abstract
- © 2019 Elsevier Ltd The non-parametric identification problem aims to estimate a suitable model based on the response produced by a given stimulus on an uncertain model. Complementary, input estimation considers a different problem where the model and the output are known. However, if neither model nor input is known, the identification problem seems to be more complicated. This is a major challenge in electrophysiological systems where the output signal can be measured, but the biological system is uncertain and the stimuli are unknown. The aim of this study was to introduce a novel methodology that may estimate the uncertain input in this more complex situation where the biological model of the system under analysis is uncertain and the input stimuli must be estimated accurately. A two step non-parametric identification scheme based on differential neural networks (DifNN) and the second order super-twisting sliding mode algorithm (STA) identified the input stimulus. The STA makes a robust exact estimation of the output signal. The DifNN produces an internal non-parametric mathematical modeling of the input-output relationship executed without the input information. Learning laws for both levels of DifNN modeling were presented as part of the uncertain input identification process. The estimation of uncertain visual stimulus applied over the retina in an ex-vivo avian model (EVAM) served to test the method proposed in this study. The output information corresponded to the electrophysiological voltage variation in the optical nerve.