BTS | HARDING PRIZE COMPETITION 2023
I decided to use the Southwark Tunnel project as a
proof-of-concept study prior to its further development on the NBRS. The Southwark Tunnel data included TBM data, which can be used to explore additional data science applications. The project involved excavation below the River Thames at depths of circa 30m, and the tunnel featured a minimum curve radius of 150m. The primary ground conditions were London Clay, Chalk, and the Lambeth Group.
What is Machine Learning? A technology that can improve civil engineering’s efficiency is machine learning.1
Neural networks are inspired by the brain, which itself
is made up of neurons – signal transmission pathways. Here, for modelling analysis, ‘neurons’ are mathematical functions – algorithms – that accept weighted inputs, sum them, pass them through an activation function, and output a result. The networks are made from several such ‘neurons’ connected together. Figure 2 shows a typical neural network structure. To learn, a neural network adjusts the weights of
This is a field of artificial
intelligence that fits data onto a model to understand its underlying patterns, effectively learning from data over time.
By using such data-driven approaches, complex
behaviours can be captured, even predicted. The arrival of vast amounts of data, powerful
computers, and a resurgence in artificial intelligence research has led to machine learning becoming a powerful tool in multiple industries. Whilst sectors like automotive2
, healthcare3 finance4 , and are active in machine learning, construction
has only recently begun to adopt artificial intelligence in select applications5
is machine vision, which is predominantly used for safety 6
, however research in construction output is
lacking and there is ample opportunity to leverage the new technology. Studies have been conducted that use machine
learning to predict certain aspects of tunnelling performance.7,8
However, a key limitation of these is
that they are inputted with the machine parameters, which are data that can only be used once a project has commenced. Another limitation is that these methods rarely
forecast sufficiently far ahead on a project, such as to the end of tunnel excavation. Instead, they are focussed on predicting the duration of the current ring being erected. This presents an opportunity to apply machine learning to tunnelling by focusing on the geological aspect and, further, to then apply it to tendering, improving the understanding of a project’s potential productivity and challenges. Geology is difficult to predict, as only a snapshot
is available at the tender stage of a project, but new modelling techniques like neural networks could be used to build an understanding of the ground conditions.9
Neural Network Theory Neural networks are one of the most popular machine learning algorithms as they can handle multiple types of problem (such as regression and classification), yield high accuracy, and can be updated as new data are discovered. One of the most popular applications for neural
networks is object detection, with facial recognition on a phone being a prime example.
34 | August 2023 . One of the primary applications
It is important to not use all the available data for training the neural network as neural network performance needs to be validated by testing it on data it has never seen before, to ensure that it has captured the underlying patterns of the data. This train-test split is usually 70:30, as this provides the neural network with enough information to generalise across new datasets without relying on memorising the data, which is called ‘overfitting’. The impact of training is shown in Figure 3.
Scope of the study Having established its potential benefits, regarding understanding productivity, Murphy Group was interested in applying machine learning for tunnel boring yet wanted to understand its capabilities and possible applications. The research question it wanted addressed was: What is the feasibility of machine learning to forecast tunnelling performance? This question was addressed by the first case study,
on the data from Southwark Tunnel. The study had the aim to: investigate the feasibility of neural networks for predicting multiple types of output, using both machine and geological data. Next, there is a gap in the literature that needed to be
addressed – projecting tunnelling performance without TBM performance data. Current literature relies on TBM performance data
that have been generated already in specific ground conditions. When the ground conditions are new, there are not much data. How the machine behaves in these conditions is unknown, requiring additional prediction and potentially lower predictive accuracy. Instead, the data available are the ground conditions and the tunnel plan. This led to my second case study, at NBRS, with the aim to: use a geology-based approach for forecasting the tunnelling productivity.
each neuron input. This is performed during training, using a process called backpropagation: ● A data point is passed forward through the neural network to generate a prediction at the output.
● This prediction is compared to the actual value of the data at that point to generate an error.
● The error is moved backward through the neural network, slightly adjusting the weights of each neuron input. The size of this adjustment is known as the learning rate.
● The same data point is passed forward again, generating a slightly closer prediction.
● This process is repeated until the weights yield predictions that are close to the actual values.
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49
