Unsupervised learning drilling and blasting method tunnel point cloud feature extraction and back break detection method

Abstract

The invention belongs to the technical field of tunnel engineering digital construction, and particularly relates to an unsupervised learning drilling and blasting method tunnel point cloud feature extraction and back break detection method. The method comprises the following steps: firstly, acquiring initial point cloud data of tunnel excavation by adopting three-dimensional laser scanning, and preprocessing by adopting point cloud thinning and registration; then, a convex hull algorithm is used for calculating an actual excavation axis, excavation sections with the thickness of 0.1 m are cut out according to the fixed interval of 10 m, a K-means clustering algorithm is used for conducting noise reduction processing on point cloud slices, and convex hull vertexes are screened to accurately extract actual excavation contour lines; and finally, the back break amount of each convex hull is calculated, and real-time quantitative analysis of the back break condition of the tunnel is achieved. According to the method, by fusing three-dimensional laser scanning, unsupervised learning and a convex hull algorithm, crossing of tunnel back break detection from discrete, lagging and manual detection to continuous, real-time and intelligent detection is achieved, and efficient, accurate and reliable digital technical support is provided for drilling and blasting method tunnel construction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of tunnel engineering digital construction, and particularly relates to a point cloud feature extraction and overbreak and underbreak detection method for drill-and-blast tunneling based on unsupervised learning, which is suitable for real-time analysis of overbreak and underbreak in drill-and-blast construction and optimization of blasting parameters. BACKGROUND

[0002] With the continuous advancement of infrastructure construction in China, tunnel engineering plays an increasingly important role in the fields of transportation, water conservancy, etc. The quality of tunnel construction is directly related to the safety, service life and economic benefits of the project. Traditional tunnel measurement methods have low precision, low efficiency, and incomplete data, which cannot meet the requirements of modern tunnel construction quality control. Therefore, accurate and efficient real-time detection and repair of tunnel overbreak and underbreak are particularly important.

[0003] The traditional overbreak and underbreak detection method mainly has the following technical bottlenecks and discrete sampling defects: total station and section instrument need to be collected point by point manually, the data density is insufficient, usually < 50 points / section (Zhang Xiangdong, Zhang Xuedong. Application of laser section instrument in tunnel section detection [J]. Transportation technology, 2014, (05): 94-96.), it is difficult to construct a continuous excavation contour, and there is a detection blind area of 20-30 cm. Lack of timeliness: the existing method needs to be measured on site and processed offline, the data lag is up to 4-6 hours, which cannot meet the real-time demand of dynamic adjustment of blasting parameters (tunnel drill-and-blast cycle is usually < 24 hours). Low degree of automation: manual measurement error can reach ±3 cm, and is significantly affected by the tunnel environment such as dust, lighting, etc. According to statistics, the efficiency of manual detection is only 15-20 sections per shift.

[0004] A tunnel overbreak and underbreak automatic calculation method based on Monte Carlo method (patent number 202411871094.X) is disclosed, which judges the positional relationship of random points with two convex hulls to judge the three-dimensional shape of the tunnel excavation surface, especially in rock tunnels or complex geological conditions, the value of random points is different, the calculated data will have slight changes, thereby affecting the calculation accuracy.

[0005] A tunnel overbreak and underbreak identification and volume calculation method based on laser point cloud (patent number 202411871097.3) is disclosed, for each region, the design contour point cloud and the actual excavation region point cloud constituting the region are constructed into polyhedrons, the polyhedrons are divided into quadrilaterals, and the volume of each quadrilateral is calculated, thereby calculating the volume of each overbreak and underbreak region. Using quadrilaterals to calculate the volume will make it difficult to calculate irregular regions, and it is difficult to avoid a certain loss of accuracy.

