Arabian Journal of Geosciences (2022) 15: 173

[Link]

ORIGINAL PAPER

Predicting blast-induced pull using regression tree

Aditya Rana1 · Narayan Kumar Bhagat1 · Atul Singh1 · Pradeep Kumar Singh1

Received: 25 September 2020 / Accepted: 6 January 2022 / Published online: 15 January 2022
© Saudi Society for Geosciences 2022

Abstract
The tunnel advance (pull) per blast considerably influences the time and expenditure incurred to construct tunnels using
conventional drill and blast methods. Blasts with high pull efficiency can considerably lessen the construction time by reduc-
ing the number of drilling, blasting and mucking cycles. Researchers have developed some empirical and theoretical models
for determining pull. However, these models are based upon hit and trial techniques. The previous works on the similar
subject suggest that the pull efficiency cannot be correlated to a variable. Nevertheless, the literature suggests that the soft
computing techniques can be used for solving such non-linear problems. The present study aims to develop a classification
and regression tree (CART) model to predict the pull using various blast design parameters, such as cross-sectional area,
number of blast holes, hole diameter, hole depth, charge per hole, maximum charge per delay, total charge and charge factor.
A dataset comprising of 100 data points was compiled from tunnelling sites of Karcham Wangtoo Hydroelectric Project
for model formulation. The dataset was divided in a ratio of 80:20 for training and testing, respectively. The coefficient of
determination (R2) and root mean square error (RMSE) of the CART model and multivariate regression analysis were com-
pared for the validation. The suitability of the CART model was further verified by conducting statistical hypothesis tests.

Keywords Tunnel · CART​· Regression tree · Drill and blast · Actual pull

Abbreviations Introduction
ANOVA Analysis of variance
CART​ Classification and regression tree Nowadays, almost all infrastructure projects, such as trans-
DBM Drill and blast method portation, irrigation, and hydel, in hilly and mountainous
HEP Hydroelectric Project terrains are constructed by tunnelling. Additionally, tunnels
HRT Head race tunnel are also excavated for the mining and the creation of under-
MCPD Maximum charge per delay ground spaces (Jimeno et al. 1995). Depending upon rock
MAT Main access tunnel types and hydrological conditions, different techniques are
MLR Multiple linear regression adopted for tunnel excavations. Commonly, drill and blast
MVRA Multivariate regression analysis method (DBM), tunnel boring machine (TBM) and road-
PPV Peak particle velocity headers are used for tunnel excavations. The applicability of
RMR Rock mass rating TBM is limited to bore circular tunnels only, whereas DBM
RMSE Root mean square error can yield tunnels of varying profiles, such as horseshoe, rec-
SCCL Singareni Collieries Company Limited tangular and D shape. Sometimes, due to the variation in
TRT​ Tail race tunnel geology and presence of folds, faults and shear zones, engi-
TBM Tunnel boring machine neers refrain from using TBM (Goel 2015). Furthermore,
USBM United States Bureau of Mines it is difficult to transport and assemble the assets involved
in TBM. However, easy portability of components used in
DBM along with low initial cost of investment results in
Responsible Editor: Murat Karakus preferability of DBM over TBM. Moreover, TBM is gener-
ally preferred for the excavation of long tunnels (kilometric
* Aditya Rana or multi-kilometric). However, in short length tunnels, the
adityarana@[Link]
installation time for TBM can exceed the excavation time.
1
Rock Excavation Engineering, CSIR-Central Institute Hence, DBM is a preferred excavation technique for short
of Mining and Fuel Research, Dhanbad, Jharkhand, India

13
Vol.:(0123456789)

Download fulltext IDR 3K - [Link]

