A Deep Learning-Based Casting Gating Design Method and System

Abstract

This disclosure relates to the field of precision casting technology, and in particular to a casting gating design method and system based on deep learning. It employs automated feature extraction to automatically extract geometric, topological, and process features from the STP format 3D model of the casting and gating system. A deep learning model is constructed to learn the complex mapping relationship between casting features and the gating system structure. Expert experience is integrated; by pre-learning standard models and structure libraries, casting process rules and expert knowledge are embedded into the AI ​​model, improving the rationality and interpretability of the design. End-to-end generation is adopted, achieving full automation from the input of a new casting model to the output of a complete 3D model of the gating system. This disclosure significantly improves design efficiency, achieving automated design from casting to gating system, shortening the traditional design cycle of several days to weeks to minutes. The design quality of this disclosure is controllable; by embedding standard models, human experience, and process rules, it ensures that the design meets casting process requirements.

Description

Technical Field

[0001] This disclosure relates to the field of precision casting technology, and in particular to a casting gating design method and system based on deep learning. Background Technology

[0002] The design of the gating system is the first step in the precision casting process design of structural components, laying the foundation for the success of subsequent process design and directly affecting the quality of the casting. Currently, in the casting industry, the design of gating systems has long relied on manual experience. Designers need to refer to gating system models of similar castings from the past for manual comparison, fine-tuning, and optimization. This process is time-consuming, inefficient, and limited by personal experience, making it difficult to guarantee the consistency and optimality of the design.

[0003] While existing technologies, particularly deep learning-based generative models, have shown promise in industrial design and have seen initial applications in 3D geometric modeling and structural optimization, still face the following challenges in complex engineering scenarios such as gating system design:

[0004] Feature extraction relies on manual methods, and existing methods often depend on manually defining features or simple geometric parameters, making it difficult to fully capture the complex geometric and topological relationships of castings and gating systems.

[0005] The model has poor interpretability. Traditional deep learning models are "black box" structures, the design decision-making process is untraceable, and it is difficult to incorporate expert knowledge and physical constraints of the casting process.

[0006] The system integration is low. Existing AI design tools are often independent of CAD systems and cannot directly read, parse and output industry standard 3D model formats (such as STP), resulting in a broken design process.

[0007] In summary, this invention proposes an intelligent design method for gating systems based on generative AI and deep integration of expert experience. By constructing an intelligent mapping model between the casting and the gating system, it achieves automated and high-precision generation from the three-dimensional model of the casting to the design of the gating system. Summary of the Invention

[0008] To address the aforementioned issues, this disclosure provides a casting pouring design method and system based on deep learning.

[0009] Firstly, a deep learning-based casting design method, the method comprising:

[0010] Obtain the three-dimensional model of the casting and the corresponding gating system, and convert the three-dimensional model into point cloud data or voxelized features of the casting, as well as point cloud data of the gating system.

[0011] Based on casting point cloud data or voxelized features, casting features are extracted and combined into a casting feature vector. Casting features include geometric features, structural features, component features, and member features. The structural features combine RANSAC and geometric threshold to determine the layer structure. The member features are used to count blades by statistically analyzing the peak values ​​of the angle histogram.

[0012] Based on the point cloud data of the casting system, the features of the casting system are extracted and composed into a casting system feature vector. The features of the casting system include casting component type, component quantity, component location, component connection relationship construction and component size variation parameters.

[0013] A neural network model is established, with the feature vector of the casting as input and the feature vector of the gating system as output. The neural network model is trained to obtain the casting casting generation model.

[0014] The casting features are extracted from the new casting and combined to form a new casting feature vector. The new casting feature vector is then input into the casting casting generation model to obtain a new casting system feature vector, which is used to generate a three-dimensional model of the casting system.

[0015] Furthermore, the 3D model is converted into casting point cloud data or voxelized features, as well as gating system point cloud data and / or voxelized features, and then includes:

[0016] The point cloud data of the casting and the point cloud data of the gating system are aligned and normalized.

[0017] Furthermore, the casting features are extracted, including:

[0018] Principal component analysis is used to align the point cloud, the principal orientation of the point cloud is calculated, and the principal axis is aligned with the Z-axis of the height direction by rotation.

[0019] Further geometric features include: center point, diameter, and height;

[0020] The center point is used to characterize the center of the casting; the average value of all points is calculated.

[0021] The diameter is used to represent the farthest distance between two points in the circumference. It is obtained by calculating the radial distance from all points to the center point and taking twice the maximum value.

