Machine learning methods for rockburst prediction-state-of-the-art
review
Yuanyuan Pu a ,Derek B.Apel a,⇑,Victor Liu a ,Hani Mitri b
tip网络用语a School of Mining and Petroleum Engineering,University of Alberta,Edmonton T6G 2R3,Canada b
Department of Mining and Materials Engineering,McGill University,Montreal H3A 2T6,Canada
a r t i c l e i n f o Article history:
Received 1October 2018
Received in revid form 28December 2018Accepted 2January 2019
Available online 24June 2019Keywords:
Rockburst prediction Burst liability
Artificial neural network Support vector machine Deep learning
a b s t r a c t
One of the most rious mining disasters in underground mines is rockburst phenomena.They can lead to injuries and even fatalities as well as damage to underground openings and mining equipment.This has forced many rearchers to investigate alternative methods to predict the potential for rockburst occur-rence.However,due to the highly complex relation between geological,mechanical and geometric parameters of the mining environment,the traditional mechanics-bad prediction methods do not always yield preci results.With the emergence of machine learning methods,a breakthrough in the prediction of rockburst occurrence has become possible in recent years.This paper prents a state-of-the-art review of various applications of machine learning methods for the prediction of rockburst poten-tial.First,existing rockburst prediction methods are introduced,and the limitations of such methods are highlighted.A brief overview of typical machine learning methods and their main features as predictive tools is then prented.The current applications of machine learning models in rockburst prediction are surveyed,with related mechanisms,technical details and performance analysis.
Ó2019Published by Elvier behalf of China University of Mining &Technology.This is an open
access article under the CC BY-NC-ND licen (creativecommons/licens/by-nc-nd/4.0/).
1.Introduction
Rockburst is a common geological hazard encountered in min-ing engineering and rock engineering and can damage equipment and lead to injuries and death [1–8].Virtually all mining countries have records about rockburst hazards.In Canada,more than 15mines reported rockburst ca histories,including the Brunswick lead-zinc mine at Bathurst,the Lake Shore mine,Teck-Hughes mine,Wright-Hargreaves mine,and Macassa gold mines at Kirk-land Lake [9].In the United States,from 1936to 1993,172rock-burst cas were recorded.The cas resulted in more than 78fatalities and 158injuries [10,11].During November 1996,rock-bursts causing three fatalities and five additional rious injuries occurred in a two-week period [12].Rockburst occurrences in Ger-many have declined in recent years,not becau of better tech-niques that can predict or limit the occurrence and verity but becau of a decrea in underground mining.Despite the decrea in underground mining activity,Germany still recorded rockbursts which led to injuries and fatalities;bet
ween 1983and 2007,more than 40cas have been recorded with injuries and deaths [13].In Australia,the first rockburst event with related fatalities and inju-
ries occurred in 1917at the Golden Mile underground working face in Kalgoorlie.Hundreds of rockbursts and mine ismicity were obrved.Between 1996and 1998,three fatalities in west Aus-tralian underground mines occurred as a result of falls of ground potentially associated with large ismic events [14].Due to high-stress mining conditions,rockburst hazards have become an increasingly frequent problem in Australia [15].China is currently the world’s largest coal producer.With its high rate of under-ground coal production,China has en a steady increa in the number of recorded rockbursts.More than 100Chine mines have recorded rockbursts [16–18].In November 2011,a rious rock-burst occurred in the Qianqiu Mine in Henan province,injuring 64miners and killing 10.Fig.1shows a historical rockburst map of 1108events during the period from 1995to 2000.All the loss are a proof that rockbursts are a rious problem and should be given attention.
2.Traditional rockburst prediction methods
Rockburst prediction may be classified into two categories:long-term prediction,which can be ud at
the design stage of engineering,and short-term prediction,which is helpful during the life of an engineering project [20].
doi/10.1016/j.ijmst.2019.06.009
2095-2686/Ó2019Published by Elvier behalf of China University of Mining &Technology.
This is an open access article under the CC BY-NC-ND licen (creativecommons/licens/by-nc-nd/4.0/).
