CN100349163C

CN100349163C - Similarity estimating method for three-dimensional CAD model based on shape

Info

Publication number: CN100349163C
Application number: CNB200410067152XA
Authority: CN
Inventors: 王玉; 何玮; 肖熠中
Original assignee: Shanghai Jiaotong University
Current assignee: Shanghai Mould Technology Research Institute Co., Ltd.
Priority date: 2004-10-14
Filing date: 2004-10-14
Publication date: 2007-11-14
Anticipated expiration: 2024-10-14
Also published as: CN1614593A

Abstract

The present invention relates to a similarity estimating method for a three-dimensional CAD model based on a shape, which belongs to the technical field of the data management technique. The method comprises the following steps that: STEP AP 203 Part 21 data of a three-dimensional CAD model is converted to attributed graph data; a characteristic invariant of the three-dimensional CAD model is extracted from the attributed graph data so as to construct a characteristic invariant vector; the semblance evaluation is carried out to the CAD model by applying the characteristic invariant vector based on construction to a self-organizing neural network. Clustering of the three-dimensional CAD model and the CAD model index based on similarity is directly carried out by the method. The method of the present invention has the important characteristics of easy and agile algorithm, less resource consumption, high computational efficiency, self-learning performance, robustness and flexibility. The contained content of the characteristic invariant vector can be flexibly defined by a subscriber according to the actual demand of semblance evaluation, the vector can be long or short, and thus, the semblance evaluation requirements of different layers are satisfied.

Description

Three-dimensional CAD model is based on the similarity estimating method of shape

Technical field

The present invention relates to the similarity estimating method of a kind of three-dimensional CAD model based on shape, specifically is the similarity estimating method of a kind of three-dimensional CAD model of STEP form based on shape.Be used for the data management technique field.

Background technology

Along with popularizing of using of three-dimensional CAD and deeply, a large amount of product designs exists with the form of three-dimensional CAD model, simultaneously because the needs of exchanges data, this neutral file presentation format of STEP is widely adopted, and how effectively to manage and even reuse the extensive interest that these existing cad datas have caused Chinese scholars.Because it is traditional based on attribute, as the search method of file name, Part No. etc. not only exist result for retrieval wide or narrow usually, require the user to know many defectives such as Property Name, attribute also might change in time in advance, this method of what is more important can not be carried out the similarity assessment of cad model, thereby be difficult to realize to effective management of CAD design resource and reuse, so proposed somely, be one of them based on the method for figure based on shape similarity or content-based retrieval method.

Find by prior art documents, people such as Regli W C are at " Computer-Aided Design " 32 (2000), " the Managing digital libraries for computer-aideddesign " that delivers on the 119-132, (" computer-aided design (CAD) ", the management in digitized computer-aided design (CAD) storehouse) this article has been introduced the method based on figure, its basic step is at first cad model to be converted into figure to represent, utilizes the attribute of figure and isomorphism to calculate the similarity assessment that carries out model.Because the accurate coupling of figure is a very complex mathematical problem, counting yield is very low, and computational resource expends huge, has the scholar also to propose the approximate match algorithm.However, the defective that this class methods exist is: the computing method complexity, resource cost is big, travelling speed is slow, do not have self-learning function, is difficult to therefore satisfy that the Large-scale CAD data of database is managed and the needs of Knowledge Discovery.

Summary of the invention

The objective of the invention is to overcome deficiency of the prior art, the similarity estimating method of a kind of three-dimensional CAD model based on shape proposed, after the cad model to the STEP form carries out the attributed graph conversion, extract its feature invariant vector, adopt the self organizing neural network model that three-dimensional CAD model is carried out similarity assessment again, thereby make this method have that algorithm is easy flexibly, resource cost is few, operational efficiency is high and can dynamically upgrade distinguishing feature such as study, therefore can finely satisfy the needs of management of Large-scale CAD data of database and Knowledge Discovery.

