A multi-algorithm fusion sintering machine head ash thallium content prediction method

Abstract

The application discloses a sintering machine head ash thallium content prediction method based on multi-algorithm fusion, adopts a stacking integrated learning framework to fuse the advantages of three different types of algorithms, namely, a Transformer, a GRU and an XGBoost, can simultaneously consider the requirements of time sequence feature extraction and nonlinear relationship fitting, fully excavates the time sequence dependent features of sintering process parameters and the nonlinear correlation between the process parameters and thallium content, and significantly improves the precision and stability of thallium content prediction compared with a single prediction model. The application can directly output the thallium content prediction result based on the real-time collected sintering process parameters, does not need long-time laboratory testing, the prediction result is real-time, can help the steel enterprises to timely adjust the sintering process parameters, realizes accurate classification and disposal of the sintering machine head ash, and meets the requirement of real-time and accurate prediction in an industrial field.

Description

Technical Field

[0001] This invention relates to the technology for predicting the thallium content in sintering solid waste, specifically to a method for predicting the thallium content in sintering machine head ash using a multi-algorithm fusion approach, belonging to the field of thallium content prediction technology for sintering solid waste. Background Technology

[0002] Thallium in steel enterprises mainly originates from iron ore. After processes such as sintering, blast furnace smelting, desulfurization, and solid waste disposal, thallium is primarily concentrated in solid products generated from each process, including sintering machine head ash, sintered ore, zinc-containing dust, and blast furnace slag, as well as gaseous products such as sintering flue gas and blast furnace gas, and liquid products such as desulfurization wastewater and ash washing water. On-site testing of thallium-containing products from each process has revealed that approximately 50% to 70% of thallium is concentrated in sintering machine head ash. As the main thallium-containing solid waste generated during the sintering process, accurate prediction of the thallium content in sintering machine head ash is fundamental for enterprises to implement targeted thallium pollution control and achieve classified disposal of thallium-containing solid waste. This can prevent high-thallium-content solid waste from entering downstream utilization stages and causing secondary pollution.

[0003] Currently, the thallium content in sintering ash is typically determined in industrial settings through periodic sampling and laboratory chemical analysis. This method suffers from long detection cycles and delayed results, failing to meet the demands for real-time adjustment of sintering process parameters to stabilize thallium enrichment and timely classification and disposal. Existing technologies have attempted to predict thallium content based on experience, but these methods largely rely on manual estimation of the correlation between thallium content in iron ore and thallium content in sintering ash. This results in low accuracy and poor adaptability, failing to accommodate changes in thallium enrichment caused by fluctuations in raw material composition and process parameters. Furthermore, existing single machine learning prediction models either fail to effectively capture the time-dependent characteristics of sintering process parameters or lack sufficient fitting ability for nonlinear correlations. When used alone, they exhibit large prediction errors and poor stability, making it difficult to meet the accuracy requirements for thallium content prediction in industrial settings. Summary of the Invention

[0004] To address the problems of large lag or poor prediction stability in existing thallium content detection methods for sintering machine heads, this invention proposes a multi-algorithm fusion method for predicting thallium content in sintering machine heads. By integrating the prediction advantages of different algorithms through a stacking ensemble learning framework, it can simultaneously leverage the feature extraction advantages of different algorithms, fully explore and capture the time-dependent features of sintering process parameters and the nonlinear correlation between process parameters and thallium content, effectively improving prediction accuracy and stability, and meeting the needs of real-time and accurate prediction in industrial settings.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A multi-algorithm fusion method for predicting thallium content in sintering die head ash, comprising the following steps:

[0007] S1: Collect historical process parameters of the sintering process and historical test values ​​of thallium content in the die head ash. After time alignment processing of the historical process parameters and historical test values, a time-series sample dataset is obtained.

[0008] S2: Using time-series sample datasets, we train the basic models based on the Transformer algorithm, GRU algorithm, and XGBoost algorithm respectively. After training, we obtain three basic prediction models.

[0009] S3: A meta-model is constructed based on the LightGBM algorithm, and the prediction results of the three basic prediction models are used as meta-feature vectors. The historical test values ​​of thallium content in the ash at the machine head are used as targets to train the meta-model, resulting in a comprehensive prediction model that integrates multiple algorithms.

[0010] S4: Real-time acquisition of current process parameters during sintering, inputting them into the comprehensive prediction model, and outputting the predicted value of the current thallium content in the die head.

[0011] Preferably, in step S1, the time alignment process specifically involves: grouping and matching historical process parameters and historical test values ​​according to timestamps; taking the sampling time of historical test values ​​as a benchmark, extracting all historical process parameter data within a preset time interval before the time node; and organizing the process parameters within the interval into a sequence with a unified dimension according to time order. At the same time, missing values ​​and outliers caused by equipment acquisition anomalies are removed, thus completing the time dimension alignment of process data and test values, so that they correspond to the same batch of sintering mixture.

[0012] Preferably, in step S2, the basic prediction model based on the Transformer algorithm includes an input layer, a positional encoding layer, an encoding layer, a residual connection and normalization layer, and an output layer. Specifically: the input layer is used to input time-aligned time-series sample data into the model. The positional encoding layer adds temporal position information to the input time-series sample data to clarify the order of process parameters at different time points. The encoding layer consists of a multi-head self-attention mechanism sublayer and a feedforward neural network, used to capture the correlation features between process parameters at different time points and to perform nonlinear mapping and enhancement of features. The residual connection and normalization layer is used to alleviate the gradient vanishing problem during model training, improve model training stability, and avoid deep network degradation. The output layer outputs the predicted value of thallium content in the machine head ash at the target prediction time.

