Lessons From A Sensor Network ExpeditionRobert Szewczyk, Joseph Polastre, Alan Mainwaring, David CullerLessons From A Sensor Network ExpeditionRobert Szewczyk1, Joseph Polastre1, Alan Mainwaring2, and David Culler1,21University of California, BerkeleyBerkeley CA 94720{szewczyk,polastre,culler}@cs.berkeley.edu2Intel Research Lab oratory at BerkeleyBerkeley CA 94704{amm,dculler}@intel-research.netAbstract. Habitat monitoring is an important driving application forwireless sensor networks (WSNs). Although researchers anticipate somechallenges arising in the real-world deployments of sensor networks, anumber of problems can be discovered only through experience. Thispap e r evaluates a sensor network system described in an earlier workand presents a set of experiences from a four month long deploymenton a remote island off the coast of Maine. We present an in-depth anal-ysis of the environmental and node health data. The close integrationof WSNs with their environment provides biological data at densitiesprevious impossible; howe ver, we show that the sensor data is also use-ful for predicting sys tem operation and network failures. Based on overone million data and health readings, we analyze the node and networkdesign and develop network reliability profiles and failure models.1 IntroductionApplication-driven research has been the foundation of e xcellent science con-tributions from the computer science community. This research philosophy isessential for the wireless sensor network (WSN) community. Integrated closelywith the physical environment, WSN functionality is affected by m any environ-mental factors not foreseen by developers nor detectable by simulators. WSNsare much more exposed to the environment than the traditional systems. Thisallows WSNs to densely gather environment data. Researchers can study the re-lationships between collected environmental data and sensor network behavior.If a particular aspect of sensor functionality is correlated to a set of environmen-tal conditions, network designers can optimize the behavior of the network toexploit a beneficial relationship or mitigate a detrimental one.Habitat monitoring is widely accepted as a driving application for wirelesssensor network research. Many sensor network services are useful for habitatmonitoring: localization [1], tracking [3,14,16], data aggregation [9,15,17], and, ofcourse, energy efficient multihop routing [5,13,25]. Ultimately the data collectedneeds to be meaningful to disciplinary scientists, so sensor design [19] and in-the-field calibration systems are crucial [2,24]. Since such applications need torun unattended, diagnostic and monitoring tools are essential [26].2 Szewczyk et al.While these services are an active area of research, few studies have beendone using wireless sensor networks in long-term field applications. During thesummer of 2002, we deployed an outdoor habitat monitoring application that ranunattended for four months. Outdoor applications present an additional set ofchallenges not seen in indoor experiments. While we made many simplifying as-sumptions and engineered out the need for many complex services, we were ableto collect a large set of environmental and node diagnostic data. Even thoughthe collected data was not useful for making scientific conclusions, the fidelityof the sensor data yields important observations about sensor network behav-ior. The data analysis discussed in this paper yields many insights applicable tomost wireless sensor deployments. We utilize traditional quality of service met-rics such as packet loss; however the sensor data combined with network metricsprovide a deeper understanding of failure modes. We anticipate that with systemevolution comes higher fidelity sensor readings that will give researchers an evenbetter understanding of s ensor network behavior.The paper is organized as follows: Section 2 provides a detailed overview ofthe application, Section 3 analyzes the network behaviors that can be deducedfrom the sensor data, Section 4 c ontains the analysis of the node-level data.Section 5 contains related work and Section 6 concludes.2 Application overviewIn the summer of 2002, we deployed a 43 node sensor network for habitat mon-itoring on an uninhabited island 15km off the coast of Maine, USA. Biologistshave seasonal field studies with an emphasis on the ecology of the Leach’s StormPetrel [12]. The ability to densely instrument this habitat with sensor networksrepresents a significant advancement over traditional instrumentation. Monitor-ing habitats at scale of the organism was previously impossible using standalonedata loggers. To assist the biologists, we developed a complete sensor networksystem, deployed it on the island and monitored its operation for over fourmonths. We used this case study to deepen our understanding of the researchand engineering challenges facing the system designers, while providing data thathas been previously unavailable to the biologists. The biological background, thekey questions of interest, and core system requirements are discussed in depthin [18,20].In order to study the Leach’s Storm Petrel’s nesting habits, nodes were de-ployed in underground nesting burrows and outside burrow entrances aboveground. Nodes monitor typical weather data including humidity, pressure, tem-perature, and ambient light level. Nodes in burrows also monitored infrared ra-diation to detect the presence of a petrel. The specific components of the WSNand network architecture are described in this section.2.1 Application SoftwareOur approach was to simplify the problem wherever possible, to minimize engi-neering and development efforts, to leverage existing sensor network platformsLessons From A Sensor Network Expedition 3and components, and to use off-the-shelf products. Our attention was focusedon the sensor network operation. We used the Mica mote developed by UCBerkeley [10] running the TinyOS ope rating system [11].In order to analyze the long term operation of a WSN, each node executed asimple periodic application that met the biologists requirements in [20]. Every70 seconds, each node sampled each of its sensors. Data readings were time-stamped with 32-bit sequence numbers kept in flash memory. The readings weretransmitted in a single 36 byte data packet using the TinyOS radio stack. We re-lied on the underlying carrier sense MAC layer protoc ol to prevent against packetcollisions. After successfully transmitting the