Docstoc

defence

Document Sample
defence Powered By Docstoc
					                                             1




An Automated Framework for Power-Efficient
  Detection in Embedded Sensor Systems




               Ari Y. Benbasat
        Responsive Environments Group
             MIT Media Laboratory
                                             ayb
                                                              2




Introduction
• Embedded sensor nodes are being used in a variety of
  applications, such as:
    Detecting the activities of housebound elders
    Monitoring wildlife in remote regions
    Tracking the state of smart assets in the supply chain
• Want to make sensor systems as power-efficient as
  possible to allow for a wider range of applications
• While any (reasonable) power savings is beneficial,
  there are two key targets:
    Mass-market applications
    Infinite life (via parasitic power)
                                                              ayb
                                                              3
Illustrative application:
Gait Measurement System
• Gait changes can be symptoms of important
  medical conditions
• Uses simple external shoe attachment rather
  than full motion tracking lab
   Motion tracked with inertial sensors and tactile insole
   Uses framework mentioned later
• Validated against MGH gait
  laboratory
• High 100mW power draw

                                                              ayb
                                                                   4
Illustrative application:
Commercial Potential
• Considerable interest in combining medical sensors
  with commonplace portable devices
• Difficult to achieve because of power use
• e.g. consider (naively) adding the above inertial sensor
  unit (65mW) to familiar portable devices:
    iPod Shuffle
         Currently 67mW - power usage almost doubles

    Motorola v60
       Currently 20mW (standby) - power usage increase of >300%

• But user is most often standing still!
                                                                   ayb
                                                                5




Overall Strategy
• First principles: reduce power usage of sensor nodes
  by reducing power usage of the sensors themselves
• Achieve savings through high-level algorithms and
  designs rather than low-level technical changes
    System will alter its sensing based on the current state

• Benefits of reduced sensing cascade upwards:
    Reduced data processing
    Reduced storage/bandwidth needs
    Reduced analysis for human experts


                                                                ayb
                                                                        6
Power Breakdown:
Field-Tested Applications

                      Baseline                 Sensing
                                                                %
                                                              Sensing
  System        µC               RF    Sensors     Response


 Gait Shoe   35 mW        N/A         65 mW       15 mW         70

 Zebra Net   15 mW        13 mW       30 mW       0.2 mW        52

Great Duck
             118 µW       465 µW      118 µW      14 µW        18.5
  Island




                                                                        ayb
                                                               7




Overall Approach
Three-part framework for the design and construction of
   power-efficient sensor systems:
   1. A modular hardware system for prototyping of power-
      efficient sensor nodes
   2. A technique for constructing state classifiers with
      parameterized power/accuracy trade-offs
   3. An embedded implementation of the classifier including
      sensor management




                                                               ayb
                                 8




Distinguishing Characteristics
Our solution is:
• Algorithmically generated
   Not ad-hoc
• Scalable
   Not limited size
• Responsive
   Not static



                                 ayb
                                                               9
Related Works:
Processor/Transceiver (Historic)
• Power reduction concentrated on the processor
  and transceiver
• Gains were achieved through either:
   Changing the rate of operation to complete the task just
    in time, or
   Completing the task as quickly as possible and sleeping
• Laptops and PDAs use these techniques



                                                               ayb
                                                                        10
Related Works:
Sensing (Recent)
•   Alter sampling rate based on measured data
       Responds in purely entropic fashion, hence cannot
        specify which states are/n’t of interest. [Jain 04/Rahimi 04]
•   Sentry nodes awaken slave nodes in network
       Slave nodes have fixed operation and do not use other
        information available to them. [He 04/Zhao 02]
•   State detection through sampling
       Binary and hand-scripted [Dutta 05]


                                                                         ayb
                                                           11
Test Application:
Real-time Gait Analysis
• Want to distinguish the following motions:
   Level gait
   Walking up or down stairs
   Walking up or down incline
   Shuffling
• Physicians often interested in only one of these
  motions:
   Parkinson’s Disease: Information about shuffling
   Total knee replacement: Information about stair climbing
                                                               ayb
                              12




Overall Approach
• Modular hardware


• Classification Algorithms


• Embedded Implementation




                               ayb
                                                         13
Modular Hardware:
Goals
• Encapsulate Knowledge
   Set designs for common structures
• Simplify Prototyping
   Architecture allows designer to quickly construct
    applications
• Expandability
   New boards can be added without redesigning others



                                                          ayb
                                                                    14
Modular Hardware:
Overall Board Design
                         Control power to sensor



                    Analog Signal
    Sensors                                    Mux             P
                     Processing
                                                   Select sensor

• Sensor boards with a variety of sensors
    Each with own signal processing
    Sensors chosen to be low-power and fast wake up
    Uses low power op amps and other components
• Microprocessor also controls the power to the individual
  sensors
                                                                     ayb
                                                            15
Modular Hardware:
Boards
• Processor board:
    Responsible for the data collection
    TI MSP430 processor with analog to digital
     converter
    Low-power, fast-wake sleep mode and hardware
     multiply
• Six-degree inertial measurement unit (IMU)
    Measures motion in all three dimensions
    Acceleration: two ADXL202 and a four-way static tilt
     switch (micropower/single-bit)
    Angular velocity: two Murata ENC03J and a
     ADXRS300 gyroscopes (allows for planar package)
                                                             ayb
                                              16
Modular Hardware:
Sample Uses of Platform
Complete applications:
• FlexiGesture: Instrument that allows
  flexible assignment from input gesture to
  output sound.
• Gait Shoe: Shoe mounted system for real-
  time measurement of subject motion
Used for prototyping:
• Huggable: Instrumented responsive plush
  bear to act as a companion animal in wide
  variety of settings.

                                               ayb
                                                                17
Test Application:
Sensors and Test costs
• We collected data using a system made up of
  the processor board and the IMU board
• Data collected at 200Hz
   Will downsample to mimic lower frequencies
                                 Power Usage (µW) at

                         25 Hz     50 Hz   100 Hz      200 Hz

          Gyroscope      22500     30000    30000      30000

         Accelerometer   396        792     1585       1983

           Tilt Switch   0.413     0.825    1.65        3.3


                                                                 ayb
                              18




Overall Approach
• Modular hardware


• Classification Algorithms


• Embedded Implementation




                               ayb
                                                           19
Classification:
Goals
• Determine the state as accurately and power-
  efficiently as possible
   System allows a trade-off between the power/accuracy
• Only sensors active are those needed to make a
  decision at a given point
• Requires a minimum of effort from the
  application designer



                                                            ayb
                                                                  20
Classification:
Data Stream
• The application designer will provide
  representative sensor data streams
   Individual states (e.g. shuffling, climbing stairs) must be
    annotated
• System will respond only when it is determined
  to be in one of these states
• Classification algorithms will generate a state
  decision tree using this data stream


                                                                   ayb
                                                        21
Classification:
Decision Trees
• Standard top down divide and
  conquer method for construction
• Each node specifies the
  information necessary to make              Mean(Az)
                                     yes
  the next decision                           <2009

                                    walk
• Each leaf specifies the
                                             Max(Gz)
  determined system state                     <1878
                                     yes


                                    Uphill     walk

                                                         ayb
                                                                     22
Classification:
Classifier Choice
Decision trees are used for the following reasons:
• Classification is structured as a sequence of
  queries:
   Only certain sensors are used based                   Mean(Az)
    on state                                      yes      <2009

   Hierarchical activation requires a           walk
    hierarchical classifier.                              Max(Gz)
                                                           <1878
• Can be easily converted into very               yes

  compact microcontroller code
                                                 Uphill     walk
   Very fast execution (based on comparisons)
                                                                      ayb
                                                                  23
Classification:
Features




• Features used are a selection of simple functions :
    Windowed minimum, maximum, mean, variance
    All constant time for incremental calculation (on average)
    Calculated at a number of sampling rates
                                                                   ayb
                                                                                  24
Classification:
Tree Construction (1)
    Number of positive (p)
    and negative (n) examples         [p,n]         Root node      Sample
                                                                    Split

               Left child       [p1,n1]   [p2,n2]        Right child


•     Decision trees are constructed in a recursive top down
      function
         All splitting points of all the features are tested
         Best division is chosen based on a purity measure C(split)
         Process continues with child nodes until all are pure (all one state)

                                                                                   ayb
                                                              25
Classification:
Test Costs
• For any given feature, the test cost is the total energy
  necessary to obtain the feature:
      Test Cost = Test Cost(sensing) + Test Cost(features)
• Test Cost(sensing) is energy necessary to collect a sample
• Test Cost(features) is the energy necessary to calculate the
  desired feature in the microcontroller
    For active sensors, the sampling cost dominates
    For passive sensors, the feature cost dominates

• Features already used in the tree are free (Test Cost = 0)

                                                                 ayb
                                                                        26
Classification:
Tree Construction (2)
•   Basic splitting criterion modified to take test cost into
    account:
                                        C ( split )
     Basic form: C ' ( split ) 
                                  (    (TestCost ))W
    where W is a parameter, which alters the power/accuracy trade-off

•   Desire two modes of operation:
    1. All activated sensors (Test Cost = 0) are equivalent
    2. Unused sensors are weighted proportional to test cost

•   Set =1 ,  such that min  · (Test Cost) = 10
•   Max. useful value of W is set by range of C(split)

                                                                         ayb
                                                             27
Test Application:
Structure
• Decision trees were built to separate each
  individual motion from the other five:
   Non-ambulatory (roughly: still) data was not used with
    this classifier
• W varied to create a population of classifiers
• Priors:           Motion         % time
                    Level Gait       80
                  Ascend Stairs      4
                  Descend Stairs     4
                  Downhill Walk      4
                   Uphill Walk       4
                  Shuffling Gait     4
                                                              ayb
                                                                        28
Test Application:
Comparison
• Support vector machines (SVM) were used for
  comparison
    Classify by dividing feature space using a linear or Gaussian kernel
    All available features are used in every classification

• SVMs are not appropriate for real-time embedded
    Too computationally expensive
    Linear: Dot product of vectors with length of number of features
    Gaussian: One dot product as above for each support vector

• SVM power usage was altered by constructing multiple
  classifiers, each with access to a subset of sensors
                                                                            ayb
 Uphill Gait     Downhill Gait   Shuffling Gait




Ascend Stairs   Descend Stairs    Level Gait
                                                      30
Test Application:
Tree for Uphill Gait (1)
           Mean(Tilt)<4.8               W = 0.020
                                        Acc = 0.980
                                       Cost = 5.4mW
walk                        Mean(Az)<2000



              SD(Gz)<72

                                    walk
  Uphill         walk

                                                       ayb
                                                    31
Test Application:
Tree for Uphill Gait (2)
           Mean(Az) <2000             W = 0.010
                                      Acc = 0.990
                                    Cost = 12.8mW
walk                   Mean(Tilt)< 4.8



              SD(Gz)<72


                                   walk
  Uphill            walk
                                                     ayb
                                                 32
Test Application:
Tree for Uphill Gait (3)
                                   W = 0.002
         Mean(Az) < 2000           Acc = 0.993
                                 Cost = 13.6mW
                  SD(Gz) < 78
walk


                            Max(Gy) < 1880
       walk



              Uphill            walk

                                                  ayb
                              33




Overall Approach
• Modular hardware


• Classification Algorithms


• Embedded Implementation




                               ayb
                                                               34
Embedded Software:
Goals
• Conversion of particular decision tree classifier
  for use in a real-time system
   Fairly compact implementation
• Software deals with two main timing issues:
   Microscopic: Power-cycling of components within a
    single data collection cycle to save power
   Macroscopic: Application of hysteresis to activity level
    transitions as a trade-off between power and latency


                                                                ayb
                                                            35
Embedded Software:
Main Loop
• Collect data from sensors current in use
• Run classifier to determine state (or next sensor
  needed)
   Respond to state (if desired by application designer)
• Handle sensor de/activation requests
• Sleep until next cycle




                                                             ayb
                                                                36
Embedded Software:
Sampling Cycle
    Actions
   Parts




                                                    Waking up
                                                     Active

• Active sensors are awakened in reverse order of wake
  up time
    Sensors with long wake-up times cannot be duty-cycled
                                                                 ayb
                                                                                  37
Embedded Software:
Sensor Activation
                                                                 Sensor needed
                                                               Sensor not needed 




• Small glitches in the data stream can change the current
  node of the decision tree
    Addition of latencies allows system to smooth over small glitches
• Gap in sampling prevents windowed features from
  being calculated for a full cycle
    Turn off latency is set long to take this into account
                                                                                      ayb
                                 39




Distinguishing Characteristics
Our solution is:
• Algorithmically generated
   Not ad-hoc
• Scalable
   Not limited size
• Responsive
   Not static



                                  ayb
                                                                       40




Contributions
• Designed an automated framework to construct power-
  efficient state detection systems
   A modular hardware platform, incorporating a number of low-power
    design techniques, to aid in the construction of embedded sensors.
   Modified the standard decision tree training algorithm to add a
    parameterized weighting relative to the cost of the features.
   Showed this parameter creates a population of trees at different
    points on the power/accuracy plane.
   In a complex sample application, these classifiers were shown to
    generally perform better on the cost/accuracy plane than support
    vector machines


                                                                         ayb
                                                             41




Future Work (1)
• Ability to add new features to classifier
   Exploit knowledge of application and structure of data
• Tree pruning algorithm to allow accuracy /
  latency tradeoff
   New sensors sometimes provide minimal increase in
    accuracy
• Add top-level trigger variables
   Binary decision which wakes system


                                                              ayb
                                                                42




Future Work (2)
• Increased automation to simplify the task of
  going from hardware to classifier to software
   Currently requires a lot of manual labeling and selection
• Addition of active sensors (such as radar/sonar)
  to the framework
   Various power levels alter operation of device
• Extend framework to sensor networks
   State or data from nearby nodes could be included in
    decision tree
                                                                 ayb
                                                       43




Acknowledgements
• Joe Paradiso
• Rosalind Picard and Mani Srivastava
• ResEnv
• Synthetic Characters
• Tangible Media and Aesthetics & Computation
• The Academic Office and our Administrative Support
• My friends inside (and out!) of the Lab
And of course,
• My parents
                                                        ayb

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:13
posted:6/14/2012
language:English
pages:43