Outliers in process modeling and identification

TLDR

Model‑based control strategies such as MPC rely on accurate process models, which are typically fitted using convenient structures like FIR, pole‑zero, Hammerstein, or Wiener models, but real measurement data often contain outliers that can severely degrade empirical identification. This paper investigates how outliers affect linear and nonlinear system identification using real datasets and discusses outlier detection and data cleaning. The authors fit convenient model structures—such as linear FIR, pole‑zero, Hammerstein, or Wiener models—to observed input‑output data and evaluate the impact of outliers on both linear and nonlinear system identification, while addressing outlier detection and data cleaning. Although no single strategy is universally applicable, the Hampel filter described here is often extremely effective in practice.

Abstract

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.