Mode Information Separated β-VAE Regression for Multimode Industrial Process Soft Sensing

Multimode processes commonly exist in real industries owing to the changes in production batches and requirements. Since different data modes are derived from the same reaction process, certain commonalities that represent the substantial characteristics of the process could exist. To explore and extract those substantial commonalities, this article develops a self-weighted
$\beta $
variational autoencoder regression (SW-
$\beta $
-VAER) model for observation and measurement of industrial processes with separated multimode information based on soft sensors. The SW-
$\beta $
-VAER dissociates modes information from multimode processes, extracting major disentangled features and mode-independent features using
$\beta $
variational autoencoder (
$\beta $
-VAE) and parameterization, respectively. A regression model is established depending on these two fused features. Moreover, to strengthen disentangled feature expression between different modes and enhance the generalization abilities of the previous model, a mode adaptive self-weighted
$\beta $
-VAE-based regression (MA-SW-
$\beta $
-VAER) model is further proposed, integrated with a mode classifier based on the gradient reversal layer (GRL) strategy. Finally, the proposed two models are equipped as soft sensors in two industrial cases, where given comparisons with other methods demonstrate their effectiveness and superiorities.