A Reinforcement Learning Control in Hot Stamping for Cycle Time Optimization.

Area of Science:

• Materials Science and Engineering
• Manufacturing Technology
• Artificial Intelligence in Manufacturing

Background:

• Hot stamping is a crucial metal forming process for ultra-high strength automotive parts.
• Reducing production cycle times is essential for improving Key Performance Indicators (KPIs) in hot stamping.
• Current cycle time management strategies may not fully optimize production efficiency or maintain product quality.

Purpose of the Study:

• To develop and evaluate a Reinforcement Learning (RL) approach for dynamic cycle time management in hot stamping.
• To optimize manufacturing production and minimize cycle time without compromising final product quality.
• To compare the RL approach against traditional control methods and dynamic programming.

Main Methods:

• Implementation of a Reinforcement Learning (RL) agent for real-time control of hot stamping parameters.
• Dynamic management of cycle time based on learned optimal behavior strategies.
• Comparative analysis of RL control with business-as-usual (BAU) methods and dynamic programming (DP).

Main Results:

• Reinforcement Learning control demonstrated a reduction in overall cycle time compared to BAU.
• The RL approach also decreased total batch time, particularly in non-stable temperature phases.
• RL achieved an optimal behavior strategy for dynamic cycle time management.

Conclusions:

• Reinforcement Learning offers a superior method for optimizing hot-stamping cycle times and batch durations.
• Dynamic management through RL enhances manufacturing efficiency and KPIs in hot stamping processes.
• The RL strategy effectively balances production speed with the critical requirement of final product quality.