Coarse-to-fine 3D Reconstruction Method of Aeroengine Blades Based on Overlap Region Guidance

Abstract

The present invention belongs to the technical field of three-dimensional reconstruction of blades, and particularly relates to a coarse-to-fine three-dimensional reconstruction method of aero-engine blades guided by an overlapping area, including S100: Coarse registration; obtaining three groups of point cloud data under different fields of view, using Kpconv to downsample the point cloud data of multiple fields of view, extracting geometric features within the field of view by combining a geometric self-attention mechanism, capturing geometric consistency features between fields of view through a cross-attention mechanism, predicting the confidence of the overlapping area, and using a multi-layer perceptron to predict the global transformation vector to complete the preliminary alignment of multi-view point clouds; S200: Fine registration; based on the coarse registration result, using an overlapping attention mechanism to mine local features, combining the overlapping score to perform weighted fusion of the features, and gradually optimizing the transformation vector in combination with a weighted cyclic update strategy. The present invention realizes the high-precision reconstruction of the three-dimensional morphology of the blade through feature enhancement guided by the overlapping area and multi-view joint optimization.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional reconstruction of blades, and particularly relates to a method for three-dimensional reconstruction of aero-engine blades from coarse to fine guided by an overlapping area. Background Technique

[0002] As the core aerodynamic component of an aero-engine and a gas turbine, the complex spatial free-form surface configuration of the blade determines the energy conversion efficiency and operation reliability of the whole machine. Research shows that even a geometric deviation of only a few micrometers on the blade surface can lead to a significant decline in aerodynamic performance. Therefore, implementing high-precision three-dimensional shape detection during the manufacturing process plays a key role in controlling the geometric accuracy of the blade and ensuring the service performance of the engine.

[0003] Blade detection techniques are mainly divided into contact type and non-contact type; among them, the coordinate measuring machine (CMM) is a typical representative of contact measurement and has been widely used in the industrial field due to its excellent measurement accuracy. However, it has obvious limitations: First, the contact probe scanning efficiency is low, and the full-profile measurement of a single blade can take several hours; second, traditional CMM measurements mostly adopt a discrete point sampling strategy, and can only obtain two-dimensional contour data of a specific target measurement section, making it difficult to achieve a complete reconstruction of the three-dimensional shape; third, in post-processing operations such as blade repair, contact measurement is likely to cause secondary damage to the machined surface. The non-contact measurement technology based on optical scanning has attracted attention in the industry due to its high efficiency and full-field measurement advantages. Three-dimensional reconstruction is achieved through multi-viewpoint cloud registration, and its technical route usually includes system calibration, viewpoint planning, sensor positioning, point cloud registration and other links. Among them, point cloud registration is the most challenging problem. Some studies focus on introducing external calibration objects (such as standard balls, calibration plates, and gauge blocks, etc.), resulting in an increased error transfer chain.

[0004] Currently, there are also registration methods that use a laser scanning sensor instead of introducing an external calibration object; they are divided into rigid paired point cloud registration and multi-viewpoint cloud registration. The former is more commonly the iterative closest point ICP. ICP is often highly sensitive to initialization and falls into an incorrect local optimum, and the point cloud density of the data is often variable, resulting in no one-to-one hard correspondence relationship. The variant of ICP, GO-ICP, finds the optimal solution through global optimization, greatly increasing the computational complexity of the algorithm. Especially when dealing with large-scale data, the computational amount and time consumption will increase significantly. Another one is the coherent point drift CPD, which performs point cloud registration based on a statistical method by maximizing the likelihood function of the data and can better avoid falling into local minima in many cases, providing higher matching accuracy. However, the effect of CPD is better when the point cloud density is high, but for sparse point clouds, especially when the distribution of the point cloud is uneven, the matching effect of CPD may not be ideal.

