Digital twinning processing method for special-shaped thin-wall cabin section based on laser scanning

Abstract

The invention relates to a digital twinning processing method for a special-shaped thin-wall cabin section based on laser scanning, which comprises the following steps of: calculating the working allowance, namely 1) acquiring data; 2) denoising; 3) simplification; 4) extracting a feature point cloud; 5) performing coarse registration; 6) performing fine registration; (7) the allowance is calculated; the wall thickness measurement comprises the following steps: 1) data acquisition; 2) denoising; 3) simplification; 4) extracting a measurement area; 5) performing European clustering segmentation; and 6) calculating the wall thickness. According to the method, automatic control over the machining process can be achieved on the basis of a computer, the precision of part measurement and calculation is ensured by means of automatic generation of a high-precision scanner and an NX program, and the problems of frequent size measurement and allowance calculation of an operator are solved. Meanwhile, the average value of the nearest points of the machining face is adopted in the allowance calculation result and the wall thickness measurement result, high robustness and high precision are achieved, and the final calculation result is not affected by tiny registration errors existing in registration.

Description

[0001] Methodology Area

[0002] The invention relates to a digital twin processing method for a special-shaped thin-wall cabin section based on laser scanning, and belongs to the field of digital twins.

[0003] Background Methods

[0004] The cabin blanks are cast, the front end thin wall thickness of the parts is 2 mm, the tolerance range is 0-0.2 mm, and the processing process is layered processing. Before each layer is processed, the wall thickness of the processed thin-walled surface is measured with an ultrasonic thickness gauge or a contact measuring device to calculate the processing allowance. For the wall thickness measurement of thin-walled surfaces, the traditional measurement method requires direct contact with the workpiece, which is prone to errors and damage. When the wall thickness is measured by measuring points, due to the roughness of the thin-walled surface before processing, there is an error between the measuring point distance and the uniform wall thickness, and the measurement time is long. Therefore, this type of parts urgently needs a digital twin processing method with high accuracy and high speed, which can be controlled by computer to automatically calculate the precise allowance and measure the wall thickness after processing. Summary of the invention

[0005] The purpose of the present invention is to find a point cloud data processing method based on laser scanning, to compensate for the measurement point error in the allowance calculation, to realize the accurate allowance calculation of the thin-walled surface of the blank to the designed surface, and to accurately measure the wall thickness of the thin-walled surface after processing to determine whether it meets the dimensional tolerance.

[0006] The present invention is achieved by the following method scheme:

[0007] A digital twin processing method for special-shaped thin-walled cabin sections based on laser scanning, including processing allowance calculation and wall thickness measurement:

[0008] (I) Calculation of machining allowance includes the following steps:

[0009] 1) Data acquisition: The point cloud data of the parts is collected by clamping the scanner with a robotic arm installed on the gantry of the machine tool;

[0010] 2) Denoising: Remove noise and outliers through statistical filtering, and smooth the point cloud data through Gaussian filtering;

[0011] 3) Streamlining: Improve the voxel streamlining algorithm based on the KD tree to reduce the amount of data while retaining part features;

[0012] 4) Extract feature point cloud: Based on the spatial distribution of the normal vector of the point cloud, randomly sample the point cloud, greatly reduce the data and extract the feature point cloud that retains the processing benchmark; extract the holes and outer cylindrical surface in the processing benchmark as feature point cloud. The selected part of the point cloud does not involve other processing parts.

[0013] 5) Rough registration: Using the obtained feature point cloud and the design model point cloud, K4PCS is used to achieve rough registration based on the processing benchmark as the registration framework, which has fast speed and high registration accuracy;

[0014] 6) Precise registration: The ICP registration method based on the symmetry of the objective function is used to achieve precise registration. At the same time, the centroid of the point cloud of the processed surface and the designed surface is coincident to achieve the coincidence of the processing origin, thus realizing the coordinate unification of the processed surface and the designed surface;

[0015] 7) Residue calculation: Use the processing point cloud file streamlined by KD tree voxels to calculate the residue, use the Euclidean matrix obtained by registration to transform the point cloud, calculate the distance between the processing point cloud and the nearest point of the design model through the KD tree, and calculate the arithmetic mean to get the residue. First, establish a KD tree for the design model point cloud, traverse the point cloud on the thin wall of the processing surface to the nearest point of the design model and calculate the distance to the nearest point, and calculate the arithmetic mean to achieve the residue calculation;

[0016] (ii) Wall thickness measurement, including the following steps:

[0017] 1) Data acquisition: The point cloud data of the parts is collected by clamping the scanner with a robotic arm installed on the gantry of the machine tool;

[0018] 2) Denoising: Remove noise and outliers through statistical filtering, and smooth the point cloud data through Gaussian filtering;

[0019] 3) Streamlining: Improve the voxel streamlining algorithm based on the KD tree to reduce the amount of data while retaining part features;

[0020] 4) Extract the measurement area: select the wall thickness part to be measured through the registerAreaPickingCallback() function in PCL;

[0021] 5) Euclidean clustering segmentation: The extracted area is divided into outer surface and inner surface point clouds. The two surface point clouds are parallel to each other and remain isolated, and each has dense and uniform points.

