Generative Bayesian Filtering for State Estimation

Short summary

The authors propose Generative Bayesian Filtering (GBF), which replaces restrictive linear-Gaussian observation models in classical filters with pretrained conditional VAEs. GBF formulates the measurement update as a posterior sampling problem combining dynamical priors with CVAE-induced likelihoods, transformed into a score-based sampling problem. Experiments on synthetic data, manufacturing monitoring, and arrhythmia diagnosis show improved state estimation accuracy and robustness over baseline approaches.

• •GBF replaces classical linear-Gaussian observation models with conditional VAEs for flexible likelihood estimation
• •Measurement update is cast as score-based posterior sampling combining dynamical prior with generative likelihood
• •Validated on manufacturing monitoring and arrhythmia diagnosis with improved accuracy over baselines

Generated with AI, which can make mistakes.