[0006] Disclosed is a tunnel overbreak and underbreak detection method and device, electronic equipment and storage medium. The original three-dimensional point cloud of a tunnel to be detected is obtained. The original three-dimensional point cloud is preprocessed to obtain a preprocessed three-dimensional point cloud. The preprocessed three-dimensional point cloud is subjected to ground plane removal processing to obtain a target three-dimensional point cloud. A tunnel center axis in the target three-dimensional point cloud is extracted. The target three-dimensional point cloud is rotated and rectified based on related parameters of the tunnel center axis to obtain a rectified three-dimensional point cloud. The rectified three-dimensional point cloud is intercepted to obtain a tunnel cross section. The cross section is used to fit an ideal tunnel profile curve. Ideal design tunnel excavation surface information is obtained based on the ideal profile curve. Irrelevant point clouds are removed from the ideal design tunnel excavation surface information to obtain a measured tunnel three-dimensional point cloud. The ideal design tunnel excavation surface information and the measured tunnel three-dimensional point cloud are used to calculate the overbreak and underbreak of each point in the point cloud using the polar coordinate method. Although "rectified point cloud" is mentioned, the specific method for extracting the actual axis and calculating the space transformation matrix is not disclosed.

[0007] Disclosed is a tunnel cross section feature point accurate extraction method based on laser point cloud measurement. The disadvantages of traditional threshold denoising are that it relies on manual experience and is easy to lose details. The present scheme uses dynamic threshold + generative model to avoid noise misjudgment and retain profile details, especially suitable for scenes with uneven tunnel point cloud density.

[0008] The above-mentioned patents only calculate the volume or local deviation, and do not establish a closed-loop correlation with construction parameter optimization. The traditional method directly assumes that the point cloud is aligned with the design axis, ignoring the systematic deviation in construction, resulting in distorted overbreak and underbreak calculation. None of the comparative documents 1-4 discloses the application of unsupervised clustering algorithm in point cloud denoising, DCGAN generative model to distinguish noise and real features, and dynamic threshold formula to solve the general denoising problem of sparse / dense point clouds. SUMMARY

[0009] The present application is based on three-dimensional scanning technology and unsupervised learning, and proposes a tunnel point cloud feature extraction and overbreak and underbreak real-time detection method. First, three-dimensional laser scanning is used to obtain tunnel excavation initial point cloud data, and point cloud thinning and registration are used for preprocessing. Then, the convex hull algorithm is used to calculate the actual excavation axis, and the excavation section with a thickness of 0.1m is intercepted at a fixed interval of 10m. The K-means clustering algorithm is used for denoising processing of the point cloud slices, and the convex hull vertices are screened to accurately extract the actual excavation contour line. Finally, the overbreak and underbreak of each convex hull are calculated to realize real-time quantitative analysis of the tunnel overbreak and underbreak.

[0010] The technical scheme of the present application is as follows: The unsupervised learning drilling and blasting tunnel point cloud feature extraction and overbreak and underbreak detection method comprises the following steps: S1, obtain tunnel initial point cloud data by three-dimensional laser scanning, and sequentially perform thinning and registration preprocessing; S2, axis correction: extract the actual excavation axis from the registration point cloud, calculate the space transformation matrix thereof and the design axis, and correct systematic deviation; S3, cross section slice extraction: intercept a slice with a thickness of δ along the design axis direction at a fixed interval Δs, Δs = 10 m, δ = 0.1 m; S4, point cloud noise reduction: the K-means clustering algorithm is used for the slice point cloud, the cluster number N is determined according to the contour coefficient, noise clusters are identified and removed through a density threshold: 3-5 for sparse point clouds and 10-20 for dense point clouds, and a deep convolutional generative adversarial network (DCGAN) is used as an auxiliary noise reduction means to distinguish noise and real features through a generative model, so as to avoid detail loss caused by threshold dependence; S5, convex hull generation: the Graham algorithm is used to screen the convex hull vertexes of the noise-reduced point cloud, and the actual excavation contour line is sequentially connected; S6, overbreak and underbreak calculation: the shortest distance di between the convex hull vertexes and the design contour line is calculated, di > 0 is overbreak, and di < 0 is underbreak; S7, trend analysis: the overbreak and underbreak amounts of each cross section along the tunnel axis direction are counted, the overall construction quality is evaluated, and adjustment suggestions are output.

