RPS grouping prediction method and device based on deep representation and probability modeling

Abstract

This invention discloses an RPS grouping prediction method and apparatus based on deep representation and probabilistic modeling. The method includes: preprocessing an input point set to obtain a first embedding representation for grouping prediction; using a Gaussian mixture model (Gaussian Mixture Model), matching each embedding point in the first embedding representation with the Gaussian components corresponding to each unit in the Gaussian Mixture Model, and determining the predicted unit to which each embedding point belongs based on the matching results; mapping each embedding point to its corresponding supergroup based on its predicted unit to form a preliminary grouping result; constructing a graph structure based on the preliminary grouping result, and using an attribute prediction model to perform neighborhood aggregation on each node in the graph structure to obtain the attribute prediction result for the embedding point corresponding to each node. This invention solves the technical problem of inconsistencies in grouping results caused by the high degree of manual reliance in the RPS point grouping process in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle design, and more specifically, to an RPS grouping prediction method and apparatus based on deep representation and probabilistic modeling. Background Technology

[0002] RPS (Reference Point System) is a professional system commonly used in fields such as automotive design. Its core function is to achieve standardized management and efficient collaboration in specific stages through standardized rules, data, or processes.

[0003] In existing technologies, the grouping of RPS points is typically done manually. Specifically, during the part design phase, engineers generally select several key feature points as RPS points based on the part's geometric features, assembly relationships, and process requirements. These points are then grouped according to the "3-2-1" positioning principle to constrain the part's six degrees of freedom. This grouping process relies heavily on the engineers' professional experience and manual judgment, such as determining which points serve as primary positioning points and which as auxiliary constraint points.

[0004] During the inspection process, surveyors typically need to manually input and confirm the RPS grouping points in a coordinate measuring machine or inspection fixture based on design drawings or process documents to establish the inspection coordinate system. When the inspection results differ from the design model, engineers must manually analyze the reasons and adjust the grouping scheme if necessary.

[0005] It is evident that existing technical solutions rely heavily on manual labor in the RPS point grouping process. This not only results in cumbersome and inefficient procedures, but also makes it easy for grouping results to be inconsistent due to differences in the experience of different personnel, thereby affecting the accuracy and stability of part positioning and inspection.

[0006] There is currently no effective solution to the above problems. Summary of the Invention

[0007] This invention provides an RPS grouping prediction method and apparatus based on deep representation and probabilistic modeling, which at least solves the technical problem that the existing technology relies heavily on manual intervention in the RPS point grouping process, resulting in inconsistent grouping results.

[0008] According to one aspect of the present invention, an RPS grouping prediction method based on deep representation and probabilistic modeling is provided, comprising: preprocessing an input point set to obtain a first embedding representation for grouping prediction; using a Gaussian mixture model, matching each embedding point in the first embedding representation with the Gaussian components corresponding to each unit in the Gaussian mixture model, and determining the predicted unit to which each embedding point belongs based on the matching result; mapping each embedding point to a corresponding supergroup based on the predicted unit to which each embedding point belongs, forming a preliminary grouping result reflecting the global structural distribution; constructing a graph structure based on the preliminary grouping result, taking embedding points belonging to the same supergroup as mutually adjacent nodes, and using an attribute prediction model to perform neighborhood aggregation on each node in the graph structure, to obtain the attribute prediction result of the embedding point corresponding to each node.

[0009] According to another aspect of the present invention, an RPS grouping prediction device based on deep representation and probabilistic modeling is also provided, comprising: a preprocessing module configured to preprocess an input point set to obtain a first embedding representation for grouping prediction; a mapping module configured to use a Gaussian mixture model to match each embedding point in the first embedding representation with the Gaussian components corresponding to each unit in the Gaussian mixture model, and determine the predicted unit to which each embedding point belongs based on the matching result; a grouping module configured to map each embedding point to a corresponding supergroup based on the predicted unit to which each embedding point belongs, forming a preliminary grouping result reflecting the global structural distribution; and a prediction module configured to construct a graph structure based on the preliminary grouping result, treat embedding points belonging to the same supergroup as mutually adjacent nodes, and use an attribute prediction model to perform neighborhood aggregation on each node in the graph structure to obtain the attribute prediction result of the embedding point corresponding to each node.

[0010] In this embodiment of the invention, the input point set is preprocessed to obtain a first embedding representation for group prediction. Using a Gaussian mixture model, each embedding point in the first embedding representation is matched with the Gaussian components corresponding to each unit in the Gaussian mixture model, and the predicted unit to which each embedding point belongs is determined based on the matching results. Based on the predicted unit to which each embedding point belongs, each embedding point is mapped to a corresponding supergroup, forming a preliminary grouping result reflecting the global structural distribution. A graph structure is constructed based on the preliminary grouping result, with embedding points belonging to the same supergroup as mutually adjacent nodes. An attribute prediction model is used to perform neighborhood aggregation on each node in the graph structure to obtain the attribute prediction result for the embedding point corresponding to each node. This solution solves the technical problem of inconsistent grouping results caused by the high degree of manual reliance in the RPS point grouping process in existing technologies. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0012] Figure 1 This is a flowchart of an RPS grouping prediction method based on deep representation and probabilistic modeling according to an embodiment of the present invention;

[0013] Figure 2 This is a flowchart of another RPS grouping prediction method based on deep representation and probabilistic modeling according to an embodiment of the present invention;

[0014] Figure 3 This is a flowchart of another RPS grouping prediction method based on deep representation and probabilistic modeling according to an embodiment of the present invention;

[0015] Figure 4 This is a flowchart of an optional model training method according to an embodiment of the present invention;

[0016] Figure 5 This is a schematic diagram of the structure of an RPS grouping prediction device based on deep representation and probabilistic modeling according to an embodiment of the present invention;

[0017] Figure 6 A schematic diagram of the structure of a computer device suitable for implementing embodiments of the present disclosure is shown. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0019] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0020] According to an embodiment of the present invention, a method embodiment of RPS grouping prediction based on deep representation and probabilistic modeling is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0021] Figure 1 This is an RPS grouping prediction method based on deep representation and probabilistic modeling according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0022] Step S102: Preprocess the input point set to obtain the first embedding representation for group prediction;

[0023] The spatial coordinates of the input point set are standardized; the dimensionality of the standardized data is reduced based on the principal component analysis model established during the training phase to obtain the first embedding representation.

[0024] Step S104: Using a Gaussian mixture model, each embedding point in the first embedding representation is matched with the Gaussian components corresponding to each unit in the Gaussian mixture model, and the predicted unit to which each embedding point belongs is determined based on the matching results.

[0025] For each embedding point in the first embedding representation, the Mahalanobis distance between the embedding point and the Gaussian components corresponding to each unit is calculated using the Gaussian mixture model. Based on the Mahalanobis distance, it is determined whether the embedding point falls into the confidence region of the corresponding unit. The unit in which the embedding point falls into its confidence region is taken as a candidate unit. The candidate units are scored. Based on the scoring results, the unit with the highest score is selected from the candidate units as the predicted unit to which the embedding point belongs.

[0026] Step S106: Based on the predicted attribution unit of each embedding point, map each embedding point to the corresponding supergroup to form a preliminary grouping result that reflects the global structure distribution.

[0027] Based on the spatial proximity and central feature similarity among the predicted attribution units, multiple adjacent predicted attribution units are merged to form a supergroup reflecting the global spatial structure. Each embedding point is mapped to the corresponding supergroup to form a preliminary grouping result reflecting the global structure distribution. When an overlap of the two-dimensional bounding boxes of two supergroups is detected, the embedding points of the overlapping part are reassigned to adjacent supergroups by calculating the nearest distance between the points of the two-dimensional bounding boxes.

[0028] Step S108: Construct a graph structure based on the preliminary grouping results, take the embedding points belonging to the same supergroup as mutually adjacent nodes, and use the attribute prediction model to perform neighborhood aggregation on each node in the graph structure to obtain the attribute prediction results of the embedding points corresponding to each node.

[0029] The attribute prediction model is used to perform neighborhood aggregation and nonlinear transformation on the second embedding representation of each node in the graph structure to obtain the node embedding; based on the node embedding, the attribute category of the embedding point corresponding to each node is predicted by the classification layer of the attribute prediction model to obtain the attribute prediction result.

[0030] In this embodiment, by processing the 3D point set data of a part and utilizing clustering, spatial geometric constraints, and grouping optimization strategies, automated grouping of RPS points is achieved. This method reduces the degree of manual intervention, ensuring the consistency and repeatability of the grouping results; it also improves overall processing efficiency and is suitable for rapid modeling and inspection of large-scale parts; furthermore, this application can automatically adjust the grouping strategy according to changes in the part's geometry, thereby improving the stability and adaptability of the grouping.

[0031] Figure 2 This is another RPS grouping prediction method based on deep representation and probabilistic modeling according to embodiments of this application, such as... Figure 2 As shown, the method includes the following steps:

[0032] Step S202: Preprocess the input point set.

[0033] First, the input point set to be predicted is obtained. Each point in the input point set includes spatial coordinate information and related attribute features. The coordinate information of the input point set is standardized in the same way as in the training phase, and the standardized high-dimensional data is projected into a two-dimensional space using the principal component analysis model established in the training phase to obtain an initial dimensionality reduction representation. Subsequently, a Transformer is used to extract features from the dimensionality reduction point set to obtain a high-dimensional embedding representation (first embedding representation) containing global context information.

[0034] This embodiment ensures that the input features in the prediction stage remain consistent with those in the training stage through a unified preprocessing and dimensionality reduction strategy. At the same time, the Transformer's self-attention mechanism can effectively capture the global dependencies between points, providing a stable and rich deep representation for subsequent probabilistic modeling.

[0035] Step S204, RPS group prediction.

[0036] For each embedding point in the embedding representation, matching is first performed within the Gaussian Mixture Model (GMM) established during training. Specifically, the Mahalanobis distance from the embedding point to each unit is calculated, and it is determined whether the embedding point falls within the confidence region of the corresponding unit. Units falling within their confidence regions are selected as candidate units. For units containing only a single point, the Euclidean distance is used to determine whether the embedding point falls within the single-point neighborhood within the allowed radius. The log-likelihood score is used to score the candidate units, and after score filtering, the unit with the highest score is selected as the predicted unit to which the embedding point belongs. Through this method, each embedding point can be accurately mapped to the unit hierarchy constructed during training.

[0037] Step S206: Perform supergroup mapping and generate preliminary groups.

[0038] After predicting the unit for each embedding point, it is mapped to the corresponding supergroup, forming a preliminary grouping result. Each supergroup contains all the embedding points of spatially adjacent units with similar features, achieving preliminary aggregation of the point set at the global structural level. When an overlap of the 2D bounding boxes of two supergroups is detected, the embedding points of the overlapping part are reassigned to adjacent supergroups by calculating the nearest distance between the points of the 2D bounding boxes.

[0039] Step S208: Perform point attribute prediction using the attribute prediction model.

[0040] After grouping the embedding points, based on the constructed graph structure, an attribute prediction model learned during the training phase, such as the GraphSAGE (Graph Sampling and Aggregation) model, is used to predict the attributes of each embedding point. Specifically, the embedding representations of each embedding point and its neighboring nodes in the graph are input into the GraphSAGE model. After information aggregation and nonlinear transformation, the node embeddings are output, and then a classification layer is used to predict the attribute category of the embedding point. Through this method, the local neighborhood information and the global graph structure of the embedding points can be combined to achieve more accurate attribute discrimination.

[0041] To address the two-dimensional spatial overlap issues present in the initial grouping results, a nearest-distance-based overlap group adjustment strategy is employed for optimization. When overlap between the two-dimensional bounding boxes of two supergroups is detected, some embedding points are reassigned to adjacent groups by calculating the nearest distance between their embedding points, thereby reducing inter-group conflicts. Simultaneously, empty groups are ignored during the splitting process, without being assigned new group numbers, ensuring the validity and stability of the prediction results. This method effectively resolves the issues of local overlap and outlier distribution, thereby improving the accuracy and consistency of grouping.

[0042] After the above steps, each embedding point is assigned to its corresponding supergroup, and the final attribute prediction result is obtained. This result reflects both the local affiliation of the embedding point and preserves the global hierarchical structure information, while also including the attribute category information of the embedding point.

[0043] Figure 3 This is yet another RPS grouping prediction method based on deep representation and probabilistic modeling according to embodiments of this application, such as... Figure 3 As shown, the method includes the following steps:

[0044] Step S302: Obtain the input point set and preprocess it.

[0045] First, the input point set data is acquired, containing the spatial coordinates and related attribute features of each point. To eliminate dimensional differences between features and improve the stability of subsequent modeling, the input data is standardized to ensure that all features are distributed within a uniform scale. Subsequently, principal component analysis is used to reduce the dimensionality of the standardized high-dimensional data to two-dimensional space, preserving the main trends while removing redundant noise information.

[0046] Step S304, cell mapping.

[0047] The input point set is divided into units. When a unit contains multiple points, a Transformer encoder is first used to extract the global context features of the point set, and these features are used as input to a Gaussian Mixture Model (GMM). Subsequently, a differentiable GMM model is used to model the probability distribution of the point set, and parameters such as the mean, covariance, and weights of the mixture components are learned through iterative optimization using EM or gradient descent. Finally, the mean of the principal components of the GMM is used as the center representation of the unit to obtain a stable and robust spatial center estimate.

[0048] For units containing only a single point, since statistical modeling is not possible, that point is directly used as the center representation. This ensures that all units obtain a consistent center representation, effectively avoiding modeling defects in cases of data sparsity or degradation. The training process of each model will be described in detail below and will not be repeated here.

[0049] Step S306, Unit Center Abstraction and Supergroup Aggregation.

[0050] After obtaining the central representations of all units, these central points are used as high-level input for secondary clustering analysis to form supergroups. During the clustering process, a merging strategy based on a distance threshold is adopted to merge spatially adjacent unit centers, thereby achieving hierarchical organization of the overall point set. This hierarchical modeling approach preserves the local distribution characteristics within units while capturing global structural information at the supergroup level.

[0051] Step S308: Perform attribute recognition.

[0052] After grouping the embedding points, an attribute prediction model trained using GraphSAGE is used to predict the attributes of each embedding point based on the constructed graph structure. For example, the embedding features of each embedding point and its neighboring nodes in the graph structure are input into the GraphSAGE model. This network, through multi-layer neighborhood information aggregation and feature fusion operations, can comprehensively consider the relationships between points and local structural features, thereby generating node embedding representations with contextual semantics. Subsequently, a nonlinear transformation layer is used to refine the features of the node embeddings, and the attribute category is predicted through the output classification layer. In this way, more accurate attribute prediction results can be achieved by utilizing neighborhood information and the global graph topology.

[0053] Figure 4 The model training process according to an embodiment of this application is illustrated. For example... Figure 4 As shown, the method includes the following steps:

[0054] Step S402: Input point set and data preprocessing.

[0055] First, the input point set data for training is obtained. Each point in this set contains three-dimensional spatial coordinate information (x, y, z) and corresponding attribute features.

[0056] To ensure consistency in data distribution between the model training and prediction phases, the coordinate information of the point set is standardized. Simultaneously, principal component analysis (PCA) or other dimensionality reduction operations can be performed on the original features to reduce noise interference and improve the stability of the embedded features.

[0057] Step S404, encoder embedding representation.

[0058] This invention utilizes the Transformer to extract deep features from an input point set. Through its self-attention mechanism, the Transformer can automatically capture the global dependencies between points, maintaining spatial structure consistency while outputting a high-dimensional embedding representation for each point. This embedding representation is a latent representation.

[0059] Step S406: Gaussian mixture modeling.

[0060] First, based on the preset number of Gaussian components The Expectation-Maximization (EM) algorithm is used for iterative optimization, and the model parameters are adjusted accordingly. Initialization is performed, where, Indicates mixed weights, Represents the mean vector. Let represent the covariance matrix.

[0061] Next, iterative optimization and parameter learning are performed. The log-likelihood score is calculated using the EM algorithm for iterative training to maximize the data likelihood. Based on the current parameters, the posterior probability of each embedding point x belonging to each Gaussian component is calculated. Update all model parameters based on posterior responsibility. Mean Updated to the weighted center of the corresponding point set, covariance It is updated to a weighted scatter matrix. Wherein, and This will be used for calculating Mahalanobis distance and defining the confidence region.

[0062] Then, model convergence is performed. Convergence is determined when the change in model parameters falls below a preset threshold. At this point, the probability density function is:

[0063]

[0064] It has stably fitted the distribution of the training point set. Each Gaussian component All are solidified into a structurally stable unit, where K is the total number of Gaussian components.

[0065] In some embodiments, when updating the covariance matrix At that time, the regularization constraints of manifold sensing are utilized. This aligns the main direction with the inherent manifold structure of the data, rather than simply relying on sample statistics.

[0066]

[0067] in, For the first The covariance matrix updated by Gaussian components The total number of training samples. For the first The sample belongs to the first The posterior probability of each Gaussian component. For the first The embedding vector of each sample. For the first The mean vector of Gaussian components. This is the regularization coefficient, used to control the strength of the regularization term. Manifold-aware regularization function acting on the covariance matrix.

[0068] Unlike traditional methods, this embodiment utilizes manifold-aware regularization constraints, so Mahalanobis distance is no longer a simple measure of isotropy, but a semantic distance that follows the geometric characteristics of the data.

[0069] Using the above method, each embedding point is mapped to a probabilistic grouping structure, thereby achieving statistical partitioning of the point set under unsupervised conditions. Compared with traditional grouping methods based on geometric distance or manual thresholds, the modeling strategy of this invention can maintain the robustness and accuracy of grouping under complex and noisy distributions.

[0070] Step S408, graph structure attribute prediction.

[0071] Based on the grouping results of the GMM, a graph structure is constructed, where each point is a node, and the adjacency relationships between points are determined according to their group affiliation and spatial proximity. The GraphSAGE model is then introduced to learn attribute classification for this graph structure.

[0072] During training, GraphSAGE fuses node features with those of its neighboring nodes using the neighborhood aggregation function (AGG). Its core update rule is as follows:

[0073]

[0074] in, Let AGG be the embedding representation of the k-th layer node v, where AGG is the neighborhood aggregation function. For the first Layer nodes Embedded representation; For nodes The set of neighboring nodes, For the first Layer nodes Embedded representation, node It is a node neighboring nodes, W (k) σ represents the weights, CONCAT represents the concatenation operation, and σ represents the activation function.

[0075] In traditional methods, grouping and point attribute classification are usually treated as independent steps, which can easily lead to error accumulation. This invention transforms the grouping results of a Geometric Matrix (GMM) into a graph structure and combines this with the GraphSAGE model to classify the attributes of nodes in the graph. GraphSAGE can learn the relationship features between points through neighborhood aggregation, thereby achieving accurate prediction of point attributes (such as support and compaction) based on the grouping results.

[0076] Step S410: Construct the joint loss function.

[0077] During training, this invention employs a multi-task joint loss function to achieve end-to-end optimization. This loss consists of three parts: log-likelihood loss (L... GMM ), used to optimize the probability distribution fitting effect of GMM in the embedding space; attribute classification loss (L GraphSAGE ), used to constrain the accuracy of GraphSAGE in point attribute prediction tasks; within-group consistency loss (L consistenty This is used to maintain the embedding compactness of points within the same group and suppress boundary blurring and loose grouping.

[0078] To avoid ambiguous boundaries and loose internal grouping in the grouping results, this invention introduces an intra-group consistency constraint into the loss function. This constraint measures the embedding similarity or variance of points within the same group and adds it as a penalty term to the total loss function. The form of the loss function is as follows:

[0079]

[0080] in, and Let G be the representation vector of points in the same group in the embedding space, where G represents the number of groups, g represents a specific group, and i and j represent the indices of two distinct embedding points belonging to the same group. The norm is used to represent the group. By employing the above method, the compactness within the group can be guaranteed, making the grouping results more consistent with the requirements of engineering applications.

[0081] Furthermore, this invention jointly models the log-likelihood loss of GMM, the attribute classification loss of GraphSAGE, and the intra-group consistency loss, constructing a unified joint loss function. During training, the parameters of the three models—Transformer, GMM, and GraphSAGE—are updated end-to-end through this joint loss, achieving coordinated optimization of grouping, attribute discrimination, and distribution constraints.

[0082]

[0083] in, and The weight hyperparameter is used to balance the optimization objectives of different tasks. Compared with the traditional step-by-step processing method, this method effectively reduces information fragmentation and error propagation, and significantly improves the stability and prediction accuracy of the overall modeling.

[0084] Step S412, parameter update and model convergence.

[0085] Driven by the joint loss function, the parameters of the Transformer, GMM, and GraphSAGE models are updated sequentially through backpropagation, enabling simultaneous convergence of grouping modeling and attribute prediction. After multiple rounds of iterative training, when the joint loss reaches the convergence condition, the final training results are output, including the Transformer encoder weight parameters, GMM distribution parameters, and GraphSAGE classifier weights, thus forming a complete grouping and attribute prediction model.

[0086] The training method proposed in this invention breaks through the limitations of traditional geometric heuristic grouping by integrating Transformer deep embedding and probabilistic modeling; it eliminates the error accumulation caused by step-by-step training by unifying graph structure attribute prediction and grouping model; and it improves the compactness of intra-group structure and prediction robustness by using intra-group consistency constraints and multi-task joint loss mechanism.

[0087] This application also provides an RPS grouping prediction device based on deep representation and probabilistic modeling, such as Figure 5As shown, the system includes: a preprocessing module 52, configured to preprocess the input point set to obtain a first embedding representation for group prediction; a mapping module 54, configured to use a Gaussian mixture model to match each embedding point in the first embedding representation with the Gaussian components corresponding to each unit in the Gaussian mixture model, and determine the predicted unit to which each embedding point belongs based on the matching result; a grouping module 56, configured to map each embedding point to a corresponding supergroup based on the predicted unit to which each embedding point belongs, forming a preliminary grouping result reflecting the global structural distribution; and a prediction module 58, configured to construct a graph structure based on the preliminary grouping result, treat embedding points belonging to the same supergroup as mutually adjacent nodes, and use an attribute prediction model to perform neighborhood aggregation on each node in the graph structure to obtain the attribute prediction result of the embedding point corresponding to each node.

[0088] It should be noted that the RPS group prediction device based on deep representation and probabilistic modeling provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the RPS group prediction device based on deep representation and probabilistic modeling provided in the above embodiments and the RPS group prediction method embodiments based on deep representation and probabilistic modeling belong to the same concept. The specific implementation process is detailed in the method embodiments and will not be repeated here.

[0089] Figure 6 A schematic diagram of a computer device suitable for implementing embodiments of the present disclosure is shown. It should be noted that... Figure 6 The computer device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0090] like Figure 6 As shown, the computer device includes a central processing unit (CPU) 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage section 1008 into a random access memory (RAM) 1003. The RAM 1003 also stores various programs and data required for system operation. The CPU 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0091] The following components are connected to I / O interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to I / O interface 1005 as needed. A removable medium 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 1010 as needed so that computer programs read from it can be installed into storage section 1008 as needed.

[0092] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An RPS grouping prediction method based on deep representation and probabilistic modeling, characterized in that, include: The input point set is preprocessed to obtain the first embedding representation for group prediction; Using a Gaussian mixture model, each embedding point in the first embedding representation is matched with the Gaussian components corresponding to each unit in the Gaussian mixture model, and the predicted unit to which each embedding point belongs is determined based on the matching results. Based on the predicted affiliation unit of each embedding point, each embedding point is mapped to the corresponding supergroup to form a preliminary grouping result that reflects the global structure distribution; Based on the preliminary grouping results, a graph structure is constructed, and the embedding points belonging to the same supergroup are taken as mutually adjacent nodes. The attribute prediction model is used to perform neighborhood aggregation on each node in the graph structure to obtain the attribute prediction results of the embedding points corresponding to each node. The preprocessing of the input point set to obtain a first embedding representation for group prediction includes: standardizing the spatial coordinates of the input point set; and reducing the dimensionality of the standardized data based on the principal component analysis model established during the training phase to obtain the first embedding representation. The input point set is the three-dimensional point set data of the part; the attributes of the embedded point include at least one of the following: support and compression.

2. The method according to claim 1, characterized in that, Using a Gaussian mixture model, each embedding point in the first embedding representation is matched with the Gaussian components corresponding to each unit in the Gaussian mixture model, and the predicted unit to which each embedding point belongs is determined based on the matching results, including: For each embedding point in the first embedding representation, the Mahalanobis distance between the embedding point and the Gaussian components corresponding to each unit is calculated using the Gaussian mixture model, and the Mahalanobis distance is used to determine whether the embedding point falls into the confidence region of each unit. The cells whose embedding point falls within their confidence region are taken as candidate cells. The candidate cells are scored and calculated. Based on the scoring results, the cell with the highest score is selected from the candidate cells as the predicted belonging cell of the embedding point.

3. The method according to claim 1, characterized in that, Based on the predicted affiliation unit of each embedding point, each embedding point is mapped to its corresponding supergroup, forming a preliminary grouping result reflecting the global structural distribution, including: Based on the spatial proximity and central feature similarity between the predicted attribution units of each embedding point, multiple mutually adjacent predicted attribution units are merged to form the supergroup that reflects the global spatial structure. Each embedding point is mapped to its corresponding supergroup to form a preliminary grouping result that reflects the global structure distribution.

4. The method according to claim 3, characterized in that, After forming preliminary grouping results that reflect the global structure distribution, the method further includes: when an overlap of two-dimensional bounding boxes of two supergroups is detected, the embedding points of the overlapping part are reassigned to adjacent supergroups by calculating the nearest distance between the points of the two-dimensional bounding boxes.

5. The method according to claim 1, characterized in that, The attribute prediction model is used to perform neighborhood aggregation on each node in the graph structure to obtain the attribute prediction results of the embedding points corresponding to each node, including: The attribute prediction model is used to perform neighborhood aggregation and nonlinear transformation on the second embedding representation of each node in the graph structure to obtain the node embedding. Based on the node embedding, the attribute prediction result is obtained by predicting the attribute category of the embedding point corresponding to each node through the classification layer of the attribute prediction model.

6. An RPS grouping prediction device based on deep representation and probabilistic modeling, characterized in that, include: The preprocessing module is configured to preprocess the input point set to obtain a first embedding representation for group prediction; The mapping module is configured to use a Gaussian mixture model to match each embedding point in the first embedding representation with the Gaussian components corresponding to each unit in the Gaussian mixture model, and determine the predicted unit to which each embedding point belongs based on the matching result. The grouping module is configured to map each embedding point to a corresponding supergroup based on the prediction attribution unit of each embedding point, forming a preliminary grouping result that reflects the global structural distribution. The prediction module is configured to construct a graph structure based on the preliminary grouping results, take the embedding points belonging to the same supergroup as mutually adjacent nodes, and use the attribute prediction model to perform neighborhood aggregation on each node in the graph structure to obtain the attribute prediction results of the embedding points corresponding to each node. The preprocessing module is further configured to: standardize the spatial coordinates of the input point set; and reduce the dimensionality of the standardized data based on the principal component analysis model established during the training phase to obtain the first embedding representation. The input point set is the three-dimensional point set data of the part; the attributes of the embedded point include at least one of the following: support and compression.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the method according to any one of claims 1 to 5.

8. A computer device, characterized in that, include: Memory and processor The memory stores computer programs; The processor is configured to execute a computer program stored in the memory, wherein the computer program, when executed, causes the processor to perform the method described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images