Soft-sensor reliability evaluation and y-analyzer fault identification with applications to vinyl acetate monomer (VAM) benchmark process

摘要

A great deal of research has been done to develop data-driven soft-sensors for quality estimation and control. However, a soft-sensor does not always function well. If estimates of the soft-sensor are blindly believed and used in a control system, the product quality and process performance will be deteriorated. To solve this issue, an on-line reliability evaluation method of soft-sensors using the k-nearest-neighbor rule (RE-kNN) was developed in this study. RE-kNN makes full use of the kNN rule to obtain local information, therefore it is suitable for nonlinear and multimodal cases. Since RE-kNN is a model-free approach, it can be applied to reliability evaluation of any inferential model. In addition, simple y-analyzer fault identification rules were also proposed by integrating RE-kNN with the 3-sigma method. The validity of the proposed methods was verified through numerical examples and a new, rigorous vinyl acetate monomer plant model. The application results demonstrate that the proposed methods are efficient to evaluate the reliability of soft-sensors and to detect y-analyzer faults.