Principal components analysis in estimation and control of paper machines

In paper machines the control objective is to maintain the paper properties (e.g. basis weight, thickness, moisture) as uniform as possible. This requires estimation of these properties from noisy data available from on-line sensors. In this work we use a particular principal components analysis technique, known as the Karhunen-Loeve (KL) expansion which does data compression and filtering. Spatiotemporally varying disturbance profiles are modeled by projecting the data on a lower dimensional subspace spanned by empirical orthogonal functions calculated from the data. The time series of the temporal modes or the coefficients for the KL expansion are modeled by an autoregressive model, and the resulting KL expansion is cast into a state-space form suitable for Model Predictive Control (MPC). We show with an example how a disturbance profile can be identified and controlled.