such as some older dampers) can only achieve larger step sizes. Typical adjustment step size settings include: kiln speed ±0.1 r / min, air volume ±50 Nm³ / h, fuel supply rate ±1%, kiln head negative pressure ±2 Pa, etc. A reasonable step size setting allows the optimization process to explore sufficiently fine parameter combinations while avoiding waste of computational resources.

[0157] Next, establishing the constraint relationships between parameters is a key innovation in parameter space construction. In actual kiln systems, there are complex interdependencies and constraints between parameters. Blindly adjusting a single parameter may lead to system instability or violation of process requirements. Constraint relationships are mainly divided into two categories: hard constraints and soft constraints.

[0158] Hard constraints are physical or safety limitations that must be strictly met; violations will result in equipment damage or process failure. Typical hard constraints include: the fuel-to-air ratio (to ensure complete combustion), the matching relationship between kiln speed and feed rate (to prevent overload or no-load), and the safety boundaries between temperature and kiln speed (to prevent damage to refractory materials). These constraints are usually expressed as inequalities or equations between parameters, such as the air-to-air ratio must be within a specific range: 0.9 ≤ (primary air volume + secondary air volume) (coefficient) / fuel supply rate ≤ 1.1.

[0159] Soft constraints are constraints that affect the optimization objective but can be appropriately relaxed, and are typically related to product quality and operational stability. Typical soft constraints include requirements for feed uniformity, stability of the calcination zone position, and the range of temperature fluctuations at the kiln tail. Soft constraints are usually added to the optimization objective in the form of a penalty function, rather than as hard boundaries. For example, when the feed rate changes by more than 10% / hour, a penalty term is added to the optimization objective, reducing the priority of that option, but not completely eliminating it.

[0160] By defining key parameters, their ranges, adjustment steps, and constraints, the system obtains a structured kiln process parameter space. Traditional methods typically treat the parameter space as a simple Cartesian product, neglecting the process characteristics and safety constraints of parameter adjustment, making the optimization results difficult to implement. This embodiment innovatively represents the parameter space as a graph structure, more realistically simulating the actual operation of the kiln.

[0161] In the graph structure representation, each node represents a feasible combination of parameters in the parameter space (i.e., parameter settings that satisfy all hard constraints). If only discrete values ​​are considered, the theoretical size of the parameter space is the product of the number of possible values ​​for each parameter; for a typical 7-8 key parameters, this could generate millions of combinations. However, in reality, due to constraints, truly feasible parameter combinations usually only occupy a small portion of the theoretical space, perhaps only tens of thousands. The system employs a constraint propagation algorithm to efficiently screen feasible combinations. It first applies the strictest constraints (usually safety-related constraints) to quickly reduce the search space, then gradually applies other constraints, ultimately generating a set of nodes that satisfy all hard constraints.

[0162] Next, edges are established between node pairs that meet specific conditions, forming a graph structure representation of the kiln process parameter space. The edge establishment conditions include three aspects: adjacency, smoothness, and constraint compatibility. Adjacency requires that at most a few parameters differ between two parameter combinations, and these differences should be small. Typically, a maximum of 2-3 parameters are limited to changing simultaneously, and each parameter change should not exceed 1-2 adjustment steps. This condition simulates the gradual adjustment process in actual operation, avoiding drastic parameter jumps. Smoothness requires that parameter changes do not cause drastic system fluctuations, usually determined by evaluating the stability of key derived variables (such as the rate of change of heat load, temperature gradient changes, etc.). When the changes in these derived variables exceed a preset threshold, edge connections are not established even if the original parameters change little. Constraint compatibility requires that all connected nodes satisfy the constraint conditions. Hard constraints must be fully satisfied, while soft constraints are allowed to a certain degree of violation, but the constraint violation degree must be calculated as the cost of the edge.

[0163] These three edge-building rules construct a sparse but connected parameter space graph structure, where nodes represent feasible parameter combinations and edges represent safe and feasible parameter adjustment operations. Compared with traditional parameter space representations, this graph structure representation has three important advantages: First, it directly models the process constraints of parameter adjustment, ensuring the feasibility of the optimized path; second, it encodes the constraints into the graph's topology, simplifying the complexity of subsequent optimization algorithms; and finally, it provides a natural framework, transforming the parameter optimization problem into a path planning or subgraph search problem on the graph, facilitating the application of efficient graph algorithms.

[0164] Based on the graph structure representation of the kiln process parameter space and the trained time-series prediction model, it is necessary to evaluate the energy consumption performance of each parameter combination. The system inputs the parameter combination corresponding to each node in the graph structure into the trained time-series prediction model, predicts the unit coal consumption under that parameter setting, and uses the predicted value as the weight of the node. This step connects the parameter space with the energy consumption prediction model, providing an evaluation criterion for subsequent optimization algorithms.

[0165] The prediction process is not a simple single-point input and output, but rather considers the dynamic impact of parameter adjustments. Specifically, it considers three aspects of dynamic characteristics: First, the time lag effect of parameter adjustments, as the thermal inertia of the kiln system means that parameter changes require a certain amount of time to be fully reflected in energy consumption, typically ranging from 30 minutes to 2 hours; second, the interaction effect of parameters, where changes in some parameters may amplify or weaken the effects of other parameters, requiring the use of time-series models to capture this nonlinear relationship; and finally, state dependence, where the same parameter adjustments may produce different effects under different initial states, requiring the current kiln state to be considered as conditional information for prediction.

[0166] To improve the reliability of predictions, a prediction interval and an uncertainty estimate are generated simultaneously during the evaluation process at each node. The prediction interval provides the possible upper and lower bounds of coal consumption, while the uncertainty estimate quantifies the confidence level of the prediction. For parameter combinations with high prediction uncertainty, the system marks them in the graph structure to remind the optimization algorithm to handle these regions with caution. Simultaneously, the system also identifies extrapolation regions of the model (i.e., parameter combinations whose distribution differs significantly from the training data) and adds corresponding uncertainty assessments to prevent the optimization algorithm from blindly exploring unknown regions.

