You are on page 1of 10
112 + 201S TRANSACTIONS OFTHE SOCIETY FOR MINING, METALLURGY & EXPLORATION Ground rippability classification by decision trees S.R. Dindarloo* Ph.D. candidate, Department of Mining and Nuclear Engineering, Missouri University of Science and Technology, Rolla, MO, USA “Corresponding author email: srdSzb@mstedu E. Siami-Irdemoosa Ph.D, candidate, Department of Geosciences and Geological and Petroleum Engineering, Missouri University of Science and Technology, Rolla, MO, USA Abstract Ease of excavation or rippability, also called diggability or excavatibility, is a critical decision-making parameter in the selection of both the overburden removal method and the required equipment in surface mining and geotechnical engineering projects. The most widely used method of estimating rip- ability is rock mass classification. Although there are different classification schemes in the literature, all of them have the two major limitations of sharp transitions at class boundaries and subjective un- certainties in decision making. In this study, the decision trees method was employed as a classification tool for estimating ground rippability, using the four rock mass parameters of weathering degree, uni axial compressive strength, joint spacing and bedding spacing. The results were compared with those of the widely used method of diggability index rating (DIR) and a fuzzy-based DIR. The problem of subjective uncertainty was resolved in the proposed decision trees method, as the tree rules were derived automatically from the training data sets. Compared with fuzzy-DIR, a decision tree elassifier is sim- per, needs less computational time and is more appreciable by industry practitioners. Nevertheless, the proposed method is not flawless and, similar to both the conventional and fuzzy DIR methods, yielded ‘poor estimations in particular instances. The limitations of the three methods are discussed. Key words: Ground rippability, Diggability Index Rating, Fuzzy set theory, Decision trees, Rock mass classification 2015 Transactions of the Society for Mining, Metallurgy & Exploration, Vol. 338 {in press) Introduction Determining the ease of ground excavation, called rippa- bility, diggability or excavatibility, sa critical decision-mak- ing slep in selecting both the overburden removal method and equipment in surface mining and geotechnical engi- neering projects. Surface preparation is @ primary phase in many geotechnical-geomechanical engineering projects and is performed in two major ways: (1) by drilling and blasting, and (2) by digging and ripping. In urban and environmet tally sensitive sites, use of the drilling and blasting method is highly restricted, leaving the mechanical digging and ripping ‘method as the only option. The proper selection and sizing of the machinery, such as dozer-rippers, is essential to project success, Economie, environmental and ground conditions are the most important constraints in equipment selection and sizing. From the technical point of view, the ground or rock mass properties are the dominant parameters in equipment selection. Many researchers have addressed the issue, using Papor number TP-15.016, Original manuscript submitted Avril 2016, Revised manuser 2018. Discussion ofthis peer different methods, Basarit and Karpuz (2004) classified the ‘ground rippability estimation methods into wo groups of di- rect and indirect methods. When proper equipment for trial tests is available, ripping tests can be implemented directly. Alternatively, if direct ripping runs cannot be conducted, the indirect methods are used to estimate ground rippability from the rack mass geomechanical properties. The major methods of indirect rippability estimations are seismic, graphical and ranking/classification, Seismic methods. Estimation of rippability from seismic wave velocity is among the earliest attempls in this field. The underlying theory of the method is that p-waves travel faster in denser rocks than in loose materials, A measurement of p- wave velocity in an area can classify the rock mass to differ cat levels of rippability. Bailey (1975) used the seismic wave velocity in a rock mass to predict ground rippability. MeCann and Fenning (1995) examined the effectiveness of using seis accepted for publication October ‘@wed and approved papers invited and must bo suom ted to SME Publications Dept. prior to Sep. 30, 2016. Copyright 2078, Society for Mining, Metallurgy and Exploration inc 201BTRANSACTIONS + Vol. 338 n2 ‘SOCIETY FOR MINING, METALLURGY & EXPLORATION 2201S TRANSACTIONS OF THE SOCIETY FOR MINING, METALLURGY & EXPLORATION + 112 mic velocity measurements for predicting rippability. More recently, Bin Mohamad et al. (2010) examined the reliability, of using seismic waves to estimate the rippability of a weath- cred sedimentary area by comparing the results with actual rippability trials, However, interpretation of the seismic test results is a key element in the reliability of the output, requir- ing the employment of highly skilled people to interoret the tests (Iphar and Goktan 2006). Moreover, predicting rippabil- ity based on only one parameter, the wave speed, might not, be a comprehensive indicator, Graphical methods. Developed in the early 1970s, the ‘graphical methods mainly consider the two parameters of dis- continuity spacing and rock strength. One of the first graphi- cal