HARDING PRIZE COMPETITION 2023 | BTS


Table 1: Data inputs and outputs for the neural network on Southwark Tunnel Inputs


Average Head Torque (MN.m) Average Propulsion (kN) Average RPM (RPM) Radius (m) Invert (m)


Ground Level (m)


Distance Previous Crossing (m) Ground Water (m) Dry Density (kg/m3 Cohesive or Gran Permeability Permeability


)


Table 2: Data inputs and outputs for the neural network on NBRS Inputs


Residual Soil


801 - Mudstone 802 - Siltstone 803 - Sandstone Siltsand


804 - Limestone 807 - Breccia


808 - Conglomerate


816 - Gypsum/Rocksalt I/II III


IV/IVa/IVb MF


FI Average


Distance from Aquifer (m) Distance from crossing (m)


Outputs Average EPB (bar)


Average Excavation Rate Excavation Time (mins)


Outputs Ring Cycle Duration (mins)


These aims were chosen to utilise new technology


in a novel application that would help improve the efficiency of a tunnelling project by providing data- driven solutions to understand productivity. This project adds value by addressing a gap in the literature, building on existing research.


METHODOLOGY Datasets Datasets from two studies were used to train the neural networks. The first study is Southwark Tunnel and the input


data is summarised in Table 1 which contains TBM performance parameters, as used in previous literature. A use-case for the Southwark Tunnel data is to look


at on-going excavation works and machine behaviour analysis. A model built on TBM performance data can help improve the operating parameters of a machine in real-time, and future work could utilise this knowledge for machine automation. The flexibility of neural networks is demonstrated by predicting three different tunnelling metrics. There are also limited ground condition variables, meaning the model will have to accurately understand the machine. The second study is from NBRS, with the data used


summarised in Table 2. This was supported by Maxwell Geosystems ground survey data, which provided rich detail of the ground being excavated. Maxwell was commissioned by Murphy Group to


provide a 3D model that assigned a probability of each geology being encountered in 20m lengths along the tunnel. This resulted in the production of models that contained sections of the expected tunnel face with a confidence level of each geology to be encountered. No TBM performance data were used, making the primary use-case a forecasting project, and only ring cycle duration needed as a prediction.


Data preparation was required before the neural


networks can be run. First, any missing datapoints had to be dropped as missing values can cause problems. It is better to drop data rather than to populate gaps as it is less likely to impact the training. Next, the data were scaled between -1 and 1, which allows different magnitudes of data inputs to be comparable on the same scale without losing the impacts of any changes. After an output is generated, it is rescaled to reflect its real-world value.


Training and Deployment Both case studies have unique methodologies, due to their data train-test split requirements. For the first case study, the model was trained on


1500 datapoints. One datapoint represents a 1m length of tunnel and 500 such datapoints were used to test the performance of the learning algorithm, where the model can make amendments to its parameters. Then, performance was finally determined on the remaining 950 datapoints, to ensure that the model had received sufficient data to capture the underlying patterns of the data. Overtraining on the same data causes models only to


memorise all the data hinders predictive performance. The training needs to be sufficient to ensure that the model can be generalised for application on new datasets, which a crucial aspect for real-world deployment. For the second case study, the model was trained on


1000 datapoints, tested on 700 datapoints, and validated on 500 datapoints. These started from Ch.1021on NBRS project, which is at the start of the tunnel. The dataset was used to train several variations of


neural networks (‘machine learners’), across different parameters, until an accuracy of >85% was achieved and that best performing network was selected for


August 2023 | 35


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