Defect prediction method and system for investment casting of aircraft engine tail nozzle regulating plate

Abstract

The present invention discloses a method and system for predicting defects in investment castings of aircraft engine tail jet flaps. This method, which belongs to the technical fields related to casting and machine learning, includes: numerically simulating the flap casting and its pouring system using finite element simulation software to obtain the location and size distribution of casting defects, and converting the results into three-dimensional point cloud data; dividing the three-dimensional point cloud data into a training set and a validation set, and using the training set to train a PointNet-based defect prediction model; evaluating the prediction accuracy and generalization capability of the defect prediction model using the validation set, and adjusting hyperparameters based on the validation results; and converting the three-dimensional model of the pouring system of the flap casting to be predicted into three-dimensional point cloud data, inputting the trained defect prediction model, and outputting the location of the casting defects and the predicted results. This invention addresses the shortcomings of existing finite element simulation methods, which are time-consuming and wasteful of computing resources.

Description

Technical Field

[0001] The present invention relates to the technical field related to casting and machine learning, and in particular to a method and system for predicting defects in investment castings of aircraft engine tail jet regulating plates. Background Art

[0002] Aircraft engine tailpipe trims are a key component at the tail of an aircraft engine, primarily used to adjust the direction and shape of the engine's exhaust jet, thereby affecting the magnitude and distribution of thrust. They are typically used to improve engine performance, optimize fuel efficiency, control exhaust noise, and enhance aircraft maneuverability. These trims are widely used on jet engines, particularly fighter jets and high-performance aircraft, which operate under complex and variable flight conditions. These trims provide flexible thrust adjustment and aerodynamic control.

[0003] Adjustment plates are typically produced using investment casting, a precision casting process known for its high precision, complex shapes, and smooth surfaces. However, during the production process, various process parameters and operating conditions often lead to defects in the casting, which can affect performance and even lead to the rejection of the casting. Investment castings are also extremely expensive to produce, and the resulting rejection of castings can result in significant financial losses for the company. Therefore, metallurgical quality control of investment castings is crucial in the production process.

[0004] Early predictions of investment casting quality relied entirely on the judgment of professionals with years of relevant experience. Actual production would then be conducted, and based on the actual defect distribution, professionals would analyze and improve the results, repeating this process until the product met the standards and was put into full production. This process was not only time-consuming and labor-intensive, but also required extremely high professional expertise from the relevant practitioners. Later, with the advancement of finite element simulation technology, researchers developed specialized casting finite element simulation software, such as ProCAST, Huazhu CAE, and anycasting, which significantly reduced product R&D costs. However, this process was still relatively time-consuming and computationally expensive. Furthermore, to achieve higher accuracy, a denser finite element mesh and smaller time steps were required for simulation calculations. This method was time-consuming and required significant computing resources. Therefore, the foundry industry urgently needed a fast, low-cost method for predicting casting defects.

[0005] In recent years, with the continuous development of artificial intelligence technology, probabilistic algorithms based on large amounts of data have gained favor with a large number of scholars. Various machine learning algorithms have been proven to be superior in many scenarios. Therefore, the application of machine learning to casting defect prediction has received increasing attention. For example, some researchers use U-Net networks to predict the temperature field of the casting process, thereby inferring the defects of the castings based on the distribution of the temperature field; some use neural networks to construct relationship models between casting process parameters and casting defects, predict casting defects, and thus adjust casting process parameters; and use fully connected convolutional neural networks to transform the casting defect prediction task into a semantic segmentation task of three-dimensional images, distinguishing defects from non-defective parts in the casting system from pixel points. The above methods all confirm the feasibility of machine learning methods in the task of casting defect prediction. However, these methods do not actually indicate the location of the defect. Therefore, it is difficult to adjust casting process parameters from the perspective of the entire casting system to reduce casting defects and improve the metallurgical quality of castings.

[0006] There are few studies on using machine learning methods to predict casting defects. Most of them are based on building neural networks between casting process parameters and defect size and number. The impact of the gating system on casting defects is rarely considered. However, in the actual casting production process, the rationality of the gating system setting greatly affects the molding quality of the casting. Summary of the Invention

[0007] To address the above technical issues, a method for predicting defects in investment castings of aircraft engine tail jet flaps was proposed. The method includes numerically simulating the flap casting and gating system using finite element simulation software to obtain the location and size distribution of casting defects, and converting the results into three-dimensional point cloud data.