methods was proposed by Franklin, Broch and Walton (1971). The Franklin method was later modified by Pettifer and Fookes (1994), who revised the original method using the results of some 100 new case studies. Weaver (1975) proposed a rating chart based on geological conditions. This ‘was among the first works that emphasized the importance of geological features along with wave speeds as the indicators of a rock mass rippability classification. Graphical methods, however, should be regularly updated with the development and deployment of new digging and ripping equipment (Ba- sarir and Karpuz, 2004), Ranking/classification methods. In ranking/classifica- tion methods, several variables, such as geotechnical proper- ties like uniaxial comprehensive strength, that influence rock mass rippability are measured, after which a classification system based on the individual variable’s values is proposed, categorizing rock masses into different classes and ranks, Singh et al. (1986) proposed a rippability classifier chart based on rock tensile strength, abrasiveness, seismic veloci- ty, weathering, and discontinuity spacing, Hadjigeorgiou and Scoble (1990) proposed a rock mass classification system based on rock strength, block size, weathering, and relative ‘ground structure. Goktan and Eskikaya (1991) found that the rock uniaxial compressive strength and coefficient of relative rock mass weakness correlate well with rippability in sedi- mentary rocks of some surface lignite mines. Macgregor et al, (1994) used a database of detailed ripping and geological data, gathered from highway and mine sites, to determine the influence of different factors on the ripping productivity of bulldozers, The dominant factors affecting productivity were the uniaxial compressive strength of the rock, the degree of ‘weathering, seismic velocity, joint roughness, strength and spacing of the joints, and bedding in the unripped rock. More recently, Basarir and Karpuz (2004) developed a rippability classification system for marls in several lignite mines based ‘on direct ripping runs, a specific energy concept, and indirect, rippability assessment methods. The proposed system includ- ced rock parameters such as uniaxial compressive strength, seismie p-wave velocity, discontinuity spacing, and Schmidt hammer hardness value, Tsiambaos and Saroglou (2010) pro- posed a classification method to assess the ease of excavation of sedimentary and metamorphic rock masses based on the geological strength index (GSI) and the point load strength, of the intact rock, Limitations. Despite their widespread application, clas- sification methods involve a number of subjective uncertain- ties stemming from: (1) qualitative (linguistic) criteria; (2) sharp class boundaries; (3) fixed rating (or weight) scales; and (4) variable input reliability (Aydin, 2004), To address SOCIETY FOR MINING, METALLURGY & EXPLORATION some of the limitations of the classification approach, sev- eral researchers tried to deal with the subjective uncertainty in decision making by applying artificial intelligence meth- ods. Probably forthe first time, Nguyen andAshworth (1985) examined the applicability of Fuzzy set theory in rock mass classification, Later, Habibagahi and Katebi (1996) proposed a fuzzy classification system based on the conventional meth- od of rock mass rating. Iphar and Goktan (2006) deployed fuzzy set theory inthe diggability index rating method (DIR) to address the limitation of sharp class transitions in the original method, Liu and Chen (2007) proposed a classifi- cation model that uses an analytic hierarchy process (AHP) and the fuzzy delphi method (FDM) to assess slope rock mass quality. Hamidi et al. (2010) used fuzzy set theory to classify rock masses based on excavatability. Chebreghani, Alipour and Eskandarzade (2011) employed artificial neural, networks (ANS) to estimate rock mass excavatability in limestone quarry. They used uniaxial compressive strength, tensile strength, and discontinuities spacing of rocks as the input parameters. Sacidi, Torabi and Ataci (2014) used ANNs, multiple regressions and fuzzy set theory to estimate the basie rock mass diggability index. Despite their relatively extensive application in rippability estimation, ANNs need substantial amounts of reliable data for model training. Fur- thermore, optimal model parameters are selected based on trial and error or the model builder's experience. Thus, the problem of subjective uncertainty arises again. Proposed by Zadeh (1965), fuzzy set theory is an exten sion of classical set theory in which elements of a set have different degrees of memberships. In classical set theory, ev- ery element either belongs or does not belong to a set, that is, a0 or 1 classification, but in fuzzy set theory all elements are assigned with a membership function that takes on real values in the interval [0, 1). This characteristic of fuzzy set theory is highly useful in dealing with the problem of sharp transitions at the boundaries of the conventional classifica- tion methods. However, the application of fuzzy set theory in ground rippability classificatc ANNs. For instance, the selection of the membership fune- tions, that is, the fuzzy if-then rules, and defuzzification methods depends on the opinion and experience of the model ‘builder, as there are several alternatives for each variable. Similar to