Knowledge base automatic identification and matching method, device and equipment and storage medium

Abstract

The application relates to the technical field of automobile manufacturing process design, and discloses a knowledge base automatic identification and matching method, device, equipment and storage medium. Original die surface data is acquired, the original die surface data is preprocessed, the preprocessed original die surface data is split into three part profiles by using a point cloud segmentation model, and profile data is obtained; a multi-modal feature fusion model is used to identify the part type of the profile data, and a part type identification result is obtained; based on the part type identification result, an improved SIFT+curvature constraint feature extraction algorithm is used to identify part features of the part, a feature map is constructed based on a GAT graph neural network, the part type identification result, the part features, a target blank holder control line and a target process supplementary surface cross section line are associated, and an associated matching result is stored in a knowledge base; the parameter deviation of manual design is greatly reduced, and the process trial and error cost is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automobile manufacturing process design, and particularly relates to a knowledge base automatic identification and matching method, device and equipment and a storage medium. BACKGROUND

[0002] In the automobile manufacturing industry, the stamping process is a key link in vehicle body manufacturing, and the generation quality of the drawing process die surface directly affects the formability of the part. At present, the generation of the drawing process die surface of the stamping part mainly depends on manual experience. Manual experience generates a control line on the product surface to fit the shape trend of the part, and then generates a complete pressure surface by continuously adjusting the control line. Manual experience generates a section line and a guide line between the product surface and the pressure surface according to experience, industry standards and part characteristics, and then generates a process supplementary surface. The generated pressure surface and process supplementary surface are used for CAE simulation calculation. If the formability is unqualified, the pressure surface or process supplementary surface needs to be adjusted according to the problem and reason for adjustment, and multiple iterations are performed. This method relies heavily on manual experience, and it takes a long time for new employees to get started, making it difficult to quickly generate qualified process surfaces. The manual generation process is complicated, and multiple changes and calculations are required, which is low in efficiency and affects the product development cycle. Knowledge extraction of control lines and section lines involves a lot of manual measurement work, and it is difficult to form systematic knowledge accumulation. The quality of the generated process surface is unstable, and formability problems are prone to occur, leading to repeated modifications and increased costs. SUMMARY

[0003] The present application provides a knowledge base automatic identification and matching method.

[0004] In a first aspect, the present application provides a knowledge base automatic identification and matching method, which comprises the following steps:

[0005] Obtain original die surface data, and after preprocessing the original die surface data, use a PointNet++ point cloud segmentation model to split the preprocessed original die surface data into three part surfaces to obtain surface data;

[0006] Perform part type identification on the surface data by using a ResNet-50+Transformer multi-modal feature fusion model to obtain a part type identification result;

[0007] Based on the part type recognition result, the part features are recognized by an improved SIFT+curvature constraint feature extraction algorithm, the surface lines are broken at the step and corner of the part product surface contour in the vertical direction of the contour to obtain the process supplementary surface section line, and the surface lines are broken at the four vertices of the part product surface contour in the parallel direction of the plane to obtain the blank holder surface control line; the straight lines and circular arcs of the broken surface lines in the process supplementary surface section line and the blank holder surface control line are extracted, the length and draft angle of the straight lines are calculated, and the radius and central angle of the circular arcs are calculated to obtain the target blank holder surface control line and the target process supplementary surface section line;

[0008] Based on the GAT graph neural network, a feature map is constructed, the part type recognition result, the part features, the target blank holder surface control line and the target process supplementary surface section line are associated, and the associated matching result is stored in the knowledge base;

[0009] The new part data is received, the part features of the new part data are obtained, and the corresponding target blank holder surface control line and target process supplementary surface section line are matched in the knowledge base according to the part features.

[0010] In a possible implementation manner, after the original mold surface data is preprocessed, the preprocessed original mold surface data is split into three part surfaces by using a PointNet++ point cloud segmentation model to obtain the surface data; including:

[0011] The original mold surface data is obtained, including the numerical model data of the part product surface, the blank holder surface and the process supplementary surface, the adaptive triangulation algorithm is used to generate the surface patch of the original mold surface data, the straight-through filtering is performed on the surface patch, and then the Poisson reconstruction algorithm is used to generate the triangular surface patch model to obtain the standard part data;

[0012] The triangular surface patch model in the standard part data is subjected to Laplace smoothing, and the feature retention degree of the smoothed model is detected; if the feature loss is too large, the iteration result of the previous time is traced back, the blank holder surface and the process supplementary surface are separately subjected to constraint smoothing to obtain the denoising part data;

[0013] The center of gravity of the part is taken as the origin, the maximum three-dimensional direction size of the part in the denoising part data is calculated, the model is scaled in the three-dimensional space in a proportional manner, and the scaling factor and the original coordinate origin are recorded to obtain the pretreated part data.

[0014] In a possible implementation manner, the original mold surface data is obtained, the preprocessed original mold surface data is split into three part surfaces by using a PointNet++ point cloud segmentation model after the original mold surface data is preprocessed to obtain the surface data; further including:

[0015] The PointNet++ point cloud segmentation model based on part surface segmentation optimization is called, the first layer sampling radius is 0.05mm, the number of neighborhood points is 32, which is used for extracting local geometric features at the patch level; the second layer sampling radius is 0.1mm, the number of neighborhood points is 64, which is used for fusing adjacent local features to form part features; the third layer sampling radius is 0.2mm, the number of neighborhood points is 128, which is used for integrating part features to obtain global surface features;

[0016] The preprocessed part data is used for feature extraction by using the PointNet++ point cloud segmentation model to construct a segmentation label library, three types of surface feature indicators are defined, K-Means clustering is used for preliminary grouping, and then a full connection layer is input to classify by using a Softmax activation function to obtain surface data;

[0017] The product surface of the surface data is located in the middle part, the pressing surface is located in the outermost circle, and the process supplement surface is located in the middle; the pressing surface is a large surface, and the radius is greater than 50mm; there is a closed and continuous circular arc surface near the pressing surface; the process supplement surface starts from the circular arc surface of the pressing surface, the circular arc surface connecting the product surface is not a complete circle, a boundary line is found through the circular arc surface, the surface between the two parts is the process supplement surface, and the middle part is the product surface.

[0018] In a possible implementation manner, the original mold surface data is acquired, a ResNet-50+Transformer multi-modal feature fusion model is used for part type recognition on the surface data, and a part type recognition result is obtained, including:

[0019] The surface data is rendered by using a Blender rendering engine to generate 10 standardized images, and a geometric feature vector is calculated based on the standardized images, the geometric feature vector at least including a part product outer contour, a punching number, a complexity and a geometric shape of a main cross section line, and an envelope size;

[0020] A pre-trained ResNet-50 model is used to extract features by using convolution layers and pooling layers, a feature vector output by a penultimate layer is acquired, average pooling is performed on the feature vector, a visual feature vector is generated by retaining feature vectors of each view;

[0021] The geometric feature vector is mapped to a feature vector by using a full connection layer, and the visual feature vector is spliced to form a multi-modal feature vector;

[0022] The geometric feature vector is input into a Transformer encoder, the Transformer encoding adopts a 6-layer encoder, the number of multi-head attention heads is 8, global average pooling is performed on the encoded feature matrix to obtain a fusion feature vector;

[0023] The multimodal feature vector and the fused feature vector are output through a fully connected layer to obtain the part type identification result. The part type identification result includes at least the outer hood panel, inner hood panel, outer front door panel, inner front door panel, fender, side panel, wheel arch, A-pillar, and B-pillar.

