Method and Application of Full-Space Deformation Monitoring of Surrounding Rock in Coal Mine Roadway Based on Mobile Three-Dimensional Laser Scanning

Abstract

Deformation monitoring of roadway surrounding rock is the key link to ensure the safety production of the coal mine. The traditional monitoring method can only obtain the displacement information of discrete measuring points, and it is difficult to fully reflect the spatial distribution characteristics and evolution law of surrounding rock deformation. Based on the engineering background of the extra-thick coal seam roadway in the Yushupo Coal Mine, Shanxi Province, China, this study proposes a set of full-space deformation monitoring methods for roadway surrounding rock based on explosion-proof mobile 3D laser scanning technology. Firstly, a hierarchical denoising method based on improved statistical filtering is established. The quality of point cloud data is effectively improved by region clipping, a connectivity analysis guided by multi-dimensional geometric features and adaptive density threshold three-level processing strategy. Secondly, a hierarchical point cloud registration method combining physical anchor geometric constraints and deep learning patch guided matching is proposed to reduce the registration error to millimeter level. Finally, the deformation evaluation of surrounding rock is carried out by combining the overall deformation identification with the quantitative analysis of local section slices. The engineering application results show that the deformation of the roadway floor is the most significant during the monitoring period, the maximum deformation is 90.0 mm, and the average deformation is 46.9 mm. The maximum deformation of the roof is 35.0 mm, and the convergence of both sides is asymmetric. Compared with the total station, the results show that the maximum displacement error in each direction does not exceed 5 mm, and the standard deviation is within 1.3 mm, which meets the engineering accuracy requirements of coal mine roadway deformation monitoring. This study provides a complete technical scheme for panoramic and high-precision monitoring of surrounding rock deformation in coal mine roadway.

1. Introduction

With the increasing depth and strength of coal mining, the problems of large deformation, long duration and serious damage of roadway surrounding rock caused by ‘three high and one disturbance’ have become increasingly prominent [1,2]. The deformation and failure of the roadway surrounding rock will hinder underground traffic, damage production equipment, and even cause casualties, which seriously threatens coal mine safety production. Therefore, it is of great significance to monitor the deformation of the roadway surrounding rock in time and grasp its state and development trend for ensuring coal mine safety production and economic benefits [3].

With the continuous progress of monitoring technology, the deformation monitoring method of surrounding rock in coal mine roadway has experienced an evolution process from single-point discrete measurement to full-field continuous perception [4]. There are many methods for measuring the deformation of the surrounding rock in underground roadway. According to the different characteristics of data acquisition, it is generally divided into three types: The first type is a measurement method based on discrete points, that is, using a laser range finder or a total station (TS) to measure the three-dimensional coordinates of the preset monitoring points, and evaluating the deformation state of the roadway by analyzing the spatial displacement of these feature points [5,6]. This method has high measurement accuracy, but it can only provide discrete information of preset points and cannot reflect the continuous spatial distribution characteristics of surrounding rock deformation. The second type is the measurement method based on a linear scale, which is represented by the cross measurement method or the multi-point displacement extensometer [7]. The structural deformation is characterized by monitoring the distance change between the tunnel lining surface and the fixed reference point. The third category is area-based measurement methods, typical techniques such as photogrammetry and 3D laser scanning [8]. Photogrammetry technology reconstructs the three-dimensional shape of rock mass surface from image sequences based on the visual principle, and shows unique advantages in geological structure mapping, rock mass quality assessment and geotechnical disaster identification. However, this kind of method has higher requirements for illumination conditions and surface texture. In the complex environment of large dust concentration, uneven illumination and single rock wall texture in coal mines, the accuracy of image matching decreases significantly, which restricts its wide application in coal mine roadway deformation monitoring. At the same time, the mobile laser radar system based on simultaneous localization and mapping (SLAM) algorithm, with the advantage of fast acquisition of dense 3D point clouds without fixed stations, has received extensive attention in spatial perception and dynamic change detection of underground engineering, especially in the early identification of hazards such as broken rock bolts, roof crack propagation and progressive spalling of coal walls, showing increasing application potential [9,10].

With the continuous development of science and technology, the urgent needs of one-key automatic measurement, dynamic tracking measurement and spatial scanning measurement have promoted the rapid development of three-dimensional laser scanning technology [11,12]. Laser scanning is a non-contact digital surveying technology [13,14] capable of efficiently acquiring the three-dimensional spatial coordinates of target surfaces through high-precision laser sensors, thereby enabling faithful reconstruction of real-world scenes [15]. It can not only realize the digital reconstruction of the real scene, but also effectively avoid the safety hazards caused by the operator’s need to go deep into the high-risk environment in the traditional measurement method [16]. Due to the significant advantages of the ground laser scanning system in the acquisition and monitoring of tunnel section morphology, its deformation measurement in underground engineering applications has also attracted wide attention [17,18,19]. For example, domestic and foreign scholars have carried out a series of studies on the characterization of rock mass structure by laser scanning technology. Wang et al. [20] combined the ground laser scanning (TLS) with the discrete element method to develop an automatic identification algorithm of discontinuities based on fuzzy K-means, which realized the non-contact acquisition of the geometric parameters of the tunnel surrounding rock. Ge et al. [21] established a digital elevation model based on a laser scanning point cloud and used artificial intelligence algorithms such as BPNN and SVM to quantitatively analyze the influence of geometric parameters on shear failure. Aiming at the underground roadway environment, Gu et al. [22] proposed a recognition method based on feature pre-separation. The point cloud features were separated by the P-RG algorithm, and the discontinuous surface was accurately extracted by DBSCAN clustering. These studies show that laser scanning technology has been fully applied to underground engineering for data acquisition and numerical analysis. With the increasing demand for intelligent perception of underground roadway space, Kajzar et al. [23] applied laser scanning technology to the monitoring of roadway and coal pillar stability in room-and-pillar mining. Through multi-phase scanning, all-round data acquisition of the entire mining space was realized, and the reliability and practicability of 3D laser scanning in mine deformation monitoring were verified. Camara et al. [24] proposed a framework for analyzing the displacement of surrounding rock based on point cloud data. This method captures the displacement information by establishing the spatial correspondence between multi-period scanning point clouds and then uses the comprehensive statistical values of cross-sectional displacement, drift degree of roadway central axis and standardized index of cross-section to describe the deformation evolution law of the surrounding rock. Kang et al. [25]. used the mobile laser scanning technology to collect the multi-temporal point cloud of the tunnel, realized the unification of the spatial datum through the centerline registration, and then established the spatio-temporal point pair correlation with the k-nearest neighbor matching to complete the global-local deformation analysis of the rectangular tunnel. Furthermore, Jiang et al. [26] proposed a deformation observation method of deep roadway surrounding rock based on 3D laser scanning technology, which was applied to the deep roadway of Jinchuan No. 2 Coal Mine, and gave the preliminary processing steps of roadway deformation point cloud data. However, there are many problems of point cloud noise and difficult registration, which reduce the accuracy of surrounding rock deformation monitoring.