[0013] Preferably, in step S2, when training the basic prediction model based on the Transformer algorithm: the Adam optimizer is used, and the initial learning rate is set to 0.001~0.01. The loss function is the mean squared error, minimizing the error between the model's predicted value and the actual test value. The batch size is set to 32~128, the number of iterations is set to 50~500 rounds, and an early stopping strategy is adopted, with 10~30 early stopping rounds.

[0014] Preferably, in step S2, the basic prediction model based on the GRU algorithm includes an input layer, a GRU hidden layer, and an output layer. Specifically: the input layer is used to input time-aligned time-series sample data into the model. The GRU hidden layer includes at least one GRU network layer, each containing multiple neurons. Through update and reset gates, it selectively retains and forgets time-series information of process parameters at different time points, capturing the long-term dependence features of process parameters over time. The output layer outputs the predicted value of thallium content in the machine head ash at the target time, based on the time-series features extracted by the GRU hidden layer. Preferably, a dropout layer is also set between the GRU hidden layers to randomly forget the output of some neurons, avoiding overfitting and improving the model's generalization ability.

[0015] Preferably, in step S2, when training the basic prediction model based on the GRU algorithm: the Adam optimizer is used, and the initial learning rate is set to 0.001~0.01. The loss function is the mean squared error, minimizing the error between the model's predicted value and the actual test value. The batch size is set to 16~64, the number of iterations is set to 30~300 rounds, and an early stopping strategy is adopted, with 5~20 early stopping rounds.

[0016] Preferably, in step S2, the basic prediction model based on the XGBoost algorithm includes an input layer, an iterative layer, and an output layer. Specifically: the input layer is used to input time-aligned time-series sample data into the model; the iterative layer performs multi-round iterations to split and grow process features, optimize the objective function, and explore the nonlinear correlation between different process parameters and thallium content, completing the selection and importance ranking of key features; the output layer is used to map the feature learning results to the predicted value of thallium content in the dust at the target time.

[0017] Preferably, in step S2, when training the basic prediction model based on the XGBoost algorithm, the Adam optimizer is used, the initial learning rate is set to 0.01~0.3, the number of trees is 100~500, the maximum tree depth is 3~10, the subsample sampling ratio is set to 0.6~0.9, the column sampling ratio is set to 0.6~1.0, the mean squared error is used as the loss function, the batch size is 32~128, the number of iterations is 50~500 rounds, and an early stopping strategy is adopted, with 10~30 early stopping rounds.

[0018] Preferably, in step S3, the meta-model based on the LightGBM algorithm includes an input layer, a training layer, and an output layer. Specifically: the input layer is used to input a three-dimensional meta-feature vector, composed of the predicted values ​​from three basic prediction models, into the model. The training layer uses the actual measured value of thallium content in the machine head ash as the prediction target. Through multiple iterations, iterative learning is performed on the fusion weights of the predicted values ​​from multiple basic models to explore the prediction advantages of different basic models under different process scenarios and output the optimal fusion relationship. The output layer outputs the predicted value of thallium content in the machine head ash at the target prediction time based on the learned fusion relationship.

[0019] Preferably, in step S3, when training the meta-model based on the LightGBM algorithm, the following settings are used: the Adam optimizer is employed, the initial learning rate is set to 0.01~0.2, the number of trees is 50~400, the number of leaf nodes is set to 10~100, the maximum tree depth is 2~10, the minimum number of leaf node samples is 5~50, the mean squared error is used as the loss function, the batch size is 16~128, the number of iterations is 100~300 rounds, and an early stopping strategy is adopted, with 10~50 early stopping rounds.

[0020] In this invention, the multi-algorithm fusion method for predicting thallium content in sintering die head ash first involves collecting historical process parameters of the sintering process and historical test values ​​of thallium content in the die head ash. The historical process parameters and test values ​​are then time-aligned (resulting in a time-series sample dataset). Three basic prediction models are constructed using a self-attention mechanism neural network prediction algorithm (Transformer), a gated recurrent unit (GRU) prediction algorithm, and an XGBoost algorithm. These three basic prediction models are then trained using the time-aligned time-series sample data. Next, a meta-model is constructed using the LightGBM algorithm. Based on the prediction values ​​of the three basic prediction models, a 3-dimensional meta-feature vector is concatenated and used as the input variable of this meta-model. That is, the meta-feature vector is used as the new feature, and the true value of thallium content in the die head ash is used as the target to train the meta-model. After the meta-model training is completed, based on the new samples (the current process parameter data values ​​of the sintering production process and the test value of the thallium content in the die head ash at the current moment), three initial prediction values ​​are first obtained through the basic prediction model, and then input into the meta-model to obtain the final predicted value of the thallium content in the die head ash. In other words, this invention first trains three types of basic prediction models based on Transformer, GRU and XGBoost respectively, and then uses LightGBM to train the meta-model to weight and fuse the outputs of the three basic models to obtain a comprehensive prediction model, ultimately achieving real-time and accurate prediction of thallium content.

[0021] In this invention, from the perspective of process mechanism, the production process parameters that affect the thallium content in the sintering machine head include, but are not limited to, the chlorine and sulfur content of the sintering mixture, the thallium content of the mixture, the flow rate of the sintering mixture, the hourly coke consumption, the air volume and ignition temperature of the sintering main exhaust fan, the thickness of the material layer, the speed of the sintering machine, the negative pressure of the main flue, the exhaust gas temperature of the main flue, the sintering endpoint position and endpoint temperature, and the sintering air box temperature.

