Bolt-grouting joint shear failure evolution monitoring method and system based on three-dimensional scanning

Abstract

The invention relates to the field of geotechnical engineering and rock mechanics research, in particular to a bolt-grouting joint shear failure evolution monitoring method and system based on three-dimensional scanning, and the method comprises the steps: carrying out the multi-view three-dimensional scanning of a target rock mass at different load stages, and obtaining a time-sharing three-dimensional point cloud reflecting crack dislocation and surface deformation; and multi-modal data collected by a strain sensor, a vibration sensor and the like are input into a multi-scale feature extraction algorithm to generate automatic regulation and control parameters for the grouting pressure and the load mode. During application, a researcher can dynamically adjust a grouting scheme or a load strategy, and the whole crack propagation process is displayed by using a three-dimensional deformation diagram and a crack distribution diagram, so that the monitoring process is more visual and operable. According to the method, the key damage symptom of the rock mass in the loading process can be captured, the plugging effect of bolting and grouting can be effectively evaluated, and more visual and accurate data support is provided for rock mechanics research and engineering field reinforcement decision making.

Description

Technical Field

[0001] The present invention relates to the fields of geotechnical engineering and rock mechanics research, and in particular to a method and system for monitoring the shear failure evolution of anchor-grout joints based on three-dimensional scanning. Background Art

[0002] The technology for monitoring the evolution of shear failure of anchor-grout joints is mainly aimed at geotechnical engineering and rock mechanics research scenarios. By capturing the crack propagation and surface deformation of rock joints during the shear process, it provides accurate data support for reinforcement design and failure mechanism analysis. This idea has attracted attention in recent years because shear failure of joint surfaces often leads to a significant weakening of the overall stability of the rock mass, and traditional methods based on a small number of measurement points or post-slice analysis are difficult to capture crack evolution in real time and with high resolution. Many existing technologies only use simple displacement meters or two-dimensional image monitoring, resulting in insufficient restoration of the true failure morphology of the joint surface. Grouting and load adjustment also rely mainly on human judgment, making it difficult to capture signs of instability or evaluate the reinforcement effect in a timely manner. Summary of the Invention

[0003] In response to the many problems existing in the above-mentioned prior art, the present invention provides a method and system for monitoring the evolution of shear failure of anchor-grouting joints based on three-dimensional scanning. The present invention uses time-sharing three-dimensional point cloud acquisition and strain and vibration signal capture, combined with multi-layer convolution or fusion algorithms, to explore the crack growth patterns of joint surfaces under shear loads. The deep learning module outputs the control parameters of the grouting pressure or load mode based on real-time deformation and mechanical characteristics, and then suppresses crack propagation through measures such as grouting or load reduction. The system visualizes the execution process and the measured deformation and signal data in the form of three-dimensional deformation maps and crack distribution maps, which can not only reduce the risk of instability, but also provide a reliable basis for on-site reinforcement and scientific research analysis.

[0004] A method for monitoring the shear failure evolution of anchor-grouting joints based on three-dimensional scanning comprises the following steps:

[0005] The perspective data obtained by multi-angle observation of the target rock mass are spliced and denoised to generate the initial three-dimensional point cloud;

[0006] Applying a shear load to the target rock mass and collecting a time-sharing three-dimensional point cloud at each loading stage, performing registration or differentiation relative to the initial three-dimensional point cloud, identifying the scope of surface cracks and recording shear failure evolution information through geometric and topological feature analysis, and synchronously collecting strain signals and vibration signals at each loading stage;

[0007] Inputting the time-sharing three-dimensional point cloud, strain signal, and vibration signal into a deep learning model for multi-scale feature extraction to generate control parameters for grouting and shear load;

[0008] Adjust the grouting pressure or loading method according to the control parameters, and visually integrate the execution process with the time-sharing three-dimensional point cloud, strain signal, and vibration signal through a three-dimensional deformation map and a crack distribution map to obtain a monitoring report on the shear failure evolution of the grouted joint.

[0009] Preferably, when applying a shear load to the target rock mass, a step-by-step incremental method is adopted. After each level of load reaches the set value, a constant load period is maintained and the time-sharing three-dimensional point cloud is collected for comparing the crack changes under different load conditions.

[0010] Preferably, when the crack propagation rate exceeds a pre-set threshold, reduce the increment of the next load level and simultaneously collect the strain signal and vibration signal to improve the acquisition density of the time-sharing data.

[0011] Preferably, the geometric and topological feature analysis is carried out by comparing the displacement differences between points in the time-sharing three-dimensional point cloud, and the part exceeding the threshold is marked as the crack dislocation area.

[0012] Preferably, the topological feature analysis performs a hole search based on the adjacency relationship on the marked area to judge the crack trend and connectivity and generate the crack propagation path.

[0013] Preferably, the deep learning model uses a multi-layer convolution structure to perform coordinate correction on the time-sharing three-dimensional point cloud, performs feature encoding on the strain signal and vibration signal, and then performs a fusion operation with the time-sharing three-dimensional point cloud.

[0014] Preferably, the area with noise or missing points in the time-sharing three-dimensional point cloud is first subjected to a local density-based correction operation and then input into the multi-layer convolution structure.

[0015] Preferably, after comparing the control parameters output by the deep learning model with the crack propagation rate, determine the adjustment step of the grouting pressure or the change interval of the loading method, and collect the time-sharing three-dimensional point cloud after grouting to determine the crack closure degree.

[0016] Preferably, superimpose the grouting process and the crack propagation trajectory through the three-dimensional deformation map and the crack distribution map, and list the time-series data of the strain signal and vibration signal in the monitoring report to show the corresponding relationship between loading and crack evolution.

[0017] A monitoring system for the shear failure evolution of grouted joints based on three-dimensional scanning, used to implement the monitoring method for the shear failure evolution of grouted joints based on three-dimensional scanning, and the system includes:

[0018] A three-dimensional scanning data acquisition module, used to observe the target rock mass from multiple angles and generate an initial three-dimensional point cloud after splicing and denoising the observation data;

[0019] The shear loading and signal acquisition module is used to apply shear loads to the target rock mass and collect time - division three - dimensional point clouds, strain signals, and vibration signals at each load stage;

[0020] The multi - scale feature extraction module is used to perform feature encoding on the time - division three - dimensional point clouds, strain signals, and vibration signals and generate regulation parameters for grouting and shear loads;

[0021] The regulation and visualization output module is used to adjust the grouting pressure or loading mode according to the regulation parameters, and display the execution process together with the time - division three - dimensional point clouds, strain signals, and vibration signals in the form of three - dimensional deformation maps and crack distribution maps, and output a monitoring report on the shear failure evolution of the grouted joint.

[0022] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0023] Through the multi - perspective three - dimensional scanning technology means, the full - field deformation detection effect of the rock mass surface at different load stages is realized, and the positions and scales of joint crack cracking and dislocation are comprehensively captured.

[0024] Through the means of fusing multi - source sensing and deep learning, the dynamic regulation effect of the grouting pressure or loading mode is realized, and it is automatically determined whether pressure reinforcement or load reduction for stability preservation is required.

[0025] Through the visualization module combined with the three - dimensional deformation map and crack distribution map technology means, the intuitive display effect of the crack propagation process and strain and vibration signals is realized, providing a more complete basis for subsequent decision - making and reporting. Brief Description of the Drawings

[0026] Figure 1 is a schematic flow chart of the method of the present invention;

[0027] Figure 2 is a structural block diagram of the system of the present invention. Detailed Embodiment

[0028] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for the sake of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure.

[0029] As Figure 1 shown, a monitoring method for the shear failure evolution of grouted joints based on three - dimensional scanning includes the following steps:

[0030] After splicing and denoising the perspective data obtained from multi - angle observations of the target rock mass, an initial three - dimensional point cloud is generated;

[0031] Obtain the complete geometric information of the surface of the target rock mass through multi-angle observations, and splice and denoise these observation data in a unified coordinate system to generate an initial three-dimensional point cloud, which serves as a reference for subsequent core steps such as deformation comparison or crack identification. The specific principles include the following key points:

[0032] By arranging pairs of projectors and imaging devices at different orientations, project beams with characteristic encodings onto the surface of the rock mass successively, and use cameras to record the pattern offsets formed by the beams on the surface. This process will produce several perspective data with overlapping areas for subsequent splicing and integration.

[0033] Align the local three-dimensional data corresponding to each observation perspective. Usually, a transformation model centered on the rotation matrix R and the translation vector t can be adopted to minimize the corresponding error between the local point cloud {p i} and the reference point cloud {q i}. One of the common methods is the iterative least squares registration algorithm. In this algorithm, set the objective function:

[0034]

[0035] where, p i is the i-th three-dimensional point in the local point cloud, q i is the three-dimensional point in the reference point cloud corresponding to p i , R is the rotation matrix to be solved, and t is the translation vector to be solved. By iteratively updating R and t multiple times, the splicing result that minimizes the error E is finally obtained.

[0036] After aligning the local point clouds, it is necessary to perform filtering or elimination on the noise points caused by overlapping or observation blind spots. Judgment criteria based on neighborhood density or curvature can be established to delete or interpolate and repair the discrete points that deviate significantly from the distribution of surrounding points. After the processing, multiple aligned sub-point clouds are merged into a complete initial three-dimensional point cloud, reflecting the original geometric shape of the target rock mass under the condition of "without applying any additional loads".

[0037] Multi-angle observations and point cloud splicing can completely cover the surface details of the target rock mass, reduce the observation blind spots, and make the subsequent deformation comparison more accurate. Through denoising and coordinate unification, the finally obtained initial three-dimensional point cloud can be used as a reference benchmark in subsequent shear processes and crack analysis, facilitating the quantification of crack offset or surface displacement. For rock masses with different scales or surface complexities, by adjusting the number of observation angles and the parameters of the splicing algorithm, the three-dimensional characteristics of different time periods or different parts can be flexibly obtained; this is particularly crucial for subsequent "monitoring of the evolution of shear failure of grouted joints" in the present invention.

