Graphical Estimation of Permeability Using RST&NFIS

Graphic al Estimation of Perme ability Using RST&N FI S H.Owla deghaffar i , K.Shahriar W. P ed rycz Department of Mi ning & Metallurgical Engineering Department of Electrical and C omputer Engineering Amirkabir Un iversity of Technology University o f Alberta Tehran, Ir an Alberta, C anada h. o. gh a ffar i @ gmai l. com ; k.sh ah ri ar @a ut. a c.ir ped r ycz@ ec e. ua l ber ta . ca Abst ract - Th i s paper pur s ues som e appl ications of Ro ugh Set Theo ry (RST) and neura l-fuzz y m odel to analy sis of " l uge on data". In the manner, us ing S e lf O rganiz i ng Map (SO M) as a pre-processin g the data are scaled and then the dom inant rules by R ST , ar e el i c it e d. B as e d on t hes e r ul e s vari at i on s of permeab ility i n the diffe rent leve l s o f Shiv ash an dam, Ira n h as be e n hi g hl ig h te d . The n, vi a u si n g a comb i n i ng of SO M an d an ad a p ti ve N e ur o-F uzzy I nf e r e nc e S ys t e m ( NFI S ) a n ot h er an al ys i s on the data w as carri ed out. Fina ll y, a br ief co mparison betwe en t h e ob t a i ne d re s u lt s of R ST a n d SO M - NF I S ( b ri e f l y SO NF I S) ha s been render ed. I. INTRODUCTION Dur in g th e dam str uct ur es des ign , on e of th e most sig n if ica n t is sue s is t he e stim atio n o f pe rme ability v ariatio n s in d iffe r en t le ve ls o f the d am si te . H owev er, p red ictio n of pe rmeab ility , usi ng o btai ned data , f rom i n-s itu te sts is a b ig ch al l eng e. R elati ng to the d ete rmi n atio n of po tentia l w ate r flo w p aths wit h in the r ock ma ss, un der l y i n g a pot ent ia l dam str uct ur e i s es pecially impo r ta nt and this has an e xte n siv e imp act o n the planning o f grouting p r oce dures [1] . Due to asso ciation o f unce r tainty and vague n ess w ith the mo nito r ed data set, pa rtic ula rly , res ul ted f r o m the in -situ te sts (such lugeo n t est), acco un ting r elev an t approac h es such prob ab ility , Fuzzy Se t Theo r y (FST) a n d R ough Se t Theo r y (R ST) to kno w led ge acq uisitio n, ex tr actio n o f rules and pr e di cti on of u n kn own ca ses , m or e th an th e pa st ha ve b e en distinguis hed. T he RS T i ntroduce d by Pawlak has o ften pro ve d to b e an e xc elle nt ma them atic al to ol fo r the an aly sis of a vague de scription o f obj ect [2], [3] . App l i cat i on of RST in differ ent fi el d s of th e app li ed sci en ces h a s been r epor ted [ 4], but d evelopi n g of such syste m (ba sed on a pprox i mat e an al y si s) i n r ock en gin eer in g ha s no t been out stan d in g, r ela ti vel y. Th e ma in r ea son i s devel opi n g and applic ation of n ume r ical mode lling (1-1 mapping mode ls), fr om bi r t h d a y, in th i s fi e l d. Fr om th e ot h er ha n d, em be dd i n g of t h e r ock en gin eerin g da ta with th e sev er a l of t h e un cert ain ties and am big u iti es , per su a de to con sider a p pr ox i ma t e an a l ysi s m et h od s . Un d er th i s id e a an d ba s ed on Informatio n Gra n ulatio n Theo r y (IGT), w e ( are) dev el oped (d evel op in g) a set of a l gor ith m s in rela ti on to th e trad i ti ona l mo de lling in roc k mec hanics [5] . Fig 1 s h o w s a g ene ral pr ocedur e, in whi ch th e IGT accom pan i es b y a pr edefi n ed proje ct b ased r ock e