Abstract Predicting and optimizing extrusion molding process parameters is a significant and challenging research topic in the field of manufacturing, but machine learning-based approaches remain relatively scarce. In this study, orthogonal experiments were designed to simulate the extrusion process using Simufact Forming software. The thinning rate, stamping speed, and friction coefficient were used as independent variables. Subsequently, prediction models for extrusion force and equivalent force were established based on a Radial Basis Function (RBF) neural network. The data obtained from the simulation software were used as the training set, and additional experiments outside the training set were designed for validation. The model achieved a relative error below 6%, demonstrating its reliability. This study not only proposes a novel method for predicting process parameters in steel pipe extrusion but also holds significant value for optimizing these parameters and improving product quality. The application of this model can benefit various manufacturing sectors, including automotive, aerospace, and construction, where precise control of extrusion parameters is critical for enhancing product performance and reducing material waste.
Introduction: The Challenge of Extrusion Molding
Extrusion molding—a process where metal is pushed through a die to create complex cross-sections—is a cornerstone of the automotive, aerospace, and biomedical industries. However, finding the optimal process parameters, such as speed and friction, is notoriously difficult. Traditionally, engineers have relied on expensive, time-consuming physical experiments or complex manual calculations to determine necessary extrusion forces. Inaccurate predictions lead to poor product quality, material waste, and excessive equipment wear. While machine learning offers a potential solution, its application in this specific niche of metallurgy has remained relatively scarce.
Methodology
To bridge the gap between simulation and machine learning, researchers developed a high-precision predictive model using a three-step approach:
- Digital Simulation: Using Simufact Forming software, a "digital twin" of a C15_c steel pipe extrusion process was created. This allowed for the generation of high-fidelity data without the cost of physical prototypes.
- Orthogonal Experimental Design: To ensure a robust dataset despite a small sample size, an orthogonal design was used to systematically vary three independent variables: thinning rate, stamping speed, and the friction coefficient.
- RBF Neural Network: A Radial Basis Function (RBF) neural network was selected for the predictive engine. Unlike standard deep learning models that require massive datasets, RBF networks excel at "local approximation," learning complex, nonlinear relationships from a relatively small number of data points. The network utilized a Gaussian function as the kernel and was optimized via a grid search algorithm.
Data and Training
The study utilized synthetic data generated through simulation, consisting of a training set of 7 experimental groups and a validation set of 5 unique scenarios. Input features included thinning rate (%), punching speed (mm/s), and the coefficient of friction (N/m²). The target outputs were equivalent force (MPa) and extrusion force (KN). Data normalization to a [0, 1] range was critical for the RBF network to converge and perform reliably.
Key Results and Insights
The model demonstrated impressive accuracy when tested against the validation set, achieving a relative error of less than 6% for both extrusion and equivalent forces. For stress predictions specifically, relative errors were consistently below 5%.
A significant finding was the effectiveness of the RBF network on a very small dataset. The hidden layer successfully mapped inputs into a higher-dimensional space where complex relationships became manageable, making the prediction much more accurate than traditional statistical methods. Visual analysis confirmed that the predicted values closely tracked the "true" values from the simulation across different experimental conditions.
Conclusion and Future Work
While the results are promising, the model was validated on simulation data. Real-world factory conditions involve "noise" like temperature fluctuations and material impurities that simulations might not fully capture. Future research aims to expand the model's applicability to more complex extrusion shapes and integrate it into real-time production monitoring systems. This would allow for "on-the-fly" adjustments during the manufacturing process, further reducing waste and improving product consistency.
Summary updated