[0022] Height, used to characterize the farthest distance between two points in the height direction, is obtained by the difference between the maximum and minimum values ​​in the height direction.

[0023] Further structural features include: number of layers, layer thickness, interlayer spacing, and layer height difference;

[0024] The structure combines RANSAC and geometric threshold decision layer, including: performing point cloud split volume segmentation, using RANSAC plane fitting to extract planar regions, identifying corners by detecting changes in the angle of neighboring points or protrusions, and segmenting the point cloud into multiple split volumes with the corners as boundaries;

[0025] Explore radially outward from the center to identify the number of layers. For each split body, calculate the radial width and height. If the radial width is less than the width threshold and the Z height is greater than the height threshold, it is determined to be one layer.

[0026] Layer thickness, layer spacing, and layer height difference are obtained as follows: layer thickness is the radial width of the split body; layer spacing is the distance between two adjacent layers, with each split body surrounded by a convex hull, and the distance between the convex hulls is used as the layer spacing; layer height difference is the height difference between the split bodies.

[0027] Furthermore, component characteristics, including: number of blades, presence of flanges, flange width, and flange thickness, are obtained through the following methods:

[0028] The number of blades is counted by peak statistics of the angle histogram. Specifically, this includes: identifying circumferential repeating structures; for each segment, calculating the angle of each point relative to the center, constructing a 360-degree angle histogram, counting the number of peaks by peak detection, and accumulating the peaks of all segments as the number of blades.

[0029] Flange inspection is based on square ring structure identification. For each split body, the numerical range of radial width and height is checked, and the neighborhood normal vector of the sampling point of the split body is calculated. If the consistency of the normal vector exceeds the normal vector threshold, it is determined to be a flange.

[0030] The radial width of the split body is used as the flange width, and the height of the split body is used as the flange thickness.

[0031] Furthermore, the component features, including the location and quantity of flanges, inner rings, outer rings, and blades, are obtained through the following methods:

[0032] By calculating the point cloud of each standard structure—flange, inner ring, outer ring, and blade—and the directed Hausdorff distance to the point cloud of the casting, the similarity is converted using the following formula:

[0033] Similarity = 1 / (1 + directed Hausdorff distance);

[0034] If the similarity exceeds the similarity threshold, it is determined to be a match, and the structure location and number are recorded; the point cloud center is used as the structure location.

[0035] Furthermore, based on the point cloud data of the casting system, features of the casting system are extracted, including:

[0036] The existing gating system is divided into gating system, riser, and inlet pipe, and standardized by classification to construct a standard point cloud model of gating system, riser, and inlet pipe;

[0037] Calculate the nearest distance from each point in the gating system point cloud to the casting point cloud; if the distance is less than the distance threshold, it is considered to overlap with the casting, and after filtering, a pure gating system point cloud is obtained;

[0038] Calculate the directed Hausdorff distance between the point cloud distribution of the pure gating system and the standard point cloud models of the gating system, risers, and inlet pipes;

[0039] Randomly sample several points, calculate the convex hull volume consistency and surface normal vector consistency, and obtain the geometric consistency by weighted summation of the convex hull volume consistency and surface normal vector consistency;

[0040] The Hausdorff distance is weighted using geometric consistency to obtain a weighted distance, and the model with the smallest weighted distance is selected as the best match.

[0041] Furthermore, extracting features of the gating system also includes:

[0042] The number of components is calculated using a clustering algorithm based on a distance threshold. A KD tree is constructed for the point cloud, and a neighborhood search is performed. Points with a distance less than the threshold are grouped into the same cluster by depth-first traversal. The number of clusters and their center positions are counted as the number of components.

[0043] The component position is determined by establishing a three-dimensional spatial coordinate system with the Z-axis representing the height direction, and the component position is represented by the three-dimensional spatial coordinate system.

[0044] For each runner, calculate the Euclidean distance to all risers, select the nearest riser, and establish the component connection relationship between the runner and the riser.

[0045] Dimensional variation parameters are recorded for each component, including scaling factor, bending angle, and elongation factor, to describe the dimensional changes of the actual component relative to the standard component.

[0046] Furthermore, a neural network model is established, using the casting feature vector as input and the gating system feature vector as output. The neural network model is trained to obtain a casting casting generation model, including:

[0047] A multi-layer fully connected neural network is used, with the input being the feature vector of the casting and the output being the feature vector of the gating system;

[0048] A supervised learning dataset is constructed using the extracted feature vectors of the casting and the feature vectors of the gating system.

[0049] By using the mean squared error loss function and optimizing the network parameters through backpropagation, the mapping from casting features to gating system features is achieved.