On the whole, the total station has high precision but limited discrete coverage; TLS can achieve full-field acquisition but relies on fixed site deployment and cannot be directly applied to the coal mine environment; the cost of photogrammetry is low, but it is obviously restricted by coal mine dust and a weak texture environment. The mobile SLAM lidar takes into account both high efficiency and full-field coverage, but its Inertial measurement unit (IMU) cumulative drift error and registration accuracy in low overlap scenarios need to be further studied. In summary, domestic and foreign scholars have carried out some research on the application of laser scanning technology in underground engineering, but there are still the following limitations: (1) lack of a complete workflow for surrounding rock deformation analysis; (2) point cloud denoising, registration method is not good, the accuracy is not high; (3) it is less used in coal mines, and the use of explosion-proof performance instruments needs further research and application.

In view of the above technical gaps, this study innovatively introduces the explosion-proof mobile 3D laser scanning technology into the field of deformation monitoring of surrounding rock in coal mine roadways. The measured data of the total station are used as reference data to compare and verify the system accuracy of SLAM LiDAR monitoring results, and to evaluate the applicability of mobile 3D laser scanning technology in deformation monitoring of mine roadways. In this study, the super-thick coal seam roadway of Yushupo Coal Mine in Shanxi Province, China, was used as the engineering background. The GHJS12 mobile 3D laser scanner was used to scan the roadway in the underground coal mine, and a complete set of technical links covering data acquisition, hierarchical denoising, hierarchical registration and deformation analysis was constructed. By comparing and analyzing the two high-quality point cloud data, the deformation evolution characteristics of the surrounding rock during the excavation of the roadway are systematically evaluated, and the accuracy of the monitoring results is verified by the measured data of the total station. The main innovations of this paper are as follows:

(1)

A three-level progressive point cloud denoising method suitable for the complex environment of a coal mine is proposed. The three-level processing strategy of region clipping, improved statistical filtering and adaptive density threshold effectively improves the quality of point cloud data in an underground multi-interference source environment.

(2)

A hierarchical point cloud registration method that combines the geometric constraints of physical anchors and the guidance of deep learning features is established. Through the two-stage strategy of ‘coarse registration-fine registration’, the high-precision registration problem of the mobile laser scanning system in low overlap rate and weak texture scenes of coal mine roadway is solved.

(3)

The whole space deformation monitoring technology system of coal mine roadway surrounding rock based on explosion-proof mobile 3D laser scanning is constructed. A complete technical link from data acquisition to deformation analysis is formed, which breaks through the limitations of traditional discrete point monitoring methods and realizes high-precision deformation monitoring of the roadway surrounding rock in the whole field.

2. Engineering Background

Yushupo Coal Mine is a large-scale coal production mine with an annual output of more than 5 million tons, located in Xinzhou City, Shanxi Province, China. The mining depth of the mine is about 300–500 m. The comprehensive mechanized top coal caving (fully mechanized caving) mining technology is adopted. The roadway system is composed of air inlet roadway, air return roadway and transportation roadway, forming a complete ventilation and transportation network. The 5# coal seam of the mine is a typical ultra-thick coal seam deposit with an average thickness of 14.8 m, which has the advantages of high coal production and significant economic benefits. However, due to the large mining thickness and high mining intensity, the coal seam mining faces technical challenges such as high mining pressure and serious deformation of the surrounding rock. The mechanical properties of the extremely thick coal seam and the roof and floor strata are significantly different. Under the combined action of complex in situ stress and potential concealed structure, it is easy to produce large deformation, floor heave, long-term creep and other surrounding rock deformation problems. At present, the combined support scheme of long and short anchor cables is adopted for the roof of a deep roadway serving mining, and all anchor cables are anchored in low-strength coal seams. This support system is still difficult to completely avoid the occurrence of local abnormal deformation under the premise of overall stability (Figure 1). More seriously, the existing surrounding rock monitoring methods, such as the convergence meter and the roof abscission layer instrument, can only obtain the displacement data of discrete points, and cannot realize the continuous and full coverage monitoring of the overall deformation evolution, convergence development trend and surface damage state of the roof surface area. This technical bottleneck leads to the existence of monitoring blind spots in roadway safety management. In order to break through the limitations of traditional monitoring methods of roadway, it is urgent to introduce panoramic and high-precision three-dimensional deformation monitoring technology to obtain continuous and real deformation data of the roadway surface and improve the risk identification ability of surrounding rock stability.

In this study, a three-dimensional laser scanning monitoring test was carried out on the roadway of the 5111 return air roadway in Yushupo Coal Mine to observe the development of surrounding rock deformation during the excavation of the roadway. As a special return air channel of the 5111 fully mechanized caving face, the 5111 return air gateway undertakes the core functions of ventilation, exhaust and safe ventilation in the working face. It belongs to the roadway of an extra-thick coal seam, and the cross-section of the roadway is a rectangle of 5.5 m × 3.9 m.

There is a slope section in the roadway, and the longitudinal slope of this section is about 30°, which belongs to the area where the geometric conditions change obviously in the test area. The slope section has been deformed before 3D laser monitoring. Therefore, 3D laser scanning deformation monitoring is applied to this section to evaluate its applicability under complex roadway geometric conditions. It should be noted that the change in slope will have a certain impact on the line of sight distribution, local occlusion and sampling density in the process of point cloud acquisition, thus increasing the difficulty of multi-period point cloud alignment in local areas. However, the ‘coarse registration-fine registration’ hierarchical point cloud registration strategy adopted in this paper is to uniformly register the point cloud under the complete three-dimensional coordinate framework and does not depend on the local level of the roadway, so it does not introduce the systematic geometric bias caused by the slope itself. We apply the deformation measurement of three-dimensional laser scanning to the deformation monitoring of the surrounding rock in this section of roadway and measure the deformation of the roadway for one month. The specific monitoring roadway and monitoring section are shown in Figure 2. The H776 measuring point in the figure is the permanent anchorage point, which is the distance reference point for subsequent scanning and analysis. Figure 2 also shows the relative position relationship between 5111 intake and return air crossheadings. The two roadways are arranged in parallel to serve the same fully mechanized caving face, and the intake and return air systems are independent, which ensures the safety of underground ventilation.

Through field observation and coal rock physical and mechanical experiments, it is found that the rock mass quality of the test section is poor. The floor is sandy mudstone, low-strength soft rock, and the other three sides are solid coal walls. The strength of the surrounding rock is low, which will cause the deformation of the surrounding rock during the excavation of the roadway.

It should be noted that the mine mainly provides test scenarios and safe operating conditions. The point cloud processing flow design, parameter setting and deformation analysis method construction involved in this paper are all independently completed by the research team.

