IBM RESEARCH
© 2008 IBM Corporation
Johnathan M. Reason

IBM RESEARCH
Outline

 Why wireless sensor networks for railroads
 Railroad sensor network solution
 Some Results
 Next Steps
Johnathan M. Reason
© 2008 IBM Corporation 2

IBM RESEARCH
The North America Railroad Industry
 40% of U.S. freight travels by rail
• Major contributors are coal, chemicals, food, and machinery
• Intermodal rev. has been consistently growing
 Railroads are three times as fuel-efficient as trucks
 7 Class 1 railroads represent 90% of total freight revenue (each with over $320M in annual sales)
• Burlington Northern, Union Pacific,
Canadian National Railway, Norfolk
Southern, CSX, Kansas City Southern,
Canadian Pacific Railway
 30 Regional railroads
• e.g Florida East Coast Industries, …
 Hundreds of locals (short line operators )
Johnathan M. Reason
© 2008 IBM Corporation 3

IBM RESEARCH
Union Pacific Railroad Fast Facts (2007 data)
• Largest railroad in NA
• Op. Revenue $15.5B
• Industrial, energy, intermodal, agricultural, chemicals, auto, etc.
• Route Miles 32,300
• Employees 50,000
• Annual Payroll $3.7 billion
• Purchases Made $6.9 billion
• Locomotives 8,500
• Freight Cars 104,700
• Fuel efficiency 780 ton-mile/g
• More than 70% of IT budget is spent on supporting the operations
Johnathan M. Reason
© 2008 IBM Corporation 4

IBM RESEARCH
Railroad track-side sensors: railcar identification and fault prevention
 AEI: Automatic Equipment Identification
• NA railroad standard: identify railroad equipment while enroute
• passive UHF RFID tags mounted on each side of rolling stock
• trackside readers
• Adopted since early 1990’s
• As of 2000, over 95% railcars were tagged with 3000+ trackside readers
 In addition to AEI readers, additional sensors are deployed along the track, including
• Hot Box Detectors (bad bearings)
• WILDs or Wheel Impact Load Detectors (bad wheels)
• TADs or Trackside Acoustical Detectors
(cracked or flat wheels)
AEI Reader
AEI tag affixed to the side of a freight car.
Hot-box Detector
Johnathan M. Reason
Acoustic Sensor
Wheel Impact Load Detector
© 2008 IBM Corporation 5

IBM RESEARCH
Outline Progress
 Background on N.A. Freight Railroads

 Railroad sensor network solution
 Some Results
 Next Steps
Johnathan M. Reason
© 2008 IBM Corporation 6

IBM RESEARCH
Problem Summary
 Data from trackside sensors are sparse
• Does not provide timely information to prevent or mitigate all problems (sample every 45 min, on avg.)
• Each technology is one-dimensional; not capable of supporting all the operational needs
• Does not scale well for multiple sensor modalities
 Proposed next-generation infrastructure requires
• On-board telemetry for real-time visibility, using wireless sensor nodes or motes
• One infrastructure supporting multiple sensor modalities
• One infrastructure for communicating data, control, and events
• Localized analytics
• Demonstrable ROI
• Large-scale deployment
Johnathan M. Reason
Railcar
Tracking
Brake control
Bearing temperature
Weight distribution
© 2008 IBM Corporation 7

IBM RESEARCH
Capabilities of a Wireless Sensor Node (or mote)?
 Computation:
• Low-power microProcessor
(e.g. TI MSP430)
• Small amount of memory
(e.g. 10KB RAM, 48KB ROM)
 Sensing:
• Temperature and light onboard
• Embedded A/D converter
• SPI bus for expansion
 Communication:
• low-power energy efficient radio
(e.g. 802.15.4)
Computation Sensing
Communication
Mica2 from Crossbow
Iris mote from Crossbow
Design Tradeoffs:
Energy Vs Performance
Cost Vs Computational power and reliability
Telosb from Crossbow
© 2008 IBM Corporation 8
Johnathan M. Reason

IBM RESEARCH
Outline Progress
 Background on N.A. Freight Railroads
 Why wireless sensor networks for railroads

 Some Results
 Next Steps
Johnathan M. Reason
© 2008 IBM Corporation 9

IBM RESEARCH
SEAIT: Sensor-Enabled Ambient-Intelligent Telemetry for Trains
 SEAIT is a WSN-based architecture and framework for building advanced railroad applications.
 The framework provides a collection of protocols, services, and a data model that serve as the building blocks to enable intelligent telemetry through
• timelier sensing,
• localized analytics, and
• robust communications.
 The architecture specifies an onboard infrastructure to facilitate realtime data capture and analysis for better visibility and in-field management of the rolling stock.
 At the heart of the architecture are intelligent wireless sensing nodes that form the on-board WSN and continuously monitor the health of critical components (e.g., wheel bearings).
© 2008 IBM Corporation 10
Johnathan M. Reason

