State estimations and noise identifications with intermittent corrupted observations via Bayesian variational inference
| Authors | Peng Sun et al. |
| Year | 2026 |
| Field | Statistics / ML |
| arXiv | 2604.02738 |
| Download | |
| Categories | stat.ML, cs.LG, stat.CO |
Abstract
This paper focuses on the state estimation problem in distributed sensor networks, where intermittent packet dropouts, corrupted observations, and unknown noise covariances coexist. To tackle this challenge, we formulate the joint estimation of system states, noise parameters, and network reliability as a Bayesian variational inference problem, and propose a novel variational Bayesian adaptive Kalman filter (VB-AKF) to approximate the joint posterior probability densities of the latent parameters. Unlike existing AKF that separately handle missing data and measurement outliers, the proposed VB-AKF adopts a dual-mask generative model with two independent Bernoulli random variables, explicitly characterizing both observable communication losses and latent data authenticity. Additionally, the VB-AKF integrates multiple concurrent multiple observations into the adaptive filtering framework, which significantly enhances statistical identifiability. Comprehensive numerical experiments verify the effectiveness and asymptotic optimality of the proposed method, showing that both parameter identification and state estimation asymptotically converge to the theoretical optimal lower bound with the increase in the number of sensors.
Engineering Breakdown
Plain English
This paper addresses a real-world challenge in distributed sensor networks where data is unreliable: packets get dropped during transmission, sensor measurements contain outliers or corruption, and you don't know how much noise to expect from the environment. The authors propose VB-AKF (variational Bayesian adaptive Kalman filter), a new algorithm that simultaneously estimates three things — the true system state, the noise characteristics, and the network reliability — all within a unified Bayesian framework. The key innovation is a dual-mask approach using two separate Bernoulli random variables to explicitly model both communication losses and corrupted sensor readings, rather than treating them as separate problems like existing methods do.
Core Technical Contribution
The core novelty is the dual-mask generative model that decouples packet dropouts from measurement corruption. Traditional adaptive Kalman filters handle missing data or outliers separately, which creates blind spots — you can't distinguish between a lost packet and a genuinely bad measurement. VB-AKF treats these as independent latent variables: one Bernoulli mask captures whether the packet arrived, another captures whether the measurement is trustworthy given that it arrived. The authors then formulate joint posterior inference over system states and noise parameters using variational Bayesian approximation, enabling tractable inference without requiring manual noise covariance tuning or separate outlier detection pipelines.
How It Works
The system treats the observed sensor readings as a mixture of two hidden problems: communication reliability and measurement authenticity. For each timestep, two Bernoulli random variables indicate (1) did the packet arrive from the sensor to the fusion center, and (2) is the measurement free of corruption/outliers. The algorithm maintains variational distributions over the system state, noise covariance matrices (process and measurement noise), and the probability parameters of both Bernoulli masks. The inference loop updates these distributions iteratively: given current beliefs about noise levels and network reliability, it estimates the true state; given the state estimate, it refines noise parameter estimates and reliability assessments. The Kalman filter update equations incorporate uncertainty from both missing-data and corruption channels, with adaptive gain matrices that automatically reduce influence from likely-corrupted or missing observations.
Production Impact
In production sensor fusion systems (autonomous vehicles, industrial IoT, environmental monitoring), this directly improves robustness without requiring manual hyperparameter tuning or separate anomaly detection modules. Instead of engineers specifying noise covariances offline and building separate outlier filters, VB-AKF automatically learns both from data, adapting as network conditions degrade or sensor health changes. The trade-off is computational cost: variational inference requires iterative updates of multiple posterior distributions per timestep, making this slower than standard Kalman filters by a factor that depends on iteration count and state dimensionality — this is acceptable for systems with millisecond-scale latency budgets (most sensor fusion) but not microsecond-critical applications. Integration complexity is moderate: you need to specify prior distributions over noise parameters and Bernoulli probabilities, and tune the VB convergence criteria, but the payoff is significant robustness when real-world packet loss (WiFi: 5-20%, cellular: 1-10%, wired industrial: <1%) and sensor degradation occur simultaneously.
Limitations and When Not to Use This
The paper assumes noise follows Gaussian distributions with unknown but time-invariant covariances — non-Gaussian heavy-tailed noise (sensor saturation, multipath effects) would break this assumption and likely hurt performance. The dual-mask model assumes independence between packet loss and measurement corruption, which may not hold if network congestion causes both simultaneously or if sensor failures are network-correlated. Scalability to high-dimensional state spaces (>1000 dimensions) is unclear since the method requires maintaining and updating full covariance matrices; the paper likely demonstrates results on systems with <100 state dimensions. The approach also requires enough clean training data to initialize the noise covariance estimates — in severely degraded environments with >50% packet loss and frequent outliers, the EM-like inference loop might converge to poor local optima.
Research Context
This work extends classical adaptive Kalman filtering (AKF literature dating to the 1970s) by incorporating Bayesian model selection and variational inference techniques popular in modern probabilistic ML. It sits between two research communities: the controls/signal-processing world focused on state estimation, and the Bayesian inference community focused on scalable posterior approximation. The dual-mask generative model is inspired by mixture models and missing-data frameworks in statistics, but applies them systematically to the coupled problems of network reliability and measurement quality. This likely opens directions toward hierarchical Bayesian sensor fusion (modeling entire networks rather than single fusion centers) and federated variants where edge nodes run local VB inference and exchange sufficient statistics.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::