Accuracy improvement of the traditional rolling force and torque model powered by an industrial data platform
2022-10-28
Research Institute,Baoshan Iron & Steel Co.,Ltd.,Shanghai 201999,China
Abstract: The calculation results of the rolling force and torque model based on Orowan’s differential equation numerical solution method do not fit with the industrial measurements very well.In particular,a quite large deviation on the torque model was found.On the basis of analyzing the shortcomings of the existing method,an improved rolling force and torque model algorithm aided by the Process Integrated Data Application System platform is proposed.Accordingly,the calculation accuracy of the rolling torque model is improved.The improved models are verified by 1 711 136 records of a data platform.The improved models are also based on Orowan’s differential equation.Two coefficients,namely,friction factor and forward slip,are recognized as the crucial factors to be determined from industrial measurements to improve the accuracy.Therefore,the proposed method is a hybrid method that can be used to deeply understand the rolling process and improve the model’s accuracy by combining traditional plastic mechanics and data-driving global optimization algorithms.This paper proposes a new approach to studying theoretical rolling deformation models powered by the industrial data platform.
Key words: data platform; rolling force; torque; numerical method; hybrid strategy
1 Introduction
The rolling force and torque are the important rolling process parameters in plate mills.The cal-culation accuracy of the rolling force and torque model directly determines the process control level and represents the data analysis ability of the rolling process.Considerable research has been performed on the calculation of the rolling force and torque,among which the numerical method based on the rolling deformation differential equation is an important approach.With the help of the Process Integrated Data Application System (PIDAS) indus-trial data platform,the preliminary data exploration shows that the torque model calculation results from the existing numerical methods[1-4]on the rolling differential equation approach have a quite large devi-ation from plant measurements.Hence,it is necessary to develop a new calculation method for the rolling force and torque with improved accuracy.
Taking advantage of the rich data resources in the PIDAS data platform,our study approach is the soft measurement for the key parameters of rolling force and torque models based on mass data in the real world.First,the rolling deformation equation is estimated based on the principle of material deformation.Then,the shortcomings of existing methods on the rolling equation are shown and explained.To improve the model accuracy,a hybrid optimization method is developed,in which the fri-ction factor and neutral angle are the independent variables to be optimized by searching for the best fit with plant measurements.After obtaining the appro-priate model parameters,an improved rolling force and torque model calculation results are obtained.During the research and modeling procedure,the advantages of the availability of mass data,abound-ing algorithm packages,and high computing power provided by the PIDAS are fully utilized.
2 Approach
The symbols used in Orowan’s equation are shown in Fig.1.

Note:R-rolling radius,φ-angular coordinate of the arc of contact,p-horizontal pressure,q-vertical pressure,s-normal rolling pressure,τ-shear stress,h-plate thickness,φN-angular coordinate of the neutral plane,hN-thickness of the neutral plane.
2.1 Existing numerical methods and calculation results
OROWAN established the plate slice rolling differential equation as follows:

(1)
ALEXANDER[3]assumed that the horizontal pressure distribution of the plate slice was constant along the thickness,i.e.,Y=hp,and adopted the yield conditionq-p=2k.Accordingly,p=s-2k∓τtanφwas derived,and the differential equation is expressed as

(2)
where,Yis horizontal force per unit of the plate width;andkis yield stress in the plane strain.
Substituting the geometric relation betweenhandφinto Equation (2),we can get the following differential equation:
(3)
By settingτ=τ(s, 2k,φ),Equation (3) is converted to
(4)
The rolling force per unit of width is
(5)
The torque per unit of width is

(6)
On the basis of ALEXANDER’s study,we attempted to build models with the data provided by the PIDAS data platform and adopted the friction coefficient method from the literature[4].For the flow stress,the modified Miyasaka-Masuke formula[10]was adopted:
(7)

The comparisons of the calculated rolling forces and torques with the measured ones are shown in Fig.2.

Fig.2 Comparisons of the model calculation and measured values
The results show that the calculated rolling forces are close to the measured values,but the torque deviations are quite large.
Fig.3 shows the overall idea of this study.
2.2 Analysis of the torque model deviation
The deviation of the torque calculation can be attributed to at least two reasons.First,the tem-perature varies along the rolling contact arc length during hot-plate rolling.There is a steeply unsteady heat transfer between rolls and the plate,including heat conduction and deformation heat of the workpiece.The steel plate’s surface oxidation and pass descaling water make the rolling conditions highly complicated.Hence,it is difficult to know the temperature distribution exactly.Moreover,the plate temperature is not uniform along the rolling direction.As a result,the derivative of the material deformation resistance to the rolling angle is also affected by the temperature,which is not taken into account in the existing model.Second,the hori-zontal forces along the plate thickness are not com-pletely the same,especially at the rolling exit side.This case is shown in Fig.4,which also presents the two-dimensional finite element analysis results of the rolling process.

