Blade profile optical detection method, system, device and terminal for point cloud registration

Abstract

This invention belongs to the field of optical detection technology for blade contours. It discloses a method, system, device, and terminal for point cloud registration of blade contours. A deep feature interaction network (DFInet) is constructed for fine registration of multi-field-of-view data. A dual-branch structure is designed in the DFInet to integrate global and local features to encode point cloud features. A feature interaction module is designed to enhance the information correlation between two point clouds during feature extraction. An attention mechanism is used to fuse the matching information between two matching matrices based on global and local features. This invention's novel deep learning-based point cloud registration algorithm reduces the impact of geometric accuracy and motion stability on the developed multi-field-of-view system, improves the registration accuracy of blade contours, and strengthens feature communication between the source and target point clouds using the feature interaction module. Experimental results on labeled and measured data demonstrate the effectiveness of the method.

Description

Technical Field

[0001] This invention belongs to the field of optical detection technology for blade profiles, and particularly relates to a method, system, device and terminal for optical detection of blade profiles by point cloud registration. Background Technology

[0002] Currently, blades are critical components of aero-engines, determining their energy conversion efficiency, operational safety, and service life. The manufacturing process of blades is exceptionally complex, requiring high dimensional accuracy and surface integrity. Therefore, accurate measurement of blade profiles is crucial for ensuring machining quality and operational reliability. Currently, blade profile inspection methods based on coordinate measuring machines (CMMs) are widely used due to their high accuracy. However, this method is a contact measurement and has inherent limitations when measuring surfaces with high curvature variations due to the influence of contact radius compensation. Furthermore, the complexity and low efficiency of the measurement process conflict with agile manufacturing. Therefore, many researchers have dedicated themselves to studying optical measurement methods to overcome these problems.

[0003] Currently, optical measurement methods for blade profiles are generally classified into point source methods, line source methods, and area source methods, based on the type of light source and the corresponding data acquisition method. The implementation of systems based on optical detection methods involves several common steps, including field-of-view planning, sensor positioning, and blade profile registration. Among these, blade profile registration is the most challenging problem. Current research on blade profile registration methods mainly focuses on introducing external calibration objects, such as standard spheres, calibration plates, and gauge blocks. To our knowledge, Wang et al. first proposed using a laser sensor based on the blade's self-features to achieve blade profile measurement without introducing external calibration objects, thereby improving registration accuracy and reducing accumulated errors. However, these methods all share a common problem: the measurement accuracy is still affected by the geometric accuracy and motion stability of the developed multi-field-of-view system.

[0004] The most representative point cloud registration method is Iterative Closest Point (ICP), which is considered the standard method for solving point cloud registration. Subsequently, many variants have been proposed based on the basic concepts of ICP. However, ICP-type methods employ a one-to-one hard correspondence and, due to non-convexity, are prone to getting trapped in local minima. To achieve a "soft assignment" of correspondence between two point clouds, researchers have proposed probabilistic models to improve adaptability to uncertain data. Among these, Coherent Point Drift (CPD) is the most well-known method. However, probabilistic models focus more on the correspondence between points and ignore the geometric features of the point cloud. Therefore, feature-based methods have been designed to solve the point cloud registration problem. Generally, feature-based methods can be divided into two categories: global feature-based methods and local feature-based methods, the main difference being the features they use. Global features describe the shape of the entire model but cannot fully express the features of overlapping regions. Therefore, these algorithms have limited recognition capabilities and are only suitable for extracting features from regions with prominent features. Local features only encode neighboring points. For local regions that are almost planar or spherical, it is difficult to establish a local reference coordinate system. Therefore, for blades with special geometric features, these algorithms have limited ability to extract useful information for fine registration. Recently, Yang et al. reformulated point cloud registration as Trunceed Least Squares Estimation (TEASER). Although it is robust to erroneous correspondences between points, it still cannot solve the problem of extracting appropriate features.

