A multistage approach to decision fusion using a distributed network of non-identical nodes

A distributed sensor network deployed to detect a binary event using a hard decision fusion scheme is considered. Earlier works show that the K-out-of-N counting rule is optimum in minimizing the total error rate when the sensor nodes participating in the event detection have identical performance indexes (detection and false-alarm probability pairs). In most scenarios, this optimum rule turns out to be the half-voting or majority rule. However, when the sensor nodes have non-identical performance indexes, the optimum value of K that gives the maximum correct decision probability may be anything from 1 to N and no general expression for it is available. Trying all possibilities before choosing the best value of K is computationally very demanding and the complexity grows rapidly as the number of sensor nodes increases, making it an impractical approach for a large sensor network. We propose a multistage decision fusion scheme for such networks using unanimity fusion rules and show that it gives an improved performance over unanimity, half-voting, majority, and some soft fusion rules and a comparable performance to the best K-out-of-N fusion rule in many cases with considerably reduced computational complexity in determining the global detection and false-alarm probabilities.