ANNS, in fuzzy sets, the selection of the system's topology and optimal parameters is more art than science (Wilamowski, 2007). Thus, for a similar dataset, different analyzers might get different results Classification by decision trees. To address these limi- tations, we used a decision trees classification technique to estimate rock mass rippability for several sets of field and laboratory data obtained from the literature (Bozdag, 1988), Decision trees have been widely used in different geotechni- cal, civil and mining sectors as a versatile, fast and inexpen- sive decision-making aid, Shahriar, Sharifzadeh and Hamidi (2008) used decision trees to perform a geotechnical risk assessment for tunnel-boring-machine selection in difficult ground conditions. Glastonbury and Fell (2008) used deci- sion tree analysis to propose a framework for the assessment of the post-failure velocity of large natural rock slopes. Kaur and Pulugurta (2008) proposed a fuzzy decision tree mode! for pavement treatment prediction. Hiring et al. (2012) pro- posed a decision tree method for the spatial disaggregation ‘of complex soil map units. Lee and Park (2013) employed a decision tree model for ground subsidence hazard mapping 13 Vol. 338 + 2015TRANSACTIONS. 114 + 2018 TRANSACTIONS OFTHE SOCIETY FOR MINING, METALLURGY & EXPLORATION Table 1 — Rock mass samples gathered by Bozdag (1988) Site Weathering degree Uniaxial compressive strength (MPa) Bedding spacing (rm) 7 Slightly on 7 2 Slightly 103.8 os 3 Slightly 103.8 3 o7 4 Slightly oe 3 07 5 Slightly om 15 15 ‘ Slightly 29 os os 7 Slightly 424 " 128 8 Medium 227 os 07 9 Slightly a8 125 15 10 Slightly 314 035 03 ” Slightly ar 18 16 2 Slightly 385 15 1 3 Slightly 188 07 o7 4 Unweathered 308 or o7 6 Slightly 314 2 16 6 Medium 15 os os w Slightly 218 1 1 1% Slightly 447 12 1 19 Slightly 383 18 1 20 Unweathered waa 2 2 2 Slighty 58 os 06 2 Slightly ne os 04 23 Slightly 22 os 0.45 24 Slightly 39 2 og 26 Slightly v8 os 1 26 Slightly Ww 08 06 27 Slightly 203 os 1 28 Unweathered 52 2 29 Unweathered 26 2 16 30 Unweathered 20 13 1 2 Slightly 198 os os 2 Unweathered 60 1 1 3 Medium 30 08 1 Pa Slightly 24 1 18 38 Unweathered 263 2 16 6 Slightly 10s 12 os. 37 Medium 36 08 0.45 38 Medium 25 035 os 39 Stightly 2 0.35 04s 40 Medium 308 or 0.28 a Medium 162. 0.35 03 a Medium 36 04 oa 43 Slightly 268 055 058 “ Medium 36 088 os 45 Slightly 188 0.25 0.25 46 Medium 15 03 os 2015TRANSACTIONS + Vol. 338 4 ‘SOCIETY FOR MINING, METALLURGY & EXPLORATION 201S TRANSACTIONS OFTHE SOCIETY FOR MINING, METALLURGY & EXPLORATION +115 near abandoned underground coal mines. Emmer andVilimek (2014) used decision trees to assess the susceptibility of gla- ial lakes to outburst floods in the Peruvian region of the Cor- dillera Blanca. In Dindarloo and Siami-Irdemoosa (2015), the use of decision trees to propose a novel shallow tunnel classification scheme based on maximum surface settlement in urban soft grounds is presented. Although there are numerous applications of decision trees in the geotechnical and geological engineering fields, this is the first time that the method is applied to the problem of rock mass rippability prediction, In decision tress, clas- sification rules are extracted directly from the historical data, (training data sets) by decision rules. The classes of unknown, rock masses (test data sets) are identified automatically based. ‘on the derived tree rules, without subjective uncertainties, as the user does not interfere with rules parameters, Thus, the limitation of subjective uncertainties is resolved in this study by the application of decision tree algorithms. To validate the method, the results of analysis with the decision tree method ‘were compared with those of both the original DIR method anda fuzzy-based DIR method. In addition, the generalizabil- ity of the model was confirmed by estimating the rippability class of an “unseen” data set, The data set was obtained from the literature and is considerably different from the training data sct that was used in model building in the sense that the two data sets were gathered at different locations and mines, and hence have different geomechanical properties, Methods and materials Stille and Palmstrm (2003) listed the main requirements for a true classification system capable of solving rock-engi- neering problems as: (1) The reliability of the classes to as- sess the given rock-engineering problem must be estimated. (2) The classes must be exhaustive (every object belongs to a class) and mutually exclusive (no object belongs to more than ‘one class). (3) The principles of division (rules) governing as- signment into the classes must be based on suitable indicators (e.g, ground parameters, etc.), and must include the possibil- ity of being updated during construction using the experience gained. (4) These rules must also be so flexible that additional indicators can be incorporated. (5) The uncertainties, or the Table 2 — Diggability index rating (DIR) rock mass parameters and ratings (Scoble and Muftuoglu, 1984). Parameter Scores (ihWeathering Completely Highly ‘Moderately Slightly Unweathered Rating ° 5 6 20 25 {2| Uniaxial compressive strengthiMPe) <20 20-40 40-60 60-100 > 100 Rating ° 0 6 20 28 (3) Joint spacing <03 03-06 0618 182 22 Rating 5 6 30 48 50 (4) Bedding spacing <04 01-03 208 ots 218 Rating ° 5 10 20 20 Table 3 — DIR diggability