[0022] In this invention, historical process parameters of the sintering process and historical test values ​​of thallium content in the die head ash are generally acquired through a sintering process parameter acquisition module and a die head ash thallium content test value acquisition module. Specifically, the real-time values ​​of historical process parameters of the sintering process are recorded in the PLC system and can be accessed via OPC. The UA communication protocol reads the process parameters from the PLC. For example, after configuring the channel connecting to the PLC through Kepware software, the host computer reads the process parameters from the PLC through Kepware software and stores them in the host computer's database system. The collected data includes, but is not limited to: chlorine and sulfur content of the sintering mixture, thallium content of the mixture, flow rate of the sintering mixture, lime consumption per unit hour, coke consumption per unit hour, water addition to the sintering mixture, air volume of the sintering main exhaust fan, ignition temperature, material layer thickness, sintering machine speed, negative pressure of the main flue, exhaust gas temperature of the main flue, end temperature of the main flue, sintering end point location and end temperature, and sintering wind box temperatures (temperature of wind box #1 on the south side, wind box #3 on the south side, wind box #5 on the south side, wind box #7 on the south side, wind box #9 on the south side, wind box #1 on the north side, wind box #3 on the north side, wind box #5 on the north side, wind box #7 on the north side, and wind box #9 on the north side). The historical test values ​​of thallium content in the ash from the machine head are obtained by the server where the prediction model is located, which is directly connected to the production inspection and testing MES system via Ethernet. The server reads the test values ​​of thallium content in the ash from the machine head and the sampling time according to the data communication interface of the MES system, and stores them in the local database system of the server for the prediction model to call.

[0023] In this invention, the sintering production process is a continuous process with multiple variables, large time delays, and strong coupling. If the production process data is not time-aligned, the prediction model will be "causally reversed," leading to inaccurate predictions. This invention uses the sampling time of thallium content detection in the sintering machine head ash as the benchmark to calculate the time delay of each process parameter. According to the process flow, the process parameters can be divided into four steps, with the following process parameters for each step: Step 1: Chlorine and sulfur content in the sintering mixture, thallium content in the mixture, lime usage per hour, coke usage per hour, and average coke powder particle size. Step 2: Water addition to the sintering mixture. Step 3: Ignition temperature, sintering machine speed, and material layer thickness. Process 4: Main flue gas negative pressure, main flue gas temperature, main flue end temperature, south side #1 wind box temperature, south side #3 wind box temperature, south side #5 wind box temperature, south side #7 wind box temperature, south side #9 wind box temperature, north side #1 wind box temperature, north side #3 wind box temperature, north side #5 wind box temperature, north side #7 wind box temperature, north side #9 wind box temperature, etc.

[0024] Using the sampling time for thallium content detection in sintering mill head ash as a baseline, and based on the material flow time of the mixture from batching to mill head ash, assuming the data lag times for processes 1 to 4 are T1, T2, T3, and T4, and considering the actual production situation, the sampling times for processes 1 to 4 are calculated backward from the sampling time for the final pellet cold pressing strength test (in minutes): Process 1 = Sampling time for thallium content detection in mill head ash - T1; Process 2 = Sampling time for thallium content detection in mill head ash - T2; Process 3 = Sampling time for thallium content detection in mill head ash - T3; Process 4 = Sampling time for thallium content detection in mill head ash - T4. In other words, the time series of all process parameters are shifted "past" according to their respective lags, so that they correspond to the same batch of mixture. Only after time alignment of the sintering production process data can it be used for the prediction model.

[0025] In this invention, the input data of the self-attention mechanism neural network model (Transformer) consists of time-aligned historical data of the sintering production process and historical test values ​​of thallium content in the sintering machine head ash (i.e., time-series sample data). Specifically, the sintering production process parameters of n consecutive time nodes before the target prediction time T are combined with the historical test values ​​of thallium content in the machine head ash at the corresponding time to form a time-series sample. The training label value is set to the thallium content value of the machine head ash at time T+10min, thus forming a supervised learning mode of "historical time-series input - future target output". Its network structure includes: Input layer: The time-aligned sintering production process parameters and historical values ​​of thallium content in the machine head ash are combined to form a multi-dimensional feature sequence input model. Position encoding layer: Time position information is added to the input sequence, enabling the model to recognize the chronological order between different time nodes. Encoding Layer: Multiple Transformer encoding units are set up, each consisting of a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism is used to extract the correlation between different process parameters at different time scales, identifying key temporal features affecting the change in thallium content in the ash at the die head. The feedforward neural network is used for nonlinear mapping and enhancement of the features extracted by attention. Residual Connection and Normalization Layer: Residual connections and layer normalization are introduced in each encoding layer to improve model training stability and avoid degradation of deep networks. Output Layer: A fully connected layer is used to output the predicted value of thallium content in the ash at the target prediction time. The training process is as follows: The optimizer uses the Adam optimizer with an initial learning rate of 0.001~0.01; the loss function uses mean squared error (MSE) to minimize the error between the model's predicted value and the actual test value; in the training strategy, the batch size is set to 32~128, the number of epochs is set to 50~500, and an early stopping strategy is adopted, stopping training when the validation set error does not decrease significantly for 10~30 consecutive epochs, and saving the current optimal model parameters. The prediction process is as follows: First, the target prediction time T is received; second, the data sequence of n consecutive time nodes before time T is extracted from the time-aligned production process database and the historical database of thallium content in the machine head ash; then, this sequence is input into the trained Transformer prediction model, and the self-attention mechanism is used to complete the temporal correlation analysis and feature extraction; finally, the predicted value of thallium content in the machine head ash at time T+10min is output.

