Intelligent inspection method for a coordinate measuring machine

Abstract

The embodiment of the present application provides a kind of intelligent detection method for three coordinate measuring machine, comprising: the non-contact scanning of measuring machine workbench is carried out, and initial three-dimensional data is acquired;The data is processed to segment and identify the workpiece object on the table, and extract its geometric feature information;Based on the feature information, at least one workpiece object is generated according to the preset decision rule Measurement scheme, the scheme at least includes selecting at least one from multiple measuring heads and planning detection path for the selected measuring head;Control measuring machine to execute the scheme to complete detection;And based on the execution and result of this detection, the decision rule or related database is optimized, the present application realizes the whole process closed loop automation from workpiece autonomous discovery, intelligent identification, scheme planning to accurate execution and learning, without prior information and artificial programming of workpiece, significantly improve detection efficiency, equipment utilization and the adaptive ability of system.

Description

Technical Field

[0001] This invention relates to an intelligent detection method for a coordinate measuring machine. Background Technology

[0002] Coordinate measuring machines (CMMs) are widely used precision measuring devices that obtain spatial coordinate information of contact points through contact measurement, reflecting the actual manufacturing dimensions of parts and information such as surface flatness.

[0003] In typical inspections, operators usually need to position the workpiece on the worktable and formulate a measurement program based on the workpiece's CAD model or after performing basic measurements to form its basic shape. These measurement programs clearly specify the measuring head, probe type, measurement path, and safety plane used for each feature to be measured, and then the measuring machine automatically executes the program to complete the inspection.

[0004] However, the above-mentioned typical form of detection has the following problems: 1. Each time a new workpiece is inspected or the workpiece's placement changes, it is necessary to perform positioning and reprogramming by an experienced engineer or manual teaching. The preparation work is time-consuming and requires high skill levels from personnel.

[0005] 2. When there are multiple workpieces of different types or randomly placed on the workbench, the system cannot automatically identify and plan them, making it difficult to achieve efficient batch inspection.

[0006] 3. When multiple measuring heads are configured, the selection of the measuring head may not be the optimal choice, or because the measuring head is fixed by the program, it is impossible to dynamically optimize it according to the real-time status and feature type of the workpiece, which further leads to suboptimal measurement efficiency or accuracy.

[0007] To address or solve the aforementioned technical problems, several new technical solutions have been proposed. For example, patent documents CN121255178A and CN115112018A propose that, based on a known CAD model, offline measurement programs can be automatically generated through feature recognition, adaptive algorithms to select probes, and path planning, thereby improving detection efficiency. Patent document CN121074140A proposes that, based on deep reinforcement learning, the measurement path and action sequence of a single probe can be planned online according to information such as the features to be measured and probe status obtained from the design drawings, thereby optimizing the motion path and improving efficiency.

[0008] However, current technical solutions still rely heavily on accurate pre-modeling, which cannot handle situations where there is no CAD model for the workpiece or the incoming material does not match the model. They also cannot quickly and accurately measure one or more workpieces set on the table on site, nor can they intelligently decide the optimal detection method, such as selecting the appropriate measuring head and the order of measuring workpieces. Therefore, a better technical solution is needed to solve the above technical problems. Summary of the Invention

[0009] To address these issues, the present invention provides an intelligent detection method for a coordinate measuring machine to solve the aforementioned technical problems.

[0010] A smart inspection method for a coordinate measuring machine, characterized by comprising the following steps: S1: Perform non-contact scanning on the worktable of the coordinate measuring machine to obtain initial three-dimensional data; S2: Process the initial three-dimensional data to identify one or more workpiece objects on the worktable and obtain information about each workpiece object related to the detection plan; S3: Based on the information of each workpiece object, a corresponding measurement plan is generated for each workpiece object according to the preset decision rules; the measurement plan includes at least measurement head selection and detection path planning; S4: Control the coordinate measuring machine to execute the measurement plan and complete the inspection of the workpiece object; S5: Based on the execution and results of this test, optimize the preset decision rules or related database.

[0011] Step S1 includes the following specific steps: S11: Control the non-contact scanning sensor to perform the first stage scan at the first resolution and acquire the first point cloud data of the worktable surface; S12: Based on the first point cloud data, identify one or more target areas on the workbench surface; S13: For each target area, control the non-contact scanning sensor to perform a second-stage scan with a resolution higher than the first resolution to acquire the corresponding second point cloud data; S14: Generate the initial three-dimensional data based on the first point cloud data and each of the second point cloud data.

[0012] Step S12 includes the following specific steps: S121: Point cloud preprocessing, performing outlier filtering on the first point cloud data; S122: Workbench reference extraction and background removal; S123: Point cloud clustering and segmentation and initial selection of target regions: A clustering algorithm based on Euclidean distance is used to group spatially adjacent points into multiple independent point cloud clusters, and each point cloud cluster is regarded as a candidate target region; S124: Target region screening and parameterization: Calculate the three-dimensional axis-aligned bounding box of each point cloud cluster, and at the same time, based on prior knowledge or preset rules, screen out clusters that do not conform to the workpiece characteristics. The parameterization refers to calculating and recording the key parameters of each candidate target region that passes the screening. S125: Output target region: Output target region is a structured list of target regions.