IBM RESEARCH
Goals, Applications, and Benefits of SEAIT
 Goal
• Improve operational business objectives by providing real-time visibility into the rolling stock
 Some Enabled Applications
• Real-time Fault Detection with Closed-loop Notification
• Train Configuration Monitoring
• Asset Tracking
• Predicative Maintenance
• Continuous Health Monitoring
 Some Key Business Benefits
• Schedule Optimization
• Accident Prevention
• Asset Utilization
• Customer Satisfaction
© 2008 IBM Corporation 11
Johnathan M. Reason

IBM RESEARCH
Basic Approach of SEAIT Illustrated through a Hot Bearing Detection Solution
 WSN nodes
• perform timelier sensing of wheel bearing temperature,
• local analytics to detect overheated bearings, and
• robust communications to relay “hot” bearing events to the gateway
 Gateways
• aggregate hot bearing events with other situational awareness data,
• perform train-wide analytics, and then
• provide closed-loop event notification directly to the engineer
 WSN nodes communicate to gateways on locomotives or trackside gateways
 Locomotive gateways communicate to the enterprise via an uplink (Cellular, WiFi,
Satellite, proprietary RF bands, etc.) e) gateway f) locomotive
...
a) hopper car b) WSN node c) thermocouple d) bearing adapter
© 2008 IBM Corporation 12
Johnathan M. Reason

IBM RESEARCH
Key Technology Components
 Gateway Software
• Information model called the Railroad Business Object Model (RRBOM)
• RRBOM is the meta-model for all railroad objects (trains, cars, axles, wheels, bearings, motes, sensors, etc.)
• Uniform information model for enterprise applications to configure, query, and control the mote network
• Performs onboard, train-specific analytics (enables closed-loop control)
• Supervise railroad communication protocols and services
 WSN Node Software
• Uniform information and messaging model for managing and reporting sensor, configuration, and application data; provides hooks to gateway to map into RRBOM
• All communication protocols and services to realize railroad applications and support application requirements
• Low Latency
• High Reliability
• Long Life
© 2008 IBM Corporation 13
Johnathan M. Reason

IBM RESEARCH
Key Technical Challenges to Realizing the Benefits of WSN for Railroads
 Detection and Measurement Accuracy
• Reliable detection and prediction of catastrophic faults (e.g., over heated bearing) with low false positive rate
• Accurate reporting of train consist and parameters for operational optimization
 Alert Latency
• Predictable, low end-to-end latency from detection of a fault to alerting the engineer of such an event over many hops
 End-to-end Data Reliability
• End-to-end reliability over many communication hops under various conditions
(weather, speed, terrains, ...)
 Service Lifetime
• The energy source for each mote must last at least the maintenance cycle of its associated car (> 5 years)
© 2008 IBM Corporation 14
Johnathan M. Reason

IBM RESEARCH
Gateway-to/from-Railroad WSN Architecture
 Applications and services send and receive messages through the interface to the communication stack
 The information and reporting services realize the execution a uniform information model for managing and reporting sensor, configuration, and application data
 The synchronization service realizes simple and robust management of a software RTC
 Network features time-scheduled queues and cross-layer optimized routing
 Link features semantic-based wakeups and delay measurements
Railroad Applications
Car-to-Train Hot Bearing
Association Detection
Consist
Identification
Dark Car
Detection
...
Node Services
Information Reporting
Synchronization
Router
Next
-hop
Proto
1
List
...
Proto
N
Message Manager
Sender Receiver
Rx/Tx Queues
Dispatcher Scheduler
Neighbor List nodeId Addr Position Hops Cost
Packet Delivery
ARQ Multisend Wakeups
Packet Measurements
Delay LQI RSSI PSR
IEEE 802.15.4
© 2008 IBM Corporation 15
Johnathan M. Reason

IBM RESEARCH
Consist Identification: Car Disjoining from the Train
 Problem: Dynamic join/disjoin of rail cars
• No real-time or near real-time visibility of what cars are actually on the train
 Possible Solution: Periodic car ID reporting via a Mote network
• If one or more motes are uniquely associated with each car, then dynamic join/disjoin is a simply application that detects the presence/absence of a car-specific mote in the network
• Motes can detect the status of their car and change their mode of operation: join => active reporting, disjoin => hibernation
Car D’s motes leave the network as car D is disjoined from the train and the train in no longer in range motes gateway
Wayside
E D C B A
© 2008 IBM Corporation 16
Johnathan M. Reason

