Automatic generation method of industrial spline surface selfing detection data set
By using parameterized sample generation and ADASYN algorithm optimization, a high-quality self-intersection detection dataset is automatically generated, solving the problem of low computational efficiency in traditional methods and achieving dataset availability and fairness. It is applicable to self-intersection detection of various surface models.
Patent Information
- Application Number
- CN202511068198.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-31
- Publication Date
- 2025-11-14
AI Technical Summary
When dealing with large-scale complex surface data, existing technologies suffer from low computational efficiency of traditional self-intersection detection methods and lack high-quality, well-structured self-intersection detection datasets, making it difficult to support the training of machine learning models.
We employ parametric sample generation, automatic label assignment, and the ADASYN algorithm to optimize the number of samples, generating a self-intersection detection dataset covering Bézier and NURBS surfaces. We automatically label samples through high-density grid sampling and 3D spatial distance detection, and combine the ADASYN algorithm to balance sample categories.
It achieves automated generation and strong controllability of self-crossing detection datasets, improves the usability and fairness of datasets, solves the problem of uneven distribution of self-crossed and non-self-crossed samples, and is suitable for tasks such as CAD modeling, geometric quality assessment, and design optimization.
Smart Images
Figure CN120951407A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of unmanned aerial vehicle (UAV) shape design technology, and in particular to an automatic generation method for industrial spline surface self-intersection detection datasets. Background Technology
[0002] Surface self-crossing detection is one of the core problems in computer-aided geometric design (CAGD), and it is widely used in CAD modeling, finite element mesh quality assessment, geometric optimization, and manufacturing process simulation. Traditional self-crossing detection methods rely on precise geometric calculations, such as surface sampling and mesh Boolean detection, interval analysis, and Bézier / NURBS surface parameter domain traversal. Although these methods have high accuracy, they face computational efficiency bottlenecks when processing large-scale complex surface data.
[0003] Leveraging the automatic geometric feature extraction capabilities of deep learning, researchers are attempting to introduce machine learning methods into the self-crossing detection problem. To support this direction, high-quality, well-structured self-crossing detection datasets containing both positive and negative samples are fundamental. However, real-world industrial data often suffers from uneven distribution, privacy restrictions, and poor accessibility. Therefore, there is an urgent need for an automated and highly controllable data generation method to construct a training dataset containing a large number of self-crossing and non-self-crossing samples. Summary of the Invention
[0004] In view of this, the purpose of this invention is to provide an automatic generation method for industrial spline surface self-intersection detection datasets, aiming to provide large-scale, highly diverse, and structurally clear data support for self-intersection detection algorithms.
[0005] The present invention solves the technical problem by adopting the following technical solution:
[0006] An automatic generation method for industrial spline surface self-intersection detection dataset includes the following steps:
[0007] S1, Parametric sample generation, the generated samples cover two types of parametric surfaces, including Bézier surfaces and NURBS surfaces;
[0008] S2, Automatic Label Assignment: For each generated surface, high-density grid sampling and three-dimensional spatial distance detection are used to determine whether there are self-intersecting regions on the surface. The results of this stage are used to automatically assign self-intersecting or non-self-intersecting labels to the samples.
[0009] S3, Sample quantity optimization: ADASYN synthesizes auxiliary samples for difficult-to-learn regions based on the distribution of self-interleaved samples in the feature space, thereby achieving class balance.
[0010] Furthermore, in S1, the order of the surface is set to 3 to 5.
[0011] Furthermore, in S1, parameterized sample generation also includes:
[0012] For control point layout, the NURBS surface control point mesh size is set to 6×6 to 10×10, where the control point coordinates are generated by normal perturbation and the nodes use uniform node vectors; the Bézier surface control point mesh size is set to 4×4 to 6×6 according to the order.
[0013] NURBS weighting uses a uniform weight in the initial stage.
[0014] Furthermore, in S2, the logic for determining whether a surface has a self-intersecting region includes:
[0015] S21, Are there two different parameter points whose spatial distance is less than a certain threshold?