[0008] The three-dimensional point cloud data is divided into a training set and a validation set. The training set is used to train a PointNet-based defect prediction model, including: local feature extraction and global feature aggregation, and the local features and global features are fused; the local feature extraction is to use the farthest point sampling method to extract index points of the three-dimensional point cloud data, and obtain a local point cloud set through a spherical neighborhood query method, extract the local geometric features of each point through a multi-layer perceptron with shared weights, and use a set abstraction layer to hierarchically capture multi-scale local features;

[0009] Evaluate the prediction accuracy and generalization ability of the defect prediction model using the validation set, and adjust hyperparameters based on the validation results.

[0010] The 3D model of the casting system of the regulating plate to be predicted is converted into 3D point cloud data, which is then input into the trained defect prediction model to output the location of the casting defect and the prediction results.

[0011] As a preferred embodiment of the method for predicting defects in investment castings of aircraft engine tail jet regulating pieces according to the present invention, the three-dimensional point cloud data includes three-dimensional regulating piece gating system point cloud data and defect point cloud data;

[0012] The three-dimensional regulating plate gating system point cloud data is a complete casting and gating system structure, providing overall geometric information of the casting and gating system;

[0013] The defect point cloud data is a defect area simulated and predicted to complete the defect prediction task;

[0014] The converting of the result into three-dimensional point cloud data includes performing three-dimensional point cloud conversion on the position and size distribution results of the casting defects simulated by finite element simulation using a uniform sampling method.

[0015] As a preferred solution of the method for predicting defects in investment castings of aircraft engine tail nozzle adjustment plates described in the present invention, the method of extracting index points of three-dimensional point cloud data using the farthest point sampling method includes randomly determining a point in the three-dimensional point cloud data as an initial point, using the farthest point sampling method to obtain the point with the farthest square Euclidean distance from the initial point, and obtaining an index point set and spatial structure characteristics.

[0016] As a preferred embodiment of the method for predicting defects in investment castings of aircraft engine tail nozzle regulating plates according to the present invention, the method of obtaining a local point cloud set by using a spherical neighborhood query method includes selecting neighborhood points of the sampling point from the three-dimensional point cloud data using a spherical neighborhood query method with a specified radius, and obtaining a local point cloud set centered on each index point;

[0017] The local geometric features of each point are extracted through a multi-layer perceptron with shared weights, and the multi-scale local features are captured hierarchically using the set abstraction layer. The three-dimensional coordinates, color information and normal vector information of each point in the point set are independently processed through the multi-layer perceptron layer with shared weights. The input dimension of each point is 9, including the three-dimensional coordinates , color information , normal vector information ; It is mapped to 64 dimensions and 128 dimensions through a shared multi-layer perceptron layer to capture the local features of each point.

[0018] As a preferred solution of the method for predicting defects in investment castings of aircraft engine tail nozzle control panels described in the present invention, global feature aggregation includes stacking set abstraction layers and aggregating the local features and spatial structure features of each point using a global pooling method to form a 1024-dimensional global feature required for casting defect prediction.

[0019] As a preferred solution of the method for predicting defects of investment castings of aircraft engine tail nozzle adjustment plates described in the present invention, the fusing of local features and global features includes using a multi-layer perceptron to perform feature splicing on the 1024-dimensional global features and the 128-dimensional local features extracted from each set abstraction layer to form a 1088-dimensional feature vector to obtain a fused feature.

[0020] As a preferred solution of the defect prediction method for the investment casting of the aircraft engine tail nozzle adjustment plate described in the present invention, the fusing of local features and global features also includes inputting the fused features into a shared multi-layer perceptron layer, generating prediction results point by point, and using the prediction results of each point to calculate the defect prediction model loss and update the model parameters.

[0021] Another object of the present invention is to provide a defect prediction system for investment castings of aircraft engine tail jet flaps, including a numerical simulation module that uses finite element simulation software to perform numerical simulations on the flap casting and gating system, obtain the location and size distribution of casting defects, and convert the results into three-dimensional point cloud data.

[0022] A model training module divides the three-dimensional point cloud data into a training set and a validation set, and uses the training set to train a PointNet-based defect prediction model;

[0023] The model validation module evaluates the prediction accuracy and generalization ability of the defect prediction model through the validation set and adjusts the hyperparameters based on the validation results;

[0024] The execution module converts the three-dimensional model of the casting system of the regulating plate to be predicted into three-dimensional point cloud data, inputs the trained defect prediction model, and outputs the location of the casting defect and the prediction result.