[0024] In one possible implementation, based on the part type identification result, the part features are identified using an improved SIFT+curvature constraint feature extraction algorithm. At the steps and corners of the part's product surface contour, cut lines are made perpendicular to the contour to obtain the process supplementary surface section lines. Based on the four vertices of the part's product surface contour, cut lines are made parallel to the plane to obtain the blank holder control lines. The straight lines and arcs of the cut lines from the process supplementary surface section lines and blank holder control lines are extracted, the lengths and draft angles of the straight lines are calculated, and the radii and central angles of the arcs are calculated to obtain the target blank holder control lines and the target process supplementary surface section lines, including:

[0025] Based on the part type identification results, the part features are identified using an improved SIFT+curvature constraint feature extraction algorithm. A scale space is constructed using Gaussian difference, increasing the number of Gaussian pyramid layers to 8. The calculation formula is as follows:

[0026]

[0027]

[0028] in, Indicates the pixel values ​​of the input image. Represents a two-dimensional Gaussian function. This represents the standard deviation of the Gaussian blur. This represents the image after Gaussian blurring. Indicates the scale factor. Represents a Gaussian difference image;

[0029] Sub-pixel level feature point coordinates are determined through extreme point detection and interpolation calculations. Low-contrast points and edge response points are removed. The directional gradient of the feature point neighborhood is calculated using the following formula:

[0030]

[0031]

[0032] in, This represents the gradient magnitude, and indicates the pixel value. The intensity of the brightness change at that location. Indicates the gradient direction, representing the pixel. The direction and range of brightness change (0°-360°). , , and respectively represent the Gaussian blur values of the right, left, top and bottom of the pixel point ;

[0033] A discrete surface curvature calculation method is used to take 30 neighborhood points of each feature point to fit a local surface, calculate Gaussian curvature and average curvature, compare the curvature value of the feature point with the rule library, retain the feature points meeting the conditions and eliminate the false detection points to obtain the part features.

[0034] In a possible implementation, based on the part type identification result, the part features of the part are identified by using the improved SIFT+curvature constraint feature extraction algorithm, the surface lines are broken at the step and corner of the part product surface contour in the part features along the contour vertical direction to obtain the process supplementary surface section line; the surface lines are broken based on the four corner points of the part product surface contour along the plane parallel direction to obtain the blank holder surface control line; the straight lines and circular arcs of the broken surface lines in the process supplementary surface section line and the blank holder surface control line are extracted, the length and draft angle of the straight lines are calculated, and the radius and central angle of the circular arcs are calculated to obtain the target blank holder surface control line and the target process supplementary surface section line, and the method further comprises:

[0035] The surface lines are broken at the step and corner of the part product surface contour in the part features along the contour vertical direction to obtain the process supplementary surface section line; the surface lines are broken based on the four corner points of the part product surface contour inwardly by 20 mm along the plane parallel direction to obtain the blank holder surface control line; the plane includes the XOZ plane and the YOZ plane;

[0036] The RANSAC algorithm fitting is performed on the broken surface lines in the process supplementary surface section line and the blank holder surface control line, the iteration number of straight line fitting is 1000, the inlier threshold is 0.01 mm, and the calculation formula is as follows:

[0037]

[0038]

[0039] wherein, , , represent the parameters of the straight line, and represent the pixel coordinates of the broken surface line, represents the distance from the point to the straight line, represents the inlier threshold of the straight line, represents the index of different pixel points on the broken surface line;

[0040] The iteration number of circular arc fitting is 1500, the inlier threshold is 0.015 mm; the average distance between the fitting elements and the original point cloud is calculated, and the calculation formula is as follows:

[0041]

[0042]

[0043] wherein, and denotes the coordinate of the center of the arc, denotes the radius of the arc, denotes the distance from the point to the arc, denotes the inner point threshold of the arc, denotes the coordinate difference of the pixel point in the x-axis from the center of the arc, denotes the coordinate difference of the pixel point in the y-axis from the center of the arc;

[0044] The straight line parameters are calculated by the distance formula between two points, and the draft angle is calculated by the angle between the straight line and the direction perpendicular to the parting surface, and the calculation formula is as follows:

[0045] }

[0046]

[0047] wherein, denotes the number of inner points of the arc, and denotes the coordinate of the inner point of the arc, denotes the coordinate difference of the m-th inner point in the horizontal direction from the center of the arc, denotes the coordinate difference of the m-th inner point in the vertical direction from the center of the arc, denotes the draft angle, denotes the vector perpendicular to the direction of the parting surface, denotes the direction vector of the fitting straight line of the section line, denotes the dot product of two vectors, , denotes the length of the two vectors; The center angle is calculated by the vector angle between the two end points of the arc and the center, and the calculation formula is as follows:

[0048]

[0049]

[0050] wherein, denotes the center angle, denotes the horizontal and vertical coordinates of the starting end point of the arc, denotes the horizontal and vertical coordinates of the ending end point of the arc, denotes the center coordinate of the arc, and denotes the x and y axis components of the vector from the center to the starting end point of the arc, and​​ denotes the component of the vector from the center of the circle to the end point of the arc in the x-axis, denotes the inverse cosine function, and the feature points are clustered by the Euclidean distance to obtain the target blank holder control line and the target process supplementary surface section line.

[0051] In a possible implementation, based on the part type identification result, a feature graph is constructed based on a GAT graph neural network, the part type identification result, part features, the target blank holder control line and the target process supplementary surface section line are associated, and the association matching result is stored in a knowledge base, including:

[0052] The feature graph is constructed based on the GAT graph neural network, the part type node attributes include part type ID, name, confidence and key feature list; and the target feature node attributes include feature ID, type, belonging surface, geometric parameter, extraction confidence and fitting accuracy.

[0053] The association edge type is part type-target feature, target feature-blank holder control line and target feature-process supplementary surface section line, the GAT graph neural network is executed for association to obtain an association result, and the association matching result is stored in the knowledge base.

[0054] In a second aspect, the present application provides a knowledge base automatic identification and matching device, which comprises the following modules:

[0055] A surface data splitting module is configured to obtain original surface data, split the original surface data into three parts of surface after preprocessing the original surface data by using a PointNet++ point cloud segmentation model, and obtain surface data.

[0056] A part type identification module is configured to identify part types by using a ResNet-50+Transformer multi-modal feature fusion model to obtain a part type identification result.

[0057] A part feature extraction module is configured to identify part features based on the part type identification result by using an improved SIFT+curvature constraint feature extraction algorithm, break the surface line in the vertical direction of the contour at the step and corner of the part product surface contour in the part features, obtain a process supplementary surface section line, break the surface line in the parallel direction of the plane based on the four corner points of the part product surface contour, obtain a blank holder control line, extract straight lines and circular arcs of the broken surface lines in the process supplementary surface section line and the blank holder control line, calculate the length and draft angle of the straight lines, calculate the radius and central angle of the circular arcs, and obtain the target blank holder control line and the target process supplementary surface section line.

[0058] The feature map construction module is used to construct feature maps based on the GAT graph neural network, associate the part type identification results, part features, target blanking surface control lines and target process supplementary surface cross-sectional lines, and store the association matching results in the knowledge base;

[0059] The part association and matching module is used to receive new part data, obtain the part features of the new part data, and match the corresponding target pressure surface control line and target process supplementary surface section line in the knowledge base according to the part features.

[0060] Thirdly, this application provides an automatic identification and matching device for a knowledge base, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the automatic identification and matching device for the knowledge base to perform the various steps of the above-described automatic identification and matching method for the knowledge base.

[0061] Fourthly, this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the various steps of the above-described automatic identification and matching method for the knowledge base.