[0038] In one embodiment: A geotechnical laboratory obtains three-dimensional scanning reference data for a jointed rock mass with dimensions of approximately 1.5 meters in length, 1.2 meters in width, and 0.8 meters in height:

[0039] Six observation positions are deployed around the rock mass. Each time, a fringe grating with phase-shift encoding is projected onto the surface, and a single industrial camera records the pattern deformation; a total of six sets of local three-dimensional data files are obtained, with each set containing 200,000 to 300,000 vertices.

[0040] Arbitrarily select any one set of data as the reference point cloud in sequence, and use the iterative registration method solved by the rotation matrix and translation vector to splice it with the other five sets of data one by one. To improve efficiency, feature points or markers can be used to assist registration in the initial stage, and the number of non-linear iterations is controlled within 10 times.

[0041] Perform neighborhood density screening on the spliced point cloud, delete low-density outliers, and finally obtain a three-dimensional model file containing approximately 1.5 million points, with an accuracy of approximately ±0.2 mm. In subsequent shear tests, this point cloud is used for differential comparison to accurately observe crack movement and rock mass surface dislocation under each load level.

[0042] The initial three-dimensional point cloud formed by the above process is the reference data for subsequent time-sharing monitoring under shear load. Once the rock mass deforms or crack propagates, it can be registered or differenced with this initial point cloud in the same coordinate system to accurately judge the amount of deformation, crack direction, and scale.

[0043] Apply a shear load to the target rock mass and collect time-sharing three-dimensional point clouds at each load stage, register or difference them with respect to the initial three-dimensional point cloud, identify the surface crack range through geometric and topological feature analysis, record the shear failure evolution information, and synchronously collect strain signals and vibration signals at each load stage;

[0044] Apply a shear load to the target rock mass and collect time-sharing three-dimensional point clouds at each load stage. This process aims to observe and quantify the surface deformation and crack evolution of the rock mass under the action of gradually increasing or continuously maintained shear forces. Before obtaining the time-sharing three-dimensional point clouds, an initial three-dimensional point cloud has been established as the reference data for subsequent comparison. By registering or differencing the three-dimensional point clouds obtained at each load stage with the initial three-dimensional point cloud, the local displacements, surface dislocations, and crack paths of the rock mass under different stress conditions can be determined. The core value of this method lies in: through repeated observations and cumulative comparisons, the gradually increasing rupture range of the joint surface under the action of shear forces can be revealed from the geometric and topological dimensions, while recording the evolution of crack morphology. In order to make these deformation characteristics and mechanical response information corroborate each other, strain signals and vibration signals are synchronously collected at each load stage to obtain richer physical indicators.

[0045] At the principle level, when analyzing the time - shared three - dimensional point cloud from a geometric perspective, a set of rotation and displacement parameters can be used to align the point cloud at that moment to the coordinate system where the initial three - dimensional point cloud is located, and then the difference calculation is performed between the two sets of point clouds.

[0046] From a topological perspective, the three - dimensional point cloud can be presented in the form of a grid or a point cloud map. Geometric and topological feature analysis, on the one hand, checks the distance or normal vector differences between points and their neighboring points, and on the other hand, conducts a connectivity search to see if some suspicious high - displacement grid cells form an entire continuous crack band. When a series of closely adjacent high - displacement grid cells form a channel, it can be judged that there is a significant crack trend there. Moreover, if the high - displacement area forms an unclosed cavity in the point cloud, it may mean the presence of penetrating fissures. In the present invention, automated crack identification and connectivity determination are achieved through geometric and topological feature analysis, and the results can be cross - verified with strain signals and vibration signals in subsequent steps. For example, if the analysis shows that the topological cavity in a certain area expands rapidly, and there is also a sudden increase in amplitude in the strain record, it implies that new joint surfaces may be opened or existing cracks may further expand in this area under shear loading.

[0047] In practical applications, in order to make the loading and data acquisition targeted, it is necessary to obtain time - shared three - dimensional point clouds at different load stages, aiming to track the evolution of the rock mass during the load - increasing or steady - load - maintaining stages. If each level of load is maintained for a specific time interval, three - dimensional scans can be performed at each stage interval, thus obtaining a series of time - sequenced three - dimensional data. Compared with the application of a single large load, this operation of gradually increasing the load or maintaining the load in stages can capture more transition information. If it is found that geometric and topological feature analysis shows that cracks are quickly connected at a certain level of load, and once combined with the peak values of strain signals and vibration signals, it is very likely that it is at the verge of approaching failure at this moment, sounding a warning for subsequent safety control.

[0048] The synchronous acquisition of strain signals and vibration signals brings multi-angle mechanical and dynamic indicators to the present invention. Strain signals can be captured by strain gauges, fiber Bragg gratings or other sensors pasted or embedded on the surface of the rock mass to obtain the strain distribution inside the material. If at a certain load level, the values of the strain sensors near the crack soar, it can be mutually corroborated with the three-dimensional point cloud differential result: there may be a large dislocation or significant displacement at the corresponding position in the three-dimensional point cloud. Vibration signals can indicate whether there is obvious energy release and intense friction in the rock mass during loading. Some studies have shown that when a large number of microcracks inside the rock mass begin to form and penetrate each other, the vibration signal often has specific spectral characteristics, and several high-frequency energy peaks can be located through fast Fourier transform. If these peaks match the crack network revealed by geometric topology analysis, it can indicate that the crack is growing at a relatively fast speed during this period. For example, researchers can simultaneously plot the peak frequency band in the vibration signal energy spectrum and the crack connectivity index of the point cloud as a time series curve to evaluate when entering the critical interval of rapid crack expansion.

[0049] Preferably, when applying shear load to the target rock mass, a step-by-step incremental method is adopted. After each level of load reaches the set value, a constant load period is maintained and time-sharing three-dimensional point clouds are collected for comparing crack changes under different load conditions.

[0050] When applying shear load to the target rock mass, a step-by-step incremental method is adopted. After each level of load reaches the set value, a constant load period is maintained and time-sharing three-dimensional point clouds are collected for comparing crack changes under different load conditions. This idea aims to observe the failure process of the rock mass at different times by gradually increasing the shear force, and collect high-fidelity three-dimensional data at each stage to provide detailed basis for subsequent deformation and crack evolution analysis. By performing differential comparison between the collected time-sharing three-dimensional point clouds and the initial three-dimensional point cloud or the time-sharing three-dimensional point clouds under adjacent load conditions, the deformation increment, crack initiation and expansion characteristics of the rock mass at each load level can be grasped. With the support of three-dimensional scanning technology, this step-by-step incremental loading method can significantly improve the detection accuracy of the rock mass failure process and has been effectively verified in experimental and engineering applications.

[0051] In terms of principle, the step-by-step incremental loading method can split an overall shear loading process into several continuous and observable stages. Assuming that the maximum planned value of the shear load is denoted as F max , it can be divided into several levels, such as F1, F2,..., F n , and each level of load maintains a stable period, usually denoted as Δt. During the Δt of maintaining this load, the surface of the rock mass is three-dimensionally scanned to obtain the time-sharing three-dimensional point cloud of this stage. At the same time, the aforementioned deep learning model or geometric and topological feature analysis method can be used to compare between the two consecutive loads (for example, from F i-1 to Fi ) Differences in the coordinates of the rock mass surface.

[0052] Specifically, denote the point cloud obtained at the end of the i-th level of load as {P i}, and the point cloud obtained at the end of the (i - 1)-th level of load as {P i-1}. After coordinate registration, the displacement vector Δv (i-1) →i of each corresponding point can be defined, and then analyze how many points have displacement amounts exceeding the threshold, so as to determine what kind of law the crack increment shows in these high-displacement regions. When a large number of displacements exceeding the threshold are detected to gather in a certain area, it can be inferred that the crack in this area is rapidly expanding or there is significant dislocation along the joint surface.

[0053] In the application of the present invention, the hierarchical increment method can enable researchers to better capture the non-linear failure characteristics of the rock mass in the critical load interval. If the shear load is rapidly increased to F max at one time, the crack will often expand or penetrate rapidly in a short time, resulting in the lack of opportunity for researchers to observe its gradual evolution path. On the contrary, the hierarchical increment approach can maintain for a period of time at each critical load level. This period of time not only ensures the slow adjustment of the rock mass stress field balance, but also provides a sufficient recording window for three-dimensional scanning and signal acquisition devices such as strain and vibration. High-precision point cloud data can be obtained in a relatively stable environment. Once it is observed that at a certain level of load, the crack connectivity suddenly increases and typical high-amplitude fluctuations also appear in the strain or vibration signals, it indicates that the rock mass has entered the instability interval, and it is necessary to decide in a timely manner whether to continue increasing the load or take safety measures such as grouting and anchoring according to the monitoring data.

[0054] At the same time, the hierarchical increment loading also avoids the risk of the rock mass being subjected to excessive shear impact in a short time and reduces the uncertain factors in the experimental process. For some target rock masses with more complex joint development or uneven rock quality, the crack may show a sudden acceleration of expansion in a certain load interval, while it is relatively stable in other load intervals. If the hierarchical increment method is adopted, the specific load range of this accelerated expansion can be located interval by interval, and the main crack direction, branch trend and potential hidden cracks that may penetrate within this range can be automatically identified by combining time-sharing three-dimensional point clouds. For example, if from the first level of load {F1} to the second level of load {F2}, the three-dimensional point cloud difference shows only sporadic crack increments, it indicates that the overall rock mass is still relatively safe; but when from {F2} to {F3}, the crack network suddenly becomes connected in patches, which indicates that {F3} is the critical load demarcation point and requires subsequent plans to pay high attention or introduce more reinforcement means.

