Modelling for Multi-Phase Batch Processes using Steady State Identification and Deep Recurrent Neural Network

Deep learning is widely used in soft sensors to monitor product quality in complicated batch processes. Most soft sensoring models are formed as auto-regressive with exogenous inputs (ARX) to deal with time series problems. However, it is difficult to identify the order of ARX model without any empirical knowledge. In order to avoid this issue, a novel modelling method combined the multi-phase deep recurrent neural network (RNN) and a steady state identification (SSID) is proposed in the study. First, the process is divided into several sub-phases by using SSID analysis with a moving window. Then, a RNN model is built in each sub-phase. The final state of each phase is used as the initial memory for next phase, which ensures the continuity of the entire process. The proposed algorithm is evaluated by the application in the penicillin fermentation process.