[0062] Based on the method provided in this application, its beneficial effects are as follows: By combining the PointNet++ point cloud segmentation model with a standardized preprocessing workflow, the three types of surfaces of parts are accurately segmented, solving the problem of boundary ambiguity in traditional segmentation and improving the accuracy of surface segmentation. The ResNet-50+Transformer multimodal fusion architecture integrates visual and geometric features, achieving a part type recognition confidence level of over 0.85, which is 15% more accurate than a single-modal model. Utilizing an improved SIFT+curvature constraint algorithm, target features are filtered through curvature thresholds. Combined with precise sectioning and parameter calculation, the fitting error of the blank holder control line and the process supplementary surface section line is controlled within 0.03mm, significantly reducing parameter deviations in manual design and reducing process trial-and-error costs. The feature map constructed based on the GAT graph neural network enables multi-dimensional data association, improving the structured nature of knowledge storage. When matching new parts, based on standardized feature extraction and map retrieval, the matching response time is shortened to the second level, improving matching accuracy and avoiding the time-consuming and missed matching problems of traditional manual retrieval. This significantly improves the efficiency of part process design. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating an embodiment of the automatic identification and matching method for the first knowledge base provided in this application.

[0064] Figure 2 This is a flowchart illustrating an embodiment of the automatic identification and matching method for the second type of knowledge base provided in this application.

[0065] Figure 3 An embodiment flow chart of a third knowledge base automatic identification and matching method provided in the embodiments of the present application;

[0066] Figure 4 A structural schematic diagram of a knowledge base automatic identification and matching device provided in the embodiments of the present application;

[0067] Figure 5 A structural schematic diagram of a knowledge base automatic identification and matching device provided in the embodiments of the present application. DETAILED DESCRIPTION

[0068] The embodiments of the present application provide a knowledge base automatic identification and matching method.

[0069] The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application. In the specification and claims of the present application and the above drawings, the terms first, second, third, fourth, etc. exist for distinguishing similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the term includes or has and any variants thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0070] It can be understood that any part related to data acquisition or collection in the present application has obtained the authorization of the user.

[0071] It can be understood that the execution subject of the present application can be a knowledge base automatic identification and matching device, and can also be a mobile terminal or a server, which is not limited here.

[0072] The specific flow of the embodiments of the present application will be described below. Please refer to Figure 1 , Figure 1 An embodiment flow chart of a third knowledge base automatic identification and matching method provided in the embodiments of the present application;

[0073] 101, obtaining original mold surface data, and after preprocessing the original mold surface data, using a PointNet++ point cloud segmentation model to split the preprocessed original mold surface data into three parts of the surface, to obtain surface data;

[0074] It can be understood that the original mold surface data is obtained, at least including the numerical model data of the part product surface, the pressing surface and the process supplementary surface, the surface patches of the original mold surface data are generated by using the adaptive triangulation algorithm, the straight-through filtering is performed on the surface patches, and then the triangular patch model is generated by using the Poisson reconstruction algorithm to obtain standard part data;

[0075] The triangular patch model in the standard part data is subjected to Laplace smoothing, the feature retention degree of the model after smoothing is detected, if the feature loss is too large, the iteration result of the previous time is traced back, the pressing surface and the process supplementary surface are separately subjected to constraint smoothing to obtain denoising part data;

[0076] The three-dimensional direction maximum size of the part in the denoising part data is calculated with the center of gravity of the part as the origin, the model is scaled in the three-dimensional space in a proportional manner, and the scaling factor and the original coordinate origin are recorded to obtain preprocessed part data.

[0077] The PointNet++ point cloud segmentation model based on part surface segmentation optimization is called, the first layer sampling radius is 0.05 mm, the number of neighborhood points is 32, which is used for extracting patch-level local geometric features; the second layer sampling radius is 0.1 mm, the number of neighborhood points is 64, which is used for fusing adjacent local features to form part features; the third layer sampling radius is 0.2 mm, the number of neighborhood points is 128, which is used for integrating part features to obtain global surface features;

[0078] After the preprocessed part data is subjected to feature extraction by using the PointNet++ point cloud segmentation model, a segmentation label library is constructed, three types of surface feature indicators are defined, the K-Means clustering is used for preliminary grouping, and then the full connection layer is inputted and classified by using the Softmax activation function to obtain surface data;

[0079] The product surface of the surface data is located in the middle part, the pressing surface is located in the outermost circle, and the process supplementary surface is located in the middle; the pressing surface is a whole surface, and the radius is greater than 50 mm; there is a whole circle of closed continuous arc surface near the pressing surface; the process supplementary surface starts from the arc surface of the pressing surface, the arc surface connecting the product surface is not a whole circle, the boundary line continuous with the contour is found through the arc surface, the surface between the two parts is the process supplementary surface, and the middle part is the product surface.

[0080] 102. The part type recognition result is obtained by identifying the part type of the surface data by using the ResNet-50+Transformer multi-modal feature fusion model;

[0081] It can be understood that the surface data is subjected to multi-view rendering based on the Blender rendering engine to generate 10 standardized images, and the geometric feature vector is calculated based on the standardized images, and the geometric feature vector at least includes the part product outer contour, the punching number, the complexity and geometric shape of the main cross section line, and the envelope size;

[0082] The pre-trained ResNet-50 model is adopted to extract features step by step through convolution layers and pooling layers, obtain a feature vector of an output of a penultimate layer, perform average pooling on the feature vector, and reserve feature vectors of each view to generate a visual feature vector;

[0083] The geometric feature vector is mapped to a feature vector through a fully connected layer, and the visual feature vector is spliced to form a multi-modal feature vector;

[0084] The geometric feature vector is input into a Transformer encoder, the Transformer encoding adopts a 6-layer encoder, the number of multi-head attention heads is 8, global average pooling is performed on the encoded feature matrix to obtain a fusion feature vector;

[0085] The multi-modal feature vector and the fusion feature vector are output through a fully connected layer to obtain a part type recognition result, and the part type recognition result at least includes a cover outer plate, a cover inner plate, a front door outer plate, a front door inner plate, a fender, a side wall, a wheel cover, an A column and a B column.

[0086] First, the product surface of the part is automatically split by the PointNet++ point cloud segmentation model, and then the part type is recognized based on the ResNet-50+Transformer multi-modal feature fusion model, and the key features such as the product surface contour shape, corner feature and trimming angle of the part are automatically recognized by the improved SIFT+curvature constraint feature extraction algorithm. The product surface is in the middle part, the pressing surface is in the outermost circle, and the supplementary surface is in the middle. Relying on the surface features, the pressing surface is generally a large surface, and the radius is generally not less than 50. There is a whole circle of closed continuous arc surface near the pressing surface. From this circle of arc surface, it is a supplementary surface, and there is a circle of arc surface connecting the product surface, but it may not be a whole circle of arc surface, and there may be a part of straight surface, but the boundary line between the two parts can be found through the arc surface, and the surface between the two parts is the supplementary surface, and the middle part is the product surface. The recognition result of the model here is the part type recognition result of the product surface.

[0087] 103、Based on the part type recognition result, the part features of the part are recognized by the improved SIFT+curvature constraint feature extraction algorithm. The step and corner of the part product surface contour are broken along the contour vertical direction to obtain the process supplementary surface section line. The four vertices of the part product surface contour are taken as the basis to break the surface line along the plane parallel direction to obtain the pressing surface control line. The straight lines and arcs of the process supplementary surface section line and the pressing surface control line are extracted, the length and draft angle of the straight line are calculated, and the radius and central angle of the arc are calculated to obtain the target pressing surface control line and the target process supplementary surface section line.

