A manufacturing process strategy prediction method based on b-rep self-supervised learning

Abstract

The application relates to the technical field of manufacturing process strategy prediction, in particular to a manufacturing process strategy prediction method based on B-rep self-supervised learning, which comprises the following steps: preprocessing input B-rep data, extracting the geometry and topological features of each face, ring, edge and vertex, and encoding the features into a feature matrix; inputting the feature matrix into an encoder-decoder model, training the model through a self-supervised learning framework, and learning the embedding representation of each face in a latent space from the B-rep data; and utilizing the embedding representation of each face in the latent space obtained by training the model through the self-supervised learning framework. Through the self-supervised neural network framework, the application effectively learns the geometry and topological information of the model in unmarked CAD data, can effectively learn on a small amount of samples and unlabeled data, and thus provides accurate and rapid process strategy prediction for part machining.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of manufacturing process strategy prediction, in particular to a manufacturing process strategy prediction method based on B-rep self-supervised learning. BACKGROUND

[0002] In the field of manufacturing process strategy prediction, traditional manufacturing process decision-making mainly relies on experienced engineers and their professional knowledge. However, this method is not only time-consuming and laborious, but also prone to errors, especially when it comes to complex or high-precision manufacturing tasks. In addition, as manufacturing gradually shifts from mass production to flexible manufacturing mode with small batches and multiple varieties, the complexity of process planning increases significantly. The traditional decision-making process is inefficient in this case and difficult to quickly adapt to changing production demands.

[0003] To address this issue, some machine learning-based manufacturing process prediction methods have been proposed in recent years. These methods typically use the geometric data of 3D CAD models by converting them into point clouds, voxels, or meshes for analysis.

[0004] However, the above-mentioned technologies have problems such as information loss, difficult data labeling, and poor model generality when used in practice, so they still face challenges when dealing with complex manufacturing process planning tasks. SUMMARY

[0005] The purpose of the present application is to provide a manufacturing process strategy prediction method based on B-rep self-supervised learning to solve the problems of information loss, difficult data labeling, and poor model generality, so that it still faces challenges when dealing with complex manufacturing process planning tasks.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solution: a manufacturing process strategy prediction method based on B-rep self-supervised learning, comprising the following steps:

[0007] S1, CAD data preprocessing: pre-process the input B-rep data, extract the geometric and topological features of each face, loop, edge, and vertex, and encode them into a feature matrix;

[0008] S2, self-supervised pre-training: input the feature matrix pre-processed in step S1 into an encoder-decoder model, train the model through a self-supervised learning framework, and learn the embedding representation of each face in the latent space from the B-rep data;

[0009] S3, process strategy prediction: use the embedding representation of each face in the latent space obtained by training the model in step S2 through a process strategy prediction network, calculate and output the process strategy label of each face;

[0010] S4, process planning result output: output the model segmentation result in the form of graphics or text.

[0011] Preferably, the B-rep data includes different levels of topological entities of faces, loops, edges and vertices.

[0012] Preferably, in step S1, different coding processes are performed on different types of topological entities of faces, loops, edges and vertices to form corresponding feature matrices.

[0013] Preferably, in the encoder-decoder model of step S2, the encoder realizes information aggregation through a multi-layer residual graph convolution layer, and the decoder realizes self-supervised learning of the latent representation through a signed distance function reconstruction.

[0014] Preferably, in step S3, the embedding representation of each face in the latent space is used to predict the process strategy on a small sample dataset, thereby realizing segmentation and recommendation of different manufacturing process strategies.

[0015] Preferably, the process strategy label in step S3 includes bottom wall milling, cavity milling, depth contour milling, adaptive milling and drilling milling processing strategies.

[0016] Preferably, in step S3, the face embedding representation and the connection matrix between faces are used as network inputs in the process strategy prediction network, and the process strategy prediction network containing a residual graph convolution layer and a multi-layer perception module is used to predict the corresponding process strategy label of each face.