[0167] By using predicted coal consumption values ​​as node weights, the kiln parameter space is transformed into a weighted graph structure, and the optimization problem becomes finding the node or subgraph with the minimum weight. Compared with traditional direct optimization methods, this graph structure optimization has two important advantages: First, it naturally integrates parameter adjustment process constraints into the optimization framework, ensuring that the obtained optimal parameter combination is practically executable; second, it allows the system to plan the adjustment path from the current state to the optimal state, focusing not only on the final result but also on the safety and smoothness of the transition process.

[0168] In practical applications, the focus is not usually on finding a single optimal node, but rather on identifying "advantageous regions" with lower energy consumption—that is, a set of connected low-energy-consumption nodes in the graph. These regions typically represent a class of similar optimization strategies, providing operators with diverse options and increasing the system's flexibility in dealing with uncertainty. For example, under different external conditions (such as changes in raw material properties or fluctuations in ambient temperature), specific parameter combinations suitable for the current situation can be flexibly selected within the advantageous region, without having to rerun the entire optimization process each time.

[0169] Based on graph-structured parameter space representation and energy consumption assessment, subsequent optimization algorithms can efficiently find the optimal parameter combination for energy consumption while ensuring process safety and stability, providing reliable decision support for intelligent control and energy optimization of the kiln. Compared with traditional optimization methods, this method demonstrates higher feasibility and stability in practical applications, achieving an average additional energy consumption reduction of 3-5%, while reducing parameter adjustment failures or rollback events by more than 70%, significantly improving the actual operating efficiency of the system.

[0170] Example 11: In this embodiment, the application optimization algorithm searches for the optimal solution with the goal of minimizing the predicted coal consumption value to obtain the optimal combination of process parameters. This includes: based on the graph structure representation of the kiln process parameter space and the current process parameter value, using the node corresponding to the current process parameter value as the starting point, employing an adaptive branching strategy to select branch variables and branch order to obtain a candidate branch set; based on the candidate branch set and the predicted coal consumption value, applying a multi-level bounding strategy to estimate the lower bound of energy consumption, perform reachability analysis, and constraint propagation on the candidate branch set, pruning branches in the candidate branch set whose predicted coal consumption value is higher than the current optimal solution to obtain a retained branch set; based on the retained branch set, introducing a dynamic constraint strategy to adjust the constraint width, dynamically controlling the number of nodes retained at each level according to the search state, and continuously updating the optimal solution set with the goal of minimizing the predicted coal consumption value to obtain the optimal combination of process parameters.

[0171] Specifically, kiln energy consumption optimization is a challenging problem with a high-dimensional parameter space, complex constraints, and multi-objective balance. Traditional optimization methods such as gradient descent or evolutionary algorithms often struggle to efficiently handle this complexity. The branch-and-bound algorithm based on graph structure search proposed in this embodiment, combining adaptive branching strategies and multi-level bounding techniques, can significantly improve search efficiency while ensuring solution quality.

[0172] First, based on the graph structure representation of the kiln process parameter space and the current process parameter values, the node corresponding to the current parameter value is selected as the search starting point. This starting point selection strategy has two important considerations: on the one hand, it ensures that the optimization starts from a practically feasible operating condition, avoiding theoretically feasible but practically unattainable solutions; on the other hand, it allows the system to consider the transition path from the current state to the target state, which is crucial for the smooth adjustment of industrial processes. Compared to traditional optimization methods that start from random or fixed initial points, the strategy of starting from the current operating condition is more in line with actual operational needs and can generate more practical optimization solutions.

[0173] Based on the starting point, an adaptive branching strategy is adopted to select branch variables and branch order, generating a candidate branch set. The branching strategy is a core component of the branch and bound algorithm, determining how the search space is partitioned and the efficiency of the search process. Traditional branching strategies typically use a fixed order or simple heuristics to select branch variables, making them difficult to adapt to the complex characteristics of kiln systems. The adaptive branching strategy proposed in this embodiment comprehensively considers three key factors: parameter sensitivity, constraint correlation, and uncertainty distribution.

[0174] Parameter sensitivity analysis assesses the impact of each process parameter on coal consumption. A time-series forecasting model is used to calculate the coal consumption sensitivity index for each parameter, which measures the expected change in coal consumption caused by a unit change in the parameter. Highly sensitive parameters (such as fuel supply rate and main air volume) have a greater impact on the optimization objective and are preferentially used as branch variables, helping to quickly narrow down the search area for high-quality solutions. Sensitivity analysis not only considers the first-order effects of parameters but also evaluates parameter interaction effects through second-order derivative estimation or cross-term analysis, identifying parameter combinations with strong nonlinear relationships.

[0175] Constraint correlation analysis assesses the degree of correlation between process parameters and system constraints. A constraint-parameter correlation matrix is ​​constructed to record the participation and weight of each parameter in each constraint. Parameters with high constraint correlation (such as those related to multiple key constraints) are given priority branching, which can more quickly eliminate infeasible solutions and reduce computational resource waste. For example, the oxygen content at the kiln tail is usually associated with multiple combustion-related constraints, and determining its value range will simultaneously affect the feasible region of multiple parameters such as fuel rate and air volume; therefore, it is often given priority branching.

[0176] Uncertainty distribution analysis identifies uncertain regions in the parameter space of the prediction model. Using variance estimates or confidence intervals of the prediction model, it identifies regions of high uncertainty (e.g., sparse data regions, model extrapolation regions) and regions of low uncertainty (e.g., data-rich regions, well-fitting models). In the branching strategy, it tends to first explore low-uncertainty regions to ensure a reliable optimization solution is found, and then gradually expand to high-uncertainty regions based on search progress, balancing exploration and utilization.

