Adaptive virtual sensors using SNPER for the localized construction and elastic net regularization in nonlinear processes

Abstract Traditional data-driven virtual sensors were constructed upon the macro-perspective of the manifold structure using generalized models. They did not focus on the local relationships among the data samples through micro-perspective of manifold proximity indicating the local relationships among the data samples. In the case with the quantity of data points fewer than the dimensions of the data variables, the virtual sensor model is likely to be unstable, ill-conditioned, and computationally expensive. And the real-world data often vary with time. It is difficult, in the long term, to sustain good performance by a single fixed model. This paper addresses the aforementioned issues by proposing three algorithms (NPER, SNPER, and LW-SNPER) to successively improve the virtual sensor modeling performance for nonlinear, high-dimensional and time-varying processes. It is shown through the numerical cases and a real semiconductor process that the proposed algorithms perform better than the other regular regression algorithms.