A point cloud-based shaft repair operation auxiliary system

Abstract

This invention relates to the field of image processing technology, specifically disclosing a point cloud-based auxiliary system for well repair operations. This system addresses the problem that fixed morphological operators for patching holes in well walls can easily cross the true boundary or fail to propagate sufficiently, leading to morphological hole patching errors. The system includes a point cloud image data acquisition module and a data analysis module. The data analysis module incorporates the characteristics of the well point cloud data for adaptive morphological hole patching correction, correcting these errors. Through design drawing reference registration, deep extraction of well features, and adaptive morphological correction with three-dimensional recommended constraints, the system ensures stable point cloud image morphological hole patching correction even under conditions of dust obstruction, reflective stripes, and missing holes, preventing errors caused by hole patching crossing the true boundary or insufficient propagation.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a point cloud-based auxiliary system for well chute repair operations. Background Technology

[0002] Over time, various types of damage will occur on the walls of chutes. Current technologies mostly use 3D point cloud image data to identify the damage information of the chute walls and help guide the chute repair work. By using 3D point cloud image data, the spatial morphology and geometric deviation of the inner surface of the chute wall can be realistically restored within the entire elevation range of the chute. By aligning with the original CAD design of the chute wall and diameter benchmark, the damage boundary, maximum damage depth, maximum damage width and damage volume of each segment can be extracted according to the elevation. This provides a quantitative basis for determining the cross-section, arranging the anti-impact section and selecting support parameters such as anchor bolts. It can also further help determine the elevation of the key working surface in confined space operations, thereby guiding the landing position and target air volume configuration of ventilation ducts, local fans and sealed air ducts to ensure ventilation and dust extraction during the repair process. However, in point cloud processing, if voxelization or distance field construction is chosen followed by closed-loop reconstruction for hole repair and connected component correction, existing technologies often use fixed structural element scales and static thresholds to generate marker fields and mask fields to control reconstruction propagation. However, the dust carried by the ore-falling airflow above the main chute can cause instantaneous occlusion and high-frequency floating points. The steel rails used for external support or the manganese steel plates used as lining plates in the chute can form striped pseudo-boundaries due to strong reflection. Furthermore, the combined effects of backflow dust caused by the vacancy or blockage of the lower inclined chute and blind spots can lead to missed scan holes and noise spikes in the point cloud data, and further contribute to the problem at distance... The noise in the field is amplified, causing the reconstruction closing operation to cross the actual damage boundary, mistakenly sticking adjacent defects together and overestimating the damage volume and width. It also causes the reconstruction closing operation to stop prematurely at weak spalling boundaries and cracks, underestimating the damage depth and continuous height. This results in the segmented damage level and key treatment area identification in the 3D point cloud recognition information, leading to the mismatch of treatment strategies. It also causes the work surface elevation and high dust risk section positioning to be misaligned, resulting in repeated adjustments and rework of the ventilation ducts, local fans, air duct layout, dust extraction path, and air volume margin configuration of the ventilation auxiliary system, resulting in redundant investment in support materials and ventilation materials and wasted time. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide an auxiliary system for well repair operations based on point cloud. Through design drawing reference registration, deep extraction of well features and adaptive morphological correction with three-dimensional recommendation constraints, the system enables stable morphological hole filling of point cloud images under conditions of dust obscuration, reflective stripes and missing holes, thereby reducing morphological hole filling errors caused by mis-adhesion overestimation and insufficient propagation underestimation.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A point cloud-based auxiliary system for well pass repair operations includes a point cloud image acquisition module and a data analysis module that are interconnected. The point cloud image acquisition module acquires well wall point clouds, images, timestamps, and camera posture. The data analysis module imports the well diameter and well wall reference from the design drawings, establishes the well shaft axis, and performs registration mapping. It calculates roundness, eccentricity, cavity expansion ratio, profile gradient, branch entry azimuth angle, and reflective strip intensity using a grid with elevation and circumferential angle steps, and calculates the probability of the first collision zone from the azimuth angle and cavity expansion ratio. The image occlusion rate is the percentage of pixels within the grid whose brightness or contrast is below a threshold, and the point cloud missing rate is the ratio of the actual number of points within the grid to the expected number of points. For example, the expected number of points is converted from the design well diameter and step size; the image quality weight is one minus the weighted sum of occlusion rate and missing detection rate and truncated to zero to one; the data analysis module processor calls the trained model to calculate and determine the target side length, principal axis length of the structuring element, direction vector, direction weight, collision zone priority area identifier, strip suppression area identifier, marking threshold, mask threshold, propagation radius and iteration upper limit based on the above parameters, the direction vector is normalized, the direction weight is truncated to zero to one, and the iteration upper limit is rounded down; and the correction field is obtained by scaling the propagation radius and iteration upper limit according to the quality weight and constrained by the priority area and suppression area on the voxel field or distance field and performing morphological reconstruction closing operation.

[0006] It should be noted that, compared to existing methods for wellbore wall inspection, which often rely on manual measurement using single laser point clouds or 2D images, and employ fixed voxel scales and structural elements with static thresholds for filtering and morphological closing operations to fill holes in the point cloud image, this invention addresses the problems of inconsistent cross-sectional references and inconsistent segmented statistical calibers by establishing the wellbore axis based on design drawings and completing point cloud registration and mapping. Furthermore, it utilizes point cloud feature parameters from segmented wellbores to fuse images. This method also addresses issues such as inaccurate segmented elevation positioning due to misalignment with existing CAD design benchmarks, amplification of noise and missed holes in the distance field, and false boundary adhesion caused by reflective stripes and floating points, or premature truncation and reconstruction leading to morphological hole filling errors. Occlusion rate and point cloud missing measurement rate form quality weights to solve point cloud quality fluctuations caused by impact dust occlusion, missed scanning, and reflective stripes. The initial values ​​of the triangular mesh unit scale and the principal axis scale of the structuring element are determined by three-stage recommendation adaptively. Furthermore, the orientation and weight of the structuring element, the identification of the collision zone priority area and the stripe suppression area, and the propagation constraint parameters of the labeling threshold, mask threshold, propagation radius, and iteration upper limit are given. This solves the problem of morphological hole filling errors in point cloud image morphological hole filling caused by dust occlusion, reflective stripes, and missing point cloud data in the existing technology, thereby improving the stability of morphological hole filling of point cloud images by the 3D point cloud data correction field.

[0007] As a further embodiment of the present invention, the point cloud image acquisition module includes a mounting frame for installing and fixing image acquisition and point cloud acquisition equipment set within the mounting frame. A LiDAR scanner is installed on the inner side of the mounting frame. The LiDAR scanner is used to scan the internal structure of the ore pass and present a three-dimensional point cloud model of the ore pass. A camera is installed below the LiDAR scanner for real-time detection of damage to the inner wall of the ore pass. A high-intensity flashlight facing the ore pass wall is installed on one side of the camera to assist in omnidirectional imaging of high-definition images. A wire bridge protection device is installed on the top of the mounting frame to protect the wire bridge from mechanical damage caused by falling ore or the inner wall of the ore pass. A wire bridge device is provided on the inner side of the wire bridge protection device to realize data communication between the point cloud image acquisition module and the data analysis module. The point cloud image acquisition module communicates with the data analysis module through the wire bridge device. A winch is connected to the mounting frame through a suspension rope. The winch is used to wind the suspension rope to ensure that the point cloud scanning range of the point cloud image acquisition module and the shooting range of the camera are within the entire ore pass section.

[0008] As a further embodiment of the present invention, the data analysis module includes a design drawing import module, which imports the original chute design drawings from the design institute. This module is connected to a registration and mapping module, which imports the well diameter and well wall references from the original chute design drawings, establishes the well shaft axis, and determines the correspondence between the design coordinate system and the point cloud coordinate system. It performs translation, rotation, and scale correction on the point cloud data to complete the registration with the design reference, and projects the registered point cloud onto the elevation and circumferential angle coordinate systems according to the well shaft axis to form a unified spatial reference for segmented analysis. The registration and mapping module is connected to a chute feature extraction module, which extracts the chute features required by the present invention. This module is connected to a first recommendation module, a second recommendation module, and a third recommendation module. The first recommendation module recommends grid and morphological correction scale parameters based on the chute geometry and point cloud resolution features, and outputs the initial values ​​of the target side length of the triangular mesh unit and the principal axis length of the structural element. The first recommendation module is used to determine the basic analytical scale for voxelization, distance field construction, and subsequent reconstruction operations. The second recommendation module is used to recommend morphological correction direction and regional constraint parameters based on the collision zone information formed by the branch entry orientation and cavity expansion characteristics, combined with the reflective strip features. It outputs the structuring element direction vector, direction weight, collision zone priority area identifier, and strip suppression area identifier to control the main direction of reconstruction propagation and the priority / suppression spatial region. The third recommendation module is used to recommend morphological correction threshold and propagation constraint parameters based on the image quality weight formed by the image occlusion rate and point cloud missing rate, combined with the leakage geometry factor. It outputs the marker field threshold, mask field threshold, reconstruction propagation radius, and iteration upper limit to determine the seed, propagable range, and stopping condition of the reconstruction closing operation. The first, second, and third recommendation modules are all connected to the reconstruction correction module. The reconstruction correction module is connected to the segmented parameter output module. The registration and mapping module is also connected to the field construction module, which is connected to the reconstruction correction module.

