Internal defect prediction method based on fusion of GNN and point cloud transformer

Abstract

The application discloses an internal defect prediction method based on GNN and point cloud Transformer fusion, relates to the technical field of deep learning, and comprises the following steps: acquiring foundry casting pouring process parameters and three-dimensional point cloud data of a pouring system structure, and constructing a Geo-GPCT-DeepONet model.The model comprises a branch network for processing the process parameters and a trunk network for processing the point cloud structure, the trunk network extracts local topological and edge features of the point cloud through GNN, accesses point cloud Transformer to enhance spatial feature expression through self-attention, and then performs high-frequency nonlinear coding through a SIREN network; finally, the coding results of the trunk network and the branch network are point multiplied and fused to predict the internal defect distribution of the casting.The application can couple the process parameters, the spatial structure and the defect result, improve the defect prediction precision of shrinkage holes, shrinkage and the like, and reduce the repeated simulation of finite elements and the dependence on artificial experience.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to an internal defect prediction method based on the fusion of GNN and point cloud Transformer. Background Technology

[0002] With the development of high-end equipment manufacturing, the quality requirements for investment castings are constantly increasing. Investment castings are prone to internal defects such as shrinkage cavities and porosity during the forming process, which significantly affect the mechanical properties and service reliability of the material. Therefore, accurate prediction of defects during the casting process design stage is of great significance for improving casting quality and shortening the research and development cycle.

[0003] Existing methods for predicting casting defects mainly include finite element method (FEM) numerical simulation and machine learning methods. The FEM method can analyze the temperature, flow, and stress fields during solidification to predict the location and distribution of defects. However, this method suffers from high computational costs, complex modeling processes, and long computation cycles, making it difficult to meet the needs of rapid optimization under multiple process parameters. Machine learning methods establish a mapping relationship between process parameters and defect results; neural network models can predict casting defects to some extent. However, existing machine learning methods typically consider relatively simple casting process parameters and have limited ability to handle the complex geometry of the gating system, thus failing to accurately predict defects based on the gating system settings. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an internal defect prediction method based on the fusion of GNN and point cloud Transformer to solve the problem of low accuracy in existing casting defect prediction methods based on PointNet models.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an internal defect prediction method based on the fusion of GNN and point cloud Transformer, which includes acquiring casting process parameter data of investment casting and three-dimensional point cloud data of casting system structure;

[0008] A Geo-GPCT-DeepONet model for defect prediction in investment castings is constructed, including a branch network for processing casting process parameter data and a backbone network for processing 3D point cloud data of the casting system structure.

[0009] In the backbone network, a graph-enhanced point cloud Transformer module is set up: after extracting high-dimensional feature vector information of point cloud through GNN, the point cloud Transformer module is connected to enhance the expressive ability of point cloud features through self-attention mechanism.

[0010] In the future, the encoded data of the backbone network and branch networks will be combined in the form of dot products to form the prediction results of the Geo-GPCT-DeepONet model.

[0011] As a preferred embodiment of the internal defect prediction method based on the fusion of GNN and point cloud Transformer described in this invention, wherein: the dedicated dataset of the Geo-GPCT-DeepONet model includes casting process parameters, three-dimensional point cloud data of the casting system structure, and casting defect results;

[0012] The casting process parameters include converting the casting process parameters into an array as training data, and setting the casting temperature and mold shell preheating temperature range according to the technical specifications of the investment casting casting process.

[0013] The three-dimensional point cloud data of the casting system structure includes discretizing the three-dimensional model of the casting and the casting system structure into point cloud form, and generating training data through uniform sampling.

[0014] The casting defect results include labeling casting defects as binary labels based on the defect distribution in the finite element simulation results, where 0 indicates no defects and 1 indicates the presence of defects.

[0015] As a preferred embodiment of the internal defect prediction method based on the fusion of GNN and point cloud Transformer described in this invention, the branch network processes the casting process parameters into an array of shape (N, 2) as input according to the finite element simulation, and encodes them through a multi-layer fully connected network to extract the high-dimensional feature representation of the process parameters; where N represents the number of points in each point cloud.

[0016] As a preferred embodiment of the internal defect prediction method based on the fusion of GNN and point cloud Transformer described in this invention, the input of the backbone network is a tensor of shape [B, N, 6], where B is the batch size, N is the number of points in each point cloud, and 6 is the feature dimension of each point, including:

[0017] The three-dimensional coordinates (x, y, z) of each point in the point cloud are in the shape [B, N, 3].

