An intelligent analysis method for shape and position deviation of aircraft composite thin-walled components based on a double-branch network

Abstract

The application provides an intelligent analysis method for shape and position deviation of an aircraft composite thin-walled component based on a double-branch network, which comprises the following steps: taking a measured three-dimensional point cloud and a theoretical CAD model as cross-modal double-input sources, constructing a multi-source heterogeneous initial data set; constructing a twin feature extraction network, extracting high-dimensional geometric semantic features through physical position coding and a hierarchical attention mechanism; introducing an anti-occlusion dynamic confidence mask to construct a cross-attention fusion module, establishing a high-robustness soft corresponding mapping between the measurement and the numerical model; based on feature reverse propagation and an engineering tolerance-oriented decoding model, combining an asymmetric penalty loss to directly solve three-dimensional manufacturing deviation; constructing a global shape and position deviation field, generating a deviation heat map and an out-of-tolerance early warning, and outputting a detection report. The application realizes end-to-end high-precision quantification of the manufacturing deviation of the thin-walled component, effectively overcomes the analysis distortion caused by physical deformation and scanning blind area, and significantly improves the automation level and reliability of the quality detection of the aviation component.

Description

Technical Field

[0001] This invention relates to the field of digital inspection and intelligent analysis technology in aerospace manufacturing, and in particular to an intelligent analysis method for shape and position deviations of thin-walled composite material components for aircraft based on a dual-branch network. Background Technology

[0002] In modern aircraft manufacturing, after demolding and curing, thin-walled composite material components such as wing panels often exhibit non-rigid warping and springback due to anisotropic shrinkage and thermal stress release. To assess whether this manufacturing deformation meets assembly tolerance requirements, engineers typically need to perform spatial geometric comparison between the measured surface point cloud obtained by high-precision 3D scanning equipment and the original theoretical CAD design model, thereby quantifying the distribution of form and position deviations across the entire field.

[0003] Currently, the mainstream approach in the industry for comparing measured data with theoretical numerical models heavily relies on traditional rigid registration algorithms such as iterative nearest point. However, traditional rigid algorithms have fatal flaws when dealing with typical non-rigid warping deformations in composite materials. Because rigid registration emphasizes minimizing the overall Euclidean distance, it is extremely sensitive to the initial spatial pose. When a component exhibits large local deformations, traditional algorithms often incorrectly force alignment of the deformed regions, causing the model to get trapped in local minima in characteristic regions such as large-curvature bends. This erroneous forced fitting not only masks the true deformation amount but also generates distorted and unreliable false bias results.

[0004] Secondly, existing geometric alignment software is essentially based on calculations of purely physical spatial distances, and the network itself lacks a deep understanding of the high-dimensional geometric semantics of thin-walled components. When the measured point cloud does not completely match the theoretical numerical model due to scanning occlusion or local processing defects, traditional methods cannot identify and ignore these topological differences like human experts do; instead, they mechanically establish incorrect point-to-point mapping relationships. To correct these distorted mappings, inspectors have to spend a lot of time on manual intervention before alignment, such as manually removing scanning noise, manually selecting registration reference planes, or specifying feature alignment points. This fragmented, non-end-to-end processing mode results in extremely poor consistency of analysis results, which cannot meet the needs of the aerospace manufacturing industry for efficient and automated quality inspection of mass-produced parts.

[0005] Therefore, traditional purely geometric rigid registration and comparison techniques have revealed their limitations in addressing the challenges of topological inconsistencies and registration distortions caused by non-rigid deformation of thin-walled composite materials. These limitations stem from poor algorithm robustness and a high reliance on manual intervention. Consequently, there is an urgent need to propose a novel intelligent analysis method with high-dimensional semantic cognitive capabilities to overcome the end-to-end computational barrier that requires dual inputs of measured point clouds and theoretical numerical models to directly output high-fidelity 3D manufacturing deviations. Summary of the Invention

[0006] To address the problems in existing technologies, such as the tendency for registration of thin-walled composite components to get trapped in local minima, heavy reliance on manual intervention in the detection process, and inaccurate deviation calculations due to distortion in local feature matching, this invention provides an intelligent analysis method for form and position deviations of thin-walled composite components for aircraft based on a dual-branch network. This method extracts high-dimensional semantic features by constructing a twin network that includes parameter-free physical position encoding and hierarchical attention, and introduces a cross-attention mechanism to achieve cross-modal fusion and soft correspondence matching between measured point clouds and theoretical models. Finally, a multilayer perceptron is used to directly regress and output a global three-dimensional form and position deviation vector. This aims to achieve end-to-end high-precision quantification of manufacturing deviations in thin-walled components, overcome analysis distortions caused by physical deformation and scanning blind spots, and significantly improve the automation level and reliability of aerospace component quality inspection.

[0007] To achieve the above-mentioned technical objectives, the present invention provides the following technical solution:

[0008] A smart analysis method for form and position deviations of thin-walled composite components for aircraft based on a dual-branch network, specifically including:

[0009] S1. Obtain the measured point cloud of the composite thin-walled component, and obtain the theoretical point cloud that matches the measured point cloud through joint geometric sampling; perform spatial reference normalization on the measured point cloud and the theoretical point cloud to obtain the dual-branch initial dataset.