[0005] In recent years, deep learning, capable of extracting abstract information from given data, has shown particular promise in point cloud registration. Elbaz et al. proposed an algorithm for 3D point cloud registration between large-scale point clouds and close-range scanned point clouds using a deep autoencoder neural network. Aoki et al. improved the Lucas & Kanade (LK) algorithm and extended PointNet and LK into a trainable deep neural network. Wang et al. proposed Deep Closest Point (DCP), which solves the coordinate transformation problem by calculating the matching probability of each point's features. Subsequently, Wang et al. designed a Partial Registration Network (PRNet) for local-to-local point cloud registration and used L2 regularization to obtain the keypoints for matching. Yew and Lee proposed Robust Point Matching (RPM), which uses differentiable sinkhorn layers and annealing parameters to obtain a "soft assignment" of correspondences between points. Li et al. proposed Iterative Distance-Aware Similarity Matrix Convolution (IDAM), which computes the matching matrix using a convolutional neural network. Zhu et al. used Representative Overlapping Points (ROPNet) with differential features for matching, transforming local-to-local matching into local-to-global matching. However, these works only focus on local or global features of the point cloud and cannot extract effective features from blades with thin walls and twisted spatial surfaces. Furthermore, these studies do not provide a clear method to associate information between the source and target point clouds; they only fuse features after feature extraction, failing to acquire sufficient preference knowledge.

[0006] Furthermore, designing an efficient and precise registration algorithm is even more challenging for blades with unique geometric features. First, to improve efficiency, the blade profile should be measured using as few fields of view as possible, ensuring that only a small portion of the data overlaps between two adjacent fields of view. Therefore, the blade profile registration problem becomes a local-to-local registration problem. Second, the thin-walled, twisted shape of the blade surface leads to a significant gap between local and global features. This increases the difficulty of extracting salient features containing both local and global information. Third, due to differences in the fields of view, the point density of the same part of the blade profile varies across different fields of view. Therefore, finding suitable matching points between measurement data from different fields of view is difficult, necessitating the design of a novel point cloud registration method and system for optical detection of blade profiles.

[0007] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0008] (1) The measurement accuracy of existing optical measurement methods for blade profiles is still affected by the geometric accuracy and motion stability of the developed multi-field system; ICP-type methods are prone to getting trapped in local minima; CPD methods based on probability models focus more on the correspondence between points and ignore the geometric features of the point cloud. TEASER reformulates point cloud registration as truncated least squares estimation, but it still cannot solve the problem of extracting appropriate features.

[0009] (2) Global feature-based methods cannot fully represent the features of overlapping regions and have limited recognition capabilities. They are only suitable for extracting features from regions with prominent features. Local feature-based methods only encode neighborhood points, and their ability to extract useful information for fine registration is limited for blades with special geometric features. At the same time, existing deep learning-based technologies only focus on the local or global features of point clouds and cannot extract effective features from blades with thin walls and twisted spatial surfaces.

[0010] (3) Existing technologies do not provide a clear method to associate information between source point clouds and target point clouds, but only fuse features after feature extraction, which is not enough time to obtain sufficient preference knowledge. Summary of the Invention

[0011] To address the problems existing in the prior art, this invention provides a method, system, device, and terminal for optical detection of blade contours based on point cloud registration, and particularly relates to a method, system, medium, device, and terminal for optical detection of blade contours based on point cloud registration using a deep feature interaction network.

[0012] This invention is implemented as follows: a method for optical detection of blade contours by point cloud registration includes:

[0013] A deep feature interaction network is constructed for fine registration of multi-view data. In the deep feature interaction network, a dual-branch structure is used to integrate global and local features and perform point cloud feature encoding. The feature interaction module is used to enhance the information association between the source point cloud and the target point cloud during the feature extraction process. The attention mechanism is used to fuse the matching information between two matching matrices based on global and local features.

[0014] Furthermore, the point cloud registration method for optical detection of blade profiles includes the following steps:

[0015] Step 1: In each iteration i, the source point cloud X undergoes a rigid transformation [R] obtained in the previous iteration i-1. i-1 ,T i -1The transformation is performed; the local feature extraction module extracts local features Fl from the transformed source point cloud X and the original target point cloud Y, respectively. X and Fl Y ;

[0016] Step 2: The global feature extraction module extracts global features Fg from the two point clouds. X and Fg Y The system uses two additional multilayer perceptrons with identical structures but different shared parameters to predict four parameters, including the exterior point parameter α. l and α g Annealing parameter β l and β g ;

[0017] Step 3, extracting features Fl X 、Fl Y 、Fg X 、Fg Y and parameter α l β l α g β g Used to calculate two matching matrices M respectively l and M g The information between the matching matrices in the local feature space and the global feature space is fused using an attention mechanism to obtain the final matching matrix M.