[0018] The normal vector information (nx, ny, nz) of each point in the point cloud has a shape of [B, N, 3].

[0019] As a preferred embodiment of the internal defect prediction method based on the fusion of GNN and point cloud Transformer described in this invention, wherein: at the head of the backbone network, a fully connected layer and a ReLU activation function are used to extract features from the normal vector information of the point cloud; the extracted normal vector features are mapped to high-dimensional features of the point cloud through the fully connected layer;

[0020] The high-dimensional features of the point cloud are input into the GNN, and the distance, direction and feature differences are calculated and concatenated to form the edge features between the points.

[0021] As a preferred embodiment of the internal defect prediction method based on the fusion of GNN and point cloud Transformer described in this invention, the self-attention mechanism includes mapping the features of each point through queries, keys, and values, and enhancing the representation of the relationship between points by calculating relative position codes;

[0022] Attention weights are assigned to each point through attention calculation, and the points are updated through weighted aggregation. Finally, the enhanced features of each point are output.

[0023] After the point cloud features are enhanced, the point cloud Transformer module extracts features for each point and inputs them into the SIREN network for high-frequency nonlinear encoding to obtain the output of the backbone network.

[0024] As a preferred embodiment of the internal defect prediction method based on the fusion of GNN and point cloud Transformer described in this invention, the encoded data of the backbone network and the branch networks are combined in the form of dot products to form the prediction result of the Geo-GPCT-DeepONet model. This indicates the predicted distribution of casting defects:

[0025]

[0026] in, Indicates the hidden dimension. This represents encoded data from the backbone network. This represents encoded data from a branch network.

[0027] Secondly, the present invention provides an internal defect prediction system based on the fusion of GNN and point cloud Transformer, including a data acquisition unit for acquiring casting process parameter data of investment casting and three-dimensional point cloud data of casting system structure.

[0028] The model building unit constructs a Geo-GPCT-DeepONet model for defect prediction in investment castings, including a branch network for processing casting process parameter data and a backbone network for processing 3D point cloud data of the casting system structure.

[0029] The analysis unit sets up a graph-enhanced point cloud Transformer module in the backbone network: after extracting high-dimensional feature vector information of the point cloud through GNN, it is connected to the point cloud Transformer module, which enhances the expressive ability of point cloud features through self-attention mechanism.

[0030] The prediction unit combines the encoded data of the backbone network and branch networks in the form of dot products to form the prediction results of the Geo-GPCT-DeepONet model.

[0031] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the internal defect prediction method based on the fusion of GNN and point cloud Transformer as described in the first aspect of the present invention.

[0032] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the internal defect prediction method based on the fusion of GNN and point cloud Transformer as described in the first aspect of the present invention.

[0033] The beneficial effects of this invention are as follows: By coupling the casting process parameter data with the three-dimensional point cloud data of the casting system structure, low-dimensional process parameters such as casting temperature and mold preheating temperature are no longer merely used as overall judgment conditions, but can participate in the generation of defect results at various spatial locations in the point cloud, thereby improving the expressive ability of process conditions to predict the spatial distribution of defects. Simultaneously, this invention combines GNN and point cloud Transformer in the backbone network. GNN is used to establish local topological relationships between points and their neighbors, and spatial distance, relative direction, and neighborhood connectivity are introduced into the attention weight generation process of the point cloud Transformer. This ensures that attention allocation not only depends on feature similarity but is also constrained by spatial location and geometric continuity, avoiding the problem of ordinary Transformers ignoring the real spatial structure of the casting. Therefore, the model can simultaneously capture defect-sensitive areas such as locally thick regions and abrupt cross-sectional changes in the casting, as well as global influencing factors such as the overall feeding path of the casting system and long-distance spatial associations, improving the prediction accuracy and stability of the spatial distribution of internal defects such as shrinkage cavities and porosity. Compared with methods that process process parameters or point cloud structures separately, this invention can more comprehensively reflect the parameter-structure-defect coupling relationship in the investment casting process, reduce the reliance on repeated finite element simulations and manual experience judgment, and improve the efficiency and reliability of defect prediction in the casting process design stage. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is an overall flowchart of an internal defect prediction method based on the fusion of GNN and point cloud Transformer.

[0036] Figure 2 A 3D point cloud data map of the casting system structure is used in the internal defect prediction method based on the fusion of GNN and point cloud Transformer.

[0037] Figure 3 This is a histogram of point data label distribution in an internal defect prediction method based on the fusion of GNN and point cloud Transformer.

