In recent decades, soft sensors have been profoundly studied and successfully applied to predict critical process variables in real-time. While dealing with various application scenarios, data-driven methods with representation learning possess great potentials. Latent features are formulated in these approaches to predict outputs from correlated input variables. In this study, a probabilistic framework of feature extraction is proposed in the context of process data analysis. To address switching behaviors in industrial processes, multiple emission models are utilized to construct latent space. To address temporal correlations from continuously operating processes, a dynamic model is implemented in latent space. Bayesian learning strategy is then developed for parameters estimation, where modeling preferences and uncertainties from multiple models are considered. The effectiveness and practicability of the proposed feature extraction algorithm are illustrated through numerical simulations, as well as an industrial case study. (c) 2018 American Institute of Chemical Engineers AIChE J, 64: 2037-2051, 2018