[0013] Step S13 includes the following specific steps: S131: Based on the target area list output in step S12, generate an independent fine scanning task for each target area in the list; S132: Perform a fine scanning task, specifically including: S1321: Drive the non-contact scanning sensor to a safe approach point outside the current target area boundary, and then lower it to a safe starting height calculated for that area; S1322: Control the non-contact scanning sensor to move along the scanning path planned for the area, while continuously acquiring data at a second resolution; S133: Acquire and output the second point cloud data.

[0014] Step S14 includes the following specific steps: S141: Coordinate system alignment; S142: Data fusion to form initial 3D data: Including in the first point cloud data, all low-resolution point clouds in the target area that have been covered by the second point cloud data are removed or marked as expired. Then, the second point cloud data units are inserted into the corresponding empty positions in the first point cloud data according to their spatial coordinates.

[0015] Step S3 includes the following specific steps: S31: Rule-based feature-measurement head matching is used for preliminary screening to form a candidate set; S32: Perform global multi-objective optimization based on the candidate set to generate a measurement scheme that includes measurement head selection and detection path; Among them, the multi-objective optimization includes minimizing the total detection time, minimizing the measurement uncertainty, minimizing the collision risk index, and minimizing the system wear uniformity.

[0016] The optimization process in step S32 includes the following constraints: (1) Each feature must be assigned one available and technically feasible measurement head.

[0017] (2) The measurement path must meet the kinematic constraints of the coordinate measuring machine.

[0018] (3) The path must completely avoid known interference areas.

[0019] (4) For the same workpiece, the measurement sequence must meet the geometric tolerance principle of datum first.

[0020] In step S31, each inferred feature of each workpiece is traversed, and predefined expert rules are applied to select a set of feasible measurement head candidates.

[0021] Before the coordinate measuring machine performs step S4, there is a step of using a measuring head to verify and correct a significant feature of the workpiece to be measured.

[0022] During the verification process, a verification scan or measurement is performed, and the actual position obtained by the scan or measurement is compared with the theoretical position in the measurement plan. If the deviation is within the allowable range, the coordinates of all measurement paths of the workpiece are subsequently compensated by overall translation and rotation. If the deviation exceeds the limit, a warning message is issued and manual intervention is required.

[0023] Beneficial Effects: This invention provides an intelligent inspection method for a coordinate measuring machine (CMM), comprising: performing non-contact scanning of the CMM's worktable to acquire initial three-dimensional data; processing the data to segment and identify workpiece objects on the worktable and extracting their geometric feature information; generating a measurement scheme for at least one workpiece object based on the feature information and according to preset decision rules, wherein the scheme includes selecting at least one from multiple measuring heads and planning a detection path for the selected measuring head; controlling the CMM to execute the scheme to complete the inspection; and optimizing the decision rules or related database based on the execution and results of the inspection. This invention achieves a closed-loop automation of the entire process from workpiece autonomous discovery, intelligent identification, scheme planning to precise execution and learning, without requiring prior workpiece information or manual programming, significantly improving inspection efficiency, equipment utilization, and system adaptability. Attached Figure Description

[0024] Figure 1 A schematic diagram of intelligent detection for precision helical worm gears. Detailed Implementation

[0025] The preferred embodiments of this application are described in detail below so that the advantages and features of this application can be more easily understood by those skilled in the art, thereby making a clearer and more explicit definition of the scope of protection of this application.

[0026] The embodiments of this application provide an intelligent inspection method for a coordinate measuring machine, which enables intelligent detection of the entire process from workpiece autonomous discovery, intelligent identification, scheme planning, precise execution to result feedback, thereby reducing reliance on operator experience and skills and improving the adaptability of inspection tasks, optimizing inspection efficiency and resource utilization.

[0027] Understandably, the coordinate measuring machine (CMM) is equipped with a connecting end that can move at least along the x, y, and z directions. Simultaneously, the CMM is equipped with various types of measuring sensors, including contact measuring heads (such as trigger probes and scanning probes) and non-contact scanning sensors (such as laser line scanners and vision sensors). The connecting end can selectively connect to one of these measuring sensors, and use that sensor to scan or measure the workpiece. The structural form of the CMM is not limited in this application. For ease of description, contact measuring heads and non-contact scanning sensors will be collectively referred to as measuring heads below, but their types will be specified when specific functional steps are involved.

[0028] Understandably, the connection end and the measuring sensor are equipped with a quick interface device, so that the connection end can automatically select and connect to one of the measuring sensors to perform scanning or contact measurement tasks according to instructions.

[0029] In a typical embodiment, the coordinate measuring machine (CMM) uses a gantry structure. The gantry consists of two vertically arranged columns and a crossbeam. The two ends of the crossbeam are supported on the columns. The columns can move in the y-direction on a base platform of the CMM. The CMM has a slide that can travel in the x-direction along the axis of the crossbeam's length. A support column is mounted on the slide, which can move in the z-direction. A connecting end is formed at the lower end of the support column and connects to the measuring head. It is understood that the x, y, and z directions are mutually perpendicular, defining the axes of a Cartesian coordinate system. A worktable for fixing the workpiece is provided on the lower side of the columns.

