Research,and,application,of,the,digitalization,of,the,production,process,design,for,plate,steels

,

Research Institute,Baoshan Iron & Steel Co.,Ltd.,Shanghai 201999,China

Abstract: There are multiple processes and corresponding parameters in steel production,and combinations of these comprise various process routes.Different steel products require distinct process routes due to variations in performance targets.Thus,how to accurately set each key process parameter in certain process routes is an ongoing conundrum,because it not only requires a wealth of expert experience but also generates additional costs from the trial productions.In this paper,a new production design system for plate steels is proposed.The proposed system consists of methodology and function development.For methodology,multi-task Elastic Net,clustering,classification,and other methods are used to design process routes.Furthermore,the results are expressed in the form of parameter confidence intervals,which are close to practical application scenarios.For function development,the steel plate process route design function is developed on the Process Intelligent Data Application System (PIDAS) intelligent big data platform.The results demonstrate the method’s practical application value.

Key words: plate steels; process design; digitalization; big data

Steel production goes through complex production procedures,including multiple processes and corre-sponding parameters,to turn raw slabs into final products.Taking the production of heavy plates as an example,the production of heavy plates primarily consists of several procedures,such as slab selection,reheating,rolling,accelerated cooling after rolling,hot leveling,air cooling,shearing,and cold leveling.The processes and corresponding parameters in the production of different products often vary,thus leading to the phenomenon of multiple process routes occurring simultaneously.Regardless of which pro-cess route is selected,the ultimate goal is to ensure that the resulting steel plates meet the requirements (e.g.,performance,dimensions,quality,etc.).In particular,the mechanical properties related to performance requirements include tensile strength,yield strength,impact energy,and so on.

There are two main factors that can affect the quality of steel products:(1) the composition of the steel material (i.e.,the contents of different elements) and (2) various process parameters on the process route,such as the finish rolling tem-perature,slab discharging temperature,rough rolling entrance and exit temperatures,and finish rolling entrance and exit temperatures,etc.

Regarding the research and development (R & D) of new steel products,the R & D personnel hope to design an effective process route according to slab information (dimensions,composition,etc.) to obtain products that meet the end-users’ needs.In addition,these products must also meet cor-responding requirements,including performance,dimensions,quality,etc.This part of the work depends largely on the experience of the R & D per-sonnel.Thus,apart from designing the required processes and the corresponding parameters of each process according to the requirements of the product before the trial production,it is also required to design remedial measures in the case of unexpected product results.This puts forward higher experience requirements on the R & D personnel and necessitates the use of computer technology to help the R & D personnel in the design process.However,many studies have focused on the performance prediction and quality judgment of steel products[1],and only a few have investigated the parameter design of the entire process route.Therefore,it is imperative to set the design function of the steel plate process route on the big data platform serving factory production.

As an intelligent big data platform for heavy plate departments,the Process Intelligent Data Appli-cation System (PIDAS) features several functions,such as data collection,data visualization,model-ing,and calculation[2].It is designed to allow operators and technicians to quickly screen data,understand real-time production conditions,and guide production through the built-in models on PIDAS.The purpose of the current work is to build a design function of the steel plate process on the PIDAS platform,using historical information to assist technicians in designing process routes before the trial production of new products.Such a tool can improve the efficiency of R & D and trial production.At the same time,the new product process route based on production history data mining has a strong reference significance to help ensure the product quality and yield of the steel plate.

In this paper,the research and application of steel plate process route design are studied,including methodology and function development.The rese-arch idea of the production design system for plate steels is shown in Fig.1.

Fig.1 Research idea of the production design system for plate steels

The core of the production design is the utilization of the chemical composition of the steel,along with slab details and some basic information about the production line to design a process route that meets product requirements.The route includes the required processes and the target parameters of each process.This research mainly explores the topic from two aspects:(1) data selection and processing and (2) process route prediction.

The first aspect is data selection and processing.The data are derived from the historical production data found in the database.These include variables in the slab-to-product process such as dimensions (e.g.,slab width and slab thickness),chemical composition (e.g.,w(C) andw(Mn)),rolling process parameters(e.g.,start rolling temperature and finish rolling temper-ature),and performance parameters (e.g.,tensile strength,yield strength,and impact energy).The types of data vary and include continuous and categorical variables.After selecting all the required data,it is necessary to perform data preprocessing on the dirty data;otherwise,the modeling accuracy may be negatively affected,leading to results that lack credibility.Common data problems include missing data,data noise,data anomaly,data dupli-cation,data redundancy,and data imbalance.Data must be observed by means of charts,statistics,etc.,to identify the common data problems.Furthermore,appropriate mechanisms must be used to solve the problems.

Next,the process route is predicted.The core of the process route prediction is to predict the key process parameters.This is divided into the follow-ing six steps:

(1) Cluster the cleaned data.Due to the wide variety of steels and the complex process lines,it is necessary to cluster the data to accurately grasp the data characteristics and obtain more accurate predic-tions of process parameters.As the product target has a great influence on the process design,the weight of the product target is increased in the pro-cess of clustering.A variety of clustering methods (Kmeans,Mean shift,DBSCAN),combined with clustering criteria (Calinski,Silhouette,Davies),can be used to obtain the best clustering result.

(2) Regression prediction for each process par-ameter.The data are divided into input variables and target variables.Input variables include chemi-cal composition,performance parameters,raw mate-rial parameters,etc.,while target variables include process parameters,product parameters,etc.The multi-task Elastic Net method is used to perform regression analysis on the input variables.The regular term in Elastic Net[3]combines the regular term of the Lasso method with that of the Ridge method,thus incorporating the advantages of both methods.Efficiently performing regression fitting ensures that important features are screened and features that have less impact are deleted.In addition,Elastic Net has a good group effect com-pared to the Lasso method,which does not generate such an effect.The basic regression model is expressed as:

Y=XTβ+ε

(1)

where,Y=(y1,y2,…,yn)T;X=(X1,X2,…,Xn);Xi=(xi1,xi2,…,xip)T(i=1,2,…,n);β=(β1,β2,…,βp)T;ε=(ε1,ε2,…,εp)T;nis the sample size;andpis the data dimension.

The expression for Elastic Net regression is as follows:

(2)

where,L1andL2are the regularization degree flags;andαis the regularization depth flag.

In the multi-task Elastic Net,it is necessary to select the appropriate ratio ofL1toL2,in which the closer it is to 1,the greater the Lasso characteri-stics,whereas the closer it is to 0,the greater the ridge characteristics.At the same time,the Bayesian method is used to select the appropriateαto achieve the best regression effect.There are multiple para-meters in the process route,and the multi-task Elastic Net method can handle multiple regression tasks with just one model.

(3) Run Step 2 on each class in the clustering results obtained in Step 1 to generate the corre-sponding prediction model.

(4) For the clustering results in Step 1,the SVM algorithm[4]is used to generate a classification model.

(5) When designing a process route for a new product,the new data are processed according to the method of data preprocessing in Step 1 and the classification model in Step 4 to determine the class to which the data belong.Next,the trained multi-task Elastic Net method is used in this class for the regression analysis.

(6) According to the engineering application scenarios and based on the model’s accuracy,the interval method is used to represent the regression results.The interval estimation method aims to increase the probability of the true value of the sample,which is used to ensure that the regression analysis falls as close as possible within the interval under the premise of a certain interval width.The final result is the interval of each process parameter having a certain degree of confidence.

The design function of the steel plate process route is integrated into the PIDAS big data platform.It is a visual analysis system that provides a means by which the steel plate process route design can be achieved via human-computer interaction (HCI).Data visualization can clearly reflect the characteristics of the data,generate statistical data,and analyze such data[5].The visual interface,as shown in Fig.2,includes three parts:model selection,parameter configuration,and visual analysis results.

Fig.2 Visual configuration interface

3.1 Model selection

The model selection module is used to determine the model for the steel plate process route,including the built-in model of the system and the model trained by the user.The model training must define the start and end time,filter the original data in the database,and then train the model through the built-in method.The model trained by the user can be saved in the model list as a user-defined model.

3.2 Parameter configuration

The parameter configuration interface includes three parts:algorithm parameter,graph type,and style configurations.The algorithm parameter configuration is used to specify the specific parameters required by the process parameter model,the graph type configur-ation is used to specify the type of visual graphics or components,and the style configuration is used to con-figure the specific parameters of each visual graphic in detail (e.g.,data corresponding to the coordinate axis) and graphic elements (e.g.,color and size).Meanwhile,algorithm parameter configuration includes configuring parameters (e.g.,raw material parameters,performance parameters,and chemical components) and the generated algorithm parameter configuration data.

The visualization graphs defined in the graph type configuration include line,column,scatter,pie,and custom graphs.Users can select appropriate visuali-zation graphics according to their needs,and the system supports adjustment to the selected graphics.With abstraction,the common parts in the visuali-zation graph are extracted,including the title,legend,range selection box,and so on.Various default layouts are provided,from which users can choose their preferred multiple visualization graphs.

Style configuration is the process of mapping data values to different visual channels through configur-ation.Different configuration items are supported for various visualization graphics,and the configur-ations can be opened by clicking the graphics edit-ing button in the main display area.Mapping between the data of different dimensions and gra-phic attributes,includingx-axis,y-axis,graphic element color,size,and so on,can be done to generate the graphic descriptions that form the gra-phic parameter configuration data (Fig.3).

Fig.3 Style parameter configuration interface

3.3 Visual analysis results

After the model selection and parameter configuration,the main display area shows the design results of the steel plate process route,as shown in Fig.4,including the numerical results and visual graphics.Vertical stacking is used as the basic layout.Here,users can make decisions with the assistance of recommended process parameters and save the ideal visualization results as PDF files.

Fig.4 Process parameter calculation result interface

3.4 System implementation

This system chooses B/S architecture and is developed in the current mainstream method,which allows front-end and back-end separation.The front-end uses the React framework,the back-end uses the python back-end stack,and the front-end and back-end use the JSON format for transmission by the HTTP protocol.Echarts,a visualization lib-rary,is also used in the system to provide a visual view and an interactive interface[6].

This section uses HSLA steel as an example to prove the effectiveness of the steel plate process route design method.First,data selection and processing were performed.According to the production experience,the HSLA steel data from August 1,2019,to August 1,2021,were extracted from the database,containing 31 variables.These variables consisted of chemical composition,as well as performance,raw material,process,and product parameters.The data also included tensile strength,yield strength,slab thickness,finishing rolling tem-perature,as well asw(C),w(Mn),andw(Si),among others.Data preprocessing was performed according to the basic situation of the data,including outlier processing,missing value filling,or deletion,to name a few.The specific data situations are described in Fig.5.

Fig.5 Basic statistics of each variable

As shown in Fig.5,many chemical element com-ponents exhibited the characteristics of a continuous distribution,while some raw material,product,or process parameters showed the characteristics of categorical variables.The categorical variable data,in this case,were all recessive categorical variables,which were processed by corresponding coding methods.The training and test sets were divided at a ratio of 4∶1.

After modeling the process route prediction using the training set,we normalized the data,added weights to the performance variables,and used a variety of clustering methods to obtain the optimal clustering results.The five clusters obtained by the Kmeans method are shown in Table 1.A classifica-tion model was generated using the support vector regression (SVR) algorithm from the clustering results.

Table 1 Results of cluster analysis

A regression parameter prediction model was established for each cluster obtained in turn.Then,the data were divided into input and target vari-ables.There are 18 input variables and 13 target variables.The multi-task Elastic Net method was used for regression analysis.Through the Bayesian method,the optimal parameters in the multi-task Elastic Net model were obtained,as shown in Table 2.The respective iterative curves of the parameters drawn by the first group of clustering are shown in Figs.6 and 7,and the ordinate is the goodness of fit of the regression.

Table 2 Optimal regression parameters in the first group of clustering

Fig.6 Relationship between the score and the ratio of L1 to L2

Fig.7 Relationship between the score and regularization factor α

Finally,the test set was used to evaluate the multi-task regression model.In turn,the data in the test set were processed according to the same data preprocessing method,and the above classification model was used to determine the group to which it belonged.The trained regression model in this group was called for regression analysis.The regression results were represented by the confidence intervals of the target variable under different confidence levels.A total of 100 samples were randomly selected from the third group of the cluster in the test set for further analysis.Then the average voting method on the selected sample was determined to obtain the confidence interval of the target parameter (Table 3).When the confidence level is above 80%,the precise process route can be achieved.

Table 3 Average recommended process parameters for 100 random samples

Based on the above process,the process parameter interval and the recommended process route were finally generated.The results are displayed on the intelligent big data platform,as shown in Fig.8.

Fig.8 Recommended process route and parameter interval visualization interface

This paper proposes a digitally aided production design method for plate steels,which includes design ideas,implementation cases,and functions on the big data platform.A clear procedure for the production design for plate steels is given,which covers three aspects:data selection and processing,production prediction,and function development on the big data platform.In the implementation case,the method is verified by taking HSLA steel as an example,and the results prove the effectiveness and innovation of the method.Finally,the proposed method is launched on the intelligent big data plat-form PIDAS,which is designed and implemented from four aspects:model selection,parameter confi-guration,visual presentation,and system imple-mentation.

Currently,more production data and wide steel varieties are needed to verify and further optimize the method.This can improve the model’s robust-ness when having excessive amounts of data and real-time update difficulties.

推荐访问:digitalization production Research