A two-stage real-time retrieval method for similar non-standard parts

Abstract

This invention relates to the field of 3D CAD model retrieval technology, specifically a two-stage real-time retrieval method for similar non-standard parts. This method includes two stages: offline database construction and online querying. In the offline stage, historical non-standard part STEP models are aligned and normalized, point cloud sampling is performed, ESF global feature extraction and compression encoding are conducted, and fine-grained matching feature calculations are performed. An Annoy approximate nearest neighbor index is constructed based on the compressed ESF vectors. In the online stage, the same preprocessing and feature calculations are performed on the model to be retrieved. First, a coarse search is conducted using compressed ESF vectors to obtain a candidate set. Then, a fine-grained ranking is performed using a fine-grained matching vector fused with principal inertial features and Zernike moments, outputting the ranking results of similar parts. This invention adopts a two-stage architecture from coarse to fine, requiring no training annotations. It can adapt to industrial scenarios where non-standard part samples are scarce and categories have long tails, achieving fast real-time retrieval while ensuring high retrieval accuracy.

Description

Technical Field

[0001] This invention relates to the field of 3D CAD model retrieval technology, specifically a two-stage real-time retrieval method for similar non-standard parts. Background Technology

[0002] In equipment manufacturing, automotive parts, aerospace, and specialized tooling design, non-standard parts typically exhibit significant structural variations, a long tail of category distribution, and high value for reusing historical models. When designing new parts, planning processes, or searching for similar structures, companies need to quickly find geometrically similar parts from a large pool of historical 3D models. 3D model similarity retrieval is a crucial means to achieve design reuse, reduce redundant modeling, and shorten product development cycles.

[0003] Existing 3D model retrieval schemes mainly include global descriptor retrieval, boundary representation or graph structure-based retrieval, local feature matching retrieval, and deep learning retrieval.

[0004] Traditional global descriptors are computationally efficient, but they are usually difficult to balance large-scale retrieval efficiency with fine-grained structural difference representation.

[0005] Schemes based on B-Rep or graph neural networks have high requirements for CAD topology analysis, data cleaning, and model training processes;

[0006] While local feature matching methods can describe local details, the computational cost of matching each feature one by one in a large-scale model library is significant.

[0007] Deep learning retrieval, such as the retrieval method, retrieval device, electronic device, and computer-readable storage medium disclosed in Chinese Patent Publication No. CN122223219A, includes a training process, a coarse screening process, and a fine sorting process. The drawback of this approach is that it relies on deep learning model training. Each time a new part category or a new structure family is introduced, the training set needs to be reorganized, the labels updated, and the model retrained or fine-tuned, resulting in high engineering maintenance costs.

[0008] Especially in scenarios where the non-standard parts library is constantly growing, newly added parts often lack category labels and similarity relationship annotations. If a deep learning approach of training before inference is adopted, each time a new part category or a new structure family is introduced, it may be necessary to reorganize the training set, update the labels, and retrain or fine-tune the model, resulting in high engineering maintenance costs. Conversely, manufacturing enterprises need a retrieval method that can directly utilize the existing STEP model library, quickly update the index after adding a new model, and stably output the ranking of similar parts during queries.

[0009] In summary, existing technical solutions struggle to simultaneously achieve training-free annotation, rapid online querying, and high-precision similar parts ranking within a real-world industrial model library of non-standard parts. Therefore, there is an urgent need for a retrieval method that can directly utilize existing STEP model libraries, rapidly update the index after adding new models, and stably output similar parts rankings during queries. Summary of the Invention

[0010] The purpose of this invention is to provide a two-stage real-time retrieval method for similar non-standard parts, so as to solve the problem that existing three-dimensional model retrieval schemes cannot simultaneously achieve training-free annotation, fast online query and high retrieval accuracy.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a two-stage real-time retrieval method for similar non-standard parts, comprising the following steps;

[0012] S1: Obtain the STEP model library of historical non-standard parts, preprocess the STEP models in the historical non-standard parts library, and obtain the normalized model;

[0013] S2: Perform point cloud sampling on the normalized model to obtain a fixed number of three-dimensional point clouds;

[0014] S3: Based on the sampling points, the global shape descriptor of each historical non-standard part is calculated using ESF, and the global shape descriptor of ESF is sequentially subjected to PCA dimensionality reduction and binary quantization encoding to obtain a compressed ESF vector; at the same time, the fine matching vector of each historical non-standard part is calculated, and the fine matching vector is a vector formed by standardizing and fusing the main inertial feature vector and the three-dimensional Zernike moment feature vector according to the preset weight coefficient α.