⇑Corresponding author.
E-mail address:dapel@ualberta.ca (D.B.Apel).
International Journal of Mining Science and Technology 29(2019)
565–570
Contents lists available at ScienceDirect
International Journal of Mining Science and Technology
journal homepage:www.el v i e r.c o m /l o c a t e /i j m s
t
Short-term rockburst prediction mainly includesfield monitor-ing such as micro-ismic method,electromagnetic radiation method,drilling cutting method,micro-gravity method,infrared thermal imaging method.Butt et al.ud high frequency micro-ismic system to capture high frequency microismic events at Creighton Mine in Sudbury,Canada[21].The high frequency events were caud by microfractures of the overstresd rock. Similarfindings were reported by Wang and his colleagues who ud microismic monitoring techniques to predict rockbursts in Jinping II hydropower station,and propod that precursory microcracking exists prior to most rockbursts,which could be cap-tured by the micro-ismic monitoring system[22].The stress con-centration is evident near structural discontinuities,which should be the focus of rockburst monitoring.Frid et al.propod that higher stress associated with incread rockburst hazard in rocks near a working mine generates an increa in their natural electro-magnetic radiation(EMR)[23].Petukhovfirst put forward the idea of using the volume of‘‘drilled coal rubble”to estimate a rockburst hazard at coal mines[24].The corresponding mechanism of this method is drilling a borehole that leads to excitation of an inten-sive process of cracking in the zone of its influence.Fajklewic explained the relationship between a micro-gravity anomaly and the occurrence of a rockburst,and pointed out that during the pro-cess of a rockburst the variation of micro-gravity anomaly changes from a positive value to a negative value[25].At the point before a rockburst,the ne
gative value of a micro-gravity anomaly will be an extreme value.Some rearchers measured moisture content in coal ams and propod that when the moisture content in a coal am is greater than3%,there is no rockburst hazard.Zhang et al. employed a thermal infrared radiation system monitoring the tem-perature variation in a tunnelfloor surface at coal mines and found that the infrared radiation temperature on the left and right walls of tunnels shows a sudden increa before rockbursts[26].Short-term rockburst prediction methods can be implemented after the completion of underground development.Only after excavating an underground tunnel or drift can the monitoring equipment be installed at underground excavations.On the other hand,in order to avoid areas with high rockburst hazard during excavation,a long-term rockburst prediction method should be employed at the design stage of future excavations.
Long-term rockburst prediction is bad on a combination of evaluating rockburst potential andfield conditions.To evaluate burst potential,some scholars put forward veral indicators.The strain energy storage index(W ET),which refers to a ratio between strain energy retained(W sp)and strain energy dissipated(W st),is propod by Kidybinski[27].Wattimena et al.ud elastic strain energy density as a burst potential indicator[28].The rock brittle-ness coefficient,bad on the ratio between UCS(uniaxial com-pressive stress)and tensile stress,is another widely ud burst liability indicator[29]fuelgauge
.A criterion of tangential stress,the ratio between tangential stress around underground excavations,r h, and the UCS of rock,r c can be ud to asss the rock burst ten-dency[30].Mitri et al.developed an energy-bad burst potential index(BPI)to diagno the burst proneness[31].Table1shows some frequently ud burst potential evaluation indicators.
However,rockburst occurrence relates to a number of factors, including geologic structure,mining or excavation methods, mechanical properties of rocks,and in-situ stress[32].Further-more,the mutual effects of the impact factors for occurrence of rockburst are still not clear.As such,current prediction methods have significant limitations in engineering.Given this situation, some scholars have tried to u machine learning methods to pre-dict
rockbursts.
Fig.1.A historical rockburst map for the period of1995–2000[19].
Table1
Frequently ud rockburst potential asssment indicators.
Number Name Description Classification criteria
1Strain energy storage index(W ET)[27]A ratio between strain energy and retained(W sp)
strain energy dissipated(W st)W ET 2.0,no burst potential;
2.0<W ET
3.5,weak burst potential;
3.5<W ET<5.0,moderate burst potential; W ET!5.0,strong burst potential.
