Nonlinear Dynamic System Identification of Separation Unit for Raw Materials

This paper focuses on nonlinear dynamic modeling of the combine harvesters cleaning unit. Therefore, first, the cleaning sub process is characterized in detail. The main properties are the nonlinear dynamic process itself as well as the upstream transportation process resulting in an overall time delayed nonlinear dynamic process. Since no first principle models exist for the dynamics of the process a black box identification of the process is carried out in a two-step procedure. First, the time delay is estimated separately following a brute-force method and the data is compensated for it. Secondly, nonlinear dynamic system identification is performed. To find an appropriate model architecture, a benchmark of four data-driven models is accomplished. The compared models are chosen to realize different dynamic realizations ranging from more complex feedback structures (internal or external) as well as pure feedforward architectures. Additionally, the nonlinear approximator is varied. The benchmark shows that different architectures are capable to model the combine harvesters cleaning unit. Depending on the limitations for the number of parameters or training time an appropriate choice has to be made. Furthermore, the investigation points out the difficulty to deal with real measurement data influenced by external conditions which are not available in the modeling phase.