A method, apparatus, and medium for predicting optimal target moisture of a sintering mix

Abstract

The application discloses a sintering mixture optimal target moisture prediction method, equipment and medium, belongs to the technical field of steel metallurgy sintering, solves the problem of improving the prediction accuracy and robustness of the sintering mixture moisture, and introduces a particle swarm optimization algorithm to optimize the LSTM super parameter to realize modeling of time series characteristics, and introduces a Bayesian tree structure parameter optimization algorithm to optimize the XGBoost super parameter to realize the reinforced fitting of static characteristics, and improve the overall performance of the model, then the high-order characteristics extracted by the optimized LSTM and XGBoost models are spliced, the adaptive weight is distributed in combination with the attention mechanism, dynamic fusion in the feature level is realized, the optimal target moisture value of the sintering mixture is output according to the fused characteristics and the weight, various process disturbances and data noises in the industrial field can be effectively inhibited, and high-precision prediction of the optimal target moisture value of the sintering mixture is realized.

Description

Technical Field

[0001] This invention belongs to the field of sintering process control technology in iron and steel metallurgy, and relates to a method, equipment and medium for predicting the optimal target moisture content of sintering mixtures based on dual-path feature fusion and adaptive optimization. Background Technology

[0002] In the sintering process, the moisture content of the sintering mixture is a core process parameter affecting the quality of the sinter. Appropriate moisture content enhances the bonding force between particles, promotes the formation of a uniform spherical structure, and thus improves the strength, permeability, and yield of the sinter. Too low a moisture content leads to loose particles, poor permeability, and increased energy consumption; too high a moisture content easily causes the formation of an overly wet layer, reducing the sintering speed and increasing the return rate. Furthermore, moisture improves the fluidity of the material layer through lubrication and participates in chemical reactions such as the separation of crystal water from limonite, affecting the thermodynamic processes of sintering.

[0003] Currently, the setting of target moisture content in sintering mixes largely relies on operator experience or historical statistical regression results. When production conditions, raw material properties, and environmental conditions change, existing manual moisture setting and passive adjustment methods are often lagging behind and cannot meet the requirements of efficient and stable production. The inability to accurately predict and adjust target moisture content in advance can easily lead to fluctuations in the physical properties of sinter, reduced energy utilization, and increased return ore volume, which can even seriously affect the safety and efficiency of continuous production. Therefore, achieving accurate prediction of the optimal target moisture content in the mix and guiding real-time adjustments to the amount of water added and mixing process parameters has become a key technical challenge for intelligent sintering control.

[0004] The prior art, such as the invention patent with publication number CN112986491A, discloses a method for correcting the moisture detection value of a mixture based on a feedback adaptive prediction model. First, the amount of water added to the mixture is calculated through a batching calculation system. Then, the input features are screened using the MIV (mean influence value) algorithm. Next, the moisture of the mixture is predicted using a BP neural network. The network weights are initialized using the PSO algorithm. Finally, the predicted value is compared with the target value for feedback correction. Although this method improves the accuracy of moisture prediction to a certain extent, it still has the following shortcomings: (1) The core model of this technology is a single-layer BP neural network, which has a weak ability to capture the long-term dependence of time series data and is difficult to effectively model the dynamic change law of moisture in sintering mixtures during the mixing process. (2) This technology uses grid search or empirical setting of model hyperparameters and lacks an adaptive hyperparameter optimization mechanism for different feature types and model structures, which limits the generalization ability of the model under different working conditions. (3) This technology uses a BP neural network and MIV feature screening, which is essentially a relatively shallow nonlinear mapping. When the input dimension is high and the working conditions are complex, it is easy to have problems such as overfitting or decreased prediction accuracy.

[0005] Therefore, among the existing sintering mixture moisture prediction technologies, there is a lack of a prediction scheme that can simultaneously and efficiently handle dynamic time-series characteristics and static slowly changing characteristics, has adaptive hyperparameter optimization capabilities, and can achieve dynamic fusion of heterogeneous model features. Developing a new method that can accurately predict the optimal target moisture content of sintering mixtures in advance is of great significance for realizing intelligent closed-loop control of the sintering process. Summary of the Invention

[0006] The technical problem to be solved by this invention is how to adaptively optimize the time series and static characteristics of sintering mixtures to improve the accuracy and robustness of moisture prediction in sintering mixtures.