3. 3D Laser Scanning Monitoring Roadway Deformation Method and Data Processing Technology

3.1. Roadway Deformation Monitoring Method

The 3D laser scanner obtains the geometric shape data of the object by emitting a laser beam to the surface of the target object and receiving its reflected signal. The current mainstream 3D laser scanning technology can be divided into three categories [27]: pulse, phase and triangulation. In this study, we used Tianqi Changyuan (Inner Mongolia) Digital Technology Co., Ltd.’s GHJS12 (Baotou, Inner Mongolia, China) explosion-proof mobile 3D laser scanner in China. In the field of coal mine safety, explosion-proof refers to the equipment that meets the intrinsic safety (Ex i) or explosion-proof (Ex d) technical requirements of the explosive hazardous environment of gas and coal dust in coal mines, and can work normally without causing gas explosion; in China, such equipment must pass the coal mine safety mark (MA) certification before it can be used in the well. GHJS12 has obtained MA certification and meets the requirements of underground explosion-proof use in coal mines. At present, some similar mobile 3D scanning systems (such as ExScan (manufactured by Suzhou Zhangcheng Electronic Instrument Co., Ltd., Suzhou, China), SLAM200E (manufactured by China Coal (Tianjin) Mining Technology Co., Ltd., Tianjin, China), etc.) have been applied in general underground engineering, but the above equipment is mainly oriented to non-explosive dangerous environments and has not yet obtained special explosion-proof certification for coal mine gas environments, and the size of some equipment is limited in the narrow roadway of a coal mine. In contrast, GHJS12 is designed for high-risk closed environments such as coal mines and has a complete explosion-proof hardware solution, which has better applicability and safety compliance under the engineering conditions of this study.

The GHJS12 scanner is based on the SLAM algorithm, which can realize autonomous scanning and mapping without external satellite positioning assistance. It is the best choice to establish 3D data in a complex, closed environment such as a coal mine roadway. At the same time, the handheld mobile scanning method overcomes the problem that the airborne three-dimensional scanning system is difficult to move under complex underground environmental conditions and does not need to consider the point cloud splicing problem. The equipment is easy to operate and can be skillfully used after simple training. The original point cloud data output by the scanner supports general formats such as LAZ and E57 and can be seamlessly connected with mainstream point cloud processing software (such as CloudCompare (version number is V2.11.0.), LiDAR360 (Version number is standard version V9.0), etc.) to facilitate subsequent denoising, registration and deformation analysis. Therefore, this study innovatively introduces the GHJS12 three-dimensional laser scanner into the field of roadway surrounding rock deformation monitoring in complex coal mine environments and develops a data processing algorithm suitable for coal mine environments to post-process point cloud data (Section 3.3 and Section 3.4) to improve the accuracy of point cloud data processing. At present, the device realizes 360° omnidirectional scanning coverage through the horizontal rotation and vertical flip of the laser head and obtains the distance information of the target object along the optical path based on the principle of laser ranging. Its core component, the laser time rangefinder, calculates the spatial distance between two points by accurately measuring the time difference between the laser pulse emission and the reflected echo, as shown in Equation (1):

$$s={\displaystyle \frac{z\times t}{2}}$$

(1)

In the formula, the parameter s is defined as the linear distance between the laser scanning system and the measured target, the parameter z represents the velocity value of the laser on the propagation path, and the parameter t quantifies the length of time that the laser signal returns to the receiving end after being reflected by the target from the transmitting end.

The accuracy of the 3D laser scanner is determined by the measurement accuracy of the pulse laser round-trip time [28]. Therefore, this pulse-based scanner is very suitable for accurately measuring engineering structures and morphologies. The GHJS12 explosion-proof mobile 3D laser scanner used in this study is designed for engineering structure measurement under complex geological conditions, such as coal mines. The horizontal and vertical scanning range of the scanner is 360° and 280° respectively, the scanning speed is 320,000 points/s, and the operating temperature range is −20 °C to 50 °C. The scanner can work independently for 4.0 h, and the built-in battery does not need to be replaced. Based on SLAM positioning technology, the scanning accuracy reaches millimeter level, which can meet the requirements of roadway surface deformation measurement in the study (Figure 3).

What needs to be explained is that the equipment manufacturers mainly provide scanning hardware, underlying SLAM capabilities and standard format data export functions, while the point cloud denoising, registration optimization and quantitative analysis process for coal mine roadway deformation monitoring requirements are further developed and implemented by this research team. Secondly, the roadway monitored in this study is located inside the extra-thick coal seam. The surrounding rock of the roadway is mainly composed of solid coal walls. Only the local floor is sandy mudstone and the cement layer is laid in the construction. Therefore, the lithology difference in the scanning area is small. This paper focuses on the use of three-dimensional laser scanning to obtain high-density point cloud data on the surface of the roadway and analyze its geometric deformation characteristics, rather than identifying lithologic units. Since 3D laser scanning obtains spatial coordinate information based on laser reflection, most coal and rock materials can provide stable reflection signals, so different lithologies have little effect on the monitoring results of roadway surface deformation.

3.2. Roadway Deformation Monitoring Process

In order to realize the global and high-precision monitoring of the deformation of the surrounding rock of the coal mine roadway, this study established a systematic deformation monitoring process. The process covers five core links: field data acquisition, original point cloud export, point cloud data denoising, multi-phase point cloud registration and deformation quantitative analysis. Among them, field scanning and original data generation are completed by mobile 3D laser scanning equipment. The original data export is realized by supporting software of the equipment, while denoising, registration and deformation analysis are completed by the data processing flow constructed in this study. A logical, progressive and mutually coupled technical closed-loop is formed between each link. The overall process is shown in Figure 4.

In the field implementation stage, the equipment installation, initialization and test roadway scanning are first completed to obtain the original three-dimensional point cloud data of the monitoring section. Then the scanning results are exported to a general point cloud format for subsequent processing. Aiming at the common problems of dust interference, moving personnel and equipment occlusion, boundary overflow and system measurement error in the coal mine underground environment, the original point cloud is processed by hierarchical denoising to improve data purity and geometric reliability. Then, aiming at the registration difficulties of multi-period scanning data in low overlap rate, weak texture and long-distance roadway scenes, a hierarchical point cloud registration strategy of “coarse registration-fine registration” is adopted to unify the point clouds in different monitoring periods to the same spatial reference system. Finally, through the combination of overall difference identification and local section slice analysis, the spatial positioning, quantitative calculation and evolution characteristic evaluation of roadway surrounding rock deformation are completed.

In order to further clarify the boundary of responsibilities between the hardware equipment, software and the self-developed processing flow of the research team in this study, it is explained as follows: The field data acquisition stage mainly relies on the GHJS12 mobile 3D laser scanner, which belongs to the existing functions of the equipment hardware system and its supporting underlying control program. After the original scanning results are exported to general formats such as LAZ and E57 by the equipment supporting software, the subsequent data processing and analysis for the deformation monitoring of the roadway surrounding rock are completed by the research team. Specifically, aiming at the boundary overflow points, independent interference clusters and sparse outliers in the complex environment of a coal mine, this study proposes a hierarchical denoising method. Aiming at the problem that multi-period point clouds are difficult to align accurately under low overlap rate and weak texture conditions, a hierarchical registration method combining physical anchor geometric constraints and a patch-guided matching network is constructed in this study. On this basis, this study further completed the overall deformation identification, section slice reconstruction and deformation quantitative analysis. In addition to the original data acquisition and standard format export, the point cloud denoising, registration optimization and deformation evaluation methods involved in this paper are the processing flow established or improved by our team for the research problem, and also the main scientific contribution of this paper.

In addition, Yushupo Coal Mine is mainly used as an engineering application and on-site verification support unit in this study, which is responsible for providing test sites, on-site operation coordination and safety assurance; mine personnel did not participate in the data processing algorithm design and core process development in this paper. Therefore, the positioning of the mine in this study is to provide a test site, rather than the algorithm co-development subject.

