Developing Soft-Sensor Models Using Latent Dynamic Variational Autoencoders

Quality variables, which are usually measured offline, play important roles in describing process behaviors. However, online data obtained from soft sensors are significant as they provide accurate and immediate information. The reliability of online soft sensors is questionable due to changes in sensors, equipment, raw material availability, and operation conditions. In addition, chemical plants have dynamic properties and complex correlations amidst a large number of process variables. This causes most of the predictions obtained from steady-state soft sensors to be inaccurate in representing the particular chemical process. In this paper, the latent dynamic variational autoencoder is proposed to provide an estimation model and supervise soft-sensors. The input data are encoded in the latent space to remove underlying noises and disturbances in the data. Afterward, the dynamical properties are learned in the latent space through the bi-directional recurrent neural network, whose output (latent variable) is used to reconstruct back the input data. A simulation case study is conducted to show the effectiveness of the proposed method.