Process control strategies for a steel making furnace using ANN with bayesian regularization and ANFIS

https://doi.org/10.1016/j.eswa.2009.06.056Get rights and content

Abstract

This paper illustrates the control strategies of an Electric Arc Furnace. It involves the prediction of the control action which aids in reduction of carbon, manganese and other impurities from the in-process molten steel. Predictive models using Artificial Neural Networks (ANN) with Bayesian Regularization and Adaptive Neuro Fuzzy Inference System (ANFIS) were developed. The control action is the amount of oxygen to be lanced at different sampling instants. The predictive models were constructed based on the values of the individual chemical constituents of the collected molten samples. Two control strategies were devised: one with full sampling and the other with limited or reduced sampling. For the full sampling case two predictive models were devised separately with ANN with Bayesian Regularization and ANFIS. For the limited sampling strategy a combination of ANN with Bayesian Regularization and ANFIS were employed. For full sampling strategy, ANFIS model performs better than ANN. The application of the limited sampling strategy gave satisfactory Mean Percentage Error (MPE) thereby justifying its practical implementation. The main advantage of reduced or limited sampling is that it helps in the reduction of cost, time and manpower associated the sample collection and its subsequent analysis.

Introduction

Proper control of a manufacturing facility is essential for the production of consistent good quality end product. A good control strategy also ensures less wastage, less breakdowns and also do guarantee enhanced safety for the personals involved. The literature on the development and use of control strategies is vast with areas ranging from mathematical model based control techniques (Billings et al., 1979, Ho and Chandratilleke, 1991) to that of intelligent control techniques involving techniques such as fuzzy logic (FL) (Kocaarslan et al., 2006, Lah et al., 2006, Shaocheng et al., 2005), Artificial Neural Networks (ANN) (Horng, 2008, Liu et al., 2007, Zarate and Bittencout, 2008) and neuro-fuzzy systems, an amalgamation of FL and ANN (Krause et al., 1994, Rao and Gupta, 1994, Tian and Collins, 2005). The choice of a particular control strategy depends on the process to be controlled, amount of information available pertaining to the process being controlled and also the practical and the economic feasibility of a particular control strategy. The implementation of mathematical model based control techniques requires the complete understanding of the process. A clear understanding of the inputs and outputs behavior and their interactions is imperative and as such one should be able to map the functional relationship between them. However, for most of the processes, a distinct mathematical functional relationship between the inputs and the outputs is hard to establish due to the complexity of the concerned process and incomplete and imprecise information pertaining to the process. In contrast to this, control strategies based on Artificial Intelligence (AI) techniques namely, FL and ANN does not require the in-depth understanding of the microscopic behavior of the concerned process. Hence, AI techniques are the preferred choice for modeling of complex and ill-understood processes. The mathematical details of AI based tools and techniques are given in Rajasekharan and Pai, 2003, Hagan et al., 1996, Hagan and Menhaj, 1994.