[0088] It can be understood that the cutting and the corner are the cutting on the contour and the corner gathering place. Based on the part type recognition result, the part features of the part are recognized by the improved SIFT+curvature constraint feature extraction algorithm, the scale space is constructed by the Gaussian difference, the number of Gaussian pyramid layers is increased to 8 layers, and the calculation formula is as follows:

[0089]

[0090]

[0091] Wherein, represents the input image pixel value, represents a two-dimensional Gaussian function, represents the standard deviation of Gaussian blur, represents the image after Gaussian blur, represents the scale factor, represents the Gaussian difference image, represents the horizontal direction pixel position of the corresponding image, corresponding vertical direction pixel position;

[0092] The sub-pixel level feature point coordinates are determined by extreme point detection and interpolation calculation, the low contrast points and edge response points are removed, the direction gradient of the neighborhood of the feature point is calculated, and the calculation formula is as follows:

[0093]

[0094]

[0095] Wherein, represents the gradient amplitude, represents the intensity of the brightness change of the pixel point , represents the gradient direction, represents the direction of the brightness change of the pixel point , the range is (0°-360°), , , and respectively represent the right, left, upper and lower Gaussian blur values of the pixel point .

[0096] The discrete surface curvature calculation method is adopted, 30 neighborhood points are fitted to the local surface for each feature point, the Gaussian curvature and the average curvature are calculated, the curvature value of the feature point is compared with the rule library, the feature points meeting the conditions are retained and the misdetected points are removed, and the part features are obtained.

[0097] The section line is broken at the step and corner of the part product surface profile in the vertical direction of the profile in the part feature to obtain a process supplementary surface section line; the section line is broken in the direction parallel to XOZ plane or YOZ plane by 10mm-20mm from the four corner points of the part product surface profile to obtain a blank holder surface control line;

[0098] The RANSAC algorithm is executed on the broken section lines in the process supplementary surface section line and the blank holder surface control line, the linear fitting iteration number is 1000, the inlier threshold is 0.01mm, and the calculation formula is as follows:

[0099]

[0100]

[0101] wherein, , , represents the linear parameter, and the pixel coordinates of the section line, represents the distance from the point to the line, represents the inlier threshold of the line, represents the index of different pixel points on the section line;

[0102] The circular arc fitting iteration number is 1500, the inlier threshold is 0.015mm; the average distance between the fitting elements and the original point cloud is calculated, and the calculation formula is as follows:

[0103]

[0104]

[0105] wherein, and represent the circular arc center coordinates, represents the circular arc radius, represents the distance from the point to the circular arc, represents the inlier threshold of the circular arc, represents the coordinate difference value of the pixel point on the x-axis and the center, represents the coordinate difference value of the pixel point on the y-axis and the center;

[0106] The linear parameter is calculated by the distance formula between two points, and the draft angle is calculated by the angle between the line and the direction perpendicular to the parting surface, and the calculation formula is as follows:

[0107] }

[0108]

[0109] wherein, represents the number of points inside the circular arc, and represents the coordinate of the point inside the circular arc, represents the coordinate difference between the th point inside the circular arc and the center of the circle in the horizontal direction, represents the coordinate difference between the th point inside the circular arc and the center of the circle in the vertical direction, represents the draft angle, represents the vector perpendicular to the parting surface direction, represents the direction vector of the section line fitting straight line, represents the dot product of two vectors, , represents the modulus of two vectors;

[0110] The center angle is calculated by the vector angle between the line connecting the two end points of the circular arc and the center of the circle, and the calculation formula is as follows:

[0111]

[0112] wherein, represents the center angle, represents the horizontal and vertical coordinates of the starting point of the circular arc, represents the horizontal and vertical coordinates of the end point of the circular arc, represents the center coordinates of the circular arc, and represents the x and y axis components of the vector from the center of the circle to the starting point of the circular arc, and represents the x and y axis components of the vector from the center of the circle to the end point of the circular arc, represents the inverse cosine function, the feature points are clustered by the Euclidean distance, and the target pressure surface control line and the target process supplementary surface section line are obtained.

[0113] The blank holder surface control line is generated, and high-precision control lines are generated in combination with process requirements and geometric parameters. Key feature point extraction: selecting feature points on the blank holder surface boundary, such as corner points, product surface connection points, and positioning hole centers, to ensure that the point spacing is ≤0.5 mm and the feature point coordinate accuracy is ≤0.01 mm; fitting algorithm selection: using a cubic B-spline curve fitting to ensure the smoothness of the curve and the curvature change rate ≤0.1 mm⁻¹, and the boundary conditions are constrained during fitting, and the tangent directions of the starting point and the ending point are consistent with the blank holder surface boundary; precision verification: calculating the maximum distance between the control line and the original blank holder surface boundary, ≤0.1 mm, checking the continuity of the control line, and the angle between adjacent segments ≤5°; process optimization: adjusting the control line shape according to the drawing coefficient, such as the first drawing coefficient ≥0.5, and the material yield strength, such as the yield strength of Q235 steel 235 MPa, to ensure uniform distribution of blank holder force, and the control line curvature change gentle area corresponds to the uniform blank holder force area. The coordinate point set of the control line is output, with an interval of 0.1 mm and B-spline curve parameters.

[0114] The process supplementary surface cross-section line is generated based on the cross-section line parameters and the forming process requirements. Cross-section shape design: according to the part forming depth, such as a drawing depth of 50 mm, the material thickness, such as 2 mm, and the fillet requirement, such as R3 mm, selecting a rectangular, trapezoidal, or arc cross-section, and the drawing depth > 30 mm prefers to use a trapezoidal cross-section; parameter determination: the width of the rectangular cross-section = forming gap + 2 × material thickness, and the height = process supplementary height; the upper base of the trapezoidal cross-section = rectangular width, and the lower base = upper base + 2 × slope × height, and the slope is usually 5°-10°; fusion processing: using geometric Boolean operation and intersection operation to fuse the designed cross-section line with the part cross-section line to ensure that there is no gap at the junction, ≤0.02 mm, and the fusion tool uses the OpenCASCADE geometric engine; output form: outputting the coordinate data of the cross-section line, with an interval of 0.05 mm, a two-dimensional contour graph, a DXF format, including size annotations and design descriptions, including cross-section type, design basis, and process parameters.

[0115] The identification basis is DP algorithm or other. The key positions are also determined according to the contour features of the product surface of the part, such as a certain distance (such as 20 mm) on both sides of the corner position, and a certain distance (such as 5 mm) on the upper and lower arcs of the stepped position. If the distance between two key points is too long, some points need to be inserted according to a certain distance (such as 100 mm).

[0116] 104、Based on the GAT graph neural network, a feature map is constructed, the part type recognition result, the part feature, the target blank holder surface control line, and the target process supplementary surface cross-section line are associated, and the associated matching result is stored in the knowledge base;

[0117] It can be understood that by constructing a feature map based on a GAT graph neural network, the part type node attribute includes part type ID, name, confidence and key feature list; the target feature node attribute includes feature ID, type, belonging profile, geometric parameter, extraction confidence and fitting accuracy;

[0118] The association edge type is part type-target feature, target feature-pressing surface control line and target feature-process supplementary surface section line, a GAT graph neural network is executed for association, an association result is obtained, and the association matching result is stored in the knowledge base.

[0119] The association relationship of the feature map is converted into multi-level structured data. The core five tuples are: part type, including ID, name, confidence-target feature list, including the ID, type, parameter, confidence of each feature-pressing surface control line parameter, ID, coordinate point set, accuracy, process constraint-process supplementary surface section line parameter, ID, section type, parameter, design basis-association confidence, and the average of all edge weights; extended information: add data processing link, processing time, tool, operator, quality score, based on feature extraction accuracy, fitting accuracy, etc. Calculate, full score 10 points, suitable process range, such as suitable for drawing process, material type Q235 / Q345; data format: use JSON-LD format, support semantic association, at the same time generate simplified version of CSV format data, used for MySQL storage, after structure, execute syntax check, ensure field integrity and format correctness.