[0010] S2. Construct a twin high-dimensional semantic extraction network for the topological features of thin-walled components. Through parameterless physical location encoding and hierarchical adaptive attention mechanism, perform high-frequency geometric feature perception and high-dimensional semantic abstraction on the measured point cloud and theoretical point cloud after spatial benchmark normalization, respectively, and extract high-dimensional geometric semantic features.

[0011] S3. Construct a cross-modal cross-attention fusion module, introduce an anti-occlusion dynamic confidence mask, and perform cross-modal fusion of the high-dimensional geometric semantic features of the measured point cloud and the high-dimensional geometric semantic features of the theoretical point cloud to obtain cross-modal fusion features;

[0012] S4. Using a multi-layer feature backpropagation and engineering tolerance-oriented decoding model, spatial upsampling regression is performed on the cross-modal fusion features; an asymmetric penalty loss function combining process features is introduced to calculate the three-dimensional manufacturing deviation vector of each measured point;

[0013] S5. Construct a global form and position deviation field based on the three-dimensional manufacturing deviation vector, generate a continuous deviation heat map through chromatographic mapping, and combine semantic segmentation to provide early warning of out-of-tolerance in key high-frequency feature regions, and output a detection report of form and position deviation of composite thin-walled components.

[0014] Furthermore, step S1 specifically includes:

[0015] A three-dimensional scanning point cloud of the surface of a thin-walled composite component is collected, and then outlier removal and filtering noise reduction are performed on the three-dimensional scanning point cloud to obtain the measured point cloud.

[0016] A theoretical CAD model of a thin-walled component made of the same composite material is obtained. The surface of the theoretical CAD model is discretized using a uniform Poisson disk sampling algorithm to generate a theoretical point cloud that matches the scale of the measured point cloud.

[0017] An initial alignment strategy is adopted to coarsely register the measured point cloud and the theoretical point cloud to the same global reference coordinate system, which is an absolute coordinate system based on the design origin of the theoretical CAD model. Subsequently, spatial reference normalization is performed to calculate the geometric center of the coarsely registered point cloud and translate it to the origin of the global reference coordinate system. At the same time, global scaling is performed using the distance of the farthest point as the scaling factor, so that both the coarsely registered measured point cloud and the theoretical point cloud are confined within a unit sphere centered at the coordinate origin.

[0018] The normalized measured point cloud and the theoretical point cloud are reorganized by channel recombination and dimension alignment to obtain the initial dataset of the two branches.

[0019] Furthermore, step S2 specifically includes:

[0020] A twin high-dimensional semantic extraction network for topological features of thin-walled components is constructed. The twin feature extraction network consists of a parameter-free physical location encoding module, a sampling and local neighborhood construction module, an adaptive attention learning module, and a multi-level feature aggregation module connected in series.

[0021] The initial datasets of the two branches are input into the parameterless physical location encoding module. The parameterless physical location encoding operation is used to map the three-dimensional coordinates of each point in the normalized point cloud to a high-dimensional space to obtain high-dimensional encoded features.

[0022] In the sampling and local neighborhood construction module, the farthest point sampling algorithm is used to determine the center point, and a local geometric neighborhood is established for each center point based on the spatial distance.

[0023] In the adaptive attention learning module, a self-attention mechanism is introduced into the local geometric neighborhood of each center point. The self-attention weight of each neighboring point in each local geometric neighborhood is calculated relative to the center point through a learnable feature similarity scoring multilayer perceptron.

[0024] In the multi-level feature aggregation module, based on the self-attention weights between nodes, the high-dimensional encoded features of the neighboring points in each local geometric neighborhood are weighted and summed to aggregate the local geometric features and update the high-dimensional geometric semantic features of the center point.

[0025] The sampling and local neighborhood construction module, the adaptive attention learning module, and the multi-level feature aggregation module are treated as a downsampling layer. L downsampling layers are stacked, and the number of samples at the center point is reduced by forward propagation layer by layer, finally obtaining the high-dimensional geometric and semantic features of the measured point cloud / theoretical point cloud.

[0026] Furthermore, step S3 specifically includes:

[0027] The high-dimensional geometric and semantic features of the measured point cloud are mapped into a query vector through a linear projection matrix; the high-dimensional geometric and semantic features of the theoretical point cloud are mapped into a key vector and a value vector through a linear projection matrix, respectively.

[0028] A spatial occlusion perception subnetwork is constructed. The spatial occlusion perception subnetwork takes the high-dimensional geometric semantic features of the measured point cloud as input and maps the high-dimensional geometric semantic features of each center point into a scalar form of effective confidence distribution through a regression branch composed of a fully connected layer, a batch normalization layer and a nonlinear activation function.

[0029] Calculate the inner product of the query vector and the key vector and perform scaling to obtain the basic cross-modal similarity association matrix that represents the spatial correlation between the high-dimensional geometric semantic features of the measured point cloud and the high-dimensional geometric semantic features of the theoretical point cloud.

[0030] The matching weights of each node in the basic cross-modal similarity association matrix are gating and filtering using the effective confidence distribution values. After gating and filtering, the anti-occlusion mask attention weight matrix is ​​obtained. The anti-occlusion mask attention weights and the value vector are weighted and summed to obtain the cross-modal fusion features.