[0026] In this invention, the input data for the gate loop unit prediction model consists of time-aligned historical data of the sintering production process and historical test values ​​of thallium content in the sintering machine head ash (i.e., time-series sample data). Similar to the Transformer model, the input sequence is composed of the process parameters and historical thallium content values ​​of the machine head ash from n consecutive time nodes before the target prediction time T. The thallium content value of the machine head ash at time T+10 min is used as the training label to establish the time-series prediction model. Its network structure is as follows: Input layer: Inputs the time-aligned sintering production process feature sequence and the historical sequence of thallium content in the machine head ash. GRU hidden layer: One or more GRU networks are set up, each layer containing multiple (e.g., 64) neurons. The GRU unit mainly includes two gating structures: update gate and reset gate. The update gate controls the degree of retention of historical information, and the reset gate controls the fusion method between the current input and the historical state, thereby achieving dynamic modeling of the time-series change pattern. Fully connected output layer: Maps the output of the GRU hidden layer to the predicted thallium content value of the machine head ash at the target time. The training process is as follows: the optimizer uses the Adam optimizer, with an initial learning rate of 0.001~0.1; the loss function uses mean squared error (MSE) to minimize the error between the model's predicted value and the actual test value; the batch size is set to 16~64, and the number of epochs is set to 30~300 rounds; to avoid model overfitting, dropout layers can be set between the GRU hidden layers, and an early stopping strategy is used to save the model parameters with the best performance on the validation set (the number of early stopping rounds is 5~20 rounds). The prediction process is as follows: first, time-aligned data for n consecutive time nodes before the target prediction time T is obtained; second, this time-series data is input into the trained GRU prediction model, and the historical information and the current process state are dynamically fused by update and reset gates; finally, the predicted value of the thallium content in the die head at time T+10min is output. Compared with the Transformer model, the GRU model has fewer parameters, faster training speed, and can better capture short-cycle fluctuations and short-to-medium-term time-series dependencies in the sintering process.

[0027] In this invention, the input data for the XGBoost algorithm prediction model consists of time-aligned historical data of the sintering production process and historical test values ​​of thallium content in the sintering die head ash (i.e., time-series sample data). Considering the strong fitting ability of the XGBoost algorithm for tabular data and nonlinear relationships, the historical process data before the target prediction time T, the historical test values ​​of thallium content in the die head ash, and their statistical characteristics are used together to form the input feature vector. The thallium content value of the die head ash at time T+10 min is used as the training label to establish the XGBoost regression prediction model. The training process is as follows: First, features are constructed on the time-aligned time-series sample data to form a supervised learning sample set of "input features - target values"; second, the squared error is used as the objective function, and multiple regression trees are used to fit the prediction residuals of the previous round in each round to continuously reduce the model prediction error; preferably, the Adam optimizer is used, the initial learning rate can be set to 0.01~0.3, the number of trees can be set to 100~500, the maximum depth of the trees can be set to 3~10, the subsample sampling ratio can be set to 0.6~0.9, and the column sampling ratio can be set to 0.6~1; during the training process, the mean squared error is used as the loss function, the batch size is 32~128, the number of iterations is 50~500 rounds, and the validation set error monitoring and early stopping strategy are adopted. When the validation set error does not decrease significantly for 10~30 consecutive rounds, the training is stopped and the current optimal model parameters are saved. The prediction process is as follows: The sintering production process data aligned with the time before the target prediction time T, the historical value of the thallium content in the die head and its statistical characteristics are input into the trained XGBoost prediction model, and the predicted value of the thallium content in the die head at time T+10min is output.

[0028] In this invention, the input data for the LightGBM algorithm prediction model consists of time-aligned historical data of the sintering production process and historical test values ​​of thallium content in the sintering machine head ash (i.e., time-series sample data). Considering that LightGBM is more suitable for processing tabular feature data, the time-aligned process parameters before the target prediction time T, the historical values ​​of thallium content in the machine head ash, and their derived features are used to construct a feature vector. These derived features include, but are not limited to, statistical features such as the difference between adjacent time points, moving average, moving maximum, moving minimum, and fluctuation amplitude. The training label value is set to the thallium content value in the machine head ash at time T+10 min. The algorithm process is as follows: First, samples are constructed from the time-aligned production process data and the historical data of thallium content in the machine head ash to form a supervised learning sample set of "feature vector - target value". Second, the LightGBM algorithm is used to establish a regression prediction model. LightGBM is an ensemble learning algorithm based on gradient boosting decision trees. It continuously improves prediction accuracy by fitting the prediction residuals of previous rounds through multiple decision trees. During tree model growth, LightGBM employs an optimal splitting method based on histograms and a leaf-first growth strategy, demonstrating strong fitting capabilities for high-dimensional, multivariate, and nonlinear industrial process data. Finally, after multiple rounds of iterative training, the optimal LightGBM thallium content prediction model is obtained, which outputs the predicted thallium content in the machine head ash at the target time. The prediction process is as follows: First, extract the time-aligned process parameter data, historical test values ​​of thallium content in the machine head ash, and derived statistical features corresponding to the target prediction time T. Second, input this feature vector into the trained LightGBM prediction model. Finally, output the predicted thallium content in the machine head ash at time T+10 min. The LightGBM model can effectively uncover the complex nonlinear mapping relationship between various process parameters and the thallium content in the machine head ash, making it particularly suitable for rapid modeling and online prediction of high-dimensional, heterogeneous, and non-stationary tabular data in industrial production.

[0029] In this invention, a meta-learning fusion method is used to fuse the prediction results of the three basic prediction models based on the Transformer algorithm, GRU algorithm, and XGBoost algorithm, specifically using the Stacking method. The specific steps are as follows:

[0030] The first step is to generate meta-features. For the three base prediction models, cross-validation is used to generate their predicted values ​​on the training set. For example, the training set is divided into K folds, and each fold is used to train the base model with the other K-1 folds. Then, the current fold is predicted to obtain the output of that base model for the training sample. In this way, each training sample will obtain the predicted values ​​of the three base models, forming a 3-dimensional meta-feature vector.