ngineeri n g des ign. Af ter dete rm inatio n of con st ra in t s an d th e as soci at ed rock en g in eer in g con si d erat i ons, the in itial gra nul atio n of info rmatio n as w e ll as nu meric al (dat a bas e) o r li ngu istic f o rmat is ac co mplis hed. Imp r ov eme nt of mo delli ng inst r uments b ased upo n IGs, w h ether in inde pende n t o r af f iliated s hape w ith hard co mputing met h o ds (su ch fu zz y fin ite el em en t, fuz z y boun d ar y elem en t, stoch as ti c fin ite el em en t …) are n ew ch a llen g es in th e curr ent di scus si on . Th u s, on e can em pl o y such m e th od a s a n ew m eth od ol og y in desi g nin g of rock en g in eer in g flowch ar t s [5 ]. Fi g.1 A gene ral met hod olog y fo r b ack analysis based on IGT In thi s s t ud y , a ccor d in g to th e “m odel l in g ins tr um ent s” in fig1, usi n g Self Organizing M ap (SOM), NFIS, and RS T, so me analy sis o n the permeab ility da ta, r ounded up fro m Shivasha n dam si te lo cated in n o r th w estern o f Iran. I n our ap pr oa ch, S OM, NF IS and RST ar e uti li z ed t o cons tr u ct IGs . II. INSTRUMENTS A. Self Organ izing feat ure Map (SOM) Ko honen self -organizing n etwo r ks (Koho n en f eature maps o r to po logy -pres erv ing map s) are c o mpe tit ion-b ase d n e t wor k pa r a di gm for d at a c l u st er in g. Th e l ear n i n g pr oce d u r e of Ko hone n fe atur e maps is sim ilar to the co mpeti tive lear ni ng netw orks . T h e m ain ide a be h ind co mpetitiv e le arnin g is simple ; t he winne r take s al l. The co mpetitiv e t r ansf er fu nct io n r e t ur n s n e ur a l ou t p u t s of 0 for a ll n e ur on s ex ce pt for t h e winn er wh i ch recei v es t h e hi gh es t net in put with out p ut 1. SOM ch an g es al l weig h t vector s of n eu r on s in the n ear vi ci n it y of t h e wi n n er n eur on towa r d s th e in p ut vec t or . D ue to this pro pe rty SO M, a r e use d to red uc e the di mens iona lity of co mple x data (da ta clus teri n g) . Co mpe t itive lay ers will auto matic ally le arn to clas sify input v ec to rs, the cl asse s that th e com pet i ti ve la yer fin ds ar e dep en d onl y o n th e di stan ces be t w een input v ecto r s [6]. B. Neu ro-fuzzy infere nce syste m (NFIS) Th er e ar e differ en t solu t i o n s of fuz z y i n fer en ce syst em s. Tw o w ell- know n fuzzy mo deling me tho ds a r e t h e Tsu k amo to fuzzy mode l and Takagi– Suge n o–Ka n g (TSK) mode l. I n the pr esent wor k, on ly th e TS K model h as been con sid er ed. Th e T SK fu zz y in fer ence s ystem s ca n be eas il y im pl an t ed in th e for m of a s o ca l l ed N eur o- fu z z y n et wor k s tr uc t ur e .in this stu dy , we h av e e mploy ed an ad aptiv e neuro -fu zzy in fer en ce syst em [7 ]. C. Rough Se t Theory (RST) Th e r ough set th eor y in tr oduced b y Pa wla k [2 ], [ 3 ] h a s oft en prov ed t o be an ex cel l e n t math em ati cal t ool for th e analy sis of a v ague de scription o f object. The adje ctive vague ref erring to the qu ality o f i n fo rmatio n mea n s i nco ns iste ncy , o r amb igui ty whic h fo llow s f r o m inf ormat ion g ranul atio n . A n in fo r ma t i on s yst em i s a pa ir S = < U, A >, wh er e U i s a no n empty f in ite set c alle d t he u nive rse and A is a none mp t y fin ite s et of at tr i but es. An at tri but e a can be r