[0031] Furthermore, step S4 specifically includes:

[0032] The cross-modal fusion features are upsampled layer by layer using a multi-layer feature propagation module, and the feature resolution is restored to the scale of the original input measured point cloud based on the inverse distance weight.

[0033] The upsampled cross-modal fusion features are concatenated with the normalized 3D coordinates of each measured point and input to the multilayer perceptron decoder. The multilayer perceptron decoder adopts a dual-output head architecture, including a rigid parameter branch for regressing the global pose transformation and a deviation vector branch for regressing the displacement residuals of each sampling point.

[0034] During the training phase of the multilayer perceptron decoder, a process-oriented asymmetric penalty loss function is constructed by applying a penalty weight greater than that for the negative deviation of the composite thin-walled component and applying a geometrically weighted high penalty constraint to the assembly critical surface.

[0035] After forward propagation and optimization using the asymmetric penalty loss function, the two output heads of the multilayer perceptron decoder directly regress the rigid transformation parameters used to compensate for global spatial alignment errors, and the three-dimensional manufacturing deviation vector for each measured point used to quantify the physical deformation of the component.

[0036] Furthermore, step S5 specifically includes:

[0037] Based on the three-dimensional manufacturing deviation vector, the absolute Euclidean distance deviation value of each measured point relative to the theoretical surface is calculated;

[0038] A pseudo-color chromatogram mapping rule is set to map the absolute Euclidean distance deviation value to the corresponding color region, and the measured 3D model is rendered point by point to generate a deviation heatmap.

[0039] By combining semantic segmentation labels, the local deviation distribution of key high-frequency feature regions is extracted; the local deviation distribution is compared with the preset engineering tolerance threshold, and the geometric coordinates and maximum deviation extreme values ​​of the out-of-tolerance regions are automatically marked and extracted.

[0040] The system integrates statistical data on deviations across the entire domain, early warning information on out-of-tolerance conditions in key high-frequency characteristic areas, and visualization results of deviation heatmaps to output a detection report on the shape and position deviations of thin-walled composite material components.

[0041] Furthermore, this application also discloses an application system for the aforementioned intelligent analysis method of form and position deviations of thin-walled composite material components for aircraft based on a dual-branch network, which specifically includes:

[0042] The data preprocessing module is used to convert the measured point cloud and theoretical CAD model of the composite thin-walled component into a two-branch initial dataset with normalized spatial coordinates.

[0043] A twin feature extraction network is used to extract high-dimensional geometric and semantic features of the processed measured point cloud and theoretical point cloud from the dual-branch initial dataset;

[0044] The cross-attention fusion module introduces an anti-occlusion dynamic confidence mask to establish a soft correspondence space mapping between measured features and theoretical numerical model priors, thereby fusing two sets of high-dimensional geometric and semantic features into cross-modal fusion features.

[0045] The deviation regression decoding module is used to upsample and decode cross-modal fusion features, and after optimization by an asymmetric penalty loss function, it calculates the rigid transformation parameters and the three-dimensional manufacturing deviation vector of each measured point end-to-end.

[0046] The visualization and early warning module is used to construct a form and position deviation field based on the 3D manufacturing deviation vector, generate a deviation heat map, and combine semantic segmentation to provide out-of-tolerance early warning and output an inspection report.

[0047] An electronic device is also disclosed, comprising a memory and a processor, wherein:

[0048] Memory is used to store computer programs that can run on a processor;

[0049] The processor is used to execute, as described above, an intelligent analysis method for form and position deviations of thin-walled composite material components for aircraft based on a dual-branch network when running the computer program.

[0050] A computer-readable storage medium is also disclosed, which stores computer instructions for causing a processor to execute an intelligent analysis method for form and position deviations of thin-walled composite material components for aircraft, as described above, based on a dual-branch network.

[0051] Based on the above technical solution, the present invention has at least the following beneficial effects:

[0052] 1. This invention solves the problem that traditional rigid registration algorithms are prone to getting trapped in local minima and have high initial alignment dependence when processing features by constructing a dual-branch twin feature extraction network that includes parameterless physical location encoding and hierarchical attention mechanism. It achieves accurate capture of high-dimensional geometric semantic features of measured scanned point clouds and theoretical CAD digital models.

[0053] 2. This invention introduces a cross-attention mechanism to construct a cross-modal feature fusion module. By introducing an effective confidence distribution to calculate the feature space similarity, it blocks the forced alignment error caused by large-area scanning blind spots and establishes a high-precision "soft correspondence" mapping relationship between measured data and theoretical models. This effectively overcomes the problem of inconsistent local topology and registration distortion caused by measurement occlusion and manufacturing deformation coupling in composite materials, and significantly improves the feature matching robustness of the intersection area and thin-walled transition area.

[0054] 3. Based on feature propagation and multilayer perceptron decoding technology, this invention realizes an end-to-end intelligent regression closed loop from "dual data input of measured / theoretical data" to "direct output of three-dimensional deviation vector of each point";

[0055] The method proposed in this invention breaks through the limitations of traditional detection processes, which rely heavily on manual noise removal, manual selection of reference alignment, and low computational efficiency. It greatly improves the automation level and engineering quantification accuracy of full-domain manufacturing deviation detection for aerospace composite material parts. Attached Figure Description

[0056] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0057] Figure 1This is a flowchart of an intelligent analysis method for form and position deviations of thin-walled composite components for aircraft based on a dual-branch network, as proposed in this invention.

