Kulau, UlfUlfKulauBüsching, FelixFelixBüschingWolf, LarsLarsWolf2021-11-112021-11-112016-04-15ACM Transactions on Sensor Networks 12 (2): 11 1-38 (2016-04)http://hdl.handle.net/11420/10893The energy consumption of Wireless Sensor Networks (WSNs) correlates with the voltage level at which the nodes are powered-a lowered voltage leads to a prolonged lifetime of nodes and networks. Operating nodes at voltage levels below the recommendation-so-called undervolting-saves energy but is accompanied by an increased risk of failures. In this article, we show that a WSN can still work reliably, even if the voltage recommendations are violated. We show that there is a correlation between temperature and error-proneness at the same voltage level and that ideal voltage levels depend on environmental conditions. Especially in outdoor deployments, temperatures fluctuate often and heavily. Additionally, we show that these ideal voltage levels are different for each individual node. To deal with these individual variations and varying temperatures, we present a supervised learning strategy that is able to keep the nodes in uncritical states even if environmental conditions are constantly and heavily changing while saving as much energy as possible by constantly adapting the voltage level of these nodes to an individually ideal level. All measurements were performed on sensor node prototypes that are also presented in this article. In detailed evaluations, it is shown (i) that a single node will never run in the same unsafe state twice, (ii) that only five known bias points are needed to predict nearly all temperature dependencies for an individual node, and, thus, (iii) that a system of undervolted nodes can be as reliable as a conventionally powered network while prolonging the lifetime by more than 40%.en1550-485920162138ACMEnergy efficiencyRobustnessSupervised learningUndervoltingVoltage scalingWireless sensor networksInformatikJournal Article10.1145/2885500Other