Aiming at the problems of low recovery rate of tailings resources and poor adaptability to dynamic working conditions, this paper proposes a reinforcement learning (RL) optimization method driven by Industrial Internet of Things (IIoT). By deploying a multi-source sensor network for real-time sensing of ore grade, flow rate and equipment status, a digital twin collaborative framework is constructed; a multi-intelligent body depth deterministic policy gradient (MADDPG) algorithm is innovatively designed to realize dynamic decision-making for agent addition and equipment control. Based on the open data set validation, the system recovery rate is increased to 92.3%, and the energy consumption per ton of ore is reduced by 18.7%, which significantly optimizes the dynamic resource recovery efficiency. The method provides key technical support for the green transformation of the mining industry.