Input selection for dynamic RBF models in process monitoring

This paper investigates the monitoring of continuous processes using dynamic nonlinear principal component analysis (NPCA). Previously, it was shown that integrating the RBF networks with principal curves significantly had increased the sensitivity of fault detection for nonlinear processes. Despite this, the previous method may not function well for processes which exhibit strong dynamic characteristics. An effective method of capturing dynamic behaviour is to consider a time-lagged data extension. However, the augmented data matrix may lead to the inclusion of a large number of variables in the RBF network input, and hence increase the computational load and network complexity. To prevent this, an input selection scheme, based on the nonlinear dynamic relationship underlying the process variables, is introduced. This selects the most important and relevant time-lagged variables before constructing the RBF network model. Consequently, a modified dynamic NPCA approach is now proposed. The advantages of this improvement are demonstrated using a benchmark simulation example from the literature.