Infrastructure Monitoring With Failure-Possible Sensor Networks

There is an every growing interest and need to deploy sensors for monitoring our various infrastructure systems for the purposes of condition monitoring, operational control and security. The subject of this thesis is the assessment of the reliability of infrastructure management systems which use sensor networks, specifically an infrastructure management system that does long-term monitoring of an infrastructure.;In order to pervasively deploy a large network of sensor nodes and analyze the anomalous behaviour, we have developed the Critter Dataport Sensor. The Critter is an analog computer-attached sensor device developed at Carnegie Mellon University as part of this research. It is an economical (and thereby pervasively deployable) alternative to all-in-one sensor devices, which include sensing, processing, memory, and networking. The Critter sensor device was developed as a means to research pervasive infrastructure sensor networks, as commercial sensors did not provide raw unmodified sensor data. Commercial sensors have been too cost-prohibitive to deploy en-masse. Four sensor data sets were collected by deploying Critter sensors through one wing of an academic building: (1) high fidelity sensing in which data is collected 10x second, (2) pervasive sensing in which many sensors are placed in different offices, (3) dense sensing in which many sensors were placed into a single room, and (4) longitudinal sensing in which data was collected for over five years.;One theme that is explored is the prediction of a sensor failure. The first possible approach is to use anomalous sensor readings as a predictor. The anomalies did not predict a failure, but did uncover a pattern that linked their existence to time of day, showing a secondary pattern in the data. The second approach to prediction of failure is based on lifetime, showing that sensor applications are prone to the same lifetime effect (especially infant mortality) as computer systems, infrastructure systems, and HVAC systems.;The next theme that has been explored is the mitigation of the effects of the missing data. For data fidelity problems in which sensor data is not consistently collected at the requested rate, the use of an average can overcome missing sensor data, however it is tempered by parameter choices that will affect how well the average can represent the data selection. For lifetime problems, changes in the service behaviour, such as when to schedule updates and repairs, can greatly impact the length and frequency of downtime in which the sensor is not collecting data.;Finally, an approach is presented for using in-exact replicas as a means for filling in missing data. When using sensor data streams for long-term infrastructure monitoring it is not possible to transparently substitute one sensor for another. Using data imputation techniques, it is possible to impute missing data based on a non-linear relationship between the sensor streams.;This thesis seeks solutions to improving the performance of a sensor network used for infrastructure monitoring, as a means to mitigate the potential problem of trading one management problem (monitoring the infrastructure) with another one (monitoring the sensor network).