TL;DR
This paper introduces three novel Bayesian filters designed to estimate the state of nonlinear systems when sensor data is intermittently lost and—crucially—the estimator doesn't know when these losses occur. By exploiting the binary nature of measurement losses, the proposed BKF-I, BKF-II, and Rao-Blackwellised Particle Filter (RBPF) provide robust estimation that approaches the performance of "ideal" filters that have perfect knowledge of sensor failures.
What Problem Does This Paper Address?
In real-world applications like GPS navigation, target tracking, or drone control, sensor data isn't always reliable. Signals drop out due to "canyon" effects in cities, network congestion, or hardware faults.
If an estimator knows exactly when a measurement is lost, it can use the Intermittent Kalman Filter (IKF) to maintain optimal estimates. However, in many practical scenarios, the estimator receives a data packet that might be pure noise or a "fake" measurement, and it has no flag indicating the data is bad. This uncertainty turns a standard linear Gaussian problem into a complex non-Gaussian, nonlinear one. Standard tools like the Extended Kalman Filter (EKF) often diverge or degrade significantly when they unknowingly ingest "garbage" data.
Methods
The authors propose three suboptimal but practical recursive filters to bridge the gap between the simple EKF and the computationally expensive Particle Filter:
- Bayesian Kalman Filter I (BKF-I): This filter uses a Maximum A Posteriori (MAP) approach. At each step, it makes a "hard decision" on whether the measurement was lost or received based on the likelihood of the incoming data. It then applies the IKF update using this estimated binary state.
- Bayesian Kalman Filter II (BKF-II): Instead of a hard decision, this filter calculates the probability of a measurement loss. It then performs a weighted update that accounts for this uncertainty, acting as a "soft" version of the BKF-I.
- Rao-Blackwellised Particle Filter (RBPF): This is the most sophisticated approach. It uses particles to represent the distribution of the discrete loss process while analytically marginalizing out the continuous state variables. This makes the filter much more efficient than a standard Particle Filter because particles are only needed for the binary loss sequence, not the entire high-dimensional state space.
The authors also introduce a "Fast RBPF" implementation that reduces computational load by removing duplicate particles that inevitably occur in discrete state spaces.
Datasets
As is common in control theory and signal processing research, the authors evaluate their methods using simulated benchmarks rather than static datasets:
- Linear Stochastic System: A 2D system used to establish a baseline against the optimal (but "cheating") IKF.
- 2D Target Tracking: A nonlinear radar tracking simulation where a base station measures distance and angle via a lossy wireless network.
- Quadrotor Path Control: A high-fidelity Matlab/Simulink model of a quadrotor drone following a triangular trajectory, where IMU and GPS data are subject to random packet losses.
Key Insights and Ideas
- Exploiting Binary Structure: The "aha moment" is realizing that measurement loss is a discrete, binary event ($\gamma=1$ or $\gamma=0$). Rather than treating the system as a general nonlinear problem, the authors specifically target this binary switch, which allows for much more efficient filtering.
- The Power of Marginalization: The RBPF works so well because it recognizes that given a specific sequence of losses, the system remains conditionally Gaussian. This allows the filter to handle the "hard" part (the losses) with particles and the "easy" part (the state) with Kalman-style math.
- Computational Efficiency: The BKF-I and BKF-II have computational complexities nearly identical to a standard EKF, making them highly attractive for real-time embedded systems with limited CPU power.
Key Results
- Accuracy: In target tracking simulations, the RBPF consistently achieved the lowest Root-Mean-Square Error (RMSE), often outperforming even the IKF (because the IKF's linearization is less robust than the RBPF's particle-based approach in high-loss scenarios).
- Robustness: While a standard EKF diverged when measurement loss probability reached 90%, the proposed filters remained stable and tracked the target effectively.
- Control Performance: In the quadrotor experiment, the proposed filters allowed the drone to complete its path even with 20% packet loss, whereas a standard EKF-based controller failed to follow the trajectory.
- Speedup: The "Fast RBPF" optimization significantly reduced computation time when using a large number of particles, proving that the method is viable for complex systems.
Limitations and Future Work
- Initial Estimates: Like most nonlinear filters (including the EKF), these methods are sensitive to the initial state estimate. If the filter starts with a very poor guess, the linearization in the BKF variants can lead to poor performance.
- Theoretical Proofs: The authors acknowledge that while the filters perform well in simulation, formal proofs for the theoretical convergence of these suboptimal filters remain an open challenge for future research.
- Parameter Tuning: The filters assume the prior probability of loss (e.g., the Bernoulli parameter $\theta$) is known. Future work could involve estimating this loss probability online alongside the state.
Summary updated