Aerospace component registration method and electronic equipment

Abstract

The invention provides a spaceflight component registration method and electronic equipment, and the method comprises the following steps: carrying out the preprocessing of an original point cloud of a spaceflight component, and obtaining a preprocessed point cloud; inputting the preprocessed point cloud into an encoder and a decoder to obtain a rough point cloud; inputting the preprocessed point cloud and the rough point cloud into a generator to obtain a complete point cloud; and according to the complete point cloud, carrying out registration on the spaceflight component. Through the self-supervised learning framework and the attention mechanism, the problem of point cloud missing is solved, it is ensured that the input point cloud is more complete, and therefore the registration precision is remarkably improved. In addition, multiple features are combined, multi-feature fusion is achieved through weight addition, and the accuracy and reliability of registration are further improved. And meanwhile, through reasonable module design and optimization, the calculation complexity is reduced, so that the method is more efficient when processing large-scale point clouds.

Description

Technical Field

[0001] The present invention relates to the field of aerospace technology, and in particular to an aerospace component registration method and electronic equipment. Background Art

[0002] With the continuous development of aerospace technology, large aerospace components have higher and higher requirements for accuracy and reliability in the manufacturing and assembly process. Traditional component registration methods usually rely on complete geometric information and complex measurement equipment. However, in actual operation, obtaining complete point cloud data often faces many challenges, such as sensor viewing angle limitations, occlusion and data noise. These factors make it an urgent problem to solve how to perform high-precision component registration in the absence of information. Existing point cloud registration methods mainly rely on geometric features and topological structures for registration. The iterative closest point (ICP) algorithm achieves registration by iteratively minimizing the distance between two point clouds, but the registration effect is often not ideal when dealing with partially missing point cloud data. The normal distribution transformation (NDT) algorithm represents the geometric structure of the point cloud by constructing a probability distribution. Although it can handle partially missing data to a certain extent, its computational complexity is high and its effect is limited when facing complex scenes. Summary of the invention

[0003] The embodiments of the present application provide a method and electronic device for aerospace component registration, which innovatively solves the problem of missing point cloud information through self-supervised learning and attention mechanism, and combines multiple features such as edge curvature similarity and contour matching to achieve high-precision, low-computational complexity aerospace component registration.

[0004] An embodiment of the present application provides a method for registering a spacecraft component, comprising the following steps: preprocessing an original point cloud of the spacecraft component to obtain a preprocessed point cloud; inputting the preprocessed point cloud into an encoder and a decoder to obtain a rough point cloud; inputting the preprocessed point cloud and the rough point cloud into a generator to obtain a complete point cloud; registering the spacecraft component according to the complete point cloud; wherein the original point cloud is a residual point cloud, and the generator includes at least one attention feature enhancement layer.

[0005] In one embodiment, the preprocessing of the original point cloud of the aerospace component to obtain the preprocessed point cloud includes: sampling the farthest point of the original point cloud to obtain a sampled point cloud; normalizing the sampled point cloud to obtain a normalized point cloud; and dividing the normalized point cloud into regions to obtain a preprocessed point cloud.

[0006] In one embodiment, the farthest point sampling of the original point cloud to obtain the sampled point cloud includes: arbitrarily selecting one or more point clouds in the original point cloud as an initial sampled point cloud set, and the remaining original point clouds as an initial unsampled point cloud set; for each point cloud in the current unsampled point cloud set, calculating the minimum distance from the unsampled point cloud to all point clouds in the current sampled point cloud set, and adding the unsampled point cloud corresponding to the maximum value of the minimum distance to the current sampled point cloud set; when the number of point clouds in the current sampled point cloud set reaches a preset threshold, using the point cloud in the current sampled point cloud set as the sampled point cloud.

[0007] In one embodiment, the normalized point cloud is divided into regions to obtain a preprocessed point cloud, including: randomly initializing M center points according to the normalized point cloud; extracting neighborhood regions for the M center points based on the K nearest neighbor algorithm, and dividing the normalized point cloud into M parts; for each of the M parts, dividing it into a first part, a second part and a third part; using the first part, the second part and the third part as the preprocessed point cloud; wherein the first part is an area observable by the encoder, the second part is a masked area, and the third part is an area unobservable by the encoder.

[0008] In one embodiment, the inputting the preprocessed point cloud into an encoder and a decoder to obtain a rough point cloud comprises: inputting the preprocessed point cloud into an encoder to extract important feature information of the point cloud; and inputting the important feature information of the point cloud into a decoder to obtain the rough point cloud.

