Journal of JSCE, Vol.

10, 260-274, 2022

EVALUATING TUNNEL ROCK MASS

USING DEEP LEARNING

Koji HATA1

1Member of JSCE, Technology Division, Technology Research Institute, Obayashi Corporation

(4-640, Shimokiyoto, Kiyose-shi, Tokyo 204-8558, Japan)
E-mail: hata.koji.ro@obayashi.co.jp

The design of support structures for Japanese mountain tunnels is difficult owing to the limitations in
investigating techniques and the complexity of the geological structures involved. In general, during exca-
vation, the size of a support structure is modified to best suit the geological conditions. As such, engineers
must carefully observe the rock mass conditions to understand their mechanical and hydraulic properties as
accurately as possible, and to predict their deformation behavior and hydraulic structure. In recent years,
artificial intelligence (AI) has been adopted in various fields. In this study, a deep neural network (DNN)
was applied to evaluate a mountain tunnel’s rock mass. The input was a photo of the excavation surface
(face) of the mountain tunnel, and the output was the rock mass properties such as degree of weathering,
alteration, and fracture. Based on past excavation records, the DNN was tested using supervised learning,
and the results showed that the AI judgments were consistent with the engineers’ judgments, having a 73%–
97% accurate answer rate. Therefore, practically applying the method of rock mass evaluation using AI
was determined as being feasible. Furthermore, to allow ease in its field-based application, a cloud com-
puter system using a tablet computer device was used to enable evaluations, creating a system that contrib-
uted to increased productivity.

Key Words : mountain tunnel, artificial intelligence, deep learning, deep neural network, diagnostic im-
age

1. INTRODUCTION Yoshioka13) developed an expert system based on the

results of their research. In this system, image pro-
Artificial intelligence (AI) has garnered significant cessing and pattern classification by a neural network
research attention since 1997 when IBM®’s Deep were applied to observe and evaluate rock mass. To
Blue® defeated the world chess champion1). In 2011, evaluate the in situ measurement results, the final de-
on the American quiz show Jeopardy!, IBM’s Wat- formation of the inner cavity and the extent of the in-
sonTM defeated the record holders of the most consec- elastic region were predicted using the Weibull dis-
utive wins and the highest accumulated prize win- tribution curve. Moreover, two-dimensional elasto-
ner2). However, Watson is not defined as AI but as a plasticity based on finite element analysis was easily
cognitive computing system3). In the industrial world, performed as a numerical analysis method using a
applying AI has expanded to include automated driv- personal computer.
ing4), medical and nursing care robots5), and other The three important methods of rock mass evalua-
fields. tion, namely observation, measurement, and numeri-
The period from the 1980s to the early 1990s is cal analysis, are designed to be functionally linked.
known as the second generation of AI6), 7). During this Specifically, Bieniawski14) calculated the rock mass
period, various research and development activities rating from rock mass observations and selected in-
were performed in the Japanese construction indus- put properties for the numerical analysis of the rock
try. For example, many AI techniques were devel- mass. This device enabled parametric analysis in vi-
oped, such as simple neural networks, fractal dimen- cinity of the excavation site. However, personal com-
sion8), fuzzy theory9), 10), and genetic algorithms11). puters then were not as powerful as the ones used to-
Among them, the expert system12) has attracted atten- day, and their computational speed and capacity were
tion as a practical system for engineering. Hata and poor. Therefore, neural networks were only used for

260
pattern classification and not for solving complex Tanimoto and Yoshioka22),23) explained the im-
problems. portance of measuring the displacement inside the
In recent years, AI efforts have advanced to the cavern in mountain tunnels and proposed a new rock
third generation and have once again become the mass classification technique and a support classifi-
topic of considerable interest for application in vari- cation method. Furthermore, Tanimoto et al.24) pre-
ous fields because of the increasing speed of comput- sented a method for evaluating the rock strength. The
ers and innovations in algorithms. Lawal and Kwon15) initial stress, which is the external force for support
surveyed and summarized AI applications in rock structure design, is an important rock mass property
mechanics. From this survey, it can be seen that there that determines the size of the support structure. The
is a wide range of applications such as blasting ef- initial stress measurement methods are broadly clas-
fects, strength and brittleness prediction of rock sified into in situ and laboratory tests. Michihiro et
mass, development to numerical analysis, and rock- al.25),26) and Hata et al.27) focused on the acoustic
burst prediction in mines. However, there are few emission (AE) method. The initial stress test method
studies that lead to tunnel face evaluation. On the by AE was established by clarifying the fundamental
other hand, Ohtsu et al.16) and Hasegawa et al.17) at- phenomena related to the Kaiser effect in rocks28),29).
tempted to evaluate tunnel face by using AI to learn The Kaiser effect is a phenomenon in which, when a
seismic wave velocity and resistivity of the ground in historically loaded material is reloaded, almost no AE
a preliminary survey of mountain tunnels. Tsuchiya occurs until the prior load value is exceeded. Further-
et al. 18) pointed out the importance of increasing the more, elastic and elastoplastic analyses are consid-
number of training data and including a variety of ered as standard methods for the numerical analysis
rock types in order to apply AI to tunnel face evalua- of mountain tunnels. Nakaoka et al.30) focused on the
tion. In their study, AlexNet, a multilayer deep neural strain-softening phenomenon that enabled accurate
network (DNN) proposed by Krizhevsky et al.19), was predictions of stress deformation behavior under tri-
used to: axial stress. In the case of mountain tunnels, it is not
• evaluate the observation of the surface of the possible to conduct observations inside the rock
mountain tunnel, mass; therefore, the nature of the tunnel cannot be de-
• operate the technology in the field, and termined until it is excavated. For this reason, various
• deploy a cloud system. rock mass exploration methods have been adopted to
Its applicability and issues are described in detail. understand the rock mass properties before excava-
tion. Kuwahara and Hata31) , as well as Kuwahara et
al.32),33) developed an exploration method using a
2. EVALUATING ROCK MASS IN drilling machine, that is now commonly used at many
MOUNTAIN TUNNELS sites to reveal the rock mass properties for mountain
tunnels. Furthermore, Hata et al.34), 35), 36) developed a
In the 1960s in Austria, Rabcewicz20),21) proposed method to evaluate the damage area around a cavern
a mountain tunneling method called the New Aus- using AE. Recently, Hata et al.37) modified this
trian tunneling method (NATM). This method rap- method into an optical AE method and extended its
idly gained worldwide importance and was intro- function to long-term monitoring technology for
duced in Japan in the 1980s. Currently, the Japanese- high-level radioactive waste and underground stock-
based NATM has been deemed as the standard piling of liquefied petroleum gas. Currently, rock
method for mountain tunnel construction. In this mass is evaluated comprehensively based on tests,
method, shotcrete is applied to the excavated wall measurements, and numerical analysis.
after excavating the rock mass. Then, rock bolts are The direct observation of rock excavation condi-
inserted in a radial direction from the rock wall. Dur- tions contributes to a comprehensive evaluation as
ing this process, the support structure and rock wall relevant inputs are provided based on the knowledge
are in close contact with each other, and the inherent and experience of the engineer conducting the obser-
resistance of the rock is fully utilized. If the rock mass vation. In Japan, the excavation surface (face) is care-
is heterogeneous or fragile owing to significant fully observed immediately after excavation, and a
weathering and alteration, a steel set support is added. face observation record is prepared as described in
To exploit the inherent load-bearing capacity of a the Standard Specification for Tunneling-201638).
rock mass, its geological characteristics must be ac- Figure 1 illustrates a railway tunnel, and Fig.2 a road
curately evaluated. Therefore, observations of the tunnel. Fig.1’s A to I and Fig.2’s A to G were used
rock mass are essential along with tests, measure- to evaluate the condition and behavior of the rock
ments, and numerical analyses to comprehensively mass, and to determine the optimum support required.
evaluate the characteristics of the rock mass. The following seven items were common among
them:

261
 Distance from starting point Provide additional information about the following items for unconsolidated ground,:
Name of tunnel ○○Tunnel Location
Distance from portal
Comprehensive Evaluation of ground classification or pattern Stratum 1. Single soil layer 2. Alternation (1. Horizontal 2. Inclined)
Overburden evaluation classification conditions 3. Lens (1. No 2. Yes) 4. Others ( )
Name of rock
△△m

Type of rock G and geological sandstone and shale

age Special 1. Talus layer 2. Mudflow layer 3. Boundary with rock mass
Status of special Swelling pressure, unsymmetrical pressure, fluidity, small overburden depth ( m), conditions 4. Cross section overlain by 5. Fill or
conditions important neighboring structures, directly under valley, others ( ) soft layer embankment 6. Others ( )
1. Alternation 2. Unconformity 3. Intrusion 4. Microfolding 5. Fault Discontinuity 1. Well-developed cracks or 2. Seam 3. Fault 4. Others ( )

Ground conditions
fissures
6. Others 1. Cohesive soil 2. Sandy soil 3. Gravelly soil
Soil type 4. Special soil (1. Decomposed granite 2. Volcanic ash 5. Others ( )

Mode of
occurrence
3. Shirasu 4. Organic soil)
Condition and behavior of ground at excavation site 1. Soft 2. Medium 3. Hard 4. Very dense
4. Unable to (4 > N) (8 > N ≥ 4) (15 > N ≥ 8) (30 > N ≥ 15)
support itself, Cohesive soil
A Face condition 1. Stable 2. Rocks fall off 3. Squeezing
occurs at the the face 5. Others 5. Consolidated (N ≥ 30)
the face face collapses or
flows out Sandy soil 1. Loose 2. Medium 3. Dense 4. Very dense
2. Loosening of 3. Barely able to (10 > N) (30 > N ≥ 10) (50 > N ≥ 30) (N ≥ 50)
Condition of ground and stand and 1. 2. 3.
1.Standing (no rock falls requires early 4. Support before Loose 2-5cm 5-20cm 20-75cm Gravel 1. ≤ 30%
B unsupported support occur from support after excavation 5. Others Gravelly soil 1. 2. 30–50%
excavation 2. Compacted Gravel size content 3. 50% ≤
needed) time to time excavation needed 4. 75-300cm 5. 300cm ≤

Soil conditions (for

special soil, select most
applicable items.)
surface (ordinary (immediate
support) support) 1. Impervious layer 2. Impervious to
3. 20 > σc ≥5 relatively impermeable
1. σc ≥100 MPa 2. 100 > σc ≥20 4. 5MPa>σc Ground properties N-value Permeability
Compressive Withstands Crumbles Crumbles 3. Alternation of both 4. Others ( )
C strength hammer when hit by a when hit A hammer head 5. Others
ends up lodged
blows hammer lightly by a in
hammer Groundwater head m above FL Remarks
1. No 2. Discoloring 3. General Sketch & Photographs:
Weathering/ weathering; along rifts; discoloration, 4. Soil, clay,
D slightly considerably fractured, 5. Others

262
alteration sound reduced in reduced in unconsolidated
condition collapsed rock at the top
strength strength
Crack/ fissure 1. Spacing d 2. 1 m> d ≥20 3. 20 cm> d ≥5 4. 5 cm>d
E frequency ≥1m cm cm Fracture 5. Others
unconsolidated discontinuity
Crack/ fissure 4. Interbeds of collapsed rock
F conditions 1. Tight 2. Partially open 3. Open clay, 5. Others
unconsolidated
1. Random 2. Columns 3. Layers,
Mode of squares fragments, 4. Soil,
G cracking/ plates fragmented, 5. Others
fissuring unconsolidated

H Water inflow 1. None, seepage 2. Dripping 3. Concentrated 4. Entire face 5. Others

I Deterioration 1. None 2. Loosening 3. Softening 4. Collapse, 5. Others
due to water outflow Article:
1. Horizontal (10º>θ>0º) 2. Dipping into slope (30º>θ≥10º, 80º>θ≥60º) Alternate layers of sandstone and shale.
Longitudinal The parent rock is solid, but has well-developed fractures.
direction 3. Dipping into slope (60º>θ≥30º) 4. Dipping out of slope (60º>θ≥30º) Some rocks have collapsed.
(cutting face) Brownish discoloration on some of the cut wings.
5. Dipping out of 6. Vertical (θ≥80º) [in terms of maximum angle Unsafe due to rock fall at the top.
slope(30º>θ≥10º, 80º>θ≥60º) of slope]
1. Horizontal (10º>θ>0º) 2. Right to left (30º>θ≥10º, 80º>θ≥60º)
Transverse
direction 3. Right to left (60º>θ≥30º) 4. Left to right (60º>θ≥30º)
(cutting face) Name of
5. Left to right (30º>θ≥10º, 6. Vertical (θ≥80º) [in terms of apparent angle of Koji Hata
observer

Crack/fissure orientation
(presence of a discontinuous
cutting face)
80º>θ≥60º) slope of cutting face]

Fig.1 Examples of face observation records in a railway tunnel in Japan.

Tunnel name Observation date:
Measuring station + Distance from portal: m Cross section: No. Supporting pattern:
Overburden: m Rock name/Geological age: Rock group (1-5): Rock code:
Specifications of auxiliary method Measurement A / B
(inc. face shotcrete and bolts) Specifications of additional supports Specify A or B if it is the closest cross section

Special conditions, status, etc .

Crown
.
Left Right
shoulder shoulder
Presence/absence and status of collapse .
Presence/absence of early invert closure .