classification and equipment suggestion (Scoble and Muftuoglu, 1984), Ease of digging index _—_—Excavation method Equipment type 1. Ripping (A) Ripperscraper, Caterpillar D8 Very easy <40 2. Dragline cast (8) Dragline > 5 m, Lima 2400 3. Shovel digging (c} Rope shovel » 3 m, Ruston Bucyrus 71 RB 1. Ripping (A) Ripper-seraper, Caterpillar D8 Easy 40-50 2, Dragline cast (8) Dragline > 8m, Mation 195 3. Shovel digging (Cl Rope shovel > 3m, Ruston Bucyrus 150 RB Moderately 50-60 1. Ripping {A) Ripper-shovel, Caterpillar D8 difficult 2.Shovel digging (8) Hydraulic shovel > 3 m, Caterpillar 245, i ' 1. Ripping (A) Rippershovel, Caterpillar D10 Dieu 50-702. Shovel digging (8) Hydraulic shovel > 3 m, Cat. 245 or O&K RH 40 Very afficutt 70-95 Shovel digging (A) Hydraulie shovel > 3 m®, Cat, 245 or ORK RH 40 Extremely (A) Hydraulic shovels > 7 m8, Demag H111, Poclain 1000CK, difficult es t00 Shovel digging P&H 1200, O&K RH 75 Marginal without 100 Shovel digging (A) Hydraulic shovels > 7 m?, Demag Hit, H241, O&K RH 300 blasting SOCIETY FOR MINING, METALLURGY & EXPLORATION "5 Vol. 338 + 2015 TRANSACTIONS. 16 + 2015 TRANSACTIONS OFTHE SOCIETY FOR MINING, METALLURGY & EXPLORATION Table 4 — Comparison of field data (Bozdag, 1988), fuzzy DIR lIphar and Goktan, 2006) and decision trees. Fuzzy ease of Fuzzy ease of Ease sea Ease of eas sin aggg—88U ocontne te digging SPE Cron ne (field data) emorne {field data) omens 1 Very difficult vo Very difficult 24 Ditficule Difficult (0.82) Difficult exvemely Moderately 2 veya EMH ery ginauk 25 amour Meseate,nu arene Modertaly Moder 2 veryaiiout OMY ey iscan 26 MOBY say iggay Moder 4 Very difficult Extremely Very difficult 2 Very difficult Moderately Very difficult ery cites difficult (0.93) ‘ery aitfenl ‘ery eificul difficult (0.57) fery dificul Marina wit Marina it 5 Verydifeuk—“eutbasing — Vorysiteuk ae utamety NGS ramet oan io Marina wt Marginal 6 Difficult dint Difficult 29 Merit wie out blasting without Very ist Ver cies 7 veryaseun MEU Vey gti 20 ryt YOSE very gicu Moerstay Modesty Modesty Ey Madera 8 difficult Easy (0.62) difficult o difficult (0.78) difficult ._Bremay eu sven vere stu 9 vewyancut EAM, veya a2 ryt MEME oy git we yo cosy 8 vatican Hay ic Margi ihe Magia Marin wit ,Vediiosk 1 Merial sling whet st 34 Vegi YEE oy pe a io Maria vite Marginal , erate an Mari wine Marie ite Marina 12 veryaiteat YOM ry ainea a5 Matin wh easing” without a ina 13 Diout anyon) ious Mose ee art Marinate Marginal 18 Marginal with- ‘out blasting without blast 38 Very easy Very easy Very easy a ia Very ita esey 17 Weryaeun EE ry git 40 cay bie Very teat Very ey 8 vervaieut Veryaiieuk 81 eyes Very oss ” (ost) ” ny easy 0 ryessy Ver situ Very ony ” (0.65) ” ny easy 0 ny easy Marinate Magia Modersey tbiaing ot vu sie it sit 2 Easy Easy (0.79) Easy 44 Very easy ‘Very easy (1) Very easy, 22 Eay aye) Eaty #5 yomny—Veyeanyit)—_Veryenty 2 kay ays) any Heyy, Veyenuyit)—_Veryesty Z2015TRANSACTIONS + Vol. 338 16 ‘SOCIETY FOR MINING, METALLURGY & EXPLORATION 201S TRANSACTIONS OF THE SOCIETY FOR MINING, METALLURGY & EXPLORATION + 117 quality, of the indicators must be established so that the prob- ability of misclassification can be estimated, (6) The useful system should be practical and robust, and give an economic and safe design, There is no rock mass classification system that meets all these requirements, but keeping them in mind, ‘we investigated the applicability of decision trees to the spe- cific problem of rock mass rippability classification, using 46 samples of raw geotechnical field data (Bozdag, 1988). Table 1 lists the weathering degree, uniaxial compressive strength (UCS), joint spacing and bedding spacing of the 46 samples, which were gathered from a lignite mine, These four param- eters are used in the DIR classification (Tables 2 and 3). The DIR arose from a series of experimental rippability evalua- tions in British surface coal mines (Scoble and Muftuoglu, 1984), Table 2 classifies a rock mass into five categories based on the scores assigned to the four parameters. Weather- ing is the only linguistic term in Table 2, and itis quantified by a score between 0 for completely weathered and and 25 for completely unweathered rock masses, The summations of the scores for the four rock mass parameters arc used as & decision index in Table 3, where rock masses are assigned to different categories of ease of diggability and corresponding, excavation method and appropriate equipment type. Methodology Classification and regression trees are machine-leaming methods for constructing prediction models from data, Old- en, Lawler and PofT (2008) listed the advantages of decision trees as: (1) inherently nonparametric and, therefore, not af fected by heteroscedastcity that affects parametric proce- dures; 2) invariant to monotonie transformations ofthe data, thus eliminating the need for data transformations; (3) able to handle mixed data, including categorical, interval, and con- tinuous variables; (4) able to deal with missing variables; (5) not affected by outliers (outliers are isolated into a node, and have only a minimal effect on splitting); (6) able to detect and reveal interactions in the data set; (7) able to effectively deal with higher dimensionality (that i, it can identify a reduced sot of important variables from a large number of submitted variables); and (8) relatively simple to interpret