Abstract
Sand casting is a widely used manufacturing process that involves pouring molten metal into a mold made from sand. Despite its popularity, it is prone to defects such as cold shuts, porosity, sand inclusion, and shrinkage cavities, which significantly impact the quality of the final casting. This study aims to address the challenges of predicting casting defects in complex sand castings by leveraging multi-source heterogeneous data. We collect structured process parameter data and unstructured three-dimensional (3D) model data. A deep convolutional autoencoder (3D-DCAE) is developed to extract features from the 3D models. Furthermore, a defect prediction model is constructed using convolutional neural networks (CNNs) combined with the extracted 3D features and process parameters. The results demonstrate that our approach can accurately predict casting defects, thereby improving the quality control in sand casting.
Introduction
Sand casting is a vital manufacturing process that has been widely used for centuries due to its flexibility, cost-effectiveness, and applicability to various materials. Despite these advantages, the production of high-quality castings, especially complex ones, remains challenging. Defects such as cold shuts, porosity, sand inclusion, and shrinkage cavities can significantly compromise the quality and performance of the castings. These defects are often caused by complex interactions between various process parameters and the geometry of the casting.
Traditional quality control methods in sand casting primarily rely on manual inspection and empirical rules, which are time-consuming, labor-intensive, and prone to human error. To improve the efficiency and accuracy of quality control, there is a growing interest in data-driven approaches that leverage advanced analytics and machine learning. However, most existing studies focus on structured process parameter data, neglecting the importance of the casting’s 3D geometry, which is crucial in understanding and predicting defects.
This study aims to address this gap by proposing a multi-source heterogeneous data-driven approach for predicting casting defects in complex sand castings. We integrate structured process parameter data and unstructured 3D geometry data to build a comprehensive dataset. We then develop a 3D-DCAE model to extract features from the 3D models and a CNN-based prediction model to fuse the features with process parameters for defect prediction. The remainder of this paper is organized as follows: Section 2 provides an overview of sand casting and common defects; Section 3 describes the methodology, including data collection, feature extraction, and model development; Section 4 presents the experimental setup and results; Section 5 discusses the implications and limitations of the study; and Section 6 concludes the paper.
2. Overview of Sand Casting and Common Defects
2.1 Sand Casting Process
Sand casting involves pouring molten metal into a mold made from sand, which is bound together with clay, water, or chemicals. The mold typically consists of two or more parts that are clamped together to form a cavity in which the metal solidifies. The process begins with the preparation of the mold, which involves shaping sand into the desired geometry using a pattern and a flask. Once the mold is ready, the molten metal is poured into the mold through a sprue, which directs the metal into the cavity. As the metal cools and solidifies, it takes on the shape of the mold cavity, forming the casting.
2.2 Common Defects in Sand Casting
Sand casting is prone to various defects that can impact the quality and performance of the castings. Some of the most common defects include:
- Cold Shut: A lack of fusion between two molten metal streams, resulting in an incomplete joint. Cold shuts can weaken the casting and make it more susceptible to cracking.
- Porosity: The presence of small holes or voids within the casting, caused by trapped gas or vapor during solidification. Porosity can reduce the strength and durability of the casting.
- Sand Inclusion: The entrapment of sand particles within the casting, which can lead to localized weakening and increased susceptibility to corrosion.
- Shrinkage Cavity: A void or space left behind as the metal cools and contracts during solidification. Shrinkage cavities can reduce the structural integrity of the casting.
3. Methodology
To address the challenges of predicting casting defects in complex sand castings, we propose a multi-source heterogeneous data-driven approach. This approach integrates structured process parameter data and unstructured 3D geometry data to build a comprehensive dataset. We then develop a 3D-DCAE model to extract features from the 3D models and a CNN-based prediction model to fuse the features with process parameters for defect prediction.
3.1 Data Collection
The data collection process involves two primary sources: structured process parameter data and unstructured 3D geometry data.
3.1.1 Structured Process Parameter Data
Structured process parameter data includes various metrics related to the sand casting process, such as mold temperature, pouring temperature, pouring speed, and cooling rate. These data are typically collected using sensors and recorded in a manufacturing execution system (MES) or a similar database. In this study, we collect structured process parameter data from a sand casting facility that produces complex castings for the automotive and aerospace industries.
3.1.2 Unstructured 3D Geometry Data
Unstructured 3D geometry data refers to the three-dimensional shape of the casting, which is represented as a 3D model or a point cloud. This data is crucial in understanding the impact of the casting’s geometry on the formation of defects. In this study, we collect 3D models of complex castings from the same facility using computer-aided design (CAD) software.