[0120] Adopting a hybrid architecture of relational + graph database ensures efficient data storage and query. Knowledge base design: create 5 core tables, part type table, store type information and key features, target feature table, feature parameters and associated part ID, pressing surface control line table, coordinate parameters and accuracy, process supplementary surface section line table, section parameters and design basis, data link table, processing flow information, each table is associated through part ID, establish B+ tree index of part type and feature type, improve query efficiency by 50%; Neo4j graph database design: import node table and edge table, node label is set to part type, target feature, etc., edge type is set to association type, establish node attribute index, such as feature type, part ID; data security: sensitive parameters, such as process parameters, customer customization information, use AES-256 encryption, key rotation period is 30 days, database access log is turned on, record access IP, time, operation content, backup strategy is daily full backup + real-time incremental backup, backup data retention is 90 days; access interface: develop RESTful API interface, support HTTP / HTTPS, realize data addition, deletion, modification and query, interface response time ≤500ms.

[0121] An automatic updating process of incremental updating + quality checking + deduplication optimization is established. Incremental updating: a timing task is triggered at 2 a.m. every day, and the incremental query of the database is performed based on the processing timestamp to extract the newly added part data processing results on the same day. The results are updated to the knowledge base according to the principle of verification first and then import, and the MySQL performs the INSERT operation and the Neo4j performs the CREATE operation. Quality checking: the quality score of the newly added data is calculated, and the data with a score of ≥8 is directly imported, the data with a score of 6-7 is imported after manual audit, and the data with a score of <6 is returned for reprocessing. Deduplication optimization: the cosine similarity algorithm is used to calculate the similarity of the newly added data and the existing data in the knowledge base, and the data with a feature vector similarity of ≥0.95 is considered as repeated data. The repeated data triggers a reminder and sends a deduplication notification to the administrator, and the data with a higher quality score is retained. Model feedback: the matching accuracy of the newly added data is calculated every month, and if the accuracy decreases by ≥5%, the GAT model is fine-tuned, the newly added data is used as the training set, and the model is updated after 50 rounds of iteration. The updated model is deployed to the production environment.

[0122] 105. receiving new part data, obtaining part features of the new part data, and matching corresponding target blank holder control lines and target process supplemental surface section lines in the knowledge base according to the part features.

[0123] It can be understood that the new part data is received, the part features of the new part data are obtained, and the standardized preprocessing process identical to steps 101 and 102 is performed. Format adaptation: STL, STEP, PLY, and other input formats are supported, the format type is automatically identified and corresponding conversion is performed, and the conversion parameters are consistent with historical data, such as Poisson reconstruction depth 10. Noise removal: the same two-stage denoising strategy is adopted, the statistical filtering neighborhood point is 50, the Laplacian smoothing iteration is 3 times, and global noise detection is additionally performed for possible scanning noise of the new part, such as consistent error of batch scanning. If the proportion of noise points is >5%, re-scan is prompted. Normalization: the barycenter of the new part is taken as the origin, the same scaling factor calculation method as the historical data is adopted, and the data is scaled to [-1, 1] space. The scaling parameters are recorded for subsequent result inverse transformation. Consistency checking: the deviation of the data format, point cloud density, and other indicators after preprocessing from the historical data is compared, and the deviation ≤5% is considered as qualified, and the unqualified data is adjusted and reprocessed.

[0124] Feature extraction and standardization: Perform feature extraction with full-process replication + real-time verification. Feature preprocessing: Generate 10 standardized images, render parameters, enhancement methods consistent with history, calculate 18-dimensional geometric feature vector, parameter definition consistent with history; Type identification: Call the same ResNet-50+Transformer model, use the latest fine-tuned weights, output new part type and confidence, confidence <0.7, generate feature enhancement suggestions, such as adding scanning angles; Target feature extraction: Improved SIFT algorithm, parameters unchanged + curvature constraint, rule base updated synchronously, extract target features and label; Section line parameter calculation: The cutting position strategy, fitting algorithm, and parameter calculation method are completely consistent with step 3.2; Feature standardization: Use standardized parameters stored in the knowledge base, such as historical mean and standard deviation, to ensure feature space consistency, output new part complete feature set, including type, target feature, geometric parameter, confidence, quality score.

[0125] Initial matching, type screening: Use the type identification result of the new part as the search condition to filter out historical part data of the same type in the knowledge base. Types with a confidence score of 0.8 or higher are considered valid search conditions. If the confidence score is below the threshold, expand the search to similar types.

[0126] Precise matching, feature similarity calculation: Extract the feature set of the historical part after screening and the feature set of the new part, and calculate the similarity using the weighted cosine similarity algorithm. The part type weight is 0.3, the target feature weight is 0.4, and the geometric parameter weight is 0.3. Set the similarity threshold to 0.85 and filter out historical data that meets the conditions.

[0127] Optimal matching result determination: Sort the results obtained by precise matching in descending order of similarity, take the top 3 results as candidate matching items, and perform secondary screening based on process feasibility, such as material compatibility and forming equipment adaptability. Finally, output the optimal matching result with the highest similarity and process feasibility.

[0128] Output structured matching report, including new part information, ID, type, feature list, optimal matching historical part information, matching similarity, pressure surface control line parameters, coordinates, length, draft angle, process supplementary surface section line parameters, radius, central angle, contour graph, support PDF format export and database storage.

[0129] The knowledge base stores according to part type-geometric feature-control line-pressing surface, part type-geometric feature-section line / guide line-process supplementary surface, and the like. If a new part comes, the corresponding control line and process supplementary surface section line / guide line are matched according to the part features in the knowledge base, and some adaptive change algorithms are also needed. The corresponding boundary line is obtained by mapping with the control line. The guide line also has a set of corresponding rules to connect the nodes of the section line. The final pressing surface and supplementary surface are process surfaces matched according to the existing knowledge base. The data in the knowledge base are experience values that have been verified for formability, so compared with artificial generation, the situation of many formability problems in artificial generation is greatly improved.

[0130] The beneficial effects are that the PointNet++ point cloud segmentation model combined with the standardized preprocessing process accurately splits the three types of part surfaces, solves the boundary ambiguity problem in traditional segmentation, and improves the surface segmentation accuracy. The ResNet-50+Transformer multi-modal fusion architecture integrates visual and geometric features, and the part type recognition confidence is above 0.85, which improves the accuracy of single modal model by 15%. The improved SIFT+curvature constraint algorithm screens the target features through the curvature threshold, and the cross section line is accurately cut and calculated by combining the parameters, so that the fitting error of the pressing surface control line and the process supplementary surface section line is controlled within 0.03 mm, which greatly reduces the parameter deviation of artificial design and reduces the process trial and error cost. The feature atlas constructed based on the GAT graph neural network realizes multi-dimensional data association, and the structured degree of knowledge storage is improved. When matching new parts, based on standardized feature extraction and atlas retrieval, the matching response time is shortened to seconds, the matching accuracy is improved, and the time-consuming and missing problems of traditional manual retrieval are avoided. The part process design efficiency is significantly improved.

[0131] Please refer to Figure 2 , Figure 2 The embodiment flowchart of the third knowledge base automatic recognition and matching method provided in the embodiments of the application includes:

[0132] 201, call the PointNet++ point cloud segmentation model based on part surface segmentation optimization, the first layer sampling radius is 0.05mm, the number of neighborhood points is 32, which is used to extract local geometric features at the patch level; the second layer sampling radius is 0.1mm, the number of neighborhood points is 64, which is used to fuse adjacent local features to form component features; the third layer sampling radius is 0.2mm, the number of neighborhood points is 128, which is used to integrate component features to obtain global surface features;

[0133] 202. After extracting features from the preprocessed part data using the PointNet++ point cloud segmentation model, a segmentation label library is constructed. Three types of surface feature indicators are defined, and the data is initially grouped by K-Means clustering. Then, the data is input into a fully connected layer and classified using the Softmax activation function to obtain the surface data.

