Batch Process Monitoring with Two-Dimensional LSTM-Variational Auto-Encoder

Batch process monitoring is crucial in production loss prevention and process safety promotion. However, the two-dimensional nonlinear dynamics in the batch process make the monitoring cumbersome. This paper proposes a novel dynamic deep model called two-dimensional LSTM variational auto-encoder (2DLSTM-VAE). Structurally, our model constructs with LSTM encoded under the conditional variational Bayes framework, which can cope with the nonlinear dynamics from two dimensions. Afterwards, a systematic fault detection framework is defined under the latent and residual domains to account for each running batch along with those run-to-run variations so as to reflect the process variability in the two-way monitoring aspects. Finally, a deep reconstruction-based contribution (DRBC) diagram is further developed to investigate root causes. For the case study, the validity of the 2DLSTM-VAE model is demonstrated in the penicillin fermentation process, and the results show that compared with other methods, the proposed deep model has better performance on fault detection and diagnosis.