[0011] In step S1, the three-dimensional laser scanner calculates the reflection time by emitting a laser pulse signal to determine the ranging observation value S between the target point and the scanner, and simultaneously records the lateral and longitudinal scanning angle observation values a and θ of each laser pulse by controlling the encoder. In the instrument internal coordinate system Pi, the X-axis is located in the lateral scanning plane, the Y-axis is perpendicular to the X-axis and is in the same lateral scanning plane, and the Z-axis is orthogonal to the scanning plane. Based on this coordinate system and the measurement data, the three-dimensional coordinates (XS, YS, ZS) of the scanning point P can be obtained, and the calculation formula is as follows: The general storage form of the original data is point cloud, which can be preprocessed by special point cloud data software, and the specific steps are as follows: (1) Thinning: due to the large amount of original point cloud data, the calculation and storage costs are increased, the density of the point cloud can be reduced by different sampling methods while retaining the original point cloud features, and common methods include thinning algorithm based on uniform distribution random number and voxel filtering method; (2) Registration: multiple groups of point cloud data obtained by scanning are combined and converted into a unified coordinate system, and common methods include ICP and FPFH.

[0012] In the step S2, the tunnel actual excavation axis is extracted from the registered point cloud, and a spatial transformation matrix between the design axis and the actual axis is calculated by the least square method to correct systematic deviation of the point cloud data.

[0013] In the step S4, the K-means is a clustering algorithm, and the main purpose is to divide the data set into N non-intersecting clusters, and the sum of squared errors in the cluster is minimized as shown in formula (2), that is, by iteratively optimizing the cluster center, the data points in the cluster are as close as possible, and the difference between the clusters is as large as possible; In the formula, J is the sum of squared errors in the cluster, Ci is the point set of the i-th cluster, p is a data point in Ci, represents the squared Euclidean distance between the data point and the cluster center, q i is the i-th cluster center; K is the number of clusters, and X is any data point in the data set; The algorithm first randomly selects N data points as initial cluster centers, and assigns each data point to the cluster with the nearest initial cluster center according to the Euclidean distance as shown in formula (3), then calculates the mean of all points in each cluster to update the position of the cluster center as shown in formula (4), and iterates until the change of the cluster center is less than a certain set threshold or the preset iteration number is reached. In the formula, d is the Euclidean distance of each point to the cluster center, (x1, y1, z1) is the three-dimensional coordinates of each point, and (x0, y0, z0) is the three-dimensional coordinates of q. After setting a reasonable threshold for the data points in the cluster divided by the K-means algorithm, the algorithm can be applied to point cloud denoising, and the specific operation is as follows: a) For each point cloud slice, use the contour coefficient method to determine its appropriate cluster number N, and determine the optimal clustering result of each cluster through the update iteration process of the K-means algorithm; b) After clustering, set a density threshold to analyze the size and density of each cluster to identify noise clusters. Generally, smaller clusters or clusters with lower density are more likely to be noise. In order to more accurately identify noise points, a reasonable density threshold can be set according to the actual situation of the initial point cloud: for sparse point clouds, the threshold is set to (3, 5); for dense point clouds, the threshold is set to (10, 20), and a deep convolutional generative adversarial network (DCGAN) is used as an auxiliary denoising means to distinguish noise from real features through a generative model, avoiding detail loss caused by threshold dependence; when the number of points in a cluster is less than the set threshold, the points in the cluster are considered as noise points; c) Remove the points in the clusters identified as noise from the point cloud to obtain the denoised point cloud data; this process effectively removes noise points in the point cloud while preserving the main features of the point cloud, thereby achieving accurate feature extraction of the point cloud.

[0014] In the step S6, overbreak and underbreak refer to the spatial relationship between the actual excavation profile and the design section: when the actual convex hull profile line exceeds the design section profile line, it is called overbreak; when the design section profile line envelopes the actual convex hull profile line, it is called underbreak; the core of overbreak and underbreak calculation is to compare the spatial position relationship between the actual convex hull and the design section, and to quantify the overbreak and underbreak; a virtual line is drawn from the design axis to the convex hull vertex, and intersects with the tunnel design section profile line at a point, and the distance between the convex hull vertex and the point is the overbreak and underbreak, and the calculation formula is as follows: wherein di is the overbreak and underbreak of a convex hull vertex i point, xi is the horizontal coordinate of i point, yi is the vertical coordinate of i point, xh is the horizontal coordinate of the point on the design section closest to i point, and yh is the vertical coordinate thereof.

[0015] The unsupervised learning drill-and-blast tunnel point cloud feature extraction and overbreak and underbreak detection method has the following significant technical effects: High detection accuracy, significantly reduced error: By combining the K-means clustering algorithm and the DCGAN generative denoising means, noise points are effectively identified and removed, overcoming the detail loss problem caused by threshold dependence in traditional methods; the convex hull algorithm is combined to accurately extract the actual excavation profile, and the systematic deviation is corrected through the space transformation matrix, greatly improving the calculation accuracy of the overbreak and underbreak. The embodiments show that, compared with the traditional method, the error of the present method is reduced by at least 11.89%, and the maximum reduction is 40.40%.

[0016] Strong real-time, supports dynamic optimization of construction: The present method can realize real-time quantitative analysis of overbreak and underbreak, greatly shorten the data processing lag time, meet the dynamic adjustment demand of blasting parameters within the drill-and-blast cycle period (usually < 24 hours), and effectively improve the construction efficiency and timeliness of quality control.

[0017] High degree of automation, reduces manual intervention: The unsupervised learning algorithm is used for point cloud denoising and feature extraction, avoiding the subjectivity and instability of setting thresholds by relying on manual experience in traditional methods, reducing the influence of environmental factors (such as dust and lighting) on measurement results, and improving the objectivity and consistency of detection results.

[0018] Strong adaptability, suitable for complex geological conditions: Through dynamic density threshold (3~5 for sparse point cloud, 10~20 for dense point cloud) and DCGAN auxiliary denoising, the uneven density of tunnel point cloud and complex noise distribution can be effectively handled, especially suitable for rock tunnels or engineering environments with variable geological conditions.

[0019] Comprehensive quantitative analysis, supports construction decision making: Not only the precise calculation of single-section overbreak and underbreak is realized, but also multi-section trend analysis along the tunnel axis is carried out to evaluate the overall construction quality, output specific suggestions such as drilling and blasting parameter optimization, support timing adjustment, etc., and form a closed-loop quality control system of "measurement-analysis-feedback".

[0020] The engineering application effect is good, and the economic benefit is improved: Through accurate identification of overbreak and underbreak distribution law (such as arch top and side wall as key control areas), targeted repair and support are guided, concrete overconsumption and rework risk are reduced, construction cost is reduced, and safety and durability of tunnel engineering are improved.

[0021] In summary, the present application realizes the leap of tunnel overbreak and underbreak detection from "discrete, lagging, manual" to "continuous, real-time, intelligent" by fusing three-dimensional laser scanning, unsupervised learning and convex hull algorithm, and provides efficient, accurate and reliable digital technical support for drilling and blasting method tunnel construction. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 Point cloud processing and overbreak and underbreak real-time detection flowchart; Figure 2 Three-dimensional laser scanner space coordinate schematic diagram; Figure 3 Tunnel three-dimensional laser scanning point cloud model; Figure 4 Tunnel cross-section point cloud slice; Figure 5 Tunnel each section overbreak and underbreak histogram. DETAILED DESCRIPTION

[0023] The technical solutions of the present application are described below in combination with the drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present application, not all embodiments.

