Assembly precision prediction method based on PointNet + + improved network

Abstract

The invention discloses an assembly precision prediction method based on a PointNet + + improved network, and belongs to the field of computer vision and industrial manufacturing. The problems that in the prior art, obtaining of the assembling parallelism is affected by the measuring environment, and it cannot be guaranteed that the assembling parallelism can be directly measured are solved. According to the method provided by the invention, when assembling is not carried out, the point cloud data of the surface is collected and is used as the input of a PointNet + +-based improved network; the class probability of the assembly parallelism is output through the trained PointNet + + improved network; and obtaining an assembly precision category. According to the technical scheme, the robustness of the measurement process and the precision and efficiency of the measurement result can be improved, and accurate and efficient three-dimensional measurement is achieved.

Description

Technical Field

[0001] The present invention relates to an assembly accuracy prediction method based on an improved PointNet++ network, and belongs to the field of computer vision and industrial manufacturing. Background Art

[0002] In precision assembly fields such as machinery manufacturing, the automotive industry, and aerospace, assembly parallelism is a crucial indicator for measuring component assembly accuracy, directly impacting the final product quality and assembly success rate. However, due to the small size, high rigidity, complex structure, high machining precision, single-piece, small-batch production, and irreversible assembly issues of parts, how to effectively predict assembly parallelism and ensure the final product accuracy and assembly success rate is a pressing issue in the precision assembly field.

[0003] Deep learning technology has broken through the limitations of traditional assembly parallelism prediction algorithms, which have difficulty balancing prediction efficiency and accuracy. Deep learning combines color coding, complex network structures, and complex input and output strategies to achieve end-to-end parallelism prediction. Its application has improved the unresolved problem of obtaining parallelism values ​​in complex environments where parallelism cannot be measured directly. However, in the research of traditional assembly parallelism prediction algorithms, experienced engineers are required to reconstruct virtual assembly models of different surfaces to complete assembly predictions, so low efficiency is inevitable. Deep learning strategies in this field are based on regression predictions, and the direct mapping relationship between input surface point cloud information and parallelism values ​​is difficult. It is also difficult to guarantee the final measurement accuracy requirements when performing assembly predictions. Summary of the Invention

[0004] In response to the problems existing in the existing technology, this application proposes an assembly accuracy prediction method based on the improved PointNet++ network, which strengthens the feature extraction capability of the deep learning network for point cloud data, strengthens the extraction of global and local features, and improves the parallelism prediction accuracy.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is an assembly accuracy prediction method based on an improved PointNet++ network, characterized in that it includes the following steps: S1) Build a physical measurement system; prepare a data set in the physical measurement system and build a neural network model; train the model using the data set prepared by the physical measurement system and classification principles; S2) In the physical measurement system, the parallelism category accuracy is divided into units of 0.01 mm to obtain the assembly parallelism category, which is used as the label of the data set and input into the neural network; S3) In the first set abstraction stage of feature extraction, convolution is improved by introducing gated linear units to extract finer features. Chebyshev polynomials are then introduced into the convolution operation to extract features at multiple scales simultaneously, improving efficiency while enriching features. S4) In the second set abstraction stage of feature extraction, an attention mechanism layer is introduced to perform convolution and attention operations on multiple scales simultaneously to capture information at different scales and enhance feature expression capabilities. S5) The output feature information is input into the third set abstraction stage for convolution and pooling operations to generate feature vectors. The parallel class probability output is completed through full connection to achieve assembly accuracy prediction.

[0006] The above-mentioned assembly accuracy prediction method based on the improved PointNet++ network is optimized. In step S1), the gated linear unit and Chevshcherby polynomial are combined with ordinary convolution and introduced into the PointNet++ network to replace the standard convolution module to form EfficientViM_CGLU and KACNConv2DLayer. In addition, in order to strengthen the relationship between the global and local, the self-attention mechanism LEGM layer is introduced to construct a more accurate and efficient network. The network has a high feature extraction capability for point cloud features and has a high prediction accuracy for assembly parallelism.

