TFNet: Few-shot identification of LTI systems based on convolutional neural networks

Abstract

Recent research in system identification aims to reduce the amount of data required for identification and the persistence of excitation while maintaining the identified model’s generalization and goodness of fit. This paper introduces TFNet11TF stands for Transfer Function. as a few-shot transfer function identification method for industrial systems. TFNet enables both structural and parametric identification of industrial linear time-invariant systems from their pulse response. Furthermore, TFNet identifies properties such as the existence of integrator and non-minimum-phase zeros in systems, as long as the order of the system is two or smaller. We compare the performance of TFNet with that of classical and state-of-the-art rival methods on syntactic data, where TFNet outperforms other methods by a wide margin. Furthermore, experiments on two laboratory setups: an air temperature control system and a two-tank control system demonstrate that TFNet estimates a close model to the ground truth despite colored measurement noise in the first experiment, and without driving the system out of its operating point in the latter case.22The python implementation of TFNet is in https://github.com/abbasnosrat/TFNet.git.