[0177] Based on the three analytical dimensions mentioned above, the branch priority score for each parameter is dynamically calculated, and the branching order is determined according to the score. During the search process, the branching strategy is adaptively adjusted based on information from the explored region (such as discovered constraint conflicts and the changing trend of the objective function). For example, if it is found that a parameter almost always causes constraint violations within a certain interval, the branch priority for that interval will be reduced; if a parameter exhibits a sensitivity different from expected during the search, its branch priority will be adjusted accordingly.

[0178] An adaptive branching strategy is used to generate a set of candidate branches, each representing a sub-region of the parameter space. Theoretically, the complete set of branches covers the entire parameter space reachable from the current state, but directly exploring all branches is computationally too costly. Therefore, a multi-level bounding strategy is applied to filter the candidate branches, pruning branches whose predicted coal consumption values ​​are higher than the current optimal solution, resulting in a set of retained branches.

[0179] The multi-level bounding strategy is another innovation of this embodiment. It quickly eliminates low-quality branches and concentrates computing resources on promising regions through a series of estimation methods with increasing computational complexity and progressively improving accuracy. This strategy includes three main levels: energy consumption lower bound estimation, reachability analysis, and constraint propagation.

[0180] Energy consumption lower bound estimation is the most basic level of bounding, calculating the theoretical lower bound of coal consumption for each branch. The system uses several techniques to calculate this lower bound: linear relaxation constructs a linear programming problem using a first-order approximation of the parameters with respect to coal consumption, and solves for its optimal solution as the lower bound; convex hull approximation constructs a convex hull of the parameter-coal consumption relationship based on historical data, and uses convex optimization to solve for the lower bound; Lagrange relaxation transforms complex constraints into penalty terms, constructing a relaxation problem to solve for the lower bound. These techniques have low computational complexity, enabling rapid assessment of the potential of branches and eliminating branches whose lower bounds have exceeded the current optimal solution.

[0181] Reachability analysis is an intermediate-level bounding function that assesses whether a feasible adjustment path exists from the current state for a given combination of parameters within a branch. In a kiln system, some parameter combinations, while theoretically feasible, may not be safely achievable from the current state due to constraints during the adjustment process (such as adjustment rate limits and process stability requirements). The system uses a path search algorithm in a graph structure to find feasible paths from the current node to the target region in the parameter space graph. If no such path exists, or if the cost of all paths (such as adjustment time and energy consumption fluctuations) is too high, the branch is pruned to avoid generating practically unfeasible optimization solutions.

[0182] Constraint propagation is the finest level of bounding, narrowing the feasible domain of parameters by reasoning and propagating the effects of constraints. The system represents hard constraints as a network of relationships between parameters. Once the range of a parameter is determined, the constraints on other parameters are derived, thus reducing the domain. Constraint propagation is an iterative process, repeatedly applying constraints until all parameter domains are stable or a constraint conflict is found (indicating that no feasible solution exists for that branch). This method can eliminate a large number of infeasible solutions through logical reasoning before actually calculating the complete coal consumption prediction, greatly improving search efficiency.

[0183] By employing a multi-level bounding strategy, the original massive set of candidate branches is reduced to a relatively small set of retained branches. These branches all have low theoretical coal consumption lower bounds, feasible adjustment paths, and satisfy all constraints. Nevertheless, the number of retained branches can still be large, especially in high-dimensional parameter spaces. To further improve search efficiency, a dynamic binding strategy is introduced to adaptively adjust the number of nodes retained at each level during the search process.

[0184] Dynamic binding strategy is a method to balance exploration breadth and computational efficiency. Its core idea is to dynamically adjust the "bundle width" (i.e., the number of nodes retained at each layer) based on the search state. Traditional bundle search uses a fixed bundle width, which may prematurely discard potentially high-quality solutions in complex problems or waste computational resources in simple problems. The dynamic binding strategy of this system adaptively adjusts the bundle width based on three factors: search depth, solution distribution characteristics, and computational resource status.

[0185] Search depth is the primary consideration for beamwidth adjustment. In the early stages of the search (layers near the starting point), the system uses a larger beamwidth to explore various directions; as the search progresses, the beamwidth is gradually reduced, concentrating resources on the most promising regions. This strategy reflects a practical observation: high-quality solutions tend to cluster in specific regions, and as the search deepens, the quality differences between different branches gradually become apparent.

[0186] Solution distribution characteristics are the second consideration. The predicted coal consumption distribution of the current layer nodes is continuously monitored, and the dispersion and skewness of the distribution are calculated. When there are significant differences in solution quality (high dispersion), the system tends to use a small bundle width to quickly focus on high-quality solutions; when the solution quality distribution is uniform (low dispersion), the system tends to maintain a large bundle width to avoid premature convergence to suboptimal solutions. This adaptive adjustment can automatically select an appropriate exploration strategy based on the difficulty characteristics of the problem.

[0187] Computational resource availability is the third consideration. Real-time monitoring of computational resource usage (such as CPU load, memory usage, and computation time) allows for adjusting the beamwidth when resources are limited to ensure the algorithm completes within an acceptable timeframe; conversely, when resources are plentiful, the beamwidth is appropriately increased to improve solution quality. This resource-aware adjustment enables the algorithm to adapt to different computing environments, fully utilizing computing power on high-performance servers and conserving resources on edge devices to ensure response time.

[0188] By employing a dynamic constraint strategy, the optimal solution set can be continuously updated with the goal of minimizing predicted coal consumption by efficiently searching and preserving the branch set. Compared to traditional branch-and-bound or heuristic search, this method demonstrates significant advantages in kiln optimization problems: search speed is increased by 3-5 times, optimization quality is improved by 5-10%, while ensuring the feasibility and stability of the solution.

[0189] Ultimately, the optimal combination of process parameters is output. This combination not only has the lowest predicted coal consumption value but also satisfies all process constraints and adjustment constraints, and can be achieved from the current state through a safe and feasible adjustment path. Typically, it not only provides a single optimal solution but also retains a set of suboptimal solutions, offering operators diverse optimization options and enhancing the system's adaptability to uncertainties and special operating conditions. In practical applications, this optimization method can generate customized energy-saving solutions based on the specific operating characteristics of the kiln, typically achieving a 3-7% reduction in coal consumption while maintaining stable product quality, creating significant economic and environmental benefits for enterprises.

