Scalar quantizers for decentralized estimation of multiple random sources

We consider a new inference-centric application for a distributed sensor network. Consider multiple signal sources, i.e., acoustic sources, in the field being monitored. Governed by physics law each sensorpsilas measurement can be modeled as a linear combination of the original multiple signal sources, corrupted by the additive measurement noise. In a non-cooperative communication scenario, each sensor transmits (a summary of) its own observations to the fusion center (FC), that is interested in reconstruction of all the original signal sources. We show that the inherent correlation among sensorspsila measurements, which is due to their spatial proximity, can be effectively exploited to compress sensorspsila data and reduce the transmission rate, without necessarily comprising the inference performance, i.e., quality of the reconstructed sources at the FC. In particular, we propose a practically simple and yet effective encoding algorithm for sensors, built on the concept of distributed source coding, two data reconstruction schemes for the FC (referred to as pairing and sequential schemes), and two corresponding rate allocation policies. The proposed compression algorithm is developed based on the side information coding technique in. We further investigate the trade off between rate and source reconstruction quality for the proposed compression/reconstruction schemes and verify their effectiveness via simulations.