Neural Luenberger state observer for nonautonomous nonlinear systems

A team of researchers including Moritz Woelk, Jarod Morris, and Wentao Tang has published a novel paper on arXiv titled 'Neural Luenberger state observer for nonautonomous nonlinear systems.' The work introduces a method to synthesize a state observer—a system that estimates internal states from external outputs—for complex, time-varying systems without requiring a pre-defined mathematical model. Instead, it leverages the structure of the Kazantzis-Kravaris/Luenberger (KKL) observer, extending it to handle systems with external inputs (nonautonomous) by adding an input-affine term. The core innovation is training two fully connected neural networks offline on historical or simulated data: one to learn this input-affine term for the observer dynamics, and another to learn the nonlinear mapping from observer states back to the true system states.

The technical approach guarantees a bounded estimation error when the trained observer is deployed on new input-output data, a critical feature for reliability in safety-sensitive applications like chemical engineering. The researchers validated their method through case studies on two classic benchmark systems in process control: a bioreactor and a Williams-Otto reactor. This demonstrates the method's potential to monitor and control complex industrial processes where first-principles models are difficult or impossible to derive. By providing a purely data-driven path to robust state estimation with performance guarantees, this work bridges advanced control theory with modern machine learning, opening doors for AI-enhanced automation in pharmaceuticals, energy, and manufacturing.