173 Page 2 of 10 Arab J Geosci (2022) 15: 173

length tunnels. The results obtained from DBM in tunnel- = 0.83 and RMSE = 0.33) for forecasting ground vibrations.
ling are generally evaluated by parameters: (i) blast-induced Recently, Rana et al. (2020) compared decision tree, artifi-
vibrations; (ii) blast-induced under-break, over-break or rock cial neural network and multivariate regression techniques
mass damage; and (iii) pull per round blast. The pull effi- for predicting blast vibrations in Indian tunnels. They con-
ciency is defined as the ratio of face advance achieved post- cluded that the decision tree precisely predicts vibrations.
blast to the length of the rock mass drilled in a blast round. The decision tree helped them in ascertaining that near-
It is generally expressed in percentage. The expenditure field and far-field vibrations are predominantly affected by
and time incurred to excavate a tunnel are inversely related total charge fired in a round and maximum charge per delay
to pull. Attainment of an optimum pull efficiency reduces (MCPD), respectively.
the number of mucking, drilling and blasting cycles, hence,
resulting in considerable time and cost savings.
In a recent study, the relationship between pull efficiency Site description and data
and different blast design parameters (tunnel cross section,
cut-holes and specific drilling) was examined using Pear- Karcham Wangtoo Hydroelectric Project (KWHEP) is a
son’s correlation coefficient (Cardu and Seccatore 2016). 1000 MW hydro-power project situated on Sutlej River,
The investigators concluded that the pull efficiency could Himachal Pradesh, India. The project is located 47 km
not be linearly correlated to a variable. Furthermore, they north-east of Jhakri City. The rock mass at the KWHEP
recommended that designing a tunnel blast for an optimum is predominantly composed of granitic gneiss with minor
progress is a complex task as many factors are involved. occurrences of augen, micaceous gneiss and gneissic schist.
The unavailability of a mathematical or numerical model The rock mass was classified as fair (class III) following
co-relating all the factors for an effective progress further Barton’s classification. The Table 1 presents the details of
complicates the scenario. In the recent years, many research- rock mass, shape and area of different excavations investi-
ers (Monjezi et al. 2013; Sayadi et al. 2013; Saadat et al. gated in the present study. The CART model was designed to
2014) have proposed the use of machine learning (data assist the engineers in finalizing the blast design for attaining
mining) for predicting blasting nuisance such as vibration, an optimum pull. A dataset comprising of 100 data points
air overpressure, over-break and back-break. Regression was collected for model generation during the excavation of
analysis is a machine learning technique that develops a adit to diversion tunnel, diversion tunnel and main access
relationship between inputs and outputs by automated learn- tunnel. Laser profiler surveying was deployed to measure the
ing for explicit prediction of a new sample. Many regres- pull after each blast. Parallel cut as shown in Fig. 1 was used
sion analysis techniques such as linear regression, multi- for the excavations. The advantage of parallel hole cut is that
layer perceptron, multivariate adaptive regression splines, trials for higher pulls can be conducted by just increasing the
k-nearest neighbour, automated learning of algebraic models hole depth. In a rational tunnel driving, an unchanged drill-
for optimization, Kriging and classification and regression ing pattern shall be applied in an excavation for a high pull
tree (CART) are reported in literature. CART constructs a efficiency (Langefors and Kihlström 1978). Hence, only the
regression or decision tree for determining the influence controllable parameters (Rana et al. 2020) were considered
of inputs on output. It is a non-parametric technique capa- for developing the models (cross section of tunnels, no. of
ble of handling highly skewed data. In recent years, few holes, hole diameter, the average depth of holes, the aver-
investigations have deployed CART models for predicting age charge per hole, MCPD, total charge and charge factor).
blast-induced ground vibrations in mining. Hasanipanah and Cartridges containing emulsion explosives (Table 2) were
Shirani (2016) compared the performance of CART mod- used during the investigation.
els with multiple regression (MR) and empirical equations. The inter-hole and inter-row connections were provided
The models were trained and tested using 69 and 17 data using non-electric detonating system. The blast was deto-
points, respectively. The performance indices suggested that nated using one electric blasting cap no. 8. The ignition
CART (R2 = 0.95 and RMSE = 0.17) outperformed MR (R2 sequence for a tunnel blast is clearly defined in the literature.
= 0.88, RMSE = 0.286) and USBM predictor equation (R2 Firstly, a central opening is created. This central opening

Table 1  Rock types, rock mass Excavations Rock types Rock mass Cross-section Shape
class, shape and cross section of class area ­(m2)
different excavations
Adit to diversion tunnel Jointed augen and micaceous gneiss III 43.70 D
Diversion tunnel Blocky augen and micaceous gneiss III 86.83 D
Main access tunnel Granitic gneiss with subordinate schist III 64.50 D

13

Download fulltext IDR 3K - [Link]

Arab J Geosci (2022) 15: 173 Page 3 of 10 173

(a) Adit to diversion tunnel (b) Diversion tunnel

(c) Main access tunnel

Fig. 1.  Parallel cut pattern adopted for drilling and blasting in a adit charged (45 mm) and dummy (89 mm) holes, respectively. Roman
to diversion tunnel, b diversion tunnel, and c main access tunnel of numbers indicate delay sequences. Arabic numbers indicate dimen-
Karcham Wangtoo HEP. Note: Black and hollow bullets represent sions in metre

is further widened using the principle of rectangularity

Table 2  Specifications of explosives used at investigated sites and approximating 90֯ breakage. Generally, the use of long
delay detonators is preferred in tunnel blasting. This prin-
Explosive Manufac- Diameter Weight VoD Density ciple was followed in toto for designing all the blasts. The
turer (mm) (g) (mm/s) (g/cc)
delay sequences were randomly used in the blasts as per their
Ideal Ideal 40 395 4500 1.20 -1.25 availability at the site. The rock mass classification (rock
power ±250 mass class III) and properties of explosives were similar for
90
the studied excavations, and the pull obtained from parallel