[0030] In addition, scales are installed on the crossbeam, base platform, and support columns, extending along the x, y, and z directions respectively, and reading heads are installed at the corresponding positions to automatically read the corresponding measurement values ​​from the scales. At the same time, the coordinate measuring machine is also equipped with a control device to drive the slide, column, or support column to move along the x, y, and z directions, ultimately enabling the measuring head to move along the x, y, and z directions.

[0031] The method described in this application is described in detail below.

[0032] S1: Perform non-contact scanning on the worktable of the coordinate measuring machine to obtain initial three-dimensional data.

[0033] Understandably, by performing non-contact scanning on the worktable of the coordinate measuring machine, the three-dimensional spatial information of the worktable can be fully acquired without prior knowledge of the workpiece's prior information, providing a data source for subsequent intelligent recognition and decision-making.

[0034] Specifically, it includes the following steps.

[0035] S11: Control the non-contact scanning sensor to perform the first stage scan at the first resolution and acquire the first point cloud data of the worktable.

[0036] It is understandable that step S11 can quickly and initially obtain the three-dimensional information of the entire work surface. In the first stage of scanning, a relatively low first resolution can be used to identify the existence and general outline of the object, thereby improving the scanning speed. It is understandable that the first stage of scanning is not sufficient for accurate size measurement or fine feature analysis.

[0037] Furthermore, the point cloud density corresponding to the first resolution is 1 to 10 points per square centimeter. For example, when using a line laser scanner, the average projection point spacing of the generated point cloud on the table can be 3 to 10 millimeters by controlling its scanning line spacing or movement speed. This resolution setting can effectively distinguish the worktable, the workpiece to be measured, and the clamping fixture while ensuring fast scanning.

[0038] Furthermore, during the first stage of scanning, the non-contact scanning sensor moves along a predetermined first motion path, which should be able to cover the entire effective area of ​​the worktable. Preferably, the first motion path is a bow-shaped path or a parallel grid path.

[0039] The bow-shaped path is also known as a reciprocating or serpentine path. Specifically, when the non-contact scanning sensor scans along the bow-shaped path, it can start from a starting point on one side of the worktable and move linearly along the X-axis at a constant speed Vx while scanning synchronously. When it reaches the end of the scan line, the scanning module steps a fixed line spacing ΔY in the Y-axis direction. The size of ΔY can be understood to match the first resolution, such as ΔY=20mm. Then, it moves at a constant speed along the opposite X-axis direction and scans the next line, repeating this process until the entire range of the worktable in the Y direction is covered.

[0040] When the non-contact scanning sensor scans along a parallel grid path, the worktable surface is divided into a regular grid in the XY plane. The movement of the scanning module consists of a series of parallel, unidirectional scanning line segments. Specifically, the scanning module starts from the beginning of the first scanning line, completes the scanning of the entire line along the X-axis, picks up the measuring head or raises it to a safe height, and quickly moves to the beginning of the next scanning line. This starting point is located at the same Y-coordinate as the starting point of the previous line, and then performs the next scan along the positive X-axis. All scanning lines maintain a constant spacing ΔY in the Y-axis direction.

[0041] It is understood that the non-contact scanning head may be a line laser scanner, a structured light scanner, or a lidar, and there are no restrictions on this.

[0042] Furthermore, before performing step S11, the following steps are also included: S111: Obtain the measurement sensor information currently installed on the measuring machine and compare it with the predefined non-contact scanning sensor information suitable for performing the first stage of scanning; S112: If the current measurement sensor meets the requirements, it will be used directly; otherwise, it will be replaced with a non-contact scanning sensor that meets the requirements.

[0043] S12: Based on the first point cloud data, identify one or more target areas on the workbench surface.

[0044] Specifically, it includes the following steps: S121: Point cloud preprocessing.

[0045] Specifically, outlier filtering is performed on the first point cloud data.

[0046] In this embodiment, a statistical filtering algorithm is used to calculate the average distance between each point and its nearest neighbor, and to remove discrete noise points whose distance exceeds 3-5 times the global average value, in order to eliminate possible errors from dust, reflection, or sensors.

[0047] Furthermore, after performing outlier filtering, voxel downsampling is performed.

[0048] Specifically, while preserving the overall shape of the point cloud, the point cloud is uniformly downsampled using the voxel grid method. Each small voxel, such as all points within a cube with a side length of 2-3 mm, is represented by its centroid, thereby reducing the total amount of data.

[0049] S122: Workbench reference extraction and background removal.

[0050] Specifically, the Random Sample Consensus (RANSAC) algorithm or other plane fitting algorithms can be used to fit the largest plane from the point cloud. This plane is defined as the reference plane of the workbench. At the same time, all points that are determined to belong to this plane, such as points whose distance from the plane is less than a threshold, such as 1 mm, will be temporarily removed or marked as background, thereby simplifying the scene and facilitating the subsequent processing of objects on the workbench.

[0051] S123: Point cloud clustering and segmentation and initial selection of target regions.

[0052] Specifically, clustering algorithms based on Euclidean distance, such as DBSCAN or Euclidean clustering extraction, can be used. A distance threshold d1 is set, such as 5mm to 15mm. At the same time, d1 matches the point spacing of the first resolution. The algorithm traverses all points and groups points with a spatial distance less than d1 into the same cluster. Through the above clustering, spatially adjacent points are grouped into multiple independent point cloud clusters. Each point cloud cluster is regarded as a candidate target region, corresponding to a potential workpiece.

[0053] S124: Target region filtering and parameterization.

[0054] Specifically, the three-dimensional axis-aligned bounding box (AABB) of each point cloud cluster is calculated, and clusters that do not conform to the workpiece characteristics are filtered out based on prior knowledge or preset rules.

[0055] The clusters that do not conform to the characteristics of the workpiece include: clusters that are too small in size and clusters that are too low in height. It is understood that clusters that are too small in size may be residual noise or small fragments, and clusters that are too low in height may be surface textures.

[0056] Furthermore, in the screening of target areas, it is also necessary to determine spatial interference. Specifically, the projection of the bounding box of each cloud cluster on the XY plane is calculated. If the overlapping area of ​​two projection areas exceeds a predetermined ratio and they are highly overlapping in the Z-axis direction, they need to be marked as suspected spatial interference areas and further processed or warned in subsequent steps.

[0057] In addition, the parameterization mentioned above refers to calculating and recording the key parameters of each candidate target region that passes the screening.

[0058] Parametric Description: For each candidate target region that passes the screening, its key parameters are calculated and recorded. These parameters include the two-dimensional projected bounding box, the region center and height, and the number of point clouds. The two-dimensional projected bounding box refers to the minimum and maximum X and Y coordinates on the XY plane, used to define the range of fine scanning in step S13. The region center and height refer to the weighted centroid position (X, Y, Z) of the point cloud and its maximum height relative to the platform.

[0059] S125: Output target area.

[0060] Understandably, the output target region is to output a structured list of target regions. The list of target regions includes several items, each of which corresponds to an identified valid target region. The valid target region should include information such as ID, two-dimensional boundary, center position, height, and possible interference markers, which are used to guide the high-resolution data acquisition in subsequent steps.

[0061] S13: For each target area, control the non-contact scanning sensor to perform a second-stage scan with a resolution higher than the first resolution to acquire the corresponding second point cloud data.

[0062] Specifically, it includes the following steps.

[0063] S131: Based on the target region list output in step S12, generate an independent fine scanning task for each target region in the list.

[0064] Understandably, each task may include: scan range, safe starting height, second resolution, and scan path, etc.

[0065] The scanning range is defined by the two-dimensional projection bounding box of the target area to ensure that the scan can completely cover the area. The safe starting height is used to avoid collision risks. In a specific embodiment, it can be limited by adding a preset safety margin to the maximum height recorded in the target area. The safety margin can be 5-10mm.

[0066] Understandably, the second resolution should be significantly higher than the first resolution. In one specific embodiment, the point cloud density corresponding to the second resolution is approximately 50 to 200 points per square centimeter. By controlling the moving speed and scanning frequency of the scanning module, the average spacing of the point cloud formed on the object surface is 0.5 mm to 2 mm. Using this resolution, the geometric details, edges and surface features of the workpiece can be accurately captured.

[0067] The scanning path is designed for the shape and extent of each target area to obtain higher coverage density within the two-dimensional boundary of the target area. Typically, for regular areas, the scanning path can be a fine grid path. When using a fine grid path, the row spacing ΔY1 is much smaller than ΔY in the first stage, and the data sampling interval within the scan line is also correspondingly encrypted. For complex boundary areas, a contour tracking path or an adaptive spiral path can be used to ensure denser data acquisition of the boundary and feature areas.

[0068] S132: Perform a fine scanning task.

[0069] Specifically, detailed scanning tasks for each target area can be executed sequentially according to the task queue order, which may include: S1321: Drive the non-contact scanning sensor to a safe approach point outside the current target area boundary, and then lower it to a safe starting height calculated for that area.

[0070] S1322: Controls the non-contact scanning sensor to move along the scanning path planned for the area, while continuously acquiring data at a second resolution.

[0071] Furthermore, during or after scanning, the coverage integrity and noise level of the acquired point cloud can be calculated in real time. If local data is missing, a supplementary scanning path for that area can be planned and executed to ensure the integrity of the point cloud data for a single target area.

[0072] Furthermore, after completing the scanning of a target area, the high-density point cloud data collected in that area can be temporarily stored and associated with its target area ID and coordinate transformation information.

[0073] S133: Acquire and output the second point cloud data.

[0074] Understandably, the output second point cloud data is a collection of second point cloud data that corresponds one-to-one with the target area. Each data unit is a high-precision, high-density digital representation of the surface morphology of its corresponding target area.

[0075] S14: Generate the initial three-dimensional data based on the first point cloud data and each of the second point cloud data.

[0076] Understandably, the first point cloud data obtained in step S11 is fused with the multiple local second point cloud data obtained in step S13 to construct a unified, complete, and multi-resolution three-dimensional digital model of the workbench, providing a data foundation for subsequent accurate workpiece identification and inspection planning. Specifically, this includes the following steps.

[0077] S141: Coordinate system alignment.

[0078] It is understandable that the first point cloud data and each of the second point cloud data should be data under the same coordinate system, so as to facilitate fusion. Preferably, the first point cloud data and each of the second point cloud data are based on the inherent machine coordinate system of the coordinate measuring machine, so that no additional coordinate system transformation is required.