[0007] The present invention solves the above-mentioned technical problems through the following technical solutions:

[0008] A method for predicting the optimal target moisture content of sintering mixtures includes the following steps: S1. Collect historical data on the sintering mixture blending process and clean the historical data; the historical data includes dynamic process parameters, static parameters and slowly changing process parameters. S2, feature extraction from historical data, including: Using dynamic process parameters as input, the hyperparameters of the first time-series prediction model are adaptively optimized. A high-order time-series feature vector reflecting the target moisture content is then extracted using the optimal first time-series prediction model. ; Using static parameters and slowly varying process parameters as input, the hyperparameters of the second static prediction model are adaptively optimized. A high-order static feature vector reflecting the target moisture content is then extracted using the optimal second static prediction model. ; S3, based on the attention mechanism and Calculate their respective weighting coefficients; S4, based on weight coefficients and Perform a weighted summation to output the final moisture prediction value.

[0009] Furthermore, as described in S2, a high-order temporal feature vector reflecting the target moisture content is extracted using the optimal first time-series prediction model. Includes the following: Using dynamic process parameters as the time series features of the model input, we define the LSTM hyperparameter space, map the hyperparameter space to the particle swarm, initialize the particle swarm, and each particle represents a complete combination of LSTM hyperparameters. Within a given hyperparameter space, the particle begins its search. Based on the hyperparameters corresponding to the particle's position, the LSTM model structure is compiled and the model is trained. The model loss or fitness is calculated and fed back to the particle. Based on the particle's current fitness, the particle's individual optimal position and global optimal position are updated using the velocity-position update formula, generating a new combination of hyperparameters. The above logic is repeated until convergence to the global optimum. The optimal LSTM model is constructed using the optimal combination of hyperparameters. The final state of the hidden layer in the trained LSTM model Feature vectors considered as target water .

[0010] Furthermore, the velocity-position update formula is as follows:

[0011]

[0012] in, , This represents the velocity of the particle at time steps t+1 and t. , The position of the particle at time step t+1, t represents the hyperparameter combination of the LSTM model. and These are the individual optimal position and the global optimal position of the particle, respectively. It is inertial weight. and It is a learning factor. and It is a random number. This represents the LSTM hyperparameter combination corresponding to the i-th particle.

[0013] Furthermore, as described in S2, a high-order static feature vector reflecting the target moisture content is extracted using the optimal second static prediction model. Includes the following: Using static parameters and slowly varying process parameters as static features for model input, the range of hyperparameters for the XGBoost model is defined, and the TPE algorithm is initialized. Candidate hyperparameters are randomly sampled and an XGBoost model is trained based on the current hyperparameters. The fitness is evaluated, and the distribution is adjusted according to the fitness to search for the optimal combination of hyperparameters. ; Based on optimal hyperparameter combination Construct an optimal XGBoost model and extract static features of the target optimal water content using the XGBoost model.

[0014] Furthermore, the search for the optimal combination of hyperparameters Specifically: In each iteration, the TPE algorithm adjusts the hyperparameter combination based on the currently sampled parameters. Calculate the objective function The calculation formula is:

[0015] in, This is test set data. The model's prediction results are used to minimize the RMSE. The TPE algorithm gradually converges and finds the optimal combination of hyperparameters. .

[0016] Furthermore, the static features for extracting the target optimal water content using the XGBoost model specifically include: First, construct the optimal XGBoost model. Using static parameters and slowly varying process parameters as inputs, let the first... The static and slowly varying process parameter feature vectors of each sample are: , The number of characteristics for static parameters and slowly varying process parameters. Indicates the first The first sample A static or slowly changing feature The XGBoost model optimized based on the TPE algorithm can be represented as:

[0017] in, This is the intermediate prediction output of the XGBoost branch for the target moisture content. The number of classification and regression trees, Indicates the first Tree of classification and regression, This represents the optimal combination of XGBoost hyperparameters obtained through the TPE algorithm search; Next, the gain value of each feature during the node splitting process is calculated as the feature contribution, and the first... The importance scores for each feature are:

[0018] in, Indicates the first Normalized importance scores for each feature. Indicates the first Among the trees, the first The set of nodes with each feature serves as a splitting feature. Indicates the first The first of the trees The decrease in loss function after splitting each node; obtaining the feature importance vector. ; Subsequently, the input static features are weighted according to the feature importance score to obtain the static feature representation with enhanced contribution. ,in, This indicates element-wise multiplication; High-order static eigenvectors are obtained through a nonlinear mapping function:

[0019] in, This is the static feature mapping weight matrix. For bias terms, Nonlinear activation function Furthermore, in S3, respectively... and Attention scores are obtained by mapping through different fully connected layers. , :

[0020]

[0021] in, , , , Both are weight matrices. and For attention vectors, and For bias terms; Attention score is obtained through the Softmax function. , Convert to normalized weights , This can be represented using the following logic:

[0022]

[0023] in, , Attention scores , The corresponding normalized weights satisfy the weight , The sum is 1. This indicates exponentiation.

