Sie sind auf Seite 1von 7

Tunnel Blast Design using Artificial Neural Network a Case Study

V M S R Murthy, Member K Dey, Non-member R R Chimankar, Non-member


Artificial intelligent research has produced several tools for commercial applications. Some of the techniques that are widely used today include neural network, fuzzy logic and expert systems. Artificial neural network (ANN) is an excellent predictive and data analysis tool. In the mining industry, ANN techniques are being used commercially for real-time process-control applications. Blast design in tunnel construction, is still accomplished on trial and error basis which is not only time consuming but also leads to sub-optimum results in many situations. In order to design blasts based on field data and standardize tunnel blasting pattern, development of an automated design program is necessary. Thus, in this study an attempt has been made to develop a new method using Artificial Intelligence (AI), such as Artificial Neural Network (ANN), based on the data generated carefully from a metal mine. To serve as a general-purpose design model more data needs to be included.This paper presents the application of standard back propagation algorithm for tunnel blast design using parallel hole cut. The developed ANN model, after training with actual field data, is used to design different tunnel blast design parameters. Performance of the model has been evaluated by comparing the values obtained using the ANN model and actual values observed in the field. Results indicate that the average error in prediction of blast design and performance parameters is less than five per cent1. Thus, the developed neural network model can be applied for intelligent tunnel blast design.
Keywords : Artificial neural network; Back propagation; Tunnel blasting; Parallel hole cut; Blast design

INTRODUCTION Tunnelling is an important activity in the field of underground construction. Tunnels are driven for road and rail transportation, sewer and water supply, gaining access to underground hydroelectric power projects, mountain caverns for industrial, recreational and storage purpose. Drilling and blasting technique, owing to its capability to meet wide variations in geology, is predominantly used to excavate majority of the tunnels. Therefore, proper blast design in tunnels is the key to enhance blasting efficiency and reduce damage to the tunnel wall, vibration and noise level. Important tunnel blast design variables are burden, spacing, drilling dia, empty hole dia, face advance and tunnel area2. In order to standardize the tunnel blasting pattern, development of an automated design program based on field data, which serves as knowledgebase, is necessary. This article reports the development of one such self-learning tunnel blast design3 method based on Artificial Intelligence (AI), namely, Artificial Neural Network (ANN). The ANN networks do not require an explicit functional model for relating system variables. It is therefore, expected that limitations associated with traditional method may be overcome by using ANN. Additionally, due to their robustness to imperfect data and because of their capability of modelling complex data set ANN presents an attractive alternative to the traditional designing techniques. TUNNEL BLAST DESIGN Tunnel blast design, in recent times, is being accomplished by parallel hole cuts for achieving higher pulls. The blasts in tunnels and drifts are characterized by lack of adequate free surfaces towards which
V M S R Murthy, K Dey and R R Chimankar are with Mining Engineering Department, Indian School of Mines, Dhanbad 826 004. This paper (re-modified) was received on October 05, 2005. Written discussion on this paper will be received untill April 29, 2006. Contour Holes

Roof Holes

Stoping Hole

Cut Cut Spreader Hole Lifters

Wall Hole

Floor Holes Figure 1 Different zones in tunnel blasting

breakage can occur effectively. The principle behind tunnel blasting is to create an opening by means of a cut ( a set of holes that provide initial free face) and then stoping to enlarge the opening. The cut, usually, has a surface area of 1m2 2m2, although with large drilling dia holes it can reach up to 4m2. The different zones in tunnel blasting4 are shown in Figure 1. Stoping can geometrically be compared to bench blasting although it requires powder factors that are four to ten times higher5. Such a high explosive consumption is mainly due to drilling error, the demand made by swelling, the absence of hole inclination, the lack of cooperation between adjacent charges and also blasting against gravity in case of lifter holes. Contour holes are those, which establish the final shape of the 39

