Transcript Sphere of Influence Project owner: Mark Yarvis Intel

Copyright© Intel Corporation 2000-2004
Industrial Applications
for Sensor Networks
Condition based monitoring pilot project
Lama Nachman
Lakshman Krishnamurthy
Researcher
Hans Mulder
Intel Research and System
Technology Lab
Ralph Kling
Mark Yarvis
Jasmeet Chhabra
Carl Dellar
®
Copyright© Intel Corporation 2000-2004
Agenda
 Introduction & Problem Statement
 Equipment Health monitoring Pilot in Intel FAB
 Application requirements
 Current Implementation (Phase 2)
 MICA & iMote clusters
 Reliability protocol
 Network Configuration
 Power Saving Protocol
 Status & Next steps
 Key learnings
Intel Research
•2•
Copyright© Intel Corporation 2000-2004
Fab Pre-emptive Maintenance
Application at Intel
 Use vibration signatures to
identify problems with equipment
 Avoid failure




~5000 Sensor points in each fab
4 years of archived data
Done by sneaker net today
Move to wireless sensor network
 Demonstrate a commercially feasible ROI for sensor
and mesh network deployments
Intel Research
•3•
Copyright© Intel Corporation 2000-2004
Problem Statement
 Equipment failures in live production fabs is
extremely costly ($Millions)
 Shutdown results in opportunity loss
 Cost of evacuation and requalifying all the tools
 Possible loss of wafer lots in the pipeline
 Need to predict equipment failures early
enough and perform preemptive maintenance
during pre-scheduled down-time
 Monitor equipment health using vibration
signatures
Intel Research
•4•
Copyright© Intel Corporation 2000-2004
Case study (RA FAB)
 ~5000 sensing points already instrumented
 40% permanent sensors, 60% portable sensors
 Vibration and RPM sensors (Wilcoxon & Honeywell)
 Manual data collection using handheld devices
 Time domain data is collected, spectrum and magnitude
plots are generated
 Data is downloaded to Rockwell Enshare software
 Sensors are manually configured in DB
 Type, location, direction, collection frequency, etc
 Alarms are generated, further manual collection is
performed on specific sensors
Intel Research
•5•
Copyright© Intel Corporation 2000-2004
Case Study (RA FAB)
 Prevention estimates
 Once per month -> catch (80-85)%
 Once per week, and selective daily collections -> catch
~97%
 Manual collection method is currently used
 Target is once per month
 Headcount cost : ~$500,000 in one FAB
 Rockwell based solution (EnWatch)
 Ethernet based on-line system (~$5000)
 16 channels, data collection and analysis
 Controlled by EnShare backend
Intel Research
•6•
Copyright© Intel Corporation 2000-2004
Application Requirements
 Interface to Wilcoxon vibration sensors and Honeywell RPM
sensors
 0.5 Hz – 5KHz range
 3000 Samples, 16 bits each
 Collect once per week (optional selective collection)
 Battery life
 No access to power or Ethernet at sensing locations
 6 months @ 1 collection per month
 4 months @ 1 collection per week
 Reliability
 MTBF : 6 months
 Identify bad data (especially false good data)
 Interface to Rockwell EnShare backend
 Automatic network configuration and maintenance
Intel Research
•7•
Copyright© Intel Corporation 2000-2004
Pilot Network Architecture
Fab Equipment
Intranet
Intranet isolation
Root Node
Ad Hoc Mote
Network
Cluster Heads
Mote + Vibration Sensors
Intel Research
•8•
Copyright© Intel Corporation 2000-2004
Solution components
 Ad hoc Mote network
 MICA based clusters
 Imote based clusters
 End to End Reliable datagram transport protocol
(sensor node -> Root Node)
 802.11 overlay mesh network using stargates
 Cluster head manages data collection and
sleep/wake schedule
 Root Node collects the raw data, stores in EnShare
format and sends it to server
 EnShare data base imports the raw data
Intel Research
•9•
Copyright© Intel Corporation 2000-2004
Reliability Protocol
 Runs on Mica motes, iMotes and Stargates
 TinyOS implementation
 Provides VarSend, VarRecv interfaces to app layer
 Uses Generic Packet interface to abstract network layer
 Sliding window protocol
 Connection parameter negotiation (fragment size,
window size, timeout info)
 Receiver sends an ACK bitmap within window
 Sender retransmits NACK’d fragments
 3 phases
 Connection setup (light weight, 2 packets)
 Data exchange (data and NACK packets)
 Final ACK (2 packets)
Intel Research
• 10 •
Copyright© Intel Corporation 2000-2004
Mote Cluster
Implementation
Intel Research
• 11 •
Copyright© Intel Corporation 2000-2004
Data Collection / Power Saving
 Cluster head sends a command to each
sensor node to start data collection
 Sensor node initiates reliable transport
protocol with RootNode for each connected
sensor
 Sensor node informs cluster head when data
transfer is complete
 Cluster repeats the process for each sensor
node
 When all sensors have been collected, the
complete cluster is put to sleep until next
collection
• 12 •
Intel Research
Copyright© Intel Corporation 2000-2004
iMote Cluster Details
Intel Research
• 13 •
Copyright© Intel Corporation 2000-2004
Intel Mote: an enhanced wireless
network research platform
 Intel Mote is a modular,
stackable design
 Hardware features
 High platform integration level (core,
radio, memory…)
 Main board (ARM core, SRAM,
FLASH, BT radio)
 Low power operation
 Power supply board (battery,
AC, solar, …)
 Small physical size
 Modular HW/SW design
 Sensor board(s)
 Low cost and volume production
potential
 Other boards (alternate radio,
debug, actuator, …)
TinyOS applications
TinyOS base components
Network layer (multihop)
Sensor board
Main board
Intel Mote layer
Power board
Firmware (BT-LLS)
Backbone interconnect
Hardware
Intel Research
• 14 •
Copyright© Intel Corporation 2000-2004
Network Configuration
 Automatic scatternet formation algorithm