[0025] The present invention solves the shortcomings of the existing finite element simulation method, which is time-consuming and wastes computing resources.

[0026] A computer device includes a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of the method for predicting defects of an aircraft engine tail nozzle adjustment plate investment casting are implemented.

[0027] A computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the method for predicting defects of an aircraft engine tail nozzle adjustment plate investment casting are implemented.

[0028] The beneficial effects of the present invention are as follows: the position and size distribution of casting defects are obtained by finite element simulation software and converted into three-dimensional point cloud data, which not only retains the spatial information of the defects but also reduces the detection cost. The point cloud data is divided into a training set and a validation set, so that the model can fully learn the characteristics of the casting and defects, avoid overfitting, and improve the generalization ability. In the feature extraction stage, the farthest point sampling and spherical neighborhood query are used to evenly select index points, and the local geometric features are extracted by a multi-layer perceptron with shared weights to ensure that the model can accurately capture the subtle defects of the casting. The multi-scale local features are captured hierarchically through the set abstraction layer, and the overall structural information is extracted in combination with the global pooling operation to achieve the fusion of local and global features, thereby improving the stability and accuracy of defect prediction. In the classification process, the fused features are further extracted by the convolution layer and the fully connected layer, and normalized by the Softmax function to achieve accurate classification of each point. At the same time, the weighted cross entropy loss function is combined to optimize the model to deal with the problem of class imbalance. The validation set is used to evaluate the model performance and adjust the hyperparameters to ensure that it can maintain high-precision prediction under different casting structures. Ultimately, the trained model can convert the new casting system into point cloud data and quickly output defect location and category prediction results, significantly improving quality inspection efficiency, reducing production costs, and providing strong support for casting process optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0030] Figure 1 An overall flow chart of a method for predicting defects in investment castings of aircraft engine tail nozzle trim panels provided by one embodiment of the present invention.

[0031] Figure 2 A flowchart of a method for predicting defects in investment castings of an aircraft engine tail jet trim flap provided in one embodiment of the present invention.

[0032] Figure 3 A finite element numerical simulation process diagram of a method for predicting defects in investment castings of aircraft engine tail nozzle adjustment panels provided in one embodiment of the present invention.

[0033] Figure 4 This is a diagram showing defect prediction results of a finite element simulation of a defect prediction method for an aircraft engine tail nozzle adjustment plate investment casting provided by one embodiment of the present invention.

[0034] Figure 5The finite element simulation results of the method for predicting defects in investment castings of aircraft engine tail nozzle adjustment plates provided in one embodiment of the present invention are more likely to produce SEM scanning electron microscope results of defective parts.

[0035] Figure 6 A flowchart of acquiring point cloud data required for defect prediction model training of an aircraft engine tail nozzle adjustment plate investment casting defect prediction method provided by one embodiment of the present invention.

[0036] Figure 7 This is an overall structural diagram of a PointNet-based casting defect prediction model for an aircraft engine tail nozzle adjustment plate investment casting defect prediction method provided by one embodiment of the present invention.

[0037] Figure 8 This is a diagram showing the training results of a casting defect prediction model for an aircraft engine tail nozzle adjustment plate investment casting defect prediction method provided by one embodiment of the present invention.

[0038] Figure 9 A comparison chart of the finite element simulation results of a method for predicting defects in investment castings of aircraft engine tail nozzle trim panels provided by one embodiment of the present invention and the casting defect prediction results predicted by the model. DETAILED DESCRIPTION

[0039] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0040] Example 1, with reference to Figures 1-9 , which is the first embodiment of the present invention, provides a method for predicting defects in investment castings of aircraft engine tail nozzle regulating pieces, comprising:

[0041] S1: Use finite element simulation software to perform numerical simulation of the regulating plate casting and gating system to obtain the location and size distribution results of casting defects. A portion (80% of the total finite element simulation) is used as a training set and the remaining 20% ​​is used as a validation set. The results are converted into three-dimensional point cloud data.

[0042] Finite element simulation software is used to numerically simulate the regulating plate casting and the pouring system to obtain the location and size distribution results of the casting defects. The principle of numerical simulation of the casting process is to geometrically discretize the pouring system and the casting mold of the casting. With the support of physical-mathematical models, the changes in the temperature field, flow field and stress field during the casting process are analyzed to predict the defects that may occur in the casting. The results are converted into three-dimensional point cloud data, and the casting defect results of the finite element simulation are exported into a three-dimensional CAD file format. Point cloud sampling is performed together with the three-dimensional CAD pouring system of the casting to obtain the data required for training this model.