2Strain energy density(SED)[28]SED¼r2c
2E S SED50,No burst potential;
50SED<100,Weak burst potential; 100SED200,Moderate burst potential; SED>200,Strong burst potential.
r c is uniaxial compression stress(MPa),and E S is unloading elastic modulus(GPa)
3Rock brittleness(B)[29]B¼r c r
T B>40,No burst potential;
26.7<B40,Weak burst potential;
14.5<B26.7,Moderate burst potential; B14.5,Strong burst potential.
r c is uniaxial compression stress(MPa),and r T is the tensile strength(MPa)
4Criterion of tangential stress(Ts)[30]T s¼r h r
c Ts<0.3,no burst potential;
0.3Ts<0.5,Weak burst potential; 0.5Ts<0.7,Moderate burst potential; Ts!0.7,Strong burst potential.
r h is the tangential stress in rockmass surrounding the excavations and r c is the UCS of rock
5Failure duration time(D t)[30]The failure duration time from the beginning of
peak strength to total failure D t>500ms,No burst potential;
50ms<D t500ms,Moderate burst potential;
D t100ms,Strong burst potential.
6Energy-bad burst potential index(BPI)[31]BPI¼ESR
e c
Â100%
ESR(kJ/m3)is the energy storage density in the rock
mass and e c(kJ/m3)is the maximum SED of the rock
566Y.Pu et al./International Journal of Mining Science and Technology29(2019)565–570
3.Brief introduction of machine learning methods
Machine learning can be initially dated back to the rearch about artificial neural network.Warren McCulloch et al.propod a hierarchical model of a neural network,which was ud as a cal-culation theory for neural networks [33].Frank Ronblatt put for-ward the concept of ‘‘Perceptron”[34].Furthermore,he designed the first computer neural network in the world.This perceptron algorithm became a pioneer of machine learning methods.Hubel and Wiel put forward the famous ‘‘Hubel-Wiel biological visual model”from rearch on the cerebral cortex of cats [35].This model effectively lowered study complexity,enlightening a few subquent neural network models.However,the inability of perceptron to solve the XOR problem placed neural network rearch into the background during the 1970s.Rumelhart et al.published backpropagation algorithm (BP),which significantly decread computation burden in solving optimization problems and solved the XOR problem by adding a hidden layer in the neural network model [36].This rearch immediately activated neural network rearch again.Yann LeCun et al.propod a prevailing Convolutional Neural Network (CNN),and he derived an efficient training method for CNN bad on BP algorithm [37]N was the first successfully trained artificial neural network,which was one of the most successful and most widely ud neural network models.After the 1990s,various shallow m
achine learning models were developed such as logistic regression (LR),support vector machine (SVM),boosting [38,39].The shallow machine learning models can be regarded as a simple neural network with only one hidden layer (SVM,Boosting)or even with no hidden layer (LR).Compared with traditional machine learning methods bad on rules,the methods bad on statistical laws are more easily to be trained and simpler to analyze.However,the models only have limited learning capability,which usually fails to reprent complex functions and extract basic features [40].As computer hardware technologies improved,operational capability of the computer was no longer a barrier for machine learning model con-struction.Hinton and Salakhutdinov propod a deep learning model that utilized a multi-layer neural network to approximate functions [41].This propod model opened a new era for machine learning.Deeping learning is an intelligent learning method that is most similar to the human brain.Supported by cloud computing,big data,and other computer technologies,deep learning repre-
nts the future of machine learning [42].Fig.2demonstrates a development process of various machine learning models.4.Rockburst prediction with machine learning methods Rockburst prediction is a complex and nonlinear procedure that is influenced by model and parameter uncertainty,which is restricted by insufficient knowledge,lack of characterizing infor-mation and noisy
data [43,44].Taking advantage of machine learn-ing in dealing with nonlinear problems,some rearchers apply machine learning methods in rockburst prediction.