[0009] As a further aspect of this invention, in the data analysis module, the registration and mapping module establishes the wellbore axis based on the well diameter and well wall reference output by the design drawing import module and maps the point cloud to the elevation and circumferential angular coordinate system; the well chute feature extraction module generates point cloud feature parameters of the segmented well chute based on the registration and mapping results. The point cloud feature parameters of the segmented well chute include roundness, eccentricity, expansion rate, elevation profile gradient, branch entry azimuth, reflective strip intensity, and empty inclined chute discharge geometric factor, and outputs the probability of the first collision zone and image quality weight; the first recommendation module receives the roundness, eccentricity, expansion rate, and elevation profile gradient output by the well chute feature extraction module, and outputs the initial values ​​of the target side length of the triangular mesh unit and the principal axis length of the structural element required for morphological hole filling; the second recommendation module receives the branch entry azimuth, expansion rate, and probability of the first collision zone, and outputs the structural element required for morphological hole filling. The system comprises three modules: a first module receives the image quality weights, a second module receives the image quality weights, a collision zone priority area identifier, and a stripe suppression area identifier; a third recommendation module receives the image quality weights, the intensity of the reflective stripe, and the geometric factor for the discharge of the empty inclined chute, and outputs the marker field threshold, mask field threshold, reconstruction propagation radius, and iteration upper limit required for morphological hole filling; a field construction module receives the point cloud output by the registration and mapping module and generates a voxel field or a range field; a reconstruction correction module receives the voxel field or range field generated by the field construction module and the output parameters of the first recommendation module, the second recommendation module, and the third recommendation module, and performs a reconstruction closing operation under the constraints of the image quality weights, the collision zone priority area identifier, and the stripe suppression area identifier to obtain a correction field to correct morphological hole filling errors; and a segmented parameter output module receives the correction field and outputs the maximum damage depth, maximum damage width, and damage volume in segments according to elevation, evaluates the damage level, and obtains a set of ventilation parameters.

[0010] As a further aspect of the present invention, the first recommendation module employs a constrained gradient boosting regression tree model for recommendation. The model inputs are roundness, eccentricity, cavity expansion rate, elevation profile gradient, and point density. The model outputs the initial values ​​of the target side length of the triangular mesh unit and the three principal axis lengths of the structural element. The model uses the well diameter in the design drawing as the upper bound of the scale and the local point spacing as the lower bound of the scale. Monotonicity constraints are set on the target side length and the three principal axis lengths. The monotonicity constraints are set according to the coupling scale index of the cavity expansion rate and the point spacing. The coupling scale index is the cavity expansion rate multiplied by the point spacing. In the gradient boosting regression tree, monotonically increasing constraints are applied to the cavity expansion rate, the point spacing, and the coupling scale index simultaneously, so that the recommended target side length of the triangular mesh and the three principal axis lengths of the structural element increase with the increase of the cavity expansion rate and the point spacing, and are guaranteed not to decrease with the increase of the coupling scale index.

[0011] It should be noted that the first recommendation module transforms the scale of triangulation and structural elements from fixed empirical values ​​into constrained adaptive outputs: when the local cavity expansion of the chute is more severe or the point cloud is sparser, the model will inevitably provide target side lengths and principal axis lengths that do not decrease, and these are constrained by the upper limit of the designed well diameter and the lower limit of the point spacing. This avoids taking too small a scale on the sparse point cloud, which would cause noise protrusions to be treated as real damage boundaries, the distance field to be amplified, and false adhesion to occur in the closing operation. It also avoids taking too small a scale in areas with significant cavity expansion, which would cause reconstruction propagation to cross cracks and weak spalling boundaries. This ensures that the scale, geometric deformation degree, and sampling resolution of voxelization, distance field construction, and subsequent morphological reconstruction are consistent, reducing the risk of scale mismatch caused by dust obstruction, missed scan holes, and reflective strips. This improves the stability of damage depth, width, volume calculation, and level judgment from the source, thereby reducing repeated adjustments to repair sections, support parameters, and ventilation parameters and the waste of materials and time caused by misjudgment.

[0012] As a further aspect of this invention, the second recommendation module employs a graph attention network model oriented towards a circumferential ring topology for recommendation. Circumferential sectors at the same elevation are used as graph nodes, and adjacent circumferential sectors and the first and last sectors are used as graph edges to construct a circumferential ring graph. The model input includes the branch entry azimuth, cavity expansion rate, probability of the first collision zone, intensity of the reflective strip, and geometric feature vectors of the circumferential sectors. The model uses a dual-channel periodic representation of the branch entry azimuth using sine and cosine encoding, sets attention bias weights for the probability of the first collision zone, and sets a propagation suppression mask for the intensity of the reflective strip. The model output includes the structural element direction vector, direction weights, collision zone priority area identifier, and strip suppression area identifier. Unit length normalization constraints are applied to the structural element direction vector, and interval mapping constraints are applied to the direction weights.

[0013] It should be noted that the morphological school is transformed from isotropic and globally weighted to a circumferentially continuous, zone-specific, and directionally controllable constraint output: the circumferential loop connects sectors at the same elevation into closed loops based on their adjacency, ensuring continuity and consistency of direction vectors and direction weights in the circumferential direction; the branch entry azimuth angle is periodically encoded using sine and cosine to avoid azimuth jumps at 0° and 360°; the collision zone probability serves as an attention bias, enabling the model to automatically focus information on high-risk sectors of the first collision and output priority zone identifiers; the reflective strip intensity serves as a propagation suppression mask, causing the corresponding sectors to be downweighted during information transmission and correction propagation and output suppression zone identifiers; and the normalization of direction vectors and weight interval mapping ensure that the output is directly used for... The structural elements are oriented without scale drift or divergence; when the ore falls and impacts during dumping, the impact airflow instantly picks up dust, causing obstruction. At the same time, the strong reflection of the steel rails or manganese steel plates installed on the outside of the shaft wall forms strip-shaped pseudo-boundaries in the point cloud. Furthermore, when laser point cloud scanning also misses some holes, it can avoid the reconstruction closure operation crossing the real damage boundary at the strip pseudo-boundary, causing false adhesion. At the same time, it prioritizes the completion of cavity repair and connectivity correction along the main impact direction near the collision zone, reducing the overestimation of damage width and volume and the underestimation of damage depth and continuous height. This improves the reliability of segmented damage level determination and key treatment area positioning, and reduces rework and waste caused by repeated adjustments to repair sections, support parameters, and ventilation parameters.

[0014] As a further aspect of the present invention, the third recommendation module employs a quantile regression neural network model with uncertainty for recommendation. The model inputs are image occlusion rate, point cloud missing rate, image quality weight, vacant inclined chute discharge geometric factor, reflective strip intensity, collision zone priority area identifier and strip suppression area identifier. The model outputs are label field threshold, mask field threshold, reconstruction propagation radius and iteration upper limit. The model applies upper and lower bound interval mapping constraints to the label field threshold and mask field threshold, applies non-negativity constraints to the reconstruction propagation radius, applies integerization and upper limit truncation constraints to the iteration upper limit, and uses image quality weight as sample weight in loss calculation during training.

[0015] It should be noted that the most sensitive thresholds and propagation boundaries in morphological reconstruction closing operations are transformed from fixed thresholds and fixed iteration steps into outputs that adapt to data reliability and have uncertainty control: the quantile regression neural network provides predictions for different quantiles under the same input conditions, making the label field threshold, mask field threshold, propagation radius, and iteration upper limit controllable in terms of conservatism; the upper and lower bound mapping of the threshold ensures that the threshold always falls within the usable range, and the non-negativity constraint of the propagation radius and the integerization and truncation of the iteration upper limit ensure that the propagation will not produce invalid values ​​or infinite growth; during training, image quality weights are used to weight the loss, so that the impact of low-quality samples with high occlusion, high missing values, and strong banding on model parameter updates is suppressed. High-quality samples drive learning, resulting in more stable thresholds and stopping conditions. When instantaneous occlusion caused by impact dust coexists with floating points, missed scan holes, and reflective strip pseudo-boundaries, the system can automatically tighten or loosen the marker and mask thresholds and adjust the propagation radius and iteration steps based on the occlusion rate and missing detection rate. This avoids the situation where the threshold is too loose, causing the reconstruction to cross the real damage boundary and mis-adhere, or overestimating the volume and width. It also avoids the situation where the threshold is too tight or the propagation is insufficient, resulting in unrepaired voids and underestimated damage depth and continuous height. This improves the consistency between the correction field and the segmented damage level output, and provides more reliable segmented input for the ventilation parameter set, reducing repeated modifications and rework waste in ventilation configuration and repair strategies.