[0043] S1.1. Analyze the structural characteristics of the regulating plate parts, combine the principles of fluid mechanics, thermodynamics and solidification feeding, avoid turbulence or air entrainment of the molten metal during the pouring process, and ensure that the casting achieves sequential solidification and directional feeding during the solidification process, so as to establish a suitable casting pouring system, use finite element simulation software to simulate the pouring of the regulating plate casting, the finite element simulation process is as follows Figure 3 As shown. Through finite element meshing of castings and pouring systems, pouring parameter setting, and finite element simulation calculation, in the finite element simulation process, if you want to improve the simulation accuracy, you will inevitably need to reduce the finite element mesh size. Reducing the mesh size means that a lot of computing resources are consumed. Therefore, different mesh sizes are used to mesh the castings and pouring systems. Compared with castings, the pouring system is not our focus, so the mesh size of the casting is set to 1mm, and the pouring system is set to 5mm. When setting parameters, create a user's personal material library based on the material composition used in the actual production of castings, so that the finite element simulation is closer to the actual production process. The defect distribution of the adjustment plate casting is thus obtained, as shown Figure 4 By combining the finite element simulation results with the actual analysis of the casting structure, two locations in the adjustment plate casting where the finite element simulation results are more likely to cause defects were sampled and used for high-order material characterization to verify the accuracy of the finite element simulation method in predicting casting defects. The results showed that there were indeed dense hole defects under SEM scanning at the locations where the finite element simulation results are more likely to cause defects. Therefore, the finite element numerical simulation method has a high accuracy rate in predicting casting defects. The scanning results are shown in Figure 2. Figure 5 shown.

[0044] S1.2. Save the casting 3D model and the finite element prediction defect results in S1.1 as a 3D CAD model, and perform point cloud sampling on the 3D CAD model. When sampling the 3D CAD model, it is necessary to ensure uniform sampling so that the point cloud can contain the position information of each part of the gating system and most of the structural features. The gating system point cloud is used as a data sample, and after point clouding, the gating system and defect distribution results are merged into data labels. In order to meet the processing requirements of the model, the point cloud needs to be downsampled so that its point number is 100,000, and each point contains 9 attributes, namely, the 3D coordinates , color information , normal vector information The process of obtaining point cloud data required for model training is as follows: Figure 6 shown.

[0045] Furthermore, finite element simulation software was used to numerically simulate the regulating plate casting and gating system, determining the location and size distribution of casting defects. The simulation results were then converted into three-dimensional point cloud data. The core purpose of this process is to accurately predict defects that may occur during the casting process, such as shrinkage cavities and porosity, through numerical simulation, thereby providing a scientific basis for subsequent defect detection and analysis. By combining the finite element simulation results with the actual casting structure, the accuracy of the simulation is ensured, providing high-quality annotated data for subsequent machine learning model training, further optimizing the accuracy of defect prediction.

[0046] Furthermore, by combining finite element simulation with 3D point cloud data sampling, this step effectively overcomes the traditional inspection method's reliance on actual castings, allowing model training to be completed in a virtual environment. Uniform sampling and downsampling techniques ensure the quality and balance of the point cloud data, while also assigning multiple attributes (such as 3D coordinates, color information, and normal vector information) to the point cloud data, enhancing the model's ability to learn different defect types. This approach not only improves the accuracy and efficiency of defect prediction but also reduces actual production testing costs. It helps to foresee potential quality issues before production begins and provides data support for process adjustments during casting production.

[0047] S2: Divide the three-dimensional point cloud data into a training set and a validation set, and use the training set to train a PointNet-based defect prediction model.

[0048] All processed data are randomly divided into training set and validation set.

[0049] Based on the advancement of finite element simulation technology, researchers have developed specialized casting finite element simulation software, such as ProCAST, Huazhu CAE, and anycasting, significantly reducing product R&D costs. However, this process remains relatively time-consuming and computationally expensive. Furthermore, achieving higher accuracy requires a denser finite element mesh and a smaller time step for simulation calculations. This method is time-consuming and requires significant computing resources. Therefore, the foundry industry urgently needs a fast, low-cost method for predicting casting defects. Finite element simulation software is used to numerically simulate the trim plate casting and gating system, determining the location and size distribution of casting defects and converting the simulation results into 3D point cloud data. The core purpose of this process is to accurately predict defects that may occur during the casting process, such as shrinkage cavities and porosity, through numerical simulation, thereby providing a scientific basis for subsequent defect detection and analysis. By combining finite element simulation results with the actual casting structure, simulation accuracy is ensured, providing high-quality annotated data for subsequent machine learning model training, further optimizing defect prediction accuracy.

