Sintered ore FeO prediction method based on machine learning algorithm
By expanding the dataset with WGAN and introducing CNN and RA into BiLSTM, the problems of insufficient data and insufficient model robustness in FeO prediction of sintered ore were solved, and high-precision FeO content prediction was achieved.
Patent Information
- Application Number
- CN202511643931.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-11
- Publication Date
- 2026-02-13
AI Technical Summary
Existing technologies for predicting FeO content in sinter suffer from limited data sample size, simplistic model structure, and insufficient robustness, resulting in low prediction accuracy and an inability to meet real-time control requirements.
The dataset is expanded using the WGAN method, and CNN and RA are introduced into BiLSTM to construct a BiLSTM fusion model. Through data preprocessing and feature extraction, accurate prediction of the sintering process is achieved.
It improves the diversity of the dataset and the robustness of the model, enhances the ability to capture long-term dependencies in the sintering process, and improves prediction accuracy and stability.
Smart Images
Figure CN121528347A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of steel smelting, and in particular to a sinter FeO prediction method based on a machine learning algorithm. BACKGROUND
[0002] As an important raw material for blast furnace ironmaking, the quality of sinter directly affects the ironmaking efficiency and energy consumption control. Among them, the FeO (ferrous oxide) content is one of the key indicators for evaluating the quality of sinter, which directly affects the reducibility, strength and softening performance of sinter. Accurate prediction of FeO content is of great significance for optimizing sintering process parameters, improving product quality and reducing production cost. Traditional FeO content prediction methods mainly rely on artificial experience formula or numerical calculation based on thermodynamic model, which has strong lag and poor adaptability, and is difficult to meet the real-time control demand.
[0003] In recent years, with the rapid development of machine learning technology, data-driven FeO content prediction methods have gradually become a research hotspot. Such methods collect process parameters (such as mixture ratio, sintering temperature, wind speed, etc.) and online detection data during the sintering process to build a prediction model for FeO content, so as to realize fast and non-destructive composition analysis. However, existing technologies still have significant limitations in practical application, which can be summarized as follows:
[0004] On the one hand, in terms of data samples, the number of data samples in existing research is limited. Due to the complexity of the sintering production process, which is affected by many factors such as raw material composition fluctuation, equipment operation state difference, operation process change, etc., it is difficult to obtain a large number of effective data samples. Limited data samples make it difficult for the model to learn comprehensive feature information during training, greatly limiting the generalization ability of the model, resulting in a significant decrease in prediction accuracy when the model faces new production conditions, and failing to meet the demand for accurate prediction of sinter FeO under different conditions in actual production.
[0005] On the other hand, from the perspective of model performance, existing prediction models have a single structure and generally lack robustness and accuracy. The common model architecture is a traditional simple neural network or linear regression model. Such models are difficult to fully capture the non-linear relationships between numerous influencing factors when faced with complex sintering processes. In actual sintering production, the production environment and conditions are in dynamic change, and the model is easily disturbed by noise data, abnormal conditions and other factors. When existing models deal with these disturbances, they often fail to maintain stable prediction performance, resulting in overfitting, making the model overly dependent on training data and having poor adaptability in real production scenarios. SUMMARY
[0006] In view of the deficiencies in the prior art, the application provides a sinter FeO prediction method based on a machine learning algorithm.
[0007] The application achieves the above technical object through the following technical means.
[0008] A sinter FeO prediction method based on a machine learning algorithm comprises the following steps:
[0009] Data related to sinter FeO is collected from a sinter production process, and after preprocessing of the collected data, a data set is formed; the data set is expanded by using a WGAN method; CNN and RA are introduced into BiLSTM to construct a BiLSTM fusion model; the improved BiLSTM model is trained by using the expanded data set; and the trained BiLSTM fusion model is used to predict the FeO content.
[0010] Further, the production data includes raw material composition, sintering temperature, sintering time, atmosphere condition and equipment operation parameter.
[0011] Further, the data preprocessing process is as follows: the collected data is cleaned, missing value is processed, abnormal value is identified and processed, and the data is normalized.
[0012] Further, the discriminator D objective function of the WGAN method is as follows:
[0013]
[0014] wherein the real data x is derived from a real data distribution , the random noise z is derived from a noise distribution , E(·) is an expectation, D(x) is a judgment result of the discriminator on a real sample, G(z) is a sample generated by the generator according to the noise, and D(G(z)) is a judgment result of the discriminator on the generated sample.
[0015] Further, the generator G objective function of the WGAN method is as follows:
[0016]
[0017] wherein the real data x is derived from a real data distribution , the random noise z is derived from a noise distribution , E(·) is an expectation, D(x) is a judgment result of the discriminator on a real sample, G(z) is a sample generated by the generator according to the noise, and D(G(z)) is a judgment result of the discriminator on the generated sample.
[0018] Further, the BiLSTM fusion model comprises an input layer, a CNN layer, a BiLSTM layer, an RA layer, a full connection layer, and an output layer.
[0019] Further, the CNN layer extracts features from the data through convolution and pooling operations.
