international journal of mathematical …finally, the steel is run through a molten-zinc coating...
TRANSCRIPT
1
Improving scheduling methodologies in a Hot-DipGalvanizing Line combining non-linear projectors
and clusteringAndrés Sanz García, Francisco J. Martínez de Pisón Ascacibar, Rubén Lostado Lorza, Roberto Fernández
Martínez and Julio Fernández Ceniceros
Abstract—An improving methodology for the scheduling coils
of a Hot Dip Galvanizing Line (HDGL) is presented. This
method uses a non-linear projector which has been selected
from various techniques to generate a coil map from the most
significant parameters of the coils database: process variables,
chemical composition of steel, measurements, etc. The created
bidimensional map helps experts to decide which are the more
fitting groups showing the distances between all coils. After that,
the expert can select with an end-user application one group and
identify other coils that can complicate the scheduling purposes.
Finally, the methodology uses hierarchical clustering to obtain a
list of effective sequences of coils. A decrease of the number of
shutdowns and irregular heat treatments failures can be obtained
by using this scheduling method.
Index Terms—Scheduling Methodology, Hot Dip Galvanizing
Line, Hierarchical Clustering, Sammon Mapping, Kruskal’s non-
metric projector
I. INTRODUCTION
THERE The constant effort to increase product quality[27] and reduce the expenses caused by failures in the
manufacturing process is ongoing with Hot-Dip GalvanizingLines, which are always trying to optimize their operativecosts.
A. Hot Dip Galvanizing LineThis industrial process is composed by several stages. The
initial material is the steel coils from the cold rolling with therequired thickness.
Broadly speaking, the process within a Hot-Dip GalvanizingLines (HDGL) (Figure 1) can be described in the following
Manuscript received January 31, 2011: Revised version received March8, 2011. This work was supported in part by La Rioja University throughFPI fellowships and the Autonomous Government of La Rioja under GrantFOMENTA 2010/13.
A. Sanz-Garcia is with EDMANS Group, Mechanical Engineering Depart-ment at La Rioja University, Logroño, La Rioja, 26004, Spain. (phone: +34-941-299-273; fax: +34-941-299-794; [email protected]).
F. J. Martinez-De-Pison-Ascacibar is with EDMANS Group, MechanicalEngineering Department at La Rioja University, Logroño, La Rioja, 26004,Spain. ([email protected]).
R. Lostado-Lorza is with EDMANS Group, Mechanical EngineeringDepartment at La Rioja University, Logroño, La Rioja, 26004, Spain.([email protected]).
R. Fernandez-Martinez is with EDMANS Group, Mechanical Engineer-ing Department at La Rioja University, Logroño, La Rioja, 26004, Spain.([email protected]).
J. Fernandez-Ceniceros is with EDMANS Group, Mechanical Engineer-ing Department at La Rioja University, Logroño, La Rioja, 26004, Spain.([email protected]).
Figure 1. Basic scheme of a HDGL
manner: The first stage in the line consists of the formationof a continuous strip using steel coils that come from a millprocess. Next, the strip passes through a preliminary cleaningsection at the entrance of the annealing furnace, where itreceives a heat treatment; the stage prior to its immersion in theliquid zinc bath. This treatment is essential in order to improvethe properties of the strip and its coating. After the bath, thestrip exits vertically, passing through blade-like currents of airwhich regulate the thickness of the cover. The temperature andcooling rates are controlled to obtain the desired mechanicalproperties for each steel type.
Finally, the steel is run through a molten-zinc coating bath,followed by an air stream wipe that controls the thickness ofthe zinc finish (see Fig. 1). This description, with minor mod-ifications, is valid and applicable to all Hot Dip GalvanizingLines in operation throughout the world.
B. The Quality of the Galvanized Product
The quality of the galvanized product, according to severalauthors [2], [19], can be divided into two fundamental aspects:the anti-corrosive characteristics and the properties of the steel.
First, the anti-corrosive characteristics: steel comes markedby the thickness and uniformity from the zinc coating andbasically depends on the superficial preparation of the metalbases, the control of the heat treatment, the bath compositionand temperature, the air-blade control, the speed of the stripand the amount of time available for alloying at the specifiedtemperature after the air-blade treatment.
Secondly, the properties of the steel which fundamentallydepend on the composition of the steel, the surface roughness,the process of smelting, the processes of rolling and finally,
INTERNATIONAL JOURNAL OF MATHEMATICAL MODELS AND METHODS IN APPLIED SCIENCES
Issue 5, Volume 5, 2011 890
2
the heat treatment that is applied to the strip before its passageby immersion in the liquid-zinc bath.
By examining previously exposed coils, it can be deducedthat the control of the heat treatment in a HDGL is funda-mental for the process of coating and for the improvement ofthe properties of the steel. One of the key aspects concerningtemperature control consists in assuring that the suitablethermal cycle is reached for each coil while considering thelimitations of the HDGL furnace inertia. This is fundamentalfor the process of coating and for the improvement of theproperties of the steel and can only be accomplished whenthe flow of the steel mass for unit of time is constant andthe changes therein are carried out gradually. A fundamentalcoil sequence constraint is the need to limit changes in themass flow rate from one coil to the next. But we must alsoexamine other inconveniences. For example, coils within eachgroup have to be sequenced based largely on decreasing sortby width or coils that have to be selected with similar steelproperties and thermal requirements.
Also, it must be borne in mind that defective coils canappear, or coils with characteristics that differ greatly fromthe others, and therefore, must be separated from the rest.For these reasons, it is crucial to order the coils that formthe strip in such a way that the varying thickness, widths andtypes of steel do not differ too greatly [23], [31]. When thisis accomplished, the temperatures generated by the furnaceprogressively fit each individual coil.
We see, therefore, that one of the most important tasks inthe process in an HDGL is centered on obtaining sequencesof coils that do not contain abrupt changes in the dimensionsand types of steel in the consecutively treated coils. Logically,depending on the type of process, the demands of planningwill be different.
As has been mentioned above, one of the primary aims ofthe HDGL is to determine the best sequence of coils, afterwhich it conforms to the changes in the new dimensions andtypes of steel. This is essential in assuring that the quality ofthe coating of each coil and the production are optimal.
The present systems of planning in the HDGL divide theprocess into two general phases:◦ The coils are grouped according to the type of steel (de-
pending on its chemical composition) and its specifications(specifications of production, date of delivery of the product,specifications of quality, etc.).
◦ Each group of coils is ordered such that the changes in widthand thickness are minimalized. Normally, the process of coilscheduling begins by selecting those coils of greater width,and gradually moving towards those that are more narrow.Also, the HDGL looks for variances which are minimal, soas to create a smooth transition.
II. PREVIOUS WORK
We review some of the present scheduling [4], [25] tech-niques in steel industry. There are several publications that dealwith the problem of scheduling in an HDGL. The majorityof the publications that address applications and techniquesfor scheduling in the steel industry are based on continuous
casting machines. The systems presently employed in theseprocesses are generally based on either heuristic schedulingstrategies or are implemented using knowledge software tech-nology [23], [26].
Tang et al. [31] offer a review of the more commonlyused methods and scheduling systems in the steel industry.In this work, different methods of optimization are describedaccording to many planning systems (actually integrated inseveral industrial plants in the steel industry).
In this article, the optimization techniques are groupedaccording to the following classification:◦ Mathematical or heuristic models extracted directly from the
analysis of the process.◦ Methods [11] of artificial intelligence (AI) divided as well
into: a) Expert systems: based on the experience of theexperts and the observed process restrictions. b) Randommethods in search intelligence: genetic algorithms (GA),simulated annealing (SA) and tabu-search algorithms (TS).c) Constraints satisfaction methods (CSP) [32].
◦ Human-machine coordination methods: consisting of thecreation of scheduling by means of the iterative dialoguebetween human and machine.
◦ Multi-agents methods (A-teams methods): which rely on thecooperation of multiple models with different algorithms(each one of them with different constraints), so that thebest solutions are found [10], [15].For each kind of problem, multiple mathematical ap-
proaches and heuristic algorithms of optimization are selected[18], [29]. Other authors [9] focus the resolution towardsproblems of the type Traveling Salesman Problem (TSP)using the dimensions of each coil and other parameters. Thismethodology is based on the calculation of a triangular matrixof distances between all the coils to determine the best possiblesequence. The most critical point of this consists in finding atrustworthy formulation that indicates the degree of existingsimilarity between two coils according to their physical andchemical properties, as well as restrictions in the process.
Other works, however, are based on the use of graphicalmethods for scheduling. For example, Tamura et al. [28] looksfor optimal ways in distribution maps where each coil isrepresented according to two coordinates: thickness and width.
At the moment, these techniques are complemented withnew techniques [6][17], [20]: Prize Colleting VRP models,IA, multi-agents, methods of man-machine iteration, or com-binations of all of the above.
Finally, other works study when scheduling problems aresubject to unexpected events. In these cases, the steel plantneeds a scheduling decisión that must be made in real timeabout the possible rescheduling or reordering of tasks. Re-cently years, new scheduling techniques, as known as hybriddynamic scheduling systems, have been developed por solvingthis hard problem. Aytug et al. [1] used an evolutionaryalgorithm to modify an initial knowledge base, using resultstaken from the simulation of the manufacturing line. In thatcontribution they achieved that the system could react tounexpected events. Ikkai et al. [13] proposed a status selectionmethod based on evolutionary algorithms to induce scheduling
INTERNATIONAL JOURNAL OF MATHEMATICAL MODELS AND METHODS IN APPLIED SCIENCES
Issue 5, Volume 5, 2011 891
3
knowledge from manufacturing lines.
!
1
Data Base (Coils):
1. Dimensional Parameters. 2. Chemical Composition. 3. Process Parameters. 4. ...
3
2
Non-Linear Projection
Visual Selection Groups
Modify Zone B
4 Automatic
Scheduling Group B
5
6 Reselect Groups
and Optimise Zone B.
Recalculate Scheduling.
Optimal Scheduling
Zone B?
NOT
YES
7
Saving Scheduling Zone B.
Bidimensional Mapping Coils
Figure 2. The diagram of the complete methodology
In this work consists of a several-step approach to the prob-lem using multi-dimensional scaling and hierarchical cluster-ing. In the following sections, another experience is described,results obtained are discussed and final conclusions are drawn.
III. METHODOLOGY
The methodology that considers in this work, consist ofan approach to the problem in seven steps. Figure 2 presentsthe complete and comprehensive flow chart explaining theproposed methodology to select the optimal coils schedulingsequence.
The first step consists in developing a data base with then weighted parameters of each coil (considered importantwithin the classification process). These parameters are thefollowings:
• Width and thickness of the coil.• Objective temperature at the process.• Chemical composition of the steel (C, Mn, Si, S, P, Al,
Cu, Ni, Cr, Nb, V, Ti, B, N, Ceq).
Second, the present work makes use of multidimensionalscaling (MDS) for exploring similarities or dissimilarities inorder to generate a bidimensional map of the coils and then,a clustering technique is utilized to find the local clusters.The clustering technique is able to reduce the amount of dataitems by grouping them, but projection methods are necessaryto reduce the dimensionality of the data in the first place.In addition, the projection can be used to show groups if aenough small output dimensionality (two or three dimensions)is chosen.
A. Principal component analysisThis mathematical procedure is used to show a set of data as
a linear projection [21] on such a subspace of the initial dataspace. This conversion is defined in such a way that the firstprincipal component accounts for as much of the variabilityin the data as possible. This transformation is sensitive to thescaling of the initial data values and to outliers that producelarge errors.
Table IIMPORTANCE OF COMPONENTS FOR PCA
PC1 PC2 PC3Standard deviation 2.1394 1.6382 1.3419
Proportion of Variance 0.2692 0.1579 0.1059Cumulative Proportion 0.2692 0.4271 0.5330
Thanks to the PCA and the knowledge based on expertise,it can be defined a set of uncorrelated and relevant variables(some chemical components of the steel). These are shown inFigure 3.
-0.10 -0.05 0.00
-0.10
-0.05
0.00
PC1
PC2
123
4567891011121314151617 18192021
2223 24252627
28
2930313233343536373839404142 434445
46474849505152
53
545556575859
6061
62
63646566
6768697071
72
7374
757677
78798081828384 8586
87
88
899091
92
93949596
979899100
101
102103 104105106107108
109
110111112113114115116117118119120121122123124125
126
127128
129130131
132133
134
135136
137138139140
141142143144145
146147148149
150151
152153
154155156157
158159
160161162163
164165166167
168
169
170
171
172173174175
176177178179
180
181182183184185186
187188189190191192193194195196197198
199200
201202
203204205206207208
209
210 211212213214215216217
218
219220
221222223224
225 226
227228229230
231232233
234235236
237
238
239
240
241242
243244245
246247 248249250251252253254255256257258259 260261262263
264265
266267268269270
271272273274275276
277278279280
281282
283
284285286
287288
289290
291292
293
294295
296
297
298
299
300301302
303
304305306307308309310311
312
313314 315316317318 319320321
322
323324325326 327328
329
330
331
332333
334335336337338
339340 341342343344345346
347348349
350351352353354
355356357358
359360361
362363364365366367
368369
370
371372373374
375376377378379
380
381382383384385386387388389390391
392393394395396397398399
400401402403404405406
407408409410411412413
414
415416417
418419420421
422423424425
426427428429430431432433434435436
437438
439
440441442
443444445446447448 449450451
452453 454
455
456
457
458459460461462463
464
465
466467
468
469470471472473
474
475476477478479480481
482
483484485486487
488489490491492493494495496
497498499500
501502
503
504
505
506
507508509510511512
513
514515516
517518519520521522523
524525526527528 529
530531532
533534
535 536537 538539540541542
543544545546
547548
549550551552553554555
556557558559560
561562563564565566567568 569570571
572573574575576
577578579
580
581
582583584
585586
587
588589
590
591 592593
594
595596597
598599
600
601
602
603604605
606607608609610611612613614615616
617
618619620621622623
624625
626
627
628629
630631632633634
635
636637638
639
640641
642643644645
646647648649650651652
653654655656657658659660
661662663664665666667
668669670
671672673674675676677
678679 680681682683684
685
686687688689 690691692
693694695
696697
698
699700701702703
704705
706707
708709710711
712713714715716717718719720721722723724725726
727
728
729730731
732733
734735736737738
739740741
742743744745746747748749750 751752753754
755756
757
758759
760
761
762763
764
765766
767768769
770771772
773774
775776777778779780781
782783
784785
786
787
788789790
791792
793794795796797
798
799800
801 802
803804805806
807
808809810811812813814
815
816
817
818819820
821
822823824825826
827828829
830
831
832
833834
835
836
837
838
839840841842843844845
846847
848
849850851852
853854855856857858
859
860861862863864865866867
868869870871
872873874875876
877878879880
881882
883884885
886887
888889890
891892
893894
895896897898899900 901
902
903904
905
906907908909910
911
912913914915916 917918919920921922923924925926
927928929930931932933934935936937938939940941942943944945946947948949950
951
952
953954955956
957958959
960
961
962963
964965966967968
969970
971
972973974975976977
978979980981
982983
984
985986987988989990
99199299399499599699799899910001001
1002100310041005
10061007
10081009101010111012
101310141015
1016
101710181019
1020
1021
10221023102410251026
1027 1028
1029103010311032
1033
10341035
103610371038103910401041
104210431044
10451046
1047
1048
10491050
10511052105310541055
1056
10571058
105910601061
10621063
1064
1065
106610671068
106910701071
1072107310741075
107610771078
10791080108110821083
10841085 1086
1087
10881089
109010911092
109310941095
1096
10971098
1099
1100
1101
11021103110411051106
11071108
1109
111011111112
11131114
1115
1116
11171118111911201121
1122
11231124
1125
1126
11271128112911301131
11321133
11341135
1136
1137
1138113911401141
1142114311441145
1146114711481149
1150115111521153115411551156
115711581159
11601161
1162116311641165
1166116711681169
11701171
117211731174117511761177
1178
11791180
11811182
1183
1184118511861187
1188
118911901191119211931194
119511961197
119811991200
120112021203
1204
1205120612071208
120912101211121212131214121512161217
1218121912201221
1222122312241225122612271228
12291230123112321233
1234
12351236
12371238
1239124012411242124312441245
1246
1247
12481249
1250
125112521253
125412551256
-80 -60 -40 -20 0 20
-80
-60
-40
-20
020
ESPENTANCHO
MATTEMPMED2C
Mn
Si
S
P
AlTot
CuNi
Cr
Nb VTi
B
N
Figure 3. Coils data set with 17 main components
As shown Figure 4, the first three principal components(PCA1, PCA2, PCA3) doesn´t explain most of the coils groups
INTERNATIONAL JOURNAL OF MATHEMATICAL MODELS AND METHODS IN APPLIED SCIENCES
Issue 5, Volume 5, 2011 892
4
Figure 4. Tridimensional plot of the steel coils
in this work. Table I demonstrates that three principal axes ofthe PCA cover less than 55% of the cumulative proportion ofvariance, and therefore can not be used to project every of thecoils.
It can be seen that the linear projection doesn´t workbecause , so we propose two alternative that does.
B. Kruskal’s non-metric projectorThis is another form of non-metric MDS method [5]. First, a
n-dimensional configuration to minimize the stress is chosen.Then, an iterative algorithm is used, which will converge mostof times in few iterations (about 10), but due the complexityof the algorithm this procedure is slow for large datasets.
-0.5 0.0 0.5 1.0 1.5
-1.0
-0.5
0.0
0.5
1.0
1.5
MATISO$points[,1]
MATISO$points[,2]
Figure 5. Bidimensional plot of Kruskal’s Non-metric MDS
Finally, this MDS uses the power for Minkowski distancein the configuration space, so results can vary considerablyfrom one computer to another. Basically, the rank correla-tion between calculated dissimilarities and plotted distancesis maximized, allowing tied distances to not have identicalplotted distances, only sequential ranks. In Figure 5, thebidimensional map generated represents the distances betweenall the coils to organize.
C. Sammon non-linear projector
The Sammon non-linear projector [24] is suited for use inexploratory data analysis. In the present work, it is used tovisualize the equivalent Euclidean distances [3] of the coils tobe treated within a �2 space.
This algorithm employs non-linear transformations in orderto map the original space, from the significant n weightedparameters so that they can influence the classification process,onto a low-dimensional visual space (�2 or �3), attemptingto preserve the Euclidean distances between coordinates ofpatterns (logically, other linear or non-linear multi-dimensionalscaling methods (MDS) can be used) [8].
Let us denote the distance between two objects i and j in thespace �n by d̂ij , and the distance between their projections bydij ; Sammon mapping intends to minimize the error functionknown as Sammon´s stress:
Ei =1
�i<j d̂ij
�
i<j
�d̂ij − dij
�2
d̂ij(1)
The next step consists in reducing the original data basewith the n parameters of each coil to the most important pparameters. These chosen parameters are the following:
• Thickness of the coil.• Objective temperature at the process.
-0.5 0.0 0.5 1.0
-0.5
0.0
0.5
1.0
1.5
MATSAM$points[,1]
MATSAM$points[,2]
Figure 6. First bidimensional plot of Sammon’s mapping
INTERNATIONAL JOURNAL OF MATHEMATICAL MODELS AND METHODS IN APPLIED SCIENCES
Issue 5, Volume 5, 2011 893
5
• Chemical composition of the steel (C, Mn, Si, S, P, V,Ti, B, N).
This new data set has better projection because there is onlythe most relevant variables for the first two axes of principalcomponent analysis performed previously. In Figure 7, thefinal bidimensional plot of Sammon’s mapping with the newdata is represented.
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
MATNORMPROY[,1]
MATNORMPROY[,2]
Figure 7. Second bidimensional plot of Sammon’s mapping
D. Grouping of dataOnce defined, the coils like points within a space �n are
projected by means of Sammon projector in a bidimensionalspace. This bidimensional generated map represents the equiv-alent Euclidean distances between all the coils to organize.
Figure 8. Defined groups visually using a Sammon´s map
In this way, the degree of similarity between the coils andthe obtaining of the standardized distance of its projectionscan be determined visually.
This map of points allows one to visualize in an objectivemanner the equivalent distances between all the coils, and toclassify them according to all the considered parameters prop-erly weighted. Logically, the coils with very close parameterswill appear tightly-packed, whereas those coils whose valuesvary greatly amongst themselves will appear more loosely-grouped.
Once all the coils in a two-dimensional graph have beenprojected, the integrated software system allows for iterationwith a human expert for the selection of main groups of coilswith similar characteristics (see Fig. 8).
From the map of coils generated with software package,the new groups of coils with similar characteristics are definedvisually with a contour of lines that can be adapted by the user.Those different coils that do not correspond to any definedgroup are detected (Fig. 9).
Figure 9. Defined groups visually using a Sammon´s map
E. Cluster analisys of dataWhen the main groups of coils are all selected, the coils
of each group are locally grouped according to hierarchicalclustering (see Fig. 10).
This is one of the large number of methods that havebeen proposed for cluster analysis. Presently, cluster analy-ses appearing in the literatures are mainly divided in fourgroups: hierarchical methods, optimal partitioning, distributionmixtures and non-parametric estimation of local densities.A strictly hierarchical method depends on dissimilarity orsimilarity measures, providing a complete scheme of division.This algorithm [14], [7] uses a set of dissimilarities for then objects being clustered. Initially, each object is assigned toits own cluster and then the algorithm proceeds iteratively, ateach stage joining the two most similar clusters, continuinguntil there is just one single cluster.
Figure 10 shows a classification tree or dendrogram whereeach node beneath the tree represents a coil from group B
INTERNATIONAL JOURNAL OF MATHEMATICAL MODELS AND METHODS IN APPLIED SCIENCES
Issue 5, Volume 5, 2011 894
6
Figure 10. Dendrogram of Group B formed from hierarchical clustering withsix clusters
and each straight line represents the grouping of two clusterspositioned by a Euclidean distance equal to the “height” of thatstraight line. The main advantage in using this type of graphis that the search for subgroups can be automated such thateach individual coil does not exceed a determined distance.Partitioning into a given number of set of coils is found bymerely cutting the dendogram at an suitable height. Also,abrupt jumps between clusters can be predicted.
Figure 11. Different subclusters into group B
In Figure 11 are shown 6 subgroups whose Euclideandistance between coils does not exceed 0.15 and are distributedwithin a bidimensional space. The final scheduling is madeautomatically for each one of the defined subclusters accordingto one or several of the parameters employed: width, thickness,
Figure 12. Local scheduling method applied starting of the greatest densityin zone B
chemical components of the steel, target temperature, etc.Once the local scheduling is made, it is necessary to verify
the number of times that it surpasses a defined distance(number of gaps), as well as the maximum distances obtainedbetween ordered coils. If the number of gaps is big or thedistances between consecutive coils great, the human expertcan take different courses of action until the results of eachlocal scheduling are optimal: i.e. modify the contours thatdefine the groups, divide or fuse groups, choose anotherdistance of clustering, etc.; or by means of techniques widelyknown: mathematical models, TSP, heuristic techniques, func-tions relating several variables, etc.
One of the techniques that yields the greatest results consistsin beginning the classification by starting at the point wherethe greatest density of coils exists (top of the mountain ofdensities) and continuing downward with those points thatare less dense (see Figure 12). The process is analogous todescending from a mountain, starting at the peak and theending at the foot.
Finally, once all of the groups are ordered, those coils thatcould not be classified in any group can be eliminated, orderedseparately, or included in some of the already existing groups.
IV. RESULTS
A. Analysis of a previous historical data using the proposedmethodology
The first interesting results arose when analyzing the pro-jection of the 14 chemical components of the coils (C, Mn, Si,S, P, Al, Cu, Ni, Cr, Nb, V, Ti, B, N) for each one of the coilsfrom previous scheduling and upon comparing them with theresults obtained after the galvanizing process.
The use of the created map allowed, with a single look andan objective form, the groups of existing coils to be determinedas well as to detect those coils whose chemical compositionswere completely different. Figure 13 presents the non-linear
INTERNATIONAL JOURNAL OF MATHEMATICAL MODELS AND METHODS IN APPLIED SCIENCES
Issue 5, Volume 5, 2011 895
7
Figure 13. Sammon coils projection according to chemical composition
Sammon projection according to chemical composition andoutliers coils that clearly appeared projected separately fromthe others. In addition, by extending the right part of themap, two main groups (cluster A and B) and three secondaryones (clusters C, D and E) are defined (see Fig. 14). Thismap can easily be used to classify new coils in one groupor another. The method we present is expected to become auseful tool in HDGL applications as Figure 14 shows with 6groups generated by clustering the coils.
B. Decreasing of unexpected situations and process interrupts
One of the most important problematic events, that worriesto the engineers, happens when the industrial process ofunexpected way is interrupted such as: breakage of the band,desaleaning, failures in the system, etc.
However, the main advantage when using the proposedmethodology is observed when we compare the distancesbetween consecutive coils within the scheduling, since it canbe used to help in the prediction of potential problems in theprocess.
One of these possible causes that can be attributed toproblems in an HDGL is produced when one of the weldedcoils that forms a strip consists of a steel with mechanicalcharacteristics different from the others. Thanks to the useof the map, the expert can easily detect the presence of twoconsecutive coils with different steel compositions that aregoing to be processed one after another one. Also, the mapcan serve as a useful tool to visualize in each moment the typeof steel in the coil that is being processed.
The system can also be used to detect when a cross-sectiondiffers substantially from the coils that precede and follow it,or has a defective weld, etc.
Another important problem is the breakage of the band. Thisevent is one of the most dangerous contingencies by the higheconomic losses that produces. These fundamentally must to
Figure 14. Groups of coils according to the chemical composition of thesteel
the lost time in removing the defective band and feeding theline with a new band, as well as, by the lost material and thedeterioration that takes place in the machinery. The proposedmethodology can help to avoid this event by detecting all stripswith mechanical characteristics very different from the others.
C. Two cases of study with problematic events with chemicalcomposition
In Figure 15, are shown two cases of potential, problematicevents is shown a case of potential, problematic events due tothe inclusion of a coil with a steel different from the otherswithin scheduling. Clearly, one can appreciate the utility inmeasuring and charting the chemical composition of the steelcoils which are involved in the industrial process.
D. Results of the ExperienceAs has been mentioned, this methodology takes into account
not only the chemical parameters of the steel coils, butalso the dimensional parameters (thickness and width), targettemperature, and other parameters of the process.
In the tests, the following parameters of each coil wereincluded: width and thickness, chemical composition of thesteel and desired final temperature; all of them properlyweighted.
Introducing the database composed of 2, 436 different coils,a nonlinear Sammon projection was made by obtaining a mapsimilar to the one of Figure 9.
The handling of the map of projected coils was very user-friendly due to the following reasons:◦ The selection of the different groups from coils was ex-
tremely simple.◦ It allowed the expert to detect those coils with parameters
which differed greatly from the others.◦ It helped to classify new coils easily within the process.
INTERNATIONAL JOURNAL OF MATHEMATICAL MODELS AND METHODS IN APPLIED SCIENCES
Issue 5, Volume 5, 2011 896
8
Once the groups of coils were selected, the schedulingfor each one was generated semi-automatically according tohierarchical clustering. The results were such that each groupwas easily ordered according to the desired requirements.The number of gaps between coils was considerably reduced,as well as maximum distances between them. Even in thepresence of a great number of gaps or distance between coils,this methodology proved highly effective in generating smallergroups to accommodate the gaps, and in adapting to thecontours of greater distances.
Figure 15. A dangerous hypothetical event due to the presence of a coilfrom another steel family in scheduling
V. CONCLUSIONS
This article presents a successful experience in the use ofan MDS to generate bidimensional maps from the multitudeof parameters that factor into the process of the creation ofscheduling within an HDGL.
The use of these bidimensional maps and the dendrogramsallows one to graphically group the coils with the least distancebetween one another, thus ensuring that their schedulingpresents no significant problems.
An iterative process consisting of the following steps wasproposed: projecting the coils according to the parametersselected in a bidimensional map, defining the groups of coilswith the closest distances between them, selecting subclustersusing hierarchical clustering and separately ordering themaccording to algorithms of already known scheduling.
In this way, lists of safer coil combinations are obtained,abrupt jumps between consecutive coils are reduced andpotential problems spotted before occurring.
REFERENCES
[1] H. Aytug, G.H. Koehler, J.L. Snowdon, “Genetic learning of dynamicscheduling within a simulation enviroment,”Comput. Oper. Res. 21(8),1994, pp. 909–925.
[2] V. L. Baum, J.M Stoll, “LTV stell hot dip galvanizing line upgrade forexposed automotive products,”Iron and Steel Engineer 71, 1994, pp. 32–37.
[3] S.-H. Cha, “Comprehensive Survey on Distance/Similarity Measuresbetween Probability Density Functions.” International Journal Of Math-ematical Models And Methods In Applied Sciences, Issue 4, vol. 1, 2007,pp. 300–307.
[4] J. I.-Z. Chen, C.-C. Yu, “Estimate to the Trajectory of ManeuveringTargets by Combining Sensor Scheduling with Energy Efficient inWSNs.” International Journal Of Mathematical Models And MethodsIn Applied Sciences, Issue 2, vol. 3, 2009, pp. 77–85.
[5] T. Condamines, “The noisy multidimensional scaling problem: an opti-mization approach.” In CSCC 2000 Multiconference Proc., ISBN: 960-8052-19-X, 2000, pp. 2411–2416.
[6] P. Cowling, M. Johansson, “Optimization in Steel Hot Rolling. Opti-mization in Industry.” European Journal of Operational Research, 1391,2002, pp. 230–244.
[7] R. D’andrade, “U-Statistic Hierarchical Clustering.” Psychometrika 4,1978, pp. 58–67.
[8] M. Daszykowski, B. Walczak, D.L. Massart, “Projection meth-ods in chemistry.” Chemometrics and intelligent laboratory sys-tems 1345, 2003, pp. 1–16.
[9] J. Dorn, M. Girsch, G. Skele, W. Slany, “Comparison of iterativeimprovement techniques for schedule optimization.” European Journalof Operational Research 25, 1996, pp. 349–361.
[10] D. Frey, J. Nimis, H. Wörn, P. Lockemann, “Benchmarking and robustmulti-agent-based production planning and control.” Engineering Appli-cations of Artificial Intelligence 16, 2003, pp. 307–320
[11] B. Grabot, “Artificial intelligence and soft computing for planning andscheduling: how to efficiently solve more realistic problems.” Engineer-ing Applications of Artificial Intelligence 14, 2001, pp. 265–267.
[12] I. Harjunkoski, I.E. Grossmann, “A Decomposition Approach for theScheduling of a Steel Plant Production”, Computer & Chemical Engi-neering 25, 2001, pp. 647–1660.
[13] Y. Ikkai, M. Inoue, T. Ohkawa, N. Komoda, “A learning method ofscheduling knowledge bu genetic algorithms”, In: IEEE Symp. EmergingTechnol. Factory Automation, 1995, pp. 641–648.
[14] S. Johnson, “Hierarchical clustering schemes.” Psychometrika 2, 1967,pp. 241–254
[15] A. Karageorgos, N. Mehandjiev, G. Weichhart, A. Hämmerle, “Agent-based optimisation of logistics and production planning.” EngineeringApplications of Artificial Intelligence 16, 2003, pp. 335–348
[16] B. Lally, L. Biegler, H. Henein, “A model for sequencing a continuouscasting operation to minimize costs.” Iron and Steelmaker 14, 1987,pp. 63-70.
[17] L. Lopez, M. W. Carter, M. Gendreau, “The hot strip mill productionscheduling problem: A tabu search approach.” European Journal ofOperational Research 106, 1998, pp. 317–335.
[18] S. Moon, A. N. Hrymak, “Scheduling of the batch annealing pro-cess, deterministic case.” Computer & Chemical Engineering 23, 1999,pp. 1193–1208.
[19] J.B. Ordieres Meré, A. González Marcos, J. A. González, V. LobatoRubio, “Estimation of mechanical properties of steel strip in hot dipgalvanising lines.” Ironmaking and Steelmaking 1, 2004, pp. 43–50.
[20] H. Park, Y. Hong, S. Y. Chang, “An efficient scheduling algorithm forthe hot coil making in the steel mini-mill.” Production planning &control 13, 2002, pp. 298–306.
[21] N. Pessel, J.-F. Balmat, “Principal Component Analysis for GreenhouseModelling.” WSEAS Transactions on Systems Issue 1, vol. 7, 2008,pp. 24–30.
[22] P. Petr, J. Krupka, S. R. Provazníkova, “Multidimensional Modeling ofCohesion Regions.” International Journal of Mathematical Models andMethods in applied techiques Issue 1, vol. 5, 2011, pp. 298–306.
[23] D. J. Renn, J. L. Stott, J. F. Vasko, “Penalty-based sequencing strategyimplemented within a knowledge-based system. Journal of the Opera-tional Society.” Journal of the Operational Society 50, 1999, pp. 205–210
[24] J. W. Sammon, “A nonlinear mapping for data structure analysis.” IEEETransactions on Computers 18, 1969, pp. 401–409.
[25] B. Scholtz-Reiter, J. Heger, M. Lütjen, A. Schzweizer, “A MILP forinstallation scheduling of offshore wind farms.” International JournalOf Mathematical Models And Methods In Applied Sciences, Issue 2,vol. 5, 2011, pp. 371–378.
[26] M. Soo Suh, Y. Jae Lee, Y. Kwan, “A two-level hierarchical approachfor raw material scheduling in steelworks.” Engineering Applications ofArtificial Intelligence 10, 1997, pp. 503–515.
INTERNATIONAL JOURNAL OF MATHEMATICAL MODELS AND METHODS IN APPLIED SCIENCES
Issue 5, Volume 5, 2011 897
9
[27] J.. Srisertpol, S. Tantrairatn, P. Tragrunwong, V. Khomphis, “Estimationof the Mathematical Model of the Reheating Furnace Walking HearthType in Heating Curve Up Process” International Journal Of Mathe-matical Models And Methods In Applied Sciences, Issue 1, vol. 5, 2011,pp. 167–174.
[28] R. Tamura, M. Nagai, Y. Nakagawa, T. Tanizaki, H. Nakajina, “Syn-chronized scheduling method in manufacturing steel sheets.” IntelligentTransactions Operational Research 3, 1998, 189–199.
[29] L. X. Tang, J. Y. Liu, A. Y. Rong, Z. Yang, “A mathematical program-ming model for scheduling steelmaking-continuous casting production.”European Journal of Operational Research 120, 2000, pp. 423–435.
[30] L. X. Tang, J. Y. Liu, A. Y. Rong, Z. Yang, “A multiple travelingsalesman problem for the hot rolling scheduling in Baoshan Iron andSteel Complex.” European Journal of Operational Research 124, 2000,pp. 267–282.
[31] L. Tang, J. Liu, A. Rong, Z. Yang, “A review of planning and schedulingsystems and methods for integrated steel production.” European Journalof Operational Research 133, 2001, pp. 1–20.
[32] H. Wai Chun, “Scheduling as a multi-dimensional placement problem.”Engineering Applications of Artificial Intelligence 9, 1996, pp. 261-273.
Andrés Sanz García , received his B.S degree in Industrial Engineeringin 1999 and M.S degree in Industrial Engineering from Universidad de LaRioja in 2002. Currently he is engaged on his PhD in the area of industrialoptimization through soft computing techniques. He is registered ProfessionalEngineer until 2011 in La Rioja, Spain. Recently, he is a professor andresearcher of the Universidad de La Rioja in the Mechanical Department withEDMANS research group. He has published several papers in conferences. Hisresearch area includes but not limited to the use of finite element analysis tosolve engineering problems, materials science, fracture mechanics, and failureanalysis of engineering materials.
Francisco J. Martínez de Pisón Ascacíbar , PhD - graduated in IndustrialEngineering from Universidad de La Rioja in 1999. In 2003 he completeda PhD degree at the same university in the area of industrial optimizationthrough data mining techniques. He is teacher and researcher of the Univer-sidad de La Rioja in the Mechanical Department. His main research interestsare data mining; soft computing; pattern recognition; artificial intelligence andindustrial optimization. He has published many papers in journals and is authorof several books and industrial patents. He has been the main research ofseveral national projects and has participated in other nationals and Europeandata mining Projects. Currently supervises several MSc and PhD students inthe area of data mining and soft computing.
Rubén Lostado Lorza , received the M.Sc. degree in engineering in 2003 andthe Ph.D. degree in engineering from the University of La Rioja, Logroño,Spain in 2010. From 2003 to 2005 he was working on numerical simulation(Finite element method) at the department of Wind Energy in the companyM-torres industrial designs. From 2005 to 2007 he was working on numericalsimulation (Finite element method) in the Automotive Technological Center ofNavarra (CITEAN). He is currently working at the University of La Rioja asprofessor of engineering projects and environmental technology. His currentresearch is the combined modeling of industrial processes and products usingthe Finite Element Method and Data Mining.
Roberto Fernández Martínez , MSc - graduated in Industrial Engineeringfrom Universidad de La Rioja in 2006. He is working in his PhD degreeat the same university in the area of industrial optimization through datamining techniques. He is researcher of the Universidad de La Rioja in theMechanical Department. His main research interests are data mining; softcomputing; pattern recognition; artificial intelligence; industrial optimizationand feature selection. He has been researcher of several data mining Projects.Currently is finishing his thesis in the area of data mining and soft computing.
Julio Fernández Ceniceros , received the M.Sc. degree in IndustrialEngineering in 2009 and the Master Degree in Project Management fromthe University of La Rioja (Spain) in 2010. He is working at the Universityof La Rioja as a fellowship and his current research is applied numericalsimulations (Finite Element Method) and Data Mining techniques in steel andconcrete structures. His interests include the behavior of bolted connections,FEM simulations, failures modes, plasticity and damage.
INTERNATIONAL JOURNAL OF MATHEMATICAL MODELS AND METHODS IN APPLIED SCIENCES
Issue 5, Volume 5, 2011 898