3.3. Data Processing Method

3.3.1. Grading Denoising of Roadway Point Cloud Data

Based on the roadway point cloud data obtained by laser SLAM technology, through in-depth analysis of its environmental characteristics, the noise data can be summarized into three categories: boundary overflow points, independent interference clusters and sparse anomaly points [29,30].

(1)

Boundary overflow point: This kind of data is essentially the real reflection information on the surface of the roadway, but it is located outside the boundary of the preset collection range, resulting in the lack of integrity of the reconstructed three-dimensional model. The salient features of this category include: a large amount of data, wide distribution and high similarity with the target data in terms of geometric attributes.

(2)

Independent interference cluster: It refers to the non-target object point cloud that interferes with subsequent data processing. In the mine environment, it mainly comes from obstacles such as mobile personnel and transportation vehicles. Its typical characteristics are: limited space occupation range and relatively isolated location.

(3)

Sparse outliers: It is mainly due to the systematic measurement bias of the scanning device, which is randomly distributed in the periphery of the target data. The key characteristics of this type of data are: sparse quantity, low density and strong spatial dispersion.

In view of the differentiated characteristics of the above three types of noise, the corresponding hierarchical processing strategy should be adopted: for the boundary overflow points, according to the actual scene, manual screening, region clipping or conditional constraints can be used to eliminate them, and the region clipping scheme is adopted in this study; the processing of independent interference clusters usually relies on Gaussian smoothing algorithm, and density-based clustering methods (such as DBSCAN) can also be used to achieve, but such clustering techniques generally have the limitation of poor computational efficiency. Sparse outliers are suitable for effective filtering by the neighborhood radius detection algorithm.

In view of the characteristics of the above-mentioned roadway point cloud data and the problems of denoising, this study proposes a hierarchical point cloud denoising method based on improved statistical filtering. The specific implementation steps are as follows:

The first level: boundary overflow point denoising.

(1)

Establishment of roadway boundary space model: According to the geometric structure characteristics of the roadway, a three-dimensional space boundary constraint model is constructed to determine the geometric parameters of the effective roadway area.

(2)

Set the spatial threshold range: Based on the actual roadway dimensions and the expected measurement accuracy, the boundary threshold R_b is determined to define the effective measurement region.

(3)

Perform region clipping operation: detect the spatial range of the original point cloud data, remove the overflow points beyond the boundary threshold R_b, and obtain the effective point cloud data set within the boundary. The principle of boundary cutting is shown in Formula (2).

$$P_{\left(1\right)}=\left\{p_i\in P_{\mathit{input} }\mid d\left(p_i,p_{\mathit{center} }\right)\le R_b\right\}$$

(2)

Among them, $P_{\left(1\right)}$ is the point cloud data set after the first-level denoising; $p_i$ is the three-dimensional coordinates of the i-th point; $P_{\mathit{input}}$ is the input original point cloud data set; $d\left(p_i,p_{\mathit{center}}\right)$ is the distance from the point to the central axis of the roadway.

The second level: independent interference cluster denoising.

(1)

Constructing an enhanced KD-tree spatial index: Based on the original KD-tree construction, the neighborhood connection information is added to establish a spatial index structure with topological relations. The k nearest neighbor query is shown in Formula (3).

$$N_k\left(p_i\right)=\left\{p_j\in P_{\left(1\right)}\mid ,p_j, \mathit{is} k \mathit{nearest} \mathit{neighbors} \mathit{of} ,p_i\right\}$$

(3)

where $N_k\left(p_i\right)$ is the k-nearest neighbor point set of point $p_i$; $p_j$ is the j-th point in the neighborhood.

(2)

Calculate the multi-dimensional geometric feature vector: For each point p_i, a three-dimensional geometric feature vector is constructed as F_i = [c_i, n_i, ρ_i], including curvature, normal vector, and local point density. The covariance matrix is calculated as follows:

$$C_i={\displaystyle \frac{1}{k}}{\displaystyle \sum _{p_j\in N_k\left(p_i\right)}\left(p_j-{\overline{p}}_i\right)}{\left(p_j-{\overline{p}}_i\right)}^T$$

(4)

where the neighborhood centroid ${\overline{p}}_i$ is:

$${\overline{p}}_i={\displaystyle \frac{1}{k}}{\displaystyle \sum _{p_j\in N_k\left(p_i\right)}{p}_j}$$

(5)

The curvature characteristic $c_i$ is:

$$c_i={\displaystyle \frac{{\lambda }_3}{{\lambda }_1+{\lambda }_2+{\lambda }_3}}$$

(6)

The normal vector $n_i$ is:

$$n_i=v_3,‖n_i‖=1$$

(7)

The local point density ${\rho }_i$ is:

$${\rho }_i={\displaystyle \frac{k}{\frac{4}{3}\pi r_k^3}}, r_k=‖p_i-p_k‖$$

(8)

where ${\lambda }_1$, ${\lambda }_2$, and ${\lambda }_3$ are the three eigenvalues of the covariance matrix, arranged in descending order; $v_3$ is the unit eigenvector corresponding to the minimum eigenvalue ${\lambda }_3$; and $r_k$ is the distance of the k-th nearest neighbor point.

(3)

Set the feature similarity threshold: establish the feature similarity measure standard and set the threshold to judge the geometric feature consistency of adjacent points. The feature similarity measure is calculated as follows:

$$S\left(F_i,F_j\right)=\sqrt{w_c{\left(c_i-c_j\right)}^2+w_n{\left(1-\left|n_i\cdot n_j\right|\right)}^2+w_{\rho }{\left({\displaystyle \frac{\left|{\rho }_i-{\rho }_j\right|}{\max \left({\rho }_i,{\rho }_j\right)}}\right)}^2}$$

(9)

Among them, the feature similarity distance of point $S\left(F_i,F_j\right)$ p_i and p_j; $w_c$, $w_n$ and $w_{\rho }$ are the weight coefficients of curvature, normal vector and density, respectively, satisfying $w_c+w_n+w_{\rho }=1$; $n_i\cdot n_j$ is the normal vector dot product, which represents the cosine value of the angle.

(4)

Perform connectivity analysis: Based on the KD-tree neighborhood relationship and feature similarity, a neighborhood connected graph G is constructed to identify point clusters with similar features. The neighborhood connected graph G is calculated as follows:

$$G=\left(V, E\right)$$

(10)

Among them:

$$V=P_{(1)}$$

(11)

$$E=\left\{\left(p_i,p_j\right)\mid p_j\in N_k\left(p_i\right)\wedge S\left(F_i,F_j\right)\text{<}{\theta }_{\mathit{sim}}\right\}$$

(12)

Based on Equations (10)–(12), the connected cluster decomposition is carried out, and the following results are obtained:

$$G=G_1\cup G_2\cup \dots \cup G_M$$

(13)

where V is the vertex set, that is, all points in the point cloud; E is the edge set, which connects the adjacent point pairs with similar features; ${\theta }_{\mathit{sim}}$ is the feature similarity threshold; $G_M$ is the M-th connected cluster.

(5)

Identifying interference clusters based on feature consistency: Analyze the feature statistics of each cluster in the connected graph, identify independent interference clusters that do not match the surface features of the roadway, and set the cluster size threshold θ_cluster. Remove the interference cluster markers that are smaller than the threshold, retain the large-scale effective clusters, and complete the removal of independent interference clusters.

