An assembly precision prediction method based on a PointNet++ improved network

Abstract

The application discloses an assembly precision prediction method based on a PointNet++ improved network, and belongs to the fields of computer vision and industrial manufacturing. In the prior art, the acquisition of assembly parallelism is affected by the measurement environment, and it cannot be ensured that direct measurement can be performed. In the method, point cloud data of a surface is collected when the assembly is not performed, and is used as input of the PointNet++ improved network. The trained PointNet++ improved network outputs category probability of the assembly parallelism. The assembly precision category is obtained. The technical scheme can improve the robustness of the measurement process and the accuracy and efficiency of the measurement result, and realizes accurate and efficient three-dimensional measurement.

Description

TECHNICAL FIELD

[0001] The application relates to an assembly precision prediction method based on a PointNet++ improved network and belongs to the fields of computer vision and industrial manufacturing. BACKGROUND

[0002] In the field of precision assembly, such as mechanical manufacturing, automobile industry and aerospace, assembly parallelism is an important index for measuring the assembly precision of parts, and is directly related to the final quality of products and the assembly success rate. However, due to the small size, large rigidity, complex structure, high machining precision, small batch production of single parts and the emergence of irreversible assembly problems, how to effectively predict the assembly parallelism to ensure the final precision of products and the assembly success rate is a problem to be solved in the field of precision assembly.

[0003] The deep learning technology breaks through the limitation that the traditional assembly parallelism prediction algorithm cannot simultaneously consider the prediction efficiency and precision. Deep learning combines color coding, complex network structure and complex input and output strategy to realize end-to-end parallelism prediction, and its application has improved the unsolved problem of obtaining parallelism value in the scene where the parallelism cannot be directly measured in a complex environment. However, in the research on the traditional assembly parallelism prediction algorithm, an experienced engineer is required to reconstruct a virtual assembly model of different surfaces to complete the assembly prediction, so that the low efficiency is inevitable. The deep learning strategy in the field is based on regression prediction, and it is difficult to directly map the input surface point cloud information and the parallelism value, so it is difficult to ensure the final measurement precision requirement when performing assembly prediction. SUMMARY

[0004] In view of the problems in the prior art, the application provides an assembly precision prediction method based on a PointNet++ improved 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 precision.

[0005] To solve the above technical problems, the technical scheme adopted by the application is an assembly precision prediction method based on a PointNet++ improved network, characterized in that it comprises the following steps:

[0006] S1) Build a physical measurement system; prepare a data set in the physical measurement system, and build a neural network model; train the model through the data set prepared by the physical measurement system and the classification principle;

[0007] S2) In the physical measurement system, the parallelism category precision is divided in units of 0.01 mm, the assembly parallelism category is obtained as the label of the data set, and is input into the neural network;

[0008] S3) In the first Set Abstraction stage of feature extraction, the convolution is improved by introducing a Gated Linear Unit to extract finer features, and then introducing Chebyshev polynomials in the convolution operation, while extracting features at multiple scales, improving efficiency while enriching features;

[0009] 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 and enhance feature expression ability.

[0010] S5) The feature information output above is input into the third Set Abstraction stage, a convolutional pooling operation is performed to generate a feature vector, and a parallel degree class probability output is completed through full connection to realize assembly accuracy prediction.

[0011] The improved assembly accuracy prediction method based on the PointNet++ improved network, in step S1), introduces the Gated Linear Unit and Chebyshev polynomial into the PointNet++ network by combining them with ordinary convolution, replacing the standard convolution module, to form EfficientViM_CGLU and KACNConv2DLayer; In addition, to strengthen the relationship between global and local, the self-attention mechanism LEGM layer is introduced to build a network with higher accuracy and efficiency, which has high feature extraction capability for point cloud features and high prediction accuracy for assembly parallelism.