graphically. Logie. In decision trees, the models are obtained by re- cursively partitioning the data space and fitting a simple prediction model within each partition. As a result, the par- titioning can be represented graphically as a decision tree, Regression trees are for dependent variables that take con- tinuous or ordered discrete values, with prediction error typically measured by the squared difference between the observed and predicted values (Loh, 2011). Classification trees are for dependent variables that take linguistic values, such as rock mass diggability, In decision tree modeling, an. ‘empirical trec represents a segmentation of the data that is created by applying a series of simple rules. These models ‘generate a set of rules that can be used for prediction through, the repetitive process of splitting (Tso and Yao, 2007). The decision tree approach is built upon the implicit assumption that the relationship between features and target objects is cither linear or nonlinear. Therefore, it can be used to handle complex nonlinear relationships. In'a decision tree, features that carry maximum information are automatically selected for classification/regression and the remaining features are rejected, thereby raising computational efficiency as well as eliminating subjective uncertainty, The construction of a tree is also based on binary recursive partitioning, which is an SOCIETY FOR MINING, METALLURGY & EXPLORATION iterative process that splits the data into partitions. Initially, all the training samples are used to determine the structure of the tree, The algorithm then breaks the data using every pos- sible binary split and selects the split that partitions the data into two parts such that it minimizes the sum of the squared deviations from the mean in the separate parts. The splitting process is then applied to each of the new branches. The pro- cess continues until each node reaches a user-specified mini- ‘mum node size (Xu et al, 2005). The term “binary” implies that each group of observations, represented by a node in a decision tee, is split into two child nodes, a process through which the original node becomes a parent node. The term "re- cursive” refers to the fact that the binary partitioning process can be applied repetitively. Thus, each parent node can give rise to two child nodes and, in turn, each of these child nodes may themselves be split, forming additional children, The ‘term “partitioning” refers to the fact that the data set is split into sections or partitioned (Olden, Lawler and Poff, 2008), Algorithm, C4.5 is one of the most-used decision tree algorithms in classification tasks. Given a set S of cases, CAS first grows an initial tree using the divide-and-conquer algorithm as follows: (1) If all the cases in S belong to the same class or $ is small, the tree is a leaf labeled with the most frequent class in S. (2) Otherwise, choose a test based on a single attribute with two or more outcomes. Make this test the root of the tree with one branch for each outcome of the test, partition S into corresponding subsets Sy, Sy. according to the outcome for each case, and apply the same procedure recursively to each subset. (Wu et al, 2008). C4.5 consists of tree-growth and tree-pruning steps. In the former, tree growth begins from a node, which is then split by select- ing the attribute that best classifies a set of examples on the basis of an attribute selection measure. ‘The attribute selection measure uses the concept of en- tropy, which is defined as the degree of disorder. Thus, a tree grows by selecting an attribute with the smallest Entropy. At anode N, Entropy is calculated by: Entropy (1) = ~¥ , (GIN) log, (CN) 0 where p(GIN) is the relative frequency of V. Of the & attri- ‘utes of N; the entropy for selecting attribute A is given by: Entropy (N) = YF) (NAVIN) * Entropy (N;) (2) InfoGain is a gain from differences between the Entropy of the original node and the Entropy of the newly split nodes. The equation is as follows: InfoGain (A) = Entropy (N) ~EntropyyN) 3) ‘Thus, C4.5 selects an attribute with the smallest Entropy or biggest InfoGain. The C4.5 algorithm was initially devel- oped by Quinlan (1993). For more detailed information on the theory of the algorithm, see Quinlan (1993, 1996) Implementation. The structure of the tree constructed for this study, along with its associated rules are illustrated in Fig. 1, The results of classification of the 46 samples using fuzzy DIR and decision trees are compared against field labo- ratory measurements in Table 4. Fast and easy application of the derived tree is demonstrated in the following example showing the estimation of the diggability of sample No. 13 (Table I and Fig. 1). The joint spacing of sample No. 13 is 17 Vol. 338 + 2015TRANSACTIONS. 118 + 201STRANSACTIONS OFTHE SOCIETY FOR MINING, METALLURGY & EXPLORATION eae — aa pore) geben ees (Sei mmae at Figure 1 — Decision tree rules, showing the case of sample No. 13 of fable 1. 