Vol 86 , February 2006

tunnel and are spaced closely (0.2m 0.4m apart) and directed outwards to make room for the drill in collaring and advance. The position of the cut has influence on rock projection, fragmentation and also on the number of blast holes. Of the three positions, namely, corner, lower centre and upper centre, the latter is usually chosen as it avoids the free fall of the material, the profile of the broken rock is more extended, less compact and better fragmented. Design of Parallel Hole Cuts The blasts in tunnels and drifts are much more complex than bench blasting owing to the fact that the only free surface is the tunnel heading. The powder factors are elevated and the charges are highly confined. On the other hand, burdens are small, which require sufficiently insensitive explosive to avoid sympathetic detonation and at the same time have enough detonation velocity (above 3000m/s) to prevent channel effect in the cartridge explosive placed in large dia blast holes. This phenomenon consists of the explosion gases pushing the air that exists between the column charge and the wall of the blasthole, compressing the cartridges in front of the shock wave, destroying the hot spots or excessively increasing the density of the explosive6. Drilling has become more mechanized in the last decade with the introduction of fully automatic and computer assisted hydraulic jumbos with multiple booms. Thus, there is a general shift towards parallel hole cuts as they are easier to drill, do not require a change in the feed angle and the advances are not as conditioned by the width of the tunnels, as happens with angled cuts7. Some of the problems that can arise in blasting with parallel blasthole cuts are sympathetic detonation and dynamic pressure desensitization. The first phenomenon can appear in adjacent holes to the detonating hole when the explosive used has a high degree of sensitivity, such as, all those with nitroglycerine in their composition. On the other hand, the dynamic pressure desensitization takes place in many explosives and, especially, in ANFO. This is due to the compression of explosive charge ahead of the shock wave and resultant increase in the density of the adjacent charge above the critical density. Desensitization problem can be attenuated by correctly designing the initiation sequence, sufficiently delaying the successive detonation of each blasthole so that the shock wave from the last shot disappear, allowing the explosive to recuperate its normal density and degree of sensitivity. Hagan8 suggests that, in order to minimise these problems, the parallel hole cut may be carried out by placing three relief holes in such a manner that they act as a shield between the charged holes. Hagan has proved that fine-grained rock is more susceptible to cut failure than coarse grained, due to the larger volume of relief opening that is needed for the expulsion of the material. BACK-PROPAGATION ARTIFICIAL NEURAL NETWORK Neural networks are simplified model of the biological structure found in human brains. Their layered structure is composed of a large number of interconnected elementary processing elements to mimic the biological neurons. The characteristics of a neural network come from the activation function and connection weights. Since the neural network stores data as pattern in a set of processing elements 40

by adjusting the connection weights, it is possible to realize complex mapping through its characteristics of distributed representations. The neural network can find automatically the closest match through its content-addressable property, even if the data are incomplete or vague9. Even in the case that a few processing elements malfunction or fail completely, the network can still function through its faulttolerance attribute. The neural network has the ability for extracting a generalized correlation (or regularity) from many individual examples (or experiences). Neural network models are set up by learning and training. If a network model is trained with a large number of inputoutput pairs (or pairs of input and target vectors), it can produce an appropriate output for untrained inputs. More than 50 neural network models have been devised so far, and it is found that the back-propagation learning algorithm based on the generalized delta rule proposed by Rumelhart, et al 10 is the most popular and efficient learning procedure for multi-layer neural networks. The back-propagation neural network generally consists of many sets of nodes arranged in layers (for example input, hidden and output layer) as shown in Figure 2. The output signals from one layer are transmitted to subsequent layer through links that amplify or attenuate or inhibit the signals using weighting factors. Except for the nodes in the input layer, the net input to each node is the sum of the weighted output of the nodes in the prior layer. An activation function, such as, sigmoidal logistic function is used to calculate the output of the nodes in the hidden and output layers. In the calculation, both the input and output are usually normalized to give a value between 0 and 1. The number of hidden layers and the number of nodes in each layer depends on the complexity of the pattern and the nature of the problem to be solved. The back-propagation neural network operates in two different phases: one for learning and the other for production. The learning phase is composed of forward and reverse passes. In the forward pass, a set of learning pattern (or training pairs) is presented to the system and the system calculates the output from the input patterns. In the reverse pass, the system calculates the sum-squared error for each training pair, comparing the output with the target pattern. If the error summed overall learning pattern converge within a certain limit, the learning phase is ceased. Otherwise, the system follows the error-back propagation process where the calculated error signals
1 1

