Rockburst Laboratory Tests Database- Application of Data Mining Techniques

15
Rockburst laboratory tests database Application of data mining techniques Manchao He a , L. Ribeiro e Sousa a,b, , Tiago Miranda c , Gualong Zhu a a State Key Laboratory for Geomechanics and Deep Underground Engineering, Beijing, China b University of Porto, Porto, Portugal c University of Minho, Guimarães, Portugal abstract article info Article history: Received 13 March 2014 Received in revised form 22 September 2014 Accepted 13 December 2014 Available online 18 December 2014 Keywords: Rockburst Experimental tests Data mining Rockburst index Rockburst is characterized by a violent explosion of a block causing a sudden rupture in the rock and is quite com- mon in deep tunnels. It is critical to understand the phenomenon of rockburst, focusing on the patterns of occur- rence so these events can be avoided and/or managed saving costs and possibly lives. The failure mechanism of rockburst needs to be better understood. Laboratory experiments are undergoing at the Laboratory for Geomechanics and Deep Underground Engineering (SKLGDUE) of Beijing and the system is described. A large number of rockburst tests were performed and their information collected, stored in a database and analyzed. Data Mining (DM) techniques were applied to the database in order to develop predictive models for the rockburst maximum stress (σ RB ) and rockburst risk index (I RB ) that need the results of such tests to be deter- mined. With the developed models it is possible to predict these parameters with high accuracy levels using data from the rock mass and specic project. © 2014 Elsevier B.V. All rights reserved. 1. Introduction A large number of accidents and other associated problems occur during construction and exploration of underground structures, and are very often related to uncertainties concerning ground conditions. Many researchers have collected, analyzed and published re- ports on accident cases in tunnels during construction and explora- tion (HSE, 1996; Vlasov et al., 2001; Sousa, 2006, 2010). In the study conducted by Sousa (2010), data on accidents were collected from the technical literature, newspapers and correspondence with experts in the underground domain. The data were stored in a database and the accidents were classied into different catego- ries, providing an evaluation on their causes and their conse- quences. The main goal was to determine the major undesirable events that may occur during tunnel construction, their causes and consequences and ultimately present mitigation measures to avoid accidents on tunnels during the construction. Different types of events were identied and classied (Rock Fall, Collapse and Daylight Collapse, Rockburst, Excessive Deformation, Water Inow, Fires and Explosions). The accidents can cause loss of lives, equipment damage and damage to the tunnel structure that may lead to collapse. In deep underground structures, rockburst is a frequent type of event caused by the overstressing of rock mass or intact brittle rock, when stresses exceed the local compressive strength of the material. It can cause spalling or in the worst cases, sudden and violent failure of the rock mass. Rockbursts can cause serious, and often fatal, injuries. They are mainly dependent on the stress exerted on the rock, which in- creases with depth. Rockburst is also common in deep underground mines. This phenomenon can also occur in tunnels for transportation systems and hydroelectric projects (Sousa, 2012a). For deep underground engineering rockburst is one of the most im- portant accidents. They are not easy to predict. Rockburst hazard assess- ment is therefore a very important task and the major topic of this paper. Laboratory tests are one way to analyze the rockburst phenome- non. In this paper, and after a comprehensive presentation of rockburst occurrences, a description of a unique laboratory system developed at the SKLGDUE, Beijing, of China University of Mining and Technology is presented. In this work, the information regarding 139 of these tests was gathered in a database. Then Data Mining techniques (Multiple Regression, Articial Neural Networks and Support Vector Machines) were applied to the database in order to develop predictive models for rockburst parameters associated to these tests, namely the rockburst maximum stress (σ RB ) and rockburst risk index (I RB ). The developed models are based on data concerning the rock mass and specic project and can be used to predict these rockburst indexes when it is not possi- ble to carry out these rockburst tests or in preliminary stages of design of underground works to obtain an approximate prediction of their values. Different sets of input parameters were considered so that the models can adapt to the available data. Also, the importance of each input parameter was assessed. Engineering Geology 185 (2015) 116130 Corresponding author at: University of Porto, Porto, Portugal. Tel.: +351 966012385. E-mail address: [email protected] (L.R. e Sousa). http://dx.doi.org/10.1016/j.enggeo.2014.12.008 0013-7952/© 2014 Elsevier B.V. All rights reserved. Contents lists available at ScienceDirect Engineering Geology journal homepage: www.elsevier.com/locate/enggeo

description

Rockburst is characterized by a violent explosion of a block causing a sudden rupture in the rock and is quite commonin deep tunnels. It is critical to understand the phenomenon of rockburst, focusing on the patterns of occurrenceso these events can be avoided and/or managed saving costs and possibly lives. The failure mechanism ofrockburst needs to be better understood. Laboratory experiments are undergoing at the Laboratory forGeomechanics and Deep Underground Engineering (SKLGDUE) of Beijing and the system is described. A largenumber of rockburst tests were performed and their information collected, stored in a database and analyzed.Data Mining (DM) techniques were applied to the database in order to develop predictive models for therockburst maximum stress (σRB) and rockburst risk index (IRB) that need the results of such tests to be determined.With the developed models it is possible to predict these parameters with high accuracy levels usingdata from the rock mass and specific project.

Transcript of Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Page 1: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Engineering Geology 185 (2015) 116–130

Contents lists available at ScienceDirect

Engineering Geology

j ourna l homepage: www.e lsev ie r .com/ locate /enggeo

Rockburst laboratory tests database — Application of datamining techniques

Manchao He a, L. Ribeiro e Sousa a,b,⁎, Tiago Miranda c, Gualong Zhu a

a State Key Laboratory for Geomechanics and Deep Underground Engineering, Beijing, Chinab University of Porto, Porto, Portugalc University of Minho, Guimarães, Portugal

⁎ Corresponding author at: University of Porto, Porto, PE-mail address: [email protected] (L.R. e Sousa)

http://dx.doi.org/10.1016/j.enggeo.2014.12.0080013-7952/© 2014 Elsevier B.V. All rights reserved.

a b s t r a c t

a r t i c l e i n f o

Article history:Received 13 March 2014Received in revised form 22 September 2014Accepted 13 December 2014Available online 18 December 2014

Keywords:RockburstExperimental testsData miningRockburst index

Rockburst is characterized by a violent explosion of a block causing a sudden rupture in the rock and is quite com-mon in deep tunnels. It is critical to understand the phenomenon of rockburst, focusing on the patterns of occur-rence so these events can be avoided and/or managed saving costs and possibly lives. The failure mechanism ofrockburst needs to be better understood. Laboratory experiments are undergoing at the Laboratory forGeomechanics and Deep Underground Engineering (SKLGDUE) of Beijing and the system is described. A largenumber of rockburst tests were performed and their information collected, stored in a database and analyzed.Data Mining (DM) techniques were applied to the database in order to develop predictive models for therockburst maximum stress (σRB) and rockburst risk index (IRB) that need the results of such tests to be deter-mined. With the developed models it is possible to predict these parameters with high accuracy levels usingdata from the rock mass and specific project.

© 2014 Elsevier B.V. All rights reserved.

1. Introduction

A large number of accidents and other associated problems occurduring construction and exploration of underground structures, andare very often related to uncertainties concerning ground conditions.

Many researchers have collected, analyzed and published re-ports on accident cases in tunnels during construction and explora-tion (HSE, 1996; Vlasov et al., 2001; Sousa, 2006, 2010). In thestudy conducted by Sousa (2010), data on accidents were collectedfrom the technical literature, newspapers and correspondencewith experts in the underground domain. The data were stored ina database and the accidents were classified into different catego-ries, providing an evaluation on their causes and their conse-quences. The main goal was to determine the major undesirableevents that may occur during tunnel construction, their causesand consequences and ultimately present mitigation measures toavoid accidents on tunnels during the construction. Differenttypes of events were identified and classified (Rock Fall, Collapseand Daylight Collapse, Rockburst, Excessive Deformation, WaterInflow, Fires and Explosions). The accidents can cause loss oflives, equipment damage and damage to the tunnel structure thatmay lead to collapse.

In deep underground structures, rockburst is a frequent type ofevent caused by the overstressing of rock mass or intact brittle rock,

ortugal. Tel.: +351 966012385..

when stresses exceed the local compressive strength of the material. Itcan cause spalling or in the worst cases, sudden and violent failure ofthe rock mass. Rockbursts can cause serious, and often fatal, injuries.They are mainly dependent on the stress exerted on the rock, which in-creases with depth. Rockburst is also common in deep undergroundmines. This phenomenon can also occur in tunnels for transportationsystems and hydroelectric projects (Sousa, 2012a).

For deep underground engineering rockburst is one of the most im-portant accidents. They are not easy to predict. Rockburst hazard assess-ment is therefore a very important task and the major topic of thispaper. Laboratory tests are one way to analyze the rockburst phenome-non. In this paper, and after a comprehensive presentation of rockburstoccurrences, a description of a unique laboratory system developed atthe SKLGDUE, Beijing, of China University of Mining and Technology ispresented. In this work, the information regarding 139 of these testswas gathered in a database. Then Data Mining techniques (MultipleRegression, Artificial Neural Networks and Support Vector Machines)were applied to the database in order to develop predictive models forrockburst parameters associated to these tests, namely the rockburstmaximum stress (σRB) and rockburst risk index (IRB). The developedmodels are based on data concerning the rockmass and specific projectand can be used to predict these rockburst indexes when it is not possi-ble to carry out these rockburst tests or in preliminary stages of designof underground works to obtain an approximate prediction of theirvalues. Different sets of input parameters were considered so that themodels can adapt to the available data. Also, the importance of eachinput parameter was assessed.