[0134] 203. The product surface of the profile data is located in the middle part, the pressing surface is located in the outermost circle, and the process supplement surface is located in the middle. The pressing surface is the entire large surface with a radius greater than 50mm. There is a closed and continuous arc surface close to the pressing surface. The process supplement surface starts from the arc surface of the pressing surface. The arc surface connecting the product surface is not a complete circle. Find the boundary line that is continuous with the contour through the arc surface. The surface between the two parts is the process supplement surface, and the middle part is the product surface.

[0135] Based on the method provided in this application, significant performance improvements are achieved through precise design. The optimized PointNet++ point cloud segmentation model employs three-layer differential sampling, extracting local to global features layer by layer, balancing detail capture with overall feature integration. Combining a segmentation label library with a clustering + classification process, and using K-Means for initial grouping and Softmax for classification, the accuracy of surface segmentation is significantly improved, resolving the problem of ambiguous segmentation boundaries in traditional methods. The spatial distribution and characteristics of the three types of surfaces are clearly defined, enabling precise positioning of product surfaces, pressing surfaces, and process supplementary surfaces. This provides highly reliable data support for subsequent feature extraction and process parameter design, reducing manual correction costs and significantly improving the efficiency and accuracy of subsequent processes.

[0136] Please see Figure 3 , Figure 3 This is a flowchart illustrating an embodiment of the automatic identification and matching method for the second knowledge base provided in this application, including:

[0137] 301. Render the surface data from multiple perspectives using the Blender rendering engine to generate 10 standardized images. Calculate the geometric feature vectors based on the standardized images. The geometric feature vectors should include at least the outer contour of the part, the number of holes, the complexity and geometry of the main cross-section line, and the size of the envelope.

[0138] 302. Using a pre-trained ResNet-50 model, features are extracted step by step through convolutional layers and pooling layers. The feature vector output by the penultimate layer is obtained. Average pooling is performed on the feature vector to retain the feature vectors from each viewpoint and generate a visual feature vector.

[0139] 303. Map the geometric feature vectors to feature vectors through a fully connected layer, and concatenate them with the visual feature vectors to form a multimodal feature vector;

[0140] 304、input the geometric feature vector into the Transformer encoder, the Transformer encoding adopts a 6-layer encoder, the number of multi-head attention heads is 8, global average pooling is performed on the encoded feature matrix to obtain a fusion feature vector;

[0141] 305、the multi-modal feature vector and the fusion feature vector are output through a fully connected layer to obtain a part type recognition result, the part type recognition result at least includes a hood outer plate, a hood inner plate, a front door outer plate, a front door inner plate, a fender, a side wall, a wheel cover, an A column and a B column.

[0142] Based on the method provided in the embodiments of the present application, multi-link optimization is used to achieve accurate identification of part types. Blender rendering generates 10 standardized images, combined with multi-dimensional geometric feature calculation, to ensure comprehensive and standardized feature coverage and avoid missing single features. Pre-trained ResNet-50 extracts visual features, average pooling retains perspective information, and improves feature representation capability. Geometric and visual feature splicing and Transformer encoding, 6-layer encoder + 8-head attention strengthens feature association, and the fused features after global pooling are more recognizable. Finally, the fully connected layer outputs adapt to various body part types, significantly improves the recognition accuracy, and reduces the cost of manual feature design.

[0143] The above describes the automatic identification and matching method of the knowledge base in the embodiments of the present application, and the automatic identification and matching device of the knowledge base in the embodiments of the present application is described below. Please refer to Figure 4 , Figure 4 The structure diagram of the automatic identification and matching device of the knowledge base provided in the embodiments of the present application includes:

[0144] The surface data splitting module is used to obtain original surface data, and after preprocessing the original surface data, the preprocessed original surface data is split into three parts of surface by using the PointNet++ point cloud segmentation model to obtain surface data.

[0145] The part type identification module is used to identify the part type of the surface data by using the ResNet-50+Transformer multi-modal feature fusion model to obtain a part type recognition result.

[0146] The part feature extraction module is configured to identify part features of the part based on the part type identification result by using an improved SIFT+curvature constraint feature extraction algorithm, break the surface line in the part features along the contour vertical direction at the step and corner of the part product surface contour, and obtain a process supplementary surface cross-section line; break the surface line along the plane parallel direction based on the four corners of the part product surface contour, and obtain a blank holder surface control line; extract straight lines and circular arcs of the broken surface line in the process supplementary surface cross-section line and the blank holder surface control line, calculate the length and draft angle of the straight lines, and calculate the radius and central angle of the circular arcs, to obtain a target blank holder surface control line and a target process supplementary surface cross-section line;

[0147] The feature map construction module is configured to construct a feature map based on a GAT graph neural network, associate the part type identification result, the part features, the target blank holder surface control line, and the target process supplementary surface cross-section line, and store the associated matching result in a knowledge base.

[0148] The part association matching module is configured to receive new part data, obtain part features of the new part data, and match corresponding target blank holder surface control lines and target process supplementary surface cross-section lines in the knowledge base according to the part features.

[0149] Based on the device provided in the present application, the original mold surface data is obtained by the mold surface data splitting module, the PointNet++ point cloud segmentation model is used in combination with the standardized preprocessing process to accurately split the three types of part surfaces, the boundary ambiguity problem in traditional segmentation is solved, and the accuracy of surface segmentation is improved. The ResNet-50+Transformer multi-modal fusion architecture in the part type identification module integrates visual and geometric features, and the part type identification confidence is above 0.85, which is 15% higher than the accuracy of a single modal model. The improved SIFT+curvature constraint algorithm is used to screen target features through a curvature threshold, and the blank holder surface control line and the process supplementary surface cross-section line are accurately cut and parameter calculated, so that the fitting error is controlled within 0.03 mm, the parameter deviation of manual design is greatly reduced, and the process trial and error cost is reduced. The feature map constructed based on the GAT graph neural network realizes multi-dimensional data association, and the structured degree of knowledge storage is improved. When matching new parts, based on standardized feature extraction and graph retrieval, the matching response time is shortened to seconds, the matching accuracy is improved, the time-consuming and missing problems of traditional manual retrieval are avoided, and the part process design efficiency is significantly improved.

[0150] Please refer to Figure 5 , Figure 5is a structural schematic diagram of the knowledge base automatic identification and matching device provided in the embodiments of the present application. The knowledge base automatic identification and matching device 500 can have great differences in configuration or performance, and can include one or more processors 510, for example, one or more processors and memories 520, one or more storage media 530 storing application programs 533 or data 532. The memories 520 and the storage media 530 can be temporary storage or persistent storage. The programs stored in the storage media 530 can include one or more modules, which are not shown in the figure, and each module can include a series of instruction operations in the knowledge base automatic identification and matching device 500. Furthermore, the processor 510 can be configured to communicate with the storage media 530 and execute the series of instruction operations in the storage media 530 on the knowledge base automatic identification and matching device 500.

[0151] The knowledge base automatic identification and matching device 500 can further include one or more power supplies 540, one or more wired or wireless network interfaces 550, one or more input and output interfaces 560, and / or one or more operating systems 531, such as Windows Server, MacOS X, Unix, Linux, FreeBSD, and the like. Those skilled in the art can understand that the knowledge base automatic identification and matching device 500 can further include other components, which are not shown in the figure, for example, bus lines, a memory, a processor, and the like. Figure 5 The structure of the knowledge base automatic identification and matching device shown in the figure does not constitute a limitation on the knowledge base automatic identification and matching device, and can include more or fewer components than those shown in the figure, or combine certain components, or arrange different components.

