Tubular part classification method based on deep learning
By using a deep learning-based method to acquire 3D point cloud data of tubular parts, local cylinder fitting and topology mapping are performed, solving the problems of scattered feature representation and poor robustness in the identification of tubular parts in the existing technology, and achieving efficient and accurate classification of tubular parts.
Patent Information
- Application Number
- CN202511978316.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies for identifying tubular parts suffer from fragmented feature representations, a lack of unified semantic representation, an inability to automatically learn and infer topological connections, poor robustness, low computational efficiency, and difficulty in meeting the needs of real-time automated production.
A deep learning-based approach is adopted to acquire 3D point cloud data of tubular parts, perform preprocessing and local cylinder fitting, extract local geometric features, construct a topological structure map, and use graph neural networks to fuse global topological features to generate semantic feature representations, and finally classify tubular parts.
It achieves efficient and accurate classification of tubular parts, significantly reduces the amount of computation, improves recognition accuracy and robustness, and meets the needs of real-time automated production.
Smart Images
Figure CN121962713A_ABST
Abstract
Description
A Deep Learning-Based Classification Method for Tubular Parts Technical Field
[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a method for classifying tubular parts based on deep learning. Background Technology
[0002] Although tubular parts are all macroscopically classified as "cylindrical structures," their shapes vary greatly in actual manufacturing and application. Common types include straight pipes, bends, tees, crosses, reducers, and branch pipe assemblies. Different types differ only in their geometric topology, while their surface texture, material, and even color may be completely identical. Therefore, identifying tubular parts is difficult and inaccurate.
[0003] The following are some existing identification and classification methods:
[0004] Traditional geometric fitting-based methods analyze point clouds using algorithms such as RANSAC cylinder fitting and curve fitting. They are acceptable for identifying simple structures, but they have poor stability for complex multi-branch structures and are sensitive to noise and highly dependent on parameters.
[0005] Shape recognition methods based on two-dimensional images use edge detection, Hough transform, and other techniques to process two-dimensional projections. However, they are difficult to distinguish between pipe fittings that are topologically similar but structurally different in three-dimensional space, such as L-shaped bends and tees that look similar from the side.
[0006] The CAD template matching method aligns and matches the scanned data with the standard CAD model. However, it involves a large amount of computation, is sensitive to tolerances and deformations, and has difficulty in identifying non-standard or composite structures.
[0007] Traditional machine learning methods use models such as SVM and KNN, but rely on manual extraction of geometric features, lack end-to-end learning capabilities, and have insufficient generalization ability when new forms appear.
[0008] In summary, the existing technologies have the following main drawbacks: feature representation is scattered and lacks a unified semantic representation; they cannot automatically learn and infer topological connection relationships; they have poor robustness to noise, occlusion and data heterogeneity; and their computational efficiency is low, making it difficult to meet the needs of real-time automated production. Summary of the Invention
[0009] To address the shortcomings of existing technologies, this invention provides a deep learning-based method for classifying tubular parts, thereby solving the technical problem that existing technologies do not perform combined monitoring of dual hosts.
[0010] This invention provides a deep learning-based method for classifying tubular parts, comprising the following steps:
[0011] Step 1: Obtain the spatial information of the tubular part and generate 3D point cloud data;
[0012] Step 2: Preprocess the data obtained in Step 1, and perform local cylindrical fitting and geometric segmentation based on the preprocessed data to obtain several cylindrical segments and their corresponding parametric representations;
[0013] Step 3: Extract the local geometric features of each cylindrical segment and generate a high-dimensional vector descriptor;
[0014] Step 4: Construct a topology diagram based on the spatial relationships between the cylindrical segments;
[0015] Step 5: Perform global topological feature fusion on the topological graph using a graph neural network to generate semantic feature representations of each node and the overall structure;
[0016] Step 6: Classify tubular parts based on semantic feature representation.
[0017] Furthermore, in step 2, the specific method for preprocessing is as follows: preprocessing is performed through outlier removal, normal estimation, and voxel downsampling.
[0018] Furthermore, in step 2, the multi-scale RANSAC fitting algorithm is used to perform local cylinder fitting.
[0019] Furthermore, in step 2, the parameterized representation of the cylindrical segment includes: centerline vector, radius, length, direction, and fitting confidence level.
[0020] Furthermore, in step 3, a point cloud-based deep learning network extracts local geometric features from each cylindrical segment, and then a PointNet++ network is used to extract high-dimensional feature vectors that reflect the geometric shape of the cylindrical segments.
[0021] Furthermore, in step 4, in the constructed topology graph, cylindrical segments are used as nodes, and the connections between cylindrical segments are used as edges.
[0022] Furthermore, in step 5, a multi-task loss function is used during the training of the graph neural network, specifically as follows:
[0023] ;
[0024] In the formula, Llocal is the local loss term for constraining the local geometric features of the cylindrical segments, Ltopo is the topology loss term for constraining the consistency of topological relationships between the cylindrical segments, and Lclass is the classification loss term for classifying tubular parts. , , These are the weighting coefficients for the corresponding loss terms.
