A belt tearing detection method and device based on three-dimensional vision

Abstract

The present application provides a belt tear detection method and device based on three-dimensional vision. Among them, the method includes: calculating the spatial distance values between each target point and adjacent points in the three-dimensional point cloud data in the direction perpendicular to the tangent of the laser arc, screening the points whose spatial distance values exceed a preset distance threshold to form an initial candidate point set, starting from each point, expanding the space through the breadth-first search algorithm to obtain a target candidate point set, and calculating the length and depth of each connected component; taking the connected components that meet the threshold conditions as the corresponding potential cracks, forming a potential crack set with the spatial position information of all potential cracks, and generating a belt tear detection result including the spatial position information of the real cracks and the corresponding alarm level. The technical solution provided by the present application realizes high-precision dynamic identification and alarm of cracks on the belt surface, effectively reduces the false detection rate, and improves the efficiency of industrial safety monitoring.

Description

Technical Field

[0001] This application relates to the technical field of belt anomaly detection, and particularly to a belt tear detection method and device based on three-dimensional vision. Background Art

[0002] In steel production, longitudinal tears often occur in material conveyor belts (i.e., belts). Once a longitudinal tear accident occurs in a material conveyor belt, it will be completely damaged in a very short time, causing huge economic losses. Even if it can be repaired, it requires a considerable amount of manpower and time, which has a great impact on normal production. In recent years, the usage of material conveyor belts has been increasing, and their application scope has been expanding. However, so far, there is no relatively ideal and mature method and device for effectively detecting longitudinal tears that can be widely promoted.

[0003] Existing solutions are based on two-dimensional vision technology to achieve automatic detection of longitudinal tears in material conveyor belts, which can replace manual detection, avoid potential personnel safety accidents, and improve the safe operation level of conveyor belts.

[0004] However, existing solutions based on two-dimensional vision methods are limited by the extraction of planar information from a single perspective, and it is difficult to effectively distinguish the three-dimensional morphological differences between stains, shadows, etc. on the belt surface and real cracks. For example, when there is reflection or dust coverage on the belt surface, pseudo-edges are easily generated in two-dimensional image features, resulting in an increase in the false detection rate. At the same time, due to the lack of depth information, the actual depth of the crack cannot be accurately quantified, and potential tear risks that are small but have depth may be missed. In addition, the two-dimensional threshold setting is sensitive to lighting conditions and lacks robustness in complex industrial environments, and frequent manual parameter calibration is required to maintain detection accuracy. Summary of the Invention

[0005] This application provides a belt tear detection method and device based on three-dimensional vision to solve the problem of low accuracy of belt tear detection in the prior art due to the difficulty in effectively distinguishing real cracks from surface interference, the inability to quantify the depth of cracks, and insufficient environmental robustness.

[0006] In a first aspect, this application provides a belt tear detection method based on three-dimensional vision, including:

[0007] Collecting images of the belt surface through a line laser and an industrial camera deployed below the belt, and calculating the three-dimensional point cloud data of the laser arc in the belt surface image;

[0008] Calculating the spatial distance values between each target point and its adjacent points in the three-dimensional point cloud data in the direction perpendicular to the tangent of the laser arc, screening the points with spatial distance values exceeding a preset distance threshold, and forming an initial candidate point set;

[0009] Starting from each point in the initial candidate point set, points that are spatially adjacent and belong to the initial candidate point set are expanded through the Breadth-First Search (BFS) algorithm to obtain a target candidate point set, so as to form multiple non-connected connected domains, and the length and depth of each connected domain are calculated;

[0010] Filter out the connected domains that do not meet the threshold conditions among all the connected domains, retain the connected domains that meet the threshold conditions, regard all the connected domains that meet the threshold conditions as the corresponding potential cracks, and form a potential crack set with the spatial position information of all the potential cracks. The non-compliance with the threshold conditions means that both the length and the depth do not reach the corresponding preset thresholds;

[0011] During the belt cyclic operation, if there is a real crack, a belt tearing detection result including the spatial position information of the real crack and the corresponding alarm level is generated. The real crack is a potential crack at the same spatial position that is recorded in the potential crack set in multiple consecutive operation cycles.

[0012] Optionally, the starting from each point in the initial candidate point set, expanding points that are spatially adjacent and belong to the initial candidate point set through the breadth-first search algorithm to obtain a target candidate point set includes:

[0013] Traverse each target point in the initial candidate point set, mark the target point as a member of the current connected domain, and insert it into a preset queue to be viewed. According to the belt running direction and the belt running speed, calculate the dynamic neighborhood range of the current target point;

[0014] Search for adjacent points that are spatially adjacent and belong to the initial candidate point set along the direction perpendicular to the laser arc tangent line within the dynamic neighborhood range;

[0015] Judge whether the adjacent point of the target point belongs to the initial candidate point set and has not been marked;

[0016] If so, mark the adjacent point as a member of the current connected domain, insert it into the end of the preset queue to be viewed, regard the adjacent point as a new target point, and repeat the steps of marking, inserting, calculating, searching, and judging until all the target points in the preset queue to be viewed form a connected domain, empty the preset queue to be viewed, and repeat the formation process of the new connected domain until each target point in the initial candidate point set is marked;

[0017] Delete all the target points in the initial candidate point set that are not in any connected domain to obtain a target candidate point set.

[0018] Optionally, the determination condition for spatial adjacency is that the three-dimensional Euclidean distance between an adjacent point and the target point is less than a preset connectivity distance threshold, and the preset connectivity distance threshold is dynamically set according to the mapping relationship between the spot diameter of the line laser calibrated by experiments and the belt running speed;

[0019] The adjacent points that are spatially adjacent and belong to the initial candidate point set include:

[0020] Calculate the tangent direction vector of the laser arc of the target point;

[0021] Generate a vector perpendicular to the tangent direction vector of the laser arc according to the tangent direction vector of the laser arc and the unit vector in the belt thickness direction;

[0022] Extend a preset connectivity distance threshold in both the positive and negative directions of the tangent direction vector of the laser arc to form a strip-shaped search area, and determine the points other than the target point in the strip-shaped search area as adjacent points.

[0023] Optionally, the calculation of the spatial distance value between each target point and the adjacent point in the three-dimensional point cloud data in the direction perpendicular to the tangent direction of the laser arc, and the screening of the points whose spatial distance value exceeds the preset distance threshold to form the initial candidate point set includes:

[0024] Obtain the cross-sectional data of the multi-layer composite structure through a preset belt cross-sectional scanning device, and establish an association model between the belt material and the belt thickness;

[0025] Dynamically generate a layer distance threshold for each hierarchical structure where the target point is located according to the association model;

[0026] Calculate the spatial distance difference between the target point and the adjacent points in the anisotropic neighborhood in the belt thickness direction; compare the spatial distance difference with the layer distance threshold, screen out the abnormal points whose spatial distance difference exceeds the layer distance threshold, and determine the maximum penetration depth and layer position information of each abnormal point;

[0027] Construct a density field based on the spatial distribution characteristics of the abnormal points, and calculate the spatial aggregation degree of each abnormal point in the density field within a preset volume element;

[0028] Determine the product result of the spatial aggregation degree and the maximum penetration depth as the comprehensive abnormality score, and screen out the abnormal points whose comprehensive abnormality score exceeds the preset score threshold to form the initial candidate point set.

[0029] Optionally, the determination of the maximum penetration depth and layer position information of each abnormal point includes:

[0030] According to the spatial position information of the abnormal points and in combination with the geometric model of the layering boundary, calculate the maximum deviation amount of each abnormal point from the layering boundary to generate the maximum penetration depth.

[0031] In combination with the layering index table, perform layering position parsing on the spatial position information of the abnormal points to obtain the layering position information.

[0032] Optionally, the forming the set of spatial position information of all potential cracks includes:

[0033] Extract crack feature points according to the spatial positions of all the potential cracks and form an initial set of potential crack points.

[0034] Perform spatial clustering on the initial set of potential crack points and merge adjacent points to form potential crack segments.

[0035] Calculate the geometric features of all the potential crack segments, analyze the adjacency and intersection relationships between the crack segments, and construct a crack topology graph.

[0036] Perform global optimization based on geometric continuity and the crack topology graph, and eliminate low-confidence crack segments to obtain the set of potential cracks.

[0037] Optionally, the calculating the length and depth of each connected domain includes:

[0038] Analyze the spatial positions and boundary information of all the connected domains to generate an initial description information set for each connected domain.

[0039] According to the initial description information set and a preset spatial geometric relationship model, perform boundary extraction for each connected domain to obtain the coordinate of all boundary points of each connected domain.

[0040] According to the coordinates of all boundary points of each connected domain, analyze the relative position relationship between all boundary points of the connected domain, determine two first endpoints corresponding to the longest path, and use the distance value between the two first endpoints as the length of the connected domain.

