Abstract
We consider a state estimation problem where observations are made by multiple sensors. These observations are communicated over a lossy wireless network to a central base station that computes estimates via a Kalman filter. The goal is to determine the optimal location of the base station under a certain class of packet loss probability models. It is shown in the two sensor case that the base station is optimally located at one of the sensor locations. Empirical evidence suggests that the result holds in some generality.
1. Introduction
Wireless sensor networks (WSNs) offer significant advantages for monitoring dynamic systems in environments where cabling is impractical. However, the wireless medium is inherently unreliable, leading to packet loss that can severely impact the performance of state estimation. This paper addresses a fundamental spatial optimization problem: given a set of sensors at fixed locations, where should the central base station be placed to minimize the expected estimation error?
The quality of the Kalman filter estimates at the base station depends on the arrival of measurement packets. Since the probability of successful packet delivery typically decreases as the distance between the sensor and the base station increases, the base station's location is a critical factor in system performance.
2. Mathematical Framework and Methodology
The system state is estimated using a standard Kalman filter. The researchers model the packet arrival process as independent Bernoulli trials. For a sensor at distance $d$ from the base station, the probability of a packet arriving is $f(d)$, where $f$ is assumed to be a convex, decreasing function.
In the fundamental case of two sensors located on a line of unit length, the base station is positioned at distance $d$ from the first sensor and $1-d$ from the second. The possible sets of measurements received at any time step $k+1$, denoted by $P_{k+1}$, and their associated probabilities are summarized in the following table:
The objective is to minimize the expected error covariance matrix. The authors demonstrate that this cost function is matrix concave with respect to the base station's position $d$. In optimization theory, a concave function over a closed interval reaches its minimum at the boundaries of that interval.
3. Key Results and Analysis
The primary theoretical contribution of this work is the proof that for two sensors in a linear arrangement, the optimal base station location is not in the center, but rather co-located with one of the sensors. By eliminating packet loss for one sensor (setting $d=0$ or $d=1$), the system gains a perfectly reliable data source that provides a better overall estimate than two unreliable sources.
3.1 Noise and Multi-Sensor Extensions
The research further explores variations of this setup:
- Sensor Heterogeneity: If sensors have different measurement noise levels, the base station should be placed at the location of the more accurate sensor to maximize the reliability of the highest-quality data.
- Linear Multi-Sensor Arrays: For $N$ sensors arranged colinearly, empirical results indicate that the error function is piecewise concave. This suggests that the optimal location will always coincide with one of the $N$ sensor positions.
- Dimensionality Effects: The "boundary-only" rule is specific to 1D topologies. In 2D setups, such as sensors at the corners of an equilateral triangle, the researchers found that the optimal location can reside in the interior, such as at the centroid.
4. Discussion and Conclusion
This study provides a rigorous mathematical basis for base station placement in 1D wireless sensor networks. The counter-intuitive finding that the optimum lies at the sensors rather than the geometric center highlights the importance of modeling packet loss non-linearly.
Future work may address more complex scenarios, including sensors with local processing capabilities (smart sensors) or decentralized estimation architectures. Additionally, developing a generalized theory for optimal placement in arbitrary 2D and 3D geometries remains an important open challenge for the field.
Summary updated