[0190] Example 12: In this embodiment, the step of generating parameter adjustment instructions based on the optimal process parameter combination and issuing them to the control unit for adjustment includes: based on the optimal process parameter combination and the current process parameter value, finding the optimal path from the node corresponding to the current process parameter value to the node corresponding to the optimal process parameter combination in the graph structure representation of the kiln process parameter space; designing a timing adjustment curve for each process parameter in the optimal path to obtain a parameter adjustment scheme; according to the parameter adjustment scheme, performing a safety verification of the parameter adjustment scheme through a thermal process simulation model to check the transient response and parameter cross-influence of the parameter adjustment scheme to obtain a verified parameter adjustment scheme; generating the parameter adjustment instructions based on the verified parameter adjustment scheme, encoding the parameter adjustment instructions into an industrial control protocol format to obtain an executable parameter adjustment instruction.

[0191] Specifically, in a kiln energy consumption optimization system, the process of transforming the theoretical optimal solution into practically executable control commands is crucial and complex, involving path planning adjustments, safety verification, and control protocol conversion. Traditional optimization systems often neglect this transition process, leading to poor implementation of the optimization scheme or causing process fluctuations. The method proposed in this embodiment ensures that the optimization scheme can be implemented safely, smoothly, and effectively through meticulously planned parameter adjustment paths, rigorous safety verification, and standardized command conversion.

[0192] First, based on the optimal combination of process parameters and the current process parameter values, the optimal path from the node corresponding to the current parameter value to the node corresponding to the optimal parameter combination is found in the graph structure representation of the kiln process parameter space. The core of this step lies in path planning, the goal of which is not only to find a feasible path connecting the starting and ending points, but also to select the optimal adjustment sequence from multiple possible paths. In industrial processes, the selection of parameter adjustment paths directly affects the smoothness, safety, and economy of the transition process, and is a key factor in the successful implementation of the optimization scheme.

[0193] The optimal path search uses a modified version of A This algorithm is a heuristic search algorithm capable of efficiently finding near-optimal paths in large-scale graph structures. Traditional A The algorithm uses the sum of the known path cost and the estimated target cost as the evaluation function. This system extends it in three aspects: multi-objective evaluation function, dynamic heuristic function, and adaptive search strategy.

[0194] The multi-objective evaluation function comprehensively considers multiple quality dimensions of the path, including energy cost, adjustment time, process fluctuation, and safety margin. Energy cost is calculated as a weighted sum of predicted coal consumption at each node along the path, typically assigning higher weights to nodes closer to the target to encourage rapid achievement of low energy consumption. Adjustment time reflects the actual operation time required to complete the entire adjustment sequence, considering adjustment rate limitations for different parameters and the response time of executing equipment. Process fluctuation quantifies the amplitude and frequency of fluctuations in key process indicators (such as temperature and pressure) during the adjustment process; paths with low fluctuations are more conducive to product quality stability and equipment lifespan. Safety margin assesses the distance between the path and hazardous conditions (such as high temperature and high pressure safety boundaries), prioritizing paths with sufficient safety redundancy. These objectives often conflict; a weighted summation or Pareto optimization method is used to balance these objectives and construct a comprehensive evaluation function.

[0195] Dynamic heuristics refer to target estimation methods that adaptively adjust based on search progress. Traditional A The algorithm uses a fixed heuristic function to estimate the cost from the current node to the target. However, in complex systems like kiln control, a single heuristic function is insufficient to accurately estimate costs under different conditions. The system employs a multi-level heuristic function library, selecting the most suitable estimation method based on the distance between the current node and the target node, parameter differences, and system state. In regions far from the target, a coarse but computationally efficient straight-line distance or parameter difference is used as an estimate; as the target approaches, a refined estimation method considering constraints and dynamic characteristics is switched to improve path quality; in special regions (such as areas with dense constraints or high uncertainty), specialized heuristic functions are used to guide the search and avoid potential risks.

[0196] Adaptive search strategies are techniques that adjust the search process based on real-time conditions. The system dynamically adjusts the open list sorting rules, node expansion strategies, and memory management schemes to adapt to the characteristics of the search space. For example, when the search is found to be trapped in a local loop, a random perturbation term is added to break the loop; when the search direction deviates significantly from the expected goal, heuristic weights are temporarily adjusted to refocus; when memory usage is close to the limit, node merging or pruning strategies are applied to free up space. These adaptive adjustments enable the algorithm to efficiently handle path planning problems of varying complexity and characteristics.

[0197] Through the above-mentioned enhanced A The algorithm finds the optimal adjustment path from the current state to the optimal parameter combination. This path balances multiple objectives, including energy consumption, adjustment time, process stability, and safety, while ensuring reachability. Once the path is determined, the system designs timing adjustment curves for each process parameter along the path, generating a complete parameter adjustment scheme.

[0198] The timing adjustment curve is a function describing how parameters change over time, determining the speed, rhythm, and smoothness of parameter adjustments. Simple linear adjustments can lead to process shocks and oscillations; therefore, the system employs advanced curve design methods to generate optimal adjustment curves based on parameter characteristics and process dynamics. Different curve templates are used for different types of parameters: slow-response parameters (such as kiln temperature) use smooth S-shaped curves to avoid overshoot and oscillations; fast-response parameters (such as damper position) can use ramp or step adjustments to improve response speed; highly coupled parameters (such as fuel rate and air volume) use coordinated curves to ensure that key ratios remain within a reasonable range during adjustment.