In terms of parameter selection, the cluster size threshold θ_cluster is determined by empirical parameter adjustment combined with scene prior constraints. Specifically, this paper first sets the candidate range of θ_cluster according to the characteristics that the point cloud of the main surface of the roadway usually has a large-scale continuous distribution, while the point cloud of mobile personnel, transportation equipment and local suspended interference usually shows a small-scale isolated cluster. Subsequently, the visual comparison and result inspection were carried out in combination with multiple sets of original point cloud samples in the test section, and the final parameters were determined under the principle of “retaining the main surface of the roadway as completely as possible while effectively eliminating small-scale isolated interference clusters”. This parameter is not automatically given by a single statistical formula, but an engineering robust setting for coal mine underground scenes. The test results show that the threshold can effectively suppress the independent interference clusters caused by the occlusion of mobile personnel, transport vehicles and local equipment in the test area of this paper, and no obvious false deletion of the point cloud of the main structure of the roadway is observed.

Among them, the cluster size statistics are:

$$N_m=\left|G_m\right|$$

(14)

Identify and remove interference clusters based on statistical information:

$$P_{(2)}={\displaystyle \underset{m\text{:}{N}_m\ge {\theta }_{\mathit{cluster} }}{\cup }{G}_m}$$

(15)

where $N_m$ is the number of points (cluster size) of the m-th cluster; $\left|G_m\right|$ is the number of points in the cluster $G_m$; θ_cluster is the cluster size threshold used to identify interference clusters; $P_{(2)}$ denotes the point cloud dataset obtained after the second-level denoising.

The third level: sparse outlier denoising.

(1)

Neighborhood radius detection: establish a neighborhood range with a fixed radius r for each point and count the number of points in the neighborhood. The calculation formula of the fixed radius neighborhood is:

$$N_r\left(p_i\right)=\left\{p_j\in P_{(2)}‖p_i-p_j‖\le r,j\ne i\right\}$$

(16)

where $N_r\left(p_i\right)$ is the set of neighborhood points within the radius r of point $\left(p_i\right)$.

(2)

Statistical neighborhood density: Calculate the point density ${\rho }_r\left(p_i\right)$ of each point p_i within the neighborhood radius r, and quantify the density of the local point cloud.

Among them, the neighborhood point density is:

$${\rho }_r\left(p_i\right)={\displaystyle \frac{\left|N_r\left(p_i\right)\right|}{\frac{4}{3}\pi r^3}}$$

(17)

where $\left|N_r\left(p_i\right)\right|$ is the number of points in the neighborhood.

(3)

Set the density threshold determination: determine the minimum density threshold ρ_min, as the standard to distinguish between normal points and sparse abnormal points. The global density statistics are:

$$\overline{\rho }={\displaystyle \frac{1}{\left|P_{(2)}\right|}}{\displaystyle \sum _{\mathit{pi}\in P_{(2)}}{\rho }_r}\left(p_i\right)$$

(18)

$${\sigma }_{\rho }=\sqrt{{\displaystyle \frac{1}{\left|P_{(2)}\right|}}{\displaystyle \sum _{\mathit{pi}\in P_{(2)}}{\left({\rho }_r\left(p_i\right)-\overline{\rho }\right)}^2}}$$

(19)

The density threshold is:

$${\rho }_{\mathit{min}}=\overline{\rho }-\beta \cdot {\sigma }_{\rho }$$

(20)

where $\overline{\rho }$ denotes the global mean point density of the point cloud; ${\sigma }_{\rho }$ is the standard deviation of the point density and $\beta$ is a scaling coefficient used to control the density threshold.

(4)

Abnormal point identification and removal: For the points whose density value is lower than the threshold, the sparse abnormal points are marked and removed, and the remaining points are retained as normal points, and finally, the optimized denoising point cloud is obtained.

The final denoising point cloud is:

$$P_{\mathit{final} }=\left\{p_i\in P_{(2)}\mid {\rho }_r\left(p_i\right)\ge {\rho }_{\mathit{min} }\right\}$$

(21)

The denoising effect of the roadway point cloud in the test section of 5111 return air roadway in Yushupo Coal Mine is shown in Figure 5. After three-stage progressive denoising, the boundary overflow noise, interference object noise and sparse outliers in the original point cloud are effectively eliminated, and the quality of point cloud data is significantly improved. The denoised roadway point cloud completely retains the geometric feature information of the roadway surface, the noise interference is minimized, and the data purity and geometric accuracy meet the requirements of subsequent analysis. The high-quality point cloud data lays a reliable data foundation for subsequent multi-stage roadway point cloud registration and surrounding rock deformation analysis and ensures the accuracy and reliability of deformation detection results.

3.3.2. Roadway Point Cloud Registration

Point cloud registration is the core technology to realize multi-period roadway deformation monitoring. Its goal is to unify the point cloud data collected at different times into the same coordinate reference system, so as to provide a reliable spatial reference for subsequent deformation quantitative analysis. However, the underground roadway environment is special: on the one hand, the IMU of the mobile laser scanning system has cumulative drift, resulting in systematic deviation of the scanning data; on the other hand, due to the high geometric similarity and low feature discrimination of the roadway structure, the traditional registration method is easy to fall into a local optimum. In view of the above challenges, this study proposes a hierarchical point cloud registration method that combines physical anchor constraint and deep learning feature guidance and achieves high-precision point cloud alignment through a two-stage strategy of ‘coarse registration-fine registration’.

The overall framework of this method is shown in Figure 6. In the first stage, the geometric constraint is carried out by using the physical anchor points laid in the roadway, and the rigid body transformation is solved by weighted SVD to eliminate the large-scale error caused by IMU drift. In the second stage, the patch-guided matching method based on deep learning is used to further optimize the registration accuracy through multi-scale feature extraction and a hierarchical matching strategy. The organic combination of the two stages gives full play to the global positioning ability of physical constraints and the local fine matching ability of deep learning.

(1)

Rough registration method based on multi-anchor geometric constraints

The mobile 3D laser scanning system relies on the inertial measurement unit (IMU) for real-time positioning during data acquisition. Due to the inherent bias error, scale factor error and random walk characteristics of IMU, its position estimation error increases polynomially with scanning time. Let $\delta p\left(t\right)$ be the position drift at time t, then:

$$\delta p\left(t\right)=\delta p_0+\delta v_{0}t+{\displaystyle \frac{1}{2}}{b}_a{t}^2+{\displaystyle \frac{1}{6}}{\stackrel{.}{b}}_a{t}^3$$

(22)

Here, $\delta p_0$ and $\delta v_0$ are the initial position and velocity errors, respectively, $b_a$ is the accelerometer bias, and ${\stackrel{.}{b}}_a$ is the time-varying drift term caused by the bias instability. The error model shows that in the absence of external constraints, long-distance roadway scanning will produce significant geometric distortion. In order to effectively suppress the above cumulative errors, this method sets up several physical anchor points with known absolute coordinates in the target roadway as spatial control benchmarks (Figure 7).