The present invention is achieved by the following technical solutions, and the inventive method comprises three basic steps:

1. the STEP AP203 Part21 data with three-dimensional CAD model are converted to the attributed graph data

Use the string matching technology, method based on search key from STEP AP203 Part21 physical file is set up attributed graph G={N, E, Φ, Ω }, wherein, N is the set of the node of figure, E is the set on the limit of figure, and Φ is about the set of node set with the limit set relation of linking to each other, and Ω is the node of figure and the association attributes set on limit.Concrete grammar is: the type formation aggregated(particle) structure that utilizes the STEP file, find key word OPEN_SHELL or CLOSED_SHELL earlier, then search its following structure ADVANCED_FACE, and then search SURFACE below the ADVANCED_FACE etc., search corresponding EDGE by SURFACE again, so the whole STEP file of searching loop progressively successively finds every limit of each face correspondence, and writes down its mutual relationship and association attributes.Nodal community comprises face type, area of face etc., and the attribute on limit comprises the line style on limit, length etc.Last with specific file layout record attribute diagram data.

2. the dependency diagram data extracts the feature invariant of three-dimensional CAD model and constructs its feature invariant vector with this

The association attributes of computation attribute figure, as number of vertex, limit number, the maximum dimension in summit, summit smallest dimension, the average dimension in summit, diameter etc., extract the association attributes on summit and the limit of figure simultaneously, face type and area as the summit, the line style on limit and length etc., the aforementioned feature invariant that calculates is connected in series the feature invariant vector that forms the cad model that comprises a large amount of topological sum geological informations, and its dimension or content can be determined flexibly according to the actual needs of cad model similarity assessment.

3. use self organizing neural network based on the feature invariant vector of structure cad model is carried out similarity assessment

With the input of the feature invariant vector of cad model as self organizing neural network, neural network is carried out training study, the relative distance when utilizing convergence between the neural network output layer manifold neuron is measured the similarity of cad model.Distance is near more, and representation model is similar more; Otherwise the expression difference is big more.Concrete grammar is: the feature invariant that all cad models in the database are correlated with one by one extracts and structure feature invariant vector separately according to actual needs, and all feature invariant vectors finally constitute the network training sample set.Employing increases one-dimension method with vector input vector is carried out normalized.Selected network topology and correlation parameter are trained until convergence the input sample set, the relative distance in the record neural network output layer manifold between each feature invariant vector, thus the similarity between the cad model is made assessment.

Can directly carry out the cluster of three-dimensional CAD model and based on the retrieval of the cad model of similarity according to this method.The inventive method can be used for fields such as content-based three-dimensional CAD model search engine, CAD database Knowledge Discovery, the product design based on example, process planning and assessment of cost.The important feature of the inventive method is that it learns by oneself habit, robustness and dirigibility, the content that the user can comprise according to the flexible defined feature invariant vector of the actual needs of similarity assessment, and vector is changeable, thereby satisfies the similarity assessment needs of different levels.

Description of drawings

Fig. 1 the inventive method theory diagram

Fig. 2 network training cad model synoptic diagram

Embodiment

Specific implementation method of the present invention is followed following steps:

1. develop based on the search key method and read and change the attributed graph switching software module that STEP AP203 Part21 file is the attributed graph file.Utilize this module that the STEP file of all cad models that need train in the database is changed into the attributed graph data file.

2. develop the association attributes of figure to be calculated with the feature invariant of extraction figure and with this and construct the software module of cad model feature invariant vector based on the attributed graph data file.Invariant is except the number of vertex that comprises figure, limit number, the maximum dimension in summit, summit smallest dimension, the average dimension in summit, diameter etc., the association attributes that also comprises summit and the limit of figure simultaneously, contain 22 kinds of face types as the summit, the limit contains 28 kinds of line styles, adds two above-mentioned types of ancillary statistic statistics and line style face, line style quantity in addition at last again.So the aforementioned correlated characteristic invariant that calculates is connected in series and forms a feature invariant vector that comprises a large amount of topological sum geological informations of cad model, its dimension can be according to actual needs determines flexibly, and its length is generally between 1-60 when the type attribute on the summit of only considering figure and limit.Utilizing feature invariant vector extraction module that all cad models in the database are carried out the feature invariant one by one extracts to form feature invariant vector separately, the final input sample set that constitutes network training is the matrix of " (feature invariant vector dimension+1) * sample number " after the normalization.

3. develop the self organizing neural network training module.To network configuration parameters, comprise that output layer topology (rectangle or hexagon) and node number, input layer number, initial learn rate, field function parameter, field time parameter, learning rate time parameter, frequency of training etc. are provided with by this module.With one 9 * 1 dimension matrix stores all-network configuration parameter.Utilizing a random function initialization weight matrix, is the matrix of " output layer node number * (feature invariant vector dimension+1) ".Training module fan-in network training sample set Network Based begins training until convergence, the relative distance in the record neural network output layer manifold between each feature invariant vector, thus the similarity between the cad model is made assessment.