[0038] Figure 4 This is a schematic diagram of the enhanced point feature representation method in the GPCT module of the internal defect prediction method based on the fusion of GNN and point cloud Transformer.

[0039] Figure 5 This is a diagram of the GPCT point cloud feature extraction network architecture in an internal defect prediction method based on the fusion of GNN and point cloud Transformer.

[0040] Figure 6 This is a diagram showing the overall structure of the Geo-GPCT-DeepONet model in the internal defect prediction method based on the fusion of GNN and point cloud Transformer.

[0041] Figure 7 The graph shows the training and testing loss curves of the Geo-GPCT-DeepONet model in the internal defect prediction method based on the fusion of GNN and point cloud Transformer.

[0042] Figure 8 The graph shows the changes in various metrics during the training of the Geo-GPCT-DeepONet model in the internal defect prediction method based on the fusion of GNN and point cloud Transformer. Here, a represents the change in recall, b represents the change in precision, and c represents the change in accuracy.

[0043] Figure 9 The graph shows the F1 score variation during the training process of the Geo-GPCT-DeepONet model in the internal defect prediction method based on the fusion of GNN and point cloud Transformer.

[0044] Figure 10 This is a comparison chart of the prediction results of casting model 1 and the actual labels in the internal defect prediction method based on the fusion of GNN and point cloud Transformer.

[0045] Figure 11 This is a comparison chart of the prediction results of casting model 2 and the actual labels in the internal defect prediction method based on the fusion of GNN and point cloud Transformer.

[0046] Figure 12 The figure shows the finite element simulation verification results of casting 1 in the internal defect prediction method based on the fusion of GNN and point cloud Transformer.

[0047] Figure 13 The figure shows the verification results of the finite element simulation of casting 2 in the internal defect prediction method based on the fusion of GNN and point cloud Transformer.

[0048] Figure 14 The image shows the X-ray verification results of the defective part of casting 2 in the internal defect prediction method based on the fusion of GNN and point cloud Transformer.

[0049] Figure 15 The image shows the SEM characterization verification results of the defective part of casting 2 in the internal defect prediction method based on the fusion of GNN and point cloud Transformer. In the image, a and b are SEM scan results of different regions at the same scale, and c is a magnified SEM view of the area within the box in the image b.

[0050] Figure 16 This is a comparison of the accuracy of an internal defect prediction method based on the fusion of GNN and point cloud Transformer with a PointNet-based method. Detailed Implementation

[0051] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0052] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0053] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0054] Reference Figures 1-16 This is one embodiment of the present invention, which provides an internal defect prediction method based on the fusion of GNN and point cloud Transformer, including the following steps:

[0055] S1: Obtain the casting process parameter data and the three-dimensional point cloud data of the casting system structure of the investment casting.

[0056] A dedicated dataset for predicting defects in investment castings was constructed, which includes casting process parameter data and three-dimensional point cloud data of the casting system structure.

[0057] Specifically, by performing finite element modeling of the casting gating system and gating parameters, selecting appropriate gating parameters, and conducting finite element simulation, the gating parameters and casting defect results during the simulation process are obtained.

[0058] Correlation analysis revealed that the main influencing factors were the pouring temperature and the mold shell preheating temperature. The pouring temperature and the mold shell preheating temperature were processed as pouring process parameters into an array of shape (N,2), which was used as one of the input parameters of the model.

[0059] Point cloud processing was performed on the casting and gating system structures. Specifically, the casting defect results were exported from the finite element simulation software as 3D CAD files, and then discretized using homogenized point cloud sampling technology to obtain 3D point cloud data of the gating system structure. Each point cloud, after homogenization sampling, contains 25,000 points. Defects were labeled on the point cloud based on the defect results. The portions of the casting and gating system that were identified as defects in the finite element simulation results were assigned a value of 1, and the normal portions were assigned a value of 0. In the 3D point cloud visualization, white points represent defects, and red points represent normal portions. The visualized point cloud is shown below. Figure 2 As shown in the figure, in the casting defect prediction task, the number and volume of defects are relatively small, there are few defect points and many normal points, which is a case of data imbalance, such as... Figure 3 As shown, defects account for only about 5%.

[0060] Furthermore, different gating systems and gating process parameters are designed according to the shape and structure of different castings. A large number of finite element simulations are carried out to obtain the three-dimensional geometric information of the castings and gating systems, gating process parameter information, and casting defect result data from the finite element simulation results.