Assuming that the anchor observation coordinate of the i-th scanning point cloud in the local coordinate system is $P_i^{\mathit{local} }=\left\{p_1^l,p_2^l,\dots ,p_n^l\right\}$, and the corresponding global real coordinate is $p^{\mathit{global} }=\left\{p_1^g,p_2^g,\dots ,p_n^g\right\}$, the goal of coarse registration is to solve the optimal rigid body transformation $T_c=\left\{R_c,t_c\right\}$, and the following results are obtained:

$$T_c^{*}=\underset{R_c\in \mathrm{SO}\left(3\right),t_c\in R^3}{\mathit{argmin}}{\displaystyle \sum _{j=1}^n{w}_j}{‖R_c{p}_j^l+t_c-p_j^g‖}_2^2$$

(23)

Among them, w_j is the confidence weight of the j-th anchor point, which can be adaptively determined according to the observation conditions of the anchor point (such as distance from the scanning path, observation angle, etc.). The above optimization problem can be efficiently solved by weighted singular value decomposition (Weighted SVD). Firstly, the weighted centroid is calculated:

$${\overline{p}}^l={\displaystyle \frac{{\displaystyle \sum _{j=1}^n{w}_j}{p}_j^l}{{\displaystyle \sum _{j=1}^n{w}_j}}},{\overline{p}}^g={\displaystyle \frac{{\displaystyle \sum _{j=1}^n{w}_j}{p}_j^g}{{\displaystyle \sum _{j=1}^n{w}_j}}}$$

(24)

Then the covariance matrix H is constructed and SVD decomposition is performed:

$$H={\displaystyle \sum _{j=1}^n{w}_j}\left(p_j^l-{\overline{p}}^l\right){\left(p_j^g-{\overline{p}}^g\right)}^{\top }=U\Sigma V^{\top }$$

(25)

The optimal rotation matrix and translation vector are obtained as follows:

$$R_c^{*}=V{U}^{\top },t_c^{*}={\overline{p}}^g-R_c^{*}{\overline{p}}^l$$

(26)

After coarse registration, the macroscopic geometric error of the multi-period point cloud can be reduced to the centimeter level, which lays a good foundation for subsequent fine registration.

(2)

Point cloud fine registration method based on patch-guided registration network

Although coarse registration can eliminate large-scale geometric deviations, the registration accuracy still has millimeter-level residuals due to factors such as a limited number of anchor points, sparse distribution, and local coordinate measurement noise. In order to further improve the registration accuracy, this study introduces a patch-guided fine registration method based on deep learning [31,32]. The core idea of this method is that compared with single-point feature matching, local patch (Patch) contains richer geometric structure information, which has higher matching confidence and stronger robustness.

The roadway environment has the characteristics of low overlap rate, weak texture and linear extension, which puts forward higher requirements for the registration method. The fine registration network used in this study includes two core modules: feature pyramid network (FPN) and matching pyramid network (MPN). The structure is shown in Figure 8. FPN uses an encoder–decoder structure to extract multi-scale geometric features. MPN realizes the hierarchical matching of ‘large patch → small patch → point’, and the matching results of high confidence in the upper layer guide the fine matching in the lower layer.

Assuming that the source point cloud after coarse registration is X and the target point cloud is Y, the goal of fine registration is to learn a high-quality corresponding point pair set ${\Theta }^{*}$, and then estimate the fine rigid body transformation T_f = {R_f, t_f}. Aiming at the characteristics of the large-scale roadway point cloud, the matching module adopts a linear attention mechanism:

$$\mathit{Linear}-\mathit{Attention}(Q,K,V)={\displaystyle \frac{\varphi \left(Q\right)\left(\varphi {\left(K\right)}^{\top }V\right)}{\varphi \left(Q\right){\displaystyle \sum _j\varphi }{\left(k_j\right)}^{\top }}}$$

(27)

where $\varphi (\cdot )$ is the activation function. This mechanism reduces the computational complexity from O(N²C) to O(NC²), making real-time processing of large-scale roadway point clouds possible. At the same time, the cross-layer context aggregation (CCA) module realizes the transmission of upper-layer matching information to the lower layer through the attention mechanism and improves the matching accuracy of lower-layer features. In order to ensure the reliability of the final corresponding point pairs, the matching consistency judgment strategy is used to fuse the multi-layer matching results. The calculation process is shown in Figure 9. Firstly, the score matrix of each layer is binarized, and the high-confidence matching pair is retained. Then the upper patch matching relationship is mapped down to the point level; finally, the intersection of each layer matching is calculated by the product of elements:

$$S_c={\widehat{S}}_b^1\odot {\widehat{S}}_b^2\odot \dots \odot {\widehat{S}}_b^L$$

(28)

Only point pairs that are judged to be matched at all levels are retained, and the final corresponding point pairs are selected accordingly:

$${\Theta }^{*}=\left\{\left(x_i,y_j\right)\mid (i,j)\in {\mathit{top}}_k\left(S_c\odot S^L\right)\right\}$$

(29)

This strategy effectively filters out outliers in single-layer matching and significantly improves the interior point rate, which is particularly important for scenes with high repeatability of geometric features such as roadways.

(3)

Fusion strategy of hierarchical registration

The innovation of this method lies in the organic integration of physical anchor constraints and deep learning feature matching to form a hierarchical registration framework with complementary advantages. The physical anchor provides an absolute spatial reference to ensure that the multi-period point clouds are aligned to a unified global coordinate system, avoiding the matching failure of the deep learning method under large-scale errors. The deep learning method makes up for the limitation of the sparse distribution of anchors and realizes fine registration of regions between anchors. The final rigid body transformation is a composite of two-stage transformations:

$$T^{*}=T_f\circ T_c=\left\{R_f{R}_c,R_f{t}_c+t_f\right\}$$

(30)

The point cloud registration effect of the roadway in the test section is shown in Figure 10. The red and blue in the figure represent the point cloud data of the two phases of scanning. (a) Pre-registration state: Due to the cumulative drift of IMU, the two-stage point clouds have obvious overall offset and cannot be directly used for deformation analysis; (b) after coarse registration: after the anchor point constraint correction, the macro position of the two phase point clouds is basically aligned, but there are still visible dislocations in the local area; (c) after fine registration: after deep learning fine registration processing, the two phases of point cloud height overlap. From the error distribution, it can be seen that the registration error RMSE is reduced to 8.5 cm after coarse registration, which is 83.7% higher than that before registration. After the fine registration, the error RMSE is reduced to 0.6 cm, and the accuracy is improved by 92.9% compared with the coarse registration stage. The experimental results show that the proposed two-stage registration method can effectively deal with complex scanning scenes with low overlap rate and weak texture features, such as coal mine roadways. The registration accuracy meets the engineering requirements of high-precision deformation monitoring of surrounding rock and provides a reliable data basis for subsequent multi-stage point cloud difference analysis and quantitative evaluation of surrounding rock convergence deformation.