Page 2: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Fig. 1. Jinping II hydropower scheme.

Table 1Statistics of rockburst occurrence at Jinping II.

Tunnel No rockburst(%)

Level I(%)

Level II(%)

Level III(%)

Level IV(%)

Auxiliary A 81.01 12.54 4.13 1.73 0.09Auxiliary B 83.72 10.32 4.67 1.12 0.17HP no. 1 92.62 6.76 0.62 0.00 0.00HP no. 2 88.39 7.62 3.36 0.50 0.12HP no. 3 92.48 7.20 0.32 0.00 0.00HP no. 4 87.44 10.52 1.78 0.26 0.00Total HP 90.10 8.04 1.62 0.21 0.04

HP — High Pressure tunnel.

117M. He et al. / Engineering Geology 185 (2015) 116–130

2. Rockburst occurrences

Rockburst is characterized by a violent explosion of a block causing asudden rupture in the rock mass and can be common in very deep tun-nels. This phenomenon can occur in tunnels for transportation systems,hydroelectric projects and mining operations, and has been more asso-ciated with mine excavations for a long time. Therefore it is critical tounderstand rockburst, focusing on the patterns of occurrence so theseevents can be avoided and/or managed saving costs and possibly lives(Camiro, 1995; Kaiser, 2009; Tang et al., 2010).

All rockbursts produce seismic waves and seismic disturbances.These seismic events have been recorded at seismological stationsmany times hundreds of miles away from the origin (Tang et al.,2010). Rockburst first became a recognized problem in the Kolar FoldField of India in 1898 and by the end of 1903, 75 rockbursts had oc-curred with fatalities and serious injuries at a depth of about 450 mbelow the surface. Investigations concluded that rockbursts werecaused by great pressures on the mine pillars (Blake, 1972). Rockburstaccidents were also reported in gold mines of Witwaterstrand, SouthAfrica, in the early 1900s at lower depths damaging pillars and othermine workings. Rockbursts occur frequently in South Africa, thereforelong-term researches have been carried out systematically in the coun-try on the mechanisms of rockbursts. In 1975, 680 accidents took placein 31 gold mines which claimed a large number of fatalities and loss ofproduction (Tang et al., 2010).

Also in other types of underground structures rockburst have beenreported for instance during the construction of the Simplon hydraulictunnel in the Alpes region at depths greater than 2200m, in the Shimizutunnel in Japan, for depths between 1000 m and 1300 m and in theKanestu tunnel, constructed mainly in quartz diorite, rockburst oc-curred at an overburden depth between 730 m and 1050 m (Tanget al., 2010). Norwegian tunneling experience includes a significant

Fig. 2. Simplified geological profile of the high pressure t

number of tunnels subjected to high rock stresses. The majority of theproblems are associated with spalling due to anisotropic stressesbelow steep valleys. This is found normally in road tunnels along or be-tween the fjords under high overburden. An example is the 24.5 kmlong Laerdal tunnel where moderately intense spalling and slabbingwas encountered most of the times. In some areas heavy rockburstscaused violent ejection of sharp edged rock plates (Sousa, 2010).Rockburst also occurred at the Gothard base railway tunnel inSwitzerland.

In China the first rockburst was recorded in 1933 at Shengli Mine inFushun. Based on the available data it is estimated that over 2,000 coalbursts occurred in 33 mines in China during 1949–1997. In the periodfrom 2001 to 2007 rockburst in deep metal mines caused more than13,000 accidents and at least 16,000 casualties (Adoko et al., 2013).

In China many rockbursts occurred also during excavation of highpressure, drainage and auxiliary tunnels of the Jinping II hydropowerscheme (Figure 1), (Wu et al., 2010; Feng and Hudson, 2011; Fenget al., 2012a). This scheme is composed by four high pressure tunnels,each with 16.67 km in length, 60 m spacing between them, two parallelauxiliary tunnels A and B, 17.5 km longwith a span of 6m excavated byDrill & Blast (D&B) and a drainage tunnel of about 16.73 km, with a di-ameter of 7.2m excavated by a Robbins TBMand byD&B. The high pres-sure tunnel nos. 1 and 3 were excavated, respectively, by a Robbin TBMand by a Herrenknecht TBM both 12.4m in diameter. The high pressuretunnel nos. 2 and 4 were excavated by D&B method with an equivalentdiameter of about 13m. They were excavated in marble, sandstone andslate strata, with amaximum overburden up to 2500m (Figure 2), (Wuet al., 2010; Wang et al., 2012).

A consultingWorkshop took place in 2009 organized by the ChineseSociety for Rock Mechanics and Engineering about the Jinping II longtunnels. Several reports were elaborated mainly focusing on rockburstproblem (He, 2009; Hudson, 2009; Kaiser, 2009; Qian, 2009; Sousa,2009). Conclusions were established and suggestions were madeabout TBM problems, the estimation of in situ stresses, the geometryof the excavations when using D&B method, and modeling speciallyfor rockburst prediction. For rockburst the suggestions included the

unnel no. 1 of Jinping II project (Wang et al., 2012).

Page 3: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Table 2Classification of rockburst as used at Jinping II.

Rockburstlevel

Description Type of sound Characteristics of duration Depth of the block (m) Impact in excavation

I Light Sound of cracking Sporadic explosion b0.5 SmallII Moderate (mild) Clear sound of cracking Long duration and not progressive with time 0.5 — 1.0 A certain impactIII Intensive (strong) Sound of a strong explosion Fast with increase of overburden 1.0–2.0 Reasonable impactIV Very strong (excess of loads) Sound of an intensive explosion Sudden with increase of overburden N2.0 With large impact

118 M. He et al. / Engineering Geology 185 (2015) 116–130

establishment of the Rockburst Vulnerability Index (RVI) to be calibrat-ed with the rockburst experience (Hudson, 2009), the establishment ofa database containing information about rockburst and the descriptionof the events, application of DM techniques and development ofBayesian Networks (BN) for predicting the probability of occurrence ofrockburst, its location, depth andwidth, and time delay for the differenttypes of rockburst (Sousa, 2009).

During construction different types of rockbursts were observed atJinping II which permitted to describe the mechanism and to settlecriteria for rockburst (Feng et al., 2012b; Wang et al., 2012; Yan et al.,2012). According toWang et al. (2012), the statistics of rockburst eventsis represented in Table 1. The drainage tunnel was partly excavated bythe TBM until around 6 km, and after by D&B due to the occurrence ofvery strong rockbursts (Liu et al., 2011). The classification of therockburst is presented in Table 2, corresponding level I to lightrockburst, level II to moderate, level III to intensive and level IV tovery strong (Peixoto et al., 2011).

Rockbursts along auxiliary tunnels mainly occurred in the strata T2zand T2b (marbles). The most intensive rockbursts in T2b are very strong,intensive in T3 (sandstones), and moderate in the other strata. Mostrockbursts occurred within 6–12 m from the face and 5–20 h afterexcavation. For the high pressure tunnels, the number and intensityincrease in spite of a higher percentage of no rockburst length as it ispresented in Table 1. The fractured face was rough or dome-shaped.Tensile and shear failure resulted in wedge or dome-shaped fracturesurfaces with a depth that sometimes reached 1.6 m. Since the start ofthe high pressure tunnels, 77 rockbursts of several levels occurred attunnel no. 1; about 200 happened at tunnel no. 2; about 100 at tunnelno. 3; and more than 100 at tunnel no. 4. Since the tunnels were exca-vated in marbles, rockburst of levels III and IV occurred with increasingdepth. Themaximumejection distance of level IV rockbursts reached 5mand the depth of the crater ranged from 3 m to 5 m. Most intensiverockburst occurred within 10–30 m from the working face. The majority

Fig. 3. Schematic representation of potential for rockbursts and the effect of confinement(Castro et al., 2012).

of the rockbursts occurred on the north spandrel (1–3 clock position)and at the south arch corner (7–10 clock position),mainly due to theprin-cipal stress directions and the geological structures (Wang et al., 2012).

There are several mechanisms by which the rock fails, originatingrockburst. The main source mechanisms are usually associated withlocal underground geometry of the cavities, structural elements like pil-lars and the existing geology (Ortlepp and Stacey, 1994; Camiro, 1995;M. He et al., 2012). The representation of potential rockburst phenomenais indicated in Fig. 3. The rockburst is usually classified as a strainburst,pillar burst or fault slip burst (Castro et al., 2012; M. He et al., 2012).These topologies normally occur in large scale mining operations, but,in civil works themost commonphenomenon is strainbursting, althoughbuckling and face crushing may also occur. Also impact-inducedrockburst has to be considered for less stressed and deformed rock for-mations, created by blasting, caving and adjacent tunneling.