Enter the
Observation item Rating ratings here
Left CenterRight
Unconfined
A compressive 100 or more 100 - 50 50 - 25 25 - 10 10 - 3 3 or less
strength
Point load 4 or more 4-2 2-1 1 – 0.4 0.4 or less
strength (N/mm2)

Hard to break Breakable 3 2 3

Compressive

Rough guide for when placed on when placed on Breakable Breakable Partially Crushable
strength in the ground and the ground and when held by when hit with breakable using when force is
hammering strongly hand and another rock of two hands applied with a
strongly hammered the same type fingertip
hammered hammered
Rating 1 2 3 4 5 6
Rough guide for Generally fresh Weathered or Weathered or altered Weathered into soil or
B weathering altered along the crack to the rock core unconsolidated soil
Weathering and

Rough guide for Decreased strength

Clay observed in the
alteration

hydrothermal No alteration observed crack due to alteration observed at the rock

Remarkable alteration 3 2 3
into soil or clay
alteration core due to alteration

Rating 1 2 3 4
d≥1m 1 m > d ≥ 50 cm 50 cm > d ≥ 20 20 cm > d ≥ 5 cm
C Crack intervals cm
5 cm > d

4 3 4
intervals

RQD 80 or more 80 - 50 60 - 30 40 - 10 20 or less

Crack

Rating 1 2 3 4 5
Some of the cracks Many of the cracks
D Crack opening Tightly closed open (less than 1 open (less than 1 Open Open to the width of
mm wide) mm wide) (1-5 mm wide) 5mm or more
Substance in None None Thin clay Thick clay
None 4 4 4
conditions

cracks (5 mm or less) (5 mm or more)

Crack

Crack roughness A limited A highly polished

and slickenside Rough Smooth slickenside slickenside
Rating 1 2 3 4 5
Strike 1. Reverse dip of 2. Reverse dip of 3. Regular/reverse 4. Regular dip of 5. Regular dip of 45
E perpendicular to 45 to 90 degrees 20 to 45 degrees dip of 0 to 20 20 to 45 degrees to 90 degrees 4 3 3
tunnel axis degrees
Strike and

Parallel with 1. Dip of 0 to 20 2 .Dip of 20 to 45 3. Dip of 45 to 90

tunnel axis degrees degrees degrees 2 2 2
dip

Evaluation by water inflow and water-induced deterioration within 10 m from the face
(Deterioration shall be evaluated considering existing and possible futuredeterioration.)
No inflow or water Water extensively
F Status seeping not exceeding Water dripping at
1 to 20 liters/min
Water locally inflowing inflowing at 100 liters or
at 20 to 100 liters/min.
1 liter/min. more per min.
2 2 2
inflow
Water

Rating 1 2 3 4

G Water-induced None Loosened Weakened Washed out

deterioration
Water-induced
deterioration

3 3 3
Rating 1 2 3 4

Fig.2 Examples of face observation records in a Japanese road tunnel.

263
 (a) Compressive strength, ture of Japan is very complex and the size and distri-
(b) Weathering and alteration, bution of fractures need to be considered in detail.
(c) Crack intervals, Therefore, these international rock mass determina-
(d) Crack conditions, tion methods are currently only used for research or
(e) Strike and dip, specific projects. During the construction of road and
(f) Water inflow, as well as rail tunnels in Japan, the rock mass determination
(g) Water-induced deterioration. methods shown in Figs.1 and 2 are used. Moreover,
In addition to these parameters, other information, sufficient training data are not available for AI re-
such as crack/fissure orientation, overburdening, type of search.
rock and geological age, and status of special conditions In recent years, it has been recognized that it is im-
were considered necessary and had to be taken into ac- portant to share information obtained from observa-
count. The descriptions in Figs.1 and 2 were the cases tion, measurement, and numerical analysis through
judged important from the attached tunnel face images; information and communication technologies. Hata
the corresponding judgment results are shaded in or- et al.41) were among the first to recognize the need for
ange. such a system and conducted pioneering research in
To accurately evaluate the geological properties of the field of mountain tunnels in Japan, finally, de-
a rock mass, it is essential to have a thorough ploying a practical system in the field.
knowledge of geology and rock mechanics. However,
most engineers working in mountain tunnels do not
necessarily have sufficient knowledge to make such 3. DEEP NEURAL NETWORK OVERVIEW
judgments. Moreover, when conducting observations
of the rock mass or geology under construction for Human beings make judgments and inferences
the first time, engineers find it difficult to make deci- about events based on the knowledge and experience
sions. However, if a computer can determine the state they have acquired in the past. Machine learning
and behavior of Fig.1’s A to I with appropriate accu- (ML) is a technology and method that attempts to
racy from the face images, the task of selecting sup- realize the same or better functions in a computer by
ports will be made more efficient. In fact, while con- emulating the natural learning ability of humans.
structing mountain tunnels, Obayashi Corporation of- Deep learning (DL) is a method for realizing ML and
ten requests rock mechanics and geology experts to corresponds to an algorithm. Neural networks, which
help their field engineers in preparing face observa- have recently become a popular topic, have been
tion records (see Figs.1 and 2). However, there are recognized as methods for realizing DL; for example,
insufficient professional engineers in rock mechanics DNN or convolutional neural networks. Figure 3
and geology to handle all the requests from the field. shows the basic structure of DNN.
If geological conditions can be automatically deter- This section describes the features of DNN
mined and evaluated, the productivity of each site can processing from the perspective of image analysis.
be significantly improved. Therefore, for the AI eval- The DNN outputs the results by passing the input im-
uation system aimed at practical use, the input infor- age through feature extraction phase and classifica-
mation should include tunnel face images, while the tion phases. In Fig.3, the feature extraction phase
seven parameters mentioned concerning the condi- consists of convolution and pooling layers. The con-
tion and behavior of the rock mass will be obtained volution layer extracts local features of the image
as output. The rock mass rating by Bieniawski14), ge- based on red–green–blue (RGB) changes via sharp-
ological strength index by Hock et al.39), and rock ening filters, noise processing filters, etc. In other
mass quality by Barton et al.40) are well known rock words, geometric information that is only available
mass rating systems. However, the geological struc-

Fig.3 Basic structure of DNN.

264
 fc 6

Input
conv 1 conv 2 conv 3 conv 4 conv 5 fc 7
Image
Output
fc 8
pool 1 pool 2 pool 5 Classification

fully connected
convolution layer／pooling layer layer
Feature extraction phase Classification phase

Fig.4 Layered structure of AlexNet – a DNN algorithm.