[0012] The improved assembly accuracy prediction method based on the PointNet++ improved network, in step S2), obtains the point cloud data of the assembly surface and the feature surface of the part by a three-coordinate measuring machine, and outputs the parallelism value; Then the deep learning strategy based on classification prediction classifies the parallelism value according to the national standard principle, with an interval of 0.01mm, and divides the parallelism class into T1-T10. This method has a tolerance of 0.01mm in assembly prediction accuracy, further improving the accuracy of assembly parallelism class prediction.

[0013] In step S3), in the first Set Abstraction stage of feature extraction, in the first stage of feature extraction, small radius parameters are set for grouping, and EfficientViM_CGLU blocks (convolution + Gated Linear Unit) are added to dynamically adjust the activation function of the feature channel.

[0014] Then the KACNConv2DLayer (Convolution + Chebyshev Polynomial) is used at the second and third radius scales to enhance feature representation.

[0015] The features of the three scales are concatenated to obtain more features.

[0016] In the second feature extraction stage, the LEGM layer is introduced in step S4) to optimize the above-mentioned assembly accuracy prediction method based on the improved PointNet++ network, which combines convolution operation with attention mechanism.

[0017] On this basis, the WATT module is added to the attention layer to scale the feature map.

[0018] Convolution operation can effectively extract local features, while attention mechanism can highlight important features and suppress unimportant features. The WATT module learns the importance weight of the features and adaptively scales the feature map, further enhancing the expression ability of global and local features.

[0019] This design enables the model to pay more attention to key features, thereby improving the quality of feature representation.

[0020] The improved assembly accuracy prediction method based on the improved PointNet++ network,

[0021] The first Set Abstraction layer in the feature extraction stage is the feature extraction stage of PointNet++.

[0022] First, the farthest point sampling is used to downsample the point cloud to obtain a new point set.

[0023] Then, the ball query algorithm is used to query the points in the neighborhood of each new point to form a local point cloud.

[0024] The first group of features with smaller radius is subjected to convolution operation, and the activation degree of the feature channel is dynamically adjusted through the EfficientViM_CGLU block, expressed as . F group1 represents the input feature matrix, which represents the local point cloud features obtained by the ball query algorithm. These features include the coordinate information of the local points and their corresponding feature vectors, and each row in the matrix corresponds to the features of a point.

[0025] Conv represents a standard convolution operation, which is used to extract higher-level features from the input features. By setting the size and sliding step of the convolution kernel, the dimension of the output features will be affected.

[0026] EfficientViM_CGLU represent a high-efficiency feature channel activation adjustment module, and finally output a feature matrix Fi .

[0027] the second group of local features with a larger radius F group2 and the third group of local features with the largest radius F group3 respectively apply KACNConv2DLayer layers, respectively expressed as F group2 represent an input feature matrix, indicating the local point cloud feature geometry obtained by the ball query algorithm within the second group of radii, which includes the coordinate information of the local points and their corresponding feature vectors;

[0028] KACNConv2DLayer is a two-dimensional convolution layer used to extract higher-level features from the input features, and then output a feature matrix F 2 ;

[0029] F group3 is an input feature matrix, indicating a set of local point cloud features obtained by the ball query algorithm within the third group of radii. Then F 3 represent an output feature matrix, which is the feature matrix processed by the KACNConv2DLayer layer;

[0030] The features at three scales are spliced to obtain comprehensive features: F 1 is the first group of feature matrices, indicating local point cloud features within a small radius;

[0031] F 2 is the second group of feature matrices, indicating local point cloud features within a medium radius;

[0032] F 3 is the third group of feature matrices, indicating local point cloud features within a large radius;

[0033] the final feature contains rich feature information at different scales (within the local range of three different radii);

[0034] 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;

[0035] The input feature map is subjected to a convolution operation to extract local features, and the features after the convolution are input into the LEGM layer for attention mechanism processing to highlight important features and suppress unimportant features, and the features output by the LEGM layer are input into the WATT module to learn the importance weights of the features and perform adaptive scaling on the feature map: F conv The feature map after the convolution operation; LEGM The Local Enhancement Global Modulation Layer is a kind of 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 the local features, so that the model pays more attention to the feature regions that are important to the task target.

