State estimations and noise identifications with intermittent corrupted observations via Bayesian variational inference
New AI filter tackles missing data and corrupted signals simultaneously, achieving asymptotic optimality.
A team of researchers has introduced a new AI-powered filtering method designed to bring order to chaotic sensor networks. In a paper titled "State estimations and noise identifications with intermittent corrupted observations via Bayesian variational inference," Peng Sun, Ruoyu Wang, and Xue Luo address a critical real-world problem: how to accurately track a system's state when sensor data is unreliable. This unreliability comes from intermittent packet dropouts, corrupted observations (outliers), and unknown noise characteristics, all common issues in distributed networks like those used in autonomous vehicles or industrial IoT. Their solution, the Variational Bayesian Adaptive Kalman Filter (VB-AKF), frames this as a joint Bayesian inference problem, simultaneously estimating the hidden system state, unknown noise parameters, and the network's reliability.
Unlike prior adaptive filters that handle missing data and outliers separately, the VB-AKF's innovation is its dual-mask generative model. This model uses two independent Bernoulli random variables to explicitly characterize both observable communication losses and the latent authenticity of each data point. Furthermore, it integrates multiple concurrent observations into its framework, significantly boosting statistical identifiability. Comprehensive numerical experiments validate the VB-AKF's effectiveness, demonstrating that both its parameter identification and state estimation performance asymptotically converge to the theoretical optimal lower bound as the number of sensors in the network increases. This represents a significant step toward more robust and trustworthy data fusion in complex, noisy environments.
- Proposes a Variational Bayesian Adaptive Kalman Filter (VB-AKF) for joint state and noise estimation.
- Uses a novel dual-mask model with Bernoulli variables to handle packet loss and data corruption simultaneously.
- Achieves asymptotic optimality, with estimates converging to the theoretical lower bound as sensor count grows.
Why It Matters
Enables more reliable autonomous systems and industrial monitoring by making sense of flawed sensor data.