[0018] Step 4: From the source point cloud X and the matching average point Obtain rigid transformation [R] i ,T i ], and used for the next iteration i+1; wherein, the matching average point It is generated from the matching matrix M and the target point cloud Y.

[0019] Another objective of this invention is to provide a blade profile optical inspection system based on point cloud registration using the aforementioned point cloud registration method. The point cloud registration system consists of three translation axes X / Y / Z and one rotation axis R, used to acquire point cloud data of a specific blade cross section and simultaneously record motion parameters, including the linear displacements X / Y of the translation axes and the angle R of the rotation axis.

[0020] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the point cloud registration optical detection method for blade profile.

[0021] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the point cloud registration optical detection method for blade profile.

[0022] Another objective of this invention is to provide an information data processing terminal for implementing the aforementioned point cloud registration blade profile optical detection system.

[0023] Based on the above technical solutions and the technical problems solved, please analyze the advantages and positive effects of the technical solution to be protected by this invention from the following aspects:

[0024] This invention provides a Deep Feature Interaction Network (DFInet) for fine registration of multi-view data. The DFInet employs a dual-branch structure to integrate global and local features to encode point cloud features. Furthermore, a feature interaction module is proposed to enhance the information correlation between two point clouds during feature extraction. Next, an attention mechanism is used to fuse matching information between two matching matrices based on global and local features. Experimental results demonstrate the feasibility of the proposed method and its promising practical application prospects.

[0025] This invention proposes a novel deep learning-based point cloud registration algorithm to reduce the impact on the geometric accuracy and motion stability of the developed multi-view system. First, this invention proposes a blade contour point cloud registration algorithm based on local and global feature extraction, improving the registration accuracy of the blade contour. Furthermore, this invention designs a feature interaction module to enhance feature communication between the source and target point clouds. Based on this, this invention designs a simple yet effective attention mechanism module to fuse matching information between the matching matrices of the local and global feature spaces. Extensive experimental results on labeled and measured data demonstrate the effectiveness of the proposed method.

[0026] The learning-based blade profile registration algorithm provided by this invention, namely the Deep Feature Interaction Network (DFInet), mainly consists of four parts: Local Feature Extraction Module (LFE), Global Feature Extraction Module (GFE), Feature Interaction Module (FI), and Matching Matrix Generation Module (MMG).

[0027] Therefore, the main contributions of this invention can be summarized as follows:

[0028] (1) This invention proposes a novel blade profile registration method based on a deep learning system. The local feature extraction module focuses on the local geometric features in the neighborhood space (r-RN) of the sphere radius r, which is used to embed all points into the local feature space, while the global feature extraction module is designed in parallel and is used to embed all points into the global feature space.

[0029] (2) The present invention proposes a feature interaction module to promote feature interaction between source point cloud and target point cloud, thereby providing more discriminative features for subsequent modules.

[0030] (3) This invention proposes a simple and effective attention mechanism that can fuse information between matching matrices obtained from local feature space and global feature space.

[0031] The precision registration algorithm provided by this invention reduces the impact on the geometric accuracy and motion stability of the developed multi-field system, thereby further improving the accuracy of blade profile measurement. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of the blade profile optical detection method for point cloud registration provided in an embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram of the blade profile optical detection system for point cloud registration provided in an embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of the overall architecture of DFInet provided in an embodiment of the present invention;

[0036] Figure 4(a) shows the point coordinates provided in the embodiment of the present invention. (Below) Features of the calculation points (above); f θ (·) represents a series of fully connected layers used, and the learnable parameter θ is its associated weight;

[0037] Figure 4(b) shows the max-pooling operation provided in this embodiment of the invention; the output of Max-pooling is calculated by aggregating the features of all points connected to the center point;

[0038] Figure 5 This is a schematic diagram of the Feature Interaction Module (FI) in the network structure provided in an embodiment of the present invention;

[0039] Figure 6 These are schematic diagrams of three different blades provided in embodiments of the present invention;

[0040] Figure 7 (a) shows the three fields of view provided in the embodiments of the present invention, illustrating the overall overview of section 1 obtained by the detection system;

[0041] Figure 7 (b) is a schematic diagram of a coarse registration method based on the prior art provided in an embodiment of the present invention;

[0042] Figure 7 (c) is a schematic diagram of fine registration based on a matching algorithm provided in an embodiment of the present invention;

[0043] Figure 8 This is a schematic diagram showing the reconstructed blade cross section provided in an embodiment of the present invention and its comparison with the detection results of CMM. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0045] To address the problems existing in the prior art, the present invention provides a method, system, device and terminal for optical detection of blade profiles by point cloud registration. The present invention will be described in detail below with reference to the accompanying drawings.