Oi Input Layer (i)

Wji

Oj Hidden Layer (j)

Wkj

Ok Output Layer (k)

Ep

tk

Figure 2 Architecture of back propagation in artificial neural network

IE(I) JournalMN

Blast Design Database

propagate backward through the network and adjust the connection weights (or strengths). In completion of this learning procedure, the weights are stored in the neural network for next production phase. In production phase, the network produces appropriate output pattern for newly presented input patterns. DEVELOPMENT OF INTELLIGENT TUNNEL BLAST DESIGN The artificial neural network described above is applied to tunnel blast design. A neural network model has been developed using Neural Planner Software(Easy NN Plus). The Neural Planner is a menu-driven system in which three layers for input, hidden, and output can be developed. It has the capability to interact with the database developed in MsAccess/MsExcel for training purpose. A large number of data prepared on MS Excel have been prepared and the entire set is available with the authors. The developed network consists 31 inputs, 29 hidden(three layer system) and seven output layers. The input parameters consist of tunnel area, relief hole dia, blast hole dia, P-wave and the burden, charge per hole and number of holes in different holes. The output parameters are vibration, volume of rock, pull, overbreak, powder factor, detonator factor and specific drilling. The procedural steps of the working of the model are shown in Figure 3. Initially, a database of the previous blast records has been developed from the actual blasts conducted in mines11. This database contains a sufficient number of blast datasets so that during training the model can converge. Thus, the model developed in Neural Planner software has been trained. During training, software assigns

Selection of Important Parameters

Selection of Training Cases

Blast Design Calculation by Conventional Method

Processing of Data

Initial Estimation of Weight


Forward Calculattion of Network

Adaptive Revision of Weights

Yes

Whether a Reasonable error for all the Training Cases Obtained ?

No

Simulated Weights Testing Estimation

Determination of Different Output Figure 3 Flowchart showing the sequence of operation of the model

IX

IX

IX

IX

IX

IX

IX

IX

800

VIII

VI

V III 2 II IV 8 III

IV II 1

V III

VI

VII

VIII
800 3200

VIII

VI

II VI 5 II IV X X III V

VI

VII

VIII

800

VIII

VI

VI

VII X

VIII
800

All Dimension in mm Not to Scale 5000 700 720 720 720 720 720 700 Reamer Hole Charge Hole

1, 2,... represent short delay, I, II,... represent long delay

Figure 4 Blast pattern for drift with mechanised drilling

Vol 86 , February 2006

41

Table 1 Charging and delay pattern for blasting in mechanised drift (4.5 m x 3.2 m) Hole(s) Delay Number* Number of Charge/Hole Total Charge Holes Center hole First square 0 1, 2, 5, 8 1 4 8 8 9 6 8 8 52 (Cartridge) 10 + 1P 10 + 1P 11+ 1P 11 + 1P 12 + 1P 11+1P 10 + 1P 12 + 1P (Cartridge) 11 44 96 96 117 72 88 104 628 Dia of cartridge : 32 mm Weight of cartridge : 0.220 kg Total explosive : 138.16 kg Total Yield : 157.5 t (expected)

Table 2 Laboratory test results on rock samples Rock Property Rock density, R Q D, % RMR Q Index UCS, MPa Tensile strength, MPa Youngs modulus, GPa P-wave velocity, km/s S-wave velocity, km/s t/m3 Value 2.9 81.67 66 5.11 77.64 10.27 28.66 4.50 6.10 2.50 3.50

Second square II 4, III 4 Third square IV 4, V 4 Easers Side holes Top holes Bottom holes Total Total number of holes : 52 + 4 R Depth of round : 3.2 m Dia of blasting holes : 38 mm Dia of reamer holes : 64 mm Total number of cartridges : 628 VI 6, VII 3 VIII 6 IX 8 X8

analysis of the model results are discussed in the following paragraphs. GENERATION OF FIELD DATA Field visits were made to collect representative data to generate actual dataset on both input and output parameters pertaining to tunnel blast design for training, validation and testing of the model developed using ANN. Trial blasts were done in horizontal drifts of a metal mine located in eastern India for this purpose where parallel cut is being practiced on a large scale (Figure 4). The details of charging and delay pattern for blasting in mechanized draft (4.5m x 3.2m) is given is Table 1. Parallel cut demands a careful blast design with the best possible delay sequence to achieve satisfactory blast results. The rock type was chlorite-sericite-schists of massive metamorphic formation. Some of the geo-technical studies conducted are given in Table 2. The mine has both mechanised and manual faces. The field
Mechanised face 53.2 m