[0007] The optimized assembly accuracy prediction method based on the improved PointNet++ network, in step S2), uses a three-dimensional coordinate measuring machine to obtain point cloud data of the assembly surface and feature surface of the part, and outputs the parallelism value. Then, the deep learning strategy uses classification-based prediction to classify the parallelism values ​​into T1-T10 categories based on the national standard principle, with an interval of 0.01mm. This method has a tolerance rate of 0.01mm in assembly prediction accuracy, further improving the accuracy of assembly parallelism category prediction. In step S3), in the first set abstraction stage of feature extraction, in the first stage of feature extraction, a small radius parameter is set to group the features, and then an EfficientViM_CGLU block (convolution + gated linear unit) is added to the smaller radius scale to dynamically adjust the activation function of the feature channel; Then, KACNConv2DLayer layer (convolution + Chev-Schneider polynomial) is used simultaneously at the second and third radius scales to enhance feature expression; The features of the three scales are spliced ​​together to obtain more features.

[0008] Optimized,the assembly accuracy prediction method based on the improved PointNet++ network,in step S4), in the second feature extraction stage, the convolution operation is combined with the,attention mechanism, and the LEGM layer is introduced; On this basis, the WATT module is added to the attention layer to scale the feature map; Convolution operations can effectively extract local features, while the attention mechanism can highlight important features and suppress unimportant ones. The WATT module learns the importance weights of features and adaptively scales the feature map, further enhancing the expressiveness of global and local features.

[0009] This design enables the model to focus on key features more precisely, thereby improving the quality of feature representation.

[0010] Optimized, the above assembly accuracy prediction method based on the improved PointNet++ network, The first set abstraction layer in feature extraction is the feature extraction stage of PointNet++; First, the point cloud is downsampled by sampling the farthest point to obtain a new point set; Then, the ball query algorithm is used to query the points in the neighborhood of each new point to form a local point cloud; The convolution operation is performed on the first set of features with a smaller radius, and the activation degree of the feature channel is dynamically adjusted through the EfficientViM_CGLU block, which is expressed as ; F group1 Represents the input feature matrix, which represents the local point cloud feature geometry obtained by the ball query algorithm. These features include the coordinate information of the local point and its corresponding feature vector, and each row in the matrix corresponds to the feature of a point; Conv Represents a standard convolution operation, which is used to extract higher-level features from input features. Setting parameters such as the size of the convolution kernel and the sliding step size will affect the dimension of the output features. EfficientViM_CGLU Represents an efficient feature channel activation adjustment module, and finally outputs the feature matrix Fi .

[0011] For the second set of local features with a larger radius F group2 and the third group of local features with the largest radius F group3 Apply KACNConv2DLayer layers respectively, expressed as F group2Represents the input feature matrix, which represents the local point cloud feature geometry obtained within the second set of radii by the ball query algorithm. These features include the coordinate information of the local points and their corresponding feature vectors; KACNConv2DLayer It is a two-dimensional convolutional layer that extracts higher-level features from input features and then outputs a feature matrix F 2 ; F group3 is the input feature matrix, which represents the local point cloud feature set obtained within the third group radius by the ball query algorithm. F 3 Represents the output feature matrix, which is the feature matrix after processing by the KACNConv2DLayer layer; The features at the three scales are spliced ​​together to obtain comprehensive features: F 1 is the first set of feature matrices, representing the local point cloud features within a smaller radius; F 2 is the second set of feature matrices, representing the local point cloud features within the medium radius; F 3 is the third set of feature matrices, representing the local point cloud features within a larger radius; Final Features Contains rich feature information at different scales (local ranges at three different radii); In the second feature extraction stage, the convolution operation is combined with the attention mechanism, the LEGM layer is introduced, and the WATT module is added to the attention layer; The input feature map is convolved to extract local features, and the convolved features are input into the LEGM layer for attention mechanism processing to highlight important features and suppress unimportant features. The features output by the LEGM layer are input into the WATT module to learn the importance weights of the features and adaptively scale the feature map: F conv It is the feature map after convolution operation; LEGM The Local Enhancement Global Modulation Layer (Local Enhancement Global Modulation Layer) is an attention mechanism module that combines local and global information. It models the global information of the feature map and then uses the obtained global features to modulate local features, so that the model pays more attention to feature areas that are important to the task objectives.

