A nuclear power welding process adaptive control method based on reinforcement learning

Abstract

This invention relates to the field of intelligent welding, and more particularly to an adaptive control method for nuclear power welding processes based on reinforcement learning. The method comprises: obtaining a real-time process dataset; defining the state space and action space of the nuclear power welding process, mapping the welding process state to environmental state inputs, and mapping welding control parameters to action outputs; constructing a reinforcement learning model to model the state transition relationships of the welding process online, obtaining a welding process model; introducing a DODT architecture to construct a policy generation structure, forming a welding process control policy model; inputting the current welding process state into the welding process control policy model, and outputting the corresponding welding process control decision; collecting welding process feedback data during welding execution to form a welding process execution feedback dataset; and updating the welding process model and the welding process control policy model. This invention improves the consistency and adaptability between welding control decisions and the dynamic changes in the welding process.

Images