Principles and applications for monitoring and assessing underground cavities sites

Dr. Zili Li

​

Lecturer (tenured Assistant Professor)

in Geotechnical Engineering

University College Cork

Irish Centre for Applied Geoscience (iCRAG)

​

zili.li@ucc.ie

1

Who am I ?

Zili Li

Professional areas:

Underground space technology, Innovative monitoring technology,

Internet of things, big data analytics in geotechnics

Content

​

​

​

​

​

• Wireless sensor network
• Distributed fibre optic sensing
• Drone inspection and image-based automatic assessment

​

​

​

​

Innovative monitoring and assessment technologies for underground sites

Content

​

​

​

​

​

• Wireless sensor network
• Distributed fibre optic sensing
• Drone inspection and image-based automatic assessment

​

​

​

​

Innovative monitoring and assessment technologies for underground sites

Wireless Sensor Network

(in real time or continuous time)

Wireless sensor network

Data

transmitting

Gateway

Displacement gauge

/ Crackmeter

Dublin Port Tunnel monitoring in partnership with TII

Mobile network

Gateway

Tunnel surface spalling

Cracks at the tunnel springline

water leakage and crack at the tunnel crown

Structural deterioration at Dublin Port Tunnel

A Cross passage at Dublin Port Tunnel

Dublin Port Tunnel

WSN Monitoring the performance of Dublin Port Tunnel

The long-term ageing behaviour of a cross passage at Dublin Port Tunnel

Tunnel deteriorations

Water leakage; concrete cracks; concrete corrosion; concrete spalling

WSN monitoring

Structural inclination

Structural movement

Crack development

14 tiltmeters

2 crackmeters

2 gateways

Monitoring the performance of Dublin Port Tunnel

The long-term ageing behaviour of a critical section of Dublin Port Tunnel

WSN monitoring of aging road tunnel

Inclinometer ring on central Layby Tunnel (Northbound)

T4: B axis: 0.0100°/month, counter clockwise

T6: A axis: 0.0048°/month, clockwise

T5: B axis: 0.0049°/month, counter clockwise

The upper part of layby tunnel lining is tilting inwards the tunnel.

Observed tunnel deformation rate is much greater than that predicted by finite element models!

WSN monitoring of Underground World Heritage

WSN Monitoring for Dolmen de Antelas in Aveiro, Portugal

Dolmen de Antelas in Aveiro, Portugal

9

Structural parameter: inclination, crack development, distance change

Inclination development

inclination development

Environmental sensor: CO2, temperature, relative humidity

Wireless Sensors

CO₂ change during experiment

Crack development with time

10

Content

​

​

​

​

​

• Wireless sensor network
• Distributed fibre optic sensing
• Drone inspection and image-based automatic assessment

​

​

​

​

Innovative monitoring and assessment technologies for underground sites

Fibre optic cable

Distributed fibre optical sensing (in 3D continuous space)

Optical Analyser

​

• Continuous strain measurement
• Low cost of the optic fibre cables
• Resistant to corrosion and water
• Easy to integrate/install

…

​

Main advantages:

Diaphragm wall cage

Spool for cables

Fibre optical cables

Comparison between DFOS and other field data

Incremental lateral displacement along the D-wall

Fibre optic cable deployment at CERN Tunnel

the European Organization for Nuclear Research

near Geneva, Switzerland

CERN: Conseil Européen pour la Recherche Nucléaire

2013

Higgs boson

Installation of Fibre Optic Sensors: CERN tunnel

Optical analyser

Cables loops

Shaft

OUTPUT

Axial strain

Tunnel side

CUT DATA

INPUT

Fibre optical data analysis

Crown

Tunnel side

Development of tunnel strain with structural deterioration in the long term

Content

​

​

​

​

​

• Wireless sensor network
• Distributed fibre optic sensing
• Drone inspection and image-based automatic assessment (by Ms. Ran Zhang)