[0046] I. Explanatory and Illustrative Embodiments. To enable those skilled in the art to fully understand how the present invention is specifically implemented, this section provides an explanatory and illustrative description of the embodiments described in the claims.

[0047] like Figure 1 As shown, the point cloud registration method for optical detection of blade contours provided in this embodiment of the invention includes:

[0048] S101, Construct the DFInet, a deep feature interaction network for fine registration of multi-view data;

[0049] S102, a dual-branch structure is designed in the deep feature interaction network to integrate global and local features to encode point cloud features;

[0050] S103, Design a feature interaction module to enhance the information association between two point clouds during the feature extraction process;

[0051] S104 utilizes an attention mechanism to fuse matching information between two matching matrices based on global and local features.

[0052] To accurately measure the blade profile, this invention develops a four-axis inspection system, which consists of three translation axes (X / Y / Z) and one rotation axis (R), as follows: Figure 2 As shown. In existing coordinate measuring machine (CMM)-based measurement standards, the geometric accuracy of a blade is evaluated based on several industry-defined specific cross-sections. The measurement datum of the CMM is first determined by the datum plane A and the height S of the specific cross-section. Then, the CMM obtains the point cloud data of the blade profile through point-to-point contact measurement. Therefore, it is necessary to accurately calibrate the sensor and blade attitude and establish a measurement datum before data acquisition. Based on this, the developed system can accurately acquire point cloud data of specific blade cross-sections. Simultaneously, motion parameters (i.e., linear displacement of the translation axis (X / Y) and angle of the rotation axis (R)) can be recorded.

[0053] The deep feature interaction network provided in this embodiment of the invention includes:

[0054] To mitigate the impact of geometric accuracy and motion stability issues on multi-field-of-view four-axis detection systems, this invention will elaborate on the proposed point cloud fine registration algorithm in this section. The following symbols will be used in this invention. It is assumed that this invention represents the source point cloud. and target point cloud These points are collected from two different adjacent fields of view by a laser scanning sensor. n and m represent the number of points in the source and target point clouds, respectively. For each point, the dimension is 2. The goal of the point cloud registration algorithm is to recover the rigid transformation [R,T] that aligns X to Y.

[0055] Example

[0056] The overall architecture of DFInet proposed in this embodiment of the invention is as follows: Figure 3 As shown. DFInet mainly consists of four parts: Local Feature Extraction (LFE), Global Feature Extraction (GFE), Feature Interaction (FI), and Matching Matrix Generation (MMG). In each iteration i, the source point cloud X is first transformed by the rigid transformation [R] obtained in the previous iteration i-1. i-1 ,T i-1 The transformation is performed. Then, the local feature extraction module extracts local features Fl from the transformed source point cloud X and the original target point cloud Y, respectively. X and Fl Y Simultaneously, the global feature extraction module extracts global features Fg from the two point clouds. X and Fg Y The study uses two additional multilayer perceptrons (MLPs) with identical structures but different shared parameters to predict four parameters, namely the exterior parameter α. l and α g Annealing parameter β l and βg Extracted features (Fl) X 、Fl Y 、Fg X and Fg Y ) and parameters (α) l β l α g and β g These are used to calculate the two matching matrices M respectively. l and M g Then, an attention mechanism (AM) is used to fuse the information between the matching matrices in the local and global feature spaces to obtain the final matching matrix. Finally, the source point cloud X and the average matching points are analyzed. (Generated from the matching matrix M and the target point cloud Y) yields the rigid transformation [R] i ,T i This is used for the next iteration i+1. Table 1 provides a more detailed setup.

[0057] Table 1 Detailed settings for DFInet

[0058]

[0059]

[0060] C and K represent the number of channels and the number of k-nearest neighbors, respectively, and N represents the number of points in the source or target point cloud, i.e., n or m. For Conv1d and Conv2d, the kernel size is 1×1.

[0061] In this embodiment of the invention, the provided local feature extraction module includes:

[0062] Blade outline point cloud registration is a local-to-local registration problem. Therefore, focusing on local geometric features is crucial.

