Decision tree-based machine learning algorithm for in

Decision tree-based machine learning algorithm for in
Development of Micro Wireless
Sensor Platforms for Collecting
Data of Passenger-Freight
Interactions
Mohammad Mozumdar
Graduate Students: Kyle Ying, Alireza Ameri, Ankit Trivedi,
Dilip Ravindra, Darshan Patel, Varun Sharma, Ashutosh Parhad
Department of Electrical Engineering
California State University Long Beach
The thing we don’t like but have to face
everyday – Road Traffic!
But we can make smarter turn with real
time traffic!
Google Maps showing optimized route from USC to Orange County
How exactly does Google Maps/Garmin/TomTom know how
clogged the highway is on your way out to home or office?
The traffic information comes from a variety of sources:
• Commercial traffic data providers (INRIX, Tele Atlas, HERE, ..)
• Departments of Transportation
• State agency – Caltrans
Raw data is collected from:
• Mobile users (Google Maps)
• Road sensors
• Traffic cameras, and even through aircraft
This information is compiled and delivered via radio frequency
(FM/HD Radio™ or satellite) to your navigation system.
Road Sensors: Inductive Loop
Source: US DoT Federal Highway Administration
Physical Representation
•
•
•
•
Loop Detector Schematic
Existing traffic/vehicle detection is determined with “Inductive Loop”
technologies
These loops generate a magnetic field that operates at frequencies typically
less than 1kHz
Large rectangular loops (4’ x 8’, 6’ x 8’, 6’x 12’ are used to detect larger
vehicles
Small size loops (i.e. 2’ x 5’, 3’ x 6’, 6’ x6’) are used to detect smaller vehicles,
such as motorcycles and automobiles
Road Sensors: Inductive Loop
Source: US DoT Federal Highway Administration
Inductive Loop Pros & Cons
Advantages
• Detects ferrous objects precisely
• Typically immune from environmental effects such
as weather, temperature, a terrain variations
Disadvantages
• Expensive to install and maintain ($$$)
• Relatively significant power usage for the
generation of the magnetic field.
• Large area usage (greater than 10 sq.ft.)
Proposed Solution For Smart Road
Proposed VS Traditional Inductive Loop Based System
Anisotropic Mangeto-Resistive
Sensors (AMR) IC Sensor
AMR Sensor IC (Honeywell HMC5883L 3-axis
magnetometer- 3mm in size )
• Wheatstone bridge variable resistor
network that changes resistance w.r.t.
changes to the magnetic field
• Provides the same advantages to inductive
loop technologies without the power and
area disadvantages
• Power consumption extremely low
(~200uA at lower sampling rates )
Source: Honeywell
Microcontroller (CC430F5137 )
• Low power modes (LPM) for sleep
between computational and
communication operations
• Single package μproc and RF core for low
area wireless transmissions
Source: Texas Instruments
Machine Learning Based Vehicle
Classification
• Useful when the sets of data is large enough that
human observations for extracting patterns in
data become impractical.
• Typically associated with the field of data mining
• Pattern recognition based on a set of rules
General Idea:
• Collect vehicle data crossing
the AMR sensor
• Utilize ML tools to generate a
model for classification
Our Lab Testbed Setup
• 7 different RC Vehicles with a variety of similar and different
attributes
• 7 ft straight track for each vehicle to make passes
• 2 sensors roughly 4’ apart to take gather readings and
classify
Data Collection for supervised machine
learning
• We collected data for each of the 7 vehicles
across 350 runs over 2 sensors
• Total : 700 samples, 100/class for training
Why a decision-tree based algorithm?