[0022] 6) Wall thickness calculation: Build a KD tree for the outer surface point cloud, then traverse the inner surface point cloud to the nearest point of the outer surface, calculate the distances of all the nearest points, and use the average of the nearest point distances to calculate the wall thickness.

[0023] When a layering process is completed, the wall thickness of the processed parts needs to be measured to verify the accuracy of the allowance calculation, and the measured wall thickness can be used to guide the next layering process.

[0024] The method of the present invention can realize automatic control of the machining process based on a computer, relying on high-precision scanners and automatic generation of NX programs, ensuring the accuracy of part measurement and calculation, and reducing the operator's frequent dimensional measurement and allowance calculation problems. At the same time, the results of the allowance calculation and wall thickness measurement are both based on the average value of the nearest point on the machining surface, which has strong robustness and high precision, and the slight registration error in the registration does not affect the final calculation result. It should be noted that when the operator selects the box, he should ensure that the selected part of the point cloud does not involve other machining parts, otherwise it will have a greater impact. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of the processing unit;

[0026] Figure 2 Scanning flowchart;

[0027] Figure 3 KD tree establishment flow chart;

[0028] Figure 4 KD tree nearest neighbor search flow chart;

[0029] Figure 5 Improved voxel downsampling process based on KD tree;

[0030] Figure 6 Point cloud preprocessing flow chart;

[0031] Figure 7 Schematic diagram of K4PCS registration principle;

[0032] Figure 8 Schematic diagram of the ICP principle based on the symmetry of the objective function;

[0033] Fig. 9 Flow chart of stock calculation under digital twin processing method;

[0034] Fig.10 Registration errors do not affect the theoretical graph of wall thickness calculation;

[0035] Fig.11 Schematic diagram of the Euclidean clustering segmentation principle;

[0036] Fig.12 Flowchart of wall thickness measurement under the digital twin processing method. DETAILED DESCRIPTION

[0037] The method of the present invention performs coordinate coincidence of the blank surface and the design surface based on the machining datum and the machining origin of the thin-walled surface of the part compartment, realizes the feature registration of the blank uniformly containing the design model, and establishes a KD tree of the design surface point cloud set, traverses the nearest point of the blank thin-walled surface to the design surface and calculates the Euclidean distance at this time, calculates the average value of the distance of all the nearest points in the blank surface as the actual machining allowance, and uses the distance of all the nearest points as the next layer of machining constraint conditions to judge whether the machining conditions are met; after layered machining, the KD tree is established by segmenting the outer surface point cloud set and the inner surface point cloud set through Euclidean clustering, and the arithmetic mean of the nearest distance of the inner and outer surfaces is used as the wall thickness size, and the variance and standard deviation of the nearest point are calculated to guide the machining quality.

[0038] The digital twin processing method is as follows:

[0039] 1. First, connect and install the automated processing unit consisting of a laser scanner, a robotic arm, and a processing machine. In this processing center, the robotic arm is installed in the reference correction unit of the gantry machine tool in a suspended working state. The movement of the robotic arm on the X and Y axes can be achieved by operating the gantry of the machine tool. Clamp the laser scanner on the end of the robotic arm, and program the motion trajectory of the robotic arm to achieve the scanning path planning of the part. Figure 1 .

[0040] In the production process, the blank is first positioned and clamped according to the processing reference, and the laser scanner scans the point cloud data. The processing allowance is calculated by the computer, and the NX program is automatically generated and sent to the machine tool. The machine tool starts processing. After processing, the scanner scans the part again, automatically calculates the wall thickness, determines whether it meets the processing requirements, and feeds back to the machine tool to complete a layered processing. Figure 2 .

[0041] After obtaining the point cloud, it is first necessary to denoise and filter the data to remove the influence of the environment and scanning errors, and different filtering methods are used according to different noise types; after filtering, the design model is discretized into point cloud data, and feature extraction is performed on the two point clouds at the same time. The processing reference uses the hole features of the front and rear end faces and the cylindrical surface of the cabin. The size of the point cloud normal vector at this type of feature is higher than the average normal vector of the entire point cloud model. Therefore, the present invention uses the normal vector space as the feature index to extract the holes in the processing reference and the outer cylindrical surface features.