[0063] The present invention first introduces a local feature extraction branch, which embeds the measured point cloud X and Y into the metric feature space to extract the local geometric features of the points in vector form.

[0064] Next, the local feature extraction branch of the source point cloud X is described in this embodiment of the invention. The local feature extraction branch of the target point cloud Y is the same. Since K-NN is not robust under non-uniform sampling conditions, the convolution method of the present invention ensures robustness of point clouds with different densities in the neighborhood space of the sphere radius r. In particular, for a point x in the source point cloud X... i A circle of radius r with x i This point is the center of the circle. Point x i The neighborhood of x can be determined by the radius r. (Except for x) i The point inside the outer circle is x.i The nearest neighbor, represented as For efficient computation, the number of nearest neighbors of any point must be consistent. If the number of points in a point's neighborhood is less than k, a fixed point or a randomly sampled point is used as a supplement. Conversely, if the number of points in a point's neighborhood is more than k, k points are randomly sampled. Based on this, the present invention first considers each neighboring point... The original coordinates are input into a series of fully connected layers, as shown in Figure 4. The features in the l-th layer... Represented as:

[0065]

[0066] Among them, f θ :R 2 →R F It is a nonlinear function with a set of learnable parameters. Due to the invariance of point cloud arrangement, this invention uses a simple symmetric function to converge local geometric features:

[0067]

[0068] in, It is a max-pooling layer that can pool any number of unordered and irregular points into a single descriptor, as shown in Figure 4(b). Then, the features obtained from two adjacent point clouds in the l-th layer are... and Send it to the FI module for information exchange. Finally, send the descriptor. The data is sent to a one-dimensional convolutional layer for L2 normalization to obtain the local feature vector Fl. X .

[0069] In this embodiment of the invention, the provided global feature extraction module includes:

[0070] The unique geometric structure of point clouds is beneficial for extracting prominent features. However, the global information of the point cloud is still not considered. Therefore, this invention introduces a global feature extraction branch to supplement this information. First, for a point x in the source point cloud X... i The global feature extraction branch extracts the global features of a point from its coordinates, that is:

[0071]

[0072] In other words, this invention does not require defining an r-RN space to extract global features from individual points. Furthermore, by performing the remaining operations—FI and l2 normalization—based on the local feature extraction branch, this invention obtains the global feature vector Fg. X .

[0073] In this embodiment of the invention, the provided feature interaction module includes:

[0074] Pairwise registration of point clouds is a local-to-local problem, meaning that only features extracted from overlapping regions are conducive to pairwise registration. Therefore, enhancing the feature interaction between the source and target point clouds can enable the designed network to extract more effective features and better understand preference perception. Figure 5 As shown, the feature interaction module is introduced.

[0075] This invention proposes a novel feature operator for obtaining features from adjacent point clouds at the l-th layer. and This process involves exchanging rich information between the points. In this operation, firstly, max pooling is applied along the channel axes of the feature matrix to pool the features. The pooled features are then broadcast. Finally, the broadcast features of one point cloud are concatenated with the features of another point cloud at the same level. In short, at layer (l+1), the features in the source point cloud X can be updated as follows:

[0076]

[0077] Where max(·) represents finding the maximum value, broadcast(a,N) Z ) indicates that vector N is repeated along the channel direction axis. Z Next, N Z This represents the number of points in the point cloud Z.

[0078] In this embodiment of the invention, the provided matching matrix generation module includes:

[0079] Since it is difficult to find a one-to-one correspondence between points, this invention proposes a soft assignment method. To achieve a balance between local and global features, this invention proposes a novel attention mechanism to fuse matching information between matching matrices obtained from the local and global feature spaces.

[0080] Specifically, such as Figure 3 As shown. This invention first addresses M. l and M g These two matching matrices are concatenated. Then, the present invention extracts M. l and M g The maximum value for each row. Based on this, the present invention uses the Softmax operation to obtain a confidence score for the match, where a higher value means higher confidence. In particular, the present invention defines the operator.

[0081]

[0082] in This indicates a splicing operation. It is to take M l and M g For each row's maximum value, φ{·} is the Softmax function. Finally, this invention performs element-wise multiplication and matrix addition to obtain the final matching matrix M, i.e.:

[0083]

[0084] in This indicates element-wise multiplication.