[0009] In one embodiment, the encoder includes a first shared multi-layer perceptron layer, a first maximum pooling layer, an expansion layer, a second shared multi-layer perceptron layer and a second maximum pooling layer; the pre-processed point cloud is input into the encoder to extract important feature information of the point cloud, including: inputting the pre-processed point cloud into the first shared multi-layer perceptron layer, performing linear transformation and nonlinear activation on each point cloud, and extracting shallow features; inputting the shallow features into the first maximum pooling layer, extracting the maximum value of each point cloud in the feature dimension, and obtaining significant features; inputting the significant features into the expansion layer to obtain extended features; wherein the extended features have the same feature dimension as the shallow features; inputting the shallow features and the extended features into the second shared multi-layer perceptron layer to obtain comprehensive features; inputting the comprehensive features into the second maximum pooling layer to extract important feature information of the point cloud.

[0010] In one embodiment, the encoder includes a third shared multi-layer perceptron layer and a reshaping layer; the inputting of the important feature information of the point cloud into the decoder to obtain the rough point cloud includes: inputting the important feature information of the point cloud into the third shared multi-layer perceptron layer to extract decoding features; inputting the decoding features into the reshaping layer to convert the decoding features into features of a preset shape and format to obtain the rough point cloud; wherein the rough point cloud includes at least the main geometric features of the original point cloud.

[0011] In one embodiment, the generator includes: a first multi-layer perceptron layer, a second multi-layer perceptron layer, three attention feature enhancement layers and a third multi-layer perceptron layer; the preprocessed point cloud and the rough point cloud are input into the generator to obtain a complete point cloud, including: inputting the preprocessed point cloud and the rough point cloud together into the first multi-layer perceptron layer to extract first multi-level features; inputting the important feature information of the point cloud into the second multi-layer perceptron layer to extract second multi-level features; superimposing the first multi-level features and the second multi-level features, and sequentially inputting them into three attention feature enhancement layers to obtain attention features; superimposing the multi-level features and the attention features, and inputting them into the third multi-layer perceptron layer to obtain the complete point cloud.

[0012] In one embodiment, the registering of the aerospace component according to the complete point cloud includes: obtaining edge curvature similarity and contour matching of the aerospace component according to the complete point cloud; obtaining a multi-feature registration result of the aerospace component through weighted fusion according to the curvature similarity and contour matching; and registering the aerospace component when the multi-feature registration result meets preset conditions.

[0013] An embodiment of the present application further provides an electronic device, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to execute the above-mentioned aerospace component registration method.

[0014] The solution provided by the above-mentioned embodiments of the present application innovatively solves the problem of missing point clouds through a self-supervised learning framework and an attention mechanism, ensuring that the input point cloud is more complete, thereby significantly improving the registration accuracy. In addition, the present application combines a variety of features (such as edge curvature similarity and contour matching), and realizes multi-feature fusion through weighted addition, further improving the accuracy and reliability of registration. At the same time, through reasonable module design and optimization, the computational complexity is reduced, making the present application more efficient when processing large-scale point clouds. Compared with traditional methods, the present application can not only adaptively extract important feature areas, but also comprehensively consider the importance of each feature through a multi-feature fusion strategy, thereby achieving high-precision, low-computational complexity aerospace component registration. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solution of the embodiments of the present application, the drawings required for use in the embodiments of the present application are briefly introduced below.

[0016] Figure 1 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application;

[0017] Figure 2 It is a flowchart of a method for aligning aerospace components provided in an embodiment of the present application;

[0018] Figure 3 It is a flow chart of another aerospace component registration method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0020] Similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0021] As mentioned above, how to perform high-precision aerospace component registration has become an urgent problem to be solved. In recent years, point cloud completion methods based on machine learning have begun to attract attention. Through deep learning models, complete point cloud information can be learned and reconstructed from partial point cloud data, thereby improving the accuracy of registration. In particular, the self-supervised learning paradigm has shown great potential in the field of point cloud completion. This method does not require complete point cloud information and is suitable for scenarios where it is difficult to obtain complete point cloud information for large-size components in the aerospace field. By combining the encoder-decoder architecture, it is possible to efficiently learn and extract multiple features of residual point cloud information, thereby achieving point cloud completion. However, existing self-supervised learning methods still have certain limitations when dealing with large-scale aerospace component registration with high-precision requirements, such as insufficient reconstruction accuracy and limited ability to capture complex geometric shapes. In addition, as an important research direction in the field of deep learning in recent years, the attention mechanism has performed well in dealing with tasks that require capturing long-distance dependencies. Introducing the attention mechanism into the field of point cloud processing can effectively reveal the fine geometric features of three-dimensional shapes and enhance the understanding of the spatial relationship between points, thereby providing support for high-precision point cloud registration. However, the application of attention mechanism to point cloud registration is still in its early stages, and its specific application effects and optimization strategies need further exploration. Combining the self-supervised learning framework with the attention mechanism to achieve point cloud reshaping and thus improve registration accuracy is the focus of current research.