Sun ployed the knowledge of fuzzy mathematics and neural network to build a rockburst prediction model that trained with the improved BP algorithm bad on typical rockburst data [45].This model us fuzzy mathematics to improve comprehen-sive index and multi-index judgement.Finally,the model was suc-cessfully ud to predict rockburst in the Sanhejian coal mine in China,which showed this model is not only preci and simple but also intelligent.
Jia et al.propod a rockburst prediction method bad on the swarm optimization algorithm and general regression neural net-work (GRNN)[46].The characteristic of this model is to u the swarm optimization algorithm to determine optimal parameters of GRNN,which avoids the influences of human factors on GRNN.This model was successfully employed in rockburst prediction in the Cangshanling highway tunnel and Dongguanshan Copper Mine in China.Similarity,Zhang et al.provided a radial basis function (RBF)neural network optimized by a genetic algorithm to predict rockbursts [47].
Zhao ud a support vector machine to reprent nonlinear relationship between a rockburst and its
factors [48].This model learned from ca histories and then could be ud for a quick clas-sification of a rockburst for similar conditions.
Zhou ployed a support vector machine (SVM)to deter-mine the classification of long-term rockburst for underground openings [49].Two optimization methods:genetic algorithm and swarm optimization algorithm,are adopted to automatically determine the optimal hyper-parameters for SVMs.The results indicated that the heuristic algorithm of GA and PSO can speed up SVM’s parameter optimization arch.This propod method might hold high potential to become a uful tool in rockburst
prediction.
Fig.2.History of machine learning development.
Y.Pu et al./International Journal of Mining Science and Technology 29(2019)565–570567
Su et al.propod a new method bad on k-Nearest Neighbor ca reasoning technology [50].The results of the prediction of a mining induced rockburst at a great depth in South Africa show that this method is feasible and reliable for rockburst prediction with high precision.
A Bayes discriminant analysis model is ud by Fu pre-dict the possibility and classification of rockbursts [51].Three fac-tors are the discriminating factors of the model.Rockbursts in Dongyu Mine and Pingdingshan deep development opening were predicted using this model.The predicted results are consistent with the obrved ones.
Cai bined a principal component analysis and fuzzy comprehensive evaluation model for coal burst potential asss-ment [1].The two methods can decrea the correlation of orig-inal data,which can avoid interaction among original data.
Zhou et al.summarized 12machine learning algorithms includ-ing artificial neural network (ANN),dist
ance discriminant analysis (DDA),support vector machine (SVM),Bays discriminant analysis (BDA),Fisher linear discriminant analysis (LDA),etc.,in long-term rockburst prediction and compared their prediction accura-cies [44].The algorithms ud different rockburst indicators as input features and their training samples sizes were different.Fig.3shows various accuracies obtained from 12machine learning algorithms.
The applied machine learning methods in rockburst predic-tion mainly focus on burst potential evaluation which can be regarded as long-term rockburst prediction.Some other scholars applied machine learning in microismic monitoring,a field mon-itoring method that can predict rockburst within a short-term.Microismic signals are critical evidence for rockburst occur-rence.However,many noi sources characterized by an abrupt amplitude,including human walking,passing vehicles,and espe-cially blasting,increasing give the appearance of microismic events [48].Hence,the first step to u microismic signal to pre-dict rockburst is extracting genuine rock microismic signals from received monitoring signals.
Zhao and Gross demonstrated how to u a support vector machine (SVM)to distinguish genuine microismic from noi events [48].16input attributes were extracted from 71original time-domain and frequency-domain features to train the SVM model bad on a dimensionality reduction method
called neigh-borhood component analysis (NCA).Four different kernel functions (linear,Gaussian,Quadratic,Cubic)were embedded in the SVM model to compare accuracy.However,SVM is a binary classifier which can only distinguish microismic events and non-microismic events.We can anticipate a multi-classifier to further classify noi events into more subclass such as walking noi,vehicle noi and blasting noi
Dong pared the three machine learning models (Fisher classifier,naive Bayesian classifier,and logistic regression)in dif-
ferentiating ismic events and blasts generate ismic waveform [52].The results showed that the logistic regression model had the best discriminating performance in the three mines.How-ever,databa from three mines were ud as training as well as testing.The generalization performance of the model might be doubtful.In other words,this model only guaranteed empirical risk minimization instead of structural risk minimization.
Shang et al.ud a BP neural network to distinguish rock mass fracturing signals and blasting vibration signals [53].A combined method:frequency slice wavelet transform (FSWT)plus singular value decomposition (SVD)was adopted to extract relevant infor-mation from original microismic si
gnals as input parameters for BP neural network.The results showed 86.67%of the signals could be precily identified.This BP neural model had 70training samples as well as 50test samples,which was not an optimal pro-portion.In general,the proportion between training samples and test samples should be 2:1.
Yıldırım et al.ud three different neural network models (feed-forward neural networks,adaptive neural fuzzy inference system,and probabilistic neural network)to discriminate between ismic events and quarry blasts [54].He found that the feedforward neu-ral network performs better than other two neural networks with a classification accuracy 99%against 96%for adaptive neural net-work and 97%for probabilistic neural network under a support of 175ismic events data.
5.Conclusion,discussion and future rearch
A significant number of lab tests and engineering projects show a highly nonlinear relationship between rockburst occurrence and corresponding control factors.Machine learning is a prospective way to simulate such relationship becau it does not need any prior knowledge about the nature of the relationship between the input/output variables,which is one of the benefits that machine learning has over most empirical and statistical methods.However,problems do exist with the application of machine learn-ing in rockburst prediction.
Most existing machine learning methods in rockburst predic-tion u a shallow machine learning model like SVM,decision tree,and logistic regression.Although the models enjoy computation convenience,they can only show relatively simple relationships between rockburst control factors and rockburst occurrence.Highly nonlinear relationship may not locate in the scope of shal-low machine learning models.In future rearch,deep learning can be involved in rockburst prediction.A multi-layer neural network can tackle any function with arbitrary precision,which ensures a more accurate relationship between rockburst control factors and rockburst occurrence.Furthermore,deep learning has demon-strated its superior performance with more supportive data.Fig.4compares performances between deep learning and tradi-tional shallow machine learning methods under different data sizes.Traditional machine learning methods perform
slightly
Fig.4.Performance comparison between traditional machine learning and deep
learning.
Fig.3.Rockburst classification accuracy from twelve machine learning algorithms [42].
568Y.Pu et al./International Journal of Mining Science and Technology 29(2019)565–570
authorware是什么better than deep learning with small data size.However,this advantage will totally rever when data size is large enough.
Another advantage of deep learning is automatic feature lec-tion.In general,when we apply machine learning to rockburst pre-diction,we must assign features for input vector.For example,in burst potential asssment with machine learning,some indicators in Table1are determined as features for training samples.How-ever,manual feature determination usually cannot reveal all char-acteristics of a problem.Feature engineering in deep learning is able to decide features automatically,create features and improve features,which is very helpful in microismic signal identification in short-term rockburst prediction.As of now,there is no rearch referring to this topic.
Bad on the machine learning principle‘Garbage in,garbage out’,the lection of training samples directly influences the suc-cess or failure of prediction.Some deficiencies exist in current long-term r
ockburst prediction with machine learning.Firstly,a common problem is lack of training samples.A small amount of training samples cannot feed enough features into a machine learning model,which leads to inferior prediction accuracy. Another problem for training samples is imbalance,which means there are significant differences among the number of labels in the training t.For example,if the expected prediction results are‘rockburst happen’and‘no rockburst’,an optimal training t should consist of50%‘rockburst happen’labelled samples and 50%‘no rockburst,labelled samples.However,the‘‘rockburst hap-pen”records in engineering projects are much less than‘no rock-burst’records,which usually results in an unbalanced training t.One common solution to solve training t imbalance is over-sampling for‘rockburst happen’samples and under-sampling for‘no rockburst’samples.