[0061] The 3D model of the casting and the structure of the gating system are discretized into point clouds. Training data is generated through uniform sampling to ensure that the 3D point cloud data contains all structural features of the casting and gating system. The point cloud data obtained by uniform sampling is downsampled so that each point cloud contains 25,000 points, ensuring that the point cloud data contains as much structural information as possible about the casting and gating system without affecting the smooth progress of the model training process.

[0062] The casting process parameters are converted into arrays as training data, and reasonable casting temperature and mold shell preheating temperature ranges are set according to the technical specifications of investment casting casting process.

[0063] By fusing casting process parameters with 3D point cloud data of the casting system structure, a dedicated dataset for casting defect prediction is obtained. This dataset is randomly divided into training and testing sets, with 70% of the data used as the training set and the remaining 30% as the testing set. The training set data is used to train the model, and the testing set data is used to validate the model's performance.

[0064] S2: Construct a Geo-GPCT-DeepONet model for defect prediction in investment castings, including a branch network for processing casting process parameter data and a backbone network for processing 3D point cloud data of the casting system structure.

[0065] In order for the model to effectively process the casting process parameter data, the branch network processes the casting process parameter data of shape (N,2) through three stacked fully connected layers. Then, it is fused with the high-dimensional point cloud features of the backbone network and then input into the three fully connected layers for processing to obtain the high-dimensional encoded data of the casting process parameters.

[0066] Each point in the geometric point cloud data of the casting system contains spatial coordinates (x, y, z) and normal vector information (nx, ny, nz). Therefore, the backbone network first extracts features through three fully connected layers, and then inputs the high-dimensional features of the casting process parameter data into the Graph Augmentation Point Cloud Transformer (GPCT) module for point cloud feature capture. Subsequently, the high-dimensional point cloud encoded data obtained by GPCT is multiplied by the high-dimensional encoded data of the casting process parameter data to obtain the casting defect prediction results of the model.

[0067] S3: In the backbone network, a graph-enhanced point cloud Transformer module is set up: After extracting the high-dimensional feature vector information of the point cloud through GNN, the point cloud Transformer module is connected to enhance the expressive ability of the point cloud features through the self-attention mechanism.

[0068] It's worth mentioning that the GPCT module, which combines GNN and point cloud Transformer modules, enhances the representation capability of point cloud data through graph-augmented point cloud feature transformation. The network is designed to effectively fuse the geometric information and local features of the point cloud, and further enhances the model's focus on important features through an attention mechanism.

[0069] First, feature extraction is performed on the normal vector information of the point cloud: In the initial stage of the network, fully connected layers and ReLU activation function are used to perform nonlinear transformation on these input features to generate a more abstract feature representation. The initial feature extraction part of the network is dedicated to processing normal vectors.

[0070] Next, high-dimensional feature vector information of the point cloud is extracted: the normal vector is mapped to a high-dimensional feature vector through a fully connected layer (Dense Layer). Specifically, the normal vector is first transformed by three 50-dimensional fully connected layers, and then mapped to a HIDDEN-dimensional feature representation. Here, the HIDDEN dimension is set to 32. The network then uses GNN to process the features of the input point cloud.

[0071] In GNNs, to capture the relationships between points, the k-nearest neighbor (k-NN) algorithm is first used to find each index point and its neighboring points in the point cloud. To reflect the local features of the point cloud data to some extent without ignoring global features, the value of k is set to 16. After retrieving the neighboring points of each point, the edge features between the neighboring points are calculated, including distance, direction, and feature differences. For the input point cloud data, the index point... (x) i , y i , z i ) and neighboring points (x) j , y j , z j The formula for calculating the distance feature is:

[0072]

[0073] The method for calculating the directional characteristic unit vector is as follows:

[0074]

[0075] The formula for calculating the difference between the feature vectors of the index point and its neighbors is:

[0076]

[0077] in, This is the feature vector of the current point; is the feature vector of the neighboring points.

[0078] Finally, the calculated distances, directions, and feature differences are concatenated into a single edge feature, which is used in subsequent GNN iterations. This concatenated edge feature contains the distance, direction, and differences between each pair of points in the feature space, providing rich contextual information for the graph neural network, enhancing its ability to model relationships between points, and improving point feature representation methods. Figure 4 As shown, the formula for calculating edge features is:

[0079]

[0080] Next, after completing GNN message passing, the network connects to the point cloud Transformer module. In the Transformer module, a self-attention mechanism is used to enhance the expressive power of point cloud features. The features of each point are first mapped using query (Q), key (K), and value (V), and then the relative position encoding is calculated to enhance the representation of relationships between points. During the calculation, the relative position encoding is incorporated into the attention mechanism to more accurately capture the spatial relationships between points in the point cloud data. Through attention calculation, the network can assign an attention weight to each point. These weights reflect the model's level of attention to different points. In investment casting production tasks, this specifically refers to locations on the casting that are relatively thick or have abrupt changes in cross-section. After attention calculation, the features are updated through weighted aggregation, ultimately outputting the enhanced features for each point.