[0055] In practice, the difference between each load level and the next can be set to a fixed value, ΔF, or dynamically adjusted based on the previously observed crack growth rate. If the crack growth rate is detected to be below a certain threshold, the next load increment can be increased appropriately. If the crack growth rate significantly exceeds the threshold, the next load increment can be reduced to prevent sudden failure of the rock mass. For example, suppose an indoor shear test is conducted on a rock mass with an overall target maximum shear force of 45kN. This can be divided into nine levels, each of 5kN. After each shear force level is applied, a 5-minute constant load period is maintained, during which point cloud data is collected using a 3D scanner and combined with strain gauges or acoustic emission sensors to record stress, strain, and vibration information. If clear signs of crack penetration are observed at load level 6, the load level 7 can be temporarily reduced from 5kN to 2.5kN for subsequent safety considerations, and more intensive 3D scanning can be performed during this period. In this way, more detailed crack evolution data can be obtained within a safe and controllable range. Ultimately, a complete crack propagation time series diagram can be constructed through a series of time-sharing 3D point clouds, reflecting the local deformation zones and overall failure modes of the rock mass in the form of a 3D model.

[0056] This operation process brings significant results: first, the experiment can continuously track the entire process of cracks from tiny cracks to large-scale connections or dislocations; second, it allows researchers to suspend force application or adjust the loading scheme at critical moments, reducing damage caused by unexpected instability and ensuring the integrity of experimental data; third, by comparing the three-dimensional point cloud difference results under each load level, a "load-crack area" or "load-crack extension rate" curve can be output. These data are particularly valuable for deep learning or numerical simulation because they contain information on the progressive failure mechanism of rock masses under different stress levels. In this way, the monitoring process based on the present invention can not only provide the final failure form, but also reveal the process behavior and critical signs of rock masses under shear loads.

[0057] Preferably, when the crack growth rate exceeds a preset threshold, the increment of the next load level is reduced and the strain signal and the vibration signal are collected simultaneously to increase the collection density of the time-sharing data.

[0058] When the crack growth rate exceeds a preset threshold, the increment of the next load level is reduced and strain signals and vibration signals are collected simultaneously to increase the acquisition density of time-sharing data. This is a key means to ensure the safety of the rock mass and obtain detailed monitoring information in multi-level loading tests. In the process of monitoring the evolution of shear failure of anchor-grout joints based on 3D scanning, the crack growth rate is usually obtained by comparing the time-sharing 3D point cloud with the previous reference point cloud, and is obtained by measuring the displacement difference of the surface point coordinates in adjacent load stages. Assuming that the crack width change rate of a certain area is defined as where Δw represents the increment of crack width within the time interval Δt. When exceeds a critical value obtained through prior experiments or accumulated experience, it indicates that the crack is rapidly expanding and there is a possibility of triggering rock mass instability. In this case, in order to avoid large-scale damage caused by continuing with the same or higher load increment in the next stage, it is necessary to specifically reduce the increment of the next load level, subjecting the rock mass to a smaller mechanical impact, so that scientific research or engineering personnel can deeply observe the formation mechanism and subsequent development trajectory of the crack at this location.

[0059] Since 3D scanning can capture the geometric information of the entire rock mass surface, while strain signals and vibration signals can reflect the stress concentration and energy release conditions inside or on the surface of the rock mass, the combination of these three enables the monitoring personnel to evaluate the progress of the crack from multiple perspectives. If only judging the crack development from the 3D point cloud, it is difficult to know whether it is accompanied by a large amount of stress accumulation or periodic microseismic activities; but after incorporating strain and vibration data, a complete "geometry - mechanics - dynamics" feedback system can be formed. Once the crack propagation rate breaks through the threshold, it indicates that significant deformation has occurred at the geometric level, often corresponding to internal stress accumulation. At this time, reducing the load increment can avoid instability failure and give the monitoring device more time to densely observe the details on the surface or inside of the rock mass during this critical stage.

[0060] To illustrate the implementation of this approach more specifically, an example of a laboratory shear load test is given. On a certain geotechnical testing machine, scientific research personnel loaded a jointed rock mass in a stepped increment manner, with the initial increment set at 5 kN, maintaining each level for five minutes and obtaining time - shared 3D point clouds, strain signals, and vibration signals. As the load was increased step by step, the research team aligned the time - shared 3D point clouds with the initial 3D point cloud through image registration technology, and then found that the increment value of the crack width at a certain location on the rock mass surface increased rapidly. For example, at the end of the 3rd load level, by comparing the point clouds at the end of the 2nd and 3rd levels, it can be measured that the average crack width in this area increased by 0.25 mm within the Δt time, and after conversion If it was previously determined in the safety assessment that If it is 0.03 mm / min, it can be determined that the crack propagation rate has exceeded the threshold. At this time, the strain gauge also shows a significant change in the crack surrounding area compared with before, and the vibration sensor captures the pulses of several high-amplitude microseismic events during this period. Multiple indicators indicate that the crack is entering the unstable expansion stage. To avoid catastrophic expansion caused by the next-level load based on the original 5 kN increment, the operator will, according to the above judgment, reduce the next-level load increment to 2 kN or 3 kN, and maintain a longer constant load observation after applying it to the new load value. At the same time, the frequencies of three-dimensional scanning, strain detection, and vibration acquisition are increased. This can not only protect the rock mass and equipment but also record more intensive time-sharing data at the critical stage, thus providing more microscopic details for subsequent data processing and model analysis.

[0061] The process of mutual verification formed by three-dimensional scanning, strain signals, and vibration signals is particularly significant in this link: if the growth rate of the crack slows down significantly after reducing the load increment and the number of peak values of the vibration signal drops substantially, it indicates that this measure effectively inhibits crack propagation; if it is still found that the crack continues to grow at an excessive speed, it may indicate that the crack has spread in deeper or more concealed joint surfaces. At this time, further plugging and anchoring measures need to be taken in on-site tests or engineering applications. These judgments are based on the high-precision geometric observation results provided by time-sharing three-dimensional point clouds and also rely on the mechanical support of strain and vibration data feedback.

[0062] At the same time, increasing the acquisition density of time-sharing data is particularly helpful for subsequent numerical simulation and deep learning model training. Many deep learning models require stable and dense input sequences for time series prediction. If more frames of three-dimensional point clouds and sensing signals can be obtained during the crack mutation stage, more reliable time series information can be brought to the model. For example, before the transient crack penetration, the system will record a more detailed crack extension process and may be able to capture the rapid breakthrough phenomenon after a short pause of the crack. This kind of phenomenon is likely to be ignored or blurred by relying only on the conventional low acquisition frequency mode, thus missing extremely crucial evidence of the fracture mechanism.

[0063] It can be seen that the idea reflected in "when the crack propagation rate exceeds the pre-set threshold, reduce the increment of the next load level and simultaneously collect strain signals and vibration signals" is not only to "step on the brakes" for the rock mass failure process at the load control level but also to capture more critical and complex crack evolution details, thereby further deepening the understanding of the joint failure process. Through this measure, the present invention plays a greater potential in accurately revealing the spatio-temporal distribution of rock mass failure and crack dynamics characteristics, enabling researchers to dynamically balance safety and data requirements, both effectively preventing instability risks and bringing more observation periods for three-dimensional scanning and multi-source sensing means, accumulating richer deformation and signal evidence for subsequent systematic analysis.

[0064] Preferably, the geometric and topological feature analysis compares the displacement differences between points of the time - shared three - dimensional point cloud, and marks the part exceeding the threshold as the crack dislocation area.

[0065] The geometric and topological feature analysis compares the displacement differences between points of the time - shared three - dimensional point cloud, and marks the part exceeding the threshold as the crack dislocation area. This is a highly practical determination method in the process of monitoring the shear failure evolution of grouted joints based on three - dimensional scanning. In principle, the time - shared three - dimensional point cloud refers to the three - dimensional data of the rock mass surface collected at different times or different load stages. To ensure that the deformation differences can be compared within the same coordinate system, the point clouds obtained at each stage are usually registered first to align them with the point cloud under the initial or previous load. After registration, the displacement differences between the coordinates of each grid point or each sampling point can be directly defined at the geometric level. If certain areas in the difference results are found to have values significantly larger than other areas or exceed a certain threshold, it can be determined that these areas have experienced obvious dislocation or tensile cracking at this load stage.

[0066] When further understanding this process, it is necessary to clarify what displacement difference is. In most three - dimensional scanning applications, each point cloud consists of a set of three - dimensional coordinates, denoted as {x i}. If the coordinates of the same position in the reference point cloud are defined as and are denoted as in the time - shared point cloud, then a basic deformation vector can be given. The larger the value of ||Δx i ||, the more significant the spatial movement of this grid point or this pixel point. To simplify the determination, a threshold ∈ can be set. When ||Δx i ||>∈, this part is marked as the "crack dislocation area". The value of ∈ depends on the specific experimental object, sensor accuracy, and acquisition noise level. For example, in some small - scale rock experiments, ∈ can be set at 0.1 mm, while in a large - scale geotechnical engineering environment, the threshold may need to be increased to several millimeters or even larger.

[0067] When this threshold determination is extended to the local neighborhood, the connectivity of the high-displacement region can be detected by combining topological feature analysis. Specifically, there are adjacency relationships in many 3D point clouds or mesh data. If some high-displacement points are close to each other and form a connected band, it represents a possible crack trend or crack activity band. When this band shows a concentrated, high-displacement "strip-shaped" or "sheet-shaped" distribution in the point cloud map, combined with the experience of rock mechanics and the analysis of joint contact surfaces, it can be basically confirmed as the area where shear failure activity is most active. If it is further compared with sensor information such as strain or vibration and it is found that there are peaks of stress concentration or energy release at the same time, it can further confirm its potential damage degree. The advantage of this method is that it can not only obtain the distribution of cracks in the spatial geometry level, but also help researchers quantify the speed and scale of the fracture process from the perspective of deformation, without the need to additionally disassemble or break the rock mass.