[0036]

[0037] wherein, F legm represents the feature map after processing by the LEGM layer; WATT The adaptive feature weight learning module is used to learn the importance weights of the features and to perform adaptive scaling on the feature map according to 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 make important contributions to the current task, while reducing the influence of unimportant features. The output feature map WATT after processing by the F module is the final feature map LEGM watt , which integrates the local features extracted by the convolution operation, the important feature information enhanced by the

[0038] layer, and the structure of adaptive weight learning, contains more rich, accurate and discriminative features, providing high-quality feature representation for subsequent classification or other tasks;

[0039] The above assembly accuracy prediction method based on the improved network of PointNet++ is optimized,

[0040] The main principle of the EfficientViM_CGLU block is to dynamically adjust the activation degree of the channel through convolution operation and gated linear unit; when the input feature tensor X , the dimension is B , C ,H , W ), wherein B is the batch size, C is the number of channels, H and W are the height and width of the feature map, using 1x1 convolution to map the input feature to twice the hidden dimension space,

[0041] (7)

[0042] wherein, W is the weight of the convolution kernel, b is the bias term, D is the number of output channels;

[0043] is then divided into two parts F and M , each part has a shape of I , B , D , H , W ): (8)

[0044] apply a 3x3 depth convolution to M , while the depth convolution applies a convolution kernel to each channel respectively, (9)

[0045] wherein, K is the radius of the convolution kernel, and then pass through the GELU activation function: (10)

[0046] wherein, is the cumulative distribution function of the standard normal distribution;

[0047] The above formula is combined with I through element-level multiplication to form a gating mechanism; through 1x1 convolution, the gated features are mapped back to the original number of channels.

[0048] The above improved network based on PointNet++ assembly accuracy prediction method, KACNConv2DLayer layer combines convolution operation and Chebyshev polynomial transformation; when the input feature vector x , using 3x3 convolution to extract local features is recorded as L 1 ;

[0049] Chebyshev formula transformation maps the input feature to a higher dimensional space to capture more complex patterns. And the higher dimensional space refers to the space with higher dimension than the original data space, that is, the space dimension corresponds to the number of features of the data points;

[0050] For the polynomial order K, the transformation is represented as: (12)

[0051] where, T n is n a Chebyshev polynomial of order K, α n is the learned weight coefficient;

[0052] The Chebyshev polynomial is recursively defined as: , (13)

[0053] The value of each pixel position in the feature map (size of the center point x the neighborhood points around each center point) is transformed by this polynomial.

[0054] The above assembly accuracy prediction method based on the improved network of PointNet++ is optimized. In the second stage of SetAbstraction, after sampling and grouping; by performing convolution operation on each neighborhood, more extensive feature patterns are captured, and attention mechanism operation is performed to generate QK and V are spliced into a feature map;

[0055] The feature map window is divided, the feature map size is adjusted to adapt to the window division, and the attention score is calculated by self-attention mechanism for each window (14)

[0056] Here, Q , K are query and key vectors respectively, V is the value vector, d k is the feature channel dimension, SoftMax the function is used to normalize the attention score to a probability distribution;

[0057] Get the updated features, then in the WATT module, by relative position coding of the features, get the relative position information, add the relative position weight into QK update the attention score matrix, and multiply it with V output the feature map (15) (16) where, R i,j represents the relative position information between points, x i and x j represent the point cloud positions of the first i and j points respectively.

[0058] The beneficial effects of the present application are:

[0059] In the technical solution of the present application, a parallelism prediction model is constructed by applying deep learning technology to fuse the physical model and the data model, and the mapping relationship between the point cloud new information of the part feature surface and the assembly surface, the deep network model and the assembly parallelism is determined. The robustness of the prediction process and the accuracy and efficiency of the measurement results are improved by the improved strategy based on PointNet++, so as to realize accurate and efficient parallelism prediction.

[0060] In the technical solution of the present application, a mapping relationship model of point cloud and parallelism is built. The proposed strategy does not need the assistance of additional modes, and realizes the preparation of the assembly parallelism data set. The problems of data set preparation and network model construction in the training of the current deep learning technology in the field of precision assembly are solved.

