HARDING PRIZE COMPETITION 2023 | BTS
Hidden Dendrites Nucleus
Cell body Axon
Axon terminals in1 in2 in3 Above, figure 2: A neuron in the brain (top)10 , modelled for a neuron (bottom) that is part of a neural network (right) Out Input Output
the situation of data availability at the pre-tender stage of a project, where having such a system to support making a reliable forecast on tunnelling progress would have been valuable to a business looking to bid on such a contract. The results of the studies showed neural networks to
be successful at forecasting tunnel boring completion and the system was also able to capture spikes in excavation performance. This was particularly impressive as the spikes captured were from geological data points not been previously collected: gypsum and sandstone. Additionally, neural networks demonstrated excellent
generalisation when applied to Southwark Tunnel and were again able to accurately forecast the behaviour of the TBM. This project demonstrated the feasibility of artificial intelligence and highlighted promising applications in tunnel boring to address productivity and improve efficiency, which are both being taken forward by Murphy into further work.
INTRODUCTION Project Background I studied an MEng in Product Design Engineering at Loughborough University, with a placement year at JCB Excavators as an engineer. I later pursued a PhD in Artificial Intelligence for Robotic Excavation, and it was at this time that I gained interest in tunnelling. Despite
having no tunnelling experience at the time, I was driven to apply my newfound skillset in artificial intelligence to tunnelling. Murphy Group engaged with me to undertake a study
to investigate the feasibility of machine learning for forecasting tunnelling. This is a new area for Murphy Group, as well as the tunnelling industry, but the potential benefits are numerous, with more accurate, data-driven forecasting the main benefit. The TBM used in this exercise – for both projects
discussed – was an earth pressure balance (EPB) Lovat RMP131SE machine, capable of building a 2.85m i.d. lining. The EPB machine is shown in Figure 1. Murphy Group were engaged on the NBRS project
for Wessex Water during the period of this work. The project had been awarded in early 2018 and involved construction of a 5.5km-long, 2.85m i.d. tunnel, requiring excavation of two 30m-deep 6m-i.d. shafts, 14 manholes, 1km of pipe jacking and auger boring drives, and 1km of open cut digging for pipework. The ground was predominantly mudstone, siltstone,
and conglomerate. Murphy Group also had the tunnelling data for
Southwark Tunnel, a previous tunnelling project undertaken that involved construction of a 9.0m diameter shaft along with the 3.0km-long 2.85m i.d. tunnel.
Underfitting
Ideal
Overfitting
Above, figure 3: If a model isn’t trained enough, it underfits and creates a poor function (left). If it is overtrained, the model tries to fit to all the data and makes a brittle model (right). A training-testing balance is needed to ensure an optimal model (centre)
August 2023 | 33
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
