Industrial Virtual Sensing for Big Process Data Based on Parallelized Nonlinear Variational Bayesian Factor Regression
Zeyu Yang, Zhiqiang Ge
Abstract
Virtual sensors are mathematical methods that describe the dependence of primary variables on secondary variables. For the majority of industrial processes with particularly nonlinear characteristics, traditional linear virtual sensors may not function well. Based on probabilistic modeling, this article aims to extend the linear probabilistic virtual sensor to the nonlinear form, with incorporation of the nonlinear mapping technique. Especially, an enhanced nonlinear variational Bayesian factor regression (NVBFR) algorithm is proposed for virtual sensing of nonlinear processes. Meanwhile, with the ever increasing data size collected from the processes, the era of big data has arrived in the industrial process. Since the complexity of parameter updating is highly related to both sample size and number of dimensionality, intractable computing problems often occur in practice. To this end, a parallel framework-based NVBFR (P-NVBFR) is further proposed to tackle the big data problem. To evaluate the feasibility and efficiency of the developed virtual sensors, a real industrial example is demonstrated.