Double-hyperplane fuzzy classifier design for tendency prediction of silicon content in molten iron
Introduction
A blast furnace is the key unit for continuous molten iron production under various conditions such as use of coke and hot air. The detailed structure of the blast furnace and its detailed process are shown in Fig. 1. It is very difficult to control a blast furnace because of its complex nonlinear and long time delay characteristics. The silicon content in molten iron not only reflects the thermal state of the blast furnace but also indicates the quality of molten iron. Therefore, if the development trend of the silicon content in molten iron can be predicted accurately, it is of great significance for the automatic control of a blast furnace. With the continuous enhancement of data availability regarding the blast furnace, many data-driven methods have emerged for modeling the silicon content in molten iron.
In the early days, some linear data-driven methods were used to model the silicon content in molten iron [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13]. Then many nonlinear methods were proposed considering the complex nonlinearity of the blast furnace. A neural network is the method most often adopted. Chen et al. [15] proposed the use of a genetic algorithm to optimize the neural network. The complexity of a neural network is closely related to the performance of the model. So to balance them, Saxén et al. [14] presented a predator-prey model that modifies the weight between layers of the neural network by a Kalman filter. The generalization ability of the final model was improved. Obviously, it did not consider that the silicon content of molten iron is affected by the lag time of the blast furnace. In [16], a sparse neural network was obtained with the pruning technique considering the lag time. Unfortunately, sparse neural networks cannot be generated automatically. Later, Nurkkala et al. [17] further improved the performance of the above-mentioned algorithm by using the Pareto boundary. David et al. [18] used a simplified neural network to show that the silicon content of hot metal is positively correlated with the theoretical flame temperature, the pressure blow, and the coke rate, and is negatively correlated with the hot metal rate. Zhou et al. [19] presented a robust data-driven modeling method for the online estimation and control of molten iron quality. In [20], the silicon content prediction model for molten iron is approximated by a Wiener model identification using a linear subspace identification algorithm. In [21], a piecewise linear switching system is used for the blast furnace, and a self-organizing-based model is proposed.
The generalization ability of the model using the empirical risk minimization mentioned above cannot be guaranteed. So some modeling methods based on structural risk minimization are used, such as least-squares support vector machines (SVMs) with radial basis kernels [22], support vector regression with a multiple-function kernel [24], a sliding window [25] or improve its performance by chaos particle swarm optimization [23], ν-SVM [26] that its output with probability, SVM that its output class is coded by binary [27], adaptive least-squares SVM [28]. The fact proved that structural risk minimization–based models are better than empirical risk minimization–based models.
Some fuzzy modeling methods have been proposed that make a model has better interpretability. In [29], the fuzzy method using the Stone-Wilstrath theorem is used to model the relationship between input and output. In [30], the fuzzy c-means clustering model is used to search for the optimal control center of the blast furnace thermal state. To have a more intuitive understanding of the changing trend of silicon content in molten iron, a three classification fuzzy model–based SVM was proposed [31]. Saxén et al. [32] systematically reviewed the development of silicon content prediction models. In addition, there have been some studies on the simultaneous modeling of multiple indexes in molten iron [33], [34].
Recently, blast furnace data were visualized in two dimensions by means of t-distributed stochastic neighbor embedding [35], [36]. It was found that the data for blast furnace problems have a crossing feature, which leads to it being hard to classify them accurately by traditional classifiers such as least-squares and large margin–based classifiers. The discovery of relevant blast furnace data characteristics can well guide the design of classifiers. Li et al. [35] presented a fuzzy classifier with nonparallel hyperplanes (NHFC) that can conquer the intractable problem but that preserves the advantages of fuzzy classifiers. Here we also only focus on whether the silicon content in molten iron increases or decreases. The tendency prediction for the blast furnace is to judge whether the blast furnace is cooling or warming, which is enough for the control of the blast furnace. Therefore, the modeling process can be regarded as a data-driven binary classification modeling problem, and the model has a great guiding role for blast furnace operators. On the basis of the advantage of the fuzzy model, we propose a new nonparallel hyperplane–based fuzzy classifier (i.e., a double-hyperplane fuzzy classifier [DHFC]) that solves two smaller quadratic programming problems by conquering the cross-classification problem.
The contributions of this article are twofold: (1) A DHFC is proposed. The design of the classifier fully takes into account the cross-characteristics of blast furnace data, so it has a good classification effect. (2) The formulation of the proposed classifier is a pair of smaller quadratic programming problems that work faster than an SVM-based fuzzy classifier, which is a large quadratic programming problem.
The remainder of this article is organized as follows. In Section 2, the basic principle of the fuzzy classifier is recalled. In Section 3, the DHFC is introduced. The application and a discussion of the new fuzzy classifier in the blast furnace are given in Section 4. Section 5 provides a summary.