Rockburst phenomena have been extensively investigated by manyresearchers based on in situ and laboratory tests and also by theoreticalapproaches. Laboratory tests play an important role in understandingrockburst mechanisms, calibration of numerical models, evaluation ofmechanical parameters, and identification of the stress state where a dy-namical event may be initiated. This event can be classified according tothe potential damage, scale and violence (M. He et al., 2012; Wanget al., 2012). In terms of damage the classification referred in Table 2can be used. In terms of scale, rockbursts can be divided into sparse(length of rockburst L ≤ 10 m), large-area (10 b L≤ 20 m), and continu-ous rockbursts (L≥20m). According to the failure pattern, it canbedivid-ed in slabby spalling, bending failure, dome/wedge-shaped failure, and incavern collapse. In terms of severity of rockburst damage as a function offailure depth rockburst can be classified as represented in Fig. 4.

The construction method seems to have an influence on thebehavior of the excavation in what regards to rockburst. Not only theexistence of a support system that stops the violent ejection of frag-ments of rock is essential to guarantee safety, but also the type of con-struction process seems to have an effect on the severity of therockburst. According to experience, for the same type of conditionsand for the same rock, strain bursting is more likely to occur in a TBMtunnel than in a D&B tunnel, as happened at the Jinping II hydroelectricscheme (Feng et al., 2012b). However, for the surrounding rocks withlower stress levels, external disturbances, such blasting, caving and ad-jacent tunneling, can also trigger rockbursts. Rockbursts can be classi-fied in strainbursts and impact-induced bursts as indicated in Fig. 5(M. He et al., 2012). Strainbursts can be divided into the followingsub-types: instantaneous burst, delayed burst and pillar bursts. Theimpact-induced burst can also be divided into the three sub-types,like rockburst induced by blasting or excavation, by roof collapse andby fault slip, according to their formation mechanisms. An influence di-agram containing the factors that affect the occurrence of a rockburst ispresented in Fig. 6. The type and the strength of the rock are other im-portant factors affecting rockburst and its severity. Rockburst occursmore likely and with greater severity in brittle rocks. Geometry of thecavity, particularly the shape, the existence of faults and the in situstate of stress are also important parameters to be considered in the in-fluence diagram of rockburst.

Many indexes and predictive models related to rockburst can befound in literature (Adoko et al., 2013). As stated, in the Workshop

Page 4: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Fig. 4. Severity of rockburst damage as a function of depth of failure.Adapted by Kaiser, 2009.

119M. He et al. / Engineering Geology 185 (2015) 116–130

held for Jinping II, Hudson (2009) proposed the index RVI which canprovide an indication of the likelihood of rockburst occurrence. As inother empirical systems, like RMR and Q systems, for the key variablescertain values are allocated and the index is established from thesevalues. The RVI index is calibrated with experience and can be refinedwith future knowledge. For the Jinping II hydroelectric scheme theparameters proposed were the following: height of overburden;compressive strength of the intact rock; intact rock brittleness (brittle,semi-brittle, semi-ductile, ductile); presence of folding; presence ofbedding plane separations or faults; and inputs from numerical models(when available). Also, it is important for the calibration of the RVI in-dexes to take into consideration the scale effect since rockburst is stron-ger when the equivalent dimension of the tunnel increases. This factwas also evident at Jinping II where the rockburst was stronger in thehigh pressure tunnels than in the access tunnels (Hudson, 2009). Thedepth of the rockburst failure and the volume of rock involved can becorrelated with the RVI value. Adoko et al. (2013) proposed accuratepredictivemodels for rockburst intensity (RBI) based on fuzzy inferencesystem and adaptive neuro-fuzzy inference systems (ANFIS), and fieldmeasurements data. Based on a literature review a database was gath-ered composed by 174 rockbursts. The input parameters for themodelswere: the maximum tangential stress, the uniaxial compressivestrength, the uniaxial tensile strength of the rock and the elastic strainenergy index and in some cases the stress coefficient and the rock brit-tleness coefficient, while the RBI was the output. The ANFIS model indi-cates the best performance. This work used a data-driven approach i.e.the models were derived from real data whereas many other modelspresent in literature use only knowledge from experts to derive rulesand models. In the present work this data-driven approach was also

Fig. 5. Laboratory experiments methods based o

used since a databasewas gathered and themodels were developed ap-plying specific intelligent algorithms.

As shown, brittleness of the rock is one of the major issuesconcerning rockburst. Its characterization and description is essentialfor the understanding of the rockburst phenomenon. However, the def-inition and themeasurement of brittleness are not yet standardized andmany empirical relations with different approaches can be found in lit-erature to predict rock brittleness (Gong and Zhao, 2007). Normally, it ismeasured indirectly as a function of rock strength parameters like theuniaxial compressive and the Brazilian tensile strength (Yagiz andCandan, 2010). Yagiz (2009) proposed a brittleness index (BI) basedon the punchpenetration test. Later, Yagiz and Candan (2010) proposedtwomodels to predict BI based on a fuzzy inference system and nonlin-ear regression analysis. To develop the models, which presented a rele-vant accuracy level, the authors gathered a database with 48 samplesusing as input parameters the uniaxial compressive strength, theBrazilian tensile strength and the unit weight of the rock.

Rockbursts are not easy to predict. Investigations using AcousticEmission (AE) monitoring are sometimes recommended since theyallow one to monitor the accumulation of cracking and to evaluate thetendency for the rock to suffer rockburst (Tang et al., 2010; Wanget al., 2012). Rockburst is affected by several factors. The stress releaseis one of the factors affecting the level of these events. Therefore, thecharacteristics of blasting vibration and effects should be investigatedin detail, which happened in the case of Jinping II. By optimizing theblasting parameters and the network used in the blasting, the risk ofrockburst can be reduced (Wang et al., 2012). It is important to developappropriatemethodologies for the support design in the tunnel sectionscharacterized by potential rockburst. Special reference is made for the

n rockburst classification (He et al., 2012a).

Page 5: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Type and rock strengthGeometry

(diameter)

Faults (Folding)

Stress state(Overburden & K=σh/σv)

RockburstConstruction method(Support & advanced

rate)

Severity(Time delay)

Dimensions of burst (Location)

Damage of tunnel

Fatalities & injuries

Shape and equivalent

Fig. 6. Influence diagram of rockburst.Adapted from Sousa, 2010.

120 M. He et al. / Engineering Geology 185 (2015) 116–130

principles of rockburst prevention followed at Jinping II (Wang et al.,2012) and also to the work developed in Canada by Kaiser and Cai(2012) in implementing an interactive design tool for conductingrockburst support design in tunnels under burst-prone ground. Finallyspecial emphasis is made for a new Constant-Resistance and Large-Deformation (CRLD) bolts developed at SKLGDUE in order to mitigatethe disturbance impacts and control rockburst (M. He et al., 2012).These bolts can bear many abrupt loads maintaining a good supportingperformance. They have an ideally elastoplastic behavior comparedwith a traditional bolt which was corroborated by experimental tests(He et al., submitted for publication).

Risk analysis and risk management for rockburst should follow theguidelines established by ITA (Eskesen et al., 2004), and proposed byother authors (Sousa, 2010; Brown, 2012). One process suggested forDUSEL (Deep Underground Science and Engineering Laboratory) cavi-ties by Popielak andWeining (2010), involves the establishment of con-ceptual models, to perform numerical analysis in order to study thepotential impacts of rockburst and risk management plans as well asstrategies for its implementation. For instance, during the excavationsof high pressure tunnels of the Jinping II scheme by TBMs some con-struction processes were implemented for the purpose of rockburstcontrol namely to avoid the occurrence of very strong rockbursts andto prevent negative impacts on safety (Wang et al., 2012). The mainidea of the planwas to release the high in situ stresseswith pilot tunnelsand simultaneously these pilot tunnels served as geological exploratorypits and theworking faces were used formicro-seismicmonitoring. Thepilot tests were classified in three types as represented in Fig. 7. This

Fig. 7. Pilot tunnels used at Jinp

approach allowed reducing the risk of occurrence of strong rockbursts.For the high pressure D&B tunnels preventive measures were alsotaken including blasting control, use of shotcrete and bolt support(Wang et al., 2012).

The risk of rockburst can be also analyzed by using BN probabilistictools. The methodology for risk assessment and decision making pro-posed by Sousa (2012b) for tunnel projects during the design and con-struction phases can be implemented for the different hazards includingof course the rockburst scenario. The decision support framework canconsist of two models, namely: a rockburst (geologic) predictionmodel and a decisionmodel. Thismethodologywas successfully appliedto tunnels of Porto Metro in Northern Portugal where three collapsesoccurred during the construction (Sousa and Einstein, 2012).

3. Rockburst laboratory tests

3.1. True triaxial system

Laboratory experiments on rockburst have been carried out by re-searchers using several integrated static and dynamic loading and uni-axial and biaxial testing machines. However, among these laboratorytests, no one has simulated physically the rockburst process and repro-duce the circumstances of the occurrence of the rockburst. A modifiedtrue-triaxial experimental system was developed at SKLGDUE, formodeling in a more realistic way the rockburst process (He et al.,2012a, 2012b). It is a system that includes the main machine, hydraulicpressure controlling unit and data acquisition for forces, displacements,

ing II (Wang et al., 2012).