[0024] Furthermore, in S4, the weight coefficients are used to pair... and Perform a weighted summation to obtain the fused feature vector. :

[0025] The fused feature vector The input is fed into a fully connected layer, and the final moisture prediction value is output. :

[0026] in, It is the transpose of the final prediction layer's weight vector. It is a bias term.

[0027] An electronic device includes a memory and a processor, the memory being used to store a program that supports the processor in executing the above-described method for predicting the optimal target moisture content of sintered mixtures, the processor being configured to execute the program stored in the memory.

[0028] A storage medium storing a computer program, which, when executed by a processor, performs the steps of the above-described method for predicting the optimal target moisture content of sintered mixtures.

[0029] The advantages of this invention are: To address the challenge of predicting the optimal moisture content of sintering mixtures, this invention constructs a dual-path hybrid model architecture combining "time series and static features." This architecture leverages the advantages of LSTM networks in capturing dynamic temporal information and XGBoost models in processing static features, achieving synergistic optimization of dynamic time series modeling and static feature analysis. By introducing PSO and TPE algorithms to drive dual-model hyperparameter synergistic optimization, it overcomes the limitations of traditional single-model tuning. Furthermore, by efficiently fusing multi-source heterogeneous data and adaptively weighting the prediction of this data, the reliability of predicting the optimal target moisture content is improved. This addresses the limitations of traditional single models in complex production environments, providing a more accurate and stable moisture prediction solution for sintering mixtures.

[0030] Specifically, in the sub-model, this invention introduces the Particle Swarm Optimization (PSO) algorithm to optimize the LSTM hyperparameters to model the time series features; and the Bayesian Tree Structure Parameter Optimization (TPE) algorithm to optimize the XGBoost hyperparameters to enhance the fitting of static features and improve the overall performance of the model. Then, the high-order features extracted by the optimized LSTM and XGBoost models are concatenated, and adaptive weights are assigned using an attention mechanism to achieve dynamic fusion at the feature level. Based on the fused features and weights, the optimal target moisture value of the sintering mixture is output, which can effectively suppress various process disturbances and data noise in the industrial field and achieve high-precision prediction of the optimal target moisture value of the sintering mixture. Attached Figure Description

[0031] Figure 1This is a flowchart illustrating the optimal target moisture content prediction method for sintered mixtures according to Embodiment 1 of the present invention. Figure 2 This is a flowchart illustrating the extraction of high-order temporal feature vectors according to Embodiment 1 of the present invention; Figure 3 This is a flowchart illustrating the extraction of high-order static feature vectors according to Embodiment 1 of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments: Example 1 Currently, the prediction of the optimal target moisture content of the mixture in the sintering process in factories usually relies on manual experience or traditional linear regression methods. These two methods are insufficient to meet the precise control requirements of complex production environments. To address the challenge of predicting the optimal moisture content of the mixture, this invention constructs a dual-path hybrid model architecture combining "time series and static features." It leverages the advantages of LSTM networks in capturing dynamic temporal information and XGBoost models in processing static features, achieving synergistic optimization of dynamic time series modeling and static feature analysis. By introducing PSO and TPE algorithms to drive the synergistic optimization of hyperparameters in both models, it overcomes the limitations of traditional single-model tuning. Furthermore, by efficiently fusing multi-source heterogeneous data and adaptively weighting the prediction of multi-source heterogeneous data, it improves the reliability of the optimal target moisture content prediction.

[0034] Specifically, in the sub-model, this invention introduces the Particle Swarm Optimization (PSO) algorithm to optimize the LSTM hyperparameters to model the time series features; and the Bayesian Tree Structure Parameter Optimization (TPE) algorithm to optimize the XGBoost hyperparameters to enhance the fitting of static features and improve the overall performance of the model. Then, the high-order features extracted by the optimized LSTM and XGBoost models are concatenated, and adaptive weights are assigned using an attention mechanism to achieve dynamic fusion at the feature level. Based on the fused features and weights, the optimal target moisture value of the sintering mixture is output, which can effectively suppress various process disturbances and data noise in the industrial field and achieve high-precision prediction of the optimal target moisture value of the sintering mixture.

[0035] like Figure 1As shown, this invention discloses a method for predicting the optimal target moisture content of sintering mixtures, comprising the following steps: S1, Data Acquisition and Preprocessing: Collect historical data on the blending process of the sintering mixture and clean the historical data; the historical data includes dynamic process parameters, static parameters and slowly changing process parameters, and the cleaning includes handling outliers and missing values.

