Location-Aware Sensing in Smart Networks João Pedro Gomes, Pinar Oguz Ekim, João Xavier, Paulo Oliveira Institute for Systems and Robotics, LA Instituto Superior Técnico

Outline Localization in wireless sensor networks (WSN) Implications for smart energy management Technologies for localization Localization algorithms  Blind initialization  Outlier rejection Numerical results

Localization and tracking in sensor networks Geolocation is key in WSN applications  Data are often meaningful only when georeferenced  Locations may be the main variables of interest  Spatial information used in SN middleware Preview  Key features of our algorithms: Robust convergence Robust to outlier measurements  Advantages: Avoid node calibration Simple measurement HW

WSNs and smart energy management Why care about WSNs for energy management? Smart energy policies need to know where “customers” are, and how they behave  Focus thermal/lighting resources in occupied areas  Switch off unnecessary appliances in unused areas  Predict future energy needs from known patterns of behavior

Example: User tracking in an office

Technologies for localization GPS  Unusable in indoor applications RFID  Cheap tags, expensive readers  Coarse resolution RSS (Received Signal Strength)  Complex relation to distance  Build RSS tables, then index

Range-based localization TDOA (Time-difference of arrival)  Linearly related to distance  Accurate (with line-of-sight)  Limited range (~10m)  Extra hardware needed TXRX EM pulse Acoustic pulse TT  T’

Range-based source localization Ping! (x,y)

An experiment on range-based localization

Challenges in range-based localization Can we avoid calibration? (anchors) How to deal with outlier measurements? Assumption: Centralized processing ? ?

Simultaneous localization and tracking (SLAT) Collect target sightings, regenerate target and sensor positions ? ? (x1,y1) (x5,y5) (x2,y2) (x3,y3) (x4,y4)(x6,y6)

Simultaneous localization and tracking Our approach to SLAT  Initialization 1.EDM completion 2.ML refinement  Recursive updating 1.Source localization 2.ML refinement Non-bayesian approach Target dynamics ignored ? ?

Key features of our SLAT approach Use convex relaxations for blind position estimation  EDM completion for initialization CompleteFactorize

Key features of our SLAT approach Use convex relaxations (cont.)  SDP relaxation for single-source localization (SLCP) Majorization/minimization for ML optimization y x1x1 x2x2 x3x3

Numerical results: Gaussian noise

Handling outliers Model outliers as Laplacian noise Extend all the subproblems to the new noise model

Numerical results: Laplacian noise