[0024] EMBODIMENT

[0025] The newly-built Xining-Chengdu Railway is located in the junction of Qinghai, Gansu and Sichuan provinces. The line starts from Xining, the capital of Qinghai Province, and goes south through Ping'an District of Haidong City, Huai Long County, Jianzha County and Tongren County of Huangnan Tibetan Autonomous Prefecture, then enters Gannan Tibetan Autonomous Prefecture of Gansu Province, and goes south through Xiahe County, Hezuo City and Luqu County of Aba Tibetan and Qiang Autonomous Prefecture of Sichuan Province, and connects to Huangshengguan Station of the under-construction Chengdu-Lanzhou Railway, and then goes into Chengdu Hub along the under-construction Chengdu-Lanzhou Railway. The total length of the line is 833.1 km, the project uses 307.75 km of the under-construction Chengdu-Lanzhou Railway and 26.3 km of the Lanzhou-Xinjiang High-speed Railway, and the newly-built main line of Haidongxi to Huangshengguan section is 498.913 km long, of which 172.651 km is newly-built in Sichuan Province, 183.198 km is newly-built in Gansu Province, and 143.064 km is newly-built in Qinghai Province. The XCTJ10 section is located in Tongren County and Jianzha County of Huangnan Tibetan Autonomous Prefecture in Qinghai Province, with an elevation of 2100m-2900m, starting and ending at DK435+043 and DK454+583.76, with a main line length of 19.558km. The beginning of the section is located at the right bank of Longmu River in Gualangdi Village of Dangshun Township in Jianzha County, and the end of the section is located at the right bank of the Yellow River in Dangshun Township in Jianzha County. This paper takes a tunnel in this section as the research object to calculate the overbreak and underbreak.

[0026] 1. Instruments and equipment In this project, a three-dimensional laser scanner is used to collect tunnel point cloud data. According to the size of the tunnel cross section and the performance index of the scanner, the stations are arranged in the longitudinal direction of the tunnel according to the principle of equal interval. The station spacing is determined to be 15 meters through theoretical calculation and field verification, ensuring that there is enough overlapping area between adjacent stations to facilitate subsequent point cloud registration.

[0027] 2. Point cloud preprocessing Twelve groups of initial point cloud data sets and spatial coordinates are obtained by scanning the tunnel 12 times with a three-dimensional laser scanner. Figure 2 Firstly, the point cloud data is thinned by using a downsampling algorithm based on uniformly distributed random numbers to reduce data redundancy. Then, the ICP algorithm is used to accurately register the multi-station scanning data to realize the spatial unification of the point cloud data. Due to the influence of construction equipment shielding, the point cloud data of the entrance and exit sections of the tunnel is significantly missing, so this study excludes this section and selects the 20-110m range with good point cloud integrity as the analysis object. Finally, a tunnel point cloud reference model is constructed based on the actual axis as shown in Figure 3 , providing a geometric reference for subsequent overbreak and underbreak analysis. This preprocessing process effectively improves the geometric accuracy and structural integrity of the point cloud data, meeting the precision requirements of engineering analysis.

[0028] 3. Cross-section feature extraction and convex hull calculation Based on preprocessing, a spatial transformation matrix was used to transform the coordinate system, making the point cloud perpendicular to the design axis. An equidistant cross-section analysis method was employed, with an analysis cross-section set every 10m along the tunnel's longitudinal direction. The thickness of each cross-section slice was set to 0.1m to ensure comprehensive coverage of the tunnel structure while maintaining the uniformity and systematic nature of the analysis. To further improve the quality of the point cloud data, a K-means denoising algorithm was used to denoise the point cloud slices, with a threshold set to (5, 20) to eliminate measurement noise and outliers, while ensuring that effective information in the point cloud was preserved to the maximum extent. This completed the feature extraction of the cross-section point cloud slices, providing an accurate data foundation for convex hull calculation.

[0029] For each denoised cross-section slice, the Graham convex hull algorithm was used to select convex hull vertices. The calculated number of convex hull vertices for each cross-section fluctuated between 60 and 110, primarily due to spatial differences in over- and under-excavation levels across different sections. The extracted convex hull vertices were connected clockwise to form a closed curve, constructing the actual excavation contour line for that tunnel cross-section, as shown below. Figure 4 The diagram shows the tunnel's designed cross-sectional outline in green, and the actual cross-sectional outline in red, with the points on the red line representing the vertices of the convex hull. Spatial comparison of these two lines visually reveals the distribution characteristics of over-excavation and under-excavation.