[0058] Figure 2 This is a schematic diagram of the dual-branch feature extraction and cross-modal fusion matching network architecture of measured point cloud and theoretical numerical model in an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of a full-domain three-dimensional manufacturing deviation heatmap and a key feature out-of-tolerance early warning system in an embodiment of the present invention. Detailed Implementation

[0060] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. This will allow for a full understanding of how technical means are applied to solve technical problems and achieve technical effects, and to facilitate implementation of the present application.

[0061] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0062] like Figure 1 As shown, this embodiment discloses an intelligent analysis method for form and position deviations of thin-walled composite components for aircraft based on a dual-branch network. Through dual-branch input data spatial normalization, high-dimensional semantic feature extraction via Siamese networks, cross-modal feature cross-attention fusion, and deviation direct regression calculation based on a multilayer perceptron, it achieves high-precision, end-to-end automatic quantification and intelligent analysis of the global spatial form and position deviations of thin-walled composite components for aircraft (such as wing panels and air intakes). The method proposed in this invention specifically includes:

[0063] S1. Obtain the measured point cloud of the composite thin-walled component, and obtain the theoretical point cloud that matches the measured point cloud through joint geometric sampling; perform spatial reference normalization on the measured point cloud and the theoretical point cloud to obtain the dual-branch initial dataset.

[0064] In a preferred embodiment, step S1 specifically includes:

[0065] A three-dimensional scanning point cloud of the surface of a thin-walled composite component is acquired. Then, outlier removal and filtering are performed on the three-dimensional scanning point cloud containing noise to obtain the measured point cloud.

[0066] A theoretical CAD model of a thin-walled component made of the same composite material is obtained. The surface of the theoretical CAD model is discretized using a uniform Poisson disk sampling algorithm to generate a theoretical point cloud that matches the scale of the measured point cloud.

[0067] An initial alignment strategy (such as principal component analysis or feature matching) is adopted to coarsely register the measured point cloud and the theoretical point cloud to the same global reference coordinate system, which is an absolute coordinate system based on the design origin of the theoretical CAD model. Subsequently, spatial reference normalization is performed, the geometric center of the coarsely registered point cloud is calculated and translated to the origin of the global reference coordinate system, and global scaling is performed using the distance of the farthest point as the scaling factor, so that both the coarsely registered measured point cloud and the theoretical point cloud are confined within a unit sphere centered at the coordinate origin.

[0068] The normalized measured point cloud and the theoretical point cloud are reorganized by channels and aligned by dimensions to obtain a two-branch initial dataset. This dataset is used to construct a two-branch initial tensor data structure that meets the input requirements of the subsequent twin feature extraction network, so that the subsequent network can extract spatial geometric features in parallel.

[0069] S2. Construct a twin high-dimensional semantic extraction network for topological features of thin-walled components. Through parameterless physical location encoding and hierarchical adaptive attention mechanism, refine the perception of high-frequency geometric features (such as large curvature transition and intersection region) and high-dimensional semantic abstraction of the measured point cloud and theoretical point cloud after spatial benchmark normalization, and extract high-dimensional geometric semantic features.

[0070] As a preferred embodiment, such as Figure 2 As shown, in this embodiment, step S2 specifically includes:

[0071] A twin high-dimensional semantic extraction network for topological features of thin-walled components is constructed. The twin feature extraction network consists of a parameter-free physical location encoding module, a sampling and local neighborhood construction module, an adaptive attention learning module, and a multi-level feature aggregation module connected in series.

[0072] The initial dataset from the two branches is input into the parameterless physical location encoding module. The parameterless physical location encoding operation maps the 3D coordinates of each point in the normalized point cloud to a high-dimensional space, obtaining high-dimensional encoded features while preserving the absolute 3D spatial position information of the components in the original scene. The formula is expressed as:

[0073] ;

[0074] in, Initialize the generated high-dimensional encoded feature vector for the i-th point. It is a parameterless spatial geometric mapping function; For the normalized point cloud The three-dimensional coordinates of each point; if a "parameterized" linear layer is used directly, random perturbations are likely to occur in the early stage of training, resulting in geometric distortion. Therefore, this application uses parameterless physical position encoding.

[0075] In the sampling and local neighborhood construction module, the farthest point sampling algorithm is used to determine the center point, and a local geometric neighborhood is established for each center point based on spatial distance; thus, a hierarchical attention learning framework containing multiple point set abstract attention modules is constructed; the mathematical definition of local geometric neighborhood is:

[0076] ;

[0077] in, Indicates The local geometric neighborhood of the center point This represents the Euclidean distance ascending sort function. Representing neighborhood points and The Euclidean distance between them This represents the number of neighboring points; The input is a point cloud (measured point cloud or theoretical point cloud);

[0078] In the adaptive attention learning module, a self-attention mechanism is introduced into the local geometric neighborhood of each center point. This adaptively evaluates and assigns higher feature weights to high-curvature regions (such as the intake lip) and connecting structures (such as rib junctions), suppressing redundant information in flat regions. A learnable feature similarity scoring multilayer perceptron is used to calculate the self-attention weights of each neighboring point relative to the center point within each local geometric neighborhood. The formula is expressed as:

[0079] ;

[0080] in, For the local geometric neighborhood of the th Neighboring points Relative to the center point Self-attention weights; Multilayer perceptron for scoring learnable feature similarity; , Neighboring points Center point High-dimensional encoded features; For Any point in the local geometric neighborhood of the center point. for High-dimensional encoded features;

[0081] In the multi-level feature aggregation module, based on the self-attention weights between nodes, the high-dimensional encoded features of the neighboring points in each local geometric neighborhood are weighted and summed to aggregate the local geometric features and update the high-dimensional geometric semantic features of the center point.

