Delay-aware identification for online controller tuning: A deep learning-based experimental modal analysis approach for nanopositioning systems

Real-time control of nanopositioning systems requires robust and accurate identification of dynamic parameters, especially when system properties vary due to changes in the scanning environment. This paper proposes a novel deep learning-based methodology for rapid modal parameter extraction to enable online tuning of integral resonant controllers (IRC). The approach builds upon a previous data-driven tool, NeuroID, which employs convolutional neural networks to estimate modal parameters from frequency response functions of single-degree-of-freedom systems.
Given the high natural frequencies of nanopositioners, time delays arising from the actuation and measurement chain significantly affect the system response. To address this limitation, the previous tool is extended into a Delay-Aware NeuroID, denoted DANID, which is both robust against time delays and capable of estimating them. A detailed analysis of the discrete domain implications of peak-picking techniques and acquisition conditions is conducted to establish a reliable identification framework.
The proposed methodology is trained exclusively using synthetic data. Subsequently, simulation analyses are performed under varying acquisition conditions and varying modal-property scenarios. Finally, the methodology is experimentally validated on a real nanopositioner platform. The new tool achieves identification accuracy comparable to conventional curve-fitting techniques while reducing the computation time by approximately 62%. Compared with NeuroID, the proposed approach reduces the average mean normalised error by a factor of 6.52. Finally, when integrated into a delay-sensitive IRC design, DANID can increase the achievable closed-loop bandwidth by up to 12.3 times with respect to NeuroID, successfully handling delay and enabling automatic IRC retuning in response to dynamic changes.