[0030] 4. Cross-sectional over-excavation and under-excavation analysis The permissible over- or under-excavation amounts for tunnels in Q / CR 9604—2015 "Technical Specification for Construction of High-Speed ​​Railway Tunnel Engineering" are shown in Table 1 below.

[0031] Table 1. Specifications for Permissible Over- and Under-excavation Volume in Tunnels

[0032] Over-excavation and under-excavation are calculated at each cross-section measuring point based on the convex hull vertices. Values ​​greater than 0 indicate over-excavation, and values ​​less than 0 indicate under-excavation. The distribution of over-excavation and under-excavation across ten cross-sections is visualized, as shown below. Figure 5 As shown in the figure. The results show that: 1) all sections exhibit significant over-excavation characteristics; 2) under-excavation occurs only locally and the values ​​are generally small; 3) the areas with severe over-excavation have obvious spatial distribution patterns, mainly concentrated in the sidewalls and the top of the arch.

[0033] The overbreak and underbreak of each section of the tunnel were statistically analyzed, as shown in Table 2. The analysis showed that there was a serious overbreak and underbreak in this tunnel section: from the extreme value distribution analysis, the extreme points appeared at the 20+0.1 m and 110+0.1 m sections, reaching the maximum overbreak of 0.473 m and the maximum underbreak of 0.334 m, respectively; from the statistical analysis of the proportion, the overbreak rate of each section exceeded 80%; the underbreak rate showed a fluctuating distribution of 0-20%; from the mean value distribution analysis, the average overbreak and underbreak exceeded 0.11 m. Compared with the traditional measurement method, the method proposed in this study greatly improves the detection speed and reduces the detection error by at least 11.89% and a maximum of 40.40%, indicating that the method proposed in this study is effective.

[0034] Table 2 Statistical indicators of overbreak and underbreak of single section of tunnel

[0035] Since the tunnel vault and the vault shoulder are the key parts of the tunnel stress, overbreak and underbreak can easily cause lining deformation and other diseases, and the repair difficulty is great, so the overbreak value of the convex part is analyzed separately. From the compliance analysis, the average overbreak value of the ten sections does not exceed the allowable average linear overbreak limit of surrounding rock (15 cm) specified in the Technical Specification for High-speed Railway Tunnel Engineering Construction (Q / CR 9604-2015), but the maximum overbreak value is seriously out of standard (see Table 3 for details), and the over-standard rate of the 70+0.1 m section is as high as 17%, and the risk of repair of the vault and shoulder is prominent.

[0036] The above analysis shows that there is a systematic deviation in the current construction quality control. It is necessary to optimize and adjust the drilling and blasting parameters and the timing of temporary support.

[0037] Table 3 Statistical indicators of overbreak of arch part of each section of tunnel

[0038] 5. Overbreak and underbreak trend analysis The histogram distribution characteristics and statistical parameters of the overbreak and underbreak of each section of the tunnel were systematically analyzed to obtain the overall overbreak and underbreak trend of the tunnel. The analysis is as follows: 1) from the spatial distribution of the histogram, the overbreak and underbreak shows a clear decreasing trend along the tunnel excavation direction; 2) the average overbreak and underbreak of each section generally decreases with the increase of the tunnel depth; 3) the standard deviation of the overbreak and underbreak also shows a decreasing trend with the increase of the excavation distance, reflecting the gradual reduction of the construction dispersion. In summary, with the continuous advancement of the tunnel project, the construction party effectively improves the control accuracy of the excavation profile by dynamically optimizing the drilling and blasting parameters and timely adjusting the support scheme, significantly improves the similarity between the actual excavation section and the design section, and gradually reduces the overbreak and underbreak as a whole, effectively improving the stability and reliability of the construction quality.

