Abstract: Aiming at the problem of lack of effective intelligent control methods in coal slurry flotation industrial site, an AS-DDPG control method based on improved deep reinforcement learning is proposed. The method introduces an attention mechanism on the basis of Actor-Critic network, which can accurately extract the key features in the temporal data and solve the problem of slow convergence speed faced by the classical DDPG algorithm when it is applied individually to the process control of time-varying objects. In terms of controller design, a multi-dimensional state space containing key parameters such as slurry concentration, ash, flow rate, etc., and a multi-objective reward function that takes into account the quality of tailings and the recovery rate of pharmaceuticals are established, which solves the control problems caused by the lack of model accuracy in traditional intelligent control based on the mechanistic model. The experimental phase uses field data to drive model training and achieve iterative strategy optimisation, and the simulation results show that the AS-DDPG algorithm reduces the training error by 27% compared with the classical DDPG method. The results of the field industrial test show that the standard deviation of ash is reduced by 32.3%, the consumption of trapping agent is reduced by 23.2%, and the consumption of frothing agent is reduced by 10.7%, which demonstrates the effectiveness of this algorithm for the intelligent control of coal slurry flotation, and provides a useful reference for the intelligence of coal slurry flotation.