As a short-term prediction method,microismic monitoring can relatively predict the burst location and burst time,which sug-gests a hopeful prospect in rockburst control.However,current rearch about machine learning in microismic monitoring mainly focus on distinguishing burst signals from other nois. There is little rearch about the subquent steps on how to build a model between burst signals and rockburst occurrence.This job is pending for future rearch.The new breakthrough of rockburst prediction applying machine learning bad onfield monitoring may reside
in the monitoring signal anomaly detection.All types offield monitoring signal are expected to show anomalies before a real rockburst happens.If we can build the relation between sig-nal anomalies with rockburst,somehow can determine the rock-burst happen time.Now,anomaly detection with machine learning is a prospective technology with many propod methods. The introduction of this kind of technology may provide an alterna-tive method for rockburst prediction.
In summary,rearch about machine learning in rockburst pre-diction have made some achievements although someflaws exist in current rearch.Future achievements can be obtained if more advanced machine learning methods are involved in addressing this issue.
References
[1]Cai W,Dou L,Si G,Cao A,He J,Liu S.A principal component analysis/fuzzy
comprehensive evaluation model for coal burst liability asssment.Int J Rock Mech Min Sci2016;81:62–9.
[2]He J,Dou L,Gong S,Li J,Ma Z.Rock burst asssment and prediction by
dynamic and static stress analysis bad on micro-ismic monitoring.Int J Rock Mech Min Sci2017;
93:46–53.
[3]He M,Ren F,Liu D.Rockburst mechanism rearch and its control.Int J Min Sci
Technol2018;28(5):829–37.
[4]He MC,Miao JL,Feng JL.Rock burst process of limestone and its acoustic
emission characteristics under true-triaxial unloading conditions.Int J Rock Mech Min Sci2010;47(2):286–98.
[5]Pu Y,Apel DB,Lingga B.Rockburst prediction in kimberlite using decision tree
with incomplete data.J Sustain Min2018;17(3):158–65.
[6]Pu Y,Apel DB,Wang C,Wilson B.Evaluation of burst liability in kimberlite
using support vector machine.Acta Geophys2018:1–10.
[7]Dou L,Cai W,Cao A,Guo WC.Comprehensive early warning of rock burst
utilizing microismic multi-parameter indices.Int J Min Sci Technol2018;28
(5):767–74.
[8]Yang Z,Liu C,Tang S,Dou LM,Cao JL.Rock burst mechanism analysis in an
advanced gment of gob-side entry under different dip angles of the am and prevention technology.Int J Min Sci Technol2018;28(6):891–9.
[9]Blake W,Hedley DG.Rockbursts:ca studies from North American hard-rock
mines.SME2003.
[10]Mark C.Coal bursts in the deep longwall mines of the United States.Int J Coal
primeurSci Technol2016;3(1):1–9.
[11]Williams TJ,Brady TM,Bayer DC,Bren MJ,Pakalnis RT,Marjerison JA,et al.
Underhand cut andfill mining as practiced in three deep hard rock mines in the United States.Paper prented at the proceedings of the CIM conference exhibition,2017.
[12]Ellenberger JL,Heasley KA.Coal mine ismicity and bumps:historical ca
studies and currentfield activity.Paper prented at the proceedings of the 19th international conference ground control in mining,Morgan Town,2000.
[13]Baltz R,Hucke A.Rockburst prevention in the German coal industry.Paper
prented at the Proceedings of the27th international conference on ground control in mining.Morgantown,WV:West Virginia University;2008.
助理会计师考试时间[14]Potvin Y,Hudyma M,Jewell RJ.Rockburst and ismic activity in underground
Australian mines-an introduction to a new rearch project.Paper prented at the ISRM international symposium,2000.
[15]Wondrad M,Chen D.Application of mine ismicity monitoring technology in
mitigating geotechnical risks at Barrick’s Darlot Gold Mine.Paper prented at the Golden Rocks2006,The41st US symposium on rock mechanics(USRMS), 2006.
[16]Li T,Mu Z,Liu G,Du J,Lu H.Stress spatial evolution law and rockburst danger
induced by coal mining in fault zone.Int J Min Sci Technol2016;26(3):409–15.