[0022] In order to solve the above problems, the embodiment of the present application at least provides a method for aerospace component registration, including the following steps: preprocessing the original point cloud of the aerospace component to obtain a preprocessed point cloud; inputting the preprocessed point cloud into an encoder and a decoder to obtain a rough point cloud; inputting the preprocessed point cloud and the rough point cloud into a generator to obtain a complete point cloud; and registering the aerospace component according to the complete point cloud; wherein the original point cloud is a residual point cloud, and the generator includes at least one attention feature enhancement module. The embodiment of the present application integrates the self-supervised learning framework and the attention enhancement module, and innovatively solves the problem of missing complete point cloud data in the field of large aerospace components, thereby improving the reliability and accuracy of large aerospace component registration.

[0023] Figure 1 1 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The electronic device 100 is used to execute the aerospace component registration method provided in an embodiment of the present application. Figure 1 As shown, the electronic device 100 includes: one or more processors 102, and one or more memories 104 storing processor executable instructions. The processor 102 is configured to execute the aerospace component registration method provided in the following embodiment of the present application.

[0024] It is important to note that Figure 1 The electronic device in is only one possible device structure, and there are other possible device structures, which are not limited in this application.

[0025] The processor 102 can be a gateway, a smart terminal, or a device including a central processing unit (CPU), a graphics processing unit (GPU), or other forms of processing units with data processing capabilities and / or instruction execution capabilities. It can process data of other components in the electronic device 100 and control other components in the electronic device 100 to perform desired functions.

[0026] The memory 104 may include one or more computer program products, and the computer program product may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 102 may execute the program instructions to implement the aerospace component registration method described below. Various applications and various data may also be stored in the computer-readable storage medium, such as various data used and / or generated by the application.

[0027] In one embodiment, Figure 1 The electronic device 100 shown may also include an input device 106, an output device 108, and a data acquisition device 110, which are interconnected via a bus system 112 and / or other forms of connection mechanisms (not shown). Figure 1 The components and structures of the electronic device 100 shown are merely exemplary and non-limiting. The electronic device 100 may also have other components and structures as required.

[0028] The input device 106 may be a device for a user to input instructions, and may include one or more of a keyboard, a mouse, a microphone, a touch screen, etc. The output device 108 may output various information (e.g., images or sounds) to the outside (e.g., a user), and may include one or more of a display, a speaker, etc. The data acquisition device 110 may acquire the original point cloud of the aerospace component. Exemplarily, the data acquisition device 110 may be an electromagnetic ultrasonic thickness measuring device.

[0029] In one embodiment, the various components in the example electronic device 100 for implementing the aerospace component alignment method of the embodiment of the present application may be integrated or dispersed, such as integrating the processor 102, memory 104, input device 106 and output device 108 into one, while separating the data acquisition device 110.

[0030] In one embodiment, the example electronic device 100 for implementing the aerospace component registration method of the embodiment of the present application may be implemented as an artificial intelligence (AI) server, a supercomputer, etc.

[0031] Figure 2 The present invention is a flowchart of an aerospace component registration method according to an embodiment of the present invention. The method is executed by an aerospace component registration system, which includes an encoder, a decoder, a generator, and a multi-feature registration module.

[0032] like Figure 2 As shown, the aerospace component registration method includes the following steps 201-204.

[0033] Step 201: pre-processing the original point cloud of the aerospace component to obtain a pre-processed point cloud.

[0034] Regarding aerospace components, they may be parts of various spacecraft (such as rockets), for example, they may be melon segments of a spacecraft.

[0035] Regarding the original point cloud of the aerospace component, the aerospace component can be measured and collected by an electromagnetic ultrasonic thickness measuring device. In the embodiment of the present application, the point cloud is also referred to as point cloud data or point cloud information. The original point cloud comes from different aerospace components, which will be assembled together later. The original point cloud includes at least the three-dimensional coordinates (x, y, z) of the points on the surface of the measured aerospace component, the main geometric features of the measured aerospace component, and the detailed features of the measured aerospace component. The main features may be, for example, the contour and surface shape of the aerospace component, and the detailed features may be, for example, the curvature of the non-edge part of the aerospace component. Since it is difficult to obtain a complete point cloud of aerospace components, especially large aerospace components, by measuring, the original point cloud is a residual point cloud. Steps 201-203 of the embodiment of the present application are the process of changing the residual point cloud into a complete point cloud through a self-supervised learning framework and an attention mechanism.

[0036] The preprocessed point cloud is the point cloud obtained after a series of processing of the original point cloud, which is the input of the subsequent model (also called "network"). The model combines the self-supervised learning framework and the attention mechanism, specifically including: encoder, decoder and generator.

