Physics-Informed ensembles of dual extended kalman filters for online state and parameter estimation of mechatronic systems

Kalman filters are widely used for real-time sensor data processing in mechatronic and production systems, but their effectiveness depends on accurate system models. These models also need updating in order to handle changing or unknown operating conditions. Data-driven methods excel in adapting models to match measurements without heavy reliance on expert knowledge, but they suffer from drawbacks such as reduced interpretability, limited extrapolatability, and poor reliability for long-term predictions. Hybrid methods that integrate traditional and data-driven modeling aim to combine the strengths of both approaches, yet achieving accurate convergence between the distinct physical and data-driven components remains a significant challenge during joint optimization. This paper introduces a novel methodology based on three key ideas. First, a dual extended Kalman filter is used to jointly estimate and adapt model parameters in real-time. Second, a hybrid model combining physics-based and data-driven parameters is embedded in the filter, resulting in a more generalizable and accurate model. Finally, an extension to ensembling techniques is proposed to improve convergence toward correct physical parameters. The proposed ensemble of hybrid dual extended Kalman filters is evaluated on a cam follower system working under a wide range of operating conditions. Results demonstrate its ability to learn both the cam shape and operating voltage during operation and converged models show a reduced error in physical parameter estimates of more than 60% compared to alternative filtering techniques. Moreover, these models outperform other methods in multi-step prediction, showcasing an improvement of more than 48% in prediction accuracy. With this work, we aim to demonstrate the efficacy of physics-informed models in a filtering context and address the challenge of correct parameter learning despite the overparameterization inherent in these models. • Joint parameter and state estimation is achieved using dual kalman filtering. • Ensemble methods improved parameter convergence, avoiding local minima. • Experimental validation on a cam follower setup is performed. • The hybrid model improved parameter identification accuracy by over 60%. • 1000-step-ahead prediction shows more than 48% improvement.