Neural Networks for Data-driven Modeling of Active Magnetic Bearing Suspended Rotor System
Niko Nevaranta, Aleksandr Shishkov, Ibrahim Abubakar, Atte Putkonen, Marek Rehtla, Gyan Ranjan, Tuomo Lindh
Abstract
Active magnetic bearing (AMB) suspended rotor systems are often modeled and validated through system identification experiments, aiming to derive parametric or nonparametric models for analysis. Given the complexity of this type of system compromising non-linear electrical and speed-varying mechanical systems, neural network based prediction models are worth pursuing for to increase the modeling options. This paper addresses issues in the training of feedforward neural network (FNN) for SISO and MIMO identification cases of AMB suspended rotor system, when artificially generated excitation signals are superposed to the position controller output. Particular attention is given to the predictive capability of the trained models when applied to unseen datasets. The trained networks are validated by a test rotor system using the data collected.