[0081] It should be noted that although Transformer can learn global relationships, it is easy to ignore local geometric continuity. Therefore, GNN is needed to build a neighborhood graph structure, and KNN is used to find k points to form a neighborhood, so that features only propagate within the neighborhood, thus maintaining local topological structure and geometric continuity.

[0082] Furthermore, the ordinary Transformer was originally used for natural language processing, where words are ordered and represented by positional encoding. However, point cloud data is unordered. Therefore, the relative relationship between two points in Euclidean space is introduced. Relative positional encoding is essentially an explicit modeling of the spatial geometric relationship between points.

[0083] Transformer assigns weights based on correlation. In internal defect prediction tasks, defect points usually have characteristics such as abrupt curvature changes, drastic normal changes, and broken geometric continuity, which leads to high correlation between two points and correspondingly larger weights, thus this region will be of interest.

[0084] After the point cloud Transformer module completes the feature enhancement, it extracts features for each point and inputs them into the SIREN network (SIREN is a neural network that uses the sine function (Sin) as the activation function) for high-frequency nonlinear encoding to obtain the output of the backbone network. The network generates the final feature representation of each point. These feature representations will then be fused with the casting process parameter data and used directly for subsequent investment casting defect prediction tasks.

[0085] The design of this point cloud data feature extraction network fully considers the geometric structure and spatial relationships of point cloud data. It combines a GNN and a point cloud Transformer-based point cloud feature extraction network, GPCT. Figure 5The structure of the GPCT network is demonstrated, which can efficiently extract local and global features from point clouds, providing strong support for casting defect prediction tasks. GPCT first represents the input point cloud as two types of data: one type is point feature information, with a dimension of (N, 6), including the spatial coordinates and normal vectors of the points; the other type is point coordinate information, with a dimension of (N, 3), used to represent the spatial geometric position of the point cloud. In the local geometric feature extraction stage, GPCT first uses the k-nearest neighbor (KNN) algorithm to perform a neighborhood search on the point cloud, establishing topological connections between points and their neighbors. Subsequently, by constructing edge feature functions, the distance (Dt), direction (Dr), and feature difference (Fd) between the center point and its neighbors are encoded, thus forming an edge feature representation with local geometric constraints. Subsequently, edge features are sequentially input into a multilayer perceptron (MLP), a fully connected layer (Dense), and a ReLU activation function for nonlinear mapping to obtain local enhanced features Ef'. To further achieve neighborhood information aggregation, this invention introduces a graph neural network message passing mechanism. Through information interaction and feature propagation between nodes, local structural information is updated to obtain features Ef after fusing neighborhood context. Simultaneously, the spatial location information (POS) of points is jointly encoded with the updated features, enabling the network to maintain geometric spatial continuity during feature propagation, thereby enhancing the model's ability to express complex structural regions. GPCT constructs a feature enhancement branch based on a point cloud self-attention mechanism. This branch first performs convolution and local response normalization operations on the input point features to extract basic semantic features; then, it generates Q, K, and V feature maps respectively through the Enhance Point Feature Representation (EPFR) method. The deep fusion of GNN and point cloud Transformer constructs a local neighborhood graph structure of the point cloud, models the topological relationships between points using graph neural networks, and combines the self-attention mechanism in the point cloud Transformer to achieve joint learning of local geometric features and global dependencies. In this method, spatial geometric constraints between points are introduced through relative position encoding, and neighborhood edge features are integrated during the attention weight calculation process. This enhances the model's ability to perceive geometric abrupt changes in defect regions and enables high-precision prediction of internal defect regions.

[0086] The Geo-GPCT-DeepONet model combines the advantages of operator learning frameworks, graph neural networks, and point cloud Transformers in deep learning. While processing geometric point cloud data of casting gating systems, it can effectively predict casting defects by incorporating casting process parameters. The main structure of this model consists of a backbone network, branch networks, and a graph-enhanced point cloud Transformer module. Its overall structure is as follows: Figure 6 As shown.

[0087] To address the class imbalance problem in casting defect prediction tasks, a weighted binary cross-entropy loss function based on the focus loss function is constructed. This loss function combines the focus loss and the class weight term, and its definition is as follows:

[0088]

[0089]

[0090]

[0091]

[0092] in: This is a predicted value; This is a real label; To illustrate the relationship between predicted values ​​and true labels, The larger the value, the more accurate the prediction; As an adjustment coefficient, a value is assigned according to the category label (0 or 1); This is an adjustment parameter for focus loss, used to adjust the weights of samples that are difficult to classify; For category weight items; Category weights; This is the binary cross-entropy loss.

[0093] Evaluation metrics such as F1 score, precision, and recall are introduced to quantitatively evaluate the model's performance. The expression for precision is as follows:

[0094]

[0095] Where: TP is the number of samples correctly predicted as positive; FP is the number of samples incorrectly predicted as positive.

[0096] The expression for recall is as follows:

[0097]

[0098] Where: FN is the number of samples that were incorrectly predicted as negative.

[0099] The F1 score is the harmonic mean of precision and recall, and its expression is:

[0100]

[0101] S4: In the future, the encoded data of the autonomous backbone network and branch networks will be combined in the form of dot products to form the prediction results of the Geo-GPCT-DeepONet model.

[0102] It should be noted that after the dedicated dataset for investment casting defect prediction and the Geo-GPCT-DeepONet model are constructed, the training data is input into the model for training. This step mainly includes the following steps:

[0103] The backbone network receives the three-dimensional point cloud data of the casting system. First, it maps the input point cloud to the hidden feature space with a dimension of 32 through three consecutive fully connected layers, thus obtaining the high-dimensional features of the three-dimensional point cloud data of the casting system structure.

[0104] The branch network is used to process the casting process parameter inputs. In the investment casting process, the pouring temperature and the mold shell preheating temperature directly affect the filling, solidification process, and defect formation of the casting. The input to the branch network is set as an array consisting of the pouring temperature and the mold shell preheating temperature. The branch network extracts features from the pouring process parameters through three fully connected layers.

[0105] After the backbone network and branch networks process the relevant data to obtain high-dimensional features, the branch network maps the casting process parameters into the hidden feature space and aligns the dimensions with the input point cloud features. Then, the input parameter data and the input point cloud data are multiplied element-wise to fuse and generate a high-dimensional process feature vector, which is then input into the backbone network and branch networks for further feature extraction.

[0106] Subsequent feature extraction in the backbone network fuses global geometric information and local features of the point cloud through the embedded GPCT module. Since thicker or more distant areas from the riser in castings are more prone to defects during investment casting, an attention mechanism is added to further enhance the model's focus on these areas, improving the model's accuracy in predicting casting defects. GPCT, an embedded module in the backbone network, enhances the model's sensitivity to local regions, combining a GNN with a point cloud Transformer. GPCT establishes relationships between each point in the point cloud and enhances the model's sensitivity to local regions of the gating system's geometric point cloud through convolution operations and self-attention mechanisms. By learning the geometric relationships between points and capturing the correlation between casting defect formation and the gating system, GPCT improves the model's performance when handling castings with complex geometries and irregular defect distributions. Subsequently, the node features extracted by GPCT are input into a three-layer SIREN layer for high-frequency nonlinear encoding, improving the model's ability to represent local details and high-frequency changes in the casting and gating system.

[0107] After obtaining the high-dimensional process feature vector, the branch network processes the high-dimensional process feature vector through three fully connected layers to obtain the corresponding encoded data.

[0108] Subsequently, the branch network and the backbone network fuse the casting process parameters with the point cloud data of the casting system through dot products. In casting defect prediction, the Geo-GPCT-DeepONet model comprehensively considers the geometric features of the casting system and the casting process parameters. The encoded data from the backbone network and the branch network are then combined through dot products to form the prediction results of the Geo-GPCT-DeepONet model. This allows the Geo-GPCT-DeepONet model to accurately predict the distribution of casting defects, as shown below:

[0109]

[0110] in Indicates the hidden dimension. This represents encoded data from the backbone network. This represents the encoded data from the branch network. The dot product operation here has actually completed "implicit decoding". Specifically, it couples the global features of process parameters extracted by the branch network with the local features of spatial point cloud extracted by the backbone network through the dot product operator, thereby realizing the implicit function reconstruction of the spatial defect field and obtaining the spatial distribution prediction results of internal defects of the casting.

[0111] By fusing encoded data through dot product, the Geo-GPCT-DeepONet model can learn how casting process parameters and the gating system affect the defect distribution of castings under different casting conditions, providing comprehensive and accurate defect prediction. At the same time, considering the influence of the gating system geometry and casting process parameters on casting defects is beneficial to improving the accuracy of model prediction.

[0112] After the Geo-GPCT-DeepONet model is trained, the training results are analyzed to evaluate the model's accuracy and reliability. The analysis primarily focuses on the changes in loss value, precision, recall, accuracy, and F1 score during training. The changes in loss value during model training are shown below. Figure 7 As shown, the changes in precision, recall, and accuracy are as follows: Figure 8 As shown, the F1 score changes as follows Figure 9 As shown.

[0113] Subsequently, the model's performance was validated using two sets of data not used in model training. Finite element simulation and actual casting were used to further verify the effectiveness of the prediction model. Finite element simulation and actual casting were performed on castings 1 and 2 to analyze the casting quality and defects, verifying the predictive performance of the Geo-GPCT-DeepONet model and ensuring that the model's predictive performance reached its optimal level.

[0114] To intuitively evaluate the model's performance in actual prediction tasks, this study uses point cloud visualization to show the comparison between the predicted casting defects and the actual labels. After model training, data not seen during training was input into the model for prediction. The model prediction results and actual labels for castings 1 and 2 are shown below. Figure 10 and Figure 11 As shown, the left side displays the prediction results of the casting defect prediction model, and the right side displays the actual defect situation.

[0115] To assess the accuracy of the model's predictions, finite element simulation was first employed to verify their precision. Finite element simulations were performed on castings 1 and 2, yielding the following defect results: Figure 12 and Figure 13 As shown.

[0116] In addition to finite element simulation verification, this invention uses actual casting to verify the accuracy of model predictions. Actual casting was performed on casting 2, and the resulting X-ray projection inspection results are as follows: Figure 14 As shown, the casting defects predicted by the model were verified using scanning electron microscopy, with a focus on analyzing the microstructure of the defect regions predicted by the Geo-GPCT-DeepONet model. Figure 15 The image shows SEM images of the defective areas, revealing a high-density defect distribution that matches the model's predictions.

[0117] Therefore, the method proposed in this invention, which combines GNN and point cloud Transformer to design GPCT and construct the Gom-GPCT-DeepONet model to optimize the prediction accuracy of defects in investment castings, yields accurate prediction results and high prediction accuracy.

[0118] The method proposed in this invention will be compared with other methods below.

[0119] The proposed method was compared with the PointNet-based method for predicting shrinkage porosity and voids in aero-engine tail nozzle control plate castings. The PointNet-based method showed an accuracy of 0.6 to 0.7 on the test set during training, while the proposed method maintained an accuracy of around 0.97 during training. Furthermore, the addition of various evaluation metrics demonstrated the superiority of the proposed method in defect prediction. Figure 16 The figure shows a comparison of the accuracy of the proposed method and the PointNet-based method. It is clear from the figure that the accuracy of the proposed method in the defect prediction task is much higher than that of the PointNet-based method.

[0120] This embodiment also provides an internal defect prediction system based on the fusion of GNN and point cloud Transformer, including: a data acquisition unit for acquiring casting process parameter data of investment casting and three-dimensional point cloud data of the casting system structure.

[0121] The model building unit constructs a Geo-GPCT-DeepONet model for defect prediction in investment castings, including a branch network for processing casting process parameter data and a backbone network for processing 3D point cloud data of the casting system structure.

[0122] The analysis unit includes a graph-enhanced point cloud Transformer module in the backbone network: after extracting high-dimensional feature vector information of the point cloud through GNN, the point cloud Transformer module is connected to enhance the expressive ability of the point cloud features through a self-attention mechanism.

[0123] The prediction unit combines the encoded data of the backbone network and branch networks in the form of dot products to form the prediction results of the Geo-GPCT-DeepONet model.

[0124] This embodiment also provides a computer device applicable to the internal defect prediction method based on the fusion of GNN and point cloud Transformer, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the internal defect prediction method based on the fusion of GNN and point cloud Transformer as proposed in the above embodiment.

