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

226

7

Fig. 4. Algorithm outputs on a section of the test data compared with the ground truth.

Rigid KPCONV did not achieve an IOU substantially superior to XGBOOST, although the IOU improved when applied to the unrolled point cloud. This was likely because the network did not need to characterize features in as many orientations. Applying flexible convolutions also improved the KPCONV IOU, as the convolutions were able to adapt to the single surface nature of the data. Despite the substantial training time and accurate segmentation in some areas, the IOU of the best performing KPCONV network did not surpass the U-Net, as it failed in areas where the joints were very narrow. It is likely that the superior transfer learning ability of the U-Net is a key factor in its performance. While it is possible that a deeper version of KPCONV would be better able to characterize the multi scale features, this may lead to overfitting given the limited training data. With the small volume of training data and limited amount of training data augmentations, none of the methods could accurately segment the smaller brick joints. 6. Conclusion This study investigated three machine learning algorithms for masonry joint segmentation. A key challenge was achieving acceptable performance with limited training data. For the dataset analysed, a 3D point cloud deep learning framework, KPCONV was shown to achieve results approaching those of a 2D image-based U-Net. This suggested that small scale surface features were equally well characterized when viewed as a 2D depth map rather than as the raw 3D data. As a result, due to the substantial computational costs of KPCONV training, an unrolling and rasterizing method using off the shelf 2D image based semantic segmentation algorithms such as U-Net is recommended for analysing surface features in tunnel structures. While the U-Net and KPCONV methods perform well on the stone tunnel lining case study, it is possible that performance would decrease when the method is applied to brick masonry, due to the joint widths approaching accuracy limits of the TLS system used. We suggest that further work should be conducted to investigate how well each method operates on different masonry block materials and geometries and how effectively it can generalize when applied to types of masonry unseen in the training data. Acknowledgements The authors would like to thank Network Rail and Bedi Consulting Ltd. for collecting and sharing point cloud data of operational railway tunnels. This project was funded by EPSRC Environment DTP grant EP/T517860/1.