Download toward Intelligent, Personal Air Quality Monitoring David B. Ramsay PDF

Titletoward Intelligent, Personal Air Quality Monitoring David B. Ramsay
LanguageEnglish
File Size50.6 MB
Total Pages215
Document Text Contents
Page 1

LearnAir: toward Intelligent, Personal Air Quality Monitoring

David B. Ramsay

BSEE and BA, Case Western Reserve University (2010)

Submitted to the Program in Media Arts and Sciences, School of Architecture and
Planning in partial fulfillment of the requirements for the degree of Master of Science at
the Massachusetts Institute of Technology

September 2016

©Massachusetts Institute of Technology 2016. All rights reserved.

Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
MIT Media Lab
August 6, 2016

Certified by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Joseph A. Paradiso

Professor of Media Arts and Sciences
Thesis Supervisor

Accepted by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Pattie Maes

Academic Head
Program in Media Arts and Sciences

Page 107

107

minimized quantization errors, smoothed data to match the EPA ref-
erence, or whose averaging gave evidence of longer term trends that
might also be important in analyzing sensor state and measurement
quality.

Figure 27: Temperature Derivative
Feature Creation

Several API’s were evaluated for use in this project, and Forecast.IO
stood out as a well-reputed option. They use machine learning
to combine many weather forecast APIs into one highly accurate
dataset. The Forecast.IO data comes in hourly intervals, so a running
60 minute average was used to interpolate the values to minute res-
olution (most of the measured fields, like temperature, are relatively
slow-moving). For class-based indicators (for instance, the ‘weather-
summary’ field indicating ‘cloudy’, ‘windy’, ‘foggy’, ‘rainy’, etc) we
converted the API value into binary fields that matched each class.

We created features such as ‘temperature differential’ and ‘humidity
differential’ – a constructed feature that corresponds to the differ-
ence in measurement between the ambient conditions (as measured
by Forecast.IO) and the conditions in the box (as measured by the
SmartCitizen Kit). While these features are linear combinations
of other features (and thus won’t improve our model’s predictive
power), they serve an intuitive purpose, and may help reduce the
feature set (mapping two features to one) if they turn out to be im-
portant indicators of performance.

Finally, we added some features to represent other potentially im-
portant quantities. These features include the time of day (including
features for morning and evening rush hours), the day of the year
(mapping to the season), and the elapsed time since the device was
plugged in (for ‘warming up’ effects).

Page 108

108

Figure 28: Temperature Inside vs.
Outside the Device during Test Period

Figure 29: Humidity Inside vs. Outside
the Device during Test Period

In general, some small subset of features was removed for each train-
ing session. For instance, the higher quality Alphasense CO sensor
was removed as a feature when training the less capable SmartC-
itizen CO sensor. (Training with this feature gives incredibly high
accuracy at predicting failure, because it is effectively training itself
with the answer.) By training with comparable or cheaper sensors,
we can assess the likelihood of a cheap system working predictably.

In most cases, the EPA reference black carbon sensor data was left
as a feature– while this is not a feasible measurement for a portable,
cheap device, it is still useful to know if black carbon is a strong
predictor of a sensor’s failure. This knowledge allows us to infer
something about how the sensor is failing, and what type of sensor
we might want to pair it with. With this technique, the machine
learning process is not evaluative of a current system’s success–

Page 214

214

[57] Giuseppe Anastasi, Paolo Brushci, et al. ‘u-sense’, a coopera-
tive sensing system for monitoring air quality in urban areas.
ERCIM News, 2014.

[58] Airbeam project. http://aircasting.org/, 2014.

[59] Elm project. http://elm.perkinelmer.com/map/, 2013.

[60] Breezometer project. https://breezometer.com/, 2015.

[61] Cleanspace project. https://our.clean.space/, 2016.

[62] Clarity pm2.5 sensor. http://clarity.io, 2014.

[63] Tzoa enviro-tracker. http://www.tzoa.com, 2014.

[64] uhoo indoor sensor. https://uhooair.com/s, 2016.

[65] Air quality egg project. http://airqualityegg.com/, 2012.

[66] Speck sensor. https://www.specksensor.com/, 2015.

[67] Atmotube. http://atmotube.com/, 2015.

[68] Airbox lab. http://foobot.io/, 2013.

[69] Citizenair.io. http://www.citoyenscapteurs.net/2014/03/31/
citizenair-io/, 2014.

[70] Public lab. https://publiclab.org/, 2010.

[71] Django framework. https://www.djangoproject.com/, 2005.

[72] Apache hadoop. http://hadoop.apache.org/, 2011.

[73] RDF Working Group. Rdf 1.1 schema, concepts and abstract
syntax. Technical report, World Wide Web Consortium, 2014.

[74] JSON-LD. http://json-ld.org/, 2010.

[75] HAL - hypertext application language. http://stateless.co/
hal_specification.html, 2013.

[76] Iotivity standard. https://www.iotivity.org/, 2015.

Page 215

215

[77] Alljoyn framework. https://allseenalliance.org/framework,
2011.

[78] Evan Lynch and Joseph Paradiso. Sensorchimes: Musical map-
ping for sensor networks. New Interfaces for Musical Expression,
2016.

[79] QianSheng Li, Gershon Dublon, Brian Mayton, and Joseph A.
Paradiso. Marshvis: Visualizing real-time and historical ecologi-
cal data from a wireless sensor network. IEEE VIS, 2015.

[80] Cheng Liu et al. A directional anemometer based on mems
differential pressure sensors. 9th International Conference on
Nano/Micro Engineered and Molecular Systems (NEMS), pages
517–520, 2014.

[81] Steven G. Brown, Paul T. Roberts, et al. Characterization of 2001
ozone event-triggered voc and carbonyl samples in houston.
Technical report, Sonoma Technology, Inc. Presented at TCEQ
State of the Science Meeting, 2002.

[82] Eduardo P. Olaguer. Near-source air quality impacts of large
olefin flares. Journal of the Air Waste Management Association,
62(8):978–988, 2012.

[83] C. Treviño and F. Méndez. Simplified model for the prediction
of ozone generation in polluted urban areas with continuous
precursor species emissions. Atmospheric Environment, 33(7):1103
– 1110, 1999.

Similer Documents