IBM RESEARCH
Consist Identification: Basic Operation
 Iterative application that has four major phases:
1.
Associate cars to the train
2.
Measure closeness between each neighboring car (or pair of nodes)
3.
Report closeness measurements
4.
Apply the ordering algorithm
 Ordering Algorithm
Considering n cars in a train {N | i = 1,...,n}, the ordering algorithm operates in three steps:
1.
Compute a car closeness metric {d ij
} from the node measurements
2.
Refine the car closeness metric using a correlation based operator
3.
Construct a weighted digraph, G= (N,E), where each edge has a weight of d ij
.
The closeness metric reflects the closeness between two cars N i
The closer the two cars are, the greater the value for d ij
& N j
.
.
Consist ID is equivalent to finding the max. Hamiltonian path for graph G.
We use a greedy algorithm to construct this path.
The gateway in the locomotive serves as an anchor node.
© 2008 IBM Corporation 17
Johnathan M. Reason

IBM RESEARCH
Outline Progress
 Background on N.A. Freight Railroads
 Why wireless sensor networks for railroads
 Railroad sensor network solution

 Next Steps
Johnathan M. Reason
© 2008 IBM Corporation 18

IBM RESEARCH
Proof-of-Concept (PoC) Testbed Deployed on the Roof our Yorktown Facility
 Deployed 32 WSN platforms along the front metal railing of the roof to emulate a 16-car train
 WSN platform:
• TmoteSky node, a sensor board, batteries, an embedded antenna, an input/output connection board, and a weatherproof enclosure. The sensor board included temperature, light, and accelerometer sensors.
 On average, freight railroad cars are about 60 feet long, ranging from as little as 40 feet up to 90 feet
 Two WSN platforms per car (one at each end), each car 60 feet long and an inter-car node spacing of 10 feet
 Sample segment of the deployment showing four cars. The entire deployment spans about one fourth of a mile.
 The curvature of the front face of the building is such that, from any point along the front edge, no more than 300 feet are visible via line-of-sight.
Segment of Deployment
WSN Platform
© 2008 IBM Corporation 19
Johnathan M. Reason

IBM RESEARCH
PoC Results for Consist Identification
Setup :
• Used periodic reporting with hop-based routing. Period was every 2 minutes
• During slot time, each node measured closeness to its neighbors and reported these measurements to the gateway
• Closeness measurements consume most of the time during each slot
• Gateway runs Consist Identification algorithm
• Error = # of cars that need to be moved to match the actual consist
Key Observations :
• Algorithm is robust within 1-car transpositions or flips
• A flip is equivalent to a 2-car error
• Ignoring flips, the algorithm is 100% accurate
4
3 including flips ignoring flips
Error 2
1
0
1 6 11 16 21 26 31 36 41 46
Report Cycle
Flips ignored
Accuracy (%)
0 1 2
93.0
99.0
100.0
© 2008 IBM Corporation 20
Johnathan M. Reason

IBM RESEARCH
Graphical view of a consist being constructed
(a screenshot of the research prototype)
Johnathan M. Reason
© 2008 IBM Corporation 21

IBM RESEARCH
Outline Progress
 Background on N.A. Freight Railroads
 Why wireless sensor networks for railroads
 Railroad sensor network solution
 Some Results

Johnathan M. Reason
© 2008 IBM Corporation 22

IBM RESEARCH
Next Steps: Continue the conversation about industry standardization
Timeline: North America Railroads AEI Deployment
1/88 1500 cars were tagged
8/88 reported 99.99% accuracy
Burlington Northern started a testing program and selected two vendors for full-scale testing
AAR formed std. committee
More RRs started testing
8/89 Amtech selected
Data format defined
10/90 AEI std approved
8/91 AEI mandate rectified by AAR
3/92 - 12/94
Mandatory rollout
1.4 million railcars tagged
1986 1987 1988 1989 1990 1991 1992 1993 1994 1995
Application Layer
Presentation Layer
Transport Layer
Network Layer
Link Layer
Physical Layer
Johnathan M. Reason
 PoC was a good starting point
 PoC touched many areas requiring standardization
• Communication (mote-mote, mote-gateway)
• Message/Query
• Industry semantic model/ontology
• Power
• SW life-cycle management
 Like RFID, broad adoption of WSN will be driven by industry applications and require industry collaboration
© 2008 IBM Corporation 23

IBM RESEARCH
Next Steps: Some Possibilities
 Continue PoC investigation by conducting field tests on real trains
 Quantify the value proposition of real-time visibility with research study
• Does more timely data really yield greater efficiencies in operations?
• If so, how much?
• What localized analytics are needed?
 Explore how WSN technology can complement positive train control
• As the PTC industry standard develops, what conversation should the industry be having about a path to on-board sensing and actuation?
© 2008 IBM Corporation 24
Johnathan M. Reason

IBM RESEARCH
Acknowledgements
 Union Pacific Railroad
• Lynden Tennison, Dan Rubin
 IBM
• Co-authors: Han Chen and Sastry Duri (IBM Research), Riccardo
Crepaldi (Intern)
• Contributors: Maria Ebling and Paul Chou (IBM Research), Xianjin Zhu
(Intern), Keith Dierkx (GRIC)
© 2008 IBM Corporation 25
Johnathan M. Reason

IBM RESEARCH
Johnathan M. Reason
© 2008 IBM Corporation 26