[0017] Compared with the prior art, the present application has the following advantages:

[0018] 1. The present application directly uses boundary representation as input, completely retains the geometry and topology information of the CAD model, greatly improves the accuracy and reliability of the process strategy prediction, learns the geometric features from large-scale unlabeled data, and can realize high-precision strategy prediction under the condition of a small amount of samples, effectively reduces the demand and cost of data labeling, and the latent representation obtained through self-supervised learning has stronger generalization ability, can quickly adapt to new machining tasks in small-batch and multi-variety production scenes, meets the needs of modern manufacturing industry, significantly improves the process planning efficiency compared with the traditional method relying on artificial experience, reduces the artificial intervention and decision-making time, and reduces the risk of human error, thereby providing accurate and fast process strategy prediction for part machining. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The present application is a manufacturing process strategy prediction method based on B-rep self-supervised learning. DETAILED DESCRIPTION

[0020] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0021] Embodiment one

[0022] Please refer to Figure 1 The present application provides a technical solution: a manufacturing process strategy prediction method based on B-rep self-supervised learning, comprising the following steps:

[0023] S1, CAD data preprocessing: preprocessing the input B-rep data, extracting the geometry and topological features of each face, loop, edge and vertex, and encoding them into a feature matrix;

[0024] Wherein the B-rep data includes various different levels of topological entities such as face, loop, edge and vertex, and there is correlation between them. High-level topological entities can retrieve low-level topological entities, such as: face topology is a closed area surrounded by loop topology, loop topology is composed of an unequal number of edge topologies, and edge topology has vertex topology to define the starting point and end point.

[0025] Face topology entity is usually composed of plane, cylindrical surface, conical surface, spherical surface, torus or b-spline surface, etc., which is used to construct the surface shape of the actual CAD model;

[0026] Loop is composed of multiple edges connected end to end, defining the boundary of the face;

[0027] Edge can be composed of straight line, circular arc, circle or b-spline curve;

[0028] The relationship between these entities forms a hierarchical graph, from which face topology entity can retrieve loop topology entity, loop topology entity can retrieve edge topology entity, and edge topology entity can retrieve vertex topology entity. There is message passing between different topological entities.

[0029] Wherein the unused parameters are encoded as zero. By extracting the geometric feature information of the face, loop, edge and vertex of the B-rep data, and passing it to the network in the form of feature matrix X f , X l , X e and X v , the geometric information is extracted by the CAD kernel, and the MLP is used to compress it into a feature vector. In addition, in order to reduce the conversion, the coordinates involved are extracted using the same reference coordinate system.

[0030] The input encoding for 4 different topological types is explained in detail as follows:

[0031] Face encode: For the encoding of the CAD model face, the first five bits of the feature matrix are encoded in One-hot mode for the type of surface (such as plane, cylindrical surface, conical surface, spherical surface and torus), and b-spline surface is not within the scope of the present application. The normal direction and area parameter of each face are extracted, and the radius parameter is extracted for cylindrical and conical surfaces, and the radius parameter is 0 for the plane. At the same time, the coordinates of the centroid of the model face and the envelope of the face are also encoded to facilitate subsequent matching.

[0032] Loop encode: For the encoding of the CAD model ring, the first 10 bits of the feature matrix are encoded in One-hot mode for the type of ring (inner ring, outer ring, etc.), and the length of the ring is set to the number of edges contained by the ring. The coordinates of the face centroid and the envelope formed by the ring are encoded to facilitate the matching of the face and the ring.

[0033] Edge encode: For the encoding of the CAD model edge, the first three bits of the feature matrix provide One-hot encoding of a possible edge geometry type (straight line, circle, ellipse, spiral, intersecting curve), and the concave-convex property and length of the edge are extracted as size parameters, and the coordinate information of the two endpoints of the edge is recorded.

[0034] Vertices encode: For the encoding of the CAD model vertex, its coordinate orientation is related to the relative position of the model face and edge, and is independent of the modeling coordinate system, so by establishing a relative coordinate system at the center of the model envelope, the position information of the vertex is re-parameterized.

[0035] For the CAD model, the geometric information is extracted, and the face matrix, loop feature matrix X l ={l1, l2,..., l |l|}, edge matrix X e ={e1, e2,..., e |e|}, and vertex matrix X v ={p1, p2,..., p |v|} are constructed respectively. and matrix are constructed as attribute connection graphs, representing the corresponding relationship between vertices and edges, edges and loops, and loops and faces respectively. The input features are standardized on the training set, and the same scaling is applied to the validation set and the test set. At this point, by representing the CAD model as a lightweight matrix, the geometric and topological information of the CAD model is processed into a structure suitable for neural network input.

[0036] S2, self-supervised pre-training: input the feature matrix after step S1 preprocessing into the encoder-decoder model, train the model through the self-supervised learning framework, and learn the embedding representation of each face in the latent space from the B-rep data;

[0037] The encoder-decoder model specifically includes:

[0038] Encoder: in the encoder structure, the Structured BREP Graph Convolutional Network is referred to for processing the heterogeneous input data of the CAD model. The Structured BREP Graph Convolutional Network is a hierarchical graph convolutional neural network, and the BREP graph nodes are divided into four layers according to their topological entity types. On this basis, the present application focuses on the face topology of B-Rep, therefore, only the upward transmission part of the SBGCN network is retained as the network structure of the encoder. Three graph convolutional layers (MRGCN) are used in the encoder to complete the aggregation of the different hierarchical topological information of B-Rep. The process of message passing is specifically as follows:

[0039] Formally, the extracted input feature matrix is defined as a set of node sets, and the value of the node set is N=X f ∪X l ∪X e ∪X v , X f , X l , X e , X v are the feature matrices of the face, ring, edge and vertex respectively. In addition, the connection matrices of the face topology and ring topology connection matrix of ring topology and edge topology connection matrix of edge topology and vertex topology are obtained to obtain the connection relationship between different topological entities.

[0040] Firstly, a shared MLP linear layer is used to convert all input feature vectors into the same 64-dimensional feature vector:

[0041]

[0042] N (l) =H f ∪H l ∪H e ∪H vThe new node set has a feature vector dimension of 64. Then three graph convolution layers (MRGCN) are used to realize message passing, which are denoted as MRGCNLayer1, MRGCNLayer2, and MRGCNLayer3. MRGCNLayer1 first aggregates the information of vertices to the edges to which they are bound, MRGCNLayer2 aggregates the bounded edge information to the topological ring, and finally MRGCNLayer3 aggregates the ring information to the topological surface. The information set obtained after message passing by a single graph convolution layer MRGCN is

[0043]

[0044] wherein, is the feature vector matrix in N (l) after message passing;

[0045] τ(v) represents a certain feature vector matrix in N (l) ;

[0046] u∈Ν τ(u):τ(v) (v) represents that the two topological features are connected, that is, the connection matrix exists;

[0047] MLP (l) is a standard fully connected structure composed of linear layers, batch normalization, and ReLU. The main reason for using this structure is that the relationship of convolution operation depends on the layer, and only convolution operation is performed between adjacent layers (different layers are represented by the subscripts of the neighborhood operator). The order of the entire message passing is according to the hierarchy of the BREP topological entities: vertex to edge, edge to ring, and ring to surface.

[0048] Through the message passing of the encoder, the aggregation of the BREP different topological structure information is realized, and the embedded representation of the surface is obtained as part of the input of the decoder of the application.

[0049] Decoder: Generally, the representation of a geometric model can be divided into display representation and implicit representation:

[0050] Display representation (Explicit Representation) means that all points on the geometric model can be directly represented or represented by parameter mapping. Common representations such as point cloud, mesh, and voxel, etc. Implicit representation (Implicit Representation) of a geometric model means that the distance relationship between points is usually described, and the shape is defined by an equation or inequality, rather than directly giving the coordinates of the points.

[0051] Implicit representation usually uses an equation F(x, y, z) = 0 to define a surface or volume, where the points satisfying the equation constitute the geometric shape. Common implicit representations are constructive solid geometry (CSG), distance function representation.

[0052] Signed distance function (SDF), abbreviated as SDF, is a continuous function that outputs the shortest distance from a given spatial point to a given surface, and its positive and negative represent whether the point is inside (negative) or outside (positive) the closed surface. The SDF reconstruction representation of the model surface is completed using the decoder to serve as the self-supervised learning task of the present application.

[0053] In the SDF representation, the shape S can be represented as:

[0054] S = {f θ (x, z) = 0 | x e R 3 , z e R m} (3)

[0055] Where f θ (x, z): R 3 x R m → R is a signed surface distance function, which can be implemented using a neural network. This neural network model determines the output shape through the latent vector z and the sampling coordinates. Through the following formula, f θ Simultaneously capture and process the overall and local shape features of the CAD model.

[0056]

[0057] Where, represents the conversion of x from the global coordinate system to the coordinate system corresponding to x after x is converted to the local coordinate system of R i ;

[0058] represents the latent representation vector of the local shape;

[0059] d(x, p) is a distance function, shape S represents the surface of the model, and latent vector represents the output of the decoder of the present application, and the local coordinates are obtained by UV sampling on the surface. At the same time, the UV sampling is normalized, so that the UV coordinate values are all between [-0.1, 1.1], which is convenient for coding.

[0060] Overall, the decoder is building a function that maps the (u, v) coordinates of a CAD model face to the three-dimensional coordinate position of that point (x, y, z). At the same time, in the parametric domain, the same (u, v) coordinates are mapped to the value d of the signed distance function of the point to the boundary. By combining these two functions, the encoder of the present application realizes the reconstruction representation of (u, v) coordinates in the SDF representation of (x, y, z, d) four-dimensional space. Through experiments, the decoder is set to three fully connected residual layers.

[0061] S3, process strategy prediction: using the embedding representation of each face in the latent space obtained by training the model in the step S2 self-supervised learning framework, through the process strategy prediction network, calculating and outputting the process strategy label of each face, specifically:

[0062] After pre-training is completed, the process strategy prediction module takes the face embedding representation as input, and completes the process strategy segmentation task of the part through the process strategy prediction network. The specific steps are as follows:

[0063] Input feature representation: taking the embedding representation of each topological face in the CAD model and the connection matrix between the faces as the network input.

[0064] Predicting process strategy: the network structure contains a residual graph convolutional layer and a multi-layer perception (MLP) module. Under the action of the residual graph convolutional layer, the information on each face is further aggregated to the adjacent face, and then the probability value of process strategy prediction is calculated through the MLP.

[0065] Strategy recommendation: the model finally outputs the process strategy label corresponding to each face. For example, rough machining strategies such as bottom wall milling (FM), cavity milling (CM), depth profile milling (ZM), adaptive milling (AM), drilling (HD), etc. Different colored labels are used to indicate the segmentation result of the model, so as to intuitively display the area suitable for each process strategy in the CAD model.

[0066] S4, process planning result output: output the model segmentation result in the form of graphics or text.

[0067] The present application innovatively embeds process information into the topological faces of the three-dimensional model, reducing the granularity of process planning.

[0068] Milling processing strategy: Milling processing strategy refers to the method of selecting processing tool path generation according to different processing shape regions. In actual process planning, the user selects a group of faces or edges as the input of the processing strategy according to the characteristics of the milling strategy, and then the CAM software calls the tool path generation module to generate the tool path according to the selected milling processing strategy and face group information. In the processing process, selecting the correct processing strategy is a key planning task, and the appropriate milling strategy is crucial to ensure the efficiency and accuracy of part processing. Common milling strategies and their characteristics are shown in Table 1.

[0069] Table 1: Description of milling strategy and its characteristics

[0070]

[0071]

[0072] Definition 1: Roughing: Roughing is to remove the blank quickly to make the geometric shape of the blank close to the finished product, and to prepare for finishing. The main purpose of roughing is to pursue the material removal rate per unit time and to prepare for subsequent semi-finishing. The characteristics of roughing are to improve the processing efficiency as much as possible, to select higher cutting amount as much as possible without affecting the processing quality, to use reasonable processing technology, and to minimize tool wear.

[0073] Definition 2: Finishing: Finishing stage completes the final processing of each main surface, which is to process the semi-finished product parts completed by roughing with high-precision tools. The most important consideration in finishing is the surface quality of the workpiece. According to the characteristics and purposes of rough and finish machining, the selected processing strategy also differs.

[0074] Through the analysis of the machining data in the data set, the machining strategy is limited to five kinds, and according to the different machining stages, it is divided into rough milling machining data set and fine milling machining data set. Among them, the rough milling machining data set contains five kinds of machining strategies: including floor wall milling (Floor Wall Milling FM), cavity milling (Cavity Milling CM), depth profile milling Z-level Profile Milling ZM, adaptive milling (Adaptive Milling AM) and hole drilling (Hole Drilling HD). The fine milling data set includes three kinds of machining strategies: floor wall milling (Floor Wall Milling FM), depth profile milling (Z-level Profile Milling), and hole drilling (Hole Drilling HD). Table 1 details the advantages and disadvantages of each strategy and the applicable scenarios. Among them, the floor wall (FW) milling strategy is suitable for processing flat bottom straight wall area, while the Z-level profile milling (ZP) is usually used for processing workpiece area with steep side wall. Cavity milling (CM) is used for cavity processing, while adaptive milling (AM) is usually used for rough machining of complex profile area. Drilling (DR) is used for machining holes and threaded holes.

[0075] Different machining strategies are combined with the faces of the model to predict the machining strategy for the faces of the model. By providing a B-rep model segmentation data set for milling, the data set not only contains the geometric information of the B-rep model, but also contains the milling process strategy corresponding to each topological face, providing valuable data resources for subsequent process strategy prediction and model optimization.

[0076] By using the boundary representation (B-rep) as the core data structure of the CAD model as the direct input of the model trained by the self-supervised learning framework, the problem of relying on manual experience, time-consuming and laborious, and prone to errors in traditional manufacturing process decision-making is effectively solved, and the shortcomings of existing machine learning methods in information preservation, data labeling requirements and model generality are overcome.