[0041] According to the coordinates of all boundary points of each connected domain, analyze the distance distribution of all boundary points of the connected domain from a preset center point of the connected domain, determine two second endpoints perpendicular to the longest path, and use the distance value between the two second endpoints as the depth of the connected domain.

[0042] In a second aspect, the present application provides a belt tearing detection device based on three-dimensional vision, including:

[0043] An acquisition module, configured to collect a belt surface image through a line laser and an industrial camera deployed below the belt, and calculate three-dimensional point cloud data of a laser arc in the belt surface image.

[0044] A screening module, configured to calculate the spatial distance values between each target point and its adjacent points in the three-dimensional point cloud data in the direction perpendicular to the tangent of the laser arc, and screen out the points whose spatial distance values exceed a preset distance threshold, so as to form an initial candidate point set;

[0045] A calculation module, configured to start from each point in the initial candidate point set, expand the points that are spatially adjacent and belong to the initial candidate point set through a breadth-first search algorithm, obtain a target candidate point set, so as to form a plurality of non-connected connected domains, and calculate the length and depth of each connected domain;

[0046] A filtering module, configured to filter out the connected domains that do not meet the threshold conditions among all the connected domains, retain the connected domains that meet the threshold conditions, regard all the connected domains that meet the threshold conditions as corresponding potential cracks, and form a potential crack set with the spatial position information of all the potential cracks. The non-compliance with the threshold conditions means that both the length and the depth do not reach the corresponding preset thresholds;

[0047] A generation module, configured to generate a belt tear detection result including the spatial position information of a real crack and the corresponding alarm level during the belt cyclic operation, where the real crack is a potential crack at the same spatial position that is recorded in the potential crack set in multiple consecutive operation cycles.

[0048] In a third aspect, the present application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a three-dimensional vision-based belt tear detection method according to any one of the first aspect.

[0049] In a fourth aspect, the present application provides a computer storage medium storing a computer program, and when the computer program is executed by a computer, it implements a three-dimensional vision-based belt tear detection method according to any one of the first aspect.

[0050] In an embodiment of the present application, a belt tear detection method based on three-dimensional vision is provided. The method includes: collecting an image of the belt surface through a line laser and an industrial camera deployed below the belt, and calculating the three-dimensional point cloud data of the laser arc in the belt surface image; calculating the spatial distance values between each target point and adjacent points in the three-dimensional point cloud data in the direction perpendicular to the tangent of the laser arc, screening out the points whose spatial distance values exceed a preset distance threshold, and forming an initial candidate point set; starting from each point in the initial candidate point set, expanding the points that are spatially adjacent and belong to the initial candidate point set through the breadth-first search algorithm to obtain a target candidate point set, so as to form multiple non-connected connected domains, and calculating the length and depth of each connected domain; filtering out the connected domains that do not meet the threshold conditions among all the connected domains, retaining the connected domains that meet the threshold conditions, taking all the connected domains that meet the threshold conditions as corresponding potential cracks, and forming a potential crack set with the spatial position information of all the potential cracks. The condition that does not meet the threshold is that both the length and the depth do not reach the corresponding preset thresholds; during the cyclic operation of the belt, if there is a real crack, a belt tear detection result including the spatial position information of the real crack and the corresponding alarm level is generated. The real crack is a potential crack at the same spatial position that is recorded in the potential crack set in multiple consecutive operation cycles.

[0051] Based on the three-dimensional point cloud data of the laser arc jointly collected by the line laser and the industrial camera, the present application breaks through the dependence of traditional two-dimensional vision technology on plane information, can accurately characterize the depth change of the belt surface topography, and improves the detection sensitivity of the three-dimensional spatial characteristics of micro-cracks (especially longitudinal tears). By calculating the spatial distance between adjacent points in the three-dimensional point cloud in the direction perpendicular to the tangent of the laser arc and combining the preset distance threshold to screen candidate points, it effectively distinguishes real cracks from plane interferences such as surface stains and reflections, reduces the false detection rate caused by environmental noise, and improves the accuracy of the belt tear detection result. The breadth-first search algorithm is used to generate non-connected connected domains, and combined with the double threshold filtering of length and depth, the pseudo-crack areas formed by isolated noise points are accurately removed; at the same time, the potential cracks at the same spatial position are verified through multi-cycle continuous detection to avoid false alarms caused by instantaneous interference, ensuring the reliability and stability of the detection result. By generating the crack spatial position and alarm level information in real time, it provides accurate decision-making support for the operation and maintenance of the belt, meeting the requirements of unmanned and high-efficiency tear risk warning in industrial scenarios.

[0052] Furthermore, this method realizes the accurate extraction of crack connected components through dynamic neighborhood range calculation and breadth-first search algorithm. Specifically: First, based on the running direction and speed of the belt, the dynamic neighborhood range of the target point is determined, and spatially adjacent points that meet the three-dimensional Euclidean distance threshold are searched along the direction perpendicular to the tangent of the laser arc; through the processes of iterative marking, queue insertion, and connected component expansion, combined with the limitation of the strip-shaped search area, the initial candidate points are clustered into unconnected connected components, and finally the isolated points are removed to form a set of target candidate points. This solution designs a dynamic neighborhood and an adaptive connected distance threshold, solves the problems of regional fracture or over-connection caused by the change of the belt movement speed in traditional connected component search, and ensures the integrity and continuity of the crack shape; combines the strip-shaped search strategy and three-dimensional Euclidean distance constraint to accurately distinguish real cracks from discrete noise points, and avoids misjudgment caused by local point cloud anomalies; dynamically optimizes the search range through the parameter mapping of the spot diameter and the belt speed, improves the adaptability of the algorithm to complex working conditions, and enhances the recognition robustness of the crack spatial topology structure.

[0053] These aspects or other aspects of the present application will be more clearly understood in the following description of the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application or in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0055] Figure 1 It is a flowchart of a belt tear detection method based on three-dimensional vision provided by an embodiment of the present application;

[0056] Figure 2 It is a schematic diagram of the positions of a line laser and a camera provided by an embodiment of the present application;

[0057] Figure 3 It is a schematic diagram of the structure of a belt tear detection device based on three-dimensional vision provided by an embodiment of the present application;

[0058] Figure 4 It is a schematic diagram of the structure of a computing device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] In order to enable those skilled in the art to better understand the solutions of the present application, the following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application.

[0060] In some processes described in the specification, claims, and the above-mentioned drawings of the present application, a plurality of operations appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order in which they appear herein or may be executed in parallel. The sequence numbers of the operations, such as 11, 12, etc., are only used to distinguish different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions such as "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent a sequence, and do not limit that "first" and "second" are of different types.

[0061] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0062] To solve the problem of low accuracy of belt tear detection in the prior art due to the difficulty in effectively distinguishing real cracks from surface interference, the inability to quantify the crack depth, and insufficient environmental robustness, the embodiments of the present application provide a belt tear detection method based on three-dimensional vision. This method constructs a detection framework with three-dimensional point cloud data as the core and adopts the following concept: First, a line laser and an industrial camera are used in cooperation to collect images of the belt surface, and then the three-dimensional point cloud data of the laser arc is calculated, breaking through the dependence of two-dimensional vision technology on planar features; based on the spatial mutation characteristics of cracks in the direction perpendicular to the tangent of the laser arc, a set of candidate points is screened to initially eliminate planar interference; further, combined with dynamic neighborhood breadth-first search and three-dimensional connected component analysis, the continuous crack morphology is extracted from the dimension of the spatial topological structure, and isolated noise is filtered through double thresholds of length and depth; finally, the detection results of multiple cycles are fused to verify the crack persistence, ensuring the anti-interference and reliability of the detection results in a dynamic scenario, and forming a closed-loop detection logic from three-dimensional feature extraction to dynamic verification.

[0063] Figure 1 The flowchart of a belt tear detection method based on three-dimensional vision provided by the embodiments of the present application is as Figure 1 shown, and this method includes:

[0064] S11. Use a line laser and an industrial camera deployed under the belt to collect images of the belt surface, and calculate the three-dimensional point cloud data of the laser arc in the belt surface image.

[0065] Among them, a line laser is a device that emits a linear laser beam and is used to form measurable light stripes on the surface of an object. An industrial camera belongs to high-precision imaging equipment and is used to capture laser arc images and transmit them to a processing system. Exemplarily, the industrial camera can be a high-speed industrial camera. The three-dimensional point cloud data is a set of discrete points composed of three-dimensional space coordinates, which characterizes the surface topography of the belt. Calculating the three-dimensional point cloud data of the laser arc in the belt surface image can generate the three-dimensional point cloud data based on the baseline distance between the camera and the laser, the angle between the laser ray and the baseline, and the angle between the camera and the baseline through the principle of optical triangulation.

