Multi-Agent Path Finding (MAPF) is crucial for achieving efficient coordination in large-scale robotic systems, particularly within the context of industrial intelligent decision-making and smart manufacturing. Although progress has been made in exploring decentralized policies using reinforcement learning (RL) in partially observable environments, developing collision-free strategies in dense environments remains challenging. This paper proposes a novel approach that integrates communication with deep Q-learning, enhancing inter-agent collaboration through graph convolution techniques. Unlike traditional methods, this study introduces an improved Efficient Channel Attention (ECA) mechanism to dynamically optimize channel responses within the network, significantly improving the model’s ability to respond to critical environmental features. Furthermore, a position-encoding-enhanced multi-head attention mechanism is employed to optimize information exchange among agents, thereby improving decision efficiency and accelerating learning speed. The model is trained independently from the perspective of a single agent, and the learned policies are subsequently applied to each agent, ensuring decentralized execution. System-level training adopts a distributed architecture optimized via curriculum learning strategies. In densely obstructed testing environments, our approach demonstrates not only a high success rate but also reduced average steps, contributing to more effective intelligent decision-making processes in industrial applications.