[0061] In the technical solution of the present application, a data set is constructed according to the measurement data and the classification standard by relying on a physical measurement system. The data set is input into the improved PointNet++. The EfficientViM_CGLU and KACNConv2DLayer modules are used to improve the extraction ability of the network to the point cloud features, and the self-attention mechanism method is combined to enhance the connection between local and global features. According to the experimental verification, compared with the existing deep learning network model, the network model has stronger learning ability to the point cloud information, and the accuracy of the predicted assembly parallelism is higher. At the same time, the network also has strong robustness and generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 is a schematic diagram of a physical measurement system;

[0063] Figure 2 is an assembly parallelism prediction process based on the improved PointNet++ network of the present application;

[0064] Figure 3 is a schematic diagram of the assembly parallelism prediction process based on the improved PointNet++ network (PAC-Net) model of the present application;

[0065] Fig. 4(a) is a schematic diagram of a two-stage one-stage feature extraction of the improved convolution and attention mechanism fusion optimization in the present application;

[0066] Fig. 4(b) is a flowchart of a two-stage feature extraction process of the improved convolution and attention mechanism fusion optimization in the present application;

[0067] Figure 5 is a comparison diagram of the assembly parallelism prediction results of each model. DETAILED DESCRIPTION

[0068] The technical features of the present application will be further described in combination with specific embodiments.

[0069] The overall process of predicting the assembly parallelism by the network model integrating multiple convolution methods and self-attention mechanisms built by the assembly precision prediction method based on the improved PointNet++ network of the present application is shown in Figure 2

[0070] The assembly precision prediction method based on the improved PointNet++ network of the present application first builds a physical measurement system, prepares a data set in the physical measurement system, divides the parallelism measured by the measuring machine into ten categories by a standard precision classification standard, prepares a data set, converts the parallelism value prediction problem into a category prediction, and further improves the prediction accuracy to avoid large errors caused by value deviation.

[0071] Specifically, the method comprises the following steps:

[0072] S1) Build a physical measurement system, prepare a data set in the physical measurement system, and build a neural network model; train the model by the data set prepared by the physical measurement system and the classification principle;

[0073] S2) In the physical measurement system, divide the parallelism category precision to obtain the assembly parallelism category and input it into the neural network as a label of the data set;

[0074] S3) In the first set abstraction and multi-scale convolution layer stage of feature extraction, the convolution is improved by introducing a gated linear unit; Chebyshev polynomials are introduced in the convolution operation, and features are extracted for multiple scales;

[0075] 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 operation on multiple scales at the same time to capture information of different scales;

[0076] S5) The feature information of the above two stages is input into the third Set Abstraction stage, convolution pooling operation is performed to generate a feature vector, the category probability of parallelism is output through full connection, and the assembly precision prediction is realized.

[0077] In one embodiment of the present application, the network architecture integrating multiple convolutions and attention mechanisms is referenced to PointNet++, as shown in Figure 3 ​The efficient ViM_CGLU and KACNConv2DLayer are constituted by introducing the gated linear unit and the Chebyshev polynomial combined with the ordinary convolution into the PointNet++ network to replace the standard convolution module therein. In addition, the LEGM layer is introduced to construct a network with higher precision and efficiency to enhance the relationship between the global and local, and the network has high feature extraction capability for point cloud features and high prediction accuracy for assembly parallelism.

[0078] In one embodiment of the present application, the data set preparation and network training process is as follows.

[0079] The data set used for PAC-Net network training is prepared by a physical measurement system and a virtual assembly system. The physical measurement system includes a Hexagon GLOBAL S 05.07.05 three-coordinate measuring machine and PC-DMIS supporting software, and the virtual debugging system takes surface point cloud data as input, is built based on MATLAB software, and calculates the assembly parallelism by setting corresponding theoretical algorithms. The category accuracy of the point cloud provides the target true value for network training and result evaluation.

