This invention discloses a method for non-disruptive switching of redundant control based on
deep learning and
reinforcement learning, belonging to the field of redundant control technology in industrial
automation. To address the problems of switching output jumps, process quantity fluctuations, and erroneous switching caused by network
latency jitter, state asynchrony, and inconsistent integral filtering states in dual-
machine hot standby or multi-controller redundant architectures, as well as the lack of reliable and rapid
rollback after online upgrades, this invention collects primary and
backup process quantities, control outputs, synchronization states, and latency data. It utilizes
deep learning to assess consistency and switching risks, employs meta-
reinforcement learning to generate parameterized progressive fusion takeover curves and evaluates them in shadow mode. Combining feature representation drift and control performance drift
hysteresis gating, it determines takeover or
rollback, achieving state alignment, non-disruptive takeover, and automatic
rollback in case of anomalies, reducing switching deviations and shortening
recovery time.