• Simple and computationally efficient tree
• Simplicity of implementation in software
Implementation Flowchart
Normalize
Raw Data
Vehicle
Threshold
Detection
Record Data
in Detection
Window
Output
Vehicle
Classification
Classify
Vehicle using
Decision
Tree Model
Feature
Extraction at
Window End
Adaptive Baseline
2000
Magnetic Field Change (uT)
0
-1000
-2000
-3000
-4000
-5000
0
5
10
15
20
25
30
35
40
45
50
35
40
45
50
Time (seconds)
1500
1000
500
Magnetic Field Change (uT)
• Zeroing the background
environmental magnetic
field by offset
• Allows for the reuse of
the same vehicle
detection and
classification algorithm in
multiple environments
• Noise removal can be
implemented at this
stage
1000
0
-500
-1000
-1500
-2000
-2500
-3000
-3500
0
5
10
15
20
25
30
Time (seconds)
Threshold Detection
• Once a vehicle
passes the threshold
the detection flag
triggers and a certain
number of samples
are recorded for
processing
Magnetic Field Change (uT)
Magnitude with vehicle overhead
Time (seconds)
Detection Window
Threshold
Features Collected from Vehicles
Interesting Features:
• Min: minimum value of an axis during the detection
window
• Max: maximum value of an axis in the window
• Mean: average of all axis values in the window
• Range: Maximum – Minimum
Using a 3-axis sensor this results in 12 unique features
These Features are very simple to calculate and compute
Example Plot and Feature Extraction
Feature Extraction
1500
DataX
DataY
DataZ
MinX
MaxX
MinY
MaxY
MinZ
MaxZ
1000
Magnetic Field Change (uT)
Features
Data (X,Y,Z)
500
MinX=-1097
MinY=-256
MinZ=-834
MaxX=1054
MaxY=267
MaxZ=1011
MeanX=4.01
MeanY=17.13
MeanZ=7.48
RangeX=2151
RangeY=523
RangeZ=1845
0
-500
-1000
-1500
0
10
20
30
40
50
60
Time (seconds)
70
80
90
100
Example Car Data
Big Racer Car Data
Buggy Car Data
1000
200
X data
Y data
Z data
800
150
100
Magnetic Distortion (uT)
Magnetic Distortion (uT)
600
400
200
0
-200
50
0
-50
-100
-400
-150
-600
-800
X data
Y data
Z data
0
10
20
30
40
50
60
70
Samples (1/75 seconds)
80
90
100
-200
0
10
20
30
40
50
60
70
Samples (1/75 seconds)
80
Similar Sizes but Different Signatures!
90
100
Machine Learning: decision tree
learning (J48)
J48 is the open source Java
implementation of C4.5/ID3
developed by John Quinlan
Inputs: multiple features
corresponding to a single
classifier
Note: higher # samples per classifier results in a
more accurate output tree
Output: a decision tree with
the highest classification rate
given the features
Source: University of Waikato, WEKA
WEKA output
4 Features Selected
Fast!
Accuracy
Rate
Graphical Tree
Yellow
BigGreen
Red
Green
Diecast
Buggy
Red
BigGreen
BigGreen
Big Car
Buggy
Node
Num.
Min Y
1
2
3
4
5
6
7
8
9
10
0.9527
0.3339
0.3060
0
0.7108
0.9317
0.2047
0.1386
0.5724
0
Information Gain Attribute
Max X
Mean Z
0.8002
0.9707
0.8411
0.1735
0
0.1864
0.1601
0.4467
0.3189
0
0.9031
0.3978
0.2155
0
0.7108
0.7574
0.3717
0.8798
0.9625
0.0066
Range Z
0.9868
0.3846
0.3068
0.0379
0.0570
0.4976
0.4249
0.0084
0.8676
0.6627
Best Attribute
Range Z
Max X
Max X
Max X
Min Y
Min Y
Range Z
Mean Z
Mean Z
Range Z
Brute Force Search for Best Results
• The output tree doesn’t always generate the
best results given a large number of features
• Due to fast processing time to generate the
output tree, we can easily calculate all
combinations (n choose k)
𝑛
( ) or 𝑛 𝐶𝑘
𝑘
• We use n=2,3,4 where k=12
Feature Performance
2 features (66 combinations):
3 features (220 combinations):
4 features (495 combinations):
.
