CN116401785A - MLP-Mixer-based assembly data management method and system - Google Patents
MLP-Mixer-based assembly data management method and system Download PDFInfo
- Publication number
- CN116401785A CN116401785A CN202310310928.9A CN202310310928A CN116401785A CN 116401785 A CN116401785 A CN 116401785A CN 202310310928 A CN202310310928 A CN 202310310928A CN 116401785 A CN116401785 A CN 116401785A
- Authority
- CN
- China
- Prior art keywords
- mixer
- mlp
- rep
- neural network
- rep model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 50
- 238000013523 data management Methods 0.000 title claims abstract description 15
- 238000013528 artificial neural network Methods 0.000 claims abstract description 44
- 238000012549 training Methods 0.000 claims abstract description 15
- 238000005457 optimization Methods 0.000 claims abstract description 11
- 230000008569 process Effects 0.000 claims description 19
- 239000000463 material Substances 0.000 claims description 10
- 239000011159 matrix material Substances 0.000 claims description 10
- 238000009826 distribution Methods 0.000 claims description 9
- 239000013598 vector Substances 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 4
- 238000007726 management method Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 230000011218 segmentation Effects 0.000 claims description 2
- 230000007306 turnover Effects 0.000 claims description 2
- 238000011960 computer-aided design Methods 0.000 abstract description 17
- 230000000694 effects Effects 0.000 abstract description 8
- 230000006870 function Effects 0.000 description 14
- 239000003086 colorant Substances 0.000 description 5
- 238000013527 convolutional neural network Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 238000000429 assembly Methods 0.000 description 2
- 230000000712 assembly Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000017105 transposition Effects 0.000 description 2
- 102100028162 ATP-binding cassette sub-family C member 3 Human genes 0.000 description 1
- 102100028187 ATP-binding cassette sub-family C member 6 Human genes 0.000 description 1
- 101000986633 Homo sapiens ATP-binding cassette sub-family C member 3 Proteins 0.000 description 1
- 101000986621 Homo sapiens ATP-binding cassette sub-family C member 6 Proteins 0.000 description 1
- 101001014572 Homo sapiens MARCKS-related protein Proteins 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 230000037396 body weight Effects 0.000 description 1
- 230000002860 competitive effect Effects 0.000 description 1
- 230000008094 contradictory effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000004806 packaging method and process Methods 0.000 description 1
- 230000035515 penetration Effects 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 238000012384 transportation and delivery Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/17—Mechanical parametric or variational design
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/18—Manufacturability analysis or optimisation for manufacturability
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Physics & Mathematics (AREA)
- Geometry (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Computer Hardware Design (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computational Mathematics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses an assembly data management method and system based on MLP-Mixer, the method includes: establishing a corresponding attribute adjacency graph according to the B-Rep model part information of the three-dimensional CAD model; taking the problem of link prediction between B-Rep model parts into consideration, carrying out optimization training on the MLP-Mixer neural network based on the attribute adjacency graph, outputting the characteristics of the B-Rep model parts and obtaining the trained MLP-Mixer neural network; considering the connection standard between the nodes of the B-Rep model part, inputting the characteristics of the B-Rep model part into the trained MLP-Mixer neural network for identification connection to obtain a model connection result; and establishing an assembly BOM structure by taking the predicted connection as a task view. By using the invention, the B-Rep format assembly parts are applied to the MLP-Mixer network, and the assembly pairing of the parts and the assembly data thereof are output so as to achieve the effect of simplifying assembly. The method and the system for managing the assembly data based on the MLP-Mixer can be widely applied to the technical field of complex assembly of computer aided design.
Description
Technical Field
The invention relates to the technical field of complex assembly of computer aided design, in particular to an assembly data management method and system based on an MLP-Mixer.
Background
The assembly function provided in Computer Aided Design (CAD) software is used to reduce the assembly cost of engineering experiments, and the assembly can be realized by means of CAD software only by using joint constraint, so that each part can be aligned with each other to complete the construction of the complex assembly.
Disclosure of Invention
In order to solve the technical problems, the invention aims to provide an assembly data management method and system based on an MLP-Mixer, which can output assembly pairing of parts and assembly data thereof by applying B-Rep format assembly parts to an MLP-Mixer network so as to achieve the effect of simplifying assembly.
The first technical scheme adopted by the invention is as follows: an assembly data management method based on MLP-Mixer comprises the following steps:
establishing a corresponding attribute adjacency graph according to the B-Rep model part information of the three-dimensional CAD model;
taking the problem of link prediction between B-Rep model parts into consideration, carrying out optimization training on the MLP-Mixer neural network based on the attribute adjacency graph, outputting the characteristics of the B-Rep model parts and obtaining the trained MLP-Mixer neural network;
considering the connection standard between the nodes of the B-Rep model part, inputting the characteristics of the B-Rep model part into the trained MLP-Mixer neural network for identification connection to obtain a model connection result;
and determining a task view according to the model connection result, and obtaining a material view, a process view and a quality view by combining the property of the part, thereby establishing an integrated assembly BOM structure taking the task as a core.
Further, the expression for establishing the corresponding attribute adjacency graph according to the B-Rep model part information of the three-dimensional CAD model is specifically as follows:
G(V,E)
in the above formula, G represents an attribute adjacency graph, V represents a face or point of the B-Rep model, E represents a set of edges, and the attribute adjacency graph edge E is defined by adjacency.
Further, the step of optimizing training the MLP-Mixer neural network based on the attribute adjacency graph by considering the problem of link prediction between the B-Rep model parts, outputting the characteristics of the B-Rep model parts and obtaining the trained MLP-Mixer neural network specifically comprises the following steps:
constructing an MLP-Mixer neural network;
inputting the attribute adjacency graph into an MLP-Mixer neural network for segmentation processing to obtain an attribute adjacency graph block;
mapping the attribute adjacent image blocks to obtain corresponding attribute adjacent image vectors;
combining the attribute adjacency graph vectors to obtain a corresponding matrix;
and (3) predicting a joint in the presence of a joint by considering the problem of link prediction between the B-Rep model parts, combining the matrix, outputting the characteristics of the B-Rep model parts and obtaining the trained MLP-Mixer neural network.
Further, the MLP-Mixer neural network includes a first multi-layer perceptron, a second multi-layer perceptron, and a shared encoder, wherein:
the first multi-layer perceptron is used for representing the vertex of the B-Rep model part surface;
the second multi-layer perceptron is used for representing the vertexes of the edges of the B-Rep model part;
the shared encoder is configured to perform message passing to obtain each vertex embedding of two attribute adjacency graphs.
Further, the loss function used by the MLP-Mixer neural network comprises two terms, namely a first term tau CE Is the edge prediction t uv And live edge label j uv Cross entropy error between E {0,1}, the second term is the symmetric cross entropy loss τ sym Wherein:
the expression of the first item cross entropy error is specifically as follows;
the expression of the second term symmetrical cross entropy loss is specifically as follows;
in the above, j uv Representing the positive solution of the label on the edge,
represent j uv Normalized probability distribution, t uv Representing prediction of edges +.>
Representing t uv T 2D Prediction representing 2D plane,/->
Normalized probability distribution of forward labels representing 2D planes,/->
Representation ofNormalized probability distribution of row positive solution tags, +.>
Normalized probability distribution, τ, representing a column positive solution label sym Representing the cross entropy of the system.
Further, the step of inputting the features of the B-Rep model part to the trained MLP-Mixer neural network for recognition connection to obtain a model connection result by considering the connection standard between the nodes of the B-Rep model part specifically comprises the following steps:
constructing a cost function of connection between B-Rep model parts by considering connection standards between nodes of the B-Rep model parts;
minimizing the cost function according to the principle of encouraging a larger contact area and punishing the overlapping volume, thereby obtaining a minimized cost function;
searching optimal part node pairing auxiliary parameters by adopting a Nelder-Mead algorithm as a standard in a derivative-free optimization mode, wherein the auxiliary parameters comprise offset distance along a joint axis, rotation around the joint axis and a turnover parameter for reversing the direction of the joint axis;
and inputting the characteristics of the B-Rep model part into the trained MLP-Mixer neural network, and connecting the B-Rep model part by combining the auxiliary parameters and the minimized cost function to obtain a model connection result.
Further, the connection standard between the nodes of the B-Rep model part comprises an overlapped volume and a contact area between parts, and the expression is specifically as follows:
in the above, V 1 、V 2 Representing the volume of two B-Rep model parts, V 1∩2 Representing the volume of overlap of two B-Rep model parts, A 1 、A 2 Representing the surface area of two B-Rep model parts, A 1∩2 Representing the contact area of two B-Rep model parts.
The second technical scheme adopted by the invention is as follows: an MLP-Mixer based assembly data management system, comprising:
the acquisition module is used for establishing a corresponding attribute adjacency graph according to the B-Rep model part information of the three-dimensional CAD model;
the training module is used for taking the problem of link prediction between the B-Rep model parts into consideration, carrying out optimization training on the MLP-Mixer neural network based on the attribute adjacency graph, outputting the characteristics of the B-Rep model parts and obtaining the trained MLP-Mixer neural network;
the connecting module is used for inputting the characteristics of the B-Rep model part into the trained MLP-Mixer neural network to carry out identification connection in consideration of the connecting standard between the nodes of the B-Rep model part, so as to obtain a model connecting result;
and the management module is used for determining a task view according to the model connection result, and obtaining a material view, a process view and a quality view by combining the property of the part, thereby establishing an integrated assembly BOM structure taking the task as a core.
The method and the system have the beneficial effects that: the invention uses the B-Rep format to assemble the parts, the format completely accords with the excellent performance of MLP-Mixer in CV, and uses the supervision of parameter CAD file, the file contains the B-Rep surface selected by the user and the edge defining the joint, and further uses the MLP-Mixer neural network to learn the B-Rep format parts of the mechanical product and output the assembly pairing of the parts and the assembly data thereof, thereby achieving the effect of simplifying the assembly.
Drawings
FIG. 1 is a flow chart of steps of an assembly data management method based on MLP-Mixer of the present invention;
FIG. 2 is a block diagram of an assembly data management system based on an MLP-Mixer of the present invention;
FIG. 3 is a schematic diagram of the structure of an MLP-Mixer neural network constructed in accordance with the present invention;
FIG. 4 is a schematic flow chart of an embodiment of the present invention;
FIG. 5 is a graphical representation of experimental data using the method of the present invention;
FIG. 6 is a diagram of the assembly effect obtained by applying the method of the present invention.
Detailed Description
The invention will now be described in further detail with reference to the drawings and to specific examples. The step numbers in the following embodiments are set for convenience of illustration only, and the order between the steps is not limited in any way, and the execution order of the steps in the embodiments may be adaptively adjusted according to the understanding of those skilled in the art.
3D visualization of complex assemblies by CAD software, convolutional Neural Networks (CNNs) are the preferred model of computer vision, but attention-based networks such as Vision Transformer have also become popular as Computer Vision (CV) techniques have evolved to use CV learning methods to attempt to find new solutions for CAD complex assemblies. Among these recently proposed MLP-Mixer is a completely multi-layer perceptron (MLP) based architecture that is simpler than but more effective than CNN, and the MLP-Mixer contains two types of layers: one is that the MLP is applied independently to the image blocks (i.e. "mix" each location feature), the other is that the MLP is applied across blocks (i.e. "mix" spatial information), and when trained on large datasets or using modern regularization schemes, the MLP-Mixer gets a competitive score on the image classification basis, with pretraining and reasoning costs comparable to the most advanced models. The nerve network applied to the CAD field at present is complex and the required calculation cost is also very high, the MLP-Mixer network has simple structure and relatively low calculation cost, but can achieve an effect comparable with the mainstream complex network, and the MLP-Mixer is combined with the invention, so that the calculation cost and the network construction cost can be reduced, and a better result can be achieved;
aiming at the defects of time consumption, labor consumption, data management confusion and the like of the traditional complex assembly and the advantages of simplicity and high performance of the MLP-Mixer, the invention uses a B-Rep format to assemble parts, wherein the B-Rep format is a graphic format, the excellent performance of the MLP-Mixer in CV is completely matched with the format, and the supervision of a parameter CAD file is used, wherein the file comprises a B-Rep surface selected by a user and edges defining joints.
Referring to fig. 1, the present invention provides an assembly data management method based on an MLP-Mixer, the method comprising the steps of:
s1, establishing a corresponding attribute adjacency graph according to the B-Rep model part information of a three-dimensional CAD model;
specifically, the invention uses standard parameter CAD files to perform weak supervision training without object classes. CAD assembled parts are typically represented in B-Rep format, containing watertight sets of trimmed parameter surfaces connected together by well-structured maps. Each surface comprises a parametric surface and is defined by edges that define the clipping range of the surface using parametric curves (e.g., straight lines, circular arcs, and circles). The B-Rep format is used for all mechanical CAD tools, and the selection of the B-Rep entities (i.e., facets and edges) is a critical but time-consuming manual task required to set the joint. The B-Rep entity on each part is selected to define the joint axis of each part consisting of the origin and the direction vector. The joint axis is determined by the type of geometric selection, and for a circle, the center point becomes the origin and the normal becomes the direction vector. The two parts may then be aligned along their axes to the assembled state. Constructing a graph representation G (V, E) from the B-Rep topology, wherein the graph vertex V is a B-Rep plane or point, and the graph edge E is defined by an adjacency;
further, 100 entities are extracted from the assembled part B-Rep entity dataset as a training set, and 100 different entities are used as a testing set. A graph representation G (V, E) is then constructed from the B-Rep entity topology, where the graph vertex V is the B-Rep plane or point and the graph edge E is defined by the adjacency, such that each entity is represented by a graph, giving two parts G that it is desired to assemble 1 、G 2 Having n and m vertices, respectively, forming a third "joint connection graph" G j It is tightly connected with G 1 And G 2 Apex between G j Having n m edges, by identifying G 1 And G 2 And a joint is formed between the two to make a link prediction problem.
S2, considering the problem of link prediction between B-Rep model parts, performing optimization training on the MLP-Mixer neural network based on the attribute adjacency graph, outputting the characteristics of the B-Rep model parts and obtaining the trained MLP-Mixer neural network;
specifically, mixer is composed of multiple layers of the same size, each layer being composed of two MLP blocks. The first is a label mix MLP: it acts on the columns of X (i.e., it applies to transposed input table X T ) Mapping R.fwdarw.R S And shared among all columns.
The present invention first delivers vertex feature V in two B-Reps through two separate multi-layer perceptrons (MLPs) 1 And V 2 To create graph vertex embedding. One MLP is used to represent the vertices of the B-Rep face and the other is used to represent the vertices of the B-Rep edge; the generated vertices are then connected together in an embedded manner. Next, messaging is performed in the graph of each part using the two-layer graph annotation network C to obtain each vertex embedding t of the two graphs 1 And t 2 。
t 1 ＝f(v 1 ，G 1 )
t 2 ＝f(v 2 ，G 2 )
The invention sets the joint axis prediction during assembly as a link prediction problem, and aims to correctly identify G 1 And G 2 The connection between the two parts, that is to say between the two fitting parts, aligns the two parts along the ground true joint axis; this is by along G j Is done using information between the edge convolution aggregation portions. Graph G 1 And G 2 Node characteristics V of (1) 1 And V 2 Delivered through the shared encoder network f of the present invention to obtain 384-dimensional embedded t 1 And t 2 Then for graph G j Medium density connection G 1 And G 2 The invention predicts that one indicates the presence of a joint:
wherein the method comprises the steps of
R 768 R is a 3-layer MLP, -/->
Is a join operator, and t u And t v According to G j The source and target vertices of each edge in (a) are from t 1 And t 2 And collecting, so that possible connection of the assembled parts can be better obtained.
The present invention applies the use of the loss function to two MLP blocks. First term τ CE Is the edge prediction t uv And live edge label j uv Cross entropy error between E {0,1}, normalized probability distribution as
The expression is as follows:
the subscript in the softmax operation here indicates that it applies to G j And CE (p, q) = - Σip i logq i . This loss encourages true joints to have higher values while suppressing non-joints, enabling all connected joints of the part to be learned.
In order to better focus the loss term, to better compare the joint with the more likely non-joint, the present invention uses symmetric cross entropy loss τ sym As a second term in the loss function, the expression is specifically as follows:
subscript of Softmax hereinIndicating that it is uniaxial, the 2D subscript indicates G j The predictive and forward labels on the edges are remodeled into an n x m matrix.
S3, considering connection standards among nodes of the B-Rep model part, inputting characteristics of the B-Rep model part into a trained MLP-Mixer neural network for identification connection, and obtaining a model connection result;
specifically, the network predicted B-Rep entity allows the present invention to query ground truth B-Rep data to obtain joint axis predictions for each part. Once the axes are aligned together, three auxiliary parameters define a rigid joint and can be used for joint pose prediction. Offset distance along the joint axis, rotation about the joint axis, and flipping parameters for reversing the direction of the joint axis. The present invention finds these parameters using a neural guided search that allows the present invention to enumerate top-k joint axis predictions and directly consider interactions between the two parts. To evaluate candidate joint configurations, the cost function S of the present invention joint ＝S overlap +λS contact It considers two general criteria for a well-defined joint: the overlap volume and the contact area between the parts are formulated as follows:
here, V 1 And V 2 Is the volume of two parts, V 1∩2 Indicating their possible overlapping volumes (i.e. the contact areas of the two parts). Similarly, A 1 And A 2 Is the surface area of the two parts, and the contact area of the two parts is A 1∩2 . Intuitively, to closely align the two parts to each other, minimizing the cost function should encourage a larger contact area while penalizing the overlap volume to prevent penetration. Thus, if a Coverlap<0.1, let λ= -10. Otherwise, the present invention sets λ=0 to increase the overlap penalty. Given this setThe present invention uses the Nelder-Mead algorithm as a standard derivative-free optimization to search for the optimal joint (pairing) pose, and once the pairing part and pairing pose are determined, the part can be automatically assembled.
Combining the step S2 and the step S3, and further;
the invention adopts G j G is input as j After matrix transposition, inputting MLP1 by taking columns as units, obtaining output with the same shape, then inputting MLP2 after transposition, and finally obtaining a result. The result is a matrix of co-input shapes, the presence of all possible pairs of joints can be estimated by subsequent processing, and finally, joint parameters are found by searching according to the predicted joint axis to complete the assembly. The goal of this process is to correctly identify G 1 And G 2 The connection between them, the two parts are truly aligned along the predicted joint axis. Graph G 1 And G 2 Node characteristics V of (1) 1 And V 2 Delivered through the shared encoder network f of the present invention to obtain 384-dimensional embedded t 1 And t 2 。
Then predicting the existing joint, the loss function is the first term τ cE Is the edge prediction t uv And live edge label j uv Cross entropy error between e {0,1 }.
A pair of parts in B-Rep format has a limited number of faces and edges which can mate to form a joint, particularly G j N x m sides of (c). Each combination produces a positive label in the n x m prediction space, with all remaining combinations being negative labels. For complex parts, for example, mechanical gears that may contain thousands of discrete B-Rep entities, this can lead to extreme imbalances between positive and negative labels. To increase the number of positive labels, the joints between the same pair of parts are combined into a joint set. This method presents a single data sample, i.e., a joint set, to the network that contains all known joints between a pair of parts. Importantly, the joint merging avoids presenting multiple contradictory data samples to the network, where negative labels in one sample may be positive labels in another sample.
S4, CAD complex assembly data management.
Specifically, the invention establishes an integrated assembly BOM structure with a task as a core according to the logic association relation among materials, processes, tasks and quality view information based on the thought of a single product data source. The assembly Process data can be mapped into corresponding information on Materials (M) view, process (P) view, quality (Q) view and Task (Task, T) view, which answer questions about what to assemble with, how to assemble, which information to record and by whom, respectively. In the assembly Process, the information of a single engine on the M, P, Q and T views can be respectively represented by a Material BOM (MBOM), a Process BOM (PBOM), a Quality BOM (QBOM) and a Task BOM (Task BOM, TBOM), and the nodes of each tree structure have logic association relations. The logical association between the BOM nodes in the four views can be described as: each task node corresponds to a unique component node (e.g. steering wheel assembly task t 11 Corresponding steering wheel m 11 ) The method comprises the steps of carrying out a first treatment on the surface of the Each task node may correspond to one or more quality table nodes (e.g., steering wheel assembly task t 11 Corresponding steering wheel inspection table q 11 Steering wheel assembly report q 12 Etc.); one-to-one correspondence between task nodes and process nodes (e.g. steering wheel assembly task t 11 Corresponding steering wheel assembly process p 11 ). During assembly, tasks drive the execution of the process and deduce the complete assembly data. Thus, with the task as the core, an assembled BOM of an integrated tree structure can be built by correlating the single tree structure of four views based on the concept of a single product data source. According to the different demands of personnel with different roles on data in the assembly process, a multi-role data demand network model is established, and a multi-color set theory is used for carrying out concise and visual mathematical description on the model on the basis, so that the storage and the processing of a computer are facilitated, and finally, a flow meeting the industrial requirements is provided;
further, the assembly process data can be mapped to corresponding information on the M-view, P-view, Q-view, and T-view, which answer questions about what to use for assembly, how to assemble, which information to record, and by whom, respectively.
According to polychromatic set theory, the polychromatic set element of MDDNM is T, M, P, Q four views, and a is used in sequence 1 ，a 2 ，a 3 ，a 4 Denoted as a= { a 1 ，a 2 ，a 3 ，a 4 }. The individual colors of the multi-color set represent the characteristics that the nodes in T, M, P, Q view may have, i.e. the task has the characteristics of start and finish time, executives, tooling equipment status, and the like, sequentially with a 11 ，a 12 ，a 13 A representation; the material has the characteristics of out-of-tolerance parts, vulnerable parts and external cooperative parts, and a is used in sequence 21 ，a 22 ，a 23 A representation; the process has the characteristics of working hour quota and key working procedures, and a is used in sequence 31 ，a 32 A representation; the quality has important quality inspection information, unqualified item control condition and quality item delivery characteristics, and a is used for the following steps 41 ，a 42 ，a 43 Denoted as F (a) = { a 11 ，a 12 ，a 13 ，a 21 ，a 22 ，a 23 ，a 31 ，a 32 ，a 41 ，a 42 ，a 43 }. The unified colors of the multicolor set are different role personnel with data requirements: operator, artist, quality inspector, user, manager, use b in order 1 ，b 2 ，b 3 ，b 4 ，b 5 Denoted as F (A) = { b 1 ，b 2 ，b 3 ，b 4 ，b 5 }. The correlation between the elements and the unified color is a boolean matrix a×f (a), each column representing which elements the corresponding unified color relates to. The correlation between the individual colors of the elements and the unified colors is a boolean matrix F (a) x F (a), each column representing which individual colors the corresponding unified color relates to. After A×F (a), A×F (A), F (a) ×F (A) matrixes are obtained, whether the requirement is related to a task is firstly judged, if yes, task characteristics (pointing out starting and finishing time, task executors and task specific tool equipment) are designated, and all task nodes which are assembled on a BOM and accord with the designated characteristics are searched; otherwise, all task nodes of the current engine are obtained. Then sequentially traversing all the coincident fingersAnd (3) determining the task node of the characteristic, acquiring the material, process and quality information conforming to the specified characteristic, packaging and returning, and finally providing a flow conforming to the industrial requirement.
In summary, the present invention aims to take the B-Rep format assembled parts as input, output the predicted assembly pairing information, etc., and manage the information of the assembly process, if CNN is used to learn the assembled parts, it may be relatively heavy, and to highlight the novelty of the present invention, the solid selection effect is relatively similar and the body weight is very light, and the new CV learning frame is an MLP-Mixer, which takes a series of S non-overlapping image blocks as input, each block is projected to the required hidden dimension C, all patches are linearly projected using the same projection matrix, the Mixer is composed of a plurality of layers with the same size, each layer is composed of two MLP blocks, and the present invention uses an MLP-Mixer neural network to learn the B-Rep format parts of the mechanical product, and output the assembly pairing of the parts and the assembled data thereof, so as to achieve the effect of simplifying the assembly.
Referring to fig. 2, an MLP-Mixer based assembly data management system, comprising:
the acquisition module is used for establishing a corresponding attribute adjacency graph according to the B-Rep model part information of the three-dimensional CAD model;
the training module is used for taking the problem of link prediction between the B-Rep model parts into consideration, carrying out optimization training on the MLP-Mixer neural network based on the attribute adjacency graph, outputting the characteristics of the B-Rep model parts and obtaining the trained MLP-Mixer neural network;
the connecting module is used for inputting the characteristics of the B-Rep model part into the trained MLP-Mixer neural network to carry out identification connection in consideration of the connecting standard between the nodes of the B-Rep model part, so as to obtain a model connecting result;
and the management module is used for determining a task view according to the model connection result, and obtaining a material view, a process view and a quality view by combining the property of the part, thereby establishing an integrated assembly BOM structure taking the task as a core.
The drawings of the present invention are described:
as shown in fig. 3, in the MLP-Mixer, the MLP is used by the Mixer Layer to map columns and rows successively, so as to realize information fusion of a space domain and a channel domain;
FIG. 4 shows the invention, wherein the characteristics of the assembled parts are extracted through an MLP-Mixer neural network, the assembled data are predicted through the characteristics of the parts, and the assembled joints and the assembly method are tested;
fig. 5 is a graph showing the experimental results showing the convergence curve of the loss function during training and testing. The experimental result shows that the MLP-Mixer neural network has better performance in the experiment, and can quickly realize the convergence of the loss function;
fig. 6 is a diagram showing the assembly effect, showing the flow of the assembly application.
The content in the method embodiment is applicable to the system embodiment, the functions specifically realized by the system embodiment are the same as those of the method embodiment, and the achieved beneficial effects are the same as those of the method embodiment.
While the preferred embodiment of the present invention has been described in detail, the invention is not limited to the embodiment, and various equivalent modifications and substitutions can be made by those skilled in the art without departing from the spirit of the invention, and these modifications and substitutions are intended to be included in the scope of the present invention as defined in the appended claims.
Claims (8)
1. An assembly data management method based on MLP-Mixer is characterized by comprising the following steps:
establishing a corresponding attribute adjacency graph according to the B-Rep model part information of the three-dimensional CAD model;
taking the problem of link prediction between B-Rep model parts into consideration, carrying out optimization training on the MLP-Mixer neural network based on the attribute adjacency graph, outputting the characteristics of the B-Rep model parts and obtaining the trained MLP-Mixer neural network;
considering the connection standard between the nodes of the B-Rep model part, inputting the characteristics of the B-Rep model part into the trained MLP-Mixer neural network for identification connection to obtain a model connection result;
and determining a task view according to the model connection result, and obtaining a material view, a process view and a quality view by combining the property of the part, thereby establishing an integrated assembly BOM structure taking the task as a core.
2. The method for managing assembly data based on MLP-Mixer according to claim 1, wherein the expression for establishing the corresponding attribute adjacency graph according to the B-Rep model part information of the three-dimensional CAD model is specifically as follows:
G(V,E)
in the above formula, G represents an attribute adjacency graph, V represents a face or point of the B-Rep model, E represents a set of edges, and the attribute adjacency graph edge E is defined by adjacency.
3. The method for managing assembly data based on MLP-Mixer according to claim 2, wherein the step of taking into account the problem of link prediction between B-Rep model components, performing optimization training on the MLP-Mixer neural network based on the attribute adjacency graph, outputting the characteristics of the B-Rep model components, and obtaining the trained MLP-Mixer neural network specifically comprises the steps of:
constructing an MLP-Mixer neural network;
inputting the attribute adjacency graph into an MLP-Mixer neural network for segmentation processing to obtain an attribute adjacency graph block;
mapping the attribute adjacent image blocks to obtain corresponding attribute adjacent image vectors;
combining the attribute adjacency graph vectors to obtain a corresponding matrix;
and (3) predicting a joint in the presence of a joint by considering the problem of link prediction between the B-Rep model parts, combining the matrix, outputting the characteristics of the B-Rep model parts and obtaining the trained MLP-Mixer neural network.
4. The method of claim 3, wherein the MLP-Mixer neural network comprises a first multi-layer sensor, a second multi-layer sensor, and a shared encoder, wherein:
the first multi-layer perceptron is used for representing the vertex of the B-Rep model part surface;
the second multi-layer perceptron is used for representing the vertexes of the edges of the B-Rep model part;
the shared encoder is configured to perform message passing to obtain each vertex embedding of two attribute adjacency graphs.
5. The method of claim 4, wherein the loss function used by the MLP-Mixer neural network comprises two terms, a first term τ CE Is the edge prediction t uv And live edge label j uv Cross entropy error between E {0,1}, the second term is the symmetric cross entropy loss τ sym Wherein:
the expression of the first item cross entropy error is specifically as follows;
the expression of the second term symmetrical cross entropy loss is specifically as follows;
in the above, j uv Representing the positive solution of the label on the edge,
represent j uv Normalized probability distribution, t uv Representing prediction of edges +.>
Representing t uv T 2D Prediction representing 2D plane,/->
A normalized probability distribution of the forward label representing the 2D plane,
normalized probability distribution representing row positive solution tags, < ->
Normalized probability distribution, τ, representing a column positive solution label sym Representing the cross entropy of the system.
6. The method for managing assembly data based on MLP-Mixer according to claim 5, wherein the step of inputting the features of the B-Rep model parts to the trained MLP-Mixer neural network for recognition connection to obtain the model connection result, taking into consideration the connection standard between the nodes of the B-Rep model parts, comprises the following steps:
constructing a cost function of connection between B-Rep model parts by considering connection standards between nodes of the B-Rep model parts;
minimizing the cost function according to the principle of encouraging a larger contact area and punishing the overlapping volume, thereby obtaining a minimized cost function;
searching optimal part node pairing auxiliary parameters by adopting a Nelder-Mead algorithm as a standard in a derivative-free optimization mode, wherein the auxiliary parameters comprise offset distance along a joint axis, rotation around the joint axis and a turnover parameter for reversing the direction of the joint axis;
and inputting the characteristics of the B-Rep model part into the trained MLP-Mixer neural network, and connecting the B-Rep model part by combining the auxiliary parameters and the minimized cost function to obtain a model connection result.
7. The method for managing assembly data based on the MLP-Mixer according to claim 6, wherein the connection standard between nodes of the B-Rep model parts comprises overlapping volumes and contact areas between parts, and the expression is as follows:
in the above, V 1 、V 2 Representing the volume of two B-Rep model parts, V 1∩2 Representing the volume of overlap of two B-Rep model parts, A 1 、A 2 Representing the surface area of two B-Rep model parts, A 1∩2 Representing the contact area of two B-Rep model parts.
8. An assembly data management system based on an MLP-Mixer is characterized by comprising the following modules:
the acquisition module is used for establishing a corresponding attribute adjacency graph according to the B-Rep model part information of the three-dimensional CAD model;
the training module is used for taking the problem of link prediction between the B-Rep model parts into consideration, carrying out optimization training on the MLP-Mixer neural network based on the attribute adjacency graph, outputting the characteristics of the B-Rep model parts and obtaining the trained MLP-Mixer neural network;
the connecting module is used for inputting the characteristics of the B-Rep model part into the trained MLP-Mixer neural network to carry out identification connection in consideration of the connecting standard between the nodes of the B-Rep model part, so as to obtain a model connecting result;
and the management module is used for determining a task view according to the model connection result, and obtaining a material view, a process view and a quality view by combining the property of the part, thereby establishing an integrated assembly BOM structure taking the task as a core.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310310928.9A CN116401785B (en) | 2023-03-28 | 2023-03-28 | MLP-Mixer-based assembly data management method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310310928.9A CN116401785B (en) | 2023-03-28 | 2023-03-28 | MLP-Mixer-based assembly data management method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116401785A true CN116401785A (en) | 2023-07-07 |
| CN116401785B CN116401785B (en) | 2023-12-08 |
Family
ID=87008439
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310310928.9A Active CN116401785B (en) | 2023-03-28 | 2023-03-28 | MLP-Mixer-based assembly data management method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116401785B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116611850A (en) * | 2023-07-14 | 2023-08-18 | 浙江春风动力股份有限公司 | System for detecting and tracing engine assembly quality curve |