Forms a tree structure
Clusterhead is the root of the tree (Master Role)
Intermediate nodes have dual Master/Slave roles
Leaf nodes are slave only nodes
 Free nodes alternate between BT Inquiry & scan
modes to discover other nodes
 Free nodes can join at different levels in the tree,
depending on which node they connect to
 Connected nodes only scan to eliminate the
possibility of creating loops
 Simple routing algorithm
Intel Research
• 15 •
Copyright© Intel Corporation 2000-2004
Power Saving Protocol
 Leverage low power modes in Bluetooth
 Cluster head broadcasts a “network sleep” message
down the tree.
 Once the message reaches a leaf node, a response
is sent up the tree
 When a master hears a response from all its slaves,
it will put all the links on hold, and propagate the
response up the tree
 Messages can still flow through the network in
between hold intervals (20 second response time
per level in the tree)
 The cluster will broadcast a “network wake up”
message down the tree
Intel Research
• 16 •
Copyright© Intel Corporation 2000-2004
Network Observations
 1 minute to form a cluster of 16 nodes
 BT links are very stable once established
 Network formation overhead is amortized
over long connection time
 BT link layer reliability is very effective,
hence reducing the end to end NACKs
 Need to optimize the scatternet formation
algorithm to select connections based on link
quality, and reducing hop count
Intel Research
• 17 •
Copyright© Intel Corporation 2000-2004
iMote Cluster integration
 iMote cluster
 Simple routing algorithm
 iMote MHOP header (src, dest, channel)
 TOS Message is not used
 MICA cluster & RootNode
 DSDV & flood protocols
 iMote cluster head translates between domains
 Route update messages from rootnode intercepted to get
the RootNode ID
 Reliability protocol hdr/data is repackaged
 Sensor -> RooNode (iMote packet -> DSDV packet)
 RootNode -> Sensor (Flood -> iMote packet)
Intel Research
• 18 •
Copyright© Intel Corporation 2000-2004
Sensor Board
 18V power supply
 10kHz+ 24bit A/D
 Programmable antialiasing filter
 PLD bridges SPI to UART interface
Wilcoxon sensor
Voltage
output
Intel Research
Sensor
voltage
supply,
A/D, filter
• 19 •
Intel Mote
UART
900kb/s
SRAM 64kB
FLASH 512kB
Copyright© Intel Corporation 2000-2004
Time domain data
Intel Research
• 20 •
Copyright© Intel Corporation 2000-2004
Frequency domain data
80Hz
reference
signal
Intel Research
• 21 •
Copyright© Intel Corporation 2000-2004
Status
 Phase 2 development complete
 Testing will begin in the JF3 chiller room with
MICA & iMote clusters next week
 Hardware is installed CUB3
 CUB3 deployment is scheduled for mid June
 Collecting performance and power data for
platform comparison by end of June
Intel Research
• 22 •
Copyright© Intel Corporation 2000-2004
JF3 Pilot Deployment
Facilities
Rooms
Intel Research
• 23 •
Copyright© Intel Corporation 2000-2004
Next Steps
 Finalize phase 3 requirements
 Choose one Mote platform based on the
phase 2 data
 Move to TinyDB/TASK
Intel Research
• 24 •
Copyright© Intel Corporation 2000-2004
Key learnings (platform)
 Size requirements
 Not very sensitive
 Current solutions are much larger
 Mote size is negligible
 Power Consumption
 Sensor + A/D consume a lot (~60 mW)
 Can use large batteries
 Large RAM is very useful
 Adding more capabilities to the mote simplifies the
sensor board design
 Fast I/O on the mote is useful
Intel Research
• 25 •
Copyright© Intel Corporation 2000-2004
Key learnings (Network)




Automatic configuration of the network is required
Reducing hop count is key
Heterogeneous networks are very useful
Simple power saving protocols are sufficient
(Cluster based)
 Matching radio bandwidth to application
requirements can save power
 Fast network response time is needed, even if
collection frequency is infrequent
 Polling specific sensors and adding streaming
modes will be very useful
Intel Research
• 26 •
Copyright© Intel Corporation 2000-2004
Key learnings (Network)
 Debug modes would be very useful
 Tracing network topology and data flows
 Performance and power monitoring
 Isolating bad data and recovering from
failures
Intel Research
• 27 •
Copyright© Intel Corporation 2000-2004
Key Learnings (backend)
 Interfacing into existing tools is extremely
important
 Want to use Rockwell EnShare for network
control/command
 Getting info into/out of Rockwell was very painful
 Easing the installation process is very
desirable
 Automatically recognizing sensors and their
location is very useful (sensor -> equipment
mapping)
Intel Research
• 28 •