[0020] Further, the BiLSTM layer includes a forward LSTM unit and a reverse LSTM unit, respectively capturing the information flow from the past to the future and from the future to the past in the time series.
[0021] Further, the RA layer dynamically allocates the importance of different time step features in the time series through learnable attention weights, and adds the BiLSTM layer output to the attention weighted features through residual connection.
[0022] Further, the performance evaluation indicators of the BiLSTM fusion model include mean absolute error, root mean square error, and mean absolute percentage error.
[0023] The beneficial effects of the present application are:
[0024] (1) The present application uses the method of WGAN to expand the data set. WGAN can deeply learn the latent distribution characteristics of the real data through the adversarial training of the generator and the discriminator, and solve the problem of insufficient data in the prediction of sinter FeO.
[0025] (2) The present application introduces convolutional neural network and residual attention mechanism into BiLSTM, constructs BiLSTM fusion model, considers the relative importance of different data, efficiently captures the internal relationship between long sequence data, effectively alleviates the gradient disappearance problem, accelerates model convergence, and has higher accuracy and better robustness. BRIEF DESCRIPTION OF DRAWINGS
[0026] Figure 1 The step flowchart of the present application is shown in the figure.
[0027] Figure 2 The network structure diagram of the BiLSTM fusion model in the embodiment is shown in the figure. DETAILED DESCRIPTION
[0028] The present application will be further described in conjunction with the drawings and specific embodiments, but the protection scope of the present application is not limited thereto.
[0029] Step 1: Data collection: Collect data related to sinter FeO from the sintering process, including raw material composition, sintering temperature, sintering time, atmosphere condition (such as oxygen content), equipment operation parameters, etc., to ensure that the data is representative and diverse, covering different production conditions and conditions. At the same time of data collection, artificial detection of FeO content is used as subsequent verification.
[0030] Data preprocessing: First, the collected data is cleaned, missing values are removed or filled, such as using interpolation method, mean substitution or filling according to the logic of the upper and lower content. Second, identify and process outliers, such as detecting outliers through statistical methods or machine learning algorithms, and decide whether to delete or correct. Finally, normalize the data and construct the data set.
[0031] Step 2: To solve the problem of insufficient data in sinter FeO prediction, a data augmentation method based on generative adversarial networks (GAN) is used to expand the data set. GAN is a deep learning model architecture composed of a generator and a discriminator, through the adversarial game between the two, the generator learns the data distribution and generates new data samples as realistic as possible, while the discriminator tries to distinguish between real data and generated data, eventually reaching a dynamic balance to achieve high-quality data generation and other applications. The objective function is as follows:
[0032] (1)
[0033] where, is the objective function of GAN optimization, D is the discriminator, G is the generator, the real data x comes from the real data distribution , the random noise z comes from the noise distribution , E(·) is the expectation, D(x) is the judgment result of the discriminator on the real sample, G(z) is the sample generated by the generator according to the noise, D(G(z)) is the judgment result of the discriminator on the generated sample.
[0034] However, in actual situations, GAN has the problem of gradient disappearance and unstable generated data. GAN introduces EM (Wasserstein) distance to measure the difference between the real data distribution and the generated data distribution , forming WGAN, the EM distance formula is:
[0035] (2)
[0036] where, is the joint distribution and is the set of , is a distribution in the set, x is the real data, y is the generated data, (x,y) is the data sampled from the distribution , is the distance between x and y.
[0037] According to the KR (Kantorovich-Rubinstein) dual principle, formula (2) can be equivalent to the following function form:
[0038] (3)
[0039] wherein, represents a 1-Lipschitz function, and f represents a distance mapping function.
[0040] The role of the discriminator D in the WGAN is to approximate the f function in the above dual form, and the goal is to maximize the difference between the real data and the generated data in the output, so as to estimate the Wasserstein distance. The objective function is:
[0041] (4)
[0042] The goal of the generator G in the WGAN is to minimize the output difference of the discriminator to the generated sample and the real sample, so that the generated distribution is as close as possible to the real distribution. The objective function is:
[0043] (5)
[0044] By inputting different random noises into the trained WGAN, the generator can generate various sinter-related data, thereby realizing the expansion of the data set.
[0045] Step 3: During the production process of the sintering furnace, the change of FeO content is affected by many complex factors, and presents a nonlinear change trend in the time process. Therefore, when performing the FeO content prediction task, not only the complex relationship between parameters such as raw material composition, sintering temperature, sintering time, etc. needs to be considered, but also the dynamic change in the time dimension must be fully considered when constructing the model. BiLSTM (Bidirectional Long Short-Term Memory) can efficiently capture the internal relationship between long sequence data and effectively alleviate the gradient disappearance problem. In addition, existing researches directly input data into the model for calculation, thereby ignoring the relative importance of evaluating different data. Therefore, the convolutional neural network (CNN) and residual attention mechanism (RA) are introduced into the BiLSTM to construct a BiLSTM fusion model.