[0082] In this embodiment, the formula for the process of aggregating local geometric features is expressed as follows:

[0083] ;

[0084] in, No. High-dimensional geometric semantic features of the center point output by layer aggregation; It is a non-linear activation function. This is the learnable weight matrix used for feature mapping in this layer of the network. For the first Neighborhood points passed from layer High-dimensional encoded features;

[0085] The sampling and local neighborhood construction module, the adaptive attention learning module, and the multi-level feature aggregation module are treated as a downsampling layer. L downsampling layers are stacked and propagated forward layer by layer (when propagating between the front and back layers, only the features of the center point are updated, and only the center point is retained to participate in the sampling and neighborhood construction of the next layer → self-attention weight calculation → local geometric feature aggregation process). In this way, the number of center points is reduced layer by layer, and finally the high-dimensional geometric semantic features of the measured point cloud / theoretical point cloud are obtained.

[0086] S3. Construct a cross-modal cross-attention fusion module, introduce an anti-occlusion dynamic confidence mask, and shield measurement blind zone interference by evaluating the effective confidence distribution of the measured point cloud. Establish a highly robust "soft correspondence" spatial mapping network between the measured spatial features and the theoretical numerical model prior, thereby fusing the high-dimensional geometric semantic features of the measured point cloud with the high-dimensional geometric semantic features of the theoretical point cloud across modalities to obtain cross-modal fusion features.

[0087] In a preferred embodiment, step S3 specifically includes:

[0088] The high-dimensional geometric and semantic features of the measured point cloud are mapped into a query vector using a linear projection matrix. The high-dimensional geometric and semantic features of the theoretical point cloud are mapped to key vector K and value vector V through linear projection matrices, respectively; the formula is expressed as:

[0089] , , ;

[0090] in, , These represent the high-dimensional geometric semantic features of the measured point cloud and the theoretical point cloud, respectively. , , This transforms the features into linear projection matrices of query vectors, key vectors, and value vectors;

[0091] A spatial occlusion perception subnetwork is constructed. This subnetwork takes the high-dimensional geometric semantic features of the measured point cloud as input and maps the high-dimensional geometric semantic features of each center point to a scalar form of effective confidence distribution through a regression branch consisting of a fully connected layer, a batch normalization layer, and a nonlinear activation function. The effective confidence distribution quantifies the probability that the point belongs to the valid measurement area rather than a missing blind spot. In this embodiment, the effective confidence distribution... ;

[0092] Calculate the inner product of the query vector and the key vector and perform scaling to obtain the basic cross-modal similarity correlation matrix representing the spatial correlation between the high-dimensional geometric semantic features of the measured point cloud and the high-dimensional geometric semantic feature space of the theoretical point cloud. The formula is expressed as:

[0093] ;

[0094] in, Scaling factor For feature dimensions;

[0095] The matching weights of each node in the basic cross-modal similarity association matrix are gated using the effective confidence distribution values. This forces the network to match only in areas where the data is reliable, preventing feature alignment errors caused by large-area scanning blind spots. After gated filtering, an anti-occlusion mask attention weight matrix A is obtained. In this embodiment, the formula for the anti-occlusion mask attention weight is expressed as:

[0096] ;

[0097] in, Indicates the first The measured point and the first The mask attention weights for preventing occlusion between theoretical points, i.e., the th... The measured point and the first The probability of a true match between theoretical points; To prevent smooth terms with a denominator of zero; In the basic cross-modal similarity correlation matrix, the first The measured point and the first The basic similarity score between theoretical points The total number of sampling points in the theoretical point cloud participating in the matching, For the first The measured point and the first The basic similarity scores between theoretical points are used for local normalized summation calculations in the denominator.

[0098] The cross-modal fusion feature is obtained by weighting and summing the anti-occlusion mask attention weights with the value vector, as expressed by the formula:

[0099] ;

[0100] in, For cross-modal fusion features For the first The value vector is obtained by linear mapping of the high-dimensional geometric semantic features of a theoretical point.

[0101] Through step S3, the method proposed in this invention effectively overcomes the problems of inconsistent local topology and registration distortion caused by measurement occlusion and manufacturing deformation coupling.

[0102] S4. Using a multi-layer feature backpropagation and engineering tolerance-oriented decoding model, spatial upsampling regression is performed on the cross-modal fusion features; an asymmetric penalty loss function combining process features is introduced (introduced during the overall network training process) to calculate the three-dimensional manufacturing deviation vector of each measured point.