[0016] As a further embodiment of the present invention, the segmented parameter output module determines a low-damage segment as one that simultaneously satisfies the following conditions: maximum damage depth ≤ first depth threshold, maximum damage width ≤ first width threshold, damage volume ≤ first volume threshold, first collision zone probability ≤ first collision threshold, discharge geometric factor ≤ first discharge threshold, and image quality weight ≤ first quality threshold. The module outputs the target air volume as the number of workers multiplied by the minimum air supply constant per unit number of workers. The target negative pressure is based on the benchmark negative pressure. The dimensionless negative pressure demand index, obtained by combining the first collision zone probability and discharge geometric factor according to weights, is converted into a negative pressure correction amount by a preset negative pressure conversion coefficient and then added to the benchmark negative pressure. The allowable wind speed range is the target air volume divided by the effective cross-section with an upper and lower deviation coefficient applied. The dust extraction air volume is the target air volume multiplied by the first dust extraction ratio coefficient, which is a dimensionless coefficient corresponding to the low-damage segment and stored in the dust extraction configuration mapping table of the low-damage segment. The dustproof door sequence includes the pre-closing time, the unloading holding time, and the delayed dust extraction time.

[0017] It should be noted that the criteria for low-damage sections are determined by a combination of thresholds for damage depth, damage width, damage volume, probability of the first collision zone, discharge geometric factor, and image quality weight. The ventilation parameters are determined by the minimum air supply volume per person, the conversion correction amount of the baseline negative pressure and the dimensionless negative pressure demand index, the effective cross-sectional wind speed constraint, and the dust extraction configuration mapping table of the low-damage section. At the same time, the sealing sequence of pre-closing, unloading maintenance, and delayed dust extraction is output. This allows the low-damage section to achieve quantitative matching of dust extraction intensity and timing under the premise of meeting basic air supply and steady-state negative pressure. This reduces the risk of increased energy consumption and dust escape from the control area caused by excessive air volume margin, insufficient negative pressure, or excessive dust extraction due to empirical settings, and reduces the waste of materials and time caused by repeated adjustments to ventilation and dust extraction parameters.

[0018] As a further aspect of the present invention, the segmented parameter output module determines a segment as a medium-damage segment if the elevation segment meets the following criteria: maximum damage depth between a first depth threshold and a second depth threshold; maximum damage width between a first width threshold and a second width threshold; damage volume between a first volume threshold and a second volume threshold; collision zone probability between a first collision threshold and a second collision threshold; and image quality weight between a second quality threshold and a first quality threshold. The module then uses the cavity expansion rate, elevation profile gradient, collision zone probability, and reflective strip intensity to predict ventilation correction coefficients and negative pressure correction coefficients through a machine learning model. The output target air volume is the target air volume of the medium-damage segment multiplied by the ventilation correction coefficient. The target negative pressure is the target negative pressure of the low-damage segment multiplied by the negative pressure correction coefficient. The allowable wind speed range is corrected according to the speed limit coefficient and the minimum allowable coefficient. The dust extraction air volume is the target air volume multiplied by the second dust extraction ratio coefficient, which is a dimensionless coefficient corresponding to the low-damage segment and stored in the dust extraction configuration mapping table of the medium-damage segment. The module also includes the dustproof door timing pre-closing duration, ore unloading holding duration, delayed dust extraction duration, closure confirmation duration, and release delay duration.

[0019] It should be noted that the intermediate damage section is determined by using interval thresholds based on damage depth, damage width, damage volume, collision zone probability, and image quality weights. The expansion rate, elevation profile gradient, collision zone probability, and reflective strip intensity are input into the machine learning model to predict ventilation correction coefficients and negative pressure correction coefficients. The target air volume and target negative pressure are then adaptively corrected in segments. At the same time, the allowable wind speed range is constrained by the speed limit coefficient and the minimum guarantee coefficient. The second dust extraction ratio coefficient is determined by the dust extraction configuration mapping table of the intermediate damage section, and the dust extraction air volume is output. The sealing sequence consisting of pre-closing, unloading retention, delayed dust extraction, sealing confirmation, and release delay is then superimposed. This ensures that the negative pressure and dust extraction intensity are matched even under stronger impact dust and reflective strip interference in the intermediate damage section. This reduces the risk of dust escaping from the control area and insufficient air volume, and reduces repeated adjustments, rework, and waste of materials and time caused by mismatch in ventilation and dust extraction parameters.

[0020] As a further aspect of the present invention, the segmented parameter output module determines a high-damage segment as one that meets any of the following conditions: maximum damage depth > second depth threshold, maximum damage width > second width threshold, damage volume > second volume threshold, first collision zone probability > second collision threshold, discharge geometry factor > second discharge threshold, and image quality weight < second quality threshold. The output parameters for the high-damage segment are divided according to the start and end times of the unloading event and the wellhead sealing status: Before unloading, the output parameters include pre-ventilation air volume, pre-negative pressure, pre-wind speed range, pre-dust extraction air volume, and pre-door closing sequence; during unloading, the output parameters include increased operating air volume, enhanced negative pressure, dust extraction air volume, and wind speed range after speed limit, along with the wellhead dustproof door closing sequence; after unloading, the output parameters include decreased and stabilized air volume, negative pressure, wind speed range, delayed dust extraction air volume, and delayed door opening sequence, with minimum holding time constraints applied during parameter switching across the three stages.

[0021] It should be noted that high-damage sections are determined by triggering any condition based on thresholds for damage depth, damage width, damage volume, probability of the first collision zone, discharge geometric factor, and image quality weight. The ventilation and dust extraction parameters are decomposed into a preset stage of pre-ventilation, pre-negative pressure, and pre-dust extraction before unloading, a dust suppression stage during unloading by increasing the operating air volume by a preset ratio and strengthening negative pressure and dust extraction with speed control, and a cleaning stage after unloading by decreasing the air volume by a preset ratio, stabilizing, delaying dust extraction, and delaying door opening. At the same time, a minimum holding time constraint is applied to the stage switching, so that the risks caused by dust carried by the impact airflow, backflow and overflow dust, and missed sweeping are suppressed in segments in the time dimension. This avoids dust from escaping from the control area and energy consumption runaway caused by instantaneous high wind speed and insufficient negative pressure in high-damage sections, and reduces rework and material and time waste caused by repeated adjustments of ventilation and dust extraction parameters.

[0022] Compared with existing technologies, the technical advantages of this invention are:

[0023] This invention establishes the shaft axis based on the original design drawings and completes the registration of the point cloud with the design benchmark. It extracts the point cloud feature parameters of the segmented chute in the elevation-circumferential angular coordinate system, and combines the collision zone probability and image quality weights. It adopts a three-stage recommendation to adaptively determine the scale, orientation, and priority suppression region of the triangular mesh and structural elements required for morphological hole filling, as well as the threshold and propagation constraints of the marker field and mask field. Then, it performs morphological reconstruction closing operation under constraints on the voxel field or distance field to obtain a more stable and consistent correction field to correct morphological hole filling errors, improve the reliability of segmented damage parameters and level output, and match the ventilation parameter set with the damage level and dust risk of each elevation segment, reducing rework and redundant investment. Attached Figure Description

[0024] Figure 1 This is a system diagram of the present invention;

[0025] Figure 2 This is a block diagram of the data analysis module of the present invention;

[0026] Figure 3 This is a schematic diagram of the scanning operation of the camera and lidar of the present invention under the action of a winch; Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] like Figure 1 As shown, this invention proposes a point cloud-based auxiliary system for well pass repair operations, comprising a point cloud image acquisition module and a data analysis module that are interconnected. The point cloud image acquisition module acquires well wall point clouds, images, timestamps, and camera posture. The data analysis module imports the well diameter and well wall reference from the design drawings, establishes the well shaft axis, and registers and maps it. It calculates roundness, eccentricity, cavity expansion rate, profile gradient, branch entry azimuth angle, and reflective strip intensity using a grid with elevation and circumferential angle steps, and calculates the probability of the first collision zone from the azimuth angle and cavity expansion rate. The image occlusion rate is the percentage of pixels within the grid whose brightness or contrast is below a threshold, and the point cloud missing rate is the ratio of the actual number of points within the grid to the expected number of points. The insufficient proportion, the expected number of points is converted from the design well diameter and step size; the image quality weight is one minus the weighted sum of occlusion rate and missing detection rate and truncated to zero to one; the data analysis module processor calls the trained model to calculate and determine the target side length, structural element principal axis length, direction vector, direction weight, collision zone priority area identifier, strip suppression area identifier, marking threshold, mask threshold, propagation radius and iteration upper limit based on the above parameters, the direction vector is normalized, the direction weight is truncated to zero to one, and the iteration upper limit is rounded down and truncated; and the correction field is obtained by scaling the propagation radius and iteration upper limit according to the quality weight and constrained by the priority area and suppression area on the voxel field or distance field.