Feature extraction phase：AlexNet

conv1 conv2 conv3 conv4 conv5

Output of filter 1 ~ 96
11×11×3 5×5×48
Output of filter 3×3×256
1 ~ 256 3×3×192
3×3×192
・・・
13×13×384 13×13×256
27×27×256 13×13×384
55×55×96

fc6 fc7 fc8 Weathering/ alteration 1

Input Image 2
・・・

227×227×3 1
Crack/ fissure frequency
2
4096×1 4096×1 1000×1

Fig.5 Feature extraction phase of AlexNet.

for a particular pixel within a pixel region is con- 4. APPLIED DNN MODEL (ALEX NET)
verted into analytical information that represents the
features. With this operation, if the input image In this research, a multilayered DNN19) developed
slightly shifts in parallel, it is recognized as the same at the University of Toronto was used. This model,
type of image representing a feature; thus, highly ro- called “AlexNet” after its developer, was highlighted
bust training data were obtained. By contrast, the in the 2012 International Large Scale Visual Recog-
pooling layer compresses the image size (reduced the nition Challenge (ILSVRC) for its remarkably low
image to a few pixels area) using the average and discrimination error rate.
maximum values based on the local features obtained Figure 4 shows the layer structure of the feature
in the convolution layer. extraction and classification phases. In the figure,
In the subsequent classification phase, the fully “conv” denotes the convolution layer and “pool” de-
connected layer provides data on the number of neu- notes the pooling layer. The initial model had a five-
rons output in the feature extraction phase, and the layer hierarchical structure. The pooling layer was
recognition accuracy is improved by overlapping not incorporated in the third and fourth layers. The
these data. In early neural networks, the convolution “fc” in the classification phase indicates a fully con-
and pooling layers comprised one or few layers. nected layer. In this research, the model consisted of
However, in recent years, with the dramatic increase two layers of 4,096 neurons and one layer of 1,000
in the computational speed and memory capacity of neurons.
computers, complex information processing consist- Figure 5 shows the specific processing flow of the
ing of many more layers has become possible. first through the fifth convolution layers. The input
image data were denoted as 227 × 227 × 3. Here, the
size of the image was 227 pixels in width and 227
pixels in height; the “3” at the end indicated the three

265
 Fig.6 Convolutional filter in the first layer. Fig.7 Convolutional filter in the second layer.

RGB. In conv1, the first layer, 96 different convolu-

Pre-stage processing
tional filters were used to extract the features.
Figure 6 shows the visualization of these filters, ・Standardize image size of face
where the size was 11 × 11 × 3. In other words, it was ・Cut out only the face
understood that the input image was convolved by the ・Divide the cutout into three parts
convolutional filter to extract similar parts. The pro- ・Segment image to size
cessed image was passed on to the next layer, which 227×227×3
had an image size of 55 × 55 × 96. In the second layer, ・Remove less than 50% of face
more complex features were extracted using 256 dif-
ferent convolutional filters. The filter size was AI Analysis
5 × 5 × 48. At this later stage, it was difficult to visu-
・Feature extraction by AlexNet
alize the filter (see Fig.6). Therefore, Deep Dream42) ・Classification by
was used for further filter visualization. Deep Dream support vector machine
is a DL algorithm developed by Google.
The algorithm emphasizes the extracted features Generating a face learning model
and replaces images based on the training infor-
mation. Using the features of Deep Dream, it was Fig.8 Process of generating a face learning model.
possible to visualize the processing contents in the
multilayered structure of DL.
Figure 7 shows the visualization results of 100 object itself from the image but also for distin-
randomly selected features from 256 convolutional guishing it from other objects.
filters. Features that were more complex than conv1 (c) The model is built using ImageNet, an image da-
were extracted by combining the simple features ex- tabase, with 1.2 million images for training,
tracted by conv1. For example, in the conv1 layer, 150,000 images for testing, and 50,000 images
basic features, such as “vertical stripes” and “yellow for validation. In addition, the tunnel face image
and purple gradation,” were extracted. In the conv2 information described in Chapter 6 below is
layer, features, such as “speckled pattern” and com- added to the already trained model, and the
bination of “red and green gradation,” were extracted. model is trained again for use in this research.
The third and subsequent layers were omitted owing
to limited space.
AlexNet was adopted here because it has excellent 5. TRAINING DATA FOR MOUNTAIN
features over other similar models: TUNNEL FACE EVALUATION
(a) It has a simple layered structure with five layers
in the feature extraction phase and three layers In AI used for face evaluation in mountain tunnels,
in the classification phase. This structure re- the training data were created from observational
quires less time to learn from the training data data obtained from mountain tunnels constructed in
without using a GPU. the past by the Obayashi Corporation. The training
(b) It extracts features not only for recognizing the data comprised the face images and the rock mass
conditions and behaviors Fig.1’s A to I, as shown in

266
 Fig.9 Three-section evaluation area of a tunnel face.

Fig.11 Example of input tunnel face image taken

at various tunnel sites.
Fig.10 Segmentation of a tunnel face in a crown.

However, to extract features accurately, it was essen-

the previous section. In other words, supervised tial to use only images with more than 50% coverage
learning was peformed based on these data. in a 227 × 227 area. Therefore, there were originally
In general, when developing an AI system, a large 14 images, but the four images on either side of the
amount of training data is required; the correct an- first and second rows shaded in gray were deleted.
swer rate increases in proportion to the amount of DL with AlexNet was performed based on the pre-
training data. Obayashi Corporation has significant processed segmented image data. As for the evalua-
experience in mountain tunnel construction and a tion of the rock mass condition and behavior, the road
vast amount of face observation data. It would take tunnel in Fig.2 has more detailed markings; hence, all
an enormous amount of time for the AI system to the data were reevaluated to match the face judgment
learn all these data. The main reason for adopting of the road tunnel. The aforementioned process re-
AlexNet was the advantageous features of AlexNet lated the features of the images to the evaluation cat-
described in the previous section. In other words, the egories. Although AlexNet had the capability to han-
model had already learned about rocks and rock dle all phases, support vector machines were used for
masses to some extent. Therefore, it was determined classification calculations that linked image features
that adding the tunnel face image data to the already to evaluation categories. There are various types of
existing trained model to create a modified training classifiers, such as decision trees and logistic regres-
model could be used in this research. sion. However, support vector machine was used be-
The learning model was created from the training cause of its high analytical accuracy and its capability
data based on the flow in Fig.8. The face images in to identify the boundaries of classification categories.
Fig.1 have a different number of pixels at each work
site. First, the input image size had to be standardized.
As Fig.9 shows, the face was usually divided into 6. RESULTS
three sections, the crown and lower sections on the
left and right; the rock mass was evaluated in each Seventy face observation records from six project
section. The grades of each section were averaged or sites constructed by the Obayashi Corporation were
weighted to calculate the rock mass grade of the en- selected as training data. Initially, field engineers pre-
tire face; i.e., each of the 227 × 227 regions had its pared the face observation records. Then, they were
own individual evaluation results. However, because reviewed by experts in rock mechanics, tunnel engi-
the observation record of the face made at the work neering, and geology to accurately assess the condi-
site was used as the basis of the training data, the tion and behavior of the rock mass. The face photo-
evaluation values representing each of the three areas graphs us ed were close t o full HD im ages
of the face were used as the training data. Then, the (1920 × 1080 pixels). The geological grade at which
image was segmented into 227 × 227 × 3 (pixels, the rock mass was evaluated ranged from poor to rel-
RGB) based on the usage conditions of AlexNet (see atively good. For example, the H-200 grade steel set
Fig.10). In this case, in the area above the face, the was added to the shoring in poor ground; whereas,
14 images in Fig.10 were recognized as input images. the steel set was not required in relatively good