[0016] S22, the distance between point pairs in the parameter space is greater than a set range to eliminate adjacent sampling errors.
[0017] Furthermore, in S3, methods for optimizing the sample size include:
[0018] S31, it is necessary to determine whether there is a significant class imbalance in the current training set. This process is evaluated by calculating the ratio of the number of minority class samples to the number of majority class samples, i.e., calculating the imbalance degree d. If it is lower than the set imbalance ratio, a sampling operation is triggered.
[0019]
[0020] Where d is the degree of imbalance, m s m is the number of minority class samples. l The number of samples in the majority class;
[0021] S32, for each minority class sample x i Calculate the proportion r of the majority class samples in its neighborhood. i The density distribution is obtained by normalization. Once the number of synthetic samples to be generated for each sample is determined, ADASYN begins the specific synthesis process; each new sample is generated through random interpolation, that is, linear interpolation between the target minority class sample and its neighboring minority class samples.
[0022]
[0023] Where G is the total number of samples to be generated, g i To generate the number of synthetic samples, It is a density distribution;
[0024] S33, Construct new non-self-interleaved samples using a random interpolation method:
[0025] s i =x i +λ·(x si -x i ), λ∈[0,1]
[0026] Where, x si For x i One of the neighbors.
[0027] The present invention discloses an automatic generation method for industrial spline surface self-intersection detection dataset, which has the following beneficial effects:
[0028] This invention provides an automatic generation method for industrial spline surface self-intersection detection datasets, achieving automated and highly controllable data generation. By combining ADASYN technology for sample class balancing, the usability and fairness of the generated dataset are significantly improved, solving the problem of uneven distribution of self-intersection and non-self-intersection samples. The designed process is compatible with both Bézier and NURBS mainstream surface models and can be widely applied to tasks such as CAD modeling, geometric quality assessment, and design optimization. Attached Figure Description
[0029] Figure 1 This is a flowchart of the method of the present invention;
[0030] Figure 2 To generate self-crossing sample curves randomly for this invention;
[0031] Figure 3 This is a schematic diagram of the sampling points on the curved surface.
[0032] Figure 4 A schematic diagram of point pairs that meet the conditions;
[0033] Figure 5 A schematic diagram of the original test set for the Bézier surface;
[0034] Figure 6 A schematic diagram of the final test set for Bézier surfaces;
[0035] Figure 7 A schematic diagram of the original test set for NURBS surfaces;
[0036] Figure 8 This is a schematic diagram of the final test set for NURBS surfaces. Detailed Implementation
[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0038] In actual generated surfaces, the number of self-intersecting samples often far exceeds that of non-self-intersecting samples. To address the problem of uneven class distribution, this invention introduces the ADASYN (Adaptive Synthetic Sampling) algorithm for sample augmentation.
[0039] refer to Figure 1 This invention provides an automatic generation method for industrial spline surface self-intersection detection datasets, comprising the following steps:
[0040] S1, Parametric Sample Generation: The generated samples cover two types of parametric surfaces, including Bézier surfaces and NURBS surfaces. The surface order is set from 3 to 5, covering common design complexities, as shown in Table 1.
[0041]
[0042]
[0043] Table 1
[0044] For control point layout, the NURBS surface control point mesh size is set to 6×6 to 10×10, where the control point coordinates are generated by normal perturbation and the nodes use uniform node vectors; the Bézier surface control point mesh size is set to 4×4 to 6×6 according to the order.
[0045] NURBS weight settings initially use uniform weights (all set to 1), but can also be extended to random weight distribution to enhance diversity.
[0046] S2, Automatic Label Assignment: For each generated surface, the existence of self-intersecting regions is determined through high-density mesh sampling (e.g., 100×100) and 3D spatial distance detection. The determination logic includes:
[0047] 1) Are there two different parameter points whose spatial distance is less than a certain threshold?
