Small-scale workpiece three-dimensional reconstruction method based on laser line scanning and application

Abstract

The invention discloses a small-scale workpiece three-dimensional reconstruction method based on laser line scanning and application. The three-dimensional reconstruction method comprises the following steps: analyzing a key surface, and planning a scanning path and a scanning angle; performing multi-view scanning to obtain multi-view point cloud data; preprocessing to obtain preprocessed point cloud data; carrying out point cloud registration and splicing to obtain registration and splicing data; and generating a reconstructed high-precision triangular mesh surface as a three-dimensional reconstruction model. According to the technical scheme of multi-view scanning path planning, collection of point cloud data is more accurate and reliable, and parameters do not need to be frequently adjusted due to working condition changes; meanwhile, a multi-view point cloud registration and splicing mechanism and a point cloud feature extraction architecture of a U-net architecture are adopted, more details are reserved for surface geometric features of workpiece point cloud, meanwhile, a curved surface is implicitly expressed through a symbol distance function, and the method has remarkable advantages for digital-analog comparison of a workpiece with a complex profile. Three-dimensional shape matching and error evaluation can be efficiently and accurately carried out.

Description

Technical Field

[0001] The present invention relates to the field of precision machining technology, and particularly to a three-dimensional reconstruction method and application of small-scale workpieces based on laser line scanning. Background Art

[0002] In the field of industrial precision machining, real-time control of the machining quality of workpieces is crucial. This is because precision machining usually involves dimensional tolerances at the micron or even nanometer level, and any slight deviation may have a significant impact on the functionality and service life of the workpiece. With the development of modern manufacturing towards high precision and high efficiency, the demand for machining quality monitoring is increasing continuously. However, due to various factors such as the process limits of machining equipment, environmental impacts, tool wear, and operation errors, quality deviations are inevitable during the production process. If such deviations cannot be detected and corrected in a timely manner, it may lead to the unqualified of batch products, causing huge economic losses and resource waste.

[0003] The method for realizing workpiece machining quality monitoring can plan the scanning path and angle according to the geometric characteristics of the workpiece to ensure that all key surfaces are covered and blind spots are avoided. Scan each view one by one to generate local point cloud data. Through point cloud registration, align the point cloud data of each view to a unified coordinate system. Import the CAD digital model generated in the design stage as a comparison benchmark. Use the global registration algorithm to accurately align the reconstructed model with the design model to ensure that both are in the same coordinate system. Calculate the deviation between the reconstructed model and the design model to generate the magnitude and distribution of the error.

[0004] For example, Chinese invention patent CN114820750A provides a method for collecting point clouds during the machining process of a numerically controlled machine tool, determining workpiece machining status data, and generating a 3D model in real time based on the parameter data of the machining process, and comparing it with a reference model.

[0005] In the above technical solution, a binocular camera is used for sparse point cloud collection during point cloud collection, and then a dense point cloud is generated through an algorithm. The accuracy of the binocular camera mainly depends on the resolution and calibration accuracy of the camera. For precision workpieces with complex curved surfaces or tiny features, the generated sparse point cloud may not meet the requirements of micron-level or higher accuracy, thereby affecting the quality of the dense point cloud. At the same time, in a dynamic acquisition environment (such as a vibrating or moving workpiece during machine tool machining), it is easily affected by motion blur, reducing the quality of the point cloud.

[0006] In addition, there are also some technical solutions that use laser line scanning for three-dimensional monitoring, but they can only monitor the dimensions of some key points and lack global monitoring of the workpiece, which is not conducive to improving the machining quality. Summary of the Invention

[0007] Aiming at the deficiencies of the prior art, the purpose of the present invention is to provide a three-dimensional reconstruction method and application for small-scale workpieces based on laser line scanning.