[0005] Some end-to-end deep learning methods have gradually emerged with the proposal of the PointNet model that can handle the disorder of point clouds. For example, DCP uses a deep neural network to learn the feature descriptors of point clouds. Given the source point cloud and the target point cloud as inputs, it directly outputs a transformation matrix to align the source point cloud. Another excellent soft matching strategy, RPM, establishes an adaptive soft-to-hard point matching relationship between point clouds through the Skhorn algorithm, effectively solving the registration problem of partially overlapping point clouds. In addition, PREDATOR borrows KPConv as the backbone for encoding-decoder feature extraction, predicts the overlap scores of corresponding points between point clouds, and uses these scores as importance assignment weights, enabling the model to focus on the features of the overlapping regions and solve the 3D point cloud registration problem in the case of low overlap. However, these pairwise point cloud registration methods face a common problem: when registering multi-viewpoint clouds, errors accumulate due to successive rigid transformations. Summary of the Invention

[0006] The object of the present invention is to provide a coarse-to-fine three-dimensional reconstruction method for aeroengine blades guided by the overlapping region. This method extracts point cloud features, enhances the features, predicts the overlapping regions of adjacent viewpoints and uses them as auxiliary features for screening, and uses a multi-layer perceptron to predict the transformation vector for multi-view simultaneous registration, thereby realizing an accurate three-dimensional reconstruction method for the pressure surface of the blade.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A coarse-to-fine three-dimensional reconstruction method for aeroengine blades guided by the overlapping region, comprising the following steps:

[0009] S100: Coarse registration; Obtain cloud data under different fields of view, downsample the point cloud data of multiple fields of view using Kpconv, extract the geometric features within the field of view in combination with the geometric self-attention mechanism, capture the geometric consistency features between fields of view through the cross-attention mechanism, predict the overlapping region confidence, and use a multi-layer perceptron to predict the global transformation vector to complete the preliminary alignment of multi-viewpoint clouds;

[0010] S200: Fine registration; Based on the coarse registration result, use the overlapping attention mechanism to mine local features, perform weighted fusion on the features in combination with the overlap scores, and gradually optimize the transformation vector in combination with the weighted cyclic update strategy.

[0011] Further, the coarse registration specifically includes the following steps:

[0012] S101: Obtain point cloud data from multiple different fields of view , representing the viewpoints under the data coordinates; Perform downsampling operations using Kpconv to generate a super point cloud dataset and its corresponding feature matrix ;

[0013] S102: Mine the geometric features within the field of view using the geometric self-attention mechanism and output a geometric feature matrix ;

[0014] S103: Capture the geometric consistency features of adjacent fields of view using the cross-attention mechanism and achieve information interaction of features, then output a mixed feature matrix , and splice the mixed features of multiple fields of view along the point dimension to obtain a total mixed feature matrix , denotes splicing along the point dimension;

[0015] S104: Map the total mixed feature matrix extracted in S103 to an overlap score vector = , which identifies the confidence of the overlapping region, is a learnable non-linear function;

[0016] S105: Combine the total mixed feature matrix obtained in S103 and S104 and the overlap score vector through NN-uppersampling and linear layers connected to the Kpconv downsampling skip connection to output a feature matrix and an overlap score vector with the same resolution as the point cloud data ;

[0017] S106: Perform a symmetric operation on the feature matrix to output a global feature matrix , denotes max pooling operation along the point dimension, denotes repeating the vector along the point dimension times; Pass the global feature matrix through a multi-layer perceptron to predict the transformation vector for coarse registration;

[0018] The coordinate frames of the linear sensors and satisfy the following equation: , where, denotes the viewpoint coordinates in the coordinate system of the linear sensor under the first field of view; denotes the viewpoint coordinates in the coordinate system of the linear sensor under the nth field of view; denotes the transformation vector from the first measurement position to the th measurement position, denotes the viewpoint based on the rotation angle to Rotation matrix, represents matrix multiplication, translation vector and rotation angle can be obtained from the measurement system; H is the height of the cross-section of the target measurement of each blade from its bottom reference plane;

[0019] Apply the predicted to the point cloud data of multiple fields of view to transform it into the rotation coordinate system to achieve rough alignment.

[0020] Furthermore, the geometric feature matrix The feature in , represents the total number of points in the nth field of view, is the ith feature in the geometric feature matrix, is the feature matrix The jth feature in is the geometric weight coefficient, is the learnable mapping matrix of the value vector.

[0021] Furthermore, the hybrid feature matrix The feature in , represents the total number of points in the nth field of view, is the ith feature in the hybrid feature matrix, is the feature concatenated along the point dimension of the geometric feature matrices of two adjacent fields of view, is the hybrid weight coefficient; is the learnable mapping matrix of the value vector.