[0066] In the embodiment of the present application, a line laser deployed under a high-speed running conveyor belt emits a laser beam to form a laser arc on the surface of the belt. At the same time, a high-speed industrial camera synchronously captures the arc image at a preset angle (that is, a fixed angle is formed between the high-speed industrial camera and the laser baseline). Based on the principle of optical triangulation, combined with the known parameters of the baseline distance between the camera center and the laser center, the angle between the laser ray and the baseline, and the angle between the camera and the baseline, the three-dimensional coordinates of the laser arc in each frame of the image are calculated. The specific process is as follows: Through the geometric relationship between the laser plane equation and the camera projection model, the two-dimensional coordinates of each pixel point in the belt surface image are converted into three-dimensional space coordinates, and finally the three-dimensional point cloud data of the belt surface is generated.

[0067] S12. Calculate the spatial distance value between each target point and its adjacent point in the three-dimensional point cloud data in the direction perpendicular to the tangent of the laser arc, and filter out the points whose spatial distance value exceeds the preset distance threshold to form an initial candidate point set.

[0068] Among them, the tangent direction of the laser arc is the vector direction of the local geometric tangent of the laser arc, which is used to define the normal plane. The spatial distance value is the Euclidean distance between two points in the normal plane. Each target point in the three-dimensional point cloud data is the current point. The initial candidate point set is a candidate point set formed by calculating and filtering the spatial distance value between each target point and its adjacent point in the three-dimensional point cloud data in the direction perpendicular to the tangent of the laser arc.

[0069] In the embodiment of the present application, for each point in the three-dimensional point cloud data, its spatial feature (X vector) is calculated, and the spatial feature is used to reflect the spatial distance difference between the point and its surrounding adjacent points in the direction perpendicular to the tangent of the laser arc. The specific process is as follows: First, the local tangent direction of the laser arc is fitted through the coordinate difference of adjacent points to determine the normal plane perpendicular to the tangent direction; then the Euclidean distance between the target point and the adjacent points in this normal plane is calculated. If the distance difference exceeds the preset threshold (such as 1 mm), the point is marked as a candidate point. All candidate points that meet the conditions form set A (the initial candidate point set), where set A may contain crack points or noise points.

[0070] S13. Starting from each point in the initial candidate point set, expand the points that are spatially adjacent and belong to the initial candidate point set through the breadth-first search algorithm to obtain the target candidate point set, so as to form multiple non-connected connected domains, and calculate the length and depth of each connected domain.

[0071] Among them, the breadth-first search algorithm is a graph traversal algorithm that expands adjacent nodes by level to discover connected regions. A connected domain refers to an independent region composed of interconnected points in space. The target candidate point set is a set of connected region points obtained by expanding the initial candidate point set through the breadth-first search algorithm. Exemplarily, the formula for calculating the length of a connected domain is: , where L is the length of the connected domain, is the maximum and minimum x-axis coordinates of the projections of all points in the connected domain on the belt surface, is the maximum and minimum y-axis coordinates of the projections of all points in the connected domain on the belt surface. Exemplarily, the formula for calculating the depth of a connected domain is: , where D is the depth of the connected domain, is the height value of each point in the connected domain, is the reference height of the belt surface.

[0072] In the embodiment of the present application, for each point in the initial candidate point set, the breadth-first search algorithm is used to expand the connected domain: starting from the current point, check whether its spatially adjacent points belong to the initial candidate point set. If so, classify them into the same connected domain, and recursively search for the adjacent points of the adjacent points until no new candidate points can be found. All non-connected connected domains constitute the target candidate point set, and each connected domain represents an independent suspected crack region.

[0073] S14. Filter out the connected domains that do not meet the threshold conditions among all connected domains, retain the connected domains that meet the threshold conditions, use all the connected domains that meet the threshold conditions as the corresponding potential cracks, and form a potential crack set with the spatial position information of all potential cracks. Not meeting the threshold conditions means that both the length and depth do not reach the corresponding preset thresholds.

[0074] Among them, the length threshold generally refers to the major axis length or contour perimeter of the connected component. Cracks usually exhibit an elongated shape, and connected components that are too short may be noise rather than real cracks. The depth threshold refers to the gray value difference or actual three-dimensional depth information. Connected components with insufficient depth may indicate low contrast or shallow cracks, making it difficult to determine them as valid cracks. Only when both the length and depth of the connected component do not reach the threshold are they determined to be "not eligible". This means that the retention condition is length ≥ length threshold or depth ≥ depth threshold, and either one is sufficient. The filtering condition is length < length threshold and depth < depth threshold, and when both are not satisfied, it means not meeting the threshold condition. The spatial position of the potential crack is the set of coordinates of the detected crack connected component in the image or three-dimensional space.

[0075] In the embodiment of the present application, threshold screening is performed on each connected component in the target candidate point set, and the screening process can be as follows: Calculate the length (i.e., the straight-line distance through its farthest two endpoints) and depth of the connected component. If both the length and depth do not reach the preset threshold, it is determined as noise and eliminated; otherwise, it is retained as a potential crack, and its spatial position information is stored in the potential crack set.

[0076] S15. During the cyclic operation of the belt, if there is a real crack, a belt tear detection result including the spatial position information of the real crack and the corresponding alarm level is generated. The real crack is a potential crack at the same spatial position that has been recorded in the potential crack set for multiple consecutive operating cycles.

[0077] Among them, the real crack is a potential crack that appears repeatedly across cycles and has passed the spatio-temporal consistency verification. The alarm level is the severity of the alarm divided according to the crack depth (such as low risk, medium risk, and high risk). The finally generated spatial position information of the real crack can be in the following two forms: Merge the three-dimensional point cloud data of the crack detected in all cycles to form a spatio-temporal fusion point cloud, reflecting the complete morphology and expansion trend of the crack. Take the weighted average of the crack center coordinates in all cycles to ensure that the position information reflects the spatial position of the crack in real time.

[0078] In the embodiment of the present invention, combined with the cyclic operation characteristics of the conveyor belt, cross-cycle verification is performed on the potential crack set. If a crack at a certain spatial position is detected in multiple consecutive operating cycles, it is determined as a real crack. An alarm level is generated according to the crack depth, and a belt tear detection result including the position, size, and alarm level is output.

[0079] The following is a specific example:

[0080] In the scenario of steel belt transportation, a line laser projects a green laser line onto the running belt, and an industrial camera captures images at a rate of 30 frames per second. The system calculates the three-dimensional point cloud data in real time and finds that the distance of a certain point cloud in the normal plane suddenly changes to 3 mm, which is marked as an initial candidate point. The point is expanded through the breadth-first search algorithm to form a connected domain with a length of 15 cm and a depth of 4 mm. After filtering, this connected domain is retained as a potential crack due to exceeding the threshold. In the subsequent three belt cycles, the system detects this crack at the same coordinates, and finally triggers a high-risk alarm, prompting for shutdown and maintenance.

[0081] By executing steps S11 - S15, the embodiment of the present application improves the accuracy and robustness of belt tear detection through three-dimensional point cloud analysis combined with multi-cycle verification: the line laser and the camera cooperate to achieve high-resolution topography reconstruction; through spatial distance screening and connected domain extraction, real cracks and instantaneous noise can be effectively distinguished; the cross-cycle matching mechanism further reduces the false alarm rate and ensures the credibility of the alarm result.

[0082] Specifically, based on the three-dimensional point cloud data of the laser arc collected by the cooperation of the line laser and the industrial camera, the present application breaks through the dependence of traditional two-dimensional vision technology on plane information, can accurately characterize the depth change of the belt surface topography, and improves the detection sensitivity of the three-dimensional spatial characteristics of micro-cracks (especially longitudinal tears). By calculating the spatial distance between adjacent points in the three-dimensional point cloud in the direction perpendicular to the tangent of the laser arc and combining a preset distance threshold to screen candidate points, real cracks can be effectively distinguished from plane interferences such as surface stains and reflections, reducing the false detection rate caused by environmental noise and improving the accuracy of the belt tear detection result. Using the breadth-first search algorithm to generate non-connected connected domains and filtering with double thresholds of length and depth, pseudo-crack regions formed by isolated noise points can be accurately removed; at the same time, by continuously detecting and verifying potential cracks at the same spatial position in multiple cycles, false alarms caused by instantaneous interference are avoided, ensuring the reliability and stability of the detection result. By generating the spatial position and alarm level information of the crack in real time, precise decision-making support is provided for the operation and maintenance of the belt, meeting the requirements of unmanned and high-efficiency tear risk warning in industrial scenarios.