[0037] Step 201 specifically includes: performing farthest point sampling on the original point cloud to obtain a sampled point cloud; performing normalization processing on the sampled point cloud to obtain a normalized point cloud; and performing region division on the normalized point cloud to obtain a preprocessed point cloud.

[0038] The farthest point sampling is performed on the original point cloud to obtain a sampled point cloud, specifically including: arbitrarily selecting one or more point clouds in the original point cloud as the initial sampled point cloud set, and the remaining original point clouds as the initial unsampled point cloud set; for each point cloud in the current unsampled point cloud set, calculating the minimum distance from the unsampled point cloud to all point clouds in the current sampled point cloud set, and adding the unsampled point cloud corresponding to the maximum value of the minimum distance to the current sampled point cloud set; when the number of point clouds in the current sampled point cloud set reaches a preset threshold, the point cloud in the current sampled point cloud set is used as the sampled point cloud. It can be understood that the initial sampled point cloud set is the first current sampled point cloud set, and the initial unsampled point cloud set is the first current unsampled point cloud set.

[0039] The normalized point cloud is divided into regions to obtain a preprocessed point cloud, specifically comprising: randomly initializing M center points according to the normalized point cloud; extracting neighborhood regions for the M center points based on the K nearest neighbor algorithm, and dividing the normalized point cloud into M parts; for each of the M parts, dividing it into a first part, a second part and a third part; taking the first part, the second part and the third part as the preprocessed point cloud; wherein the first part is an area observable by the encoder, the second part is a mask area, and the third part is an area unobservable by the encoder.

[0040] In one embodiment, the specific implementation of step 201 is:

[0041] First, the farthest point of the original point cloud is sampled, and the minimum value of the unsampled point cloud and the sampled point cloud set is continuously retained. The point corresponding to the maximum value of these minimum distances is selected as the new sampling point cloud, and it is added to the sampling point cloud set until enough points are sampled to ensure uniform sampling. The specific mathematical expression is:

[0042]

[0043] Among them, Q is the set of point clouds that have not been sampled, that is, the unsampled point cloud set, P is the set of point clouds that have been sampled, that is, the sampled point cloud set, p i is the unsampled point cloud, p is the sampled point cloud, P next is the next sampled point cloud. argmax represents the independent variable that maximizes the objective function.

[0044] Then, the sampled point cloud is normalized and the range is set to [-1,1] to obtain a normalized point cloud, which helps the model remain stable during training. The specific mathematical expression is:

[0045]

[0046] x, y, and z are the coordinate values ​​of the sampling point cloud respectively; center ,y center , z center are the coordinate averages of all sampling point clouds; max ,y max , z max are the maximum coordinates of the sampling point cloud; min ,y min , z min are the minimum coordinate values ​​of the sampling point cloud.

[0047] Next, based on the normalized point cloud, M center points are randomly initialized, and the neighborhood area is extracted for each center point based on the K nearest neighbor (KNN) algorithm, and then divided into M parts. Each part represents different local feature information of the point cloud. Each local area can capture different geometric shapes and texture features, helping the network understand the structural information of each position on the surface of the object. Each of the M parts is further divided into three parts: the first part, the second part, and the third part. The first part P obs Represents the observable area of ​​the network, forcing the decoder to output the predicted rough point cloud P c Align with this area. In order to further standardize point cloud completion, P mis , represents the masked area that needs to be predicted, encouraging the network to fill in the intentionally masked and naturally missing areas. latent Including segments not observed by the encoder allows for additional regularization. Effective point cloud region segmentation ensures that the network can learn diverse feature information on the surface of the object, thereby improving the performance of the completion task. It should be noted that the area ratio between the first part, the second part, and the third part can be 20:40:4.

[0048] Finally, the first part, the second part and the third part are used as the preprocessed point cloud.

[0049] Step 202: Input the preprocessed point cloud into an encoder and a decoder to obtain a rough point cloud.

[0050] Regarding the encoder, it may include a first shared multi-layer perceptron layer (ShareMLP), a first maximum pooling layer (Point-wise Maxpool), an expansion layer (Expand), a second shared multi-layer perceptron layer and a second maximum pooling layer.

[0051] Regarding the decoder, a third shared multi-layer perceptron layer and a reshape layer (Reshape) may be included.

[0052] Regarding the rough point cloud, it retains the main geometric features of the original point cloud, such as the contour and surface shape of the component, but the detailed information is reduced.

[0053] Step 202 specifically includes: inputting the preprocessed point cloud into an encoder to extract important feature information of the point cloud; and inputting the important feature information of the point cloud into a decoder to obtain the rough point cloud.