267
 Input
conv1 conv2 conv3 conv4 conv5 classifier
Image

pool1 pool2 pool5

Fig.12 Example of output image at each convolution layer.

ground. A variety of rock types were considered, in- discrimination results that humans cannot discrimi-
cluding sandstone, slate, mudstone, shale, granite, nate. A DNN is not necessary if the results were
rhyolite, tuff, and tuff-angular conglomerate. equivalent to the visual observation of an engineer. It
would be more effective to use ML in combination
(1) Processing results in the convolution layer with other image-processing methods that require
Figure 11 shows an example of the input image, less computation.
and Fig.12 the processed images at each level of
AlexNet. Figure 12 shows the output values of acti- (2) Correctness of training data
vations processed by AlexNet for nine different fac- In the classification phase, the relationship be-
eted images, expressed as luminance distributions. tween the image features and evaluation categories
The output of each convolutional layer is drawn as a was verified. In this project, 70 observation records
grayscale image with no RGB information. Origi- from six sites were used as the training data. As
nally, the analysis output was based on a 227 × 227 Fig.10 shows, the total number of divided images
segmented image, but the images were combined to was approximately 3,000. Approximately 2,000 data
show the progress of the process over the entire face. points from the 3,000 preprocessed data points were
In DL, the further the processing, the more abstract used for training. The remaining 1,000 data points
the image becomes. The output will be an image that were used as the test dataset. Then, the correct answer
is difficult to distinguish from what it represents. This rate was verified with the remaining 1,000 data points.
trial analysis also showed that the abstraction of the This process was repeated 10 times. In other words,
image progressed as it was processed. The composite the variability of the correct answer rate was deter-
image in conv1 was more difficult to see than the mined from 10 trial calculations, and the possibility
other layer analysis results because of the luminance of practically applying this technology was deter-
distribution, but it could be recognized as represent- mined.
ing a tunnel face. However, in the case of conv3 and The results of calculating the accuracy of the cor-
later, the composite image was barely recognizable rect answers were as follows: Table 1 shows the re-
as a tunnel face, and it was not clear what the sults of the comparison between the real and pre-
227 × 227 segmented image represented. In other dicted evaluations of weathering alteration. For ex-
words, conv3 and later output similar images made it ample, in the category classified as “Generally fresh”
impossible to determine whether feature extraction in Table 1, the number of correct answers was 183.
was in progress. Thus, it seemed that DL output Because the total number of images in this category
seemingly contradictory results: as if the processing was 262 images (183 + 67 + 11 + 1), the correct an-
by the convolutional layer was in progress, but the swer rate was 70%. In the case of the training data,
feature extraction was not. the overall average correct answer rate was 75%. In-
The significance of using DL is that it can output dividually, the number of correct answers far ex-
ceeded the number of incorrect answers, except for

268
 Table 1 Contrasting results of real and predicted assessments of weathering alteration.

Predictive evaluation
Weathered or altered Weathered or altered Weathered into soil or
Evaluation category Generally fresh
along the crack to the rock core unconsolidated soil
Generally fresh 183 67 11 1
Real evaluation

Weathered or altered
along the crack
72 451 45 3
Weathered or altered
to the rock core
11 39 130 6
Weathered into soil or
unconsolidated soil
0 1 5 8
correct answer rate=772/1033=75%

Table 2 Contrasting results of real and predicted assessments of crack intervals.

Predictive evaluation

Evaluation category d≥1m 1 m > d ≥ 50 cm 50 cm > d ≥ 20 cm 20 cm > d ≥ 5 cm 5 cm > d

d≥1m 0 0 0 0 0
Real evaluation

1 m > d ≥ 50 cm 0 412 113 1 0

50 cm > d ≥ 20 cm 0 120 312 7 2
20 cm > d ≥ 5 cm 0 3 10 42 3
5 cm > d 0 0 0 6 2
correct answer rate=768/1033=74%

Table 3 Contrasting results of real and predicted assessments of crack conditions.

Predictive evaluation
Some of the cracks Many of the
Open (1-5 mm Open to the width
Evaluation category Tightly closed open (less than 1 cracks open (less
wide) of 5 mm or more
mm wide) than 1 mm wide)
Tightly closed 61 37 1 0 0
Some of the cracks open
Real evaluation

(less than 1 mm wide)

35 647 27 2 1
Many of the cracks open
(less than 1 mm wide)
2 37 146 0 6
Open (1-5 mm wide) 0 2 6 2 0
Open to the width of
5 mm or more
1 1 6 0 8
correct answer rate=864/1028=84%

the category indicating “weathered into soil or un- the number of images corresponding to the predicted
consolidated soil.” The low percentage of correct an- rating category. In the case of the training data, the
swers in the category of “weathered into soil or un- overall average correct answer rate was 74%. The
consolidated soil” was due to the relatively small confusion in classifying and discriminating between
amount of training data in this category (14 images). “1 m > d ≥ 50 cm” and “50 cm > d ≥ 20 cm” was
Therefore, it was assumed that there was insufficient considered to be the main reason the accuracy rate
information to properly identify them in the classifi- remained in the 70% range. It was not difficult for an
cation phase. engineer to distinguish between “1 m > d ≥ 50 cm”
Table 2 shows the results of the comparison be- and “50 cm > d ≥ 20 cm” visually, but it might be
tween the real and predicted crack interval evalua- difficult to distinguish them from the images. There-
tions. As Table 1 shows, the vertical column was the fore, it was necessary to examine the suitability of
real rating category, and the horizontal column was this classification.

