Design of structured meshes of mining excavations based on variability trends of real point clouds from laser scanning for numerical airflow modeling

Various technologies are used to acquire and process 3D data from mining excavations, such as Terrestrial Laser Scanning (TLS), photogrammetry, or Mobile Mapping Systems (MMS) supported by Simultaneous Localization and Mapping (SLAM) algorithms. Due to the often difficult measurement conditions, the data obtained are often incomplete or inaccurate. There are gaps in the point cloud due to objects obscuring the tunnel. Data processing itself is also time-consuming. Point clouds must be cleaned of unnecessary noise and elements. On the other hand, accurate modeling of airflows is an ongoing challenge for the scientific community. Considering the utilization of 3D data for the numerical analysis of airflow in mining excavations using Computational Fluid Dynamics (CFD) tools, this poses a considerable problem, especially the creation of a surface mesh model, which could be further utilized for this application. This paper proposes a method to create a synthetic model based on real data. 3D data from underground mining tunnels captured by a LiDAR sensor are processed employing feature extraction. A uniformly sampled tunnel of given dimensions, point cloud resolution, and cross-sectional shape is created for which obtained features are applied, e.g. general trajectory of the tunnel, shapes of walls, and additional valuable noise for obtaining surfaces of desired roughness. This allows to adjust parameters such as resolution, dimensions, or strengths of features to obtain the best possible representation of a real underground mining excavation geometry. From a perspective of Computational Fluid Dynamics (CFD) simulations of airflow, this approach has the potential to shorten geometry preparation, increase the quality of computational meshes, reduce discretization time, and increase the accuracy of the results obtained, which is of particular importance considering airflow modeling of extensive underground ventilation networks.


Introduction
The affordability of laser scanners has led to their growing popularity in the 3D reconstruction of various environments.Currently, the most commonly used technologies in underground mines are Terrestrial Laser Scanning (TLS) [1,2] and Mobile Laser Scanning (MLS) [3,4].The first method is used to map areas that require high resolution and accuracy, but over smaller areas [5].For larger areas, MLS is a more effective technique, providing quick and easy measurement, but with a lower accuracy of a few centimeters.Due to the compactness and ease of use of MLS solutions, they are increasingly used in demanding mining environments [6].1295 (2024) 012006 IOP Publishing doi:10.1088/1755-1315/1295/1/012006 2 Due to various reasons related to the specific environment of underground mining plants, the acquisition of good quality measurement data is a challenge.These include environmental issues, such as high temperature, humidity, dust, and harmful gases in the air [7,8], that affect the ability to perform measurements over a longer period, both from the perspective of the equipment and the operator.Technical issues include the lack of developed power supply infrastructure, lack of GPS or WiFi coverage, etc.Even if the measuring devices have a power source and can be moved between the measurement points relatively easily, other issues should also be taken into account, such as conducting measurements covering the technological process of the mine and limitations related to the impossibility of interfering with the production process, the scale and spatial extent of the tested objects [9,10], or the possibility of unpredictable, potentially even dangerous situations [11,12] in such a hazardous environment [13].
For the reasons mentioned above, the 3D data acquired in underground tunnels using LiDAR technology is often incomplete.Sensors capture moving vehicles, people, or other objects that obscure the tunnel, causing gaps or gaps in the point cloud.The resulting point cloud may include noise as a result of its precision.In this case, it is necessary to clean up the point cloud by filtering and manually removing unnecessary elements.All the processing steps mentioned are time-consuming and require considerable effort.
CFD applications are commonly used in different fields of the mining industry, particularly in the health and safety area.To ensure the safety and comfort of miners, mine ventilation systems undergo CFD testing to assess their design [14].CFD can be used to examine the most effective location tactics for identifying spontaneous heating in longwalls [15].The planning of the removal of gases resulting from combustion is essential to eliminate and prevent fires in underground workings, where numerical modeling plays a vital role in achieving this goal [16][17][18][19][20]. Problems related to the distribution of methane in the working face can also be investigated using CFD [21], as well as the airflow and temperature distribution in the cabins of mining machinery [22].
Successful applications of both laser scanning and CFD modeling in the mining industry created the intention to combine the results of these techniques [23] to obtain a synergy effect.As a result of increasing the accuracy of mapping the geometry of mining excavations through measurements using laser scanning, it is possible to increase the accuracy of airflow simulation using CFD.However, the execution of this concept generates some issues that require a solution.Creating a high-quality mesh model from 3D data is a challenge.To ensure a successful CFD airflow simulation, the model must have uniform face sizes and a homogeneous structure without sharp edges.Ideally, the faces should be equilateral triangles, which aids in the generation of computational meshes.
After considering the issues mentioned above, the authors recognized the need for a method to address them.The developed methodology involves creating synthetic point clouds and applying features from real measurement point clouds.This results in a uniform base without any gaps or measurement errors, which can be used to generate surface meshes made up of regular elements.From a further perspective, the geometry prepared in this way makes it possible to omit the need to adjust it at the stage of preparation for discretization of computational domains for numerical calculations, while maintaining quality parameters at the appropriate level.

