Event-detection in sensor fields by adaptive distributed computations

Consider a sensor-field deployed to detect intrusions into a security zone, with each sensor monitoring more than one site. We develop an efficient inference algorithm for determining, from the sensor readings, the most probable set of locations where intrusions have occurred. With N potential points of intrusion, there are 2 ^{* *}N subsets of points where intrusions could occur, and searching by direct enumeration is not a scalable method. We present an adaptive distributed algorithm that drastically reduces the computational effort by several orders of magnitude by partitioning the problem into virtual `computational domains´, where the domains are adapted to the actual sensor readings. The algorithm either finds a provably optimal global solution, or provides a bound on the deviation of the solution from optimality. Its effectiveness is demonstrated on two examples that model several hundred locations being monitored for intrusion, with the possibility of multiple simultaneous breaches. In effect, our distributed approach makes it feasible to solve much larger intrusion-detection problems than can be solved by a centralized algorithm.