[0050] Input the point cloud data in the training set into Figure 7 In the defect prediction model for the adjustment plate investment casting shown in FIG, the model uses a shared MLP to extract the local features of each point, and then uses the maximum pooling method to aggregate the local features to form the global features of the point cloud.

[0051] S2.1. Input the training data into the SA feature extraction module. Each point's data is nonlinearly mapped through a shared MLP. This step aims to extract high-dimensional features for each point while ensuring that the operation is insensitive to point order. The features are mapped to 64 dimensions to obtain an Nx64 local feature matrix, which is then remapped to 1024 dimensions to obtain the features of each point in the point cloud.

[0052] S2.2: Use max pooling to aggregate the local features obtained in S2.1 into global features for the point cloud. This step combines the local features of all points into a fixed-size global feature vector, representing the entire point cloud. Ultimately, a 1024-dimensional global feature vector is obtained.

[0053] The captured Nx64-dimensional local feature matrix is ​​concatenated with the 1024-dimensional global feature vector, and the global features are copied and connected to the local features of each point to obtain an Nx1088 feature matrix containing the local features of each point and the global features of the point cloud.

[0054] The fused feature matrix is ​​input into the shared MLP layer, which generates prediction results point by point. The prediction results of each point are used to calculate the model loss and update the model parameters. First, the softmax operation is used to convert the original output of the model into a probability distribution. For the original output vector z of a sample, the softmax calculation formula is:

[0055]

[0056] in, Indicates the corresponding category The original output value of Indicates the number of categories for classification; is a category The predicted probability of .

[0057] Then the true label of each sample Calculate the negative log-likelihood loss for the label, i.e.:

[0058]

[0059] in, Representation sample The true label, is the predicted probability of the true label after softmax.

[0060] Finally, a weighting function is used to adjust the impact of category imbalance. The weighting function is:

[0061]

[0062] in: The number of points representing each gating system point cloud; Represents the indicator function, which is 1 if the label of sample i is category j, and 0 otherwise.

[0063] Furthermore, by dividing the 3D point cloud data into a training set and a validation set, and using the training set to train a PointNet-based defect prediction model, the model can effectively extract features from the point cloud data and perform defect prediction. Specifically, the point cloud data in the training set is subjected to local feature extraction through a shared MLP (multi-layer perceptron) network and aggregated into global features through a maximum pooling method. This allows each point to not only reflect the local geometric structure but also incorporate the global information of the entire point cloud. In addition, through feature fusion, the local features are concatenated with the global features and then input into the MLP layer, further improving the model's performance in handling complex casting defects. This design enables the model to comprehensively consider the detailed features and overall structural features of the point cloud when predicting defects, thereby improving the accuracy and robustness of the prediction.

[0064] Furthermore, firstly, the combination of local feature extraction and global feature aggregation enables the model to take into account both the detailed information and the overall morphological features of the casting point cloud, thus enhancing the model’s ability to identify different defects. Secondly, the use of maximum pooling to aggregate local features to form global features ensures that the information of all points in the point cloud can be effectively integrated while avoiding information loss. In addition, by using softmax operations and cross-entropy loss, combined with a class weighting function to adjust the impact of class imbalance, it is ensured that the model can still maintain good performance even when the number of samples in different categories is uneven, especially when dealing with rarer defect types in castings. Finally, the use of a validation set ensures the generalization ability of the model, allowing it to maintain a high prediction accuracy even on unseen data, thereby enhancing the industrial application value of the model.

[0065] S3: Evaluate the prediction accuracy and generalization ability of the defect prediction model through the validation set, and adjust the hyperparameters based on the validation results.

[0066] The validation set is not used in the model training process. Therefore, the data in the validation set is data that the model has never seen before. It is used to evaluate the performance of the model and provide feedback to the model to adjust the model hyperparameters based on the validation results.

[0067] The validation set is not used in the model training process. Therefore, the data in the validation set is data that the model has never seen before. It is used to evaluate the performance of the model and provide feedback to the model to adjust the model hyperparameters based on the validation results.

[0068] Repeatedly update the model parameters and verify the detection ability of the model, and finally take the model parameters when the accuracy reaches the maximum value as the final training result of the model. Figure 8 This shows the changes in loss and accuracy during the model training process under the technical solution of the present invention.