[0029] It should be noted that, in order to solve the technical problem, the system proposed in this invention performs the following operations:

[0030] Step 1, Point Cloud Image Synchronous Acquisition and Attitude Marking: The point cloud image acquisition module acquires point cloud data and image data of the well wall along the axial direction inside the well chute. It records the acquisition timestamp for each frame of point cloud and image synchronously, and records the camera attitude parameters. The point cloud frame, image frame and corresponding timestamp and attitude parameters are bound together to form the original acquisition data packet.

[0031] Step 2, Design datum import and wellbore axis establishment: The data analysis module imports the well diameter and well wall datum information from the original chute design drawings, establishes the design coordinate system, fits the wellbore axis based on point cloud data and determines the axis direction, and establishes the coordinate definition relationship of axis, elevation and circumferential angle, providing a unified geometric datum for subsequent registration and segmented statistics;

[0032] Step 3, Point cloud registration with design datum and elevation and circumferential angle mapping: Perform translation and rotation correction on the point cloud to align the point cloud coordinate system with the design coordinate system, and complete the registration of the point cloud with the design well diameter / well wall datum under axis constraints; project and map the registered point cloud onto the elevation and circumferential angle coordinate system to obtain a point cloud representation that can be segmented by elevation and positioned by circumference;

[0033] Step 4, Calculation of ore sluice collection and surface features: Perform layered / segmented statistical analysis on the point cloud within the elevation-circumferential angular coordinate system, and calculate the point cloud feature parameters for the following segmented ore sluices:

[0034] 1) Roundness: The statistical measure of the radial deviation from each elevation section point to the fitted circle of the section;

[0035] 2) Eccentricity: The offset of the center of the fitted section circle relative to the design axis;

[0036] 3) Expansion ratio: The enlargement ratio of the equivalent diameter of the cross-section relative to the designed well diameter;

[0037] 4) Elevation profile gradient: The rate of change of the cavity expansion ratio or equivalent radius along the elevation direction;

[0038] 5) Branch entry azimuth: Circumferential azimuth information corresponding to the branch chute;

[0039] 6) Reflection strip intensity: A strip index extracted from the reflection intensity and brightness characteristics of the image and point cloud;

[0040] 7) Geometric factor for vacant inclined chute discharge: A geometric quantitative index describing the impact of a vacant lower inclined chute on airflow discharge;

[0041] Step 5, Initial Collision Zone Probability Assessment: Based on the branch entry azimuth and expansion rate, calculate the possible area where the ore collides with the shaft wall for the first time after entering the main ore pass from the branch; express this area as a probability distribution in the circumferential and elevation dimensions to obtain the probability of the first collision zone, and form a candidate range of priority for the collision zone.

[0042] Step 6, Image quality weight calculation: Calculate the occlusion rate (such as dust occlusion, and the decrease in the proportion of effective pixels caused by blind spots) from the image; calculate the missing detection rate (such as missed scan holes, sparse area proportion) from the point cloud; fuse the occlusion rate and the missing detection rate to form the image quality weight, which is used to characterize the credibility of this data segment and serve as the weight input for subsequent recommendation and reconstruction constraints;

[0043] Step 7, First Recommendation: Input the roundness, eccentricity, cavity expansion ratio, elevation profile gradient and point cloud resolution related quantities into the first recommendation model, and output the initial values ​​of the target side length of the triangular mesh unit and the principal axis length of the structural element;

[0044] Step 8, Second Recommendation: Input the branch entry azimuth, cavity expansion rate, first collision zone probability, reflective strip intensity and circumferential sector geometric features into the second recommendation model, and output the structural element direction vector, direction weight, collision zone priority zone identifier, and strip suppression zone identifier.

[0045] Step 9, Third Recommendation: Input the image occlusion rate, point cloud missing rate, image quality weight, leakage geometry factor, reflective strip intensity, and priority / suppression region identifier into the third recommendation model, and output the label field threshold, mask field threshold, reconstruction propagation radius, and iteration upper limit.

[0046] Step 10, Voxel field or distance field construction: Based on the registered and mapped point cloud, construct the field representation for morphological operations: The voxel field divides the point cloud into voxels and generates voxel attributes of occupancy / density / distance, while the distance field calculates the distance values ​​from the point cloud surface to the spatial grid to form a scalar field.

[0047] Step 11, Constrained Reconstruction Closure Operation and Correction Field Generation: On the voxel field or distance field, constraints are applied according to image quality weights, collision zone priority region identifiers, and stripe suppression region identifiers. Combined with the scale parameters obtained from the first recommendation, the orientation and region parameters obtained from the second recommendation, and the threshold and propagation control parameters obtained from the third recommendation, morphological reconstruction closure operation is performed. Connectivity correction and hole repair control are applied to the reconstruction results to generate a correction field for destructive quantization.

[0048] Step 12, Segmented damage parameter calculation and level determination: Divide the correction field into segments according to elevation, extract the damage boundary of each segment, calculate the maximum damage depth, maximum damage width and damage volume of each segment, and output the damage level according to the preset threshold rules or level mapping table.

[0049] Step 13, Generation and Output of Ventilation Parameter Set: Take the segmented damage level and segmented damage parameters as input to generate a ventilation parameter set. The ventilation parameter set includes the target air volume, target negative pressure, allowable wind speed range and dust extraction air volume, and can generate corresponding parameter groups according to the ore unloading stage or operation status; output the ventilation parameter set bound to the corresponding elevation segment.

[0050] The point cloud image acquisition module includes a mounting frame. A LiDAR scanner is mounted inside the mounting frame. A camera is mounted below the LiDAR scanner. A high-powered flashlight, facing the well wall, is mounted to one side of the camera. A network bridge protection device is mounted on the top of the mounting frame. A network bridge device is located inside the network bridge protection device. Figure 2As shown, the point cloud image acquisition module is connected to the data analysis module via a network bridge device, and the mounting frame is connected to a winch via a suspension rope.

[0051] like Figure 3 As shown, the data analysis module includes a design drawing import module, which is connected to a registration and mapping module. The registration and mapping module is connected to a chute feature extraction module, which is connected to a first recommendation module, a second recommendation module, and a third recommendation module. The first recommendation module, the second recommendation module, and the third recommendation module are all connected to a reconstruction and correction module. The reconstruction and correction module is connected to a segmented parameter output module. The registration and mapping module is also connected to a field construction module, which is connected to the reconstruction and correction module.

[0052] Specifically, the data analysis module is implemented by an industrial computer or edge computing server, including a central processing unit, a graphics processing unit, a memory, and a storage medium; the central processing unit performs registration, feature and rule calculations, and the graphics processing unit is used for point cloud rendering, voxel or distance field construction, and deep learning inference acceleration; the storage medium saves design drawing files, point cloud files, image files, model parameter files, and segmented output results.

[0053] It should be noted that in the data analysis module, the registration and mapping module establishes the wellbore axis based on the well diameter and well wall reference output by the design drawing import module and maps the point cloud to the elevation and circumferential angle coordinate system; the registration and mapping module is implemented by running a registration algorithm on an industrial computer to establish the wellbore axis, align the point cloud with the design reference, and output the data index in the elevation-circumferential angle coordinate system.