3.4. Comparative Analysis of Deformation of Roadway Point Cloud Data

3.4.1. Overall Deformation Recognition Based on Point Cloud Distance

Based on the ‘point cloud/point cloud distance calculation’ method, this study performs a global deviation analysis on the two-phase scan data after spatial registration. The mathematical principle is that for each sampling point in the point cloud to be detected, the nearest neighbor point with the closest spatial position is searched in the reference point cloud, and the Euclidean distance between the two is calculated. The specific expression is as follows:

$$d_i=\underset{j}{\min }\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2+{(z_i-z_j)}^2}$$

(31)

where d_i denotes the nearest distance from the i-th point in the target point cloud to the reference point cloud, (x_i, y_i, z_i) is its coordinate, and (x_j, y_j, z_j) is the coordinate of the j-th point in the reference point cloud. After the calculation is completed, a distance value is assigned to each point and visualized by a color gradient. Through this chromatogram, the position and approximate scale of the significant deformation area can be identified intuitively and quickly, which provides the target for the fine analysis of the subsequent key areas.

3.4.2. Section Slice Deformation Analysis

After identifying the potential deformation area, further quantitative analysis is carried out by section slicing. The three-dimensional point cloud model is divided into a series of continuous cross sections along the longitudinal direction of the roadway. At each section position, the local point set is projected bidirectionally to the XOY and YOZ planes, respectively, and the geometric centroid is determined by covariance analysis. Taking the obtained centroid as the reference origin, a cutting plane orthogonal to the roadway axis is established. At the same time, the slice thickness is adaptively adjusted according to the distribution density of the point cloud in this area to ensure that sufficient effective data points are intercepted. In the contour reconstruction stage, the non-uniform B-spline method is introduced, and the smooth node distribution sequence is generated by the chord length parameterization strategy. The spline basis function is recursively solved, and the weight of the control vertex is iteratively optimized to achieve a high-fidelity curve approximation of the discrete point cloud of the section. On this basis, the area enclosed by the fitting contour of each section is calculated, and then the quantitative evaluation of the convergence degree of the roadway is carried out with the area change rate as the index.

4. Engineering Application and Deformation Result Analysis

The point cloud data were collected from the test roadway mileage of 103–230 m (Figure 2). The first phase of point cloud data was collected on 5 July 2025. At this time, the heading face was located at about 245 m, and the second phase of point cloud data was collected on 5 August 2025. Through the differential comparison of the two-phase point cloud data, the surrounding rock deformation evolution characteristics of the roadway in this area during the gradual transition to the stable stage of the roadway after the excavation disturbance are systematically evaluated.

4.1. Comparative Analysis of the Overall Deformation of the Roadway

Figure 11 shows the three-dimensional deformation analysis results of the two-stage point cloud data after registration. Based on the color rendering method, the visualization of the global deformation field is realized. The spatial distribution characteristics of the color reflect the deformation state of each section of the roadway, and the light and dark levels of the color correspond to the size of the deformation amplitude. The analysis results show that the peak deformation during the test within the monitoring range is within 90 mm. At the same time, two deformation concentration areas are identified: one is located at the shoulder angle of the coal side of the roadway, and the other is located at the floor of the inclined section of the roadway. In addition, the roof surrounding rock also shows a certain degree of deformation response, but its deformation amplitude is relatively limited. In the future, fine slice analysis will be carried out on the deformation concentrated area to further evaluate its deformation evolution law.

It should be noted that the point cloud difference analysis shown in Figure 11 covers the monitoring section of 103–230 m in front of the H776 measuring point.

4.2. Slice Quantitative Analysis of Roadway Deformation Area

Through the above analysis of the overall deformation of the roadway, it can be seen that there is a significant deformation area in the range of 167–194 m from the H776 measuring point. In order to further explore the specific deformation, deformation amplitude and deformation development trend of the region, the section slice fitting method is used to quantitatively analyze the morphological changes in the roadway. As shown in Figure 12, 10 sets of reference monitoring sections (PS1–PS10) are arranged along the direction of roadway excavation at a distance of 3 m, and 10 sets of monitoring sections are evenly arranged in the aforementioned deformation concentration Section (167–194 m away from the H776 measuring point). The section spacing is 3 m, and the 27 m monitoring section completely corresponds to the deformation concentration area marked in the point cloud difference analysis in Figure 11. Among them, PS1 is close to the anchoring point H776 side, and PS10 is closer to the heading side.

It should be noted that the section layout shown in Figure 12 and the slice deformation results in Figure 13 have a clear spatial correspondence with the local strong deformation zone in Figure 11.

The quantitative analysis results of the two-phase point clouds are shown in Figure 13. The green section is the point cloud slice on 5 July 2025, and the purple section is the point cloud slice on 5 August 2025. In order to enhance the engineering interpretability of the results, this paper uses a symbolic representation of the section displacement: the movement of the inner contour of the roadway relative to the first-stage monitoring to the inner side of the roadway clearance is defined as a negative value, and the movement to the outer side of the surrounding rock is defined as a positive value. Among them, the floor uplift (floor heave) is recorded as positive displacement, the floor depression is recorded as negative displacement, the roof subsidence is recorded as negative displacement, the roof arch is recorded as positive displacement, and the convergence of the two sides to the roadway is recorded as negative displacement.

The analysis results show that the deformation of the surrounding rock in the slice contrast section has significant spatial heterogeneity, in which the uplift deformation of the floor is the main one, which is manifested as continuous floor heave. The maximum positive displacement is +90.0 mm, and the average positive displacement is +46.9 mm, which is the most important deformation mode in this section. The roof is dominated by weak subsidence, with the maximum negative displacement of −35.0 mm and the average negative displacement of −19.3 mm. The side wall of the solid coal wall is convergent to the roadway as a whole, the maximum negative displacement is −33 mm, and the average negative displacement is −21.2 mm; the convergence of the side wall of the coal pillar is slightly larger, the maximum negative displacement is −39 mm, and the average negative displacement is −28 mm, indicating that the convergence of the two sides has asymmetric characteristics.

In general, the current support system in the monitoring section has generally played the expected control role. Although the deformation of the surrounding rock has developed, it is still in the controllable range of the project as a whole, and there is no sign of penetrating damage or structural instability. However, the continuous uplift of the floor constitutes the most important safety hazard in this section. Combined with the results of the field investigation and coal rock mechanics test, it can be seen that the lithology of the 5111 return air trough floor is mainly sandy mudstone, which belongs to medium-low strength soft rock, and the uniaxial compressive strength is relatively low. At the same time, there is water leakage in the local roof during the excavation process, which further aggravates the water softening effect of the floor rock mass and weakens its bearing capacity. During the continuous advancement of the roadway, the surrounding rock of the two sides showed a gradual introverted trend under the high stress environment, while the floor experienced continuous floor heave and gradually accumulated development under the combined action of strength weakening, stress redistribution and long-term pressure bearing. The deformation response is consistent with the full-space point cloud monitoring results obtained in this study, that is, the floor deformation is the most significant, followed by the two sides, and the roof is relatively small.

4.3. Consistency of Monitoring Methods and Engineering Reliability Verification

In order to verify the engineering reliability of the proposed monitoring method, a total station was set up at the corresponding position of the slice PS1–PS10 to measure the deformation of the surrounding rock, and the measured data of the total station were compared with the point cloud slice monitoring results. The accuracy index evaluated in this section is the consistency between the point cloud monitoring results and the total station measured results; that is, the total station is used as the spatial reference benchmark to evaluate the SLAM point cloud slicing method. The deviation level between the deformation measurement results of each monitoring section and each part (roof, floor, two sides) and the total station results, rather than the absolute accuracy in the strict metrological sense or the precision of multiple independent repeated observations. The total station is selected as the reference datum because it has been widely recognized in the field of mine surveying and has a nominal measurement accuracy of millimeter level. The total station (Leica TZ08) used in this study has been certified by MA and can be regarded as a reliable datum under the background of this project.