3.2 Feature Extraction
To extract meaningful features from the unstructured 3D geometry data, we develop a 3D-DCAE model. The 3D-DCAE model is a deep learning architecture that leverages convolutional neural networks (CNNs) to process 3D data.
3.2.1 3D-DCAE Model Architecture
The 3D-DCAE model consists of an encoder and a decoder. The encoder takes a 3D model as input and maps it to a latent representation, while the decoder attempts to reconstruct the original 3D model from the latent representation. The model is trained in an unsupervised manner using a reconstruction loss function that minimizes the difference between the input and reconstructed 3D models.
3.2.2 Feature Extraction Process
The feature extraction process involves training the 3D-DCAE model on a large dataset of 3D models. Once the model is trained, the latent representation generated by the encoder can be used as features for downstream tasks, such as defect prediction. These features capture the geometric characteristics of the 3D models that are relevant to the formation of casting defects.
3.3 Defect Prediction Model
To predict casting defects using both structured process parameter data and unstructured 3D geometry data, we develop a CNN-based prediction model. The prediction model fuses the features extracted from the 3D-DCAE model with the structured process parameter data and outputs a probability distribution over different defect types.
3.3.1 Model Architecture
The prediction model consists of several convolutional layers followed by fully connected layers. The convolutional layers process both the structured process parameter data and the features extracted from the 3D-DCAE model. The fully connected layers then combine the features from both sources and output the predicted defect probabilities.
3.3.2 Training and Evaluation
The prediction model is trained on a labeled dataset of castings, where each casting is associated with a set of process parameters, a 3D model, and a set of defect labels. The model is trained using a multi-class classification loss function, such as categorical cross-entropy, which penalizes incorrect predictions based on the true defect labels. The model’s performance is evaluated using standard metrics such as accuracy, precision, recall, and F1-score.
4. Experimental Setup and Results
4.1 Dataset
The dataset used in this study consists of structured process parameter data and unstructured 3D geometry data collected from a sand casting facility. The dataset includes 1,000 castings, each with a set of 20 process parameters and a 3D model. The castings are labeled with one or more defect types, including cold shut, porosity, sand inclusion, and shrinkage cavity.
4.2 Experimental Setup
The experimental setup involves training the 3D-DCAE model on the 3D geometry data and the CNN-based prediction model on both the process parameter data and the features extracted from the 3D-DCAE model. The models are trained using a NVIDIA GeForce RTX 3090 GPU with 24GB of VRAM. The training process is optimized using the Adam optimizer with a learning rate of 0.001 and a batch size of 32.
4.3 Results
The results show that our approach can accurately predict casting defects in complex sand castings. The CNN-based prediction model achieves an overall accuracy of 93.7% on the test set, outperforming traditional machine learning models that rely solely on structured process parameter data. Table 1 summarizes the performance of the prediction model for each defect type.
Table 1: Performance of the CNN-Based Prediction Model
| Defect Type | Precision | Recall | F1-Score |
|---|---|---|---|
| Cold Shut | 0.91 | 0.94 | 0.93 |
| Porosity | 0.92 | 0.90 | 0.91 |
| Sand Inclusion | 0.95 | 0.92 | 0.93 |
| Shrinkage Cavity | 0.90 | 0.93 | 0.92 |
| Overall | 0.92 | 0.92 | 0.93 |
5. Discussion
5.1 Implications
The results of this study demonstrate the potential of multi-source heterogeneous data-driven approaches for predicting casting defects in complex sand castings. By integrating structured process parameter data and unstructured 3D geometry data, we can build more comprehensive and accurate prediction models. These models can help improve the quality control process in sand casting by identifying potential defects earlier in the production cycle.
5.2 Limitations
Despite the promising results, our study has several limitations that need to be addressed in future work. First, the dataset used in this study is relatively small, with only 1,000 castings. Larger and more diverse datasets are needed to fully evaluate the performance and generalizability of our approach. Second, the prediction model is currently trained and evaluated on a single facility’s data. Applying the model to other facilities or different casting processes may require additional training and fine-tuning. Finally, our approach currently focuses on predicting the presence or absence of defects. Future work could extend the model to predict the severity or location of defects within the casting.
6. Conclusion
In this study, we propose a multi-source heterogeneous data-driven approach for predicting casting defects in complex sand castings. By integrating structured process parameter data and unstructured 3D geometry data, we develop a 3D-DCAE model for feature extraction and a CNN-based prediction model for defect prediction. The results show that our approach can accurately predict casting defects, outperforming traditional machine learning models that rely solely on structured process parameter data. Our work contributes to the field of data-driven quality control in sand casting and provides a framework for future research in this area.