IEEE AI4IM 2026

Artificial Intelligence (AI) enables the creation of virtual/soft sensors, by enabling the prediction of the desired quantity in locations otherwise not accessible by traditional sensors due to systems functional or other restrictions. This enables a better understanding of the systems and their further enhancements. In this paper, we comprehensively evaluate several AI-models towards their performance as virtual temperature sensors within a Nanopositioning System (NPS).

The goal is to build reliable virtual sensors to later enhance the positioning precision of the NPS where the positioning accuracy and precision are negatively affected due to environmental disturbances such as variations in temperature. We select different AI architectures as possible solutions for predicting temperature as the disturbance variable and evaluate those based on real-world data of the coil assembly component of the NPS.

Our results reveal a number of challenges related to the creation of such virtual sensors that need further attention in the future. One of the challenges are the observed residual trends of the deep learning models as often unnoticed problem. Further challenges are non-fulfilling the required measurement accuracy for later compensation for the disturbances caused by the temperature variations and transfer of the models to the final system with unsatisfactory performances in the source systems.