[0015] S4: Construct a coarse-grained Annoy index library based on the compressed ESF vector, and associate and store the fine-grained matching vector of each historical non-standard part with the corresponding part number;

[0016] S5: For the STEP model of the non-standard parts to be retrieved, perform the same preprocessing as step S1, the same point cloud sampling as step S2, and the same feature calculation as step S3 in sequence to obtain the query compressed ESF vector and the query fine matching vector.

[0017] S6: Use the query compressed feature vector to perform a coarse search in the approximate nearest neighbor index to obtain a set of candidate parts;

[0018] S7: Use the query fine matching vector to perform fine distance calculation and sorting on each candidate part in the candidate part set, and output the Top-K similar non-standard parts sorting results.

[0019] Preferably, the preprocessing in step S1 includes geometric analysis, coordinate alignment, scale normalization, and pose normalization, specifically:

[0020] S11: Analyze the three-dimensional geometric boundary of the STEP model and convert it into a sampleable geometric representation;

[0021] S12: Calculate the model's bounding box, centroid, or point cloud center, and translate the model to a unified coordinate origin;

[0022] S13: Scale normalization is performed based on the longest side of the bounding box, the diagonal length of the model, or the maximum radius of the point cloud, so that models of different sizes can enter a unified scale space.

[0023] S14: Obtain the three principal directions of the model through principal component analysis or principal axis inertia analysis, and rotate and align them according to the preset principal axis order;

[0024] S15: Combine the bounding box direction, point cloud distribution offset, or fixed direction rules to determine the positive and negative directions of the principal axis, and avoid feature inconsistencies caused by flipping the principal axis direction sign.

[0025] Preferably, the point cloud sampling in step S2 includes:

[0026] Candidate points are generated on the normalized STEP model surface.

[0027] A fixed number of point clouds is obtained by using uniform sampling, farthest point sampling, or hierarchical sampling methods, wherein the fixed number is 4096 points.

[0028] Preferably, in step S2, the point cloud sampling adopts the farthest point sampling method. The farthest point sampling is achieved by iteratively selecting the point farthest from the selected point set, so that the sampling points are evenly distributed on the model surface and the shape contour differences are preserved.

[0029] Preferably, the point cloud sampling further includes: adjusting the sampling weights based on local curvature, boundary lines, or local geometric changes to increase the sampling probability of holes, slots, bosses, and edge transition regions, so as to improve the ability of subsequent retrieval to express fine-grained structural differences.

[0030] Preferably, the PCA dimensionality reduction processing and binary quantization encoding in S3 include:

[0031] Principal component analysis is performed on the original 640-dimensional ESF features to reduce the dimensionality of the 640-dimensional features to a 512-dimensional PCA projection matrix. The original ESF vectors are then projected onto the low-dimensional space determined by the reduced projection matrix.

[0032] The binary quantization encoding uses an iterative quantization method to learn the orthogonal rotation matrix, converting the dimensionality-reduced 512-dimensional real-valued vector into a compact binary code, resulting in a compressed ESF vector.

[0033] Preferably, the calculation of the fine-grained matching vector in step S3 includes:

[0034] Calculate the principal inertial feature vector v of the sampled point cloud. prin The principal inertial eigenvector is composed of three eigenvalues ​​λ1, λ2, and λ3 of the covariance matrix or the inertial matrix.

