Faults in Sensory Readings: Classification and Model Learning

Faults in wireless sensor networks are a common occurrence and their accumulation can have a significant negative influence on the reliability of the network. Accuracy of sensory readings decreases over time. We focus on detection of faults as they can be observed in sensory readings. Trace that faults leave in data can be used for classification of faults independently of the underlying cause. We propose a complete and consistent fault classification based on two aspects. The first aspect is continuity and frequency of the occurrence, and the second is the existence of observable and learnable patterns. The network can learn a model of a fault from the past behavior and patterns visible in the data. We rely on centralized and straightforward detection methods using neighborhood vote and time series analysis. For the full classification phase, we propose the use of statistical pattern recognition, specifically, decision trees and regression. Current results show that this method works comparatively well when applied to dense data-centric wireless sensor network.