Outliers in process modeling and identification

Model-based control strategies like model predictive control (MPC) require models of process dynamics accurate enough that the resulting controllers perform adequately in practice. Often, these models are obtained by fitting convenient model structures (e.g., linear finite impulse response (FIR) models, linear pole-zero models, nonlinear Hammerstein or Wiener models, etc.) to observed input-output data. Real measurement data records frequently contain "outliers" or "anomalous data points," which can badly degrade the results of an otherwise reasonable empirical model identification procedure. This paper considers some real datasets containing outliers, examines the influence of outliers on linear and nonlinear system identification, and discusses the problems of outlier detection and data cleaning. Although no single strategy is universally applicable, the Hampel filter described here is often extremely effective in practice.