[0199] The S-curve is a commonly used adjustment template, characterized by a slow start, acceleration in the middle, and a slow finish, which can reduce the shock caused by adjustment. Systems typically use the sigmoid function or cubic splines to construct the S-curve, and its smoothness is determined by the parameter response characteristics and process requirements. For slow-response variables such as temperature, a gentler S-curve is used; for medium-response variables such as pressure, a relatively steeper S-curve can be used; and for rare cases such as emergency adjustments, a near-linear adjustment curve can be used to shorten the adjustment time.

[0200] Coordination curves solve the challenge of synchronous adjustment of multiple parameters. In kiln systems, multiple parameters typically need to maintain specific relative relationships, such as the air-fuel ratio and the ratio of kiln speed to feed rate. The system uses constrained optimization methods to design coordination curves, ensuring that the combination of relevant parameters meets process requirements at any adjustment point. For example, when increasing the fuel rate, a matching adjustment curve for the air volume is automatically calculated to keep the air-fuel ratio within a safe range; when increasing the kiln speed, the feed rate adjustment is planned simultaneously to maintain stable material load. This coordinated adjustment significantly reduces the risk of process fluctuations caused by parameter mismatch.

[0201] Curve design also needs to consider the characteristics and limitations of the executing equipment. Different control devices have different response characteristics: electric valves may have dead zones and hysteresis; frequency converters have acceleration limitations; PLC control systems have scan cycles and communication delays, etc. The curve design is adjusted according to the equipment model. For example, for valves with significant dead zones, a slightly larger initial step size is used at the beginning of the adjustment to overcome the dead zone; for drive systems with lag, deceleration is performed in advance when approaching the target to avoid overshoot. These equipment adaptation optimizations ensure that the theoretical curve can be accurately executed by the actual equipment.

[0202] Using the curve design method described above, time-series adjustment curves were generated for each parameter, forming a complete parameter adjustment scheme. However, due to the complexity of the kiln system and the uncertainty of the model, the theoretically designed scheme may still have potential risks. Therefore, the system uses a thermal process simulation model to conduct a comprehensive safety verification of the parameter adjustment scheme, checking transient responses and parameter cross-influences to ensure the practical feasibility of the scheme.

[0203] Thermal process simulation models are a type of digital twin technology that uses mathematical equations to describe the physical and chemical processes of a kiln, enabling the simulation and prediction of system behavior under different operating conditions. Unlike simplified predictive models, simulation models are typically based on detailed physical principles, including mechanistic models of heat transfer, fluid dynamics, chemical reactions, material transport, and other aspects. This allows for more accurate prediction of the dynamic effects of parameter changes, particularly the system response during transient processes.

[0204] The safety verification process consists of three main stages: transient response analysis, parameter cross-influence assessment, and extreme case testing. Transient response analysis simulates the entire parameter adjustment process, calculating the time trajectories of key process indicators (such as temperature, pressure, and oxygen content) to check for any exceeding limits or drastic fluctuations. The system pays particular attention to indicators that may affect equipment safety and product quality, such as temperature gradients, pressure fluctuations, and changes in airflow distribution, ensuring that these indicators remain within reasonable ranges throughout the adjustment process.

[0205] Cross-parameter impact assessment analyzes the interactions between different parameter adjustments. In complex kiln systems, a change in one parameter can affect other parameters through multiple pathways, and these indirect effects are sometimes difficult to predict with simple models. Simulation models can capture these complex cross-influences, such as the impact of fuel adjustments on temperature distribution, leading to changes in material flowability; and airflow adjustments altering airflow patterns, thus affecting the location of the combustion zone. By analyzing these cross-influences, the system can identify potential parameter conflicts or instabilities, allowing for further optimization of adjustment schemes.

[0206] Extreme case testing verifies the robustness of a solution by simulating extreme conditions or adding disturbances. It simulates various possible anomalies, such as device response delays, sensor reading deviations, and sudden changes in external conditions, evaluating the solution's safety margins and corrective capabilities under these circumstances. Only solutions that perform stably in extreme case testing are considered truly safe and reliable; otherwise, it is necessary to return to the path planning stage and redesign a more conservative or robust solution.

[0207] Through comprehensive verification via thermal process simulation, a validated parameter adjustment scheme was obtained, which theoretically can safely and stably achieve the energy consumption optimization target. The next step is to translate the abstract adjustment scheme into concrete control commands, which requires encoding and converting industrial control protocols.

[0208] Based on the validated parameter adjustment scheme, parameter adjustment instructions are generated and encoded into an industrial control protocol format. Industrial control systems typically use specific communication protocols and instruction formats, such as Modbus, PROFIBUS, and OPC UA. These protocols define data encoding methods, transmission rules, and security mechanisms. It is necessary to translate the high-level adjustment scheme into a low-level instruction sequence that conforms to the specific protocol to ensure that the control unit can correctly understand and execute it.

[0209] The instruction generation process is divided into three layers: the scheduling layer, the conversion layer, and the communication layer. The scheduling layer is responsible for discretizing the timing adjustment curve into a series of target values ​​at different time points, and determining the frequency and rhythm of instruction transmission based on the system's control cycle and communication capabilities. For example, rapidly changing parameters may require updates multiple times per second, while slowly changing parameters may only need updates once per minute. The scheduling layer also handles instruction priorities and dependencies, ensuring that the adjustment order of related parameters is reasonable and avoiding unsafe combinations made temporarily.

[0210] The conversion layer converts engineering units of parameters (such as temperature, pressure, and flow rate) into signal units (such as voltage, current, and digital values) used by the control system. This conversion is typically based on calibration formulas or lookup tables, taking into account the characteristics of sensors and actuators. For example, it converts gas flow rate (Nm³ / h) into the opening percentage of a flow control valve, or a temperature setpoint (°C) into a burner control signal (mA). The conversion process also includes signal limiting, significant bit processing, and unit conversion to ensure that the generated instructions are physically executable and of appropriate accuracy.