[0083] In a possible embodiment, S13: Starting from each point in the initial candidate point set, expand the points that are spatially adjacent and belong to the initial candidate point set through the breadth-first search algorithm to obtain a target candidate point set, including:

[0084] Step 131: Traverse each target point in the initial candidate point set, mark the target point as a member of the current connected domain, and insert it into a preset queue to be viewed. Calculate the dynamic neighborhood range of the current target point according to the belt running direction and the belt running speed.

[0085] Among them, the search distance extended along the belt running direction is greater than that perpendicular to the tangent direction of the laser arc. The extended distance of the dynamic neighborhood range in the belt running direction is: L = v ×Δ t + L 0, where v is the real-time speed of the belt, Δ t is the image acquisition period, L 0 is the reference neighborhood length.

[0086] In the embodiment of the present application, the target point is marked as a member of the current connected domain through a specified field corresponding to the current connected domain. For example, the marking information of the target point A1 is the specified field LTY1, indicating that the target point A1 is a member of the connected domain 1; the marking information of the target point A2 is the specified field LTY2, indicating that the target point A2 is a member of the connected domain 2. The extended length of the dynamic neighborhood in the belt running direction is the speed multiplied by the time window, and the lateral width is fixed as a preset value, forming a dynamic rectangular search area. This step ensures the spatio-temporal consistency of the neighborhood search by dynamically adjusting the neighborhood range to adapt to the spatial displacement of candidate points caused by belt movement.

[0087] Step 132: Search for adjacent points that are spatially adjacent and belong to the initial candidate point set along the direction perpendicular to the tangent of the laser arc within the dynamic neighborhood range.

[0088] Among them, the judgment condition for spatial adjacency is that the Euclidean distance in three-dimensional space between the adjacent point and the current point is less than the preset connected distance threshold, and the point whose three-dimensional distance from the current point within the dynamic neighborhood is less than the resolution threshold. The initial candidate point set is a set of points with a spatial distance exceeding the threshold screened out, including potential crack points and noise points. The belt running direction is the direction in which the belt conveys materials, usually measured by an encoder or a speed sensor. The tangent direction of the laser arc is the local tangent direction of the laser line on the belt surface, calculated by the coordinate difference of adjacent points.

[0089] In the embodiment of the present invention, the specific process of step 132 is: centered on the current target point, divide the dynamic neighborhood into three-dimensional grids, traverse all grid points, and if there is an initial candidate point in the three-dimensional point cloud data corresponding to a certain grid point, it is determined as a spatially adjacent point. This step reduces the interference in non-related directions by constraining the search direction and range and focusing on the deformation characteristics in the vertical direction of the crack.

[0090] Step 133: Determine whether the adjacent point of the target point belongs to the initial candidate point set and has not been marked.

[0091] In an embodiment of the present invention, the specific implementation of step 133 is as follows: Check whether the unique identifier of the adjacent point (such as the coordinate hash value) exists in the hash table of the initial candidate point set, and query the marking status table to confirm that it is not marked as a member of any connected domain. If the conditions are met, proceed to step 134; otherwise, skip this adjacent point and continue to search for other neighboring points.

[0092] Step 134: If so, mark the adjacent point as a member of the current connected domain, insert it at the end of the preset queue to be viewed, use the adjacent point as the new target point, and repeat the steps of marking, inserting, calculating, searching, and judging until all the target points in the preset queue to be viewed form a connected domain. Then empty the preset queue to be viewed, and repeat the process of forming a new connected domain until each target point in the initial candidate point set is marked.

[0093] Among them, the preset queue to be viewed is a first-in-first-out data structure for temporarily storing points of the connected domain to be expanded. The marking status table is a hash table that records whether each point has been assigned to a certain connected domain.

[0094] In an embodiment of the present invention, if the adjacent point meets the conditions, mark it as a member of the current connected domain, insert it at the end of the preset queue to be viewed, and use the adjacent point as the new target point. Until the preset queue to be viewed is empty, indicating that the expansion of the current connected domain is completed. Then empty the queue, select the next unmarked point from the initial candidate point set, and repeat the above process until all points are marked. This process ensures the integrity of the connected domain through iterative expansion.

[0095] Step 135: Delete all target points in the initial candidate point set that are not in any connected domain to obtain the target candidate point set.

[0096] Among them, the target candidate point set is the candidate point set retained after the expansion of the connected domain, excluding isolated noise points.

[0097] In an embodiment of the present invention, the specific operation of step 135 is as follows: Traverse the initial candidate point set. If the marking status of a certain point is "unassigned", remove it from the set. The finally retained points form the target candidate point set, and each connected domain thereof represents a suspected crack area for subsequent threshold filtering and cross-cycle verification.

[0098] The following is a specific example:

[0099] In the iron ore belt conveyor system, when the belt runs at a speed of 2 m / s, the system detects the initial candidate points. According to the belt running speed of 2 m / s and the time window of 0.1 second, the dynamic neighborhood range of this point is calculated as [4.8 m, 5.2 m] in the x direction and fixed as [0.15 m, 0.25 m] in the y direction. It is marked as a member of the current connected domain and inserted into the queue to be viewed. Search is carried out in the dynamic neighborhood along the normal plane direction (y-axis), and 3 adjacent points belonging to the initial candidate point set are found. After verifying that none of the 3 points are marked, it enters the expansion process. The 3 points are marked and inserted into the end of the queue, and recursive expansion is carried out with the new points as the targets. Finally, a connected domain containing 12 points is formed. After emptying the queue, the remaining unmarked points are traversed, and 20 unmarked isolated points in the initial candidate point set are deleted, and 3 connected domains are retained to form the target candidate point set. In subsequent processing, one connected domain triggers a high-risk alarm due to exceeding the threshold, and the others are filtered. The system accurately locates the crack position and triggers shutdown for maintenance to avoid belt tearing accidents.

[0100] By executing steps 131 to 135, the embodiments of the present application improve the robustness and accuracy of crack detection through the dynamic neighborhood range and the directional search strategy, adapt to the displacement of candidate points caused by belt movement, avoid missed detection caused by speed changes, focus on the vertical deformation characteristics of cracks, suppress the interference of transverse textures on the belt surface, ensure the integrity of the connected domain, and efficiently filter instantaneous noise while being applicable to industrial scenarios with different belt speeds and crack morphologies.

[0101] In a possible embodiment, the judgment condition for spatial adjacency in step 132 is that the three-dimensional spatial Euclidean distance between the adjacent point and the target point is less than the preset connected distance threshold, and the preset connected distance threshold is dynamically set according to the mapping relationship between the spot diameter of the line laser calibrated by experiments and the belt running speed.

[0102] Searching for adjacent points that are spatially adjacent and belong to the initial candidate point set includes:

[0103] Step a1, calculating the tangent direction vector of the laser arc of the target point.

[0104] Among them, the tangent direction vector of the laser arc is the unit vector of the local geometric tangent of the laser arc, which characterizes the trend of the crack to expand along the belt running direction.

[0105] Step a2, generating a vector perpendicular to the tangent direction vector of the laser arc according to the tangent direction vector of the laser arc and the unit vector in the belt thickness direction.

[0106] Among them, the unit vector in the belt thickness direction is a preset direction vector perpendicular to the belt surface, which is used to define the belt thickness direction.

[0107] Step a3: Extend a preset connected distance threshold in both the positive and negative directions of the tangent direction vector of the laser arc to form a strip-shaped search area, and determine the points other than the target point within the strip-shaped search area as adjacent points.

[0108] Among them, the preset connected distance threshold is a preset length value (such as 10 cm) for expanding the search range along the tangent direction, which controls the maximum extension distance of the connected domain. The strip-shaped search area is a narrow three-dimensional area with the tangent direction as the main axis and the normal plane direction as the secondary axis, which is used for directional search of adjacent points.

[0109] The following is a specific example:

[0110] In the detection of the steel belt, the system detects the target point, takes 3 adjacent points before and after this point, and obtains the tangent direction vector by fitting with the least squares method =(0.98, 0.02, 0), cross product and the belt thickness direction vector (0, 0, 1)(0, 0, 1) to generate the normal plane direction vector =(﹣0.02, 0.98, 0), extend 10 cm in both the positive and negative directions along the tangent direction, and the width in the normal plane direction is ±2 mm, construct a strip-shaped area (x range [2.9 m, 3.1 m], y range [0.498 m, 0.502 m]), screen out 5 adjacent points, and through the expansion of the connected domain, form a crack with a length of 12 cm and a depth of 3 mm. After cross-cycle verification, an alarm is triggered to guide the maintenance personnel to accurately locate and repair.

[0111] By executing steps a1~a3, the embodiment of the present application reduces the interference of non-related areas by restricting the search direction, and improves the crack detection efficiency. The strip-shaped area fits the characteristic that the crack extends along the running direction of the belt, avoiding missing detection of oblique or curved cracks. The preset connected distance threshold allows dynamic adjustment of the detection sensitivity according to the actual working conditions (such as the belt speed, crack length).