Due to the ease with which they can be developed and applied in real life situation, AI based control strategies have found wide range of applications in from aeronautics (Savran, Tasaltin, & Becerikli, 2006) to roller kiln involved in the manufacturing of ceramic tiles (Dinh & Afzulpurkar, 2007). ANN with linear filters and trained with back propagation (BP) algorithm was used for the designing of a controller for an unmanned research aircraft (Suresh & Kannan, 2008). ANN based control strategies have found widespread application in the area of electrical engineering. It has been used for modeling the behavior of an AC servo motor (Horng, 2008). ANN has also been used for development of a controller for controlling the speed of a synchronous motor drive (Elmas, Ustun, & Sayan, 2008) and induction motor drive (Ren & Chen, 2006). Apart from that ANN was employed for temperature control of a high voltage DC resistive drive (Yilmaz, Dincer, Eksin, & Kalenderli, 2007). Robotics is another field where ANN based control scheme has been widely used (Chatchanayuenyong and Parnichkun, 2006, Huang et al., 2008, Sato and Ishii, 2006, Thanh and Ahn, 2006). Other usage of ANN based control strategy include prediction of the parameters of gas metal arc welding process (Ates, 2007), development of control system for gun fire (Lee, 2007), development of a controller for an air conditioner (Liu et al., 2007), control of batch reactors in a chemical industry (Mujtaba, Aziz, & Hussain, 2006), and controlling air fuel ratio in a spark ignition engine (Wang, Yu, Gomm, Page, & Douglas, 2006). Control strategies based on the combination of neural network and fuzzy logic or neuro-fuzzy approaches were also cited in the literature (Rao & Gupta, 1994). Application of neuro-fuzzy controller ranges from airbag controller design for automobiles (Mon, 2007) to design of controllers for power plants (Alturki & Abdennour, 1999). Apart from these, neuro-fuzzy controllers have found application in refuse incineration plant, flexible manipulator system and in the field of robotics (Aguilar et al., 2003, Naadimuthu et al., 2007, Rao and Gupta, 1994, Tian and Collins, 2005).

Table 1 lists some of the control strategies adopted in Electric Arc Furnace (EAF) and other steel making furnaces. As evident, majority of those are based on mathematical model developed from first principles. A couple of them dealt with the control strategy of electrode positioning controller (Billings et al., 1979, Nicholson and Roebuck, 1972). Control of EAF off-gas process has also been carried out (Bekker et al., 2000, Kirschen et al., 2006). Other applications include estimation of tap temperature (Fernandez, Cabal, Montequin, & Balsera, 2008) and designing of set-point controllers for an EAF cooling system (Shinohara & Goodall, 2004). Estimation of the chemical composition of the final steel alloy (Ekmekci, Yetisken, & Camdali, 2007) with the aid of mass balance modeling and estimation of the various output variables of a Basic Oxygen Furnace (BOF) (Kubat, Taskin, Artir, & Yilmaz, 2004) were also carried out.

In this study, a control strategy for the EAF has been attempted with the aid of AI tools such as FL and ANN. ANN model which is an emulation of the biological neuron system and Adaptive Neuro Fuzzy Inference System (ANFIS) (Melin and Castillo, 2005, Mon, 2007, Shing and Jang, 1993), a fusion between ANN and FL were employed for prediction of the control action of the EAF. Following the developed AI tools, several models have been developed for predicting amount of oxygen to be lanced into the EAF. The methodology for the developments is presented in Section 2. Two types of models are developed using full sampling and limited sampling scheme. The application of these models is described in Section 3. The possible implications of the study to the industry have been discussed in Section 4 followed by the conclusions in Section 5.

Section snippets

Research setting

The work carried out consists of the prediction of control action in the EAF of a Steel Making Shop (SMS) for an integrated steel plant. An EAF is used for refining the molten metal which is subsequently converted into steel in the form of billets, blooms or slabs Fig. 1 depicts the process flow of the concerned SMS. The EAF under consideration is of 40 ton capacity and uses molten metal, Direct Reduced Iron (DRI) and scrap as the main raw material or charge composition. Along with molten metal,

Data collection and analysis

The methodology developed above is employed for prediction of the control action, the amount oxygen to be lanced at each sampling instant. The subsequent subsections delve in detail about the data being collected and analysis being carried out.

Industrial implications

The usage of the ANN and ANFIS models in prediction of the control action may ultimately aid in the abolition of the manual regulatory control action being carried out by the operators. Manual regulatory control based on human expertise is considered a safe and proven approach provided the expert opinion is emanated from an experienced operator or supervisor well conversant with the process. Thus manual regulatory control is totally subservient to human cognition which has the potential to

Conclusions

The study involved devising control strategies for a EAF. ANN with Bayesian Regularization and ANFIS were used for development of the control strategies. AI techniques can model the process based on process history data without requiring deep and intense knowledge about the process concerned. Two control strategies were proposed, one based on full sampling and the other based on limited sampling. For the full sampling case, two predictive models were developed, one based on ANN with Bayesian

Acknowledgement

The authors gratefully acknowledge the help and cooperation provided by the management and operating staff of the SMS during the course of this study and Mr. V. V. Khanzode, Research Scholar, IEM, IIT Kharagpur during the analysis.