[0022] Furthermore, the fine registration specifically includes the following steps:

[0023] S201: Based on the roughly aligned point cloud data obtained from rough registration, successively pass through the geometric self-attention mechanism, cross-attention mechanism, and upsampling operation to output the per-point feature matrix ;

[0024] S202: Based on the per-point feature matrix and the overlap score vector mine the local feature matrix through the overlap attention mechanism;

[0025] S203: Perform a symmetric operation on the local feature matrix and pass it through a multi-layer perceptron to predict the transformation vector and its corresponding weight ;

[0026] S204: Gradually optimize the transformation vector and improve the prediction accuracy by using a weighted cyclic update strategy , represents the total number of iterations.

[0027] Furthermore, the transformation vector , represents the weight corresponding to the predicted transformation vector for each iteration, represents the element-wise product; satisfies that the sum of each dimension is 1, , represents the th dimension.

[0028] Furthermore, the loss function of the weighted cyclic update , where is the loss between the predicted transformation vector and the ground truth transformation vector; is the self-supervised loss function; is the cross-entropy loss function; is the penalty strength in the iterative allocation; is the weight coefficient of the self-supervised loss function, is the weight coefficient of the cross-entropy loss function.

[0029] Furthermore, the features in the local feature matrix , represents the total number of points in the nth field of view, is the ith feature in the local feature matrix, is the overlap weight coefficient, is the point-wise feature matrix the jth feature in; is the learnable mapping matrix of the value vector.

[0030] The present invention has the following beneficial effects:

[0031] (1) In the coarse registration stage, Kpconv is used to downsample the multi-viewpoint cloud, geometric features within the field of view are extracted by combining the geometric self-attention mechanism, geometric consistency features between fields of view are captured through the cross-attention mechanism, and the confidence of the overlapping area is predicted. The global transformation vector is predicted using a multi-layer perceptron to complete the preliminary alignment;

[0032] (2) In the fine registration stage, local features are mined based on the attention mechanism guided by the overlap score, and the transformation vector is gradually optimized by combining the weighted cyclic update strategy to reduce error accumulation;

[0033] (3) The loss function fuses the transformation vector distance, chamfer distance, and cross-entropy loss of the overlapping area to further improve the registration accuracy. Brief Description of the Drawings

[0034] Figure 1 It is a schematic flow diagram of the present invention.

[0035] Figure 2 It is a deviation result diagram at the target cross-section of the blade. Detailed Embodiments

[0036] As Figure 1 shown, a coarse-to-fine three-dimensional reconstruction method for aero-engine blades based on overlap region guidance provided in this embodiment includes the following steps:

[0037] S100: Coarse registration; Obtain point cloud data under different fields of view, downsample the point cloud data of multiple fields of view using Kpconv, extract geometric features within the field of view by combining geometric self-attention mechanism, capture geometric consistency features between fields of view through cross-attention mechanism, predict the confidence of the overlap region, and use a multi-layer perceptron to predict the global transformation vector to complete the preliminary alignment of multi-view point clouds;

[0038] S200: Fine registration; Based on the coarse registration result, use the overlap attention mechanism to mine local features, combine the overlap score to weight and fuse the features, combine the weighted cyclic update strategy (RUS) to gradually optimize the transformation vector, and reduce error accumulation through multiple rounds of iteration.

[0039] The specific steps of the coarse registration include the following:

[0040] S101: Obtain point cloud data from multiple different fields of view , and perform downsampling operation using Kpconv to generate a super point cloud data set and its corresponding feature matrix .

[0041] S102: In order to extract context information, use the geometric self-attention mechanism to mine geometric features within the field of view and output a geometric feature matrix ;

[0042] Among them, , is the i-th feature in the geometric feature matrix, represents the total number of points in the n-th field of view, n represents the n-th field of view, represents the feature matrix the j-th feature in, is the geometric weight coefficient, is a learnable mapping matrix of the value vector.