[0039] 6. Construction suggestions Based on the analysis of the overbreak and underbreak of the tunnel project, the following improvement suggestions are proposed for the subsequent construction: (1) Optimize the design of drilling and blasting parameters. For the side walls and the crown where overbreak is concentrated, it is suggested to appropriately reduce the hole spacing and correspondingly reduce the charge weight. For the underbreak section, the reverse adjustment strategy is adopted.

[0040] (2) Strengthen the partition control and implement the precise blasting mode of "section partition and charge segmentation", that is, the "one shot one design" dynamic adjustment mechanism, to adapt to different segment surrounding rock monitoring data.

[0041] (3) Perfect the measurement and monitoring system, measure the overbreak and underbreak during construction, and deal with the problems early to reduce the subsequent rework.

[0042] (4) Develop a grading repair scheme: for overbreak areas, the principle of "support first, then process" must be followed. If the overbreak is small, directly spray concrete to smooth to the design contour line, if it is large, you can hang a reinforced steel mesh or make a steel arch support before spraying concrete in layers. For local underbreak areas, if the underbreak is small, use small machinery for precise repair and combine anchor shotcrete support for fine adjustment of the section, if the underbreak is large, use secondary blasting repair technology.

[0043] (5) Implement the quality responsibility system, link the section deviation rate with the performance of the construction team, form a virtuous cycle of "measurement-construction-evaluation", and promote fine construction management.

Claims

1. A method for point cloud feature extraction and over-excavation / under-excavation detection in drill-and-blast tunnels using unsupervised learning, characterized in that... Includes the following steps: S1. Obtain the initial point cloud data of the tunnel through three-dimensional laser scanning, and perform thinning and registration preprocessing in sequence; S2. Axis Correction: Extract the actual excavation axis from the registration point cloud, calculate its spatial transformation matrix with the design axis, and correct systematic deviations. S3. Section slice extraction: Cut slices with a thickness of δ at fixed intervals Δs along the design axis direction, where Δs=10m and δ=0.1m; S4. Point Cloud Denoising: K-means clustering algorithm is used for sliced ​​point clouds. The number of clusters N is determined according to the contour coefficient. Noise clusters are identified and removed by using density thresholds: 3~5 for sparse point clouds and 10~20 for dense point clouds. Deep convolutional generative adversarial network DCGAN is used as an auxiliary denoising method. The generative model distinguishes noise from real features and avoids the loss of details caused by threshold dependence. S5. Convex Hull Generation: Graham algorithm is used to filter vertices of the convex hull after noise reduction of the point cloud, and they are connected sequentially to form the actual excavation outline. S6. Over-excavation and under-excavation calculation: Calculate the shortest distance di between the vertex of the convex hull and the design contour line. di>0 indicates over-excavation, and di<0 indicates under-excavation. S7. Trend Analysis: Statistically analyze the over-excavation and under-excavation of each section along the tunnel axis, assess the overall construction quality, and output adjustment suggestions.

2. The unsupervised learning method for point cloud feature extraction and over-excavation / under-excavation detection in drill-and-blast tunnels according to claim 1, characterized in that, In step S1, the 3D laser scanner calculates the reflection time by emitting laser pulse signals to determine the distance observation value S between the target point and the scanner. Simultaneously, it controls the encoder to synchronously record the horizontal and vertical scanning angle observation values ​​α and θ for each laser pulse. In the instrument's internal coordinate system Pi, the X-axis is defined as lying in the horizontal scanning plane, the Y-axis is perpendicular to the X-axis and also lies in the horizontal scanning plane, and the Z-axis is orthogonal to the scanning plane. Based on this coordinate system and the measurement data, the 3D coordinates (XS, YS, ZS) of the scanning point P can be obtained, calculated using the following formula: The raw data is generally stored in the form of point clouds, which can be preprocessed using specialized point cloud data software. The specific steps are as follows: (1) Thinning: Due to the large amount of original point cloud data, the computation and storage costs increase. Different sampling methods can be used to reduce the density of the point cloud while preserving the original point cloud features. Common methods include thinning algorithms based on uniformly distributed random numbers and voxel filtering. (2) Registration: Merge multiple sets of point cloud data obtained from scanning and convert them to a unified coordinate system. Common methods include ICP and FPFH.

