2012 Inaugural Retreat MIT Center for Wireless Networks and Mobile Computing Efficient and Reliable Ultra Low Power RFID Networks Jue Wang, Haitham Hassanieh and Dina Katabi Reader shines an RF signal on nearby RFIDs Tag reflects the reader’s signal using ON‐OFF keying Motivation • Wireless protocols require power and computation • RFIDs are very flaky • No power source • Ultra‐low cost not much circuitry • RFIDs can’t perform typical functions like carrier sense or rate adaptation Machine‐Generated Data RFID will be a major source of such traffic • “number of RFID tags sold globally is projected to rise from 12 million in 2011 to 209 billion in 2021.” – McKinsey Big Data Report 2011 • In Oil & Gas – about 30% annual growth rate • In Healthcare – $1.3B revenue annually Are Our Wireless Protocols Ready? Background Backscatter Communication Node Identification Problem Challenge: RFIDs cannot hear each other Collisions Applications: • Inventory management • Shopping cart Each object has an ID Reader learns IDs of nearby objects Data Communication Problem Data communication in RFID networks performs poorly because it lacks rate adaptation RFIDs always send 1 bit/symbol Can’t exploit good channels to send more bits Inefficiency Can’t reduce rate in bad channels Unreliability ID = 1 ... ID = 2 ID = 4 ID = 3 ID = 5 ID = N ID = 6 System is represented by a vector if node with ID = is in cart Network‐Based Compressive Sensing 0 500 1000 1500 2000 4 8 12 16 Number of Tags Our Design Slotted Aloha Number of Symbols to Identify Nodes 5.5× reduction in symbols needed for identification Reliability Message Loss Rate TDMA CDMA Our Design 0.57 bits/symbol 1.7 bits/symbol 3.2 bits/symbol 0% 10% 20% 30% 40% 50% Medium SNR (5dB െ 9dB) High SNR (10dB െ 20dB) Low SNR (0dB െ 4dB) Efficiency Evaluation • Reader implementation on GNURadio USRP • UMass Moo programmable RFIDs Network‐Based Rate Adaptation • Nodes transmit messages and collide • Reader collects collisions until it can decode • good channel decode from few collisions • worse channel decode from more collisions b 1 b 2 b 3 ⁞ b K y 1 y 2 y 3 y 1 =h 2 b 2 +h K b K ⁞ y 2 =h 1 b 1 y 3 =h 2 b 2 +h 3 b 3 +h K b K Collisions act as a distributed rateless code Adapts bit rate to channel quality without feedback Collisions act as a sparse random code Quickly decode using a Belief Propagation decoder 0 1 0 0 1 0 … 0 Node with ID = transmits Collisions mix bits on the air is distributed across all nodes! is Sparse Want the network to emulate a compressive sensing sender Reader decodes using a Compressive Sensing decoder Bits transmitted by different nodes Received wireless symbols