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

223

4

at the expense of being able to operate on outlying geometries such as side passages. As the current state of the art, the workflow developed by (Smith et al., 2023b) was also analysed in this study. Further to these methods, neural networks have been developed that can operate directly on 3D point clouds. These must deal with three additional complications over 2D images: • Points are unordered. As a result, a network must be able to identify and group similar features for varying permutations of different shapes. • Clouds can be sparse. There must be an efficient way of grouping points when there are substantially varying point densities across a cloud. • Point clouds are unstructured. This means that computationally expensive nearest neighbour calculations must be conducted to characterize spatial relationships. Many network architectures have been developed to deal with these issues. KPCONV (Thomas et al., 2019) has been chosen for this study, as it achieves state of the art performance, while being easily tuneable to a target application. KPCONV applies a 3D convolution that can operate with a varying number of points directly to the point cloud. These convolutions can also be made deformable, which allows them to adapt their shape to different local geometries. A comparison of the key methods trialled in this study is shown in Table 1.

Table 1. Comparison between different masonry joint segmentation algorithms. Method Cloud preprocessing requirement Description

Relative cost of computation

Background

XGBOOST

K Nearest neighbour points must be calculated within specified radius. 3D roughness features must be calculated for each point, as outlined by Hackel et al. (2016). Tunnel point cloud must be unrolled and then rasterized into a 2D Image. Workflow developed by (Smith et al., 2023b) K Nearest neighbour points must be calculated within specified radius.

Medium

Initially released in 2014, the XGBOOST library (Chen and Guestrin, 2016) was developed to improve the performance of decision tree methods. Developed by Ronneberger et al., (2015), the U-Net was the first semantic segmentation network to show excellent transfer learning performance. Thomas et al., (2019) developed KPCONV in order to investigate how image convolutions could be adapted for sparse 3D point clouds

Boosting decision tree ensemble method. (point feature based machine learning)

U-NET

2D encoder-decoder convolutional neural network (2D image deep learning)

Medium with GPU

KPCONV

High with GPU

3D encoder-decoder convolutional neural network with adjustments to work with sparse and uneven point clouds (3D point cloud deep learning)

3. Dataset A 3D point cloud of a 19.88m section of masonry lined tunnel obtained by TLS was provided by Bedi Consulting Ltd. for this research. The dimensions of the tunnel segment are shown in Fig. 2a. The tunnel is located in the west of England and carries a double track railway line through an urban area. The tunnel is predominantly lined with limestone blocks dating from its construction in the 1850s and contains multiple historic repairs. As a result, some areas of the sidewalls have been infilled with blocks of smaller brick masonry and there is a significant volume of spalling present. Spalling severity is classified into multiple categories depending on the spalling depth. For Stone masonry, low severity is typically defined as <40mm deep and its location only needs to be monitored. Higher severities such as Medium (>40mm) and high (>100mm) will require surface repairs or local reconstruction. This tunnel contains large areas of low severity spalling and small areas of medium severity spalling. Efflorescence is also present, but with no other major defects the overall condition is acceptable for continued railway operation. The masonry joint locations were labelled onto this point cloud as shown in Fig. 2b.