Data processing method for three-dimensional model of multi-degree-of-freedom inkjet forming of curved surface

Abstract

The present invention belongs to the field of multi-degree-of-freedom additive manufacturing of curved surfaces, and specifically relates to a three-dimensional model data processing method for spatial motion planning in multi-degree-of-freedom inkjet forming of curved surfaces. The method comprises extracting a target spatial surface point cloud from an STL model, fitting the target spatial surface, generating a two-dimensional path for the target spatial surface, and generating a spatial path for inkjet forming of the target spatial surface. The method rapidly extracts the target surface spatial point cloud by calculating the convex or concave hull of the spatial point cloud. This surface is then fitted based on deep learning to obtain a spatial function. The extracted target surface spatial point cloud is then projected onto a two-dimensional plane to generate a path, which is then mapped onto a three-dimensional spatial surface to obtain the curved surface inkjet forming path. This method can solve the current problems in the field of curved surface additive manufacturing, such as the difficulty in path planning, the increase in the number of model vertices and solution points of single-point solution speed when accessing STL model data, the single surface forming path planning method and the difficulty in realizing complex surface forming path planning methods such as surface fractal and surface contour offset. It provides an efficient and convenient data access method for complex path planning methods of curved surface multi-degree-of-freedom inkjet forming, and the single-point solution speed does not increase with the increase in the number of model vertices and solution points, which is of great significance for realizing efficient curved surface multi-degree-of-freedom inkjet forming.

Description

Technical Field

[0001] The present invention belongs to the field of multi-degree-of-freedom additive manufacturing of curved surfaces, and relates to a path generation method combining three-dimensional model data processing with deep learning. Specifically, it is a three-dimensional model data processing method for spatial motion planning in multi-degree-of-freedom inkjet forming of curved surfaces. Background Art

[0002] Curved multi-degree-of-freedom inkjet forming technology is an additive manufacturing technology. It uses a multi-axis motion mechanism, such as a multi-joint serial robot or parallel robot, as the motion mechanism, and mounts an inkjet forming device at the end of the multi-axis motion mechanism for follow-up stacking and forming. The forming process involves key technologies such as 3D model data processing, multi-axis motion mechanism control, follow-up control of the inkjet forming device, and precision closed-loop control. 3D model data processing is used to read 3D STL models and plan forming motions. It has prominent applications in the field of curved electronic printing, and is also expanding its application in areas such as curved wire and tape laying, curved welding, and curved surface spraying.

[0003] During the multi-degree-of-freedom inkjet forming process on curved surfaces, the three-dimensional model data processing process also faces many difficulties, making it difficult to meet complex path planning requirements.

[0004] (1) There is no topological relationship between the model point data stored in the general three-dimensional model file STL for additive manufacturing, and the forming process only focuses on the outer surface of a fixed area without the need for other model data points.

[0005] (2) In order to realize complex motion path planning during the forming process, data points need to be repeatedly located. The disordered storage of data points requires all data points to be sorted when locating a single point, so as to locate the triangle where the target point is located, and then interpolate and calculate the spatial distance z of the target point. Therefore, the efficiency of single-point solution in the forming process decreases with the increase of the number of model vertices and the number of solution points.

[0006] (3) Existing methods all use methods based on data topology reconstruction and data correlation, which are difficult to effectively solve the above problems and difficult to implement complex path planning methods. The current forming path design method is relatively simple and is still based on the traditional slicing and layering sweeping method, which cannot meet the diverse forming path design requirements of various performances.

[0007] In recent years, with the continuous development of computer technology and the gradual improvement of artificial intelligence frameworks, it has become possible to use neural networks to fit any spatial function. This can effectively avoid the repeated sorting and calculation of point data when locating spatial points, and is robust to the increase in the number of model vertices and solution points. Complex path planning methods can be implemented by repeatedly accessing the model to meet various performance requirements. Today's surface forming motion planning methods are still derivatives of data-based models and data processing methods, and have not proposed effective solutions to the various difficulties mentioned above. Summary of the Invention

[0008] To solve the above problems, the present invention discloses a method for processing three-dimensional model data of multi-degree-of-freedom inkjet forming of curved surfaces, so as to solve the limitations of existing motion planning of multi-degree-of-freedom inkjet forming of curved surfaces, that is, it is difficult to implement complex path planning methods and the efficiency of single-point solution decreases with the increase of the number of model vertices and the number of solution points, and provide an effective solution for efficient multi-degree-of-freedom inkjet forming of curved surfaces.

[0009] To achieve the above objectives, the technical solutions adopted in this plan are as follows:

[0010] A method for processing data of a curved surface multi-degree-of-freedom inkjet-formed three-dimensional model comprises the following steps:

[0011] Read the STL model file used in additive manufacturing, obtain the target spatial surface point cloud data in the model through convex hull or concave hull calculation methods, and use the spatial point cloud data as the basic data set for spatial surface fitting;

[0012] Construct a point cloud data processing network model suitable for spatial surface fitting;

[0013] In the established point cloud data processing network model, the plane coordinates and spatial coordinates of the model points in the basic data set are trained to obtain the training completion weights;

[0014] Project the target space surface onto a two-dimensional plane and generate a two-dimensional path based on the two-dimensional point cloud;

[0015] The obtained two-dimensional path is mapped to the point cloud data processing network model and the trained weights are used to output the surface multi-degree-of-freedom inkjet forming spatial path.

[0016] As a further design of this solution, the method for reading STL files and obtaining the target surface is as follows: after reading the spatial coordinates of the unordered points on the surface of the STL model in ASCII code format, the convex hull or concave hull calculation in the three-dimensional point cloud processing method is used to quickly extract the spatial point cloud data of the target surface in the model and use (x, y) in the point cloud data as input and z as the label as the training data set.

[0017] As a further design of this scheme, the method of constructing a point cloud data processing network model for spatial surface fitting is as follows: a point cloud segmentation module is constructed to segment the target surface point cloud into several block point clouds that are easy to fit; the block point cloud rotation module is used to rotate the block point cloud obtained in the point cloud segmentation module to obtain a rigid rotation matrix with the maximum projection area; and the deep neural network fitting module is used to fit the rigidly rotated block point cloud obtained in the block point cloud rotation module.

[0018] As a further design of this solution, the point cloud segmentation module is used to segment the obtained two-dimensional coordinates into multiple block point clouds according to the projection grid. The construction method is as follows: the target surface point cloud is projected onto a two-dimensional plane, and the minimum square enclosure algorithm is used to calculate the minimum square enclosure of the point cloud, and then the point cloud is divided into multiple block point clouds of equal size by isometric interpolation in the square enclosure.

[0019] As a further design of this solution, the block point cloud rotation module is used to rotate the block point cloud obtained in the point cloud segmentation module to obtain the rigid rotation matrix of the maximum two-dimensional projection, the rotated point cloud data and the rigid inverse rotation matrix. The construction method is as follows: read in several block point clouds obtained in S2, take the two points with the largest Euclidean distance in each point cloud as the rotation axis and the X vector of the rigid rotation matrix, take the midpoint O of the two points as the coordinate origin, and generate 720 Y vectors passing through point O and perpendicular to the X axis in the plane passing through point 0 and perpendicular to the X axis, with the plane intersection line angle difference of 0.5°. The outer product of the X vector and each Y vector is calculated to obtain the Z vector of the rotation matrix, calculate the projected area under the point cloud rotation and iterate the rigid rotation matrix under the maximum projected area, and output the rotated point cloud data and the rigid inverse rotation matrix.

[0020] As a further design of this solution, a deep neural network fitting module is used to fit the rotated point cloud output by the block point cloud rotation module and add a rigid inverse rotation module. The construction method is as follows: read the rotated point cloud data output by the block point cloud rotation module, capture the spatial topological relationship of the point cloud through the deep neural network, and form a mapping relationship between the two-dimensional plane points and the three-dimensional spatial positions in the surface. The spatial position can be directly output by inputting the plane path points.

[0021] As a further design of this scheme, the training of the point cloud data processing network model includes the following steps: reading the target surface point cloud data and processing it through the point cloud segmentation module and the block point cloud rotation module to obtain several rotated block point clouds and their rigid inverse rotation matrices; normalizing the input point cloud data, calculating the spatial distance of the current input (x, y) calculated through the network and multiplying it by the rigid inverse rotation matrix to obtain the calculated spatial distance, comparing the calculated spatial distance with the actual spatial distance in the label, so that the calculated spatial distance is equal to the actual spatial distance; repeating the second step to optimize and adjust the parameters of each network node of the model, and finally obtaining an optimal solution. The optimal solution can make the calculated spatial distance equal to the actual spatial distance. The optimal solution is the weight of the point cloud data processing network model.

[0022] As a further design of this scheme, the method for generating a two-dimensional path for curved surface forming is as follows: read the target space point cloud and project it into a two-dimensional plane, and calculate the minimum enclosing area of ​​the two-dimensional point cloud through a convex hull or concave hull algorithm, and then generate a two-dimensional forming path through calculation methods such as fixed-angle straight line scanning, contour offset, parting curve and grid paving.

[0023] As a further design of this solution, the method for outputting the curved surface multi-degree-of-freedom inkjet forming spatial path is as follows: read the obtained two-dimensional forming path and the trained point cloud data processing model, and import the two-dimensional forming path points into the point cloud data processing model in batches to generate the curved surface inkjet forming spatial path.

[0024] Beneficial effects of the present invention:

[0025] (1) In the method for processing data of a curved surface multi-degree-of-freedom inkjet forming three-dimensional model designed by the present invention, the target surface point cloud is segmented and rotated, then introduced into a network fitting model based on deep learning and then inversely rotated and outputted, thereby achieving direct mapping of two-dimensional data to three-dimensional data and effectively avoiding repeated sorting and calculation of point position data when locating spatial points;

[0026] (2) The method for processing data of a curved surface multi-degree-of-freedom inkjet forming three-dimensional model designed by the present invention can realize complex path planning methods for curved surface forming motion that are difficult to achieve with traditional computing methods, such as curved surface contour offset, curved surface fractal, and curved surface meshing;

[0027] (3) In the data processing method for curved surface multi-degree-of-freedom inkjet forming three-dimensional models designed by the present invention, the efficiency of spatial point positioning calculation is robust to the increase in the number of model vertices and the number of solution points, is suitable for large-scale point cloud computing, and can provide technical support for efficient curved surface multi-degree-of-freedom inkjet forming. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1This is a schematic diagram of the target surface point cloud extraction of the STL model described in an embodiment of the present invention, wherein the left picture is the initial point cloud three-dimensional image, the middle picture is the surface point cloud annotation calculation three-dimensional image, and the right picture is the surface point cloud three-dimensional image after removing the internal point cloud.

[0029] Figure 2 This is a diagram of the point cloud data processing network model architecture for spatial surface fitting described in an embodiment of the present invention.

[0030] Figure 3 Schematic diagram of the point cloud segmentation module described in an embodiment of the present invention.

[0031] Figure 4 This is a schematic diagram of the block point cloud rotation module described in an embodiment of the present invention, wherein the left figure is the projection of the maximum projection area and the minimum envelope of the block point cloud rotation, and the right figure is the curve of the change of the projection area of ​​the block point cloud rotation.

[0032] Figure 5 This is a schematic diagram of the deep neural network fitting module described in an embodiment of the present invention, where the left figure is the three-dimensional image output by helmet model fitting, and the right figure is the error change curve.

[0033] Figure 6 This is a schematic diagram of three-dimensional path generation described in an embodiment of the present invention.

[0034] Figure 7 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0035] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only intended to illustrate the present invention and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0036] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0037] like Figure 7 As shown, the embodiment of the present invention discloses a method for processing data of a curved surface multi-degree-of-freedom inkjet forming three-dimensional model, comprising the following steps:

[0038] (1) Read the STL model file used in additive manufacturing, obtain the target spatial surface point cloud data in the model through the convex hull calculation method, and use the spatial point cloud data as the basic data set for spatial surface fitting;

[0039] (2) Constructing a point cloud data processing network model suitable for spatial surface fitting;

[0040] (3) In the established point cloud data processing network model, the plane coordinates and spatial coordinates of the model points in the basic data set are trained to obtain the training completion weights;

[0041] (4) Projecting the target space surface onto a two-dimensional plane and generating a two-dimensional path based on the two-dimensional point cloud;

[0042] (5) The obtained two-dimensional path is mapped to the network fitting model and the trained weights are used to output the surface multi-degree-of-freedom inkjet forming spatial path.

[0043] As a preferred embodiment of the present invention, refer to Figure 1 ,In step (1), the data set acquisition process is as follows:

[0044] (1-1) Read the ASCII code format STL model through the mesh interface;

[0045] (1-2) Calculate the surface point cloud using the convex hull algorithm in the PIL interface;

[0046] (1-3) Take the point cloud data (x, y) as input and z as label as the training data set.

[0047] As a preferred embodiment of the present invention, refer to Figure 2 In step (2), the point cloud data processing network model for spatial surface fitting includes: a point cloud segmentation module for segmenting the target surface point cloud into several block point clouds that are easy to fit; a block point cloud rotation module for rotating the block point cloud obtained in the point cloud segmentation module to obtain a rigid rotation matrix with the maximum projection area; and a deep neural network fitting module for fitting the rigidly rotated block point cloud obtained in the block point cloud rotation module.

[0048] As a preferred embodiment of the present invention, refer to Figure 3 The process of the point cloud segmentation module includes: projecting the target surface point cloud onto a two-dimensional plane, calculating the minimum square enclosure of the point cloud using the minimum square enclosure algorithm, and then dividing the point cloud into block point clouds of 50×50 size using equidistant interpolation with a spacing of 50 in the square enclosure.

[0049] As a preferred embodiment of the present invention, refer to Figure 4The process of the block point cloud rotation module includes: reading several block point clouds obtained in S2, taking the two points with the largest Euclidean distance in each point cloud as the rotation axis and the X vector of the rigid rotation matrix, taking the midpoint O of the two points as the coordinate origin, generating 720 Y vectors passing through point O and perpendicular to the X axis in the plane passing through point O and with an angle difference of 0.5° between the plane intersection lines, calculating the outer product of the X vector and each Y vector to obtain the Z vector of the rotation matrix, calculating the projected area under the point cloud rotation and iterating the rigid rotation matrix under the maximum projected area, and outputting the rotated point cloud data and the rigid inverse rotation matrix. Figure 4 The projection area is the largest when the point cloud rotation angle is 157°.

[0050] As a preferred embodiment of the present invention, refer to Figure 5 The process of the deep neural network fitting module includes: reading the rotated point cloud data output by the block point cloud rotation module, capturing the spatial topological relationship of the point cloud through a multi-layer perceptron, in which the batch size is 256, the epoch is 10, the optimizer is 'adam', the oss_funct i on is 'mse', and the activation function is 'sigmo i d'. The mapping relationship between the two-dimensional plane points on the surface and the three-dimensional spatial positions is formed, and the spatial position can be directly output by inputting the plane path points.

[0051] As a preferred embodiment of the present invention, refer to Figure 2 ,In step (3), the point cloud data processing network model training process is as follows:

[0052] (3-1) Reading the target surface point cloud data and processing it through the point cloud segmentation module and the block point cloud rotation module to obtain several rotated block point clouds and their rigid inverse rotation matrices;

[0053] (3-2) Normalize the input point cloud data, calculate the spatial distance of the current input (x, y) calculated by the network and multiply it by the rigid inverse rotation matrix to obtain the calculated spatial distance, compare the calculated spatial distance with the actual spatial distance in the label, and make the calculated spatial distance equal to the actual spatial distance;

[0054] (3-3) Repeat (3-2) to optimize and adjust the parameters of each network node of the model, and finally obtain an optimal solution. The optimal solution can make the calculated spatial distance equal to the actual spatial distance. The optimal solution is the weight of the point cloud data processing network model and is stored as an h5 file together with the network model.

[0055] As a preferred embodiment of the present invention, refer to Figure 6 ,In step (4), the point cloud data processing network model training process is as follows:

[0056] (4-1) Read the target space point cloud and project it into a two-dimensional plane and calculate the minimum enclosing area of ​​the two-dimensional projection point cloud of the radome, aircraft nose and helmet model using the convex hull algorithm;

[0057] (4-2) Then, a two-dimensional forming path is generated by using fixed-angle linear scanning for the radome and the front nose of the aircraft and a contour offset calculation method for the helmet.

[0058] As a preferred embodiment of the present invention, refer to Figure 6 ,In step (5), the point cloud data processing network model training process is as follows:

[0059] (5-1) Read the two-dimensional forming path of the radome, aircraft nose and helmet models, the trained point cloud data processing model and the corresponding weights;

[0060] (5-2) Importing two-dimensional forming path points in batches into the point cloud data processing model to generate the surface inkjet forming spatial path.

[0061] Finally, 10,000 path points were solved for the radome, aircraft nose, and helmet models. The average single-point solution speeds for 20 experiments were 0.0244ms / point, 0.0243ms / point, and 0.0246ms / point, respectively. The single-point solution speed did not increase with the increase in the number of vertices, but decreased with the increase in the number of solution points.

[0062] The technical means disclosed in the solution of the present invention are not limited to the technical means disclosed in the above-mentioned embodiment, but also include technical solutions composed of any combination of the above technical features.

Claims

1. A method for processing data of a curved surface multi-degree-of-freedom inkjet-formed three-dimensional model, characterized in that: The steps include: S1: Read the STL model file used in additive manufacturing, obtain the target spatial surface point cloud data in the model by convex hull or concave hull calculation method, and use the spatial surface point cloud data as the basic data set for spatial surface fitting; S2: Construct a point cloud data processing network model suitable for spatial surface fitting; S3: In the point cloud data processing network model established in S2, the plane coordinates and spatial coordinates of the model points in the basic data set obtained in S1 are trained to obtain the trained weights; S4: Project the target spatial surface obtained in S1 onto a two-dimensional plane and generate a two-dimensional path based on the two-dimensional point cloud; S5: Mapping the two-dimensional path obtained in S4 to the point cloud data processing network model constructed in S2 and using the weights obtained in S3 to output the surface multi-degree-of-freedom inkjet forming spatial path; In S2, the point cloud network processing model for spatial surface fitting includes: point cloud segmentation module, block point cloud rotation module and deep neural network fitting module; Among them, the point cloud segmentation module is used to segment the target surface point cloud into several block point clouds that are easy to fit; the block point cloud rotation module is used to rotate the block point cloud obtained in the point cloud segmentation module to obtain a rigid rotation matrix with the maximum projected area; the deep neural network fitting module is used to fit the rigidly rotated block point cloud obtained in the block point cloud rotation module; the point cloud segmentation module divides the obtained two-dimensional coordinates into multiple block point clouds according to the projection grid. The process of generating the block point cloud includes: projecting the point cloud obtained in S1 onto a two-dimensional plane, and using the minimum square enclosure algorithm to calculate the minimum square enclosure of the point cloud, and then isometrically interpolating the point cloud into multiple block point clouds of equal size in the square enclosure; The block point cloud rotation module rotates the block point cloud obtained in the point cloud segmentation module to obtain the rigid rotation matrix of the maximum two-dimensional projection, the rotated point cloud data, and the rigid inverse rotation matrix. The process includes: reading in several block point clouds obtained in S2, taking the two points with the largest Euclidean distance in each point cloud as the rotation axis and the X vector of the rigid rotation matrix, taking the midpoint O of the two points as the coordinate origin, generating 720 Y vectors passing through O and perpendicular to the X axis in a plane passing through O and with an angle difference of 0.5° between the plane intersection lines, calculating the outer product of the X vector and each Y vector to obtain the Z vector of the rotation matrix, calculating the projected area under the point cloud rotation and iterating the rigid rotation matrix under the maximum projected area, and outputting the rotated point cloud data and the rigid inverse rotation matrix; The deep neural network fitting module is used to fit the rotated point cloud output by the block point cloud rotation module and add a rigid inverse rotation module. The process includes: reading the rotated point cloud data output by the block point cloud rotation module, capturing the spatial topological relationship of the point cloud through the deep neural network, forming a mapping relationship between the two-dimensional plane points in the surface and the three-dimensional spatial position, and directly outputting the spatial position by inputting the plane path point.

2. The method for processing data of a curved surface multi-degree-of-freedom inkjet forming three-dimensional model according to claim 1, characterized in that: In S1, the spatial coordinates of the unordered points on the surface of the STL model in ASCII code format are read, and then the spatial point cloud data of the target surface in the model is quickly extracted through the convex hull or concave hull calculation in the three-dimensional point cloud processing method, and the (x, y) in the point cloud data is used as input and z as the label as the training data set.

3. The method for processing data of a curved surface multi-degree-of-freedom inkjet-formed three-dimensional model according to claim 1, characterized in that: In S3, the process of training the point cloud data processing network model includes the following steps: S3-1, the target surface point cloud data is processed by the point cloud segmentation module and the block point cloud rotation module in S2 to obtain several rotated block point clouds and their rigid inverse rotation matrices; S3-2, normalize the input point cloud data, calculate the spatial distance of the current input (x, y) calculated by the network and multiply it by the rigid inverse rotation matrix to obtain the calculated spatial distance, compare the calculated spatial distance with the actual spatial distance in the label, and make the calculated spatial distance equal to the actual spatial distance; S3-3, repeat S3-2 to optimize and adjust the parameters of each network node of the model, and finally obtain an optimal solution. The optimal solution can make the calculated space distance equal to the actual space distance. This optimal solution is the weight of the point cloud data processing network model.

4. The method for processing data of a curved surface multi-degree-of-freedom inkjet-formed three-dimensional model according to claim 1, characterized in that: In S4, the target space point cloud obtained in S1 is read and projected into a two-dimensional plane, and the minimum enclosing area of ​​the two-dimensional point cloud is calculated by the convex hull or concave hull algorithm. Then, a two-dimensional forming path is generated by calculation methods such as fixed-angle linear scanning, contour offset, parting curve and grid paving.

5. The method for processing data of a curved surface multi-degree-of-freedom inkjet-formed three-dimensional model according to claim 1, characterized in that: In S5, the two-dimensional forming path obtained in S4 and the point cloud data processing model trained in S3 are read, and the two-dimensional forming path points are batch-imported into the point cloud data processing model to generate a curved inkjet forming spatial path.

Images