[0077] The present application directly uses B-rep as input, completely retains the geometry and topology information of the CAD model, provides a more rich and accurate data basis for the prediction of manufacturing process strategy, thereby significantly improving the accuracy and reliability of the prediction, by constructing a self-supervised pre-training model, learning the geometric features of the CAD model using large-scale unlabeled data, effectively reducing the dependence on data labeling, reducing the time and cost of data preparation, the learned latent representation has stronger generalization ability, can adapt to different types and shapes of parts, improves the adaptability and flexibility of the process strategy; in addition, through automatic process strategy prediction, reduces manual intervention and decision-making time, significantly improves the efficiency of process planning, provides strong technical support for the intelligentization and automation of manufacturing industry, provides a new solution and technical path for process planning, production optimization and intelligent upgrading of manufacturing industry, is expected to play an important role in the future development of manufacturing industry, and promote the development of manufacturing industry to be more efficient, more intelligent and more sustainable.

[0078] Embodiment two

[0079] Unlike embodiment one, this embodiment uses point cloud or voxel representation as input of CAD model in step S1. This method converts three-dimensional model into regularly sampled point cloud or voxel to adapt to existing deep learning framework. Through supervised learning on these discrete representation data, process strategy prediction task can be completed. However, due to the partial loss of geometry and topology information caused by point cloud and voxel conversion, and the dependence of accuracy on sampling density, this method is not as good as embodiment one in information fidelity and prediction accuracy.

[0080] Embodiment three

[0081] Unlike embodiment one, this embodiment uses graph neural network (GNN) to learn the graph structure of CAD model in step S2. The CAD model is represented as a graph structure composed of nodes and edges, the nodes represent the faces or vertices of the model, and the edges represent the geometric or topological relationship. By learning node features through graph neural network, the process strategy of different regions can be predicted. However, the implementation of this method is usually more complex, and the capture of model topology information is not as direct and efficient as embodiment one, so it may be limited when dealing with complex geometry and diversified process strategy.

[0082] Embodiment four

[0083] Unlike the first embodiment, the present embodiment uses a conventional supervised learning method to train the manufacturing process strategy in step S2 by a large-scale labeled dataset. This method requires a large amount of CAD model data to be manually labeled with the corresponding machining strategy for each face, and then a deep learning model is trained for strategy prediction. The accuracy of this scheme depends on the size and quality of the dataset, but it is difficult to compare with the self-supervised learning scheme of the first embodiment in terms of data annotation cost, time requirement, and adaptability.

[0084] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0085] While embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, changes, and alterations can be made by those skilled in the art without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for manufacturing process strategy prediction based on B-rep self-supervised learning, characterized in that: The method comprises the following steps: S1, CAD data preprocessing: preprocessing the input B-rep data, extracting the geometric and topological features of each face, loop, edge and vertex, and encoding them into a feature matrix; S2, self-supervised pre-training: input the feature matrix pre-processed in step S1 into an encoder-decoder model, train the model through a self-supervised learning framework, learn the embedding representation of each face in the latent space from the B-rep data, in the encoder-decoder model, the encoder realizes information aggregation through multiple layers of residual graph convolutional layers, and the decoder realizes self-supervised learning of the latent representation through a signed distance function reconstruction; S3, process strategy prediction: using the embedding representation of each face in the latent space obtained by training the model through the self-supervised learning framework in step S2, calculating and outputting the process strategy label of each face through a process strategy prediction network, the embedding representation of each face in the latent space is used for process strategy prediction on a small sample data set, realizing the segmentation and recommendation of different manufacturing process strategies, and the process strategy label includes bottom wall milling, cavity milling, deep profile milling, adaptive milling and drilling milling processing strategies; S4, process planning result output: outputting the model segmentation result in the form of graphics or text.

2. The method of claim 1, wherein the method is based on B-rep self-supervised learning for manufacturing process strategy prediction. The B-rep data includes topological entities of different levels of faces, loops, edges and vertices.

3. The method of claim 1, wherein: In step S1, different types of topological entities of faces, loops, edges and vertices are processed differently to form corresponding feature matrices.

4. The method of claim 1, wherein: In step S3, the face embedding representation and the connection matrix between faces are input into the process strategy prediction network, and the process strategy label corresponding to each face is predicted through the process strategy prediction network comprising a residual graph convolutional layer and a multilayer perception module.

Images