[0031] The second step is to train the meta-model, which is constructed using the LightGBM algorithm. The meta-feature vector is used as the new feature, and the thallium content in the ash from the machine head is used as the target to train the meta-model.

[0032] In the algorithm fusion module, the meta-model is constructed using the LightGBM regression algorithm. First, K-fold cross-validation is used to generate training samples for the meta-model. Specifically, the training set is divided into K non-overlapping subsets. Each time, one subset is selected as the validation subset, and the remaining K-1 subsets are used to train the Transformer prediction model, GRU prediction model, and XGBoost prediction model, respectively. Then, the three trained basic prediction models are used to predict the current validation subset, obtaining the (Transformer prediction value, GRU prediction value, XGBoost prediction value) for the corresponding sample. These three prediction values ​​are combined to form a 3D meta-feature vector for the sample, with the true thallium content in the aircraft nose ash corresponding to that sample as the target value. This process is repeated until all subsets have completed one validation prediction, ultimately obtaining the complete meta-model training sample set. Next, the LightGBM meta-model is trained using the 3D meta-feature vector as input features and the true thallium content in the aircraft nose ash as the output target. Preferably, the LightGBM meta-model uses mean squared error (MSE) as the loss function, and its formula is as follows:

[0033] .

[0034] in, This represents the true value of thallium content in the machine head ash. These are the predicted values ​​from the meta-model. The sample size is denoted as _. Model parameters can be set as follows: initial learning rate (learning_rate) of 0.01~0.2, number of decision trees (n_estimators) of 50~400, number of leaf nodes (num_leaves) of 10~100, maximum tree depth (max_depth) of 2~10, and minimum number of data in a leaf node (min_data_in_leaf) of 5~50. During training, the LightGBM meta-model uses gradient boosting to train the decision trees round by round, continuously fitting the current prediction residuals. Simultaneously, it employs validation set error monitoring and early stopping strategies. When the validation set MSE shows no decreasing trend for 10~50 consecutive rounds, training stops and the current optimal meta-model parameters are saved. Finally, in the prediction phase, new samples are first processed by the Transformer prediction model, GRU prediction model, and XGBoost prediction model to obtain Transformer prediction values, GRU prediction values, and XGBoost prediction values, respectively. These three prediction values ​​are then combined to form a meta-feature vector, which is input into the trained LightGBM meta-model, ultimately outputting the predicted value of thallium content in the aircraft head.

[0035] The third step is to predict the new sample data. The new sample (the current value of the sintering production process data and the test value of the thallium content in the die head at the current moment) is first processed through all the basic models to obtain three output results, and then the results are input into the meta-model to obtain the final predicted value.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] 1. This invention employs a stacking ensemble learning framework that integrates the advantages of three different types of algorithms: Transformer, GRU, and XGBoost. Transformer and GRU can effectively extract the temporal correlation and long-term dependency features of process parameters, while XGBoost can fully explore the nonlinear correlation between process parameters and thallium content. Finally, the optimal fusion relationship of the three basic models is obtained through LightGBM meta-model learning, which can simultaneously meet the needs of temporal feature extraction and nonlinear relationship fitting. Compared with a single prediction model, it significantly improves the accuracy and stability of thallium content prediction, and can provide real-time and accurate data support for thallium pollution prevention and control in the sintering process and the classification and disposal of thallium-containing solid waste.

[0038] 2. This invention can directly output thallium content prediction results based on real-time collected sintering process parameters, without the need for lengthy laboratory testing. The prediction results are highly real-time, which can help steel companies adjust sintering process parameters in a timely manner, achieve precise classification and disposal of sintering machine head ash, and effectively reduce the risk of thallium pollution spread. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the prediction method of the present invention.

[0040] Figure 2 This is a comparison chart of the predicted values ​​of thallium content by the model and the actual test values ​​during a production period in Example 1 of the present invention. Detailed Implementation

[0041] The technical solution of the present invention will be illustrated below with examples. The scope of protection sought by the present invention includes, but is not limited to, the following embodiments.

[0042] Example 1

[0043] A multi-algorithm fusion method for predicting thallium content in sintering die head ash, comprising the following steps:

[0044] S1: Collect historical process parameters of the sintering process and historical test values ​​of thallium content in the die head ash. After time alignment processing of the historical process parameters and historical test values, a time-series sample dataset is obtained.

[0045] S2: Using time-series sample datasets, we train the basic models based on the Transformer algorithm, GRU algorithm, and XGBoost algorithm respectively to obtain three basic prediction models.

[0046] S3: A meta-model is constructed based on the LightGBM algorithm, and the prediction results of the three basic prediction models are used as meta-feature vectors. The historical test values ​​of thallium content in the ash at the machine head are used as targets to train the meta-model, resulting in a comprehensive prediction model that integrates multiple algorithms.

[0047] S4: Real-time acquisition of current process parameters during sintering, inputting them into the comprehensive prediction model, and outputting the predicted value of the current thallium content in the die head.

[0048] Example 2

[0049] Repeat Example 1, except that in step S1, the time alignment process is specifically as follows: historical process parameters and historical test values ​​are grouped and matched according to timestamps. Based on the sampling time of the historical test values, all historical process parameter data within a preset time interval before the time node are extracted, and the process parameters within the interval are arranged into a sequence of uniform dimensions according to time order. At the same time, missing values ​​and abnormal values ​​caused by equipment acquisition anomalies are removed, and the time dimension alignment of process data and test values ​​is completed so that they correspond to the same batch of sintering mixture.