[0085] In this embodiment of the invention, the provided singular value decomposition (SVD) and loss function include:

[0086] After obtaining the matching matrix M, this invention uses it to generate the average matching point in the target point cloud Y for each point in the source point cloud X:

[0087]

[0088] in, It is the average point cloud of the match. This invention can convert point clouds... Each point is considered as an ideal matching point of the point cloud X.

[0089] In this case, the pairwise registration problem can be described by minimizing the following error to compute its rigid transformation:

[0090]

[0091] This invention uses SVD to solve equation (8). In this part, the invention only needs to solve the 2×2 feature problem, which is easily solved using simple algorithms. Before proceeding to the next iteration i, [R] is used. i-1 ,T i-1 The source point cloud X will be moved to a new location.

[0092] The local feature extraction module and the global feature extraction module are neural networks with a set of learnable parameters, which must be optimized during the training phase. This invention uses the following loss function to minimize the distance of the overlapping portion between the source point cloud X and the target point cloud Y:

[0093]

[0094] Among them, [R g ,T g ] represents a real rigid transformation, [R p ,T p ] indicates the predicted rigid transformation.

[0095] II. Evidence of the Relevant Effects of the Embodiments. The embodiments of the present invention have achieved some positive effects during research and development or use, and indeed possess significant advantages compared to existing technologies. The following description, in conjunction with data, charts, and other materials from the experimental process, illustrates these advantages.

[0096] A. Experimental Design

[0097] To verify the effectiveness of the algorithm, experiments were conducted on labeled and measured data. Labeled data was obtained from the numerical model of the blade profile, and measured data was obtained from the 3D entity of the blade profile using the developed quadaxial detection system. The labeled data was divided into training, validation, and test sets. The proposed algorithm was compared with seven point cloud registration algorithms, including three traditional algorithms: ICP, CPD, and TEASER, and four deep learning-based algorithms: PointNetLK, DCP, IDAM, and ROPNet. Since the deep learning-based algorithms are data-driven, they were first trained on the training set and then validated on the validation set. Furthermore, to ensure the stability of the output of the learning-based algorithm in each training iteration, a random seed was used. The eight point cloud registration algorithms were compared on the test set, and the reconstruction results of the blade profile measurement data were also compared.

[0098] B. Labeled data and measurement data

[0099] To generate labeled data, this invention first extracts the entire target cross-sectional profile from the numerical model of the blade. Then, points are taken at 0.02mm intervals on the profile. During actual measurement, this invention randomly extracts 256 consecutive points from the overlapping portion as the source point cloud X, and the point cloud offset by ±16 points near X is used as the target point cloud, i.e., the target point cloud Y contains 272 consecutive points. In this way, the labeled data can simulate the overlapping portion of the real measurement data from two adjacent angles. Finally, this invention performs a random transformation on the target point cloud Y along any axis, with a rotation angle of [-3, 3] degrees and a translation of [-0.2, 0.2] mm. Repeating the above operations, a total of 72,000 samples are obtained, of which 48,000 samples are used as the training set, 12,000 samples as the validation set, and the remaining samples as the test set.

[0100] For actual measurement data, multiple original fields of view of the target cross-section are first acquired using a calibrated four-axis detection system. These fields of view should encompass the entire cross-section of the target, with the overlap between adjacent fields of view exceeding (272*0.02) mm. Then, coarse registration is performed on the entire cross-section data. The original field of view data is interpolated, and points are taken at 0.02 mm intervals to obtain new field of view data. Finally, this invention uses the overlap of the new adjacent fields of view as the source point cloud X and the target point cloud Y. A rigid transformation obtained from the new field of view data is applied to the original field of view data to obtain the complete contour.

[0101] C. Method comparison on the test set

[0102] First, experiments were conducted on a test set of labeled data. All algorithms were run on a desktop computer with an Intel i7-6700K CPU, an NVIDIA GTX 1080 GPU, and 32GB of RAM. This invention introduces mean squared error (MSE) and mean absolute error (MAE) to represent the difference between the predicted rigid transformation and the true rigid transformation. As shown in Table 2, the MSE and MAE of the proposed algorithm are both lower than those of other algorithms. This indicates that the proposed algorithm has higher registration performance compared to other algorithms. Furthermore, by analyzing the computation time of the algorithm, although the proposed algorithm is slightly slower than the DCP algorithm, its efficiency is still verified.

[0103] Table 2 shows the experimental results on the test set.