[0054] The well pass feature extraction module generates roundness, eccentricity, cavity expansion rate, elevation profile gradient, branch entry azimuth, reflective strip intensity, and empty inclined chute discharge geometric factor based on the registration and mapping results, and outputs the probability of the first collision zone and image quality weights. For each elevation segment (e.g., every half meter or one meter), the module takes a set of points and performs robust circle fitting (commonly using random sampling consistency + least squares or median residual minimization). Roundness can be represented by the maximum minus the minimum of the radius residuals or the residual quantile difference. On industrial machines, this involves statistically analyzing the radius residuals of each segment. Using the center of the circle obtained from the previous fitting step, the offset distance of the center relative to the wellbore axis is calculated, then normalized using the design radius, outputting an eccentricity value for each segment. The eccentric direction is output according to the circumferential sector (for subsequent direction vector recommendation); the expansion rate is obtained by comparing the equivalent radius / equivalent cross-section of the segment with the corresponding radius / cross-section of the design well diameter. The equivalent radius is usually the median radius or the weighted average radius of the circumferential sector, which is more noise-resistant. The expansion rate curve (which varies with elevation) is output, and the expansion distribution of each segment (which circumferential sectors expand the most significantly) can also be output; the sequence of equivalent radius (or equivalent cross-section) - elevation is differentially / slidingly fitted to obtain the rate of change, and then the median filter is used to suppress jump points, outputting the profile gradient of each segment to identify sudden expansion, sudden contraction, steps, and funnel transition zones; the branch entry azimuth is obtained through two-way fusion to avoid misjudgment caused by point cloud noise. The drawing path directly reads the branch entry in the design coordinate system. Azimuth angle is mapped to the circumferential coordinate system. Point cloud path: Within the branch elevation range, the peak sector of circumferential expansion rate or the sector with the strongest local concavity and convexity changes is statistically analyzed to obtain the observed azimuth angle; consistency is checked with the drawing path, and if the difference exceeds the threshold, it is marked as low confidence. The azimuth angle and confidence level of each branch entering the well are output; The intensity of reflective strips (typical strip pseudo-boundaries of light rail / manganese steel plates) is jointly estimated using point intensity and image texture. On the point cloud side, the high quantile value of point intensity and the proportion of continuous high intensity points are statistically analyzed in each sector, and circumferential connectivity is checked (strips are usually "long, narrow, and extend along the well wall"). On the image side, the bright narrow strip texture energy is extracted from the image area of ​​the corresponding azimuth (e.g., using directional filtering or frequency domain energy), and time stamp alignment is performed with the point cloud intensity peak. The system outputs the strip intensity of each segment and sector for subsequent strip suppression zone identification. The vacant inclined chute discharge geometric factor is a dimensionless representation of the lateral opening's ability to discharge compressed airflow from the wellbore. In engineering, it is usually calculated as follows: first, the geometric range of the inclined chute opening in the circumference and elevation is given using drawings; then, the effective opening ratio within this range is determined using point cloud (significant missing points in the point cloud at the opening, clear edges, and discontinuity with the inner surface of the wellbore). The discharge geometric factor is formed by combining the ratio of the effective opening area to the main well cross-sectional area and the ratio of the opening length along the path to the characteristic diameter. Corrections are given for the blocking status (whether there is continuous point cloud backfill / structural occlusion features). The discharge geometric factor of each segment is output, directly serving the segmented configuration of subsequent ventilation and dust extraction parameters.The first collision zone in a chute is usually strongly correlated with the branch entry azimuth, cavity expansion morphology, and profile gradient. For industrial machines, creating a probability map of circumferential sectors is most practical. Specifically, candidate center sectors are determined by the branch entry azimuth, with sectors showing peak expansion rate and abrupt changes in profile gradient used as gain terms, and sectors with high stripe intensity used as suppression terms (to prevent false boundaries from dragging the collision zone off-center). The final output is the collision zone probability for each segment and sector, along with a high-probability continuous sector formation priority area identifier. The image quality weights simultaneously reflect whether the image is clear and whether the point cloud has any defects. The measurement should guide subsequent morphological reconstruction parameter tightening / loosening. This is achieved by calculating the effective texture / contrast attenuation percentage (dust and fogging reduce contrast) of the image region matched to that sector, thus obtaining the image occlusion rate. The number of points in that grid is compared with the expected number of points (given by scan trajectory, distance, incident angle, and historical density statistics) to obtain the point cloud missing rate. The image occlusion rate and point cloud missing rate are mapped to quality scores of the same dimension, then synthesized into quality weights (lower weights for lower quality), and written into the constraint input for subsequent reconstruction and correction.

[0055] The first recommendation module receives the roundness, eccentricity, expansion rate, and elevation profile gradient output by the well feature extraction module, and outputs the initial values ​​of the target side length of the triangular mesh unit and the principal axis length of the structural element. The second recommendation module receives the branch entry azimuth, expansion rate, and probability of the first collision zone, and outputs the direction vector of the structural element, direction weight, and the collision zone priority area identifier and strip suppression area identifier. The third recommendation module receives the image quality weight, reflective strip intensity, and empty inclined chute discharge geometric factor, and outputs the marker field threshold, mask field threshold, reconstruction propagation radius, and iteration upper limit. The field construction module receives the point cloud output by the registration and mapping module and generates a voxel field or range field. The reconstruction correction module receives the voxel field or range field generated by the field construction module and the output parameters of the first, second, and third recommendation modules, and performs reconstruction closing operation under the constraints of image quality weight, collision zone priority area identifier, and strip suppression area identifier to obtain the correction field. The segmented parameter output module receives the correction field and outputs the maximum damage depth, maximum damage width, and damage volume in segments according to elevation, evaluates the damage level, and obtains the ventilation parameter set.

[0056] It should be specifically noted that the first recommendation module uses a constrained gradient boosting regression tree model for recommendation. The model inputs are roundness, eccentricity, cavity expansion ratio, elevation profile gradient, and point density. The model outputs the initial values ​​of the target side length of the triangular mesh unit and the three principal axis lengths of the structural element. The model uses the well diameter in the design drawing as the upper bound of the scale and the local point spacing as the lower bound of the scale. Monotonicity constraints are set on the target side length and the three principal axis lengths. The monotonicity constraints are set according to the coupling scale index of the cavity expansion ratio and the point spacing. The coupling scale index is the cavity expansion ratio multiplied by the point spacing. In the gradient boosting regression tree, monotonically increasing constraints are applied to the cavity expansion ratio, point spacing, and coupling scale index simultaneously, so that the recommended target side length of the triangular mesh and the three principal axis lengths of the structural element increase with the increase of the cavity expansion ratio and the point spacing, and are guaranteed not to decrease with the increase of the coupling scale index.

[0057] It should be specifically noted that the second recommendation module adopts a graph attention network model oriented towards a circumferential ring topology for recommendation. It uses circumferential sectors at the same elevation as graph nodes and adjacent circumferential sectors and the first and last sectors as graph edges to construct a circumferential ring graph. The model inputs are the branch entry azimuth angle, cavity expansion rate, probability of the first collision zone, intensity of the reflective strip, and geometric feature vectors of the circumferential sector. The model uses a dual-channel periodic representation of the branch entry azimuth angle with sine and cosine encoding, sets attention bias weights for the probability of the first collision zone, and sets a propagation suppression mask for the intensity of the reflective strip. The model outputs are the structuring element direction vector, direction weights, collision zone priority area identifier, and strip suppression area identifier. The structuring element direction vector is subject to a unit length normalization constraint, and the direction weight is subject to an interval mapping constraint.

[0058] It should be specifically noted that the third recommendation module uses a quantile regression neural network model with uncertainty for recommendation. The model input includes image occlusion rate, point cloud missing rate, image quality weight, vacant inclined chute discharge geometric factor, reflective strip intensity, collision zone priority area identifier and strip suppression area identifier. The model output includes the label field threshold, mask field threshold, reconstruction propagation radius and iteration upper limit. The model applies upper and lower bound interval mapping constraints to the label field threshold and mask field threshold, applies non-negativity constraints to the reconstruction propagation radius, and applies integerization and upper limit truncation constraints to the iteration upper limit. During training, the image quality weight is used as the sample weight to participate in the loss calculation.

[0059] Specifically, this can be implemented by using a unified multi-task network that integrates all three recommendation tasks into a single recommendation model infrastructure: the underlying structure shares a common encoding backbone, and different recommendation tasks output results separately using three output heads (one for scale, one for direction and partitioning, and one for threshold and uncertainty). The specific network structure includes:

[0060] 1) Input organizational layer

[0061] Segment-level global feature vectors: roundness, eccentricity, expansion ratio, elevation profile gradient, point density, point spacing, coupling scale index of expansion ratio and point spacing, and geometric factor for vacant inclined chute discharge.

[0062] Circumferential sector feature sequence: Circumferential sectors are divided according to the same elevation segment. Each sector includes the sine and cosine values ​​of the branch entry azimuth angle, the probability of the first collision zone, the intensity of the reflective strip, the local geometric statistics of the sector, the image occlusion rate, and the point cloud missing rate.

[0063] 2) Shared backbone encoder

[0064] Segment-level encoder: The multilayer perceptron embeds segment-level global features into segment-level latent vectors.

[0065] Circumferential Loop Encoder: The circumferential sector is used as a graph node, and adjacent sectors and the first and last sectors are used as graph edges. The graph attention layer is stacked to obtain the latent vector of each sector. An attention bias is injected into the probability of the first collision zone, and a propagation suppression mask is injected into the intensity of the reflective strip, so that the circumferential start and end are continuous and the strip does not dominate the propagation.