269
 Table 4 Correct answer rate on training data of tunnel face.
Rock mass condition Number of Average of correct Range of correct
and behavior item categories answer rate (%) answer rate (%)
Compressive strength 6 96 95 - 97
Weathering and alteration 4 73 71 - 75
Crack intervals 5 73 71 - 74
Crack conditions 5 83 82 - 84
Strike and dip 5 81 79 - 84
Water inflow 4 88 86 - 90
Water-induced deterioration 4 97 96 - 97

Table 3 shows the results of the comparison be- the number of training data was considered as an ef-
tween the real and predicted evaluations of crack con- fective means. As for deviating correct answer values,
ditions. In Tables 1 and 2, the real evaluation there was a possibility that there was an error in the
categories are shown in the vertical column, and the discrimination of the training data. In other words,
number of images corresponding to the predicted based on the three sections of the face observation
evaluation category are shown in the horizontal record operated in the field in Fig. 9, the training data
column. In this case, the overall correct answer rate were created with the same evaluation for the entire
was 84%, which was a high fit rate compared to area. It was possible that there was a problem with
Tables 1 and 2. However, the number of incorrect the method used to create the training data. Improv-
answers was higher in the two categories of “Open ing the rate of correct answers and resolving the dis-
(1–5 mm wide)” and “Open to the width of 5 mm or crepancy between fitted values is a future task.
more.” Moreover, a large discrepancy was observed,
resulting in a difference between the two steps. This
large discrepancy raised questions regarding the suit- 7. FIELD APPLICATION OF TUNNEL
ability of the training data. A major challenge was to FACE EVALUATION AI SYSTEM
find ways to improve the predictive fit of this cate- (CLOUD-BASED)
gory.
Iterative learning using training set, test set, and At a mountain tunnel site, a face observation is
validation set are commonly used in DNN training. conducted immediately after the rock mass is re-
The AlexNet used in this study is a trained network, moved from the excavation. Normally, the results of
and since it is used to convert images into feature vec- the observation and sketching are brought back to the
tors, it does not undergo iterative learning. However, work office, transcribed into the prescribed face ob-
the variability of the estimated values was confirmed servation record. However, it is important to be able
by 10 trial and error runs. This method evaluates the to make an immediate evaluation while looking at the
variability of the estimated values, which is thought face in situ after excavation. Visualization by images
to lead to the verification of accuracy. was the most convenient way for construction man-
Table 4 presents the results. No significant differ- agers, workers, and contractors to share face infor-
ence was found among the seven rock mass charac- mation accurately. The degree of partial structural in-
teristics and evaluation categories, with the error of stability was identified by evaluating the cracks and
the target rate in the ten trials being less than 5%. The weathering that appear on the rock excavation sur-
minimum correct answer rate was 73% for weather- face. Then, depending on the needs of the site, a suit-
ing alteration and crack intervals, and the maximum able solution for the deformation characteristics of
correct answer rate was 97% for water-induced dete- the rock mass was discussed. Considering these re-
rioration. Surprisingly, the percentage of correct an- quirements, it was necessary to have an analysis sys-
swers was over 70% for the seven evaluation items. tem that could be accessed anywhere, anytime, by an-
Each evaluation item was classified into four to six yone, and that could discriminate rock masses with
categories; the many choices had been a source of the same accuracy. In recent years, computer com-
concern for field engineers. The ability to judge using munication network environments have been devel-
a machine is thus significant. However, there have oped even in mountainous regions far from urban ar-
been cases where some predictions deviated by as eas, and an environment to promote the use of cloud
much as two ranks from the correct answer. To im- computing for analysis systems has been prepared.
prove the percentage of correct answers, increasing

270
 Server/Cloud
Tablet PC GUI
Internet Web
screen
Apps

VPN
Web browser
Multiple locations
Web server Multiple users
Simultaneous processing of access
Fig.13 Block diagram of the AI cloud system for tunnel face evaluation.

(a) (b)
Fig.14 Example of use of face evaluation AI system at tunnel site. (a) Photo taken at the tunnel face and
(b) tunnel face AI evaluation response on a tablet device.

Figure 13 shows a block diagram of the system. A precision correction method for distortion in their re-
picture of the face to be evaluated with a tablet device, search on tunnel displacement measurement technol-
such as an iPad, is taken and sent to a web server con- ogy using a digital camera. Nakamura et al.44) inves-
nected to the Internet. tigated the effect of the tunnel face and the image sen-
The analysis results were immediately sent back to sor not being in parallel planes and proposed a
the tablet device. Although it depended on the com- method to compensate for this. The new tunnel face
puter communication network environment and the AI system evaluated the rock mass condition and be-
capacity of the captured tunnel face images, the speed havior in the observation area; hence, high-precision
was several tens of seconds for one or two items and image correction was not required to calculate the de-
less than two minutes for all seven items. Moreover, formation. Therefore, advanced correction tech-
multiple processes at multiple sites could be analyzed niques, such as distortion and optical axis correction,
concurrently; the system was designed to minimize were not considered necessary.
the stress experienced by field engineers when they Figure 14 shows an example of the use of this
were on standby. The aim of this AI cloud system for method at a mountain tunnel site. After the excava-
tunnel face evaluation was to make the evaluation tion of one span of the rock mass (in this case, by dy-
system available to anyone, anywhere, and anytime. namite blasting), the rock mass was removed. After
Photographic equipment used ranges from inex- confirming that the tunnel face had not collapsed, the
pensive devices such as iPads and iPhones to a single engineer entered the vicinity of the tunnel face to ob-
lens reflex digital camera. However, what they all serve and take photographs (Fig.14(a)). Then, the
had in common was that the captured image was af- photographs were sent to the designated web server.
fected by lens distortion. Moreover, the center of the By running the face evaluation AI application on the
image sensor and the center of the optical axis did not tablet, the transmission to the web server was com-
coincide, resulting in nonlinearity in the distortion. In pleted as soon as the face was photographed. After
the case of small lenses, such as those used on the sending the data, the analysis results were sent back
iPad and iPhone, the impact was greater than that of to the sending tablet device as the observation pro-
large aperture lenses. Hata et al.43) proposed a high- cess continues (Fig.14(b)). Usually, it was sufficient