[0068] In this analysis method, "geometric and topological features" play an important role: geometry focuses on numerical distance or displacement difference, while topology emphasizes indicators such as point cloud adjacency, connectivity, and hole detection. When there are enough points with displacement differences exceeding the threshold and they are determined to be within the same connected component, it is very likely that there is significant dislocation in this area or a new through-crack is being formed. If some scattered high-displacement points are isolated, it may be caused by noise or just the shedding of loose particles on the rock surface, and it does not necessarily constitute a complete crack. In the present invention, by using both difference determination and neighborhood connectivity search, cracks and noise points can be effectively distinguished, thereby improving the interpretation accuracy.

[0069] This process has obvious effects in practical applications, especially suitable for rock masses with uneven joint development or complex crack trends. The following situation can be referred to as an example: A jointed rock mass with a length of 1.2 meters, a width of 0.8 meters, and a height of 0.6 meters is loaded in a shear testing machine, and the initial 3D point cloud records the surface coordinates of the rock mass under zero load. Subsequently, whenever the load increases to a specified value or a crack propagation signal is observed, a new 3D scan is performed and aligned to the initial coordinate system. Suppose in the third scan, the coordinate differences in some areas reach or exceed 3 mm, while most of the other areas are less than 1 mm. Mark the set of coordinates with differences exceeding 3 mm as R. Then perform a neighborhood-based connectivity search on R (such as starting from a high-displacement point and retrieving all adjacent points that also exceed 3 mm). If the search result shows a continuous band-shaped connected set, it indicates that an obvious crack band or shear dislocation area has been formed at this load stage. At the same time, the strain signal and vibration signal collected at different times can also be compared with this high-displacement area. If the strain sensor measures a large stress concentration at this place, or the vibration sensor records a peak of energy release in the same time period, it proves that this crack band is becoming the main channel of shear failure and it is necessary to be vigilant about the subsequent load evolution.

[0070] From a monitoring perspective, these high-displacement areas marked as crack dislocation regions are not only dangerous hidden trouble areas but also the priority target areas for subsequent grouting reinforcement or bolt strengthening. In engineering, if it is found that this high-displacement area continues to expand, it is very likely a sign that the rock mass is approaching instability, and it is necessary to modify the loading strategy in real time or strengthen the grouting and bolting materials to avoid sudden large-scale failures such as sudden breakthroughs. If the crack dislocation zone gradually stabilizes, it indicates that it has entered a relatively stable stage, and it can be considered to continue to advance the loading or maintain the original load-bearing state for subsequent tests.

[0071] In addition to the laboratory environment, the present invention can also be extended to underground engineering or slope monitoring: Periodically perform three-dimensional scanning on the rock wall at the geological exploration site. Once a large range of displacement differences exceed the threshold value during comparison, professional personnel should be quickly dispatched for site investigation or safety protection. This approach can significantly reduce the risk of sudden rockfalls or landslides.

[0072] Geometric and topological feature analysis is an evaluation method that combines theoretical rationality and operational feasibility by comparing the displacement differences between points in the time-sharing three-dimensional point cloud and marking the parts that exceed the threshold as crack dislocation regions. Its basic principle lies in three-dimensional geometric difference, threshold judgment, and the search for local connectivity, which can significantly improve the visualization and quantification level of the rock mass shear failure process, and form a complement with other monitoring methods in the present invention (such as strain signal and vibration signal acquisition), bringing more reliable data support for the research of joint failure mechanisms and actual reinforcement decisions.

[0073] Preferably, the topological feature analysis performs a void search based on the adjacency relationship on the marked area, judges the crack direction and connectivity, and generates a crack propagation path.

[0074] The topological feature analysis performs a void search based on the adjacency relationship on the marked area, judges the crack direction and connectivity, and generates a crack propagation path, which is a key means for conducting research on rock mass failure based on three-dimensional scanning. In this method, the point cloud areas that exceed the displacement or dislocation threshold have been marked in advance, indicating that obvious displacements or cracks have occurred at these positions under shear loads. Next, it is necessary to analyze the adjacency relationship between these high-displacement points to further divide the crack connection area and confirm the specific crack direction and possible expansion boundary. If it is detected that an internal void or a large-area discontinuous zone has formed between some adjacent high-displacement points, it also means that the crack has penetrated in this area or a relatively complete corridor-like failure area has been formed. Through this process, researchers can generate curves or network structures reflecting the crack direction, providing precise basis for subsequent grouting and bolting material layout or shear load control.

[0075] Three-dimensional point clouds are often stored in a grid or scatter point pattern. The grid pattern means that each data point has fixed adjacent units in space, while the scatter point pattern records the neighbor information of each point through data structures such as KD-trees or octrees to calculate the distance relationship between each point and its surrounding points. If, in a previous geometric analysis, all points exceeding the displacement threshold are marked as "high-displacement points", then these high-displacement points can be used as input in the topological analysis stage to compare whether they are directly or indirectly connected in the grid or scatter point topology. A common approach may be to start from a certain high-displacement point and search layer by layer along its neighbor points. If the neighbor point also meets the "high-displacement" criterion and the spatial distance between the two is lower than an adjacency threshold, it means that they are located in the same potential crack band. Spreading layer by layer along the adjacency relationship between points, the boundary of the entire high-displacement area can be identified. If this area forms an unclosed cavity or contains other intersection segments, it indicates that there are signs of crack penetration or multi-channel extension here.

[0076] In the void search based on the adjacency relationship, a "void" does not only refer to a geometric cavity, but can also represent the detachment between entities in a local grid surface. If a large number of continuous units have become disconnected at this location, it may mean that a gap has formed between rock blocks and the crack penetrates in this direction. When such voids intersect with other marked areas, the scale of the crack network is further expanded. Researchers usually visualize the results of these void searches, showing one or several continuous crack paths on the rock mass surface. In actual engineering or experiments, cracks often do not follow a single path, but are branched or network-connected. At this time, topological analysis can identify the intersection points between branches, helping to infer the main crack axis and the main trend.

[0077] To make this process more intuitive, an example can be given: Suppose in a shear test, after differential analysis of the three-dimensional point cloud at different times, it is found that the coordinates of about thousands of grid cells have changed beyond the threshold. After inputting these grid cells into the topological feature analysis program, the program first pairs their adjacency relationships to determine which ones are close enough in distance or have little difference in normal vectors and can be regarded as connected. If two significant connected components are finally obtained, each component has a relatively tight void structure or a penetrating edge inside. Based on the spatial distribution of these components, it can be judged that the crack is currently divided into two branches, one extending upward towards the rock mass and the other spreading towards a certain joint surface inside. Combining strain and vibration data can further confirm whether these two branches of cracks are active at the same time or whether they will connect into a larger fracture area under the action of the next-level load. The researcher can then make the analysis results into crack propagation path data, such as drawing a curve or a boundary band on a three-dimensional model to mark the crack segments.

[0078] Accurately determining the direction and connectivity of cracks is not only valuable at the research level, but also provides direct evidence for engineering decisions. If it is detected that the crack has passed through a critical surface, it may be necessary to subsequently implement anchoring, grouting, or change the loading strategy to avoid instability and loss of control. In addition, in some cases, if the topological feature analysis shows that the cracks are concentrated in a few isolated small blocks and there are no through-channels, it means that the rock mass as a whole has not yet suffered large-scale damage, and the load can be maintained or slightly increased to check its ultimate bearing capacity. It can be seen that the present invention makes shear failure monitoring more quantitative and visual: it can clearly show the spatial morphology of the cracks and the evolution order of the cracks.

[0079] The cavity search algorithm used in the present invention often draws on standard methods in graph theory or computer graphics, such as depth-first search, breadth-first search, or union-find. The core concept is to traverse all marked points, assign points that can reach each other within a local distance to the same component, and then determine whether a self-contained hole or channel is formed. Of course, for continuous media such as rock cracks, the change in the normal vector of the grid or scattered points must be considered in conjunction with geometric analysis. If a completely independent sub-surface is formed in certain areas, it may indicate that the rock mass has a macroscopic separation or a tendency for block detachment at that location. For geotechnical engineers, such separation areas often mean potential risks of rockfall or collapse, which require special attention.

[0080] In one example, consider a square specimen whose initial joint distribution is not prominent. However, after a certain shear force is applied, a new crack gradually penetrates from the side. After time-sharing 3D point cloud comparison, dozens of grid cells are identified as "exceeding threshold displacement points." When these are input into topological feature analysis, the program detects a unidirectionally extending void strip within these high-displacement cells, ultimately generating a thin crack propagation path, only a few centimeters from the outer edge. Combined with mechanical monitoring, it can be inferred that if the load is further increased, this crack is likely to extend to the rock surface, forming a complete shear slip plane. In this case, the load increment can be reduced, or grouting can be injected into the channel to prevent further crack propagation. Therefore, by combining adjacency-based void search with crack connectivity assessment, the present invention provides a highly operational and visual criterion for rock failure processes. Researchers can visually visualize the crack "network" in the 3D model and track its development, effectively applying it to the reinforcement, monitoring, and safety assessment of jointed rock masses.

[0081] Inputting the time-sharing three-dimensional point cloud, strain signal, and vibration signal into a deep learning model for multi-scale feature extraction to generate control parameters for grouting and shear load;

[0082] The time-sharing 3D point cloud, along with strain and vibration signals, is fed into a deep learning model for multi-scale feature extraction, generating control parameters for grouting and shear load. This approach aims to integrate crack information from a purely geometric perspective with strain and vibration changes from a mechanical perspective. A deep learning architecture is used to explore potential correlations between high-dimensional data, ultimately assisting in determining strategies for anchoring or adjusting shear loads in the rock mass. The core process can be divided into four parts: data preparation, model construction, feature extraction, and control parameter generation.