[0046] The BiLSTM fusion model uses a CNN layer to extract spatial features of the data, a BiLSTM layer to capture dynamic changes in time series data, and a RA layer to capture the reinforcement ability of key spatiotemporal features, thereby improving the precision and stability of the model in handling long-term dependencies in the sintering process and achieving accurate prediction of the FeO content of sinter.
[0047] First, input the data through the input layer, and use the CNN layer to extract features from the input data. The convolutional layer slides along the time dimension with a one-dimensional convolution kernel to capture local feature patterns, and the pooling layer reduces the dimensionality of the features while retaining key features.
[0048] Second, input the features extracted by the CNN layer into the BiLSTM layer. The BiLSTM layer consists of forward and reverse LSTM (Long Short-Term Memory) units, which capture information flow from past to future and from future to past in time series, respectively. This bidirectional structure enables BiLSTM to more comprehensively understand the dynamic changes of data in time series and capture long-term dependencies.
[0049] Next, introduce the RA layer at the output end of the BiLSTM, dynamically allocate the importance of features at different time steps in the time series through learnable attention weights, and design a residual connection to add the BiLSTM layer output and the attention-weighted features. This not only strengthens key period information but also alleviates the gradient vanishing problem through the residual structure.
[0050] Finally, flatten the enhanced time series features and input them into the fully connected layer for feature compression and nonlinear combination, and then output the final prediction results through the output layer.
[0051] Step 4: Train and test the BiLSTM fusion model in step 3, with the following steps:
[0052] First, divide the expanded data set into training and test sets in an 8:2 ratio.
[0053] Second, train the BiLSTM fusion model parameters using the training set until the predetermined training rounds are reached; input the data in the test set into the BiLSTM fusion model to obtain the prediction results of the test set; and select the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) as performance evaluation indicators for the improved BiLSTM model, with the following formulas:
[0054] (6)
[0055] (7)
[0056] (8)
[0057] N represents the number of samples, and respectively represent the predicted value and the true value (artificially detected value), N, and The value range of N, MAE and MAPE is [0, ∞), and the smaller the RMSE value, the better the prediction ability of the model, and the smaller the MAE and MAPE values, the better the stability of the model.
[0058] Step 5: Collecting data in the actual sintering production process, inputting the BiLSTM fusion model tested in step 4 to obtain the FeO content in the actual sintering production process.
[0059] The embodiments are preferred embodiments of the present application, but the present application is not limited to the above embodiments, and any obvious improvements, replacements or modifications made by those skilled in the art without departing from the essential content of the present application shall fall within the protection scope of the present application.
Claims
1. A sinter FeO prediction method based on a machine learning algorithm, characterized in that: data related to sinter FeO is collected from the sinter production process, and after preprocessing the collected data, a data set is formed; the WGAN method is used to expand the data set; CNN and RA are introduced into BiLSTM to build a BiLSTM fusion model; the improved BiLSTM model is trained using the expanded data set; and the trained BiLSTM fusion model is used to predict the FeO content.
2. The machine learning algorithm based sinter FeO prediction method according to claim 1, characterized in that, The production data includes raw material composition, sintering temperature, sintering time, atmosphere condition, and equipment operating parameters.
3. The machine learning algorithm based sinter FeO prediction method according to claim 1, characterized in that, The data preprocessing process includes cleaning, missing value processing, identifying and processing outliers, and normalizing the data.
4. The machine learning algorithm based sinter FeO prediction method according to claim 1, characterized in that, The discriminator D objective function of the WGAN method is: where real data x comes from a real data distribution , random noise z comes from a noise distribution , E(·) is the expectation, D(x) is the discriminator’s judgment result on real samples, G(z) is the sample generated by the generator according to the noise, and D(G(z)) is the discriminator’s judgment result on the generated sample.
5. The machine learning algorithm based sinter FeO prediction method according to claim 1, characterized in that, The generator G objective function of the WGAN method is: where real data x comes from a real data distribution , random noise z comes from a noise distribution , E(·) is the expectation, D(x) is the discriminator’s judgment result on real samples, G(z) is the sample generated by the generator according to the noise, and D(G(z)) is the discriminator’s judgment result on the generated sample.
6. The machine learning algorithm based sinter FeO prediction method according to claim 1, characterized in that, The BiLSTM fusion model includes an input layer, a CNN layer, a BiLSTM layer, a RA layer, a fully connected layer, and an output layer.
7. The machine learning algorithm based sinter FeO prediction method according to claim 6, characterized in that, The CNN layer extracts features from the data through convolution and pooling operations.
8. The machine learning algorithm based sinter FeO prediction method according to claim 7, characterized in that, The BiLSTM layer includes forward and reverse LSTM units that capture information flow from past to future and from future to past in time series, respectively.
9. The machine learning algorithm based sinter FeO prediction method according to claim 8, characterized in that, The RA layer dynamically allocates the importance of different time step features in time series through learnable attention weights and adds the BiLSTM layer output to the attention weighted features through residual connection.
10. The machine learning algorithm based sinter FeO prediction method according to claim 9, characterized in that, The performance evaluation indicators of the BiLSTM fusion model include mean absolute error, root mean square error, and mean absolute percentage error.