Fig.3 Overall idea of the present study
The inconsistency between the model’s sim-plified assumptions and the actual situation would affect the distribution of the friction stress calcula-tion and the location of the neutral plane.Subse-quently,it would affect the calculation accuracy of the torque integration formula.
As shown in the experimental curves of the fric-tion stress in the literature[9],the measured friction stress was continuously distributed along the rolling arc length and approached zero near the neutral plane,rather than switching from the maximum positive stress to the maximum negative one near the neutral plane in a cliff-like style as obtained in ALEXANDER’s method.

Fig.4 Two-dimensional finite element analysis results in the horizontal stress distribution of the rolling process
The shortcomings mentioned above could be solved by introducing a neutral angle as the undeter-mined variable to be optimized and improving the calculation steps.
2.3 Theoretical and data-driven hybrid method
2.3.1 Improved numerical solution method
The friction model proposed by WANHEIM and BAY et al.[5-8]was adapted to correlateτands.The model has been well verified from theoretical and experimental data.
In the rolling process,the critical condition parameters to determine sticking or sliding are the critical pressures′ and critical shear stressτ′:
(8)
where,fis the friction factor.
Whens≤s′,sliding friction occurs:
(9)
Whens>s′,sticking friction occurs:
τ=k[A1+(f-A1)(1-eAE)]
(10)
Based on the analysis and on-site observation of the variation of the forward slip speed,the unrea-sonable neutral angular causes the torque calculation result deviations.Hence,Grutev’s hysteresis zone was introduced into the models,as shown in Fig.5.The hysteresis zone is near the neutral plane,and its lengthlncan be calculated with the following formula:
(11)
where,ldis the contact length;andhmis the mean thickness.
In the hysteresis zone,the shear stressτis assumed to linearly change,and the angular coordinate of the neutral planeφNis to be determined.

Fig.5 Illustration of Grutev’s hysteresis zone
(12)
The calculation steps of the improved numerical solution method are as follows:
(1) The rolling angleφ1is cut intoMslices,and there are (M+1) calculation points.

(3) The positive pressures1at the entry side and positive pressures2at the exit side are calculated.
(4) Starting froms1to the exit side,point shear stressτand the corresponding partial derivatives are calculated using Equation (9) or (10),denotingτasτen(φ).Then,the ordinary differential equation (Equation (4)) is solved.The first symbol is taken for ± and ∓,andsis calculated,denoted assen(φ).
(5) Starting froms2to the entry side,the point shear stressτand corresponding partial derivatives are calculated according to Equation (9) or (10),denotingτasτex(φ),and then the ordinary differential equation (Equation (4)) is solved.The second symbol is taken for ± and ∓,andsis calculated,denoted assex(φ).
(6) According to Equation (12),pointsAandBare obtained fromτenandτex(negative),respectively,and the calculation points between pointsAandBare calculated to obtain the shear stressτand its partial derivative in the hysteresis zone.
(7) Then,the contact arc length is divided into three zones:the entry zone (from the entry toA),the hysteresis zone (fromAtoB),and the exit zone (fromBto the exit).In the entry zone,s=sen;in the hysteresis zone,sis recalculated according to Equation (4);in the exit zone,s=sex.
(8) By integrating Equation (5),the rolling pressurePpointing to the center of the rolling is obtained.
(9) By integrating Equation (6),the torqueGis obtained.
The examples of the calculation results are shown in Fig.6.
2.3.2 Optimization solution
In the preceding calculation steps,two variables need to be determined,namely,friction factorfand neutral angleφN.How to get the appropriate friction factor and neutral angle is a nonlinear constrained optimization problem.According to the collected rolling pass measurement values,i.e.,rolling forcePmeasand torqueGmeas,denoted as the rolling forcePcalcand torqueGcalccalculated in the preceding steps,the objective function of the optimization is built as follows:

Fig.6 Examples of the calculation results
(13)
When a local optimization algorithm is adopted,the optimization results sometimes fail to achieve the desired accuracy,which means obj>tolerance (according to the plate-mill common sense,the tolerance is set as 0.01).In this case,one of the global optimization algorithms is subsequently adopted.The following hybridized steps were designed to improve the optimization calculation procedure.The hybridized steps were described as UML activity diagram in Fig.7.

Fig.7 Hybridized steps for improving the optimization calculation procedure
Generally,a local optimization algorithm has a short computation time,but it sometimes converges to the local minimum solution.By contrast,a global optimization algorithm has a high computation time cost.Accordingly,hybridized steps were adopted.
The local optimization algorithm is Sequential Least Squares Programming[11-12]from the open-source package SciPy[13].The global optimization algorithm adopts the following types(Table 1) from open-source packages SciPy and NLopt[14].

