Abstract
The set-membership information fusion problem is studied for general multi-sensor dynamic systems. Based on set-membership theory, three centralized state fusion estimation algorithms in the presence of bounded disturbances are proposed, namely augmented algorithm, combined measurement filtering algorithm and pseudo-sequential filtering algorithm. Theoretical discussions on the convergence and boundedness of the proposed fusion algorithms are provided and their stability is proved. The estimate accuracy, computational complexity and flexibility of these three fusion algorithms are compared through theoretical analysis and simulation. And their exchanging property of measurement update order is discussed. Results show that these algorithms are functionally equivalent in terms of the estimation accuracy and the exchangeability of the measurement update order can be guaranteed as long as the parameters satisfy certain conditions. Meanwhile the simulation results prove the role of the proposed algorithms in improving state estimation accuracy. In addition, the combined measurement filtering algorithm has the highest calculation speed due to lower dimension. But it is less flexible because the sensor measurement matrices need to satisfy some additional conditions. These conclusions are valuable in applications.
Introduction
Most traditional data fusion techniques, such as the ubiquitous Kalman Filter, rely on probabilistic models that assume known statistical properties (like mean and variance) of process and measurement noise. In many real-world engineering scenarios—such as target tracking or navigation in complex environments—these statistics are often impossible to determine accurately. However, physical limits or "bounds" of the noise are often known.
This is where Set-Membership (SM) estimation comes in. Instead of a point estimate with a probability cloud, SM estimation produces an ellipsoid that is guaranteed to contain the true state. This paper addresses how to efficiently fuse these ellipsoidal sets from multiple sensors in a centralized system to achieve the most accurate "bounding box" possible.
Proposed Methods
The authors propose and compare three centralized SM fusion architectures:
- Augmented Algorithm: Stacks all sensor measurements into one high-dimensional vector and processes them in a single filtering step.
- Combined Measurement Filtering Algorithm: Uses a weighted average to merge multiple sensor measurements into a single "fused" measurement before filtering.
- Pseudo-sequential Filtering Algorithm: Processes measurements from each sensor one by one, sequentially updating the state estimate.
While these architectures exist for probabilistic systems, this paper derives them specifically for set-membership theory using ellipsoidal bounding. Crucially, the authors provide a formal proof of Input-to-State Stability (ISS), ensuring the estimation error remains bounded even in the presence of disturbances. They also establish the mathematical conditions under which these three seemingly different methods yield identical results.
Experimental Evaluation
The algorithms were evaluated using a synthetic target tracking simulation involving a moving object and three sensors. Sensors 1 and 2 provided 2D position data with uniform noise distributed within ellipsoids, while Sensor 3 provided 1D acceleration data with noise bounded within a specific interval. This setup was specifically designed to show that even if one sensor (Sensor 3) is "divergent" on its own, the fusion algorithms can still achieve a stable, convergent estimate by leveraging the other sensors.
Key Results
- Accuracy: All three fusion methods significantly outperformed single-sensor estimation. In the simulation, the average Mean Square Error (MSE) for the fused state was drastically lower than any individual sensor's performance.
- Computational Speed:
- Combined Measurement: ~3.32 ms per step (Fastest).
- Augmented: ~3.81 ms per step.
- Pseudo-sequential: ~5.70 ms per step (Slowest).
- Convergence: The fusion algorithms successfully "bounded" the state even when the initial error was large, proving the practical utility of the stability theorems.
Discussion and Insights
The most significant finding is the functional equivalence of the three methods regarding estimation accuracy, provided parameters are chosen correctly. This allows developers to select an algorithm based on hardware constraints (like memory or speed) without worrying about losing precision.
However, there are trade-offs. The Combined Measurement method is the fastest because it reduces the dimensionality of the problem before filtering, but it is the least flexible because it requires the sensor measurement matrices to satisfy specific positive-definite conditions. In the Pseudo-sequential method, the order in which sensors are processed can change the result if parameters are optimized at each step, highlighting a hidden complexity in sequential processing.
Conclusion
This research demonstrates that the fusion process effectively "rescues" poor-quality sensors, maintaining stability even when individual sensor-pair combinations are not uniformly observable. The proposed algorithms offer a robust and flexible toolkit for state estimation in unpredictable environments. Future work will address the optimal update order for sequential methods and the computational scaling of the Augmented method in massive sensor networks.
Reference: Shen, Q., et al. "Centralized Fusion Methods for Multi-Sensor System With Bounded Disturbances." IEEE Access, vol. 7, 2019. DOI: 10.1109/ACCESS.2019.2943163.
Summary updated