Page 6: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Fig. 8. Rockburst testing system.

121M. He et al. / Engineering Geology 185 (2015) 116–130

acoustic emission (AE), high speed digital recording and also infrared(IR) thermal monitoring system (Figure 8). During the test procedure,one surface is unloaded immediately from the triaxial condition,which simulates in an approximate way the stress path in the excava-tion face during tunneling. The evolution model for a rockburst can betranslated by three phases as illustrated in Fig. 9, with cracking in a ver-tical plan, followed by a buckling deformation and the final rockburstejection (He et al., 2010b). Fig. 10 illustrates the dropping system usedin the system for the unloading of a facewhich is implemented througha change of a pistonmovement. The stress paths used by the testing sys-tem simulate the different types of rockburst that can occur (He et al.,2010b, 2012c).

The development of the triaxial rock test machine makes it possibleto have a better understanding of the rockburst phenomena. To illus-trate the rockburst loading path, the results of a rockburst test for asandstone sample is presented (Figure 11). The sample was first loadedwith three principal stresses at the same loading rate step by step tosimulate the in situ original stress state (51.8; 39.3 and 29.5 MPa). Ateach loading level, the interval of loading timewas 5min. Then themin-imum horizontal stress on one surface of the sample was unloaded.Therefore the simulation of the rockburst started. A detailed analysisof the results of this test is presented in the publication of He et al.

Fig. 9. Evolution model for a ro

(2012a), regarding rockburst time, classification and typical results ofthe failure process.

3.2. Measuring systems

The measuring system for the rockburst experiments includes(M. He et al., 2012):

3.2.1. Data acquisition systemThe changes in forces and displacements that occurred during the

rockburst process are registered by a dynamic strain amplifier and aportable data acquisition instrument, which consists of sensors, ampli-fiers, a data acquisition instrument, a computer and the appropriateprocessing software. The dynamic data collection device can achieve adata acquisition speed as high as 100 kps. Fig. 12 presents the dataacquisition apparatus.

3.2.2. High-speed image recording systemThe system is equipped with a high-speed digital camera in order to

record the kinetic characteristics of the rock fragments ejected duringthe rockburst event, providing excellent material for the analysis of

ckburst (He et al., 2012a).

Page 7: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Fig. 10. Illustration of the dropping system for load bar and loading plate (He et al., 2010b).

Fig. 12. Data acquisition system (He et al., 2012a).

122 M. He et al. / Engineering Geology 185 (2015) 116–130

the mechanism of rockburst. It consists of high-speed cameras, capturecards, disk for storage of the images and the control software.

3.2.3. AE monitoring systemThe system is equippedwith an AEmonitoring instrument in order to

obtain the features of the development and changes in microcracks andthe characteristics of energy release during the rockburst process. TheAE includes an acquisition card, a continuous current source, a preampli-fier and an AE sensor (Figure 13) (He and Zhao, 2013). The AE character-istics are very important and can be used and analyzed in order tounderstand the crack propagation phenomena in rocks, as representedin Fig. 14 for the instantaneous rockburst of the Carraramarble. The signaland images can be investigated from the tests. The evolution and frequen-cy is important to investigate the failure characteristics of the samples. Bysignal processing using fast Fourier transformation analysis, the time fre-quency spectrum for the AE signals can be obtained (Jia, 2013).

3.2.4. IR thermal monitoring systemThe IR thermography system shows the changes in the surface

temperature of a sample which lead to uneven heat transfer. Thechanges of temperature can be analyzed using an IR thermal moni-toring system providing data for the changes in the characteristicsof the sample and consequently to better understand the mechanismof rockburst. IR thermography has been extensively used as a nonde-structive and non-contact technique to inspect cracks or defects in-side the materials (Gong et al., 2009; He et al., 2009, 2010a). IRthermography, together with such image processing procedures asfeature extraction and frequency-spectral analysis, were used for vi-sualization and characterization of the mechanical and structural re-sponses of rock masses (He et al., 2010a). The IR detection schemeused in the rockburst tests is illustrated in Fig. 15.

Fig. 11. Sandstone rockburst loading path (He et al., 2012b).

4. Database with laboratory test results

4.1. Organization of the database

From 2004 to 2012 a large number of rockburst laboratory testswere performed at SKLGDUE. A survey was conducted through severalreports containing results from rockburst tests processed during theseyears and a database was elaborated with 139 tests from 4 countries,mainly from China (88.5%), and also from Italy (5.0%) with samplesfrom Carrara quarries, from Canada (5.0%) with granite samples fromCreighton mine, and from Iran (1.5%) from samples obtained in petro-leum wells. The tested samples were mainly coal (43), sandstone (28)and granite (25), but other rocks were also tested. Fig. 16 illustratesthe relative distribution of the tests for the different types of rocks.

A formwas elaborated and a great number of fieldswere considered,namely: (1) location of the test; (2) dimensions of the sample in length,width and height; (3) rock material; (4) main minerals and cracks;(5) stresses before loading; (6) stresses during the test; (7) characteris-tics of the rockburst test; and (8) critical depth. In Table 3, the differentinformation inserted for each field is referred.

The samples had a prismatic shape with an average length of 59 mm,with a minimum of 39 mm and a maximum of 111 mm; the widthcorresponding to the face to be unloaded had 33 mm in average,61 mm as maximum and as 19 mm minimum; the height was 148 mmin average, 229 mm in maximum and 98 mm in minimum. The volumewas 312 cm3 in average, 1186 cm3 inmaximumand 72 cm3 inminimum.

The main minerals present in the samples were clay, carbon,quartz, calcite and feldspar. The information about the percentageof these minerals in the samples existed for almost all the cases.For the existence of cracks the following designation was adopted:0 — no information; 1 — few cracks; 1.5 — cracks in some parts;2 — a lot of cracks.

All the tests were of the strainburst type. The rockburst criticaldepth (He) was calculated assuming a simplified circular shape inthe crown of the tunnel, a concentration factor for the stressesequal to 2 and a specific weight of 27 kN/m3 for the overburdenrock mass, by the following expression:

He ¼ 18:52σRB ð1Þ

where σRB is the rockburst maximum stress obtained in the test.A rockburst risk index (IRB) was also calculated following the formu-

la (He, 2009):

IRB ¼ HHe

¼ 0:054HσRB

: ð2Þ

Page 8: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Fig. 13. AE detection system (Jia, 2013).

123M. He et al. / Engineering Geology 185 (2015) 116–130

4.2. Statistical results and classifications

Some of themost relevant statistics concerning the database param-eters are presented in Table 4, namely the mean values of some of themain parameters for different rock types.

UCS values have an average of 65.5 MPa for all samples, with a mini-mum average value of 8.0 MPa for shale and a maximum of 234.1 MPafor peridotite. The average values for the most representative rockswere 11.9 MPa for coal, 113.9 MPa for granite, and 83.4 MPa forsandstone.

The deformability modulus (E) has an average of 22.3 GPafor all samples, with a minimum average value of 2.0 GPa for mud-stone and a maximum of 74.1 GPa for peridotite. The average values

Fig. 14. The AE spectral on points and micro-crack structure charact

for the most representative rocks were 2.4 GPa for coal, 33.8 GPa forgranite, and 24.5 GPa for sandstone. In relation to the Poisson ratio,the average value is equal to 0.25, with a minimum average valueof 0.18 for marble and a maximum of 0.37 for mudstone and forshale.

Regarding the depth (H) of the samples, the average value isequivalent to 678 m, with a minimum average value of 250 m for dolo-mite and a maximum of 3375 m for limestone. For the critical depth He,the average value is equal to 1529m, with a minimum average value of343 m for shale and a maximum of 2989 m for granite. For averagevalues and the most representative rocks H and He were equal,respectively, to 507 m and 352 m for coal, 700 and 2989 m for granite,and 854 and 1878 m for sandstone.

eristics after an instantaneous rockburst test (He et al., 2012c).

Page 9: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Fig. 15. Scheme of an IR detection system on a rockburst tests (He et al., 2012a).

Table 3Fields considered in the database.

Field Topics

Location of the test Location sample; depth (m); country; date.Dimensions of sample Code, length; width; height (mm); volume (cm3)Rock material Type of rock; RQD; UCS (MPa); specific weight

(g/cm3); E (GPa) elastic modulus; ν — Poisson ratio.Main minerals and cracks % clay; % feldspar; % calcite; % carbon; existence of cracksStresses before loading(MPa)

σv — vertical in situ stress; σh1 — horizontal in situstress; σh2 — horizontal in situ stress in the face to beunloaded; I1 (first invariant of the stresses); I2(second invariant of the stresses); I3 (third invariantof the stresses).

Stresses during tests(MPa)

Rockburst maximum stress (σRB); maximum stressaxis; loading rate in MPa/s; unloading rate forvertical stresses in MPa/s; unloading times.

Characteristics ofrockburst test

Type of burst position; duration of the test inminutes; time of burst delay (minutes); mainlyshape of fragments.