[0035] Calculate the three-dimensional Zernike moment eigenvector v of the sampled point cloud. zer ;

[0036] The principal inertial feature vector and the three-dimensional Zernike moment feature vector are standardized and fused according to a preset weight coefficient α to obtain the fine matching vector. ,in and These are the standardized principal inertial eigenvectors and the three-dimensional Zernike moment eigenvectors, respectively. This indicates a vector connection.

[0037] Preferably, the construction parameters of the Annoy index library in step S4 include: vector dimension d. vec =16, Number of trees N tree =50, Number of retrieval nodes N node =6.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] 1. This invention addresses the challenges of non-standard parts in equipment manufacturing, automotive parts, aerospace, and specialized tooling design scenarios, characterized by significant structural differences, long-tailed category distribution, and high reuse value of historical models. It employs a two-stage architecture combining offline database construction and online querying, moving from coarse to fine. This architecture eliminates the need for category labels and deep learning model training, enabling direct retrieval of similar parts using the existing STEP model library. This completely avoids the high engineering maintenance costs associated with reorganizing the training set, updating labels, and retraining or fine-tuning the model when adding new part categories or structural families. The first stage rapidly recalls candidate parts by compressing global shape descriptors, while the second stage refines the candidate set using fine-grained matching vectors. This ensures both rapid online query efficiency with a large-scale model library and effective differentiation of fine-grained structural differences such as holes, slots, and steps. In industrial scenarios where the non-standard parts library is constantly growing, newly added parts can be quickly deployed with only incremental index updates, significantly improving design reuse efficiency, reducing repetitive modeling workload, and shortening product development cycles. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the overall process of the two-stage real-time retrieval method for non-standard parts of the present invention;

[0041] Figure 2 This is a schematic diagram illustrating the process of establishing the offline database and Annoy index library of this invention;

[0042] Figure 3 This is a schematic diagram of the online query and candidate ranking process of the present invention;

[0043] Figure 4 This is a schematic diagram of the global local feature vector construction process of the present invention. Detailed Implementation

[0044] 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, and 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.

[0045] This invention proposes a two-stage, coarse-to-fine real-time retrieval method for non-standard parts similarity. The first stage is the offline database construction stage, which performs geometric analysis, alignment normalization, point cloud sampling, global feature extraction, compression encoding, and fine feature calculation on historical non-standard parts STEP models, and establishes the feature vectors into an Annoy index library for fast nearest neighbor queries. The second stage is the online query stage, which performs the same preprocessing and feature calculation as the offline stage on the STEP model to be retrieved. First, a coarse search is performed using the compressed global ESF vectors to obtain a set of candidate parts, and then fine vector matching is performed on the candidate set to finally output a ranking of similar non-standard parts.

[0046] The core idea of ​​this solution is to move the feature calculation and index building of a large number of historical models to the offline stage. In the online stage, it is only necessary to calculate the corresponding features of the query parts and perform nearest neighbor retrieval. There is no need to train deep models for category labels, nor is it necessary to perform high-cost registration or local matching of the entire database models one by one during the query.

[0047] For ease of description, let the historical non-standard parts library be D={M1, M2, ..., M... n}, where M i Let represent the i-th historical STEP model, where in a specific embodiment n≈23000; the model to be retrieved is Q. The offline stage is for each M... i Construct the coarse search vector bi and the fine ranking vector v i And create an index I={b i ->i};In the online phase, the coarse retrieval vector b of the query model Q is calculated. q And the finely sorted vector v q First, I returns a candidate set C, and then the sorting results are output in C based on the fine-ranking distance.

[0048] Specifically, please refer to Figure 1-4 The present invention provides a technical solution: a two-stage real-time retrieval method for similar non-standard parts, comprising the following steps;

[0049] Offline database building stage:

[0050] S1: Obtain the historical non-standard parts STEP model library, preprocess the STEP models in the historical non-standard parts library to obtain normalized models, where preprocessing includes geometric analysis, coordinate alignment, scale normalization and posture standardization, specifically:

[0051] S11: Analyze the 3D geometric boundaries of the STEP model and convert them into a sampleable geometric representation;

[0052] S12: Calculate the bounding box, centroid, or point cloud center c of the model, where , To determine the number of point clouds, the model is translated to a unified coordinate origin. The translation formula is: , where p i This refers to the i-th point cloud data point on the model surface before or after sampling.

[0053] S13: Scale normalization is performed based on the longest side of the bounding box, the diagonal length of the model, or the maximum radius of the point cloud, so that models of different sizes enter a unified scale space. Specifically, the longest side of the bounding box, the diagonal length of the model, or the maximum radius of the point cloud is used as the scale normalization factor s, and the formula is: The model coordinates are divided by s for scale normalization, so that models of different sizes are placed in a unified scale space. The scale normalization formula is: ;

[0054] S14: Obtain the three principal directions of the model through principal component analysis or principal axis analysis. These three principal directions are the three eigenvector directions of the covariance matrix or inertia matrix, ordered from largest to smallest eigenvalue as the first, second, and third principal directions. Align the first, second, and third principal directions with the X, Y, and Z axes respectively according to a preset principal axis order to ensure consistent model attitude. Construct a rotation matrix R based on the determined principal axis directions to rotate the model to a standard attitude. The rotation formula is: ,in The points are normalized.

[0055] Calculate the point cloud covariance matrix C:

[0056] ;

[0057] Perform eigenvalue decomposition on the covariance matrix C:

[0058] ;

[0059] ;

[0060] The covariance matrix C mentioned above is used to estimate the principal directions of the model. U is the eigenvector matrix, Λ is the eigenvalue matrix, and the eigenvectors are arranged in descending order of their corresponding eigenvalues, forming the three principal axis directions of the model. Through the above alignment and normalization process, the feature calculations of historical parts and queried parts are under the same coordinate, scale, and orientation conditions.

[0061] S15: To avoid feature inconsistencies caused by flipping the principal axis direction sign, the positive and negative directions of the principal axis can be determined by combining the bounding box direction, point cloud distribution offset, or fixed direction rules.

[0062] S2: Perform point cloud sampling on the normalized model to obtain a fixed number of 3D point clouds. Point cloud sampling includes:

[0063] Candidate points are generated on the normalized STEP model surface.

[0064] A fixed number of point clouds is obtained by uniform sampling, farthest point sampling or hierarchical sampling. The fixed number is 4096 points. In this embodiment, the farthest point sampling method is used for point cloud sampling. The fixed number of point clouds can ensure the stability of the dimension of subsequent feature calculation. The farthest point sampling can reduce data redundancy while preserving the overall outline of the part, hole and groove boundaries and local abrupt areas as much as possible.

[0065] The specific process of sampling at the farthest point is as follows: First, randomly select an initial sampling point x1 from the model surface. The formula is: , This can represent the distance from a point to the model center, curvature weights, or randomly initialized weights. Then, for the t-th sampling point (t=2,3,...,m), the point furthest from the set of selected sampling points is chosen as the next sampling point: x t This represents the sampling result for the t-th farthest point, where m is the number of target sampling points (m=4096 in this embodiment). This represents the j-th sampling point in the selected point set;

[0066] For complex non-standard parts, the sampling weights can be adjusted based on curvature, boundary lines, or local geometric changes to give higher sampling probabilities to areas such as holes, slots, bosses, and edge transitions. This improves the ability of subsequent retrieval to express fine-grained structural differences. For each candidate initial sampling point x, its local curvature is calculated. and boundary indicators (e(x) = 1 when the point is located at a boundary, hole edge, or sharp transition region; otherwise, e(x) = 0). Adjust the sampling weights as follows: , where β and γ are preset weighting coefficients. Weighted sampling can enhance the expression of key local structures while maintaining overall uniformity.

[0067] S3: Based on sampling point cloud computing, the global shape descriptor of ESF for each historical non-standard part is obtained. The ESF descriptor can be obtained by statistical analysis of shape functions such as the distance between points, the area of ​​the triangle formed by random points, and the angle between random points. It is used to describe the overall geometric shape of the non-standard part. Since the original ESF feature dimension is high and the direct full-database matching is costly, the global shape descriptor of ESF is subjected to PCA dimensionality reduction and binary quantization encoding in sequence to obtain a compressed ESF vector. At the same time, the fine matching vector of each historical non-standard part is calculated. The fine matching vector is a vector formed by standardizing and fusing the main inertial feature vector and the three-dimensional Zernike moment feature vector according to the preset weight coefficient α.

[0068] PCA dimensionality reduction and binary quantization encoding include:

[0069] Principal component analysis (PCA) was performed on the original 640-dimensional ESF features of the historical parts to obtain a 512-dimensionality-reduced projection matrix. The original ESF vectors were then projected onto the low-dimensional space determined by the reduced projection matrix. The dimensionality reduction formula is as follows:

[0070] ;

[0071] ;

[0072] Where P is the normalized point cloud, x is the original ESF vector, μ is the mean vector obtained from offline statistics, and W pca Let W be the PCA dimension reduction projection matrix, z be the dimension reduction real-valued vector, and W be the vector of the PCA dimension reduction. pca It is learned from the historical parts library during the offline phase and is used consistently during the online phase;

[0073] Binary quantization encoding uses Iterative Quantization (ITQ) to learn the rotation matrix, transforming the dimensionality-reduced 512-dimensional real-valued vector into a 16-dimensional binary encoded vector, resulting in a compressed ESF vector. The goal of ITQ is to minimize the quantization error.

[0074] ;

[0075] Where Z is the historical part dimensionality reduction feature matrix, R is the orthogonal rotation matrix, B is the target binary encoding matrix, and b is the compressed ESF binary vector of a single part. Through this compression process, high-dimensional continuous ESF descriptors can be transformed into low-dimensional binary codes, reducing storage space and improving coarse retrieval speed.

[0076] In the online phase, the same projection matrix and quantization rules are used to generate a 16-dimensional binary code for the queried parts. This 16-dimensional binary code vector is used for fast calculation of the Hamming distance in the coarse retrieval phase. Hamming distance The smaller the value, the closer the global geometry of the two non-standard parts are. Similarity can be quickly calculated using Hamming distance. The formula for Hamming distance is:

[0077] ;

[0078] in, For Hamming distance, and For the compressed ESF binary encoded vectors of the two parts, The dimension of the binary encoded vector. and For the k-th bit of the vector, 1 This is an indicator function; it returns 1 if the condition is true, and 0 otherwise.

[0079] In this embodiment, the fine-matching vector is a standardized fusion vector of the principal inertial feature vector and the three-dimensional Zernike moment feature vector. The calculation of the fine-matching vector includes:

[0080] Calculate the principal inertial eigenvector v of the sampled point cloud prin The principal inertial eigenvector is composed of three eigenvalues ​​λ1, λ2, and λ3 of the covariance matrix, and is used to describe the overall spatial distribution characteristics of the part. The formula is as follows: ;

[0081] Calculate the 3D Zernike moment eigenvector v of the sampled point cloud zer Specifically, the sampling point cloud in step S2 is triangulated, and then the triangular facets are mapped onto a voxel mesh of a preset resolution to obtain surface voxels. The interior is determined by ray parity rules or 3D flooding filling. Local multi-point sampling is used in holes, slots, bosses, thin walls, and edge transition areas to improve the accuracy of voxel occupancy determination. Finally, a complete voxel model is obtained. Based on this voxel model, the spatial occupancy value of each voxel is weighted and summed with the value of the 3D Zernike polynomial basis function at the corresponding position to obtain the 3D Zernike moment eigenvector v. zer This feature vector can capture shape details of different orders, is used to describe the local topological details of the part, and has rotation invariance;

[0082] The principal inertial feature vector and the three-dimensional Zernike moment feature vector are standardized and fused according to a preset weight coefficient α to obtain a fine-grained matching vector. ,in and These are the standardized principal inertial eigenvectors and the three-dimensional Zernike moment eigenvectors, respectively. Indicates vector concatenation. The preset weighting coefficients are 0 < α ≤ 1.

[0083] S4: Construct a coarse-grained Annoy index based on compressed ESF vectors, and associate and store the fine-grained matching vectors of each historical non-standard part with the corresponding part number. The construction parameters of the Annoy index include: vector dimension d. vec =16, Number of trees N tree =50, Number of retrieval nodes N node =6;

[0084] At the same time, the fine matching vectors of each historical part The model is bound to the corresponding part number and stored as a key-value pair in an additional data file. When a new model is added to the historical parts library, only the same normalization, sampling, and feature calculation are performed on the new model, and the Annoy index is updated incrementally, without the need to retrain the classification model.

[0085] Online query stage:

[0086] S5: Input the STEP model of the non-standard part to be retrieved, and perform normalization, sampling, and feature vector calculation exactly the same as in the offline stage to obtain the compressed ESF vector of the query part. and fine matching vector .

[0087] S6: Compress feature vectors using queries A coarse search is performed in the near nearest neighbor index to obtain a candidate parts set. The top M nearest neighbor parts are then returned to form the candidate parts set. The coarse search formula is:

[0088] ;

[0089] in, This indicates the offline-built Annoy index library, and M represents the number of candidates returned in the coarse search stage (M=10*K in this embodiment). The recall rate and query time of the Annoy index can be adjusted by parameters such as the number of trees and the number of search nodes.

[0090] Annoy indexes divide a high-dimensional vector space by constructing multiple random projection trees, enabling them to quickly return near nearest neighbor results during queries, making them suitable for real-time retrieval of large-scale historical parts databases.

[0091] S7: Use the query fine matching vector to perform fine distance calculation and sorting on each candidate part in the candidate part set, and output the Top-K similar non-standard parts ranking results. The fine distance calculation adopts L2 Euclidean distance calculation, and the candidate part set is sorted in ascending order of distance.

[0092] The fine distance is calculated using L2 Euclidean distance, and the query fine matching vector is used. The candidate parts set obtained from the coarse search Fine-grained distance calculations are performed on each candidate part in the query. Specifically, the fine-grained matching vector of the query is calculated. Fine matching vectors of candidate parts The system is divided into principal inertia features and Zernike moment features. The L2 Euclidean equations for each feature are calculated separately, and then weighted and fused according to a weighting coefficient α to obtain the refined ranking distance.

[0093] ;

[0094] in, To find parts With candidate parts The distance between them, To query the fine-grained matching vector of a part, For the first Fine matching vector of candidate parts The L2 norm of a vector. The candidate parts set obtained from the coarse search For the part index in the candidate set;

[0095] This formula calculates the Euclidean distance between two fine-grained matching vectors in high-dimensional space. The smaller the L2 distance value, the closer the fine-grained feature vectors of the two parts are, meaning the more similar their geometric shapes and topological structures. Finally, the final result is determined according to... Sort the parts from smallest to largest and output the Top-K=10 similar non-standard parts.

[0096] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A two-stage real-time retrieval method for similar non-standard parts, characterized in that: Includes the following steps; S1: Obtain the STEP model library of historical non-standard parts, preprocess the STEP models in the historical non-standard parts library, and obtain the normalized model; S2: Perform point cloud sampling on the normalized model to obtain a fixed number of three-dimensional point clouds; S3: Based on the sampling points, the global shape descriptor of each historical non-standard part is calculated using ESF, and the global shape descriptor of ESF is sequentially subjected to PCA dimensionality reduction and binary quantization encoding to obtain a compressed ESF vector; at the same time, the fine matching vector of each historical non-standard part is calculated, and the fine matching vector is a vector formed by standardizing and fusing the main inertial feature vector and the three-dimensional Zernike moment feature vector according to the preset weight coefficient α. S4: Construct a coarse search approximate nearest neighbor index library based on the compressed ESF vector, and associate and store the fine matching vector of each historical non-standard part with the corresponding part number; S5: For the STEP model of the non-standard parts to be retrieved, perform the same preprocessing as step S1, the same point cloud sampling as step S2, and the same feature calculation as step S3 in sequence to obtain the query compressed ESF vector and the query fine matching vector. S6: Use the query compressed feature vector to perform a coarse search in the approximate nearest neighbor index to obtain a set of candidate parts; S7: Use the query fine matching vector to perform fine distance calculation and sorting on each candidate part in the candidate part set, and output the sorting results of the top K similar non-standard parts, where K is a positive integer.

2. The two-stage real-time retrieval method for similar non-standard parts according to claim 1, characterized in that: The preprocessing in step S1 includes geometric analysis, coordinate alignment, scale normalization, and pose normalization, specifically: S11: Analyze the three-dimensional geometric boundary of the STEP model and convert it into a geometric representation; S12: Calculate the model's bounding box, centroid, or point cloud center, and translate the model to a unified coordinate origin; S13: Scale normalization is performed based on the longest side of the bounding box, the diagonal length of the model, or the maximum radius of the point cloud, so that models of different sizes can enter a unified scale space. S14: Obtain the three principal directions of the model through principal component analysis or principal axis inertia analysis, and rotate and align them according to the preset principal axis order; S15: Combine the bounding box direction, point cloud distribution offset, or fixed direction rules to determine the positive and negative directions of the principal axis, and avoid feature inconsistencies caused by flipping the principal axis direction sign.

3. The two-stage real-time retrieval method for similar non-standard parts according to claim 2, characterized in that: The point cloud sampling in step S2 includes: Candidate points are generated on the normalized STEP model surface. A fixed number of point clouds is used, which is 4096 points.

4. The two-stage real-time retrieval method for similar non-standard parts according to claim 1, characterized in that: In step S2, the point cloud sampling adopts the farthest point sampling method. The farthest point sampling is achieved by iteratively selecting the point farthest from the selected point set, so that the sampling points are evenly distributed on the model surface and the shape contour differences are preserved.

5. The two-stage real-time retrieval method for similar non-standard parts according to claim 2, characterized in that: The point cloud sampling also includes: adjusting the sampling weights based on local curvature, boundary lines, or local geometric changes to increase the sampling probability of holes, slots, bosses, and edge transition regions, so as to improve the ability of subsequent retrieval to express fine-grained structural differences.

6. The two-stage real-time retrieval method for similar non-standard parts according to claim 1, characterized in that: The PCA dimensionality reduction and binary quantization encoding in S3 include: The original ESF features are 640-dimensional. Principal component analysis is performed on the original 640-dimensional ESF features to reduce the dimensionality of the 640-dimensional features to a 512-dimensional PCA projection matrix. The original ESF vector is then projected onto the low-dimensional space determined by the reduced projection matrix. The binary quantization encoding uses an iterative quantization method to learn the orthogonal rotation matrix, converting the dimensionality-reduced 512-dimensional real-valued vector into a compact binary code, resulting in a compressed ESF vector.

7. The two-stage real-time retrieval method for similar non-standard parts according to claim 1, characterized in that: The calculation of the fine-matching vector in step S3 includes: Calculate the principal inertial feature vector v of the sampled point cloud. prin The principal inertial eigenvector is composed of three eigenvalues ​​λ1, λ2, and λ3 of the covariance matrix or the inertial matrix. Calculate the three-dimensional Zernike moment eigenvector v of the sampled point cloud. zer ; The principal inertial feature vector and the three-dimensional Zernike moment feature vector are standardized and fused according to a preset weight coefficient α to obtain the fine matching vector. ,in and These are the standardized principal inertial eigenvectors and the three-dimensional Zernike moment eigenvectors, respectively. This indicates a vector connection.

8. The two-stage real-time retrieval method for similar non-standard parts according to claim 1, characterized in that: The construction parameters of the approximate nearest neighbor index library in step S4 include: vector dimension d. vec =16, Number of trees N tree =50, Number of retrieval nodes N node =6.

Images