[0036] When using the PSO algorithm to optimize the hyperparameters of the LSTM model, the selection of parameters related to the mixture directly affects the accuracy of moisture prediction. Considering that the LSTM model is adept at processing time-series data, and the moisture changes of sintering mixtures are usually time-dependent, dynamic process parameters such as mixture spray flow rate, spray time, mixing time, initial moisture content of raw materials, mixer speed, pressure, temperature, and humidity are collected from historical data and used as input to the LSTM model. These dynamic process parameters help the LSTM model capture the pattern of moisture changes over time and accurately model the dynamic changes of moisture in the sintering mixture. See Table 1 below: Table 1 Dynamic Process Parameters

[0037] By optimizing the hyperparameters of the LSTM through PSO, the model can more accurately adapt to these complex time-series features, ultimately improving the accuracy and robustness of predictions. In particular, the LSTM model, through its memory mechanism, can learn and maintain long-term dependencies, adapt to the delay effects in the sintering process, and thus effectively predict future moisture states, ensuring high accuracy and stability of the prediction results.

[0038] When optimizing the XGBoost model using the TPE algorithm, this embodiment uses the static and slowly varying process parameters of the sintering mixture as input features. These parameters include at least Fe content, SiO2 content, CaO content, Al2O3 content, MgO content, return ore ratio, blending ratio, initial moisture content of raw materials, total amount of water sprayed during mixing, average water spray temperature, mixing time, average mixing and stirring power, ambient temperature, and humidity. These parameters are highly representative. They accurately reflect the composition of the mixture and the initial conditions of the production process, determining the effectiveness of moisture control during sintering. (See Table 2 below.) Table 2 Parameter Table for Static and Slowly Changing Processes

[0039] As a tree-based ensemble learning method, the XGBoost model can handle complex nonlinear relationships and effectively capture the interactions between these input parameters through its tree structure. By optimizing the XGBoost hyperparameters using the TPE algorithm, the model can find the optimal combination based on these static features, thereby improving the accuracy of moisture prediction. Furthermore, XGBoost's feature importance assessment capability allows it to more clearly identify which parameters are most critical for moisture prediction, thus providing an efficient and interpretable solution for moisture prediction of sintered mixtures and ensuring the adaptability and accuracy of the prediction results under different process conditions.

[0040] S2, extract the high-order features of both paths, including the following: S21. The Particle Swarm Optimization (PSO) algorithm is used to search for the optimal combination of hyperparameters (gBest) of the Long Short-Term Memory (LSTM) network model, and an LSTM prediction model is constructed. The hidden state of the model is extracted as a high-order time-series feature vector reflecting the target moisture content, using dynamic process parameters as input. .

[0041] like Figure 1 As shown, this embodiment uses the Particle Swarm Optimization (PSO) algorithm to optimize the hyperparameters of the Long Short-Term Memory (LSTM) network model, demonstrating significant superiority in predicting the optimal target moisture content of the sintering mixture. S21 specifically includes the following: S211 uses dynamic process parameters as the time series features of the model input, defines the LSTM hyperparameter space, maps the hyperparameter space to the particle swarm, and initializes the particle swarm.

[0042] In this embodiment, each "particle" in the particle swarm optimization (PSO) represents a complete combination of LSTM hyperparameters. The optimal hyperparameters searched using the PSO algorithm in this embodiment include: learning rate, number of long short-term memory (LSTM) network units, batch size, optimizer, time step, number of LSTM network layers, dropout rate, sequence length, momentum, and batch normalization. The PSO algorithm, by simulating swarm intelligence to search the hyperparameter space, can efficiently explore the optimal combination of hyperparameters for the LSTM model. The hyperparameters and related parameters in the PSO-optimized LSTM model are shown in Table 3, avoiding the computational redundancy of traditional grid search and improving the model's prediction accuracy. The optimized LSTM model can capture the long-term temporal dependence of moisture changes in sintering mixtures, fully utilizing the dynamic features in time series data, and can predict moisture fluctuations more accurately than traditional models.

[0043] Table 3. Optimal Hyperparameters Searched Based on Particle Swarm Optimization Algorithm

[0044] Preferably, in this embodiment, when using the particle swarm optimization algorithm to search for the optimal hyperparameters of the long short-term memory network model, the hyperparameters to be optimized and their value ranges include: learning rate of 0.0001~0.1, number of long short-term memory network units of 32~512, batch size of 16~128, optimizer can be any one of 'adam', 'sgd' or 'rmsprop', time step of 5~30, number of long short-term memory network layers of 1~3, dropout rate of 0.0~0.5, sequence length of 10~50, momentum of 0.0~0.9, and batch normalization of 0 or 1.

[0045] S212, based on the iterative drive of the Particle Swarm Optimization (PSO) algorithm: Within a given hyperparameter space, the particles begin their search, each randomly initializing its position and velocity, and then proceed as follows: Based on the hyperparameters corresponding to the particle positions, the LSTM model structure is compiled and the model is trained. The model loss (LOSS) is calculated and fed back to the particle. The particle calculates its current fitness and updates its individual optimal position and global optimal position using the velocity-position update formula, generating a new combination of hyperparameters. The above logic is repeated until it converges to the global optimum. The optimal LSTM model is then constructed using the optimal combination of hyperparameters.