[0152] The present application also provides a knowledge base automatic identification and matching device, which includes a memory and a processor. The memory stores computer readable instructions. When the computer readable instructions are executed by the processor, the processor executes the steps of the knowledge base automatic identification and matching method in each of the embodiments.

[0153] The present application also provides a computer readable storage medium, which can be a non-volatile computer readable storage medium or a volatile computer readable storage medium. The computer readable storage medium stores instructions. When the instructions are executed on a computer, the computer executes the steps of the knowledge base automatic identification and matching method.

[0154] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the system, device and unit described above can refer to the corresponding processes in the foregoing method embodiments, which will not be described herein.

[0155] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device, which can be a personal computer, a server, or a network device, to execute all or part of the steps of the methods of the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory, ROM, a random access memory, RAM, a magnetic disk or an optical disk, and various program codes that can be stored in the medium.

[0156] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in a distributed computing environment in which tasks are performed by remote processing devices connected by a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0157] The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for automatic identification and matching of knowledge bases, characterized in that, The automatic identification and matching method of the knowledge base includes the following steps: The original model surface data is obtained. After preprocessing the original model surface data, the PointNet++ point cloud segmentation model is used to split the preprocessed original model surface data into three parts to obtain the surface data. The surface data is used to identify the part type, and the part type identification result is obtained by using the ResNet-50+Transformer multimodal feature fusion model. Based on the part type identification results, the part features are identified using an improved SIFT+curvature constraint feature extraction algorithm. At the steps and corners of the part's product surface contour, cut lines are made along the vertical direction of the contour to obtain the process supplementary surface section lines. Using the four vertices of the part's product surface contour as a basis, cut lines are made along the plane parallel direction to obtain the blanking surface control lines. The straight lines and arcs of the cut lines from the process supplementary surface section lines and blanking surface control lines are extracted. The length of the straight lines and the draft angle are calculated, and the radius and central angle of the arcs are calculated to obtain the target blanking surface control lines and the target process supplementary surface section lines. Based on the GAT graph neural network, a feature map is constructed, and the part type identification result, part features, target blanking surface control line and target process supplementary surface section line are associated. The association matching result is stored in the knowledge base. Receive new part data, acquire the part features of the new part data, and match the corresponding target blanking surface control line and target process supplementary surface section line in the knowledge base according to the part features; The process involves acquiring the original model surface data, preprocessing the original model surface data, and then using the PointNet++ point cloud segmentation model to split the preprocessed original model surface data into three parts, resulting in surface data, including: The PointNet++ point cloud segmentation model based on part surface segmentation optimization is called. The first layer has a sampling radius of 0.05 mm and 32 neighboring points, which is used to extract local geometric features at the surface level; the second layer has a sampling radius of 0.1 mm and 64 neighboring points, which is used to fuse adjacent local features to form part features; the third layer has a sampling radius of 0.2 mm and 128 neighboring points, which is used to integrate part features to obtain global surface features. After extracting features from the preprocessed part data using the PointNet++ point cloud segmentation model, a segmentation label library is constructed. Three types of surface feature indicators are defined, and K-Means clustering is used for preliminary grouping. The data is then input into a fully connected layer and classified using the Softmax activation function to obtain the surface data. The product surface of the profile data is located in the middle part, the blank holder surface is located in the outermost circle, and the process supplement surface is located in the middle. The blank holder surface is the entire large surface with a radius greater than 50mm. There is a closed and continuous arc surface close to the blank holder surface. The process supplement surface starts from the arc surface of the blank holder surface. The arc surface connecting the product surface is not a complete circle. The boundary line that is continuous with the contour is found through the arc surface. The surface between the two parts is the process supplement surface, and the middle part is the product surface.

2. The method of automatic identification and matching of knowledge bases of claim 1, wherein, The original mold surface data is acquired, the original mold surface data is preprocessed, and the preprocessed original mold surface data is split into three part surfaces by using a PointNet++ point cloud segmentation model to obtain surface data, including: Obtaining original mold surface data, including part product surface, blank holder surface and process supplementary surface data, generating surface patches of the original mold surface data by using an adaptive triangulation algorithm, performing straight-through filtering on the surface patches, and then generating a triangular patch model by using a Poisson reconstruction algorithm to obtain standard part data; Performing Laplace smoothing on the triangular patch model in the standard part data, detecting the feature retention degree after smoothing, and if the feature loss is too large, backtracking to the previous iteration result, separately performing constraint smoothing on the blank holder surface and the process supplementary surface to obtain denoised part data; Taking the part barycenter as the origin, calculating the maximum three-dimensional direction size of the part in the denoised part data, scaling the model to the three-dimensional space in a proportional manner, recording the scaling factor and the original coordinate origin, and obtaining preprocessed part data.

3. The method of claim 1, wherein the step of automatically identifying and matching knowledge bases comprises the steps of: identifying a plurality of knowledge bases; and matching the plurality of knowledge bases to one another. 5 The part type recognition result is obtained by identifying the part type of the surface data by using a ResNet-50+Transformer multi-modal feature fusion model, including: Based on the Blender rendering engine, the surface data is rendered in multiple perspectives to generate 10 standardized images, and a geometric feature vector is calculated based on the standardized images, including at least part product contour, punching number, complexity and geometric shape of the main cross section line, and envelope size; A pre-trained ResNet-50 model is used to extract features step by step through convolutional layers and pooling layers, and a feature vector is obtained from the output of the second-to-last layer, and average pooling is performed on the feature vector to generate a visual feature vector; The geometric feature vector is mapped to a feature vector through a fully connected layer, and the visual feature vector is spliced to form a multi-modal feature vector; The geometric feature vector is input into a Transformer encoder, the Transformer encoding uses a 6-layer encoder, and the number of multi-head attention heads is 8. Global average pooling is performed on the encoded feature matrix to obtain a fusion feature vector; The multi-modal feature vector and the fusion feature vector are output through a fully connected layer to obtain a part type recognition result, including at least a hood outer plate, a hood inner plate, a front door outer plate, a front door inner plate, a fender, a side wall, a wheel cover, an A-pillar, and a B-pillar.

4. The method for automatic identification and matching of knowledge bases of claim 1, wherein, Based on the part type recognition result, the part features of the part are identified by using an improved SIFT+curvature constraint feature extraction algorithm. The step and corner of the part product contour are broken along the contour vertical direction to obtain a process supplementary surface cross section line. The vertices around the part product contour are used as the basis to break the surface line in the parallel direction of the plane to obtain a blank holder surface control line. The straight lines and circular arcs of the process supplementary surface cross section line and the blank holder surface control line are extracted, the length and draft angle of the straight lines are calculated, and the radius and central angle of the circular arcs are calculated to obtain target blank holder surface control lines and target process supplementary surface cross section lines, including: Based on the part type identification results, the part features are identified using an improved SIFT+curvature constraint feature extraction algorithm. A scale space is constructed using Gaussian difference, increasing the number of Gaussian pyramid layers to 8. The calculation formula is as follows: wherein, denotes an input image pixel value, denotes a two-dimensional Gaussian function, denotes a Gaussian blur standard deviation, denotes a Gaussian blurred image, denotes a scale factor, denotes a Gaussian difference image, denotes a horizontal direction pixel position of a corresponding image, a corresponding vertical direction pixel position; Sub-pixel level feature point coordinates are determined through extreme point detection and interpolation calculations. Low-contrast points and edge response points are removed. The directional gradient of the feature point neighborhood is calculated using the following formula: in, This represents the gradient magnitude, and indicates the pixel value. The intensity of the brightness change at that location. Indicates the gradient direction, representing the pixel. The direction and range of brightness change (0°-360°). , , and Representing pixels Gaussian blur values ​​for right, left, top, and bottom; The discrete surface curvature calculation method is adopted. For each feature point, 30 neighborhood points are selected to fit the local surface, and the Gaussian curvature and average curvature are calculated. The curvature values ​​of the feature points are compared with the rule base. Feature points that meet the conditions are retained and false detection points are removed to obtain the part features.