[0112] Exemplarily, the embodiment of the present application gives a specific embodiment of a belt tear detection method based on three-dimensional vision, which is described as follows:

[0113] Such as Figure 2As shown in the figure, the laser rays emitted by the line laser irradiate on the conveyor belt running at high speed, forming an arc. The camera captures this arc at a certain angle, and the algorithm processes each frame of the belt surface image correspondingly. Under the conditions of knowing the baseline distance between the camera center and the line laser center, the angle between the laser ray and the baseline, and the angle between the camera and the baseline, according to the principle of optical triangulation, the three-dimensional spatial coordinates of the arc can be calculated. The three-dimensional spatial coordinates of the arc can be represented in the form of point cloud data. If there are cracks on the belt surface, there will be uneven undulations on the point cloud. However, since the belt cracks may be very small, such as about 1 mm wide, the undulations of the point cloud are very small. In order to be able to detect such tiny undulation changes without frequent false detections, the embodiment of this application proposes a crack detection method based on the calculation of point cloud connected components. This method can perform the following processes:

[0114] First, calculate the spatial feature (X vector) of each point on the point cloud. The spatial feature is the spatial distance difference between each point and its surrounding adjacent points in the vertical direction (i.e., the direction perpendicular to the arc tangent). Find the points whose distance difference (X vector) is greater than a certain threshold, and form a set A with these points. These points may be crack points or noise points. Noise points are generally not continuous, while crack points may generally be continuous for several points.

[0115] Second, for each point in set A, calculate its surrounding connected component. That is, starting from each point in A, check whether its spatially adjacent points belong to A. If so, classify the adjacent points into the same connected component of this point, and then continue to expand the search starting from the adjacent points. If the adjacent points of the adjacent points also belong to A, they are also classified into the same connected component of this point, and then continue to expand the search starting from the adjacent points of the adjacent points until the spatially adjacent points encountered do not belong to A, and the expansion of the connected component stops. All non-connected connected components form set B.

[0116] Third, calculate the size scale of each connected component in set B. If the length and depth are greater than a certain threshold, it is considered that this connected component may be a crack, and record its spatial position information. These spatial position information form set C.

[0117] Finally, since the conveyor belt rotates cyclically, if the information of cracks is periodically recorded at a certain spatial position (i.e., this position already exists in set C), it is considered that the belt is torn and an alarm prompt is output.

[0118] In this experiment, the hardware includes a line laser and an industrial camera, both placed under the belt. Among them, the distance between the line laser and the camera is 500 mm, the angle between the ray direction of the line laser and the baseline is about 90 degrees, the angle between the camera and the baseline is about 45 degrees, the width of the belt is about 1000 mm, the distance between the belt and the laser is about 400 mm, and there are several cracks with different widths from 1 mm to 10 mm on the belt.

[0119] When the crack on the belt runs to the position irradiated by the laser line, it can be seen that the image is disconnected or undulating, and the point cloud shows a jump, indicating that there is a large difference in the vertical distance of the point cloud in space. At this time, the system outputs the spatial position information of the crack. The belt runs in a cycle. If there is a crack record at the same position every time, the system will alarm. The experimental results show that the crack detection accuracy can reach the millimeter level.

[0120] Among them, the specific steps of connected component calculation are as follows:

[0121] Step 1: Starting from any point x that conforms to the detection feature attributes of this detection (belonging to A), mark x as belonging to a connected component S. Create a queue Q to be viewed, and insert this point x into the queue Q. Currently, the queue Q has only one element x.

[0122] Step 2: Take out an element from the head of the queue Q (starting with x), set it as the current point, and check the features of the spatial left and right adjacent points of the current point. If the adjacent point has not been viewed and its features are the same as the feature attributes of the current point (belonging to A), then this point is marked as belonging to the connected component S and inserted into the tail of the queue Q to be viewed. The current point is marked as viewed and deleted from the Q queue.

[0123] Step 3: Repeat Step 2 until the queue Q is empty. Then the connected component S is found.

[0124] Step 4: Repeat the steps starting from 1 until all point cloud points have been viewed.

[0125] Finally, for each connected component, calculate the spatial shape feature parameters of the connected component, including length, depth, etc. If it meets certain threshold conditions, it is classified as a crack.

[0126] The embodiment of this application is based on computer three-dimensional vision technology to realize automatic detection of longitudinal tearing of the conveyor belt, replacing manual detection, avoiding potential personnel safety accidents, identifying in time, alarming in time, stopping the belt in time, and repairing the damage in time to avoid safety hazards, greatly improving the safe operation level of the belt.

[0127] In a possible embodiment, S12: Calculate the spatial distance values between each target point and its adjacent points in the three-dimensional point cloud data in the direction perpendicular to the tangent of the laser arc, and screen the points whose spatial distance values exceed the preset distance threshold to form an initial candidate point set, including:

[0128] Step b1: Obtain the cross-sectional data of the multi-layer composite structure through a preset belt cross-section scanning device, and establish an association model between the belt material and the belt thickness.

[0129] Among them, the belt cross-section scanning device is a detection device used to obtain the internal parameters of the multi-layer structure of the belt. The multi-layer composite structure can include a rubber covering layer, a fiber reinforcement layer, and a steel wire rope core layer. The correlation model is a mathematical model that describes the relationship between the material properties and thickness of the belt and is established through regression analysis.

[0130] Step b2: Dynamically generate a stratification distance threshold for each stratification structure where the target points are located according to the correlation model.

[0131] Among them, the stratification distance threshold is a depth difference tolerance value dynamically adjusted according to the characteristics of the material layer and is used to screen out abnormal points.

[0132] Step b3: Calculate the spatial distance difference between the target point and the adjacent points within the anisotropic neighborhood in the belt thickness direction. Compare the spatial distance difference with the stratification distance threshold, screen out the abnormal points whose spatial distance difference exceeds the stratification distance threshold, and determine the maximum penetration depth and stratification position information of each abnormal point.

[0133] Among them, the anisotropic neighborhood is a neighborhood defined with different expansion ranges in the X-axis, Y-axis, and Z-axis directions to adapt to the characteristics of the belt layered structure. The maximum penetration depth is the maximum coordinate difference in the thickness direction between the abnormal point and the target point, reflecting the severity of deformation.

[0134] Step b4: Construct a density field based on the spatial distribution characteristics of the abnormal points and calculate the spatial aggregation degree of each abnormal point in the density field within the preset volume element.

[0135] Among them, the density field is a three-dimensional grid data field representing the spatial distribution density of the abnormal points. The spatial aggregation degree is the proportion of the number of abnormal points per unit volume and is used to quantify the aggregation characteristics. The size of the volume element is proportional to the search range of the anisotropic neighborhood.

[0136] Step b5: Determine the product result of the spatial aggregation degree and the maximum penetration depth as the comprehensive abnormality score, screen out the abnormal points whose comprehensive abnormality score exceeds the preset score threshold, and form an initial candidate point set.

[0137] Among them, the comprehensive abnormality score is a weighted index that fuses the aggregation degree and depth, improving the reliability of the crack criterion.

[0138] The following is a specific example:

[0139] In the detection of the steel transportation belt, the laser tomography scanner obtains the belt cross-section data, establishes the correlation model between the rubber layer (thickness 8 mm) and the fiber layer (thickness 5 mm), and sets the stratification distance threshold of the fiber layer to = 1.2 mm. Three abnormal points are found within the neighborhood of the target point, with a maximum penetration depth of 2.0 mm, exceeding the threshold of 1.2 mm. The abnormal points are aggregated within a 1 cm³ volume element, and the spatial aggregation degree = 0.85. The comprehensive score S = 0.85 × 2.0 = 1.7 (exceeding the threshold of 0.7), and it is included in the initial candidate point set. In subsequent steps, this area is expanded by connected components and verified across cycles, confirmed as a real crack, and an alarm is triggered to avoid delamination accidents.

[0140] By executing steps b1~b5, the embodiment of the present application avoids misjudgment caused by material differences through multi-layer material dynamic adaptation and multi-feature fusion, adapts to the multi-layer structure of the composite belt, focuses on the deformation in the thickness direction, suppresses the interference of plane texture, distinguishes random noise and continuous cracks through spatial aggregation features, and combines physical depth and distribution characteristics to reduce the missed detection rate.

[0141] In a possible embodiment, step b3, determining the maximum penetration depth and delamination position information of each abnormal point, includes:

[0142] Step c1, according to the spatial position information of the abnormal point, combined with the geometric model of the delamination boundary, calculate the maximum deviation of each abnormal point from the delamination boundary to generate the maximum penetration depth.