[0211] The communication layer encodes the converted instructions into data packet formats specific to industrial protocols. This process involves technical details such as address mapping, data encapsulation, checksums, and calculations to ensure that instructions are accurately transmitted to the target device over the industrial network. For different control systems, multiple industrial protocols are supported. For example, connecting to a PLC system typically uses Modbus TCP or Profinet, connecting to a DCS system may use OPC UA or a proprietary protocol, and connecting to an edge controller may use MQTT or RESTAPI. The system automatically selects the appropriate protocol and communication interface, handling protocol conversion and compatibility issues.

[0212] In addition to basic instruction encoding, several security mechanisms are implemented: instruction validity verification ensures that instructions are within a reasonable range and meet process requirements; a communication confirmation mechanism requires the executing device to return to the execution status to verify that the instruction was correctly received and executed; a rollback mechanism provides emergency handling solutions in case of communication failure or execution anomalies; and access control ensures that only authorized systems and users can send instructions to adjust critical parameters. These mechanisms together constitute a security barrier for instruction execution, preventing communication failures or malicious interference from causing system out of control.

[0213] Through the above steps, executable parameter adjustment instructions are generated. These instructions are then sent to the kiln control unit for adjustment, realizing the implementation of the energy consumption optimization scheme. Compared with the traditional method of directly issuing target values, the method in this embodiment considers the safety of the adjustment path, the dynamic response of the process, and the characteristics of the industrial control system, greatly improving the implementation effect and operational safety of the optimization scheme.

[0214] In practical applications, this refined parameter adjustment scheme demonstrates significant advantages: First, through path optimization and curve design, process fluctuations during the transition process are reduced by an average of 40-60%, ensuring stable product quality. Second, through safety verification and protocol adaptation, the command execution success rate is increased to over 98%, significantly higher than the 70-80% success rate of traditional methods. Finally, a comprehensive safety assurance mechanism eliminates the risk of safety accidents during the adjustment process, increasing operators' trust in the automatic optimization system. These improvements enable the kiln energy consumption optimization system not only to generate theoretically feasible optimization schemes but also to safely and efficiently translate them into actual energy consumption reductions and economic benefits.

[0215] Example 13: like Figure 2 As shown, the present invention also provides a rotary kiln energy consumption optimization system incorporating visual recognition, comprising: The data acquisition and preprocessing module 10 is used to acquire real-time video data and multi-dimensional process parameter data during the operation of the kiln, and to preprocess the real-time video data and the multi-dimensional process parameter data to obtain a preprocessed kiln operation dataset. The working condition identification module 20 is used to segment the flame area and extract flame features based on the flame video data in the preprocessed kiln operation dataset, and obtain the kiln firing working condition classification result. The model training and prediction module 30 is used to construct and train a time series prediction model based on the process parameter data in the preprocessed kiln operation dataset and the kiln firing condition classification results, to obtain the trained time series prediction model, and input the process parameter data in the preprocessed kiln operation dataset into the trained time series prediction model for calculation to obtain the predicted unit coal consumption of the kiln. The parameter space construction and optimization module 40 is used to construct the kiln process parameter space based on the value range of the process parameter data in the preprocessed kiln operation dataset and the production constraints, apply the trained time series prediction model to calculate the predicted coal consumption value of the candidate parameter combination in the process parameter space, and apply the optimization algorithm to search for the optimal solution with the minimum predicted coal consumption value to obtain the optimal process parameter combination. The closed-loop control module 50 is used to generate parameter adjustment instructions based on the optimal combination of process parameters and send them to the control unit for adjustment, monitor the kiln operation status after adjustment in real time, obtain the actual kiln unit coal consumption result, compare and verify the actual kiln unit coal consumption result with the kiln unit coal consumption prediction result, and feed it back to the trained time series prediction model for model update, thus forming a closed-loop intelligent control system.

[0216] In this embodiment, the data acquisition and preprocessing module collects kiln operation data through a distributed sensor network and a high-definition camera system. It uses a timestamp alignment algorithm for data synchronization and applies an adaptive anomaly detection algorithm to handle outliers and missing values, ultimately outputting a high-quality preprocessed dataset. The operating condition identification module segments the flame region based on the DeepLabV3+ semantic segmentation network. It uses a superpixel algorithm to divide the flame region into multiple sub-regions and constructs a graph structure representation. An improved Tomita algorithm is applied to extract the feature of isolating the largest clique, and a graph neural network is used to achieve accurate classification of kiln operating conditions. The model training and prediction module adopts a hybrid architecture of temporal convolutional networks and long short-term memory networks. It uses double normalization optimization and online load balancing strategies for model training and integrates multiple basic prediction models through Bayesian model averaging to output high-precision energy consumption prediction results. The parameter space construction and optimization module transforms the parameter optimization problem into a problem of maximizing the lowest diameter dense subgraph on a graph structure. An improved branch and bound algorithm is used to solve this problem, and a dynamic constraint strategy accelerates convergence to obtain the optimal parameter combination. The closed-loop control module is designed to find the optimal adjustment path from the current parameters to the target parameters. It is then verified for safety using a thermal process simulation model. The adjustment commands are encoded into an industrial control protocol format and sent out for execution. The adjustment effect is monitored in real time, forming a self-optimizing closed-loop control system.

[0217] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing the energy consumption of a rotary kiln by combining visual recognition, characterized in that, include: Real-time video data and multi-dimensional process parameter data are collected during the operation of the kiln. The real-time video data and the multi-dimensional process parameter data are preprocessed to obtain a preprocessed kiln operation dataset. Based on the flame video data in the preprocessed kiln operation dataset, an image segmentation algorithm is applied to segment the flame region and extract flame features to obtain the kiln firing condition classification result. Based on the process parameter data in the preprocessed kiln operation dataset and the kiln firing condition classification results, a time series prediction model is constructed and trained to obtain a trained time series prediction model. The process parameter data in the preprocessed kiln operation dataset is then input into the trained time series prediction model for calculation to obtain the predicted unit coal consumption of the kiln. Based on the range of process parameter data in the preprocessed kiln operation dataset, a kiln process parameter space is constructed in conjunction with production constraints. The trained time-series prediction model is used to calculate the predicted coal consumption value of candidate parameter combinations in the process parameter space. An optimization algorithm is applied to search for the optimal solution with the minimum predicted coal consumption value as the objective, and the optimal process parameter combination is obtained. Based on the optimal combination of process parameters, parameter adjustment instructions are generated and sent to the control unit for adjustment. The kiln operation status after adjustment is monitored in real time to obtain the actual kiln unit coal consumption result. The actual kiln unit coal consumption result is compared and verified with the kiln unit coal consumption prediction result and fed back to the trained time series prediction model for model update, forming a closed-loop intelligent control system.

