Target Learning for Wireless Sensor NetworksMotivation and Problem FormulationKey IdeasCurrent Status and Future PlansTarget Learning for Wireless Sensor NetworksTarget Learning for Wireless Sensor NetworksPrasanth JeevanPrasanth JeevanMotivation and Problem FormulationMotivation and Problem FormulationDetection and classification methods application specificSignificant human-in- t he-loop componentLearning aspect is done by the userRequires gathering lots of data for offline processingSemi-supervised learning at the node-level to learn target signatures for detection and classificationsubset of person, person with ferrous object, vehicleDetection and classification methods application specificSignificant human-in- t he-loop componentLearning aspect is done by the userRequires gathering lots of data for offline processingSemi-supervised learning at the node-level to learn target signatures for detection and classificationsubset of person, person with ferrous object, vehicleKey IdeasKey IdeasNode-level capabilities are sufficient to detect and classify targetsMultiple sensors on each moteAggregation to higher levels will improve confidenceExpectation Maximization algorithmEach node will develop models of the different targets and “no-target”Adaptation to changing environmental conditions will pose a significant challengeNode-level capabilities are sufficient to detect and classify targetsMultiple sensors on each moteAggregation to higher levels will improve confidenceExpectation Maximization algorithmEach node will develop models of the different targets and “no-target”Adaptation to changing environmental conditions will pose a significant challengeCurrent Status and Future PlansCurrent Status and Future PlansCurrentReal-world data from XSM/Trio motesNeed to process the data into useful features for classificationImplementation of EM (Matlab)Must adapt to current problem and make robust to real-world dataFutureBring learning in-network (on-line), adaptiveExpand from node-level and exploit correlation Learn other aspects of detection/classification such as how to automatically manipulate data in the most effective way to bring out features for classificationCurrentReal-world data from XSM/Trio motesNeed to process the data into useful features for classificationImplementation of EM (Matlab)Must adapt to current problem and make robust to real-world dataFutureBring learning in-network (on-line), adaptiveExpand from node-level and exploit correlation Learn other aspects of detection/classification such as how to automatically manipulate data in the most effective way to bring out features for
View Full Document