[0080] In this embodiment, the data set is divided into 2000, 200 and 200 groups according to training, verification and testing, and the 200 group in the test group is prepared by the physical measurement system. The network training is performed on a Dell Delta workstation Precision 7920, the training is based on the Pytorch deep learning framework, the batch size is 32, and the training is performed for 200 rounds. The Adam optimizer is used to optimize the network parameters in the back propagation process, and the learning rate is initially set to 0.001. In this paper, the method of combining the gated linear mechanism and the Chebyshev polynomial with convolution is adopted, and the self-attention mechanism layer is combined, and the original feature extraction stage of PointNet++ is improved to ensure that the network training process can be smoothly fitted, and the accuracy of the inference result of the network model is further improved.

[0081] The principle of the technical scheme of the present application will be described in detail in combination with a specific application.

[0082] The principle of data set preparation is as follows:

[0083] Based on 5,184 dense points, 2,000 assembly examples are calculated, the parallelism value and the relative position of the surface point cloud of each assembly are obtained by the global series three-coordinate measuring machine of Hexagon, and the point cloud of the assembly surface is collected into 36x11 points, and the point cloud of the feature surface is collected into 4x4 points, which are specifically as follows: Figure 1As shown, the features of the parts 1, 2 and the assembly are measured by a three-coordinate measuring machine, and then the corresponding parallelism values are directly obtained through the PC-DIMS software, and then the operation is repeated to obtain 2000 assembly cases, and then divided into ten levels (T1 to T10) with 0.01 mm. 2000 examples are divided into training set, validation set and test set in the ratio of 7:1:2, and then input into the neural network for prediction. The specific process is as shown in Figure 2 As shown.

[0084] The principle of the network (PAC-Net) that fuses the improved convolution and attention mechanism is as follows.

[0085] As shown in Figure 3 PAC-Net network is based on PointNet++ network, combined with multiple convolution methods and self-attention mechanism, taking point cloud classification task as the goal, in order to build a network structure with high efficiency, high accuracy and strong generalization ability. The main input of this network is the three-dimensional coordinates of point cloud, and the output is the assembly parallelism category. It is composed of shared multi-layer perception (S-MLP) layer, set abstraction and multi-scale convolution (Set Abstraction MSG Convolution) layer, set abstraction and multi-scale attention mechanism (Set Abstraction MSG Attention) layer, set abstraction (Set Abstraction) layer and fully connected layer. Through the input of three-dimensional information of surface point cloud, the feature extraction is completed through three stages of SetAbstraction layer, after the maximum value pooling layer, it is mapped into a feature vector, and then the fully connected layer is processed, the probability value of the assembly parallelism in the category is output, and the classifier is constructed to complete the category prediction of the parallelism.

[0086] 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 input point cloud of the neural network training is represented by a three-dimensional array with N elements. Each element Pi Records the 3D point cloud coordinates. Through shared multi-layer perception (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 three stages (b), (c), (d). The dimension number N is reduced to 1 / 16 of the original, and the coordinate dimension is increased by 16 times.

[0087] In the feature extraction process, three types of set abstraction layers are designed to build global and local relationships between different points. First, the input point cloud data is sampled by the farthest point, assuming that the point cloud data set is , where each point P iis 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 its distance to the nearest sampling point as d i Initially, the distance of all points to q 1 is formula (17). (17)

[0088] Select the point farthest from the current sampling point set as the next sampling point , that is:

[0089] (18)

[0090] Then for each point P i , update its distance to the nearest sampling point as the minimum of the current distance and the new distance: (19)

[0091] Continuously perform the above operation, effectively retain the main structural features of the point cloud, while reducing the data amount, select 512 points in the first feature extraction stage, then perform spherical neighborhood query, define three query radii r . Then calculate the distance of all points to the center point, perform neighborhood division, ensure that there are k points in the neighborhood (here k need to be set in advance).

[0092] As shown in FIG. 4(a), in the neighborhood under the first radius, the input feature X 1 ∈R N×C , where N is the number of points, C is the feature dimension is sequentially convolved, EfficientViM_CGLU layer and three layers of convolution operation. Specifically, first perform linear transformation, formula as shown in 20.