271
and convenient to analyze only the face area. How-
Legend：weathering alteration classification
ever, to do so, it was necessary to select the necessary 1:Generally fresh
and unnecessary areas in the face image and crop the 2:Weathered or altered along the crack
photographs. Considering the workability in the field, 3:Weathered or altered to the rock core
it was best to make the operation as simple as possi- 4:Weathered into soil or unconsolidated soil
ble. Therefore, the entire area of the captured image
was analyzed and sent back to the tablet. In a formal
report submission, the appearance was arranged as in
Fig.15, which shows an example of the results of the
determination of weathering alteration. The rock
strength was about 110 MPa, and primarily com-
posed of andesite. Conventionally, Fig.9 shows how
the rock mass was divided into three parts.
In this case, the center area corresponded to cate-
gory 2, “weathering alteration along the fracture,”
and the left and right areas were considered to corre-
spond to category 3, “weathered or altered to the rock Fig.15 Output heat map of tunnel face evaluation.
core.” By contrast, in Fig.15, the entire face was sub-
divided into 227 × 227 pixels for evaluation. For ex- cases where two-rank category deviations were ob-
ample, the area to the right of the center region was served between the real and predicted evaluations. As
rated as category 3, indicating that weathering altera- there might be some inappropriateness in the creation
tion had progressed locally. As the example shows, it of the training data, this would remain an issue for
was possible to predict the degree of weathering lo- future research. Moreover, the Japanese standard for
cally and specify the parts of the face that have to be rock mass determination was adopted here. In the fu-
considered in detail. During the actual construction, ture, it will be necessary to adapt the system to the
a slide collapse of a small rock mass was observed on global rock mass determination technology such as
the right side of the center, which was caused by rock mass rating, geological strength index, and rock
weathering and alteration. Therefore, the visualiza- mass quality. In the future, it is hoped that the ratio
tion of the prediction of local conditions by this sys- of the training set to the test set will be changed and
tem was considered significant for the safety man- the validation set will be used for re-training to create
agement of construction.
an optimal learning model for rock mass evaluation.
In parallel, this research has attempted visualization
with Google’s Deep Dream. However, at this point,
8. SUMMARY AND CONCLUSION it is not possible to answer with sufficient certainty
what physical quantities in the face image were
In this research, AI technology was tested for tun- determined to be features. In recent years, there has
nel face evaluation. The analysis algorithm used was been a lot of research on explainability using methods
AlexNet, which is a DNN. Moreover, supervised such as Grad-CAM45). This research will continue to
learning was conducted based on the observation rec- focus on explainability.
ords of the face of mountain tunnels constructed by The output of the analysis results was a heat map
the Obayashi Corporation in the past. The condition for each area of 227 × 227 pixels. Normally, the heat
and behavior of the seven categories of rock masses map is characterized by representative values in the
were predicted with a high rate of correct answers. In three regions of the crown, left, and right areas. In
general, supervised learning requires training with a mountain tunnel faces in Japan, it is rare for the entire
large amount of original data. However, here, a high surface to show the same characteristics; in many
rate of correct answers was obtained through feature cases, local differences appear. Therefore, it was im-
analysis using only 3,000 face images. The high cor- portant to evaluate the face by subdividing it as in this
rect answer rate result was attributed to the high apti- system. Local support structures were rationally se-
tude ability of the AlexNet algorithm and transfer lected by visualizing the local stability and instability,
learning. Eight types of rock masses were analyzed which contributed to safe and secure construction.
For example, if the crack intervals and crack condi-
including sandstone, but these were still insufficient.
tions were accurately evaluated, it was possible to
To expand the range of applications and improve the
predict the rock masses that slid at the top and other
rate of correct answers, it was necessary to increase
parts of the lateral walling and prevent rock mass col-
the amount of training data. Furthermore, there were

272
lapses, such as key blocks, ahead of time by spot bolt- 15) Lawal, A. I. and Kwon, S. : Application of artificial intelli-
gence to rock mechanics: An overview, J. Rock Mech. and
ing.
Geo. Eng., Vol. 13, pp. 248-266, 2021.
An AI system for tunnel face evaluation using tab- 16) Ohtsu, H., Kobayashi, T. and Hasegawa, N. : Evaluation us-
let terminals and cloud computing was developed to ing artificial neural network in mountain tunnel preliminary
predict rock mass properties and behavior easily and geological investigation, J. Jpn. Soc. Civ. Eng. Ser. F4 (Con-
quickly at construction sites. In the future, the system struction Management), Vol. 70 (4), pp. I_83-I_94, 2014. (in
Japanese)
will be improved and developed into a practical sys- 17) Hasegawa, N., Hasegawa, S., Kitaoka, T. and Ohtsu, H. :
tem. Applicability of Neural network in rock classification of
mountain tunnel, J. Soc. Mater. Sci. Jpn., Vol. 67(3), pp.
ACKNOWLEDGMENT: Technical assistance 354-359, 2018. (in Japanese)
18) Tsuchiya, Y., Kitaoka, T., Hasegawa, N., Pipatpongsa, T.
from MathWorks Japan was provided for this re- and Ohtsu, H. : Influence of characteristics of training data
search and development. The author is grateful for set on the accuracy of rock classification estimation for
this support. mountain tunnels by artificial neural network, J. Jpn. Soc.
Civ. Eng. Ser. F1 (Tunnel Engineering), Vol. 75 (1), pp.
REFERENCES 130-142, 2019. (in Japanese)
1) Campbell, M., Joseph Hoane Jr., A. and Feg-Hsiung, H. : 19) Krizhevsky, A., Sutskever, I. and Hinton, G. E. : ImageNet
Deep blue, Artif. Intell., Vol. 134, pp. 57-83, 2002. classification with deep convolutional neural networks in
2) Lally, A., Bachi, S., Barborak, M. A. and Buchnan, D. W. : NIPS2012, Conference and Workshop on Neural Infor-
Watson paths: Scenario-based question answering and in- mation Processing Systems (NIPS), Lake Tahoe, Nevada,
ference over unstructured information. IBM Research Re- USA, 2012.
port, Computer Science, IBM Research Division, Thomas J. 20) Rabcewicz, L.V. : The new Austrian tunnelling method. Wa-
Watson Research Center, Yorktown Heights, New York ter Power November, pp.453-515, 1964.
(NY), 2014. 21) Rabcewicz, L.V. : The new Austrian tunnelling method. Wa-
3) Hurwitz, H., Kaufman, K. and Bowles, A. : Cognitive Com- ter Power January, pp.19-24, 1965.
puting and Big Data Analytics, Wiley, Hobken (NJ), 2015. 22) Tanimoto, C. and Yoshioka, H. : Meaning of convergence
4) Kong, J. S. and Managi, S. : Consumer demand for fully au- measurement in tunnelling, J. Soc. Mater. Sci. Jpn., Vol. 40
tomated driving technology: Evidence from Japan, RIETI (452), pp. 630-636, 1990. (in Japanese)
(Research Institute of Economy, Trade and Industry), RIETI 23) Tanimoto, C. and Yoshioka, H. : Rock mass classification
Discussion Paper Series 17-E-032, 1-33, 2017. based on convergence measurement in tunneling, J. Soc.
5) Takahasi, T. : Artificial intelligence/robots and social im- Mater. Sci. Jpn., Vol. 41 (463), pp. 424-429, 1992. (in Jap-
pacts: Is human first innovation wishful thinking?, J. Inf. anese)
Sys. Soc. Jpn., Vol. 14 (2), pp. 7-17, 2019. (in Japanese) 24) Tanimoto, C., Fujiwara, T., Yoshioka, H. and Hata, K. : De-
6) Terano, T. : Survey of artificial intelligence research, past, termination of rock mass strength through convergence
present, and future―From artificial intelligence to popula- measurements in tunneling, Proceedings of the 2nd Interna-
tion intelligence, J. Phys. Soc. Jpn., Vol. 74 (7), pp. 454- tional Symposium on Field Measurements in Geomechan-
462, 2019. (in Japanese) ics, Kobe, Japan, pp. 1069-1078, 1987.
7) The Japanese Society for Artificial Intelligence. Resource. 25) Michihiro, K., Hata, K., Fujiwara, T. and Zenitani, Y. :
What’s AI (JP), 2017. https://www.ai-gakkai.or.jp/whatsai/ Study on estimating the initial stress and residual stress of
AIhistory.html (Accessed 1 October 2021) rock masses, Proceedings of the Third International Con-
8) Mandelbrot, B. B. : The Fractal Geometry of Nature, W. H. ference on Residual Stress, Tokushima City, Japan, pp. 399-
Freeman, New York (NY), 1983. 409, 1991.
9) Zadeh, L. A. : Syllogistic reasoning in fuzzy logic and its 26) Michihiro, K., Hata, K., Fujiwara, T. and Yoshioka, H. : Pre-
application to reasoning with dispositions, Berkeley Cogni- diction of failure in rock masses by acoustic emission, Ma-
tive Science Report, No. 34, 1984. https://www2.eecs. ter. Sci. Res. Int., Vol. 3 (2), pp. 106-111, 1997.
berkeley.edu/Pubs/TechRpts/1984/ERL-m-84-17.pdf (Ac- 27) Hata, K., Michihiro, K., Yoshioka, H. and Sugihara, K. :
cessed 1 October 2021) Application of acoustic emission technique for estimation of
10) Hirota, K. and Sugeno, M. : Industrial applications of fuzzy initial stress, J. Soc. Mater. Sci. Jpn., Vol. 44 (502), pp. 885-
technology in the world, Tokyo Institute of Technology, 890, 1995. (in Japanese)
Singapore World Scientific, 1995. 28) Kaiser, J. : Erkenntnisse und Folgerungen aus der Messung
11) Holland, J. H. : Adaptation in Natural and Artificial Systems von Gerauschen bei Zugbeanspruchung von metallischen
- An Introductory Analysis with Applications to Biology, Werkstoffen, Arechiv fur das Eisenhuttenwesen, Vol. 24,
Control, and Artificial Intelligence, MIT Press, Cambridge pp. 43-45, 1953.
(MA), 1992. 29) Goodman, R. E. : Subaudible noise during compression of
12) Ignizio, P. J. : Introduction to Expert Systems: The Develop- rocks, Geol. Soc. Am. Bull., Vol. 74 (4), pp. 487-490, 1963.
ment and Implementation of Rule-Based Expert Systems, 30) Nakaoka, K., Hata, K., Akutagawa, S. and Takahashi, S. :
McGraw-Hill Education, New York (NY), 1991. Application of strain softening analysis to evaluate the avail-
13) Hata, K. and Yoshioka, H. : Development of support system ability of tunnel face bolts, Proceedings of the EURO: TUN
for selection of tunnel work, Tunn. Undergr., Jpn. Tunn. As- 2009, 2nd International Conference on Computational
soc., Vol. 27(1), pp. 65-71, 1996. (in Japanese) Methods in Tunnelling, Bochum, Germany, pp. 1-6, 2009.
14) Bieniawski, Z. T. : The geomechanics classification in rock 31) Kuwahara, T. and Hata, K. : Geological prediction ahead of
engineering applications, Proceedings of the 4th Interna- the tunnel face by the exploration drilling system during the
tional Congress on Rock Mechanics, Montreux, Switzer- tunnel excavation, ITA-AITES 2010 World Tunnel Con-
land, pp. 41-48, 1979. gress, Vancouver, British Columbia, Canada, 2010.
32) Kuwahara, T., Hata, K., Nakaoka, K. and Inagawa, Y. : New