[0025] Furthermore, in step 6, the classification process also includes optimizing the classification results using a low-confidence automatic re-judgment mechanism, specifically as follows:
[0026] When the confidence level is lower than the preset value, the historical frame cache or point cloud data from adjacent angles are called for re-identification. Alternatively, when the confidence level is lower than the preset value, the topology of the tubular part is checked for consistency through the structural reasoning process to correct the classification result. The structural reasoning process is a process of rule-based reasoning of the structural type of the tubular part based on the number of connections, connection angles, and branching relationships between cylindrical segments.
[0027] The beneficial effects of this invention are:
[0028] This invention makes targeted improvements at the level of computer algorithms and model design, and focuses on solving the problems of high computational complexity of traditional geometric rule algorithms, difficulty in expressing topological relationships by deep learning models, and poor compatibility with heterogeneous inputs (point clouds, images):
[0029] This invention is based on a unified data representation mechanism using cylindrical segment parameterization. First, a cylindrical parameterized data structure is established in computer memory. Local geometric units are represented by quintuples of cylindrical segments, automatically fitted from point cloud or image data. This approach is both compact and fully expresses geometric features, providing a unified input format for subsequent network learning. Through this quintuple-based parameterization mechanism, the system can significantly reduce redundant point cloud computing load (reducing data processing overhead by approximately 70%), while simultaneously achieving format unification for multi-source data (such as laser scanning, structured light, and camera data).
[0030] This invention presents a hybrid feature learning framework integrating PointNet++ and Graph Neural Networks (GNNs): it can simultaneously capture local geometric features and global connectivity semantics. The first stage uses a PointNet++ network to extract the local feature vector for each cylindrical segment; the second stage treats the cylindrical segments as graph nodes, maps connectivity relationships to graph edges, and utilizes a Graph Neural Network (GNN) for global topological feature inference. This hybrid dual-network framework, when implemented on a computer, reduces memory usage and improves model inference speed by sharing the feature encoder and sparse adjacency matrix computation.
[0031] This invention employs a multi-task joint optimization learning strategy. During the training phase, the system designs a multi-task joint optimization objective function that includes local reconstruction loss, topological consistency loss, and classification loss, ensuring that the model maintains geometric accuracy while possessing topological consistency and classification accuracy. This method achieves stable convergence under a GPU parallel computing architecture and improves recognition accuracy by approximately 15% compared to single-task training on the same dataset. Attached Figure Description
[0032] The features and advantages of the invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the invention in any way. In the drawings:
[0033] Figure 1 is a flowchart illustrating a specific embodiment of the present invention. Detailed Implementation
[0034] 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.
[0035] The present invention will be further illustrated below with reference to specific embodiments. Those skilled in the art should understand that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Modifications to the present invention in various equivalent forms all fall within the scope defined by the appended claims.
[0036] As shown in Figure 1, this invention provides a deep learning-based method for classifying tubular parts, comprising the following steps:
[0037] Step 1: In the industrial field, use a laser scanner, structured light 3D camera or industrial binocular camera to acquire the spatial information of the tubular part, generate 3D point cloud data or multi-view image data; and store the data in the cache area using a unified format (such as PLY, OBJ or image sequence).
[0038] Step 2: Preprocess the data obtained in Step 1. Specifically, outlier removal, normal estimation, and voxel downsampling can be used to obtain point cloud data with complete structure and sparse uniformity, thereby reducing the computational burden and improving the feature extraction accuracy.
[0039] The multi-scale RANSAC fitting algorithm is used to divide the overall point cloud of the pipe fitting into several cylindrical segments, and the centerline, radius, length, orientation, and confidence level of each segment are calculated. The generated geometric segments are in the form of quintuples.
[0040] {v, r, l, θ, conf}
[0041] Where v is the centerline; r is the radius; l is the length; θ is the direction; and conf is the confidence level.
[0042] The quintuple form is stored as a cylindrical segment and its corresponding parameterized representation, which serves as the input for subsequent feature extraction.
[0043] All cylindrical segment features are converted into a unified parameter vector table, normalized, and then encapsulated into a standard data block (DataBlock) to complete data standardization and format encapsulation, achieving compatibility between different acquisition devices.
[0044] Step 3: Use a point cloud-based deep learning network to extract local geometric features of each cylindrical segment. Use the PointNet++ network to extract high-dimensional feature vectors that reflect the geometric shape of the cylindrical segment and generate a high-dimensional vector descriptor (LocalDescriptor). Local geometric features include information such as roundness, curvature distribution, and axial direction consistency.
[0045] Step 4: Using each cylindrical segment as a node, automatically determine the connection, intersection, or branch relationship between the cylindrical segments, and establish a topology diagram based on the spatial distance and connection relationship;
[0046] Step 5: Perform global topological feature fusion on the topological graph using a graph neural network to generate semantic feature representations of each node and the overall structure. These semantic feature representations can describe the complete geometric and topological semantic structure, providing high-level input for the subsequent classification process.