[0066] Fusion layer: Cross attention is used to fuse segment-level latent vectors and sector-level latent vectors to obtain a unified fused latent representation; thus, a forward inference simultaneously possesses global scale information and circumferential partition information.

[0067] 3) The three output heads correspond to the three recommendation modules.

[0068] First output head (scale recommendation head, corresponding to the first recommendation module)

[0069] Input: The segment-level part of the fused latent representation.

[0070] Output: Initial values ​​for the target side length of the triangular mesh element and the lengths of the three principal axes of the structuring element.

[0071] Constraint implementation method: A monotonically increasing constraint is applied to the cavity expansion rate, point spacing, and coupling scale index inside the head using a monotonically increasing constraint layer (which can be implemented in engineering by weight nonnegation and accumulation structure), and an interval mapping layer is used to clamp the output within a range with the design well diameter as the upper bound and the local point spacing as the lower bound.

[0072] Second output header (direction and partition recommendation header, corresponding to the second recommendation module)

[0073] Input: The sector portion of the fused hidden representation.

[0074] Output: Structural element direction vector, direction weight, collision zone priority region identifier, and stripe suppression region identifier.

[0075] Constraint implementation method: The direction vector is forced to a unit length using a normalization layer; the direction weight is limited to zero to one using an interval mapping layer; the priority zone identifier and the suppression zone identifier are output with thresholding and additional circumferential continuity regularization, so that no discontinuity or jump occurs in the circumferential direction in the same elevation segment.

[0076] The third output header (threshold and uncertainty recommendation header, corresponding to the third recommendation module)

[0077] Input: The parts of the latent representation related to occlusion, missing measurements, and quality are fused.

[0078] Outputs: Marker field threshold, mask field threshold, reconstruction propagation radius, iteration upper limit, and uncertainty quantile information for each output quantity.

[0079] Constraint implementation: The threshold is limited by an interval mapping layer; the propagation radius is limited by a non-negative constraint layer; the upper limit of iteration is limited by an integerization and upper limit truncation layer; during training, image quality weights are used as sample weights in the quantile loss, so that low-quality segments automatically tend towards a more conservative threshold and a shorter propagation.

[0080] In point cloud data processed using existing techniques, dust obscuring and floating points near the wellhead / funnel mouth form a splash point cloud. Fixed-scale closed-loop operations easily treat these discrete noises as propagable entities, leading to an outward expansion of the wellhead profile, flattening or bulging of the upper section. The reflective strips of the external light rail / manganese steel plate appear as continuous high-contrast bands in the distance field. Static thresholds may misjudge these strips as damage boundaries or entity boundaries, causing false adhesion along the strip direction. This connects what should be separate damage areas into one, further overestimating the width and volume of the damage. If the missed scan holes happen to fall on weak spalling boundaries or... Near the crack, the fixed threshold will stop propagation too early, resulting in the underestimation of the damage depth and continuous height, and the position of the survey line will drift with the change of the viewing angle. This invention introduces a collision zone priority area marker during the correction, so that the hole repair / reconstruction propagation is completed first in the high-risk zone of the first collision, but does not spread uncontrollably to the non-collision zone. It also introduces a strip suppression zone marker, and the area corresponding to the reflective strip is downweighted / masked in the propagation and connected domain correction, so that the strip no longer dominates the boundary. It introduces a quality weight constraint, and the propagation radius and iteration upper limit of the segments with high occlusion and high missing measurement are tightened to avoid low-quality segments dragging the contour off.

[0081] Under existing technology, circumferential sectors are prone to loop breaks or abrupt changes (zero-degree / 360-degree discontinuities) at the beginning and end junctions, causing the circular wellbore to appear elliptical / squared in the upward view. Near the branch entry point, stripes or missed scan holes can cause incorrect connectivity in the closure operation, mistaking defects at the branch interface as part of the main well, and thus misjudging the cavity expansion range. After processing by the system of this invention, the boundary of the circular wellbore will be closer to a closed loop, the boundary of the branch interface will be clearer, and the beginning and end sectors will no longer be broken. This is because the circumferential loop topology attention of the second recommended module brings two direct improvements:

[0082] 1) The branch entry azimuth is represented by a sine / cosine period to eliminate circumferential start and end jumps;

[0083] 2) Apply attention bias to the collision zone probability and apply a propagation suppression mask to the strip intensity to make the priority / suppression zone of the model output continuous and consistent in the circumferential direction, so that the subsequent morphological school propagates continuously in the correct circumferential direction and automatically shrinks at the strip.

[0084] Under existing technology, if the patching is too strong, the pit boundary is filled, leading to an underestimation of the damage depth. If noise protrusions are not suppressed, they are treated as boundaries, resulting in an overestimation of the damage height or width. The landing points of the three measurement lines (depth / width / height) are highly dependent on the boundary pixels / points. Noise and missed scans can cause the endpoints of the measurement lines to drift, and the same damage will give different values ​​under different truncation methods. After processing with the system of this invention, the local damage boundary will fit the real pit contour more closely, the endpoints of the measurement lines will be more consistent, the landing point of the damage depth measurement line will no longer be pulled off by the floating point, the damage width measurement line will no longer cross the strip pseudo boundary, and the damage height measurement line will be more stable on the peeling continuous segment and will not be cut off by the hole. This is because the third recommended module makes the marker field threshold, mask field threshold, propagation radius, and iteration upper limit into an adaptive output with uncertainty. When the occlusion rate / missed measurement rate is high, the threshold and propagation will be automatically tightened. When the quality weight is high, it will allow sufficient repair of real holes, thereby reducing both the two types of deviations: overestimation of volume due to false adhesion and underestimation of depth due to insufficient propagation.

[0085] It should be noted that the segmented parameter output module determines a low-damage segment as one that simultaneously meets the following criteria: maximum damage depth ≤ first depth threshold, maximum damage width ≤ first width threshold, damage volume ≤ first volume threshold, first collision zone probability ≤ first collision threshold, discharge geometric factor ≤ first discharge threshold, and image quality weight ≤ first quality threshold. The module outputs the target air volume as the number of workers multiplied by the minimum air supply constant per unit number of workers. The target negative pressure is based on the benchmark negative pressure. The dimensionless negative pressure demand index, obtained by combining the first collision zone probability and discharge geometric factor according to weights, is converted into a negative pressure correction amount by a preset negative pressure conversion coefficient and then added to the benchmark negative pressure. The allowable wind speed range is the target air volume divided by the effective cross-section with an upper and lower deviation coefficient applied. The dust extraction air volume is the target air volume multiplied by the first dust extraction ratio coefficient, which is a dimensionless coefficient corresponding to the low-damage segment and stored in the dust extraction configuration mapping table of the low-damage segment. The dustproof door timing includes the pre-closing time, the unloading holding time, and the delayed dust extraction time.

[0086] It should be noted that the segmented parameter output module determines the elevation segment as a medium-damage segment if the maximum damage depth is between the first and second depth thresholds, the maximum damage width is between the first and second width thresholds, the damage volume is between the first and second volume thresholds, the collision zone probability is between the first and second collision thresholds, and the image quality weight is between the second and first quality thresholds. The module then uses the cavity expansion rate, elevation profile gradient, collision zone probability, and reflective strip intensity to predict ventilation correction coefficients and negative pressure correction coefficients through a machine learning model. The output target air volume is the target air volume of the medium-damage segment multiplied by the ventilation correction coefficient, the target negative pressure is the target negative pressure of the low-damage segment multiplied by the negative pressure correction coefficient, the allowable wind speed range is corrected by the speed limit coefficient and the minimum guarantee coefficient, the dust extraction air volume is the target air volume multiplied by the second dust extraction ratio coefficient, the second dust extraction ratio coefficient is a dimensionless coefficient corresponding to the low-damage segment and stored in the dust extraction configuration mapping table of the medium-damage segment, and the dust door timing includes the pre-closing duration, ore unloading holding duration, delayed dust extraction duration, closure confirmation duration, and release delay duration.

[0087] It should be noted that the segmented parameter output module determines a high-damage segment as one that meets any of the following conditions: maximum damage depth > second depth threshold, maximum damage width > second width threshold, damage volume > second volume threshold, first collision zone probability > second collision threshold, discharge geometry factor > second discharge threshold, and image quality weight < second quality threshold. The output parameters for high-damage segments are divided according to the start and end times of the unloading event and the wellhead sealing status: Before unloading, the output includes pre-ventilation air volume, pre-negative pressure, pre-wind speed range, pre-dust extraction air volume, and pre-door closing sequence; during unloading, the output includes increased operating air volume, enhanced negative pressure, dust extraction air volume, and wind speed range after speed limiting, along with the wellhead dust door closing sequence; after unloading, the output includes decreased and stabilized air volume, negative pressure, wind speed range, delayed dust extraction air volume, and delayed door opening sequence, with minimum hold time constraints applied during parameter switching across all three stages.