[0069] The schematic diagram of defect prediction results of the adjustment plate investment casting obtained by the technical solution of the present invention is as follows Figure 9 As shown in the figure, the left side is the finite element numerical simulation result, and the right side is the prediction result obtained by the present invention. The red part is the normal part, and the white part is the defective part.

[0070] It can be found that the prediction results of the present invention correspond to the traditional finite element numerical simulation results. Since the finite element simulation method has been proven to be reliable in predicting casting defects in step S1.1, and the defect distribution in the model prediction results is roughly the same as the finite element results, the present invention achieves rapid and accurate prediction of defects in investment castings of aircraft engine tail nozzle adjustment plates.

[0071] Furthermore, the accuracy and generalization of the defect prediction model are evaluated using a validation set. Hyperparameters are adjusted based on the evaluation results to ensure the model's robust predictive capabilities across diverse data distributions. Repeatedly updating and validating model parameters avoids overfitting and ensures the stability and reliability of the final model in practical applications. Furthermore, selecting model parameters that maximize accuracy ensures optimal performance on the test set. This gradual adjustment process allows the model to adapt to diverse casting structures and defect types, thereby enhancing the comprehensiveness and accuracy of defect prediction.

[0072] Furthermore, the accuracy and generalization of the defect prediction model are evaluated using a validation set. Hyperparameters are adjusted based on the evaluation results to ensure the model's robust predictive capabilities across diverse data distributions. Repeatedly updating and validating model parameters avoids overfitting and ensures the stability and reliability of the final model in practical applications. Furthermore, selecting model parameters that maximize accuracy ensures optimal performance on the test set. This gradual adjustment process allows the model to adapt to diverse casting structures and defect types, thereby enhancing the comprehensiveness and accuracy of defect prediction.

[0073] S4: Convert the three-dimensional model of the casting system of the adjustment plate to be predicted into three-dimensional point cloud data, input the trained defect prediction model, and output the location of the casting defect and the prediction result.

[0074] Furthermore, the 3D model of the gating system of the regulating plate casting to be predicted is converted into 3D point cloud data. This process integrates actual engineering data (such as 3D CAD models) with the input format of the machine learning model, achieving a seamless transition from traditional engineering design to a data-driven predictive model. By converting the 3D model into point cloud data, deep learning models (such as PointNet) can be used to efficiently process and analyze the casting's geometry and defect information. This step ensures that the model fully utilizes the casting's structural features and defect information, thereby improving the accuracy and reliability of defect prediction.

[0075] Furthermore, point cloud data preserves the casting's three-dimensional spatial structure, facilitating in-depth analysis of the model's local and global characteristics at each point. Secondly, by inputting the trained defect prediction model, the location and type of defects in the casting can be automatically predicted, avoiding the time-consuming and human-incorrect nature of traditional manual inspection. Ultimately, this process improves the efficiency of casting defect detection, making quality control more precise and efficient during production, reducing production costs, and enabling the early detection and resolution of potential problems, thereby improving overall productivity and casting quality.

[0076] Example 2 is the second embodiment of the present invention, which provides a method for predicting defects in investment castings of aircraft engine tail nozzle regulating plates. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0077] Finite element simulation software was used to numerically simulate the flap casting and gating system. The simulation focused on the gating system's geometry, temperature field, flow field, and solidification process. Two different flap casting types (Casting A and Casting B) were selected as experimental subjects. The simulated defect locations and size distributions were exported and converted into 3D point cloud data using a 3D visualization tool. Each point's data contained 3D coordinates, color information, and normal vector information. To ensure data uniformity and representativeness, a uniform sampling method was used to sample the point cloud, resulting in 50,000 points.

[0078] During the data preparation phase, point cloud data from 100 different adjustment sheet castings was used. Point clouds from 80 castings were used as the training set, and point clouds from 20 castings were used as the validation set. Validation was performed after each training round. Specifically, 80% of the casting data (i.e., 80 castings) were used for training, and 20% (i.e., 20 castings) were used for validation. The training data was indexed using the farthest point sampling (FPS) method. Neighboring points were selected using the spherical neighborhood query method, and local geometric features were extracted for each point using a multi-layer perceptron (MLP) with shared weights. The input features for each point included 3D coordinates, color information, and normal vector information. After mapping, 64- and 128-dimensional local feature representations were obtained. After the MLP extracted local features, a set abstraction layer was used to capture multi-scale local features. Global pooling was then used to aggregate the local features of each point into a 1024-dimensional global feature.