[0047] During the training of the graph neural network, a multi-task loss function is employed to jointly optimize local feature reconstruction error, topological consistency constraints, and classification accuracy, achieving simultaneous learning of geometric structure recognition and type determination. Model weights are updated in a distributed parallel manner on a computer GPU architecture, ensuring convergence stability and training efficiency. The specific formula for the multi-task loss function is as follows:
[0048] ;
[0049] In the formula, Llocal represents the local loss term used to constrain the learning of local geometric features of the cylindrical segment; Ltopo represents the topology loss term used to constrain the consistency of topological relationships between cylindrical segments; and Lclass represents the classification loss term used for classifying tubular parts. , , These are the weighting coefficients for the corresponding loss terms.
[0050] Step 6: Classify tubular parts based on semantic feature representation and output confidence scores.
[0051] During the classification process, the fused semantic feature representation is input into the Classifier Head, and the probability distribution of various pipe fittings is output through Softmax to determine the categories such as straight pipes, bends, tees, crosses, and reducers. Simultaneously, the confidence score and prediction variance for each classification are generated to facilitate subsequent uncertainty management.
[0052] The classification results are optimized using a low-confidence automatic re-judgment mechanism, specifically as follows:
[0053] When the confidence level of the classification result is lower than the preset threshold, the historical frame cache or point cloud data from adjacent angles are called for re-identification. Alternatively, when the confidence level of the classification result is lower than the preset threshold, the topology of the tubular part is checked for consistency through the structural reasoning process to correct the classification result. The structural reasoning process is a process of rule-based reasoning of the structural type of the tubular part based on the number of connections, connection angles and branching relationships between cylindrical segments.
[0054] When a new identification sample is discovered, the new identification sample and the result of manual confirmation are stored in the experience sample pool. The sample is then continuously optimized through a self-supervised comparative learning mechanism to achieve online learning and model adaptation.
[0055] The final identification results (including category labels, structural parameters, and confidence levels) are written into the production line control database and fed back to the host computer system via API interface for subsequent sorting, assembly, or inspection processes.
[0056] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.
Claims
1. A deep learning-based method for classifying tubular parts, characterized in that, The process includes the following steps: Step 1: Acquire spatial information of the tubular part and generate 3D point cloud data; Step 2: Preprocess the data acquired in Step 1, and perform local cylindrical fitting and geometric segmentation based on the preprocessed data to obtain several cylindrical segments and their corresponding parametric representations; Step 3: Extract the local geometric features of each cylindrical segment and generate high-dimensional vector descriptors; Step 4: Construct a topology diagram based on the spatial relationships between the cylindrical segments; Step 5: Perform global topology feature fusion on the topology diagram using a graph neural network to generate semantic feature representations of each node and the overall structure; Step 6: Classify the tubular part based on the semantic feature representations.
2. The deep learning-based tubular parts classification method as described in claim 1, characterized in that, In step 2, the specific method for preprocessing is as follows: preprocessing is performed by outlier removal, normal estimation, and voxel downsampling.
3. The deep learning-based tubular parts classification method as described in claim 1, characterized in that, In step 2, the multi-scale RANSAC fitting algorithm is used to perform local cylinder fitting.
4. The deep learning-based tubular parts classification method as described in any one of claims 1-3, characterized in that, In step 2, the parameterized representation of the cylindrical segment includes: centerline vector, radius, length, direction, and fitting confidence level.
5. The deep learning-based tubular parts classification method as described in claim 1, characterized in that, In step 3, a deep learning network based on point cloud extracts local geometric features of each cylindrical segment, and then a PointNet++ network is used to extract high-dimensional feature vectors that reflect the geometric shape of the cylindrical segments.
6. The deep learning-based tubular parts classification method as described in claim 1, characterized in that, In step 4, in the constructed topology graph, cylindrical segments are used as nodes, and the connections between cylindrical segments are used as edges.
7. The deep learning-based tubular parts classification method as described in claim 1, characterized in that, In step 5, a multi-task loss function is used during the training of the graph neural network, specifically: In the formula, Llocal is the local loss term for constraining the local geometric features of the cylindrical segments, Ltopo is the topology loss term for constraining the consistency of topological relationships between the cylindrical segments, and Lclass is the classification loss term for classifying tubular parts. 、 、 These are the weighting coefficients for the corresponding loss terms.
8. The deep learning-based tubular parts classification method as described in claim 1, characterized in that, In step 6, the classification process also includes optimizing the classification results using a low-confidence automatic re-judgment mechanism. Specifically, when the confidence level is lower than a preset value, historical frame cache or point cloud data from adjacent angles are called for re-identification. Alternatively, when the confidence level is lower than a preset value, the topological structure of the tubular part is checked for consistency through a structural reasoning process to correct the classification results. The structural reasoning process is a process of rule-based reasoning of the structural type of the tubular part based on the number of connections, connection angles, and branching relationships between cylindrical segments.