S. Lee, L. Quagliato, J. Sun, N. Kim
Keywords: preform design, forging, piston forging, convolutional neural network, deep learning, AI
Summary:Preform design has largely been based on experience rather than engineering, a fact that leads to expensive and time-consuming trial-and-error procedures. In order to define a more engineering-oriented approach to preform design, a methodology based on the convolutional neural network is introduced to design optimal preform in forging processes. The proposed approach has been defined by means of numerical simulation of a piston head where the requirements have been set on the full filling of the die and no folding of the material flow. The learning model for prediction is based on the U-Net (with 19 convolutional and 8 pooling layers). Convolutional layers were used to extract features of piston and preform shape and more complicated features are calculated as the shape varies in size by pooling layer. The learning model could predict the preform shape robustly through the extracted features. In the pre-processing step, the piston and corresponding preform were made using FEM. the FEM database consists of 24 training data, 6 tuning data, and 2 verification data. The piston and preform cad files were converted into a 3D voxel file. In the learning process, the final piston shape for X and the preform shape for Y was used where X means input and Y means golden truth in the U-Net model. The model was trained with 4 batch sizes, 0.0001 learning rate, and 300 epochs. After the training of the model, a final piston design, not used in training, has been used to predict the preform design. The predicted preform shape was rendered and converted into cad file. FEM analysis was conducted to show that the predicted preform design allows minimizing the defects during the forging process. According to the validation results, the proposed model, based on the convolutional neural network of deep learning, has shown to be reliable in defining the preform design for piston, which allows minimizing the defects during the hot forging process. In the future, the deep learning based methodology could be used to predict the optimal preform design of parts in aircraft in the forging or preform design for shape rolling process.