PSI - Issue 64

Jack Smith et al. / Procedia Structural Integrity 64 (2024) 220–227 Smith, J., Paraskevopoulou, C./ Structural Integrity Procedia 00 (2019) 000 – 000

224

5

Fig. 2. Tunnel point cloud used for training and testing the algorithms.

The tunnel was split into a training and a testing section, each 9.94m long and with 20% of the training section reserved as a validation set. Four algorithms were tested: pointwise XGBOOST, 2D U-Net with tunnel unrolling and 3D KPCONV with and without tunnel unrolling. In the unrolled cases, A cylinder was fit to the tunnel profile using principal component analysis. The tunnel was then unrolled around a cylinder to leave the profile as shown in Fig. 3. While this is not completely flattened, this representation enables rasterisation onto a 2D image without substantial distortions. As the distortion is present in both the training and testing data, the neural networks learnt to operate on images with the distortions present. Early stopping was used for all of the algorithms and training was conducted until there was no improvement in accuracy within the trailing 20 epochs.

Fig. 3. Unrolled 3D tunnel point cloud with Z co-ordinate set as scalar field.

4. Methods 4.1. XGBOOST

Features were created for the XGBOOST classifier using the method outlined by Hackel et al. (2016). 0.05m and 0.01m were selected as the nearest neighbor radii., as the average joint width within the dataset is 0.01m. These distances should therefore characterize both features solely within a joint and those that represent how a joint appears within the context of the broader point neighborhood. In addition to the spatial features developed by Hackel et al. (2016), surface roughness was used. This is defined as the offset of the target point from a 2D plane fit to the remaining points within the neighborhood. After the features were calculated, the number of block and joint points was balanced to ensure equal importance during the training. The XGBOOST model was then trained with a learning rate of 0.1 for 599 epochs.