[0008] To achieve the foregoing invention purpose, the technical solutions adopted by the present invention include:

[0009] In the first aspect, the present invention provides a three-dimensional reconstruction method for small-scale workpieces based on laser line scanning, which includes:

[0010] Analyze the key surfaces according to the workpiece processing model, and plan the scanning path and scanning angle based on the key surfaces;

[0011] Based on the scanning path and scanning angle, use a laser line scanning camera to perform multi-view scanning on the workpiece to obtain multi-view point cloud data;

[0012] Preprocess the multi-view point cloud data to obtain preprocessed point cloud data;

[0013] Adopt the iterative closest point matching method to perform point cloud registration and splicing on the data of different views in the preprocessed point cloud data to obtain registered and spliced data;

[0014] Adopt a three-dimensional reconstruction algorithm based on the U-Net architecture, use the registered and spliced data as input, retain the multi-level features of the point cloud through skip connections, based on the core features output by the decoder, convert the point cloud data into a continuous implicit signed distance representation through a multi-layer perceptron and interpolation technology, and finally use a surface extraction algorithm to extract the isosurface from the implicit field to generate a reconstructed high-precision triangular mesh surface as the three-dimensional reconstruction model of the workpiece.

[0015] In the second aspect, the present invention also provides a processing quality monitoring method for small-scale workpieces based on laser line scanning, which includes:

[0016] Adopt the above three-dimensional reconstruction method for small-scale workpieces to perform three-dimensional reconstruction on the workpiece to be detected to obtain a three-dimensional reconstruction model;

[0017] Based on the difference between the three-dimensional reconstruction model and the workpiece processing model of the workpiece to be detected, evaluate the processing error of the workpiece to be detected.

[0018] In the third aspect, the present invention also provides a workpiece machining method, which includes:

[0019] Adopt the above processing quality monitoring method to monitor the workpiece to be detected to obtain the processing error;

[0020] Based on the processing error, adjust the compensation parameters.

[0021] In the fourth aspect, the present invention also provides a three-dimensional reconstruction system for small-scale workpieces based on laser line scanning, which includes:

[0022] A scanning planning module, configured to analyze key surfaces according to a workpiece machining model and plan scanning paths and scanning angles based on the key surfaces;

[0023] A multi-view scanning module, configured to drive a laser line scanning camera to perform multi-view scanning on the workpiece based on the scanning paths and scanning angles, and obtain multi-view point cloud data;

[0024] A preprocessing module, configured to preprocess the multi-view point cloud data to obtain preprocessed point cloud data;

[0025] A registration and stitching module, configured to perform point cloud registration and stitching on data from different views in the preprocessed point cloud data by using the iterative closest point matching method to obtain registered and stitched data;

[0026] A 3D reconstruction module, configured to adopt a 3D reconstruction algorithm based on the U-Net architecture, use the registered and stitched data as input, retain multi-level features of the point cloud through skip connections, based on the kernel features output by the decoder, convert the point cloud data into a continuous implicit signed distance representation through a multi-layer perceptron and interpolation technology, and finally extract an isosurface from the implicit field by using a surface extraction algorithm to generate a reconstructed high-precision triangular mesh surface as the 3D reconstruction model of the workpiece.

[0027] Based on the above technical solutions, compared with the prior art, the beneficial effects of the present invention at least include:

[0028] The present invention provides a method for monitoring the machining quality of workpieces. By combining the selected laser line scanning camera during the production process, it can realize the error analysis of workpieces with complex surfaces or high-precision requirements. Through the technical solution of multi-view scanning path planning, the acquisition of point cloud data is more accurate and reliable, and there is no need to frequently adjust parameters due to changes in working conditions. At the same time, by adopting the mechanism of multi-view point cloud registration and stitching and the point cloud feature extraction architecture of the U-net architecture, more details of the surface geometric features of the workpiece point cloud are retained. At the same time, the surface is implicitly represented by a signed distance function, which has significant advantages in the digital model comparison of workpieces with complex surfaces and can perform 3D shape matching and error evaluation efficiently and accurately.

[0029] The above description is only an overview of the technical solutions of the present invention. In order to enable those skilled in the art to understand the technical means of the present application more clearly and implement it according to the content of the specification, the following takes the preferred embodiments of the present invention and combines detailed drawings to illustrate as follows. Brief Description of the Drawings

[0030] Figure 1 It is a schematic flowchart of a machining quality monitoring method provided by a typical embodiment of the present invention. Detailed Embodiments

[0031] In view of the deficiencies in the prior art, through long-term research and a large number of practices, the inventors of this case were able to propose the technical solution of the present invention. The following will further explain the technical solution, its implementation process, principles, etc.