[0054] The preprocessed point cloud is input into the encoder to extract important feature information of the point cloud, specifically including: inputting the preprocessed point cloud into the first shared multi-layer perceptron layer, performing linear transformation and nonlinear activation on each point cloud, and extracting shallow features; inputting the shallow features into the first maximum pooling layer, extracting the maximum value of each point cloud in the feature dimension, and obtaining significant features; inputting the significant features into the extension layer to obtain extended features; wherein the extended features have the same feature dimension as the shallow features; inputting the shallow features and the extended features into the second shared multi-layer perceptron layer to obtain comprehensive features; inputting the comprehensive features into the second maximum pooling layer to extract important feature information of the point cloud.

[0055] Inputting the important feature information of the point cloud into a decoder to obtain the rough point cloud specifically includes: inputting the important feature information of the point cloud into a third shared multi-layer perceptron layer to extract decoding features; inputting the decoding features into a reshaping layer to convert the decoding features into features of a preset shape and format to obtain the rough point cloud.

[0056] Step 203: Input the preprocessed point cloud and the rough point cloud into a generator to obtain a complete point cloud.

[0057] Regarding the generator, it may include a first multi-layer perceptron layer (MLP), a second multi-layer perceptron layer, three attention feature enhancement layers (Attention Feature Augmenttation) and a third multi-layer perceptron layer.

[0058] The complete point cloud is the original point cloud, or residual point cloud, obtained after being processed by the model. It is the point cloud predicted by the model. Compared with the original point cloud, the complete point cloud can characterize the features of aerospace components (such as contour features and surface features) in more detail, so as to complete the subsequent component registration operation more accurately.

[0059] Step 203 specifically includes: inputting the preprocessed point cloud and the rough point cloud into the first multi-layer perceptron layer together to extract the first multi-level features; inputting the important feature information of the point cloud into the second multi-layer perceptron layer to extract the second multi-level features; superimposing the first multi-level features and the second multi-level features, and sequentially inputting them into three attention feature enhancement layers to obtain attention features; superimposing the multi-level features and the attention features, and inputting them into the third multi-layer perceptron layer to obtain the complete point cloud.

[0060] Step 204: aligning the aerospace component according to the complete point cloud.

[0061] Regarding the registration, the aerospace components may be assembled, such as assembling a plurality of melon segments. As mentioned above, the original point clouds come from different aerospace components, and therefore, the complete point clouds obtained from the original point clouds also come from different aerospace components. These aerospace components will be assembled in step 204, and the edge curvature similarity and contour matching between the aerospace components to be assembled can be obtained through the complete point clouds of these aerospace components, and then the final registration is completed based on these two features.

[0062] Step 204 specifically includes: obtaining edge curvature similarity and contour matching of the aerospace component based on the complete point cloud; obtaining a multi-feature registration result of the aerospace component through weighted fusion based on the curvature similarity and contour matching; and registering the aerospace component when the multi-feature registration result meets preset conditions.

[0063] Figure 3 1 is a flow chart of another aerospace component registration method according to an embodiment of the present application. The method is executed by an aerospace component registration system, which includes an encoder, a decoder, a generator, and a multi-feature registration module. The method specifically includes the following steps:

[0064] Step 301: Measure and collect data on a plurality of aerospace components to be assembled using an electromagnetic ultrasonic thickness measuring device to obtain an original point cloud.

[0065] The multiple aerospace components to be assembled may be multiple melon segments of a rocket tank, and these melon segments need to be assembled, that is, aligned.

[0066] Step 302: preprocess the original point cloud to obtain a preprocessed point cloud.

[0067] The specific method for preprocessing the original point cloud is the same as step 201 and will not be repeated here.

[0068] Step 303: Input the preprocessed point cloud into the encoder to extract important feature information of the point cloud.

[0069] Step 303 may specifically include the following steps:

[0070] Step 3031: The preprocessed point cloud enters the first shared multi-layer perceptron layer (ShareMLP1), performs linear transformation and nonlinear activation on each point, and extracts shallow features.

[0071] Step 3032: Input the shallow features into the first maximum pooling layer (Point-wise Maxpool1), perform maximum pooling on the features of each point, extract the maximum value of each point in the feature dimension, reduce the feature dimension, and retain key feature information.

[0072] Step 3033: The pooled features are restored to the same feature dimension as the original point cloud through the expansion layer (Expand), ensuring the consistency of the features in space so that each point can maintain the feature representation of its original position.

[0073] Step 3034: The shallow features and the expanded features are processed again by the shared multi-layer perceptron layer (ie, input into the first shared multi-layer perceptron layer (ShareMLP2)) to further extract features.