[17]Shi Q,Pan Y,Li Y.The typical cas and analysis of rockburst in China.Coal Min
Technol2005;10(2):13–7(in Chine).
[18]Zhang J,Jiang F,Yang J,Bai W,Zhang L.Rockburst mechanism in soft coal am
建造师考试科目within deep coal mines.Int J Min Sci Technol2017;27(3):551–6.
[19]Bennett TJ,Marshall ME.Identification of rockbursts and other mining events
using regional signals at international monitoring system stations.Sci Appl Int Corp Mclean VA2001.
[20]Zhang J,Fu B.Rockburst and Its criteria and control.Chin J Rock Mech Eng
2008;27(10):2034–42(in Chine).
[21]Butt SD,Apel DB,Calder PN.Analysis of high frequency microismicity
recorded at an underground hardrock mine.Pure Appl Geophys1997;150(3–
4):693–704.
[22]Wang J,Zhang J.Preliminary engineering application of microismic
monitoring technique to rockburst prediction in tunneling of Jinping II project.J Rock Mech Geotech Eng2010;2(3):193–208.
in
[23]Frid V,Rabinovitch A,Bahat D.Fracture induced electromagnetic radiation.J
Phys D Appl Phys2003;36(13):1620.
[24]Petukhov I,Lin’kov L.Mechanics of rock bursts and outbursts.Moskva:Nedra;
1983(in Russian).
[25]Fajklewicz Z.Application of microgravimetry method to detection of
subsurface cavities and prediction of rock bursts-advances in coal geophysics EAG,Hyderbad,India;1988.p.1–11.
[26]Zhang Y,Li J,Liu X,Tian B.Infrared radiation portentous characteristics of rock
burst in roadway.J Min Saf Eng2015;32(5):7(in Chine).
[27]Kidybin´ski A.Bursting liability indices of coal.International journal of rock
mechanics and mining sciences&geomechanics abstracts,vol.18,no.
4.Pergamon;1981.p.295–304.
[28]Wattimena RK,Sirait B,Widodo NP,Matsui K.Evaluation of rockburst
potential in a cut-and-fill mine using energy balance.Int J JCRM2012;8
(1):19–23.
[29]Altindag R.Correlation of specific energy with rock brittleness concepts on
rock cutting.J South Afr Inst Min Metall2003;103(3):163–71.
[30]Wang JA,Park H.Comprehensive prediction of rockburst bad on analysis of
strain energy in rocks.Tunn Undergr Space Technol2001;16(1):49–57. [31]Mitri H,Tang B,Simon R.FE modelling of mining-induced energy relea and
storage rates.J-South Afr Instit Min Metall1999;99(2):103–10.
[32]Kabwe E,Wang Y.Review on rockburst theory and types of rock support in
rockburst prone mines.Open J Saf Sci Technol2015;5(04):104.
ship
[33]McCulloch WS,Pitts W.A logical calculus of the ideas immanent in nervous
activity.Bull Math Biophys1943;5(4):115–33.
[34]Ronblatt F.The perceptron:a probabilistic model for information storage
and organization in the brain.Psychol Rev1958;65(6):386.
[35]Hubel DH,Wiel TN.Receptivefields,binocular interaction and functional
advice
architecture in the cat’s visual cortex.J Physiol1962;160(1):106–54.obvious是什么意思
[36]Rumelhart DE,Hinton GE,Williams RJ.Learning internal reprentations by
error propagation(No.ICS-8506).California Univ San Diego La Jolla Inst for Cognitive Science;1985.
[37]LeCun Y,Bor B,Denker JS,Henderson D,Howard RE,Hubbard W,et al.
Backpropagation applied to handwritten zip code recognition.Neural Comput 1989;1(4):541–51.
[38]Cortes C,Vapnik V.Support-vector networks.Mach Learn1995;20(3):273–97.
[39]Freund Y,Schapire RE.Experiments with a new boosting algorithm.Paper
prented at the ICML,1996.
Y.Pu et al./International Journal of Mining Science and Technology29(2019)565–570569