Critical depth (m) Critical depth; rockburst risk index.

124 M. He et al. / Engineering Geology 185 (2015) 116–130

The rockburst maximum stresses (σRB) obtained in the tests have anaverage value for all samples of 82.6 MPa, with a minimum of 16.5 MPafor mudstone and a maximum of 161.4 MPa for granite. The averagevalue of coal was equal to 19.0 MPa and for sandstone 101.4 MPa.Fig. 17 shows the distribution of the rockburst stresses obtained in allsamples. A large number of samples for soft rocks are recognized withvalues lower than 40 MPa.

The percentage of some minerals namely clay, calcite, feldspar,quartz and carbon were obtained from the tested samples. The averagevalues for all samples and for the highest average values were: clay —

11.9% and 31.1% (mudstone); calcite — 8.7% and 99.5% (limestone);feldspar — 15.4% and 78.7% (peridotite); quartz — 22.1% and 51.6%(sandstone); carbon— no value and 93.2% (coal).

The rockburst risk index (IRB) was calculated for all the samples inaccordance with Eq. (2). The average value for all the samples is around1.08. Very different values were calculated for the existent rock types,with low values for basalt, granite and slate, and high values for shale,coal, mudstone and limestone, varying from 1.5 to 3.5. For themost rep-resentative rocks the following average values were obtained: 1.65(coal); 0.25 (granite); and 0.62 (sandstone). Taking these values intoconsideration the rockburst index was classified in four classes as indi-cated in Table 5. The distribution for the classes for all tests and for themost representative rock formations (coal, granite and sandstone) areindicated in Table 6. Low IRB valueswere obtained in 56% of the samples,and very high values were 13% of the total. Fig. 18 shows a plot with thenumber of samples distributed by IRB.

Some representative graphs were also obtained relating the IRB withother parameters. Fig. 19 represents the relation between IRBwithσRB. A

9%3%

31%

1%

18%1%

5%3%

1%

20%

4%

4%

Basalt

Cloud Schist

Coal

Dolomite

Granite

Limestone

Marble

Mudstone

Peridotite

Sandstone

Shale

Slate

Fig. 16. Rockburst tests in different rock types.

relation between IRB and K (ratio between average horizontal stressesand vertical stress due to overburden) is represented in Fig. 20. The fol-lowing relation was achieved, with a correlation coefficient of −0.904:

logK ¼ −0:126−0:780logIRB ð3Þ

5. Application of DM techniques

5.1. General

Nowadays, due to the advances in information and communicationtechnologies, there is an extraordinary expansion of data generationthat needs to be stored. The data can hold valuable information, suchas trends and patterns that can be used to improve decision makingand optimize processes. Due to the great potential of this subjectthere has been an increasing interest in the Knowledge Discoveryfrom Databases (KDD) and DM fields that led to the fast developmentof electronic data management methods (Fig. 21).

DM is a relatively new area of computer science that lies at the inter-section of statistics, machine learning, datamanagement and databases,pattern recognition, Artificial Intelligence (AI) and other areas. DMconsists in the searching and inference of patterns or models in thedata which can represent useful knowledge. There are several DMtechniques, each one with its own purposes and capabilities, namelyDecision Trees and Rule Induction, Neural Networks, Support VectorMachines and BN, Learning Classifier Systems and Instance-Based Algo-rithms (Berthold and Hand, 2003; Miranda et al., 2011). The increasinginterest on DM led to the necessity of defining standard procedures tocarry out this task. In this context, the two most used methodologiesin DMare the CRISP-DM(Cross-Industry Standard Process for DataMin-ing) and the SEMMA (Sample, Explore, Modify, Model, and Assess),(Bulkley et al., 1999; Chapman et al., 2000; Miranda, 2007).

In the field of geotechnics many advances have been observed, likein constitutive models, test equipment and instruments, probabilisticmethodologies to deal with uncertainty, numerical tools and manyothers. However, the tools and techniques normally used to analyzegeotechnical data did not undergo significant development during thelast decades. The establishment of standard organization and represen-tation methods of geotechnical data in an electronic format is a subjectwhose importance was recognized by the geotechnical community,particularly with the creation of Joint Committee JTC2 from the SocietiesISRM, ISSMGE and IAEG. In the events organized by this Committeesome developments were presented and the relevance of the applica-tion of DM techniques was emphasized.

Page 10: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Table 4Statistical values for different relevant parameters. 1— Basalt; 2— Schist; 3— Coal; 4—Dolomite; 5—Granite; 6— Limestone; 7—Marble; 8—Mudstone; 9— Peridotite; 10— Sandstone;11— Shale; 12— Slate.

Parameters All rocks Rocks

1 2 3 4 5 6 7 8 9 10 11 12

UCS (MPa) Mean 65.5 143.9 – 11.9 – 113.8 24.1 58.4 11.5 234.1 83.4 8.0 58.8E (GPa) Mean 22.3 62.2 – 2.4 – 33.8 9.7 43.1 2.0 74.1 24.5 3.0 13.5ν Mean 0.25 0.21 – 0.28 – 0.24 0.24 0.18 0.37 0.20 0.24 0.37 –

H (m) Mean 678 400 1000 507 250 700 3375 500 910 1554 854 500 500He (m) Mean 1529 2585 1413 352 2315 2989 1278 1561 306 2317 1878 343 1587σRB (MPa) Mean 82.6 139.6 76.3 19.0 125.0 161.4 69.0 84.3 16.5 125.1 101.4 18.5 85.7IRB Mean 1.08 0.16 0.81 1.65 1.11 0.27 3.49 0.36 3.02 0.68 0.62 1.54 0.34

125M. He et al. / Engineering Geology 185 (2015) 116–130

However studies concerning the application of formal KDD process-es are rare in geotechnical engineering activities. The results of KDDprocesses were recently presented concerning geotechnical data gath-ered in two important underground works in predominantly graniterock masses, Venda Nova II and Bemposta II hydroelectric schemes(Miranda and Sousa, 2012). New alternative regression models weredeveloped using several DM techniques for the analytical calculationof strength and deformability parameters and geomechanical indexes.These models were built up considering different sets of input data,allowing their application in different scenarios of data availability.Most of the models use less information than the original formulationsbut maintain a high predictive accuracy, which can be useful in the pre-liminary design stages in any case where geological/geotechnical infor-mation is limited. The application of DM also provided insight to themost influential parameters for the behavior of the rockmass of interest.

An important application was performed for the DUSEL laboratory,located at the former Homestake gold mine (McPherson et al., 2003).The laboratory is seen as a multi-discipline facility with particle physicsproviding the lead but other disciplines being a significant part of thefacility, including geomicrobiology, geosciences, and geoengineering.At DUSEL, a large database of geotechnical data was already produced.The geotechnical database was analyzed using these innovative DMtools, including BNmodels, and new and useful models were developedfor the prediction of geomechanical parameters (Sousa et al., 2012).

The main issue of the DM task is building a model to represent data.In this step of the KDD process, learning occurs by adopting a searchalgorithm for training. This process occurs over a training set until agiven criteria is met. After training, the model is built and its quality isnormally evaluated over a test set not used for training. There areseveral different models but there is no universal one to efficientlysolve all the problems. Each one presents specific characteristics(advantages and drawbacks) which make them better suited in a cer-tain case. Themodeling techniques applied in this study were: MultipleRegression—MR, Artificial Neural Networks— ANN and Support VectorMachines — SVM.

Fig. 17. Histogram with the distribution of the rockburst maximum stress σRB.

In regression problems, the goal is to estimate themodelwhichmin-imizes an error measurement between real and predicted values con-sidering N examples. In this work the used error measures were themean absolute error (MAE) and the root mean squared error (RMSE),as defined by equations:

Mean Absolute Error : MAE ¼

XNi¼1

eij j

Nð4Þ

Root Mean Squared Error : RMSE ¼

ffiffiffiffiffiffiffiffiffiffiffiffiffiXNi¼1

e2i

N

vuuuut ð5Þ

where N is the total number of samples and ei is the difference betweenthe real value and the estimated value by the model.

TheMR is similar to the simple regression being themain differencethe number of independent variables involved. The simple regressioninvolves only one independent variable whereas the MR involvesseveral independent variables and establishes a relationship amongthem and the dependent variable (Berry and Linoff, 2000). ANN wasconceived to imitate the biological networks of neurons found in thebrain. They are formed by groups of connected artificial neurons in asimplified but very similar structure to the brain neurons. Like the bio-logical structures, these networks can be trained and learn from a setof examples to find solutions to complex problems, recognize patternsand predict future events. The acquired knowledge can then be general-ized to solve new problems. This means that they are self-adaptive sys-tems. Multi-layer networks are the most common type of network andare composed by different parallel layers of neurons (Haykin, 1999).Fig. 22 shows the scheme of a multi-layer network used in this work.The first is the input and the last the output layer. Intermediate onesare called hidden layers. There are several architectures or topologiesfor the network, each one with its own potentialities, but the mostused is themultilayer feed-forward. In this type of network connectionsare unidirectional (from input to output) and there are no connectionsbetween neurons in the same layer forming an acyclic network.