[0032] In the following description, many specific details are set forth in order to provide a thorough understanding of the present invention. However, the present invention may be practiced in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited by the specific embodiments disclosed below.

[0033] Moreover, relative terms such as "first" and "second" are only used to distinguish one component or method step with the same name from another, and do not necessarily require or imply any actual relationship or order between these components or method steps.

[0034] See Figure 1 As shown, an embodiment of the present invention provides a three-dimensional reconstruction method for small-scale workpieces based on laser line scanning, which includes the following steps:

[0035] Analyze the key surfaces according to the workpiece processing model, and plan the scanning path and scanning angle based on the key surfaces;

[0036] Based on the scanning path and scanning angle, use a laser line scanning camera to perform multi-view scanning on the workpiece to obtain multi-view point cloud data;

[0037] Preprocess the multi-view point cloud data to obtain preprocessed point cloud data;

[0038] Adopt the method of iterative closest point matching to perform point cloud registration and splicing on the data of different views in the preprocessed point cloud data to obtain registered and spliced data;

[0039] Adopt a three-dimensional reconstruction algorithm based on the U-Net architecture. The encoder extracts multi-scale features (such as 16× / 8× / 4× downsampling) through 3D sparse convolution. Taking the registered and spliced data as input, fuse the low-level geometric details (edges / curvatures) with the high-level semantic features through skip connections, retain the multi-level features of the point cloud, based on the core features output by the decoder, convert the point cloud data into a continuous implicit signed distance representation through a multi-layer perceptron, map the spatial coordinates to the signed distance values, and interpolation techniques (such as trilinear interpolation) generate smooth transitions in the sparse point cloud region. Finally, use a surface extraction algorithm to extract the isosurface from the implicit field to generate a reconstructed high-precision triangular mesh surface as the three-dimensional reconstruction model of the workpiece.

[0040] In some embodiments, the key surfaces include a machining reference surface, a functional surface, and a complex surface; the scanning path and scanning angle cover all of the key surfaces. Among them, the determination method of the complex surface can refer to various definitions of surface complexity in the prior art, which is common knowledge in the field of machining.

[0041] In some embodiments, the workpiece machining model is represented as M = {v k , f l}, where v represents the workpiece machining model and f represents the points that make up the surface of the workpiece machining model; the key surface is represented as a collection of key points, that is, Q = {q j};

[0042] The constraint conditions for the planning of the scanning path and scanning angle are expressed as:

[0043]

[0044] The planning goal is to minimize the number of scanning positions, and the number of scanning positions is expressed as:

[0045]

[0046] Among them, V = [v ij , representing the visibility matrix; v ij = 1 indicates that the key point q j can be observed by the scanning position P i , v ij = 0 indicates a blind area; Z represents the number of scanning positions; xi ∈ {0, 1}, representing a decision variable. When x i = 1, it indicates that the scanning position P i is selected; k, l, m, n represent natural numbers, where m represents the total number of key points and n represents the total number of scanning positions.

[0047] In some embodiments, the preprocessing includes noise removal, outlier removal, and outlier correction, where the outlier correction includes correcting the detected outlier points using interpolation or least squares fitting.

[0048] In some embodiments, the process of point cloud registration is expressed as:

[0049] P = p1, p2,... p r

[0050] Q = q1, q2,... q s

[0051] Or Q' = RQ + t

[0052] R represents the rotation matrix; t represents the translation vector; Q′ represents the point cloud after registration of point cloud Q, and q. represents the data points in point cloud Q; P represents another point cloud, and p. represents the data points in point cloud P; r represents the number of data points in point cloud P; s represents the number of data points in point cloud Q; b represents the serial number;

[0053] The point cloud registration is performed by the ICP algorithm, continuously iterating for nearest point matching, and continuously updating the rotation matrix R and the translation vector t by minimizing the distance, so that the distance between point clouds is minimized.