egar ded a s a f unctio n fro m the do main U to so me v alue se t a V . A n in for m at i on syst em ca n be rep r es ent ed a s an a t tri but e- val u e tab le, in w hic h row s are labe led by ob jec ts o f the u nive r se and colu m n s by at tr i but es. With ever y su bset o f a ttr i but es B ⊆ A, on e can ea sil y asso ci a t e an eq ui va l ence r el at ion B I on U: {( , ) : , ( ) ( )} B Ix y U f o r e v e r y a B a x a y =∈ ∈ = (1) Th en , Ba B a II ∈ = ∩ . If XU ⊆ ,t he se ts [] {: } B x Ux X ∈⊆ and [] {: } B xU x X ϕ ∈≠ ∩ , w h e r e [ ] B x den otes th e equ i val en ce cla s s of th e object x U ∈ r ela ti ve t o B I , ar e cal l ed th e B- low er and th e B-upper a ppr oxi ma t i on of X in S an d d en ot e d by BX an d BX , r esp ect i vel y. Con si der 12 n {x , x , ..., x } U = an d 12 n {a , a , ..., a } A = in th e infor mati on sy stem S= U, A ≺ . B y the d isc ernib ility ma trix M (S ) o f S is mea nt a n n*n mat rix such th at { } :( ) ( ) ij i j ca A a x a x =∈ ≠ ( 2) A d is c er nib il ty f unc tio n s f i s a fu n cti on of m Bool e a n va r ia bl e s 1 ... m aa cor r e sp on din g t o a t tr i bu t e 1 ... m aa , res pe ctiv ely , and de f ined as fo llow s: 1 ( , ..., ) { ( ) : , , , } s m ij ij f aa c i j n j i c ϕ =∧ ∨ ≤ ≠ ≺ (3 ) Wh er e () ij c ∨ is the dis ju nctio n o f all v a riab le s w ith ij ac ∈ . Usin g such dis cri min ant ma tr i x th e appr opria t e r u l es are elicited. In t hi s study w e have dev elo p ed de pendency r ule generatio n –RS T- i n MatL ab7, and o n this a dded too l bo x oth er ap pr opr i ate al g ori th ms have been pr epar ed . III. T HE PROPOSED PROCEDURE Deve l oped algo r ithms use f our basic axioms upon the bal an c in g of th e s uccess i ve g ran ules a ssu m pti on: Step (1): d ividing t h e mo n i to r ed data i nt o g r oups o f tr aining an d t es t in g da ta Step (2) : first g r anulatio n (crisp) by SOM o r ot h er c risp gra n ula tion me thod s S t ep ( 2- 1 ): sel ect in g th e l evel o f gr anu lar it y ran dom l y or de p en d on th e obt a i n e d er r or fr om th e NF I S or RS T (r eg ula r n e ur on gr owt h ) Step (2 - 2): con str uct i on of t h e gran ul es (cr i sp ). Step (3) : sec o n d gra nulation (f uz zy or rough IG s) by NFIS o r RST Step (3 -1 ): cr i sp gran u l es as a n ew data. S tep (3 -2) : sele cti ng t h e le v el o f g ranula rity ; (Erro r le v e l, numbe r of rule s, s trength t hres h old...) Step (3- 3): c hec king the sui tabil ity . ( Clo se -ope n ite ratio n: r eferr in g t o th e r ea l data an d rein spect cl osed wor l d ) Ste p (3-4): c onstruc tion o f fuzzy/rough g ran ule s. Step (4): ext raction o f know ledge rules T h i s st ud y in vol ve s on l y SOM an d N FI S in th e men t i on e d proce dur e a nd ot h er figu res o f suc c ess i ve gr anulatio n ca n b e fo llow ed in [8], [9] . Se lec tion o f initial c risp g ranule s ca n b e suppo sed as “ Cl os e Wo r ld Assu mptio n (CWA)” .But in m any appl icat io ns, t h e as su mp tio n o f co mple te inf ormat ion is no t feasi bl e, an d onl y cann ot be us ed. In such ca ses, an “Op en Wor l d As s um p ti on ( OW A) ’, wh er e in for m at i on n ot kn own b y an ag e n t i s as sum ed t o be un kn own, is often a ccep ted [ 10 ]. B alanc i ng as sum ptio n is satis fie d by the c los e -o pen ite ratio ns: this proc