The SVM (Cortes and Vapnik, 1995) were originally developed to beused in classification problems i.e., to model discrete labeled outputs.After the introduction of the -insensitive loss function, it was possibleto apply SVM to regression problems. The basic idea of the SVM is totransform the input data into a high-dimensional feature space byusing a nonlinear mapping, which is normally unknown, using a set of

Table 5Classification in accordance with the rockburst index.

Value of IRB Classification

IRB b 0.6 Low0.6 b IRB ≤ 1.2 Moderate1.2 b IRB ≤ 2.0 HighIRB ≥ 2.0 Very high

Page 11: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Table 6Distribution of samples for the different classes of rockburst index.

Classification for IRB All samples Coal Granite Sandstone

Low 77 3 25 22Moderate 17 10 0 3High 26 18 0 2Very high 18 12 0 1

σ RB

/ M

Pa

126 M. He et al. / Engineering Geology 185 (2015) 116–130

functions known as kernels. Then, the SVM finds the best hyperplane oflinear separation within the feature space (Miranda et al., 2011; Sousaet al., 2012).

0.6 1.2 2.0

Fig. 19. Distribution of IRB vs. σRB.

5.2. Modeling and evaluation

Among the DM algorithms used in this study, only MR provides anequation relating to the output and the input variables. The modelingsoftware was the R program environment (R Development Core Team,2010) which is an open source freeware statistical package. Withinthis framework a specific library RMiner (Cortez, 2010) was usedwhich allows applying several algorithms and evaluating their behaviorunder a different set of metrics. In this work the goal was to predict theσRB and the IRB using different groups of input variables. Hence, themodels were developed considering two groups of input parametersnamely amain and a secondary group. Themain group contains the var-iables thatmay have a significant influence on the prediction capacity ofthemodels and in the secondary group the oneswith expectedmarginalinfluence. In Table 7 the four data groups considered as input parame-ters are presented.

Before fitting theANNand SVMmodels, the datawere normalized toa zero mean and one standard deviation and the outputs were post-processed with the inverse transformation (Hastie et al., 2009).

The performance of themodelswas accessed by using 20 runs under20-fold cross validation approach (Hastie et al., 2009). Under thisscheme, the data are divided into 20 different subsets, being one usedto test the model and the remaining to fit it, which means that all dataare used for training and testing. The mean and confidence intervalsfor the error measures are then computed considering the results ofall the runs and a 95% confidence interval of a T-student distribution.These statistical measures define the range of expected errors for futurepredictions of the final model, which is estimated using all the data fortraining.

Often, complex models, such as ANN and SVM, are viewed as black-boxes which is their main drawback. However, it is possible to havean insight on how they work by applying a sensitivity analysis andstudying the most important input variables on the prediction of thetarget variable (Kewley et al., 2000; Cortez and Embrechts, 2011).Such procedure is carried out analyzing the model responses when agiven input is changed. Such quantification is determined by keepingall the inputs constant, except one that is varied through its range of

0

20

40

60

80

<0.6 0.6-1.2 1.2-2 >2

CoalOthers

IRB

Nu

mb

er o

f S

amp

les

Fig. 18. Number of samples distributed by rockburst index IRB.

values. A parameter with a strong influence induces a high variance inthe model output whereas a parameter with low importance inducesa low variance.

5.3. Model for σRB

The results for group G1, with themain variables, show good perfor-mances for all the developed models with error measures fluctuatingapproximately between 20 MPa and 32 MPa for a variable that rangesfrom 10.6 MPa to 255.5 MPa (with a mean value of 82.6 MPa). Alsothe values of the correlation coefficient are rather high and near 0.9.All the models present similar results as shown in Table 8, where themetrics for each model are presented for the groups of variables G1

and G2, considering the parameters MAE and RMSE, and R as the corre-lation coefficient. However, the model based on the SVM slightly out-performs the remaining. Fig. 23 presents the plot of experimental vs.predicted values for the SVMmodel. A fairly good distribution of valuesaround the 45° slope line can be observed pointing out for a good per-formance of the model.

Themost important parameters in the prediction of σRB are UCS andσh1, followed by Depth and σv and finally by E and σh2 with relative im-portance levels below 10%. Fig. 24 presents the importance of the vari-ables for the SVM model. The evaluation of σRB for group G1 ispresented in Table 9.

Eq. (6) translates the obtained multiple regression model for σRB.

σRB ¼ 9:132−0:013 �Hþ 0:381 � UCSþ 0:364 � Eþ 1:211� σh1−0:069 � σh2 þ 0:365 � σv ð6Þ

Considering also the secondary set of input parameters (group G2)the results slightly improve and the SVM model continues to have the

Fig. 20. Relation between the rockburst indices IRB and K.

Page 12: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Fig. 21. Phases of a KDD process (Fayyad et al., 1996).

127M. He et al. / Engineering Geology 185 (2015) 116–130

best performance. In terms of importance, the main parameters are al-most the same as in the previous case, namely UCS and σh1, followedby Depth, σv and Q. As it can be stated, in this case the importance ofE and σh2 drops and Q appears with a significant relative importance.The results are presented at Table 10.

5.4. Model for IRB

For Group G3 and considering all the tests, the model based in theANN presents excellent results (Table 11). The SVM model also hasvery good results and theMRmodel presents the worst results. It is im-portant to emphasize that the relation between the input variables andIRB is highly non-linear which explains the excellent prediction capacityof the ANNmodel. Fig. 25 presents the plot of experimental versus pre-dicted values for the ANN model. An excellent distribution of valuesaround the 45° slope line can be observed pointing out for a good per-formance of the model. The most important variables are H and σRB

followed by E, K and UCS. The importance of the variables in the ANNmodel is illustrated in Fig. 26.

The evaluation of IRB for group G3 is presented at Table 12.

Table 7Generated groups for evaluation of σRB and IRB.

Parameter Symbol σRB IRB

G1 G2 G3 G4

Depth (m) H Y Y Y YUniaxial compressive strength (MPa) UCS Y Y Y YDeformability modulus (GPa) E Y Y Y YHorizontal in situ stress (loading face) (MPa) σh1 Y Y – –

Horizontal in situ stress (unloading face) (MPa) σh2 Y Y – –

Vertical stress due to overburden σv Y Y – –

Percentage of clay (%) Cl – Y – –

Percentage of quartz (%) Q – Y – –

Percentage of feldspar (%) F – Y – –

Percentage of calcite (%) Ca – Y – –

Percentage of carbon (%) Cb – Y – YVolume of the sample Vol – Y – –

Rockburst maximum stress (MPa) σRB – – Y YRatio between average horizontal stresses and σv K – – Y Y

Y — Yes; – — No; G1 — Group 1; G2 — Group 2; G3 — Group 3; G4 — Group 4.

Fig. 22. Scheme of the multi-layer ANN.

Eq. (7) translates the obtained multiple regression model for IRBgroup G3.

IRB ¼ 1:432þ 8:035 � 10−4 �H−8:429 � 10−4 � UCS−0:009� E−0:007 � σRB−0:074 � K ð7Þ

Considering additional secondary variables (groupG4) the same ten-dency is observed in terms of the performance of the models. Table 13illustrates for this group the evaluation of IRB. The most important vari-ables are also H and σRB, followed by the volume of the samples, UCSand K. Comparatively with the previous group E loses importance andin the other hand the secondary variable volume of the sample presentsa significant importance.

Concerning group G3 the performance of the models is quitedifferent. TheMRmodel presents theworst performancewith consider-ably high error values, between 0.458 and 0.602 for MAE and RMSE,respectively for a parameter with a mean value of 0.954 and rangingfrom 0.046 to 5.207. On the other hand, the ANN model presentsexcellent results translated by very low error values and an R coefficientvery near to the unity. The SVMmodel also presents good results but itis clearly outperformed by the ANNmodel. For group G4, that considersalso a secondary set of parameters, the results are similar in relationto the previous case, although a slight decrease on the model's perfor-

Fig. 23. Experimental versus predicted σRB values for SVM model.

Table 8Metrics for the evaluation of σRB.

DM tech. σRB

G1 G2

MAE RMSE R MAE RMSE R

MR 21.20 31.08 0.873 20.46 30.43 0.880ANN 21.32 32.54 0.864 21.21 31.48 0.636SVM 20.49 30.51 0.879 19.64 29.51 0.888

Page 13: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Fig. 24. Importance of variables for predicting σRB using SVM algorithm.

Table 11Metrics for the evaluation of IRB.

DM tech. IRB

G3 G4

MAE RMSE R MAE RMSE R

MR 0.458 0.602 0.789 0.446 0.606 0.786ANN 0.069 0.114 0.989 0.094 0.207 0.976SVM 0.191 0.408 0.991 0.246 0.479 0.872

128 M. He et al. / Engineering Geology 185 (2015) 116–130

mance is observed. In fact, the input parameters that were added to theprocess could not improve the model's performance due to their lowimportance in relation to IRB. The only exception is the input parametervolume of the sample that replaced E in the order of importance.

