Content uploaded by Michael Smith
Author content
All content in this area was uploaded by Michael Smith on Jan 16, 2022
Content may be subject to copyright.
Innovation in individual vehicle
noise management
Presentation to TEKH Acoustics 2019
Altissimo Consulting
Introduction
Contents
•Problem definition
•Solutions in Aus/NZ
•Analysis techniques
•Alternate hardware
•Classification
Individual vehicle noise
Annoyance
Engine
braking
Motorbikes
Sirens
Audio tactile
profile
Clangs and
bangs
Acceleration
•Complaints
•Investigations required
•Is the source actually what
people think?
•Limited ability to mitigate
Engine braking
•Supplementary braking system
that vents during the compression
cycle
•Not just using gears to slow down
•Less of an issue with modern trucks
•Culture of drivers
Current hardware
Installation details Features / issues
•High resolution camera with
Number Plate Recognition
•Uses “in-service engine braking
algorithm”
•Limited fixed cameras with high
relocation costs
•Long lead time to address
community concern
Current processing algorithm
RMS Modulation
•A-weighted
•5ms average
•200 Hz resample
•Band-pass filter (5-80Hz)
Variables
•RMS threshold level
•Bandpass filter
Accuracy
Sources of error
•False positives
•Birds
•Rain
•Wind
•Truck vs bike
•Unable to differentiate
Limitations
•Too many even ts to manually
view
•Commercial system restricts
this to samples with valid
number plate
Our solution -hardware
•Pole mounted
•Solar powered
•Low-power processor
(Raspberry Pi)
•Camera
•Remote access (3G)
altissimo.nz
Read from sensor
Analysis code
Standard OS functions
Interface with
Altissimo server
General concept
Problems
Power
•Solar panel capacity
•Charging circuits
•Reboot cycle corrupt disk
image
•Newer hardware has higher
power draw
Remote access / communications
•USB modems targeted at
consumers (Windows)
•Support for linux
•Bulk availability
•No remote ‘reboot’ facility
Hardware details
MPPT solar charger
Power management
Processor
Hologram IoT modems
•Full size 3B/3B+
•Pi Zero
•Balance between
processor power and
current draw
•Shuts down on low
battery voltage
•Watchdog to restart
Pi if software stops
running
Hardware -progress
Machine learning
Training dataset
Feature extraction
Train/test
Result
Machine learning –training dataset
Machine learning –feature extraction
•32 different analyses
•Spectral
•Energy
•Chroma
•Beats
Machine learning – test/train
Source: mapr.com
Machine learning -classification
Engine Braking classified as:
Engine
braking
Trucks
Motorbikes
Other
98%
1%
0%
1%
Motorbikes classified as:
Trucks classified as:
Engine
braking
Trucks
Motorbikes
Other
2%
6%
83%
9%
Engine
braking
Trucks
(accel/
decel)
Motorbikes
Trucks
–body slap
/ curbing
Other
3%
63%
4%
18%
11%
Results
Differences in analysis techniques
Analytical
•Come up with a metric
•Level
•Frequency
•Modulation
•Compare with criteria
Machine learning
•Define output categories
•Manually classify training set
(~2000 events)
•Train
•Test and tweak pa rameters
Conclusion
•Site selection is important
•Cheaper hardware allows
more units to be installed
•Ability to mount directly to
pole significantly reduces lead
time
•Machine learning can give
better outcomes than trying
to come up with analytical
solutions
Contacts
Michael Smith
michael@altissimo.nz
Rob Wareing
rob@altissimo.nz