[0093] (20)

[0094] Where W∈R C×C is a weight matrix, b∈R C is a bias vector, then input to the EfficientViM_CGLU layer, the core of the layer is a gating mechanism, as shown in formula 21. (21)

[0095] That is, the feature extraction in the first scale is completed by the activation function and the feature transformation. The KACNConv2DLayer operation is used to learn the sample features in the neighborhoods of the second and third radii. That is, the Chebyshev polynomial transformation is performed on the local neighborhood, as shown in equation 22. (22)

[0096] wherein, k represents the order and the Chebyshev polynomial transformation is an operation on the features of each point, which does not change the number of points but changes the values of the features, and the feature dimension of each point does not change but the feature value is transformed into T k (x) Then, the convolution operation is performed, as shown in equation 23. (23)

[0097] wherein, W conv is the convolution kernel weight matrix, b is the bias vector, and the features extracted in the two ways are spliced as new features input to the second stage. (24)

[0098] In the feature extraction stage of the second layer, as shown in FIG. 4(b), the LEGM and WATT modules are introduced to increase the attention mechanism. In order to enhance the feature interaction between dimensions, the channel cross-attention module is introduced. Then, the features are extracted through a series of convolution operations, and are aggregated through pooling. The aggregated features are connected with the initial input features, and then input to the third layer to abstract the point cloud data into a global feature vector. The vector is input to the fully connected layer for step-by-step dimension reduction, and finally the classification result is output.

[0099] Specifically, after the sampling and grouping operations, the feature matrix of the points in the field of the new sampling points is obtained, which has a shape of [B,C+D,nsample,npoint] wherein 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.

[0100] The new sampling points in the local neighborhood are input into the attention mechanism layer, the input features are mapped to a dimension suitable for attention mechanism processing through convolution, and then in the LEGM layer, the input features are normalized to a range of features easy to extract, and the average value and variance value of each column are taken. And the normalized features pass through the attention mechanism module, that is, the feature map first passes through two linear layers to be QK and VThe two parts are spliced to form a new feature map QKV The new feature map is obtained, the feature map is divided into a plurality of fixed size windows, then the feature map is rearranged, then the feature map is divided into windows, and each window is flattened into a one-dimensional vector, and the WATT module is applied to each window. In the WATT module, the input feature QKV is divided into Q , K and V , and the attention score is calculated: (25)

[0101] wherein C represents the number of feature channels. Then, the initial input feature is subjected to relative position encoding, a relative position offset is dynamically generated by calculating the distance of the points, the relative position is added as a bias to the attention score, and the attention score is applied through a Softmax layer, and the V is weighted and summed. The output features of each window are merged back to the original feature map size, and in addition, the output of the attention mechanism is fused with the output of the convolution operation through a projection layer, completing the extraction process 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 convolution operation.

[0102] After feature extraction, the unordered input of the point cloud is processed using the symmetry layer, 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 symmetry function is defined as:

[0103] (26)

[0104] Equation (26) defines an idealized end-to-end training process e , which can directly convert the input point cloud data into a prediction label. In this process, PAC-Net serves as the backbone of the network, responsible for feature extraction and processing. Through the max-pooling layer, the feature map is compressed into a 512-dimensional vector. Subsequently, the vector is input into a classifier composed of three fully connected layers (FC layers), with neuron numbers of 256, 128 and C , respectively, where C represents the number of categories. In order to accelerate the training process and prevent the occurrence of gradient vanishing and overfitting, normalization layers and dropout layers are added between the fully connected layers. Finally, the classifier adopts the sigmoid function as the activation function, and the cross-entropy loss function is used to calculate the difference between the prediction result and the true label.

[0105] Based on the PAC-Net network, a detailed introduction and analysis are carried out. PAC-Net realizes the effective extraction and fusion of different scale features in point cloud data by introducing attention mechanism (such as LEGM and WATT modules) and multiple feature extraction methods (such as EfficientViM_CGLU layer and KANConv2DLayer). This design not only enhances the model's perception of local and global features, but also improves the accuracy of feature expression by dynamically adjusting the feature weight. In order to verify the effectiveness of PAC-Net, model comparison verification is carried out, and Table 1 shows the performance comparison of PAC-Net and other models (such as PointNet2 and its variants) on the parallelism prediction task. The accuracy can reach 92.18%, which meets the industrial requirements and can be used.

[0106] Table 1 Comparison of training results

[0107] 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%

[0108] Of course, the above description is not a limitation of the present application, and the present application is not limited to the above examples. Within the scope of the present application, changes, modifications, additions or substitutions made by ordinary skilled in the art should be within the scope of the present application.

Claims

1. An assembly precision prediction method based on a PointNet++ improved network, characterized in that: The method comprises the following steps: S1) building a physical measurement system; preparing a data set in the physical measurement system, and building a neural network model; training the model through the data set prepared by the physical measurement system and the classification principle; S2) in the physical measurement system, the parallelism category accuracy is divided to obtain the assembly parallelism category as the label of the data set and input into the neural network; S3) in the first set of feature extraction abstraction stage, the convolution is improved by introducing a gated linear unit; Chebyshev polynomials are introduced in the convolution operation, and features are extracted at multiple scales; S4) in the second set of feature extraction abstraction stage, an attention mechanism layer is introduced to perform convolution and attention mechanism operation on multiple scales at the same time to capture information at different scales; S5) after the feature information obtained in the first set of feature extraction abstraction stage and the second set of feature extraction abstraction stage is spliced, it is input into the third set of abstraction stage to perform convolution pooling operation to generate a feature vector; the category probability of parallelism is output through full connection to realize assembly accuracy prediction; In step S1), when building the neural network model, the gated linear unit, Chebyshev polynomial and ordinary convolution are combined and 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 construct the network; In step S2), the point cloud data of the assembly surface and the feature surface of the part are 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 of feature extraction abstraction stage, grouping is performed by setting a small radius parameter, and then the EfficientViM_CGLU block is added to the smaller radius scale to dynamically adjust the activation function of the feature channel; KACNConv2DLayer layers are used at the second and third radius scales to strengthen feature expression; The features of the three radius scales are spliced to obtain more features; In step S4), in the second set of feature extraction abstraction stage, 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 for scaling the feature map; The convolution operation extracts local features, and the attention mechanism highlights important features and suppresses unimportant features; The WATT module learns the importance weight of the features to adaptively scale the feature map and enhance the expression ability of global and local features.