[0042] In order to further improve the speed of model registration, the amount of calculated data is further reduced by downsampling. Voxels are cubes that divide three-dimensional space into extremely small units. Several or dozens of points in the cube are represented by voxel centers, which can better retain features and avoid dense or sparse data areas. KD tree is a data structure that divides point clouds into K-dimensional space, which can achieve efficient nearest neighbor search and greatly reduce calculation time. Therefore, the present invention improves the voxel uniform downsampling algorithm based on KD tree. In a voxel, KD tree will find the closest point to the voxel center to replace the voxel center. At the same time, with the help of KD tree, the voxel search efficiency and processing speed are more ideal. Figure 3 , 4 、5.

[0043] Using this point cloud preprocessing method, the more than two million scanned points are reduced to 10,000 to 100,000 selectable feature point clouds, and the required calculation time is less than one minute. Figure 6 .

[0044] 3. The ideal margin calculation scenario is that the blank point cloud uniformly includes the design model. The processing reference is used as the registration framework to achieve feature-level precision registration, focusing on the specific processing of thin walls. The processing origin is used as the registration condition to coincide the coordinates of the blank processing surface and the design surface.

[0045] The point cloud registration process is divided into coarse registration and fine registration. In the coarse registration stage, the feature point cloud extracted from the design model and the blank is used as the target point cloud and source point cloud for registration. The K4PCS registration algorithm extracts key points from the point cloud for matching, and based on the affine invariance of the four-point set between the key points, finds the optimal matching point pair of the two point clouds, and realizes rapid registration based on local features. Therefore, based on the machining benchmark, the present invention uses K4PCS coarse registration to achieve efficient and high-precision registration that takes into account both the machining benchmark as the global feature and the key points as the local feature. This method is used to achieve coarse registration of 10,000 points, and more than 90% overlap between the two can be achieved in more than ten seconds. Figure 7 .

[0046] In the precise registration stage, the traditional ICP algorithm uses the point-to-point distance metric as the objective function, that is, minimizing the sum of the Euclidean distances of the corresponding points between the source point cloud and the target point cloud. However, the blank and the design model in mechanical processing belong to asymmetric point cloud registration, and the two cannot completely overlap. Therefore, the present invention adopts an ICP algorithm based on the symmetry of the objective function to optimize the registration result, requiring the objective function of the sum of the Euclidean distances to be symmetrical between the source point cloud and the target point cloud and between the target point cloud and the source point cloud, so that the blank can evenly contain the design model. In addition, by aligning the machining origin of the machined thin-walled surface with the design surface, the scanning coordinate system and the machining coordinate system are overlapped at this time, see Figure 8 .

[0047] 4. Residue calculation: Fig. 9 , 10 As shown, the casting is positioned and clamped in the processing unit according to the processing datum. At this time, the robotic arm holds the laser scanner to fully scan the entire model of the blank according to the preset scanning trajectory. The point cloud data is exported by the scanning software VX. The point cloud at this time is preprocessed, and the discrete design model is designed. Both are simultaneously subjected to denoising, feature extraction and downsampling to obtain a feature point cloud that meets the registration conditions.

[0048] The design model and the scanned model are coarsely and finely registered with feature-level accuracy, and the specific processing parts are selected through PCL interactive operation. It should be ensured that the point cloud selected in the box is part of the processing part to avoid non-extracted processing parts, which will lead to the selection of other parts of the point cloud, otherwise it will cause margin calculation errors. The centroids of the two point clouds are overlapped to achieve the coincidence of the processing origins of the two. At this time, the processed thin-walled surface evenly includes the design surface, and the scanning coordinates and the processing coordinates completely coincide.

[0049] Under the above conditions, the uniform machining allowance can be calculated by establishing a KD tree to traverse the nearest point. First, a KD tree is established for the design model after the coordinates are overlapped, and the point cloud on the thin-walled surface is traversed to the nearest point of the design model and the distance is calculated. Finally, the distance a of the nearest point on all the thin-walled surface point clouds is calculated. i Find the arithmetic mean a 平均 In addition, the distance a after the two point cloud coordinates overlap is used as a constraint condition. i When <0, it is judged that the processing conditions are not met and subsequent processing is terminated.

[0050] The traditional method of calculating the allowance characterizes the wall thickness by measuring the distance of several points with a thickness gauge or a contact measuring device. This method will inevitably produce measurement point errors. The present invention calculates the distances of all the nearest points on the processed surface to the design surface, and the calculated allowance essentially compensates for the measurement point errors caused by the small number of measurement samples. At the same time, the accuracy of the allowance calculation is inevitably affected by the accuracy of the coincidence of the scanning coordinates and the processing coordinates, that is, the registration error generated when the scanning model and the design model are accurately registered at the feature level. The present invention uses the establishment of a KD tree to traverse the nearest point from the processed thin-walled surface to the design surface, calculates the distances of all the nearest points at this time, and avoids the allowance calculation error caused by the registration error by obtaining the allowance through homogenization.