[0103] In a preferred embodiment, step S4 specifically includes:

[0104] The cross-modal fusion features are upsampled layer by layer using a multi-layer feature propagation module, and the interpolation points are calculated based on Euclidean distance. With nearest neighbor control points The inverse distance weights between the control points (i.e., the sparse center points retained in the previous layer) restore the feature resolution layer by layer to the scale of the original input measured point cloud. The formula is expressed as:

[0105] ;

[0106] in, For the i-th measured point Interpolated features for Inverse distance weights; For the first nearest neighbor control points The known feature vectors at the current network layer; This represents the number of nearest neighbor control points during the upsampling interpolation process.

[0107] The upsampled cross-modal fusion features are concatenated with the normalized 3D coordinates of each measured point and input to the multilayer perceptron decoder. The multilayer perceptron decoder adopts a dual-output head architecture, including a rigid parameter branch for regressing the global pose transformation and a deviation vector branch for regressing the displacement residuals of each sampling point.

[0108] During the training phase of the multilayer perceptron decoder, considering the process characteristics of thin-walled composite parts—that "excessive thickness (positive deviation) can be corrected by grinding, while excessive thinness (negative deviation) easily leads to part scrap"—a penalty weight greater than that for the positive deviation of the thin-walled composite parts is applied to the negative deviation. Furthermore, geometrically weighted high-penalty constraints are applied to critical assembly surfaces (such as rib intersection areas), thus constructing an asymmetric penalty loss function. The formula is expressed as:

[0109] ;

[0110] in, The global rigid registration loss is used to characterize the alignment error between the global spatial pose predicted by the network and the true pose. It is calculated from the mean square error between the predicted rigid transformation parameters and the true rigid transformation parameters. For the i-th measured point The deviation between the predicted deviation and the actual deviation; and The asymmetric penalty coefficient and , and To balance the hyperparameters; For the i-th measured point Geometric weights at the location; The total number of points in the measured point cloud used to calculate this loss.

[0111] After forward propagation and optimization using the asymmetric penalty loss function, the two output heads of the multilayer perceptron decoder directly regress the rigid transformation parameters used to compensate for global spatial alignment errors, and the three-dimensional manufacturing deviation vector for each measured point of the component's physical deformation. .

[0112] S5. Construct a global form and position deviation field based on the three-dimensional manufacturing deviation vector, generate a continuous deviation heat map through chromatographic mapping, and combine semantic segmentation to provide early warning of out-of-tolerance in key high-frequency feature regions, and output a detection report of form and position deviation of composite thin-walled components.

[0113] As a preferred embodiment, such as Figure 3 As shown, in this embodiment, step S5 specifically includes:

[0114] Based on the three-dimensional manufacturing deviation vector, the absolute Euclidean distance deviation value of each measured point relative to the theoretical surface is calculated; the formula is expressed as:

[0115] ;

[0116] in, , , They are respectively Deviation on the three coordinate axes;

[0117] A pseudo-color chromatogram mapping rule is set to map the absolute Euclidean distance deviation value to the corresponding color region, and the measured 3D model is rendered point by point to generate a deviation heatmap.

[0118] By combining semantic segmentation tags, the local deviation distribution of key high-frequency feature regions (such as wing panel ribs, air intake flanges, and assembly holes) is extracted; the local deviation distribution is then compared with a preset engineering tolerance threshold. The system performs a comparison, automatically marks and extracts the geometric coordinates of the out-of-tolerance areas and the extreme values ​​of the maximum deviation;

[0119] The system integrates the statistical data of deviation across the entire domain (i.e., the root mean square error, maximum positive and negative deviation values, and tolerance pass rate calculated based on the absolute Euclidean distance deviation values ​​of the measured points across the entire domain), the early warning information of out-of-tolerance in key high-frequency feature areas (i.e., the feature semantic labels, geometric coordinates, and maximum deviation extreme values ​​of the extracted out-of-tolerance areas), and the visualization results of the deviation heatmap, and outputs a test report on the form and position deviation of thin-walled composite material components.

[0120] Furthermore, this application also discloses an application system for the aforementioned intelligent analysis method of form and position deviations of thin-walled composite material components for aircraft based on a dual-branch network, which specifically includes:

[0121] The data preprocessing module is used to convert the measured point cloud and theoretical CAD model of the composite thin-walled component into a two-branch initial dataset with normalized spatial coordinates.

[0122] A twin feature extraction network is used to extract high-dimensional geometric and semantic features of the processed measured point cloud and theoretical point cloud from the dual-branch initial dataset;

[0123] The cross-attention fusion module introduces an anti-occlusion dynamic confidence mask to establish a soft correspondence space mapping between measured features and theoretical numerical model priors, thereby fusing two sets of high-dimensional geometric and semantic features into cross-modal fusion features.

[0124] The deviation regression decoding module is used to upsample and decode cross-modal fusion features, and after optimization by an asymmetric penalty loss function, it calculates the rigid transformation parameters and the three-dimensional manufacturing deviation vector of each measured point end-to-end.

[0125] The visualization and early warning module is used to construct a form and position deviation field based on the 3D manufacturing deviation vector, generate a deviation heat map, and combine semantic segmentation to provide out-of-tolerance early warning and output an inspection report.

[0126] An electronic device is also disclosed, comprising a memory and a processor, wherein:

[0127] Memory is used to store computer programs that can run on a processor;

[0128] The processor is used to execute, as described above, an intelligent analysis method for form and position deviations of thin-walled composite material components for aircraft based on a dual-branch network when running the computer program.