273
 geological prediction ahead of the tunnel face by the explo- derground Excavations in Hard Rock, A. A. Balkema, Rot-
ration drilling system during the tunnel excavation, Int. J. terdam, 1995.
JCRM., Vol. 8(1), pp. 25-30, 2012. 40) Barton, N., Lien, R. and Lunde, J. : Engineering classifica-
33) Kuwahara, T., Hata, K., Tanaka, K., Nogashira, H. and Fu- tion of rock masses for the design of tunnel support, J. Rock
jikawa, H. : New ground evaluation and construction man- Mech. Geotech., Vol. 6 (4), pp. 189-236, 1974.
agement using geological prediction ahead of the tunnel face 41) Hata, K., Sugiura, S., Goto, N. and Fujioka, D. : Develop-
in mountain tunnelling, Proceedings of the ATC 2013, Ara- ment and application of tunnel predictive CIM using infor-
bian Tunnelling Conference & Exhibition 2013, Dubai, mation and communication technology, J. Jpn. Soc. Civ.
UAE, pp. 166-175, 2013. Eng. Ser. F3 (Civil Engineering Informatics), Vol. 71 (2),
34) Hata, K., Yoshioka, H., Kinashi, H. and Yuyama, S. : Study pp. II_78-II_85, 2015. (in Japanese)
on estimating the loosened region around rock cavern by 42) Mordvintsev, A., Olah, C. and Tyka, M. : Deep Dream – A
acoustic emission technique, Proceedings of the 11th AE code example for visualizing neural networks, https:// re-
Conferences, pp. 263-268, 1997. (in Japanese) search.googleblog.com/2015/07/deepdream-code-example-
35) Hata, K., Yoshioka, H. and Kinashi, H. : Study on evalua- for-visualizing.html (Accessed 1 August 2021), 2015.
tion of loosened area around rock cavern by in situ acoustic 43) Hata, K., Hashimoto, S. and Nakamura, S. : Development of
emission measurement, J. Soc. Mater. Sci. Jpn., Vol. 47 (5), tunnel convergence measurement method by CCD camera,
pp. 489-494, 1998. (in Japanese) Proceedings of the 11th Japan National Symposium for
36) Hata, K., Matsui, H., Kiyama, H. and Kinashi, H. : Evalua- Rock Mechanics, 2002. (in Japanese)
tion of damaged zone around underground structures due to 44) Nakamura, S., Aoki, Y., Hashimoto, S. and Hata, K. : Image
excavation by acoustic emission technique, J. Jpn. Soc. Civ. correction for high accurate image measurement with digital
Eng., No. 715/III-60, pp. 251-262, 2002. (in Japanese) image, J. Inst. Image Electron. Eng. Jpn., Vol. 31 (4), pp.
37) Hata, K., Nyunoya, S., Aoyagi, K. and Fujita, T. : Evalua- 534-541, 2002. (in Japanese)
tion of EDZ (Excavated damage zoon) by multi-optical 45) Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R.,
measurement probe in Horonobe Underground Research Parikh, R. and Batra, D. : Grad-CAM : Visual Explanations
Center, Proceedings of the 44th Symposium of Rock Me- from Deep Networks via Gradient-Based Localization, 2017
chanics, Fukuoka City, Japan, pp. 319-324, 2016. (in Japa- IEEE International Conference on Computer Vision
nese) (ICCV), Venice, Italy, pp. 618-626, 2017.
38) Standard Specification for Tunneling-2016: Mountain Tun-
nels, Japan Society of Civil Engineers, pp.208-210, 2016. (Received October 29, 2021)
39) Hock, E., Kaiser, P. K. and Bawden, W. F. : Support of Un- (Accepted February 28, 2022)

274