[0046] In this embodiment, the PSO algorithm dynamically adjusts the search direction during the optimization process, ensuring the robustness and generalization ability of the LSTM model in complex environments, making the final moisture prediction results more stable and reliable. The velocity-position update formula of the PSO algorithm is as follows:

[0047]

[0048] in, , This represents the velocity of the particle at time steps t+1 and t. , The position of the particle at time step t+1, t represents the hyperparameter combination of the LSTM model. This represents the LSTM hyperparameter combination corresponding to the i-th particle; and These are the individual optimal position and the global optimal position of the particle, respectively. It is inertial weight. and It is a learning factor. and These are random numbers. Through multiple iterations, PSO can search for an optimal combination of LSTM hyperparameters, ensuring that the LSTM model performs optimally on the water prediction task.

[0049] like Figure 2 As shown, the LSTM model includes an input layer, a hidden layer, and an output layer; dynamic process parameters are normalized in the input layer (such as Z-score normalization) to generate standardized input data, so as to eliminate differences in different physical dimensions and prevent gradient vanishing or exploding during long sequence propagation; The hidden layer comprises a stacked number of Long Short-Term Memory (LSTM) network units, each of which consists of an input gate. Forgotten Gate Output gate Cell state and hidden state Composed of the hidden state from the previous moment and current input Together, they determine the current gating state, thereby updating the cell state. Output the current hidden state. The effective long-term features are dynamically retained and invalid noise is forgotten through a gating mechanism; This invention differs from LSTM prediction models. The output layer, after iterative training, does not directly predict the final moisture value. Instead, it extracts a high-order temporal feature vector containing long-term dependencies. The hidden state at the last extracted time step is used as the high-order temporal feature vector reflecting the target moisture level. .

[0050] S213: After optimizing the LSTM model using the PSO algorithm, the optimized hyperparameters are used to train the LSTM model, and this model is used to extract the feature vector of the target optimal water content. The LSTM model learns from time-series data (records of changes in water content, temperature, humidity, etc. over time), effectively capturing the long-term dependencies of water changes. The trained LSTM model can output the hidden state. This state contains dynamic information about the water change at time step t. In the final LSTM model, the final state of the hidden layer... Feature vectors considered as target water It can be expressed by the following recursive formula:

[0051] in, This represents the hidden state at time step t. These are the input features at time step t. It is the hidden state of the previous time step, and this feature vector It captures key features of water changes, including the correlation between water and other environmental factors, and uses them as input feature vectors for subsequent prediction tasks.

[0052] S22. A sequence model-based hyperparameter optimization (TPE) algorithm is used to search for the optimal combination of hyperparameters for the extreme gradient boosting model (XGBoost model). An XGBoost prediction model is constructed, using static parameters and slowly varying process parameters as inputs. After weighting the contribution of the model output, a high-order static feature vector reflecting the target moisture content is extracted. .

[0053] like Figure 1 As shown, the TPE algorithm, through Bayesian optimization principles, explores the hyperparameter space using a probabilistic model, effectively improving the performance of the XGBoost model. This embodiment demonstrates significant superiority in predicting the optimal target moisture content of sintering mixtures using the TPE algorithm to optimize the XGBoost model. Specifically, S22 includes the following: S221 defines the range of hyperparameters for the XGBoost model by using static parameters and slowly varying process parameters as static features for model input, and initializes the TPE algorithm.

[0054] The optimal hyperparameters searched by the sequence model-based hyperparameter optimization algorithm include: learning rate, maximum depth, number of trees, subsample ratio, column sampling ratio, regularization parameter reg_alpha, regularization parameter reg_lambda, minimum sample weight, column sampling ratio, and maximum number of iterations. The hyperparameter settings for TPE-optimized XGBoost models are shown in Table 4 below. Table 4. Optimal Hyperparameters for XGBoost Model Searched Based on TPE Algorithm

[0055] Preferably, when using a sequence model-based hyperparameter optimization algorithm to search for the optimal hyperparameters of the extreme gradient boosting model in this embodiment, the hyperparameters to be optimized and their value ranges include: a learning rate of 0.01 to 0.3, a maximum depth of 3 to 10, a number of trees of 50 to 500, a subsample ratio of 0.5 to 1.0, a column sampling ratio of 0.5 to 1.0, a regularization parameter reg_alpha of 0 to 1, a regularization parameter reg_lambda of 0 to 1, and a minimum sample weight of 1 to 10.

[0056] S222: Randomly sample candidate hyperparameters and train the XGBoost model based on the current hyperparameters. Evaluate the fitness and adjust the distribution according to the fitness to search for the optimal combination of hyperparameters. .

[0057] In this embodiment, the XGBoost model consists of k classification and regression trees. Each tree is used to fit the cumulative residuals of all preceding tree sets, and the prediction result is the sum of the outputs of all trees, denoted as . , where x represents the sample feature vector input to the XGBoost model. The hyperparameters of the k-th classification and regression tree are represented by the hyperparameters of the TPE search, which directly affect each tree split and the overall integration process during the construction of the XGBoost model.

[0058] like Figure 3 As shown, the TPS algorithm loop process is as follows: For a set of hyperparameters generated in each iteration The XGBoost algorithm is trained, and the root mean square error (RMSE) is calculated on the validation set as the objective function value for the TPE algorithm. This value is used to update the probability distribution of TPE and guide the sampling direction in the next round. The TPE algorithm continuously optimizes the objective function in multiple iterations, thereby avoiding overfitting and enhancing the model's generalization ability. This makes the optimized XGBoost model more stable and accurate in predicting the moisture content of sintering mixtures, and able to adapt to changes under different production conditions. In each iteration, the TPE algorithm adjusts the hyperparameter combination based on the currently sampled data. Calculate the objective function The calculation formula is:

[0059] in, This is test set data. The model's prediction results are used to minimize the RMSE. The TPE algorithm gradually converges and finds the optimal combination of hyperparameters. .

[0060] S223, based on optimal hyperparameter combination Construct the optimal XGBoost model and output the static features of the target optimal water content extracted by the XGBoost model.

[0061] Unlike conventional XGBoost model predictions, this embodiment, based on the optimal XGBoost model obtained through TPE optimization, does not directly output the final predicted moisture value. Instead, it extracts the static feature vector of the target optimal moisture level. This embodiment uses time-series features (such as moisture, temperature, and humidity) as input, calculates the contribution of each feature in the XGBoost model, and uses the contribution-weighted feature representation as the high-order static feature vector of the target moisture level. This feature vector reflects the relationship between moisture changes and other production parameters.

[0062] Unlike conventional XGBoost models that directly output the final predicted water content, this embodiment obtains the optimal hyperparameter combination using the TPE algorithm. Then, the optimal XGBoost model is constructed first. Furthermore, using static parameters and slowly varying process parameters as inputs, we extract features that significantly contribute to the prediction of target moisture content.

[0063] Let the first The static and slowly varying process parameter feature vectors of each sample are:

[0064] in, The number of characteristics for static parameters and slowly varying process parameters. Indicates the first The first sample A static or slowly changing feature The XGBoost model optimized based on the TPE algorithm can be represented as:

[0065] in, This is the intermediate prediction output of the XGBoost branch for the target moisture content. The number of classification and regression trees, Indicates the first Tree of classification and regression, This represents the optimal combination of XGBoost hyperparameters obtained through the TPE algorithm search.

[0066] After model training is complete, the contribution of each input feature to the model's prediction result is calculated. This embodiment uses the gain value of each feature during node splitting in the XGBoost model as the feature contribution, defining the... The importance scores for each feature are:

[0067] in, Indicates the first Normalized importance scores for each feature. Indicates the first Among the trees, the first The set of nodes with each feature serves as a splitting feature. Indicates the first The first of the trees The decrease in loss function resulting from node splitting. This yields the feature importance vector. :

[0068] satisfy .

[0069] Subsequently, the input static features are weighted according to their feature importance scores to obtain the enhanced static feature representations:

[0070] in, This indicates element-wise multiplication.

[0071] To further map the weighted static features to a feature space that matches the higher-order temporal features of the LSTM branch, this embodiment obtains the higher-order static feature vector through a nonlinear mapping function:

[0072] in, This represents the high-order static feature vector extracted by the optimal XGBoost model. This is the static feature mapping weight matrix. For bias terms, This is a nonlinear activation function. This feature vector is not the final water prediction value, but rather an intermediate feature representation reflecting the nonlinear relationship between static parameters, slowly varying process parameters, and the target water level. It is used subsequently with the higher-order time-series feature vector extracted by the LSTM branch. Perform attention fusion.

[0073] Finally, scores were given based on feature importance. Sort by largest to smallest, then select the first... These important features constitute a representative feature set. ,Will and The updated result is obtained by element-wise multiplication. Substitute .