Section snippets
Fuzzy classifier
The fuzzy classifier consists mainly of two parts: the fuzzy antecedent and the fuzzy consequent [37]. Firstly, the input space is divided into specific fuzzy groups based on the fuzzy antecedent. Then each fuzzy region is described by a linear function of inputs based on the fuzzy consequent. As long as the two parts are determined, the fuzzy classifier can be obtained according to the positive and negative fuzzy output. Compared with the traditional classifier, the fuzzy classifier can
Double-hyperplane fuzzy classifier
Taking into account the crossing feature of the data for blast furnace problems, Li et al. [35] recently proposed a fuzzy classifier with nonparallel hyperplanes (NHFC) that solved the intractable problem but retained the advantage of fuzzy classifiers. In this article, we propose a new nonparallel hyperplane-based fuzzy classifier. The new classifier also aims at generating two nonparallel planes where each plane is as close as possible to one class and is as far as possible from the other.
Model input and data preprocessing
It is very critical for accurately modeling the silicon content in molten iron that determine the model input. The effect of the lag time of all related variables on the result of the fuzzy model is different. So the model input variables and the lag time of their actions must be determined accurately. The details of the input variables are shown in Table 1, with data from [18], [21], [26], [27], [28]. h is the shift operator. For example, indicates the previous moment of a variable. In the
Conclusion
Taking into account the crossing feature of the binary classification problem for a blast furnace, we have presented a new nonparallel hyperplane–based fuzzy classifier (i.e., the DHFC). The DHFC can not only conquer the intractable cross-classification problem but also retains the good interpretability of fuzzy classifiers. Besides, the two hyperplanes can be easily obtained by just solving two small quadratic programming problems. The validity of the algorithm was proved through simulation
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
This work was supported in part by the National Natural Science Foundation of China under Grant 61751309, Grant 62073276, Grant 61933009, Grant 61825304, and Grant 61703361; in part by the Key Projects of Science and Technology Plan of Colleges and Universities of Hebei Provincial Department of Education under Grant ZD2019031; and in part by the Natural Science Foundation of Hebei Province under Grants F2020203062, F2019111009.
References (48)
- et al.
Application of MIMO identification to a blast furnace
- et al.
VARMAX modelling of blast furnace process variables
Eur. J. Oper. Res.
(1996) - et al.
State realization with exogenous variables - a test on blast furnace data
Eur. J. Oper. Res.
(1996) - et al.
Improvement of identification of blast furnace ironmaking process by outlier detection and missing value imputation
J. Process Control
(2009) - et al.
Identification of switching linear systems using self-organizing models with application to silicon prediction in hot metal
Appl. Soft Comput.
(2016) - et al.
Prediction of silicon content in hot metal using support vector regression based on chaos particle swarm optimization
Expert Syst. Appl.
(2009) - et al.
Self-tuning fuzzy modeling with adaptive membership function, rules, and hierarchical structure based on genetic algorithm
Fuzzy Sets Syst.
(1995) - et al.
An incremental support vector machine trained TS-type fuzzy system for online classification problems
Fuzzy Sets Syst.
(2011) - et al.
Identification of multiinput multioutput transfer function and noise model of a blast furnace from closed-loop data
IEEE Trans. Autom. Control
(1974) - et al.
Modeling, prediction and control of blast furnace operation from observed data by multivariate time series
Blast furnace modeling and control by the DDS method
The adaptive autoregressive models for the system dynamics and prediction of blast-furnace
Chem. Eng. Commun.
Adopt three criterions to choose MISO prediction model of blast furnace process
Short-term prediction of silicon content in pig iron
Can. Metall. Q.
On the development of predictive models with applications to a metallurgical process
Ind. Eng. Chem. Res.
Prediction of silicon content in blast furnace hot metal using partial least squares
ISIJ Int.
Time-Varying event-internal trends in predictive modeling methods with applications to ladlewise analyses of hot metal silicon content
Ind. Eng. Chem. Res.
Evolving nonlinear time-series models of the hot metal silicon content in the blast furnace
Mater. Manuf. Process.
Prediction and control for silicon content in pig iron of blast furnace by integrating artificial neural network with genetic algorithm
Ironmak. Steelmak.
Nonlinear prediction of the hot metal silicon content in the blast furnace
ISIJ Int.
Nonlinear modeling method applied to prediction of hot metal silicon in the ironmaking blast furnace
Ind. Eng. Chem. Res.
Artificial neural network model for predict of silicon content in hot metal blast furnace
Mater. Sci. Forum
Data-driven robust M-LS-SVR-based NARX modeling for estimation and control of molten iron quality indices in blast furnace ironmaking
IEEE Trans. Neural Netw. Learn. Syst.
Wiener model identification of blast furnace ironmaking process
ISIJ Int.
Cited by (5)
Improved algorithm of extreme gradient boosting for predicting silicon content in large proportion pellet smelting process
2024, Metallurgical Research and TechnologySilicon Content Control for Blast Furnace: Model-free DDQN-ICM Method
2023, Proceedings - 2023 China Automation Congress, CAC 2023