[0050] Furthermore, a three-dimensional model of the gating system is generated, including:

[0051] The features of the new casting system are obtained based on the feature vector of the new casting system, and the corresponding components are selected from the standard model library;

[0052] Based on the position and size change parameters, the components are translated, rotated, and scaled.

[0053] Combine the components to generate a complete 3D model of the gating system.

[0054] Secondly, a deep learning-based casting gating design system includes:

[0055] Point cloud data processing unit, casting feature extraction unit, gating system feature extraction unit, casting gating generation model construction unit, and gating system 3D model generation unit;

[0056] The point cloud data processing unit is used to acquire the three-dimensional model of the casting and the corresponding gating system, and convert the three-dimensional model into point cloud data or voxelized features of the casting, as well as point cloud data of the gating system.

[0057] The casting feature extraction unit extracts casting features based on casting point cloud data or voxelized features, and combines the casting features into a casting feature vector. The casting features include geometric features, structural features, component features, and component features. The structural features combine RANSAC and geometric threshold decision layer structure. The component features are counted by peak statistics of angle histogram.

[0058] The gating system feature extraction unit extracts gating system features based on gating system point cloud data and assembles these features into a gating system feature vector. The gating system features include gating component type, component quantity, component location, component connection relationship construction, and component size variation parameters.

[0059] The casting pouring generation model building unit is used to establish a neural network model. It takes the feature vector of the casting as input and the feature vector of the pouring system as output, trains the neural network model, and obtains the casting pouring generation model.

[0060] The gating system 3D model generation unit is used to extract casting features from the new casting and form a new casting feature vector. The new casting feature vector is then input into the casting gating generation model to obtain a new gating system feature vector, which is used to generate a 3D model of the gating system.

[0061] This disclosure includes at least the following beneficial effects:

[0062] This disclosure employs automated feature extraction to automatically extract geometric, topological, and technological features from the STP format 3D model of the casting and gating system. A deep learning model is constructed to learn the complex mapping relationship between casting features and the gating system structure.

[0063] This publication integrates expert experience, embedding casting process rules and expert knowledge into the AI ​​model through pre-learning standard models and structure libraries, thereby improving the rationality and interpretability of the design. It employs end-to-end generation, achieving full automation from the input of a new casting model to the output of a complete 3D model of the gating system.

[0064] Other features and advantages of this disclosure will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures pointed out in the description and the accompanying drawings. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a schematic diagram of the design method flow according to an embodiment of the present disclosure;

[0067] Figure 2 This is a schematic diagram illustrating the design principle of the embodiment of the present disclosure;

[0068] Figure 3 This is a schematic diagram of casting feature extraction according to an embodiment of the present disclosure;

[0069] Figure 4 This is a schematic diagram of the gating system according to an embodiment of the present disclosure;

[0070] Figure 5 A schematic diagram of the system architecture designed for an embodiment of this disclosure. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0072] like Figure 1As shown, a deep learning-based casting design method is described, the method comprising:

[0073] S101, acquire the three-dimensional model of the casting and the corresponding gating system, and convert the three-dimensional model into point cloud data or voxelized features of the casting, as well as point cloud data of the gating system.

[0074] S102, based on casting point cloud data or voxelized features, extract casting features and assemble casting features into a casting feature vector; among which, casting features include geometric features, structural features, component features and member features; the structural features combine RANSAC and geometric threshold decision layer structure; the member features are counted by peak statistics of angle histogram;

[0075] S103, Based on the point cloud data of the casting system, extract the features of the casting system and form a casting system feature vector; wherein, the casting system features include casting component type, component quantity, component location, component connection relationship construction and component size change parameters;

[0076] S104. Establish a neural network model, using the feature vector of the casting as input and the feature vector of the gating system as output, train the neural network model to obtain the casting casting generation model;

[0077] S105: Extract the casting features of the new casting and form a new casting feature vector. Input the new casting feature vector into the casting casting generation model to obtain a new casting system feature vector, which is used to generate a three-dimensional model of the casting system.

[0078] In practical implementation, the principle is as follows: Figure 2 As shown, the introduction is as follows:

[0079] S101, 3D model reading and point cloud extraction:

[0080] The software API reads the STP format 3D model of the casting and the corresponding gating system.

[0081] Convert the model into point cloud data or voxelization, and then perform coordinate alignment and normalization.

[0082] If converted to a point cloud, when the number of points in the model point cloud exceeds 20,000, 20,000 points are randomly selected from the model point cloud to represent the model, ensuring the efficiency of subsequent processing.

[0083] S102, Automatic extraction of casting features, such as Figure 3 As shown:

[0084] Based on point cloud data and expert experience, the following key features of the casting are automatically identified using algorithmic logic:

[0085] Geometric features: diameter (the farthest distance between two points in the circumferential direction), height (the farthest distance between two points in the vertical direction), and center point.

[0086] Structural features: number of layers (exploring radially outward from the center), thickness of each layer, interlayer spacing, and layer height difference.

[0087] Component characteristics: number of blades, presence of flanges, flange width and thickness.

[0088] Standard structure matching: Similarity matching is performed with pre-learned standard structures (flange, inner ring, outer ring, blade) to identify their location and quantity.

[0089] S103, Feature identification and modeling of the gating system, such as Figure 4 As shown:

[0090] Standard model library pre-learning: Existing gating systems are manually cut and classified into runners, risers, and inlet pipes, and then uploaded to preload standard gating system models such as JD (runner), MK (riser), and ZZG (inlet pipe) to extract their point cloud features.

[0091] Feature matching and recognition: Using metrics such as Hausdorff distance, the point cloud of the actual casting system is matched with the standard model to identify its type, quantity, location and size variations.

[0092] Component relationship construction: Establish the connection relationship and spatial layout between the gating system, risers, and inlet pipes.

[0093] S104, Construction of Deep Learning Mapping Model:

[0094] Network structure: A multi-layer fully connected neural network is used, with the input being the feature vector of the casting and the output being the feature vector of the gating system.

[0095] Training data: A supervised learning dataset is constructed by pairing existing casting-gating system model data.

[0096] Training process: The mean squared error loss function is used to optimize the network parameters through backpropagation, so as to achieve an accurate mapping from casting features to gating system design.

[0097] Specifically, a multi-layer fully connected neural network is used, with 2 to 10 layers and 32 to 2048 nodes per layer.

[0098] The number of nodes in the input layer is input_dim (dynamic, depending on the dimension of the input features); the input consists of casting point cloud data + geometric features + structural features; ReLU is used as the activation function.

[0099] The number of nodes in the output layer is output_dim (dynamic, depending on the output feature dimension); linear output is used; the output includes the gating system type index, quantity, location coordinates, and size variation parameters;

[0100] The optimizer used was Adam; the loss function used was MSE (mean squared error); the validation set ratio was 20%.

[0101] S105, Generation of 3D Model of Gating System:

[0102] Based on the feature vector of the gating system output from the newly uploaded model, select the corresponding component from the standard model library.

[0103] Confirm the quantity of the injection pipe, gating system, and riser components, and perform transformations such as translation, rotation, and scaling on the components based on their position and size variation parameters.

[0104] Assemble all components, ensuring that the gating pipe, runner, and riser are connected and in contact with the casting, to generate a complete 3D model of the gating system, which can be exported in STP format.

[0105] S102, Automatic Extraction of Casting Features, specifically including:

[0106] Point cloud data preprocessing and alignment:

[0107] First, extract the 3D point cloud data from the STP file. Use Principal Component Analysis (PCA) to align the point cloud, calculate the principal orientation of the point cloud, and rotate it to align the principal axis with the Z-axis to ensure the circumferential surface is flat for subsequent analysis.

[0108] Geometric feature extraction:

[0109] Calculate the center point of the casting (average of all points). The diameter is obtained by calculating the radial distance from all points to the center and taking twice the maximum value. The height is obtained by the difference between the maximum and minimum values ​​in the Z direction.

[0110] Structural feature extraction (layer number detection):

[0111] The layer structure is identified using the split volume method. Point cloud split volume segmentation: Planar regions are extracted using RANSAC plane fitting. Corners are identified by detecting changes in the angle of neighboring points (more than 60 degrees) or protrusions (more than 2 mm). The point cloud is segmented into multiple split volumes with the corners as boundaries.

[0112] Layer identification: For each split element, calculate its radial width (width along the radial direction from the center) and Z-direction height. If the radial width is less than 15 mm and the Z-direction height is greater than 50% of the total height, it is determined to be a layer. The layer thickness is the radial width of the split element, the layer spacing is approximated by the convex hull volume, and the layer height difference is the Z-direction height difference of the split element.

[0113] Component feature extraction:

[0114] The number of blades is detected based on the identification of circumferential repeating structures. For each segment, the angle of each point relative to the center is calculated, and a 360-degree angle histogram is constructed. The number of peaks is counted by peak detection (threshold = average + 0.05 × standard deviation). The peaks of all segments are summed to obtain the number of blades, with an upper limit of 50.

[0115] Flange inspection is based on square ring structure identification. For each segmented body, check if the radial width and Z-height are within the range of 5-25 mm, and if the difference between the two is less than 8 mm. Further verification is performed using convex hull volume (range 100-10000), and a surface consistency check is conducted: calculate the neighborhood normal vector of the sampling point; if the normal vector consistency exceeds 0.7, it is determined to be a flange. Record the flange width (radial width) and thickness (Z-height).

[0116] Standard structural matching:

[0117] Preload point cloud features of standard structures (flange, inner ring, outer ring, blade). For each standard structure, calculate the directed Hausdorff distance to the casting point cloud and convert it into a similarity score (1 / (1+distance)). If the similarity score exceeds 0.3, it is considered a match, and the structure location (point cloud center) and number are recorded.

[0118] S103, Feature identification and modeling of the casting system, specifically including:

[0119] Standard model library pre-learning:

[0120] Preload the standard gating system model (JD gating system, MK riser, ZZG center pipe), extract the point cloud from the STP file, align it using PCA, and store the point cloud features in the standard model library.

[0121] Feature matching and recognition:

[0122] Point cloud differential processing uses a KD-tree fast search to calculate the nearest distance from each point in the gating system point cloud to the casting point cloud. If the distance is less than a threshold (initially 1.0 mm, increasing to 2.0 mm if insufficient), it is considered to overlap with the casting, and after filtering, a pure gating system point cloud is obtained.

[0123] Standard model matching is performed. For each model type (gating system, riser, and inlet pipe), the directed Hausdorff distance between the point cloud of the pure gating system and each standard model is calculated. Geometric consistency is also calculated: 500 points are randomly sampled, and the convex hull volume consistency (1 minus the ratio of volume difference to maximum volume) and surface normal vector consistency are calculated. Overall consistency = volume consistency × 0.6 + surface consistency × 0.4. The Hausdorff distance is weighted using geometric consistency (score = distance × (1 - 0.3 × geometric consistency)), and the model with the lowest score is selected as the best match.

[0124] Component count detection uses a distance-threshold-based clustering algorithm. A KD-tree is constructed on the point cloud, and a neighborhood search is performed with a 20 mm threshold. Points with a distance less than the threshold are grouped into the same cluster by depth-first traversal, ignoring small clusters with fewer than 10 points. The number of clusters and their center positions are counted as the component count.

[0125] Component relationship construction:

[0126] Location calculation:

[0127] There is usually one central injection pipe, located at the center of the casting, slightly higher than the top layer (center Z coordinate + layer height + 10 mm).

[0128] The number of risers is 2, 4, 6, or 8 per layer, and the flange layer has a fixed 4. Each layer is evenly distributed along the circumference, with an angular interval of 360 degrees divided by the number, and the radius is 0.6 times the radius of the casting. The Z coordinate is located at the middle height of the layer.

[0129] The number of gating channels is 2, 4, 6 or 8, located at the bottom layer, evenly distributed along the circumference, with an angular interval of 360 degrees divided by the number, and the radius is 0.8 times the radius of the casting.

[0130] Connection building:

[0131] For each gating system, calculate the Euclidean distance to all risers and select the nearest riser. If the distance is less than 0.1 mm, connect directly; otherwise, force connection to the nearest riser and fine-tune the riser position (raise it by 0.05 mm in the Z direction) to establish the connection mapping relationship between the gating system and the risers.

[0132] Dimensional variation parameters:

[0133] For each component, record the variation parameters, including scaling factors (X, Y, Z directions, range 0.8-1.2), bending angle (range 0-10 degrees), and elongation factors (range 0.8-1.2), to describe the dimensional changes of the actual component relative to the standard model.

[0134] This disclosure achieves automatic extraction and recognition of structured feature parameters from point cloud data, providing a data foundation for subsequent intelligent design of casting systems. Key technologies are described below:

[0135] Point cloud alignment: Use PCA principal axis alignment to ensure coordinate system consistency.

[0136] Volume segmentation: RANSAC plane fitting combined with corner detection to identify geometric structures.

[0137] Layer count detection: Identification of ring structures based on the ratio of radial width to height.

[0138] Blade detection: Peak detection of circumferential angle histogram to identify repeating structures.

[0139] Flange inspection: dimensional constraints, convex hull volume verification, and surface normal vector consistency check.

[0140] Hausdorff distance: used for point cloud similarity matching.

[0141] Geometric consistency: Combining convex hull volume consistency and surface normal vector consistency.

[0142] Clustering detection: Connectivity analysis based on distance thresholds to identify the number of components.

[0143] Spatial layout: Calculated based on the uniform distribution of casting geometric parameters (center, radius, layer height).

[0144] Connection relationship: Nearest neighbor matching algorithm based on Euclidean distance.

[0145] like Figure 5 As shown, a deep learning-based casting gating design system includes:

[0146] Point cloud data processing unit 501, casting feature extraction unit 502, gating system feature extraction unit 503, casting gating generation model construction unit 504, and gating system three-dimensional model generation unit 505;

[0147] The point cloud data processing unit 501 is used to acquire the three-dimensional model of the casting and the corresponding gating system, and convert the three-dimensional model into point cloud data or voxelized features of the casting, as well as point cloud data of the gating system.

[0148] The casting feature extraction unit 502 extracts casting features based on casting point cloud data or voxelized features, and combines the casting features into a casting feature vector; among which, the casting features include geometric features, structural features, component features and component features; the structural features combine RANSAC and geometric threshold determination layer structure; the component features are counted by peak statistics of angle histogram;

[0149] The casting system feature extraction unit 503 extracts casting system features based on the casting system point cloud data and assembles the casting system features into a casting system feature vector; wherein, the casting system features include casting component type, component quantity, component location, component connection relationship construction and component size change parameters;

[0150] The casting pouring generation model building unit 504 is used to establish a neural network model. The neural network model is trained with the casting feature vector as input and the pouring system feature vector as output to obtain the casting pouring generation model.

[0151] The three-dimensional model generation unit 505 of the gating system is used to extract the features of the new casting and form a new casting feature vector. The new casting feature vector is then input into the casting gating generation model to obtain a new gating system feature vector, which is used to generate a three-dimensional model of the gating system.

[0152] This disclosure significantly improves design efficiency, enabling automated design from casting to gating system, reducing the traditional design cycle of several days to weeks to minutes.

[0153] The quality of this design is controllable. By embedding standard models, human experience, and process rules, the design is guaranteed to meet the requirements of the casting process, thereby improving the first-time design pass rate.

[0154] This publicly available system is highly compatible, supports the industry-standard STP format, can be seamlessly integrated with existing CAD / CAE systems, and supports design-simulation closed-loop optimization.

[0155] This publicly available model is highly interpretable, and the design process is based on traceable feature matching and component combination, which facilitates review and adjustment by process engineers.

[0156] To enable those skilled in the art to better understand this disclosure, the principles of this disclosure are explained as follows:

[0157] Using a deep learning network connected to FreeCAD to identify point cloud data of castings and gating systems, the following is an example of a guide vane for a certain aero-engine power turbine:

[0158] Taking a certain aero-engine power turbine guide vane casting as an example: Input the STP model of the casting. The system automatically identifies and extracts its features: diameter φ320mm, height 40mm, 2-layer structure, 35 blades, flange, etc. The model infers and outputs the gating system scheme: ZZG-1 type center pouring pipe ×1, JD-1 type runner ×6, MK-1 type riser ×12. The system automatically generates a 3D model of the gating system. Simulation verification shows stable flow and uniform temperature field, meeting the process requirements.

[0159] By connecting a deep learning network to Open3D, the casting and gating system can be voxelized and features identified. The following is an example of a guide vane for a certain aero-engine power turbine:

[0160] Taking a certain aero-engine power turbine guide vane casting as an example: Input the STP model of the casting. The system automatically identifies and extracts its features: diameter φ320mm, height 40mm, 2-layer structure, 35 blades, flange, etc. The model infers and outputs the gating system scheme: ZZG-1 type center pouring pipe ×1, JD-1 type runner ×6, MK-1 type riser ×12. The system automatically generates a 3D model of the gating system. Simulation verification shows stable flow and uniform temperature field, meeting the process requirements.

[0161] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A casting pouring design method based on deep learning, characterized in that, The method includes: Obtain the three-dimensional model of the casting and the corresponding gating system, and convert the three-dimensional model into point cloud data or voxelized features of the casting, as well as point cloud data of the gating system. Based on casting point cloud data or voxelized features, casting features are extracted and combined into a casting feature vector. Casting features include geometric features, structural features, component features, and member features. The structural features combine RANSAC and geometric threshold to determine the layer structure. The member features are used to count blades by statistically analyzing the peak values ​​of the angle histogram. Based on the point cloud data of the casting system, the features of the casting system are extracted and composed into a casting system feature vector. The features of the casting system include casting component type, component quantity, component location, component connection relationship construction and component size variation parameters. A neural network model is established, with the feature vector of the casting as input and the feature vector of the gating system as output. The neural network model is trained to obtain the casting casting generation model. The casting features are extracted from the new casting and combined into a new casting feature vector. The new casting feature vector is then input into the casting casting generation model to obtain a new gating system feature vector, which is used to generate a three-dimensional model of the gating system. Structural features include: number of layers, layer thickness, interlayer spacing, and layer height difference; The structure combines RANSAC and geometric threshold decision layer, including: performing point cloud split volume segmentation, using RANSAC plane fitting to extract planar regions, identifying corners by detecting changes in the angle of neighboring points or protrusions, and segmenting the point cloud into multiple split volumes with the corners as boundaries; Explore radially outward from the center to identify the number of layers. For each split body, calculate the radial width and height. If the radial width is less than the width threshold and the Z height is greater than the height threshold, it is determined to be one layer. Layer thickness, interlayer spacing, and layer height difference are obtained as follows: layer thickness is the radial width of the split body; interlayer spacing is the distance between two adjacent layers, with each split body surrounded by a convex hull, and the distance between the convex hulls used as the interlayer spacing; layer height difference is the height difference between the split bodies. Based on the point cloud data of the casting system, features of the casting system are extracted, including: The existing gating system is divided into gating system, riser, and inlet pipe, and standardized by classification to construct a standard point cloud model of gating system, riser, and inlet pipe; Calculate the nearest distance from each point in the gating system point cloud to the casting point cloud; if the distance is less than the distance threshold, it is considered to overlap with the casting, and after filtering, a pure gating system point cloud is obtained; Calculate the directed Hausdorff distance between the point cloud distribution of the pure gating system and the standard point cloud models of the gating system, risers, and inlet pipes; Randomly sample several points, calculate the convex hull volume consistency and surface normal vector consistency, and obtain the geometric consistency by weighted summation of the convex hull volume consistency and surface normal vector consistency; The Hausdorff distance is weighted using geometric consistency to obtain a weighted distance, and the model with the smallest weighted distance is selected as the best match.

2. The casting pouring design method based on deep learning according to claim 1, characterized in that, The process involves converting the 3D model into casting point cloud data or voxelized features, as well as gating system point cloud data and / or voxelized features, followed by: The point cloud data of the casting and the point cloud data of the gating system are aligned and normalized.

3. The casting pouring design method based on deep learning according to claim 1, characterized in that, Extract casting features, including: Principal component analysis is used to align the point cloud, the principal orientation of the point cloud is calculated, and the principal axis is aligned with the Z-axis of the height direction by rotation.

4. The casting pouring design method based on deep learning according to claim 1, characterized in that, Geometric features include: center point, diameter, and height; The center point is used to characterize the center of the casting; the average value of all points is calculated. The diameter is used to represent the farthest distance between two points in the circumference. It is obtained by calculating the radial distance from all points to the center point and taking twice the maximum value. Height, used to characterize the farthest distance between two points in the height direction, is obtained by the difference between the maximum and minimum values ​​in the height direction.

5. The casting pouring design method based on deep learning according to claim 1, characterized in that, Component characteristics, including: number of blades, presence of flanges, flange width, and flange thickness, are obtained through the following methods: The number of blades is counted by peak statistics of the angle histogram. Specifically, this includes: identifying circumferential repeating structures; for each segment, calculating the angle of each point relative to the center, constructing a 360-degree angle histogram, counting the number of peaks by peak detection, and accumulating the peaks of all segments as the number of blades. Flange inspection is based on square ring structure identification. For each split body, the numerical range of radial width and height is checked, and the neighborhood normal vector of the sampling point of the split body is calculated. If the consistency of the normal vector exceeds the normal vector threshold, it is determined to be a flange. The radial width of the split body is used as the flange width, and the height of the split body is used as the flange thickness.

6. The casting pouring design method based on deep learning according to claim 1, characterized in that, Component features, including the location and quantity of flanges, inner rings, outer rings, and blades, are obtained through the following methods: By calculating the point cloud of each standard structure—flange, inner ring, outer ring, and blade—and the directed Hausdorff distance to the point cloud of the casting, the similarity is converted using the following formula: Similarity = 1 / (1 + directed Hausdorff distance); If the similarity exceeds the similarity threshold, it is determined to be a match, and the structure location and number are recorded; the point cloud center is used as the structure location.