2. The assembly accuracy prediction method based on the improved network of PointNet++ according to claim 1, wherein: In the first set of feature extraction abstraction stage, first, the farthest point sampling is used to downsample the point cloud to obtain a new point set; then, the ball query algorithm is used to query the neighborhood points of each new point to form a local point cloud; Conv represents a standard convolution operation for extracting higher-level features from input features, and the dimension of the output feature is affected by setting parameters such as the size and sliding step of the convolution kernel. ； F group1 represents the input feature matrix, which represents the local point cloud feature geometry obtained by the ball query algorithm, and these features include the coordinate information of the local points and their corresponding feature vectors, and each row in the matrix corresponds to the features of a point; EfficientViM_CGLU represents a high-efficiency feature channel activation adjustment module, and the final output feature matrix F1 is obtained. KACNConv2DLayer is a two-dimensional convolution layer for extracting higher-level features from input features. a second group of local features F with a larger radius group2 and a third group of local features F with the largest radius group3 respectively applying KACNConv2DLayer layers, respectively expressed as F group2 represents the input feature matrix, which indicates the local point cloud feature geometry obtained by the ball query algorithm within the second set of radii, and these features include the coordinate information of the local points and their corresponding feature vectors; F2 represents the output feature matrix. F3 represents the output feature matrix, which is the feature matrix processed by the KACNConv2DLayer layer. F group3 is an input feature matrix, representing the local point cloud feature set obtained by the ball query algorithm within the third set of radii; The features under three scales are spliced to obtain comprehensive features: F1 is the first feature matrix, representing the local point cloud features within a small radius. F2 is the second feature matrix, representing the local point cloud features within a medium radius. F3 is the third feature matrix, representing the local point cloud features within a large radius. In the second feature extraction stage, convolution operation is combined with attention mechanism, and LEGM layer is introduced, and WATT module is added in the attention layer. Final features Include feature information at different scales; Convolution operation is performed on the input feature map 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 feature map output by the LEGM layer is input into the WATT module to learn the importance weight of the feature and perform adaptive scaling on the feature map. F conv is the feature map after convolution operation, LEGM is a local and global information combined attention mechanism module, which models the global information of the feature map, and then uses the obtained global features to modulate the local features; wherein F legm represents the feature map after the LEGM layer processing on the basis of the convolution feature map; WATT represents an adaptive feature weight learning module, which is used to learn the importance weight of the feature and to perform adaptive scaling on the feature map according to the weight; and the output final feature map F watt processed after the WATT module comprehensively integrates the local features extracted by the convolution operation, the important feature information enhanced by the LEGM layer, and the structure of the adaptive weight learning. The feature map of the previous layer is input into the Set Abstraction layer, and direct convolution is performed for local feature extraction, and then global feature extraction is performed on the initial point cloud maximum pooling, and then spliced into a high-dimensional vector. After processing by the full connection layer, the category probability is output.

3. The assembly precision prediction method based on the improved network of PointNet++ according to claim 2, characterized in that: The EfficientViM_CGLU block dynamically adjusts the activation degree of the channel through convolution operation and gated linear unit. When the input feature tensor X has a dimension of (B, C, H, W), where B is the batch size, C is the channel number, H and W are the height and width of the feature map, a 1×1 convolution is used to map the input feature to a two-fold hidden dimension space, where W is the weight of the convolution kernel, b is the bias term, and D is the output channel number. Then F is divided into two parts M and I, each part has a shape of (B, D, H, W): where K is the radius of the convolution kernel, and then activated by a GELU activation function: wherein is the cumulative distribution function of the standard normal distribution; M activated The gating mechanism is combined with the element-wise multiplication of I, and the gated features are mapped back to the original channel number through 1x1 convolution. 3×3 deep convolution is applied to M, and the deep convolution applies a convolution kernel to each channel, 4. The assembly precision prediction method based on the improved network of PointNet++ according to claim 1, characterized in that: The KACNConv2DLayer layer combines convolution operation and Chebyshev polynomial transformation; when the input feature vector x is used, 3×3 convolution is used to extract local features denoted as L1; Chebyshev formula transformation is performed to map the input feature dimension to a higher dimensional space to capture more complex patterns; and the higher dimensional space refers to a space with higher dimension than the original data space, i.e., the space dimension corresponds to the number of features of the data points; For the polynomial order K, the transformation is represented as: ； where T n is an n-th order Chebyshev polynomial, and α n is a learned weight coefficient. The recursive definition of the Chebyshev polynomial is: ， The value of each pixel position in the feature map is transformed by the polynomial.

5. The assembly precision prediction method based on the improved network of PointNet++ according to claim 1, characterized in that: In the set abstraction of the second stage, after sampling and grouping; By performing a convolution operation on each neighborhood, more extensive feature patterns are captured, and an attention mechanism operation is performed to generate QK and V to be spliced into a feature map; The feature map window is divided, the feature map size is adjusted to adapt to the window division, and the self-attention mechanism is calculated for each window to obtain the attention score Here, Q, K are query and key vectors, V is value vector, d k is the dimension of the feature channel, and the SoftMax function is used to normalize the attention scores into a probability distribution; The updated features are obtained, and then in the WATT module, the relative position information is obtained by performing relative position coding on the features, the relative position is weighted into QK to update the attention score matrix, and the feature map is output by multiplying V. where R i,j represents the relative position information between points, x i and x j represent the point cloud positions of the i and j points, respectively.

Images