Researchers propose GBF for better state estimation

A team of researchers led by Lei Cao from Tsinghua University and collaborators introduced Generative Bayesian Filtering (GBF), a novel framework designed to address the limitations of classical state estimation methods like Kalman filters. Traditional approaches struggle with the nonlinear and heterogeneous patterns in high-dimensional sensor data, often relying on oversimplified observation models. GBF replaces these restrictive models with pretrained conditional generative models, specifically parametrized by conditional variational autoencoders (CVAE), enabling more flexible and accurate state inference.

GBF performs online inference by executing a Bayesian prediction-update recursion, where the measurement update is framed as a posterior sampling problem. This combines a dynamical prior with a CVAE-induced likelihood, transforming the filtering problem into a score-based sampling problem. This approach leverages the flexibility of generative models and the uncertainty quantification capabilities of ensembling. Experiments on synthetic datasets and real-world applications—such as manufacturing system monitoring and arrhythmia diagnosis—demonstrate that GBF significantly improves state estimation accuracy and robustness compared to baseline methods, including traditional Kalman filters.