[0054] In some embodiments, the reconstruction process of the three-dimensional reconstruction model includes an encoding stage and a decoding stage. The encoding stage includes:

[0055] Step 1: Divide the registered and spliced data into multiple voxels, and identify the data points in any voxel, expressed as:

[0056] V v ={p u |p u ∈P, p u ∈Voxel v}

[0057] where V v represents the set of data points in the voxel; Voxel v represents the voxel, and u, v represent the serial numbers;

[0058] Step 2: Perform coordinate transformation on the data points in any voxel to convert them into in-voxel coordinates, expressed as:

[0059]

[0060] where, represents the in-voxel coordinates; o v is the origin of the voxel Voxel v ;

[0061] Step 3: Connect the normal vector information about the laser scanning angle with the in-voxel coordinates into a 6D feature vector, expressed as:

[0062]

[0063] where z i represents the 6D feature vector, and C i represents the normal vector information;

[0064] Step 4: Use the residual encoder network to extract voxel features, expressed as:

[0065] f v =ResPointNet(zi |p u ∈V v )

[0066] where f v represents the voxel feature; ResPointNet(·) represents the residual encoder;

[0067] Step 5: Apply the sequential sparse convolutional sublayer in the max pooling layer to gradually process the grid features, quantize each data point into voxel-level features, and generate a sparse voxel grid, denoted as:

[0068]

[0069] where represents the feature of voxel Voxel v at the l-th layer, represents the feature of voxel Voxel v at the (l + 1)-th layer; N(v) represents the neighborhood of voxel Voxel v ; W k and B represent the convolutional kernel weights and biases; σ(·) represents the activation function.

[0070] In some embodiments, the decoding stage performs feature splicing with the feature layers of different levels in the encoding stage through skip connections, denoted as:

[0071]

[0072] where is the low-level feature output by the encoding stage, is the high-level feature output by the encoding stage, represents the spliced feature, and w represents the serial number;

[0073] The decoding stage specifically includes:

[0074] Based on the spliced feature and combined with the in-voxel coordinate plocal, an input vector x is formed:

[0075]

[0076] Through a multi-layer perceptron and interpolation technology, the spliced input is mapped to the signed distance value SDF(x), and the point cloud data is converted into a continuous implicit signed distance representation:

[0077] h1 = σ(W1x + b1)

[0078] h2 = σ(W2h1 + b2)

[0079] …

[0080] SDF(x) = (W n h n-1 + b n )

[0081] Wherein, W i , b i represent learnable parameters; h i represents the i-th perception layer.

[0082] Use the Marching Cubes surface extraction algorithm to traverse all voxels, calculate the SDF values of 8 vertices, and determine whether there is an isosurface of 0 inside the voxel according to the sign change of the vertex SDF values, extract all isosurfaces, and generate a reconstructed high-precision triangular mesh surface.

[0083] The embodiment of the present invention also provides a method for monitoring the machining quality of small-scale workpieces based on laser line scanning, which includes the following steps:

[0084] Use the three-dimensional reconstruction method of small-scale workpieces provided by any of the above embodiments to perform three-dimensional reconstruction on the workpiece to be detected, and obtain a three-dimensional reconstruction model;

[0085] Based on the difference between the three-dimensional reconstruction model and the workpiece machining model of the workpiece to be detected, evaluate the machining error of the workpiece to be detected.

[0086] The embodiment of the present invention also provides a workpiece machining method, which includes the following steps:

[0087] Use the above machining quality monitoring method to monitor the workpiece to be detected and obtain the machining error;

[0088] Based on the machining error, adjust the compensation parameters.