6. Final considerations

Failure mechanism of rockburst needs to be well understood. Thelaboratory tests developed at SKLGDUE provide important informationon the subject and were described in detail. A large number of testswere performed and it was decided to create a database with the ob-tained information. Special reference was made to the introduction ofa rockburst index IRB and a classification for the rockburst in accordancewith this index. Important relations were achieved with the index andthe rockburst maximum stresses obtained in the tests σRB, as well aswith the ratio K between the average in situ horizontal stresses andthe vertical stresses due to overburden.

Table 9Evaluation of σRB for group G1 of variables.

Parameters DM technique

MR ANN SVM

H 0.130 0.140 0.117UCS 0.276 0.212 0.254E 0.084 0.109 0.094σh1 0.354 0.256 0.272σh2 0.025 0.098 0.126σv 0.134 0.183 0.136

Table 10Evaluation of σRB for group G2 of variables.

Parameters DM technique

MR ANN SVM

H 0.144 0.122 0.104UCS 0.173 0.162 0.210E 0.029 0.042 0.046σh1 0.230 0.187 0.189σh2 0.028 0.040 0.078σv 0.079 0.093 0.073Cl 0.042 0.040 0.049Q 0.121 0.098 0.067F 0.044 0.058 0.054Ca 0.076 0.065 0.024Cb 0.017 0.038 0.035Vol 0.016 0.056 0.070

With the available data composed by 139 laboratory rockburst tests,DM techniques were applied using MR, ANN and SVM algorithms, inorder to obtain predictive models that can improve the knowledgeabout rockburst through the laboratory rockburst tests. The developedmodels can be used to infer the rupture stressσRB and also the rockburstindex IRB. Two sets of input parameters were used considering themostimportant variables and considering additional secondary ones. The re-sults for σRB emphasized the importance of UCS and the horizontal insitu stresses in the face not unloaded, followed by the depth of thesample and the vertical estimated in situ stresses. Using a multiple re-gression algorithm an equation was determined. All the developedmodels presented excellent results, however the model based on theSVM algorithm presents the best performance. The models developedfor IRB presented excellent results when ANN algorithm was used,which translates the highly non-linear relationwith the input variables.

Fig. 26. Importance of variables for predicting IRB using ANN model.

Fig. 25. Experimental versus predicted IRB values for ANN model.

Page 14: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

Table 12Evaluation of IRB for group G3 of variables.

Parameters DM technique

MR ANN SVM

H 0.449 0.321 0.261UCS 0.034 0.089 0.148E 0.120 0.057 0.212σRB 0.320 0.430 0.285K 0.077 0.104 0.094

129M. He et al. / Engineering Geology 185 (2015) 116–130

The most important variables are the depth of the samples H and σRB,

followed by UCS and K. The volume of the samples is also importantwhen considering the secondary variables which can be related withthe scale effect.

AE results obtained during laboratory rockburst tests were not yetanalyzed using these DM techniques. It is planned to perform in the fu-ture in order to bring new approaches to the study of the rockburstphenomenon.

Due to the nature of the rock and the discontinuity nature of the rockmasses, the evaluation of the mechanical strength characteristics is in-fluenced by the dimensions of the tested volumes. Although rockburstis characterized by a violent explosion of a block of rock and conse-quently this phenomenon depends more of the nature of the rock, astudy was already initiated with selected cases of rockburst thatoccurred during the construction of several tunnels covering several hy-droelectric projects and mines. A first step is intended to complete thedatabase with more cases that occurred at the megaproject of JinpingII and then DM technique will be applied to this particular project inorder to identify the more relevant parameters and establish modelsfor some of the parameters. Several rockburst tests were also performedfor Jinping II site particularly in marbles. In these formations a largenumber of in situ rockburst events occurred. Using probabilistic ap-proaches a method for extrapolating laboratory data to field scale willbe evaluated.

The application of various AI techniques on existing in situ data, inparticular DM and BN, identified the importance of various parametersinvolved in the rockburst phenomenon. Special attention was given tothe development of influence diagrams and the use of BN, whichallow replacing with advantage, other techniques to better address theuncertainties involved in the phenomenon of rockburst. With the en-larged database more complex studies will be implemented regardingthe assessment of rockburst and its consequences. The studies alreadyperformed using laboratory rockburst tests can be correlated with therockbursts that occurred in situ and the scale effect can be analyzed indetail. The particular situation that occurred at Jinping II hydroelectricscheme will be emphasized due to the existence of important informa-tion obtained with laboratory rockburst tests and the occurrence ofrockburst in situ. The establishment of a new rockburst coefficient isintended to be calibrated with the obtained results.

Table 13Evaluation of IRB for group G4 of variables.

Parameters DM technique

MR ANN SVM

H 0.422 0.219 0.201UCS 0.023 0.101 0.110E 0.065 0.059 0.053σRB 0.249 0.346 0.246K 0.074 0.071 0.028Cl 0.020 0.058 0.089Cb 0.022 0.079 0.042Vol 0.125 0.072 0.130

Acknowledgments

The authors want to express their acknowledgments to the supportfrom the State Key Laboratory for Geomechanics and Deep UndergroundEngineering (Beijing) and China University of Mining and Technology tothe project entitled Risk Assessment Activities Applied to Slope Stability,Rockburst and Soft Rocks at the State Key Laboratory for Geomechanicsand Deep Underground Engineering (Beijing) of China University ofMining and Technology. Also the authors acknowledge the supportof the State National Basic Research Program of China (973 Projectno. 2010CB226800).

References

Adoko, A., Gokceoglu, C., Wu, L., Zuo, Q., 2013. Knowledge-based and data-driven fuzzymodeling for rockburst prediction. Int. J. Rock Mech. Min. Sci. 61, 86–95.

Berry, M., Linoff, G., 2000. Mastering data mining: the art and science of customer rela-tionships management. John Wiley & Sons, Inc., USA.

Berthold, M., Hand, D., 2003. Intelligent data analysis: an introduction. Second edition.Springer.

Blake, W., 1972. Rock-burst mechanics. Q. Colorado Sch. Min. 67 (1), 64.Brown, E.T., 2012. Risk assessment and management in underground rock engineering —

an overview. J. Rock Mech. Geotech. Eng. 4 (3), 193–204.Bulkley, J., Gayle, S., Hicks, B., Stephens, R., 1999. Adding the where to the who. 24th

SUGI — SAS Users Group International conference, paper 173, Miami, p. 3.Camiro, 1995. Rockburst research handbook. CAMIRO mining division, Canadian

Rockburst Research Programe, 1990–1995 6 vol. p. 977 (Sudbury).Castro, L.M., Bewick, R.P., Carter, T.G., 2012. An overview of numerical modelling applied

to deep mining. In: Sousa, Vargas, Fernandes, Azevedo (Eds.), Innovative NumericalModeling in Geomechanics. CRC Press, London, pp. 393–414.

Chapman, P., Clinton, J., Kerber, R., Khabaza, T., Reinartz, T., Shearer, C., Wirth, R., 2000.CRISP-DM 1.0. Step-by-step data mining guide. SPSS Inc., p. 73.

Cortes, C., Vapnik, V., 1995. Support vector networks. Machine Learning 20(3). KluwerAcademic Publishers, pp. 273–297.

Cortez, P., 2010. Data mining with neural networks and support vector machines usingthe R/rminer tool. In: Perner, P. (Ed.), Advances in Data Mining. Applications andtheoretical aspects. Proc. of 10th Industrial Conference on Data Mining, Berlin,Germany, Lecture Notes in Computer Science. Springer, pp. 572–583.

Cortez, P., Embrechts, M., 2011. Opening black box data mining models using sensitivityanalysis. 2011 IEEE Symposium on Computational Intelligence and Data Mining(CIDM 2011), pp. 341–348.

Eskesen, S., Tengborg, P., Kampmann, J., Veicherts, T., 2004. Guidelines for tunnelling riskmanagement: ITA, working group no. 2. Tunn. Undergr. Space Technol. 19, 217–237.

Fayyad, U., Piatesky-Shapiro, G., Smyth, P., 1996. From data mining to knowledgediscovery: an overview. In: Fayyad (Ed.), Advances in Knowledge Discovery andData Mining. AAAI Press, The MIT Press, Cambridge MA, pp. 471–493.

Feng, X., Hudson, J., 2011. Rock engineering design. Taylor & Francis, London, p. 468.Feng, X., Jiang, Q., Sousa, L.R., Miranda, T., 2012a. Underground hydroelectric

powerschemes. In: Sousa, Vargas, Fernandes, Azevedo (Eds.), Innovative NumericalModeling in Geomechanics. CRC Press, London, pp. 13–50.

Feng, X., Chen, B., Li, S., Zhang, C., Xiao, Y., Feng, G., Zhou, H., Qiu, S., Zho, Z., Chen, D., Ming,H., 2012b. Studies on the evolution process of rockbursts in deep tunnels. J. RockMech. Geotech. Eng. 4 (4), 289–295.

Gong, M., Zhao, J., 2007. Influence of rock brittleness on TBM penetration rate inSingapore granite. Tunn. Undergr. Space Technol. 22, 317–324.

Gong, W.L., Zhao, H.Y., An, L.Q., Mao, L.T., 2009. Temporal and spatial analysis of infraredimages from water jet in frequency domain based on DFT. J. Beijing Univ. Aeronaut.Astronaut. 34 (6), 690–694 (2008).