[0125] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0126] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the internal defect prediction method based on the fusion of GNN and point cloud Transformer as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0127] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An internal defect prediction method based on the fusion of GNN and point cloud Transformer, characterized by: This includes acquiring the casting process parameter data of the investment casting and the three-dimensional point cloud data of the gating system structure; A Geo-GPCT-DeepONet model for defect prediction in investment castings is constructed, including a branch network for processing casting process parameter data and a backbone network for processing 3D point cloud data of the casting system structure. In the backbone network, a graph-enhanced point cloud Transformer module is set up: after extracting high-dimensional feature vector information of point cloud through GNN, the point cloud Transformer module is connected to enhance the expressive ability of point cloud features through self-attention mechanism. In the future, the encoded data of the autonomous backbone network and branch network will be combined in the form of dot products to form the prediction results of the Geo-GPCT-DeepONet model; At the head of the backbone network, a fully connected layer and the ReLU activation function are used to extract features from the normal vector information of the point cloud. The extracted normal vector features are mapped to high-dimensional point cloud features through a fully connected layer; The high-dimensional features of the point cloud are input into the GNN, and the distance, direction and feature difference are calculated and concatenated to form the edge features between the points; The self-attention mechanism includes mapping the features of each point through queries, keys, and values, and enhancing the representation of relationships between points by calculating relative position codes; Attention weights are assigned to each point through attention calculation, and the points are updated through weighted aggregation. Finally, the enhanced features of each point are output. After the point cloud features are enhanced, the point cloud Transformer module extracts features for each point and inputs them into the SIREN network for high-frequency nonlinear encoding to obtain the output of the backbone network.

2. The internal defect prediction method based on the fusion of GNN and point cloud Transformer as described in claim 1, characterized in that: The dedicated dataset for the Geo-GPCT-DeepONet model includes casting process parameters, three-dimensional point cloud data of the casting system structure, and casting defect results. The casting process parameters include converting the casting process parameters into an array as training data, and setting the casting temperature and mold shell preheating temperature range according to the technical specifications of the investment casting casting process. The three-dimensional point cloud data of the casting system structure includes discretizing the three-dimensional model of the casting and the casting system structure into point cloud form, and generating training data through uniform sampling. The casting defect results include labeling casting defects as binary labels based on the defect distribution in the finite element simulation results, where 0 indicates no defects and 1 indicates the presence of defects.

3. The internal defect prediction method based on the fusion of GNN and point cloud Transformer as described in claim 2, characterized in that: The branch network, based on the finite element simulation, processes the casting process parameters into an array of shape (N, 2) as input, encodes them through a multi-layer fully connected network, and extracts high-dimensional feature representations of the process parameters. Where N represents the number of points in each point cloud, and 2 represents the pouring temperature and mold preheating temperature parameters of the input branch network.

4. The internal defect prediction method based on the fusion of GNN and point cloud Transformer as described in claim 3, characterized in that: The input to the backbone network is a tensor of shape [B, N, 6], where B is the batch size, N is the number of points in each point cloud, and 6 is the feature dimension of each point, including: The three-dimensional coordinates (x, y, z) of each point in the point cloud are in the shape [B, N, 3]. The normal vector information (nx, ny, nz) of each point in the point cloud has a shape of [B, N, 3].

5. The internal defect prediction method based on the fusion of GNN and point cloud Transformer as described in claim 4, characterized in that: The encoded data of the backbone network and the branch networks are combined using a dot product to form the prediction result of the Geo-GPCT-DeepONet model. This indicates the predicted distribution of casting defects: ； in, Indicates the hidden dimension. This represents encoded data from the backbone network. This represents encoded data from a branch network.

6. An internal defect prediction system based on the fusion of GNN and point cloud Transformer, based on the internal defect prediction method based on the fusion of GNN and point cloud Transformer as described in any one of claims 1 to 5, characterized in that: This includes a data acquisition unit that acquires the casting process parameters of the investment casting and the three-dimensional point cloud data of the casting system structure. The model building unit constructs a Geo-GPCT-DeepONet model for defect prediction in investment castings, including a branch network for processing casting process parameter data and a backbone network for processing 3D point cloud data of the casting system structure. The analysis unit sets up a graph-enhanced point cloud Transformer module in the backbone network: after extracting high-dimensional feature vector information of the point cloud through GNN, it is connected to the point cloud Transformer module, which enhances the expressive ability of point cloud features through self-attention mechanism. The prediction unit combines the encoded data of the backbone network and branch networks in the form of dot products to form the prediction results of the Geo-GPCT-DeepONet model.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the internal defect prediction method based on the fusion of GNN and point cloud Transformer as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the internal defect prediction method based on the fusion of GNN and point cloud Transformer as described in any one of claims 1 to 5.

Images