ess is a guidel in e to bala n cing of crisp a nd s ub fuzzy /r oug h granules b y some r ando m/regula r sele ction of init ial g ranu les o r o the r o pti mal s truc tures an d i nc reme n t of suppo rt ing rules (f uz zy pa rtitions o r increas ing of l ow er /upp er app r ox im atio n s ) , gr adua lly . T h e over a l l s ch em a ti c of Se l f Or g an i z in g Ne ur o-F u z z y Infere n ce Sy stem -R an dom : SONF IS-R has b een show n i n f ig2. D ete rmi natio n of gra nul atio n le v el is c on tro lled w it h th r ee ma in par am e t er s: r a n g e of n e ur on gr owt h , nu m be r of rule s a n d e rro r le v el. The ma in b enef it of this alg o rithm is to l ook i n g for be st s tr u ct ur e an d r ul es for two k n own in t el li g ent sy ste m, w h ile i n i n de pe nde nt si tua tio n s e ac h o f them has so me a p pr op r i a t e pr obl em s s u ch fin d in g of s p ur i ou s pa tt er n s for th e large da ta se ts, ext ra- time tr ai n ing o f NF IS f o r l arge d ata s et. So, we ca n u se NFI S as an org aniz in g m ea sur emen t. It must be notice d by employ in g a r egu l ar neu r o n grow th /rule-or o the r a ppro priate paramet ers- a nd util izing ot her natural c omputing methods wi t hin a n int elligent c ommun ity (netw or k), ph ase tr ansi tion of such comp lex s yste m, in fac in g with the curr en t d eb it-I /O-, can be e valu ated . Fi g.2 A comb i ning of Self Organiz ing fe ature Ma p and Neur o-Fuz zy Inference System (SONFIS -R) IV. R ESU LTS S h ivas ha n h y droe le ctric ea rth da m is lo cate d 45k m no rth o f S a r d a s h t c i t y i n n o r t h w e s t e r n o f I r a n . G e o l o g i c a l inv estig atio n f o r the s ite s e lec tion o f the Sh iv asha n h ydr oelect r i c power pl an t was m ad e with in an ar ea of a bout 3 square k i lomete r. The w idth o f t he V-s haped v alley with similarly sl oping fl anks, at t h e e lev ati o n of 1185m and 131 0m wit h r esp ect to sea l evel ar e 3 8m an d 467m, r esp ect ivel y . In or d er t o obt a in en gi n e er in g g eol og i c a l in for m at i on , bo r eho les w ere d rilled in diff erent po i nts o f S hiv asha n d am ’s ar ea . Tota ll y, 2 0 bor eh ol es ha ve been dr ill ed and con seq u entl y abo ut 789 obj e cts we re r esul ted. W a te r P r ess ur e T est (W PT) has use d f o r dete rminat ion o f t his a r ea’ s pe rmeab ility . W PT is an e f f ectiv e me thod f o r wid ely de terminat ion o f ro ck m as s pe r m ea bi l i t y. Th e Lug e on va l u e, whi c h i s al s o kn own a s th e Lu ge o n n umb er (N Lu ) is de fined as fo llow s: Lu=W ate r take (lite rs/me ter/m in)*[ 10(b a rs)/ actu al test p ress u re (b ars) ] The Lu ge o n uni t is not s tate d as a ratio o f pe rmeab ility , but t o g et a se n s e of p r op or ti on , i t mi gh t be r el at ed su ch th a t : 1Luge on=1.3*10-5 c m /s. I n practice, usu a lly , t he Lugeo n tes t is ut ilize d b efo r e g r ou tin g to de ter mine qua nti tativ ely th e vo lume o f w ater t ake pe r unit of time. The max imu m meaningf ul Luge on is c o n side red 100 . A ge n eral asses sment fr om al l of t h e bor e h ol es h a s been sh own in fi g 3. To ev aluate the pe rme abili ty due to the lu ge o n v alue s w e fo llow two situatio ns: 1) utili zin g o f S ON FIS -R and R ST- 1 on the f ive chief attributes (fig3); 