Gradient Boosting Regression Trees for Nonlinear Delay Identification in a Polymer Extrusion Process

Modern polymer extrusion processes, such as pipe productions, usually consist of several interconnected processing steps exhibiting nonlinear behavior. To support the operators at the production line, elaborate control designs are required based on data-driven models. For this purpose, finite impulse response (FIR) models are used to determine the system's behavior, particular the response delay. Unfortunately, conventional linear FIR models do not account for nonlinear dependencies and are prone to high variance for long process delays. On the other hand, Gradient Boosting Regression Tree (GBRT) is an established method in the field of machine learning producing competitive results for numerous nonlinear problems. Consequently, this work proposes a nonlinear FIR model using GBRT at its core. The method is compared to the conventional linear approach and cross-validated based on actual data from a polymer extrusion process. Additionally, the bias and consistency of the GBRT estimator for process delays are examined with the means of a Monte Carlo simulation. It is shown, that the GBRT FIR model shows superior results for the investigated nonlinear, noisy and slow process with long delays.