Information fusion strategies from distributed filters in packet-drop networks

In this paper we study different distributed estimation schemes for stochastic discrete time linear systems where the communication between the sensors and the estimation center is subject to random packet loss. Sensors are provided with computational and memory resources so that they can potentially perform data processing of the measurements before sending their information. In particular, we explore three different strategies. The first, named measurement fusion (MF), optimally fuses the raw measurements received so far from all sensors. The second strategy, named optimal partial estimate fusion (OPEF), optimally fuses at the central node the last local state estimates received from each sensor. The last strategy, named open loop partial estimate fusion (OLPEF), simply sums local state estimates received from each sensor and replace the lost ones with their open loop prediction. We provide some analytical results about the performance of these three schemes in special regimes conditions, namely low and high process noise. We also show through numerical simulations that, although none of the three schemes achieves the ideal performance of a scheme with infinite bandwidth communication between sensors and the central node, the OPEF scheme provides almost ideal performance.