Information-theoretic treatment of sensor measurements in network systems

In a distributed network system, data collection devices (e.g., sensors) may operate on fuzzy inputs, thereby generating results that possibly deviate from the reference datum in physical world being sensed. The extent of deviation and the time it takes to compute an output result (i.e., inaccuracy and timeliness of event notification) depend on the number of orthogonal information elements, i.e., modes, processed from the sensed inputs. A major issue is the large dimensionality of input data and the resource-constrained system components (i.e., limited amount of processing cycles and network bandwidths). So, there is a tradeoff between the resources expended by a device algorithm to process its input data and the timeliness and accuracy of its output result. Exercising this tradeoff requires a layered construction of sensor algorithms, where each layer processes a subset of modes in the input data and the results are fused to generate a composite output event. The paper provides an information-theoretic model of such layered algorithm designs. The goal is to evaluate the tradeoff between the quality of event detection and the processing/network resources expended, so that the device algorithms can adapt their operations based on resource availability. The paper provides a case study of network topology measurements to corroborate our model.