[0043] S103: Use the cross-attention mechanism to capture the geometric consistency features of adjacent two fields of view and realize information interaction of features, and output a mixed feature matrix The mixed features of multiple fields of view are concatenated in the point dimension to obtain the total mixed feature matrix , denotes concatenation in the point dimension;

[0044] Among them, , is the mixed feature matrix The i-th feature in represents the total number of points in the n-th field of view, and n represents the n-th field of view. is the concatenated feature of the geometric feature matrices of two adjacent fields of view in the point dimension, is the mixed weight coefficient; is the learnable mapping matrix of the value vector.

[0045] S104: Extract the context information of each key point, and map the total mixed feature matrix extracted in S103 to an overlap score vector = , identifying the confidence of the overlapping region, is a learnable non-linear function.

[0046] S105: Combine the total mixed feature matrix and the overlap score vector obtained in S103 and S104 through NN-upsampling and linear layers connected to the Kpconv downsampling skip connection to output a feature matrix with the same resolution as the point cloud data and an overlap score vector .

[0047] S106: Perform a symmetric operation on the feature matrix to obtain the global feature matrix , denotes the max pooling operation along the point dimension, denotes repeating the vector along the point dimension times; Pass the global feature matrix through a multi-layer perceptron to predict the transformation vector for coarse registration.

[0048] The coordinate frame of the linear sensor and (n = 2, 3... N) satisfy the following equation:

[0049] ,

[0050] In the formula, represents the viewpoint coordinates in the coordinate system of the linear sensor under the first field of view; Denote the viewpoint coordinates at the coordinates of the linear sensor under the nth field of view; Denote the transformation vector from the first measurement position to the th measurement position, Denote the viewpoint based on the rotation angle to rotation matrix, Denote matrix multiplication, the translation vector and the rotation angle can be obtained from the measurement system; H is the height of the cross-section (TMCSs) of each blade target measurement from its bottom reference plane.

[0051] Apply the predicted to the point cloud data of multiple fields of view to transform it into the rotation coordinate system so as to achieve rough alignment.

[0052] The fine registration specifically includes the following steps:

[0053] S201: Based on the roughly aligned point cloud data obtained from the rough registration, sequentially execute S101, S102, S103, S105, that is, sequentially pass through the geometric self-attention mechanism, the cross-attention mechanism, and the upsampling operation to output the point-by-point feature matrix .

[0054] S202: Based on the point-by-point feature matrix and the overlap fraction vector mine local features through the overlap attention mechanism ;

[0055] Among them, , is the ith feature in the local feature matrix, F(n) represents the total number of points in the nth field of view, is the overlap weight coefficient, is the point-by-point feature matrix the jth feature in, is the learnable mapping matrix of the value vector.

[0056] S203: Perform a symmetric operation on the local feature matrix and predict the transformation vector and its corresponding weight through a multi-layer perceptron.

[0057] S204: Adopt a weighted cyclic update strategy to gradually optimize the transformation vector and improve the prediction accuracy , represents the total number of iterations.

[0058] Transformation vector , represents the weight corresponding to each iteration of the predicted transformation vector , denotes the element-wise product. Satisfies that the sum of each dimension is 1, i.e.: , represents the th dimension.

[0059] In the initial stage, , the predicted transformation vector , after weighting, we get ; when the m-th iteration, and The transformation vector of each iteration after weighting ; As the loop updates, the sequence of weighted transformation vectors gradually converges to the ground truth transformation vector .

[0060] The loss function of weighted loop update , where is the loss between the predicted transformation vector and the ground truth transformation vector; is the self-supervised loss function; is the cross-entropy loss function; is the penalty strength in the iterative assignment; is the weight coefficient of the self-supervised loss function, is the weight coefficient of the cross-entropy loss function.

[0061] , is the weighted transformation vector, is the ground truth transformation vector, denotes the 2-norm.

[0062] , denotes selecting from the set minimum values; denotes the point cloud data corresponding at the m-th iteration ; denotes corresponding to the point cloud data set at the m-th iteration ; denotes the 2-norm. As a self-supervised loss function, it enables the model to still be fine-tuned with real data after training, greatly optimizing the reconstruction accuracy.

[0063] To enable the model to have the ability to predict overlapping regions, the cross-entropy loss function ; F(n) represents the number of points in the n-th field of view, is the overlapping confidence for the i-th iteration, is the true overlapping score label.