2.3.3 Data preparation
Due to the integration of the PIDAS data platform for various automation systems[27],abundant plant process data need to be used.In a roughing mill,most passes belong to high-stocking rolling,and the shape factor of the deformation zone has a great influence where OROWAN’s equation is of low adaptability.In a finishing mill,passes belong to normal rolling,so we only chose the data on finishing mill passes with a shape factor greater than 1.2.
Rolling motors are accelerated or decelerated in the plate rolling thread-in or thread-out phase.Hence,the torque measurements include not only rolling torques but also acceleration torques that are not of concern.Thus,we chose the middle body of the plate rolling process data as the measurements of the rolling force and torque.
Process parameters,such as entry thickness,exit thickness,rolling radius,rolling flattening radius,and strain and strain rates,are derived from the L2 computer online model calculation.
The improved model’s key parameterKTrelated to the material in the flow stress formula was obtained by the regression statistics of the data of the same steel-grade productions.
2.3.4 Support of the PIDAS platform
The PIDAS platform provides not only rich and well-correlated production details data but also powerful computing capability support.Using mul-tiple multi-core processors,multiple global optimi-zation algorithms with a heavy CPU load can be run in parallel,which allows for running global optimization algorithms that consume time and computational loads.To make full use of the com-puting capabilities of a multi-core server,several processes of the global optimization algorithm are scheduled and undergo a parallel run.The powerful computing power support makes the algorithm calculation match the pace of on-site production.
The PIDAS platform provides rich open-source algorithm software packages,which can be con-venient in exploring and evaluating different optimization algorithms and greatly speeds up the research progress.
3 Results and discussion
3.1 Verification of the accuracy of the impro-ved torque model
The results of 1 711 136 calculations in the PIDAS data platform were statistically analyzed.The comparison results of the rolling force and torque calculation and measured values are shown in Fig.8.

Fig.8 Comparisons between the improved model calculation and measured values
Table 2 shows improved model calculation accuracy.

Table 2 Improved model calculation accuracy %
The results show that the calculated torques are very close to the measured values,but the rolling forces are slightly overall biased,which needs further study.
3.2 Comparison of the optimization algorithms
Through the analysis of the 1 711 136 calculation results in the PIDAS platform,the efficiency of the optimization algorithm was statistically analyzed.The results could be used to guide the re-factoring of the subsequent hybridized optimization strategy.
The effective acceptance rate (obj≤tol) of the local optimization SQL algorithm is records/total records=488 853/1 711 136=28.6%.
The average cost time was 5.54 s,and the standard deviation was 2.95 s.

The efficiency statistics of the global optimization algorithms are shown in Table 3.
Based on the above results of the local optimization and global optimization algorithms,considering our own needs,the subsequent hybridized optimization strategy for this problem is re-factored,which is described as UML activity diagram in Fig.9.

Fig.9 Re-factored hybridized optimization strategy
4 Conclusions
In this paper,a new approach is proposed to deeply understand the rolling process and improve the model’s accuracy by combining traditional plastic mechanics and data-driven global optimization algorithms.
(1) Through the improvement of the rolling deformation process of the numerical solution method of OROWAN’s equation using WANHEIM friction conditions,introducing the hysteresis zone,and freeing the neutral plane of certainty,the calculation accuracy can be effectively improved.The accuracy is verified through a fitting with actual plant measurements of the rolling force and torque.Hence,the proposed method is a reliable numerical method.
(2) Benefiting from the powerful support of the industrial data platform,long-term online parallel calculations based on a huge number of data were performed.Accordingly,an accurate and reliable optimization algorithm combination strategy for the classical numerical solution of the rolling deformation process was re-factored,which can provide a valuable reference for similar research works.
(3) The industrial data platform in the network and digital era provides not only rich production performance data but also rich algorithm software packages based on open-source software and sufficient computing resources.This paper shows the data analysis procedure,with the help of the industrial data platform,and the method of rich big data of plants integrated with classical metal plastic processing principles.It also digs out more reliable and exact industry domain knowledge.Furthermore,with the computing resource support of the industrial data platform,this paper shows how to identify and improve the efficiency,stability,and reliability of mathematical algorithms.
杂志排行
Baosteel Technical Research的其它文章
- Practice and numerical simulation of ingot solidification for high-performance homogeneous steel
- Effects of air-atomized water spray cooling device structure on the quenchingprocess,microstructure,and properties of wear-resistant steel
- Contributions to Baosteel Technical Research wanted
- Quantitative analysis of microstructure based on machine learning and itsapplication in product research and development
- Application of the gradient-based watershed algorithm in M/Aconstituent identification
- Intelligent decision making optimization of rolling mills through data prediction