[0074] Step 3035: Perform point-wise max pooling again (i.e., the output of step 3034 will be input into the second max pooling layer (Point-wise Maxpool2)) to extract important feature information of the point cloud.

[0075] Step 304: Input the important feature information of the point cloud into the decoder to obtain a rough point cloud.

[0076] Step 304 may specifically include the following steps:

[0077] Step 3041: The feature vector output by the encoder enters the third shared multi-layer perceptron layer (ShareMLP3) of the decoder for further processing and feature extraction;

[0078] Step 3042: Finally, the decoded feature vector is converted into the required shape and format through the reshape layer to obtain a rough point cloud.

[0079] Step 305: Input the preprocessed point cloud and the rough point cloud into a generator to obtain a complete point cloud.

[0080] Step 305 may specifically include the following steps:

[0081] Step 3051: The rough point cloud output by the decoder is input to the first multi-layer perceptron layer (MLP1) together with the pre-processed point cloud to extract multi-level features. At the same time, the important feature information of the point cloud extracted in step 303 is further extracted through a separate second multi-layer perceptron layer (MLP2). The feature matrices are combined together by point-by-point superposition to obtain a fused feature matrix, ensuring that each point feature retains the global geometric structure of the local details and shape. Among them, the local details refer to the geometric features and shape features within a small range of each point in the point cloud.

[0082] Step 3052: The fused feature matrix obtained after superposition is further fed to three attention feature enhancement layers (Attention Feature Augmenttation1, Attention Feature Augmenttation2, Attention Feature Augmenttation3) in cascade to gradually upsample point features and capture the details and global relationships in the point cloud data. Each module dynamically weights the features through the self-attention mechanism to enhance important information and suppress irrelevant interference, and finally the attention feature can be obtained.

[0083] The specific implementation of step 3052 may include the following steps:

[0084] Step 30521: Initialize query, key, and value matrices (Q, K, V);

[0085] Step 30522: Calculate the score of each attention head. For each head, the dot product of the query matrix Q and the key matrix K is calculated as the attention score. The specific mathematical expression is:

[0086]

[0087] Among them, d k / h is a scaling factor to prevent the dot product value from being too large.

[0088] Apply the softmax function to the attention score to get the attention weight, which determines the importance of each value in the weighted sum. Then, use the obtained attention weight to weight the value to get the output of each head; the outputs of all attention heads will be concatenated to form a new feature representation, and a linear transformation will be used to generate the final multi-head attention output.

[0089] Step 30523: Based on the multi-head attention output, use the residual connection to add input features and pass the original input information directly to the next layer to avoid the gradient vanishing problem and help optimize the model. The residual result calculation is as follows:

[0090] F residual =Norm(Q+MultiHead(Q,K,V))

[0091] Where Q = XW Qu ,K=XW Ku ,V=XW Vu ,X is the input feature matrix, W is the weight matrix, and Qu, Ku, and Vu are used to calculate the weight matrices of query, key, and value.

[0092] Step 30524: The residual result is input into the feedforward neural network. The feedforward neural network consists of two linear transformations and a RELU activation function. The jump connection technology is used to enable the model to learn representation more effectively. The specific form of the feedforward neural network is as follows:

[0093] FFN(x)=max(0,xW1+b1)W2+b2

[0094] Among them, W1, W2, b1, b2 are the parameters of the feedforward neural network, and x is the input of the feedforward neural network.

[0095] The specific implementation method of step 30524 is as follows:

[0096] Step 305241: The input point cloud feature vector is saved for use in subsequent jump connections to ensure that the original point cloud information is not lost during network processing.

[0097] Step 305242: Perform layer normalization to adjust the mean and variance of the input to make it more suitable for subsequent network processing, thereby improving the stability of the model and the training speed.

[0098] Step 305243: The normalized tensor is mapped to a higher dimensional space through linear mapping to capture more feature information.

[0099] Step 305244: Apply the GELU (Gaussian Error Linear Unit) activation function to the tensor mapped to the high-dimensional space to enhance the ability to express nonlinear features.

[0100] Step 305245: Map the activated high-dimensional feature vector back to the original feature dimension through linear mapping.

[0101] Step 305246: Perform a skip connection (Residual Connection) on the result obtained in the previous step and the previously saved input point cloud feature vector. Skip connection can alleviate the gradient vanishing problem and help retain the original input information, and finally return the addition result as the output of the feedforward neural network.

[0102] Step 30525: After the feedforward network is processed, the residual connection is used again to add the output of the feedforward network to the normalized input to maintain the flow of information; then, layer normalization is applied to the summed result to help stabilize the training process and reduce the risk of gradient vanishing or exploding.

[0103] F output =LayerNorm(F FFN +F residual )

[0104] Step 30526: Through the back-propagation process, the weights of the query and key are updated according to the attention weight matrix, and finally the attention feature is obtained.