[0051] 5. Wall thickness measurement: Fig.11 , 12As shown in the figure, after completing a layered processing, the wall thickness of the processed part needs to be measured to verify the accuracy of the allowance calculation, and the measured wall thickness is used to guide the next layered processing. As above, the part is scanned and preprocessed, but no registration operation is required. The thin-walled point cloud of the wall thickness at the measurement location is selected to obtain two surface point clouds. The two surface point clouds are parallel to each other, remain isolated, and each has dense and uniform points.

[0052] The Euclidean clustering method is used to segment the two surface point clouds (see Fig.11 ), the algorithm randomly selects a point as the center point by establishing a KD tree, and selects the nearest n points to the point as the cluster set, screens out the points that do not meet the nearest point distance d, and then performs this type of point selection method on the other points in the cluster set in turn, and iterates continuously until the number of points in the cluster set no longer increases, at which time a cluster segmentation is completed. Through this type of method, the thin-walled surface point cloud can be segmented into the outer surface and the inner surface point cloud. By establishing a KD tree for the outer surface point cloud, and then traversing the inner surface point cloud to the nearest point of the outer surface, the distance of all the nearest points is calculated, and the mean of the nearest point distance is used to represent the calculated wall thickness. In addition, by calculating the variance and standard deviation of the nearest point distance set, it has a certain guiding role in the processing quality.

Claims

1. A digital twin processing method for special-shaped thin-walled cabin sections based on laser scanning, characterized by: Including machining allowance calculation and wall thickness measurement: (I) Calculation of machining allowance includes the following steps: 1) Data acquisition: The point cloud data of the parts is collected by clamping the scanner with a robotic arm installed on the gantry of the machine tool; 2) Denoising: Remove noise and outliers through statistical filtering, and smooth the point cloud data through Gaussian filtering; 3) Streamlining: Improve the voxel streamlining algorithm based on the KD tree to reduce the amount of data while retaining part features; 4) Extract feature point cloud: Based on the spatial distribution of the normal vector of the point cloud, the point cloud is randomly sampled to greatly reduce the data while extracting the feature point cloud that retains the processing benchmark; 5) Rough registration: Using the obtained feature point cloud and the design model point cloud, K4PCS is used to achieve rough registration based on the processing benchmark as the registration framework; 6) Precise registration: The ICP registration method based on the symmetry of the objective function is used to achieve precise registration. At the same time, the centroid of the point cloud of the processed surface and the designed surface is coincident to achieve the coincidence of the processing origin, thus realizing the coordinate unification of the processed surface and the designed surface; 7) Residue calculation: Use the processing point cloud file streamlined by KD tree voxels to calculate the residue, use the Euclidean matrix obtained by registration to transform the point cloud, calculate the distance between the processing point cloud and the nearest point of the design model through the KD tree, and calculate the arithmetic mean to get the residue. First, establish a KD tree for the design model point cloud, traverse the point cloud on the thin wall of the processing surface to the nearest point of the design model and calculate the distance to the nearest point, and calculate the arithmetic mean to achieve the residue calculation; (ii) Wall thickness measurement, including the following steps: 1) Data acquisition: The point cloud data of the parts is collected by clamping the scanner with a robotic arm installed on the gantry of the machine tool; 2) Denoising: Remove noise and outliers through statistical filtering, and smooth the point cloud data through Gaussian filtering; 3) Streamlining: Improve the voxel streamlining algorithm based on the KD tree to reduce the amount of data while retaining part features; 4) Extract the measurement area: select the wall thickness part to be measured through the registerAreaPickingCallback() function in PCL; 5) Euclidean clustering segmentation: The extracted area is divided into outer surface and inner surface point clouds. The two surface point clouds are parallel to each other and remain isolated, and each has dense and uniform points. 6) Wall thickness calculation: Build a KD tree for the outer surface point cloud, then traverse the inner surface point cloud to the nearest point of the outer surface, calculate the distances of all the nearest points, and use the average of the nearest point distances to calculate the wall thickness.

2. According to claim 1, a digital twin processing method for special-shaped thin-walled cabin sections based on laser scanning is characterized by: When a layering process is completed, the wall thickness of the processed parts needs to be measured to verify the accuracy of the allowance calculation, and the measured wall thickness can be used to guide the next layering process.

3. According to claim 1, a digital twin processing method for special-shaped thin-walled cabin sections based on laser scanning is characterized by: When extracting feature point clouds for machining allowance calculation, the holes in the machining datum and the outer cylindrical surface are extracted as feature point clouds.

4. According to claim 1, a digital twin processing method for special-shaped thin-walled cabin sections based on laser scanning is characterized by: When extracting feature point clouds for machining allowance calculation, the selected part of the point cloud does not involve other machining parts.

Images