[0048] 2) The distance between point pairs in the parameter space is greater than a set range to exclude adjacent sampling errors. The results of this stage are used to automatically assign self-crossing or non-self-crossing labels to samples. Figure 2 (a) is a self-crossed sample. Figure 2(b) is a non-self-crossed sample.
[0049] S3, Sample Quantity Optimization: ADASYN synthesizes auxiliary samples for difficult-to-learn regions based on the distribution of self-interleaved samples in the feature space, thereby achieving class balance. Its goal is to make machine learning models pay more attention to complex boundary regions during training, improving generalization ability.
[0050] Methods for optimizing sample size include:
[0051] S31, it is necessary to determine whether there is a significant class imbalance in the current training set. This process is evaluated by calculating the ratio of the number of minority class samples to the number of majority class samples, i.e., calculating the imbalance degree d. If it is lower than the set imbalance ratio, a sampling operation is triggered.
[0052]
[0053] Where d is the degree of imbalance, m s m is the number of minority class samples. l The number of samples in the majority class;
[0054] S32, for each minority class sample x i Calculate the proportion r of the majority class samples in its neighborhood. i The density distribution is obtained by normalization. Once the number of synthetic samples to be generated for each sample is determined, ADASYN begins the specific synthesis process; each new sample is generated through random interpolation, that is, linear interpolation between the target minority class sample and its neighboring minority class samples.
[0055]
[0056] Where G is the total number of samples to be generated, g i To generate the number of synthetic samples, It is a density distribution;
[0057] S33, Construct new non-self-interleaved samples using a random interpolation method:
[0058] s i =x i +λ·(x si -x i ), λ∈[0,1]
[0059] Where, x si For x i One of the neighbors.
[0060] This invention generates a Bézier surface dataset, as shown in Table 2:
[0061]
[0062]
[0063] Table 2
[0064] The NURBS surface dataset generated by this invention is shown in Table 3:
[0065]
[0066]
[0067] Table 3
[0068] like Figure 3 and Figure 4 As shown, a high-density set of surface points (e.g., 100×100 sampling points) can be generated by uniformly sampling in the parameter domain, thus representing a three-dimensional surface as a discrete set of points. Subsequently, the three-dimensional Euclidean distance is calculated for all non-adjacent point pairs in the set. If two points are far apart in the parameter domain but their distance in three-dimensional space is less than a set threshold, then the surface is considered to have self-intersection. Based on the detection results, each sample is assigned a binary label of self-intersection or non-self-intersection, serving as a supervisory signal for subsequent machine learning training.
[0069] This invention provides an automated method for generating self-intersecting detection datasets for industrial spline surfaces, achieving automated and highly controllable data generation. By combining ADASYN technology for sample class balancing, the usability and fairness of the generated dataset are significantly improved, resolving the problem of uneven distribution of self-intersecting and non-self-intersecting samples. The designed process is compatible with both Bézier and NURBS surface models, making it widely applicable to tasks such as CAD modeling, geometric quality assessment, and design optimization. This invention can be used to explore the geometric behavior of Bézier and NURBS surfaces under different control point arrangements and parameter configurations; it achieves a balanced distribution of self-intersecting and non-self-intersecting samples, providing a training foundation for self-intersecting classification models; and it reduces the initial annotation cost of traditional detection algorithms, improving detection efficiency.
[0070] Example
[0071] This embodiment provides an automatic generation method for industrial spline surface self-intersection detection datasets, including the following steps:
[0072] Automatically generate surface datasets:
[0073] (1) Bézier surface test set
[0074] The Bézier surface sample construction is set to an order ranging from 3 to 5, corresponding to control point mesh sizes of 4×4, 5×5, and 6×6, respectively. Each order generates 150 samples, for a total of 450. (Reference) Figure 5Of the original samples, there were 412 self-crossed samples and 38 non-self-crossed samples.