[0012]

[0013] in, F legm The representative is through LEGM Feature map after layer processing; WATT It stands for Adaptive Feature Weight Learning Module, which is used to learn the importance weights of features and adaptively scale feature maps based on these weights. This module can automatically adjust the importance of different features in the feature map, so that the model pays more attention to the features that contribute most to the current task, while reducing the impact of unimportant features. WATT Output final feature map after module processing F watt , which combines the local features extracted by the convolution operation, LEGM The important feature information enhanced by the layers and the structure of adaptive weight learning contain richer, more accurate and discriminative features, providing high-quality feature representation for subsequent classification or other tasks; The feature map of the previous layer is then fed into the Set Abstraction layer, where convolution is performed directly to extract local features. Global feature extraction is then performed using max pooling on the initial point cloud, which is then concatenated into a high-dimensional vector. The fully connected layer then processes the output class probabilities.

[0014] Optimized, the above assembly accuracy prediction method based on the improved PointNet++ network, The main principle of the EfficientViM_CGLU block is to dynamically adjust the activation level of the channel through convolution operations and gated linear units; when the input feature tensor X , the dimension is ( B , C , H , W ),in B is the batch size, C is the number of channels, H and W is the height and width of the feature map, and 1×1 convolution is used to map the input features to twice the hidden dimension space. (7) in, W is the weight of the convolution kernel, b is the bias term, D is the number of output channels; Then F Divided into two parts M and I , the shape of each part is ( B , D , H , W ): (8) right M Apply 3×3 depth convolution, and depth convolution applies convolution kernel to each channel separately, (9) in, K is the radius of the convolution kernel, and then the GELU activation function is used: (10) in, is the cumulative distribution function of the standard normal distribution; The above formula and I The gating mechanism is formed by combining element-level multiplications; the gated features are mapped back to the original number of channels through 1×1 convolution.

[0015] Optimized, the assembly accuracy prediction method based on the PointNet++ improved network, the KACNConv2DLayer layer combines the convolution operation and the Chebyshev polynomial transformation; when the input feature vector x , use 3×3 convolution to extract local features and record it as L 1 ; The Chebyshev transform maps input features to a higher-dimensional space to capture more complex patterns. A higher-dimensional space is one with a higher dimension than the original data space, i.e., the spatial dimension corresponds to the number of features of the data points. For polynomial order K, the transformation is expressed as: (12) in, T n yes n Chebyshev polynomials of order, α n is the learned weight coefficient; The Chebyshev polynomials are recursively defined as: , (13) Transform the value of each pixel position in the feature map (size is the center point × the neighborhood points around each center point) through this polynomial.

[0016] Optimized, the assembly accuracy prediction method based on the improved PointNet++ network, in the second stage SetAbstraction, after sampling and grouping; by performing convolution operations on each neighborhood, a wider range of feature patterns are captured, and at the same time, the attention mechanism operation is performed to generate QK Splice with V to form a feature map; Divide the feature map window, adjust the feature map size to adapt to the window division, and calculate the attention score of each window using the self-attention mechanism (14) here, Q 、 K are the query and key vectors respectively, V is a value vector, d k is the feature channel dimension, SoftMax The function is used to normalize the attention score into a probability distribution; Get the updated features, and then in the WATT module, get the relative position information by encoding the features relative positions, and weight the relative positions. QK Update the attention score matrix in V Multiply and output feature map (15) (16) Among them, R i,j Represents the relative position information between points. x i and x j Representing the i and j The point cloud position of the point.

[0017] The beneficial effects of this application are: The technical solution of this application applies deep learning technology to construct a parallelism prediction model that deeply integrates physical and data models. This model clearly defines the mapping relationship between new point cloud information of part feature surfaces and assembly surfaces, the deep network model, and assembly parallelism. An improved strategy based on PointNet++ improves the robustness of the prediction process and the accuracy and efficiency of measurement results, achieving accurate and efficient parallelism prediction.

[0018] The technical solution of this application establishes a mapping model between point clouds and parallelism. This proposed strategy achieves assembly parallelism without the need for additional models, requiring dataset preparation. This addresses the current challenges of deep learning technology in dataset preparation and network model construction for precision assembly training.