[0089] The embodiment of the present invention also provides a three-dimensional reconstruction system for small-scale workpieces based on laser line scanning, which includes:

[0090] A scanning planning module, configured to analyze key surfaces according to the workpiece machining model, and plan scanning paths and scanning angles based on the key surfaces;

[0091] A multi-view scanning module, configured to drive a laser line scanning camera to perform multi-view scanning on the workpiece based on the scanning path and scanning angle, and obtain multi-view point cloud data;

[0092] A preprocessing module, configured to preprocess the multi-view point cloud data to obtain preprocessed point cloud data;

[0093] A registration and splicing module, configured to perform point cloud registration and splicing on different view data in the preprocessed point cloud data by using the iterative closest point matching method to obtain registration and splicing data;

[0094] A 3D reconstruction module, which is used to adopt a 3D reconstruction algorithm based on the U-Net architecture. Using the registered and spliced data as input, it retains the multi-level features of the point cloud through skip connections. Based on the core features output by the decoder, it converts the point cloud data into a continuous implicit signed distance representation through a multi-layer perceptron and interpolation technology. Finally, it uses a surface extraction algorithm to extract the isosurface from the implicit field and generate a reconstructed high-precision triangular mesh surface as the 3D reconstruction model of the workpiece.

[0095] The technical solutions of the present invention will be further described in detail below through several embodiments in conjunction with the accompanying drawings. However, the selected embodiments are only used to illustrate the present invention and do not limit the scope of the present invention.

[0096] Embodiment 1

[0097] This embodiment exemplifies a method for monitoring the quality of workpiece processing, which specifically includes the following steps:

[0098] First, according to the geometric characteristics and complexity of the workpiece, its key surfaces are analyzed, including areas such as machining datum surfaces, functional surfaces, and complex curved surfaces. This step is analyzed through the CAD design drawing or preliminary 3D model of the workpiece, and all areas that the scanning device needs to cover are clarified. The input is the 3D model M = {v k , f l} of the workpiece, and the key point set Q = {q j} represents the key points on the workpiece surface, usually selected from feature areas (such as complex curved surfaces, functional surfaces, machining datum surfaces, etc.). The field of view model of the scanning device is the scanning plane, which is determined by the position P i of the initial line laser of the scanner and the scanning direction. The direction is fixed as the normal direction of the laser line, and the scanning range is the effective length of the laser line and the working distance range, usually a rectangular projection area. For possible acquisition blind spots, it is necessary to optimize the path planning by calculating the visibility matrix V = [v ij of the device scanning path, where v ij = 1 indicates that the key point q j can be observed from the scanning position P i , and vij = 0 indicates a blind spot.

[0099] Define the decision variable x i ∈ {0, 1}, indicating whether to select the scanning position P i , x i = 1 means selecting this position, and x i = 0 means not selected.

[0100] Therefore, the final path optimization goal is to minimize the number Z of scanning positions:

[0101]

[0102] The constraint is that each key point q j is covered by at least one scanning position P i :

[0103]

[0104] The obtained Z is the initial position of the scanner for the optimal scanning path.

[0105] According to the planned path and angle, a laser line-scanning camera is used to perform multi-view scanning on the workpiece. For special parts such as areas with large curvature or local depressions, by increasing additional scanning angles or adjusting the position of the laser line-scanning camera, ensure that each key surface is covered.

[0106] Perform preliminary preprocessing on the multi-view point cloud data collected. It mainly includes:

[0107] Denoising: Use statistical filtering or bilateral filtering methods to remove noise points in the scanning

[0108] Outlier removal: Use the point cloud density threshold to judge and remove outliers.

[0109] Anomaly point correction: Correct the detected abnormal points using interpolation or least squares fitting.

[0110] Point cloud registration: Use the Iterative Closest Point (ICP) algorithm to align the point cloud data from different views to a unified coordinate system. P = p1, p2,... p r and Q = q1, q2,... q s : respectively represent the point clouds collected from two different views. The point sets P and Q each contain r and s points. For each point q in the point cloud Q j , we hope to transform it through a rotation matrix R (a 3×3 matrix) and a translation vector t (a three-dimensional vector) so that the transformed point cloud coincides with the point cloud P as much as possible. The transformation formula is:

[0111]

[0112] q′ b is the transformed point. Through the ICP algorithm, continuously perform the closest point matching iteratively, and continuously update the rotation matrix R and the translation vector t by minimizing the distance. Finally, make the point cloud Q′ = RQ + t, complete the alignment with the point cloud P, and splice the aligned point clouds.

[0113] Use a network similar to the U-Net architecture. Given the finest-level input point cloud and voxel size, divide the point cloud into voxel grids V. The point set within each voxel is defined as:

[0114] Vv = {p u | p u ∈ P, p u ∈ Voxel v}

[0115] For the voxel Voxel v , first identify the set of points located within each voxel and encode its features through the following steps:

[0116] Convert the global coordinates of the input points to local coordinates within the voxel as the input. For most datasets, the normal vector n of each point i is required as additional information to eliminate direction ambiguity. This information does not have to be very accurate and can usually be easily obtained from the sensor position. Convert the global coordinates p of the point i to local coordinates:

[0117]

[0118] where o v is the origin of the voxel Voxel v .

[0119] Concatenate the normal vector information n i with the local coordinates to form a 6D feature vector and feed it into the point encoder:

[0120]

[0121] For each voxel, run a PointNet-like residual encoder network to extract the voxel features and pool all the points within it into a feature vector:

[0122] f v = ResPointNet(z i | p u ∈ V v )

[0123] where f v is the feature vector of the voxel Voxel v .

[0124] Then, apply a max pooling layer and sequential sparse convolutional layers in the middle to gradually process the mesh features, quantize the per-point information into voxel-level features, and generate a sparse voxel grid.

[0125]

[0126] where:

[0127] represents the feature of the voxel Voxel v at the l-th layer, and N(v) represents the voxel Voxelv In the field of, W k and B are the convolutional kernel weights and biases, and σ is the activation function. Deeper layers have a larger receptive field, covering a larger geometric area, thus achieving dimensionality reduction of the spatial resolution.

[0128] The decoder upsamples by using nearest neighbor interpolation or a transposed convolution layer. For each voxel j , an upsampling operation is performed on it. In order to fuse low-level and high-level information, the decoder performs feature concatenation with the corresponding feature layers of the encoder through skip connections:

[0129]

[0130] where is the low-level feature output by the encoding stage, is the high-level feature output by the encoding stage, represents the concatenated feature.

[0131] As the inverse process of encoding, the decoder of the sparse U-Net also consists of several convolutional layers with nearest neighbor-based upsampling, and skip connections are added to encourage the fusion of low-level and high-level information.

[0132] Based on the concatenated feature output by the decoder, combined with the local coordinate plocal, the signed distance value of the point is generated through a multi-layer perceptron and interpolation techniques.

[0133] Finally, the Marching Cubes surface extraction algorithm is used to extract the isosurface from the implicit field, generating a reconstructed high-precision triangular mesh surface.

[0134] The reconstructed 3D model is compared with the CAD design model, the machining error is analyzed, and the areas with large deviations are marked, providing a basis for subsequent process adjustment.

[0135] The above technical solution provides a high-precision workpiece quality assessment method based on a laser line-scanning camera. The processed workpiece is scanned with high precision from multiple perspectives by the line-scanning camera to obtain the original high-fidelity point cloud data. The collected point cloud data is denoised, registered, and stitched through advanced point cloud processing methods to generate a dense and complete workpiece point cloud model. On this basis, a three-dimensional reconstruction algorithm based on the U-Net architecture that can preserve the original geometric features of the point cloud is adopted. The point cloud and the normal vector are used as inputs, and the multi-level features of the point cloud are retained through skip connections. Based on the kernel features output by the decoder, the point cloud data is converted into a continuous implicit signed distance representation through a multi-layer perceptron and interpolation technology. Finally, the isosurface is extracted from the implicit field by the Marching Cubes surface extraction algorithm to generate a reconstructed high-precision triangular mesh surface. By comparing and analyzing the reconstructed three-dimensional model with the designed digital model, the machining error of the workpiece can be accurately detected.