[0088] Specifically, in engineering implementation, the segmented parameter output module can directly reuse the shared backbone encoder + multi-output head network structure of the recommended model facility. The output head is expanded from the morphological school positive parameter head to a ventilation negative pressure parameter head and a three-stage destruction scheduling head. Furthermore, a regularized threshold judgment and a three-stage switching state machine are superimposed after the network output. Specifically, this includes:

[0089] 1) Shared backbone network

[0090] Segment-level coding branch: The multilayer perceptron encodes segment-level features such as cavity expansion rate, elevation profile gradient, discharge geometry factor, point density, missing detection rate, and quality weight into segment-level latent vectors;

[0091] Circumferential ring topology branch: Circumferential sectors of the same elevation segment are constructed into a circumferential ring map. The sine and cosine values ​​of the branch entry azimuth angle, the collision zone probability, the reflective strip intensity, and the sector geometric statistics are encoded into sector latent vectors using a graph attention layer. Attention bias is applied to the collision zone probability and a propagation suppression mask is applied to the strip intensity.

[0092] Fusion layer: Cross attention merges segment-level latent vectors and sector-level latent vectors into a fused latent representation, which serves as the unified input for each output head;

[0093] 2) Network structure of the output of the middle destruction section

[0094] The segmented parameter output module first determines the segment as a medium-damage segment according to the threshold rule, and then calls the ventilation negative pressure regression output head of the unified recommendation model to give the coefficients. The output head structure is: fused hidden representation → two-layer perceptron → interval mapping layer.

[0095] Ventilation correction coefficient output channel: Outputs the ventilation correction coefficient, with interval mapping constraint being greater than zero and not exceeding the preset upper limit;

[0096] Negative pressure correction coefficient output channel: Outputs negative pressure correction coefficient, with interval mapping constraint being greater than zero and not exceeding the preset upper limit;

[0097] Speed ​​limit coefficient and minimum guarantee coefficient output channels: Output speed limit coefficient and minimum guarantee coefficient, and constrain their value ranges using interval mapping;

[0098] The dust extraction ratio coefficient uses a mapping table: the second dust extraction ratio coefficient is not obtained by network regression, but is directly read from the dust extraction configuration mapping table of the middle damaged section according to the damage level index;

[0099] Subsequently, the segmented parameter output module multiplies the base air volume and base negative pressure by the above coefficients to obtain the target air volume and target negative pressure, and corrects the allowable wind speed range according to the effective cross-section, speed limit coefficient, and minimum guarantee coefficient.

[0100] 3) Network structure of the three-stage output of the high-damage segment

[0101] After being identified as a high-damage segment, the segment parameter output module uses two parts to achieve three-stage output:

[0102] Three-stage scheduling state machine (rule-based logic, not dependent on learning)

[0103] The program takes the start and end times of the ore unloading event and the wellhead sealing status as inputs, outputs the current stage identifier, and applies a minimum hold duration constraint to the stage switching.

[0104] Three-stage gain output head (learn to provide the gain multiplier for each stage)

[0105] A stage scaling output head is added to the unified recommendation model infrastructure. The structure is fused latent representation + stage identifier embedding → multilayer perceptron → interval mapping layer, which outputs three sets of scaling parameters at one time.

[0106] Pre-unloading stage ratios: pre-ventilation air volume ratio, pre-negative pressure ratio, pre-dust extraction air volume ratio, and pre-door closing sequence parameters;

[0107] Stage ratios during ore unloading: operating air volume ratio, enhanced negative pressure ratio, dust extraction air volume ratio, wind speed limit ratio, and door closing time sequence parameters;

[0108] Post-unloading stage ratios: stabilization decreasing ratio, delayed dust extraction ratio, and delayed door opening sequence parameters;

[0109] All multipliers are constrained to be positive values ​​through interval mapping and have upper limits set. Timing parameters are constrained by non-negativity constraints and upper limit truncation constraints.

[0110] Ultimately, the pre-ventilation air volume, operating air volume, and stabilization air volume of the high-damage section are obtained by multiplying the base air volume by the corresponding stage multiplier. The negative pressure and dust extraction air volume are obtained in the same way, and the allowable wind speed range is further superimposed with speed limit constraints.

[0111] In this invention, identifying the damage level of the well wall at each elevation section can also help guide the well wall repair operation: Analysis of point cloud data of the main chute in a certain area shows that the low-damage sections are concentrated in sections 3-9 and 37, which have no large-scale impact on continued use. Therefore, the treatment is mainly local repair, without widening or scraping. Only small concrete falls can be repaired with concrete. However, it is also emphasized that the risk of damage and detachment of light rail anti-collision components is more prominent in these sections. Therefore, the focus of the treatment shifts to dismantling and replacing detached or partially damaged components. The light rail detached, and measures were taken to prevent weakening of the support structure or injury from falling debris. The damaged sections were relatively dispersed (sections 1-2, 12-15, 29-30, and 32-36), with a total damaged volume of approximately 113.36 cubic meters. The damage was characterized by a relatively shallow depth and width, but a spindle shape, lighter at the top and bottom and heavier in the middle. Therefore, the treatment mainly involved clearing loose rocks, widening the area, and reinforcing the structure. Different cross-sectional supports were used according to different damage dimensions to control costs. Sections 1-2 were caused by direct impact from ore unloading at the mine entrance. Further measures were proposed to increase... The impact was absorbed and secondary impact damage was reduced by constructing a shock-absorbing section (3-4 meters high) + anchor bolt support + high-manganese steel lining + reinforced concrete reconstruction. The high-damage sections (sections 11, 16, 22-27, 31, and 34) showed more significant damage, with a maximum depth of approximately 3.3 meters and a maximum width of approximately 4.8 meters. These sections were generally accompanied by the detachment of the original support components for the light rail, fishplate, and anchor bolts. Treatment involved first removing the remaining old support components, then widening the area and rebuilding with elliptical cross-section reinforced concrete support. For sections directly opposite the branch chute, further measures were required. A stronger anti-collision section (8-10 meters high) is set up and equipped with manganese steel plates and anchor bolts of different lengths. The high-damage section (sections 17-22 and 28) has a maximum depth of about 7.91 meters, a maximum width of about 7.09 meters, and a larger volume. The treatment strategy, based on the medium damage, further allows for appropriate enlargement of the local cross-section to bring the ore trajectory closer to the center of the shaft and reduce the impact on the shaft wall. In the key anti-collision section, stronger anchoring (2-meter, 3-meter, and 5-meter anchor bolts arranged in a quincunx pattern with a row spacing of 1×1 meters) is used to ensure the integrity and impact resistance.

[0112] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0113] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A point cloud-based auxiliary system for well pass repair operations, characterized in that, The system includes a point cloud image acquisition module and a data analysis module that are interconnected. The point cloud image acquisition module acquires wellbore point clouds, images, timestamps, and camera poses. The data analysis module imports the well diameter and wellbore reference from the design drawings, establishes the wellbore axis, and performs registration and mapping. It calculates roundness, eccentricity, cavity expansion rate, elevation profile gradient, branch entry azimuth angle, and reflective strip intensity according to the elevation step size and circumferential angle step size grid, and calculates the probability of the first collision zone from the azimuth angle and cavity expansion rate. The image occlusion rate is the proportion of pixels in the grid whose brightness or contrast is lower than the threshold, and the point cloud missing measurement rate is the insufficient proportion of the actual number of points in the grid relative to the expected number of points. The expected number of points is converted from the design well diameter and step size. The image quality weight is one minus the weighted sum of occlusion rate and missing detection rate, truncated to zero to one. The data analysis module processor calls the trained model to calculate and determine the target side length, structural element principal axis length, direction vector, direction weight, collision zone priority area identifier, strip suppression area identifier, marking threshold, mask threshold, propagation radius and iteration upper limit based on the above parameters. The direction vector is normalized, the direction weight is truncated to zero to one, and the iteration upper limit is rounded down. The correction field is obtained by performing morphological reconstruction and closing operations on the voxel field or distance field with mass weight scaling of the propagation radius and iteration upper limit, and constrained by the priority region and suppression region. The correction field is used to correct the morphological hole filling error of the segmented chute point cloud image. The segmented parameter output module receives the correction field and outputs the maximum damage depth, maximum damage width, and damage volume according to the elevation segment, and evaluates the damage level to obtain the ventilation parameter set.

2. The point cloud-based well pass repair operation auxiliary system according to claim 1, characterized in that, The point cloud image acquisition module includes a mounting frame, with a LiDAR scanner mounted on the inside of the frame. A camera is mounted below the LiDAR scanner, and a high-intensity flashlight facing the well wall is mounted on one side of the camera. A network bridge protection device is mounted on the top of the mounting frame, and a network bridge device is located inside the network bridge protection device. The point cloud image acquisition module communicates with the data analysis module through the network bridge device. The mounting frame is connected to a winch via a suspension rope. The module outputs a set of maximum damage depth, maximum damage width, damage volume, damage level, and ventilation parameters in segments.

3. The point cloud-based well pass repair operation auxiliary system according to claim 1, characterized in that, The data analysis module includes a design drawing import module, which is connected to a registration and mapping module. The registration and mapping module is connected to a chute feature extraction module, which is connected to a first recommendation module, a second recommendation module, and a third recommendation module. The first recommendation module, the second recommendation module, and the third recommendation module are all connected to a reconstruction and correction module. The reconstruction and correction module is connected to a segmented parameter output module. The registration and mapping module is also connected to a field construction module, which is connected to the reconstruction and correction module.

4. The point cloud-based well pass repair operation auxiliary system according to claim 3, characterized in that, In the data analysis module, the registration and mapping module establishes the wellbore axis based on the well diameter and well wall reference output by the design drawing import module and maps the point cloud to the elevation and circumferential angular coordinate system; the well pass feature extraction module generates point cloud feature parameters of segmented wells based on the registration and mapping results. The point cloud feature parameters of segmented wells include roundness, eccentricity, cavity expansion rate, elevation profile gradient, branch entry azimuth, reflective strip intensity, and empty inclined chute discharge geometric factor. It also outputs the probability of the first collision zone and image quality weight by analyzing the point cloud feature parameters of segmented wells; the first recommendation module receives the roundness, eccentricity, cavity expansion rate, and elevation profile gradient output by the well pass feature extraction module, and outputs the initial values ​​of the target side length of the triangular mesh unit and the principal axis length of the structural element; the second recommendation module receives the branch entry azimuth, cavity expansion rate, and probability of the first collision zone, and outputs the structural element direction vector, direction weight, and collision zone priority area identifier and stripe suppression area identifier. The third recommendation module receives the image quality weight, the intensity of the reflective stripe and the empty inclined chute discharge geometry factor, and outputs the marker field threshold, the mask field threshold, the reconstruction propagation radius and the iteration upper limit; The field construction module receives the point cloud output by the registration and mapping module and generates a voxel field or a range field; the reconstruction and correction module receives the voxel field or range field generated by the field construction module and the output parameters of the first recommendation module, the second recommendation module, and the third recommendation module, and performs a reconstruction closing operation under the constraints of image quality weight, collision zone priority region identifier, and stripe suppression region identifier to obtain the correction field.

5. The point cloud-based well pass repair operation auxiliary system according to claim 4, characterized in that, The first recommendation module uses a constrained gradient boosting regression tree model for recommendations. The model inputs are roundness, eccentricity, cavity expansion ratio, elevation profile gradient, and point density. The model outputs the initial values ​​of the target side length of the triangular mesh unit and the three principal axis lengths of the structural element. The model uses the well diameter in the design drawing as the upper bound of the scale and the local point spacing as the lower bound of the scale. Monotonicity constraints are set on the target side length and the three principal axis lengths. The monotonicity constraints are set according to the coupling scale index of the cavity expansion ratio and the point spacing. The coupling scale index is the cavity expansion ratio multiplied by the point spacing. In the gradient boosting regression tree, monotonically increasing constraints are applied to the cavity expansion ratio, point spacing, and coupling scale index simultaneously, so that the recommended target side length of the triangular mesh and the three principal axis lengths of the structural element increase with the increase of the cavity expansion ratio and the point spacing, and do not decrease with the increase of the coupling scale index.

6. The point cloud-based well pass repair operation auxiliary system according to claim 4, characterized in that, The second recommendation module uses a graph attention network model oriented towards a circumferential loop topology for recommendation. It constructs a circumferential loop graph by using circumferential sectors at the same elevation as graph nodes and adjacent circumferential sectors and the first and last sectors as graph edges. The model inputs are the branch entry azimuth angle, cavity expansion rate, probability of the first collision zone, intensity of the reflective strip, and geometric feature vectors of the circumferential sector. The model uses a dual-channel periodic representation of the branch entry azimuth angle with sine and cosine encoding, sets attention bias weights for the probability of the first collision zone, and sets a propagation suppression mask for the intensity of the reflective strip. The model outputs are the structuring element direction vector, direction weights, collision zone priority area identifier, and strip suppression area identifier. The structuring element direction vector is subject to a unit length normalization constraint, and the direction weight is subject to an interval mapping constraint.

7. The point cloud-based well pass repair operation auxiliary system according to claim 4, characterized in that, The third recommendation module uses a quantile regression neural network model with uncertainty for recommendation. The model input includes image occlusion rate, point cloud missing rate, image quality weight, vacant inclined chute discharge geometric factor, reflective strip intensity, collision zone priority area identifier and strip suppression area identifier. The model output includes the label field threshold, mask field threshold, reconstruction propagation radius and iteration upper limit. The model applies upper and lower bound interval mapping constraints to the label field threshold and mask field threshold, applies non-negativity constraints to the reconstruction propagation radius, and applies integerization and upper limit truncation constraints to the iteration upper limit. During training, the image quality weight is used as the sample weight to participate in the loss calculation.

8. The point cloud-based well pass repair operation auxiliary system according to claim 4, characterized in that, The segmented parameter output module determines a low-damage segment as one that simultaneously meets the following criteria: maximum damage depth ≤ first depth threshold, maximum damage width ≤ first width threshold, damage volume ≤ first volume threshold, first collision zone probability ≤ first collision threshold, discharge geometric factor ≤ first discharge threshold, and image quality weight ≤ first quality threshold. It outputs the target air volume as the number of workers multiplied by the minimum air supply constant per unit number of workers. The target negative pressure is based on the baseline negative pressure. The dimensionless negative pressure demand index, obtained by combining the first collision zone probability and discharge geometric factor according to weights, is converted into a negative pressure correction amount using a preset negative pressure conversion coefficient and then added to the baseline negative pressure. The allowable wind speed range is the target air volume divided by the effective cross-section with an upper and lower deviation coefficient applied. The dust extraction air volume is the target air volume multiplied by the first dust extraction ratio coefficient, which is a dimensionless coefficient corresponding to the low-damage segment and stored in the low-damage segment dust extraction configuration mapping table. The dustproof door sequence includes the pre-closing duration, unloading holding duration, and delayed dust extraction duration.

9. A point cloud-based auxiliary system for well pass repair operations according to claim 4, characterized in that, The segmented parameter output module determines a segment as a medium-damage segment if the elevation segment meets the following criteria: maximum damage depth between the first and second depth thresholds; maximum damage width between the first and second width thresholds; damage volume between the first and second volume thresholds; collision zone probability between the first and second collision thresholds; and image quality weight between the second and first quality thresholds. It then uses a machine learning model to predict ventilation and negative pressure correction coefficients based on cavity expansion rate, elevation profile gradient, collision zone probability, and reflective strip intensity. The output target airflow is the target airflow of the medium-damage segment multiplied by the ventilation correction coefficient. The target negative pressure is the target negative pressure of the low-damage segment multiplied by the negative pressure correction coefficient. The allowable wind speed range is corrected using a speed limit coefficient and a minimum guarantee coefficient. The dust extraction airflow is the target airflow multiplied by a second dust extraction ratio coefficient, which is a dimensionless coefficient corresponding to the low-damage segment and stored in the dust extraction configuration mapping table of the medium-damage segment. The module also includes the dustproof door timing pre-closing duration, ore unloading holding duration, delayed dust extraction duration, closure confirmation duration, and release delay duration.

10. A point cloud-based auxiliary system for well pass repair operations according to claim 4, characterized in that, The segmented parameter output module determines a high-damage segment as one that meets any of the following conditions: maximum damage depth > second depth threshold, maximum damage width > second width threshold, damage volume > second volume threshold, first collision zone probability > second collision threshold, discharge geometry factor > second discharge threshold, or image quality weight < second quality threshold. Based on the start and end times of the unloading event and the wellhead's sealing status, the output parameters for high-damage segments are divided into: Pre-unloading stage outputs pre-ventilation air volume, pre-negative pressure, pre-velocity range, pre-dust extraction air volume, and pre-door closing sequence; During unloading, outputs increased operating air volume, enhanced negative pressure, dust extraction air volume, and speed-limited velocity range, along with the wellhead's dustproof door remaining closed; Post-unloading stage outputs decreased and stabilized air volume, negative pressure, velocity range, delayed dust extraction air volume, and delayed door opening sequence, with minimum hold-time constraints applied during parameter switching across all three stages.

Images