[0050] Example 3

[0051] Repeat Example 2, except that in step S2, the basic prediction model based on the Transformer algorithm includes an input layer, a positional encoding layer, an encoding layer, a residual connection and normalization layer, and an output layer. Specifically: The input layer is used to input time-aligned time-series sample data into the model. The positional encoding layer adds temporal position information to the input time-series sample data to clarify the order of process parameters at different time points. The encoding layer consists of a multi-head self-attention mechanism sublayer and a feedforward neural network, used to capture the correlation features between process parameters at different time points and to perform nonlinear mapping and enhancement of features. The residual connection and normalization layer is used to alleviate the gradient vanishing problem during model training, improve model training stability, and avoid deep network degradation. The output layer outputs the predicted value of thallium content in the machine head ash at the target prediction time.

[0052] Example 4

[0053] Repeat Example 3, except that in step S2, when training the basic prediction model based on the Transformer algorithm: the Adam optimizer is used, and the initial learning rate is set to 0.001~0.01. The loss function is the mean squared error, minimizing the error between the model's predicted value and the actual test value. The batch size is set to 32~128, the number of iterations is set to 50~500 rounds, and an early stopping strategy is adopted, with 10~30 early stopping rounds.

[0054] Example 5

[0055] Repeat Example 4, except that in step S2, the basic prediction model based on the GRU algorithm includes an input layer, a GRU hidden layer, and an output layer. Specifically: the input layer is used to input time-aligned time-series sample data into the model. The GRU hidden layer includes at least one layer of a GRU network, each layer containing multiple neurons. Through update and reset gates, it selectively retains and forgets time-series information of process parameters at different time points, capturing the long-term dependence features of process parameters over time. The output layer, based on the time-series features extracted by the GRU hidden layer, outputs a predicted value of thallium content in the machine head ash at the target time.

[0056] Example 6

[0057] Repeat Example 5, except that a dropout layer is added between the GRU hidden layers to randomly forget the output of some neurons, thereby avoiding overfitting and improving the model's generalization ability.

[0058] Example 7

[0059] Repeat Example 6, except that in step S2, when training the basic prediction model based on the GRU algorithm: the Adam optimizer is used, and the initial learning rate is set to 0.001~0.01. The loss function is the mean squared error, minimizing the error between the model's predicted value and the actual test value. The batch size is set to 16~64, the number of iterations is set to 30~300 rounds, and an early stopping strategy is adopted, with 5~20 early stopping rounds.

[0060] Example 8

[0061] Repeat Example 7, except that in step S2, the basic prediction model based on the XGBoost algorithm includes an input layer, an iterative layer, and an output layer. Specifically: the input layer is used to input time-aligned time-series sample data into the model. The iterative layer performs multi-round iterations to split and grow process features, optimize the objective function, and explore the nonlinear correlation between different process parameters and thallium content, completing the key feature selection and importance ranking. The output layer maps the feature learning results to the predicted value of thallium content in the machine head ash at the target time.

[0062] Example 9

[0063] Repeat Example 8, except that in step S2, when training the basic prediction model based on the XGBoost algorithm: the Adam optimizer is used, the initial learning rate is set to 0.01~0.3, the number of trees is 100~500, the maximum tree depth is 3~10, the subsample sampling ratio is set to 0.6~0.9, the column sampling ratio is set to 0.6~1.0, the mean squared error is used as the loss function, the batch size is 32~128, the number of iterations is 50~500 rounds, and an early stopping strategy is adopted, with 10~30 early stopping rounds.

[0064] Example 10

[0065] Repeat Example 9, except that in step S3, the meta-model based on the LightGBM algorithm includes an input layer, a training layer, and an output layer. Specifically: The input layer is used to input a three-dimensional meta-feature vector, composed of the predicted values ​​from three basic prediction models, into the model. The training layer uses the actual measured value of thallium content in the machine head ash as the prediction target. Through multiple iterations, iterative learning is performed on the fusion weights of the predicted values ​​from multiple basic models to discover the prediction advantages of different basic models under different process scenarios and output the optimal fusion relationship. The output layer outputs the predicted value of thallium content in the machine head ash at the target prediction time based on the learned fusion relationship.

[0066] Example 11

[0067] Repeat Example 10, except that in step S3, when training the meta-model based on the LightGBM algorithm: the Adam optimizer is used, the initial learning rate is set to 0.01~0.2, the number of trees is 50~400, the number of leaf nodes is set to 10~100, the maximum depth of the trees is 2~10, the minimum number of samples in the leaf nodes is 5~50, the mean squared error is used as the loss function, the batch size is 16~128, the number of iterations is 100~300 rounds, and an early stopping strategy is adopted, with 10~50 early stopping rounds.

[0068] Application Example 1

[0069] After configuring the connection channel to the PLC using Kepware software, the host computer reads the sintering historical process parameters recorded in the PLC through Kepware software and stores them in the host computer's database system. The collected data includes, but is not limited to: chlorine and sulfur content of the sintering mixture, thallium content of the mixture, lime consumption per unit hour, coke consumption per unit hour, water addition to the sintering mixture, air volume of the sintering main exhaust fan, ignition temperature, material layer thickness, sintering machine speed, negative pressure of the main flue, exhaust gas temperature of the main flue, temperature at the end of the main flue, temperature of the south side #1 wind box, temperature of the south side #3 wind box, temperature of the south side #5 wind box, temperature of the south side #7 wind box, temperature of the south side #9 wind box, temperature of the north side #1 wind box, temperature of the north side #3 wind box, temperature of the north side #5 wind box, temperature of the north side #7 wind box, and temperature of the north side #9 wind box, etc. The system connects to the production inspection and testing MES system via Ethernet, and reads historical test values ​​and sampling times of thallium content in the machine head ash according to the data communication interface of the MES system, and stores them in the local database system of the server.