[0064] In this embodiment, three typical blades are selected as experimental objects, named Blade-1, Blade-2, and Blade-3 respectively, and two different types of data are collected. The first is the labeled data extracted from the numerical model as training samples, which are divided into a training set, a validation set, and a test set; the second is the measurement data obtained from the physical model through a four-axis measurement system. To evaluate the geometric accuracy of blade reconstruction, the CMM measurement results are used as a reference to calculate the relative deviation. CMM is widely used in the blade manufacturing process and is a high-precision contact measurement technology.

[0065] In this embodiment, a theoretical data set is constructed to optimize a set of learnable parameters in the proposed learning framework. This process is divided into three stages. First, the numerical model is imported into Geomagic Studio, and the reference A is aligned to the XOY plane of the system coordinate system. Then, the model is loaded into Geomagic Control for discretization. Third, by using Python to simulate the real-world acquisition process, theoretical samples are obtained. In the second stage, the blade profile is sliced along the Z direction with a slice thickness of dz = 0.06 mm, and then discretized with a sampling interval of d = 0.01 mm to obtain point cloud data. According to the measurement standard of the blade profile, special attention is paid to the blade pressure surface (TMCSs) during the scanning path planning process to ensure the integrity of the data. Therefore, in the third stage, only the adjacent areas of these TMCSs are extracted, and the height from their bottom reference plane is shown in Table 1. In addition, by rotating around the Z axis, the point cloud is adjusted to simulate the initial posture of the blade in actual measurement. Then, according to the scanning path ( and ) are transformed in sequence, where and are sampled in the ranges of [-2°, 2°] and [-0.2 mm, 0.2 mm] respectively. Next, the point cloud is sparsely sampled and divided into perspectives according to the parameters of the LSS line, including the sampling point interval (dx), the field of view (FOV), and MR. Finally, the farthest point sampling (FPS) method is used to downsample the perspective to obtain samples with 2048 points per perspective. Finally, 20,000 samples are generated for this blade, of which 16,000 samples are randomly selected for training, 2,000 for validation, and the remaining are used for testing.

[0066] Table 1 Heights of the cross-sections (TMCSs) of the target measurements of each blade from its bottom reference plane

[0067]

[0068] Since the entire contour of the blade is ensured by measuring specific target cross-sections, based on the reconstruction results of all 3 algorithms, the target cross-section contour is extracted and compared with the CMM measurement results for accuracy evaluation, and then the comparison is made graphically. The deviation results at the blade target cross-section are shown, as Figure 2 shown, the deviation of this embodiment is less than that of all other algorithms. This means that the accuracy and robustness of this embodiment are superior when performing blade contour registration.

[0069] In addition to presenting the results in the form of deviation maps, in order to comprehensively and quantitatively evaluate the performance of this embodiment, the mean and standard deviation are introduced in this embodiment; as shown in Table 2, this embodiment achieves the lowest mean and standard deviation on all target cross-sections, indicating that its excellent performance and robustness are superior to other methods. In other words, this embodiment is in good agreement with the CMM measurement results, and the conclusion that this embodiment has high measurement accuracy.

[0070] Table 2: Evaluation results of evaluation parameters

[0071]

[0072] The above are only the preferred embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any modification and replacement based on the technical solutions and inventive concepts provided by the present invention should be covered within the protection scope of the present invention.

Claims

1. A coarse-to-fine three-dimensional reconstruction method for aeroengine blades based on overlapping area guidance, characterized in that It includes the following steps: S100: Coarse registration; Obtain point cloud data under different fields of view, downsample the point cloud data of multiple fields of view using Kpconv, extract in-field geometric features by combining the geometric self-attention mechanism, capture inter-field geometric consistency features through the cross-attention mechanism, predict the confidence of the overlapping area, and use a multi-layer perceptron to predict the global transformation vector to complete the initial alignment of multi-view point clouds; S200: Fine registration; Based on the coarse registration result, use the overlapping attention mechanism to mine local features, combine the overlapping score to perform weighted fusion of the features, and gradually optimize the transformation vector in combination with the weighted cyclic update strategy; Among them, the transformation vector , represents the weight corresponding to the predicted transformation vector for each iteration, represents the element-wise product; satisfies that the sum of each dimension is 1, that is: , represents the th dimension; represents the total number of iterations; Loss function with weighted cyclic update , where is the loss between the predicted transformation vector and the ground truth transformation vector; is the self-supervised loss function; is the cross-entropy loss function; is the penalty strength in iterative assignment; is the weight coefficient of the self-supervised loss function, is the weight coefficient of the cross-entropy loss function.