2) direct appl i catio n of RST - 2 an d NF I S on t h e loca l coor d in at es of d am si t e (as co nditio n al attrib utes) an d luge on v alues (as dec isio n pa rt) to depict 3D Iso-s ur fac es of l ugeo n variatio ns diagra ms. A naly sis of first situ atio n is sta r ted of f by setti ng n umb e r of close -open i teratio n and max imum nu mbe r of rules equal to 10 and 8 (n umber of ru l es = 5 to 8) in SO NFIS-R, resp ecti vely. T he er ror measu r e cr iterio n in SO NF IS is R o ot Me a n Sq u ar e E rro r (R M SE ), gi ve n as be l o w: *2 1 () m ii i tt RMS E m = − = ∑ ; Wh er e i t i s ou t p u t of S ON FI S an d * i t is r ea l an swer ; m i s the nu mbe r o f te st d ata (te st o bj ects) . I n the res t o f pape r, le t m=93 and number o f training da ta se t =600 . Figs 4, 5, 7 in d i cate th e r esu lt s of th e afor esa i d system . In di cat ed posi t ion in fi g 5 st at e s mi n i mu m RM SE over th e i t er a t i on s an d u sin g 5 r ul es. Fi g 4 sh ows ou r m ean a bout str u ct ur e d et ec t i on . With 63 neu r o n s in SO M, w e acqui r e so me do mina nt patte rn s on th e pr oblem ’ s sp a ce. Fig .3 Real data set- Z,L,RQ D ,T .W.R &lu geo n- in matri x plot for m (as train i ng data set) TA BLE I T HE R E VE LE D CODE S OF T YPE OF W EATH ERIN G R OCK (T WR) , MW: M EDIUM W EA THER IN G , SW : S LI GHTLY W EATH ERIN G , CW: C L AY W E ATH ERIN G , HW: H IGH W E ATH ERIN G ; Typ e of we athe ring Asc ribed code Fr esh-M W 1. 5 SW -MW 2 Fres h -SW .5 Fre sh 0 MW 3 CW 2.5 SW 1 HW- MW 3. 5 HW 4 Fi g. 4 Mat rix plo t of crisp g ra nul es by 7 * 9 grid to polog y SOM afte r 500 epo c h s on th e tra i ni ng da t a s et (F i g3 ) Fig . 5 a) S ON FI S - R r e su lt s w it h ma xi mu m nu mb er o f ru l e s 8 an d cl o s e -o p en it e rations 10; b) answe r of se lected SONF IS-R on the t es t dat a Fig .6 Te sting re sult R ST (1 ) on th e test data Fi g.7 Fi n al mem bers hip f unc tions of i nputs in SON FIS- R Fi g.8 Res ults of trans fe rring attrib utes(X, Y, Z and l ugeon) in five catego r ies by 1 -D SO M Fi g .9 3D views of luge on var iations by RST (2) ; accom plish ed by RS T (2) and five sc al ing of attribut es. Numbe r 6(mo re th a n 5) char acte rizes amb iguity and u nk n o w n cas e s With augm en tin g of clos e-op en it era ti on s SONF IS -R eme r ge s mo re n ea r mi n RM SE v alue s, w hile it is le d to t he dif feren t o utc o mes , not c ert ain ly low est mi n imu m RM SE. F ig 6 s h o w s t h e p e r fo r manc e of elic ited r ule s b y RS T-1 o n the clas sif icatio n o f tes t data. To tal ex trac ted r ule s o n t h e t raini n g da ta set, in thi s cas e, wer e 63. Now , w e inve stigate d ire ct ap plic atio n of RS T and N FI S o n the lo cal c o o rdinate s o f dam site (as c o ndit io nal attr ib ute s) an d l ug eon va lu es (a s d eci si on par t ) t o d epi ct 3 D Is o- su r fa c es of l u ge on var ia t i o n s dia gr a m s . Fig 9 sh ows t h e var ia ti on of the lugeo n data i n Z*= {1} to {5} whic h has b een acqui r ed b y se r ving f iv e co n dit io n a ttr ib utes in R ST ( f ig 8; the sy mbo lic val u es by 1-D SOM - 5 n eur ons ). Th e cat egor i es 1 t o 5 state: ver y lo w, l ow, me d i um , hi gh, an d ver y h i gh , r es pe ct i ve l y. Num ber 6 (m or e than 5 ) char a ct er iz es am big ui t y and unkn own cases. T o cl