[0143] Among them, the abnormal point is a discrete point deviating from the normal range in the detection data, which may represent defects such as cracks and holes. The spatial position information is the coordinate data of the abnormal point. The delamination boundary is the geometric boundary that divides different material layers or structural layers. The geometric data accurately describes the spatial position, shape, and range of each delamination through, for example, plane equations, surface parametric equations, point clouds, or three-dimensional mesh models. In the present application, the geometric model of the delamination boundary is the core basis for calculating the maximum deviation of the abnormal point. The maximum deviation is the maximum vertical distance of the abnormal point relative to the delamination boundary, distinguishing internal and external penetrations.

[0144] Step c2, combined with the delamination index table, perform delamination position analysis on the spatial position information of the abnormal point to obtain the delamination position information.

[0145] Among them, the delamination index table is a mapping table that records delamination numbers, spatial ranges, and attributes. The delamination position analysis is a spatial matching and attribute analysis process that associates abnormal points with specific delaminations. The delamination position information is structured data that includes delamination numbers, penetration status, and depth.

[0146] The following is a specific example: The conveyor belt of a steel plant consists of three layers of composite materials. Among them, the outer layer (layer 1): wear-resistant rubber layer, with a thickness of 5 mm (z range: 0 mm - 5 mm). The middle layer (layer 2): steel wire reinforcement layer, with a thickness of 5 mm (z range: 5 mm - 10 mm). The inner layer (layer 3): anti-tear fiber layer, with a thickness of 5 mm (z range: 10 mm - 15 mm). Through a three-dimensional vision detection system, the coordinates of an abnormal point are found to be (x = 200 mm, y = 50 mm, z = 7 mm), and it is necessary to determine its maximum penetration depth and layer location information. The lower boundary plane equation of the outer rubber layer: z = 5 mm, the lower boundary plane equation of the middle steel wire layer: z = 10 mm, the abnormal point z = 7 mm, is located in layer 2 (steel wire reinforcement layer), and the vertical distance relative to the lower boundary of the outer layer (z = 5 mm) is 7 - 5 = 2 mm. The vertical distance relative to the lower boundary of the middle layer (z = 10 mm) is 10 - 7 = 3 mm. The penetration depth from the outer layer to the middle layer is 2 mm, the layer is the steel wire reinforcement layer, the rubber layer has been penetrated (2 mm), the steel wire layer has not been completely penetrated, and the anti-tear fiber layer has not been touched. If only relying on two-dimensional images, the abnormal point at z = 7 mm may be misjudged as a surface scratch (the depth cannot be distinguished), resulting in missed detection of damage to the steel wire layer. The three-dimensional detection combined with the layer model can identify internal structure risks and improve the reliability of detection.

[0147] By performing steps c1~c2, the embodiments of the present application achieve precise spatial positioning and layer management of abnormal points through geometric penetration depth calculation and layer attribution analysis. Quantify the degree of damage of abnormal points to the boundary to avoid subjective misjudgment; combine with the layer index table to dynamically associate abnormal points with the layer attributes of engineering structures, support applications such as crack propagation trend analysis and safety hazard grading, and improve the interpretability and decision-making efficiency of detection data.

[0148] In a possible embodiment, S14. The spatial position information of all potential cracks forms a potential crack set, including:

[0149] Step 141. According to the spatial positions of all potential cracks, extract crack feature points and form an initial set of potential crack points.

[0150] Among them, the crack feature points are the points with geometric changes in the crack contour. The initial set of potential crack points is a discrete point set composed of feature points, representing the rough spatial distribution of the cracks.

[0151] Step 142. Perform spatial clustering on the initial set of potential crack points, and merge adjacent points to form potential crack segments.

[0152] Among them, spatial clustering is an algorithm that merges neighboring points into the same group based on distance metrics. The maximum Euclidean distance threshold for merging neighboring points to determine whether two points belong to the same cluster. Potential fracture segments are continuous curve segments generated by clustering and fitting, representing the local fracture morphology.

[0153] Step 143: Calculate the geometric features of all potential fracture segments, analyze the adjacency and crossing relationships between fracture segments, and construct a fracture topology graph.

[0154] Among them, geometric features are parameters that describe fracture segments. The adjacency relationship is the spatial connection relationship with extremely close endpoint spacing between fracture segments. The crossing relationship is the relationship where fracture segments intersect or overlap in a two-dimensional plane or three-dimensional space. The fracture topology graph is a graph structure model with fracture segments as nodes and connection relationships as edges.

[0155] Step 144: Perform global optimization based on geometric continuity and the fracture topology graph, eliminate low-confidence fracture segments, and obtain a set of potential fractures.

[0156] Among them, geometric continuity is the smooth transition characteristic of fracture segments in terms of features such as direction and curvature. Low-confidence fracture segments: suspected noise segments that do not meet the continuity or topological connection conditions.

[0157] The following is a specific example: In the three-dimensional visual inspection of a conveyor belt in a steel plant, three-dimensional point cloud data of the belt surface is collected through a line laser and an industrial camera, and 52 initial potential fracture feature points distributed in the middle of the conveyor belt are detected. These points show an irregular distribution and obvious local curvature changes. Subsequently, the system uses a spatial clustering algorithm based on the Euclidean distance (threshold of 3 mm) to merge the 52 feature points into 6 continuous fracture segments. For example, one of the fracture segments consists of 15 feature points, extends 28 cm along the conveyor belt running direction, and has an average curvature of 0.05 / mm. Then, the system analyzes the geometric features and spatial relationships of each fracture segment. When it is found that the endpoint spacing between segment 1 (length 32 cm) and segment 2 (length 18 cm) is only 1.2 mm at the coordinate (x = 1.5 m, y = 0.3 m), it is determined as an adjacency relationship, while segment 3 (length 12 cm) and segment 4 (length 9 cm) cross each other in three-dimensional space to form an "X" structure, and a topology graph with fracture segments as nodes and connection relationships as edges is constructed. Finally, based on the principle of geometric continuity (direction deviation < 5°, curvature change rate < 15%), the system eliminates segment 5 (curvature mutation 27%) and segment 6 (isolated and unconnected), retains 4 high-confidence fracture segments (total length 71 cm) to form an optimized set of potential fractures, and triggers a secondary alarm to guide maintenance personnel to focus on inspecting the steel wire reinforcement layer in the area from 1.2 m to 1.9 m in the middle of the conveyor belt to avoid production accidents caused by the expansion of longitudinal tears.

[0158] By performing steps 141 to 144, the embodiments of the present application achieve the separation of real cracks from discrete noises, extraction of feature points, transformation of disordered connected regions into discrete point sets, generation of continuous crack segments through clustering to enhance morphological consistency, and revelation of crack connection rules by quantifying geometric properties and constructing topological relationships, by integrating physical laws and topological logic, eliminating abnormal segments, and finally outputting a high-precision set of potential cracks. This process improves the accuracy, continuity, and interpretability of crack detection and is applicable to the health monitoring of complex engineering structures such as bridges and tunnels.

[0159] In a possible embodiment, S13, calculating the length and depth of each connected region includes:

[0160] Step d1, parsing the spatial positions and boundary information of all connected regions to generate an initial description information set for each connected region.

[0161] Among them, the spatial position is the position parameter of the connected region in the image coordinate system. The boundary information is the geometric property of the connected region contour. The initial description information set is a structured data set containing the basic attributes of the connected region.

[0162] Step d2, according to the initial description information set and a preset spatial geometric relationship model, performing boundary extraction for each connected region to obtain the coordinate sequences of all boundary points of each connected region.

[0163] Among them, the preset spatial geometric relationship model is an algorithm or rule for describing geometric relationships. The boundary point coordinates are the pixel coordinate sequences arranged in order of the connected region contour.

[0164] Step d3, according to the coordinate sequences of all boundary points of each connected region, analyzing the relative position relationship between all boundary points of the connected region, determining two first endpoints corresponding to the longest path, and taking the distance value between the two first endpoints as the length of the connected region.

[0165] Among them, the relative position relationship is the geometric relationship between boundary points. The longest path is the line segment with the maximum Euclidean distance among the boundary point pairs of the connected region. The first endpoints are the two endpoints of the longest path. The length of the connected region is the Euclidean distance between the first endpoints.

[0166] Step d4, according to the coordinate sequences of all boundary points of each connected region, analyzing the distance distribution of all boundary points of the connected region from the preset center point of the connected region, determining two second endpoints perpendicular to the longest path, and taking the distance value between the two second endpoints as the depth of the connected region.

[0167] Among them, the distance distribution is the statistical distribution of the projection distances from the boundary points to the center point or the axis. The center point of the connected component is the arithmetic mean position of the boundary point coordinates. The second end points are the pair of points with the maximum spacing perpendicular to the direction of the longest path. The depth of the connected component refers to the Euclidean distance between the second end points.