Methodology
The object subjected to measurements was mining excavations of the Polkowice -Sieroszowice mine, belonging to KGHM Polska Miedź S.A.They are used as transport excavations not only for people, materials, and equipment but also for air as a direct connection between the mining areas and the intake shaft.Laser scanning was carried out in one of the three main dips with a length of approximately 4800 meters each, and a representative part of approximately 200 meters in length was selected for further analysis.The point clouds were acquired with a Livox Avia laser scanner using SLAM technology.This technique is used in the field of robotics [24] and computer vision.It is used to map an unknown environment while estimating the exact location of the sensor.It can be mounted on a robot [25] or a UAV [26], as well as a hand-held [4].The main sensors used in SLAM technology are LiDARs [27] and cameras [28].In the case of this article, the captured data were processed using the Fast LiDAR Odometry and Mapping (FAST LIO SLAM) algorithm [29].The range precision of a single point measurement is 2 cm [30].The raw point cloud is shown in Figure 1.The proposed approach to creating synthetic models is shown in Figure 2. It is based on two stages: feature extraction and synthesis.In the first stage, processing begins by aligning the point cloud along the X-axis using the Principal Component Analysis (PCA) algorithm.This alignment ensures the correct orientation of the point cloud, which will facilitate further analysis.Then the point cloud is sampled into longitudinal cross-sections, which allows a more accurate collection of its features.To ensure consistency of representation, the outermost points of each cross section are computed.The tunnel trajectory is then determined by calculating the centers of the cross sections.This step provides crucial information about the object's location and direction.To simplify the analysis, the centered cross sections are converted to a polar coordinate system.To ensure standardization and consistency, sections are interpolated in the angle domain, resulting in each section containing the same number of points.After normalization of the sections, they are segmented in the angle domain into four separate walls.The segmented and normalized sections are then converted back to the Cartesian system.This conversion allows the assembly of four-wall matrices, providing a complete representation of the tunnel structure.The results of this stage are interpolated trajectories along the walls, shown in Figure 3.These represent the complex shape of each wall and provide information about the longitudinal and transverse changes of the object.This entire process has a significant impact on understanding the characteristics and behavior of the tunnel represented by the point-cloud data.In the synthesis stage, processing begins by determining the dimensions and resolution of the point cloud to ensure adequate precision of representation.In this case, a distance between points of 10 cm is taken.Performing numerical simulations of the airflow in the complicated domains defined by a laser scanning technique is associated with certain limitations, which was elaborated in [23], where the optimal number of elements for the surface mesh model was established for 100,000 triangular elements to determine a single simulation result.Mesh models with lower resolutions could not be easily processed due to computing power limitations.Therefore, when considering a 200-meter long tunnel as in the case of this paper, the authors took as a reference to limit the number of elements in the generated mesh to the suggested number of elements.This was possible by establishing a 20 cm resolution of the resulting synthetic point cloud.However, it was noticed that during the generation of a synthetic point cloud with a resolution of 20 cm, holes appeared in some areas of the model with higher geometric variability (higher ceiling height).Therefore, a 10 cm resolution was assumed in the first stages, and this will be resampled at 20 cm in the final stage.Next, a base point cloud is created, consisting of four separate walls and a specific cross-sectional shape (i.e.rectangular, trapezoidal, etc.).The case analyzed has a trapezoidal cross-sectional shape with a longer base on the roof and a shorter base at the bottom (Figure 4a).In the next step, the trajectory determined in the feature extraction stage is applied to the prepared base point cloud (Figure 4b).Each wall is interpolated with the trajectory data to fit the specified grid of that wall (Figure 4c).A faithful representation of the tunnel geometry is important, so noise is added to each wall to introduce imperfections and random fluctuations.Then, the inverse of the rotation matrix obtained from principal component analysis (PCA) is applied to return the point cloud to its original position.In addition, additive noise is added for each wall to increase the realism of the real excavation projection.Finally, the noise structures are filtered to smooth out the interference in a controlled manner and achieve an accurate and true representation of the tunnel.The resulting point cloud is shown in Figure 4d.