[0075] refer to Figure 6 After ADASYN processing, the size and class distribution of the Bézier surface dataset were updated to 412 self-intersecting surfaces and 412 non-self-intersecting surfaces. The self-intersecting surfaces were the original samples, and the non-self-intersecting surfaces were a mixture of the original and synthetic samples. The final dataset contained a total of 824 samples. The class ratio of self-intersecting to non-self-intersecting surfaces was increased from 90.8%:9.2% to 50%:50%, which significantly improved the class imbalance.
[0076] (2) NURBS surface
[0077] The NURBS surface sample construction is set to an order ranging from 3 to 5, with control point mesh sizes from 6×6 to 10×10, resulting in 5 combinations. Each combination generates 200 samples, for a total of 1000 samples. (Reference) Figure 7 Of the original samples, there were 863 self-crossed samples and 137 non-self-crossed samples.
[0078] refer to Figure 8 After ADASYN processing, the size and class distribution of the NURBS surface dataset were updated to 863 self-intersecting surfaces and 863 non-self-intersecting surfaces. The self-intersecting surfaces are the original samples, and the non-self-intersecting surfaces are a mixture of the original samples and the synthetic samples. The final dataset contains a total of 1726 samples, and the ratio of self-intersecting to non-self-intersecting surface classes has increased from the original 13.7%:86.3% to 50%:50%.
[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for automatically generating a dataset for detecting self-intersection of industrial spline surfaces, characterized in that, Includes the following steps: S1, Parametric sample generation, the generated samples cover two types of parametric surfaces, including Bézier surfaces and NURBS surfaces; S2, Automatic Label Assignment: For each generated surface, high-density grid sampling and three-dimensional spatial distance detection are used to determine whether there are self-intersecting regions on the surface. The results of this stage are used to automatically assign self-intersecting or non-self-intersecting labels to the samples. S3, Sample quantity optimization: ADASYN synthesizes auxiliary samples for difficult-to-learn regions based on the distribution of self-interleaved samples in the feature space, thereby achieving class balance.
2. The method for automatically generating an industrial spline surface self-intersection detection dataset according to claim 1, characterized in that, In S1, the surface order is set to 3 to 5.
3. The method for automatically generating an industrial spline surface self-intersection detection dataset according to claim 2, characterized in that, In S1, parameterized sample generation also includes: For control point layout, the NURBS surface control point mesh size is set to 6×6 to 10×10, where the control point coordinates are generated by normal perturbation and the nodes use uniform node vectors; the Bézier surface control point mesh size is set to 4×4 to 6×6 according to the order. NURBS weighting uses a uniform weight in the initial stage.
4. The method for automatically generating an industrial spline surface self-intersection detection dataset according to claim 2, characterized in that, In S2, the logic for determining whether a surface has a self-intersecting region includes: S21, Are there two different parameter points whose spatial distance is less than a certain threshold? S22, the distance between point pairs in the parameter space is greater than a set range to eliminate adjacent sampling errors.
5. The method for automatically generating an industrial spline surface self-intersection detection dataset according to claim 3, characterized in that, In S3, methods for optimizing the number of samples include: S31, it is necessary to determine whether there is a significant class imbalance in the current training set. This process is evaluated by calculating the ratio of the number of minority class samples to the number of majority class samples, i.e., calculating the imbalance degree d. If it is lower than the set imbalance ratio, a sampling operation is triggered. Where d is the degree of imbalance, m s m is the number of minority class samples. l The number of samples in the majority class; S32, for each minority class sample x i Calculate the proportion r of the majority class samples in its neighborhood. i The density distribution is obtained by normalization. Once the number of synthetic samples to be generated for each sample is determined, ADASYN begins the specific synthesis process; each new sample is generated through random interpolation, that is, linear interpolation between the target minority class sample and its neighboring minority class samples. Where G is the total number of samples to be generated, g i To generate the number of synthetic samples, It is a density distribution; S33, Construct new non-self-interleaved samples using a random interpolation method: s i =x i +λ·(x si -x i ),λ∈[0,1] Where, x si For x i One of the neighbors.