[0104]

[0105] D. Results of blade profile reconstruction

[0106] Considering the differences between labeled data and actual application scenarios, this invention compared the reconstruction results of actual measurement data. To ensure the reliability of the experiment, this invention selected three different types of blades as experimental subjects. Figure 6 As shown, the three target sections defined for these blades in industry are marked with gray lines.

[0107] This invention uses section 1 of blade 1 as an example to demonstrate the precise registration of overlapping portions in adjacent fields of view. For example... Figure 7 As shown in (a), the detection system obtained three fields of view encompassing the entire contour of section 1, V AB This represents the overlapping portion of field of view A and field of view B. Due to the fixed resolution of the laser's X-axis, different fields of view (V) of the same portion... AB and V BA The point densities are different. Figure 7(b) shows the coarse registration result of the point cloud for the entire cross-section. Due to errors in the geometric accuracy and motion stability of the developed multi-field-of-view system, registration errors will occur to some extent in the overlapping parts. This invention, based on Section B, extracts the overlapping parts of prominent features in adjacent fields of view, generates measurement data samples, and completes fine registration according to the point cloud registration algorithm, such as... Figure 7 As shown in (c).

[0108] Since CMM is a professional and high-precision contact inspection method, this invention uses the CMM detection results as the reference results for reconstruction. The reconstruction results of eight algorithms are matched with the reference results, and deviations are compared. Due to page limitations, this invention only displays the deviation comparison results for each blade section 1, such as... Figure 8 As shown in Table 3, DFInet exhibits significantly lower deviation than other methods. To quantitatively compare the degree of deviation between this method and the measurement results from a coordinate measuring machine (CMM), the present invention uses the mean, standard deviation, deviation range, and root mean square (RMS) of the deviation as four evaluation indicators. As shown in Table 3, the algorithm proposed in this invention demonstrates almost the best performance and is stronger than other algorithms.

[0109] Furthermore, ICP also exhibits good performance in certain situations. This is because the acquired data has a good initial position after coarse registration of the entire contour data. However, ICP employs a one-to-one correspondence search strategy and is prone to getting trapped in local minima. While CPD can perform soft assignment based on point correspondences, it ignores the features of the point cloud. TEASER, PointNetLK, DCP, IDAM, and ROPNet are all feature-based algorithms. They only focus on local or global features of the point cloud, without interacting with the information between the source and target point clouds. Therefore, they cannot effectively extract features for blades with thin walls or twisted surfaces.

[0110] Table 3 shows the deviation between the measurement results obtained by this method and the measurements obtained by CMM.

[0111]

[0112]

[0113]

[0114] The best results are shown in bold, and the second-best results are shown in both bold and underline.

[0115] E. Ablation test

[0116] In this section, ablation experiments were conducted on labeled data to verify the importance of each component in the proposed DFInet. The ablation studies tested four main components: LFE, GFE, FI, and AM. The experimental results are shown in Table 4, with "√" indicating the operation performed. Models 1 and 2 show that registration performance is unsatisfactory when using only LFE or GFE. Models 3 and 4 show that FI improves registration performance by facilitating information exchange between the source and target point clouds during feature extraction. Models 5 and 7 show that the features of LFE and GFE complement each other, which also improves registration performance. Furthermore, AM integrates information from the matching matrices obtained from the local and global feature spaces, achieving a balance between local and global features and further improving registration performance. In summary, the above analysis demonstrates the effectiveness of each component of DFInet.

[0117] Table 4 Ablation experiments conducted on the test set

[0118]

[0119]