Results
The resulting synthetic point cloud was resampled to have a spatial resolution of 20 cm.Synthetic point cloud, especially after noise introduction, was limited in inhomogeneity.Having a uniform and consistent point cloud is crucial to accurately create a surface mesh model from the point cloud data.The resolution of 20 cm is considered sufficient for model processing in software tools to generate volumetric mesh model and conduct airflow simulation calculations.The resulting point cloud is represented in Figure 5.The surface mesh model, shown in Figure 6a, was created using the Ball-Pivoting Algorithm (BPA) [31].To assess the precision of the synthetic mesh model, cloud-to-mesh (C2M) distances were calculated.A visual representation of these distances on the raw point cloud is shown in Figure 6b.Statistical analysis was performed by determining the mean, median, and standard deviation.The results obtained are presented in Table 1.Analyzing the statistics obtained, they indicate an average distance of the mesh model from the original point cloud of about 6 cm.Taking into account the spatial resolution of 20 cm, the value obtained is negligible.However, it is worth mentioning that the standard deviation is equal to approximately 46 cm, which suggests that there are significant outliers.This indicates that not every part of the synthetic grid model accurately represents the original point cloud.The applied methodology is a good way to reconstruct the geometries of excavations based on actual measurement data, which is confirmed by the calculated statistics.Taking into account the C2M distances, the agreement of the synthetic model with the measurement cloud of points was obtained in the overwhelming majority.However, in some places there are deviations caused by insufficiently accurate reconstruction of the synthetic model with reference to real data.This is particularly noticeable in areas of the tunnel with greater geometric variability (e.g., higher height).In addition, a high standard deviation error is observed in areas where the tunnel merged with other perpendicular excavations.This is caused by the incompleteness of the data in the form of gaps in the raw point cloud and filling and rounding in the synthetic point cloud.After all, taking into account the overarching goal of using this methodology, which is to unify the point cloud for creating surface models, this is considered a good result.
Consequently, it is possible to obtain a model merging the advantages of having data from a real scan, e.g.realistic shapes, sizes, and features.At the same time, the model is fully controlled in terms of the regular distribution of points, elimination of the missing data issue, and management of all parameters.The benefits include the ease and practicality of controlling the model for CFD flow simulations, as well as the possibility to create multiple variants of the same model for comparative analysis (i.e., noise vs. no noise, stronger vs. weaker wall features).
Future research directions will focus on better reconstruction of a synthetic model that best represents the geometry of the actual data.Research will be undertaken to develop synthetic models for long, multi-kilometer tunnels and more complex excavations with crossroads.

Figure 1 :
Figure 1: Raw point cloud acquired from measurements using Livox (a) and view inside the object (b)

Figure 2 :
Figure 2: Diagram of the methodology for creating synthetic models

Figure 3 :
Figure 3: Feature extracton stage: view of interpolated cross-sections with wall feature acquisition (displayed for the roof)

Figure 4 :
Figure 4: Synthesis stage: (a) the base structure of the synthetic point cloud, (b) point cloud with longitudinal trajectory applied, (c) point cloud with the trajectory of the tunnel and shapes of the walls applied, (d) point cloud with the trajectory of the tunnel, shapes and roughness of the walls applied

Figure 5 :Figure 6 :
Figure 5: Final point cloud (a) and view inside the object (b)

Table 1 :
Statistical analysis values for C2M distance.In this work, the authors propose a solution to the problem of processing the 3D data of insufficient quality obtained by the SLAM technique.The most common problems caused by specific and difficult survey conditions in underground mines include non-uniform sampling, noise, the existence of additional objects in the tunnel, or even major areas of missing data.The presented approach allows to mitigate those problems by analyzing the original point cloud in search of relevant features, generating a model based on a synthetic starting shape, and finally reapplying the features onto it.