[0019] The technical solution of this application relies on a physical measurement system to construct a dataset based on measurement data and classification criteria. This dataset is then input into the improved PointNet++. The EfficientViM_CGLU and KACNConv2DLayer modules are used to improve the network's ability to extract point cloud features, while the self-attention mechanism is combined to enhance the connection between local and global features. According to experimental verification, compared with existing deep learning network models, this network model has a stronger learning ability for point cloud information and a higher accuracy rate in predicting assembly parallelism. At the same time, the network also has strong robustness and generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a schematic diagram of the physical measurement system; Figure 2 This is the assembly parallelism prediction process based on the improved PointNet++ network in this application; Figure 3 This is a schematic diagram of the improved PointNet++ network (PAC-Net) model of this application; Figure 4 (a) is a schematic diagram of one-stage feature extraction of the dual-stage improved convolution and attention mechanism integrated with optimization in this application; Figure 4 (b) is a two-stage feature extraction flow chart of the dual-stage improved convolution and attention mechanism integrated with optimization in this application; Figure 5 The figure is a comparison chart of the assembly parallelism prediction results of each model. DETAILED DESCRIPTION

[0021] The technical features of the present invention are further described below with reference to specific embodiments.

[0022] This application proposes an assembly accuracy prediction method based on an improved PointNet++ network. The network model built by integrating multiple convolution methods and self-attention mechanisms realizes the overall process of predicting assembly parallelism. Figure 2 shown.

[0023] The assembly accuracy prediction method based on the improved PointNet++ network of this application first builds a physical measurement system; prepares a data set in the physical measurement system; and classifies the parallelism measured by the measuring machine into ten categories according to standard accuracy classification standards. The data set is prepared and the parallelism value prediction problem is converted into a category prediction, thereby further improving the prediction accuracy and avoiding excessive errors caused by value deviations.

[0024] Specifically, the following steps are included: S1) Build a physical measurement system; prepare a data set in the physical measurement system and build a neural network model; train the model using the data set prepared by the physical measurement system and classification principles; S2) In the physical measurement system, the parallelism classification accuracy is divided to obtain the assembly parallelism category and use it as the label of the data set and input it into the neural network; S3) In the first set abstraction and fusion multi-scale convolutional layer stage of feature extraction, convolution is improved by introducing gated linear units; Chebyshev polynomials are introduced in the convolution operation to extract features at multiple scales simultaneously; S4) In the second set abstraction and multi-scale attention mechanism stage of feature extraction, an attention mechanism layer is introduced to perform convolution and attention mechanism operations on multiple scales simultaneously to capture information at different scales; S5) Input the feature information of the above two stages into the third Set Abstraction stage, perform convolution and pooling operations to generate feature vectors, and complete the parallel category probability output through full connection to achieve assembly accuracy prediction.

[0025] In one embodiment of the present application, the architecture of the network that integrates multiple convolution and attention mechanisms is referenced by PointNet++, such as Figure 3 As shown in the figure, by combining gated linear units and Chevshcherby polynomials with ordinary convolutions into the PointNet++ network, replacing the standard convolutional modules, we construct EfficientViM_CGLU and KACNConv2DLayer. Furthermore, to strengthen the relationship between global and local features, a self-attention mechanism (LEGM) layer is introduced to build a more accurate and efficient network. This network has a high feature extraction capability for point cloud features and a high prediction accuracy for assembly parallelism.

[0026] In one embodiment of the present application, the data set preparation and network training process are as follows.

[0027] The dataset used for PAC-Net network training is prepared using both a physical measurement system and a virtual assembly system. The physical measurement system includes a Hexagon GLOBAL S 05.07.05 coordinate measuring machine and PC-DMIS software. The virtual assembly system, built using MATLAB software, uses surface point cloud data as input and calculates assembly parallelism using corresponding theoretical algorithms. The class accuracy corresponding to the point cloud provides the target ground truth for network training and result evaluation.

[0028] In this embodiment, the data set is divided into 2000, 200 and 200 groups according to training, verification and testing. The 200 groups in the test group are collected and prepared by a physical measurement system. Network training is performed on a Dell Precision 7920 tower workstation based on the Pytorch deep learning framework. The batch size used is 32 and the training is repeated for 200 rounds. The Adam optimizer is used to optimize the network parameters during backpropagation, and the learning rate is initially set to 0.001. This paper comprehensively adopts a gated linear mechanism and Chef Sherby polynomial combined with convolution, combined with a self-attention mechanism layer, and improved according to the original feature extraction stage of PointNet++ to ensure smooth fitting during network training and further improve the accuracy of the network model inference results.

[0029] The principles of the technical solution of this application are described in detail below in conjunction with specific applications.