[0070] The obtained historical process parameters and historical test values ​​are grouped and matched according to timestamps. Based on the sampling time of the historical test values, all historical process parameter data within the preset time interval before the time node are extracted. The process parameters within the interval are arranged into a sequence with a unified dimension according to time order. At the same time, missing values ​​and outliers caused by equipment acquisition anomalies are removed. The time dimension of the process data and test values ​​are aligned so that they correspond to the same batch of sintering mixture, thus obtaining a time-series sample dataset.

[0071] The process parameters of n consecutive time nodes (sampling interval of 1 min, time window of 30 min, n=30) before the target prediction time T, along with the historical test values ​​of thallium content in the die head ash at the corresponding times, are arranged in chronological order to form a time series sample as the model input sequence. The thallium content value in the die head ash at time T+10 min is used as the training label. Transformer basic prediction model, GRU basic prediction model, and XGBoost basic prediction model are constructed respectively, where:

[0072] The basic prediction model of Transformer consists of an input layer, a positional encoding layer, an encoding layer, a residual connection and normalization layer, and an output layer. Its training process is as follows: the optimizer uses the Adam optimizer with an initial learning rate of 0.001; the loss function is the mean squared error (MSE), which minimizes the error between the model's predicted values ​​and the actual test values; in the training strategy, the batch size is set to 32, the number of epochs is set to 100, and an early stopping strategy is adopted, stopping training when the validation set error does not decrease significantly for 10 consecutive epochs, and saving the current optimal model parameters.

[0073] The basic GRU prediction model's network structure includes an input layer, GRU hidden layers, dropout layers between the GRU hidden layers, and an output layer. Its training process is as follows: The Adam optimizer is used, with an initial learning rate of 0.001. The loss function is Mean Squared Error (MES), minimizing the error between the model's predicted values ​​and the actual test values. The batch size is set to 32, the number of epochs is set to 100, and an early stopping strategy is employed: training stops when the validation set error does not significantly decrease for five consecutive epochs, and the current optimal model parameters are saved.

[0074] The XGBoost basic prediction model's network structure includes an input layer, an iterative layer, and an output layer. Its training process is as follows: The Adam optimizer is used, with an initial learning rate of 0.05, 300 trees, a maximum tree depth of 6, a subsample ratio of 0.8, a column sampling ratio of 0.8, mean squared error as the loss function, a batch size of 32, and 100 epochs. An early stopping strategy is employed: training stops when the validation set error does not significantly decrease for 10 consecutive epochs, and the current optimal model parameters are saved.

[0075] In the construction of the LightGBM metamodel: K-fold cross-validation is used to generate metamodel training samples. Specifically, 70% of the time-series sample dataset is taken as the training set, which is divided into K (K=5) non-overlapping subsets. Each time, one subset is selected as the validation subset, and the remaining K-1 subsets are used to train the Transformer, GRU, and XGBoost base prediction models, respectively. Then, the three trained base prediction models are used to predict the current validation subset, obtaining the Transformer, GRU, and XGBoost prediction values ​​for the corresponding samples. These three prediction values ​​are combined to form the 3D meta-feature vector of the sample, with the true value of thallium content in the aircraft head ash corresponding to that sample as the target value. This process is repeated until all subsets have completed one validation prediction, ultimately yielding the complete metamodel training sample set.

[0076] The LightGBM meta-model's network structure includes an input layer, a training layer, and an output layer. Its training process is as follows: The Adam optimizer is used, with an initial learning rate of 0.05, a decision tree count (n_estimators) of 200, a leaf node count (num_leaves) of 31, a maximum tree depth (max_depth) of 5, and a minimum number of data samples per leaf node (min_data_in_leaf) of 20. The mean squared error (MSE) is used as the loss function (the formula is: ,in, This represents the true value of thallium content in the machine head ash. These are the predicted values ​​from the meta-model. The sample size is 32, the batch size is 32, and the number of iterations is 200. During training, the LightGBM metamodel uses gradient boosting to train the decision tree round by round, continuously fitting the current prediction residual. At the same time, a validation set error monitoring and early stopping strategy are adopted. When the validation set MSE shows no decreasing trend for 30 consecutive rounds, training is stopped and the current optimal metamodel parameters are saved.

[0077] The prediction process is as follows:

[0078] The current sintering process parameters and historical test values ​​of thallium content in the die head ash before the current moment were first input into the trained Transformer, GRU, and XGBoost basic prediction models, yielding three preliminary predictions of thallium content in the die head ash after 10 minutes: 145.1 mg / kg, 152.4 mg / kg, and 144.9 mg / kg, respectively. These three preliminary predictions were then used as meta-feature vectors and input into the trained LightGBM meta-model, ultimately predicting a thallium content of approximately 150.3 mg / kg in the die head ash after 10 minutes. Sampling and testing revealed an actual thallium content of approximately 148.6 mg / kg at this moment. Therefore, the absolute error of the LightGBM meta-model was approximately 1.7 mg / kg, and the relative deviation was approximately 1.14%. The prediction deviations of the three basic prediction models were all greater than the prediction deviation of the LightGBM meta-model, indicating that the multi-algorithm fusion prediction method used in this invention can effectively improve prediction accuracy and stability.

[0079] Repeat the above process to predict the thallium content in the machine head ash during a production period, as shown in the attached figure. Figure 2 As shown, comparing the model predictions with the actual test values ​​reveals that the multi-algorithm fusion method for predicting thallium content in sintering machine head provided by this invention has high accuracy, with a maximum relative deviation of no more than 2.5%.