[0168] The following is a specific example: In the three-dimensional vision inspection of the conveyor belt in a steel plant, the system identifies a certain connected component on the belt surface. First, it parses its spatial position as a rectangular area in the image coordinate system (the upper left corner coordinates are (150, 80), and the lower right corner coordinates are (320, 120)). The boundary contour contains 58 discrete points, generating an initial description information set including area, perimeter, and circumscribed rectangle parameters. Subsequently, based on the Moore neighborhood tracking algorithm, the coordinate sequence of the boundary points is extracted and arranged in a clockwise order as [(150,80),(155,82),..., (320,120)]. Then, it traverses all pairs of boundary points to calculate the Euclidean distance, determines that the end points of the longest path are (160,85) and (310,115), and the straight-line distance between the two is 152 mm, which is marked as the length of the connected component. Further, it calculates the arithmetic mean of the boundary points to obtain the center point (235, 100), establishes a vertical axis along the direction of the longest path, filters out the two second end points (230,70) and (230,130) that are farthest from the center point, and measures the depth of 60 mm. Combining the length and depth data (152 mm × 60 mm), the system determines that this connected component conforms to the characteristics of longitudinal cracks, triggers a first-level alarm, and outputs the three-dimensional coordinates (1.5 m, 0.6 m), guiding the maintenance personnel to give priority to checking the joint of the rubber covering layer and the steel wire layer in the roller area of the 3rd section of the conveyor belt to avoid belt breakage and shutdown caused by deep tearing.

[0169] By performing steps d1~d4, the embodiments of the present application first parse the spatial positions and boundary information of all connected components, generating an initial description information set including parameters such as circumscribed rectangle coordinates, area, and perimeter. Subsequently, according to the initial description information set and the preset spatial geometric relationship model, the coordinate of the boundary points of each connected component is extracted to form an ordered set of contour points. Then, it analyzes the relative position relationship between the boundary points, calculates the Euclidean distance between all pairs of points, filters out the two first end points with the largest spacing, and determines the length of the connected component. Finally, based on the center point of the connected component and the direction of the longest path, it analyzes the projection distance distribution of the boundary points, determines the two second end points perpendicular to the longest path, and calculates the depth value. This process transforms the image connected component into quantifiable length and depth parameters through layer-by-layer data conversion and geometric calculation, avoiding the error of manual intervention, and is applicable to the precise detection and safety assessment of crack sizes in engineering structures such as bridges and tunnels.

[0170] Figure 3 The structural schematic diagram of a belt tear detection device based on three-dimensional vision provided by the embodiments of the present application is as Figure 3 shown, and the device includes:

[0171] The acquisition module 31 is used to acquire the belt surface image through a line laser and an industrial camera deployed under the belt, and calculate the three-dimensional point cloud data of the laser arc in the belt surface image.

[0172] The screening module 32 is used to calculate the spatial distance value between each target point and the adjacent points in the 3D point cloud data in the direction perpendicular to the tangent of the laser arc, and screen the points whose spatial distance value exceeds the preset distance threshold to form an initial candidate point set.

[0173] The calculation module 33 is used to start from each point in the initial candidate point set, expand the points that are spatially adjacent and belong to the initial candidate point set through the breadth-first search algorithm, obtain the target candidate point set, so as to form multiple disconnected connected domains, and calculate the length and depth of each connected domain.

[0174] The filtering module 34 is used to filter out the connected domains that do not meet the threshold conditions in all connected domains, retain the connected domains that meet the threshold conditions, regard all the connected domains that meet the threshold conditions as corresponding potential cracks, and form a potential crack set with the spatial location information of all potential cracks. The domains that do not meet the threshold conditions are those whose length and depth do not reach the corresponding preset thresholds.

[0175] The generation module 35 is used to generate a belt tear detection result including the spatial position information of the real crack and the corresponding alarm level if there is a real crack during the belt circulation operation. The real crack is a potential crack recorded in the same spatial position in the potential crack set in multiple consecutive operation cycles.

[0176] Figure 3 The belt tear detection device based on three-dimensional vision can be performed Figure 1 The implementation principles and technical effects of the 3D vision-based belt tear detection method described in the illustrated embodiment are not further elaborated. The specific manner in which the various modules and units in the 3D vision-based belt tear detection device described in the aforementioned embodiment perform their operations has been described in detail in the related embodiments of the method and will not be further elaborated here.

[0177] In one possible design, Figure 3 The belt tear detection device based on three-dimensional vision of the embodiment shown can be implemented as a computing device, such as Figure 4 As shown, the computing device may include a storage component 41 and a processing component 42 .

[0178] The storage component 41 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 42 .

[0179] The processing component 42 is configured to: collect images of the belt surface through a line laser and an industrial camera deployed under the belt, and calculate the three-dimensional point cloud data of the laser arc in the belt surface image. Calculate the spatial distance values between each target point and its adjacent points in the three-dimensional point cloud data in the direction perpendicular to the tangent of the laser arc, and filter out the points whose spatial distance values exceed a preset distance threshold to form an initial candidate point set. Starting from each point in the initial candidate point set, expand the points that are spatially adjacent and belong to the initial candidate point set through the breadth-first search algorithm to obtain a target candidate point set, so as to form multiple non-connected connected domains, and calculate the length and depth of each connected domain. Filter out the connected domains that do not meet the threshold conditions among all the connected domains, retain the connected domains that meet the threshold conditions, regard all the connected domains that meet the threshold conditions as corresponding potential cracks, and form a potential crack set with the spatial position information of all the potential cracks. The condition that does not meet the threshold is that both the length and the depth do not reach the corresponding preset thresholds. During the cyclic operation of the belt, if there is a real crack, a belt tear detection result including the spatial position information of the real crack and the corresponding alarm level is generated. The real crack is a potential crack at the same spatial position that is recorded in the potential crack set in multiple consecutive operating cycles.

[0180] Among them, the processing component 42 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components for executing the above method.

[0181] The storage component 41 is configured to store various types of data to support the operations of the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Random Access Memory (RAM), Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), magnetic memory, flash memory, magnetic disk or optical disc.

[0182] Of course, the computing device may necessarily further include other components, such as an input / output interface, a display component, a communication component, etc. The input / output interface provides an interface between the processing component and the peripheral interface module, and the above-mentioned peripheral interface module may be an output device, an input device, etc.

[0183] The communication component is configured to facilitate communication between the computing device and other devices in a wired or wireless manner, etc.

[0184] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. At this time, the computing device can refer to a cloud server, and the above-mentioned processing component, storage component, etc. can be basic server resources leased or purchased from a cloud computing platform.

[0185] The embodiment of the present application also provides a computer storage medium storing a computer program, and when the computer program is executed by a computer, it can implement the above-mentioned Figure 1 belt tearing detection method based on three-dimensional vision shown in the embodiment.

[0186] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0187] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0188] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0189] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A belt tearing detection method based on three-dimensional vision, characterized in that, Including: Collecting the surface image of the belt by a line laser deployed under the belt and an industrial camera, and calculating the three-dimensional point cloud data of the laser arc in the belt surface image; Calculating the spatial distance value between each target point and its adjacent point in the three-dimensional point cloud data in the direction perpendicular to the tangent of the laser arc, and screening out the points whose spatial distance value exceeds the preset distance threshold to form an initial candidate point set; Starting from each point in the initial candidate point set, expand the points that are spatially adjacent and belong to the initial candidate point set through the breadth-first search algorithm to obtain the target candidate point set, so as to form multiple non-connected connected domains, and calculate the length and depth of each connected domain; the formula for calculating the length of the connected domain is: , where L is the length of the connected domain, is the maximum and minimum x-axis coordinates of the projections of all points in the connected domain on the belt surface, is the maximum and minimum y-axis coordinates of the projections of all points in the connected domain on the belt surface; the formula for calculating the depth of the connected domain is: , where D is the depth of the connected domain, is the height value of each point in the connected domain, is the reference height of the belt surface; Filtering out the connected domains that do not meet the threshold conditions among all connected domains, retaining the connected domains that meet the threshold conditions, taking all the connected domains that meet the threshold conditions as the corresponding potential cracks, and forming a potential crack set with the spatial position information of all potential cracks. The non-compliance with the threshold conditions means that both the length and the depth do not reach the corresponding preset thresholds; During the cyclic operation of the belt, if there is a real crack, generating a belt tear detection result including the spatial position information of the real crack and the corresponding alarm level. The real crack is a potential crack at the same spatial position that has been recorded in the potential crack set in multiple consecutive operation cycles; The spatial position information of the real crack adopts the following form: merging the three-dimensional point cloud data of the cracks detected in all cycles to form a spatio-temporal fusion point cloud; or, taking the weighted average of the central coordinates of the real cracks in all cycles; Starting from each point in the initial candidate point set, expanding the points that are spatially adjacent and belong to the initial candidate point set through the breadth-first search algorithm to obtain a target candidate point set, including: Traversing each target point in the initial candidate point set, marking the target point as a member of the current connected domain, and inserting it into a preset queue to be viewed. According to the belt running direction and the belt running speed, calculate the dynamic neighborhood range of the current target point; The search distance extended along the belt running direction is greater than the direction perpendicular to the tangent of the laser arc; The extended distance of the dynamic neighborhood range in the belt running direction is: L = v×Δt + L0, where v is the real-time speed of the belt, Δt is the image acquisition period, and L0 is the reference neighborhood length; Searching for adjacent points that are spatially adjacent and belong to the initial candidate point set within the dynamic neighborhood range in the direction perpendicular to the tangent of the laser arc; Judging whether the adjacent point of the target point belongs to the initial candidate point set and has not been marked; If so, marking the adjacent point as a member of the current connected domain, inserting it into the end of the preset queue to be viewed, taking the adjacent point as a new target point, and repeating the steps of marking, inserting, calculating, searching, and judging until all the target points in the preset queue to be viewed form a connected domain, clearing the preset queue to be viewed, and repeating the formation process of the new connected domain until each target point in the initial candidate point set is marked; Deleting all target points in the initial candidate point set that are not in any connected domain to obtain a target candidate point set.