3. The unsupervised learning method for point cloud feature extraction and over-excavation / under-excavation detection in drill-and-blast tunnels according to claim 1, characterized in that, In step S2, the actual tunnel excavation axis is extracted from the registered point cloud, and the spatial transformation matrix between the design axis and the actual axis is calculated by the least squares method to correct the systematic deviation of the point cloud data.

4. The unsupervised learning method for point cloud feature extraction and over-excavation / under-excavation detection in drill-and-blast tunnels according to claim 1, characterized in that, In step S4, K-means is a clustering algorithm whose main purpose is to divide the dataset into N non-overlapping clusters. The sum of squared errors within the clusters is minimized as shown in equation (2), that is, by iteratively optimizing the cluster centers, the data points within the clusters are made as close as possible while the differences between the clusters are as large as possible. Where: J is the sum of squared errors within the cluster, Ci is the point set of the i-th cluster, and p is the data point in Ci. q represents the squared Euclidean distance between a data point and the cluster center. i is the i-th cluster center; K is the number of clusters, and X is any data point in the dataset; The algorithm first randomly selects N data points as the initial cluster centers, and then assigns each data point to the cluster where the nearest initial cluster center is located according to the Euclidean distance, as shown in Equation (3). Then, it calculates the mean of all points in each cluster to update the position of the cluster center as shown in Equation (4), and iterates until the change of the cluster center is less than a certain set threshold or the preset number of iterations is reached. Where: d is the Euclidean distance from each point to the cluster center, (x1, y1, z1) are the three-dimensional coordinates of each point, and (x0, y0, z0) are the three-dimensional coordinates of q; After setting a reasonable threshold for the data points in the clusters divided by the K-means algorithm, it can be applied to point cloud denoising. The specific operation is as follows: For each point cloud slice, the appropriate number of clusters N is determined using the silhouette coefficient method, and the optimal clustering result for each cluster is determined through the update and iteration process of the K-means algorithm. After clustering, a density threshold is set to analyze the size and density of each cluster to identify noise clusters. Generally, smaller clusters or clusters with lower density are more likely to be noise. To more accurately identify noise points, a reasonable density threshold can be set according to the actual situation of the initial point cloud: for sparse point clouds, the threshold is set to (3, 5); for dense point clouds, it is set to (10, 20). A deep convolutional generative adversarial network (DCGAN) is used as an auxiliary noise reduction method to distinguish noise from real features through a generative model, avoiding the loss of details caused by threshold dependence. When the number of points in a cluster is less than the set threshold, the points in that cluster can be considered as noise points. Points identified as noise in the clusters are removed from the point cloud to obtain denoised point cloud data.

5. The unsupervised learning method for point cloud feature extraction and over-excavation / under-excavation detection in drill-and-blast tunnels according to claim 1, characterized in that, In step S6, over-excavation and under-excavation refer to the spatial relationship between the actual excavation outline and the design cross-section: when the actual convex hull outline exceeds the design cross-section outline, it is called over-excavation; When the design section outline encloses the actual convex hull outline, it is called under-excavation; The core of over- and under-excavation calculation is to compare the spatial relationship between the actual convex hull and the design cross-section to quantify the over- and under-excavation amounts. A dashed line is drawn from the design axis to the vertex of the convex hull, intersecting the tunnel design cross-section outline at a point. The distance between the vertex of the convex hull and this point is the over- and under-excavation amount, calculated using the following formula: Where di is the over- or under-excavation amount of a vertex i of a certain convex hull, xi is the x-coordinate of i, yi is the y-coordinate of i, xh is the x-coordinate of the point on the design section closest to i, and yh is its y-coordinate.

Images