[0030] The principles of dataset preparation are as follows: Based on 5,184 dense points, 2,000 assembly examples were calculated. The parallelism value and relative position of the surface point cloud of each assembly were obtained using Hexagon's global series three-dimensional coordinate measuring machine. The point cloud of the assembly surface was collected into 36×11 points, and the point cloud of the feature surface was collected into 4×4 point cloud. Figure 1 As shown in the figure, a three-dimensional coordinate measuring machine is used to measure the features of part 1, part 2, and the assembly. The corresponding parallelism values ​​are then directly obtained through PC-DIMS software. Then, the operation is repeated to obtain 2000 sets of assembly cases, which are then divided into ten levels (T1 to T10) with a 0.01mm interval. The 2000 examples are divided into training set, validation set, and test set in a ratio of 7:1:2, and then input into the neural network for prediction. The specific process is as follows Figure 2 shown.

[0031] The principle of the network (PAC-Net) with fusion optimization and improved convolution and attention mechanism is as follows.

[0032] like Figure 3 As shown in the figure, the PAC-Net network is based on the PointNet++ network and combines various convolutional methods and a self-attention mechanism. It targets point cloud classification tasks, aiming to build a network architecture that is efficient, highly accurate, and has strong generalization capabilities. The network's primary input is the 3D coordinates of the point cloud, and its output is the assembly parallelism category. It is internally composed of a shared multi-layer perceptron (S-MLP) layer, a set abstraction and fusion multi-scale convolution (MSG Convolution) layer, a set abstraction and multi-scale attention (MSG Attention) layer, a set abstraction (SetAbstraction) layer, and a fully connected layer. The network takes the 3D information of the surface point cloud as input, extracts features through three stages of SetAbstraction layers, maps them into feature vectors through a maximum pooling layer, and then processes them through a fully connected layer. The output is the probability of the assembly parallelism within its category, which constitutes the classifier's prediction of the parallelism category.

[0033] The specific principle is as follows: a projection relationship is established between the point cloud of the feature surface and the assembly surface and the assembly parallelism. The number of points in the neural network training input point cloud is represented by a three-dimensional array with N elements. Each element Pi Recording 3D point cloud coordinates. Through the shared multi-layer perceptron (S-MLP), the coordinate features are increased from 3 to 64. After the receptive field is expanded, the feature map is rescaled three times, corresponding to the three stages (b), (c), and (d). Number of dimensions N It is reduced to 1 / 16 of the original size, while the coordinate dimension is increased by 16 times.

[0034] In the feature extraction process, three types of set abstraction layers are designed to construct global and local relationships between different points. First, the farthest point sampling is performed on the input point cloud data. Assuming that the point cloud data set is , where each point P i Is a three-dimensional coordinate, first select a point as the first sampling point q 1 , then for each point in the point cloud P i , calculate the distance to the nearest sampling point as d i . Initially, all points to q 1 The distance is formula (17). (17) Select the point farthest from the current sampling point set as the next sampling point ,Right now: (18) Then for each point P i , update its distance to the nearest sampling point to the minimum of the current distance and the new distance: (19) The above operations are continued to effectively retain the main structural features of the point cloud while reducing the amount of data. 512 points are selected in the first feature extraction stage, and then a spherical neighborhood query is performed, defining three query radii r Then calculate the distance from all points to the center point and divide the neighborhood to ensure that there is k Points (here k The number needs to be set in advance).

[0035] As shown in Figure 4(a), in the neighborhood under the first radius, for the input feature X 1 ∈R N×C ,in N is the number of points, C The feature dimension is sequentially subjected to convolution, EfficientViM_CGLU layer, and three-layer convolution operations. Specifically, a linear transformation is first performed, as shown in formula 20.