[0079] S142: Data fusion forms the initial three-dimensional data.

[0080] Specifically, this involves removing or marking as expired low-resolution point clouds within the target area already covered by the second point cloud data in the first point cloud data, and then inserting the second point cloud data units into the corresponding empty positions in the first point cloud data according to their spatial coordinates.

[0081] Furthermore, in the boundary region between high- and low-resolution data, distance-based weighted averaging algorithms or local surface reconstruction algorithms can be used to smoothly transition the point cloud density, avoiding visual or geometric abruptness at the data boundary and ensuring the overall consistency of the fused model.

[0082] Furthermore, after the data is fused to form the initial 3D data, statistical filtering is applied again to the initial 3D data to remove any isolated noise points that may exist after fusion.

[0083] Furthermore, in the transition area where high and low resolution data overlap, there may be points with excessively high density. By using voxelized grid filtering for appropriate downsampling, the amount of data can be optimized while maintaining shape accuracy.

[0084] Understandably, the initial 3D data is in the form of a uniform 3D point cloud dataset.

[0085] This 3D point cloud dataset forms a multi-resolution scene digital model: that is, in the background of the workbench and non-key areas, it maintains the first resolution to represent the scene layout, and in all identified workpiece target areas, it has a second resolution to represent specific geometric details.

[0086] Furthermore, prior to step S142, the system may include a step of positioning the second point cloud data unit to ensure that the high-resolution local data is precisely aligned with the global frame.

[0087] Specifically, within the corresponding target area, a corresponding low-resolution point cloud subset is extracted from the first point cloud data as a reference. The Iterative Closest Point (ICP) algorithm or a similar fine registration algorithm is used to calculate the optimal rigid transformation matrix of the second point cloud data relative to the first point cloud data subset. The rigid transformation matrix is ​​then applied to accurately locate and correct the second point cloud data unit.

[0088] S2: Process the initial three-dimensional data to segment and identify one or more workpiece objects on the worktable, and obtain information about each workpiece object related to the detection plan.

[0089] Understandably, step S2 involves analyzing the multi-resolution 3D model representing the entire workbench scene generated in step S1, thereby separating each independent physical entity, such as the workpiece, and extracting its geometric, spatial, and semantic features to provide structured input information for subsequent inspection planning.

[0090] The processing of the initial 3D data includes steps of data preprocessing and scene simplification to optimize the performance of subsequent segmentation and recognition.

[0091] In a specific embodiment, the preprocessing can be based on the target area identified in step S1, and the high-resolution point cloud data corresponding to the target area can be processed first, while ignoring the low-resolution area marked as background, thereby reducing the amount of data to be processed.

[0092] In addition, lightweight denoising, such as statistical outlier removal and normal estimation, is performed on the focused point cloud to prepare for geometric feature-based segmentation.

[0093] The segmentation can be performed using Euclidean distance-based clustering segmentation, normal / curvature-based segmentation, region growing segmentation, or deep learning segmentation.

[0094] In this process, the Euclidean distance-based clustering segmentation can set a distance threshold d2, and use a KD-Tree-accelerated Euclidean clustering algorithm to group points whose spatial distance is less than the distance threshold d2 into the same cluster. Each generated cluster is considered as a candidate workpiece object.

[0095] The normal / curvature-based segmentation can identify the contact boundaries between different workpieces by analyzing abrupt changes in the normal direction or curvature of the point cloud surface. The region growth segmentation starts from the seed point and grows regions according to criteria such as normal consistency and curvature similarity, thereby separating surfaces with different geometric features. At the same time, for repetitive workpieces of known types, a trained 3D deep learning model, such as PointNet++, can be used for instance segmentation.

[0096] Furthermore, after the segmentation is completed, information about each workpiece object related to the detection plan is extracted from the point cloud cluster of each segmented candidate workpiece object.

[0097] The information about each workpiece object related to the inspection plan includes its position, orientation, external dimensions, and shape category determination.

[0098] Specifically, the position can be calculated by taking the three-dimensional centroid (X, Y, Z) of the point cloud cluster as the approximate center of the workpiece. The attitude can be determined by performing principal component analysis (PCA) on the point cloud to obtain its principal axis direction, and then calculating the yaw, pitch, and roll angles of the workpiece in space, or fitting a minimum orientation bounding box (OBB) to describe its attitude. For the outer dimensions, the length, width, and height of the workpiece can be obtained based on the OBB or axis-aligned bounding box (AABB). For the shape category determination, the workpiece can be initially classified into major categories such as plate-shaped, shaft-shaped, block-shaped, rotating body, or irregular-shaped parts based on features such as the length-width-height ratio and point cloud distribution.

[0099] In addition, the information on each workpiece object related to the inspection plan may also include planar features, cylindrical / hole features, edges and contours, etc.

[0100] The planar feature is a large planar region detected within a point cloud cluster. The planar region can serve as a reference surface or a plane to be measured in subsequent measurements. The position, normal vector, and area of ​​the planar feature can also be recorded.

[0101] Among them, the cylindrical / hole features can be fitted to the cylinder using algorithms such as RANSAC to identify the features of the inner hole or outer cylinder, while recording its axis, diameter, depth and position.