[0129] A computer-readable storage medium is also disclosed, which stores computer instructions for causing a processor to execute an intelligent analysis method for form and position deviations of thin-walled composite material components for aircraft, as described above, based on a dual-branch network.

[0130] In summary, this invention addresses the problems of overall geometric deviation and local deformation of key feature areas caused by molding process fluctuations, material springback, and local structural coupling during the manufacturing of thin-walled composite components. It provides an intelligent analysis method for form and position deviations of thin-walled aircraft composite components based on a bi-branch network. This method focuses on solving technical challenges such as insufficient registration accuracy between measured 3D point clouds and theoretical models, difficulty in establishing correspondences between feature areas, and difficulty in quantifying manufacturing deviations with high precision. By constructing a bi-branch feature extraction network and a cross-modal attention fusion mechanism, high-precision spatial semantic matching between measured point clouds and theoretical models is achieved. Combined with a deviation regression decoding model, the method automatically predicts, identifies, and visualizes the full-domain 3D manufacturing deviations of composite components.

[0131] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0132] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for intelligent analysis of form and position deviations in thin-walled composite components for aircraft based on a dual-branch network, characterized in that, Specifically, the following steps are included: S1. Obtain the measured point cloud of the composite thin-walled component, and obtain the theoretical point cloud that matches the measured point cloud through joint geometric sampling; perform spatial reference normalization on the measured point cloud and the theoretical point cloud to obtain the dual-branch initial dataset. S2. Construct a twin high-dimensional semantic extraction network for the topological features of thin-walled components. Through parameterless physical location encoding and hierarchical adaptive attention mechanism, perform high-frequency geometric feature perception and high-dimensional semantic abstraction on the measured point cloud and theoretical point cloud after spatial benchmark normalization, respectively, and extract high-dimensional geometric semantic features. S3. Construct a cross-modal cross-attention fusion module, introduce an anti-occlusion dynamic confidence mask, and perform cross-modal fusion of the high-dimensional geometric semantic features of the measured point cloud and the high-dimensional geometric semantic features of the theoretical point cloud to obtain cross-modal fusion features; S4. Using a multi-layer feature backpropagation and engineering tolerance-oriented decoding model, spatial upsampling regression is performed on the cross-modal fusion features; an asymmetric penalty loss function combining process features is introduced to calculate the three-dimensional manufacturing deviation vector for each measured point; specifically including: The cross-modal fusion features are upsampled layer by layer using a multi-layer feature propagation module, and the feature resolution is restored to the scale of the original input measured point cloud based on the inverse distance weight. The upsampled cross-modal fusion features are concatenated with the normalized 3D coordinates of each measured point and input to the multilayer perceptron decoder. The multilayer perceptron decoder adopts a dual-output head architecture, including a rigid parameter branch for regressing the global pose transformation and a deviation vector branch for regressing the displacement residuals of each sampling point. During the training phase of the multilayer perceptron decoder, a process-oriented asymmetric penalty loss function is constructed by applying a penalty weight greater than that for the negative deviation of the composite thin-walled component and a geometrically weighted high penalty constraint to the assembly critical surface. After forward propagation and optimization using the asymmetric penalty loss function, the two output heads of the multilayer perceptron decoder directly regress the rigid transformation parameters used to compensate for global spatial alignment errors, and the three-dimensional manufacturing deviation vector for each measured point used to quantify the physical deformation of the component. S5. Construct a global form and position deviation field based on the three-dimensional manufacturing deviation vector, generate a continuous deviation heat map through chromatographic mapping, and combine semantic segmentation to provide early warning of out-of-tolerance in key high-frequency feature regions, and output a detection report of form and position deviation of composite thin-walled components.

2. The intelligent analysis method for form and position deviations of thin-walled composite material components for aircraft based on a dual-branch network according to claim 1, characterized in that, Step S1 specifically includes: A three-dimensional scanning point cloud of the surface of a thin-walled composite component is collected, and then outlier removal and filtering noise reduction are performed on the three-dimensional scanning point cloud to obtain the measured point cloud. A theoretical CAD model of a thin-walled component made of the same composite material is obtained. The surface of the theoretical CAD model is discretized using a uniform Poisson disk sampling algorithm to generate a theoretical point cloud that matches the scale of the measured point cloud. An initial alignment strategy is adopted to coarsely register the measured point cloud and the theoretical point cloud to the same global reference coordinate system, which is an absolute coordinate system based on the design origin of the theoretical CAD model. Subsequently, spatial reference normalization is performed to calculate the geometric center of the coarsely registered point cloud and translate it to the origin of the global reference coordinate system. At the same time, global scaling is performed using the distance of the farthest point as the scaling factor, so that both the coarsely registered measured point cloud and the theoretical point cloud are confined within a unit sphere centered at the coordinate origin. The normalized measured point cloud and the theoretical point cloud are reorganized by channel recombination and dimension alignment to obtain the initial dataset of the two branches.