* 0,1,2,3,.. : Short delays ; I, II, III, : Long delays

a weightage to various inter-related parameters and attempts to limit the error. This process is repeated until the error converges to the set limits. The final weightages are obtained after training. The model has been tested using query process and the data predicted with the trained ANN model and actual data compared has yielded acceptable correlation. This has been done for randomly selected data one each for mechanized and manual drilling and blasting. The details of the field investigations conducted in the mines and the
Table 3 Details of mining sub-systems in mechanized face Parameters Face size Dia of blasthole, mm Dia of reamer hole, mm Drilling Number of reamer holes Drilling length, m Machine used for drilling

38 64 4 3.2 Jumbo Drill (4 nos) Explosive used: Powergel 801, Nobel gel, Belmx, Indorock

Explosive and detonator used Blasting Short and long delay used Mucking Loading and transportation Transportation

Short and long delay detonators manufactured by Indian Explosive Ltd are used. Each increment in short delay number increase a delay time of 25ms whereas for long delay it is 300ms. Charging pattern is given in Table 4. Pattern is shown in Figure 4. LHD and Scoop Tram Mine truck of 25t capacity or low profile dump truck of 10t capacity dumped in ore pass or directly in stope for filling.

The suggested support system used in the mine is rock bolting. Rock bolts are used as the permanent support for the drifts and declines and as well as for raise and winze. For drift/decline: 1.6m 1.6m Support system Length of bolt : 1.6 m Shotcrete/grouting mixture: Strength of bolt : - 16t direction of bolt is perpendicular to dip of rock. 32 mm dia with twisted surface. 1:1:0.5 (cement: sand : water) Maximum distance of row of support from face : 2.5m

Large permanent excavation/junctions : 1.2m 1.2m

42

IE(I) JournalMN

tbdann1.tvq 485 cycles. Target error 0.0500 average training error 0.055826 The first 31 of 31 Inputs in descending order. Column Input Name Importance 26.1111 18.2798 12.2427 12.2106 11.8021 11.7615 11.5500 11.4684 9.9268 9.6952 9.6333 9.5208 9.3074 9.1327 8.9300 8.8474 8.8163 8.7882 8.5337 8.5087 8.3091 7.9535 7.8870 7.8613 7.7236 6.9571 6.7330 6.4771 5.7552 5.6257 3.6936 Relative Error

Figure 5 The neural network for tunnel blast design (input, three hidden and output layers) tbdann1.tvq Normalised Error 1.0000 0.9000 0.8000 0.7000 0.6000 0.5000 0.4000 0.3000 0.2000 0.1000 37
Maximum Average Minimum Learning rate : 0.600000 Momentum : 0.800000 Maximum error : 0.086093 Average error : 0.055826 Minimum error : 0.021437 Target error : 0.050000 Validating examples : 0 Not validating or scoring

75 112 150 188 225 270 346 420 485

Layer : Input Hidden 1 Hidden 2 Hidden 3 Output Nodes 31 14 8 7 7 Weights 434 112 56 49

23 Qe 12 N2 3 P-wave 9 N1 5 Hole length 15 N3 2 Blast hole DIA 10 B2 1 Relief hole DIA 7 B1 26 Qr 24 Ne 25 Br 13 B3 30 NI 27 Nr 32 Scaled dist 4 Number of holes 16 B4 31 Total charge 22 Be 17 Q4 14 Q3 0 Area 19 B5 21 Q5 28 BI 29 QI 8 Q1 6 Number of reamer 11 Q2

Figure 8 Relative importance of input parameters. Table 4 Comparison of ANN predicted and actual results