[0102] Among them, edges and contours can extract regions with abrupt curvature changes in point clouds as possible edges or contour lines to be measured.

[0103] In addition, the information on each workpiece object related to the detection plan may also include features such as chamfers and spheres identified based on point cloud geometry.

[0104] Furthermore, after step S2, there is also a step of outputting a list of workpiece object information, which contains structured information of all successfully segmented and identified workpiece objects on the workbench, thereby facilitating multi-objective optimization and planning in step S3.

[0105] S3: Based on the information of each workpiece object, and according to the preset decision rules, generate a measurement scheme for at least one of the one or more workpiece objects; the measurement scheme includes at least selecting at least one measuring head from multiple types of measuring heads, and planning a detection path for the selected measuring head.

[0106] Understandably, before step S3, a measurement head performance parameter library is constructed for use when generating a measurement plan. The measurement head performance parameter library is pre-set with technical parameters of available measurement heads, including but not limited to measurement accuracy, repeatability, scanning speed, working distance, minimum contact diameter, applicable feature types such as planes, holes, curved surfaces, and physical dimensions.

[0107] Furthermore, it also includes building a stylus parameter library, including the size, length, and bulb diameter of available styluses.

[0108] Furthermore, it also includes constructing fixture and clamping information, including the known position and geometry of the fixture, for collision avoidance.

[0109] In this embodiment, the preset decision rules include obtaining a measurement scheme through a hybrid approach of rule-based filtering and multi-objective optimization, specifically including the following steps.

[0110] S31: Rule-based feature-measurement head matching is used for preliminary screening to form a candidate set.

[0111] Specifically, each inferred feature of each workpiece is traversed, and predefined expert rules are applied to select a set of feasible measurement head candidates.

[0112] The expert rules are mapping relationships between several characteristic conditions and results. For example, characteristic condition: deep hole with diameter < Φ2mm, mapping result: long rod small diameter ruby ​​trigger probe; characteristic condition: freeform surface profile + tolerance zone ≤ 0.01mm, mapping result: high density scanning probe, etc.

[0113] S32: Perform global multi-objective optimization based on the candidate set to generate a measurement scheme that includes measurement head selection and detection path.

[0114] Understandably, this step, based on the candidate feature set output by S31, simultaneously determines the following by solving a multi-objective optimization problem: the final measurement head and probe assigned to each feature; the detailed detection path planned for each feature (intra-feature path); and the optimal execution order and idle movement path (inter-feature path) among all feature detection tasks.

[0115] The multiple objectives include minimizing the total detection time, minimizing the measurement uncertainty, minimizing the collision risk index, and minimizing the system wear uniformity.

[0116] The total detection time includes the measurement time of each feature, the automatic switching time of the measuring head, and the idle travel time of the machine between features. The collision risk index is based on the spatial relationship between the measuring head size, the planned path and known obstacles, and the risk value evaluated through motion simulation or geometric calculation. The system wear balance is used to avoid overuse of individual measuring heads and extend the overall system life.

[0117] Understandably, the optimization process includes the following constraints: (1) Each feature must be assigned one available and technically feasible measurement head.

[0118] (2) The measurement path must meet the kinematic constraints of the coordinate measuring machine.

[0119] (3) The path must completely avoid known interference areas.

[0120] (4) For the same workpiece, the measurement sequence must meet the geometric tolerance principle of datum first.

[0121] Understandably, when performing global multi-objective optimization, improved genetic algorithms and priority-based greedy algorithms can be used to strengthen constraint backtracking.

[0122] The improved genetic algorithm uses the encoding of the measurement head and path order as genes, and the weighted sum of multiple objectives as the fitness function to find an approximate optimal solution through iterative evolution.

[0123] The priority-based greedy algorithm strengthens the constraints and prioritizes the processing of key and high-precision features through backtracking, while simultaneously optimizing the global path and backtracking adjustments when conflicts are encountered.

[0124] Understandably, when forming a measurement scheme, it is also necessary to generate a detailed sequence of probe points or a continuous scanning trajectory for each feature based on its selected measurement head type and preset measurement mode. For example, for a plane selected for measurement with a trigger probe, the optimization algorithm will determine the specific sampling point coordinates on its surface; for a curved surface selected for measurement with a scanning probe, the specific scanning line path and density will be planned.

[0125] In addition, the execution order of all feature detection tasks is optimized, and the globally optimal travel path for the measuring head to move between different feature positions is calculated. Understandably, this path planning needs to avoid known interferences, such as other workpieces and fixtures, and may be decomposed into a series of segmented straight lines or curves passing through safe space points.

[0126] S4: Control the coordinate measuring machine to execute the measurement scheme and complete the inspection of the workpiece.

[0127] Furthermore, before the coordinate measuring machine executes the measurement scheme, a step is included to verify and correct a significant feature of the workpiece to be measured using a measuring head.

[0128] Specifically, one or two significant features of the workpiece, such as edges and corners or known holes, are scanned or measured for verification. The actual positions obtained from the scan or measurement are compared with the theoretical positions in the plan. If the deviation is within the allowable range, such as ±0.1mm, then the coordinates of all measurement paths of the workpiece are subsequently compensated by overall translation and rotation. If the deviation exceeds the limit, a warning message is issued and manual intervention is requested to avoid unexpected losses.