0.7 m, which directs the tree to the left branch, The next node selects the bedding spacing as the next decision-making pa- rameter for tree spliting. A value of 0.7 m, from Table 1 directs the tree to the right-hand-side branch, as illustrated in Fig. 1. At this point a UCS of 18.9 MPa guides the tree to the right-hand side, which is a bedding spacing decision node, Since the bedding spacing is larger than the critical value of 0.65 m, the category of “difficult” is estimated for this sample. Table 4 shows “difficult” and “easy” categories for the actual measurement and fuzzy-based DIR estimation, respectively. Table 4 demonstrates the higher accuracy of the decision tree algorithm than the fuzzy-based model when the actual rippabilities are set as reference. Model generalizability. The most important characteris- tic of prediction/estimation models is their generalizability Generalizability is defined as the capability of a mathemati- cal model to predict the behavior of “unseen” data. Unseen Table §— Tost of model gnoraliabiiy, sing th 10“unavor dat ats rom ho uly of Scoble avd Moga (a) (6S unin compresive stent) esting orn das A SUM aoe Mainalthourning Moria without sing 8 Moderisy 4322 Masnithou ting Magra iho ting © Unventond 784-22, 12—_‘Magnawthourlastng Ver aie © Sluhly 668-2 Moinahwthoulstng Moria without sing ; lime F Signy m2 AB Magrath ting Ver teu GS Siohy 688118 Marinat thou isting Ver atu si ee Ver aes Ver ite J Unenteed 628 «2 —_‘Magnwihou bling Magna! wihou lasting Z2015TRANSACTIONS + Vol. 338 ne ‘SOCIETY FOR MINING, METALLURGY & EXPLORATION 20:STRANSACTIONS OFTHE SOCIETY FOR MINING, METALLURGY & EXPLORATION 18, data are data not used in the process of model building or training, also called the testing data set. The examination of ‘model generalizability is performed after the steps of mode! building, verification and validation, Loosely speaking, the decision tree model was verified and validated using the in- spection of the structure of the tree in Fig, 1 to assure it is oth geomechanically logical and practical (verification). In the next step, the model results were compared against other available methods, such as DIR, to assure it yields reasonable ‘output (validation). To examine the modei generalizability, the rippabilities of 10 different mines were obtained from the seminal work of Scoble and Muftuoglu (1984), who devel- ‘oped the original DIR method, The weathering degree, UCS joint spacing and bedding spacing of the 10 mines, along ‘with both the DIR and decision tree classifications are pre- sented in Table 5. In six (A, B, D, H, 1, and J) out of the 10 sites, the estimations from both models are the same, In the remaining cases, the estimations of the decision trees are one for bwo classes below those of the DIR method, for example, the decision trecs estimation gave a “difficult” classification while the DIR estimation gave a “very difficult” one. In gen- eral, the DIR method yields more conservative estimates for the rippabilities than the decision tree estimates (see C, F, and G in Table 5). As the decision tree was constructed solely based on the limited raw data of 46 cases in Table 1, the estimation accuracy in Table 5 of 60 percent is accept- able. Particularly, locations, mine types and rock masses, and hence, the rock mass parameters of the testing (unseen) data, sel are very different from those of the training (seen) data set (compare Tables | and 5). Furthermore, for the sake of sim plicity and applicability, both methods employ only four pa- rameters of a rock mass to estimate its rippability class, [fone considers other parameters, such as porosity, tensile strength, abrasivity, anisotropy and water conditions, the two ms ‘ods’ estimations might be more consistent with each other, In other words, the decision trec’s classification accuracy, when the original DIR is set as the reference, may be higher than 60 percent, The estimation accuracy of the model for the ‘seen’ training dataset is 100 percent (Table 4. Discussion In order to compare the original DIR method with the pro- posed decision tree method, the specifications of two generic rock masses are summarized in Table 6 and in Table 7. In Table 6, the original DIR model categorized the two different rock masses in the same class of “difficult” diggability (see Table 3), while the decision tree method was able to discern the slight difference between the two (using data in Table 6, follow Fig. 1). Different ripping equipment would therefore be recommended, that is, a more appropriate decision is ob- tained than that suggested by the original DIR method, which was the same equipment for both cases. In Table 7, the con- ventional original DIR method reached the same rating (dig- gability index = 50) and ripper recommendation for A and 1B, However, from the point of view of an experienced field engineer, itis expected that excavating Ground A would be casier than excavating Ground B. On the other hand, follow- ing the decision tree rules of Fig. 1, the diggability of Ground A would be classified as “very easy,” as confirmed by prac- titioners, and Ground B would be classified as “moderately difficult,” as would be expected by field engineers. The DIR scores for A and B in Table 7 are both $0, A score, or index, of 50 is at the boundary of the “easy” and “moderately difficult” rippability classes (see Table 3), which imposes the problem of subjective uncertainty in decision making with regard to equipment selection ‘Accloser inspection of Table 4 reveals more deficiencies, in the fuzzy-based DIR method. For instance, in the fuzzy approach, the two considerably different Sites 21 and 40 (s Table 1) are treated equally as grounds that are “easy” to rip. In contrast, both the conventional original DIR method and