Figure 6 Normalised error against the iterating cycles (with layers, nodes and weights) tbdann1.tvq cycle 485. Target error 0.0500 Average training error 0.055826 The first 14 of 14 example rows in descending order Above target Below target Row 5 4 8 13 15 6 14 12 7 9 2 11 3 10 Example Mechanised Mechanised Manual Mechanised Mechanised Mechanised Manual Mechanised Mechanised Mechanised Mechanised Mechanised Mechanised Manual Normalized Error (0-1) 0.032827 0.030889 0.030322 0.029939 0.028391 0.027595 0.020672 0.015072 0.015050 0.011680 0.010460 0.008478 0.003391 0.000011 Relative Error

Parameter

Mechanised Face Predicted Actual 29.3 72.00 1.40 15.00 0.72 1.60 0.47

Manual Face Predicted 251.83 48.00 1.23 19.00 1.204 0.84 0.54 Actual 254 50.00 1.3 20.00 1.24 0.83 0.55

Vibration, mm/s Muck Volume, t Pull, m Overbreak, % Powder factor, t/kg Detonator factor, t/det Specific drilling, t/m

48.98 73.00 1.49 15.00 0.73 1.58 0.49

ANALYSIS OF ANN MODEL OUTPUT Blasting data, pertaining to a metal mine of eastern India, have been collected for the training of the software. All the 14 data sets generated, have been used for training the ANN model and two data sets were put in the Query mode for testing purpose. For learning purpose following information was provided to obtain the desired results Learning rate : 0.60

Figure 7 Errors case-wise after training the ANN model

investigations have been carried out both in the mechanized and manual faces of the mine. The unit operations carried out, such as, drilling, blasting, mucking and transportation, in the mechanized face are described in Table 3. Vol 86 , February 2006

Momentum : 0.80 Target error : 0.05 Cycle/refresh : 100. 43

tbdann1.tvq 485 cycles. Target error 0.0500 average training error 0.055826 The first 31 of 31 inputs in descending order. Output column 33 vibration (PPV)
ANN Predicted Value

300 250 200 150 100 50 0 0 50

y =1.008 + 0.38, R2 = 1

Press refresh to update Column Input Name 12 23 10 16 31 13 7 30 5 3 26 19 32 2 24 4 9 1 6 22 21 27 14 8 11 29 25 17 0 28 15 Change from to Sensivity 0.00444162 0.00235918 0.00150934 0.00109052 0.00099188 0.00095354 0.00087944 0.00080367 0.00065139 0.00063096 0.00060552 0.00058311 0.00048237 0.00046502 0.00040134 0.00036149 0.00031301 0.00031030 0.00021434 0.00020951 0.00020773 0.00019278 0.00016050 0.00015771 0.00009467 0.00008296 0.00008229 0.00006087 0.00002802 0.00001584 0.00000071 Relative Sensitivity

N2 2 4 Qe 1.8000 2.4000 B2 14 20 B4 17 40 Total 38.0000 131.0000 charge B3 17 31 B1 8 15 NI 5 8 Hole 1.3000 3.2000 length P-wave 4300 5700 Qr 0.5000 2.2000 B5 29 50 Scaled 5.3100 16.7400 Dist Blast 32 38 Hole Dia Ne 9 28 No of Holes 30 64 N1 1 5 Relief Hole 32 89 Dia No of 1 5 Reamer Be 40 58 Q5 0.6250 2.4000 Nr 5 8 Q3 0.6250 2.8000 Q1 0.6250 2.2000 Q2 0.6250 2.2000 QI 0.7500 2.8000 Br 45 68 Q4 0.6250 2.6000 Area 12.0000 16.0000 BI 55 85 N3 2 4

100

150 Actual Value

200

250

300

Figure 11 Validation of actual and predicted values for manual face

80 70 60 50 40 30 20 10 0 0

y =1.88 - 0.53, R2 = 0.94

20

40

60

80

Figure 12 Validation of actual and predicted values for mechanized face

Figure 9 Relative sensitivity of inut parameters tbdann1.tvq 1 485 cycles. Target error 0.0500 average training error 0.055826 Output columns (min to max values) 33 34 35 36 37 38 39 Vibration (PPV) (8.4800 to 576.000) Volume (46 to 118) Pull (1.2000 to 2.4000) Ovrebreak (12 to 20) Powder Factor (0.6700 to 1.2400) Detonator Factor (0.8200 to 2.6400) Sp Drilling (0.4500 to 0.7600)