2. The method according to claim 1, wherein The judgment condition for spatial adjacency is that the three-dimensional spatial Euclidean distance between the adjacent point and the target point is less than the preset connected distance threshold, and the preset connected distance threshold is dynamically set according to the mapping relationship between the spot diameter of the line laser calibrated by experiments and the belt running speed; Searching for adjacent points that are spatially adjacent and belong to the initial candidate point set, including: Calculate the laser arc tangent direction vector of the target point; Generate a vector perpendicular to the laser arc tangent direction vector based on the laser arc tangent direction vector and the unit vector in the belt thickness direction; Extend a preset connected distance threshold in both the positive and negative directions of the laser arc tangent direction vector to form a strip-shaped search area, and determine the points other than the target point within the strip-shaped search area as adjacent points.

3. The method according to claim 1, wherein Calculate the spatial distance values between each target point and its adjacent points in the three-dimensional point cloud data in the direction perpendicular to the laser arc tangent direction, and filter out the points whose spatial distance values exceed the preset distance threshold to form an initial candidate point set, including: Obtain the cross-sectional data of the multi-layer composite structure through a preset belt cross-section scanning device, and establish an association model between the belt material and the belt thickness; Dynamically generate a layer distance threshold for each layer structure where the target point is located according to the association model; Calculate the spatial distance difference between the target point and its adjacent points in the anisotropic neighborhood in the belt thickness direction; compare the spatial distance difference with the layer distance threshold, filter out the abnormal points whose spatial distance difference exceeds the layer distance threshold, and determine the maximum penetration depth and layer position information of each abnormal point; Construct a density field based on the spatial distribution characteristics of the abnormal points, and calculate the spatial aggregation degree of each abnormal point in the density field within a preset volume element; Determine the product result of the spatial aggregation degree and the maximum penetration depth as the comprehensive abnormality score, and filter out the abnormal points whose comprehensive abnormality score exceeds the preset score threshold to form an initial candidate point set.

4. The method according to claim 3, wherein The determination of the maximum penetration depth and layer position information of each abnormal point includes: According to the spatial position information of the abnormal point, combined with the geometric model of the layer boundary, calculate the maximum deviation amount of each abnormal point from the layer boundary to generate the maximum penetration depth; Combined with the layer index table, perform layer position analysis on the spatial position information of the abnormal point to obtain the layer position information.

5. The method according to claim 1, wherein The formation of a potential crack set from the spatial position information of all potential cracks includes: Extract crack feature points according to the spatial positions of all the potential cracks and form an initial set of potential crack points; Perform spatial clustering on the initial set of potential crack points, and merge adjacent points to form potential crack segments; Calculate the geometric characteristics of all the potential crack segments, analyze the adjacency and intersection relationships between the crack segments, and construct a crack topology graph; Perform global optimization based on geometric continuity and the crack topology graph, and eliminate low-confidence crack segments to obtain a potential crack set.

6. The method according to claim 1, characterized in that, The calculation of the length and depth of each connected domain includes: Analyze the spatial positions and boundary information of all the connected domains to generate an initial description information set for each connected domain; According to the initial description information set and a preset spatial geometric relationship model, perform boundary extraction for each connected domain to obtain the coordinate values of all boundary points of each connected domain; According to the coordinate values of all boundary points of each connected domain, analyze the relative position relationships between all the boundary points of the connected domain, determine the two first endpoints corresponding to the longest path, and use the distance value between the two first endpoints as the length of the connected domain; According to the coordinates of all boundary points of each connected domain, analyze the distance distribution between all boundary points of the connected domain and the preset center point of the connected domain, determine two second endpoints perpendicular to the longest path, and use the distance value between the two second endpoints as the depth of the connected domain.

7. A belt tearing detection device based on three-dimensional vision, characterized in that Including: An acquisition module, configured to collect the belt surface image through a line laser and an industrial camera deployed under the belt, and calculate the three-dimensional point cloud data of the laser arc in the belt surface image; A screening module, configured to calculate the spatial distance value between each target point and its adjacent point in the three-dimensional point cloud data in the direction perpendicular to the tangent of the laser arc, and screen out the points whose spatial distance value exceeds a preset distance threshold to form an initial candidate point set; A calculation module, which is used to start from each point in the initial candidate point set, expand the points that are spatially adjacent and belong to the initial candidate point set through the breadth-first search algorithm to obtain a target candidate point set, so as to form multiple non-connected connected domains, and calculate the length and depth of each connected domain; the formula for calculating the length of the connected domain is: , where L is the length of the connected domain, is the maximum and minimum values of the x-axis coordinates of the projections of all points in the connected domain on the belt surface, is the maximum and minimum values of the y-axis coordinates of the projections of all points in the connected domain on the belt surface; the formula for calculating the depth of the connected domain is: , where D is the depth of the connected domain, is the height value of each point in the connected domain, is the reference height of the belt surface; A filtering module, configured to filter out the connected domains that do not meet the threshold conditions among all connected domains, retain the connected domains that meet the threshold conditions, use all the connected domains that meet the threshold conditions as corresponding potential cracks, and form a potential crack set with the spatial position information of all potential cracks. The non-compliance with the threshold conditions means that both the length and the depth do not reach the corresponding preset thresholds; A generation module, configured to generate a belt tear detection result including the spatial position information of the real crack and the corresponding alarm level during the cyclic operation of the belt. The real crack is a potential crack at the same spatial position that is recorded in the potential crack set in multiple consecutive operation cycles; the spatial position information of the real crack adopts the following form: merge the three-dimensional point cloud data of the cracks detected in all cycles to form a spatio-temporal fusion point cloud; or, take the weighted average of the center coordinates of the real cracks in all cycles; Starting from each point in the initial candidate point set, expanding the points that are spatially adjacent and belong to the initial candidate point set through the breadth-first search algorithm to obtain a target candidate point set, including: Traverse each target point in the initial candidate point set, mark the target point as a member of the current connected domain, and insert it into a preset queue to be viewed. According to the belt running direction and the belt running speed, calculate the dynamic neighborhood range of the current target point; the search distance for expansion along the belt running direction is greater than the direction perpendicular to the tangent of the laser arc; the expansion distance of the dynamic neighborhood range in the belt running direction is: L = v×Δt + L0, where v is the real-time speed of the belt, Δt is the image acquisition period, and L0 is the reference neighborhood length; Search for adjacent points that are spatially adjacent and belong to the initial candidate point set within the dynamic neighborhood range in the direction perpendicular to the tangent of the laser arc; Judge whether the adjacent points of the target point belong to the initial candidate point set and are not marked; If so, mark the adjacent point as a member of the current connected domain, insert it into the end of the preset queue to be viewed, use the adjacent point as a new target point, and repeat the steps of marking, inserting, calculating, searching, and judging until all the target points in the preset queue to be viewed form a connected domain, empty the preset queue to be viewed, and repeat the formation process of the new connected domain until each target point in the initial candidate point set is marked; Delete all target points in the initial candidate point set that are not in any connected domain to obtain a target candidate point set.

8. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a three-dimensional vision-based belt tear detection method according to any one of claims 1 to 6.

9. A computer storage medium, characterized in that, A computer program is stored, and when the computer program is executed by a computer, it implements a three-dimensional vision-based belt tear detection method according to any one of claims 1 to 6.

Images