the proposed decision tree method confirm that Site 21 is in- deed much easier to rip than Site 40, Furthermore, the mem- bership degrees for several fuzzy-classified sites, for exam- ple, Sites 13, 17, 22 and 33, in Table 4 are around 0.5, Again, the problem of subjective uncertainty in decision making arises in these cases. It is not obvious, for instance, whether Site 13, which has 2 membership degree of 0.51, should be considered “very easy,” “easy” or “moderately difficult.” In Table 6 — Specifications of two generic rock masses and comparison between DIR and decision trees: Example 1. Generic Weathering —_Uniaxialcom- Joint spac edding yp Decision rock degree pressive strength ing (m) spacing OF iRclassiication tree clase mass (seore) (ma) (score) (score) (m) (score) @ sifieation Unweath- 5 14 05 Moderately ‘red (25), co} (30) (10) & Piffieat dificult Slohtly 59 078 0.08 2 6 Diticut Difficult (20) 19) (30) o Table 7 — Specifications of two generic rock masses and comparison between DIR and decision trees: Example 2 Generic Weathering ——_—Uniaxlalcom- Joint spac jedding yg Decision rock degree pressive strength ing (m) spacing OIF iRclassification tree mass (score) (mPa) (score) (score) (m) (score) 8 sification A Moderately 5 0.38 0.65 go Esvormoder—y, (18) i} (13) (20) ately difficult ery easy 2 Moderately 35 0.35 029 so -EasVormoder Moderately (19) (5) 19) 6 ately diffeutt difficult SOCIETY FOR MINING, METALLURGY & EXPLORATION ng Vol. 228 + 2015TRANSACTIONS: 120 + 201STRANSACTIONS OFTHE SOCIETY FOR MINING, METALLURGY & EXPLORATION ‘any case, its class willbe different from those of the conven- tional original DIR method and the decision tree method, as both classify Site 13 as “difficult.” Indeed, with a DIR index ‘of 70, this site camnot be treated as “easy,” as is suggested by the fuzzy approach, It should be emphasized that despite the mentioned advan- tages, the proposed decision tree method is neither compre- hensive nor flawless, The main reason is thatthe tre rules had ‘been derived from a relatively small data sample in this study, which does not cover all ranges of the input variables. For instance, in Table 1, there is no site with “highly” or “com- pletely weathered” weathering degrees. Consequently, inthe ‘cision tree of Fig. 1, the weathering degree nodes split only to a maximum of three classes: “unweathered,” “slightly” and “moderately.” This model shortcoming ean be corrected by employing a more comprehensive dataset. Another major problem of the three models is demonstrated in the cases of Sites 10 and 43 of Table 1. Both sites have weathering de- sees, UCS, joint spacing and bedding spacing parameters ‘within the same ranges, that is, both are slightly weathere have UCS in the 20 to 40 range, joint spacing in the 0.3 t0 0.6 range and bedding spacing inthe 0.3 o 0.6 range. Thus, both sites should be classified in the same group in terms of ease of rippability. However, all ofthe three modelsDIR, fuzzy DIR and decision tree-classify Site 10 as “easy” or “very easy” and Site 43 as “moderately difficult.” Therefore, though each ‘of the three methods have some merits, none can be a univer- sal solution tothe problem of rippability estimation, and fur- ther research is required to obtain @ more comprehensive and efficient model of rippability classification. To achieve this ‘goal, the key issue will be the preparation of a sufficient! large and reliable database that encompasses a wide range of input parameters. It is worthwhile to note that for moderately weathered ‘grounds with joint spacing < 0.65 m, there is a sharp tra sition between the “very easy” and “moderately difficult classes at 19.5 MPa (see Fig. 1) It is expected that a rock mass with higher UCS will be more difficult to rip at this point. Ths is due tothe fact thatthe decision tree learned its rules ffom the data set of Table 1. Sites 10 and 43 in Table 1 ate both moderately weathered with joint spacing < 0.65 1m, but although the UCS of Site 43 (26.8 MPa) is less than that of Site 10 (31.4 MPa), it was observed to be “moderately 4ifficult” while Site 10 was “easy.” The same results were obtained from both the decision tree model and fuzzy DIR ‘model in these cases. Conclusions ‘The applicability of decision tree classification in cat egorizing the rippability of rock masses was investigated. Adccision tree model was able to classify 46 rock masses ‘comparably to the field measurements. The main advantage of the proposed decision tree method is that it resolves the major limitation of subjective uncertainties suffered by the conventional classification approaches. Particularly, decision, ‘ees are more versatile than ANNs and fuzzy set theory in re- solving the problem of subjective uncertainty. This is because the applications of ANNs and fuzzy sets need the subjective determination of several parameters during model building and training, Furthermore, a decision tree is simpler, needs less computational time and is more appreciable by industry practitioners. One drawback of decision trees, like ANS, is, their high dependency on the quality and quantity of the raw ata, 201BTRANSACTIONS + Vol. 3238 There is no flawless model that can be used universally for all kinds of grounds. All available models, including the pro- posed one in this study, have some disadvantages that, occa- sionally, yield erroncous estimations. Thus, prior