[0036] (20) in W∈R C×C is the weight matrix, b∈R C is the bias vector, which is then input into the EfficientViM_CGLU layer. The core of this layer is a gating mechanism, as shown in Formula 21. (twenty one) That is, feature extraction at the first scale is achieved through activation functions and feature transformation. Sample features are learned using the KACNConv2DLayer operation in the neighborhoods of the second and third radii. This is achieved by applying a Chebyshev polynomial transformation to the local neighborhood, as shown in Equation 22. (twenty two) in, k Represents the order and the Chebyshev polynomial transformation operates on the features of each point. It does not change the number of points, but changes the value of the feature. The feature dimension of each point remains unchanged, but the feature value is transformed into T k (x) , and then undergoes a convolution operation, as shown in Formula 23. (twenty three) in, W conv is the convolution kernel weight matrix, b It is the bias vector, and the two feature extraction methods are concatenated and input into the second stage as new features. (twenty four) In the feature extraction stage of the second layer, as shown in Figure 4(b), the LEGM and WATT modules are introduced to enhance the attention mechanism. To enhance feature interaction between dimensions, a channel-crossing attention module is introduced. Features are then extracted through a series of convolution operations and aggregated through pooling. These aggregated features are concatenated with the initial input features and then fed into the third layer, where the point cloud data is abstracted into a global feature vector. This vector is fed into a fully connected layer for progressive dimensionality reduction, ultimately outputting the classification result.

[0037] Specifically, after sampling and grouping operations, the feature matrix of the points in the domain of the new sampling points is obtained, and the shape is [B,C+D,nsample,npoint] ,in B represents the batch size, C is the coordinate dimension of the point, D is the dimension of the input feature, nsample represents the number of points in each neighborhood, npoint is the number of sampling points.

[0038] The new sampling points in the local neighborhood are input into the attention mechanism layer, and the input features are mapped to the dimensions suitable for attention mechanism processing through convolution. Then, in the LEGM layer, the input features are normalized, and the feature map is normalized to the range of features that are easy to extract. At the same time, the average and variance values ​​of each column are taken. The normalized features are passed through the attention mechanism module, that is, the feature map is first passed through two linear layers as QK and V Value, composed of two parts QKV Get a new feature map, divide the feature map into multiple fixed-size windows, then rearrange the feature map, divide the feature map into windows, flatten each window into a one-dimensional vector, and apply the WATT module to each window. In the WATT module, input feature QKV Divided into Q 、 K and V , and then calculate the attention score: (25) in C Represents the number of feature channels. Then the initial input features are encoded relative to each other, and the relative position offset is dynamically generated by calculating the distance between the points. The relative position is added to the attention score as a bias and applied to the Softmax layer. V The weighted summation is performed. The output features of each window are merged back to the original feature map size. In addition, the output of the attention mechanism is fused with the output of the convolution operation through the projection layer to complete the extraction of local and global features. The design of this module enables the model to perform fine-grained feature interaction within the local window while maintaining the continuity of global features through the convolution operation.

[0039] After feature extraction, a symmetric layer is used to process the disordered input of the point cloud, so that the network can produce the same prediction regardless of the input order of the points, thereby reducing the cost of dataset construction. The symmetric function is defined as: (26) Formula (26) defines the idealized end-to-end training process e , which can directly convert the input point cloud data into predicted labels. In this process, PAC-Net serves as the backbone of the network, responsible for feature extraction and processing. Through the maximum pooling layer, the feature map is compressed into a 512-dimensional vector. Subsequently, this vector is input into a classifier consisting of three fully connected layers (FC layers), with the number of neurons in these three fully connected layers being 256, 128, and 128, respectively. C , ( C( represents the number of categories). To accelerate training and prevent vanishing gradients and overfitting, normalization and dropout layers are added between the fully connected layers. Finally, the classifier uses a sigmoid activation function and a cross-entropy loss function to calculate the difference between the predicted results and the true labels.

[0040] This paper presents a detailed introduction and analysis based on the PAC-Net network. By introducing attention mechanisms (such as the LEGM and WATT modules) and multiple feature extraction methods (such as the EfficientViM_CGLU layer and KANConv2DLayer), PAC-Net effectively extracts and fuses features at different scales in point cloud data. This design not only enhances the model's ability to perceive local and global features but also improves the accuracy of feature representation by dynamically adjusting feature weights. To verify the effectiveness of PAC-Net, a model comparison was conducted. Table 1 shows the performance comparison of PAC-Net with other models (such as PointNet2 and its variants) on the assembly parallelism prediction task. The results achieved an accuracy of 92.18%, meeting industrial requirements and suitable for application.