5. A method of automatic identification and matching of knowledge bases as claimed in claim 4, wherein, Based on the part type identification result, the part features are identified by the improved SIFT+curvature constraint feature extraction algorithm. In the part features, the part product surface contour is cut along the vertical direction of the contour at the step and corner to obtain the process supplement surface section line; the part product surface contour is cut along the plane parallel direction based on the four vertices to obtain the pressure surface control line. Extract the straight lines and arcs from the process supplement surface section lines and the blank holder surface control lines, calculate the lengths and draft angles of the straight lines, and calculate the radii and central angles of the arcs to obtain the target blank holder surface control lines and the target process supplement surface section lines. This also includes: In the part features, cut the surface lines at the steps and corners of the part product surface contour along the direction perpendicular to the contour to obtain the process supplement surface section lines; based on the four vertices of the part product surface contour, cut the surface lines 20mm inward along the direction parallel to the plane to obtain the pressure surface control lines; the planes include the XOZ plane and the YOZ plane; The RANSAC algorithm was applied to fit the cross-sectional lines in the process supplementary surface section line and the pressure surface control line. The number of iterations for straight line fitting was 1000, and the interior point threshold was 0.01 mm. The calculation formula is as follows: wherein, , , denotes a straight line parameter, and pixel coordinates of the section line, denotes the distance of a point to a straight line, denotes an inlier threshold value of a straight line, denotes an index of different pixel points on the section line; The circular arc fitting iterations were 1500, and the inlier threshold was 0.015mm. The average distance between the fitted element and the original point cloud was calculated using the following formula: wherein, and denotes the center coordinate of the circular arc, denotes the radius of the circular arc, denotes the distance from the point to the circular arc, denotes the inner point threshold of the circular arc, denotes the coordinate difference of the pixel point from the center of the circle in the x-axis, denotes the coordinate difference of the pixel point from the center of the circle in the y-axis; The line parameters are calculated using the distance formula between two points, and the draft angle is calculated using the angle between the line and the direction perpendicular to the parting surface. The calculation formulas are as follows: } wherein, represents the number of points inside the circular arc, and represents the coordinates of the points inside the circular arc, represents the coordinate difference in horizontal direction between the th point inside and the center of the circle, represents the coordinate difference in vertical direction between the th point inside and the center of the circle, represents the draft angle, represents the vector perpendicular to the parting plane direction, represents the direction vector of the tangent line of the section line, represents the dot product of two vectors, , represents the modulus of two vectors; The central angle is calculated using the angle between the vectors of the lines connecting the two endpoints of the arc to the center of the circle. The formula is as follows: wherein, denotes the central angle of the circle, denotes the horizontal and vertical coordinates of the start point of the arc, denotes the horizontal and vertical coordinates of the end point of the arc, denotes the center coordinates of the arc, and denotes the components of the vector from the center to the start point of the arc on the x and y axes, and denotes the components of the vector from the center to the end point of the arc on the x and y axes, denotes the inverse cosine function, and the feature points are clustered by the Euclidean distance to obtain the target blank holder control line and the target process compensation surface section line.

6. The method for automatic identification and matching of knowledge bases of claim 1, wherein, Based on the part type identification result, a feature map is constructed using a GAT graph neural network. This map associates the part type identification result, part features, target blank holder control line, and target process supplementary surface cross-sectional line. The association and matching results are then stored in a knowledge base, including: By constructing a feature map based on the GAT graph neural network, the attributes of the part type node include part type ID, name, confidence level, and list of key features; the attributes of the target feature node include feature ID, type, type of surface, geometric parameters, extraction confidence level, and fitting accuracy. The associated edge types are part type-target feature, target feature-pressing surface control line and target feature-process supplementary surface cross section line. The GAT graph neural network is executed to perform the association, and the association result is obtained. The association matching result is stored in the knowledge base.

7. An apparatus for automatic identification and matching of knowledge bases, characterized in that, The automatic identification and matching device for the knowledge base includes the following modules: The surface data splitting module is configured to obtain original surface data, and split the original surface data into three parts of surface data by using a PointNet++ point cloud segmentation model after preprocessing the original surface data. The PointNet++ point cloud segmentation model based on part surface segmentation optimization is called, the first layer has a sampling radius of 0.05 mm and a neighborhood point number of 32, and is used to extract local geometric features at the patch level; The second layer has a sampling radius of 0.1 mm and a neighborhood point number of 64, and is used to fuse adjacent local features to form part features; The third layer has a sampling radius of 0.2 mm and a neighborhood point number of 128, and is used to integrate the part features to obtain global surface features; the PointNet++ point cloud segmentation model is used to extract features from the preprocessed part data to construct a segmentation label library, three types of surface feature indicators are defined, K-Means clustering is used for preliminary grouping, and a full connection layer is inputted to classify by using a Softmax activation function, so as to obtain surface data; the product surface of the surface data is located in the middle part, the pressing surface is located in the outermost circle, and the process supplement surface is located in the middle; the pressing surface is a large surface, and the radius is greater than 50 mm; there is a closed and continuous circular arc surface near the pressing surface; the process supplement surface starts from the circular arc surface of the pressing surface, the circular arc surface connected with the product surface is not a complete circle, a boundary line is found through the circular arc surface, the surface between the two parts is the process supplement surface, and the middle part is the product surface; The part type identification module is configured to identify the part type of the surface data by using a ResNet-50+Transformer multi-modal feature fusion model, so as to obtain a part type identification result; The part feature extraction module is configured to identify part features of a part based on the part type identification result by using an improved SIFT+curvature constraint feature extraction algorithm, to break the surface line in the vertical direction of the contour at the step and corner of the part product surface contour in the part features, so as to obtain a process supplement surface cross-section line; and to break the surface line in the parallel direction of the plane based on the four vertices of the periphery of the part product surface contour, so as to obtain a pressing surface control line; Straight lines and circular arcs of the broken surface lines in the process supplement surface cross-section line and the pressing surface control line are extracted, the length and draft angle of the straight lines are calculated, and the radius and central angle of the circular arcs are calculated, so as to obtain a target pressing surface control line and a target process supplement surface cross-section line; The feature atlas construction module is configured to construct a feature atlas based on a GAT graph neural network, to associate the part type identification result, the part features, the target pressing surface control line and the target process supplement surface cross-section line, and to store the associated matching result in a knowledge base; The part association matching module is configured to receive new part data, to obtain part features of the new part data, and to match corresponding target pressing surface control lines and target process supplement surface cross-section lines in the knowledge base according to the part features.

8. An apparatus for automatic identification and matching of knowledge bases, characterized by The automatic identification and matching device of the knowledge base comprises a memory and at least one processor, and instructions are stored in the memory. The at least one processor invokes the instructions in the memory to cause the knowledge base auto-identification and matching device to perform the individual steps of the knowledge base auto-identification and matching method of any of claims 1-6.

9. A computer-readable storage medium having stored thereon instructions, the computer-readable storage medium comprising: The instructions, when executed by a processor, implement the individual steps of the knowledge base auto-identification and matching method of any of claims 1-6.

Images