Hastie, T., Tibshirani, R., Friedman, J., 2009. The elements of statistical learning: data min-ing, inference, and prediction. 2nd edition. Springer-Verlag.

Haykin, S., 1999. Neural networks — a comprehensive foundation. 2nd edition. Prentice-Hall, New Jersey.

He, M.C., 2009. The mechanism of rockburst and its countermeasure of support. Int. Con-sultation Report for the Key Technology of Safe and Rapid Construction for Jinping IIHydropower Station High Overburden and Long Tunnels, Jinping II, pp. 23–28.

He, M.C., Zhao, F., 2013. Laboratory study of unloading rate effects on rockburst. J. DisasterAdv. 6 (9), 11–18 (September).

He, M.C., Gong, W.L., Li, D.J., Zhai, H.M., 2009. Physical modeling of failure process of theexcavation in horizontal strata based on IR thermography. Int. J. Rock Mech. Min.Sci. Technol. 19 (6), 689–698.

He, M.C., Jia, X., Gong, W.L., Faramarzi, L., 2010a. Physical modeling of an undergroundroadway excavation in vertically stratified rock using infrared thermography. Int.J. Rock Mech. Min. Sci. 47, 1212–1221.

He, M.C., Miao, J.L., Feng, J.L., 2010b. Rockburst process of limestone and its acoustic emis-sion characteristics under true-triaxial unloading conditions. Int. J. Rock Mech. Min.Sci. 47, 286–298.

He, M., Xia, H., Jia, X., Gong, W., Zhao, F., Liang, K., 2012a. Studies on classification, criteriaand control of rockbursts. J. Rock Mech. Geotech. Eng. 4 (2), 97–114.

He, M.C., Jia, X.N., Gong, G.J., Zhao, F., 2012b. A modified true triaxial test system that al-lows a specimen to be unloaded on one surface. In: Kwasniewski, Li, Takahashi (Eds.),True Triaxial Testing of Rocks. CRC Press, London, pp. 251–266.

Page 15: Rockburst Laboratory Tests Database- Application of Data Mining Techniques

130 M. He et al. / Engineering Geology 185 (2015) 116–130

He, M.C., Jia, X.N., Coli, M., Livi, E., Sousa, L.R., 2012c. Experimental study on rockbursts inunderground quarrying of Carrara marble. Int. J. Rock Mech. Min. Sci. 52, 1–8.

He, M.C., Gong, W., Wang, J., Qi, P., Tao, Z., Du, S., 2014. Development of a novel energy-absorbing bolt with extraordinarily large elongation and constant resistance. Int.J. Rock Mech. Min. Sci. (submitted for publication).

HSE, 1996. Safety of New Austrian Tunnelling Method (NATM) tunnels. Health & SafetyExecutive, London, p. 86.

Hudson, J., 2009. Prediction rockburst occurrence and development of the rockburst vul-nerability index (RVI). Int. Consultation Report for the Key Technology of Safe andRapid Construction for Jinping II Hydropower Station High Overburden and LongTunnels, Jinping II, pp. 25–31.

Jia, X.N., 2013. Experimental study on acoustic emission Eigen-frequency spectrum ofstrainbursts (in Chinese). PhD Thesis. China University of Mining and Technology,Beijing, p. 208.

Kaiser, P.K., 2009. Failure mechanisms and rock support aspects. Int. Consultation Reportfor the Key Technology of Safe and Rapid Construction for Jinping II Hydropower Sta-tion High Overburden and Long Tunnels, Jinping II, pp. 62–71.

Kaiser, P., Cai, M., 2012. Design of rock support system under rockburst condition. J. RockMech. Geotech. Eng. 4 (3), 215–227.

Kewley, R., Embrechts, M., Breneman, C., 2000. Data strip mining for the virtual design ofpharmaceuticals with neural networks. IEEE Trans. Neural Networks 11 (3), 668–679.

Liu, L., Wang, X., Zhang, Y., Jia, Z., Duan, Q., 2011. Tempo-spatial characteristics and influ-ential factors of rockburst: a case study of transportation and drainage tunnels inJinping II hydropower station. J. Rock Mech. Geotech. Eng. 3 (2), 179–185.

McPherson, B., Elsworth, D., Fairhurst, C., Kelsler, S., Onstott, T., Roggenthen,W., Wang, H.,2003. EarthLab: A subterranean laboratory and observatory to study microbial life,fluid flow, and rock deformation. A Report the National Science Foundation. NSF,Washington, DC, p. 60.

Miranda, T., 2007. Geomechanism parameters evaluation in underground structures:artificial intelligence, Bayesian probabilities and inverse methods. PhD thesis.University of Minho, p. 317.

Miranda, T., Sousa, L.R., 2012. Application of data mining techniques for the developmentof geomechanical characterization models for rock masses. In: Sousa, Vargas,Fernandes, Azevedo (Eds.), Innovative Numerical Modeling in Geomechanics. CRCPress, London, pp. 245–264.

Miranda, T., Correia, A.G., Santos, M., Sousa, L.R., Cortez, P., 2011. Newmodels for strengthand deformability parameters calculation in rock masses using data mining tech-niques. Int. J. Geomech. 11 (44), 44–58.

Ortlepp, W.D., Stacey, T.R., 1994. Rockburst mechanisms in tunnels and shafts. Tunn.Undergr. Space Technol. 9 (1), 59–65.

Peixoto, A., Sousa, L.R., Sousa, R.L., Feng, X.T., Miranda, T., Martins, F., 2011. Prediction ofrockburst based on an accident data base. 11th ISRM Congress, Beijing, pp. 1247–1252.

Popielak, R., Weining, W., 2010. Engineering and design services for excavation — DUSEL.Preliminary Design, Preliminary Report #2, Contract D10-04, Lakewood, Coloradop. 104.

Qian, Q., 2009. The strategy for controlling water inflow. Int. Consultation Report for theKey Technology of Safe and Rapid Construction for Jinping II Hydropower StationHigh Overburden and Long Tunnels, Jinping II, pp. 15–18.

R Development Core Team, 2010. R: a language and environment for statistical computing. RFoundation for Statistical Computing, Vienna, Austria (http://www.R-project.org).

Sousa, L.R., 2006. Learning with accidents and damage associated to underground works.In: Matos, Sousa, Kleberger, Pinto (Eds.), Geotechnical Risks in Rock Tunnels. CRCPress, London, pp. 7–39.

Sousa, L.R., 2009. Continuing site investigation and risk assessment. Int. Consultation Re-port for the Key Technology of Safe and Rapid Construction for Jinping II HydropowerStation High Overburden and Long Tunnels, Jinping II, pp. 1–7.

Sousa, R.L., 2010. Risk analysis for tunneling projects. PhD Thesis. Massachusetts Instituteof Technology, Cambridge, p. 589.

Sousa, L.R., 2012a. Report for the State Administration of Foreign Experts Affairs.SKLGDUE, Beijing, p. 54.

Sousa, R.L., 2012b. Risk assessment in tunnels using Bayesian Networks. In: Sousa, Vargas,Fernandes, Azevedo (Eds.), Innovative Numerical Modelling in Geomechanics. CRSPress, London, pp. 211–244.

Sousa, R.L., Einstein, H., 2012. Risk analysis during tunnel construction using BayesianNetworks: Porto Metro case study. Tunn. Undergr. Space Technol. 27, 86–100.

Sousa, L.R., Miranda, T., Roggenthen,W., Sousa, R.L., 2012. Models for geomechanical char-acterization of the rock mass formations at DUSEL using data mining techniques. USRock Mechanics Symposium, Chicago, ARMA 12-120, p. 14 (in CD-Rom).

Tang, S., Tong, M., Hu, J., He, X., 2010. Characteristics of acoustic emission signals in dampcracking coal rocks. Min. Sci. Technol. 20, 143–147.

Vlasov, S.N., Makovsky, L.V., Merkin, V.E., 2001. Accidents in transportation and subwaytunnels. Construction to operation. Russian Tunneling Association, Moscow, p. 198.

Wang, J., Zeng, X., Zhou, J., 2012. Practices on rockburst prevention and control inheadrace tunnels of Jinping II hydropower station. J. Rock Mech. Geotech. Eng. 4(3), 258–268.

Wu, S., Feng, X., Sousa, L.R., 2010. Jinping II mega hydropower project, China. Int.Conference on Hydroelectric Schemes in Portugal. A New Cycle, Porto, pp. 223–231.

Yagiz, S., 2009. Assessment of brittleness using rock strength and density with punch pen-etration test. Tunn. Undergr. Space Technol. 24 (1), 66–74.

Yagiz, S., Candan, G., 2010. Application of fuzzy inference system and nonlinear regressionmodels for predicting rock brittleness. Expert Syst. Appl. 37 (2010), 2265–2272.

Yan, P., Lu, W., Chen, M., Shan, Z., Chen, X., Zhou, Y., 2012. Energy release process of sur-rounding rocks of deep tunnels with two excavation methods. J. RockMech. Geotech.Eng. 4 (2), 160–167.