7. The casting pouring design method based on deep learning according to claim 1, characterized in that, Extracting features of the gating system also includes: The number of components is calculated using a clustering algorithm based on a distance threshold. A KD tree is constructed for the point cloud, and a neighborhood search is performed. Points with a distance less than the threshold are grouped into the same cluster by depth-first traversal. The number of clusters and their center positions are counted as the number of components. The component position is determined by establishing a three-dimensional spatial coordinate system with the Z-axis representing the height direction, and the component position is represented by the three-dimensional spatial coordinate system. For each runner, calculate the Euclidean distance to all risers, select the nearest riser, and establish the component connection relationship between the runner and the riser. Dimensional variation parameters are recorded for each component, including scaling factor, bending angle, and elongation factor, to describe the dimensional changes of the actual component relative to the standard component.

8. The casting pouring design method based on deep learning according to claim 1, characterized in that, A neural network model is established, using the casting feature vector as input and the gating system feature vector as output. The neural network model is trained to obtain a casting casting generation model, including: A multi-layer fully connected neural network is used, with the input being the feature vector of the casting and the output being the feature vector of the gating system; A supervised learning dataset is constructed using the extracted feature vectors of the casting and the feature vectors of the gating system. By using the mean squared error loss function and optimizing the network parameters through backpropagation, the mapping from casting features to gating system features is achieved.

9. The casting pouring design method based on deep learning according to claim 1, characterized in that, Generate a 3D model of the gating system, including: The features of the new casting system are obtained based on the feature vector of the new casting system, and the corresponding components are selected from the standard model library; Based on the position and size change parameters, the components are translated, rotated, and scaled. Combine the components to generate a complete 3D model of the gating system.

10. A casting gating design system based on deep learning, characterized in that, include: Point cloud data processing unit, casting feature extraction unit, gating system feature extraction unit, casting gating generation model construction unit, and gating system 3D model generation unit; The point cloud data processing unit is used to acquire the three-dimensional model of the casting and the corresponding gating system, and convert the three-dimensional model into point cloud data or voxelized features of the casting, as well as point cloud data of the gating system. The casting feature extraction unit extracts casting features based on casting point cloud data or voxelized features, and combines the casting features into a casting feature vector. The casting features include geometric features, structural features, component features, and component features. The structural features combine RANSAC and geometric threshold decision layer structure. The component features are counted by peak statistics of angle histogram. The gating system feature extraction unit extracts gating system features based on gating system point cloud data and assembles these features into a gating system feature vector. The gating system features include gating component type, component quantity, component location, component connection relationship construction, and component size variation parameters. The casting pouring generation model building unit is used to establish a neural network model. It takes the feature vector of the casting as input and the feature vector of the pouring system as output, trains the neural network model, and obtains the casting pouring generation model. The gating system 3D model generation unit is used to extract casting features from the new casting and form a new casting feature vector. The new casting feature vector is then input into the casting gating generation model to obtain a new gating system feature vector, which is used to generate a 3D model of the gating system. Structural features include: number of layers, layer thickness, interlayer spacing, and layer height difference; The structure combines RANSAC and geometric threshold decision layer, including: performing point cloud split volume segmentation, using RANSAC plane fitting to extract planar regions, identifying corners by detecting changes in the angle of neighboring points or protrusions, and segmenting the point cloud into multiple split volumes with the corners as boundaries; Explore radially outward from the center to identify the number of layers. For each split body, calculate the radial width and height. If the radial width is less than the width threshold and the Z height is greater than the height threshold, it is determined to be one layer. Layer thickness, interlayer spacing, and layer height difference are obtained as follows: layer thickness is the radial width of the split body; interlayer spacing is the distance between two adjacent layers, with each split body surrounded by a convex hull, and the distance between the convex hulls used as the interlayer spacing; layer height difference is the height difference between the split bodies. Based on the point cloud data of the casting system, features of the casting system are extracted, including: The existing gating system is divided into gating system, riser, and inlet pipe, and standardized by classification to construct a standard point cloud model of gating system, riser, and inlet pipe; Calculate the nearest distance from each point in the gating system point cloud to the casting point cloud; if the distance is less than the distance threshold, it is considered to overlap with the casting, and after filtering, a pure gating system point cloud is obtained; Calculate the directed Hausdorff distance between the point cloud distribution of the pure gating system and the standard point cloud models of the gating system, risers, and inlet pipes; Randomly sample several points, calculate the convex hull volume consistency and surface normal vector consistency, and obtain the geometric consistency by weighted summation of the convex hull volume consistency and surface normal vector consistency; The Hausdorff distance is weighted using geometric consistency to obtain a weighted distance, and the model with the smallest weighted distance is selected as the best match.

Images