(Figure 5), normalized errors(Figure 6), errors case-wise(Figure 7) after training the ANN model, relative importance of input parameters(Figure 8), relative sensitivity (Figure 9) of input parameters and predictions of output for training examples (Figure 10) have been arrived at using the Easy NN Plus Neural Network Planner. For assessing the predictability of the ANN model query data sets were used and the predicted output was compared with the actual values observed in the field. A comparison has been given in Table 4. One can infer from the analysis of errors that the blast design results obtained from ANN model were acceptable. The proposed model can be refined further by feeding more varied data to make it a close to general purpose ANN Model. CONCLUSION Proper tunnel blast design is mandatory to achieve faster drivage rate in order to minimize development cost in mining and tunnelling work. Blast design includes a number of inter-dependent parameters, which make it more complex to design an optimum blast. An ANN model has been developed on the theory of back propagation learning method. Altogether 14 numbers of blast data have been collected from a metal mine located in eastern India. It has been found that for the prediction of number of holes, length of hole, different burdens for cut, stoping and roof holes, powder IE(I) JournalMN

14 training examples

Figure 10 Predictions of output for the training examples

The blast design input has been given to the trained network model and the output was obtained at the end when the neural network model stopped after 485 cycles. The ANN network developed 44

factor, pull, total charge, maximum charge, etc the model could predict the output well within the range of five percent error. However, for parameters like vibration it needs modifications. Probably more data sets could find a closer prediction in these values. The database of the software is of self-learning type and over the time with large number of data it can act as an expert designer. ACKNOWLEDGMENT The authors are thankful to mine authorities for permitting to undertake the investigations. Authors also acknowledge the support of Indian School of Mines (ISM) for pursuing the research and utilize the facilities of the School. This paper reports a part of the investigations carried out under a MHRD sponsored project to ISM. The authors gratefully acknowledge MHRD for the same. REFERENCES
1. R R Chimankar. 'Development of an Intelligent Software for Tunnel Blast Design using Artificial Neural Network (M Tech Dissertation)'. Indian School of Mines, Dhanbad, December 2002, p 48. 2. U Langefors and B. Kihlstrom. 'The Modern Techniques of Rock Blasting'. John Wiley and Sons, 1973, p 23.

3. B R Yama and G T Lineberry. 'Artificial Neural Network Application for a Predictive Task in Mining'. Mining Engineering, February 1999, p 59. 4. C L Jimeno, E L C Jimeno and J A Francisco. 'Drilling and Blasting of Rocks'. A A Balkema / Rotterdam/ Brookfield, 1995, p 217. 5. R Gustafson. 'Swedish Blasting Technique'. SPI, 1997, p 23. 6. R Holmberg. 'Charge Calculation for Tunnelling'. Underground Mining Methods Handbook, AIME, 1982, p 73. 7. J Johansen. 'Modern Trends in Tunneling and Blast Design', Rotterdam, 2002, p 42. 8. T N Hagan. 'Larger Diameter Blastholes A Proposed Means of Increasing Advance Rates'. Fourth Australian Tunnelling Conference, Melbourne, 1981, p 283. 9. V Rao and H Rao. 'C++ Neural Networks And Fuzzy Logic'. BPB Publications, New Delhi, 1996, p 104. 10. D E Rumelhart and J A MeClelland (eds). 'Parallel Distributed Processing. Learning Internal Representation by Error Propagation' by D E Rumelhart, G E Hinton and R J Williams'. Explotion in the Microstructure of Cognition, 1986, p 318. 11. V M S R Murthy and K Dey. 'Development of Predictive Models for Blast Induced Damaged Assessment (BIRD) in Tunnels'. Interim Report, MHRD, Project No: MHRD, (27)/99-00/111/ME, 2002, p 12.

Vol 86 , February 2006

45

Das könnte Ihnen auch gefallen