​

​

​

​

Innovative monitoring and assessment technologies for underground sites

Tunnel Cracks

Camera on a robotic arm at CERN

Image data acquisition methods

Manual inspection and take pictures

Drone tunnel inspection

Tunnel shaft inspection at CERN

Drone inspection

Centreline location for drone flight

Pilot certificate for drone operation

• Tunnel condition for drone inspection
• No wind
• Low luminance (Compensated with spotlight)
• Low / denied GPS signal
• Obstacle avoidance (obstacle distance, stop at 1m)

​

Drone tunnel inspection

• Drone tunnel inspection procedure
• Crack and surface inspection using optic camera
• Water seepage inspection using infrared camera
• Flight at different heights
• Flight at constant distance to the tunnel wall

infrared camera

No spotlight

With spotlight

Photos taken at different heights

Content

​

​

​

​

​

• Wireless sensor network
• Distributed fibre optic sensing
• Drone inspection and image-based automatic assessment

​

​

​

​

Innovative monitoring and assessment technologies for underground sites

Image pre-processing procedures

Fig 1 original image

Fig 2 after radical filter

Fig 3 after dehaze

Apply the same setting to all images in one task

Image stitching

4 Original images

1 Stitched image

• ~350 tunnel images
• Each image was cropped to 256 x 256 pixels in uniform size
• A database of 7600 small images was created

Create a Data Base- Image segmentation

Content

​

​

​

​

​

• Wireless sensor network
• Distributed fibre optic sensing
• Drone inspection and image-based automatic assessment

​

​

​

​

Innovative monitoring and assessment technologies for underground sites

In Deep Learning, the most commonly adopted method of image analyzation is based on Convoluted Neural Networks(CNN).

Convoluted Neural Networks(CNN)

Unlike conventional CNN development from scratch, the proposed CNN incorporates transfer learning in the form of the VGG16 model with weights transferred from ImageNet.

Transfer Learning

The transfer learning models are trained using over 10k+ images validated on 2’500 images with 256 x 256 pixels in uniform size

ImageNet:  More than 14 million images

VGG16 model

Original Image

Scanning window Pass 1

Original Image

Prediction Made

CNN for crack Detection

The CNN achieved a accuracy of high results in training of 98.7 %, and out of sample accuracy of 97.9%. The integration of scanning window method into the CNN increased the accuracy.

Content

​

​

1. Innovative field monitoring technologies for big data acquisition

​

​

2. Image processing and deep learning

• Image pre-processing
• Crack detection
• Crack classification

​

​

​

Automatically classify the cracks into 4 categorise

​

following road tunnel civil engineering inspection guidebook:

​

1. Vertical structural cracks
2. Horizontal structural cracks
3. Diagonal structural cracks
4. Crescent/Complex structural cracks

​

Vertical

Horizontal

Diagonal

Complex

Crack types

Crack Classification

1. Vertical structural cracks= Blue box
2. Horizontal structural cracks= Red Box
3. Diagonal structural cracks =Green Box
4. Complex structural cracks = Black Box

The test metrics for the crack classification model

• CNN model training : 1.5h (55 epochs)
• Scanning window prediction:
• 11 min for task 3: 33 images of 501 MB
• 25 min for task 4: 74 images of 1.24GB
• 50 min for task 5: 150 images of 2.18 GB
• 5 min for task 6: 17 images of 239 MB
• 20 s per image

Crack classification result

Summary

​

• Demonstrate the acquisition of comprehensive geotechnical & environmental ‘big’ data in real time using wireless sensor network (e.g. inclination, crack development, temperature)

​

• Distributed fibre optic sensing acquires ‘big’ strain data of tunnel infrastructure in 3D space and reveals the time-dependent tunnel displacement profile

​

​

• Drone inspection enables to gather large numbers of tunnel images following a standard procedure at a lower cost of time and labour.

​

• The CNN algorithm is trained upon small images in uniform size. The scanning window technique facilitates the crack classification in a large image and increased the CNN prediction accuracy.

​

• Compared to conventional CNN model trained from scratch, the transfer learning CNN model achieves a higher validation accuracy of 97.9% at a lower computational cost (e.g. 55 epochs).

Thank you for your attention.

Dr. Zili Li

​

zili.li@ucc.ie

​

40