[0120] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0121] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optical detection of blade profiles based on point cloud registration, characterized in that, The blade contour optical detection method for point cloud registration includes: constructing a deep feature interaction network for fine registration of multi-field data; integrating global and local features and performing point cloud feature encoding in the deep feature interaction network using a dual-branch structure; enhancing the information association between the source point cloud and the target point cloud during feature extraction using a feature interaction module; and fusing matching information between two matching matrices based on global and local features using an attention mechanism. The point cloud registration method for optical detection of blade profiles includes the following steps: Step 1, in each iteration i In the middle, Source Cloud X, through the last iteration i-1 The rigid transformation obtained The transformation is performed; the local feature extraction module extracts local features from the transformed source point cloud X and the original target point cloud Y, respectively. and ; Step 2: The global feature extraction module extracts global features from the two point clouds. and The system uses two additional multilayer perceptrons with identical structures but different shared parameters to predict four parameters, including outlier parameters. and Annealing parameters and ; Step 3, Extracted Features , , , and parameters , , , Used to calculate two matching matrices respectively and The attention mechanism is used to fuse information between the matching matrices in the local and global feature spaces to obtain the final matching matrix. ; Step 4: From the source point cloud X and the matching average point Obtain rigid transformation And used in the next iteration. i+1 Wherein, the matching average point From the matching matrix Generate the target point cloud Y; The local feature extraction in step one includes: A local feature extraction branch is introduced, and the point clouds X and Y are embedded into the metric feature space to extract the local geometric features of the points in vector form. Specifically, this includes: for a point in the source point cloud X A circle with radius r Point is the center of the circle; point The neighborhood is determined by the radius r; except The inner point of the outer circle is The nearest neighbor is represented as If the number of points in the neighborhood of a point is less than 100, then ... If a point's neighborhood contains more than 100 points, then a fixed point or a random sampling point will be used as a supplement; Then random sampling 1 point; 2 adjacent points The original coordinates are input into a series of fully connected layers, and in the th... l Features of the layer Represented as: ; in, : It is a nonlinear function with a set of learnable parameters; due to the invariance of point cloud arrangement, symmetric functions are used to converge local geometric features: ; in, : It is a max-pooling layer that aggregates any number of unordered and irregular points into a single descriptor; it will be in the first... l Features obtained from two adjacent point clouds in a layer and Send the descriptor to the feature interaction module for information exchange; Send to a one-dimensional convolutional layer for processing. Normalization to obtain local feature vectors ; For the target point cloud Y, the same method is used to obtain the local feature vector. .

2. The method for optical detection of blade contours by point cloud registration as described in claim 1, characterized in that, The global feature extraction in step two includes: Introducing a global feature extraction branch, for a point in the source point cloud X... The global feature extraction branch extracts the global features of a point from its coordinates: ; Feature interaction based on local feature extraction branches and Normalization to obtain global feature vectors For the target point cloud Y, the same method is used to obtain the global feature vector. ; Max pooling is applied along the channel axis of the feature matrix to pool the features; the pooled features are then broadcast; the broadcast features of one point cloud are concatenated with the features of another point cloud at the same level; in ( l+1） In the layer, the features in the source point cloud X are updated as follows: ; in, This indicates finding the maximum value. Indicates repetition on the channel direction axis vector Second-rate, Point cloud The number of points in the target point cloud Y is obtained by updating the features of the target point cloud Y using the same method. .

3. The method for optical detection of blade contours by point cloud registration as described in claim 1, characterized in that, The acquisition of the matching matrix in step three includes: right and These two matching matrices are concatenated to extract... and The maximum value in each row; using Softmax The operation yields a credibility score for the match, with higher scores indicating greater credibility. Define operators: ; in, This indicates a splicing operation. Is to take and Operations on the maximum value of each row, yes Softmax The function performs element-wise multiplication and matrix addition to obtain the final matching matrix M. ; in, This indicates element-wise multiplication.

4. The method for optical detection of blade contours by point cloud registration as described in claim 1, characterized in that, In step four, the singular value decomposition and loss function include: Use the matching matrix M to generate the average matching point in the target point cloud Y for each point in the source point cloud X: ; in, It is the average point cloud of the matching, which will... Each point is viewed as a point cloud. X ideal matching point; The pairwise registration problem is described by minimizing the following error to compute a rigid transformation: ; Solve the equations using SVD and proceed to the next iteration. i Previously, using Transform the source point cloud X to a new location; Minimize the distance of the overlapping portion between the source point cloud X and the target point cloud Y using the following loss function: ; in, Represents a real rigid transformation. This indicates a rigid transformation in the prediction.

5. A blade profile optical inspection system for point cloud registration using the blade profile optical inspection method for point cloud registration as described in any one of claims 1 to 4, characterized in that, The point cloud registration blade profile optical detection system consists of three translation axes X / Y / Z and one rotation axis R, used to acquire point cloud data of a specific blade cross section, and simultaneously record motion parameters, including the linear displacement X / Y of the translation axes and the angle R of the rotation axis.

6. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor performs the point cloud registration optical detection method for blade profiles according to any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the point cloud registration optical detection method for blade profiles according to any one of claims 1 to 4.

8. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the blade profile optical detection system with point cloud registration as described in claim 5.

Images