Adaptive State Estimation and Continual Learning under Data Distribution Shift

In state estimation, purely model-based and datadriven observers each have inherent limitations. In modelbased observers, finding an accurate mathematical model for complex dynamic systems is challenging. In contrast, data-driven methods often achieve superior accuracy under known conditions but exhibit poor generalization to unseen data distributions, particularly in dynamic and evolving environments where system behavior changes unpredictably. To overcome these challenges, we propose an adaptive hybrid estimator that dynamically switches between a model-based observer and a neural-network predictor. Additionally, it continually updates the neural network based observer parameters through continual learning as new data are observed, enabling robust adaptation to novel operating regions. To detect the distribution change in the input space for each individual data point, a memory selection strategy is used to store the most informative data points from historical data. The proposed approach is validated using experimental data from an electric Equinox vehicle under diverse driving scenarios. Results demonstrate that, compared to purely model-based or data-driven methods, the hybrid estimator significantly reduces estimation errors, thereby enhancing both accuracy and adaptability.