[0136] In the process of point cloud acquisition, this method makes full use of the high-resolution characteristics of the line-scanning camera and combines multi-perspective scanning path planning to ensure the complete coverage of the point cloud data in key areas and complex surfaces. In the point cloud processing stage, the iterative closest point matching algorithm is used for registration to effectively eliminate the multi-perspective stitching error and retain the fine details on the workpiece surface. The three-dimensional reconstruction algorithm fuses the low-level information and high-level information through the U-Net architecture network to retain the point cloud features. At the same time, the decoder uses a multi-layer perceptron and interpolation technology to optimize the mesh generation and surface smoothing process to ensure the accuracy and coherence of the reconstructed model and restore the true geometric features of the workpiece.

[0137] It should be understood that the above embodiments are only used to illustrate the technical concept and characteristics of the present invention, and their purpose is to enable those who are familiar with this technology to understand the content of the present invention and implement it accordingly, and it cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit and essence of the present invention should be covered within the protection scope of the present invention.

Claims

1. A three-dimensional reconstruction method for small-scale workpieces based on laser line scanning, characterized in that, Including: Analyze the key surfaces according to the workpiece processing model, and plan the scanning path and scanning angle based on the key surfaces; Based on the scanning path and scanning angle, use a laser line scanning camera to perform multi-view scanning on the workpiece to obtain multi-view point cloud data; Preprocess the multi-view point cloud data to obtain preprocessed point cloud data; Adopt the iterative closest point matching method to perform point cloud registration and stitching on the data of different views in the preprocessed point cloud data to obtain registered and stitched data; Adopt a 3D reconstruction algorithm based on the U-Net architecture, use the registered and stitched data as input, retain the multi-level features of the point cloud through skip connections, based on the kernel features output by the decoder, convert the point cloud data into a continuous implicit signed distance representation through a multi-layer perceptron and interpolation technology, and finally use a surface extraction algorithm to extract the isosurface from the implicit field to generate a reconstructed high-precision triangular mesh surface as the 3D reconstruction model of the workpiece.

2. The three-dimensional reconstruction method for small-scale workpieces according to claim 1, characterized in that The key surfaces include the processing reference surface, the functional surface, and the complex surface; The scanning path and scanning angle cover all the key surfaces.

3. The three-dimensional reconstruction method of small-scale workpieces according to claim 2, characterized in that The workpiece machining model is represented as M = {v k , f1}, where v represents the points in the workpiece machining model, and f represents the points that form the surface of the workpiece machining model; the key surface is represented as the set of key points, that is, Q = {q j}; The constraint conditions for the planning of the scanning path and scanning angle are expressed as: The planning objective is to minimize the number of scanning positions, and the number of scanning positions is expressed as: Among them, V = [v ij , representing the visibility matrix; v ij = 1 indicates that the key point q j can be observed from the scanning position P i , v ij = 0 indicates the blind area; Z represents the number of the scanning positions; xi ∈ {0, 1}, representing the decision variable, when x i = 1, it indicates to select the scanning position P i ; k, l, m, n represent natural numbers, where m represents the total number of the key points, and n represents the total number of the scanning positions.

4. The three-dimensional reconstruction method for small-scale workpieces according to claim 1, characterized in that, The preprocessing includes noise removal, isolated point removal, and abnormal point correction, where the abnormal point correction includes correcting the detected abnormal points by interpolation or least squares fitting.

5. The three-dimensional reconstruction method for small-scale workpieces according to claim 1, wherein The process of point cloud registration is expressed as: P = p1, p2,... p r Q = q1, q2,...q s or Q′ = RQ + t Where, R represents the rotation matrix; t represents the translation vector; Q′ represents the point cloud after registration of point cloud Q, q. represents the data points in point cloud Q; P represents another point cloud, p. represents the data points in point cloud P; r represents the number of data points in point cloud P; s represents the number of data points in point cloud Q; b represents the serial number; The point cloud registration is performed by the ICP algorithm, continuously iterating to perform closest point matching, and continuously updating the rotation matrix R and the translation vector t by minimizing the distance to minimize the distance between point clouds.