[0041] Table 1 Comparison of training results Model Name ACC Precise Recall F1-score PointNet2 (PointNet++) 83.13% 78.2% 82% 80% PointNet2+Conv 75.78% 77.45% 76% 76.54% PointNet2+CABM 84.21% 85.3% 84% 84.64% Ours (PAC-Net) 92.18% 90% 90% 89.7% Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Any changes, modifications, additions or substitutions made by ordinary technicians in this technical field within the essential scope of the present invention should fall within the scope of protection of the present invention.

Claims

1. A method for assembly accuracy prediction based on an improved PointNet++ network, characterized by: The following steps are involved: S1) Build a physical measurement system; prepare a data set in the physical measurement system and build a neural network model; train the model using the data set prepared using the physical measurement system and classification principles; S2) In the physical measurement system, the parallelism classification accuracy is divided to obtain the assembly parallelism category and use it as the label of the data set and input it into the neural network; S3) In the first set abstraction stage of feature extraction, convolution is improved by introducing gated linear units; Chebyshev polynomials are introduced in the convolution operation to extract features at multiple scales simultaneously; S4) In the second set abstraction stage of feature extraction, an attention mechanism layer is introduced to perform convolution and attention mechanism operations on multiple scales simultaneously to capture information at different scales; S5) After concatenating the feature information obtained in the first set abstraction stage of feature extraction and the second set abstraction stage of feature extraction, the convolution operation is performed on the third set abstraction stage to generate a feature vector. The class probability output of the parallelism is completed through full connection to achieve assembly accuracy prediction.

2. The assembly accuracy prediction method based on the improved PointNet++ network according to claim 1 is characterized in that: In step S1), when building the neural network model, the gated linear unit, Chevshcherby polynomial, and ordinary convolution are introduced into the PointNet++ network to replace the standard convolution module to form EfficientViM_CGLU and KACNConv2DLayer. The self-attention mechanism LEGM layer is introduced to build the network.

3. The assembly accuracy prediction method based on the improved PointNet++ network according to claim 1 is characterized in that: In step S2), the point cloud data of the assembly surface and feature surface of the part is obtained, and the parallelism value is output; the deep learning strategy is based on the classification prediction and the parallelism category is divided into T1~T10; In step S3), in the first set abstraction stage of feature extraction, grouping is performed by setting a small radius parameter, and then adding EfficientViM_CGLU blocks to the scales with smaller radius to dynamically adjust the activation function of the feature channel; KACNConv2DLayer layers are used simultaneously at the second and third radius scales to enhance feature expression; The features of three radius scales are spliced ​​together to obtain more features.

4. The assembly accuracy prediction method based on the improved PointNet++ network according to claim 1 is characterized in that: In step S4), in the second set abstraction stage of feature extraction, the convolution operation is combined with the attention mechanism to introduce the LEGM layer; On this basis, the WATT module is added to the attention layer to scale the feature map; The convolution operation extracts local features, and the attention mechanism highlights important features and suppresses unimportant features; The WATT module adaptively scales the feature map by learning the importance weights of the features, thereby enhancing the expressive power of global and local features.

