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

Abstract: Steel rolling mills have complex processes,specifications,and varieties,along with certain process quality fluctuations and complex production events,making production management decisions difficult.With the development of industrial big data technology,several industrial event solutions based on data have been proposed.These solutions are supported by predictive data and remarkably improve the production level.Taking a heavy plate production line as the research object,through scientific calculations based on historical big data,this paper establishes an optimization logic for plan arrangement,forecasts the quality through the stable relationship between data and quality,intelligently optimizes the subsequent process flow,improves the production line capacity,and reduces the process bottlenecks.

Key words: intelligent decision making; big data; forecast; steel rolling mill

1 Introduction

With the application of big data technology and scientific computing in the manufacturing field,a rising number of successful cases emerged,and the application of ICT technology in various industries also increased.The development of new-generation information and communication technology drives the manufacturing industry to a new stage of trans-formation and upgrading,that is,a new data-driven stage[1-2].

Heavy plate rolling plants have a complex defor-mation process,which involves a complicated system of material,temperature,reduction process,cutting technology,and equipment state.Owing to the characteristics of engineering applications,heavy plate products are characterized by their numerous varieties,small-batch and complex specifications,intricate processes,and large logistics manage-ment[3].The production plan must be closely linked and adjusted with the rolling and shearing processes to maximize the capacity of the production line.In addition,the quality fluctuations of the steel rolling process,such as size defects and plate flatness,will increase the logistics pressure in the finishing pro-cess,improve the crane operations,and induce retention.Thus,the production plan must also be adjusted timely.The application of the big data method in rolling mills is introduced on the basis of two examples of rolling plan optimization,plate flat-ness prediction,and finishing process optimization.This paper aims to establish a prediction model based on historical data and prepare the adjustment and optimization method according to the current production status.

2 Processing logic and valid data status of the production process

2.1 Demand for production process optimization

The demand for production process optimization lies in the diverse product specifications of heavy plates and the different paces of the production pro-cess.Under the complex product specification mode,managers must maximize production efficiency.Plan-ning and process monitoring,as well as capacity,quality,and cost control,need timely data support.The traditional data approach is mainly based on the support of the Enterprise Resource Planning (ERP) system,multilevel report structure,and the moni-toring system of the production process.The adjust-ment is based on empirical judgment.However,the adjustment efficiency is low,and the effect is poor due to the uncertainty of quality fluctuation and the delay in data acquisition.For example,the plate flatness defects might induce bottlenecks in the inspection and cold plate leveler processes.Further-more,obtaining information through general data is difficult due to the hidden flatness defect of many products or the delayed label.Therefore,intelli-gently labeled big data is of substantially high necessity to reflect the details of the product quality.

2.2 Database of production lines

The heavy plate production line of the current study has rich data sources.Its control system is designed as a four-layer data architecture(L1,L2,L3,and L4),which is used for basic automation,process control,production and manufacturing control,and production and marketing manage-ment.Special measurement sensors are also occa-sionally regarded as a separate L0.A rolling mill is a production line with high levels of automation and database.The data center of the production line is established to collect and sort L1 data (including special sensor measurement data),L2 data,and L3 and L4 data systems rapidly and store them on a unified platform.The data storage scale can be more than 100 terabytes.The L1 system data mainly rely on the PLC and DCS automatic data recorded by PDA,which can reflect the process data and equipment status.L2 data mainly record the process control model settings and actual measurements,which can reveal most of the process and resulting information of the product.L3 and L4 systems mainly use ERP and MES system data,res-pectively,thus reflecting material data.The data center has various functions;it logically cleans and labels multilevel data,demonstrates powerful data integration function,associates production details with the overall production situation,and supports the timely decision making of process adjustment and optimization.

3 Application of data-aided intelligent deci-sion making

3.1 Planning optimization based on historical data

In a heavy plate factory,the rolling plan is arranged before the start of all processes.The quality of the plan directly determines productivity.The big data center can realize the digitization of product specification,composition,and process.It can also digitalize the processing time of each process of each steel,including discharging time,roughing mill rolling time,TMCP temperature waiting time,finishing mill rolling time,cooling time,hot leveling time,stacking cooling time,and shearing time.The quality data affected by the processes,including additional gas cutting and cold leveling,can also be obtained.The clustering+con-straint method is used in the current study for the division and prediction of rolling plans[4-5].

The first step aims to summarize the main characteristics of historical products (chemical composition,steel types,specifications,and cooling process) to form data sets,such as FV={F1,F2,F3,…,Fn}.Professional knowledge and experience can be used to delimit the FV of large categories.For instance,the classification example is pipeline category and specification[6].

The second step further summarizes the data,including the detailed characteristics of each process in the production method,according to the pro-cessing time,and performs cluster analysis based on the production process time.The classification is mainly obtained by weighting factors,such as process time and characteristics.Fig.1(a) shows that the total distance from cluster points to center points has reached the same order of magnitude when it is divided into five or more categories[7].Combined with the rationality of clustering,thet-distributed stochastic neighbor embedding(TSNE) algorithm is adopted for judgment as shown in Figs.1(b)-(d),which are the projection results when the algorithm randomly chooses 5,6 and 7 points as the initial clustering center,respectively.Herein,seven cate-gories are reasonable than other categories.Next,con-sidering FV classification,a few products with simila-rities in FV are divided into different categories,which are classified into new categories according to the majority principle.This classification mainly aims to eliminate the disturbance caused by nonproduction factors,and the optimized results are shown in Table 1.

Fig.1 Results of data clustering

By comparing the roll matching cycle in the historical data,the third step is to find the result with the highest output and the average unit weight close to the average value of the production line as the optimal plan.Then the ratio of cluster numbers will be delimited in roll matching.

The fourth step establishes the constraint con-ditions through the data process law of the finishing process,mainly defining the offline stacking cooling ratio,gas cutting ratio,and the ratio of the shear line pace lower than the average value.

The fifth step aims to optimize the plan arran-gement when the contract specifications are known and plan the appropriate quantity and proportion in the roll matching in combination with the above model.The batch quantity of similar specifications is expanded under limited conditions,the products with small differences in production time are brought close to those with different specifications,and the products with extreme specifications are enabled to realize a reasonable transition.Fig.2 shows the actual plan arrangement after a calcul-ation.

Fig.2 Actual plan arrangement after optimization

3.2 Quality prediction analysis based on his-torical data

The fluctuation of quality will influence the subsequent process.The unknown quality fluc-tuation of the rolling process will reduce the logi-stics throughput of the finishing process.Through the high-quality prediction of abnormity,early adjust-ment and optimization of operating procedures will have beneficial effects on the production line.For example,a certain width allowance is required in the process of rolling plate shearing.If the rolling width of a plate is smaller than the target or camber occurs during rolling,then the shearing machine operator must adjust the shearing parameters;how-ever,such an adjustment only wastes working time.The situation could be changed by using the rolling line measuring system,which can be utilized to judge whether cutting can be conducted smoothly in advance[4].When shearing cannot be performed,the plate can be shifted to the gas cutting process in advance (the gas cutting allows minimal trimming) to reduce the invalid occupation of the shearing process.

The paper will introduce the process optimization of flatness quality prediction.The flatness defect in the rolling mill severely impacts the process.For example,the plates of a product grade must be stacking cooled for 24 h after rolling.Poor flatness might appear and affect the efficiency of the shear-ing process when plates are lifted to the finishing line.Adjusting the logistics process is difficult when managers find the aforementioned problem.If predic-ting plate flatness after stacking cooling is possible,then serious bottlenecks in the finishing process can be prevented by controlling the number of poor flatness plates lifted to the finishing line per unit time.

Rolling theory and empirical knowledge are used in the data center to summarize the data that may affect the flatness quality to realize the function of flatness prediction,and a certain eigenvalue is established.In the historical production data,the relationship between Frel and Flat is determined through scientific calculation considering the summary data set Frel={X1,X2,X3,…,Xn},the historical flatness measurement,and the additional flatness result data set Flat={Y1,Y2,Y3,…,Yn},and a prediction model is established.

First,the preparation of data sets should be combined with industry knowledge and production status.For example,the production data set is the data of rolling plates,the temperature of each field is the temperature or temperature distribution of rolling plates,and the flatness result data are the results of sheared plates (the sheared plate is only part of the rolling plate).Herein,a concept of plate combination is provided.Two solutions are then presented:divide the data of the rolling plate by the results of the sheared plate and combine the sheared plate data to form the rolling plate data.Not all sheared plates have plate flatness result data,and blind data areas will be observed in combination;therefore,the rolling plate is divided into regions to reflect the data.The plate flatness fluctuation area is marked by regional memory results,which are divided into head and tail waves and left and right side waves to reduce mutual interference.Fig.3 is an example of the temperature distribution splitting of the rolled plate.The temperature distribution of the rolling plate after cooling is shown in Fig.3(a),and the splitting results are shown in Figs.3(b) and (c).Figs.3(d) and (e) show the measured results of flatness of plate after shearing.

Fig.3 Resolution process of rolling plate temperature distribution

First,classification and prediction are performed in accordance with similar varieties of specifications and processes to predict accuracy.For the key factors affecting plate flatness,the cooling mode can be divided into the following three categories:non-cooling,direct cooling,and slow cooling types.

In establishing the model,the weight of the data field should be initially optimized to prevent the low accuracy of the initial prediction results (caused by excessively scattered influencing factors).Freln={x1,x2,x3,…,xn} is the data set andλ=(λ1,λ2,λ3,…,λn) is the corresponding correlation coefficient of the set.When the relationship coefficient cannot be determined,the data shall be standardized first,and the main cause analysis shall then be performed to determine the correlation coefficientλ1,λ2,λ3,…,λn.Some irrelevant factors can be eliminated,and the model can finally be established[7].Through the comparison of various prediction methods,the prediction results of the random forest model are superior.Herein,the random forest method is used for demonstration.Fig.4 shows the correlation of factors in the model.The prediction results are generated within 1 h after completion of the steel rolling process.Table 2 shows the prediction accuracy of flatness at different positions of high-strength pipeline products.

Fig.4 The correlation of factors in the prediction model

A constraint model is established on the basis of the prediction results:

(1)

where,τ(rf,ep) andη(rf,elp) are the influence functions of shape defect rates of the current and the last processes on process efficiency,respectively;rfis the shape defect rate;epis the process efficiency;elpis the efficiency of the previous process;andRuis the confirmation condition of process efficiency bottleneck,and rules are established on the basis of experience.For example,the decision logic is as follows:if the defective rate of the predicted flatness result is substantially higher thanRh(the maximum defective rate that the finishing process can withstand),then the rolling plan must be adjusted to reduce the planned quantity of this specification in the next matching roll to correct the production procedure for the process and equipment personnel.The rolled steel plates with expected poor flatness are passed in batches according to the logistics of finishing,and the defective rate of plate flatness of each batch isRl(under this defective rate,the product flatness defect has minimal impact on the finishing process).

4 Conclusions

From the perspective of the application of process industry production lines,this paper introduces an attempt to use big data technology to assist decision making in steel rolling mills.Based on historical production data,by combining scientific methods with constraints of professional experience,the arrangement of rolling plan in a heavy plate mill can be optimized.Through the correlation rules of quality factors,scientific prediction can be carried out by using big data,and the prediction results can be applied to production decision-making.