ari f y of per m ea bili t y chan ges , in con seq uen t par t of rule s, t h e lo we r valu e o n the sy mbo lic lug eo n v alue s w h ic h hav e rel ativ ely simil ar c atego r y -fo r exa mple 1 ,2,3 o r 2 , 3 o r 3, 4, 5- h a ve been con s i der ed. With ser vin g NFI S o n s uch a t tr i bu t e s( X , Y, Z & l ug eon - wi th ou t sc al i n g) , p er m ea bi lit y variatio n s in figs 10, 11 ha s b ee n port ra ye d. I n thi s step, t hr ee MFs (Gaussian a s l ik e as SONFI S) fo r input paramete r s have be en utilized. I n Conseque nt of com p a r i s on bet w een t h e r es ul t s of RST a nd N F I S, on e ma y in t erpr ets th e var iat i ons , for in stan ce in Z= {2} i s th e supe rpositio n of su b le vels, invo lve d Z=1160 to 1200 b y NF IS , app r ox ima tely . So , the co mpatib il ity of the res ult s, de r i ved fr om RST an d NF I S c an be pr obed b y c om pa r i son of fig 9& 10. T h e f orec asted domains-g ray c olo urs- in fig 9, by RS T-2, have b een coincided by same regions in f i g 10, clos el y. It m u st be n oti ced th at th e RST mod el ha sn ’ t cov er ed th e high pe rmeability zones, b ec ause of employ in g conservativ e w a y in es tima tio n of dec isio n part w he reas t h e NFIS has expo s ed suc h pos sible territories . T h e rate o f lugeo n variatio n s, o r de n sity of pe r m eab le parts, disti nguishes the zo nes w ith c apab ility of poss ib le s pr ing o r ho le. S uc h cav iti es in th e dam str uct ur es disc uss ed a s “k ar s t s”, whi ch ar e th e main cha rac ter istic s o f the lime sto ne de po sits (f ig 12) . The e n tire of extr acte d ru les in NF IS is ac co mplis hed under subt ra ctive cluste r ing met h od [7] . To f in d out the co rr ela tio n be tw een ef fec tive parame te r s and p r o cu ri ng of val i d pat t ern s of th e r ock ma ss- in the dam si t e- on e may emp loy the si mila r p roc ess of NFIS o r R ST to es tima te alteratio n s of RQD and T .W.R (f igs 13 and 14 us ing 3, 5 M Fs in NFI S, r esp ecti vel y ). The con trar y out put s in some zones with genera l con t ext u al a s socia t ed ru l es a bout RQD an d lug eon , i m pli cat e to th e rel at i vel y compl ex st ru ct ur es a boa r d th e r ock ma ss. Apar t fr om a few det a i l s, com par i son of r esu l t s in di cat es th ree over a l l z on es in th e r ock ma s s: in fir s t z on e th e th eor eti c r ules (s uc h rev erse re late be twee n RQ D& lu ge o n) a re s atis f ie d, b ut in o t h e r z o n es, t h e s aid rule is disregarde d. T o fin di n g o ut of th e ba ckgr ou n d on th ese ma jor zon es , we r efer to th e clu st ered data set by 2D SOM w ith 7*9 w eights in c ompe titive lay er (f ig 5-b ), on t he f i rst se t of th e attrib utes . The clu ste r ed and grap hical e sti matio n d isc los e s uitab le coo r di nat io n, rela tiv ely . For ex a mpl e in fi g 5-b, we h ave h i ghl i ght ed th r ee d i stin ctive pa tt er n s among lu geon an d Z, RQD, TWR. O n e of th e main reaso ns o f b eing suc h pa tte rns in t h e inve stiga ted ro ck mass is in th e defin it i on of RQD. In mea sur em en t of RQD, th e dir ecti on of j oin t s h a s n ot been con si der ed, s o tha t the r ock mass es with ap p rop r iate jo ints may fo llow h igh R QD . Fi g .10 L ugeon var iat ions in z= 1160 to z= 1200; acqui red by NFI S Fi g. 11 A c ross