6. The three-dimensional reconstruction method for small-scale workpieces according to claim 1, characterized in that The reconstruction process of the 3D reconstruction model includes an encoding stage and a decoding stage. The encoding stage includes: Step 1, divide the registered and stitched data into multiple voxels, and identify the data points in any voxel, expressed as: V v = {p u | p u ∈ P, p u ∈ Voxel v} where V v represents a set of data points in the voxel; Voxel v represents the voxel, and u, v represent serial numbers; Step 2, perform coordinate transformation on the data points within any voxel to convert them into coordinates within the voxel, expressed as: Among them, represents the coordinates within the voxel; o v is the origin of the voxel v ​ Step 3, connect the normal vector information about the laser scanning angle with the coordinates within the voxel to form a 6D feature vector, expressed as: Among them, z i represents the 6D feature vector, and C i represents the normal vector information; Step 4, use a residual encoder network to extract voxel features, expressed as: f v = ResPointNet(z i | p u ∈ V v ) Among them, f v represents the voxel feature; ResPointNet(·) represents the residual encoder; Step 5, apply the sequential sparse convolution sublayer in the max pooling layer to gradually process the grid features, quantize each data point into voxel-level features, and generate a sparse voxel grid, expressed as: Among them, represents the voxel v feature at the l-th layer, represents the feature of voxel Voxelv at the (l + 1)-th layer; N(v) represents the neighborhood of voxel v ; W k and B represent the convolutional kernel weights and biases; σ(·) represents the activation function.

7. The three-dimensional reconstruction method for small-scale workpieces according to claim 6, characterized in that, The decoding stage performs feature stitching with the feature layers of different levels of the encoding stage through skip connections, expressed as: Among them, is the low-level feature output by the encoding stage, is the high-level feature output by the encoding stage, represents the concatenated feature; The decoding stage specifically includes: Based on the splicing feature Combined with the in-voxel coordinates plocal, an input vector x is formed: Through a multi-layer perceptron and interpolation technology, map the stitched input to the signed distance value SDF(x), and convert the point cloud data into a continuous implicit signed distance representation: h1 = σ(W1x + b1) h2 = σ(W2h1 + b2) … SDF(x) = (W n h n-1 + b n ) Among them, W i , b i represent learnable parameters; h i represents the i-th perception layer. Using the Marching Cubes surface extraction algorithm to traverse all voxels, calculate the SDF values of the 8 vertices, determine whether there is an isosurface of 0 inside the voxel according to the sign change of the vertex SDF values, extract all isosurfaces, and generate a reconstructed high-precision triangular mesh surface.

8. A method for monitoring the machining quality of small-scale workpieces based on laser line scanning, characterized in that, Including: Performing three-dimensional reconstruction on the workpiece to be detected by using the three-dimensional reconstruction method for small-scale workpieces described in any one of claims 1_7 to obtain a three-dimensional reconstruction model; Evaluating the machining error of the workpiece to be detected based on the difference between the three-dimensional reconstruction model and the workpiece machining model of the workpiece to be detected.

9. A workpiece machining method, characterized in that, Including: Monitoring the workpiece to be detected by using the machining quality monitoring method described in claim 8 to obtain a machining error; Adjusting the compensation parameters based on the machining error.

10. A three-dimensional reconstruction system for small-scale workpieces based on laser line scanning, characterized in that, Including: A scanning planning module, configured to analyze key surfaces according to the workpiece machining model and plan scanning paths and scanning angles based on the key surfaces; A multi-view scanning module, configured to drive a laser line scanning camera to perform multi-view scanning on the workpiece based on the scanning paths and scanning angles to obtain multi-view point cloud data; A preprocessing module, configured to preprocess the multi-view point cloud data to obtain preprocessed point cloud data; A registration and stitching module, configured to perform point cloud registration and stitching on different view data in the preprocessed point cloud data by using the iterative closest point matching method to obtain registered and stitched data; A three-dimensional reconstruction module, configured to use a three-dimensional reconstruction algorithm based on the U-Net architecture, take the registered and stitched data as input, retain the multi-level features of the point cloud through skip connections, based on the kernel features output by the decoder, convert the point cloud data into a continuous implicit signed distance representation through a multi-layer perceptron and interpolation technology, and finally use a surface extraction algorithm to extract an isosurface from the implicit field to generate a reconstructed high-precision triangular mesh surface as the three-dimensional reconstruction model of the workpiece.

Images