knowledge, experience, expert judgment and direct ripping/digging tests, if both available and feasible, should not be eliminated from the process of ground rippability classification and equip- ‘ment selection. In other words, indirect methods, such as de- cision tree methods, should not be considered as stand-alone decision-making tools for ground rippability classification, References Pytin, A, 2008, “Pay set approaches to dossifi Bates 0, 1978, "focetypesandseismieveleisersusrppsbily” Proceed ins of th Highasy Goolay Symnasim, rp. 135 Ignite mines” Enginearng Geology, 1, No.3, pp. 02318 Bn Mahaenad, £7, Saud, A Nost, MIM, Ma, Is, MLB. ang Mask, AN, 2010 Ercaataty assessment weathered sedentary rcknsss ng ream welsh Bozdag, T1988, Mie, Thess, Midale Eas: echnical Unversity, Avera CChenreghan, SH, Alvour A, ana Eskanderase, M201, "Rack mass oxcavar- sty estimation sng ete neural netnire Jounal ofthe Geena! Socityof nd, Vo 78, No. 2.9 27 Dinko Sf, and Sim aernagsa, E2018, “Maximum surtce setlement based clessiiaton of stalow tunnels in sok ground” Tunneling ara Ur dberarouna Space fcrnology, ol. 48, pp. 320-327 find Earth System Sconces Vol 18. No.9 pp. 2481-2479, Fann 2. toch, = araWion, 1971. "Loggingthe mechoncalenarcter of 2k” inst Min. Metal pp AS, Custrbury J and Fel, R 2008" decision anahss framework tor rent of lily postfalure vl ef vansliinal ra eomoound net ‘ope landelaes” Canacln Cootectnies!Joural Va 45 No.2 pp. 229950, Ccoktan AM, and Eskkae, 5, 198, 3 issing machine in torms of rox mass sropertes” Cl Erg SA, Val 31, NOT, Haibagon.G, and Kaeo 8, 1998, “Roskrmasseassestion usa ny ste Irinan Jours! of Seance and Teehneeay, rensscton B Tecnology, Vl 20, N9. 9.99. 2802 ‘saignsr@ou, 4 Scoble, Mi), 1990, “Groune cmarererzation fr as Symmeestum on Mine Panning and Equpment Selection, Cal ng elssfetion systems: on ius Tock ass excavalty naexe Rack Mecham ana Roce Engi #2, No 3 on 225-250, sting, Bits E, Oserstote S, Kosehtzk T, and Se ‘iraggregeton ot eorex sol man.ntsAcecisons ivan forest sale” Gaoderms Vol 18986, op. 37-17 Inhar i ana kan Rik, 2008 Trix Rating Metrod fr surace Journal! Rack Mechanics ana Mining Seonens Vol 82, N22, 9, 750-296, Itee'nd tate ronression methade for parernent ialmen| prediction 3.2012, "Spatial ‘ace speroachin WSEAS ansactons on intermation Sconce ad APpictions, Va 5, No 6, pp.970-000 fe, 81 and Park, |, 2013, hpeaton of deesin Wee model fo the aun sbundonesundergfoundeaclmines’ soul Liu Yet. ana Chen, CS, 2007 A new apsroach fr apaiation o clastieaon on rock lope stably assessment” Engineering Ge 9, No, 2 po 120142 Leh, WY. 261, “Clssifeation and regression trees” Wey Ines inwnis: Data Ms YacgtegorF, Fell R, Mostn, Gy Ho frtmaton rock ippoe iy Ouart Construction, pe 335263, Nauyen, WU. ana Ashworth, ., 198, “Rock mass clasifcton by furry a Proceedings af the Symeasium on Rack Mechanes, Vo 2, pp. 937-548, (Olten JD, Lawley, J), and P04, ALL, 2008, "acre leaning methods ‘tnt eat: A pier ‘or eealogsis* Quarteny Review of Bcioay Vel 2, No.2. po. 17198, Pair, 6S, and Foowes, 2S, 7884," revision of the graphical method for as 120 ‘SOCIETY FOR MINING, METALLURGY & EXPLORATION 201STRANSACTIONS OF THE SOCIETY FOR MINING, METALLURGY & EXPLORATION +121 sessing the excavilty of vak” Quarter Joumal of Engineering Gato, Vel 27 pp. 115-68 Cuinkn, J 1882, C45: Programs for Mecrine Learning, Margin Klann Pobishe's ‘iran, J. 196, "Bagging, boosting, and C45" Proceedings of he National “Ceonironce on Aten ineligonee, el. 1 ppv 725-730, See 0, Toad 5, and Ais) Ml, 2014, “Preston ofthe rock mase de sooly rdxby using fre cstoningnased ANN and mute egtesion ‘methods Rock Mechancs nd oct Engineering Val &7 No.2 pp. 7°P732, Seeble, Wi. and Matuogl, VW, 1983, “Bertin of o iggobly ex fo sae mine equipment selection” Min St fednnol Wa 1, so, 205022, ‘Shahar K Shared, Mand Hamid J. 2988, "Geotechncalriak assess. ment based approscn or 0k TEM selection eut gound conden Tuaelng ana Underground Space Techalagy, Vel 23, No.3, pp 318225 Singh IN, Deroy 2 Fgretl Lana Paha A, 1886, Assesamentof ground "ppasitynopercast mining operations” Min Mag Unw Notingham, Vl 38,2158. Stile, He ind PakosrOr A, 2003, “Chssteaton ae a 00» rock ena neerna Tunreling ond Underrcuna Space Technology, Vol 18, No pp. 221285, SOCIETY FOR MINING, METALLURGY & EXPLORATION “Tsamnases,G,sndSaroglu 2010, “Bxenstaltyessessmentotroccmssses “sing he Geological Song Insex(GSI" Buea of Engineering Geocony Sind the Enevonment Vel 89, No. pp. 1227 ‘Teo ks srdYou KW, 2007, "Proseingelctcty eneray consumption ‘s comparison of regression analysis, deison tee and neural networks” Eorgy Vo 2, No.9, pp. 1781-178, ‘wenver JM, 1875, Geohocal scons gr ieantin ine assessment of nna Tyr civ Eng Si, Vl 17 No. 12, po. 212-37, ‘amok, 8M. 2007 "Neural Networks and Fuzry Sts fr Noninear Apple ‘ators Tith iteration Canference or iteligest Engineering Systems, Susopest, Hungry, wu. X. Kumar Ross, G.., Ghosh, J Yang, ©, Motoda, , MeLaeon, ‘No.A bu 8.¥, PS, 2hou, 2-4. Stenouen M, Hand, DJ. ana Seber 1s 2608, “Tan 0 algartrnsm date mining” Recivedpe and Inermatin Systems val No. ap. 132 Xa, M, Watenacsturaor, P,Vrshnoy, PK. and ora, MK. 2005, “Decision ‘vee agressonfors0%chss feston of rrnote sensing deta” Remote Sens Ing of Emirennart Vo 87ND. 2. 9.222996, Zaaom Li, 1988, "usay v0” Inlet Conta Vo. 8, 9, 228-282 a1 Vol. 338 + 2015 TRANSACTIONS.

You might also like