| CN117746510A (en) * | 2024-02-19 | 2024-03-22 | 河海大学 | Real-time three-dimensional behavior recognition method based on point cloud feature map |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108305250A (en) * | 2018-01-30 | 2018-07-20 | 昆明理工大学 | The synchronous identification of unstructured robot vision detection machine components and localization method |
| CN113043275A (en) * | 2021-03-29 | 2021-06-29 | 南京工业职业技术大学 | Micro-part assembling method based on expert demonstration and reinforcement learning |
| CN114238676A (en) * | 2021-12-22 | 2022-03-25 | 芯勍(上海)智能化科技股份有限公司 | MBD model retrieval method and device based on graph neural network |
| CN114491841A (en) * | 2022-01-17 | 2022-05-13 | 浙江大学 | Machining feature recognition method based on NX secondary development and graph neural network |
| CN115049730A (en) * | 2022-05-31 | 2022-09-13 | 北京有竹居网络技术有限公司 | Part assembling method, part assembling device, electronic device and storage medium |
| CN115203935A (en) * | 2022-07-12 | 2022-10-18 | 厦门大学 | Frequency selection surface structure topology inverse prediction method and device based on deep learning |
-
2023
- 2023-03-28 CN CN202310310928.9A patent/CN116401785B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108305250A (en) * | 2018-01-30 | 2018-07-20 | 昆明理工大学 | The synchronous identification of unstructured robot vision detection machine components and localization method |
| CN113043275A (en) * | 2021-03-29 | 2021-06-29 | 南京工业职业技术大学 | Micro-part assembling method based on expert demonstration and reinforcement learning |
| CN114238676A (en) * | 2021-12-22 | 2022-03-25 | 芯勍(上海)智能化科技股份有限公司 | MBD model retrieval method and device based on graph neural network |
| CN114491841A (en) * | 2022-01-17 | 2022-05-13 | 浙江大学 | Machining feature recognition method based on NX secondary development and graph neural network |
| CN115049730A (en) * | 2022-05-31 | 2022-09-13 | 北京有竹居网络技术有限公司 | Part assembling method, part assembling device, electronic device and storage medium |
| CN115203935A (en) * | 2022-07-12 | 2022-10-18 | 厦门大学 | Frequency selection surface structure topology inverse prediction method and device based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| 张晶;崔汉国;朱石坚;: "基于人工神经网络的装配序列规划方法研究", 武汉理工大学学报(交通科学与工程版), no. 05, pages 1053 - 1056 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116611850A (en) * | 2023-07-14 | 2023-08-18 | 浙江春风动力股份有限公司 | System for detecting and tracing engine assembly quality curve |
| CN116611850B (en) * | 2023-07-14 | 2023-10-24 | 浙江春风动力股份有限公司 | System for detecting and tracing engine assembly quality curve |
| CN117746510A (en) * | 2024-02-19 | 2024-03-22 | 河海大学 | Real-time three-dimensional behavior recognition method based on point cloud feature map |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116401785B (en) | 2023-12-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN116401785B (en) | MLP-Mixer-based assembly data management method and system | |
| Sarlin et al. | Superglue: Learning feature matching with graph neural networks | |
| Shrestha et al. | Learned map prediction for enhanced mobile robot exploration | |
| Wang et al. | A parametric top-view representation of complex road scenes | |
| CN113961718A (en) | Knowledge inference method based on industrial machine fault diagnosis knowledge graph | |
| CN114240891A (en) | Welding spot quality identification method fusing knowledge graph and graph convolution neural network | |
| CN114758152A (en) | Feature matching method based on attention mechanism and neighborhood consistency | |
| Wang et al. | Contrastive consensus graph learning for multi-view clustering | |
| Zhou et al. | UTLNet: Uncertainty-aware transformer localization network for RGB-depth mirror segmentation | |
| Poli et al. | Neural hybrid automata: Learning dynamics with multiple modes and stochastic transitions | |
| Li et al. | Rearrangement planning for general part assembly | |
| Komatsu et al. | Octave deep plane-sweeping network: Reducing spatial redundancy for learning-based plane-sweeping stereo | |
| Ge et al. | Point2MM: Learning medial mesh from point clouds | |
| Jing et al. | Learning the dynamics of visual relational reasoning via reinforced path routing | |
| Wu et al. | Unsupervised joint 3d object model learning and 6d pose estimation for depth-based instance segmentation | |
| Cao et al. | WireframeNet: A novel method for wireframe generation from point cloud | |
| Hu | A novel deep learning driven robot path planning strategy: Q-learning approach | |
| Smedberg et al. | Trend mining 2.0: Automating the discovery of variable trends in the objective space | |
| CN108830860B (en) | Binocular image target segmentation method and device based on RGB-D constraint | |
| RU2791587C1 (en) | Method for providing computer vision | |
| Fang et al. | Direction-induced convolution for point cloud analysis | |
| Shen et al. | Action-conditional implicit visual dynamics for deformable object manipulation | |
| Ishibashi | Interactive texture chooser using interactive evolutionary computation and similarity search | |
| Xiang | Three-Dimensional Object Understanding and Visualization for Construction Collaborative Robotics | |
| Ping et al. | Unsupervised 3D Primitive Shape Detection using Mathematical Models |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |