Spatiotemporal Stacked Autoencoder based Soft Sensor Modeling for the Dow Data Challenge Problem

Process data with characteristics such as strong nonlinearity, high dimensionality, cross-correlations and auto correlations pose a great challenge for data-driven soft sensor modeling. Albeit the conventional stacked Autoencoder (SAE) is able to learn low dimensional nonlinear features from the process data, it fails to extract spatial and temporal features. To overcome this shortcoming, in the present work, spatiotemporal SAE (STSAE) is proposed based on convolutional and LSTM layers. Lower dimensional nonlinear features (spatial and temporal) extracted from STSAE are used to build the soft sensor. The effectiveness of the STSAE-based soft sensor is demonstrated in an industrial case study, the Dow data challenge problem. For the purpose of comparison, soft sensor models based on popular machine learning algorithms are developed and applied to the industrial case study. The obtained results demonstrate that the proposed STSAE based soft sensor yields more accurate estimations for impurity level at the primary column outlet of Dow process refining system.