[0083] Part 1, data preparation:

[0084] In the previous steps, a time-sharing 3D point cloud, strain signals, and vibration signals are collected for each load stage. By placing these data in a unified time coordinate, a set of multimodal inputs can be formed:

[0085] Time-sharing 3D point cloud: describes the surface deformation distribution and may be stored in the form of a grid or scattered points. Each sampling moment corresponds to a 3D data file.

[0086] Strain signal: shows the degree of local stress concentration in the rock mass. It is usually collected using distributed optical fiber or strain gauges to obtain discrete time series curves or two-dimensional distribution data.

[0087] Vibration signal: If a crack expands rapidly under load, the vibration energy will fluctuate significantly. By recording the vibration time series at an appropriate sampling frequency, its amplitude and spectral characteristics can be analyzed.

[0088] In the present invention, it is necessary to pay attention to the unification of time domain and space coordinates. The three-dimensional point cloud may be collected at a lower frequency (for example, a full scan is obtained at every specified load increment), while the strain and vibration signals may be collected at a higher frequency (such as milliseconds). Before inputting the data into the deep learning model, it is common practice to voxelize or mesh the three-dimensional point cloud first, and slice the strain and vibration data according to a time period close to the point cloud acquisition time to form a set of mutually aligned multimodal samples. This ensures that the information received by the model is a comprehensive situation around the same load moment.

[0089] Part 2, model structure:

[0090] The deep learning model is responsible for multi-scale feature extraction of 3D point clouds, strain curves, and vibration curves. Its core usually consists of two types of units: convolution units and fusion units:

[0091] Convolutional unit: 3D point clouds can be processed using techniques such as 3D convolution or graph convolution. For strain and vibration signals, 1D convolution or 2D convolution after time-frequency transformation can be used. Regardless of the architecture, the emphasis is on extracting features layer by layer, from the local to the global level.

[0092] Fusion unit: At a certain level or feature dimension, fuse the information from three-dimensional deformation (geometric information) and strain, vibration (mechanical-dynamic information); if the attention mechanism is used, the model can also focus on the high-deformation areas and high-energy release areas, improving the determination accuracy.

[0093] For example, a graph convolutional network (Graph Convolution Network) can be used for the three-dimensional point cloud part, and a one-dimensional convolutional network (1D CNN) can be used for the strain and vibration data part to capture waveform features. After the two feature tensors are merged in the fully connected layer or the attention fusion layer, a comprehensive vector is output, which contains the numerical description of key information such as the current stress state of the rock mass and the crack expandability.

[0094] The third part, feature extraction:

[0095] Multi-scale feature extraction is mainly reflected in the following aspects:

[0096] Spatial scale: The three-dimensional point cloud part may perform downsampling or pooling on geometric deformations at different resolutions to capture large-scale overall displacements and small-scale detailed cracks.

[0097] Time scale: If there are multi-temporal point clouds (from different loading stages), the model can fuse the deformation change trends at different stages at a high level and compare them with the short-term or long-term fluctuation characteristics of the strain and vibration signals.

[0098] Frequency scale: The vibration signal may undergo short-time Fourier transform or wavelet transform to extract the energy distribution at different frequency bands, which helps to identify the sudden vibration peaks of cracks in certain frequency bands and then cross-compare with the three-dimensional deformation information.

[0099] Through this multi-scale integration, the model can further determine the activity of the current crack extension in the rock mass and the possible failure risks under similar loading levels or small load increments in the future.

[0100] The fourth part, generation of regulation parameters:

[0101] The deep learning model finally outputs a set of regulation parameters for grouting and shear loads, which are used to guide the next step of processing:

[0102] Grouting regulation parameters: Determine where and with what pressure to perform grouting to inhibit crack propagation. For example, a spatial coordinate distribution map can be output, indicating the positions where cracks are most likely to penetrate or high-risk areas, and then numerical suggestions for grouting pressure are given (such as taking a higher pressure near the crack zone and maintaining a normal pressure in the surrounding areas).

[0103] Load control parameter: Determine whether to continue increasing the load, maintain a constant load, or reduce the next-level load increment based on the current crack growth rate and rock mass stress concentration. If the model identifies that the crack expansion trend is very obvious, it may output instructions such as "cannot continue to increase the load" or "reduce the load"; if it is identified that the rock mass still has a certain safety margin, it is prompted that the load can be steadily increased to obtain subsequent data.

[0104] Preferably, the deep learning model uses a multi-layer convolutional structure to perform coordinate correction on the time-sharing three-dimensional point cloud, performs feature encoding on the strain signal and vibration signal, and then performs a fusion operation with the time-sharing three-dimensional point cloud.

[0105] In the previous steps, whenever the target rock mass completes a load level, new time-sharing three-dimensional point clouds are collected. For ease of calculation in the neural network, coordinate correction needs to be done first, that is, align the time-sharing point cloud with the reference coordinate system to eliminate possible offsets and rotational inconsistencies. This coordinate correction process usually involves a pose transformation, which involves solving the rotation matrix and displacement vector. After completing the coordinate correction, map the three-dimensional point cloud into a suitable network input format. Common practices include: converting the point cloud into a regular voxel grid, where each voxel stores geometric occupancy information; directly using point cloud convolution techniques (such as PointNet or Graph Convolution) to process the scattered points; mapping the surface mesh topology to construct an adjacency list, and then using a graph convolutional network.

[0106] Regardless of which representation method is used, the model will extract local structures (such as sudden normal vectors or high curvatures near cracks) in the shallow convolutional kernels, and combine broader contexts (such as the connectivity relationships of multiple crack segments) in the deep convolutional kernels. After gradually fusing these features, one or more feature vectors for the geometric distribution on the rock mass surface can be obtained, which often include the crack degree, surface convexity and concavity, etc.

[0107] For the strain signal and vibration signal, it is necessary to first analyze their time series or frequency domain characteristics to align with the geometric features extracted by the three-dimensional point cloud convolution. In the specific process, the strain or vibration sequence can be input into a one-dimensional convolutional network (1D-CNN), or first perform a short-time Fourier transform (Spectrogram) and then use two-dimensional convolution (2D-CNN).

[0108] If 1D convolution is used: Slide the convolutional kernel on the continuous time axis to extract the local patterns of the strain or vibration signal, and compress the signal in the subsequent pooling layer. If the spectral components of the vibration signal are considered: Fourier transform can be performed after segmentation to obtain the amplitude spectrum or time-frequency spectrum, and then the features of the spatial-frequency distribution can be obtained in the form of two-dimensional convolution.

[0109] Regardless of the method adopted, one or more vector representations of the hidden layer will ultimately be obtained, characterizing the performance of the rock mass at this loading moment in terms of mechanical stress concentration, energy release, etc. Different from the three-dimensional point cloud features, this part mainly comes from the internal strain measured by the sensor and the microseismic or vibration information, and can better reflect the intensity of crack activity.

[0110] Fuse the "three-dimensional point cloud features after multi-layer convolution processing" and the "strain-vibration features after multi-layer convolution processing" at the high level of the network. A common implementation method is to splice or add them weighted, and then perform deeper operations through fully connected or attention mechanisms. For example, in the attention mechanism, the model can be made to automatically focus on the high-displacement areas of the three-dimensional point cloud and the moments corresponding to the strain peaks to determine whether crack penetration occurs.

[0111] If the model detects multiple crack nodes marked in the geometric feature map and there is an obvious overlap with the strain peak curve, it may output a high-risk assessment value; if the high-level fusion identifies that although the displacement on the rock mass surface is significant, but the strain and vibration signals only change smoothly, it indicates that the crack may not be unstable or not connected to the deep part.

[0112] The above fusion process will output one or more key parameters at the end of the network for subsequent grouting and load regulation. For example: crack propagation prediction value, crack direction probability distribution, grouting recommended pressure, next load increment, etc. Through this multi-layer convolution structure, the model gets rid of the dependence on single geometric or single mechanical data, but comprehensively grasps various signs of the rock mass in the shear failure evolution.

[0113] Relying solely on 3D point clouds lacks direct observation of the internal energy release of rock masses; relying only on strain and vibration signals also makes it difficult to clarify the spatial geometry trend of surface cracks. Through fusion operations, researchers can not only know "where crack dislocations occur", but also understand "whether the mechanical strain reaches the critical value during dislocation and whether the vibration waveform has high-frequency energy aggregation", and thus more accurately judge the crack failure rate and scale. The multi-layer convolution structure usually has the advantage of hierarchical perception of "first local, then global". The primary convolution layer captures micro-cracks or micro-deformations, and the middle and high-level convolution layers connect multiple micro-cracks into a possible through-crack, and finally generate high-level features representing the overall failure situation. Strain and vibration signals can also deepen the understanding of wave information layer by layer, similar to identifying when vibration peaks occur and which frequency band spectrum is more prominent. Finally, in the fusion stage, the spatial deformation and time signals are associated. This multi-layer convolution structure is not only applicable to single-type point cloud or waveform data. If other sensors are deployed in rock masses in the future, such as temperature field monitoring or laser Doppler velocimetry, corresponding convolution branches can also be added and feature fusion can be carried out using a similar idea. The architecture of the network is relatively flexible and suitable for further expansion of the present invention in large-scale rock mass field tests or higher-precision laboratory studies.

[0114] In one embodiment, in a certain laboratory application, the R & D team conducted a step-by-step shear test on a mixed rock mass with obvious joint surfaces. For every 5 kN increase in shear force, a three-dimensional scan was performed on the target rock mass, and fiber optic strain gauge data and microseismic sensor waveforms were collected synchronously. The collected point clouds were processed through a graph convolution network (Graph Convolution Network), while the strain and vibration signals were respectively input into a 1D convolution network (1D-CNN) to extract their main frequencies and energy characteristics. Finally, the two were integrated into a feature vector Z in the high-level fusion layer. According to the mechanism of pre-training or online learning, the model will output suggestions for the next control strategy. For example, if the crack activity is significantly increased in Z, the model gives hints such as "decrease the load increment" or "concentrate the grouting position at the crack connection"; if Z shows that the rock mass is still in a stable stage, the original planned load increase is maintained.

[0115] After several tests, it was found that when a large crack connection area appears on the surface of the 3D point cloud and the strain curve continues to rise during the load gap, the model successfully gives the instruction of "suggest decreasing the load increment" and increasing the grouting pressure, avoiding the rapid instability of the rock mass under subsequent high loads. The researcher thus verified the practicability of the multi-layer convolution fusion scheme - it can both "see" the spatial deformation of cracks and "perceive" the internal stress dynamics, playing a core supporting role in real-time or quasi-real-time decision-making.

[0116] Preferably, the area with noise or missing points in the time-sharing 3D point cloud is first subjected to a local density-based correction operation and then input into the multi-layer convolution structure.

[0117] First, perform a correction operation based on local density on the area with noise or missing points in the time - shared three - dimensional point cloud, and then input it into the multi - layer convolution structure. This is a key link to ensure the quality of three - dimensional data and thus improve the recognition accuracy of subsequent deep learning. In the application background of the present invention for monitoring the shear failure evolution of grouted joints, whenever a load stage is completed in the test or engineering site, a three - dimensional scan of the target rock mass is performed to obtain a time - shared three - dimensional point cloud. Due to factors such as the on - site acquisition environment, instrument noise, and irregular rock mass surface, there will inevitably be situations where the local density is too low (missing points) or there are isolated points (noise points) in the point cloud. If not corrected and directly input such a point cloud into the multi - layer convolution structure, it will often lead to distorted feature judgment in the crack recognition and subsequent regulation parameter prediction links. Therefore, the present invention adopts a local density correction strategy. Before sending the point cloud data into the multi - layer convolution structure, first perform interpolation or filtering on the suspicious area to make up for the loopholes in three - dimensional sampling.

[0118] At the principle level, local density correction can be regarded as a compensation process for the irregular distribution of three - dimensional data. For an area initially judged to have "missing points or noise points", there are generally the following possibilities:

[0119] There are occlusions or scanning blind spots during acquisition: The projector or camera fails to fully cover the rock mass surface, resulting in too few points in this area in the point cloud;

[0120] Specular reflection or surface interference: If the rock mass surface has moisture, high gloss, or contains other interfering substances, it is easy to generate isolated points or interfering patterns;

[0121] Multi - view stitching error: When registering or overlapping the time - shared three - dimensional point cloud, if the reference calibration is not precise enough, there may also be a small amount of drift or ghost points at certain positions.

[0122] To better correct the above - mentioned interferences, the "local - density - based" method can be adopted, that is, calculate the statistical distribution characteristics between adjacent points within a certain area. If the neighborhood density (which can be defined as the number of points within a certain radius δ) around a certain point is much lower than the global average level, it indicates that this point may be a noise point or a missing point; if there are completely no sampling points at a certain place, it indicates that this is a potential void area that needs to be interpolated and filled. When interpolating, the coordinate distribution and normal vector information of the neighboring area in three - dimensional space can be referred to, and the void is filled by approximating it with a local plane or surface. For example, in actual operation, researchers can preset a radius δ≈2 mm to measure whether the number of points in the surrounding neighborhood is lower than a minimum threshold T min , if it is lower than this threshold, it is determined as a missing area, and then it is supplemented by interpolation or grid smoothing.

[0123] This interpolation correction is particularly important for crack monitoring. Because if the missing points happen to fall on the crack path, the point cloud may wrongly regard this area as a "void", thus misjudging it as cracked; conversely, if there are multiple isolated noise points concentrated in a certain position, it may also form a false crack signal. After the correction operation, the point distribution of the time-sharing three-dimensional point cloud is relatively uniform in the overall space, which can significantly reduce the interference brought to the determination of crack connectivity. When specifically implemented, the algorithm for local density correction can be divided into the following steps:

[0124] Neighborhood search: For each suspicious point or suspicious grid cell, use the KD tree or Grid search to find all adjacent points within a radius of δ;

[0125] Density determination: If the number of adjacent points is less than the threshold, it is determined as a low-density area; if the number of adjacent points is much greater than the threshold and the normal vector difference of the corresponding points is very large, it is suspected to be an isolated noise point;

[0126] Correction strategy: Perform interpolation on the low-density area. For example, find several reference points around it, infer an approximate surface equation, and then add several filling points on this surface equation; eliminate or smooth the obvious isolated noise points.

[0127] Once this correction operation is completed, the processed point cloud can be input into the multi-layer convolution structure. Since the deep learning model is extremely dependent on data consistency and effectiveness, if there are large areas of voids or noise in the data, it will be difficult for the network to accurately determine the geometric distribution mutation of the crack during the convolution process, and it may also regard the noise points as real crack nodes. In the monitoring of the shear failure of grouted joints in the present invention, the time-sharing three-dimensional point cloud will be obtained multiple times at each level of load or each time period. Once a large number of noise point clouds appear without correction, it will interfere with the identification of the crack evolution path and even affect the final decision-making of grouting and load control.

[0128] In practice, there are successful implementation examples to prove the benefits of this process. For example, in the laboratory, a sandstone sample containing multiple hidden cracks is loaded and scanned multiple times. The point cloud at a certain load stage collected has the problem of coexistence of sparsity and noise at the side edge of the rock. After local density correction, the voids in the edge area are significantly reduced and the noise points are removed cleanly. After inputting it into the multi-layer convolution network, the result that the model finally determines that the crack penetrates from the lower edge of the middle part of the sample is greatly strengthened, which is consistent with the actual fracture morphology of the later load test. If this correction step is not done, the crack connectivity judgment output by the network will either be over-fractured (regarding voids as cracks) or be confused at the noise points, and sometimes there will be results with obvious deviations from the real crack distribution.

[0129] In the subsequent deep learning process, such smooth and reasonable point cloud data can ensure that multi-layer convolutions accurately identify crack tips or high-displacement areas, and can also more robustly perform feature alignment with strain and vibration signals. For example, when the vibration signal collected at a certain load level shows a waveform energy peak, if there are holes in the point cloud at this time, it will cause the model to be unable to accurately map to the location of the real crack band; but after correction, any area corresponding to large deformations can be better filled with points and smoothed, so the network can accurately associate the "high-deformation band" with the "high-vibration energy area".

[0130] Adjust the grouting pressure or loading method according to the control parameters, and visually integrate the execution process with the time-sharing three-dimensional point cloud, strain signal, and vibration signal through a three-dimensional deformation map and a crack distribution map to obtain a monitoring report on the shear failure evolution of the grouted joint.

[0131] Adjusting the grouting pressure or loading method according to the control parameters, and visually integrating the execution process with the time-sharing three-dimensional point cloud, strain signal, and vibration signal through a three-dimensional deformation map and a crack distribution map to obtain a monitoring report on the shear failure evolution of the grouted joint is the key step in transforming the analysis results into actual operations and visual results in the present invention. In the previous process, a deep learning model or other analysis methods have already fused and extracted control parameters based on the time-sharing three-dimensional point cloud, strain signal, and vibration signal. These parameters usually include suggestions for the grouting area and grouting pressure, adjustment instructions for the next-stage load level or increment, etc. Once these control parameters are obtained, the scheme can enter the adjustment and visual integration link to guide on-site tests or engineering construction operations, and at the same time generate a report file for research or safety assessment. The core here is to "implement theoretical control suggestions into specific operations" and present them in a graphical way to help researchers or engineering personnel quickly understand and trace back the overall process of shear failure.

[0132] In shear failure monitoring, grouting and loading are two key controllable factors:

[0133] Adjustment of grouting pressure: When analysis finds that a crack somewhere is about to connect or is expanding severely, appropriately increasing the grouting pressure can form a tighter filling or blockage within the joint, slow down the crack development, and increase the overall strength of the rock mass; if the crack expansion is still within an acceptable range, a slightly lower or stable grouting pressure can also be used to save materials but maintain basic reinforcement.

[0134] Adjustment of loading method: If the model determines that the current bearing capacity of the rock mass is approaching the limit, it may suggest reducing the next load increment or maintaining a constant load at the current load for an extended observation time; if it is determined that there is still a large safety margin, it is allowed to continue increasing the load or switch to other forms (such as a slower loading rate) to obtain more experimental data on the failure mechanism of the rock mass.

[0135] During actual execution, experimenters or on-site engineers will accordingly adjust parameters to make corresponding settings for the grouting system, which usually includes adjusting the pressure threshold of the grouting pump, changing the slurry concentration or dispersing materials; for the load, the speed of the shear testing machine or the target load level in the next stage will be adjusted accordingly. If the sensing system is still in the continuous acquisition state at this time, the impact of the adjustment measures on crack development in the next scanning period can be immediately observed, thus forming a closed-loop adjustment.

[0136] During each execution process, such as increasing the grouting pressure by 10%, extending the grouting time, or modifying the load increment, this operation must be recorded and matched with the time-sharing three-dimensional point cloud, strain, and vibration data, and aligned from the time node for subsequent visualization and report compilation. For example:

[0137] If grouting starts after the 5th stage of load and lasts for 3 minutes, then a three-dimensional scan is performed before the start of the 6th stage of load, and the effect of grouting on crack closure can be compared.

[0138] If the loading method changes from uniform force application to staged holding, denser point cloud, strain, and vibration data can be collected during the holding stage to observe whether the crack is stable.

[0139] At the data level, marks need to be made for each adjustment moment. For example, the execution process is recorded as an operation log "(timestamp, adjustment method, adjustment amplitude)", which corresponds to the sampling time sequence of the sensing data. In this way, an identifier at a critical moment can be inserted into the later three-dimensional deformation map and crack distribution map, helping readers or operators quickly understand "when what operation is performed on grouting or load and what crack changes are caused thereby".

[0140] In order for researchers to clearly see how cracks expand in space and which parts are blocked or the dislocation is relieved after grouting, a series of three-dimensional deformation maps need to be generated based on the time-sharing three-dimensional point cloud collected each time, as well as a crack distribution map indicating the crack connection path.

[0141] Three-dimensional deformation map: Usually, the reference point cloud is differentiated from the current moment point cloud, and then the displacement magnitude or normal vector change of each point is rendered in a chromatographic manner. For areas with high displacement, red or warm colors are used to represent, and cold colors are used to mark low displacement areas. This can intuitively show the most severe parts of crack activity or local relative dislocation of the rock mass.

[0142] Crack distribution map: The high-displacement connected areas or void search results obtained through geometric and topological feature analysis can be outlined with lines or polygons on the surface of the three-dimensional model; if the crack advances with the load stage, a series of crack trajectories can be drawn in the time series, enabling users to understand the evolution process of the crack from nothing to something and from short to long.

[0143] When the strain signal and the vibration signal are also superimposed and presented, the positions and their numerical ranges of each sensor can be marked on the visualization interface. For example, at the end of each load level, it can be seen that the point cloud displacement near the position of the strain gauge is relatively high. If there is a burst of high-frequency components in the vibration signal, add a time scale or a dynamic animation to the 3D map during this period to display the crack jump corresponding to the vibration peak in real time. In this way, the comparison and linkage analysis of multi-source data can be completed in the same visualization scenario.

[0144] Based on these visualization results and the records of the execution process, a monitoring report is finally formed. The report usually includes the following content: the time history of the load and grouting strategy: at what moment grouting was carried out or the load mode was changed, and what their respective amplitudes were. 3D deformation and crack evolution diagrams: showing the trend, severity or penetration of cracks in each acquisition stage in the form of charts or animations. Comparison of strain and vibration data: usually time series curves and energy spectra, and the crack activities corresponding to the key peaks are marked in contrast to the 3D deformation diagram or crack distribution diagram. Analysis and conclusion: convert the visualization results into text analysis, for example, point out that "the crack propagation rate increased at the sixth-level load, but the crack trend stopped spreading after the grouting pressure increased", or "when the load exceeded a certain value, large-scale cracking occurred in the rock mass", forming a systematic demonstration of the law of anchor grouting joint shear failure of the present invention.

[0145] In practice, researchers or engineers can use this report to verify experimental hypotheses, such as "can grouting effectively inhibit damage when the crack is near penetration", and can also use this to determine whether to strengthen anchor grouting, reduce load or maintain the current state on the engineering site. Thus, the monitoring scheme of the whole set of 3D scanning and multi-modal data fusion is closed-loop implemented, and the whole process management of "monitoring - analysis - regulation - verification" of joint failure is truly realized.

[0146] Example: Cracks began to spread in the rock mass side wall of a roadway in an underground mine under continuous excavation and stress environment. Technicians first obtained the point cloud by time-sharing 3D scanning, and then collected multi-modal data in cooperation with strain and vibration sensors; the system model output an instruction of "increasing the grouting pressure and slightly slowing down the load increment in the next stage". The project party carried out grouting plugging according to this instruction and reduced the load increment in the next stage from 10 kN to 5 kN. After the execution was completed, 3D scanning and sensing data collection were carried out again. The 3D deformation map showed that the originally rapidly spreading crack band was indeed controlled, and the number of vibration signal peaks decreased. The comparison between these two stages was presented in the form of 3D rendering in the report, proving the practical significance of the present invention for hazard diffusion.

[0147] Preferably, after comparing the regulation parameters output by the deep learning model with the crack propagation rate, the adjustment step of the grouting pressure or the change interval of the load mode is determined, and the time-sharing 3D point cloud is collected after grouting to determine the crack closure degree.

[0148] Compare the regulation parameters output by the deep learning model with the crack propagation rate to determine the adjustment step of the grouting pressure or the change interval of the loading method. After grouting is completed, collect the time-sharing three-dimensional point cloud to determine the crack closure degree, which is the key link to achieve crack suppression and enhanced monitoring in the present invention. Based on the previously collected time-sharing three-dimensional point cloud, strain signal and vibration signal, the deep learning model will output a set of numerical values or instructions, including the matching degree between the crack growth trend and the reinforcement measures. In order to better grasp the dynamic changes during the shear failure process of joints, it is necessary to compare these regulation parameters given by the model with the actually observed crack propagation rate, so as to achieve more accurate grouting and load management.

[0149] During this process, the crack propagation rate is generally measured by the difference result of the time-sharing three-dimensional point cloud before and after. Assuming that the width w of a certain crack area is defined and the sampling interval Δt of adjacent loading stages is defined, the increment Δw of the crack within a period of time can be calculated as Δw = w t - w0, and then the propagation rate is defined If the crack propagation rate is significantly higher than the reference threshold and coincides with the risk coefficient output by the deep learning model, it indicates that the crack is in an active period, and it is necessary to consider immediately increasing the grouting pressure or reducing the load increment to avoid larger-scale damage. On the contrary, if the current crack propagation rate is within an acceptable range, the grouting pressure can be appropriately relaxed or the load can be continuously increased according to the original plan. By cross-verifying the model parameters and the actual crack propagation rate, an adaptive regulation mechanism is formed, which can make a refined response to the activity degree of cracks in the rock mass.

[0150] Specifically, in order to determine the adjustment step of the grouting pressure, it is necessary to combine the geometric distribution of crack propagation, the stress concentration situation and the environment where the rock mass is located. If the deep learning model suggests increasing the grouting intensity and the measured crack rate is also rising, the grouting pressure will be increased by a certain amount compared with before. It can be assumed that the original grouting pressure is P0, the recommended coefficient output by the deep learning model is α, and the crack propagation rate reaches or exceeds a certain safety threshold In this case, a new grouting pressure can be set as P1 = P0 + ΔP, where ΔP is jointly determined according to the comparison result of α and the crack rate. The value of ΔP depends on the scale of crack network expansion and the evaluation of rock mass bearing capacity. If the crack rate is only slightly higher than the threshold, ΔP remains a small value; if the crack rate seriously exceeds the standard, ΔP can be increased to a relatively high level in order to form an effective barrier inside the crack as soon as possible. At the same time, when managing the loading mode, the next shear load increment or the duration of the holding stage will be controlled according to the load regulation parameters output by the deep learning model. Suppose the original load increment per time is 5 kN. Now it is found that the crack rate has approached the critical value, then the next load increment may be reduced to 2 kN, and the duration of constant load observation is extended to confirm whether the crack propagation has been inhibited.

[0151] In the present invention, for the high joint and composite crack network that may sometimes be encountered, three-dimensional point cloud acquisition at different times can be carried out immediately after grouting to judge whether the degree of crack closure reaches the expected goal. If, after grouting reinforcement, the coordinate difference measured in the original high-displacement area in the new round of scanning has been significantly reduced, it indicates that the grouting material has a good reinforcement effect on the crack, and then the loading can be continued or the original state can be maintained according to the load suggestion of the deep learning model. If there are still a large number of micro-cracks expanding in the surrounding area, the grouting strategy needs to be adjusted again, further increasing the pressure or changing to a different slurry formula.

[0152] Through this mechanism, the combination of joint shear monitoring and reinforcement process based on three-dimensional scanning can be realized: after each loading stage ends, the three-dimensional point cloud records the surface changes of the rock mass, and the strain and vibration signals reflect the internal stress and dynamic state; the deep learning model synthesizes multi-modal information to generate regulation parameters, and compares them with the measured results of the crack propagation rate, so that the judgment of the model can be timely feedback by on-site data, avoiding the possible deviation relying solely on the model output. Especially in an environment with a large engineering scale, comparing the crack propagation rate in real time or at intervals can enable researchers to quickly detect potential instability clues, so as to avoid blindly following the model results and leading to out-of-control failure risks.

[0153] In an embodiment of an indoor geotechnical experiment, the researchers concretized this idea: First, a multi-level shear test was carried out on a rock mass containing natural joints, and three-dimensional point clouds and mechanical signals were collected every 5 kN load. The deep learning model gave real-time suggestions for increasing the grouting pressure, and at the same time recorded indicators such as the average crack opening and propagation rate. In the first few load increment levels, the crack rate was still lower than the set warning value The grouting pressure only increased slightly. However, after entering the 4th level of load, the crack rate climbed instantaneously, and the model coefficient also increased significantly. The researchers immediately increased the grouting pressure by 30% and shortened the interval for the next load increment. At this time, it was found in the next-stage scan that the crack width did not continue to expand, indicating that grouting played an inhibitory role in the dislocation, creating conditions for the safe continuation of the test. If the cracks could not be inhibited during this process, the plan would continue to be analyzed automatically or manually until a control plan that could restore the crack rate to an acceptable range was found.

[0154] By collecting the time-sharing three-dimensional point cloud again after grouting and analyzing the crack closure degree, the present invention provides intuitive and reliable data support for studying the actual effect of joint grouting during the shear failure process. Once it is confirmed that the cracks are significantly closed, it means that an effective cementation or barrier layer has been generated by the grouting material in the crack zone, and the probability of continued expansion in this area during subsequent loading will be greatly reduced. If only some parts of the cracks are closed but other new cracks are derived, it indicates that some other blocks recommended by the deep learning model need to be reinforced in the same way. In short, by combining the control parameters with the crack propagation rate, implementing grouting and load management in field operations, and verifying the effect with three-dimensional point clouds, the grouting effect of rock masses and the shear load stability can be continuously optimized in the feedback loop of "observation - decision - execution - re-observation", and the present invention can also achieve a reliable, flexible and scientific failure monitoring and prevention mode in rock mechanics research and engineering protection.

[0155] Preferably, the grouting process and the crack propagation trajectory are superimposed and output through a three-dimensional deformation map and a crack distribution map, and the time-series data of the strain signal and the vibration signal are included in the monitoring report to show the corresponding relationship between loading and crack evolution.