sec tio n pe rspect ive of l u geon c hanges obt aine d by NFIS Fi g.12 Th e rate of lugeon vari atio ns-poss ible sp rings and cavi ties on the NFIS pre dictions (d ivergence of lugeon val ues ) Fi g.13 Is o-surf a ces of RQD by NFIS Fi g.14 T .W .R v ariations- X direct c ross s ect ion view V. CONCLUSION Th e r ole of un cer ta in t y in g eomech an i cal in for mati on is undeniable fe ature. I n deed, w ith dev eloping o f new ap pr oa ch es in in forma ti on th eor y an d com put a tion a l inte llig e n ce , as w ell as , s of t compu ting a pp roac h es , it is n ecess ar y to cons i der thes e ap pr oa ches to bett er un der stan d of na t ura l even t s in rock ma ss. Un d er th i s vi ew an d gran ula tion the o ry , we propo sed tw o mai n al go r it hms, to co mplete soft granules co n structio n i n not 1-1 mappi n g le ve l of m ode l ing: Sel f Or g ani zin g Neur o-Fu zz y In fer en ce S y st em (Ra n d om an d Reg u l a r n eur on gr owt h - , SO NF I S -A R- an d S el f Or ga n i zi n g Rough S et Theo r y (SOR ST). So, w e use d SON FIS-R to ana ly sis of permeab ility in a dam s ite, Ir an. S o, d ir ect imp l em enta ti on of NFI S an d RST on th e lu geon data set w as prov ed that t h e sugge sted met h ods c ould b e ap p lied , su ccess fu l ly. Fr om th e menti on ed an alysis t h e fo llow ing r es ults ca n be ded uc ed : 1- D ete ctio n of the pe rmeab ility v aria tio ns in s uc ce ssive leve l using NFIS and R ST 2- Elici tatio n o f the d o min an t s imple ru les b etw ee n ef fec ti ve pa rame ters 3- A pr e- pr oces si n g on th e sca t t er l ug e on d a t a us in g best SO M an d int er pr eta ti on of th e res ult s R EFEREN CES [1] A . C. Houlsb y, Cons truct ion, and Des ign of Cement Grouti ng ,Joh n Wi ley & Sons , Inc., New Yo rk, 1990. [2] Z. Pawl ak, ”Rough se ts ”, Int J Comput Inf orm Sci 11, pp. 341-356, 1982. [3] Z. Pa wlak, Rough Sets: Theoret ical A spects R easoning about D ata , K luwer academ ic, Bos ton, 1991. [4] S.K.Pa l an d P.M itra,” P attern R ecognit ion A lgori thms for Dat a Mi ning ”,Ch a pman& Hall/C RC,Bo ca Rato n,2004. [5] H.O wlad e gha ffar i a nd H . A g haba bea i , “ Ba ck a na ly s i s ba se d o n SO M - RS T system”, Accep ted in the 10t h Internat ional Symp osium on Lands lides and E ngine ering and En gineere d Slo pes , X i’ an , Chin a, in pr e ss. [6] T.K ohonen , Self -Organi zation and A ssociat e Memory , 2n d Ed , Sp r inge r – Ve rlag, Berlin , 1987. [7] J. S. R .Ja ng , C. T . Su n, a nd E .M izu ta ni, Neuro-F uzzy and Soft Comp utin g . New jersy , Prent ice Hal l , 1997. [8] H.O wlade gha ffar i, M.Sha r if zadeh and K.S ha hrair,” Mod ellin g in rock en gi n eer i ng ba se d on i n for ma t i on gr a nul a ti on th e or y”, a c ce pt e d i n “ 4 2 n d U.S. Rock Mechani cs Symp osi um & 2n d U.S. -Canada Rock M echanics Symposi um , San Fra n ci sco , i n pr es s. [9] H.Ow ladeg ha ff ar i and W. Pedrycz, ”Phase trans i tion i n the inte lligen t c ommuni ties “, pre par ing to Phys i cal Revi ew E, unp ublis hed. [10] P. Dohe rty, J. K achniarz, and A.Sz ata s ,” Us ing co ntex tually closed que ries fo r loc al cl osed-wo rld reas onin g in r ough know ledge dat a base . In rough-ne ural com puti ng tec hniques for c omputing wi th w ords , eds. S. K .Pal , L. Po lkows ki and A. S kowron, pp 219- 250, 2004.