Embodiment

Develop attributed graph conversion and feature invariant vector extraction procedure and utilize the three-dimensional CAD model similarity Evaluation Platform of MFC exploitation with VC++6.0 based on self organizing neural network.The STEP AP203 Part21 file of collecting 32 practical cad model (see figure 2)s of industry carries out attributed graph data-switching and the extraction of feature invariant vector.Feature invariant vector dimension is defined as 17, is respectively 5 basic statistics amounts (number of vertex, limit number, the maximum dimension in summit, summit smallest dimension, the average dimension in summit) of attributed graph, 5 vertex attributes, the attribute on 5 limits, and 2 ancillary statistics.Forming one 17 * 32 training sample set like this, is 18 * 32 after the normalization.Be 31 seconds working time on the machine of Pentium 1.6G.Network configuration is: output layer topology is 12 regular hexagon for the length of side, and the initial learn rate is 0.1, and the field function parameter is 12, and the field time parameter is 621, and the learning rate time parameter is 1000, and frequency of training is 3800.Initial weight matrix (397 * 18) is produced by random function.Be 39 seconds working time on the machine of Pentium 1.6G.Table 1 is that cad model feature invariant vector extracts example, and table 2 has provided the similarity assessment result of two parts.This embodiment shows that the inventive method is feasible and effective.

Table 1 cad model feature invariant vector extracts example

Table 2 cad model similarity assessment example

Claims

1, a kind of three-dimensional CAD model is characterized in that based on the similarity estimating method of shape, comprises three basic steps:

1. the STEP AP203 Part21 data with three-dimensional CAD model are converted to the attributed graph data;

2. the dependency diagram data extracts the feature invariant of three-dimensional CAD model and constructs its feature invariant vector with this;

3. use self organizing neural network based on the feature invariant vector of structure cad model is carried out similarity assessment.

2, three-dimensional CAD model according to claim 1 is characterized in that based on the similarity estimating method of shape, and 1. described step is implemented as follows:

Use the string matching technology, method based on search key from STEP AP203 Part21 physical file is set up attributed graph G={N, E, Φ, Ω }, wherein, N is the set of the node of figure, E is the set on the limit of figure, and Φ is about the set of node set with the limit set relation of linking to each other, and Ω is the node of figure and the association attributes set on limit, concrete grammar is: the type formation aggregated(particle) structure that utilizes the STEP file, find key word OPEN_SHELL or CLOSED_SHELL earlier, then search its following structure ADVANCED_FACE, and then search the SURFACE below the ADVANCED_FACE, search corresponding EDGE by SURFACE again, so the whole STEP file of searching loop progressively successively finds every limit of each face correspondence, and writes down its mutual relationship and association attributes, described association attributes is exactly the attribute on dactylus point and limit, wherein nodal community comprises the face type of face, area, the attribute on limit comprises the line style on limit, length is at last with file layout record attribute diagram data.

3, three-dimensional CAD model according to claim 1 is characterized in that based on the similarity estimating method of shape, and 2. described step is implemented as follows:

The number of vertex of computation attribute figure, limit number, the maximum dimension in summit, summit smallest dimension, the average dimension in summit, diameter attribute, extract the face type and the area on the summit of figure simultaneously, the line style on limit and length attribute are connected in series the aforementioned feature invariant that calculates the feature invariant vector that forms the cad model that comprises a large amount of topological sum geological informations.

4, three-dimensional CAD model according to claim 1 is characterized in that based on the similarity estimating method of shape, and 3. described step is implemented as follows:

With the input of the feature invariant vector of cad model as self organizing neural network, neural network is carried out training study, relative distance when utilizing convergence between the neural network output layer manifold neuron is measured the similarity of cad model, and distance is near more, and representation model is similar more; Otherwise, the expression difference is big more, concrete grammar is: the feature invariant that all cad models in the database are correlated with one by one extracts and structure feature invariant vector separately, all feature invariant vectors finally constitute the network training sample set, employing increases one-dimension method with vector input vector is carried out normalized, selected network topology and correlation parameter, described correlation parameter comprises output layer topology and node number, the input layer number, the initial learn rate, the field function parameter, the field time parameter, the learning rate time parameter, frequency of training, the input sample set is trained until convergence, relative distance in the record neural network output layer manifold between each feature invariant vector, thus the similarity between the cad model is made assessment.