5. The assembly accuracy prediction method based on the improved PointNet++ network according to claim 2 is characterized in that: In the first set abstraction stage of feature extraction, first, the point cloud is downsampled by farthest point sampling to obtain a new point set; then, the ball query algorithm is used to query the points in the neighborhood of each new point to form a local point cloud; The convolution operation is performed on the first set of features with a smaller radius, and the activation degree of the feature channel is dynamically adjusted through the EfficientViM_CGLU block, which is expressed as ; F group1 Represents the input feature matrix, which represents the local point cloud feature geometry obtained by the ball query algorithm. These features include the coordinate information of the local point and its corresponding feature vector, and each row in the matrix corresponds to the feature of a point; Conv Represents a standard convolution operation, which is used to extract higher-level features from input features. Setting parameters such as the size of the convolution kernel and the sliding step size will affect the dimension of the output features. EfficientViM_CGLU Represents an efficient feature channel activation adjustment module, and finally outputs the feature matrix Fi ; For the second set of local features with a larger radius F group2 and the third group of local features with the largest radius F group3 Apply KACNConv2DLayer layers respectively, expressed as F group2 Represents the input feature matrix, which represents the local point cloud feature geometry obtained within the second set of radii by the ball query algorithm. These features include the coordinate information of the local points and their corresponding feature vectors; KACNConv2DLayer It is a two-dimensional convolutional layer used to extract higher-level features from input features; F 2 represents the output feature matrix; F group3 is the input feature matrix, which represents the local point cloud feature set obtained within the third group radius through the ball query algorithm; F 3 Represents the output feature matrix, which is KACNConv2DLayer Feature matrix after layer processing; The features at the three scales are spliced ​​together to obtain comprehensive features: F 1 is the first set of feature matrices, representing the local point cloud features within a smaller radius; < i>F 2 is the second set of feature matrices, representing the local point cloud features within the medium radius; < i>F 3 is the third set of feature matrices, representing the local point cloud features within a larger radius; final Features Contains feature information at different scales; In the second feature extraction stage, the convolution operation is combined with the attention mechanism, the LEGM layer is introduced, and the WATT module is added to the attention layer; The input feature map is convolved to extract local features, and the convolved features are input into the LEGM layer for attention mechanism processing to highlight important features and suppress unimportant features. The features output by the LEGM layer are input into the WATT module to learn the importance weights of the features and adaptively scale the feature map: F conv is the feature map after convolution operation, LEGM The global modulation layer is a local enhancement layer that combines local and global information. It models the global information of the feature map and then uses the obtained global features to modulate the local features. in, F legm Represents the feature map after processing by the LEGM layer based on the convolutional feature map; WATT Represents the adaptive feature weight learning module, which is used to learn the importance weights of features and adaptively scale the feature maps according to these weights. WATT Output final feature map after module processing F watt , which combines the local features extracted by the convolution operation, LEGM Important feature information of layer enhancement and structure of adaptive weight learning; The feature map of the previous layer is input into the Set Abstraction layer, where convolution is directly performed to extract local features. The initial point cloud is then max-pooled for global feature extraction and then concatenated into a high-dimensional vector. The output category probability is then processed through a fully connected layer.

6. The assembly accuracy prediction method based on the improved PointNet++ network according to claim 5 is characterized in that: The EfficientViM_CGLU block dynamically adjusts the activation level of the channel through convolution operations and gated linear units; When the input feature tensor X has the dimension ( B , C , H , W ),in B is the batch size, C is the number of channels, H and W is the height and width of the feature map, and 1×1 convolution is used to map the input features to twice the hidden dimension space. in, W is the weight of the convolution kernel, b is the bias term, D is the number of output channels; Then F Divided into two parts M and I , the shape of each part is ( B , D , H , W ): right M Apply 3×3 depth convolution, and depth convolution applies convolution kernel to each channel separately, in, K is the radius of the convolution kernel, and then the GELU activation function is used: in, is the cumulative distribution function of the standard normal distribution; Will M activated and I The gating mechanism is formed by combining element-level multiplications; the gated features are mapped back to the original number of channels through 1×1 convolution.

7. The assembly accuracy prediction method based on the improved PointNet++ network according to claim 2 is characterized in that: The KACNConv2DLayer layer combines convolution operations and Chebyshev polynomial transformations; when the input feature vector x is used, a 3×3 convolution is used to extract local features, which is recorded as L 1 ; Perform Chebyshev transformation to map the input feature dimensions to a higher-dimensional space to capture more complex patterns. A higher-dimensional space refers to a space with a higher dimension than the original data space, that is, the spatial dimension corresponds to the number of features of the data point. For polynomial order K, the transformation is expressed as: ; in, T n yes n Chebyshev polynomials of order, α n is the learned weight coefficient; The Chebyshev polynomials are recursively defined as: , Transform the value of each pixel position in the feature map through this polynomial.

8. The assembly accuracy prediction method based on the improved PointNet++ network according to claim 2 is characterized in that: In the second stage of set abstraction, after sampling and grouping; By performing convolution operations on each neighborhood, a wider range of feature patterns are captured, and at the same time, attention mechanisms are performed to generate QK Splice with V to form a feature map; Divide the feature map window, adjust the feature map size to adapt to the window division, and calculate the attention score of each window using the self-attention mechanism here, Q 、 K are the query and key vectors respectively, V is a value vector, d k is the feature channel dimension, SoftMax The function is used to normalize the attention score into a probability distribution; Get the updated features, then in the WATT module, get the relative position information by encoding the features relative positions, weight the relative positions into QK to update the attention score matrix, and multiply it with V to output the feature map in, R i,j Represents the relative position information between points. x i and x j Representing the i and j The point cloud position of the point.

Images