Process Monitoring Using a Sequence to Sequence Model

Industrial process monitoring systems aim to mine valuable patterns from a large volume of process data for detecting abnormalities in an efficient and effective manner. To deal with nonlinearity and sequentiality of process data, various nonlinear and dynamic models have been adopted in previous research, which usually take both manipulated and measured variables into consideration to maximize the utilization of the available information. However, from the viewpoint of process engineers, changes only in manipulated variables should not be considered as faults. Instead, these are routine control actions to reject disturbances. As long as the process can be regulated, it is unnecessary to raise alarms. In this work, a sequence to sequence neural network model with gated recurrent units (GRUs) is proposed for efficient process monitoring, while superfluous alarms are suppressed. The feasibility of the proposed method is illustrated by the case study on the benchmark Tennessee Eastman (TE) process.