Proactive management of industrial alarm floods: A reinforcement learning framework for early prediction and operator support

An industrial alarm flood indicates the emergence of a major problem and requires immediate and effective measures to mitigate the situation. Without essential information on the current and upcoming alarms, it is difficult to respond efficiently, especially when the alarm rate is significantly high. Thus, a reinforcement learning (RL) approach is proposed in this work for early prediction of an industrial alarm flood so as to provide critical decision support to industrial operators in real-time. The proposed method is implemented mainly in the following steps: (1) An alarm flood pattern extraction strategy is adopted to exclude irrelevant alarms and generate potential online scenarios by exploiting the alarm relations in the historical alarm flood sequences; (2) To analyze alarm information effectively, a novel textual vectorization method based on mutual information is proposed; (3) Finally, the early prediction problem is formulated as a partially observable Markov decision process (POMDP) and the double deep Q-network (DDQN) algorithm is adopted, with modifications to the learning process to ensure both accuracy and earliness. The effectiveness of the proposed method is demonstrated using real industrial data from an oil refinery.