Abstract

Accurate state and parameter estimation is essential for monitoring and control, yet high-fidelity models are often unavailable or costly to obtain. As a result, many applications rely on data-based virtual sensors that learn mappings from time-windowed measurements or derived features to the quantities of interest. While effective, such approaches are prone to changes in process parameters and operating conditions—such as wear, load variations, or disturbances—which alter the underlying data–state relationship and introduce concept shift. Ensuring robustness therefore requires both selecting features that are insensitive to such shifts and choosing a mapping that preserves discriminative information under varying conditions. We analyse this challenge in the concrete setting of electromagnetic solenoids, where accurate estimation of the armature position is crucial for control and diagnostics but typically requires costly position sensors. However, virtual-sensor performance often degrades under varying external loads and wear-induced changes, which disrupt the current–position relationship. To mitigate this effect, we propose a virtual sensor design that (i) employs physics-derived features capturing position-dependent electromagnetic effects and (ii) trains the mapping on a mixture of data from multiple operating contexts to reduce shift sensitivity. Experimental results on a solenoid test bench demonstrate that this combined strategy effectively improves robustness and maintains estimation accuracy under changing load conditions.

Link: https://ieeexplore.ieee.org/document/11625296