13

Download fulltext IDR 3K - [Link]

173 Page 4 of 10 Arab J Geosci (2022) 15: 173

cut is less dependent on the rock mass unlike burn cut. Gen- four-section cut, three-section cut, double spiral cut, etc.)
erally, the maximum advance per round of angled cuts is and concluded that double spiral cylindrical cut yields high
influenced by width of the tunnel. It is around 60% of the pull efficiency. They further deduced some empirical equa-
width for small tunnels and reduces to 45% for larger widths. tions for determining pull. However, these equations do
Nevertheless, the tunnel advances achieved from cylindrical not relate the effects of all the blast design parameters on
hole cuts do not depend upon the tunnel width. Further, the actual pull. Nevertheless, soft computing techniques can be
tunnels with dimensions around 8 m are generally blasted in used to co-relate many variables with a non-linear output.
full section (Langefors and Kihlström 1978). Hence, these In the recent years, researchers have applied soft comput-
parameters were not considered for numerical modelling. ing techniques to predict blast-induced vibrations and over-
A statistical summary of blast data used for developing the break in tunnels. However, there is a paucity in literature
model is presented in Table 3. The co-relation matrix sug- pertaining to the prediction of blast-induced pull using soft
gested that the pull is most co-related to average depth of computing techniques. The present study aims to develop a
holes and charge factor (Table 4). CART model for predicting blast-induced pulls of tunnels
at KWHEP. Further, the classification generated by CART
using high number of controlled blast design parameters
Research significance can contribute in easy understanding of blasting phenomena
(Rana et al. 2020). The significance and applicability of the
In India, rhythmic driving is generally practiced for DBM CART model is further verified using statistical hypothesis
tunnel excavations. The pull/day in a rhythmic driving testing and comparing the performance indices of regression
depends upon pull/round. Furthermore, pull must be large tree with multivariate regression analysis (MVRA).
enough to provide muck for one whole shift in a rhythmic
driving. Hence, high value of pull (≈ 4 m) shall be attained
for each blast with a high breakage certainty for economy. Regression tree
Langefors and Kihlström (1978) regarded tunnel blasting
for optimum pull as gambling with many dices; sometimes Regression or decision tree is a simple and interpret-
all of them give one and other times all of them yield sixes. able machine learning method for predicting a depend-
They experimented 8000 rounds of cylindrical cuts and 5000 ent variable. Many algorithms such as ID3, CHAID and
rounds of burn cuts with different blast designs (Taby cut, CART can be used for developing a decision tree. In

Table 3:  Statistical summary of input and output parameters used for developing CART model
Statistical param- Cross section No. of holes Hole Avg. depth Avg. charge MCPD (kg) Total charge (kg) Charge Pull (m)
eter of tunnel (m) diameter of holes per hole factor
(mm) (m) (kg) (kg/m3)

Minimum 43.74 50 45 3.90 1 23.50 95 1 0.40

Maximum 87.61 245 45 4.20 4 124.30 575 18 4.10
Average 65.92 131.30 45 4 2.60 55.40 339.30 2.60 2.40
Median 64.50 129 45 3.90 2.70 52.50 350 2.10 2.40
Standard deviation 19.06 32.60 0 0.01 0.60 19.40 105.50 1.90 0.80
Variance 363.46 1061.60 0 0.01 0.40 377 11131.40 3.70 0.60
Category Input Input Input Input Input Input Input Input Output

Table 4  Correlation matrix

Cross-section Number Hole Avg. Depth Avg. Charge per Total Charge
MCPD Pull
of Tunnel of Holes Diameter of Holes Hole Charge factor
Cross-section of Tunnel 1
No. of Holes 0.49 1
Hole Diameter 0.60 0.34 1
Avg. Depth of Holes 0.39 0.47 -0.06 1
Avg. Charge per Hole -0.16 -0.44 -0.55 0.23 1
MCPD 0.15 0.59 -0.01 0.28 -0.04 1
Total Charge 0.35 0.66 0.02 0.61 0.14 0.62 1
Charge factor -0.61 -0.30 -0.64 0.04 0.58 0.04 0.01 1
Pull 0.35 -0.05 0.12 0.43 0.16 -0.28 0.07 -0.46 1

13

Download fulltext IDR 3K - [Link]