2. The method according to claim 1, characterized in that, The preprocessing of the real-time video data and the multi-dimensional process parameter data to obtain the preprocessed kiln operation dataset includes: Based on the real-time video data and the multi-dimensional process parameter data, a timestamp alignment algorithm is used to achieve time synchronization of data from different sources. The data is then integrated through a multi-level data fusion framework to generate a data stream in a unified format. Based on the unified format data stream, an adaptive anomaly detection algorithm is applied to identify outliers, missing values, and noise in the unified format data stream. Robust statistical methods are used to replace the outliers, temporal interpolation techniques are used to fill in the missing values, and wavelet transform is used to filter the noise, resulting in a cleaned dataset. Based on the cleaned dataset, parameters of different dimensions in the cleaned dataset are unified to the same numerical range. Based on domain knowledge, derived features are generated from the cleaned dataset to obtain the preprocessed kiln operation dataset.

3. The method according to claim 2, characterized in that, The application of an adaptive anomaly detection algorithm to identify outliers, missing values, and noise in the uniformly formatted data stream includes: Based on the unified format data stream, combined with statistical methods and machine learning techniques, outliers, missing values, and noise in the unified format data stream are identified to obtain anomaly data identification results.

4. The method according to claim 1, characterized in that, The applied image segmentation algorithm segments the flame region and extracts flame features, including: Based on the flame video data in the preprocessed kiln operation dataset, the flame video data is enhanced by applying adaptive histogram equalization technology, noise is suppressed by applying nonlocal mean filtering algorithm, and adaptive brightness correction is achieved to obtain the enhanced flame image. Based on the enhanced flame image, a semantic segmentation network based on the DeepLabV3+ architecture is used to segment the flame region in the enhanced flame image, and the fusion result of the flame region mask and the original image is obtained.

5. The method according to claim 4, characterized in that, The applied image segmentation algorithm segments the flame region and extracts flame features, including: Based on the fusion result of the flame region mask and the original image, a superpixel segmentation algorithm is applied to divide the flame region in the fusion result of the flame region mask and the original image into multiple sub-regions. Each sub-region is defined as a node in a graph structure to obtain a node set. Extract the color features, texture features, position features, and dynamic features of each node in the node set, calculate the similarity score between any two nodes in the node set, and establish an edge between any two nodes when the similarity score exceeds a preset threshold to obtain a graph structure representation of the flame region.

6. The method according to claim 5, characterized in that, The applied image segmentation algorithm segments the flame region and extracts flame features to obtain the kiln firing condition classification results, including: Based on the graph structure representation of the flame region, the Tomita recursive algorithm framework is applied, and a pruning heuristic method including edge density threshold pruning, average external connectivity pruning, feature variance analysis, and importance scoring is introduced to identify the largest faction in the graph structure representation of the flame region where there is an edge between any two nodes within the faction, the ratio of the number of edges within the faction to the number of nodes in the faction is greater than a first preset threshold, and the ratio of the number of edges between nodes within the faction and nodes outside the faction to the total number of nodes within the faction is less than a second preset threshold, thus obtaining the set of isolated largest factions; Based on the graph structure representation of the isolated maximum faction set and the flame region, the topological structure features and node attribute features of each faction in the isolated maximum faction set are extracted, and the feature vector of each faction in the isolated maximum faction set is constructed to obtain the faction feature vector set; Based on the set of faction feature vectors and the graph structure representation of the flame region, a graph neural network model is constructed. A graph convolutional layer is used to aggregate neighborhood information of the node features in the graph structure representation of the flame region. The set of faction feature vectors is weighted through an attention mechanism. After mapping through a fully connected layer, the probability distribution of the working condition category is obtained, and the classification result of the kiln firing working condition is obtained.

7. The method according to claim 1, characterized in that, The construction and training of the time series prediction model to obtain the trained time series prediction model includes: Based on the process parameter data in the preprocessed kiln operation dataset and the kiln firing condition classification results, the process parameter data in the preprocessed kiln operation dataset is encoded using a temporal convolutional network, and the kiln firing condition classification results are encoded using an embedding layer and a bidirectional long short-term memory network to obtain process parameter feature vectors and condition classification feature vectors. Based on the process parameter feature vector and the operating condition classification feature vector, a cross-modal attention mechanism is applied to interactively fuse the process parameter feature vector and the operating condition classification feature vector. The updated process parameter feature vector and operating condition classification feature vector are then fused through a multilayer perceptron to obtain a unified fused feature representation. Based on the unified fusion feature representation, a time series prediction model is constructed to obtain the initialized time series prediction model; Based on the initialized time series prediction model and training data, L1 regularization and L2 regularization terms are added to the loss function. An online load balancing strategy is used to dynamically adjust the L1 regularization coefficient corresponding to the L1 regularization term and the L2 regularization coefficient corresponding to the L2 regularization term. The initialized time series prediction model is trained until convergence, and the trained time series prediction model is obtained.

