Deep reinforcement learning-based orders splitting method for the wire bonding process in semiconductor chip packaging and testing workshops

The wire bonding process is a significant production bottleneck in the semiconductor chip packaging and testing workshops, which has the characteristics of large-scale multi-type production orders and parallel machine. Production order splitting is critical for improving the scheduling performance of the wire bonding process. This paper proposes a wire bonding process order splitting method using deep reinforcement learning (DQN), aiming to split production orders into optimized sub-batches. Firstly, factors such as order types, quantities, processing times, and machine capacities are adopted to construct the DQN’s state space. Then, an innovative sub-batch order temporary storage location is designed to build the action space. Finally, a reward function considering the makespan is built to guide DQN’s decision-making. Experimental results show that the proposed order splitting method is superior to traditional order splitting methods in terms of the complete time. Therefore, the proposed method is effective.