Claims

1. A method for predicting thallium content in sintering die head ash using a multi-algorithm fusion approach, characterized in that: The prediction method includes the following steps: S1: Collect historical process parameters of the sintering process and historical test values ​​of thallium content in the die head ash. After time alignment processing of the historical process parameters and historical test values, a time-series sample dataset is obtained. S2: Using time-series sample datasets, we train the basic models based on the Transformer algorithm, GRU algorithm, and XGBoost algorithm respectively. After training, we obtain three basic prediction models. S3: A meta-model is constructed based on the LightGBM algorithm, and the prediction results of the three basic prediction models are used as meta-feature vectors. The historical test values ​​of thallium content in the head ash are used as targets to train the meta-model, resulting in a comprehensive prediction model that integrates multiple algorithms. S4: Real-time acquisition of current process parameters during sintering, inputting them into the comprehensive prediction model, and outputting the predicted value of the current thallium content in the die head.

2. The prediction method according to claim 1, characterized in that: In step S1, the time alignment process specifically involves: grouping and matching historical process parameters and historical test values ​​according to timestamps; taking the sampling time of historical test values ​​as a benchmark, extracting all historical process parameter data within a preset time interval before the time node; and organizing the process parameters within the interval into a sequence with a unified dimension according to time order. At the same time, missing values ​​and outliers caused by equipment acquisition anomalies are removed, thus completing the time dimension alignment of process data and test values, so that they correspond to the same batch of sintering mixture.

3. The prediction method according to claim 1 or 2, characterized in that: In step S2, the basic prediction model based on the Transformer algorithm includes an input layer, a position encoding layer, an encoding layer, a residual connection and normalization layer, and an output layer. Specifically: the input layer is used to input time-aligned time-series sample data into the model; the position encoding layer adds time position information to the input time-series sample data to clarify the order of process parameters at different time points; the encoding layer consists of a multi-head self-attention mechanism sublayer and a feedforward neural network, used to capture the correlation features between process parameters at different time points and to complete the nonlinear mapping and enhancement of features; the residual connection and normalization layer is used to alleviate the gradient vanishing problem during model training, improve model training stability, and avoid deep network degradation; the output layer is used to output the predicted value of thallium content in the machine head ash at the target prediction time.

4. The prediction method according to claim 3, characterized in that: In step S2, when training the basic prediction model based on the Transformer algorithm: the Adam optimizer is used, the initial learning rate is set to 0.001~0.01; the loss function is mean squared error, which minimizes the error between the model prediction value and the actual test value; the batch size is set to 32~128, the number of iterations is set to 50~500 rounds, and an early stopping strategy is adopted, with 10~30 early stopping rounds.

5. The prediction method according to any one of claims 1-4, characterized in that: In step S2, the basic prediction model based on the GRU algorithm includes an input layer, a GRU hidden layer, and an output layer. The input layer is used to input time-aligned time-series sample data into the model. The GRU hidden layer includes at least one GRU network layer, each containing multiple neurons. Through update and reset gates, it selectively retains and forgets time-series information of process parameters at different time points, capturing the long-term dependence features of process parameters over time. The output layer outputs the time-series features extracted from the GRU hidden layer, mapping them to the predicted value of thallium content in the machine head ash at the target time. Preferably, a dropout layer is also set between the GRU hidden layers to randomly forget the outputs of some neurons, avoiding overfitting and improving the model's generalization ability.

6. The prediction method according to claim 5, characterized in that: In step S2, when training the basic prediction model based on the GRU algorithm: the Adam optimizer is used, the initial learning rate is set to 0.001~0.01; the loss function is mean squared error, which minimizes the error between the model prediction value and the actual test value; the batch size is set to 16~64, the number of iterations is set to 30~300 rounds, and an early stopping strategy is adopted, with 5~20 early stopping rounds.

7. The prediction method according to any one of claims 1-6, characterized in that: In step S2, the basic prediction model based on the XGBoost algorithm includes an input layer, an iteration layer, and an output layer. The input layer is used to input time-aligned time-series sample data into the model. The iteration layer performs splitting and growth on process features through multiple iterations, optimizes the objective function, explores the nonlinear correlation between different process parameters and thallium content, and completes the key feature screening and importance ranking. The output layer is used to map the feature learning results to the predicted value of thallium content in the machine head ash at the target time.

8. The prediction method according to claim 7, characterized in that: In step S2, when training the basic prediction model based on the XGBoost algorithm: the Adam optimizer is used, the initial learning rate is set to 0.01~0.3, the number of trees is 100~500, the maximum tree depth is 3~10, the subsample sampling ratio is set to 0.6~0.9, the column sampling ratio is set to 0.6~1.0, the mean squared error is used as the loss function, the batch size is 32~128, the number of iterations is 50~500 rounds, and an early stopping strategy is adopted, with 10~30 early stopping rounds.

9. The prediction method according to any one of claims 1-8, characterized in that: In step S3, the meta-model based on the LightGBM algorithm includes an input layer, a training layer, and an output layer. The input layer is used to input a three-dimensional meta-feature vector, composed of predictions from three basic prediction models, into the model. The training layer uses the actual test value of thallium content in the machine head ash as the prediction target, and iteratively learns the fusion weights of the predictions from multiple basic models through multiple iterations to explore the prediction advantages of different basic models under different process scenarios and outputs the optimal fusion relationship. The output layer outputs the predicted value of thallium content in the machine head ash at the target prediction time based on the learned fusion relationship.

10. The prediction method according to claim 9, characterized in that: In step S3, when training the meta-model based on the LightGBM algorithm: the Adam optimizer is used, the initial learning rate is set to 0.01~0.2, the number of trees is 50~400, the number of leaf nodes is set to 10~100, the maximum depth of the tree is 2~10, the minimum number of samples in the leaf node is 5~50, the mean squared error is used as the loss function, the batch size is 16~128, the number of iterations is 100~300 rounds, and an early stopping strategy is adopted, with 10~50 early stopping rounds.

Images