8. The method according to claim 7, characterized in that, The temporal prediction model employs a hybrid architecture of long short-term memory networks and temporal convolutional networks, including: Based on the unified fusion feature representation, long-term dependencies in the unified fusion feature representation are processed by a stacked long short-term memory network, and local temporal patterns in the unified fusion feature representation are processed by a multi-layer dilated convolutional structure, resulting in the output features of the long short-term memory network and the output features of the temporal convolutional network. Based on the output features of the Long Short-Term Memory Network (LSTM) and the output features of the Temporal Convolutional Network (TCNN), a temporal smoothness index and a temporal volatility index of the LSM output features are calculated. A first weighting coefficient and a second weighting coefficient are calculated based on the temporal smoothness index and the temporal volatility index using a self-attention mechanism. The LSM output features and the TCNN output features are then weighted and fused to obtain a combined temporal feature representation.

9. The method according to claim 8, characterized in that, The process parameter data from the preprocessed kiln operation dataset is input into the trained time-series prediction model for calculation to obtain the predicted unit coal consumption of the kiln, including: Based on the completed time series prediction model, multiple base models are trained using different initializations and different architectures to obtain multiple base prediction models. Based on the multiple basic prediction models and the process parameter data in the preprocessed kiln operation dataset, the process parameter data in the preprocessed kiln operation dataset is input into the multiple basic prediction models for prediction calculation. The prediction results of the multiple basic prediction models are combined using the Bayesian model averaging technique to obtain the point estimate, prediction interval, and uncertainty estimate of coal consumption. The point estimate, prediction interval, and uncertainty estimate are used to construct the prediction result of the unit coal consumption of the kiln.

10. The method according to claim 1, characterized in that, The step of constructing a kiln process parameter space based on the value range of process parameter data in the preprocessed kiln operation dataset and production constraints, and then applying the trained time-series prediction model to calculate the predicted coal consumption value of candidate parameter combinations in the process parameter space includes: Based on the value range of process parameter data in the preprocessed kiln operation dataset and the production constraints, a set of key process parameters is determined, a value range and adjustment step size are defined for each parameter in the set of key process parameters, and the constraint relationship between parameters in the set of key process parameters is established to obtain the kiln process parameter space. Based on the kiln process parameter space, each parameter combination that satisfies the constraints in the kiln process parameter space is taken as a node in the graph. Edges are established between node pairs that satisfy the conditions of adjacency, smoothness and constraint compatibility to obtain the graph structure representation of the kiln process parameter space. Based on the graph structure representation of the kiln process parameter space and the trained time-series prediction model, the parameter combination corresponding to each node in the graph structure representation of the kiln process parameter space is input into the trained time-series prediction model for prediction calculation to obtain the predicted coal consumption value corresponding to each node, and the predicted coal consumption value is used as the weight of each node.

11. The method according to claim 10, characterized in that, The application optimization algorithm searches for the optimal solution with the goal of minimizing the predicted coal consumption value, and obtains the optimal combination of process parameters, including: Based on the graph structure representation of the kiln process parameter space and the current process parameter value, taking the node corresponding to the current process parameter value as the starting point, an adaptive branching strategy is adopted to select the branch variable and branching order to obtain a set of candidate branches; Based on the candidate branch set and the predicted coal consumption value, a multi-level bounding strategy is applied to estimate the lower bound of energy consumption, perform reachability analysis and constraint propagation on the candidate branch set, and prune the branches in the candidate branch set whose predicted coal consumption value is higher than the current optimal solution to obtain the retained branch set. Based on the set of retained branches, a dynamic constraint strategy is introduced to adjust the bundle width. The number of nodes retained in each layer is dynamically controlled according to the search state. The optimal solution set is continuously updated with the goal of minimizing the predicted coal consumption value, thereby obtaining the optimal combination of process parameters.

12. The method according to claim 11, characterized in that, The step of generating parameter adjustment instructions based on the optimal combination of process parameters and sending them to the control unit for adjustment includes: Based on the optimal combination of process parameters and the current process parameter value, find the optimal path from the node corresponding to the current process parameter value to the node corresponding to the optimal combination of process parameters in the graph structure representation of the kiln process parameter space, and design a timing adjustment curve for each process parameter in the optimal path to obtain the parameter adjustment scheme; Based on the parameter adjustment scheme, the parameter adjustment scheme is verified by a thermal process simulation model to check the transient response and parameter cross-influence of the parameter adjustment scheme, and the verified parameter adjustment scheme is obtained. Based on the verified parameter adjustment scheme, the parameter adjustment instruction is generated, and the parameter adjustment instruction is encoded into an industrial control protocol format to obtain an executable parameter adjustment instruction.

13. A rotary kiln energy consumption optimization system incorporating visual recognition, characterized in that, include: The data acquisition and preprocessing module is used to acquire real-time video data and multi-dimensional process parameter data during the kiln operation process, and to preprocess the real-time video data and the multi-dimensional process parameter data to obtain a preprocessed kiln operation dataset. The working condition identification module is used to segment the flame area and extract flame features based on the flame video data in the preprocessed kiln operation dataset, and obtain the kiln firing working condition classification result. The model training and prediction module is used to construct and train a time series prediction model based on the process parameter data in the preprocessed kiln operation dataset and the kiln firing condition classification results, to obtain the trained time series prediction model, and input the process parameter data in the preprocessed kiln operation dataset into the trained time series prediction model for calculation to obtain the predicted unit coal consumption of the kiln. The parameter space construction and optimization module is used to construct the kiln process parameter space based on the value range of the process parameter data in the preprocessed kiln operation dataset and the production constraints, apply the trained time series prediction model to calculate the predicted coal consumption value of the candidate parameter combination in the process parameter space, and apply the optimization algorithm to search for the optimal solution with the minimum predicted coal consumption value to obtain the optimal process parameter combination. The closed-loop control module is used to generate parameter adjustment instructions based on the optimal combination of process parameters and send them to the control unit for adjustment. It monitors the kiln operation status after adjustment in real time, obtains the actual kiln unit coal consumption result, compares and verifies the actual kiln unit coal consumption result with the predicted kiln unit coal consumption result, and feeds it back to the trained time series prediction model for model update, thus forming a closed-loop intelligent control system.

Images