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

225

6

4.2. U-Net In order to apply the U-Net, the Z co-ordinate of each point was first projected onto a 2D plane and rasterized into a depth map image. This was split into 256x256 image patches which were used in a batch size of 4 for training the U-Net model. The model was pretrained on ImageNet to reduce the number of training epochs required. In order to improve the performance when trained on a small dataset, augmentations were applied to the training data; These were random vertical and horizontal flips, random brightness and contrast shifts (representing changes in tunnel profile and location of the depth map) and the addition of gaussian noise. Training was conducted for up to 400 epochs with a learning rate of 0.001, a Dice score-based loss function, and an ADAM optimizer. 4.3. KPCONV The KPCONV network was examined with both a rigid and flexible kernel. This enabled us to study the impact of kernel deformations on the results. To examine the how well KPCONV can generalize features to different plane orientations with limited but representative training data, the method was assessed with both the raw input tunnel point cloud and the unrolled point cloud created for the U-Net method. 3D versions of the 2D data augmentations applied to the 2D image data described in section 3.3 were applied to the point cloud, with the addition of random rotations around the Z axis to account for any possible tunnel orientation. The network was trained with a learning rate of 0.001. The maximum radius of the convolution was selected iteratively as 0.15m. Due to the substantial computational cost for training KPCONV, a high-performance computer capped at 48 hours activity was used for training up to 1500 epochs. 5. Results Each algorithm was tested on a 9.94m section of tunnel that was pre-processed using the same methods as the training data. Performance was assessed using the Intersection Over Union (IOU) score, which is a metric that compares the amount of overlap between the ground truth and output segmentation with the differences between them. A score of 1 is a perfect segmentation, however 0.5 is typically considered a good benchmark. Where is the number of true positives, is the number of false positives and is the number of false negatives, the IOU is calculated as shown in equation 1. = /( + + ) (1) The results are shown in Table 2 and the ground truth joint locations and algorithm outputs are shown in Fig. 4. Table 2. U-Net and KPCONV results. Network Point cloud flattened IOU Training GPU (Nvidia) Training time (mins) U-Net Yes 0.4068 GTX 970 34 XGBOOST No 0.185 None 100 KPCONV Flexible No 0.3373 V100 2880 KPCONV Flexible Yes 0.3761 V100 2880 KPCONV Rigid No 0.2472 V100 2880 KPCONV Rigid Yes 0.3001 V100 2880 While some joint locations are identified in the XGBOOST output, the IOU was substantially lower than the other methods and it can be seen in Fig. 4b that the algorithm could not distinguish between joints and other anomalous areas that effect surface roughness properties such as efflorescence and spalling. An additional downside of the XGBOOST method was the substantial computation time on the test dataset, which was 150% to 200% more than the KPCONV and U-Net methods. This was because the eigen features and surface roughness values had to be evaluated before the trained XGBOOST model could be applied.