[0156] This visualization integration link relies on the time-sharing three-dimensional point clouds and mechanical sensing information obtained in multiple rounds. Whenever a new three-dimensional scan is performed, a three-dimensional deformation map can be generated. This map can calculate the difference between the reference point cloud or the point cloud in the previous time period and the current point cloud, and then present the displacement magnitude of the specific position on the rock mass surface in the form of color mapping or numerical annotation. At the same time, the crack distribution map pays more attention to the main crack zones, branches, and void areas obtained after geometric and topological analysis of the cracks. The common practice is to mark them on the three-dimensional model surface with lines or semi-transparent color blocks, making it clear at a glance how the cracks move and spread at different stages.

[0157] For the "grouting process", if grouting is applied to the crack in the middle or at different stages of the load, it will often have a certain impact on the crack width or depth in a relatively short period of time. Sometimes the crack is partially blocked and the surface displacement is significantly reduced; it may also be that the grouting is only effective near the surface, while the deep cracks continue to expand. Superimposing the grouting process on the three-dimensional deformation map and the crack distribution map means showing at a visual level the time when the grouting started, the grouting pressure, how long the grouting lasted, etc., and comparing it with the crack changes obtained in subsequent scans. If the crack gradually closes with grouting, the three-dimensional deformation value in the corresponding area will also decrease, and the high-displacement connecting zone in the crack distribution map will become smaller. If the grouting does not achieve the desired effect, and the crack direction is still seen to be continuous or the topological voids are still increasing, it indicates that the grouting plan or loading strategy needs to be adjusted.

[0158] Including the time series data of strain signals and vibration signals in the monitoring report can fill in the possible blind spots of pure geometric observation: even if the deformation of the crack surface is temporarily slowed down, the internal stress may still be concentrated; or the surface seems stable, but the vibration signal increases significantly in certain frequency bands, indicating that a large number of microcracks have been initiated. By inserting the time series curves of strain and vibration in the report, the three-dimensional deformation state corresponding to each key moment can also be marked. For example, marking multiple key nodes (such as the start of grouting, the end of grouting, the load upgrade to a certain level, etc.) on the curve to correspond to the legend number in the three-dimensional deformation map or crack distribution map will help readers establish a clear mapping between space and time. In this way, once you see a sharp peak on the vibration map, you can immediately find where the large displacement occurred in the three-dimensional deformation map at that time, so as to understand the location and scale of crack expansion or material damage.

[0159] like Figure 2 As shown, a system for monitoring the shear failure evolution of anchor-grouting joints based on three-dimensional scanning is used to implement the above-mentioned method for monitoring the shear failure evolution of anchor-grouting joints based on three-dimensional scanning, and the system includes:

[0160] The 3D scanning data acquisition module is used to observe the target rock mass from multiple angles and then stitch and de-noise the observed data to generate an initial 3D point cloud. This module primarily consists of multi-view imaging hardware and stitching software, typically including a structured light projector or laser scanner, and a camera array for multi-angle acquisition of rock mass surface data. This module captures the target rock mass from multiple viewpoints on-site, acquiring several sets of images or point clouds. These images are then combined using a stitching algorithm to form a single 3D point cloud encompassing all surface features. Clutter and noise are then filtered or interpolated to produce a relatively complete and smooth initial 3D point cloud, facilitating subsequent deformation analysis.

[0161] Shearing load and signal acquisition module, which is used to apply shearing load to the target rock mass and acquire time-sharing three-dimensional point cloud, strain signal and vibration signal at each load stage; the shearing load and signal acquisition module generally includes a shearing testing machine or on-site compression-shearing equipment that can control the load increment and loading rate, and is used to apply graded or continuous shearing force to the target rock mass. Synchronously, strain gauges (such as fiber Bragg gratings) and vibration sensing devices (such as high-frequency accelerometers) are connected to the hardware to record strain and vibration data. Whenever the load enters the next stage or maintains for a period of time, this module will trigger three-dimensional scanning again to obtain time-sharing three-dimensional point cloud, so as to form time series data for real-time or post-event analysis.

[0162] Multi-scale feature extraction module, which is used to encode the features of time-sharing three-dimensional point cloud, strain signal and vibration signal and generate regulation parameters for grouting and shearing load; the multi-scale feature extraction module performs deep learning or convolution operations through high-computing power hardware (such as a workstation equipped with GPU or an embedded processor), extracts geometric features from the time-sharing three-dimensional point cloud, and encodes the time domain or frequency domain features of the strain and vibration signals. This module synthesizes this information and outputs a set of regulation parameters for the current state of the rock mass, including whether to increase or decrease the grouting pressure, or what loading scheme to adopt in the next load stage to inhibit crack propagation or optimize the experimental process.

[0163] Regulation and visualization output module, which is used to adjust the grouting pressure or loading mode according to the regulation parameters, and display the execution process together with the time-sharing three-dimensional point cloud, strain signal and vibration signal in the form of three-dimensional deformation map and crack distribution map, and output a monitoring report for the shear failure evolution of grouted joints. The regulation and visualization output module may include a grouting pump and valve control device, as well as a load adjustment interface linked with the testing machine in terms of hardware to execute the aforementioned regulation parameters. Meanwhile, this module will map the time-sharing three-dimensional point cloud and sensor data to a 3D visualization system to generate a three-dimensional deformation map and a crack distribution map. The operator can view the crack trend, deformation amplitude and strain-vibration curve on the on-site display screen or the main control computer, and finally export a monitoring report for the shear failure evolution of grouted joints for engineering decision-making or academic research.

[0164] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects.

[0165] The above are only the embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A monitoring method for the shear failure evolution of grouted joints based on 3D scanning, characterized in that It includes the following steps: After splicing and denoising the perspective data obtained from multi-angle observations of the target rock mass, an initial three-dimensional point cloud is generated; Apply a shear load to the target rock mass and collect time-sharing three-dimensional point clouds at each load stage. Register or differentiate with respect to the initial three-dimensional point cloud. Identify the surface crack range through geometric and topological feature analysis and record the shear failure evolution information. Synchronously collect strain signals and vibration signals at each load stage; Input the time-sharing three-dimensional point clouds, strain signals, and vibration signals into a deep learning model for multi-scale feature extraction to generate regulation parameters for grouting and shear loads; Adjust the grouting pressure or loading method according to the regulation parameters, and visually integrate the execution process with the time-sharing three-dimensional point clouds, strain signals, and vibration signals through three-dimensional deformation maps and crack distribution maps to obtain a monitoring report on the shear failure evolution of grouted joints; 2. The monitoring method for the shear failure evolution of grouted joints based on 3D scanning according to claim 1, wherein When applying a shear load to the target rock mass, use a hierarchical incremental method. After each level of load reaches the set value, maintain a constant load period and collect time-sharing three-dimensional point clouds for comparing crack changes under different load conditions; 3. The method for monitoring the shear failure evolution of grouted joints based on 3D scanning according to claim 2, wherein When the crack propagation rate exceeds a pre-set threshold, reduce the increment of the next load level and simultaneously collect strain signals and vibration signals to improve the acquisition density of time-sharing data; 4. The monitoring method for the shear failure evolution of the grouted joint based on 3D scanning according to claim 1, characterized in that Geometric and topological feature analysis compares the displacement differences between points of the time-sharing three-dimensional point cloud, and marks the part exceeding the threshold as the crack dislocation area; 5. The monitoring method for the shear failure evolution of grouted joints based on 3D scanning according to claim 4, characterized in that, Topological feature analysis performs a void search based on the adjacency relationship on the marked area to judge the crack direction and connectivity and generate a crack propagation path; 6. The method for monitoring the shear failure evolution of grouted joints based on 3D scanning according to claim 1, characterized in that, The deep learning model uses a multi-layer convolutional structure to perform coordinate correction on the time-sharing three-dimensional point cloud, performs feature encoding on the strain signal and vibration signal, and then performs a fusion operation with the time-sharing three-dimensional point cloud; 7. The method for monitoring the shear failure evolution of grouted joints based on 3D scanning according to claim 6, wherein First perform a local density-based correction operation on the area of the time-sharing three-dimensional point cloud with noise or missing points, and then input it into the multi-layer convolutional structure; 8. The method for monitoring the shear failure evolution of grouted joints based on 3D scanning according to claim 1, characterized in that Compare the regulation parameters output by the deep learning model with the crack propagation rate to determine the adjustment step of the grouting pressure or the change interval of the loading method, and collect the time-sharing three-dimensional point cloud after grouting to determine the crack closure degree; 9. The method for monitoring the shear failure evolution of grouted joints based on 3D scanning according to claim 8, characterized in that Overlay the grouting process and the crack propagation trajectory through three-dimensional deformation maps and crack distribution maps, and list the time-series data of the strain signal and vibration signal in the monitoring report to show the corresponding relationship between loading and crack evolution; 10. A monitoring system for the shear failure evolution of grouted joints based on 3D scanning, which is used to implement the monitoring method for the shear failure evolution of grouted joints based on 3D scanning according to any one of claims 1-9, characterized in that, Its system includes: A three-dimensional scanning data acquisition module for multi-angle observation of the target rock mass and generating an initial three-dimensional point cloud after splicing and denoising the observation data; A shear loading and signal acquisition module for applying a shear load to the target rock mass and collecting time-sharing three-dimensional point clouds, strain signals, and vibration signals at each load stage; A multi-scale feature extraction module for feature encoding the time-sharing three-dimensional point clouds, strain signals, and vibration signals and generating regulation parameters for grouting and shear loads; A regulation and visualization output module for adjusting the grouting pressure or loading method according to the regulation parameters, and displaying the execution process with the time-sharing three-dimensional point clouds, strain signals, and vibration signals in the form of three-dimensional deformation maps and crack distribution maps, and outputting a monitoring report on the shear failure evolution of grouted joints.

Images