[0079] Next, local and global features are concatenated and fed into a shared multi-layer perceptron layer to generate predictions point by point. To address class imbalance, a weighted cross-entropy loss function is used for optimization. The model's hyperparameters are adjusted using a validation set to ensure optimal training results. After training and validation, to further test the model's generalization capabilities, castings A and B, which had never been used in training or validation (i.e., new test data), were selected. Their corresponding 3D point clouds were fed into the trained defect prediction model, which then output the locations of the casting defects and the predicted results.

[0080] During the experiment, the following experimental data were collected:

[0081] Number of simulation points for casting A: 50,000; number of simulation points for casting B: 50,000; number of defect detection points (casting A): 2,430; number of defect detection points (casting B): 2,510; defect prediction accuracy (casting A): 92.35%; defect prediction accuracy (casting B): 93.01%.

[0082] Training time: 12.35 hours; testing time: 1.52 hours.

[0083] Prediction accuracy (training set): 90.88% (casting A), 91.33% (casting B).

[0084] Prediction accuracy (validation set): 91.10% (Casting A), 91.55% (Casting B).

[0085] The experimental data analysis shows that the method of the present invention demonstrates significant advantages in predicting casting defects. First, both defect detection and prediction accuracy are high. Especially for casting B, the defect prediction accuracy reached 93.01%, significantly higher than the accuracy of traditional finite element simulation methods. This indicates that the method of the present invention can more accurately capture the location and size of defects, avoiding the problem of missed detection by traditional methods. The defect prediction accuracy of casting A also reached 92.35%, indicating that the method has strong universality across different types of castings and good generalization capabilities.

[0086] Secondly, the average training time is 12.35 hours, and the testing time is only 1.52 hours. This significantly reduces the computing time and resource consumption of traditional finite element simulation processes, thereby improving production efficiency. Compared with traditional defect prediction methods based on finite element simulation, which often require longer simulation times and are limited in defect prediction accuracy by the model's computing power, the method of the present invention achieves fast and highly accurate defect prediction through a deep learning model based on PointNet.

[0087] In terms of prediction accuracy, a comparison of the training and validation sets showed that the model demonstrated good training results and strong generalization capabilities. The prediction accuracy of the training set was 90.88% (Casting A) and 91.33% (Casting B), while the prediction accuracy of the validation set was 91.10% (Casting A) and 91.55% (Casting B). This demonstrates that the model effectively learns defect characteristics from the training data and maintains high prediction performance on unseen data.

[0088] In summary, the present invention, through the combination of finite element simulation and PointNet, not only improves the accuracy of casting defect prediction, but also significantly improves computational efficiency. Compared with traditional methods, the method of the present invention can significantly shorten prediction time and reduce the consumption of computing resources, providing a more efficient and accurate solution for casting defect prediction and process optimization. This result shows that the defect prediction method based on PointNet has important practical application value and can effectively improve the production efficiency and quality control level of aircraft engine tail nozzle adjustment plate investment castings.

[0089] Example 3 is the third embodiment of the present invention, which differs from the first two embodiments in that:

[0090] If the functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or the portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0091] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0092] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, and then editing, interpreting, or processing in another suitable manner as necessary, and then storing it in a computer memory.

[0093] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0094] Example 4 is the fourth embodiment of the present invention, which provides an aircraft engine tail nozzle adjustment plate investment casting defect prediction system, including:

[0095] The numerical simulation module performs numerical simulation on the regulating plate casting and the pouring system through finite element simulation software, obtains the location and size distribution results of the casting defects, and converts the results into three-dimensional point cloud data.

[0096] The model training module divides the three-dimensional point cloud data into a training set and a validation set, and uses the training set to train a PointNet-based defect prediction model.

[0097] The model validation module evaluates the prediction accuracy and generalization ability of the defect prediction model through the validation set and adjusts the hyperparameters based on the validation results.

[0098] The execution module converts the three-dimensional model of the casting system of the regulating plate to be predicted into three-dimensional point cloud data, inputs the trained defect prediction model, and outputs the location of the casting defect and the prediction result.