2. The method for three-dimensional reconstruction of aeroengine blades from coarse to fine based on overlapping area guidance according to claim 1, wherein The specific steps of the coarse registration are as follows: S101: Obtain point cloud data from multiple different fields of view , indicating the viewpoints under the data coordinates; Use Kpconv for downsampling operations to generate a super point cloud dataset and its corresponding feature matrix ; n represents the nth field of view; S102: Employ a geometric self-attention mechanism to mine geometric features within the field of view and output a geometric feature matrix ; S103: The cross-attention mechanism is adopted to capture the geometric consistency features of two adjacent fields of view and realize the information interaction of the features, and a mixed feature matrix is output , and the mixed features of multiple fields of view are concatenated in the point dimension to obtain the total mixed feature matrix , denotes concatenation along the point dimension; S104: Map the total mixed feature matrix extracted in S103 to an overlapping fraction vector = , which identifies the confidence of the overlapping region, and is a learnable non-linear function; S105: Obtain the total mixed feature matrix obtained in S103 and S104 and the overlap score vector Combine with the NN-upper sampling and linear layer connected by the Kpconv downsampling skip connection to output a feature matrix with the same resolution as the point cloud data and the overlap score vector ; S106: Perform a symmetry operation on the feature matrix to output the global feature matrix , denotes performing a max pooling operation along the point dimension, denotes repeating the vector along the point dimension times; Pass the global feature matrix through a multi-layer perceptron to predict the transformation vector for coarse registration; Coordinate Frame of Linear Sensor and satisfy the following equation: , where represents the viewpoint coordinates in the coordinates of the linear sensor under the first field of view; represents the viewpoint coordinates in the coordinates of the linear sensor under the nth field of view; represents the transformation vector from the first measurement position to the th measurement position, represents the rotation matrix of the viewpoint to to based on the rotation angle represents matrix multiplication, and the translation vector and the rotation angle can be obtained from the measurement system; H is the height of the cross-section of each blade target measurement from its bottom reference plane. Apply the predicted to the point cloud data of multiple fields of view to transform it into a rotating coordinate system so as to achieve rough alignment.

3. The method for three-dimensional reconstruction of aero-engine blades from coarse to fine based on overlapping region guidance according to claim 2, wherein The geometric feature matrix The feature , represents the total number of points in the nth field of view, is the ith feature in the geometric feature matrix, is the feature matrix the jth feature in, is the geometric weight coefficient, is the learnable mapping matrix of the value vector.

4. The method for three-dimensional reconstruction of aero-engine blades from coarse to fine based on overlapping area guidance according to claim 2, wherein The mixed feature matrix The features in , represent the total number of points in the nth field of view, is the ith feature in the mixed feature matrix, is the concatenated feature of the geometric feature matrices of two adjacent fields of view along the point dimension, is the mixed weight coefficient; is the learnable mapping matrix of the value vector.

5. The method for three-dimensional reconstruction of aeroengine blades from coarse to fine based on overlapping area guidance according to claim 1, characterized in that The specific steps of the fine registration are as follows: S201: Based on the roughly aligned point cloud data obtained from rough registration, it successively passes through a geometric self-attention mechanism, a cross-attention mechanism, and an upsampling operation to output a point-by-point feature matrix ; S202: Based on the point-by-point feature matrix and the overlap score vector mine the local feature matrix through the overlap attention mechanism ; S203: Perform a symmetry operation on the local feature matrix and pass it through a multi-layer perceptron for predicting the transformation vector and its corresponding weight ; S204: Gradually optimize the transformation vector using a weighted cyclic update strategy and improve the prediction accuracy , represents the total number of iterations.

6. The method for three-dimensional reconstruction of aero-engine blades from coarse to fine based on overlapping region guidance according to claim 5, wherein The local feature matrix The feature , represents the total number of points in the nth field of view, is the ith feature in the local feature matrix, is the overlapping weight coefficient, is the point-by-point feature matrix the jth feature in is the learnable mapping matrix of the value vector.

Images