Reinforcement learning in free-form stamping of sheet-metals

Abstract

Sheet-metal free-form stamping technology deforms sheet-metals with simple and low costs universal tools on a working bench, which is normally an anvil. This traditional forming method is praised for its high forming flexibility but complained due to its reliance on individual experience thus low repeatability. In this paper, a python-based overall learning algorithm, which incorporates a reinforcement learning (RL) algorithm, for a designed sheet-metal free-form stamping case is developed. A neural network system, known as deep Q-network (DQN), was used to approximate the action-value function (Q function) in the Deep Q-learning algorithm. The DQN was trained using mini-batch training method, with the computational experiment data provided through Finite Element (FE) simulations. The overall learning algorithm was instantiated and evaluated by training the RL model to convergence, which is able to predict the optimal forming route to achieve the desired shape. This algorithm achieves the intellectualisation of the traditional free-form sheet-metal stamping process for the first time, without prior expertise for guidance.