3. The intelligent analysis method for form and position deviations of thin-walled aircraft composite material components based on a dual-branch network according to claim 1, characterized in that, Step S2 specifically includes: A twin high-dimensional semantic extraction network for topological features of thin-walled components is constructed. The twin high-dimensional semantic extraction network consists of a parameter-free physical location encoding module, a sampling and local neighborhood construction module, an adaptive attention learning module, and a multi-level feature aggregation module connected in series. The initial datasets of the two branches are input into the parameterless physical location encoding module. The parameterless physical location encoding operation is used to map the three-dimensional coordinates of each point in the normalized point cloud to a high-dimensional space to obtain high-dimensional encoded features. In the sampling and local neighborhood construction module, the farthest point sampling algorithm is used to determine the center point, and a local geometric neighborhood is established for each center point based on the spatial distance. In the adaptive attention learning module, a self-attention mechanism is introduced into the local geometric neighborhood of each center point. The self-attention weight of each neighboring point in each local geometric neighborhood is calculated relative to the center point through a learnable feature similarity scoring multilayer perceptron. In the multi-level feature aggregation module, based on the self-attention weights between nodes, the high-dimensional encoded features of the neighboring points in each local geometric neighborhood are weighted and summed to aggregate the local geometric features and update the high-dimensional geometric semantic features of the center point. The sampling and local neighborhood construction module, the adaptive attention learning module, and the multi-level feature aggregation module are treated as a downsampling layer. L downsampling layers are stacked, and the number of samples at the center point is reduced by forward propagation layer by layer, finally obtaining the high-dimensional geometric and semantic features of the measured point cloud / theoretical point cloud.

4. The intelligent analysis method for form and position deviations of thin-walled aircraft composite material components based on a dual-branch network according to claim 1, characterized in that, Step S3 specifically includes: The high-dimensional geometric and semantic features of the measured point cloud are mapped into a query vector through a linear projection matrix; the high-dimensional geometric and semantic features of the theoretical point cloud are mapped into a key vector and a value vector through a linear projection matrix, respectively. A spatial occlusion perception subnetwork is constructed. The spatial occlusion perception subnetwork takes the high-dimensional geometric semantic features of the measured point cloud as input and maps the high-dimensional geometric semantic features of each center point into a scalar form of effective confidence distribution through a regression branch composed of a fully connected layer, a batch normalization layer and a nonlinear activation function. Calculate the inner product of the query vector and the key vector and perform scaling to obtain the basic cross-modal similarity association matrix that represents the spatial correlation between the high-dimensional geometric semantic features of the measured point cloud and the high-dimensional geometric semantic features of the theoretical point cloud. The matching weights of each node in the basic cross-modal similarity association matrix are gating and filtering using the effective confidence distribution values. After gating and filtering, the anti-occlusion mask attention weight matrix is ​​obtained. The anti-occlusion mask attention weights and the value vector are weighted and summed to obtain the cross-modal fusion features.

5. The intelligent analysis method for form and position deviations of thin-walled composite material components for aircraft based on a dual-branch network according to claim 1, characterized in that, Step S5 specifically includes: Based on the three-dimensional manufacturing deviation vector, the absolute Euclidean distance deviation value of each measured point relative to the theoretical surface is calculated; A pseudo-color chromatogram mapping rule is set to map the absolute Euclidean distance deviation value to the corresponding color region, and the measured 3D model is rendered point by point to generate a deviation heatmap. By combining semantic segmentation labels, the local deviation distribution of key high-frequency feature regions is extracted; the local deviation distribution is compared with the preset engineering tolerance threshold, and the geometric coordinates and maximum deviation extreme values ​​of the out-of-tolerance regions are automatically marked and extracted. The system integrates statistical data on deviations across the entire domain, early warning information on out-of-tolerance conditions in key high-frequency characteristic areas, and visualization results of deviation heatmaps to output a detection report on the shape and position deviations of thin-walled composite material components.

6. An application system for an intelligent analysis method of form and position deviations of thin-walled aircraft composite material components based on a dual-branch network, as described in any one of claims 1-5, characterized in that, Specifically, it includes: The data preprocessing module is used to convert the measured point cloud and theoretical CAD model of the composite thin-walled component into a two-branch initial dataset with normalized spatial coordinates. A twin feature extraction network is used to extract high-dimensional geometric and semantic features of the processed measured point cloud and theoretical point cloud from the dual-branch initial dataset; The cross-attention fusion module introduces an anti-occlusion dynamic confidence mask to establish a soft correspondence space mapping between measured features and theoretical numerical model priors, thereby fusing two sets of high-dimensional geometric and semantic features into cross-modal fusion features. The deviation regression decoding module is used to upsample and decode cross-modal fusion features, and after optimization by an asymmetric penalty loss function, it calculates the rigid transformation parameters and the three-dimensional manufacturing deviation vector of each measured point end-to-end. The visualization and early warning module is used to construct a form and position deviation field based on the 3D manufacturing deviation vector, generate a deviation heat map, and combine semantic segmentation to provide out-of-tolerance early warning and output an inspection report.

7. An electronic device, characterized in that, The electronic device includes a memory and a processor, wherein: Memory is used to store computer programs that can run on a processor; A processor is configured to, when running the computer program, execute the intelligent analysis method for form and position deviations of thin-walled composite material components for aircraft based on a dual-branch network as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the intelligent analysis method for shape and position deviations of thin-walled aircraft composite material components based on a dual-branch network as described in any one of claims 1-5.

Images