Dynamic process calibration based on sparse partial least squares

This article proposes a sparse partial least squares (SPLS) for model calibration of dynamic processes. Via capturing the relationship of process inputs and measurements at different sampling instances, partial least squares (PLS) is a typical multivariable statistical process control technique to model dynamic processes. However, due to rare process measurements, large number of process variables and large time scale of process dynamics, the over-fitting problem will be obvious and the calibration performance will be degraded. With the sparse representation, SPLS produces a more reliable model to capture the process dynamics, which won't be deteriorated by the small sample size problem. Case studies on a simulation example and the Tennessee Eastman (TE) process illustrate the effectiveness of the proposed method.