[0105] Step 3053: Apply the third multi-layer perceptron layer (MLP3) to the integrated and superimposed features of the multi-level features and the output features of the attention enhancement layer (ie, the attention features) to generate the final predicted complete point cloud, thereby achieving the completion and refinement of the point cloud information.

[0106] Step 306: aligning the aerospace component to be assembled according to the complete point cloud.

[0107] Step 306 may specifically include the following steps:

[0108] Step 3061: Calculate the curvature of the complete point cloud, obtain the curvature value in the neighborhood of each point, obtain the local shape information of the point cloud, and then construct a curvature map of the point cloud, and output the point cloud edge curvature similarity results between different aerospace components to be assembled. The specific process is as follows:

[0109] Step 30611: Calculating the normal vector of each point in the complete point cloud is a prerequisite for curvature calculation. The normal vector of each point is estimated using the PCA (Principal Component Analysis) method through the covariance matrix of the neighborhood points.

[0110] Step 30612: Select a neighborhood for each point and use the K nearest neighbor method (the number of neighborhood points is K) to calculate the average degree of change of the neighborhood surface normal vector, which is suitable for describing the curvature of the surface.

[0111] Step 30613: Construct a curvature map for the point cloud according to the curvature value, display the curvature characteristics of each point in the point cloud, and reflect the local shape changes of the point cloud.

[0112] Step 30614: Use the Euclidean distance measurement method to quantify the similarity of the curvature distribution, calculate the sum of the squares of the curvature value differences between each pair of corresponding point clouds (the corresponding points are on the two aerospace components to be assembled), and obtain the total similarity score. The final edge curvature similarity calculation formula is as follows:

[0113]

[0114] Among them, C1 and C2 represent the curvature values ​​of different point clouds, respectively, and these two point clouds are located on two aerospace components to be assembled.

[0115] Step 3062: Extract contours using the geometric features of the complete point cloud data to evaluate the matching degree of the contours of the point clouds between the different aerospace components to be assembled. The specific process is as follows:

[0116] Step 30621: Project the point cloud onto a two-dimensional plane and select the XY plane as the best projection plane. The two-dimensional representation after projection can help simplify processing and highlight the contour information of the point cloud.

[0117] Step 30622: On the projected point cloud, edge detection technology is used to extract contours and represent them as a set of points with important geometric information.

[0118] Step 30623: Based on the calculation principle of Hu moment, by comparing the shape features of the contours, a similarity value is returned. The smaller the value, the more similar the contours are. The contour matching degree between the two point clouds is quantitatively evaluated, thereby achieving accurate point cloud registration.

[0119] Step 3063: adopt a multi-feature-based matching strategy, comprehensively consider the edge curvature similarity and contour matching features, and obtain a multi-feature registration result by weighted fusion of different features, so as to improve the matching accuracy.

[0120] The multi-feature registration result can be the registration accuracy of two aerospace components to be assembled. The specific calculation formula can be:

[0121] (Edge curvature similarity × weight of edge curvature similarity + contour matching × weight of contour matching) × 100%

[0122] For example, if the focus is on the matching degree of the contour of the aerospace component, the weight of the edge curvature similarity and the weight of the contour matching can be set to 0.3:0.7, and finally a multi-feature registration result can be obtained.

[0123] Step 3064: When the multi-feature registration result meets the preset conditions, the aerospace component is registered.

[0124] When the registration accuracy of the two aerospace components to be assembled is greater than the preset accuracy, it means that the two components meet the requirements for assembly and can be assembled together accurately. At this time, the two aerospace components can be registered.

[0125] The above-mentioned embodiment integrates the self-supervised learning framework and the attention enhancement module, innovatively solving the problem of missing point cloud data in large aerospace components, combining multiple features including edge curvature and contour matching, and realizing multi-feature benchmark registration in a weighted fusion manner, which significantly improves the accuracy of registration. At the same time, reasonable module design and optimization reduce the computational complexity, making the present application more efficient when processing large-scale point cloud data.

[0126] In several embodiments provided in the present application, the disclosed devices and methods can also be implemented in other ways. The device and method embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and a part of a module, a program segment or a code contains one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or the flowchart, and the combination of boxes in the block diagram and / or the flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0127] In addition, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0128] If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program codes.

Claims

1. A method for registering aerospace components, characterized in that: The following steps are involved: Preprocessing the original point cloud of the aerospace component to obtain a preprocessed point cloud; Inputting the preprocessed point cloud into an encoder and a decoder to obtain a rough point cloud; Inputting the preprocessed point cloud and the rough point cloud into a generator to obtain a complete point cloud; registering the aerospace component according to the complete point cloud; The original point cloud is a residual point cloud, and the generator includes at least one attention feature enhancement layer.