Comb#
Classification%
Features
64
94%
minx maxx
58
93%
minx rangex
30
93%
maxx rangey
60
91%
minx meany
Comb#
Classification%
219
98%
minx miny maxx
200
98%
minx maxx maxz
194
97.8571%
minx maxx rangez
57
97.8571%
maxx rangey rangez
149
97.7143%
miny maxx rangez
Comb#
Classification%
479
98.8571%
minx miny maxx rangez
270
98.8571%
miny maxx meanz rangez
390
98.7143%
minx maxx meanz rangez
78
98.7143%
maxx meany meanz rangez
Features
Features
Best Results Simulated vs. Testbed
Simulated Results
Actual Results:
Cross-Validation Percentages
Number of Features (Attributes)
Real world Classification Percentages
Accuracy
Number of Features (Attributes)
Real-world
Three Features (maxx, rangey, rangez)
97.86%
Three Features (maxx, rangey, rangez)
98.57%
Three Features (miny, maxx, rangez)
97.71%
Three Features (miny, maxx, rangez)
97.38%
Four Features (minx, miny, maxx, rangez)
98.86%
Four Features (minx, miny, maxx, rangez)
90.24%
Four Features (miny, maxx, meanz, rangez)
98.86%
Four Features (miny, maxx, meanz, rangez)
99.05%
Simulated results match real world testing values
very closely.
Note: minx results are lower due to clipping
Energy Scavenging Using Piezoelectric
Sensors
• Mechanical to Electrical
energy conversion
• Proper implementation
can help in continuous
operation of wireless
sensors
• Almost 70% of the
overall efficiency of the
energy scavenging
system depends on
Piezoelectric sensors
• Applications include
consumer electronics,
automotive, health,
WSN, etc.
Pressure generated by tires of cars
on the piezoelectric sensors
Generated power is stored in batteries
Energy Scavenging System
Piezoelectric
Sensors
Mechanical Energy
to Electrical
Energy
Switch
Rechargeable
Battery
Energy Source
ADC
Voltage Regulation
Rectification
Circuit
Charge Regulation
Circuit
Power Storage
Element
Lab Prototype of Energy Scavenging System
Advanced Energy Scavenging System
• 3 layers instead of 1 layer/ Smaller size and less implementation cost
• Increase probability of the sensors being pressed in every tap
• Increase number of sensors being pressed in a single tap
Energy Scavenging System
 1 AA rechargeable battery can be charged in
10 -12 hours with vehicles and pedestrians
passing over the sensors in every 5 seconds
using the designed hardware
 The sensors placed on crosswalks can increase
the average number of taps
 Charging rate would be better if the efficiency
and the number of sensors used are increased
Designed Smart Traffic Sensing Node
• Size: 2 ½” x 1 ½” x ¾” (with AA battery pack)
• Dimensions will change depending on the
battery pack used in future implementations
Final Remarks
The system described in this presentation can
replace current inductive loop technologies with:
– Maintain traffic/vehicle detection capabilities
– Additional features such as vehicle classification
– Lower power consumption
– Lower physical area utilization
In addition, many classifiers can be used at high
accuracy rates compared to other methods
utilizing solely novel features.
Related Publications
•
K. Ying, A. Ameri, A. Trivedi, D. Ravindra, D. Patel, M. Mozumdar,” Decision Treebased Machine Learning Algorithm for In-node Vehicle Classification”, Proceedings
of IEEE Green Energy and Systems Conference, Long Beach, November 2015, USA
•
V. Sharma, A. Parhad, M. Mozumdar, “Energy Scavenging Using Piezoelectric
Sensors to Power in Pavement Intelligence Vehicle Detection Systems”, METRANS
International Urban Freight Conference , Long Beach, 2015
Questions?
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement