Application of a KDD metodology to a real industrial process

APPLICATION OF A KDD METODOLOGY TO A REAL INDUSTRIAL PROCESS VICTORIA PACHhN. ESCUELA POI,ITkCNICA SUPERIOR, UNIVEKSIDAD DE HUE LVA. SPA IN. VPACHONhNJ H U.ES JACINTO MATA. ESCUELA POLIT~CNICA SUPERIOR. UNWERSIDAD DE NU E LV A. SPA I N. 1\1 ArlA63U HlI.ES FRANCISCO ROCHE. ESCUEIA POLIT~CNICA SUPERIOR, UNIVERSIDAO DE HUELVA. SPAIN. [email protected] INFORMA'TICOS. UNIVERSIDAD DE SEVl LLA. SPAIN. KlOUEl.~lEl~~LSl.tlS.ES JOSE MARiA TEJERA. A'I'LAN TIC COPPER SA. JOSE c: RIQUELME. ESCUELA TECNICA SUPERIOR DE INGENIEROS ABSTRACT By means of an acid plant it i.s pohsible to tranxfom sulphuric dioxide (502) into sulphuric acid. using for this a set of chemicat and physical procescs. In the procehs of smelting copper mineral a large amount of SO2 is produced. This compound would be highly pollutant if it was emitted to the atmosphere. However, there are certain situations in the procesx of smelting copper mineral. in which SO2 escape to the atmosphere. This would be avoidable if we exactly knew under which circumhtances this problem is produced. In this paper we present ii practical application of KDD technique., to the chemical industry. With this technique we obtain rules that mahe poshible the definition of proceedings that should help to optimise the functioning of the hulphuric acid production system. By means of the obtained resulth we show the viability of using automatic classifiers to improve a productive process, with an increase of the production and :I decrease of thz environmental pollution KEYWORDS: Data Mining, KDD, Induhtriiil Appliciitionh, Artificial Intelligence, Classitication Rule, Data Preparation, Ddta Preprocewing, Automatic Clahsifier, Data Selection, Data Cleaning, Coding. I. INTRODUCTION In the process of smelting copper mined a larze :imoun! of sulphuric dioxide (SO?) is produced'. Thix compound would be highly pollutant if it was emitted to the atmosphere. By means of an acid plant it is possible to tramform it into sulphuric acid. wing fur this a set of chemical and physical procehwh. In this way we cliapose of B marketable prurluct and, at the same time, the environment is protected. However. there are certain du:ition% in which the gaseh esczlpe to the atmosphere, creating pollufent situations. This would be avoidable if we exactty knew under which circumxtances thib problem is produced. In order to study the viability of the improvement cif the control of the hulphuric acid producticin system, it ix essential to knaw how the \yslem works. not only from the point of view of its use but internirlly. Since at the present time we do not have a detailed model of the system, it would be necessary to give the first steps to obtain such model. Basicnlly, an sulphuric acid production system consist> in a gas flux system that conthy sulphur in a higher or lesser quantity. By subjecting the gases to the appropriate processes, we can use this sulphur to produce sulphuric acid instead of releithing it to the atmosphere. In our case of study there :ire .senwrs that take measures of diferents properties or parameters of the industrial plant. Thih measureh are stored in files, therefore we have a Iiirge database. The KDD (Knowledge Discovery in Databases) I I I process try to know why the hystem de" work propery. In this paper we present ii prndical application of KDD technique.: to the chemical industry. By means of I The rweiirch har been rciilised incollabowlion wthr Atl:inlic Copper S.A. 495 the obtaiiied result> we show the viability of using automatic classifiers to improve a productive process, with an increase of the production and a decrease of the environmental pollution. In this research. we have specifically used classification 121 as Data Mining 131. technique. With this technique we obrain rules that make possible the definition of proceedings that should help to optimise the functioning of the syatem. With the obtaining ofthe goals the company will have: an nptimisatiun of the functioning of the sulphuric acid production xystem, il greater knowledge of the \y.\tcni that will Ix: uheful tu help in a later taking of decisions and a reduction of the production cmts at the same time that a greater protection of the environment. Using data from a real industrial process. we have developed a KDD proccess, We have obtained a set of classification rutes that approach the system behaviour. We have two stages in our KDD process. In the first stage, we worked with iwo clahhes, and a high number of parameterb. In the second stage, we worked with the most aignificative parameters of each subsyxtem. and 5 classes only. In this paper, we deacribe all the stage of rhe process and the mosr representative rules. The rulex obtained from the KDD process have helped the human expert to know the rmge of values in which the system works properly. With this analyxis we probe the viability of using authomatic classificalors in ii productive process, with an increase of the production and a decrease of the cantmiinat ion. 2. SULPHURIC ACID PRODUCTlON SYSTEM In order to study the viability of the improvement of the control of the sulphuric acid production system, it is essential to know how the system works, not only from the point of view of its use but iii!ernally~ Since at the pre>ent time we do nut have B detailed model of the sy\tem. it would be neceh\ary to give the first ateps to obrain such model. In Figure I we shuw a diagram of the aforementioned system. In the system there are producer., and consumers of gases. The proceas consixts in canalising in a suitable way the gahes generated by the producers loward5 the consumers, in xuch it way thnt here ix no production xurplux at the same time that the necehsities of the c~nxumers are cared for. The produced gases are conveyed to the Blending Chamber (BC), from which they are distributed to the different consumerx. The gssex are conveyed from a point to another by blowerx. Due to the complexity of the system and to the great number of parameters and situation5 that are continuously given, sometimes the system is not stable during a certain time (that is to say, there i.s pressure in the BC). Although :it present time dl the appropriate measure.; to avoid this have been taken, a KDD procehs would help to know which rank of values of the variables that operate in the syhtem is the most suitrlble to maintain a stability in the BC and to dihcuver, which are the reasons that provoke instability to the draught in the BC. I 3. ACID FACTORY DATA The results are obtained from the acid fictury data. In our case of study there are sensorx that take measures of tlifferent properties or parameter> of the industrial plant. This measures are 496 stored in file\, therefore we have a Iiage database. The KDD process tries to know why the system doesn't work properly. Each database record consi\ts of the meawre datatime, and the tiunieric value of 52 diferent wnsorh. We have one different record each 2 seconds. We add :I new column, called class. to indicate the system behaviour. We are working with il real system. xo a change in one of the parameters in the inxtant of time t can be reflected in the clilss in the inst:ant of time t+ t. t is a difficult value to know because it is the sum of the different phixical delays of the system. So. the value of the class in a record in a determinate instant crf time could come about a change in the pnrameters of the previous insrants of time. Changes will be recorded in previws dutabiise records. The results have been obtained without consideration of delay. The tool used to get these results is C4.5 141. It is one of' the most used in ctassification. We have chosen this tool because the resultant rules cat1 be eahily interpreted by the industriirl chemistry expert>. 4. RULE REPRESENTATION A rule has two parts. the antecedent and the conhequent. The ;intecedent ih the group of conclifions that are eslablihed on the parameter values and the run\equent, in our case, is one of the classeh wsigned to the file. In Figure 2 it is \howl1 on example of the decision rules obtained during the process. ....... " .............. . ,^ .>, " ,,. .. .I ... -. ... ...... _. ............ .. - w *.*L I 3 I> ..... The expression "Rule 3" indicates only one label assigned to the rule. The rule in Figure 2 express that if TCM22 <= l~l9X33 and TCM35 > Y.tI2952 iind TCMJll > Y.4320X nnd TCM4 7 0.0140164 then, the cla\s is Class 0. If all the mtecerlent conditions are true. the .systeni would not be working pruperly since ti Class 0 ia obtained as consequent, Moreover. parenthetically. it is shown the number of records from the training file that fulfil that rule. This value gives us il sign of the importance of the rule. If the rule is true in ii high percentage of the rccordh. this indic;ile, that we are in front of o frequent hiation for the syhtem. If a record fulfil.; the antecedent of 3 rule. but not the consequent. that is, fulfils the parameter restriction> but is of a different claw, it is considered as a learning mistake. To have a mure intuitive interpretiltion of the rule fur the industrial expert, a graph where the maximum and minin~utn values that the parameter can t;ike (grey color) and the specified rule vitlue (bldch color) are 4wwii. hiih been included. When TCM22 takes IOW and medium values (below 1.19) and TCM35 does riot take high value\. when TCM411 lakes medium and high values and TCM44 takes medium and high viiluec, the syhtem operetion is not optimum (Class 11). 497 r 5. GETTIKG RESULTS WITH A KDD PROCESS The KDD process is a cyclic prucess with six stages (Data selection, Cleaning, Enrichment, Coding, Data Mining and Reportint). The results obtained in each period change as the KDD process stages develop. In order to obtain enough accuracy in the rules extracted from the data, different preprocessing stage prototypea have ken tested and, therefore. different Data Mining process results have been obtained. 5.1. KDD first stage In the Data Selection stage, study of the correlation between columns was realized. One column was setected from those that were highly correlated (correlation higher than OX), eliminating the 0ther.s since we can consider that they provide the same information to the KDD process. In the Cleaning stage, as the data come froin a red industrial system. we have to check some probletnh that could appear, and in the1 case we have to .solve them: - It is pohrible that wine parrs of the system are not working in the moment of the data acquisition (for example, because of a failure), so errors in the sample taking can be produced. These errora are hhown as out of range viiluex for that parameter in the whole data or as a constant value. Each posxible error was detected and eliminated from the respective pdrameter slutly. During .some specified period?; of time, because nf unknown reasons, the data acquisition can be incorrect (of a11 the pirametera or only some of them). We have io look for each row of the file correspmding to inctirrect data acquisition and eliminate from the xtudy the respeciive period of time information. After the data helection and cleaning stages. 36 columnh were selected from the htarting 57. in Ihe Curling slage, a new urtlunin named label or cla.~. with a Class 1 ur Class 0 value wiix added. The viilue of this column is sei by the value of the parameter named TCMI2 (Real value of yrexsuere in BC). What we have to rtudy is when the draught exist5 in the BC. Talking to the expert, it was established that there is pressure in BC when the TCM12 value is in the interval 1- 2.7, -2.31. In this case, we can assume that the syrtem ir working in an optimum way. Then, we have a Class I record. On the other hand, it i.s said that the system is not working in an optimum way when the TCM12 value ih not between the interval 1-2.7. -2.31. In this case, we have a Class 0 record. The assignment of these clashes is neceawy tu create the Training file which the Data Mining algorithm works with, xince it ix a supervised learning process. So, before preprocessing we have a database with 43201 records and 57 columns. Before preprvcehaing we obrain o dalabaae with 43201 record\ and 36 columnh. This database has YMl4 recordh of Class I and 33397 records of Cia.;?; 0. When the data preparation *rep is finixhed, the knuwledge extracted from rhe Data Mining a[gorithm is represented as decihion rules. 5.2. Rcsults and conclusions of the first siage. In this first dag, only rules whith apparition frequency is 1% or higher have been raken into accuuiit. The file has 43201 records, xo, only rules that are true in more than 430 records will be shown. In futiher studies, rules with a lower apparition frequency can be taken into :iccount, if neccehary. The number of classification errors compared with the number of instances gives us an idea of how ”reliable” a rule ib. Having analized the rules, it is found that the parameter TCM37 (Factory 3 blower speed) is the most determining. High values or very high value> of TCM37 imply pressure in BC.This i.s the way this fact might be explained: TCM37 refers to the blower speed which extracts the gas from BC to the acid factury. Beaides, this blower is the moxt powerful one among the three blowers used ti) extract The gas from the BC. When the blower reaches ii very high speed for a long time, it is not possible to extriict more gm because it is working at it:, maximum. That produces an accumulation of gas in the BC which create> presxure, and Therefore. Class 0 records. - 498 The extrnc!ed rule$ involve parameters from different .\ubxystems in its predecessor. which makes its explanation difficult to an expert. Then, it was necessary a higher decrease in the number of parameters if we wanted that the ones which took part were the most significant ones. This idea would m;ke the rules easier to understand. To reduce the system ih much as possible. it was thought to work with the TCM to obriiin a produclive system equivalent to the 3 productive systems and another one with the consumers. This way the system wuuld only cunsist of a productive system, the blending chamber and a consuming bystein. The Clahs 1 rules that can be deduced from the data, gather il very low reeds pecentage. and therefore. not very much significant. In order to know if the .+ystem is working properlq-. more Class I data are needed, or at Iea\t, a significant number of them compared with the Clau 0 oneh, Besides, the htudy could be improved if the TCM 12 group of values could be heparated in more classes, not ody Clahs 0 and L. We could have 3,4 or 5 m ore clashes to express the different states of the system. From the results obtained in the First Stage, we can conclude that, although the\e results were satisfactory, more useful rule5 could be found if the number [if parameter.; wds reduced imd the number of classes wah increased. With these changes, a second stage of the KDD proceu could be developed. In this stiige. the preprocessing step cuuld be done again to prepare the data for !he new circumstances. 5.3. KDD Sccond Stagc first xtage. The differences between them are: - In this stage, the study is a consequence of the pobcible improvements that were found after the In the first stage, the value of the ctaxs depended on a calculus on the TCMIZ value. However. Clahi; 0 was detected corresponding with a correct behaviour of the system wax set by the expert to solve some situation*. To solve this problem, it is neccesary to think of mother wny to calculate the clahh. The new claseh are obtained calculating the error between the draught set point value in the BC (measured by TCM42) and the real draught value in the BC measured by TCM 12. We obtain: Pressure, High presbure, Draught and High Draught. To reduce the number of ptrometers, we must take into xcouiit the blowerh and valves that take the gits tu BC and the blowers and valve\ that extract gas from BC. Table I show.% the range of value; corresponding to each class artd the anloutit of instances of each clas\. We have selected only those rules that are true in more than 1% of the instance.; of its cI:iss. The fourth column exprehses the value that indicates if we can consider that a rule i'i true in il xignificative number of me>, Range of value3 (TCM42-TCM12) CI;is.; Number of instmceh (~t'e;ich class (-0.5..0,5) OK 29843 (05, I ,5) Pressure 7393 (1.5 ...+ fnf) High I295 (-lS.:O,5) Drought 3x25 (-id ..-1.5) High x45 pressure Draught 1 % 298 739 13 3K x Table 1. Number of in\tnnces of each CIB~S. 5.4. Results and conclusions of the sceond stage Developing the second stage, we can get rules that describe the optim;il behaviour (c1;iss OK). However, theae ruler are easier to interprete by the human expert. We obtain the same conclu&n than Firit Stage about TCM37. High or very high vdues of TCM37 blower imply pressure in BC. 499 In ddition, high or very high vhe\ of other blower.. (TCM30 and TCM46) imply prewre in BC. Medium or low vdum of theze blower5 imply CIaz5 OK Figure 3 \how\ the \ome uf the mmt repre\eiitdtive rule\ obtdined in the \econd stage We don't have rl lot of iii\tdnce\ ofcldwe... High pre\wre. Draught and High Drdught , zo the rulez with these cldske\ are not cdtegurid enough ""- ~~ *" ~ - -.7 " - - - j. -8%- Y CL*SPOI "*" W.". ... ......................... ~~ ..- -.. ........ - ,3~.ill. ". . ~ r-. I .b. ."', ........... .. ~ .!. ._ " ",,. -. " I" .............. " ... ".~ ..... "." L-'l+l, .... I I!%% -., Figwe 3 , Gi.upAic itpi~esei~fuficrn o/SwonJ Sfoge RuIw 6. COKCLUSIOKS Uxing data from a real industrial process, we have developed a KDD proccev,. We have obrained a set of classification rulex that approach the system behaviour. We have two stages in our KDD process. In the first stage, we worked with two c 'es, and a high number of parameters. In the second xtage, we worked with the most si ficative pardmeters of each hubsystem, and 5 clahses only. In this paper, we describe dl the xrages of the process and the most representative rules. The rulex obtained from the KDD process have helped the human expert to know the riinge of values in which the system works properly. With this analysis we probe rhe viability of using authomatic clahsifiers in a produciive process, with an increase of the production and a decrease of the contamination.Future works will study the errors due lo the hystem delays and develop a method of obiaining rules with delay. 500