2. The aerospace component registration method according to claim 1, characterized in that: The preprocessing of the original point cloud of the aerospace component to obtain the preprocessed point cloud includes: Performing farthest point sampling on the original point cloud to obtain a sampled point cloud; Normalizing the sampled point cloud to obtain a normalized point cloud; The normalized point cloud is divided into regions to obtain a preprocessed point cloud.

3. The aerospace component registration method according to claim 2, characterized in that: The farthest point sampling of the original point cloud to obtain a sampled point cloud includes: One or more point clouds are randomly selected from the original point cloud as an initial sampling point cloud set, and the remaining original point clouds are used as an initial unsampled point cloud set; For each point cloud in the current unsampled point cloud set, calculate the minimum distance from the unsampled point cloud to all point clouds in the current sampled point cloud set, and add the unsampled point cloud corresponding to the maximum value of the minimum distance to the current sampled point cloud set; When the number of point clouds in the current sampling point cloud set reaches a preset threshold, the point clouds in the current sampling point cloud set are used as the sampling point clouds.

4. The aerospace component registration method according to claim 2, characterized in that: The step of dividing the normalized point cloud into regions to obtain a preprocessed point cloud includes: According to the normalized point cloud, randomly initialize M center points; Extracting neighborhood regions for the M center points based on a K-nearest neighbor algorithm, and dividing the normalized point cloud into M parts; For each of the M parts, split it into a first part, a second part and a third part; Using the first part, the second part and the third part as the pre-processed point cloud; The first part is an area observable by the encoder, the second part is a mask area, and the third part is an area unobservable by the encoder.

5. The aerospace component registration method according to claim 1, characterized in that: The step of inputting the preprocessed point cloud into an encoder and a decoder to obtain a rough point cloud comprises: Input the preprocessed point cloud into an encoder to extract important feature information of the point cloud; The important feature information of the point cloud is input into the decoder to obtain the rough point cloud.

6. The aerospace component registration method according to claim 5, characterized in that: The encoder includes a first shared multi-layer perceptron layer, a first maximum pooling layer, an expansion layer, a second shared multi-layer perceptron layer, and a second maximum pooling layer; The step of inputting the preprocessed point cloud into an encoder and extracting important feature information of the point cloud comprises: Input the preprocessed point cloud into the first shared multi-layer perceptron layer, perform linear transformation and nonlinear activation on each point cloud, and extract shallow features; Input the shallow features into the first maximum pooling layer, extract the maximum value of each point cloud in the feature dimension, and obtain significant features; Inputting the significant features into an extended layer to obtain extended features; wherein the extended features have the same feature dimension as the shallow features; Inputting the shallow features and the extended features into a second shared multi-layer perceptron layer to obtain comprehensive features; The comprehensive features are input into the second maximum pooling layer to extract important feature information of the point cloud.

7. The aerospace component registration method according to claim 6, characterized in that: The encoder includes a third shared multi-layer perceptron layer and a reshaping layer; The step of inputting the important feature information of the point cloud into a decoder to obtain the rough point cloud comprises: Inputting the important feature information of the point cloud into a third shared multi-layer perceptron layer to extract decoding features; Inputting the decoded features into a reshaping layer so as to convert the decoded features into features of a preset shape and format to obtain the rough point cloud; The rough point cloud at least includes the main geometric features of the original point cloud.

8. The aerospace component registration method according to claim 6, characterized in that: The generator includes: a first multi-layer perceptron layer, a second multi-layer perceptron layer, three attention feature enhancement layers and a third multi-layer perceptron layer; The step of inputting the preprocessed point cloud and the rough point cloud into a generator to obtain a complete point cloud comprises: Input the preprocessed point cloud and the rough point cloud together into a first multi-layer perceptron layer to extract first multi-level features; Inputting the important feature information of the point cloud into a second multi-layer perceptron layer to extract second multi-level features; The first multi-level features and the second multi-level features are superimposed and sequentially input into three attention feature enhancement layers to obtain attention features; The multi-level features and attention features are superimposed and input into a third multi-layer perceptron layer to obtain the complete point cloud.

9. The aerospace component registration method according to claim 6, characterized in that: The registering the aerospace component according to the complete point cloud comprises: According to the complete point cloud, obtaining edge curvature similarity and contour matching of the aerospace component; According to the curvature similarity and contour matching, a multi-feature registration result of the aerospace component is obtained through weighted fusion; When the multi-feature registration result meets the preset conditions, the aerospace component is registered.

10. An electronic device, characterized in that: The electronic device comprises: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to execute the aerospace component registration method described in any one of claims 1-9.

Images