[0099] 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 the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for predicting defects in investment castings of aircraft engine tail jet regulating plates, characterized by: The method includes performing numerical simulation on the adjustment plate casting and the gating system by using finite element simulation software to obtain the location and size distribution results of the casting defects, and converting the results into three-dimensional point cloud data; The three-dimensional point cloud data is divided into a training set and a validation set. The training set is used to train a PointNet-based defect prediction model, including: local feature extraction and global feature aggregation, and the local features and global features are fused; the local feature extraction is to use the farthest point sampling method to extract index points of the three-dimensional point cloud data, and obtain a local point cloud set through a spherical neighborhood query method, extract the local geometric features of each point through a multi-layer perceptron with shared weights, and use a set abstraction layer to hierarchically capture multi-scale local features; Evaluate the prediction accuracy and generalization ability of the defect prediction model using the validation set, and adjust hyperparameters based on the validation results. Convert the 3D model of the casting system of the adjustment plate to be predicted into 3D point cloud data, input the trained defect prediction model, and output the location of the casting defect and the prediction result; The three-dimensional point cloud data includes: three-dimensional adjustment piece casting system point cloud data and defect point cloud data; The three-dimensional regulating plate gating system point cloud data is a complete casting and gating system structure, providing overall geometric information of the casting and gating system; The defect point cloud data is a defect area simulated and predicted to complete the defect prediction task; The converting of the result into three-dimensional point cloud data includes performing three-dimensional point cloud conversion on the position and size distribution results of the casting defects simulated by finite element simulation using a uniform sampling method.

2. The method for predicting defects in investment castings of aircraft engine tail jet modulators according to claim 1, wherein: The method of extracting index points of three-dimensional point cloud data using the farthest point sampling method includes randomly determining a point in the three-dimensional point cloud data as an initial point, using the farthest point sampling method to obtain the point with the farthest square Euclidean distance from the initial point, and obtaining an index point set and spatial structure features.

3. The method for predicting defects in investment castings of aircraft engine tail jet modulators according to claim 2, wherein: The method of obtaining a local point cloud set by using a spherical neighborhood query method includes selecting neighborhood points of the sampling point from the three-dimensional point cloud data using a spherical neighborhood query method with a specified radius, and obtaining a local point cloud set centered on each index point; The local geometric features of each point are extracted through a multi-layer perceptron with shared weights, and the multi-scale local features are captured hierarchically using the set abstraction layer. The three-dimensional coordinates, color information and normal vector information of each point in the point set are independently processed through the multi-layer perceptron layer with shared weights. The input dimension of each point is 9, including the three-dimensional coordinates , color information , normal vector information ; It is mapped to 64 dimensions and 128 dimensions through a shared multi-layer perceptron layer to capture the local features of each point.

4. The method for predicting defects in investment castings of aircraft engine tail nozzle fins according to claim 3, wherein: Global feature aggregation involves stacking set abstraction layers and aggregating the local features of each point with the spatial structure features using global pooling to form the 1024-dimensional global features required for casting defect prediction.

5. The method for predicting defects in investment castings of aircraft engine tail jet modulators according to claim 3, wherein: The fusion of local features and global features includes using a multi-layer perceptron to concatenate the 1024-dimensional global features with the 128-dimensional local features extracted from each set abstraction layer to form a 1088-dimensional feature vector to obtain a fused feature.

6. The method for predicting defects in investment castings of aircraft engine tail jet modulators according to claim 3, wherein: Fusing local features and global features also includes inputting the fused features into a shared multi-layer perceptron layer, generating prediction results point by point, and using the prediction results of each point to calculate the defect prediction model loss and update the model parameters.

7. A system for predicting defects in investment castings of aircraft engine tail jet flaps, using the method for predicting defects in investment castings of aircraft engine tail jet flaps according to any one of claims 1 to 6, characterized in that: include: A numerical simulation module, which uses finite element simulation software to perform numerical simulation on the adjustment plate casting and the pouring system, obtains the location and size distribution results of casting defects, and converts the results into three-dimensional point cloud data; A model training module divides the three-dimensional point cloud data into a training set and a validation set, and uses the training set to train a PointNet-based defect prediction model; The model validation module evaluates the prediction accuracy and generalization ability of the defect prediction model through the validation set and adjusts the hyperparameters based on the validation results; The execution module converts the three-dimensional model of the casting system of the regulating plate to be predicted into three-dimensional point cloud data, inputs the trained defect prediction model, and outputs the location of the casting defect and the prediction result.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for predicting defects of an aircraft engine tail nozzle adjustment plate investment casting according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting defects of an aircraft engine tail nozzle adjustment plate investment casting according to any one of claims 1 to 6 are implemented.

Images