[0129] Furthermore, when the coordinate measuring machine performs the aforementioned measurement scheme and makes long-distance idle movements, the measuring head should move along the global collision avoidance path planned in the scheme. At the same time, the current, load, and grating ruler feedback of each axis servo motor should be monitored in real time and compared with the normal motion model to predictively perceive potential slight contact and avoid unexpected losses.

[0130] S5: Based on the execution and results of this test, optimize the preset decision rules or related database.

[0131] Understandably, after step S4 is completed, a complete learning data package is collected and associated. This data package includes: the original measurement scheme generated in step S3, all actual execution paths and actual time recorded in step S4, actual usage records of the measuring head and probe, and all abnormal events, such as the type, location, and processing results of fine-tuning, retry, and collision warnings.

[0132] Understandably, new expert rules are extracted from the learning data to optimize decision-making rules.

[0133] Furthermore, based on the execution and results of this detection, this task can be used as a new training sample and added to the historical dataset. By periodically utilizing the accumulated sample data, the weight parameters, heuristic rules, or reward function of the multi-objective optimization model in step S3 can be retrained or fine-tuned to make the scheme generated by the model better.

[0134] The following description, in conjunction with specific embodiments, provides further details. Example

[0135] Please refer to the attached document. Figure 1 This specific embodiment is for intelligent detection of precision helical worm gears.

[0136] Coordinate measuring machine: LAISE HICCURA 9127; Measuring head system: High-precision five-axis trigger probe: RENISHAW PH20 (equipped with Φ1mm, Φ2mm, Φ3mm ruby ​​balls and cylindrical styluses); Line laser scanner: KEYENCE LJ-V; High-speed scanning probe: RENISHAW SP80; Workpiece: Precision helical worm gear.

[0137] Step S1: S11: Invoke the line laser scanner and scan the entire worktable at a safe height of 150mm with a spiral involute path at the first resolution (point cloud density 3 points / cm²). S12: In point cloud processing, dense regions and fixture regions of slender, rotating bodies are quickly identified by using height thresholds and density clustering.

[0138] S13: For the worm gear area, plan three scanning bands around the workpiece, and control the line laser to perform fine scanning at the second resolution (point cloud density 120 points / cm²), with special focus on the tooth surface area.

[0139] S14: Generate initial 3D data: a multi-resolution point cloud model containing approximately 2 million points, with the worm gear tooth surface having the highest point cloud density, followed by the journal and end face, and the background and fixture area being the sparsest.

[0140] Step S2: Based on color / intensity information, the worm gear point cloud (approximately 850,000 points) is segmented from the initial 3D data. Through Euclidean clustering and axial projection analysis, it is confirmed to be a single continuous object. Through principal component analysis (PCA) and cylinder fitting, the workpiece spatial axis is accurately calculated and defined as the reference axis Z' for subsequent measurements. Along the axis, the density and radius changes of the radial section of the point cloud are analyzed, and three journal cylindrical segments are automatically identified. Their diameters (Φ24.99mm, Φ29.98mm, Φ24.97mm) and positions are initially estimated. Point cloud curvature analysis extracts the helical tooth surface point cloud region (approximately 450,000 points), and the average helix angle (γ≈15°) is initially calculated.

[0141] Information output: Generates a structured workpiece object information, including: the datum axis equation, the start and end positions and estimated diameters of the three journals, the point cloud region index of the tooth surface, and the equations of both end faces.

[0142] Step S3: High-precision dimensional and shape evaluation is required. Rule: Trigger-type probe (PH20) is preferred, employing a hybrid strategy of "cross-sectional circle + generatrix scanning". Candidate styluses: Φ3mm styluse (high efficiency, suitable for Φ25 / 30mm bore diameters); Helical tooth surfaces: Dense point cloud evaluation of tooth profile and pitch is required. Rule: Compare the scanning probe (SP80) (contact type, high precision but slow speed) and the line laser scanner (non-contact, fast speed but easily affected by surface reflection). Based on tooth surface roughness and tolerance requirements, the rule engine recommends the scanning probe as the preferred option, with the line laser as an alternative.

[0143] End face: Flatness and runout measurement. Method: Trigger-type probe, 9-point grid measurement.

[0144] S32: With the objectives of minimizing total detection time (weight 0.5), minimizing overall uncertainty (weight 0.3), and minimizing collision risk (weight 0.2), the following solution is formulated: Phase 1: Using a PH20 trigger probe and a Φ3mm ruby ​​stylus, the measurements of three journals (4 cross-sectional circles for each journal, 12 points per circle, and 4 lines scanned along the generatrix) and both end faces were completed in one go.

[0145] Phase Two: Replace with an SP80 scanning probe and a Φ2mm sapphire stylus to perform a scan of the helical tooth surface. The planned path is as follows: divide the axis into 10 equally spaced sections. On each section, control the stylus to perform a zigzag scan along the tooth surface normal to cover the working surface of a single tooth.

[0146] Path planning: The optimized algorithm plans an efficient execution sequence: starting from one end of the workpiece, measuring end face A → journal 1 → journal 2 → journal 3 → end face B in sequence → changing the probe → scanning the tooth surface (from one end to the other). The idle path was verified by collision simulation, and all paths avoided V-blocks and centers.