On-site comparison and verification are carried out according to the standard process of “total station first, then 3D laser scanner”. Both measurements are completed within the same working day to ensure time synchronization. The total station is set up at the corresponding position of each monitoring section, and five standard characteristic measuring points are set up in each section: the midpoint of the roof (1), the midpoint of the floor (1), the left side (1), the right side (1) and the shoulder angle position (1); each measurement is performed only once, which is the same frequency as the SLAM point cloud acquisition. The data of the two phases were collected on 5 July 2025 and 5 August 2025, respectively. It should be noted that the deformation evaluation in this study is based on the absolute value strategy of “difference of difference”, that is, the difference (deformation) of the measurement results of the same section and the same part in the two periods is compared, rather than the absolute coordinates. Under this framework, the systematic drift error is partially offset in the two-phase subtraction process, and a single measurement can meet the engineering accuracy requirements of deformation comparison without additional repeated measurements. After the measurement of the total station was completed, the operator hand-held GHJS12 traveled at a constant speed along the axial direction of the roadway to complete the full-section SLAM scan, and the acquisition time was about 0.5 h.

The error statistical results are shown in Figure 14, and the absolute value of the error is used for analysis and comparison. The data analysis results show that the maximum displacement error in the roof direction is 3 mm (standard deviation 1.02 mm), the maximum displacement error in the floor direction is 4 mm (standard deviation 0.98 mm), the maximum displacement error in the side wall of the coal pillar is 4 mm (standard deviation 0.94 mm), and the maximum error in the side wall of the solid coal wall is 5 mm (standard deviation 1.3 mm). The above error analysis shows that the monitoring method fully meets the needs of coal mines for roadway deformation monitoring, which fully confirms that the deformation monitoring technology has good engineering reliability and practical value, and provides a new method for roadway surrounding rock deformation monitoring.

5. Conclusions

Aiming at the technical bottlenecks of traditional monitoring methods for surrounding rock of coal mine roadway, such as single information dimension, limited coverage and insufficient engineering application efficiency, this study systematically constructed a global deformation monitoring method system of surrounding rock of roadway based on explosion-proof mobile 3D laser scanning technology, and verified the effectiveness and engineering applicability of the method through field scanning test and total station comparison. The main conclusions are as follows:

(1)

A hierarchical point cloud denoising method suitable for the complex environment of coal mine roadway is proposed. Through the three-level progressive processing strategy of boundary clipping, improved statistical filtering guided by multi-dimensional geometric features and adaptive density threshold, three typical noises of boundary overflow points, independent interference clusters and sparse outliers are eliminated. On the premise of retaining the integrity of the geometric features of the roadway surface, this method significantly improves the purity and geometric accuracy of the original point cloud data, solves the key problem of poor quality of point cloud data in the multi-interference source environment of the coal mine, and provides data guarantee for subsequent high-precision registration and reliable deformation analysis.

(2)

A hierarchical point cloud registration method combining physical anchor geometric constraints and deep learning feature guidance is established, which effectively solves the registration problem of a mobile laser scanning system in low overlap rate, weak texture and high geometric repeatability scenes such as coal mine roadway. In the coarse registration stage, the pre-calibrated physical anchor points are used to achieve global coordinate alignment through weighted SVD, which reduces the registration error RMS from 52.3 cm to 8.5 cm, and the accuracy is improved by 83.7%. In the fine registration stage, the patch-guided registration network (FPN + MPN) is used, and the outlier corresponding point pairs are filtered by combining the matching consistency judgment strategy. The registration error RMS is further reduced to 0.6 cm, and the accuracy is improved by 92.9% compared with the coarse registration stage. The overall accuracy of the two-stage registration is improved by 98.9%, which ensures the engineering requirements of surrounding rock deformation monitoring.

(3)

The engineering application and consistency verification results show that this method can effectively capture the global spatial distribution information and evolution characteristics of roadway surrounding rock deformation. In the one-month monitoring of the 5111 return air crossheading test section of Yushupo Coal Mine, two deformation concentration areas were identified through global difference analysis. Quantitative analysis showed that the floor was dominated by uplift deformation, with the maximum positive displacement of +90.0 mm and the average positive displacement of +46.9 mm. The roof is mainly sinking, the maximum negative displacement is −35.0 mm, and the average negative displacement is −19.3 mm. The maximum negative displacement on the side of the coal pillar is −39 mm, and the maximum negative displacement on the side of the solid coal wall is −33 mm, showing a certain asymmetry. Compared with the total station, the results show that the maximum displacement error in each direction does not exceed 5 mm, and the standard deviation is within 1.3 mm. It shows that the method has sufficient engineering accuracy in the identification of roadway surrounding rock deformation, abnormal area positioning and safety monitoring and early warning, which can provide a reliable basis for support effect evaluation, key section review and on-site disposal decision.

The research results verify the engineering feasibility and technical superiority of three-dimensional laser scanning technology in the deformation monitoring of the surrounding rock of coal mine roadway, break through the information dimension limitation of the traditional discrete point monitoring method, and provide full coverage and high-precision data support for the stability evaluation and support scheme optimization of the surrounding rock of coal mine roadway. Further, compared with the total station, the results show that the method has sufficient engineering accuracy in the identification of roadway surrounding rock deformation, the positioning of abnormal areas and the safety monitoring and early warning, which can provide a reliable basis for the evaluation of support effect, the review of key sections and the decision-making of on-site disposal. From the perspective of engineering implementation, mobile SLAM LiDAR can complete the full-space point cloud acquisition of the test section within about 0.5 h, and the field operation efficiency is obviously better than the traditional point-by-point measurement method. However, its data processing needs to complete denoising, registration and difference analysis, and the data processing complexity is relatively high. In general, this method reflects the technical characteristics of ‘high efficiency on site and high amount of information in the industry’. For mine operators, although its initial equipment investment is higher than that of conventional discrete measurement equipment, it has obvious potential in expanding monitoring coverage, reducing artificial exposure in high-risk areas, improving anomaly recognition ability and reducing long-term operating cost Therefore, it is more suitable as a fast panoramic sensing method in the stability monitoring of surrounding rock of coal mine roadway, and complements with high-precision discrete measurement methods such as total station.

Research limitations: The method in this paper is mainly applicable to the surface deformation monitoring of surrounding rock under conventional excavation disturbance and static/quasi-static service conditions and has not been explicitly included in the transient response analysis of underground motion, strong periodic disturbance or other significant dynamic loads. For roadway scenes affected by low-frequency wave propagation, rock burst or strong dynamic load, further research is still needed in combination with dynamic monitoring methods with higher time resolution and dynamic response analysis methods. In addition, the verification of the hierarchical denoising method in this paper is mainly based on the real coal mine point cloud scene. Aiming at the synthetic noise injection experiments under controlled conditions (such as Gaussian noise, outlier and point cloud sparsity, etc.), a standardized benchmark data set can be further constructed in the follow-up work to more systematically quantify the robustness and generalization ability of the algorithm under different noise types.