References (42)

  • S.N. Huang et al.

    Adaptive neural network algorithm for control design of rigid-link electrically driven robots

    Neurocomputing

    (2008)
  • M. Kirschen et al.

    Mathematical modelling of heat transfer in dedusting plants and comparison to off-gas measurements at electric arc furnaces

    Energy

    (2006)
  • I. Kocaarslan et al.

    A fuzzy logic controller application for thermal power plants

    Energy Conversion and Management

    (2006)
  • B. Krause et al.

    A neuro-fuzzy adaptive control strategy for refuse incineration plants

    Fuzzy Sets and Systems

    (1994)
  • C. Kubat et al.

    Bofy-fuzzy logic control for the basic oxygen furnace (BOF)

    Robotics and Autonomous Systems

    (2004)
  • W. Liu et al.

    A neural network evaluation model for individual thermal comfort

    Energy and Buildings

    (2007)
  • P. Melin et al.

    Intelligent control of a stepping motor drive using an adaptive neuro–fuzzy inference system

    Information Sciences

    (2005)
  • Y.J. Mon

    Airbag controller designed by adaptive-network-based fuzzy inference system (ANFIS)

    Fuzzy Sets and Systems

    (2007)
  • I.M. Mujtaba et al.

    Neural network based modelling and control in batch reactor

    Chemical Engineering Research and Design

    (2006)
  • G. Naadimuthu et al.

    Application of an adaptive neural fuzzy inference system to thermal comfort and group technology problems

    Computers and Mathematics with Applications

    (2007)
  • H. Nicholson et al.

    Simulation and control of electrode position controllers for electric arc furnaces

    (1972)
  • Cited by (43)

    • Performance optimisation of forward-osmosis membrane system using machine learning for the treatment of textile industry wastewater

      2021, Journal of Cleaner Production
      Citation Excerpt :

      Kim and Park (Byungwhan and Jang Hyun, 2002) has concluded that ANFIS model performed better in predicting the response than RSM while, the ANN model performed better in the findings of Das et at. ( Das et al., 2010). Therefore, this paper involves modelling the FO process for the treatment of textile industry wastewater through RSM, ANN and ANFIS methods to evaluate their efficiency.

    • A new time–frequency method for identification and classification of ball bearing faults

      2017, Journal of Sound and Vibration
      Citation Excerpt :

      This approach was also successful in improving the computational time for the classification. In order to illustrate the potential application of proposed bearing fault diagnosis procedure, a comparative study between the present work and various bearing fault diagnosis techniques that use the CWRU data for bearing fault diagnosis is presented in Table 9 [67–73]. The comparing items include the machine elements used, fault type, fault severity levels, motor speed condition, feature extraction method, feature reduction method and classifier used, classified states and maximum classification efficiencies.

    • Predictive capability evaluation of RSM, ANFIS and ANN: A case of reduction of high free fatty acid of palm kernel oil via esterification process

      2016, Energy Conversion and Management
      Citation Excerpt :

      Contrary to the findings of Kim and Park [25] that ANFIS is more accurate in prediction capability than RSM, Taheri et al. [24] observed the reverse. Das et al. [26] concluded in their work that ANFIS performed better than ANN, whereas both techniques were reported to have performed equally well by Karimi et al. [23]. However, there is a dearth of information on comparison of predictive capabilities of RSM, ANN and ANFIS under the same investigation in order to resolve these inconsistent conclusions, especially as it relates to esterification and transesterification processes.

    • An online BOF terminal temperature control model based on big data learning

      2023, Journal of Iron and Steel Research International
    View all citing articles on Scopus
    View full text