[0147] Output: Generates an intelligent measurement process sheet containing 218 specific measurement actions and 3 probe replacement instructions.

[0148] Step S4: Execute the plan and record the data. The total execution time is 42 minutes, which saves 40% of the time compared to experienced engineers manually programming and executing the same task.

[0149] Step S5: Save the complete testing plan (including probe sequence, path parameters, and compensation strategy) as the case "Helical worm_Model A_V1.2" and add a new rule: For steel worm teeth with a helix angle of 10°-20°, the recommended scanning speed is ≤15mm / s.

[0150] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made using the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An intelligent detection method for a coordinate measuring machine, characterized in that, Includes the following steps: S1: Perform non-contact scanning on the worktable of the coordinate measuring machine to obtain initial three-dimensional data; S2: Process the initial three-dimensional data to identify one or more workpiece objects on the worktable and obtain information about each workpiece object related to the detection plan; S3: Based on the information of each workpiece object, a corresponding measurement plan is generated for each workpiece object according to the preset decision rules; the measurement plan includes at least measurement head selection and detection path planning; S4: Control the coordinate measuring machine to execute the measurement plan and complete the inspection of the workpiece object; S5: Based on the execution and results of this test, optimize the preset decision rules or related database.

2. The intelligent detection method as described in claim 1, characterized in that, Step S1 includes the following specific steps: S11: Control the non-contact scanning sensor to perform the first stage scan at the first resolution and acquire the first point cloud data of the worktable surface; S12: Based on the first point cloud data, identify one or more target areas on the workbench surface; S13: For each target area, control the non-contact scanning sensor to perform a second-stage scan with a resolution higher than the first resolution to acquire the corresponding second point cloud data; S14: Generate the initial three-dimensional data based on the first point cloud data and each of the second point cloud data.

3. The intelligent detection method as described in claim 2, characterized in that, Step S12 includes the following specific steps: S121: Point cloud preprocessing, performing outlier filtering on the first point cloud data; S122: Workbench reference extraction and background removal; S123: Point cloud clustering and segmentation and initial selection of target regions: A clustering algorithm based on Euclidean distance is used to group spatially adjacent points into multiple independent point cloud clusters, and each point cloud cluster is regarded as a candidate target region; S124: Target region screening and parameterization: Calculate the three-dimensional axis-aligned bounding box of each point cloud cluster, and at the same time, based on prior knowledge or preset rules, screen out clusters that do not conform to the workpiece characteristics. The parameterization refers to calculating and recording the key parameters of each candidate target region that passes the screening. S125: Output target region: Output target region is a structured list of target regions.

4. The intelligent detection method as described in claim 3, characterized in that, Step S13 includes the following specific steps: S131: Based on the target area list output in step S12, generate an independent fine scanning task for each target area in the list; S132: Perform a fine scanning task, specifically including: S1321: Drive the non-contact scanning sensor to a safe approach point outside the current target area boundary, and then lower it to a safe starting height calculated for that area; S1322: Control the non-contact scanning sensor to move along the scanning path planned for the area, while continuously acquiring data at a second resolution; S133: Obtain and output the second point cloud data.

5. The intelligent detection method as described in claim 4, characterized in that, Step S14 includes the following specific steps: S141: Coordinate system alignment; S142: Data fusion to form initial 3D data: Including in the first point cloud data, all low-resolution point clouds in the target area that have been covered by the second point cloud data are removed or marked as expired. Then, the second point cloud data units are inserted into the corresponding empty positions in the first point cloud data according to their spatial coordinates.

6. The intelligent detection method as described in claim 1, characterized in that, Step S3 includes the following specific steps: S31: Rule-based feature-measurement head matching is used for preliminary screening to form a candidate set; S32: Perform global multi-objective optimization based on the candidate set to generate a measurement scheme that includes measurement head selection and detection path; Among them, the multi-objective optimization includes minimizing the total detection time, minimizing the measurement uncertainty, minimizing the collision risk index, and minimizing the system wear uniformity.

7. The intelligent detection method as described in claim 6, characterized in that, The optimization process in step S32 includes the following constraints: (1) Each feature must be assigned one available and technically feasible measurement head; (2) The measurement path must meet the kinematic constraints of the coordinate measuring machine; (3) The path must completely avoid known interference areas; (4) For the same workpiece, the measurement sequence must meet the geometric tolerance principle of datum first.

8. The intelligent detection method as described in claim 6, characterized in that, In step S31, each inferred feature of each workpiece is traversed, and predefined expert rules are applied to filter out a set of feasible measurement head candidates.

9. The intelligent detection method as described in claim 1, characterized in that, Before the coordinate measuring machine performs step S4, there is also a step of using a measuring head to verify and correct a significant feature of the workpiece to be measured.

10. The intelligent detection method as described in claim 9, characterized in that, During the verification process, a verification scan or measurement is performed, and the actual position obtained by the scan or measurement is compared with the theoretical position in the measurement plan. If the deviation is within the allowable range, the coordinates of all measurement paths of the workpiece are subsequently compensated by overall translation and rotation. If the deviation exceeds the limit, a warning message is issued and manual intervention is required.

Images