[0074] In this embodiment, LSTM and XGBoost models, optimized separately by PSO, are applied to predict the moisture content of sintering mixtures. Feature vectors are extracted from each model. Subsequently, a dynamic weighting strategy is used to adaptively weight and fuse the feature vectors output by the two models. This strategy dynamically adjusts the weights of each model based on its performance in specific scenarios, thereby fully leveraging the advantages of LSTM and XGBoost under different data patterns. The fused feature vector is input into the target prediction module, ultimately yielding a more accurate moisture prediction value. This embodiment combines the powerful capabilities of LSTM in time series modeling with the efficient performance of XGBoost in nonlinear relationship modeling, improving the accuracy and robustness of prediction through weighted fusion. The specific steps are as follows: S3 inputs the dual-path high-order features into the attention module, assigning weights to each feature vector to characterize its contribution to the final prediction, specifically: S31, the higher-order temporal feature vector With higher-order static feature vectors By concatenating the features, a joint feature vector is obtained. , ,in Let R be the dimension of the joint feature vector, and R be the set of real numbers. Joint feature vector It is used to simultaneously characterize the dynamic temporal information and static / gradually changing operating condition information of the moisture change in the sintering mixture, and provides global context information for the subsequent attention module to calculate the relative contribution weights of the two types of features.

[0075] S32 represents the higher-order temporal feature vectors. With higher-order static feature vectors Calculate the attention score and normalize it into weight coefficients using the Softmax function.

[0076] To achieve more efficient feature fusion, this embodiment designs a learnable attention function for the LSTM and XGBoost models to dynamically adjust the weights based on the contribution of each model. Specifically, the feature vector of the LSTM model... An attention score is obtained by mapping through a fully connected layer. The feature vectors of the XGBoost model Another attention score is obtained by mapping through another fully connected layer. This can be represented using the following logic:

[0077]

[0078] in, , , , Both are weight matrices. and For attention vectors, and This is a bias term; in this embodiment and This is used to incorporate joint feature information into attention score calculation, so that the weights are not determined in isolation by a single path feature, but are judged by combining the overall relationship between the two types of features.

[0079] Attention score is obtained through the Softmax function. , Convert to normalized weights , In order to assign appropriate weights to the feature vectors of the above model, the following logical representation is used:

[0080]

[0081] in, , Attention scores , The corresponding normalized weights satisfy the weight , The sum is 1. This indicates exponentiation.

[0082] S4, in the weighted fusion module, the higher-order time-series feature vector and the higher-order static feature vector are weighted and summed based on the weight coefficients to output the final moisture prediction value.

[0083] In this embodiment, the feature vectors of the two models are weighted and fused based on the normalized weights obtained by S3 to obtain a fused feature vector. :

[0084] Furthermore, the fused feature vector is input into a fully connected layer for final prediction. By using the weighted fused feature vector, the model can effectively combine the advantages of LSTM and XGBoost in different scenarios, thus obtaining a more accurate target moisture prediction. :

[0085] in, It is the transpose of the final prediction layer's weight vector. This is the bias term, representing the prediction baseline offset. This represents the fused feature vector. Finally, by combining the weighted fusion of the feature vectors optimized by LSTM and XGBoost, a more accurate target moisture prediction value is obtained.

[0086] Example 2 An apparatus includes a memory and a processor, the memory being used to store a program that supports the processor in executing the optimal target moisture prediction method for sintered mixtures in Embodiment 1, the processor being configured to execute the program stored in the memory.

[0087] Example 3 A storage medium storing a computer program, which, when executed by a processor, performs the steps of the optimal target moisture prediction method for sintered mixtures in Embodiment 1.

[0088] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein can be considered as a ordered list of executable instructions for implementing logical functions. This can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-based system, or other system that can fetch and execute instructions from, or in conjunction with, such an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections having one or more wires (electronic devices), portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CD-ROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0089] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0090] In the description of this specification, references to terms such as "in a preferred embodiment," "preferred," "in this embodiment," "specific," or "furthermore," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0091] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0092] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the optimal target moisture content of sintering mixtures, characterized in that, Includes the following steps: S1. Collect historical data on the sintering mixture blending process and clean the historical data; the historical data includes dynamic process parameters, static parameters and slowly changing process parameters; S2, feature extraction from historical data, including: Using dynamic process parameters as input, the hyperparameters of the first time-series prediction model are adaptively optimized. A high-order time-series feature vector reflecting the target moisture content is then extracted using the optimal first time-series prediction model. ; Using static parameters and slowly varying process parameters as input, the hyperparameters of the second static prediction model are adaptively optimized. A high-order static feature vector reflecting the target moisture content is then extracted using the optimal second static prediction model. ; S3, based on the attention mechanism and Calculate their respective weighting coefficients; S4, based on weight coefficients and Perform a weighted summation to output the final moisture prediction value.

2. The method for predicting the optimal target moisture content of sintered mixtures according to claim 1, characterized in that, As described in S2, a high-order temporal feature vector reflecting the target moisture content is extracted using the optimal first-time-series prediction model. Includes the following: Using dynamic process parameters as the time series features of the model input, we define the LSTM hyperparameter space, map the hyperparameter space to the particle swarm, initialize the particle swarm, and each particle represents a complete combination of LSTM hyperparameters. Within a given hyperparameter space, the particle begins its search. Based on the hyperparameters corresponding to the particle's position, the LSTM model structure is compiled and the model is trained. The model loss or fitness is calculated and fed back to the particle. Based on the particle's current fitness, the particle's individual optimal position and global optimal position are updated using the velocity-position update formula, generating a new combination of hyperparameters. The above logic is repeated until convergence to the global optimum. The optimal LSTM model is constructed using the optimal combination of hyperparameters. The final state of the hidden layer in the trained LSTM model Feature vectors considered as target water .

3. The method for predicting the optimal target moisture content of sintered mixtures according to claim 2, characterized in that, The velocity-position update formula is as follows: in, , This represents the velocity of the particle at time steps t+1 and t. , The position of the particle at time step t+1 and t represents the hyperparameter combination of the LSTM model. and These are the individual optimal position and the global optimal position of the particle, respectively. It is inertial weight. and It is a learning factor. and It is a random number. This represents the LSTM hyperparameter combination corresponding to the i-th particle.

4. The method for predicting the optimal target moisture content of sintered mixtures according to claim 1, characterized in that, As described in S2, a high-order static feature vector reflecting the target moisture content is extracted using the optimal second static prediction model. Includes the following: Using static parameters and slowly varying process parameters as static features for model input, the range of hyperparameters for the XGBoost model is defined, and the TPE algorithm is initialized. Candidate hyperparameters are randomly sampled and an XGBoost model is trained based on the current hyperparameters. The fitness is evaluated, and the distribution is adjusted according to the fitness to search for the optimal combination of hyperparameters. ; Based on optimal hyperparameter combination Construct an optimal XGBoost model and extract static features of the target optimal water content using the XGBoost model.

5. The method for predicting the optimal target moisture content of sintered mixtures according to claim 4, characterized in that, The search for the optimal hyperparameter combination Specifically: In each iteration, the TPE algorithm adjusts the hyperparameter combination based on the currently sampled parameters. Calculate the objective function The calculation formula is: in, This is test set data. The model's prediction results are used to minimize the RMSE. The TPE algorithm gradually converges and finds the optimal combination of hyperparameters. .

6. The method for predicting the optimal target moisture content of sintered mixtures according to claim 4, characterized in that, The static features extracted using the XGBoost model to obtain the target optimal water content are specifically as follows: First, construct the optimal XGBoost model. Using static parameters and slowly varying process parameters as inputs, let the first... The static and slowly varying process parameter feature vectors of each sample are: , The number of characteristics for static parameters and slowly varying process parameters. Indicates the first The first sample A static or slowly changing feature The XGBoost model optimized based on the TPE algorithm can be represented as: in, This is the intermediate prediction output of the XGBoost branch for the target moisture content. The number of classification and regression trees, Indicates the first Tree of classification and regression, This represents the optimal combination of XGBoost hyperparameters obtained through the TPE algorithm search; Next, the gain value of each feature during the node splitting process is calculated as the feature contribution, and the first feature is defined as... The importance scores for each feature are: in, Indicates the first Normalized importance scores for each feature Indicates the first Among the trees, the first The set of nodes with each feature serves as a splitting feature. Indicates the first The first of the trees The decrease in loss function after splitting each node; obtaining the feature importance vector. ; Subsequently, the input static features are weighted according to the feature importance score to obtain the static feature representation with enhanced contribution. ,in, This indicates element-wise multiplication; High-order static eigenvectors are obtained through a nonlinear mapping function: in, This is the static feature mapping weight matrix. For bias terms, It is a non-linear activation function.

7. The method for predicting the optimal target moisture content of sintered mixtures according to claim 1, characterized in that, In S3, respectively and Attention scores are obtained by mapping through different fully connected layers. , : in, , , , Both are weight matrices. and For attention vectors, and For bias terms; Attention score is obtained through the Softmax function. , Convert to normalized weights , This can be represented using the following logic: in, , Attention scores , The corresponding normalized weights satisfy the weight , The sum is 1. This indicates exponentiation.

8. The method for predicting the optimal target moisture content of sintering mixtures according to claim 7, characterized in that, S4 is based on weight coefficient pairs and Perform a weighted summation to obtain the fused feature vector. : The fused feature vector The input is fed into a fully connected layer, and the final moisture prediction value is output. : in, It is the transpose of the final prediction layer's weight vector. It is a bias term.

9. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the optimal target moisture prediction method for sintered mixtures according to any one of claims 1 to 8, and the processor is configured to execute the program stored in the memory.

10. A storage medium storing a computer program, characterized in that, When the computer program is run by the processor, it performs the steps of the optimal target moisture prediction method for sintered mixtures according to any one of claims 1 to 8.

Images